Datasets:

bzantium

/

LongPerplexity

like 0

maximum likelihood estimation is a common , well - studied and powerful technique for statistical estimation . in the context of a statistical model with an unknown parameter , the maximum likelihood estimate of the unknown parameter is , by definition , any parameter value under which the observed data is most likely ; such parameter values are said to maximize the likelihood function with respect to the observed data . in classical statistical models , one typically thinks of the unknown parameter as a real number or possibly a finite dimensional vector of real numbers . here we consider maximum likelihood estimation for statistical models in which each parameter value corresponds to a stochastic system observed with noise . hidden markov models ( hmms ) provide a natural setting in which to study both stochastic systems with observational noise and maximum likelihood estimation . in this setting , one has a parametrized family of stochastic processes that are assumed to be markov , and one attempts to perform inference about the underlying parameters from noisy observations of the process . there has been a substantial amount of work on statistical inference for hmms , and we do not attempt a complete survey of that area here . in the 1960s , baum and petrie @xcite studied consistency of maximum likelihood estimation for finite state hmms . since that time , several other authors have shown that maximum likelihood estimation is consistent for hmms under increasingly general conditions @xcite , culminating with the work of douc et al . @xcite , which currently provides the most general conditions on hmms under which maximum likelihood estimation has been shown to be consistent . we focus here on the consistency of maximum likelihood estimation for parametrized families of deterministic systems observed with noise . inference methods for deterministic systems from noisy observations are of interest in a variety of scientific areas ; for a few examples , see @xcite . for the purpose of this article , the terms deterministic system and dynamical system refer to a map @xmath0 . the set @xmath1 is referred to as the state space , and the transformation @xmath2 governs the evolution of states over one ( discrete ) time increment . our main interest here lies in families of dynamical systems observed with noise . more precisely , we consider a state space @xmath1 and a parameter space @xmath3 , and to each @xmath4 in @xmath3 , we associate a dynamical system @xmath5 . note that the state space @xmath1 does not depend on @xmath4 . for each @xmath4 in @xmath3 , we assume that the system is started at equilibrium from a @xmath6-invariant measure @xmath7 . see section [ sectionassumptions ] for precise definitions . we are particularly interested in situations in which the family of dynamical systems is observed via noisy measurements ( or observations ) . we consider a general observation model specified by a family of probability densities @xmath8 , where @xmath9 prescribes the distribution of an observation given that the state of the dynamical system is @xmath10 and the state of nature is @xmath4 . under some additional conditions ( see section [ generaltheorem ] ) , our first main result states that maximum likelihood estimation is a consistent method of estimation of the parameter @xmath4 . we have chosen to state the conditions of our main consistency result in terms of statistical properties of the family of dynamical systems and the observations . however , these particular statistical properties have not been directly studied in the dynamical systems literature . in the interest of applying our general result to specific systems , we also establish several connections between well - studied properties of dynamical systems and the statistical properties relevant to maximum likelihood estimation . finally , we apply these results to some examples , including shifts of finite type with gibbs measures and axiom a attractors with srb ( sinai ruelle bowen ) measures . it is widely accepted in the field of ergodic theory and dynamical systems that these classes of systems have `` good '' statistical properties , and our results may be viewed as a precise confirmation of this view . there has been a substantial amount of work on statistical inference for hmms , and a complete survey of that area is beyond the scope of this work . the asymptotic consistency of maximum likelihood estimation for hmms has been studied at least since the work of baum and petrie @xcite under the assumption that both the hidden state space @xmath1 and the observation space @xmath11 are finite sets . leroux extended this result to the setting where @xmath11 is a general space and @xmath1 is a finite set @xcite . several other authors have shown that maximum likelihood estimation is consistent for hmms under increasingly general conditions , culminating with the work of douc et al . @xcite , which currently provides the most general conditions for hmms under which maximum likelihood estimation has been shown to be consistent . let us now discuss the results of douc et al . @xcite in greater detail . consider parametrized families of hmms in which both the hidden state space @xmath1 and the observation space @xmath11 are complete , separable metric spaces . the main result of @xcite shows that under several conditions , maximum likelihood estimation is a consistent method of estimation of the unknown parameter . these conditions involve some requirements on the transition kernel of the hidden markov chain , as well as basic integrability conditions on the observations . the proof of that result relies on information - theoretic arguments , in combination with the application of some mixing conditions that follow from the assumptions on the transition kernel . to prove our consistency result , we take a similar information - theoretic approach , but instead of placing explicit restrictions on the transition kernel , we identify and study mixing conditions suitable for dynamical systems . see remarks [ comparisonrmk ] and [ stochasticremark ] for further discussion of our results in the context of hmms . other directions of study regarding inference for hmms include the behavior of mle for misspecified hmms @xcite , asymptotic normality for parameter estimates @xcite , the dynamics of bayesian updating @xcite and starting the hidden process away from equilibrium @xcite . extending these results to dynamical systems is of potential interest . the topic of statistical inference for dynamical systems has been widely studied in a variety of fields . early interest from the statistical point of view is reflected in the following surveys @xcite . for a recent review of this area with many references , see @xcite . there has been significant methodological work in the area of statistical inference for dynamical systems ( for a few recent examples , see @xcite ) , but in this section we attempt to describe some of the more theoretical work in this area . the relevant theoretical work to date falls ( very ) roughly into three classes : * state estimation ( also known as denoising or filtering ) for dynamical systems with observational noise ; * prediction for dynamical systems with observational noise ; * system reconstruction from dynamical systems without noise . let us now mention some representative works from these lines of research . in the setting of dynamical systems with observational noise , lalley introduced several ideas regarding state estimation in @xcite . these ideas were subsequently generalized and developed in @xcite . key results from this line of study include both positive and negative results on the consistency of denoising a dynamical system under additive observational noise . in short , the magnitude of the support of the noise seems to determine whether consistent denoising is possible . in related work , judd @xcite demonstrated that mle can fail ( in a particular sense ) in state estimation when noise is large . it is perhaps interesting to note that there are examples of axiom a systems with gaussian observational noise for which state estimation can not be consistent ( by results of @xcite ) and yet mle provides consistent parameter estimation ( by theorem [ axiomathm ] ) . steinwart and anghel considered the problem of consistency in prediction accuracy for dynamical systems with observational noise @xcite . they were able to show that support vector machines are consistent in terms of prediction accuracy under some conditions on the decay of correlations of the dynamical system . the work of adams and nobel uses ideas from regression to study reconstruction of measure - preserving dynamical systems @xcite without noise . these results show that certain types of inference are possible under fairly mild ergodicity assumptions . a sample result from this line of work is that a measure - preserving transformation may be consistently reconstructed from a typical trajectory observed without noise , assuming that the transformation preserves a measure that is absolutely continuous ( with radon nikodym derivative bounded away from @xmath12 and infinity ) with respect to a known reference measure . in section [ sectionassumptions ] , we give some necessary background on dynamical systems observed with noise . section [ generaltheorem ] contains a statement and discussion of our main result ( theorem [ mainthm ] ) , which asserts that under some general statistical conditions , maximum likelihood parameter estimation is consistent for of dynamical systems observed with noise . the purpose of section [ sectionverification ] is to establish connections between well - studied properties of dynamical systems and the ( statistical ) conditions appearing in theorem [ mainthm ] . section [ sectionexamples ] gives several examples of widely studied families of dynamical systems to which we apply theorem [ mainthm ] and therefore establish consistency of maximum likelihood estimation . the proofs of our main results appear in section [ sectionproofs ] , and we conclude with some final remarks in section [ sectionconclusion ] . recall that our primary objects of study are parametrized families of dynamical systems . in this section we introduce these objects in some detail . first let us recall some terminology regarding dynamical systems and ergodic theory . we use @xmath1 to denote a state space , which we assume to be a complete separable metric space endowed with its borel @xmath13-algebra @xmath14 . then a measurable dynamical system on @xmath1 is defined by a measurable map @xmath0 , which governs the evolution of states over one ( discrete ) time increment . for a probability measure @xmath15 on the measurable space @xmath16 , we say that @xmath2 preserves @xmath15 ( or @xmath15 is @xmath2-invariant ) if @xmath17 for each set @xmath18 in @xmath14 . we refer to the quadruple @xmath19 as a measure - preserving system . to generate a trajectory @xmath20 from such a measure - preserving system , one chooses @xmath21 according to @xmath15 and sets @xmath22 for @xmath23 . note that @xmath20 is then a stationary @xmath1-valued stochastic process . finally , the measure - preserving system @xmath24 is said to be ergodic if @xmath25 implies @xmath26 . see the books @xcite for an introduction to measure - preserving systems and ergodic theory . let us now introduce the setting of parametrized families of dynamical systems . we denote the parameter space by @xmath3 , which is assumed to be a compact metric space endowed with its borel @xmath13-algebra . fix a state space @xmath1 and its borel @xmath13-algebra @xmath14 as above . to each parameter @xmath4 in @xmath3 , we associate a measurable transformation @xmath5 , which prescribes the dynamics corresponding to the parameter @xmath4 . finally , we need to specify some initial conditions . in this article , we consider the case that the system is started from equilibrium . more precisely , we associate to each @xmath4 in @xmath3 a @xmath6-invariant borel probability measure @xmath7 on @xmath27 . thus , to each @xmath4 in @xmath3 , we associate a measure - preserving system @xmath28 , and we refer to the collection @xmath29 as a parametrized family of dynamical systems . for ease of notation , we will refer to @xmath30 as a family of dynamical systems on @xmath27 , instead of referring to the family of quadruples @xmath31 . we would like to study the situation that such a family of dynamical systems is observed via noisy measurements . here we describe the specifics of our observation model . we suppose that we have a complete , separable metric space @xmath11 , endowed with its @xmath13-algebra @xmath32 , which serves as our observation space . we also assume that we have a family of borel probability densities @xmath33 with respect to a fixed reference measure @xmath34 on @xmath11 . the density @xmath35 prescribes the distribution of our observation given that the state of the dynamical system is @xmath10 and the state of nature is @xmath36 . finally , we assume that the noise involved in successive observations is conditionally independent given @xmath4 and the underlying trajectory of the dynamical system . thus our full model consists of a parametrized family of dynamical systems @xmath37 on a measurable space @xmath27 with corresponding observation densities @xmath38 . in general , we would like to estimate the parameter @xmath4 from our observations . maximum likelihood estimation provides a basic method for performing such estimation . our first main result states that maximum likelihood estimation is a consistent estimator of @xmath4 under some general conditions on the family of systems and the noise . in order to state these results precisely , we now introduce the likelihood for our model . for the sake of notation , it will be convenient to denote finite sequences @xmath39 with the notation @xmath40 . as we have assumed that our observations are conditionally independent given @xmath4 and a trajectory @xmath20 , we have that for @xmath41 and @xmath42 , the likelihood of observing @xmath43 given @xmath4 and @xmath20 is @xmath44 since @xmath45 given @xmath4 and @xmath21 , the conditional likelihood of @xmath43 given @xmath4 and @xmath46 is @xmath47 since our model also assumes that @xmath21 is distributed according to @xmath48 , we have that for @xmath49 and @xmath50 , the marginal likelihood of observing @xmath43 given @xmath4 is @xmath51 we denote by @xmath52 the product measure on @xmath53 with marginals equal to @xmath34 . let @xmath54 be the probability measure on @xmath55 such that for borel sets @xmath56 and @xmath57 , it holds that @xmath58 which is well defined by kolmogorov s consistency theorem . let @xmath59 denote expectation with respect to @xmath54 , and let @xmath60 be the marginal of @xmath54 on @xmath61 . before we define consistency , let us first consider the issue of identifiability . our notion of identifiability is captured by the following equivalence relation . define an equivalence relation on @xmath3 as follows : let @xmath62 if @xmath63 . denote by @xmath64 $ ] the equivalence class of @xmath4 with respect to this equivalence relation . in a strong theoretical sense , if @xmath65 is in @xmath66 $ ] , then the systems corresponding to the parameter values @xmath65 and @xmath4 can not be distinguished from each other based on observations of the system . now we fix a distinguished element @xmath67 in @xmath3 . here and in the rest of the paper , we assume that @xmath67 is the `` true '' parameter ; that is , the data are generated from the measure @xmath68 . hence , one may think of @xmath69 $ ] as the set of parameters that can not be distinguished from the true parameter . an approximate maximum likelihood estimator ( mle ) is a sequence of measurable functions @xmath70 such that @xmath71 where @xmath72 denotes a process that tends to zero @xmath73-a.s . as @xmath74 tends to infinity several notions in this article , including the definition of approximate mle above , involve taking suprema over @xmath4 in @xmath75 . in many situations of interest to us , @xmath1 and @xmath3 are compact , and all relevant functions are continuous in these arguments . in such cases , we have sufficient regularity to guarantee that suprema over @xmath4 in @xmath3 are measurable . however , in the general situation , such suprema are not guaranteed to be measurable , and one must take some care . as all our measurable spaces are polish ( complete , separable metric spaces ) ; such functions are always universally measurable @xcite , proposition 7.47 . similarly , a borel - measurable ( approximate ) maximum likelihood estimator need not exist , but the polish assumption ensures the existence of universally measurable maximum likelihood estimators @xcite , proposition 7.50 . thus all probabilities and expectations may be unambiguously extended to such quantities . in this work , we do not consider specific schemes for constructing an approximate mle . based on the existing results regarding denoising and system reconstruction ( e.g. , @xcite , which are briefly discussed in section [ sectionpriorwork ] ) , explicit construction of an approximate mle may be possible under suitable conditions . although the description and study of such constructive methods could be interesting , it is outside of the scope of this work . in principle , one could consider inference based on the conditional likelihood @xmath76 in place of the marginal likelihood @xmath77 . however , we do not pursue this direction in this work . for nonlinear dynamical systems , even the conditional likelihood @xmath76 may depend very sensitively on @xmath78 ; see @xcite , for example . thus optimizing over @xmath78 is essentially no more `` tractable '' than marginalizing the likelihood via an invariant measure . [ comparisonrmk ] the framework of this paper may be translated into the language of markov chains as follows . for each @xmath49 , we define a ( degenerate ) markov transition kernel @xmath79 as follows : @xmath80 in other words , for each @xmath49 , @xmath81 , and borel set @xmath56 , the probability that @xmath82 conditioned on @xmath83 is @xmath84 where @xmath85 is defined to be a point mass at @xmath10 . in all previous work on consistency of maximum likelihood estimation for hmms ( including @xcite ) , there have been significant assumptions placed on the markovian structure of the hidden chain . for example , the central hypothesis appearing in @xcite requires that there is a @xmath13-finite measure @xmath86 on @xmath1 such that for some @xmath87 , the @xmath88-step transition kernel @xmath89 is absolutely continuous with respect to @xmath86 with bounded radon nikodym derivative . if @xmath1 is uncountable , then the degeneracy of @xmath79 , which arises directly from the fact that we are considering deterministic systems , makes the existence of such a dominating measure impossible . in short , it is precisely the determinism in our hidden processes that prevents previous theorems for hmms from applying to dynamical systems . nonetheless , there is a special case of systems that we consider in section [ sectiongibbs ] that overlaps with the systems considered in the hmm literature . if @xmath1 is a shift of finite type , @xmath6 is the shift map @xmath90 for all @xmath4 , @xmath7 is a ( @xmath91-step ) markov measure for all @xmath4 , and @xmath35 depends only @xmath4 and the zero coordinate @xmath78 , then both the present work and the results in @xcite apply to this setting and guarantee consistency of any approximate mle under additional assumptions on the noise . in this section , we show that under suitable conditions , any approximate mle is consistent for families of dynamical systems observed with noise . to make this statement precise , we make the following definition of consistency . an approximate mle @xmath92 is consistent at @xmath67 if @xmath93 converges to @xmath69 $ ] , @xmath94-a.s . as @xmath74 tends to infinity . for the sake of notation , define the function @xmath95 , where @xmath96 also , for @xmath97 , let @xmath98 . consider the following conditions on a family of dynamical systems observed with noise : _ ergodicity_. the system @xmath99 on @xmath100 is ergodic . _ logarithmic integrability at _ @xmath67 . it holds that @xmath101 < \infty\ ] ] and @xmath102 < \infty.\ ] ] _ logarithmic integrability away from @xmath67 . _ for each @xmath103 $ ] , there exists a neighborhood @xmath104 of @xmath65 such that @xmath105 < \infty.\ ] ] _ upper semi - continuity of the likelihood_. for each @xmath103 $ ] and @xmath106 , the function @xmath107 is upper semi - continuous at @xmath65 , @xmath94-a.s . _ mixing condition_. there exists @xmath108 such that for each @xmath109 , there exists a measurable function @xmath110 such that if @xmath111 and @xmath112 , then @xmath113 furthermore , for each @xmath103 $ ] , there exists a neighborhood @xmath104 of @xmath65 such that @xmath114 < \infty.\ ] ] _ exponential identifiability_. for each @xmath115 $ ] , there exists a sequence of measurable sets @xmath116 such that @xmath117 the following theorem is our main general result . [ mainthm ] suppose that @xmath30 is a parametrized family of dynamical systems on @xmath27 with corresponding observation densities @xmath118 . if conditions hold , then any approximate mle is consistent at @xmath67 . the proof of theorem [ mainthm ] is given in section [ sectionproofs ] . in the following remark , we discuss conditions ( s1)(s6 ) . [ conditionsrmk ] conditions ( s1)(s3 ) involve basic irreducibility and integrability conditions , and similar conditions have appeared in previous work on consistency of maximum likelihood estimation for hmms ; see , for example , . taken together , conditions ( s1 ) and ( s2 ) ensure the almost sure existence and finiteness of the entropy rate for the process @xmath119 , @xmath120 condition ( s3 ) serves as a basic integrability condition in the proof of theorem [ mainthm ] , in which one must essentially show that for @xmath115 $ ] , @xmath121 conditions ( s4)(s6 ) are more interesting from the point of view of dynamical systems , and we discuss them in greater detail below . the upper semi - continuity of the likelihood ( s4 ) is closely related to the continuity of the map @xmath122 . in general , the continuous dependence of @xmath7 on @xmath4 places nontrivial restrictions on a family of dynamical systems . this property ( continuity of @xmath122 ) is often called `` statistical stability '' in the dynamical systems and ergodic theory literature , and it has been studied for some families of systems ; for example , see @xcite and references therein . in section [ statstability ] , we show how statistical stability of the family of dynamical systems may be used to establish the upper semi - continuity of the likelihood ( s4 ) . the mixing condition ( s5 ) involves control of the correlations of the observation densities along trajectories of the underlying dynamical system . although the general topic of decay of correlations has been widely studied in dynamical systems ( see @xcite for an overview ) , condition ( s5 ) is not implied by the particular decay of correlations properties that are typically studied for dynamical systems . nonetheless , we show in section [ sectionsubadd ] how some well - studied mixing properties of dynamical systems imply the mixing condition ( s5 ) . finally , condition ( s6 ) involves the exponential identifiability of the true parameter @xmath67 . we show in section [ sectionidentifiability ] how large deviations for a family of dynamical systems may be used to establish exponential identifiability ( s6 ) . large deviations estimates for dynamical systems have been studied in @xcite , and our main goal in section [ sectionidentifiability ] is to connect such results to exponential identifiability ( s6 ) . [ stochasticremark ] suppose one has a family of bi - variate stochastic processes @xmath123 , where @xmath124 is interpreted as a hidden process and @xmath125 as an observation process . if the observations have conditional densities with respect to a common measure given @xmath124 and @xmath4 , then it makes sense to ask whether maximum likelihood estimation is a consistent method of inference for the parameter @xmath4 . it is well known that the setting of stationary stochastic processes may be translated into the deterministic setting of dynamical systems , which may be carried out as follows . let @xmath126 be a family of stationary stochastic processes on a measurable space @xmath27 . consider the product space @xmath127 with corresponding @xmath13-algebra @xmath128 . each process @xmath124 corresponds to a probability measure @xmath48 on @xmath129 with the property that @xmath7 is invariant under the left - shift map @xmath130 given by @xmath131 . with this translation , theorem [ mainthm ] shows that maximum likelihood estimation is consistent for families of hidden stochastic processes @xmath124 observed with noise , whenever the corresponding family of dynamical systems @xmath132 on @xmath133 with observation densities satisfy conditions ( s1)(s6 ) . with the above translation , theorem [ mainthm ] applies to some families of processes allowing infinite - range dependence in both the hidden process @xmath124 and the observation process @xmath125 . from this point of view , theorem [ mainthm ] highlights the fact that maximum likelihood estimation is consistent for dependent processes observed with noise as long as they satisfy some general conditions : ergodicity , logarithmic integrability of observations , continuous dependence on the parameters and some mixing of the observation process . it is interesting to note that the existing work on consistency of maximum likelihood estimation for hmms @xcite makes assumptions of precisely this sort in the specific context of markov chains . in our main consistency result ( theorem [ mainthm ] ) , we establish the consistency of any approximate mle under conditions ( s1)(s6 ) . we have chosen to formulate our result in these terms because they reflect general statistical properties of dynamical systems observed with noise that are relevant to parameter inference . however , these conditions have not been explicitly studied in the dynamical systems literature , despite the fact that much effort has been devoted to understanding certain statistical aspects of dynamical systems . in this section , we make connections between the general statistical conditions appearing in theorem [ mainthm ] and some well - studied properties of dynamical systems . section [ statstability ] shows how the notion of statistical stability may be used to verify the upper semi - continuity of the likelihood ( s4 ) . section [ sectionsubadd ] connects well - known mixing properties of some measure - preserving dynamical systems to the mixing property ( s5 ) . in section [ sectionidentifiability ] , we show how large deviations for dynamical systems may be used to deduce the exponential identifiability condition ( s6 ) . proofs of statements in this section , as well as additional discussion , appear in supplementary appendix a @xcite . as discussed in remark [ conditionsrmk ] , the upper semi - continuity condition ( s4 ) places nontrivial restrictions on the family of dynamical systems under consideration . in this section , we establish sufficient conditions for ( s4 ) to hold . the continuous dependence of @xmath7 on @xmath4 is a property called statistical stability in the dynamical systems literature @xcite . let us state this property precisely . let @xmath134 denote the space of borel probability measures on @xmath1 . endow @xmath134 with the topology of weak convergence : @xmath135 converges to @xmath15 if @xmath136 converges to @xmath137 as @xmath74 tends to infinity , for each continuous , bounded function @xmath138 . the family of dynamical systems @xmath139 on @xmath27 is said to have statistical stability if the map @xmath122 is continuous with respect to the weak topology on @xmath134 . the following proposition shows that under some continuity and compactness assumptions , statistical stability of the family of dynamical systems implies upper semi - continuity of the likelihood ( s4 ) . [ usclemma2 ] suppose that @xmath1 and @xmath3 are compact , and the maps @xmath140 and @xmath141 are continuous . if the family @xmath30 has statistical stability , then upper semi - continuity of the likelihood holds . the proof of proposition [ usclemma2 ] appears in supplementary appendix a.1 @xcite . in this section , we focus on mixing condition ( s5 ) . recall that ( s5 ) involves a nontrivial restriction on the correlations of the observation densities @xmath142 along trajectories of the underlying dynamical system . although mixing conditions have been widely studied in the dynamics literature , the particular type of condition appearing in ( s5 ) appears not to have been investigated . nonetheless , we show that a well - studied mixing property for dynamical systems implies the statistical mixing property ( s5 ) . in order to study mixing for dynamical systems , one typically places restrictions on the type of events or observations that one considers ( by considering certain functionals of the process ) . for example , in some situations a substantial amount work has been devoted to finding particular partitions of state space with respect to which the system possess good mixing properties ; an example of such partitions are the well - known markov partitions @xcite . if a system has good mixing properties with respect to a particular partition , and if that partition possesses certain ( topological ) regularity properties , then it is often possible to show that the system also has good mixing properties for related function classes , such as lipschitz or hlder continuous observables . for variations of this approach to mixing in dynamical systems , see the vast literature on decay of correlations ; for an introduction , see the survey @xcite . in this section , we follow the above approach to study mixing condition ( s5 ) for dynamical systems observed with noise . first , we define a mixing property for families of dynamical systems with respect to a partition ( m1 ) . second , we define a regularity property for partitions ( m2 ) . third , we define a topological regularity property for a family of observation densities ( m3 ) . finally , in the main result of this section ( proposition [ subaddpropmain ] ) , we show how these three properties together imply the mixing condition ( s5 ) . here and in the rest of this section , we consider only invertible transformations . it is certainly possible to modify the definitions slightly to handle the noninvertible case , but we omit such modifications . we will have need to consider finite partitions of @xmath1 . the join of two partitions @xmath143 and @xmath144 is defined to be the common refinement of @xmath143 and @xmath144 , and it is denoted @xmath145 . note that for any measurable transformation @xmath0 , if @xmath146 is a partition , then so is @xmath147 . for a fixed partition @xmath146 and @xmath148 , let @xmath149 . notice that @xmath150 depends on @xmath4 through @xmath6 , although we suppress this dependence in our notation . now consider the following alternative conditions , which may be used in place of condition ( s5 ) : _ mixing condition with respect to the partition _ @xmath146 . there exists @xmath151 and @xmath108 such that for all @xmath49 , @xmath152 , @xmath153 and @xmath154 , it holds that @xmath155 furthermore , for each @xmath103 $ ] there exists a neighborhood @xmath104 of @xmath65 such that @xmath156 _ regularity of the partition _ there exists @xmath157 such that for all @xmath49 and @xmath152 , if @xmath158 and @xmath159 , then @xmath160 _ regularity of observations_. there exists a function @xmath161 such that for @xmath162 and @xmath163 , @xmath164 furthermore , for each @xmath103 $ ] , there exists a neighborhood @xmath104 of @xmath65 such that @xmath165 < \infty.\ ] ] let us now state the main proposition of this section , whose proof is deferred to supplementary appendix a.2 @xcite . [ subaddpropmain ] suppose @xmath30 is a family of dynamical systems on @xmath27 with corresponding observation densities @xmath118 . if there exists a partition @xmath146 of @xmath1 such that conditions and are satisfied , and if the observation regularity condition is satisfied , then mixing property holds . in this section , we study exponential identifiability condition ( s6 ) . we show how large deviations for dynamical systems may be used in combination with some regularity of the observation densities to establish exponential identifiability ( s6 ) . let @xmath166 and @xmath167 be metric spaces with metrics @xmath168 and @xmath169 , respectively . recall that a function @xmath170 is said to be hlder continuous if there exist @xmath171 and @xmath172 such that for each @xmath173 in @xmath166 , it holds that @xmath174 if @xmath175 is a dynamical system on @xmath27 such that @xmath176 is hlder continuous , then we refer to @xmath175 as a hlder continuous dynamical system . for many dynamical systems , the class of hlder continuous functions @xmath177 provides a natural class of observables whose statistical properties are fairly well understood and satisfy some large deviations estimates @xcite . consider the following conditions , which we later show are sufficient to guarantee exponential identifiability ( s6 ) : _ large deviations . _ for each @xmath115 $ ] , for each hlder continuous function @xmath177 , and for each @xmath178 , it holds that @xmath179 _ regularity of observations . _ [ holdercontinuitycondition ] there exists @xmath171 and @xmath161 such that for each @xmath10 and @xmath180 in @xmath1 , it holds that @xmath181 furthermore , for @xmath49 and @xmath172 , it holds that @xmath182 the following proposition relates large deviations for dynamical systems to the exponential identifiability condition ( s6 ) . [ identifiabilityfromlargedeviations ] suppose that @xmath30 is a family of hlder continuous dynamical systems on the @xmath27 with corresponding observation densities @xmath183 . further suppose that the large deviations property and the observation regularity property are satisfied . then the exponential identifiability condition holds . the proof of proposition [ identifiabilityfromlargedeviations ] appears in supplementary appendix a.3 @xcite . in this section we present some classical families of dynamical systems for which maximum likelihood estimation is consistent . we begin in section [ sectiongibbs ] by considering symbolic dynamical systems called shifts of finite type . the state space for such systems consists of ( bi-)infinite sequences of symbols from a finite set , and the transformation on the state space is always given by the `` left - shift '' map , which just shifts each point one coordinate to the left . such systems are considered models of `` chaotic '' dynamical systems that may be defined by a finite amount of combinatorial information . in this setting gibbs measures form a natural class of invariant measures , which have been studied due to their connections to statistical physics . these measures play a central role in a topic called the _ thermodynamic formalism _ , which is well described in the books @xcite . note that @xmath184th order finite state markov chains form a special case of gibbs measures . the main result of this section is theorem [ gibbsfamiliesthm ] , which states that under sufficient regularity conditions , any approximate maximum likelihood estimator is consistent for families of gibbs measures on a shift of finite type . the crucial assumptions for this theorem involve continuous dependence of the gibbs measures on @xmath4 and sufficiently regular dependence of @xmath185 on @xmath10 . additional proofs and discussion for this section appear in the supplementary appendix b @xcite . having established consistency of maximum likelihood estimation for families of gibbs measures on a shift of finite type , we deduce in section [ sectionfactors ] that maximum likelihood estimation is consistent for families of axiom a attractors observed with noise . axiom a systems are well studied differentiable dynamical systems on manifolds that , like shifts of finite type , exhibit `` chaotic '' behavior ; for a thorough treatment of axiom a systems , see the book @xcite . in related statistical work , lalley @xcite considered the problem of denoising the trajectories of axiom a systems . for these systems , there is a natural class of measures , known as srb ( sinai ruelle bowen ) measures . see the article @xcite for an introduction to these measures with discussion of their interpretation and importance . with the construction of markov partitions @xcite , one may view an axiom a attractor with its srb measure as a factor of a shift of finite type with a gibbs measure . using this natural factor structure , we establish the consistency of any approximate maximum likelihood estimator for axiom a systems . proofs and discussion of these topics appear in the supplementary appendix c @xcite . in this section , we consider the setting of symbolic dynamics , shifts of finite type and gibbs measures . we prove that any approximate maximum likelihood estimator is consistent for these systems ( theorem [ gibbsfamiliesthm ] ) under some general assumptions on the observations . finally , we consider two examples of observations in greater detail . in the first example , we consider `` discrete '' observations , corresponding to a `` noisy channel . '' in the second example , we consider making real - valued observations with gaussian observational noise . for a brief introduction to shifts of finite type and gibbs measures that contains everything needed in this work , see the supplementary appendix b @xcite . for a complete introduction to shifts of finite type and gibbs measures , see @xcite . let us now consider some families of measure - preserving systems on sfts . let @xmath186 be an alphabet , and let @xmath187 be a binary matrix with dimensions @xmath188 . let @xmath189 be the associated sft , and let @xmath14 be the borel @xmath13-algebra on @xmath1 . for @xmath190 , let @xmath191 be a continuous map , and let @xmath48 be the gibbs measure associated to the potential function @xmath192 . in this setting , we refer to @xmath193 as a continuously parametrized family of gibbs measures on @xmath194 . [ gibbsfamiliesthm ] suppose @xmath189 is a mixing shift of finite type and @xmath195 is a continuously parametrized family of gibbs measures on @xmath27 . if the family of observation densities @xmath118 satisfies the integrability conditions and and the regularity conditions and , then any approximate maximum likelihood estimator is consistent . the proof of theorem [ gibbsfamiliesthm ] is based on an appeal to theorem [ mainthm ] . however , in order to verify the hypotheses of theorem [ mainthm ] , we combine the results of section [ sectionverification ] with some well - known properties of gibbs measures . this proof appears in the supplementary appendix b @xcite . [ onesidedrmk ] there is an analogous theory of `` one - sided '' symbolic dynamics and gibbs measures , in which @xmath196 is replaced by @xmath197 and appropriate modifications are made in the definitions . the two - sided case deals with invertible dynamical systems , whereas the one - sided case handles noninvertible systems . we have stated theorem [ gibbsfamiliesthm ] in the invertible setting , although it applies as well in the noninvertible setting , with the obvious modifications . in this example , we consider families of dynamical systems @xmath198 on @xmath27 , where @xmath1 is a mixing shift of finite type , @xmath199 , and @xmath7 is a continuous family of gibbs measures on @xmath1 ( as in theorem [ gibbsfamiliesthm ] ) . here we consider the particular observation model in which our observations of @xmath1 are passed through a discrete , memoryless , noisy channel . suppose that @xmath11 is a finite set , @xmath34 is counting measure on @xmath11 and for each symbol @xmath200 in @xmath186 and parameter @xmath4 in @xmath3 , we have a probability distribution @xmath201 on @xmath11 . we consider the case that our observation densities @xmath142 satisfy @xmath202 . this situation is covered by theorem [ gibbsfamiliesthm ] , since the following conditions may be easily verified : observation integrability ( s2 ) and ( s3 ) and observation regularity ( m3 ) and ( l2 ) . [ gibbsexample ] in this example , we once again consider families of dynamical systems @xmath203 on @xmath27 , such that @xmath204 is a mixing shift of finite type , @xmath199 and @xmath48 is a continuous family of gibbs measures on @xmath1 ( as in theorem [ gibbsfamiliesthm ] ) . here we consider the particular observation model in which we make real - valued , parameter - dependent measurements of the system , which are corrupted by gaussian noise with parameter - dependent variance . more precisely , let us assume that @xmath205 , and there exists a lipschitz continuous @xmath206 and continuous @xmath207 such that @xmath208 we now proceed to verify conditions ( s2 ) , ( s3 ) , ( m3 ) and ( l2 ) . first , by compactness and continuity , there exist @xmath209 such that for @xmath4 in @xmath3 , @xmath210 in @xmath211 and @xmath10 in @xmath1 , it holds that @xmath212 from ( [ noiseinteqn ] ) , one easily obtains the observation integrability conditions ( s2 ) and ( s3 ) . furthermore , there exists @xmath213 such that for @xmath163 , it holds that @xmath214 \biggr ) \nonumber \\[-8pt ] \\[-8pt ] \nonumber & & \qquad = \exp\biggl ( -\frac{1}{2 s(\theta)^2 } \bigl [ \bigl(\varphi_{\theta } ( x ) - \varphi_{\theta}(z)\bigr ) \bigl(\varphi_{\theta}(x ) + \varphi_{\theta}(z)\bigr ) + 2 y \bigl(\varphi_{\theta}(z ) - \varphi_{\theta}(x)\bigr)\bigr ] \biggr ) \\ & & \qquad \leq\exp\bigl ( \bigl(c_4 + c_5|y|\bigr)\bigl|\varphi_{\theta}(x)- \varphi_{\theta } ( z)\bigr| \bigr ) . \nonumber \ ] ] let @xmath215 be lipschitz continuous with constant @xmath216 , and let @xmath217 . with this choice of @xmath218 and ( [ noiseregeqn ] ) , one may easily verify the observation regularity conditions ( m3 ) and ( l2 ) . similar calculations to those in example [ gibbsexample ] imply that any approximate maximum likelihood estimator is also consistent if the observational noise is `` double - exponential '' [ i.e. , @xmath219 . indeed , these calculations should hold for most members of the exponential family , although we do not pursue them here . in this section , we show how the previous results may be applied to some smooth ( differentiable ) families of dynamical systems . these results follow easily from the results in section [ sectiongibbs ] , using the work of bowen and others ( see @xcite and references therein ) in constructing markov partitions for these systems . with markov partitions , axiom a systems may be viewed as factors of the shifts of finite type with gibbs measures . for a brief introduction of axiom a systems that contains the details necessary for this work , see the supplementary appendix c @xcite . the basic fact that allows us to transfer our results from shifts of finite type to axiom a systems is that consistency of maximum likelihood estimation is preserved under taking appropriate factors . let us now make this statement precisely . suppose that @xmath37 is a family of dynamical systems on @xmath194 with observation densities @xmath220 . further , suppose that there are continuous maps @xmath221 and @xmath222 such that : for each @xmath4 , we have that @xmath223 ; for each @xmath4 , there is a unique probability measure @xmath224 on @xmath225 such that @xmath226 ; for each @xmath4 , the map @xmath227 is injective @xmath224-a.s . for @xmath10 in @xmath225 and @xmath4 in @xmath3 , define @xmath228 . then @xmath229 is a family of dynamical systems on @xmath230 with observation densities @xmath231 . in this situation , we say that @xmath232 is an isomorphic factor of @xmath233 , and @xmath234 is the factor map . the following proposition addresses the consistency of maximum likelihood estimation for isomorphic factors . its proof is straightforward and omitted . [ factorprop ] suppose that @xmath235 is an isomorphic factor of @xmath236 . then maximum likelihood estimation is consistent for @xmath237 if and only if maximum likelihood estimation is consistent for @xmath238 . for the sake of brevity , we defer precise definitions for axiom a systems to supplementary appendix c @xcite . we consider families of axiom a systems as follows . suppose that @xmath239 is a parametrized family of diffeomorphisms such that : @xmath240 is hlder continuous ; there exists @xmath171 such that for each @xmath4 , the map @xmath192 is @xmath241 ; for each @xmath4 , @xmath242 is an axiom a attractor and the restriction @xmath243 is topologically mixing ; for each @xmath4 , the measure @xmath7 is the unique srb measure corresponding to @xmath192 @xcite , theorem 4.1 . if these conditions are satisfied , then we say that @xmath244 is a parametrized family of axiom a systems on @xmath27 . [ axiomathm ] suppose that @xmath245 is a parametrized family of axiom a systems on @xmath27 . further , suppose that @xmath118 is a family of observations densities satisfying the following conditions : observation integrability and and observation regularity and . then maximum likelihood estimation is consistent . the proof of theorem [ axiomathm ] appears in the supplementary appendix c @xcite . propositions [ ergprop][strictineqprop ] are used in the proof of theorem [ mainthm ] , which is given at the end of the present section . [ ergprop ] suppose that condition ( ergodicity ) holds . then the process @xmath246 is ergodic under @xmath68 . let @xmath247 be arbitrary , and let @xmath186 and @xmath248 be borel subsets of @xmath249 . to obtain the ergodicity of @xmath250 , it suffices to show that ( see @xcite ) @xmath251 for @xmath81 , define @xmath252 and define @xmath253 similarly . for @xmath254 , by the conditional independence of @xmath255 and @xmath256 given @xmath67 and @xmath257 , we have that @xmath258 where we have used fubini s theorem . since @xmath259 is fixed , we have that @xmath260 since @xmath99 is ergodic , an alternative characterization of ergodicity ( see @xcite ) gives that @xmath261 thus we have verified equation ( [ eqn : ergver ] ) , and the proof is complete . for the following propositions , recall our notation that @xmath262 [ gsmb ] suppose that conditions and hold . then there exists @xmath263 such that @xmath264 moreover , the following equality holds @xmath94-a.s . : @xmath120 the proposition is a direct application of barron s generalized shannon breiman theorem @xcite . here we simply check that the hypotheses of that theorem hold in our setting . since condition ( s1 ) ( ergodicity ) holds , proposition [ ergprop ] gives @xmath246 is stationary and ergodic under @xmath73 . by definition , @xmath265 has density @xmath266 with respect to the @xmath13-finite measure @xmath267 . the measure @xmath52 is a product of the measure @xmath34 taken @xmath268 times . as such , the sequence @xmath269 clearly satisfies barron s condition that this sequence is `` markov with stationary transitions . '' define @xmath270 . let us show that for @xmath271 , we have that @xmath272 which clearly implies that @xmath273 . once ( [ finiteentropy ] ) is established , we will have verified all of the hypotheses of barron s generalized shannon mcmillan breiman theorem , and the proof of the proposition will be complete . observe that the first part of the integrability condition ( s2 ) gives that @xmath274 \leq(n+1 ) \e_{\theta_0 } \bigl [ \log^+ \gamma_{\theta_0}(y_0 ) \bigr ] < \infty.\ ] ] then the second part of the integrability condition ( s2 ) implies that @xmath275 & = & \e_{\theta _ 0 } \biggl [ \log\frac { p_{\theta_0}(y_0^n ) } { \prod_{k=0}^n \int g_{\theta_0}(y_k \mid x ) \,d\mu_{\theta_0}(x ) } \biggr ] \nonumber\\ & & { } + \e_{\theta_0 } \biggl [ \sum_{k=0}^n \log\int g_{\theta _ 0}(y_k \mid x ) \,d\mu_{\theta_0}(x ) \biggr ] \nonumber \\[-8pt ] \\[-8pt ] \nonumber & \geq&-(n+1 ) \e_{\theta_0 } \biggl [ \biggl| \log\int g_{\theta_0}(y_0 \mid x ) \,d\mu_{\theta_0}(x ) \biggr| \biggr ] \\ & > & -\infty , \nonumber \ ] ] where we have used that relative entropy is nonnegative . by ( [ eqn : pospartfinite ] ) and ( [ eqn : negpartfinite ] ) , we conclude that ( [ finiteentropy ] ) holds , which completes the proof . the following proposition is used in the proof of theorem [ mainthm ] to given an almost sure bound for the normalized log - likelihoods in terms of quantities involving only expectations . [ basicestimate ] suppose that conditions , and hold . let @xmath276 be as in condition . then for @xmath103 $ ] , there exists a neighborhood @xmath104 of @xmath65 such that for each @xmath277 , the following inequality holds @xmath94-a.s . : @xmath278 informally , in the proof of proposition [ basicestimate ] , we use the mixing property from condition ( s5 ) to parse a sequence of observations into alternating sequences of `` large blocks '' and `` small blocks , '' and then the ergodicity and integrability conditions finish the proof . more specifically , we break up the sequence of observations @xmath265 into alternating blocks of length @xmath259 and @xmath276 , where @xmath276 is given by condition ( s5 ) . let @xmath103 $ ] . fix a neighborhood @xmath104 of @xmath65 so that the conclusions of both condition ( s3 ) and condition ( s5 ) hold . let @xmath247 be arbitrary , and let @xmath276 be as in condition ( s5 ) . we consider sequences of observations of length @xmath74 , where @xmath74 is a large integer . these sequences of observations will be parsed into alternating blocks of lengths @xmath259 and @xmath276 , respectively , starting from an offset of size @xmath279 and possibly ending with a remainder sequence . for the sake of notation , we use interval notation to denote intervals of integers . for @xmath280 and @xmath279 in @xmath281 , let @xmath282 and @xmath283 be defined by the condition @xmath284 . then we partition @xmath285 $ ] as follows : @xmath286.\end{aligned}\ ] ] given a sequence @xmath265 of observations , we define the following subsequences of @xmath265 according to the above partitions of @xmath285 $ ] : @xmath287 for a sequence @xmath288 in @xmath289 , define @xmath290 then for @xmath4 in @xmath104 , it follows from condition ( s5 ) that @xmath291 taking the logarithm of both sides and averaging over @xmath279 in @xmath292 , we obtain @xmath293\nonumber \\ & & { } + \frac{1}{m+\ell } \sum_{s=0}^{m+\ell-1 } \sum _ { j=1}^k \log\gamma_{\theta } \bigl(v_s(j)\bigr ) \\ & & { } + \frac{1}{m+\ell } \sum_{s=0}^{m+\ell-1 } \bigl [ \log\gamma_{\theta } ( b_s ) + \log\gamma_{\theta}(e_s ) \bigr ] . \nonumber \ ] ] let us now take the supremum over @xmath4 in @xmath104 in ( [ eqn : threepieces ] ) and evaluate the limits of the three terms on the right - hand side as @xmath74 tends to infinity . let @xmath294 and @xmath295 be defined by @xmath296 with this notation , we have that @xmath297 \\ & & \qquad = \frac{1}{n } \sum_{i=0}^n \bigl [ \xi_1\bigl(y_i^{i+m}\bigr ) + \xi _ 2\bigl(y_i^{i+m}\bigr ) \bigr].\end{aligned}\ ] ] since @xmath246 is ergodic ( by proposition [ ergprop ] ) , it follows from birkhoff s ergodic theorem and conditions ( s3 ) and ( s5 ) that the following limit exists @xmath298-a.s . : @xmath299 \nonumber\\ & & \qquad = \lim_n \frac{1}{n } \sum _ { i=0}^n \bigl [ \xi_1 \bigl(y_i^{i+m}\bigr ) + \xi_2 \bigl(y_i^{i+m}\bigr ) \bigr ] \nonumber \\ & & \qquad= \e_{\theta_0 } \bigl [ \xi_1\bigl(y_0^m \bigr ) \bigr ] + \e_{\theta_0 } \bigl [ \xi_2\bigl(y_0^m \bigr ) \bigr ] \\ & & \qquad = \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log p_{\theta } \bigl(y_0^m\bigr ) \bigr]\nonumber\\ & & \qquad\quad { } + \e_{\theta_0 } \bigl [ \sup _ { \theta\in u } \log c_{m}\bigl(\theta , y_0^m \bigr ) \bigr ] . \nonumber\end{aligned}\ ] ] similarly , using birkhoff s ergodic theorem and condition ( s3 ) , we have that the following holds @xmath94-a.s . : \nonumber \ ] ] finally , birkhoff s ergodic theorem and condition ( s3 ) again imply that the following limit holds @xmath298-a.s . : @xmath301 = 0,\ ] ] where we have used that @xmath302 . combining the inequalities in ( [ eqn : threepieces])([term3 ] ) , we obtain that @xmath303 \\ & & { } + \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log c_{m}\bigl(\theta , y_0^m\bigr ) \bigr ] \\ & & { } + \frac{\ell}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log^+ \gamma_{\theta}(y_0 ) \bigr],\end{aligned}\ ] ] as desired . the following proposition is a direct application of lemma 10 in @xcite to the present setting , and we omit the proof . [ identifiabilityprop ] suppose that the following conditions hold : ergodicity , logarithmic integrability at @xmath67 and exponential identifiability . then for @xmath115 $ ] , it holds that @xmath304 < h ( \theta_0).\ ] ] the following proposition provides an essential estimate in the proof of theorem [ mainthm ] . [ strictineqprop ] suppose that conditions hold , and let @xmath276 be as in . then for @xmath305 $ ] , there exists @xmath247 and a neighborhood @xmath104 of @xmath306 such that @xmath307 \\ & & { } + \frac{\ell}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log^+ \gamma_{\theta}(y_0 ) \bigr ] \\ & & { } + \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log c_{m}\bigl(\theta',y_0^{m}\bigr ) \bigr].\end{aligned}\ ] ] suppose @xmath103 $ ] . by proposition [ identifiabilityprop ] , there exists @xmath308 such that @xmath309 < h(\theta_0)-\varepsilon.\ ] ] by conditions ( s3 ) ( logarithmic integrability away from @xmath67 ) and ( s5 ) ( mixing ) , there exists a neighborhood @xmath310 of @xmath65 and @xmath311 such that for @xmath312 , we have that @xmath313 \nonumber \\[-8pt ] \\[-8pt ] \nonumber & & { } + \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u ' } \log c_{m}\bigl(\theta , y_0^{m}\bigr ) \bigr ] . \nonumber\end{aligned}\ ] ] fix @xmath312 such that @xmath314 < \limsup_n \frac{1}{n } \e_{\theta_0 } \bigl [ \log p_{\theta ' } \bigl(y_0^{n}\bigr ) \bigr ] + \varepsilon/4.\ ] ] for @xmath315 , let @xmath316 denote the ball of radius @xmath317 about @xmath65 in @xmath3 . for @xmath317 such that @xmath318 , we have that @xmath319 the sum above is integrable with respect to @xmath94 and does not depend on @xmath317 . then ( the reverse ) fatou s lemma implies that @xmath320 \leq\e_{\theta_0 } \bigl [ \limsup_{\eta\to0 } \sup _ { \theta\in b(\theta',\eta ) } \log p_{\theta } \bigl(y_0^{m } \bigr ) \bigr].\ ] ] by condition ( s4 ) [ upper semi - continuity of @xmath321 , we see that @xmath322 \leq\e_{\theta_0 } \bigl [ \log p_{\theta ' } \bigl(y_0^{m}\bigr ) \bigr].\ ] ] now by an appropriate choice of @xmath315 , we have shown that there exists a neighborhood @xmath323 of @xmath65 such that @xmath324 < \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \log p_{\theta ' } \bigl(y_0^{m}\bigr ) \bigr ] + \varepsilon/4.\ ] ] combining estimates ( [ eqn : estimate1])([eqn : estimate4 ] ) , we obtain the desired inequality . proof of theorem [ mainthm ] let @xmath325 be defined as in proposition [ gsmb ] . we prove the theorem by showing the following statement : for each closed set @xmath326 in @xmath3 such that @xmath327 = \varnothing$ ] , it holds that @xmath328 let @xmath326 be a closed subset of @xmath3 such that @xmath327 = \varnothing$ ] . since @xmath3 is compact , @xmath326 is compact . suppose that for each @xmath329 , there exists a neighborhood @xmath104 of @xmath65 such that @xmath330 then by compactness , we would conclude that ( [ suffineq1 ] ) holds and thus complete the proof of the theorem . let @xmath65 be in @xmath326 . let us now show that there exists a neighborhood @xmath104 of @xmath65 such that ( [ suffineq2 ] ) holds . since @xmath65 is in @xmath326 , we have that @xmath103 $ ] . let @xmath331 be as in ( s5 ) . by proposition [ strictineqprop ] , there exists @xmath277 and a neighborhood @xmath310 of @xmath65 such that @xmath332 \nonumber\\ & & { } + \frac{\ell}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u ' } \log^+ \gamma_{\theta}(y_0 ) \bigr ] \\ & & { } + \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u ' } \log c_{m}\bigl(\theta , y_0^{m}\bigr ) \bigr ] . \nonumber \ ] ] by proposition [ basicestimate ] , there exists a neighborhood @xmath333 of @xmath65 such that @xmath334\nonumber \\ & & { } + \frac{\ell}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log^+ \gamma_{\theta}(y_0 ) \bigr ] \\ & & { } + \frac{1}{m+\ell } \e_{\theta_0 } \bigl [ \sup_{\theta\in u } \log c_{m}\bigl(\theta',y_0^{m}\bigr ) \bigr ] . \nonumber\end{aligned}\ ] ] combining ( [ eqn : mainproofest1 ] ) and ( [ eqn : mainproofest2 ] ) , we obtain ( [ suffineq2 ] ) , which completes the proof of the theorem . in this paper , we demonstrate how the properties of a family of dynamical systems affect the asymptotic consistency of maximum likelihood parameter estimation . we have exhibited a collection of general statistical conditions on families of dynamical systems observed with noise , and we have shown that under these general conditions , maximum likelihood estimation is a consistent method of parameter estimation . furthermore , we have shown that these general conditions are indeed satisfied by some classes of well - studied families of dynamical systems . as mentioned in the , our results can be considered as a theoretical validation of the notion from dynamical systems that these classes of systems have `` good '' statistical properties . however , there remain interesting families of systems to which our results do not apply , including some classes of systems that are also believed to have `` good '' statistical properties . in particular , the class of systems modeled by young towers with exponential tail @xcite has exponential decay of correlations and certain large deviations estimates @xcite . these families include a positive measure set of maps from the quadratic family [ @xmath335 and the hnon family , as well as certain billiards and many other systems of physical and mathematical interest @xcite . in short , the setting of systems modeled by young towers with exponential tail provides a very attractive setting in which to consider consistency of maximum likelihood estimation . unfortunately , our proof does not apply to systems in this setting in general , mainly due to the presence of the mixing condition ( s5 ) , which is not satisfied by these systems in general . a natural next step might be to obtain rates of convergence and derive central limit theorems for maximum likelihood estimation . to this end , it might be possible to build off of analogous results for hmms @xcite . we leave these questions for future work .

rare b decays induced by the flavor changing neutral current ( fcnc ) @xmath1 have received a lot of theoretical interest @xcite . these transitions provide an important consistency check of the standard model ( sm ) at loop level , since fcnc transitions are forbidden in the sm at tree level . these decays induced by the fcnc are very sensitive to the new physics beyond the sm . new physics appear in rare decays through the wilson coefficients which can take values different from their sm counterpart or through the new operator structures in an effective hamiltonian . among the hadronic , leptonic and semileptonic decays , the last decay channels are very significant , since they are theoretically , more or less , clean , and they have relatively larger branching ratio . the semileptonic decay channels is described by the @xmath1 transition and they contain many observables like forward backward asymmetry @xmath2 , lepton polarization asymmetries , etc . existence of these observables is very useful and serve as a testing ground for the standard model ( sm ) and for looking new physics beyond th sm . for this reason , many processes , like @xmath3 @xcite , @xmath4 @xcite , @xmath5 @xcite and @xmath6 @xcite@xcite have been studied comprehensively . recently , belle and babar collaborations announced the following results for the branching ratios of the @xmath6 and @xmath5 decays : ( b k^^+ ^- ) = \ { lc ( 11.5^+2.6_-2.4 0.8 0.2 ) 10 ^ -7 & @xcite , + + ( 0.88^+0.33_-0.29 ) 10 ^ -6 & @xcite , . ( b k ^+ ^- ) = \ { lc ( 4.8^+1.0_-0.9 0.3 0.1 ) 10 ^ -7 & @xcite , + + ( 0.65^+0.14_-0.13 0.04 ) 10 ^ -6 & @xcite . . another exclusive decay which is described at inclusive level by the @xmath7 transition is the baryonic @xmath0 decay . unlike mesonic decays , the baryonic decays could maintain the helicity structure of the effective hamiltonian for the @xmath8 transition @xcite . many experimentally measurable quantities such as branching ratio @xcite , @xmath9 polarization and single lepton polarization have already been studied in @xcite and @xcite , respectively . analysis of such quantities can be useful for more precise determination of the sm parameters and looking for new physics beyond the sm . it has been pointed out in @xcite that some of single lepton polarization asymmetries may be too small to be observed and hence may not provide sufficient number of observables in checking the structure of effective hamiltonian . in need of more observables , london _ et al _ @xcite take into account polarizations of both leptons , which are simultaneously measured , and construct maximum number of independent observables . it should be noted that the forward backward asymmetries due to the double lepton polarizations in the @xmath0 decay , are investigated in @xcite . in the present work we analyze the possibility of searching for new physics in the baryonic @xmath0 decay by studying the double lepton polarization asymmetries , using a general form of the effective hamiltonian , including all possible forms of interactions . note that the dependence of polarized forward backward asymmetry to the new wilson coefficients for the meson@xmath10meson transition has been investigated in @xcite and @xcite . in the same manner , an analogous analysis can be carried out to investigate the sensitivity of the double lepton polarization asymmetries to the new wilson coefficients , in the case of baryon@xmath10baryon transition . the paper is organized as follows . in section 2 , using the general , model independent form of the effective hamiltonian , the matrix element for the @xmath0 is obtained . in section 3 we calculate the double lepton polarization asymmetries . section 4 is devoted to the numerical analysis , discussions and conclusions . in this section we derive the matrix element for the @xmath11 decay using the general , model independent form of the effective hamiltonian . at quark level , the matrix element of the @xmath0 decay is described by the @xmath7 transition . the effective hamiltonian for the @xmath7 transition can be written in terms of twelve model independent four fermi interactions as @xcite : [ e6601 ] m v_tbv_ts^\ { c_sl |s_r i _ b_l |^+ c_br |s_l i _ b_r |^+ c_ll^tot |s_l _ b_l |_l ^_l + c_lr^tot |s_l _ b_l |_r ^_r + c_rl |s_r _ b_r |_l ^_l + c_rr |s_r _ b_r |_r ^_r + c_lrlr |s_l b_r |_l _ r + c_rllr |s_r b_l |_l _ r + c_lrrl |s_l b_r |_r _ l + c_rlrl |s_r b_l |_r _ l + c_t |s _ b |^ + i c_te _ |s ^ b |^ } , where @xmath12 is the momentum transfer and @xmath13 are the coefficients of the four fermi interactions , @xmath14 and @xmath15 . the terms with coefficients @xmath16 and @xmath17 describe the penguin contributions , which correspond to @xmath18 and @xmath19 in the sm , respectively . the next four terms in eq . ( [ e6601 ] ) with coefficients @xmath20 and @xmath21 describe vector type interactions , two ( @xmath22 and @xmath23 ) of which contain sm contributions in the form @xmath24 and @xmath24 , respectively . thus , @xmath22 and @xmath23 can be written as [ e6602 ] c_ll^tot c_9^eff- c_10 + c_ll , + c_lr^tot c_9^eff+ c_10 + c_lr , where @xmath25 and @xmath26 describe the contributions of new physics . additionally , eq . ( [ e6601 ] ) contains four scalar type interactions ( @xmath27 and @xmath28 ) , and two tensor type interactions ( @xmath29 and @xmath30 ) . the amplitude of the exclusive @xmath31 decay is obtained by calculating the matrix element of @xmath32 for the @xmath33 transition between initial and final baryon states @xmath34 . it follows from eq . ( [ e6601 ] ) that the matrix elements & & |s _ ( 1 _ 5 ) b _ b , + & & |s _ ( 1 _ 5 ) b _ b , + & & |s ( 1 _ 5 ) b _ b .are needed in order to calculate the @xmath31 decay amplitude . these matrix elements parametrized in terms of the form factors are as follows ( see @xcite ) [ e6603 ] |s _ b _ b |u_u__b , + [ e6604 ] |s _ _ 5 b _ b |u_u__b , + [ e6605 ] |s _ b _ b |u_u__b , + [ e6606 ] |s _ _ 5 b _ b |u__5 u__b , where @xmath35 and @xmath36 . the form factors of the magnetic dipole operators are defined as [ e6607 ] using the identity _ _ 5 = - _ ^ , and eq . ( [ e6605 ] ) , the last expression in eq . ( [ e6607 ] ) can be written as |s i _ _ 5 q^b _ b |u_u__b . multiplying ( [ e6605 ] ) and ( [ e6606 ] ) by @xmath37 and comparing with ( [ e6607 ] ) , one can easily obtain the following relations between the form factors [ e6608 ] f_2^t f_t + f_t^s q^2 , f_1^t q^2 = - f_3^t , + g_2^t g_t + g_t^s q^2 , + g_1^t q^2 = g_3^t . the matrix element of the scalar @xmath38 and pseudoscalar @xmath39 operators can be obtained from ( [ e6603 ] ) and ( [ e6604 ] ) by multiplying both sides to @xmath40 and using equation of motion . neglecting the mass of the strange quark , we get [ e6609 ] |s b _ b |u_u__b , + [ e6610 ] |s _ 5 b _ b |u_u__b . using these definitions of the form factors , for the matrix element of the @xmath31 we get @xcite [ e6611 ] + i _ q^+q _ ] u__b + 4 c_t |^ |u_u__b + 4 c_te ^ | _ i |u_u__b } , where the explicit forms of the functions @xmath41 and @xmath42 @xmath43 and @xmath44 can be written as @xcite [ e6612 ] a_1 f_1^t - g_1^t c_sl + f_1^t+g_1^t c_br + f_1-g_1 c_ll^tot + c_lr^tot + f_1+g_1 c_rl + c_rr , + a_2 a_1 1 2 , + a_3 a_1 1 3 , + b_1 a_1 g_1 - g_1 ; g_1^t - g_1^t , + b_2 b_1 1 2 , + b_3 b_1 1 3 , + d_1 c_rr - c_rl f_1+g_1 + c_lr^tot - c_ll^tot f_1-g_1 , + d_2 d_1 1 2 , + d_3 d_1 1 3 , + e_1 d_1 g_1 - g_1 , + e_2 e_1 1 2 , + e_3 e_1 1 3 , + n_1 ( f_1 m__b - m_+ f_3 q^2 ) ( c_lrlr + c_rllr + c_lrrl + c_rlrl ) , + n_2 n_1 c_lrrl - c_lrrl ; c_rlrl - c_rlrl , + h_1 ( g_1 m__b + m_- g_3 q^2 ) ( c_lrlr - c_rllr + c_lrrl - c_rlrl ) , + h_2 h_1 c_lrrl - c_lrrl ; c_rlrl - c_rlrl . from these expressions it follows that @xmath45 decay is described in terms of many form factors . it is shown in @xcite that hqet reduces the number of independent form factors to two ( @xmath46 and @xmath47 ) irrelevant of the dirac structure of the corresponding operators , i.e. , [ e6613 ] ( p _ ) |s b ( p__b ) = |u_u__b , where @xmath48 is an arbitrary dirac structure and @xmath49 is the four velocity of @xmath50 . comparing the general form of the form factors given in eqs . ( [ e6603])([e6610 ] ) with ( [ e6613 ] ) , one can easily obtain the following relations among them ( see also @xcite ) [ e6614 ] g_1 f_1 = f_2^t= g_2^t = f_1 + f_2 , + g_2 f_2 = g_3 = f_3 = g_t^v = f_t^v = , + g_t^s f_t^s = 0 , + g_1^t f_1^t = q^2 , + g_3^t m__b + m _ , + f_3^t - m__b - m _ , where @xmath51 . from eq . ( [ e6611 ] ) , we get for the unpolarized decay width [ e6615 ] _ 0 = v_tb v_ts^^2 ^1/2(1 , _ , ) v , where @xmath52 is the triangle function , @xmath53 and @xmath54 is the lepton velocity , with @xmath55 . the explicit expressions for @xmath56 and @xmath57 can be found in @xcite . in the next section , we present the expressions for the double lepton polarization asymmetries . in the present section we calculate the double lepton polarization asymmetries , i.e. , when polarizations of both leptons are considered . in order to calculate the double lepton polarization asymmetries , we define the following orthogonal unit vectors @xmath58 in the rest frame of @xmath59 ( @xmath60 or @xmath61 , stand for longitudinal , transversal or normal polarizations , respectively . [ e6616 ] s^-_l 0,_l^-= 0 , , + s^-_n 0,_n^-= 0 , , + s^-_t 0,_t^-= 0,_n^- _ l^- , + s^+_l 0,_l^+= 0 , , + s^+_n 0,_n^+= 0 , , + s^+_t 0,_t^+= 0,_n^+ _ l^+ , where @xmath62 and @xmath63 are the three momenta of the leptons @xmath64 and @xmath9 baryon in the center of mass frame ( cm ) of @xmath65 system , respectively . transformation of unit vectors from the rest frame of the leptons to cm frame of leptons can be accomplished by the lorentz boost . boosting of the longitudinal unit vectors @xmath66 yields [ e6617 ] s^_l _ cm , , where @xmath67 , @xmath68 and @xmath69 are the energy and mass of leptons in the cm frame , respectively . the remaining two unit vectors @xmath70 , @xmath71 are unchanged under lorentz boost . having obtained the above expressions , we now define the double polarization asymmetries as follows @xcite : [ e6618 ] p_ij(q^2 ) , where , the first subindex @xmath72 represents lepton and the second one antilepton . using this definition of @xmath73 , nine double lepton polarization asymmetries are calculated . their expressions are [ e6619 ] p_ll \ { 32 m__b^2 _ ( 1 + ) ( a_1+b_1 ) c_t^f_t^s + 12 ( 1 + ) ( 1 - 2+_- ) \ { _ ( d_1-e_1 ) h_2^ + 8 m__b^2 _ ( a_2+b_2 ) c_t f_t^v + 32 m__b f_t f_t^v } + 12 _ ( 1- ) ( 1 + 2+_- ) ( d_1+e_1 ) f_2^ + 32 m__b^3 _ ( a_2+b_2 ) c_t^f_t^s + 3 ( 1 + 2+_- ) + 12 m__b ( 1-_+ ) + 3 ( 1 - 2 + _ - ) + 12 m__b ( 1-_- ) + 24 ( 1+v^2 ) ( a_1 b_1^+ d_1 e_1^+ m__b^2 a_2 b_2^ ) + 48 m__b^2 _ ^2 ( 1+_- ) d_3 e_3^ + 16 m__b + 2 ( 1+v^2 ) ( a_1 ^2 + b_1 ^2 ) + 2 ( d_1 ^2 + e_1 ^2 ) + 2 m__b^2 ( 1+v^2 ) ( a_2 ^2 + b_2 ^2 ) + 64 m__b^3 + 4 m__b^2 v^2 ( d_2 ^2 + e_2 ^2 ) + 24 m__b ( 1-_- ) v^2 ( d_1 e_2^+ d_2 e_1^ ) + 24 m__b ( 1-_+ ) v^2 ( d_1 d_2^+ e_1 e_2^ ) + 48 m__b^2 ( v^2 d_2 e_2^+ 2 _ ^2 d_3 e_3^ ) + 128 m__b^2 f_t f_t^s + 32 m__b _ \ { ( c_t - 2 c_te ) - 6 ( c_t+2 c_te ) } a_2^f_t + 32 m__b _ \ { ( c_t + 2 c_te ) - 6 ( c_t-2 c_te ) } b_2^f_t + 96 _ a_1^f_t + 96 _ b_1^f_t + 128 c_t ^2 f_t ^2 + 512 c_te ^2 f_t ^2 } , + + [ e6620 ] p_ln \ { 4 _ ( 1- _ ) ( a_1^d_1 + b_1^e_1 ) + 32 m__b^2 _ ^2 ( 1- ) ( 1 + 2+_- ) ( d_1^+ e_1^ ) c_t f_t^s + 16 _ ( f_2 c_t^f_t^- 2 h_2 c_te^f_t^ ) + 4 m__b _ ( a_1^e_3 - a_2^e_1 + b_1^d_3 -b_2^d_1 ) + ( 1 + ) \{(a_1^+b_1^ ) f_2 - 16 m__b _ c_t^f_t^v } + 32 m__b _ ^2 + ( 1- ) ( a_1-b_1 ) h_2^ + + 4 m__b _ ( a_1^d_3 + a_2^d_1 + b_1^e_3 + b_2^e_1 ) + 16 m__b^2 _ ( 1 + 2 + _ - ) \ { c_t^f_t^s } + m__b ^2 + v^2 \{(d_1 + e_1 ) f_1 ^ - ( d_1 - e_1 ) h_1 ^ - } + 8 v^2 \{(a_1 - b_1 ) c_t^f_t^+ 2 ( a_1 + b_1 ) c_te^f_t^ + } + 8 m__b ( 1 - _ ) v^2 + m__b ^2 v^2 + 8 m__b^2 ^2 ( 1 - ) v^2 + 16 m__b } , + + [ e6621 ] p_nl - \ { 4 _ ( 1- _ ) ( a_1^d_1 + b_1^e_1 ) + 32 m__b^2 _ ^2 ( 1- ) ( 1 + 2+_- ) ( d_1^+ e_1^ ) c_t f_t^s + 16 _ ( f_2 c_t^f_t^- 2 h_2 c_te^f_t^ ) + 4 m__b _ ( a_1^e_3 - a_2^e_1 + b_1^d_3 -b_2^d_1 ) + ( 1 + ) \{(a_1^+b_1^ ) f_2 - 16 m__b _ c_t^f_t^v } + 32 m__b _ ^2 + ( 1- ) ( a_1-b_1 ) h_2^ + + 4 m__b _ ( a_1^d_3 + a_2^d_1 + b_1^e_3 + b_2^e_1 ) + 4 m__b _ ( a_1^d_3 + a_2^d_1 + b_1^e_3 + b_2^e_1 ) + m__b ^2 + v^2 \{(d_1 + e_1 ) f_1 ^ - ( d_1 - e_1 ) h_1^ + } + 8 v^2 \{(a_1 - b_1 ) c_t^f_t^+ 2 ( a_1 + b_1 ) c_te^f_t^ + } + 8 m__b ( 1 - _ ) v^2 + m__b ^2 v^2 + 8 m__b^2 ^2 ( 1 - ) v^2 + 16 m__b } , + + [ e6622 ] p_lt \ { 4 _ ( 1- _ ) + 4 _ ( a_1 d_1 ^ - b_1 e_1^+ 4 f_1^c_t f_t ) + 4 _ \ { 8 h_1 c_te^f_t^- m__b } + ( 1 - ) + ( 1 + ) + 8 \ { ( c_t + 2 c_te ) f_t + ( c_t - 2 c_te ) f_t } + 4 m__b ( 1- _ ) \ { 2 + m__b _ ( a_2 d_2 ^ - b_2 e_2^ ) } + 4 m__b _ + m__b ^2 \ { 8 f_t^v + ( a_2-b_2 ) h_1^ + ( a_2 + b_2 ) f_1 ^ - ( d_2-e_2 ) h_2 ^ - ( d_2+e_2 ) f_2 ^ - 4 m__b _ ( d_2 d_3^+ e_2 e_3^ ) } + 8 m__b^2 ^2 ( 1- ) f_t^v + 16 m__b^2 _ ( 1 + 2 + _ - ) f_1^c_t f_t^s } , + + [ e6623 ] p_tl \ { 4 _ ( 1- _ ) + 4 _ ( a_1 d_1 ^ - b_1 e_1^+ 4 f_1^c_t f_t ) + 4 _ \ { 8 h_1 c_te^f_t^- m__b } + ( 1 - ) + ( 1 + ) + 8 \ { ( c_t + 2 c_te ) f_t + ( c_t - 2 c_te ) f_t } + 4 m__b ( 1- _ ) \ { 2 - m__b _ ( a_2 d_2 ^ - b_2 e_2^ ) } + 4 m__b _ + m__b ^2 \ { 8 f_t^v - ( a_2-b_2 ) h_1^ + ( a_2 + b_2 ) f_1 ^ - ( d_2-e_2 ) h_2 ^ - ( d_2+e_2 ) f_2 ^ - 4 m__b _ ( d_2 d_3^+ e_2 e_3^ ) } + 8 m__b^2 ^2 ( 1- ) f_t^v + 16 m__b^2 _ ( 1 + 2 + _ - ) f_1^c_t f_t^s } , + + [ e6624 ] p_nt \ { 4 ( a_1 d_1^+b_1 e_1^ ) + 32 m__b _ + 32 m__b^2 _ ( 1 + ) + 6 _ ( 1 + ) ( 1 - 2 + _ - ) ( d_1-e_1 ) h_1^ + 6 _ ( 1- ) ( 1 + 2 + _ - ) ( d_1+e_1 ) f_1^ + 32 m__b^3 _ f_t^s + 4 m__b^2 ( a_2^d_2 + b_2^e_2 ) + 256 m__b^2 \ { - m__b ( 1 + ) } c_t c_te^ + 3 ( 1 - 2 + _ - ) + 96 m__b ( 1 + ) ( 1 - 2 + _ - ) \ { 8 c_t c_te^ + m__b _ ( a_2+b_2 ) c_te^f_t^v } + ( 1 + 2 + _ - ) \ { 128 m__b^4 f_t^s ^2 c_t c_te^ + 3 f_1^ } + 32 m__b + + 1536 f_t ^2 c_t c_te^ + 48 _ \ { ( 1- _ ) + } b_1^f_t + 16 m__b _ \ { 2(1-_)^2 ( c_t-2 c_te ) + - ^2 ( c_t-2 c_te ) } a_2^f_t + 16m__b _ \ { 2(1-_)^2 ( c_t+2 c_te ) + - ^2 ( c_t+2 c_te ) } b_2^f_t + 48 _ \ { ( 1- _ ) + } a_1^f_t } , + + [ e6625 ] p_tn - \ { 4 ( a_1 d_1^+b_1 e_1^ ) + 32 m__b _ + 32 m__b^2 _ ( 1 + ) + 6 _ ( 1 + ) ( 1 - 2 + _ - ) ( d_1-e_1 ) h_1^ + 6 _ ( 1- ) ( 1 + 2 + _ - ) ( d_1+e_1 ) f_1^ + 32 m__b^3 _ f_t^s + 4 m__b^2 ( a_2^d_2 + b_2^e_2 ) + 256 m__b^2 \ { - m__b ( 1 + ) } c_t c_te^ + 3 ( 1 - 2 + _ - ) + 96 m__b ( 1 + ) ( 1 - 2 + _ - ) \ { 8 c_t c_te^ + m__b _ ( a_2+b_2 ) c_te^f_t^v } + ( 1 + 2 + _ - ) \ { 128 m__b^4 f_t^s ^2 c_t c_te^ + 3 f_1^ } + 32 m__b + + 1536 f_t ^2 c_t c_te^ + 48 _ \ { ( 1- _ ) + } b_1^f_t + 16 m__b _ \ { 2(1-_)^2 ( c_t-2 c_te ) + - ^2 ( c_t-2 c_te ) } a_2^f_t + 16 m__b _ \ { 2(1-_)^2 ( c_t+2 c_te ) + - ^2 ( c_t+2 c_te ) } b_2^f_t + 48 _ \ { ( 1- _ ) + } a_1^f_t } , + + [ e6626 ] p_nn \ { 96 m__b^2 _ ( 1 + ) ( a_1+b_1 ) c_t^f_t^s + 96 _ ^2 ( a_1 b_1^+ d_1 e_1^ ) + 48 m__b _ ^2 ( 1-_+ ) ( a_1 a_2^+ b_1 b_2^ ) + 96 m__b _ ( 1 + ) ^2 ( 1 - 2 + _ - ) ( a_1 + b_1 ) c_t^f_t^v + 12 _ ( 1 - ) ( 1 + 2 + _ - ) ( d_1 + e_1 ) f_2^ + 96 m__b^3 _ ^2 ( a_2 + b_2 ) c_t^f_t^s + 12 _ ( 1 + ) ( 1 - 2 + _ - ) + 3 ^2 ( 1 + 2 + _ - ) + 24 m__b _ ^2 + 48 m__b _ ^2 ( 1-_- ) ( a_1 b_2^+ a_2 b_1 ^+ d_1 e_3^+ d_3 e_1^ ) + 4 [ + 2 _ ^2 ( 1 + _ ^2 - 2 _ + _ + - 2 ^2 ) ] ( a_1 ^2 + b_1 ^2 - d_1 ^2 - e_1 ^2 ) + 3 ^2 ( 1 - 2 + _ - ) + 96 m__b^2 _ ^2 ^2 ( a_2 b_2^+ d_3 e_3^ ) + 64 m__b^2 ^2 \{2 f_t f_t^s - m__b } + 4 m__b^2 ^2 v^2 ( d_2 ^2 + e_2 ^2 ) + 4 m__b^2 \ { - 2 _ ^2 [ 2 ( 1 + _ ^2 ) - ( 1 + ) - _ ( 4 + ) ] } ( a_2 ^2 + b_2 ^2 ) + 96 m__b _ ^2 + 96 _ + 96 _ + 768 \ { c_t ^2 + 4 c_te ^2 } f_t ^2 + 96 m__b _ ^2 a_2 f_t^ + 3 ^2 v^2 + 384 m__b ^2 ( 1 + ) ( 1 - 2 + _ - ) ( c_t ^2 - 4 v^2 c_te ^2 ) f_t f_t^v + 64 m__b^2 ( 1 - 2 + _ - ) \ { c_t ^2 - 4 v^2 ( 1 + 2 + _ + 2 ) c_te ^2 } f_t^v ^2 } , + + [ e6627 ] p_tt \ { 32 m__b^2 _ ( 1 + ) ( a_1+b_1 ) c_t^f_t^s + 96 _ ^2 ( a_1 b_1^+ d_1 e_1^ ) + 48 m__b _ ^2 ( 1-_+ ) ( d_1 d_3^+ e_1 e_3^ ) + 12 _ ( 1 - ) ( 1 + 2 + _ - ) ( d_1 + e_1 ) f_2^ + 32 m__b^3 _ ^2 ( a_2 + b_2 ) c_t^f_t^s + 96 m__b^2 _ ^2 ^2 ( a_2 b_2^+ d_3 e_3^ ) + 12 _ ( 1 + ) ( 1 - 2 + _ - ) + 3 ^2 ( 1 + 2 + _ - ) + 24 m__b _ ^2 + 48 m__b _ ^2 ( 1-_- ) ( a_1 b_2^+ a_2 b_1 ^+ d_1 e_3^+ d_3 e_1^ ) + 4 [ - 2 _ ^2 ( 1 + _ ^2 - 2 _ + _ + - 2 ^2 ) ] ( a_1 ^2 + b_1 ^2 ) + 4 m__b^2 \ { ^2 + _ ^2 [ 4 ( 1- _ ) ^2 - 2 ( 1 + _ ) - 2 ^2 ) ] } ( a_2 ^2 + b_2 ^2 ) + 4 \ { - 2 _ ^2 [ 5 ( 1- _ ) ^2 - 7 ( 1 + _ ) + 2 ^2 ) ] } ( d_1 ^2 + e_1 ^2 ) + 32 m__b _ + 3 ^2 ( 1 - 2 + _ - ) + 64 m__b^2 ^2 \{2 f_t f_t^s - m__b } + 32 m__b _ \ { [ 4 ( 1-_)^2 - 5 ( 1+_)+^2 ] a_2^(c_t-2 c_te ) f_t - 6 a_2^(c_t+2 c_te ) f_t } + 32 m__b _ \ { [ 4 ( 1-_)^2 - 5 ( 1+_)+^2 ] b_2^(c_t+2 c_te ) f_t - 6 b_2^(c_t-2 c_te ) f_t } + 256 \ { - 3 ^2 + _ ^2 } c_t ^2 f_t ^2 + 1024 \{3 ^2 + _ ^2 } c_te ^2 f_t ^2 + 96 _ a_1 f_t^ + 96 _ b_1 f_t^ + 4 m__b^2 ^2 v^2 ( d_2 ^2 + e_2 ^2 ) + 3 ^2 v^2 + 384 m__b ^2 ( 1 + ) ( 1 - 2 + _ - ) ( c_t ^2 - 4 v^2 c_te ^2 ) f_t f_t^v + 64 m__b^2 ( 1 - 2 + _ - ) \ { c_t ^2 + 4 v^2 ( 1 + 2 + _ + 2 ) c_te ^2 } f_t^v ^2 } . we start this section by presenting the numerical results for all possible double lepton polarization asymmetries . the values of the input parameters we use in out calculations are : @xmath74 , @xmath75 , @xmath76 . @xmath77 . for the wilson coefficients we use their sm values which are given as : @xmath78 , @xmath79 and @xmath80 . the magnitude of @xmath81 is quite well constrained from @xmath82 decay , and hence well established . moreover , we will fix the values of the wilson coefficients , i.e. , @xmath17 and @xmath16 are both related to @xmath81 as follows : @xmath83 and @xmath84 . as far as the wilson coefficient @xmath85 is considered , we take into account only the short distance contributions , and we neglect the long distance contributions , coming from the production of @xmath86 pair at intermediate states . it is well known that the form factors are the main and the most important input parameters necessary in the numerical calculations . the calculation of the form factors of @xmath87 transition does not exist at present . but , we can use the results from qcd sum rules in corporation with hqet @xcite . we noted earlier that , hqet allows us to establish relations among the form factors and reduces the number of independent form factors into two . in @xcite , the @xmath88 dependence of these form factors are given as follows f ( ) = . the values of the parameters @xmath89 and @xmath90 are given in table 1 . @xmath91 in further numerical analysis , the values of the new wilson coefficients which describe new physics beyond the sm , are needed . in numerical calculations we will vary all new wilson coefficients in the range @xmath92 . the experimental results on the branching ratio of the @xmath6 decay @xcite and the bound on the branching ratio of @xmath93 @xcite suggest that this is the right order of magnitude for the vector and scalar interaction coefficients . here , we emphasize that the existing experimental results on the @xmath6 and @xmath5 decays put stronger restrictions on some of the new wilson coefficients . for example , @xmath94 , @xmath95 , @xmath96 and @xmath97 , and all of the remaining wilson coefficients vary in the region @xmath92 . it follows from the explicit expressions of the double lepton polarization asymmetries that they depend on @xmath88 and the new wilson coefficients . for this reason there may appear difficulties in studying the dependencies of the physical properties on both parameters at the same time . hence , it is necessary to eliminate the dependence of @xmath73 on one of these parameters . here in the present work , we eliminate @xmath88 dependence of @xmath73 by performing integration over @xmath88 in the kinematically allowed region . the averaging of @xmath73 over @xmath88 is defined as [ e6628 ] p_ij = in figs . ( 1)(5 ) we present the correlation of the averaged double lepton polarization asymmetries on the branching ratio for the @xmath98 decay . from these figures we obtain the following results . * there exist regions of the new wilson coefficients where @xmath99 departs considerably from the sm result when @xmath100 is close to the sm prediction . * @xmath101 and @xmath102 are both very sensitive to the existence of the tensor and scalar interactions . we observe that @xmath103 exceeds the sm prediction 3 times and @xmath102 2 times , respectively . more essential than that , @xmath101 as well as @xmath102 both change their signs when wilson coefficients vary in the allowed region . such behaviors can serve as a good test for establishing new physics beyond the sm . * @xmath104 and @xmath105 are quite sensitive to the existence of the vector interactions , especially to @xmath26 . in the presence of this coefficient , the values of @xmath104 and @xmath105 can both exceed the sm result 34 times , and both changes their sign when @xmath26 varies in the allowed region . therefore , determination of the sign and magnitude of @xmath104 and @xmath106 can give unambiguous information about the existence of the new vector type interaction . we do not present the correlation of and @xmath107 , @xmath108 , @xmath109 and @xmath110 on the branching ratio , since their values are quite small . in figs . ( 6)(12 ) the correlation of @xmath111 on branching ratio for the @xmath112 decay are presented . similar to the @xmath113 decay , we observe that several of the double lepton polarization asymmetries are very sensitive to the existence of new physics and the presence of the new wilson coefficients can produce results that depart considerably from the sm prediction . more precisely , we can briefly comment on the results as follows : * all @xmath111 are very sensitive to the existence of tensor interactions with coefficients @xmath29 and @xmath30 . * the situation for the @xmath112 decay is slightly different from the @xmath113 decay , for which a considerable dependence on the existence of the scalar type interaction is observed , especially for the coefficients @xmath114 and @xmath115 . * @xmath116 , @xmath101 , @xmath102 and @xmath105 exhibit strong dependence on vector interaction with coefficients @xmath26 , while @xmath107 and @xmath108 do so on the vector interaction with coefficient @xmath25 . * @xmath116 and @xmath104 change sign when the new wilson coefficients vary in the allowed region . for this reason , measurement of magnitude and sign of the averaged double lepton polarization asymmetry can be a very useful tool in establishing new physics beyond the sm . the following question we would like to discuss is the possibility of the existence of such a region of the new wilson coefficients for which the branching ratio coincides with sm result , while the double lepton polarizations do not , similar to the @xmath117 decay . from the relevant figures we see that , the @xmath118 decay is far more informative on this issue . there exist regions of the new wilson coefficients @xmath26 , @xmath29 and @xmath30 , where double lepton polarization asymmetries differ from the sm result , but branching ratio coincides with that of the sm . at the end of this section , let us discuss the problem of measurement of the lepton polarization asymmetries in experiments . experimentally , to measure an asymmetry @xmath119 of the decay with the branching ratio @xmath120 at @xmath121 level , the required number of events ( i.e. , the number of @xmath122 pair ) are given by the expression n = , where @xmath123 and @xmath124 are the efficiencies of the leptons . typical values of the efficiencies of the @xmath125leptons range from @xmath126 to @xmath127 for their various decay modes ( see for example @xcite and references therein ) , and the error in @xmath125lepton polarization is estimated to be about @xmath128 @xcite . as a result , the error in measurement of the @xmath125lepton asymmetries is of the order of @xmath129 , and the error in obtaining the number of events is about @xmath126 . from the expression for @xmath61 we see that , in order to observe the lepton polarization asymmetries in @xmath113 and @xmath112 decays at @xmath130 level , the minimum number of required events are ( for the efficiency of @xmath125lepton we take @xmath131 ) : * for the @xmath113 decay n = \ { + ll 2.0 10 ^ 6 & ( p_ll ) , + 2.0 10 ^ 8 & ( p_lt = p_tl , p_nn , p_tt ) , + . * for @xmath112 decay n = \ { + ll ( 4.0 2 ) 10 ^ 9 & ( p_lt , p_nn ) , + ( 1.0 0.5 ) 10 ^ 9 & ( p_tt ) , + ( 2.0 1.0 ) 10 ^ 11 & ( p_ln , p_nl ) , + ( 9.0 4.5 ) 10 ^ 8 & ( p_tl ) . the number of @xmath122 pairs , that are produced at b factories and lhc are about @xmath132 and @xmath133 , respectively . as a result of a comparison of these numbers and @xmath61 , we conclude that , only @xmath134 in the @xmath135 decay and @xmath136 , @xmath137 and @xmath138 in the @xmath112 decay , can be detectable at lhc . in summary , we present the most general analysis of the double lepton polarization asymmetries in the @xmath139 decay using the most general , model independent form of the effective hamiltonian . the correlation of the averaged double lepton polarization asymmetries on the branching ratio have been studied . our results show that the study of double lepton polarization asymmetries can serve as good test for establishing new physics beyond the sm . moreover , we observe that there exist regions of the new wilson coefficients for which double lepton polarization asymmetries depart considerably from the sm , while the branching ratio coincides with that of the sm predictions . 99 a. ali , prep : hep ph/0312303 ( 2003 ) . s. r. choudhury and n. gaur , phys . d * 66 * , 094015 ( 2002 ) ; t. m. aliev , m. savci , phys . d * 60 * , 014005 ( 1999 ) . s. rai choudhury , n. gaur and n. mahajan , phys . d * 66 * , 054003 ( 2002 ) ; s. r. choudhury and n. gaur , prep : hep ph/0205076 ( 2002 ) ; e. o. iltan and g. turan , phys d * 61 * , 034010 ( 2000 ) . s. r. choudhury , a. s. cornell , n. gaur and g. c. joshi , phys . d * 69 * , 054018 ( 2004 ) ; t. m. aliev , v. bashiry and m. savci , eur . j. * 35 * , 197 ( 2004 ) ; t. m. aliev , m. k. akmak , a. zpineci and m. savci , phys . d * 64 * , 055007 ( 2001 ) . f. krger and l. m. sehgal , phys . b * 380 * , 199 ( 1996 ) ; j. l. hewett , phys . d * 53 * , 4964 ( 1996 ) . s. fukae , c. s. kim , t. morozumi and t. yoshikawa , phys . d * 59 * , 074013 ( 1999 ) . s. rai choudhury , a. gupta and n. gaur , phys . rev . d * 60 * , 115004 ( 1999 ) ; s. fukae , c. s. kim and t. yoshikawa , phys . d * 61 * , 074015 ( 2000 ) ; d. guetta and e. nardi , phys . d * 58 * , 012001 ( 1998 ) ; a. s. cornell , n. gaur , prep : hep ph/0408164 ( 2004 ) . a. ali , p. ball , l. t. handoko and g. hiller , phys . d * 61 * , 074024 ( 2000 ) . t. m. aliev , m. k. akmak and m. savci , nucl . b * 607 * , 305 ( 2001 ) ; t. m. aliev and m. savci , phys . b * 481 * , 275 ( 2000 ) . g. burdman , phys . d * 52 * , 6400 ( 1995 ) , t. m. aliev , v. bashiry , m. savci , jhep * 0405 * , 037 ( 2004 ) . t. m. aliev , a. zpineci , m. savci , c. yce , phys . d * 66 * , 115006 ( 2002 ) ; t. m. aliev , a. zpineci , m. savci , phys . b * 511 * , 49 ( 2001 ) ; t. m. aliev , a. zpineci , m. savci , nucl . b * 585 * , 275 ( 2000 ) ; t. m. aliev , c. s. kim , y. g. kim , phys . d * 62 * , 014026 ( 2000 ) ; t. m. aliev , e. o. iltan , n. k. pak , phys . b * 451 * , 175 ( 1999 ) ; t. m. aliev , c. s. kim , m. savci , phys . b * 441 * , 410 ( 1998 ) . a. ishikawa _ et . al _ , belle collaboration , phys . * 91 * , 261601 ( 2003 ) . b. aubert _ et . al _ , babar collaboration , phys . lett . * 91 * , 221802 ( 2003 ) . t. mannel and s. recksiegel , j. phys . g * 24 * , 979 ( 1998 ) . t. m. aliev , a. zpineci and m. savci , phys . d * 65 * , 115002 ( 2002 ) ; + t. m. aliev , m. savci , j. phys . g * 26 * , 997 ( 2000 ) . t. m. aliev , a. zpineci and m. savci , nucl . b * 649 * , 1681 ( 2003 ) . t. m. aliev , a. zpineci and m. savci , phys . d * 67 * , 035007 ( 2003 ) . w. bensalem , d. london , n. sinha and r. sinha , phys . d * 67 * , 034007 ( 2003 ) . t. m. aliev , v. bashiry and m. savci , prep : hep ph/0407217 ( 2004 ) . c. h. chen and c. q. geng , phys . d * 64 * , 074001 ( 2001 ) . t. mannel , w. roberts and z. ryzak , nucl . b * 355 * , 38 ( 1991 ) . c. s. huang , h. g. yan , phys . d * 59 * , 114022 ( 1999 ) . m. c. chang _ _ , belle collaboration , phys . d * 68 * , 111101 ( 2003 ) . g. abbiendi _ et . al _ , opal collaboration , phys . b * 492 * , 23 ( 2000 ) . a. rouge , workshop on @xmath125physics , orcay , france ( 1990 ) . * fig . ( 1 ) * parametric plot of the correlation between the averaged double lepton polarization asymmetry @xmath140 and the branching ratio for the @xmath117 decay , when both leptons are longitudinally polarized . + + * fig . ( 2 ) * the same as in fig . ( 1 ) , but for the averaged double lepton polarization asymmetry @xmath141 . + + * fig . ( 3 ) * the same as in fig . ( 1 ) , but for the averaged double lepton polarization asymmetry @xmath142 . + + * fig . ( 4 ) * the same as in fig . ( 1 ) , but for the averaged double lepton polarization asymmetry @xmath143 . + + * fig . ( 5 ) * the same as in fig . ( 1 ) , but for the averaged double lepton polarization asymmetry @xmath144 . + + * fig . ( 6 ) * the same as in fig . ( 1 ) , but for the @xmath145 decay . + + * fig . ( 7 ) * the same as in fig . ( 6 ) , but for the averaged double lepton polarization asymmetry @xmath146 . + + * fig . ( 8) * the same as in fig . ( 6 ) , but for the averaged double lepton polarization asymmetry @xmath147 . + + * fig . ( 9 ) * the same as in fig . ( 2 ) , but for the @xmath145 decay . + + * fig . ( 10 ) * the same as in fig . ( 3 ) , but for the @xmath145 decay . + + * fig . ( 11 ) * the same as in fig . ( 4 ) , but for the @xmath145 decay . + + * fig . ( 12 ) * the same as in fig . ( 10 ) , but for the averaged double lepton polarization asymmetry @xmath144 .

the physical mechanisms responsible for the generation of primordial perturbations seeding present - day large - scale structure , may leave their imprint in the form of small deviations from a gaussian distribution of the primordial perturbations . searches for primordial non - gaussianity can thereby provide key information on the origin and evolution of cosmological structures ( e.g. , ref . @xcite and references therein ) . although the standard single - field , slow - roll , canonical kinetic energy and adiabatic vacuum state inflation generates very small non - gaussianity , any inflationary model that deviates from this may entail a larger level of it ( refs . @xcite and references therein ) . deviations from gaussian initial conditions are often taken to be of the so - called local type and are parameterized by the constant dimensionless parameter @xmath7 @xcite : @xmath8 where @xmath9 denotes bardeen s gauge - invariant potential ( evaluated deep in the matter era ) and @xmath10 is a gaussian random field . is linearly extrapolated to @xmath11 , while in the cosmic microwave background ( cmb ) convention it is evaluated deep in the matter era . thus , @xmath12 f^{\rm cmb}_{\rm nl}\sim1.3f^{\rm cmb}_{\rm nl}$ ] , where @xmath13 denotes the @xmath14-induced linear growth suppression factor . in this paper we will use the cmb convention . ] recently , a method for constraining non - gaussianity from large - scale structure surveys has been proposed @xcite , which exploits the fact that the clustering of dark matter halos where galaxies form on large scales is modified in a scale - dependent way by the presence of even small amount of non - gaussianity . in particular , a non - gaussianity described by eq . ( [ eq : fnl ] ) , introduces a scale - dependent boost ( for @xmath15 and a suppression for @xmath16 ) of the halo power spectrum proportional to @xmath17 on large scales ( @xmath18mpc ) , which evolves roughly as @xmath19 . large - scale structure ( lss ) surveys covering large volumes are needed to access the scales where the signal arises ( e.g. , ref . @xcite and references therein ) . the possibility of using high redshift data to constrain non - gaussianity has also been addressed by means of hydrodynamic simulations of the intergalactic medium @xcite . among the many currently - available tracers of the lss , the radio sources from the nrao vla sky survey ( nvss ) @xcite and the quasar catalogue of the sloan digital sky survey release six ( sdss dr6 qsos ) @xcite are particularly interesting since they span large volumes extending out to substantial redshifts @xcite . indeed these source samples were shown to provide tight constraints on primordial non - gaussianity by the cross - correlation measurements between cmb temperature fluctuations and the lss number density which could be used to detect the late integrated sachs - wolfe ( isw ) @xcite effect @xcite . in recent papers @xcite it was shown that the observed nvss and sdss dr6 qsos auto - correlation function ( acf ) and cross - correlation function ( ccf ) hint at a positive value of @xmath20 at more than @xmath21 . this is because the acf is found to be still positive on angular scales @xmath22 , which , for the median source redshift ( @xmath23 ) , correspond to linear scales where the correlation function should be negative if the density fluctuation field is gaussian . a positive @xmath7 adds power on large angular scales , accounting for the observed acf . in this paper , we will revisit the constraints on the primordial non - gaussianity from these two lss tracers using the power spectrum @xmath0 , instead of the correlation function @xmath24 . while it is true that the angular correlation function and the angular power spectrum are spherical harmonic pairs and thus in principle they should enclose the same physical information , each of the two measurements is affected by systematic errors in different ways . for example the correlation function is less affected by the survey mask than the power spectrum but error estimation for the power spectrum is much easier then for the correlation function . a direct comparison of parameter constraints obtained via the correlation function and the power spectrum analyses offer therefore a way to quantify possible systematic effects . we will also use the recently updated sdss ii lrg photometric redshift survey , megaz - lrg ( dr7 ) @xcite to constrain the primordial non - gaussianity . ref . @xcite used the lrg angular power spectrum for a sample close to the sdss dr3 to constrain @xmath7 . here we revisit the analysis using an improved bias model , improved lrg catalog , updated knowledge of the sources redshift distribution , up - to - date cmb maps and complementary analysis methods . furthermore , we explore the constraints on physically motivated primordial non - gaussianity shapes that are different from the local case . we consider non - gaussianities which effects on the halo clustering is well described by the equilateral and enfolded templates . we consider the angular power spectra of these three lss tracers , as well as their cross - correlation power spectra with the cmb temperature fluctuation . finally , we consider the cubic correction on the halo bias , which can be motivated in scenarios like the curvaton model , in which a large cubic correction can be produced while simultaneously keeping the @xmath7 correction small @xcite . the structure of the paper is as follows . in section [ theory ] we review the effects of primordial non - gaussianity on the power spectrum @xmath0 . section [ data ] contains the analysis of power spectrum @xmath0 for nvss radio sources , sdss dr6 qsos and megaz - lrg . in section [ method ] we present the method used to derive constraints on @xmath7 . section [ result ] contains our main results . we conclude with a discussion in section [ summary ] . the effects of non - gaussianity on the source clustering properties arise because a non - zero @xmath7 affects the halo mass function and enhances the halo clustering on large scales . the second effect is the dominant one on large scales . in the presence of primordial non - gaussianity , the mass function @xmath25 can be written in terms of the gaussian one @xmath26 , for which a good fit to the results of simulations is provided by e.g. , the sheth - tormen formula @xcite , multiplied by a non - gaussian correction factor @xcite : @xmath27~,\end{aligned}\ ] ] where the normalized skewness of the density field , @xmath28 , is @xmath29 , @xmath30 denotes the rms of the dark matter density field linearly extrapolated to @xmath11 and smoothed on scale @xmath31 corresponding to the lagrangian radius of a halo of mass @xmath32 , and @xmath33 is the critical density for ellipsoidal collapse , calibrated on n - body simulations @xcite . for high peaks ( @xmath34 ) and small @xmath7 , @xmath33 is slightly smaller than the critical density for spherical collapse , @xmath35 where @xmath36 is the linear growth factor , and @xmath37 evolves very weakly with redshift . the large - scale halo bias is also modified by the presence of non - gaussianity @xcite : @xmath38 where the factor @xmath39 encloses the scale and halo mass dependence . here , we consider three types of non - gaussianity . in practice , we find that , on large scales , @xmath40 , is independent of the halo mass for the local type . the factor @xmath39 for the enfolded template is proportional to @xmath41 , while the equilateral template gives an almost scale - independent @xmath39 . it is important to note here that the equilateral and enfolded templates were created to have factorizable expressions that gave a good description of the actual bispectra , on average , over all configurations . for the the halo bias effect , nearly squeezed configurations dominate . as discussed e.g. , in ref . @xcite the templates may not necessarily offer a good approximation of the actual bispectra generated by higher - order derivatives or vacuum - state modifications . for all inflationary single - field models the bispectrum scales either as @xmath41 or as @xmath42 in the squeezed limit ( i.e. either like the equilateral template or like the local case ) @xcite . the orthogonal template instead scales like @xmath17 in the squeezed limit . nevertheless it is interesting to consider such an intermediate case : inflationary models that go beyond single field can show a range of scaling in the squeezed limit ( e.g. , ref . @xcite ) . we assume that the large - scale linear halo bias for the gaussian case is @xcite : @xmath43+\frac{2p}{\delta_{\rm c}(z_{\rm f})d(z_{\rm o})}\left\{1+\left[\frac{q\delta^2_{\rm c}(z_{\rm f})}{\sigma^2_{\rm m}}\right]^p\right\}^{-1}~,\end{aligned}\ ] ] where @xmath44 is the halo formation redshift , and @xmath45 is the halo observation redshift . as we are interested in massive haloes , we expect that @xmath46 . here , @xmath47 and @xmath48 account for the non - spherical collapse and are a fit to numerical simulations ( see also refs . @xcite ) . finally , the weighted effective halo bias is given by @xmath49 @xmath50 being the minimum halo mass hosting a source of the kind we are considering . , of the different types of objects considered for our analysis . nvss radio sources ( red dashed line ) , sdss dr6 qsos ( black solid line ) , and sdss dr6 lrgs ( blue dash - dotted line ) . ] in this subsection we describe three lss probes , nvss radio sources , high redshift sdss dr6 qsos and megaz dr7 lrgs sample . in figure [ fig : redshift ] we show their redshift distribution , @xmath51 . all distributions are normalized to unit integral . the nvss @xcite offers the most extensive sky coverage ( @xmath52 of the sky to a completeness limit of about 3 mjy at 1.4 ghz ) . we start by confining our analysis to nvss sources brighter than 10 mjy , since the surface density distribution of fainter sources suffers from declination - dependent fluctuations @xcite . density gradients in the nvss catalog become increasingly unimportant as the source brightness threshold is increased . figure 6 of ref . @xcite , figure 1 of ref . @xcite and figure 3 of ref . @xcite showed variations of nvss source surface density as a function of declination for sources with different flux thresholds . dim sources are strongly affected by density gradients , however this does not appreciably happen for the brighter sources . furthermore , in our previous work we used the nvss sources with different flux thresholds @xmath53 mjy and @xmath54 mjy to compute the acf of nvss source and its ccf with cmb temperature fluctuations and found that the obtained acf and ccf results are stable @xcite . also we mask the stripe @xmath55 , where the catalog may be substantially affected by galactic emissions . here , we have explicitly checked that the choices of masked stripes with @xmath56 and @xmath57 have negligible impact on our final results . in order not to excessively reduce the sample , in our analysis we consider the nvss sources brighter than 10 mjy and the masked strip with @xmath55 to reduce the contamination from galactic emissions . the nvss source surface density at this threshold is @xmath58 and the redshift distribution at this flux limit has been recently determined by ref . their sample , complete to a flux density of 7.2 mjy , comprises 110 sources with @xmath59mjy , of which 78 ( 71% ) have spectroscopic redshifts , 23 have redshift estimates via the @xmath60@xmath61 relation for radio sources , and 9 were not detected in the @xmath60 band and therefore have a lower limit to @xmath61 . we adopt here the smooth description of this redshift distribution given by ref . @xcite , which is shown in the red dashed line of figure [ fig : redshift ] : @xmath62 the mean redshift of this sample is @xmath63 and the redshift range is @xmath64 . the sdss dr6 qso catalog released by ref . @xcite contains about @xmath65 objects with photometric redshifts ranging from @xmath66 to @xmath67 over a total area of @xmath68 ( @xmath69 of the whole sky ) . we refer the reader to ref . @xcite for a detailed description of the object selection with the non - parametric bayesian classification kernel density estimator ( nbc - kde ) algorithm . we use the electronically - published table that contains only objects with the `` good '' flag with values in the range @xmath70 $ ] . the higher the value , the more probable for the object to be a real qso ( see 4.2 of ref . @xcite for details ) . furthermore we restrict ourselves to the `` uvxts=1 '' , i.e. to qsos clearly showing a uv excess which should be a signature of a qso spectrum ( in this case we have @xmath71 qsos ) . as we know , several systematics effects , including the galactic extinction by dust , sky brightness , number of point sources and poor seeing , could potentially affect both the observed auto-/cross - correlation power spectra . in our previous work we carefully checked for their contribution in calculations and in particular we considered extinction and point sources contamination . we found that the dominant systematic effect is the extinction @xcite . therefore , in order to minimize the effect of galactic extinction on the observed qso distribution , we use the extinction mask @xmath72 only in our following analysis @xcite . these masks will remove about @xmath7320% of the considered area . we fit the redshift distribution @xmath74 of the dr6 qso sample with a function of the form @xcite ( black solid line of figure [ fig : redshift ] ) : @xmath75~.\label{reddis}\ ] ] the best - fit values of the parameters are @xmath76 , @xmath77 , @xmath78 ; the mean redshift of the sample is @xmath79 . we use the updated megaz lrg dr7 sample @xcite , which contains @xmath80 galaxies from the sdss dr7 in the redshift range @xmath81 with limiting magnitude @xmath82 . to reduce stellar contamination , there is a variable of the megaz neural network estimator @xmath83 , defined such that @xmath84 if the object is a galaxy , and @xmath85 if it is a star . for a conservative analysis , we choose a cut @xmath86 , which is reported to reduce stellar contamination below 2% while keeping 99.9% of the galaxies . in addition to the sdss dr7 geometry mask , we also add two foreground masks to account for seeing ( removing pixels with median seeing in the red band larger than 1.4 arcsec ) and reddening ( removing pixels with median extinction in the red band @xmath87 ) @xcite . the redshift distribution function ( blue dash - dotted line of figure [ fig : redshift ] ) in this case is found directly from the photometric redshifts that are given in the catalogue @xcite , which is similar to a gaussian function : @xmath88}~,\ ] ] where the mean redshift @xmath89 and the deviation @xmath90 . a two dimensional density field @xmath91 defined over the full sky can be decomposed in a series of spherical harmonics @xmath92 and their corresponding coefficients @xmath93 : @xmath94 with @xmath95 on the full sky spherical harmonics are a complete and orthogonal basis set . however , in practice , a masked ( incomplete ) sky is observed . thus the observed @xmath96 coefficients on the partially cut sky become @xcite : @xmath97 where @xmath98 is the position dependent , weight function imposed by the mask and , optionally , the adopted weighting scheme . from these coefficients we can obtain the observed pseudo power spectrum @xmath99 , the pseudo-@xmath0 estimator @xcite : @xmath100 the pseudo power spectrum @xmath99 , given by the weighted spherical harmonic transform of a masked map , is clearly different from the full sky true power spectrum @xmath101 . the expectation value of @xmath99 is related to the true @xmath101 by a convolution : @xmath102 where the coupling matrix @xmath103 , describing the mode coupling resulting from the weight function @xmath98 @xcite , can be expressed in terms of @xmath104 symbols as : @xmath105 @xmath106 is the angular power spectrum of the weight function : @xmath107 where the weight coefficients @xmath108 are : @xmath109 for the lss catalogues presented above , in practice , we construct the pixelized maps using the healpix software package @xcite with high resolution @xmath110 , yielding pixel areas of @xmath111 and estimate their pseudo power spectrum @xmath112 using the healpix function `` anafast '' . we also subtract the shot noise contribution @xmath113 from the observed pseudo power spectrum , where @xmath114 is the surveyed area and @xmath115 is the observed number of sources . the pseudo power spectrum is measured up to @xmath116 , at these multipoles the magnitude of the estimated error bars is dominated by shot - noise . in order to avoid the effects of gauge corrections on the power spectrum on very large scales @xcite , we set the minimal multipole @xmath117 . the effects of non - gaussianity are large on very large scales . by ignoring @xmath118 we are neglecting the scales where the signal is maximal . however the theoretical predictions for the measured correlations is uncertain on very large scales as discussed in ref . @xcite . in ref . @xcite we estimated that these uncertainties could introduce a systematic error on @xmath7 of @xmath119 . here we take a more conservative approach and simply ignore the largest scales , even if this implies a reduction of the signal - to - noise . in the signal dominated limit , the inverse of the power spectrum covariance matrix @xmath120 is @xcite : @xmath121 where @xmath0 is the theoretical angular power spectrum , @xmath122 denotes the shot noise term . @xmath0 and @xmath123 are the necessary error contributions from both cosmic variance and shot noise , respectively . for the full sky survey , @xmath124 and the covariance matrix becomes diagonal : @xmath125 besides the angular power spectrum of each lss tracer , we also consider the cross - correlation power spectrum among them . in figure [ fig : redshift ] , we show that the redshift distributions of these three tracers overlap . we can thus expect some cross - correlation signal among them . we also use the healpix software to estimate the cross - correlation power spectrum @xmath126 between tracer @xmath127 and @xmath128 : @xmath129 where @xmath130 are the observed coefficients of tracer @xmath131 and they are computed on the common sky area only . consequently , the inverse of the covariance matrix becomes : @xmath132 where the shot noise term @xmath133 is scale independent , @xmath114 is the common area of two surveys , @xmath134 is the observed number of sources in this common area . similarly , we also estimate the cross - correlation power spectrum @xmath135 between the lss number density map and the cmb map . while the planck satellite is operating @xcite , here , we use the the 7-years internal linear combination ( ilc ) map of the cmb provided by the wilkinson microwave anisotropy probe ( wmap ) team @xcite . since an estimate of the uncertainty associated to this map is not provided , we have checked that using other cmb maps ( e.g. the 5 years ilc map by the wmap team ) does not change the results in any significant way . for the cross - check , we also use the v and w band all - sky 7-years cmb maps provided by wmap @xcite and smooth these maps to the resolution of ilc map ( 60 arcmins ) . using these maps , we perform all the calculations again and find that our results are undistinguishable . we adopt the wmap kq75 mask , excluding about 30% of the sky at low galactic latitude , to avoid most of the residual galactic contamination . we still use the estimator of cross - correlation power spectrum @xmath136 : @xmath137 where @xmath138 are the observed coefficients of the cmb map @xmath139 , also computed on the common sky area only . the inverse of the covariance matrix is : @xmath140 where @xmath141 is the theoretical angular power spectrum of cmb map , @xmath142 is measurement error of the cmb temperature power spectrum , in which the main source is the pixel noise which accounts for the imperfections in measured temperature on the sky . pixel noise for wmap is reported as @xmath143 , where @xmath144 is the noise per observation and @xmath145 is the number of observations per given pixel @xcite . after getting the theoretical @xmath146 and @xmath147 , we include the correction of gaussian beam @xmath148 ( using the healpix gaussbeam routine ) : @xmath149 , @xmath150 . we neglect the pixel window function since its contribution at @xmath151 is negligible at the resolution of the wmap ilc map which is @xmath152 arcmins fwhm . we use the publicly available package camb@xmath153sources @xcite to calculate the theoretical power spectrum . in our analysis , we use the halofit @xcite built - in routine for non - linear correction to obtain the fully - evolved , nonlinear matter power spectrum @xmath154 at any epoch . we perform a global fitting of cosmological parameters , including @xmath7 , for the data of section [ data ] including also the datasets described below , using the cosmomc package @xcite , a markov chain monte carlo ( mcmc ) code , modified to calculate the theoretical acf and ccf . we assume purely adiabatic initial conditions and a flat universe , with no tensor contribution to primordial fluctuations . the following six cosmological parameters are allowed to vary with top - hat priors : the dark matter energy density parameter @xmath155 $ ] , the baryon energy density parameter @xmath156 $ ] , the primordial spectral index @xmath157 $ ] , the primordial amplitude @xmath158 \in [ 2.7,4.0]$ ] , the ratio ( multiplied by 100 ) of the sound horizon at decoupling to the angular diameter distance to the last scattering surface @xmath159 $ ] , and the optical depth to reionization @xmath160 $ ] . the pivot scale is set at @xmath161mpc@xmath162 and do not consider massive neutrinos and dynamical dark energy . besides these six basic cosmological parameters , we have two more parameters related to the power spectrum @xmath0 data : the non - gaussianity parameter , the minimal halo mass @xmath50 . when we combine all the different probes together we have four parameters : three minimal halo masses for three lss surveys and the non - gaussianity parameter . the model power spectrum @xmath163 is compared with the observed values @xmath99 , respectively , through the gaussian likelihood function @xcite : @xmath164 where the fisher matrix @xmath120 is the inverse of the power spectrum covariance matrix . while it is true that at very low @xmath165 a gaussian likelihood is not a good approximation for cmb data , at high @xmath165 the gaussian approximation works well even for cmb , for lss it is customary to adopt a gaussian likelihood . the following cosmological data are also included in the fit : @xmath166 power spectra of cmb temperature and polarization anisotropies ; @xmath167 baryonic acoustic oscillations ( baos ) in the galaxy power spectra ; @xmath168 snia distance moduli . to deal with the 7-years wmap ( wmap7 ) cmb temperature and polarization power spectra we use the routines for computing the likelihood supplied by the wmap team @xcite . the wmap7 data are used only to improve the constraints on the six basic cosmological parameters , not to constrain @xmath7 . the baos @xcite can , in principle , measure not only the angular diameter distance , @xmath169 , but also the expansion rate of the universe , @xmath170 . however , the limited accuracy of current data only allows us to determine the ratio between the distance scale defined by ref . @xcite : @xmath171^{1/3},\ ] ] and the comoving sound horizon at the baryon - drag epoch , @xmath172 ( see ref . accurate determinations of the distance ratio @xmath173 , @xmath172 have been obtained by ref . @xcite : @xmath174 we adopt these values as gaussian priors . the snia distance moduli provide the luminosity distance as a function of redshift @xmath175 which provides strong constraints on the dark energy evolution . in this paper we will use the latest snia data sets from the supernova cosmology project , `` union2 compilation '' which consists of 557 samples and spans the redshift range @xmath176 @xcite . in this data set , they improved the data analysis method by using and refining the approach of their previous work @xcite . when comparing with the previous `` union compilation '' , they extended the sample with the supernovae from refs . the authors also provide the covariance matrix of data with and without systematic errors and , in order to be conservative , we include systematic errors in our calculations . in the calculation of the likelihood from snia we marginalize over the nuisance parameter as done in refs . @xcite . furthermore , we add a prior on the hubble constant , @xmath177 km / s / mpc given by ref . @xcite . finally , in the analyses we set the minimal halo mass at @xmath178 consistent with the results of ref . @xcite . in this section , we present constraints on the primordial non - gaussianity parameter @xmath7 from the observed lss power spectra , in combination with the external data sets introduced above : the wmap7 , bao and snia . these extra datasets are only used to constrain the underlying cosmology . in table [ tab : i ] and figure [ fig : fnl ] we show the constraints on the primordial non - gaussianity parameter from different data combinations , after marginalizing over all the other parameters . we also plot the observed power spectra data @xmath179 in figure [ fig : data1 ] and figure [ fig : data2 ] , together with the corresponding theoretical best fit models for comparison . for illustration purposes , here we bin the observed power spectra with the bin size @xmath180 . we begin by considering the constraint on the local type of non - gaussianity @xmath7 from the angular power spectrum only . in the first panel of figure [ fig : data1 ] , we plot the observed nvss angular power spectrum data , which is consistent with the previous work @xcite . when using the nvss @xmath99 data only , we obtain the marginalized constraint on the non - gaussianity parameter : @xmath181 ( @xmath182 ) and @xmath183 ( @xmath21 ) , which is consistent with zero . the minimal halo mass , @xmath184 ( @xmath182 ) , turns out to be remarkably close to our previous work @xcite . we show the best fit model in this panel ( black dashed line ) . the amplitude of the power spectrum on large scales ( small @xmath165 ) is enhanced to fit the data points , by the positive value of @xmath7 . in our previous works @xcite , we used the auto - correlation function ( acf ) of nvss radio sources to constrain @xmath7 and found that the nvss acf yields a positive @xmath7 at more than @xmath21 . the reason for this difference is that we only use the power spectrum data at @xmath185 , and the power spectrum approach does not include the information on the total number of nvss radio sources @xmath186 ( which was folded in the acf analysis ) . if we add a prior on the total number of sources @xmath187 : @xmath188 , we obtain a limit of @xmath189 at @xmath182 . this result is perfectly consistent with previous works @xcite . next , we use the estimated angular power spectrum of the sdss dr6 qso sample , which is shown in the second panel of figure [ fig : data1 ] . there is an obvious excess power at large scales , which favors a positive @xmath7 : @xmath190 ( @xmath182 ) and @xmath191 ( @xmath21 ) . similarly to our previous work @xcite , we also find evidence for positive @xmath7 at more than @xmath21 . however , ref . @xcite , using sdss qso data , obtained non - gaussianity constraints consistent with zero at 95% confidence level . as we discussed before , ref . @xcite used an extension of the dr3 qso sample that include sources that subsequently were released with the dr6 sample : we use the more complete and better calibrated final official sdss dr6 qso catalog release @xcite . @xcite found that their qso sample at @xmath192 ( qso0 ) seem to suffer from contamination which they attribute to systematic calibration errors , discarded this sample from their analysis and only used the high redshift sample ( qso1 ) . we have checked for this effect and find no evidence in the sample we use that the @xmath192 qsos have different contamination than the @xmath193 @xcite . in addition , we remove three narrow stripes in the southern hemisphere of the galaxy and set the minimal multipole @xmath117 , which could reduce the effect of star contamination further . finally , we consider the megaz dr7 lrg sample . @xcite split this sample in four redshift bins according to each source photometric redshift estimate and found a large excess of power over the lowest multipoles ( @xmath194 ) in the angular power spectrum , which is also found in their previous analysis using megaz dr4 lrg sample @xcite . in our analysis , although we do not split the lrg sample in redshift bins , we still find this excess power on large scales , which is shown in the third panel of figure [ fig : data1 ] . there are several explanations for this excess power that range from systematic errors as due to contaminants to new physics @xcite . @xcite checked some possible systematic errors of this megaz lrg sample in details and found that systematics errors do not seem to be responsible for the excess power at large scales . we also perform a similar check and find no significant changes in the power spectrum at large scales . one of the possibilities is that this excess power could be induced by the primordial non - gaussianity . here , we use the observed @xmath99 data of megaz dr7 lrg sample to constrain the primordial non - gaussianity for the first time . we calculate the angular power spectrum on largest scales and account for the redshift distortion power as described in ref . @xcite , although we set the minimal multipole @xmath117 . after marginalizing over other free parameters , we obtain the constraint : @xmath195 which is consistent with zero at @xmath21 . we also obtain a limit of @xmath196 at 68% confidence level . using eq . ( [ eq : nghalobias ] ) , we could obtain the limit of effective halo bias at the mean redshift of the survey : @xmath197 ( @xmath182 ) , which is consistent with other works @xcite , although in our analysis we directly use the full sample . when we combine these three observed angular power spectra data together to constrain non - gaussianity , we obtain a limit of @xmath198 at 68% confidence level and @xmath199 at 95% confidence level , which implies the current angular power spectra of lss tracers favor @xmath15 at the @xmath200 level , since non - gaussianity adds power on large angular scales yielding a good fit to the observed data points . as can be seen from figure [ fig : data1 ] and table [ tab : i ] , there is not a specific measurement that is driving the signal : all the different probes yield very consistent results and when combined the error bar on @xmath7 is reduced to yield a @xmath201 result . comparing with the angular power spectrum , the constraining power of cross - correlation power spectrum on @xmath7 is much weaker @xcite . in the panels of figure [ fig : data2 ] , we plot the observed cross - correlation power spectra @xmath126 among three lss tracers and @xmath136 between the cmb map and three lss tracers . these cross - correlation data points are consistent with zero within @xmath182 error bar but the best fit value @xmath202 is an excellent fit to the data points , as shown in figure [ fig : data2 ] . however , cross - correlation data give a larger error bar for @xmath7 than that from the angular power spectra data and are thus much less constraining . first , we consider the single tracer case in combination with its cross - correlation with the cmb map . when we combine the angular power spectrum and its cross - correlation power spectrum @xmath136 with cmb map to constrain the primordial non - gaussianity , we obtain @xmath203 , @xmath204 and @xmath205 at 68% confidence level for nvss radio sources , sdss dr6 qsos and megaz dr7 lrgs , respectively . the results are consistent with those obtained from the angular power spectrum only , but those cross correlations does not seem to add much information . if we consider all the angular and cross - correlation power spectra data sets of three lss tracers together , . ] we obtain the constraint on the local type primordial non - gaussianity : @xmath206 this result is compatible with previous estimates @xcite , and with the wmap7 limits @xcite . the lss data still implies a positive @xmath7 at more than @xmath21 . in figure [ fig : data1 ] and figure [ fig : data2 ] , we plot the theoretical best fit model @xmath202 of our non - gaussianity calculations ( red solid lines ) . these curves match the observed power spectra very well , especially on large scales . generally , the correction to the standard halo bias due to the presence of primordial non - gaussianity is @xmath207 , where the function @xmath208 is given by @xcite : @xmath209 where @xmath210 is the halo mass which is related to the scale @xmath31 , @xmath211 denotes the expression for the primordial bispectrum of the bardeen potential @xmath9 , @xmath212 its power spectrum and @xmath213 . in the local non - gaussian case , the expression for the correction to the standard gaussian halo bias simplify [ eq . ( [ eq : nghalobias ] ) ] . in this subsection , we explore other two different types of primordial non - gaussianity given by the equilateral and enfolded templates . the equilateral type of non - gaussianity @xcite , which can be generated from the inflationary models with higher - derivative operators of the inflaton , can be well described by the following template @xcite : @xmath214 , where @xmath215^{2/3}+(p_\phi(k_1)^{1/3}p_\phi(k_2)^{2/3}p_\phi(k_3)+5cyc)~.\ ] ] then we find that the correction @xmath208 in the equilateral non - gaussian case is almost scale - independent ( see figure 1 of ref . @xcite ) . in principle there is a mass dependence ( see discussion in ref . @xcite ) , but the template and the actual physical bispectrum yield different effective @xmath7 ( denoted by @xmath216 ) for the halo bias . here , therefore , we consider a phenomenological scale - independent scaling and neglect the mass dependence . the interpretation of the resulting constraint should be interpreted as a constraint on an effective @xmath7 , its connection to the physical bispectrum depends on the detailed model of inflation under consideration . using all datasets , we obtain the constraint on the equilateral type of primordial non - gaussianity : @xmath217 which is consistent with the wmap7 results @xcite . the template of another type of non - gaussianity `` enfolded '' , which arises in models with non - bunch - davies initial state @xcite or in effective field theories of inflation with higher - derivative interactions @xcite ( see also ref . @xcite ) , is given by ref . @xcite : @xmath218 , where @xmath219^{2/3}-(p_\phi(k_1)^{1/3}p_\phi(k_2)^{2/3}p_\phi(k_3)+5cyc)~.\ ] ] in figure 1 of ref . @xcite , it can be seen that the correction in this type is proportion to @xmath220 on large scales . no single field inflationary model has this scaling in the squeezed limit : all inflationary single field models considered so far have only the local or the template equilateral scaling in this limit @xcite . this is nevertheless an interesting intermediate case to consider . in fact non - single field models can yield all the intermediate scalings in the squeezed limit , see e.g. , ref . we therefore consider a phenomenological scaling like @xmath220 which could cover this enfolded template case and neglect the mass dependence . in this case , the whole data sets give the constraint on the effective non - gaussianity parameter for the enfolded template : @xmath221 in the wmap7 paper @xcite , a constraint on the orthogonal type of non - gaussianity @xcite is given . the correction to the halo bias of this template is also proportional to @xmath220 on large scales , but have the opposite sign and a normalization factor of @xmath222 with a weak mass dependence which we neglect here . therefore , based on the result of enfolded template , in our simple analysis , we obtain the constraint on the non - gaussianity parameter for the orthogonal template : @xmath223 ( @xmath182 ) and @xmath224 ( @xmath21 ) . the orthogonal type of non - gaussianity in the squeezed limit however scales like the equilateral one and , as shown in ref . @xcite its effect on the halo bias is almost degenerate with the equilateral non - gaussianity . the enfolded type of non - gaussianity arising for example from modified initial state , in the squeezed limit scales like the local template and as shown in ref . @xcite its effect on the halo bias is almost degenerate with the local non - gaussianity but with a different normalization @xmath225 . this re - scaling can be used to re - interpret the local constraints into constraints for the modified initial state non - gaussianity . finally , we consider the effect of the cubic correction on the halo bias model . in this case , the bardeen potential @xmath9 can be conveniently parameterized up to third order by : @xmath227 where @xmath226 is dimensionless , phenomenological parameter . @xcite explored the effect of a cubic correction @xmath5 on the mass function and halo bias model in detail and gave the expression of the scale - dependent bias correction when @xmath228 and @xmath229 : @xmath230 where the normalized skewness of the density field @xmath231 and the parameter @xmath232 obtained by the fitting in ref . @xcite . then we obtain the constraint on @xmath226 from all the data sets : @xmath233 the result is compatible with that obtained in refs . @xcite . c|c|c datasets & + wmap7+bao+sn & @xmath182 & @xmath21 + + @xmath234(nvss)&@xmath235&@xmath236 $ ] + @xmath237(nvss)&@xmath238&@xmath239 $ ] + @xmath234(qso)&@xmath240&@xmath241 $ ] + @xmath237(qso)&@xmath242&@xmath243 $ ] + @xmath234(lrg)&@xmath244&@xmath245 $ ] + @xmath237(lrg)&@xmath246&@xmath247 $ ] + @xmath234(all)&@xmath248&@xmath249 $ ] + @xmath250(all)&@xmath251&@xmath252 $ ] + + @xmath250(all)&@xmath253&@xmath254 $ ] + + @xmath250(all)&@xmath255&@xmath256 $ ] + + @xmath250(all)&@xmath257&@xmath258 $ ] + and @xmath21 limits on the local type of primordial non - gaussianity @xmath7 obtained from different data combinations . the vertical line denotes the gaussian case @xmath228.[fig : fnl ] ] searches for primordial non - gaussianity can provide key information on the physical mechanism that generate the primordial perturbations and can constrain specific inflationary models . a relatively new and promising method to detect primordial non - gaussianity exploits the fact that the clustering of dark matter halos ( which host massive galaxies and qsos ) is modified on large - scales in a scale - dependent way . this effect is called non - gaussian halo bias . to access these large scales , large - scale structure surveys covering large volumes , high redshift and tracers of massive halos are most suited . among the currently available surveys , the nvss and the sdss qso ( in particular the dr6 catalog ) are the largest . previous work had detected excess large - scale power in their auto - correlation function yielding a detection of a positive value of the non - gaussianity parameter @xmath7 at @xmath259 level . here we have revisited the issue using the angular power spectrum rather than the correlation function . while the two statistics should enclose the same physical information they are affected by systematics effects in different ways : this complementary analysis offers thus a consistency check and a test for possible systematics . we have also used the recently updated megaz - lrg survey : a previous release of the same survey had also been found to have excess large - scale power , but mostly at @xmath118 . in an effort to be as conservative as possible in our analysis , we excuded all multipoles at @xmath118 : these multipoles are affected by theoretical uncertainties and systematics such as calibration , sample contamination etc . these also probe the largest scales , thus these are the multipoles where most of the non - gaussian signal is likely to be concentrated . we combine the survey auto power spectra with the cross - power spectra among the surveys and of each of the survey with the cmb temperature map . in our analysis we marginalize over the lcdm cosmological parameters with a prior imposed by wmap7 angular power spectra , bao , supernovae and hubble constant measurement . we have investigated the constraints on the parameter @xmath7 , characterizing primordial non - gaussianity not only for the popular local shape but also for the equilateral , enfolded and orthogonal templates . we have also considered constraints on a cubic non - gaussianity correction ( parameterized by the @xmath226 parameter ) . we find no evidence for non - gaussianity for any `` shape '' which halo bias effect on large scale scales less steeply than @xmath17 and no evidence for non - zero @xmath226 . however for the local type of non - gaussianity we find a @xmath260 signal for positive @xmath7 when all data - sets are combined . the results are summarized in figure [ fig : fnl ] : no single data set drives the signal and the central @xmath7 values recovered are fully in agreement with previous analyses . due to the more conservative approach taken here ( and thus larger error - bars ) , the only single survey that yield more than @xmath261 detection is the sdss dr7 qso survey . the nvss survey central recovered @xmath7 value is also in agreement with both the qso measurement and previous analyses but the error - bar is larger . the combination of the data sets yields : @xmath262 and @xmath263 at @xmath264 and @xmath21 . thus the tantalizing hint for a positive local @xmath7 found from the auto - correlation function remains despite the different analysis . we have investigated several possible sources of systematic errors such as the sources number density and the contamination by galactic emissions for nvss sources ; different choices for the cmb temperature fluctuations templates ; possible contamination of stars for the sdss and lrg samples . our main results are stable when these systematic errors are considered . such a result would have profound implications for inflationary mechanisms . the improved statistical power of forthcoming large - scale structure surveys is needed to fully resolve the issue . moreover , it will be also important to use the cmb data which will be provided by planck . we acknowledge the use of the legacy archive for microwave background data analysis and the healpix package . numerical analysis has been performed at the university of cambridge high performance computing service ( http://www.hpc.cam.ac.uk/ ) . this research has been partially supported by the asi contract no . i/016/07/0 cofis , the asi / inaf agreement i/072/09/0 for the planck lfi activity of phase e2 and a prin miur . mv is supported by asi / aae , a prin miur , a prin inaf 2009 and the erc - stg `` cosmoigm '' . support was given by the italian space agency through the asi contracts euclid - ic ( i/031/10/0 ) . lv acknowledges support from grant fp7 erc - ideas phys.lss 240117 and micinn grant aya2008 - 03531 . lv and sm thank the centro de ciencias de benasque pedro pascual where stimulating discussions about this work took place . funding for the sdss and sdss - ii has been provided by the alfred p. sloan foundation , the participating institutions , the national science foundation , the u.s . department of energy , the national aeronautics and space administration , the japanese monbukagakusho , the max planck society , and the higher education funding council for england . the sdss web site is http://www.sdss.org/. the sdss is managed by the astrophysical research consortium for the participating institutions . the participating institutions are the american museum of natural history , astrophysical institute potsdam , university of basel , university of cambridge , case western reserve university , university of chicago , drexel university , fermilab , the institute for advanced study , the japan participation group , johns hopkins university , the joint institute for nuclear astrophysics , the kavli institute for particle astrophysics and cosmology , the korean scientist group , the chinese academy of sciences ( lamost ) , los alamos national laboratory , the max - planck - institute for astronomy ( mpia ) , the max- planck - institute for astrophysics ( mpa ) , new mexico state university , ohio state university , university of pittsburgh , university of portsmouth , princeton university , the united states naval observatory , and the university of washington .

entanglement , nonlocality and quantum discord ( qd ) are all different but somehow related aspects of the quantumness of the correlated composite quantum systems . entanglement is a kind of quantum correlation and determines whether the given state is separable or not @xcite . on the other hand , the quantum systems may contain quantum correlations other than entanglement . quantum discord , which is defined as the difference between the quantum versions of two analogous classically equivalent definitions of mutual information , captures all types of nonclassical correlations including entanglement @xcite . although entanglement and qd are equivalent for pure states @xcite , qd is found to be nonzero for some separable states @xcite . such states were shown to be useful in the deterministic quantum computations with one pure qubit context @xcite . therefore qd is believed a new resource for quantum computation . the nonlocality as measured by the violation of bell s inequalities is also a signature of the inseparability of a quantum state and signifies entangled states whose correlations can not be reproduced by a classical local hidden - variable model @xcite . however it must be emphasized here that the presence of entanglement does not always imply violation of bell s inequalities , while the violation of bell inequalities implies entanglement . such entangled states violating bell s inequalities play a central role in some applications in quantum information science @xcite , such as to guarantee the safety of device - independent key distrubition protocols in quantum cryptography @xcite . the relations among different measures of quantum correlations and their dynamical behaviour have been an active area of research with an aim to understand the fundamental question of what quantumness is as well as to characterize these correlations as useful resources for various quantum tasks @xcite . one of the most important obstacle to realizing quantum operations is the interaction between the qubits and the environment which tends to wipe out the quantumness of the system @xcite . characterization of the effect of environment on the system such as decoherence and dissipation is important . the dynamics of quantum correlations such as entanglement , quantum discord and bell nonlocality for a two - qubit system subject to various environments have been studied by many groups recently @xcite . entanglement sudden death ( esd ) is one of the new concepts that arise within the context of these studies @xcite . esd refers to death of entanglement between two qubits in a finite time while the coherence of the single qubit decays only exponentially . in contrary to entanglement , qd is more robust in a decohering environment , actually it was shown that almost all states have non - zero quantum discord @xcite . qd was shown to decay exponentially in markovian environments , even at finite temperatures @xcite and can become zero momentarily under non - markovian dynamics @xcite . qd was also experimentally investigated in an all - optical set up recently @xcite . furthermore , qd and entanglement were analyzed for a heisenberg spin chain with quantum phase transitions and qd was found to detect the critical points of the transition at finite temperatures while entanglement could not @xcite . very recently , sun _ _ investigated the induced quantum correlations between two qubits coupled to a common environment modelled by ising spin chain in a transverse field and during the time evolution , non - zero quantum discord is found to be created for specific initial states whereas entanglement can not be induced @xcite . in ref . @xcite quantum discord is found to be totally unaffected by non - dissipative noise for a long time interval under certain conditions , while entanglement decays to zero quickly . these remarkable properties of quantum discord makes it more practical than the entanglement to characterize quantum correlations . as mentioned , the nonlocality is curicial for some applications in quantum information . therefore , many theoretical studies have been devoted to the comparision of entanglement and the violations of bell inequalities @xcite . it was demonstrated by several authors that the survival time for entanglement is much larger than that of the bell - inequality violation @xcite . in ref . @xcite , the nonlocal entanglement identified by the violations of a bell inequality is found to be preserved during the time evolution of a system consisting of two qubits which are embedded independently in zero - temperature bosonic reservoirs and are initially in an entangled mixed state . _ showed that bell inequality might not be violated for a state with high entanglement for a two - qubit system subject to amplitude damping @xcite . moreover , the nonlocal entanglement between two qubits subject to classical dephasing environment modelled as ornstein - uhlenbeck noise has been investigated in refs . @xcite , very recently , and it was demonstrated that the strong non - markovian effect can protect the nonlocal entanglement @xcite . additionally , the nonlocal entanglement induced by the qubit - qubit interaction is found to be protected against the sudden death despite the strong influence of the environmental noise for a certain class of initial states regardless of whether the environment is markovian or non - markovian @xcite . although the dynamical relation between the entanglement and quantum discord as well as entanglement and bell nonlocality for a bipartite system subject to different environments have been investigated by a large number of groups as mentioned above , further investigations would help in clarifying the relation of these three correlation measures . along these lines , in the present paper , we have investigated the dynamics of quantum correlations such as quantum discord , entanglement and bell nonlocality between two qubits in a lossy cavity . very recently , entanglement , bell nonlocality and quantum discord have been studied in cavity qed @xcite . most of them assume identical atoms , or the atom - cavity resonant case , or only one excitation . here , we have investigated the effect of atom - cavity detuning and cavity decay rate on the corresponding correlation measures for a system which is initially prepared in separable or maximally entangled state . the most important results of the present study are : depending on the initial preparation of the system and the type of the detunings between the frequency of the cavity mode and the transition frequency of the qubits , the induced quantum discord is found to increase for a long time under cavity decay loss mechanism for certain conditions , while there is no increase in entanglement and bell nonlocality . in some situations , the quantum discord is induced even while the atoms remain separable at all times . also , the preservation of induced entanglement and quantum discord was observed under strong dissipation for certain type of initial states . the paper is organized as follows . in sec . [ model ] , the model is introduced and the exact differential equations governing the dynamics of two atoms and single cavity mode are derived . the correlation measures ; concurrence , chsh - bell inequality and quantum discord are also introduced , briefly . in secs . [ cavity1 ] and [ cavity2 ] , the effect of cavity damping on the corresponding correlation measures are revealed for different detunings for the atoms initially prepared in separable or maximally entangled state . the summary of the important results is given in sec . [ conclusion ] . here we consider two two - level atoms with excited states @xmath0 and ground states @xmath1 interacting with a single - mode cavity field of frequency @xmath2 . the hamiltonian of this system is given as @xcite : @xmath3 where @xmath4 and @xmath5 are the transition frequencies for the atoms @xmath6 and @xmath7 , @xmath8 and @xmath9 are the annihilation and creation operators for the field and @xmath10 and @xmath11 are the coupling constants to the cavity mode for the atoms @xmath6 and @xmath7 , respectively . when maximum two - photon is considered to be present in the cavity ( i.e. , the total excitation is equal to two ) and if there is no dissipation in the cavity modes , the closed form of the total state vector at any time may be written as @xcite : @xmath12 where @xmath13 @xmath14 are the time - dependent probability amplitudes . for example , @xmath15 represents that the atoms are in their ground states while there are two photons inside the cavity . for this case , the dynamics of the atoms and single mode cavity field can be solved by using time - dependent schrdinger equation and recently many studies have been reported for this case @xcite . on the other hand , when cavity decay is taken into account , since the cavity decay only changes the number of photons inside the cavity , the atoms and the cavity mode can be found in any of the states @xcite : @xmath16 . then the density matrix of the system of two atoms interacting with a single mode cavity field will evolve according to the master equation @xcite : @xmath17-\frac{\kappa}{2}\left(a^{\dagger}a\rho-2a\rho a^{\dagger}+\rho a^{\dagger}a\right),\end{aligned}\ ] ] where @xmath18 and @xmath19 @xmath20 are the spin inversion operators . here the first term in eq . ( [ mastereqn ] ) is generated by the hamiltonian given in eq . ( [ hamiltonian ] ) and represents the coherent evolution , while the other term denotes the incoherent dynamics ; @xmath21 is the rate of loss of photons from the cavity mode . by setting @xmath22 , @xmath23 , @xmath24 , @xmath25 , @xmath26 , @xmath27 , @xmath28 , @xmath29 as a basis and considering the eqs . ( [ hamiltonian ] ) and ( [ mastereqn ] ) , the differential equations governing the dynamics of two atoms and the single mode cavity field can be obtained easily . in our analysis we will restrict ourselves for the initial states : @xmath30 and @xmath31 , where @xmath32 . one should note for these initial states the diagonal elements and some of the off - diagonal elements form a closed evolution . they can be listed as : @xmath33 where @xmath34 and @xmath35 are called the detunings and describes the frequency difference between the energy levels of atoms and the cavity mode , @xmath36 and @xmath37 . we will consider two particular detuning types : for @xmath38 the atoms will be called identical since the transition frequencies of the atoms are the same and the atoms will be called unidentical for @xmath39 case for which the individual atom transition frequencies are different @xcite . in practice , the detunings might be slightly different due to broadening , in this work we will consider relatively large values of @xmath40 . one should note that it is quite difficult to obtain an analytic solution for this set of coupled differential equations , therefore we will solve these equations numerically for the considered initial states . the density matrix of the two atoms is obtained by taking a partial trace over the cavity degrees of freedom : @xmath41 and used to calculate the quantum correlations as measured by chsh - bell inequality , concurrence and quantum discord between two atoms interacting with a single mode cavity field . for the considered initial states whose dynamics is given by eq . ( [ solmaster ] ) , the reduced density matrix of the atoms in the two - qubit standart basis @xmath42 can be calculated as : @xmath43{cccc } a & 0 & 0 & 0 \\ 0 & b & e & 0 \\ o & e^ * & c & 0 \\ 0 & 0 & 0 & d \end{array}\right)\end{aligned}\ ] ] where @xmath44 and @xmath45 . it should be noted that the density matrix in eq . ( [ reducedden ] ) has an x - form and the dynamics preserve its form . the considered correlation measures for this type of states can be calculated easily @xcite . the concurrence @xcite as an entanglement measure is given by @xcite : @xmath46 the bell nonlocality measured by the chsh - inequality @xcite can be found as @xcite : @xmath47 where @xmath48 and @xmath49 . qd is defined as @xcite : @xmath50 where @xmath51 measures the total correlation between the atoms ; @xmath52 is the von - neumann entropy and @xmath53 is the reduced density matrix obtained by tracing @xmath54 over the subsystem @xmath55 . the other quantity @xmath56 is the amount of classical correlations between the atoms defined as the maximum information one can get about the atom @xmath6 by performing measurements on the atom @xmath7 , or vice versa . its calculation requires maximization over one of the subsystem and generally obtaining an analytic result is not an easy task @xcite . recently , for the density matrix having x - form the explicit expressions of qd and classical correlation are reported @xcite . we have used the expressions given in ref . the classical correlation @xmath56 for the density matrix eq . ( [ reducedden ] ) is given as : @xmath57 where @xmath58 . and qd is given as : @xmath59 where @xmath60 , with @xmath61 being the eigenvalues of @xmath54 , @xmath62 , @xmath63 , @xmath64 and @xmath65 . we have also calculated qd numerically by considering projective measurements @xcite and observed that the results obtained by using eq . ( [ qd2 ] ) and the numerical results agree perfectly . as is well known , concurrence is equal to 1 ( 0 ) for a maximally entangled ( separable ) state . although qd is equal to 1 for a maximally entangled state , it may or may not be equal to zero for a separable state because it was shown that even some separable states can carry non - zero quantum discord @xcite . for pure states qd is found to be equal to the entanglement of formation @xcite . on the other hand when @xmath66 the chsh - bell inequality is violated which signifies that the correlations can not be accessible by any classical local model @xcite . in this section , we will examine the effect of cavity decay rate , @xmath21 , on the quantum correlations as measured by concurrence , chsh - bell inequality and quantum discord . to do this we will solve the differential equations in eq . ( [ solmaster ] ) numerically for the initial states : @xmath30 and @xmath31 , where @xmath32 and we will set @xmath67 ( symmetric coupling ) , @xmath68 for identical atoms and @xmath69 for unidentical atoms . one should note that the detunings are chosen somewhat large because at higher detunings the qubit - cavity energy exchange is low @xcite and a high - degree of cavity - induced quantum correlations can be created @xcite . quantum discord versus @xmath70 for unidentical atoms ( @xmath69 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions entanglement and bell nonlocality are not induced , thus these are not plotted here and the inset in ( c ) represents the total @xmath77 and classical @xmath56 correlations versus @xmath70 for @xmath74.,title="fig:",width=283 ] quantum discord versus @xmath70 for unidentical atoms ( @xmath69 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions entanglement and bell nonlocality are not induced , thus these are not plotted here and the inset in ( c ) represents the total @xmath77 and classical @xmath56 correlations versus @xmath70 for @xmath74.,title="fig:",width=283 ] quantum discord versus @xmath70 for unidentical atoms ( @xmath69 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions entanglement and bell nonlocality are not induced , thus these are not plotted here and the inset in ( c ) represents the total @xmath77 and classical @xmath56 correlations versus @xmath70 for @xmath74.,title="fig:",width=283 ] quantum discord versus @xmath70 for unidentical atoms ( @xmath69 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions entanglement and bell nonlocality are not induced , thus these are not plotted here and the inset in ( c ) represents the total @xmath77 and classical @xmath56 correlations versus @xmath70 for @xmath74.,title="fig:",width=283 ] quantum discord versus @xmath70 for unidentical atoms ( @xmath69 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions entanglement and bell nonlocality are not induced , thus these are not plotted here and the inset in ( c ) represents the total @xmath77 and classical @xmath56 correlations versus @xmath70 for @xmath74.,width=283 ] we first consider the dynamics of concurrence and quantum discord for initially separable state @xmath71 and display @xmath78 and @xmath79 versus dimensionless time , @xmath70 in figs . 1 and 2 for unidentical and identical atoms , respectively . for this initial state neither non - zero concurrence nor violation in bell inequality ever occurs for unidentical qubits , so their dynamics are not plotted in fig . 1 . however , this does not indicate the loss of quantum correlations ; qd induced by the atom - field interaction oscillates nearly between 0 and 0.18 for @xmath72 ( fig . a similar result is found by sun _ et al . , _ in ref . @xcite where the authors studied the entanglement and quantum discord between two qubits which are independently coupled to a common environment modelled as an ising spin chain . it can be seen from the analysis of figs . 1(b)-1(e ) that qd is created independently of the cavity decay rate , but its maximum value depends on the decay rate @xmath21 inversely and the effect of cavity decay on the dynamics of qd is to create a damped oscillatory behaviour ; damping in qd is proportional to the cavity decay rate . also , in order to understand the nature of non - zero quantum discord , we have plotted the total correlation , @xmath77 , and classical correlation , @xmath56 , for @xmath74 in the inset of fig . it should be noted that for this initial state , the total as well as classical correlations are also induced and @xmath77 is always greater than @xmath56 as a consequence non - zero quantum discord is present . this inset also demonstrates that @xmath77 and @xmath56 are also damped oscillatorily as @xmath80 . correlations in the system are gradually lost to the reservoir because of the cavity decay as found in the system studied in ref . @xcite . quantum discord and concurrence ( insets in ( a ) , ( b ) and ( c ) ) versus @xmath70 for identical atoms ( @xmath68 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions bell nonlocality is not induced , and for @xmath81 and @xmath76 , @xmath82 has no dynamics as well , thus they are not plotted here . the insets are plotted under the same conditions as @xmath80.,title="fig:",width=283 ] quantum discord and concurrence ( insets in ( a ) , ( b ) and ( c ) ) versus @xmath70 for identical atoms ( @xmath68 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions bell nonlocality is not induced , and for @xmath81 and @xmath76 , @xmath82 has no dynamics as well , thus they are not plotted here . the insets are plotted under the same conditions as @xmath80.,title="fig:",width=283 ] quantum discord and concurrence ( insets in ( a ) , ( b ) and ( c ) ) versus @xmath70 for identical atoms ( @xmath68 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions bell nonlocality is not induced , and for @xmath81 and @xmath76 , @xmath82 has no dynamics as well , thus they are not plotted here . the insets are plotted under the same conditions as @xmath80.,title="fig:",width=283 ] quantum discord and concurrence ( insets in ( a ) , ( b ) and ( c ) ) versus @xmath70 for identical atoms ( @xmath68 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions bell nonlocality is not induced , and for @xmath81 and @xmath76 , @xmath82 has no dynamics as well , thus they are not plotted here . the insets are plotted under the same conditions as @xmath80.,title="fig:",width=283 ] quantum discord and concurrence ( insets in ( a ) , ( b ) and ( c ) ) versus @xmath70 for identical atoms ( @xmath68 ) and @xmath71 initial state for the decay rates : ( a ) @xmath72 , ( b ) @xmath73 , ( c ) @xmath74 , ( d ) @xmath75 and ( e ) @xmath76 . note that under these conditions bell nonlocality is not induced , and for @xmath81 and @xmath76 , @xmath82 has no dynamics as well , thus they are not plotted here . the insets are plotted under the same conditions as @xmath80.,width=283 ] we have calculated the dynamics of chsh inequality , concurrence and quantum discord for the same initial state @xmath71 for the identical atoms ( @xmath68 ) case to understand the effect of the type of detuning for the same leaky cavity and plot the resulting @xmath79 and @xmath78 as a function of @xmath70 in figs . 2(a)-2(e ) and insets of figs . 2(a)-2(c ) , respectively . as can be seen from the insets of figs . 2(a)-2(c ) , non - zero entanglement , although small , is created which decays to zero as cavity decay rate is increased and the induced entanglement has no dynamics for higher values of @xmath21 ( @xmath81 and @xmath76 ) . bell inequality is found to be not violated for any value of @xmath21 , which is expected because entanglement is low and generally bell inequality violation is shown to be a hallmark of high @xmath82 @xcite , so @xmath83 is not displayed in fig . 2 . one should note the difference in the dynamics of qd in the leaky cavity for unidentical and identical atom cases ( figs . 1 and 2 ) . while qd has an oscillatory decay type dynamics for the unidentical atoms ( fig . 1 ) , it increases until the time @xmath84 and then decays for the identical atoms . also , for the same decay parameters , qd is found to have a much longer lifetime for the identical atoms . the higher leakage in the cavity decreases @xmath85 and the lifetime of qd after the time @xmath84 . on the other hand , @xmath21 has no appreciable effect on the maximum value of @xmath80 . also note that the fast oscillations in qd are gradually smoothed for larger values of @xmath21 and longer times as can be seen from figs . 2(b)-2(e ) . chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] chsh inequality ( left plots ) , concurrence ( middle plots ) and quantum discord ( right plots ) versus @xmath70 for @xmath86 initial state and @xmath72 ( ( a ) , ( b ) , ( c ) ) , @xmath73 ( ( d ) , ( e ) , ( f ) ) , @xmath74 ( ( g ) , ( h ) , ( k ) ) , @xmath81 ( ( l ) , ( m ) , ( n ) ) and @xmath76 ( ( o ) , ( p ) , ( r ) ) . here the main subfigures are for identical atoms with @xmath68 , while the insets are for unidentical atoms with @xmath69 under the same decay rate parameters . note that for unidentical atoms @xmath83 is not violated , so they are not reported here.,title="fig:",width=192 ] now we consider another initial separable state of the form @xmath87 which has one excitation in the field and the other excitation in one of the atoms to see the effect of initial partition of the excitation on the dynamics of the quantum correlations . the time dependent of chsh - inequality , concurrence and quantum discord are displayed in the first , second and third columns of fig . 3 , respectively . in these figures , insets show the dynamics for unidentical atoms ( @xmath69 ) while main subfigures are for the identical atoms ( @xmath68 ) . inspection of subfigures demonstrates the pronounced effect of qubits being identical or unidentical . for identical atoms , the atom - field interaction induces high values of @xmath82 and @xmath80 and bell inequality is also violated at the high values of entanglement for a high - quality cavity ( fig . 3(a ) , 3(b ) and 3(c ) ) . in the case of unidentical atoms , @xmath82 and @xmath80 are also created , but their maximum are small ; 0.018 and 0.13 , respectively , and @xmath83 is not violated . when unidentical atoms interact with a low - quality cavity , entanglement can cease to exist very quickly and qd shows a fast oscillatory decaying behaviour ( see the insets in figs . 3(e)-3(r ) ) which is very similar to what we have found for @xmath71 initial state . on the other hand , this situation is completely and significantly different in the case of identical atoms . although chsh - bell inequality goes below 2 very quickly for @xmath88 and no longer is violated for @xmath89 , the dynamics of entanglement and quantum discord are very similar that both approach to constant values in an oscillatory manner ; the oscillations die down as the cavity decay rate is increased . in effect , cavity decay entangles initially separable state and the resulting entanglement as well as quantum discord are robust . in many studies , the dissipation of cavity fields were found to play a constructive role in the generation of quantum correlations as measured by concurrence and quantum discord @xcite . this situation can be understood by looking at the steady - state of the dynamics given in eq . ( [ mastereqn ] ) . for @xmath90 , @xmath91 and @xmath92 , the @xmath93 equation has a solution independent of @xmath21 and @xmath40 and the atom - atom reduced density matrix is given as : @xmath94 . the concurrence for this state is 0.50 and quantum discord of it can be calculated from eq . ( [ qd2 ] ) to be 0.412 which are seen clearly from figs . 3(p ) and 3(r ) . also , we want to emphasize that the dynamics of @xmath95 , @xmath78 and @xmath79 between identical atoms for the initial state @xmath96 is the same as that for @xmath97 , since we assume symmetric coupling ( i.e. , @xmath98 ) between the atoms and the single mode cavity field . the populations @xmath99 @xmath100 and the absolute value of the coherences @xmath101 given by eq . ( [ solmaster ] ) versus @xmath70 for the initial state @xmath86 and @xmath76 for identical ( @xmath68 ) ( a ) and unidentical ( @xmath69 ) ( b ) atoms . ( c ) shows the atomic purity ( @xmath102 ) for the density matrix eq . ( [ reducedden ] ) versus @xmath70 for identical and unidentical atoms . here the insets are for small time regions.,title="fig:",width=283 ] the populations @xmath99 @xmath100 and the absolute value of the coherences @xmath101 given by eq . ( [ solmaster ] ) versus @xmath70 for the initial state @xmath86 and @xmath76 for identical ( @xmath68 ) ( a ) and unidentical ( @xmath69 ) ( b ) atoms . ( c ) shows the atomic purity ( @xmath102 ) for the density matrix eq . ( [ reducedden ] ) versus @xmath70 for identical and unidentical atoms . here the insets are for small time regions.,title="fig:",width=283 ] the populations @xmath99 @xmath100 and the absolute value of the coherences @xmath101 given by eq . ( [ solmaster ] ) versus @xmath70 for the initial state @xmath86 and @xmath76 for identical ( @xmath68 ) ( a ) and unidentical ( @xmath69 ) ( b ) atoms . ( c ) shows the atomic purity ( @xmath102 ) for the density matrix eq . ( [ reducedden ] ) versus @xmath70 for identical and unidentical atoms . here the insets are for small time regions.,width=283 ] to further elucidate the asymptotic constant entanglement and quantum discord for the initial state @xmath86 , we display the time dependence of the density matrix elements of the atom - field system given by eq . ( [ solmaster ] ) and the purity of the atomic subsystem for the density matrix eq . ( [ reducedden ] ) for identical and unidentical detuning cases in figs . 4(a)-4(c ) for @xmath76 . as can be seen from fig . 4(a ) , all populations , except @xmath103 and @xmath104 , become zero for a short time while @xmath104 goes to @xmath105 and @xmath106 go to @xmath107 asymptotically . the key mechanism for the creation of the entanglement and qd in this case seems to be the coherence @xmath108 generated by the interaction of the atom with the dissipative cavity . the absolute value of @xmath108 also reaches to 1/4 asymptotically as can be seen from the fig . the evolved state is a trapping state which has the property that the probabilities of the different decay channels interfere destructively and the state is robust against the cavity loss mechanism @xcite . generation or protecting entanglement by population trapping has been discussed by a number of groups before @xcite . the magnitude of @xmath21 has no effect on the long time limit of @xmath103 and @xmath104 , it only affects how long other population components can be nonzero . the dynamics of atomic purity ( which is equal to 1 for a pure state and 1/4 for a maximally mixed state in four dimensional hilbert space ) displayed in fig . 4(c ) as the solid line for @xmath68 indicates that the state approaches to a partially mixed state which has high entanglement and quantum discord as indicated before in figs . 3(p ) and 3(r ) . for unidentical atoms , the short time dynamics is similar to the identical detuning case that all states are populated for a short time , but the asymptotic state is now @xmath109 as can be seen from fig . 4(b ) . the atomic purity ( dotted line in fig . 4(c ) ) indicates that the state of the atoms firstly goes to a mixed state in a short time and then approaches to the pure state @xmath110 which is the reason of the loss of correlations between unidentical atoms in the asymptotic limit . in this subsection , we will investigate the effect of initial correlations in the atom - atom part of the state on the dynamics of quantum correlations . chsh inequality ( a ) , concurrence ( b ) and quantum discord ( c ) versus @xmath70 for unidentical atoms , @xmath69 and @xmath31 where @xmath32 . here all subfigures include plots for @xmath72 , @xmath74 , @xmath81 and @xmath76.,title="fig:",width=283 ] chsh inequality ( a ) , concurrence ( b ) and quantum discord ( c ) versus @xmath70 for unidentical atoms , @xmath69 and @xmath31 where @xmath32 . here all subfigures include plots for @xmath72 , @xmath74 , @xmath81 and @xmath76.,title="fig:",width=283 ] chsh inequality ( a ) , concurrence ( b ) and quantum discord ( c ) versus @xmath70 for unidentical atoms , @xmath69 and @xmath31 where @xmath32 . here all subfigures include plots for @xmath72 , @xmath74 , @xmath81 and @xmath76.,width=283 ] we take @xmath111 as the initial state which is the maximally entangled bell state for the atomic system . qd and bell inequality violation are also maximal for this initial state . we display @xmath95 , @xmath78 and @xmath79 as a function of @xmath70 for several cavity decay rates for unidentical ( fig . 5 ) and identical ( fig . 6 ) atoms . at @xmath72 , all correlations show oscillatory behaviour ; for the unidentical atoms @xmath83 , @xmath82 and @xmath80 oscillate nearly in the ranges [ @xmath112 , 2.2 ] , [ 1.0 , 0.63 ] and [ 1.0 , 0.58 ] , respectively ( fig . 5 ) . on the other hand , @xmath83 and @xmath82 for @xmath72 have periodic sudden deaths and births for the identical detunings of the atoms ( figs . 6(a ) and 6(b ) ) , while @xmath80 oscillates nearly between 0.3 and 1 ( fig . 6(c ) ) . for small values of @xmath21 @xmath113 , the effect of cavity decay on the dynamics of correlations can be described as underdamped oscillations for both identical and unidentical detuning cases . as @xmath21 is increased to @xmath114 , the oscillation becomes overdamped and finally at large @xmath21 s ( @xmath76 ) damping becomes critical , i.e. @xmath82 and @xmath80 show exponential decay for the unidentical atom cases . for the identical atoms at @xmath81 and @xmath76 concurrence has sudden death , while qd decays exponentially . one should note that bell nonlocality is the most fragile quantum correlation considered here ; it suffers sudden death in a short time as long as @xmath91 . chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] chsh inequality ( a ) and ( d ) , concurrence ( b ) and ( e ) and quantum discord ( c ) and ( f ) versus @xmath70 for identical atoms , @xmath68 and @xmath31 where @xmath32 . here ( a ) , ( b ) and ( c ) are for @xmath72 , while ( d ) , ( e ) and ( f ) include plots for @xmath74 , @xmath81 and @xmath76.,title="fig:",width=192 ] we have investigated the dynamics of quantum correlations as quantified by entanglement , bell nonlocality and quantum discord for two qubits embedded in a common leaky cavity for three different initial states with double excitations . the qubits were considered as identical or unidentical based on the detuning of their transition frequency with respect to the frequency of the field mode . the main findings of the study can be summarized as : qubits being identical or not have profound influence on the dynamics of the considered correlations ; starting from the same initial state , the correlations show very different behaviour depending on whether the detunings of the individual qubits satisfy @xmath92 or @xmath115 . for unidentical qubits initially in their excited states and interacting with a high or low quality cavity , quantum discord is found to be induced for all considered cavity decay rates and it decays oscillatorily for a leaky cavity although the qubits remain unentangled for all times . for the same initial state and the leaky cavity , qd is found to increase in the long time for identical qubits , while entanglement is either `` never created '' or decays to zero in a short time . one of the most interesting finding is the constant high entanglement and quantum discord state which is obtained as the long - time limit of the dynamics of @xmath87 initial state for the identical qubit case . here the final state is independent of the cavity decay rate ( if @xmath91 ) and the magnitude of the detuning and has high concurrence ( @xmath116 ) and quantum discord ( @xmath117 ) . from an analysis of the time dependence of the density matrix elements it is seen that the considered initial state leads to a trapping state which happen to have high degree of quantum correlations . on the other hand , for the same initial state , the correlations between the unidentical atoms are found to be destroyed quickly even for a small cavity decay rate . for the atomic bell - type initial state we have found esd for the identical qubits interacting with a low quality cavity , while concurrence decays exponentially for the unidentical qubits . quantum discord is found to have asymptotical decay behaviour independent of the type of atoms for this initial state and leaky cavity . on the other hand , chsh - bell inequality violations survive only for a finite time for @xmath91 for both identical and unidentical qubits . it is interesting to note that steady - state nonzero correlation state of @xmath118 ( or @xmath97 ) is not observed for the @xmath111 . this shows that the stationary states depend strongly on the initial conditions @xcite . we have also demonstrated that bell nonlocality as manifested by chsh inequality is the most fragile correlation measure under dissipative environment and is violated only at the high values of entanglement .

in this article we study the allen - cahn / cahn - hilliard model ( ac - ch model for short ) , which combines and extends two famous diffuse interface models : the allen - cahn equation and the cahn - hilliard equation . both have been applied successfully to model segregation , precipitation and phase change phenomena in alloys and liquid mixtures in materials science , geology , physics , and biology , among others . the ac - ch system was first introduced in @xcite , and its mathematical properties have been studied extensively , see , e.g. , @xcite and references therein . we study here an extension of this model with a particular ansatz for the elastic energy . the allen - cahn equation was first introduced without elasticity in @xcite ; and the cahn - hilliard equation was first introduced without elasticity in @xcite . the allen - cahn equation is a second order partial differential equation of ginzburg - landau type for an _ unconserved _ order - parameter @xmath3 and thus can be used to model segregation and precipitation in solids , or other more general situations where a reordering of the crystal lattice occurs . conversely , the cahn - hilliard equation is a fourth - order partial differential equation for a _ conserved _ order parameter @xmath4 striving to model phase change phenomena where one physical quantity like the volume of the phases is preserved . throughout the paper we set @xmath5 as this plays a special role . the allen - cahn system with linear elasticity was studied before in @xcite , the cahn - hilliard system with linear elasticity in @xcite , @xcite and @xcite . an extension of the cahn - hilliard system with geometrically linear elasticity valid for single crystals was recently found in @xcite for @xmath6 . except for the latter work , elastic effects due to small scale microstructures within the phases have been neglected . here we treat the combined ac - ch model with elasticity in @xmath7 dimensions . we state the new extended model in in section [ secacch ] . this model provides a basis for the generalization of further existing isothermal diffuse interface models . our definition of the extended model does not require any special assumptions ( except regarding the regularity ) on the stored energy functional @xmath8 . however , for the proof of existence of weak solutions , we require the assumption ( @xmath9 ) phrased in section [ secexist ] below . we show that assumption ( @xmath9 ) is satisfied in the following cases : * materials which follow the linear theory of elasticity developed by eshelby @xcite in the context of elastic inclusions and inhomogeneities ( see a remark after assumption ( @xmath9 ) ) ; * materials in @xmath6 which are well - described by a geometrically linear theory of elasticity ( theorem [ thmcor1 ] , for @xmath10 , cf . also @xcite ) ; * materials in @xmath1 which are well - described by a geometrically linear theory of elasticity , where the deformations are assumed to be scalar - valued functions . this corresponds to the situation of anti - plane shear , see below for details and cf . theorem [ thmcor1 ] . the geometrically linear theory of elasticity allows for fine microstructures within the phases of the elastic materials taken into account . this becomes apparent in the explicit formulas for the relaxed energy functionals , which are , together with formulas for their derivatives , the basis of our proof of theorem [ thmcor1 ] . in section [ secexpl ] we recall an explicit formula in @xmath10 derived by chenchiah and bhattacharya @xcite . moreover we prove explicit formulas for @xmath2 , and for @xmath1 in the setting of scalar - valued deformations . ( in @xcite partial results for @xmath1 in the vectorial , i.e. the non - scalar , setting are shown . ) the coupling to elasticity changes significantly the morphology of the precipitates and the coarsening patterns , see , e.g. , the classification in @xcite . for an intuitive picture of the microstructures taken into account we refer the reader to figures [ fig : intuitivemacro ] and [ fig : intuitivemicro ] . macro.pstex_t micro2.pstex_t micro1.pstex_t figure [ fig : intuitivemacro ] displays exemplary macroscopic phases of a segregated material with phases @xmath11 being almost one ( dark gray ) and almost zero ( light gray ) . in figure [ fig : intuitivemicro ] we show the corresponding microscopic length scale featuring microstructures ; the left is a blow - up of a small region within the phase with @xmath11 being almost one and the right is a blow - up of a very small region within the phase with @xmath11 being almost zero . on this microscopic scale where we assume that it is sufficient to treat the elastic energies of single crystals , fine microstructures in the form of laminates occur , see , e.g. , @xcite . the methods developed here apply generally to any established phase change and segregation model provided the temperature is conserved . ( for non - isothermal settings , the validity of the second law of thermodynamics requires additional corrections which are not studied here . ) for further discussions of our model and the analytical results we refer to the conclusions in section [ secconclude ] . throughout this paper , let @xmath12 for @xmath13 be a bounded domain with lipschitz boundary which serves as an ( unstressed ) reference configuration . for a stop time @xmath14 , let @xmath15 denote the space - time cylinder . to the allen - cahn / cahn - hilliard system , first derived in @xcite , we add elasticity , possibly respecting the lamination microstructure , by introducing the system @xmath16 where the function @xmath17 is a _ conserved _ order parameter , typically a concentration , @xmath18 is an _ unconserved _ order - parameter , specifying the reordering of the underlying lattice , @xmath19 denotes the positive semi - definite mobility tensor , @xmath20 is a small constant determining the interfacial thickness . the choice of @xmath8 determines whether lamination microstructure occurs in the system , see . by @xmath21 we describe the displacement field , such that a material point @xmath22 in the undeformed body @xmath23 is at @xmath24 after the deformation . then the ( linearized ) strain tensor is defined by @xmath25 where @xmath26 denotes the transpose of a matrix @xmath27 . as usual , @xmath28 stands for the inner product in @xmath29 , that is @xmath30 , and for @xmath31 we denote the inner product in @xmath32 by @xmath33 moreover , @xmath34 for @xmath35 is the frobenius norm . we denote the symmetric matrices in @xmath32 by @xmath36 . the functional @xmath37 represents the stored elastic energy density . we will choose it either according to the linear ansatz by eshelby , @xmath38 , or as a single crystal composite lamination energy , @xmath39 . the definition of @xmath40 is given in the coming section [ ssectelast ] . the linear theory by eshelby @xcite developed in the context of elastic inclusions and inhomogeneities , can be summarized in the following ansatz for the elastic energy @xmath41 for all @xmath42 , @xmath43 , and @xmath44 with a constant @xmath45 . by @xmath46 we denote the symmetric , positive definite and concentration - dependent elasticity tensor of the system that maps symmetric tensors in @xmath32 to themselves . the system ( [ ach1])([ach3 ] ) is completed with the definition of the free energy @xmath47 see @xcite , where @xmath48 is the free energy density assumed to be @xmath49 for scalars @xmath50 . the term @xmath51 $ ] in ( [ psidef ] ) defines the entropic part of the free energy , given in the canonical bernoulli form for perfect mixing , and @xmath52 is the constant temperature . the functional @xmath53 in ( [ fdef ] ) represents energy effects due to applied forces . in the absence of body forces , the work necessary to transform the undeformed body @xmath23 into a state with displacement @xmath54 is then @xmath55 where we use that the applied stress @xmath56 is constant and symmetric . consequently , @xmath57 is the energy density of the applied outer forces . the system ( [ ach1])([ach3 ] ) has to be solved in @xmath58 subject to the initial conditions @xmath59 for given functions @xmath60 subject to the neumann boundary conditions for @xmath4 , the no - flux boundary conditions , and the equilibrium condition for applied forces @xmath61 here , @xmath62 defines the stress . in ( [ bc ] ) , @xmath63 is the unit outer normal to @xmath64 . for simplicity , body forces are neglected and it is assumed that the boundary tractions are dead loads given by a constant symmetric tensor @xmath56 . by @xmath65 we denote the mass flux , given by @xmath66 with @xmath67 the chemical potential . the valid parameter range of @xmath4 and @xmath3 is , see theorem [ theo1 ] , @xmath68 the inequalities are strict unless @xmath69 or @xmath70 . for @xmath71 in @xmath23 , we obtain the pathological solution @xmath72 in @xmath58 . for @xmath73 , @xmath74 in @xmath23 we obtain a cahn - hilliard equation in @xmath3 and a pathological equation for @xmath4 . the system ( [ ach1])([ach3 ] ) includes as special case the _ elastic cahn - hilliard system _ ( setting @xmath75 , @xcite ) @xmath76 with @xmath77 and the boundary and initial conditions correspondingly to above . moreover the system ( [ ach1])([ach3 ] ) includes as special case the _ elastic allen - cahn equations _ ( setting @xmath78 , @xcite ) which for rescaled @xmath3 with @xmath79 read @xmath80 with @xmath81 and @xmath82 as well as the boundary and initial conditions are as above . the system ( [ ach1])([ach3 ] ) is exemplary for an isothermal model that exhibits simultaneous ordering and phase transitions . equation ( [ ach1 ] ) is a diffusion law for @xmath4 governed by the flux @xmath65 and states the conservation of mass in @xmath23 . equation ( [ ach2 ] ) is a simple gradient flow in the descent direction @xmath83 . equation ( [ ach3 ] ) is a consequence of newton s second law under the additional assumption that the acceleration @xmath84 originally appearing on the left hand side can be neglected ( this can be proved formally by a scaling argument and formally matched asymptotics ) . the vector equation ( [ ach3 ] ) serves to determine the unknown displacement @xmath54 . [ rem1 ] the equations ( [ ach1])([ach3 ] ) can be generalized to vector - valued mappings @xmath4 , @xmath3 . this allows to study situations with more than two phases present . to fix ideas and for the sake of a clear presentation , we restrict ourselves throughout this paper to scalars @xmath4 and @xmath3 . [ rem2 ] the equations ( [ ach1])([ach3 ] ) with boundary conditions ( [ bc ] ) and and comply with the second law of thermodynamics , which in case of isothermal conditions reads for a closed system @xmath85 this inequality can be verified by direct inspection similar to the calculations in @xcite . in the proof of theorem [ theo2 ] we apply the following explicit formulation of ( [ ach1])([ach3 ] ) with constant mobility @xmath86 @xmath87\!\!,\\ \label{ach2e } { \partial_t}b & = & \lambda\triangle b+\frac{{\vartheta}}{2}\left[g'(a - b ) -g'(a+b)\right]+2\kappa_2b-\frac{\partial{{\cal w}}}{\partial d}(a+b,{\varepsilon}({\mathbf{u}})),\\ \label{ach3e } \mathbf{0 } & = & { \mathrm{div}}\left({\partial_\varepsilon}{{\cal w}}(a+b,{\varepsilon}({\mathbf{u}}))\right).\end{aligned}\ ] ] the subsequent section is devoted to the geometrically linear elastic energy for single crystals . this is a prerequisite to the discussion of existence and uniqueness results for the extensions of the allen - cahn / cahn - hilliard ( ac - ch ) model studied in section [ secexist ] . our main objective in this subsection is to study a geometrically linear theory of elasticity in the context of isothermal phase transitions . for systematic reasons , we first recall the linear ansatz dating back to eshelby , @xcite . as a byproduct of the existence theory proved in section [ secexist ] we obtain a new existence result for the ac - ch equations with linear elasticity . subsequently , we introduce the geometrically linear elasticity theory that takes the laminates of the material into account . in the following we assume that two phases are present in the considered material which may form microstructures as displayed , e.g. , in figures [ fig1 ] and [ fig2 ] . we refer to the energy @xmath88 , @xmath89 of each of the phases as _ microscopic energy _ , cf . , and to the energy @xmath90 in , which reflects the effective behavior of the system with microstructures , as the _ macroscopic energy_. to determine the energy @xmath90 in the geometrically linear theory we need to solve a local minimization problem , eqn . ( [ wmin ] ) below , which we shall outline now . we assume that the volumes occupied by each of the two phases in @xmath23 are measurable sets . in particular , if @xmath91 , @xmath92 characterize the two phases on the microscale , we have @xmath93 and @xmath94 a.e . in @xmath23 . the symbol @xmath95 denotes the space of functions of bounded variation , see , e.g. , @xcite . by @xmath96 we denote the average of a function @xmath97 in @xmath23 , where @xmath98 is the @xmath99-dimensional lebesgue measure of a set @xmath100 . let @xmath101 , @xmath89 , be the stress - free strain ( or eigenstrain ) of the @xmath102-th phase relative to the chosen reference configuration and @xmath103 be its positive definite elasticity tensor . then the elastic energy density of phase @xmath102 subject to a strain @xmath104 is given by @xmath105 for @xmath106 . in ( [ widef ] ) we assumed for simplicity that @xmath103 , @xmath107 and @xmath108 are constants , independent of @xmath109 and the order parameter @xmath11 . under the assumption that the elastic energy adapts infinitely fast and that the surface energy between laminates of the microstructure can be neglected , the effective elastic energy is , @xcite , @xmath110,\ ] ] where we used @xmath111 the definition ( [ wmin ] ) requires further clarification . firstly , @xmath112 , and instead of integrating over @xmath23 , one may integrate over @xmath113 , the open ball of radius @xmath114 around @xmath115 , where the mean @xmath116 is now taken over @xmath113 . by homotopy arguments or by results in @xcite , any @xmath117 yields the same value of @xmath118 , as long as @xmath119 . taking the union of such balls then leads to ( [ wmin ] ) . secondly , the infimum over @xmath120 is the result of homogenization subject to the constraint that the volume fraction of the selected phase is preset by @xmath121 , see ( * ? ? ? * chapter 10 ) . this infimum is taken over functions @xmath122 as explained above ensuring @xmath123 . this is why ( [ wmin ] ) is only meaningful for @xmath124 $ ] . the second infimum is taken over functions @xmath125 where the condition @xmath126 has to be read as @xmath127 for a.e . this originates from the requirement that the functional @xmath40 thus defined must be quasi - convex , see @xcite , and is the result of relaxation theory , @xcite , @xcite , as follows . if for prescribed @xmath5 the microscopic elastic energy density is denoted by @xmath129 , then @xmath130 is the elastic energy density of the material with macroscopic strain @xmath131 after microstructure has formed . for @xmath10 , explicit analytic formulas for @xmath40 are known , see @xcite as well as formulas of the partial derivatives of @xmath40 . these will be recalled in section [ secexpl ] . the existence of solutions to the allen - cahn / cahn - hilliard equation without elasticity was studied in @xcite with the help of a semigroup calculus . existence and uniqueness of weak solutions to the cahn - hilliard equation with linear elasticity is proved in @xcite , with geometrically linear elasticity in @xcite . existence and uniqueness of weak solutions to the allen - cahn equation with linear elasticity is shown in @xcite . subsequently we provide existence and uniqueness results for ( [ ach1])([ach3 ] ) , where @xmath8 is an elastic energy density satisfying the following assumption ( @xmath9 ) . ( @xmath9 ) the elastic energy density @xmath132 satisfies the conditions ( @xmath91 ) @xmath133 is strongly monotone uniformly in @xmath11 , i.e. , there exists a constant + @xmath134 such that for all @xmath135 and all @xmath136 @xmath137 ( @xmath92 ) there exists a constant @xmath138 such that for all @xmath136 and all @xmath42 @xmath139 , @xmath140 , @xmath141 and @xmath142 , but are independent of @xmath11 and @xmath131 . condition ( @xmath91 ) states that @xmath8 is convex in @xmath131 . the problem becomes non - convex through the dependence on @xmath11 . one prominent example satisfying assumption ( @xmath9 ) is the elastic energy @xmath143 in . in section [ secexam ] we prove that also the relaxed energy functional @xmath40 defined in ( [ wmin ] ) satisfies the assumption ( @xmath9 ) . we require the condition @xmath144 where @xmath145 is the solution of ( [ ach3 ] ) for @xmath146 , @xmath147 . [ theo1 ] let the mobility tensor @xmath148 be positive definite and continuous for all @xmath4 , @xmath3 satisfying ( [ abrange ] ) , @xmath8 fulfill ( @xmath9 ) , @xmath149 be given by ( [ psidef ] ) and the initial data @xmath150 satisfy ( [ abrange ] ) and ( [ w0bound ] ) . then there exists a weak solution @xmath151 to ( [ ach1])([ach3 ] ) that satisfies \(i ) @xmath152;\,l^2(\omega)\right)$ ] , + ( ii ) @xmath153 , @xmath154 , + ( iii ) @xmath155 , + ( iv ) the feasible parameter range of @xmath156 is given by ( [ abrange ] ) . : the statements of the theorem can be proved with the methods developed in @xcite for an allen - cahn system with linear elasticity and in @xcite for a cahn - hilliard system with linear elasticity . we sketch the main steps . first we introduce the operator @xmath157 associated to @xmath158 as a mapping from @xmath159 to its dual by @xmath160 from the poincar inequality and the lax - milgram theorem ( which can be applied since @xmath148 is assumed to be positive definite ) we know that @xmath157 is invertible and we denote its inverse by @xmath161 , the green s function . we have @xmath162 for @xmath163 , we define the inner product @xmath164 with the corresponding norm @xmath165 for a small discrete step size @xmath166 , chosen such that @xmath167 , for time steps @xmath168 with @xmath169 , and given values @xmath170 , @xmath171 , we introduce the discrete free energy functional @xmath172 where ( in case of @xmath173 ) it holds @xmath174 , @xmath175 , the initial values of @xmath4 and @xmath3 . by the direct method in the calculus of variations and assumption ( @xmath9 ) , it is possible to show that for @xmath176 sufficiently small , @xmath177 possesses a minimizer @xmath178 . this minimizer solves the fully implicit time discretization of ( [ ach1])([ach3 ] ) . next the discrete solution is extended affine linearly to @xmath179 by setting for @xmath180 with suitable @xmath181 $ ] @xmath182 the validity of the second law of thermodynamics ( cf.remark [ rem2 ] ) together with ( [ w0bound ] ) implies that @xmath183 is non - increasing in time . in combination with a higher integrability condition on @xmath54 , @xcite , this allows to derive uniform estimates for @xmath179 . compactness arguments then allow to pass to the limit @xmath184 and the limit solves ( [ ach1])([ach3 ] ) . in general , the uniqueness of solutions to is open . however , we prove it in a special case for the linear elastic energy density @xmath38 . [ theo2]@xmath185let @xmath186 be given by ( [ wlin ] ) , the material be homogeneous , i.e. the elasticity tensor @xmath187 be independent of @xmath11 , and let @xmath86 . then the solution @xmath151 of theorem [ theo1 ] is unique in the spaces stated there . : the proof is very similar to the proof of uniqueness for the cahn - hilliard system @xcite but we repeat it here because we later need to modify it . fix @xmath188 . let @xmath189 , @xmath190 be two pairs of solutions to ( [ ach1e])([ach3e ] ) and ( [ wlin ] ) . the differences @xmath191 , @xmath192 , @xmath193 with corresponding difference of the chemical potentials @xmath194 solve the weak equations @xmath195{\,\mathrm{d}x}{\,\mathrm{d}t}=0,\\ & { \int\limits_{\omega_{\cal t}}}\left[{\partial_t}b\eta+\lambda\nabla b\cdot\nabla\eta-{\underline{\varepsilon}}:c\left({\varepsilon}({\mathbf{u } } ) -(a+b){\underline{\varepsilon}}\right)\eta\right]{\,\mathrm{d}x}{\,\mathrm{d}t}{\nonumber}\\ \label{test2 } & \ ; = \!{\int\limits_{\omega_{\cal t}}}\left[\frac{{\vartheta}}{2}\left(g'(a^2\!+\!b^2)-g'(a^1\!+\!b^1 ) + g'(a^1\!-\!b^1)-g'(a^2\!-\!b^2)\right)\eta+2\kappa_2b\eta\right]{\,\mathrm{d}x}{\,\mathrm{d}t},\\ \label{test3 } & { \int\limits_{\omega_{t_0}}}c\left({\varepsilon}({\mathbf{u}})-{\underline{\varepsilon}}(a+b)\right):{\varepsilon}({\mathbf{u}}){\,\mathrm{d}x}{\,\mathrm{d}t}=0\end{aligned}\ ] ] for every @xmath196 with @xmath197 , @xmath198 , where in order to get ( [ test3 ] ) we plugged in @xmath199 as a test function and integrated by parts . as a test function in ( [ test1 ] ) we pick @xmath200 this shows @xmath201 the difference of the chemical potentials fulfils , with the help of ( [ fdef ] ) , @xmath202{\,\mathrm{d}x}{\,\mathrm{d}t}.\end{aligned}\ ] ] we pick @xmath203 . with ( [ step1 ] ) we obtain @xmath204|a|{\,\mathrm{d}x}{\,\mathrm{d}t}.\end{aligned}\ ] ] in ( [ test2 ] ) we choose @xmath205 as a test function and add the resulting equation to ( [ step2 ] ) and use ( [ test3 ] ) . we end up with @xmath206{\,\mathrm{d}x}{\,\mathrm{d}t}\\ & & \le{\int\limits_{\omega_{t_0}}}2(\kappa_1|a|^2+\kappa_2|b|^2){\,\mathrm{d}x}{\,\mathrm{d}t}\\ & & \quad+\!\!{\int\limits_{\omega_{t_0}}}\!\frac{{\vartheta}}{2}\big[\big|g'(a^1\!+\!b^1 ) -g'(a^2\!+\!b^2)\big|+\big|g'(a^1\!-\!b^1)-g'(a^2\!-\!b^2)\big|\big ] ( |a|\!+\!|b|){\,\mathrm{d}x}{\,\mathrm{d}t}.\end{aligned}\ ] ] from theorem [ theo1 ] we know that the terms @xmath207 , @xmath89 are finite , and @xmath208 is lipschitz continuous . applying first young s inequality , then gronwall s inequality , as @xmath188 was arbitrary , we find @xmath209 in @xmath58 . this finally yields @xmath210 with korn s inequality this proves @xmath211 in @xmath58 . in many situations like the numerical implementation of the extended models , the above definition ( [ wmin ] ) of @xmath40 is not practical since it is indirect and based on a local minimization . for these applications and for direct later use , we collect here some explicit formulas of the relaxed energy @xmath40 for @xmath6 and for the scalar case in @xmath1 . as shown in @xcite , it holds @xmath212 where @xmath213 , @xmath214 and @xmath215 are defined below . first we need to fix further notations following @xcite . let @xmath216 be given by @xmath217 where @xmath218 is the reciprocal of the largest eigenvalue of @xmath219 , @xmath103 is the elastic modulus of laminate @xmath102 , and the operator @xmath220 is given by @xmath221 in @xcite a recipe is given for the practical computation of @xmath222 . here , we only remark that if the space groups of the two existing laminates are cubic , it holds @xmath223 the first subscript of @xmath187 denotes here the phase , the other two indices are the coefficients of the reduced elasticity tensor in voigt notation , @xcite . as shown in @xcite , the scalar @xmath224 $ ] determines the amount of translation of the laminates defined by @xmath225 in this definition , @xmath226 is the unique solution of @xmath227 with @xmath228 defined by @xmath229,\\ \triangle{\varepsilon}^ * & = & \triangle{\varepsilon}^*(\beta , d,{\varepsilon } ) \,\,:=\,\,{\varepsilon}_2^*(\beta , d,{\varepsilon})-{\varepsilon}_1^*(\beta , d,{\varepsilon}),{\nonumber}\end{aligned}\ ] ] and the yet undefined functions are specified below . the four regimes have the following crystallographic interpretation , which follows from the construction of the optimal microstructure in the calculation of @xmath40 . * regime 0 * : the material is homogeneous and the energy does not depend on the microstructure . this occurs when @xmath230 . * regime i * : there exist two optimal rank - i laminates . this is characterized by @xmath231<0.\ ] ] * regime ii * : the unique optimal microstructure is a rank - i laminate . this regime occurs when the function @xmath232\ni\beta\mapsto\det\left[\left(d_2\alpha_1+d_1\alpha_2 -\beta t\right)^{-1}\left(\alpha_2({\varepsilon}_2^t-{\varepsilon})- \alpha_1({\varepsilon}_1^t-{\varepsilon})\right)\right]\ ] ] has a unique root ( which we denote by @xmath233 ) . * regime iii * : there exist two optimal rank - ii laminates . this regime is present if the operator @xmath234 is invertible and @xmath235>0.\ ] ] for illustration , we visualize prototypes of rank - i and rank - ii laminates in figures [ fig1 ] and [ fig2 ] . ( 13.,3.2 ) yy ( + 4.6,0.0)(9.57,1.4)(1,0)1 ( 13.,3.2 ) yy ( + 3.45,-6.1 ) to complete the definition ( [ phidef ] ) , set @xmath236 hence @xmath237^{-1}e({\varepsilon})$ ] . explicit computations of @xmath238 and @xmath239 are lengthy . we recall the following results for the partial derivatives of @xmath40 in @xmath10 which are proved in @xcite . set @xmath240 [ lem1 ] let @xmath10 and let @xmath103 and @xmath241 commute . then @xmath242 alternatively , in regime iii , @xmath243 moreover it holds @xmath244 @xmath245 \frac{\partial\beta^*}{\partial{\varepsilon } } & = \begin{cases } \frac{1}{\left ( ( d_2\alpha_1+d_1\alpha_2-\beta t)^{-1}(t\triangle{\varepsilon}^*)\right ) : t\triangle{\varepsilon}^*}(\alpha_1-\alpha_2)\alpha^{-1}t\triangle{\varepsilon}^ * & \text{in regime ii},\\ 0 & \text{otherwise}. \end{cases}\label{eq : beps}\end{aligned}\ ] ] the scalar setting in three dimensions is characterized by the ansatz @xmath246 for the deformation , where @xmath247 is a scalar function ( hence the name scalar theory , see @xcite , @xcite ) . physically , ( [ uscal ] ) corresponds to anti - plane shear in the @xmath248-plane . since @xmath249 @xmath250 determines the strain . this justifies to work with vectors @xmath251 , not with matrices @xmath252 . we replace ( [ wmin ] ) by @xmath253,\ ] ] where the domain of integration @xmath254 is now two - dimensional . therefore , @xmath255 , and it is well - understood from the context that @xmath256 denotes the two - dimensional average . in addition , in ( [ wmins ] ) we adapted the common notation for the scalar case and wrote @xmath257 instead of @xmath131 for the strain . finally , we set @xmath258 with @xmath259 the transformation strains in the scalar setting and @xmath103 positive definite matrices . in @xcite , partial results for @xmath40 in the non - scalar , i.e. vectorial setting are available for @xmath1 . yet , in general , the computation of @xmath238 , @xmath239 for @xmath1 as required for assumption ( @xmath9 ) remains currently open due to its complexity . this is the reason why we restrict ourselves to a special three - dimensional case . our main result in this subsection is the following theorem that provides us with an explicit formula for @xmath40 in the scalar case . [ theo5 ] the functional @xmath40 , given by ( [ wmins ] ) , satisfies the explicit representation formula @xmath260 with @xmath261,\\ \label{etasdef2 } \nabla\eta_2^*(d , f ) & : = ( d_2\alpha_1+d_1\alpha_2)^{-1}\big[\alpha_1f+d_1 ( \alpha_2f_2^t-\alpha_1f_1^t)\big].\end{aligned}\ ] ] theorem [ theo5 ] states in particular that the scalar elastic theory in 3d is directly related to the non - scalar geometrically linear elasticity theory in 2d which was discussed in subsection [ subexpl2d ] . : we apply the translation method , see for instance @xcite , @xcite for an overview . in general , the function @xmath262 which describes the translation is only quasi - convex . here , as a benefit of the scalar theory , we can pick @xmath262 as a convex function , cf . . starting from the inequality @xmath263 which is satisfied by any quasi - convex function @xmath262 , we obtain , similar to the reasoning in @xcite for the non - scalar case in @xmath10 dimensions , @xmath264.\end{aligned}\ ] ] consider @xmath265 then @xmath266 next we apply jensen s inequality which is possible since @xmath267 is convex for @xmath89 . furthermore we possibly enlarge the set of admissible functions in the minimization over @xmath247 in by considering the set of admissible functions @xmath268 with the constraint . then becomes @xmath269 if @xmath10 , @xmath262 is chosen to be @xmath270 . here we set analogously @xmath271 since @xmath262 is quadratic , there exists a unique linear operator @xmath272 such that @xmath273 , here simply @xmath274 . it holds @xmath275.\end{aligned}\ ] ] by a direct computation the remainder term @xmath276 can be rewritten , @xmath277\\ & = & -d_1d_2t(\nabla\eta_2-\nabla\eta_1)\!:\!(\nabla\eta_2-\nabla\eta_1)\\ & = & -2d_1d_2{\varphi}(\nabla\eta_2-\nabla\eta_1).\end{aligned}\ ] ] so we have found @xmath278 next we compute the optimal strains @xmath279 , @xmath280 . after differentiating the argument on the right in ( [ h2 ] ) for fixed @xmath281 , we obtain @xmath282 using the constraint @xmath283 , after rearrangement this yields the formulas @xmath284 setting @xmath285 ( see below ) , this coincides with ( [ etasdef1 ] ) , ( [ etasdef2 ] ) . the maximum over @xmath281 in ( [ h2 ] ) is attained at @xmath286 as @xmath287 . hence , the translation is trivial in this setting . we obtain @xmath288 it remains to estimate @xmath40 from above which amounts to showing that @xmath279 , @xmath280 yield an optimal microstructure . plugging in any microstructure @xmath120 with @xmath289 , from the definition ( [ wmins ] ) of @xmath40 , we get the upper bound @xmath290 as the domain of integration @xmath254 in the scalar case with @xmath1 is two - dimensional , finding the optimal microstructure subsumes to the non - scalar case in @xmath10 dimensions . hence , as can be seen from ( [ bsdef ] ) , @xmath286 occurs either for regime 0 ( no microstructure ) or regime i. the latter is equivalent to @xmath291 . from a well - known argument , see ( * ? ? ? * lemma 4.1 ) , this implies that @xmath214 , @xmath215 are compatible . besides this compatibility condition , any optimal microstructure must also satisfy the equilibrium condition @xmath292\vec{n}=0\ ] ] which states that the jump of the stress in the normal direction of the laminate must vanish . equation ( [ state ] ) originates from the euler - lagrange equation of the variational problem . here , it is automatically satisfied since , due to ( [ nosjump ] ) , @xmath293\vec{n}=\sigma_2^*-\sigma_1^*=0 $ ] . so there exists a unique rank - i lamination microstructure ( unique up to a sign @xmath294 of direction of the laminate ) that connects @xmath214 and @xmath215 , and the strain of phase @xmath102 is @xmath295 , @xmath89 . this microstructure is optimal , @xmath296 , which proves ( [ whwsexp ] ) . the methods of the previous subsection also apply to derive rigorously an explicit expression for @xmath40 in @xmath2 . here , quasi - convex functions are always convex , and the vector and scalar settings coincide . however , we use the notation in analogy to the vector case . [ theoa ] for @xmath2 , the functional @xmath40 , given by ( [ wmin ] ) , satisfies the explicit representation formula @xmath297 with @xmath298 the partial derivatives of @xmath40 are given by @xmath299}{d_2\alpha_1+d_1\alpha_2},\\ \frac{\partial{\widehat w}}{\partial d}(d,{\varepsilon } ) & = w_1({\varepsilon}_1^*)-w_2({\varepsilon}_2^ * ) + \frac{\alpha_1\alpha_2}{(d_2\alpha_1+d_1\alpha_2)^2}\big [ ( \alpha_1-\alpha_2){\varepsilon}^2+d_1\alpha_1({\varepsilon}_1^t)^2-d_2\alpha_2({\varepsilon}_2^t)^2{\nonumber}\\ & \hspace*{170pt } + \big(\alpha_2{\varepsilon}_2^t\!-\!\alpha_1{\varepsilon}_1^t \!+\!(\alpha_2\!-\!\alpha_1)(d_1{\varepsilon}_1^t\!+\!d_2{\varepsilon}_2^t)\big){\varepsilon}{\nonumber}\\ & \hspace*{170pt } \label{dwdd1d } + ( d_2\alpha_1-d_1\alpha_2){\varepsilon}_1^t{\varepsilon}_2^t\big].\end{aligned}\ ] ] : we apply the same methods as in the proof of theorem [ theo5 ] . first , we estimate @xmath300 , @xmath301 from below by their convexification . similar to ( [ h2 ] ) , we have for any convex function @xmath228 and @xmath302 , @xmath303 secondly , we choose the optimal @xmath281 and @xmath228 . the later is an art when @xmath304 . with arguments identical to the scalar case in @xmath1 , it holds @xmath305 , leaving us with the estimate @xmath306 next we compute the optimal strains @xmath307 , @xmath308 on the right . for @xmath309 , only @xmath310 is admissible and hence optimal . now , let @xmath311 . we can then resolve the constraint by setting @xmath312 and need to calculate @xmath313 after differentiation with respect to @xmath307 and using ( [ widef ] ) for the optimal @xmath214 in ( [ star2 ] ) , we are left with @xmath314 rearrangement and resolution of @xmath315 gives the formulas ( [ e1db ] ) , ( [ e1dc ] ) of the optimal strains . since @xmath316 , these formulas also state the correct solution when @xmath309 . in the final step , due to the definition of @xmath40 , it holds @xmath317 and @xmath122 may represent any microstructure with @xmath289 . the explicit construction of the optimal microstructure for which the upper bound and the lower bound coincide , @xmath318 is again as in the proof of theorem [ theo5 ] . this leads to ( [ wh1d ] ) . finally , the verification of ( [ dwdeps1d ] ) , ( [ dwdd1d ] ) follows from @xmath319 and @xmath320}{(d_2\alpha_1+d_1\alpha_2)^2},\\ \label{depss1d2b } \frac{\partial{\varepsilon}_2^*}{\partial d } & = \frac{\alpha_1\big[\alpha_2{\varepsilon}_2^t -\alpha_1{\varepsilon}_1^t+(\alpha_1-\alpha_2){\varepsilon}\big]}{(d_2\alpha_1+d_1\alpha_2)^2}\end{aligned}\ ] ] which all can be verified by elementary computations . using the relationship @xmath321 we find ( [ dwdeps1d ] ) and finally @xmath322 } { ( d_2\alpha_1+d_1\alpha_2)^2}.{\nonumber}\end{aligned}\ ] ] using @xmath323 this simplifies to identity ( [ dwdd1d ] ) . above we have already remarked that the energy @xmath324 satisfies assumption ( @xmath9 ) and thus theorem [ theo1 ] applies . in this section we investigate whether the results transfer to the geometrically linear theory of elasticity stated in the previous section . the following statement is a consequence of theorem [ theo1 ] . [ thmcor1 ] let @xmath39 be a function on @xmath325\times{\mathbb{r}}^{d\times d}_\mathrm{sym}$ ] . assume that the mobility tensor @xmath148 be positive definite and continuous for all @xmath4 , @xmath3 satisfying ( [ abrange ] ) . moreover , let @xmath149 be given by ( [ psidef ] ) and the initial data @xmath150 satisfy ( [ abrange ] ) and ( [ w0bound ] ) . in @xmath326 , let @xmath40 be given by ( [ whwsexp ] ) corresponding to the scalar setting . if @xmath10 , let 1 . @xmath327 be independent of @xmath131 , and 2 . @xmath103 and @xmath241 commute whenever @xmath328 . if @xmath2 , let @xmath40 be given by ( [ wh1d ] ) . then there exists a solution @xmath151 to ( [ ach1])([ach3 ] ) that satisfies \(i ) @xmath152;\,l^2(\omega)\right)$ ] , + ( ii ) @xmath329 , @xmath330 , + ( iii ) @xmath155 , + ( iv ) the feasible parameter range of @xmath156 is given by ( [ abrange ] ) . : we first consider the case @xmath0 . the idea is to apply theorem [ theo1 ] . therefore we 1 . verify that @xmath40 satisfies assumption ( @xmath9 ) for @xmath124 $ ] , 2 . extend @xmath40 to @xmath331 in such a way that ( @xmath9 ) is fulfilled for all @xmath136 . then theorem [ theo1 ] can be applied and we obtain by part ( iv ) therein that effectively @xmath332 $ ] in the system ( [ ach1])([ach3 ] ) . hence theorem [ thmcor1 ] follows with @xmath39 as asserted . we start with the second step and give the proofs of the former in dimensions @xmath2 , @xmath10 , and the scalar case in @xmath1 thereafter . we introduce the following extension @xmath333 of @xmath334 by @xmath335 for functions @xmath336 , @xmath337 that we may choose such that @xmath338 and that @xmath339 fulfils ( @xmath91)(@xmath92 ) for @xmath340 and @xmath341 . we further mention that the function @xmath339 defined in ( [ wext ] ) is positive and coercive for @xmath342 . we note that this construction is required for the first part of the proof of theorem [ theo1 ] , where polynomial approximations of @xmath149 are constructed , where @xmath343 $ ] may occur . it remains to show that @xmath40 fulfils ( @xmath9 ) for @xmath332 $ ] , which we prove independently for @xmath2 , @xmath10 , and the scalar case in @xmath1 . \(i ) _ verification of ( @xmath9 ) for @xmath2 _ : from , ( [ wh1d ] ) , and , it follows @xmath344 . from ( [ dwdeps1d ] ) , we obtain @xmath345 the condition ( @xmath91 ) , restricted to @xmath124 $ ] , follows , since by , for @xmath89 , @xmath346 where @xmath347 by assumption . from , , for @xmath89 and @xmath124 $ ] we have by similar arguments latexmath:[\[\label{visest } this leads to ( @xmath92)@xmath349 , since by ( [ wh1d ] ) , , @xmath350 for @xmath124 $ ] we easily compute , for @xmath89 , @xmath351 we use this estimate in ( [ star3 ] ) , to obtain by @xmath352 which is ( @xmath92)@xmath353 restricted to @xmath124 $ ] . for @xmath124 $ ] we derive with the help of ( [ dwdeps1d ] ) , ( [ star4 ] ) , @xmath354 showing the validity of ( @xmath92)@xmath355 restricted to @xmath124 $ ] . \(ii ) _ verification of ( @xmath9 ) for @xmath10 _ : the regularity of @xmath40 required for ( @xmath9 ) follows directly from and the related definitions . the constant @xmath222 solely depends on @xmath356 , @xmath140 since @xmath357 and @xmath218 is the reciprocal of the largest eigenvalue of @xmath219 , see . since @xmath224 $ ] , latexmath:[\[\label{bsest } independent of @xmath11 and @xmath131 . from ( [ bsest ] ) , ( [ epsistardef ] ) we obtain for @xmath124 $ ] latexmath:[\[\label{e1d2 } with and this shows for @xmath124 $ ] @xmath360 from this we easily verify ( @xmath92)@xmath349 restricted to @xmath124 $ ] , as the first two terms in ( [ wdef ] ) can be estimated as in the one - dimensional case . in order to show ( @xmath92)@xmath353 , we apply and ( [ visest ] ) to find @xmath361 finally , from ( [ betrus ] ) and the fact that @xmath241 is an isometry w.r.t . the frobenius norm , @xmath362 the terms @xmath363 , @xmath89 remaining in can be estimated as in the one - dimensional case . this verifies ( @xmath92)@xmath353 restricted to @xmath124 $ ] . now we prove ( @xmath92)@xmath355 . for the regimes 0 , i and ii this follows directly from ( [ visest ] ) . when in regime iii , we start from and find with for @xmath124 $ ] @xmath364 the validation of ( @xmath91 ) relies on . by assumption ( i ) of the theorem , @xmath365 and , as @xmath124 $ ] is fixed in ( @xmath91 ) , @xmath366 is a constant tensor . assumption ( @xmath91 ) then follows from @xmath367 a similar equality for @xmath368 , and the positive definiteness of @xmath366 and @xmath369 , @xmath89 . indeed , the latter is equivalent to @xmath370 positive definite , and since @xmath371 , this is implied by all eigenvalues of @xmath372 being positive . however , from ( [ gsdef ] ) and the definition of @xmath218 , this is true . for the scalar case in @xmath1 , comparing ( [ wdef ] ) and ( [ whwsexp ] ) and since the partial derivatives coincide , the theory subsumes to the geometrically linear theory for @xmath10 and @xmath286 . hence we can proceed as above and show that @xmath40 given by ( [ whwsexp ] ) satisfies ( @xmath9 ) . the assumptions ( @xmath3731 ) , ( @xmath3732 ) are not needed here because @xmath286 . the following theorem states the uniqueness of weak solutions for the system with geometrically linear theory . [ theo4 ] let @xmath39 be given by ( [ whwsexp ] ) for @xmath1 ( scalar setting ) , by ( [ wdef ] ) for @xmath10 , or by ( [ wh1d ] ) for @xmath2 . assume further that @xmath86 and the elastic moduli of the two phases are equal , i.e. @xmath374 . for @xmath10 , let @xmath375 . then the solution @xmath151 in theorem [ thmcor1 ] is unique in the spaces stated there . : ( i ) _ the case @xmath2 . _ for @xmath374 , the equations ( [ e1db ] ) , ( [ e1dc ] ) read @xmath376 which implies @xmath377 leading to @xmath378 . this yields @xmath379 in this special case , the partial derivatives of @xmath40 are @xmath380 now we pass through the steps in the proof of theorem [ theo2 ] and make the necessary modifications . let again @xmath381 , @xmath190 be two solutions to ( [ ach1])([ach3 ] ) and set @xmath191 , @xmath192 , @xmath382 , @xmath383 for @xmath190 and @xmath384 , @xmath385 . the equations ( [ test2 ] ) , ( [ test3 ] ) now read @xmath386{\,\mathrm{d}x}{\,\mathrm{d}t}{\nonumber}\\ \label{2test2 } & \quad = \!{\int\limits_{\omega_{\cal t}}}\left[\frac{{\vartheta}}{2}\left(g'(a^2\!+\!b^2)-g'(a^1\!+\!b^1 ) + g'(a^1\!-\!b^1)-g'(a^2\!-\!b^2)\right)\eta+2\kappa_2b\eta\right]{\,\mathrm{d}x}{\,\mathrm{d}t},\\ \label{2test3 } & { \int\limits_{\omega_{t_0}}}\alpha_1\big({\varepsilon}+d({\varepsilon}_2^t-{\varepsilon}_1^t)\big){\varepsilon}(u){\,\mathrm{d}x}{\,\mathrm{d}t}=0.\end{aligned}\ ] ] analogous to ( [ step2 ] ) we find @xmath387|a|{\,\mathrm{d}x}{\,\mathrm{d}t}.\end{aligned}\ ] ] choosing @xmath205 in ( [ 2test2 ] ) and adding the resulting equation to ( [ 2step4 ] ) , we obtain @xmath388{\,\mathrm{d}x}{\,\mathrm{d}t}\\ & \le\!{\int\limits_{\omega_{t_0}}}\!2(\kappa_1|a|^2\!+\!\kappa_2|b|^2 ) + \!\!{\int\limits_{\omega_{t_0}}}\!\!\frac{{\vartheta}}{2}\big[\big|g'(a^1\!\!+\!\!b^1 ) \!-\!g'(a^2\!\!+\!\!b^2)\big|\!+\!\big|g'(a^1\!\!-\!\!b^1)\!-\!g'(a^2\!\!-\!\ ! b^2)\big|\big](|a|\!+\!|b|){\,\mathrm{d}x}{\,\mathrm{d}t}{\nonumber}.\end{aligned}\ ] ] it is noteworthy that we are only able to estimate @xmath389 which differs from the total mechanical energy @xmath390 . with the lipschitz continuity of @xmath208 , and by applying the inequalities of young and gronwall to ( [ wnote ] ) , we obtain as in the proof of theorem [ theo2 ] @xmath391 with korn s inequality , @xmath392 in @xmath58 . \(ii ) _ the scalar case in @xmath1 . _ by theorem [ theo5 ] , for @xmath374 , @xmath393 consequently , @xmath394 leading to @xmath395 . since @xmath286 , the formula for @xmath40 has now the same structure as for @xmath2 and the proof follows exactly as in the first part ( i ) . \(iii ) _ the case @xmath10 . _ by assumption , @xmath396 , which results in @xmath397 in @xmath58 . in addition , for @xmath374 , by eqn . ( [ epsistardef ] ) , @xmath398 consequently , @xmath399 so again @xmath400 , @xmath378 , and the proof can be carried out as before . in this article we derived extensions of the allen - cahn / cahn - hilliard system to elastic materials . in particular we included ( i ) the linear elastic energy derived by eshelby , ( ii ) a geometrically linear theory of elasticity for @xmath6 , and ( iii ) a geometrically linear theory of elasticity for @xmath1 in the scalar setting . for all three cases we showed in a mathematically rigorous way the existence and the uniqueness of weak solutions , asserting the correctness of our approach . as a future goal , it is desirable to generalize the proof to the non - scalar , i.e. vectorial case in @xmath1 and to @xmath401 . the generalized ac - ch models contain as special cases both the allen - cahn @xcite and the cahn - hilliard equation @xcite with linear elasticity . in our work the existence and uniqueness of weak solutions to the ac - ch model with linear elasticity follows in a straightforward way since assumption ( @xmath9 ) required in the existence result theorem [ theo1 ] can easily be checked to hold true for the linear energy . this had not been shown before . we point out that for the new cases ( ii ) and ( iii ) , the formulas collected in section [ secexpl ] for the specific energy @xmath40 are essential for any numerical studies of the ac - ch models extended to microstructure . for related investigations and numerical methods we refer the interested reader to @xcite and @xcite . besides the limiting assumption of constant temperature ( cf . introduction ) , the most important pending restriction is the postulation of small strain , included in ( [ vedef ] ) . that is , it would be desirable to combine the ac - ch system with energies within the framework of the geometrically non - linear theory of elasticity . the technical problems of a large strain theory are striking , see , e.g. , @xcite , and it is in many cases not known how to compute explicit formulas for the relaxed energy functionals . similar mathematical difficulties arise when one wants to extend the ac - ch systems to elastic materials which are of polycrystalline structure . this is an open problem for future research . for related modelling aspects we refer to @xcite , for related mathematical aspects to @xcite we have worked on this project while we were employed or were visitors at the max planck institute for mathematics in the sciences , leipzig ; the institute of applied mathematics , university of bonn ; and the institute for mathematics , university of wrzburg . some part of this work was performed while tb was visiting the hausdorff research institute for mathematics , bonn . we acknowledge the hospitality of all these institutions . 99 allen , s. m. ; cahn , j. w. microscopic theory for antiphase boundary motion and its application to antiphase domain coarsening , _ acta metallurgica _ 1979 , * 27 * , 10851095 . ambrosio , l. ; fusco , n. ; pallara , d. _ functions of bounded variation and free discontinuity problems , _ clarendon press , new york , 2000 . barrett , j. w. ; garcke , h. ; nrnberg , r. a phase field model for the electromigration of intergranular voids . _ interfaces and free boundaries _ 2007 , * 9 * , 171210 . bhattacharya , k. comparison of the geometrically nonlinear and linear theories of martensitic transformation . _ contin . _ 1993 , * 5 * , 205242 . bhattacharya , k. ; kohn , r. v. elastic energy minimization and the recoverable strains of polycrystalline shape - memory materials . . rational mech . _ 1997 , * 139 * , 99180 . bhattacharya , k. ; schlmerkemper , a. stess - induced phase transformations in shape - memory polycrystals , _ arch . rational mech . _ 2010 , * 196 * , 715751 . blesgen , t. ; chenchiah , i. v. a generalized cahn - hilliard equation based on geometrically linear elasticity , _ interfaces and free boundaries _ 2011 , * 13 * , 127 . blesgen , t. the elastic properties of single crystals with microstructure and applications to diffusion induced segregation . _ crystal research and technology _ 2008 , * 43 * , 905913 . blesgen , t. ; weikard , u. on the multicomponent allen - cahn equation for elastically stressed solids . _ electr . j. diff . _ 2005 , * 89 * , 117 . blesgen , t. ; schlmerkemper , a. towards diffuse interface models with a nonlinear polycrystalline elastic energy . in : a. doughett and p. asnarez ( eds . ) , composite laminates : properties , performance and applications , nova science publishing 2010 , pp . 465 - 489 . brochet , d. ; hilhorst , d. ; novick - cohen , a. finite - dimensional exponential attractor for a model for order - disorder and phase separation . _ 1994 , * 7 * , 8387 . cahn , j. w. ; hilliard , j. e. free energy of a non - uniform system i. interfacial free energy . _ j. chem . _ 1958 , * 28 * , 258267 . cahn , j. w. ; larch , f. c. the effect of self - stress on diffusion in solids . _ acta metall . _ 1982 , * 30 * , 18351845 . cahn , j. w. ; novick - cohen , a. evolution equations for phase separation and ordering in binary alloys . _ j. stat . _ 1994 , * 76 * , 877909 . chenchiah , i. v. ; bhattacharya , k. the relaxation of two - well energies with possibly unequal moduli . rational mech . _ 2008 , * 187 * , 409479 . cherkaev , a. v. variational methods for structural optimization , _ appl . _ 2000 , * 140 * , springer publishing . cioranescu , d. ; donato , p. _ an introduction to homogenization , _ oxford university press , 1999 . dacorogna , b. _ direct methods in the calculus of variations , _ springer , berlin , 1989 . eshelby , j. d. elastic inclusions and inhomogeneities . solid mech . _ 1961 , * 2 * , 89140 . fratzl , p. ; penrose , o. ; lebowitz , j. l. modelling of phase separation in alloys with coherent elastic misfit . _ j. stat . _ 1999 , * 95 * , 14291503 . garcke , h. on cahn - hilliard systems with elasticity . edinburgh sec . a _ 2003 , * 133 * , 307331 . gottstein , g. physical foundations of materials science . springer berlin , 2004 . kohn , r. v. ; vogelius , m. relaxation of a variational method for impedance computed tomography . pure appl . _ 1987 , * 60 * , 745777 . kohn , r. v. the relaxation of a double - well energy . _ continuum mechanics and thermodynamics _ 1991 , * 3(3 ) * , 193236 . kohn , r. v. ; niethammer , b. geometrically nonlinear shape - memory polycrystals made from a two - variant material . _ 2000 , * 34 * , 377 - 398 . milton , g.w . a brief review of the translation method for bounding effective elastic tensors of composites , _ cont . models and discrete systems maugin , ed . ) , longman scientific and technical , 1990 , 6074 . novick - cohen , a. triple - junction motion for an allen - cahn / cahn - hilliard system . _ physica d _ 2000 , * 137 * , 124 . nye , j. f. _ physical properties of crystals : their representation by tensors and matrices , _ clarendon press ; new york : oxford university press , 1984 . onuki , a. ginzburg - landau approach to elastic effects in the phase separation of solids . _ j. phys . japan _ 1989 , * 58 * , 30653068 . salje , e. phase transitions in ferroelastic and co - elastic crystals . cambridge university press , 1990 . ziemer , w. _ weakly differentiable functions : sobolev spaces and functions of bounded variation , _ graduate texts in mathematics , springer , new york , 1989 .

charged particles of a cosmic ray shower , mostly electrons , passing through the atmosphere lose energy by inelastic collisions with air molecules . a small fraction of the total deposited energy is emitted by molecular nitrogen as uv fluorescence radiation in the @xmath0 nm spectral range . the detection of this radiation provides a precise determination of the longitudinal profile allowing the reconstruction of the cosmic ray properties , in particular its energy . this technique @xcite was first successfully used by the fly s eye experiment @xcite and later by hires @xcite . fluorescence telescopes are presently being employed by the pierre auger observatory @xcite and the telescope array experiment @xcite . on the other hand , the jem - euso project @xcite and the s - euso free - flying satellite mission @xcite are being designed for the detection of fluorescence traces of air showers viewing downward from the top of the atmosphere . the ratio between number of fluorescence photons and deposited energy , i.e. , the fluorescence yield @xmath1 , is a key calibration parameter determining the energy scale of the fluorescence telescopes . a number of absolute measurements of the fluorescence yield have been carried out in laboratory experiments in the last years . at atmospheric pressure , fluorescence emission in the spectral range of interest basically comes from the second positive ( 2p ) system of n@xmath2 and , to a much less extent , the first negative ( 1n ) system of n@xmath3 . each one results from the transition between two electronic states and consists of a set of molecular bands corresponding to different combinations of vibrational levels @xmath4-@xmath5 of the respective upper and lower states . excitation cross section of the 1n system decreases with energy in the @xmath6 ev range showing a very smooth growing behavior at larger energies . that of the 2p system peaks at about 15 ev decreasing strongly with an @xmath7 dependence . as a consequence , high - energy electrons themselves are very inefficient for the generation of air fluorescence . in fact , as is well known , fluorescence emission along the track of an energetic primary electron is mainly induced by low - energy secondary electrons ejected in successive ionization processes . the efficiency for fluorescence emission varies with the energy spectrum of these secondaries , which in principle depends on the primary energy . however , a theoretical calculation using a detailed monte carlo simulation @xcite has demonstrated that the fluorescence yield is nearly constant for energies above a few kiloelectronvolts . this @xmath8-independent behavior of the fluorescence yield has been checked experimentally @xcite . a number of experimental values of the absolute fluorescence yield are available in the literature . essentially , in these experiments a beam of electrons crosses a collision chamber filled with air at known conditions . however different techniques can be used . many authors @xcite have used electrons from a source of @xmath9sr with energy around 1 mev . other absolute measurements have been performed with higher energy electrons from accelerators . the macfly collaboration @xcite used the cern / sps - x5 test beam facility which delivers a pulsed electron beam of about 10@xmath10 electrons per spill ( 4.8 s duration ) every 16.8 s. measurements at 20 and 50 gev were carried out using this facility . the flash collaboration @xcite used the final focus test beam facility at the stanford linear accelerator center which provided 28.5 gev electrons in 3 ps pulses of about 10@xmath11 electrons at a rate of 10 hz . the airfly collaboration is carrying out absolute measurements of the air - fluorescence yield using different accelerators . preliminary results obtained with 350 mev electrons at the beam test facility of the instituto nazionale di fisica nucleare have been presented in @xcite . new measurements using a 120 gev proton beam from the meson beam test facility at fermilab were presented at the 6th afw @xcite and very likely final results will be published soon . in this work we will compare the absolute values reported by the above authors . in many occasions comparison can not be done directly . for instance , some authors measure single intense fluorescence bands while others detect the integrated fluorescence in a wide spectral range . comparison between these experimental results can be carried out if the relative intensities of the bands along the fluorescence spectrum are known . we will normalize all the results to the most intense fluorescence band at 337 nm using accurate experimental data in full agreement with theoretical intensity ratios proposed in @xcite . on the other hand , some authors report their measurements in units of photons per electron and meter while others give directly the @xmath1 parameter in photons per unit deposited energy ( e.g. , photons / mev ) . the conversion factor between both magnitudes is the energy deposited per electron and unit path length . as it has been pointed out @xcite , an accurate determination of the deposited energy requires discounting that corresponding to secondary electrons outside the field of view of the optical system used to detect the emitted fluorescence . to evaluate this effect we have carried out a mc simulation of the various experimental setups . when possible a dedicated simulation including geometrical details has been performed . also the effect of the spatial distribution of deposited energy on the optical efficiency of the experimental setup has been evaluated in some cases . as a result of this study absolute values of the fluorescence yield normalized to that of the 337 nm band at 293 k and 800 hpa in dry air will be shown with a discussion on possible corrections to be applied . a previous comparison at 1013 hpa for some of the above measurements has been published recently @xcite . in this work we will update those results and extend the comparison to other experimental results . in the literature , several parameters have been used to measure the amount of fluorescence light induced by an electron moving in the atmosphere . a detailed description of the various magnitudes can be found in @xcite . the number of photons emitted per electron and unit path length have been used extensively . we will name this parameter as @xmath12 and , for a band @xmath13 , it is proportional to the number of nitrogen molecules per unit volume @xmath14 and the reciprocal of the @xmath15 stern - volmer factor , accounting for the collisional de - excitation of the upper level @xmath4 , where @xmath16 is a characteristic pressure depending on the gas composition and temperature . if ignoring the contribution of secondary electrons , the proportionality constant would be the so - called optical cross section . while most secondary electrons have low energy , and thus , short range at atmospheric pressure , a small number of high - energy secondaries can escape the observation region . as a consequence , in a laboratory experiment a fraction of the fluorescence light can not be detected . following the parametrization in @xcite , an effective optical cross section @xmath17 can be defined in such a way that the total number of photons of the @xmath13 band detected under given experimental conditions ( including contributions from both primary and secondary electrons ) can be expressed as @xmath18 the parameter @xmath17 depends on @xmath8 , @xmath19 and the geometrical features of the experimental setup . therefore , even for the same primary energy , the measured @xmath20 value depends on the particular experiment . in addition , the pressure dependence of @xmath20 as would be predicted by the stern - volmer law is distorted in a geometry - dependent way . a further pressure dependence arises from the vibrational relaxation of excited nitrogen induced by collisions with surrounding nitrogen molecules in the ground state @xcite . the effect of vibrational relaxation is twofold . firstly , it contributes to the total quenching of a given level @xmath4 and thus is included in the stern - volmer factor ( [ eq : varepsilon ] ) . secondly , vibrational relaxation of upper levels ( @xmath21 ) provides an additional excitation channel . although , in principle , this effect may lead to a departure from the stern - volmer law , significant deviations have not been appreciated experimentally for any n@xmath2 band in air @xcite . in fact , for the 2p(0,0 ) band , for which detailed data on vibrational quenching are available @xcite , a simple calculation shows that the only relevant effect of vibrational quenching is a variation of the resulting characteristic pressure respect to the one from pure electronic quenching @xcite . the number of photons per unit energy deposited in the medium is a more appropriate parameter for use in fluorescence detectors of cosmic ray showers . following an already common usage , we will keep the term fluorescence yield to refer to this magnitude @xmath1 . for a given band , @xmath20 and @xmath22 are related by @xmath23 where @xmath24 is the energy deposited per primary electron and unit path length . in the above equation , both the emitted photons and the deposited energy should correspond to the same volume , which is defined by the field of view of the experimental detection system . note that a significant fraction of energy is deposited by secondary electrons which move away from the path of the primary electron . it has been shown @xcite that @xmath25 and @xmath24 are proportional , resulting @xmath22 to be given by @xmath26 where the @xmath27 parameter is independent of the experimental conditions and represents the fluorescence yield at null pressure , i.e. , in the absence of quenching . while measurements of the fluorescence yield at atmospheric pressure by different authors show a reasonable agreement , the available data on the @xmath16 parameters show large discrepancies due to secondary electrons @xcite . ] . therefore the extrapolation of the fluorescence yield to null pressure is unsafe , making the @xmath27 parameter unsuitable for our proposes . in this work , comparison will be thus performed using the fluorescence yields measured at given atmospheric conditions . the energy deposited by an electron per unit path length within a given volume can be expressed @xcite as @xmath28\sigma_{\rm ion}\,,\ ] ] where @xmath29 is the density of air molecules and @xmath30 is the ionization cross section , including the density correction for high energies . the parameters @xmath31 and @xmath32 are the average values of the energy deposited per ionization process within the observation volume by the primary electron itself and the secondaries , respectively . for electron energies above 410 ev , contribution from x - rays arising from k - shell ionization should be also added to the energy deposition @xcite . the @xmath31 parameter can be calculated from @xmath33 where @xmath34 and @xmath35 are the average excitation energy and the ionization potential of the ionized molecule , respectively , both contributing to the energy deposition of the primary electron after an ionizing interaction . the first term in the right hand of the equation accounts for the contribution from direct excitations without ionization , being @xmath36 the average excitation energy of the molecule and @xmath37 the excitation cross section for a primary electron . the parameters involved in the above expressions are either available in the literature or can be inferred from molecular data , except for @xmath32 , which has to be calculated from a simulation of the secondary processes inside the observation volume . next we will describe a monte carlo algorithm to calculate both the energy deposited by secondary electrons and their contribution to the fluorescence emission for a simple geometry . in section [ sec : detailed_simulation ] a more detailed simulation including the geometrical features of several experiments will be presented . ignoring fine geometrical details , the observation region of an experimental setup can be described by a sphere of radius @xmath38 . in addition , we will assume that primary electrons interact at its center from where secondaries are ejected and tracked until they escape the sphere or are stopped . a monte carlo algorithm for the calculation of both deposited energy and fluorescence emission in this simple geometry has been described in detail in @xcite . next the main features and ingredients of the simulation will be outlined and some improvements described . the algorithm generates secondary electrons with energies distributed as given by the extended opal formula proposed in @xcite , which depends on the input primary energy . each generated electron takes a step of random size with mean value @xmath39 , where @xmath40 is the total cross section for the electron energy . after this step , the type of process is randomly chosen between elastic , excitation , ionization and bremsstrahlung , being their relative weights the corresponding cross sections . while there are no energy losses in elastic interactions , in excitation and ionization processes the electron is assumed to locally deposit @xmath36 and @xmath41 , respectively . in case of ionization , the electron loses an additional amount of energy to generate a new secondary , whose energy is randomly calculated from the above - mentioned extended opal formula . with a probability corresponding to the cross section for k - shell ionization , also a 410 ev x - ray can be generated at the interaction point , producing another ionization at a random distance of mean value given by the photon attenuation coefficient from @xcite . in case of bremsstrahlung , the energy loss is estimated as 3% of the electron energy and the produced photons are assumed to escape the sphere . if an excitation ( ionization ) process occurs , the amount of 2p ( 1n ) fluorescence light emitted is calculated from the ratio of the 2p ( 1n ) cross section to the total excitation ( ionization ) one . the direction of the electron after an interaction is calculated from an approximated differential cross section @xcite . the stepped trajectory of each electron is so iteratively simulated until it escapes the sphere or its energy is below 11 ev ( i.e. , the threshold for fluorescence production ) , when the electron is considered to be stopped and deposits its remainder energy . new secondary electrons arising from ionization processes are added to the list of electrons to be simulated following the same procedure . input parameters in the algorithm are the energy of the primary electron , the density of air molecules ( or pressure at given temperature ) and the radius of the sphere . as a result of the simulation the value of @xmath32 is obtained , and then , the total energy deposited per primary path length is calculated from expressions ( [ eq : edep ] ) and ( [ eq : e0dep ] ) . likewise , the @xmath42 parameter can be calculated from the average number of fluorescence photons emitted inside the sphere in the absence of quenching resulting from the simulation . as expected , both magnitudes turn out to be a function of the product @xmath43 , that is , proportional to the number of molecules that a secondary electron encounters in its way out of the sphere , and the quotient provides the @xmath27 parameter , which is basically independent of @xmath8 and @xmath43 . as discussed in @xcite , our calculation of the fluorescence emission is subjected to a systematic uncertainty , that we estimate in about a 25% , due to the uncertainties in the various molecular parameters . however the evaluation of the energy deposited in the sphere is much more accurate . in fact , in the @xmath44 limit , for which no secondary escapes , the predictions from the simulation are in perfect agreement with the bethe - bloch formula for the collisional energy loss @xcite including the corresponding density correction . for this result , a crucial ingredient in the simulation is the energy distribution of secondary electrons . the extended opal formula proposed in @xcite provides this spectrum for a very wide energy interval ( ev - gev ) of both primary and secondary electrons . simulation results of the deposited energy for several @xmath43 values have been presented in @xcite . for the purpose of comparison , in figure [ fig : edep_comparison ] the energy deposited by an electron per unit mass thickness , which is a factor of two lower than the mass number @xmath45 for both nitrogen and oxygen . ] in dry air at atmospheric pressure is represented against the primary energy for @xmath46 cm and 10 cm . below 0.5 mev the fraction of energy deposited outside the sphere is negligible at this pressure and therefore the energy deposition for both @xmath38 values approaches the energy loss . as can be seen in the figure , the deposited energy is very weakly dependent on the size of the observation region at given pressure . for instance , the energy deposited by electrons of 1 mev ( 1 gev ) increases by 9% ( 6% ) when @xmath38 increases from 1 to 10 cm at atmospheric pressure . this justifies that a reasonable accuracy in the deposited energy can be achieved even though the fine geometrical details of the experimental setup are not taken into account . notice that in our algorithm , individual electron - molecule collisions as well as the emission of 1n and 2p fluorescence are simulated with a cutoff energy of 11 ev . on the contrary , usual mc codes ( e.g. , geant and egs ) transports particles in macroscopic steps adding up the energy deposition in each step . fluorescence emission , which is mostly generated by electrons with energies of about a few tens of electronvolts , can not be calculated by these codes unless the user assumes a certain value of the fluorescence yield . the experimental determination of the fluorescence yield requires the calculation of the energy deposited in the observation volume . consequently , any possible error in the evaluation of this magnitude translates to the fluorescence yield . several experiments have provided results on their simulations which can be compared with our predictions for typical sizes of experimental setups ( see figure [ fig : edep_comparison ] ) . cm and @xmath47 cm are shown ( solid lines ) for comparison with the energy deposition values reported by other authors ( symbols and broken lines ) . see text for details . ] in their paper of 2006 the flash collaboration @xcite showed results on deposited energy versus @xmath8 for a 1 cm thick slab of air at atmospheric pressure obtained with geant3 , which are significantly larger than our predictions for @xmath46 cm . in a more recent paper @xcite , they obtained an energy deposition for 28.5 gev electrons crossing a 1.67 cm air gap using egs4 which is closer to ours . in next section a detailed comparison is shown . the macfly collaboration @xcite , using geant4 , reported a @xmath24 curve obtained for a cylindrical chamber , 15 cm in diameter and 28.8 cm long . its results above 2 mev are in fair agreement with our predictions for a sphere of radius 10 cm at atmospheric pressure , except for the highest energies ( @xmath48 mev ) , where our values are up to 6% larger . at lower energies the values are oddly much larger than expected ( see also the next section for more details ) . finally , the airfly collaboration , using geant4 , has provided us with the results of a dedicated simulation for our simple geometry ( sphere of radius 1 cm and primary electrons are forced to interact in its center ) @xcite , which are fully compatible with ours within statistical fluctuations . in this section , a simulation including the particular geometry for several experiments as well as the propagation of beam electrons will be presented . that allowed the calculation of the energy deposited in the corresponding observation volume . when possible , the average geometrical acceptance for fluorescence detection has been evaluated and compared with that assumed by the authors of the experiment . for that purpose , the solid angle subtended by the optical system s entrance pupil is averaged along the electron trajectories , weighting with the amount of light emitted at each point . note that the effect on the detection efficiency due to the spread of electrons in the interaction volume is presumably small , as it results from the average of a distribution with axial symmetry around the beam axis . that is , the enhanced fluorescence contribution from electrons approaching the fluorescence detector is nearly compensated by the reduced one from those moving in the opposite direction . in addition , in those cases where the electron - photon coincidence technique is used ( e.g. , with a radioactive source ) , the detection of beam electrons is required in the simulation , and thus , primaries not firing the detector are rejected . this may affect the energy spectrum of contributing electrons , and so , the determination of the deposited energy . we have performed this detailed simulation for the experiments of nagano @xcite , airlight @xcite , flash-2008 @xcite and macfly @xcite . experiments of kakimoto @xcite and lefeuvre @xcite will be also discussed although a dedicated simulation was not carried out . for nagano s setup @xcite , we made a detailed simulation in order to determine the effect of several approximations applied in this experiment and then evaluate a correction to their results . the collisional chamber of this experiment is shown in figure [ fig : nagano_chamber ] , where the geometry implemented in this work is emphasized . they used the electron - photon coincidence technique , so this feature has been also implemented in our simulation . electrons fire a scintillator at a distance of about 16 cm from the @xmath9sr source . however , only the 5 cm gap between the collimator and the scintillator enclosure is in the field of view of the photon detectors . in this setup the observation volume corresponds to the chamber volume excluding the collimation channel and the scintillator enclosure . sr source are beamed and detected by a scintillator counter . between the collimator and the scintillator enclosure there is a 5 cm gap in the field of view of the photon detectors . the thick contour represents the geometry implemented in our simulation . the energy deposition obtained from our simulation including these geometrical details is fully compatible with the prediction for a sphere of 5 cm radius ( circle ) . ] as input to the simulation , we have used a truncated gaussian fit ( solid line in figure [ fig : nagano_spectrum ] ) to the measured @xmath9sr spectrum of @xcite , with mean energy of 0.85 mev , end point at about 2 mev and a threshold of 0.3 mev due to the material covering the radioactive source . according to our simulation , most beam electrons are scattered away at atmospheric pressure and so are not detected . this occurs preferably for low - energy primaries , and therefore the spectrum of contributing electrons is shifted to higher energies , with a mean value of 1.11 mev at atmospheric pressure ( see figure [ fig : nagano_spectrum ] ) . the approximations made by nagano are the following . firstly , they assumed that all the fluorescence light is emitted from the beam itself . however , as explained above , a non - negligible fraction of fluorescence is produced by secondary electrons at some distance , and thus , the average geometrical acceptance @xmath49 may be different from that which would be obtained if all the light were generated in the beam @xmath50 . secondly , they calculated the @xmath12 value assuming the photons are emitted in a length equal to the gap distance @xmath51 cm . however , due to the fluctuating trajectories of beam electrons , the mean path length @xmath52 should be somewhat larger . finally , for the calculation of the fluorescence yield from ( [ eq : y ] ) , they assumed that the energy deposited in the observation volume equals the collisional energy loss @xmath53 at 0.85 mev as given by the bethe - bloch formula . as discussed above , the deposited energy is expected to be smaller than the total energy loss . therefore , the fluorescence yield value @xmath54 reported by nagano , should be corrected by three factors accounting for the above described approximations @xmath55 from our detailed simulation , we have obtained the average geometrical acceptance relative to that assumed in @xcite , that is , the @xmath56 ratio in the above expression . for this calculation , we have taken into account that some regions in the observation volume are hidden to the optical system . we have found that this correction to the efficiency is very small , as expected , resulting @xmath49 to be about 1% lower than @xmath50 at atmospheric pressure . the correction is even smaller at lower pressure . on the other hand , according to our simulation , the mean path length of contributing electrons at atmospheric pressure is about 1% longer than the gap length . therefore , these small corrections happen to compensate each other in equation ( [ eq : nagano ] ) for this experiment . in fact , the nagano s results on @xmath12 ( photons / meter ) turn out to be unaffected by these approximations . the main correction to the fluorescence yield in equation ( [ eq : nagano ] ) arises from the approximation in the energy deposition . from our simulation , we have obtained the average energy deposited per electron and unit path length @xmath57 as the ratio between the integrated energy deposition and the total path length of beam electrons within the observation volume . simulation data were accumulated up to reach a statistical uncertainty of less than 1% in the @xmath58 value . as a result of this correction , the fluorescence yield at atmospheric pressure from @xcite should be increased by 7% . the result of the deposited energy from our simulation of the nagano s experiment is fully compatible with the predictions from the generic simulation for a simple geometry , assuming a sphere of radius @xmath59 cm ( see figure [ fig : nagano_chamber ] ) and an electron energy of around 1 mev . as already mentioned , deposited energy is very weakly dependent on @xmath38 . in addition , at these energies around the minimum of the bethe - bloch formula , the result is nearly independent of @xmath8 . therefore , for experiments similar to that of nagano , the generic simulation may provide reliable results on the energy deposited in the observation volume . the experimental technique used in airlight @xcite is similar to the one of nagano , but with a collisional chamber larger by about a factor of two . electrons from a @xmath9sr source are collimated to form a narrow beam of 5 mm diameter . between the collimator and the electron detector ( a scintillator of 4 cm diameter ) there is a 10 cm gap within the field of view of seven photomultipliers . in @xcite a detailed simulation of the experiment is carried out using geant4 , including electron scattering by elements of the chamber ( e.g. , collimator walls , scintillator ) . both beam electrons and secondary ones are thoroughly tracked allowing the authors the determination of the geometrical acceptance as well as the energy deposited in the observation volume . for this experiment we have carried out a simulation including primary propagation and the geometrical details of the setup . electron scattering by other elements of the chamber apart from air is not included . the aim of our simulation has been the calculation of the integrated energy deposition as a function of the electron energy to be compared with the results found by airlight . while at low pressure we have obtained results in fair agreement with those reported in @xcite ( within 5% at 50 hpa ) , at atmospheric pressure our simulation predicts a larger energy deposition . for instance at 800 hpa , deviations range from about 5% at low electron energy up to 20% at 2000 kev ( see figure [ fig : airlight ] ) . the origin of this discrepancy is still unclear . it should be indicated that electron scattering by components of the collision chamber ( included in their simulation , but not in ours ) would result indeed in an increase of the energy deposition , as pointed out in @xcite . from our simulation we have also calculated the integrated value of the collisional energy loss . at atmospheric pressure we obtain results somewhat larger than those reported by @xcite , in particular at low electron energy ( see figure [ fig : airlight ] ) . the energy loss computed by airlight seems to have been calculated as the product of the bethe - bloch stopping power @xmath53 @xcite times the 10 cm gap length , and thus , assuming that the stopping power is constant along the electron path . however , as already mentioned for nagano s experiment , due to the fluctuating trajectories of primaries , their mean path length is expected to be somewhat larger than the gap length . also , at low electron energy , the average stopping power should be larger than the initial value , because primaries lose a significant fraction of their energy in their path . obviously , the energy deposition must be smaller than the energy loss , and their difference is expected to diminish at low @xmath8 values . this behavior is found in our simulation as well as in that of @xcite . however , at 800 hpa , the energy loss from @xcite is smaller than our prediction of the energy deposition for electron energies below 400 kev ( figure [ fig : airlight ] ) showing an evident inconsistency between both calculations . assuming a mean electron energy of 1 mev , our detailed simulation predicts an average deposited energy around 10% higher than that of @xcite at 800 hpa , and thus , if our results were correct , their fluorescence yield value at this pressure should be reduced in principle by the same factor . however , airlight reports @xmath27 results ( together with a set of quenching parameters ) by averaging measurements carried out at several pressures . ] . in fact the average @xmath27 values are nearly equal to those obtained at 400 hpa . at this pressure and an electron energy of 1 mev , the difference between our energy deposition and that of airlight is about 7% . therefore , according to our simulation , results on fluorescence yield from @xcite should be reduced by an overall correction factor of 7% irrespective of pressure . in this setup @xcite , an intense beam of 28.5 gev electrons from a linear accelerator crosses a cylindrical collision chamber , 25 cm in length and 15 cm in diameter . the beam is coaxial to a pair of 1.6 cm diameter cylinders with a 1.67 cm gap between them , which defines the beam path length as seen by two pmts . each detector is mounted at the end of a long channel perpendicular to the beam axis and provided with baffles to suppress unwanted scattered light . this channel terminates , at 45 cm from the beam , in a 1.2 cm diameter window that forms the entrance pupil of the optical system . in the opposite side , there is a shorter channel holding a led used to monitor the pmt gain stability , being the extreme of this channel at about 27 cm from the beam ( 72 cm from the window of the pmt ) . the absolute calibration of the optical system was carried out by comparison with the measurement of rayleigh - scattered light from a narrow beam of a nitrogen laser . as pointed out above , the flash collaboration used an egs4 simulation to obtain the mean energy deposited per electron in the observation volume of their experiment @xcite . their simulation also provided the average geometrical acceptance for fluorescence detection relative to that for the calibration laser beam . we have carried out a detailed simulation of this experiment in order to evaluate the energy deposited in the chamber as well as the average geometrical acceptance for comparison with their results . in our algorithm , the observation volume has been approximated by a rectangular prism of dimensions 1.60 @xmath60 @xmath61 cm . the deposited energy is not explicitly given in @xcite , but it can be inferred from the ratio between the reported results on @xmath12 and @xmath1 at pressures ranging from 67 to 1013 hpa . in figure [ fig : flash_08 ] , the resulting values are shown as a function of pressure together with those from our detailed simulation , where statistical uncertainties were reduced to less than 0.3% . at high pressures ( @xmath62 hpa ) , both results are in good agreement ( within 2% ) . however , the pressure dependency is different , resulting in a deviation of 6% at 67 hpa . this discrepancy could be due to a different treatment of the density correction , which at such a high electron energy is very important , lowering the inelastic cross sections by about 40% at atmospheric pressure . as a simple test , we have recalculated the deposited energy using in our simulation the density correction for 1013 hpa in the whole pressure range . the resulting energy deposition , also presented in figure [ fig : flash_08 ] , shows a behavior similar to that of @xcite . in regard with the geometrical acceptance , flash @xcite found that the efficiency of their optical system for fluorescence detection is @xmath63% lower than that for the calibration laser beam , due to the wide - spread energy deposition from the electron beam . however , our simulation predicts that this effect is negligible in the whole pressure range for this setup . notice that according to our simulation both corrections , i.e. , deposited energy and optical efficiency , nearly cancel and therefore the fluorescence yield at 1013 hpa from @xcite would be almost unaffected ( see table [ tab : comparison ] in section [ sec : results ] ) . in this experiment @xcite , the collision chamber is a cylinder 28.8 cm long and 15 cm in diameter with a mirrored inner surface in order to increase the geometrical acceptance . they used the coincidence technique , with individual electrons either from a linear accelerator ( 20 and 50 gev ) or from a @xmath9sr source of mean energy 1.5 mev . as mentioned above , macfly calculated the energy deposited per electron and per unit mass thickness as a function of @xmath8 using geant4 ( blue dotted line in figure [ fig : edep_comparison ] ) . we have carried out a detailed simulation to obtain the energy deposition . primaries colliding the walls of the chamber or being stopped inside , and so not reaching the trigger system , are rejected in the simulation . the average energy deposited per electron and unit mass thickness has been obtained from the ratio between the integrated energy deposition and the total path length of triggering primaries . statistical uncertainties are reduced below 0.5% . the deposited energy obtained from our simulation is slightly lower ( 2% ) than the value reported by macfly at 1.5 mev and larger by about 6% at 20 and 50 gev , and therefore , according to our calculations the fluorescence yield of macfly should be increased by 2% at 1.5 mev and decreased by 6% at 20 and 50 gev . the above described detailed simulation is in good agreement ( better than 2% ) with the generic simulation for @xmath47 cm at energies above 0.5 mev ( see figure [ fig : edep_comparison ] ) . below 0.5 mev , primary electrons lose a significant fraction of their energy in the chamber ( e.g. , electrons with energy lower than 0.2 mev are fully stopped ) , and consequently the @xmath58 values resulting from our detailed simulation are somewhat larger ( @xmath64 ) than those obtained from the generic one . however this effect can not justify the very large discrepancies with the macfly results at low electron energy shown in figure [ fig : edep_comparison ] . apart from those discussed in the above sections other well - know experiments have been carried out , in particular those of kakimoto @xcite and lefeuvre @xcite . for these we have not performed a detailed simulation but we will use the result of our generic simulation . in the kakimoto s experiment electrons with a mean energy of 1.4 mev from a @xmath9sr source as well as those from an electron synchrotron with energies of 300 , 650 and 1000 mev were used . fluorescence light was produced and observed in a 10 cm gap . for the determination of the fluorescence yield , they assumed full energy deposition in the beam axis . according to our generic simulation , the deposited energy inside a sphere of 10 cm radius is about 6% , 25% , 28% and 29% lower than the energy loss for 1.4 , 300 , 650 and 1000 mev electrons , respectively . therefore , the fluorescence yield values at the above energies should be increased correspondingly by these factors . results obtained from applying these corrections to measurements of kakimoto are consistent with those obtained in @xcite using an alternative procedure . in the experiment of lefeuvre @xcite , electrons with mean energies of 1.1 mev ( or 1.5 mev when applied a higher energy threshold to the electron detector ) from a very intense @xmath9sr source ( 370 mbq ) were used . in order to shield the fluorescence detector from the large amount of x - radiation produced at such a high electron rate , both the radioactive source and the scintillator counter were placed inside a solid lead cylinder , 10 cm in diameter . electrons trajectories were confined within a cone dug in the lead and coaxial to the cylinder , being the @xmath9sr source at the apex and the scintillator on the base . a 4 cm diameter hole was bored through the lead cylinder , allowing fluorescence to be measured from both sides . in order to obtain the average geometrical acceptance and the mean path length of electrons , they performed a geant monte carlo simulation including electron scattering by the lead walls . they found that this feature has an important role in this experiment , since half of primaries reach the scintillator after hitting the cone walls . from their simulation , they also found that a very small fraction of secondary electrons have enough energy to escape the observation region at atmospheric pressure . in particular , they obtained that only 1% of secondaries have energies above 5 kev , and 33% of them are lost in the lead before having produced any fluorescence . from these simulation results , they inferred that the deposited energy nearly equals the total energy loss and a minor correction of 0.4% was applied . however , as discussed above , a significant fraction of energy is deposited by these few energetic secondaries , and thus , the correction is expected to be larger . from our generic simulation assuming a sphere of 4 cm radius for lefeuvre s experiment , we have obtained that the deposited energy is 9% and 10% lower than the total energy loss at mean electron energies of 1.1 and 1.5 mev , respectively . however , electron scattering by the lead shield of @xcite is not included in our simulation , which assumes that secondaries reaching the walls are absorbed , and thus , it may underestimate the deposited energy in this case . we have evaluated this effect in order to estimate a correction to our simulation results for this experiment . using casino2.42 @xcite , we have found that , for a @xmath65 incident angle and energies in the 1 kev - 1 mev range , about 60% of electrons are backscattered by lead , having lost almost half of their energy on the average . assuming that backscattered electrons deposit all their remainder energy in the air volume , about 30% of the energy of electrons reaching the lead walls is recovered , and thus , should be added to the total energy deposition . as a consequence , the fluorescence yield values at 1.1 and 1.5 mev from @xcite should be increased by about 7% according to our calculations . the wavelength normalization can be carried out if the relative intensities of the molecular bands within the experimental interval are known . as is well known , relative intensities of bands sharing the same upper level are given by the relative einstein coefficients @xmath66 . as described below fluorescence intensities of bands with different upper levels can be also calculated from molecular parameters . in the absence of other population mechanisms , the effect of vibrational relaxation of the 2p upper levels can be considered as included in the @xmath67 quenching parameters . ] apart from direct excitation of the vibrational level @xmath4 from the ground state x , the band intensities of a given system are proportional to the excitation cross sections of the corresponding upper levels . at high electron energy ( i.e. , in the born - bethe high - energy limit ) , these cross sections are known to be approximately proportional to the franck - condon factors @xmath68 , which correspond to the probability ratios for optical transitions . it has been found @xcite that this approximation can be also applied in our case , in spite of the fact that most fluorescence light is generated by low - energy secondaries . under this approximation , for a given band system , the relative intensities @xmath69 with respect to a reference one @xmath70 ( i.e. , the most intense band of the system , located at 337 nm for the 2p system and at 391 nm for the 1n system ) are given by @xmath71 at atmospheric conditions , where @xmath72 can be assumed , the above expression becomes @xmath73 where any pressure or temperature dependence is cancelled , since the same approximate @xmath74 growing behavior can be assumed for all @xmath16 parameters . relationship ( [ eq : relative_intensities_limit ] ) , and hence the applicability of the franck - condon factors , has been verified using experimental data on relative intensities . in figure 2 of ref . @xcite we plotted the relative intensities measured by airfly @xcite against @xmath75 for the 2p system , using the tabulated values of @xmath76 and @xmath77 from @xcite and the experimental characteristic pressures for dry air reported in @xcite . a remarkable linearity was found , in particular for the bands with upper vibrational level @xmath78 , contributing to almost 90% of the total 2p fluorescence . a similar analysis was previously performed in @xcite using the relative intensities and @xmath16 values measured by nagano @xcite . although linearity was also observed , significant fluctuations were found very likely due to much larger experimental uncertainties than those of @xcite . expression ( [ eq : relative_intensities_limit ] ) can be extended for molecular bands belonging to different systems including the ratio between the respective effective cross sections @xmath79 , defined as @xmath80 @xcite . unfortunately , whereas the franck - condon factors and the einstein coefficients are known very precisely , the available data on @xmath79 and @xmath16 are less reliable . effective cross sections depend on molecular parameters with significant uncertainties and are geometry dependent , and the available experimental data on characteristic pressures show large discrepancies . however , as already pointed out in @xcite , the relative intensities measured by different authors , which are proportional to @xmath81 , are in good agreement . for instance , for the @xmath82 ratio at atmospheric pressure , values of @xmath83 , @xmath84 , @xmath85 and @xmath86 result from data from @xcite , @xcite , @xcite and @xcite , respectively . even a very consistent 3.5 value was measured by bunner @xcite for 4 mev @xmath87-particles as well . therefore , the value of the @xmath88 products from experimental relative intensities is very reliable . in order to compare absolute results on fluorescence yield from different authors , we will reduce values measured in a given wavelength interval @xmath89 to the yield at the 337 nm band . experimental results are then normalized multiplying by the ratio between the relative intensity of the 337 nm band and the sum of all the bands within @xmath89 . this @xmath90 ratio can be thus written in terms of the empirically reliable @xmath82 ratio as @xmath91^{-1}\,,\ ] ] where relative intensities at atmospheric conditions follow ( [ eq : relative_intensities_limit ] ) separately for each band system . note that , due to the 2p dominance , even an error of a factor 2 in the @xmath82 ratio results in an uncertainty below 10% in @xmath90 for the spectral ranges of the experiments considered here . for this wavelength reduction , we use accurate experimental relative intensities from @xcite in the 284 - 429 nm spectral range , including some bands of the gaydon - herman ( gh ) system @xcite ( 1.6% total contribution ) . for instance , for the whole wavelength interval of @xcite a ratio @xmath92}=0.2569\pm 0.0022 $ ] is obtained , where the uncertainties on the relative intensities are treated as uncorrelated as the authors suggest . note that this value is slightly lower than those listed in table 1 of ref . @xcite , where intensities of bands of the gh system were neglected . in addition , equations ( [ eq : relative_intensities_limit ] ) and ( [ eq : ratio_337_interval ] ) allow the determination of @xmath90 for a wavelength interval including some of the 2p and 1n weak bands outside the spectral range of @xcite . in table [ tab : comparison ] absolute fluorescence yield values at atmospheric pressure reported by several authors ( first column ) are put in comparison . the wavelength interval , pressure , temperature and electron energy at which absolute values are reported are shown in columns 2 - 5 . the experimental results , in units of either photons / mev or photons / m , and the corresponding uncertainties provided by the authors are quoted in columns 6 and 7 . in order to get values normalized to that of the 337 nm band we have applied the procedure shown in section [ sec : wavelength_interval ] . the corresponding normalization factors are displayed in column 8 . the last column shows the fluorescence yield for the 337 nm band in units of photons / mev . for an easy comparison , all results are given for 800 hpa and 293 k ( dry air ) . for this normalization we have used expression ( [ eq : y0 ] ) assuming @xmath93 and @xmath94 . the numbers in the left side of this column correspond to the @xmath95 value obtained using the energy deposition and geometrical factors assumed ( or calculated ) by the authors . finally , in the right side ( in bold ) we show the results we have obtained applying the various corrections discussed in section [ sec : detailed_simulation ] . * 9c experiment & @xmath89 ( nm ) & @xmath19 ( hpa ) & @xmath96 ( k ) & @xmath8 ( mev ) & experimental result & error & @xmath90 & @xmath95 ( ph / mev ) + & 337 & 800 & 288 & 1.4 & 5.7 ph / mev & 10% & 1 & 5.8 / * 6.1 * + & & & & 1.4 & 3.3 ph / m & & & 5.7 / * 6.0 * + & & & & 300 & 4.9 ph / m & & & 5.6 / * 7.0 * + & & & & 650 & 4.4 ph / m & & & 4.8 / * 6.1 * + & & & & 1000 & 5.0 ph / m & & & 5.4 / * 6.9 * + nagano @xmath97 $ ] & 337 & 1013 & 293 & 0.85 & 1.021 ph / m & 13% & 1 & 6.4 / * 6.8 * + & & & & 1.1 & 3.95 ph / m & & & 6.5 / * 7.0 * + & & & & 1.5 & 4.34 ph / m & & & 7.1 / * 7.7 * + & & & & 1.5 & 17.0 ph / mev & & & 5.5 / * 5.6 * + & & & & @xmath98 & 17.4 ph / mev & & & 5.6 / * 5.3 * + & & & & @xmath99 & 18.2 ph / mev & & & 5.9 / * 5.5 * + flash @xmath100 $ ] & 300 - 420 & 1013 & 304 & @xmath101 & 20.8 ph / mev & 7.5% & 0.272 & 7.0 / * 6.9 * + airlight@xmath102 $ ] & 337 & - & - & 0.2 - 2 & @xmath103 ph / mev & 16% & 1 & 7.4 / * 6.9 * + airfly@xmath104 $ ] & 337 & 993 & 291 & 350 & 4.12 ph / mev & - & 1 & 5.1 / * - * + @xmath105the listed experimental result is the fluorescence yield at null pressure reported by the authors which , together with their measured quenching parameters , allows the evaluation of the fluorescence yield at any atmospheric condition . @xmath106preliminary result . [ tab : comparison ] as explained in section [ sec : simple_geometry ] , our simulation gives us both deposited energy and fluorescence emission and thus we can provide a theoretical result of the fluorescence yield @xcite . the updated calculations of this work lead us to a @xmath107 value of 390 ph / mev , nearly independent of the electron energy above 0.1 mev . this theoretical result in combination with the characteristic pressure for the 337 nm band in dry air @xcite gives rise to a @xmath95 value of 7.5 ph / mev at 800 hpa and 293 k , which is close to the experimental ones shown in the table . however , as pointed out above , uncertainties in the simulated fluorescence emission are large , and thus , this theoretical result can not be considered as a reference one . a direct comparison of the absolute measurements of the air fluorescence yield available in the literature is not possible since very often the reported data correspond to different wavelength intervals and the magnitude is expressed in different units . the conversion factors for such comparison depend on the particular features of the experimental setup . in this work we have described a procedure for the normalization of the various experimental results to a common unit and wavelength interval . in the first place , fluorescence yields have been reduced to that corresponding to the most intense band at 337 nm using accurate experimental data on intensity ratios in full agreement with theoretical relationships . absolute values can be expressed in photons / mev units as far as the energy deposited in the field of view of the setup is known . in this work we have used an upgraded simulation algorithm to calculate the energy deposition of electrons in a wide energy range . a generic algorithm has provided results as a function of the pressure and the overall size of the collision chamber . in addition , for several experiments , we have performed a detailed simulation ( in full agreement with the predictions of the generic one ) including in some cases the calculation of geometrical factors . the results of these simulations have been compared with those reported by the authors of the corresponding experiments and eventual disagreements have been discussed . as a result of this work a table of the absolute fluorescence yields normalized to 800 hpa and 293 k have been presented . according to our simulation , a non - negligible correction has to be applied to those results obtained under the assumption that the energy lost by the primary electrons is fully deposited in the field of view of the optical system ( i.e. , kakimoto @xcite , nagano @xcite ) and lefeuvre @xcite ) . in particular the nagano s fluorescence yield should be increased by about 7% . possible implications of these corrections to the @xmath1 values in the determination of the end of the energy spectrum of cosmic rays have been discussed recently in @xcite . on the other hand , our calculations of energy deposition are in general in reasonable agreement with those from experiments that carried out a detailed simulation of their setup . even though , in some cases we have found non - negligible discrepancies and correction factors have been proposed . experiments claiming high accuracy should be very careful in the evaluation of the error contribution from the deposited energy . from the above table , and assuming the corrections proposed in this work , most fluorescence yield measurements at those reference conditions are contained within the 6 - 7 ph / mev interval . notice , however that the fluorescence yield reported by macfly , as well the preliminary result of airfly , are smaller by more than 10% . this work has been supported by the spanish ministerio de ciencia e innovacion ( fpa2006 - 12184-c02 - 01 , fpa2009 - 07772 and consolider cpan csd2007 - 42 ) and comunidad de madrid " ( ref . : j. rosado acknowledges a phd grant from universidad complutense de madrid " .

extremely low mass ( elm ) wds , degenerate objects with @xmath5 ( cm s@xmath6 ) surface gravity or @xmath70.3 @xmath4 mass , are the product of common envelope binary evolution ( e.g. * ? ? ? elm wds are thus the signposts of the type of binaries that are strong gravitational wave sources and possible supernovae progenitors . the goal of the elm survey is to discover and characterize the population of elm wds in the milky way . previous elm survey papers have reported the discovery of 40 wds spanning 0.16 @xmath4 to 0.49 @xmath4 found in the hypervelocity star ( hvs ) survey , the sloan digital sky survey ( sdss ) , and in our own targeted survey @xcite . we refer to this full sample of wds as the elm survey sample , but reserve the term `` elm wd '' for those objects with @xmath5 . all of our wds are found in short - period , detached binaries , 60% of which have merger times @xmath210 gyr . three notable systems are detached binaries with @xmath240 min orbital periods @xcite . the eclipsing system j0651 + 2844 is the second - strongest gravitational wave source in the mhz range @xcite . we measured its period change in one year with optical eclipse timing @xcite . other results from the elm survey include the first tidally distorted wds @xcite and the first pulsating helium - core wds @xcite . here we present the discovery of 17 new wd binaries identified from spectra previously obtained for the hvs survey of @xcite . @xcite analyzed the visually - identified wds in the original dataset and discovered one elm wd binary @xcite . this approach failed to identify the lowest surface gravity wds . we now fit stellar atmosphere models to the entire collection of spectra not previously identified as wds , and acquire follow - up spectroscopy of new elm wd candidates . the result of this effort is that we find low surface gravity objects that might not be considered wds if not for their observed orbital motion . we chose to call objects with @xmath8 `` elm wds '' because these objects occupy a unique region of surface gravity / effective temperature space that overlaps the terminal wd cooling branch for elm wds ( see figure [ fig : teff ] ) . this region is well separated from both hydrogen - burning main sequence tracks and helium - burning horizontal branch tracks . our objects are systematically @xmath910,000 k too cool , given their surface gravities , to be helium burning sdb stars . @xcite refer to a similar type of object as a `` proto - wd . '' the lowest gravity objects in our sample may indeed have hydrogen shell burning and thus are , properly speaking , proto - wds , but here we address our sample of low gravity objects as elm wds . a variety of observations demonstrate that degenerate helium - core wds exist at our observed temperatures and surface gravities . elm wd companions to milli - second pulsars are directly observed ( e.g. * ? ? ? * ; * ? ? ? * ) at the temperatures and gravities targeted by the elm survey . the measured radius of the @xmath10 object in j0651 + 2844 , a @xmath11 @xmath12 star , demonstrates it is a degenerate wd @xcite . @xcite account for the unusually long pulsation periods of the elm wd pulsators with low mass wd models . hence , calling our low surface gravity objects elm wds is appropriate . interestingly , we find elm wds in binaries with @xmath130.9 @xmath4 companions and rapid merger times . when these detached binaries begin mass transfer , @xcite show that the extreme mass ratios will lead to stable mass transfer . for the case of massive wd accretors , theorists predict large helium flashes that may ignite thermonuclear transients dubbed `` .ia '' supernovae @xcite or may detonate the surface helium - layer and the massive wd @xcite and produce an underluminous supernova . however , the final outcome of such mergers is uncertain and they may not trigger supernovae explosions @xcite . if the companions are instead neutron stars , this would be the first time a milli - second pulsar is identified through its low - mass wd companion . such systems allow measurement of the binary mass ratio and the neutron star mass through a combination of the pulsar orbit obtained from radio timing and the wd orbit obtained from optical radial velocity observations . the importance of identifying the elm wds in the hvs survey is that the hvs survey is a well - defined and a nearly 100% complete spectroscopic survey . with a complete sample of elm wds we can measure the space density , period distribution , and merger rate of elm wds , and link elm wd merger products to populations of am cvn stars , r crb stars , and possibly underluminous supernovae . in a stellar evolution context , our elm wd survey complements studies of wd binaries with main sequence companions ( e.g. * ? ? ? * ; * ? ? ? * ) and with sdb star companions ( e.g. * ? ? ? * ; * ? ? ? sdb stars are @xmath14@xmath15 k helium - burning precursors to wds @xcite . our elm wds , on the other hand , have @xmath14@xmath16 k. we organize this paper as follows . in section 2 we discuss our observations and data analysis . in section 3 we present the orbital solutions for 17 new elm wd binaries . in section 4 we discuss the properties of the elm wd sample . we conclude in section 5 . the hvs survey is a targeted spectroscopic survey of @xmath17 stars with the colors of @xmath183 @xmath4 main sequence stars , stars which should not exist at faint magnitudes in the halo unless they were ejected there . the target selection is detailed in @xcite and spans @xmath19 , @xmath20 . this color selection fortuitously targets wds in the approximate range @xmath21 @xmath14 @xmath16 k and @xmath22 . we separate elm wds from the other stars in the hvs survey using stellar atmosphere model fits to the single - epoch spectra . our initial set of fits uses an upgraded version of the ferre code described by @xcite and synthetic da wd pure hydrogen spectra kindly provided by d. koester . the grid of wd model atmospheres covers effective temperatures from 6000 k to 30,000 k in steps of 500 k to 2000 k , and surface gravities from @xmath23=5.0 to 9.0 in steps of 0.25 dex . the model atmospheres are calculated assuming local thermodynamic equilibrium and include both convective and radiative transport @xcite . we fit 2,000 spectra , mostly from the original hvs survey @xcite but also some newly acquired data @xcite . these 2,000 spectra were not previously analyzed for the elm survey because the spectra were not previously identified as wds . we fit both flux - calibrated spectra ( for improved @xmath14 constraints ) as well as continuum - corrected balmer line profiles ( insensitive to reddening and flux calibration errors ) . for the continuum - corrected balmer line profiles , we normalize the spectra by fitting a low - order polynomial to the regions between the balmer lines . we adopt the parameters from the flux - calibrated spectra , except in cases where the spectra were obtained in non - photometric conditions . the uncertainties in our single - epoch measurements are typically @xmath24500 k in @xmath14 and @xmath25 dex in @xmath23 . from these fits we identify 57 low mass wd candidates . we obtain follow - up spectra for each of the candidate low mass wds to improve stellar atmosphere parameters and to search for velocity variability . observations were obtained over the course of seven observing runs at the 6.5 m mmt telescope between march 2011 and february 2013 . we used the blue channel spectrograph @xcite with the 832 line mm@xmath26 grating , which provides a wavelength coverage 3650 to 4500 and a spectral resolution of 1.0 . all observations were paired with a comparison lamp exposure , and were flux - calibrated using blue spectrophotometric standards @xcite . the extracted spectra typically have a signal - to - noise ( s / n ) of 7 per pixel in the continuum and a 14 km s@xmath26 radial velocity error . we obtained additional spectroscopy for @xmath27 mag elm wd candidates in queue scheduled time at the 1.5 m flwo telescope . we used the fast spectrograph @xcite with the 600 line mm@xmath26 grating and a 1.5 slit , providing wavelength coverage 3500 to 5500 and a spectral resolution of 1.7 . all observations were paired with a comparison lamp exposure , and were flux - calibrated using blue spectrophotometric standards . the extracted spectra typically have a s / n of 10 per pixel in the continuum and a 18 km s@xmath26 radial velocity error . our follow - up observations provide improved stellar atmosphere constraints , from which we determine that 24 ( 42% ) of the candidates are probable elm wds with @xmath8 ( see figure [ fig : teff ] ) . the remaining candidates have either @xmath28 and are normal da wds , or @xmath29 and are presumably halo blue horizontal branch or blue straggler stars . figure [ fig : teff ] plots the distribution of @xmath14 vs. @xmath23 for all of the wds with @xmath30 . previously published elm survey stars are marked with red circles . the observed wds overlap tracks based on @xcite models for he - core wds with hydrogen shell burning ( ( green lines ) * ? ? ? * ) , but do not overlap main sequence nor horizontal branch evolutionary tracks . for reference , we also plot @xcite solar metallicity main sequence tracks for 2 , 3 , and 6 @xmath4 stars ( cyan lines ) . a helium - burning horizontal branch star can have a surface gravity similar to a degenerate he wd ( e.g. , * ? ? ? * ) , but only at a systematically higher effective temperature than that targeted by the elm survey . this is illustrated by @xcite [ fe / h]=-1.48 horizontal branch tracks for 0.488 , 0.495 , and 0.500 @xmath4 stars ( magenta lines ) , as well as the homogenous helium - burning main sequence from @xcite . the zero - age main sequence and helium - burning horizontal branch isochrones ( thick lines ) mark the limits . seventeen of the wds show significant velocity variability , including twelve of the newly identified elm wds . the other elm wds have insufficient coverage for detecting ( or ruling out ) velocity variability . we focus the remainder of this paper on the 17 well - constrained systems . our sample of 17 well - constrained systems contains four @xmath23 @xmath31 objects , objects at the edge of our stellar atmosphere grid that might not be considered degenerate wds if not for their observed orbital motion . we also fit these stars to grids of @xcite models spanning @xmath32 } \le 0.5 $ ] , @xmath33 ( see * ? ? ? * ) but found very poor solutions . this motivated us to perform a new set of stellar atmosphere fits using @xcite hydrogen model atmospheres . the gianninas models employ improved stark broadening profiles @xcite , and the ml2/@xmath34 = 0.8 prescription of the mixing - length theory for models where convective energy transport is important @xcite . we calculate atmosphere models for low surface gravities , and use a grid that covers @xmath14 from 4000 k to 30,000 k in steps ranging from 250 to 5000 k and @xmath23 from 5.0 to 8.0 in steps of 0.25 dex . ccccccccc j0056@xmath350611 & 0:56:48.232 & -6:11:41.62 & @xmath36 & @xmath37 & 0.17 & @xmath38 & 8.0 & 0.69 + j0751@xmath350141 & 7:51:41.179 & -1:41:20.90 & @xmath39 & @xmath40 & 0.17 & @xmath41 & 8.0 & 0.75 + j0755@xmath424800 & 7:55:19.483 & 48:00:34.07 & @xmath43 & @xmath44 & 0.42 & @xmath45 & 9.7 & 0.18 + j0802@xmath350955 & 8:02:50.134 & -9:55:49.84 & @xmath46 & @xmath47 & 0.20 & @xmath48 & 8.2 & 1.19 + j0811@xmath420225 & 8:11:33.560 & 2:25:56.76 & @xmath49 & @xmath50 & 0.17 & @xmath51 & 8.0 & 1.30 + j0815@xmath422309 & 8:15:44.242 & 23:09:04.92 & @xmath52 & @xmath53 & 0.17 & @xmath54 & 6.7 & 1.53 + j0840@xmath421527 & 8:40:37.574 & 15:27:04.53 & @xmath55 & @xmath56 & 0.17 & @xmath57 & 8.0 & 1.69 + j1046@xmath350153 & 10:46:07.875 & -1:53:58.48 & @xmath58 & @xmath59 & 0.37 & @xmath60 & 10.2 & 0.36 + j1104@xmath420918 & 11:04:36.739 & 9:18:22.74 & @xmath61 & @xmath62 & 0.46 & @xmath63 & 10.3 & 0.18 + j1141@xmath423850 & 11:41:55.560 & 38:50:03.02 & @xmath64 & @xmath65 & 0.17 & @xmath66 & 8.0 & 1.56 + j1151@xmath425858 & 11:51:38.381 & 58:58:53.22 & @xmath67 & @xmath68 & 0.17 & @xmath69 & 8.0 & 2.57 + j1157@xmath420546 & 11:57:34.455 & 5:46:45.58 & @xmath70 & @xmath71 & 0.17 & @xmath72 & 8.0 & 2.29 + j1238@xmath421946 & 12:38:00.096 & 19:46:31.45 & @xmath73 & @xmath74 & 0.17 & @xmath75 & 8.0 & 0.68 + j1538@xmath420252 & 15:38:44.220 & 2:52:09.49 & @xmath76 & @xmath77 & 0.17 & @xmath78 & 8.0 & 1.28 + j1557@xmath422823 & 15:57:08.483 & 28:23:36.02 & @xmath79 & @xmath80 & 0.49 & @xmath81 & 11.2 & 0.18 + j2132@xmath420754 & 21:32:28.360 & 7:54:28.24 & @xmath82 & @xmath83 & 0.17 & @xmath84 & 8.0 & 0.96 + j2338@xmath352052 & 23:38:21.505 & -20:52:22.76 & @xmath85 & @xmath86 & 0.27 & @xmath87 & 9.0 & 1.29 + lccrrcccc j0056@xmath350611 & 33 & @xmath88 & @xmath89 & @xmath90 & @xmath91 & @xmath920.45 & @xmath930.374 & @xmath930.12 + j0751@xmath350141 & 31 & @xmath94 & @xmath95 & @xmath96 & @xmath97 & @xmath920.94 & @xmath930.182 & @xmath930.37 + j0755@xmath424800 & 26 & @xmath98 & @xmath99 & @xmath100 & @xmath101 & @xmath920.90 & @xmath930.468 & @xmath9328 + j0802@xmath350955 & 20 & @xmath102 & @xmath103 & @xmath104 & @xmath105 & @xmath920.57 & @xmath930.348 & @xmath9379 + j0811@xmath420225 & 24 & @xmath106 & @xmath107 & @xmath108 & @xmath109 & @xmath921.20 & @xmath930.142 & @xmath93160 + j0815@xmath422309 & 21 & @xmath110 & @xmath111 & @xmath112 & @xmath113 & @xmath920.47 & @xmath930.361 & @xmath93620 + j0840@xmath421527 & 19 & @xmath114 & @xmath115 & @xmath116 & @xmath117 & @xmath920.15 & @xmath930.879 & @xmath93230 + j1046@xmath350153 & 16 & @xmath118 & @xmath119 & @xmath120 & @xmath121 & @xmath920.19 & @xmath930.509 & @xmath9348 + j1104@xmath420918 & 25 & @xmath122 & @xmath123 & @xmath124 & @xmath125 & @xmath920.55 & @xmath930.831 & @xmath9339 + j1141@xmath423850 & 17 & @xmath126 & @xmath127 & @xmath128 & @xmath129 & @xmath920.76 & @xmath930.225 & @xmath9310 + j1151@xmath425858 & 17 & @xmath130 & @xmath131 & @xmath132 & @xmath133 & @xmath920.61 & @xmath930.277 & @xmath93150 + j1157@xmath420546 & 9 & @xmath134 & @xmath135 & @xmath136 & @xmath137 & @xmath920.44 & @xmath930.382 & @xmath93120 + j1238@xmath421946 & 21 & @xmath138 & @xmath139 & @xmath140 & @xmath141 & @xmath920.64 & @xmath930.266 & @xmath937.5 + j1538@xmath420252 & 16 & @xmath142 & @xmath143 & @xmath144 & @xmath145 & @xmath920.76 & @xmath930.222 & @xmath9335 + j1557@xmath422823 & 24 & @xmath146 & @xmath147 & @xmath148 & @xmath149 & @xmath920.43 & @xmath930.886 & @xmath9320 + j2132@xmath420754 & 35 & @xmath150 & @xmath151 & @xmath152 & @xmath153 & @xmath920.95 & @xmath930.179 & @xmath937.7 + j2338@xmath352052 & 25 & @xmath154 & @xmath155 & @xmath156 & @xmath157 & @xmath920.15 & @xmath930.554 & @xmath930.95 + our stellar atmosphere fits use the so - called spectroscopic technique described in @xcite . one difference between our work and @xcite is that we fit higher - order balmer lines , up to and including h12 , observed in the low surface gravity elm wds . the higher - order balmer lines are sensitive to @xmath23 and improve our surface gravity measurement . for the handful of wds in our sample with @xmath28 , we only fit balmer lines up to and including h10 since the higher order balmer lines are not observed at higher surface gravity . the lowest gravity wds in our sample show ca and mg lines in their spectra . for reference , the diffusion timescale for ca in a @xmath158 10,000 k , 0.2 @xmath4 h - rich wd is @xmath9@xmath159 yr @xcite . extreme horizontal branch stars , which can have surface gravities comparable to the lowest gravity wds , have longer @xmath9@xmath160 yr diffusion timescales @xcite . these diffusion timescales are shorter than the wd evolutionary timescale , suggesting there may be on - going accretion in these ultra - compact binary systems . near- and mid - infrared observations are needed constrain the possibility of accretion . we defer a detailed analysis of the metal abundances in elm wds to a future paper . for our present analysis we exclude the wavelength ranges where the metal lines are present in our fits . our error estimates combine the internal error of the model fits , obtained from the covariance matrix of the fitting algorithm , and the external error , obtained from multiple observations of the same object . uncertainties are typically 1.2% in @xmath14 and 0.038 dex in @xmath23 ( see * ? ? ? * for details ) . a measure of the systematic uncertainty inherent in the stellar atmosphere models and fitting routines comes from @xcite , who find a systematic uncertainty of @xmath161 dex in @xmath23 . this is corroborated by the difference we observe between the different fitting methods discussed in section 2.1 and here : the mean difference and the dispersion in @xmath14is @xmath162 and in @xmath23is @xmath163 dex . figure [ fig : teff ] plots the properties of the 17 wds with their internal errorbars . we use @xcite evolutionary tracks to estimate wd mass and luminosity . for purpose of discussion , we assume @xmath164 wds have mass 0.17 @xmath4 and absolute magnitude @xmath165 ( see * ? ? ? * ; * ? ? ? we summarize the observed and derived stellar parameters of the 17 new binaries in table [ tab : param ] . position and de - reddened @xmath166-band magnitude come from sdss @xcite and @xmath167 is our heliocentric distance estimate . we calculate orbital elements and merger times in the same way as previous elm survey papers , and so we refer the reader to those papers for the details of our analysis . in brief , we measure absolute radial velocities using the cross - correlation package rvsao @xcite and a high - s / n template . we use the entire spectrum in the cross - correlation . we then use the summed spectra ( figure [ fig : spec ] ) as cross - correlation templates to maximize our velocity precision for each individual object . we calculate orbital elements by minimizing @xmath168 for a circular orbit using the code of @xcite . figure [ fig : pdm ] shows the periodograms for the 17 binaries , and figure [ fig : vel ] plots the radial velocities phased to the best - fit orbital periods . we use the binary mass function to estimate the unseen companion mass ; an edge - on orbit with inclination @xmath169 yields the minimum companion mass @xmath170 and the maximum gravitational wave merger time . table [ tab : orbit ] summarizes the binary orbital parameters . columns include orbital period ( @xmath171 ) , radial velocity semi - amplitude ( @xmath172 ) , systemic velocity ( @xmath173 ) , time of spectroscopic conjunction ( the time when the object is closest to us ) , minimum secondary mass ( @xmath170 ) assuming @xmath169 , the maximum mass ratio ( @xmath174 ) , and the maximum gravitational wave merger time @xmath175 . the systemic velocities in table [ tab : orbit ] are not corrected for the wds gravitational redshifts , which should be subtracted from the observed velocities to find the true systemic velocities . this correction is a few km s@xmath26 for a 0.17 @xmath4 helium wd , comparable to the systemic velocity uncertainty . the orbital solutions constrain the nature of the elm wd binaries . here we discuss the systems with short merger times , massive companions , or that may be underluminous supernovae progenitors . the elm wd j0056@xmath350611 has a well - constrained orbital period of @xmath176 hr and a semi - amplitude of @xmath177 km s@xmath26 . we can calculate its likely companion mass if we assume a distribution for the unknown orbital inclination . although radial velocity detections are biased towards edge - on systems ( see section 4 ) , we will assume that we are observing a random inclination for purpose of discussion . the mean inclination angle for a random sample , @xmath178 , is then an estimate of the most likely companion mass . for j0056@xmath350611 , the most likely companion is a 0.61 @xmath4 wd at an orbital separation of 0.5 @xmath12 . this orbital separation rules out the possibility of a main sequence companion . there is no evidence for a 0.61 @xmath4 wd in the spectrum of j0056@xmath350611 , but we would not expect there to be . if the two wds in this binary formed at the same time , we would expect the 0.61 @xmath4 wd to be 15 - 100 times less luminous than the 0.17 @xmath4 wd for cooling ages of 100 myr - 1 gyr @xcite . a more plausible evolutionary scenario for an elm wd binary like j0056@xmath350611 is two consecutive phases of common - envelope evolution in which the elm wd is created last , giving the more massive wd a longer time to cool and fade ( e.g. * ? * ) . for the most probable companion mass of 0.61 @xmath4 , j0056@xmath350611 will begin mass transfer in 100 myr . @xcite discuss the many possible stellar evolution paths for such a system . this system s mass ratio @xmath1790.37 suggests that mass transfer will likely be stable @xcite and that j0056@xmath350611 will evolve into an am cvn system . the elm wd j0751@xmath350141 has a @xmath180 hr orbital period with aliases ranging between 1.85 and 2.05 hr ( see figure [ fig : pdm ] ) . the large @xmath181 km s@xmath26 semi - amplitude indicates that the companion is massive , regardless of the exact period . the minimum companion mass is 0.94 @xmath4 . assuming a random inclination distribution , there is a 47% probability that the companion is @xmath131.4 @xmath4 . given that the elm wd went through a common envelope phase of evolution with its companion , j0751@xmath350141 possibly contains a milli - second pulsar binary companion . helium - core wds are the most common type of companion in known milli - second pulsar binaries @xcite . we have been allocated cycle 14 _ chandra x - ray observatory _ time to search for x - ray emission from a possible neutron star . for @xmath178 , the companion is a 1.32 @xmath4 wd , and the system will begin mass transfer in 290 myr . the extreme mass ratio @xmath1790.18 means that this system will undergo stable mass transfer and evolve into an am cvn system . as the binary orbit widens in the am cvn phase , the mass accretion rate will drop and the mass required for the unstable burning of the accreted he - layer increases up to several percent of a solar mass . the final flash should ignite a thermonuclear transient visible as an underluminous supernova @xcite . it is also possible that the helium flash will detonate the massive wd in a double - detonation scenario ( * ? ? ? * though see dan et al . if j0751@xmath350141 has a massive wd companion , it is a probable supernova progenitor . the elm wd j0811 + 0225 has an orbital period of @xmath182 hr and a semi - amplitude of 221 km s@xmath26 . these orbital parameters yield a minimum companion mass of 1.20 @xmath4 . that means the companion probably exceeds a chandrasekhar mass . for @xmath178 , the most likely companion is a 1.70 @xmath4 neutron star at an orbital separation of 4.6 @xmath12 . there is no fermi gamma - ray detection at this location , but additional observations are needed to determine the nature of this system . the elm wd j0840 + 1527 has a best - fit @xmath23@xmath183 near the limit of our model grid . its best - fit orbital period is @xmath184 hr with a significant alias at 8.3 hr . assuming this object is 0.17 @xmath4 , its most likely companion is a wd with a comparable mass , 0.19 @xmath4 , at an orbital separation of 1.9 @xmath12 . if , on the other hand , j0840 + 1527 were a 3 @xmath4 main sequence star , its companion would have an orbital separation of 4.3 @xmath12 a separation comparable to the radius of a 3 @xmath4 star , and thus physically implausible . there is no evidence for mass transfer in this system . we conclude that j0840 + 1527 is a pair of elm wds . j1141 + 3850 , j1157 + 0546 , and j1238 + 1946 are the other systems containing wds near the low - gravity limit of our stellar atmosphere model grid , but their @xmath185 - 266 km s@xmath26 semi - amplitudes are significantly larger than that of j0840 + 1527 . if we assume that the objects are 0.17 @xmath4 elm wds , then the binary companions have minimum masses of 0.45 - 0.75 @xmath4 . if we instead assume that the objects are main sequence stars , then the required orbital separations are comparable to the radius of the main sequence star and physically impossible . there is no evidence for mass transfer in these systems . we conclude that j1141 + 3850 , j1157 + 0546 , and j1238 + 1946 are elm wds with likely wd companions . for j1141 + 3850 , there is a 36% probability that the companion is @xmath131.4 @xmath4 , possibly a milli - second pulsar . for j1141 + 3850 and j1238 + 1946 , mass transfer will begin in 7 - 10 gyr , making them am cvn progenitors and possible underluminous supernovae progenitors . the elm wd j2132 + 0754 has a well - constrained orbital period of @xmath186 hr and semi - amplitude of @xmath187 km s@xmath26 . the minimum companion mass is 0.95 @xmath4 , and there is a 48% probability that the companion is @xmath131.4 @xmath4 , possibly a milli - second pulsar . for @xmath178 , the most likely companion is a 1.33 @xmath4 wd that will begin mass transfer in 6 gyr . that makes j2132 + 0754 a likely am cvn progenitor and another possible underluminous supernovae progenitor . the elm wd j2338@xmath352052 has an orbital period of @xmath188 hr and a semi - amplitude of @xmath189 km s@xmath26 . in this case the companion is another elm wd ; for @xmath178 , the most likely companion is a 0.17 @xmath4 wd at an orbital separation of 0.56 @xmath12 . given the unity mass ratio , this system will undergo unstable mass transfer and will merge to form a single @xmath90.4 @xmath4 wd . this system will merge in less than 1 gyr . with these 17 new discoveries , plus the pulsating elm wd discovery published by @xcite , the elm survey has found 54 detached , double degenerate binaries ; 31 of the binaries will merge within a hubble time . table [ tab : summary ] summarizes the properties of the systems . eighty percent of the elm survey binaries are formally elm wd systems . we find low mass wds in compact binaries , binaries that must have gone through common envelope evolution . this makes sense because extremely low mass wds require significant mass loss to form ; the universe is not old enough to produce an extremely low mass wd through single star evolution . yet four objects published in the elm survey have no significant velocity variability , two of which are elm wds ( see table [ tab : summary ] ) . to understand whether or not these stars are single requires that we understand the significance of our binary detections . each elm survey binary is typically constrained by 10 - 30 irregularly spaced velocities with modest errors . given that we determine orbital parameters by minimizing @xmath168 , the @xmath190-test is a natural choice . we use the @xmath190-test to check whether the variance of the data around the orbital fit is consistent with the variance of the data around a constant velocity ( we use the weighted mean of the observations ) . @xmath190-test probabilities for the published elm survey binaries are @xmath191 . in other words , our binaries have significant velocity variability at the @xmath192% confidence level . in the null cases we need to calculate the likelihood of not detecting a binary . this is a trickier problem , and one that we approach with a monte carlo calculation . we start by selecting a set of observations ( times , velocity errors ) and an orbital period and semi - amplitude . we convert observation times to orbital phases using a randomly drawn zero time , and calculate velocities at those phases summed with a randomly drawn velocity error . we perform the @xmath190-test , using 0.01 as a detection threshold . we repeat this calculation 10,000 times for a given orbital period and semi - amplitude , and then select a new orbital period and semi - amplitude to iterate on . this analysis is done for each object . we find that the datasets for our 17 new binaries have a median 99.9% likelihood of detecting @xmath193 km s@xmath26 binaries , a 97% likelihood of detecting @xmath194 km s@xmath26 binaries , and a 44% likelihood of detecting @xmath195 km s@xmath26 binaries . it is no surprise that we are less likely to detect a low semi - amplitude binary , but this analysis suggests that we can be quite confident of detecting @xmath196 km s@xmath26systems . we find very similar likelihoods for detecting binaries containing elm wds in the full elm survey sample . the datasets for the null cases typically contain fewer observations and so are not as well - constrained . for j0900 + 0234 , a 0.16 @xmath4 elm wd with no observed velocity variation @xcite , the likelihoods of detecting a @xmath193 , 100 , and 50 km s@xmath26 binary are 87% , 57% , and 10% , respectively . additional observations are required to claim this elm wd as non - variable . there is , of course , an orbital period dependence to the detections , and periods near 24 hr are the most problematic . taken together , our datasets have a median 39% likelihood of detecting a @xmath194 km s@xmath26 binary at @xmath197 hr . yet we remain sensitive to longer periods : our datasets have median 99% and 98% likelihoods of detecting a @xmath194 km s@xmath26 binary at @xmath198 hr and @xmath199 hr , respectively . figure [ fig : kp ] plots the observed distribution of @xmath171 and @xmath172 for the elm survey binaries . the dashed lines indicate the approximate companion mass assuming @xmath200 is 0.2 @xmath4 and @xmath201 is 60@xmath202 . at @xmath194 km s@xmath26 , we can detect companion masses down to 0.1 @xmath4 at 2 hr orbital periods and 0.55 @xmath4companions at 2 day orbital periods . systems with @xmath203 km s@xmath26 are the realm of @xmath70.2 @xmath4 companions , and we observe a half - dozen systems with these parameters . our incompleteness at @xmath203 km s@xmath26 implies there are quite likely more elm wds with @xmath180.2 @xmath4 companions ; the remaining elm wd candidates that do not show obvious velocity change in a couple observations are possible examples of such low amplitude systems . finally , orbital inclination acts to increase the difficulty of identifying lone elm wds . consider the set of 46 elm wds in table [ tab : summary ] with @xmath5 , two of which are non - variable . their median @xmath172 is 240 km s@xmath26 , a semi - amplitude that would appear @xmath204 km s@xmath26 at @xmath205 . if the 46 objects have randomly distributed orbital inclinations , one object should have @xmath205 and a second @xmath206 . we conclude there is no good evidence for a lone elm wd in our present sample . this is in stark contrast to the population of 0.4 @xmath4 wds in the solar neighborhood , of which 20%-30% are single @xcite . one of the most exciting aspects of our elm wd binaries is that many have gravitational wave merger times less than a hubble time , and in one case as short as 1 myr @xcite . a natural question , then , is what will happen when the wds merge . notably , 5 of our 17 new binaries have probable system masses ( for @xmath178 ) in excess of a chandrasekhar mass . figure [ fig : merge ] plots the distribution of merger time and system mass for the elm survey binaries assuming @xmath178 , except when inclination is known . the majority of our binaries have probable masses below the chandrasekhar mass and are not potential supernova type ia progenitors , but six systems appear to have either neutron star or massive wd companions . the evolution of these six systems depends on the stability of mass transfer during the roche lobe overflow phase , and thus on the mass ratio of the binary . figure [ fig : tau ] plots the distribution of merger time and mass ratio @xmath174 for the elm survey binaries , again assuming @xmath178 except when inclination is known . the chandrasekhar mass systems generally have extreme @xmath207 and so will evolve into stable mass transfer am cvn systems . if the accretors are massive wds , these systems should experience a large helium flash that may appear as an underluminous .ia supernova ( e.g. * ? ? ? * ; * ? ? ? it is also possible that the helium flash will detonate the massive wd in a double - detonation scenario @xcite . the chandrasekhar mass systems are thus potential supernovae progenitors . however , the outcome of a merger of an elm wd with a massive wd is uncertain ( see * ? ? ? systems with @xmath208 , such as j2338@xmath352052 , will experience unstable mass transfer . these systems may merge @xcite to form a single low - mass wd , r crb star , or helium - burning sdb star . systems with intermediate @xmath209 , such as j0651 + 2844 , will experience either stable or unstable mass transfer depending on their spin - orbit coupling . after we constrain the remaining elm wd candidates plotted in fig.[fig : teff ] we look forward to calculating their space density and merger rate . elm wds binaries are clearly an important channel for am cvn star formation , and thus an important source of strong gravitational waves in the mhz regime . we perform stellar atmosphere fits to the entire collection of single - epoch spectra in the hvs survey and identify 57 new low mass wd candidates . follow - up spectroscopy reveals 17 wds with significant velocity variability , 12 of which are elm wds . presently , the elm survey sample is consistent with all of the elm wds being part of close binaries . elm wds are thus signposts for binaries that are strong gravitational wave sources and possible supernovae progenitors . four binaries in our sample contain @xmath1 objects , objects that might not be considered wds or proto - wds if not for their observed orbital motion . the existence of @xmath1 wds motivates us to develop a new set of stellar atmosphere models following @xcite . interestingly , all of the lowest surface gravity ( @xmath164 ) wds in our sample have metal lines in their spectra . this issue will be studied in detail in a future paper ; infrared observations are needed to constrain the possibility of debris disks as the source of ca and mg accretion in these compact binary systems . recent discoveries of pulsations in several of our elm wd targets @xcite enable us to probe the interiors of low - mass , presumably he - core wds using the tools of asteroseismology . due to the hvs color selection , all of the newly identified systems are hotter than 10,000 k and therefore too hot to show the g - mode pulsations detected in cooler elm wds @xcite . four binaries in our sample have massive @xmath30.9 @xmath4 companions and @xmath210 gyr merger times . if the unseen companions are massive wds , these extreme mass ratio binaries will undergo stable mass transfer and will evolve into am cvn systems and potentially future .ia or underluminous supernovae . thus the systems j0751@xmath350141 , j1141 + 3850 , j1238 + 1946 , and j2132 + 0754 are possible supernova progenitors . the other possibility is that the massive companions in these four binaries are neutron stars . given their past common envelope evolution , the systems quite possibly contain milli - second pulsars . a fifth system , j0811 + 0225 , appears to have a minimum companion mass close to a chandrasekhar mass . we are conducting follow - up optical , radio , and x - ray observations to establish the nature of these extreme mass ratio elm wd binaries . optically visible wd companions of neutron stars are useful for constraining the binary mass ratio and also to calibrate the spin - down ages of milli - second pulsars . given our present observations , we expect that the elm survey will grow to 100 binaries in a few years . a large sample is important for finding more systems like j0651 + 2844 , a 12.75 min period system that provides us with a laboratory for measuring the spin - orbit coupling and tidal heating of a rapidly merging pair of wds . we expect that photometric observations will discover more pulsating elm wds . finally , a larger sample of elm wds will be an important source of gravitational wave verification sources for @xmath210 and other future gravitational wave detection experiments . we thank m. alegria , p. canton , s. gotilla , j. mcafee , e. martin , a.milone , and r. ortiz for their assistance with observations obtained at the mmt observatory , and p. berlind and m. calkins for their assistance with observations obtained at the fred lawrence whipple observatory . we also thank p. bergeron and d. koester for their invaluable assistance in the computation of our stellar atmosphere model grids . this research makes use of the sao / nasa astrophysics data system bibliographic service . this project makes use of data products from the sloan digital sky survey , which has been funded by the alfred p. sloan foundation , the participating institutions , the national science foundation , and the u.s.department of energy office of science . this work was supported in part by the smithsonian institution . _ facilities : _ crcccccccc j0022@xmath351014 & 18980 & 7.15 & 0.07989 & 145.6 & 0.33 & @xmath211 & 0.23 & @xmath212 & 6 + j0056@xmath350611 & 12210 & 6.17 & 0.04338 & 376.9 & 0.17 & @xmath213 & 0.61 & @xmath214 & + j0106@xmath351000 & 16490 & 6.01 & 0.02715 & 395.2 & 0.17 & 0.43 & & 0.037 & 7 + j0112 + 1835 & 9690 & 5.63 & 0.14698 & 295.3 & 0.16 & @xmath215 & 0.85 & @xmath216 & 1 + j0651 + 2844 & 16530 & 6.76 & 0.00886 & 616.9 & 0.26 & 0.50 & & 0.0011 & 3,15 + j0751@xmath350141 & 15660 & 5.43 & 0.08001 & 432.6 & 0.17 & @xmath217 & 1.32 & @xmath218 & + j0755 + 4906 & 13160 & 5.84 & 0.06302 & 438.0 & 0.17 & @xmath219 & 1.12 & @xmath220 & 2 + j0818 + 3536 & 10620 & 5.69 & 0.18315 & 170.0 & 0.17 & @xmath221 & 0.33 & @xmath222 & 2 + j0822 + 2753 & 8880 & 6.44 & 0.24400 & 271.1 & 0.17 & @xmath223 & 1.05 & @xmath224 & 4 + j0825 + 1152 & 24830 & 6.61 & 0.05819 & 319.4 & 0.26 & @xmath225 & 0.61 & @xmath226 & 0 + j0849 + 0445 & 10290 & 6.23 & 0.07870 & 366.9 & 0.17 & @xmath227 & 0.88 & @xmath228 & 4 + j0923 + 3028 & 18350 & 6.63 & 0.04495 & 296.0 & 0.23 & @xmath229 & 0.44 & @xmath230 & 2 + j1005 + 0542 & 15740 & 7.25 & 0.30560 & 208.9 & 0.34 & @xmath231 & 0.86 & @xmath232 & 0 + j1005 + 3550 & 10010 & 5.82 & 0.17652 & 143.0 & 0.17 & @xmath211 & 0.24 & @xmath233 & 0 + j1053 + 5200 & 15180 & 6.55 & 0.04256 & 264.0 & 0.20 & @xmath221 & 0.33 & @xmath234 & 4,9 + j1056 + 6536 & 20470 & 7.13 & 0.04351 & 267.5 & 0.34 & @xmath229 & 0.43 & @xmath235 & 0 + j1112 + 1117 & 9590 & 6.36 & 0.17248 & 116.2 & 0.17 & @xmath236 & 0.17 & @xmath237 & 16 + j1141@xmath423850 & 11620 & 5.31 & 0.25958 & 265.8 & 0.17 & @xmath223 & 1.05 & @xmath238 & + j1233 + 1602 & 10920 & 5.12 & 0.15090 & 336.0 & 0.17 & @xmath239 & 1.20 & @xmath240 & 2 + j1234@xmath350228 & 18000 & 6.64 & 0.09143 & 94.0 & 0.23 & @xmath241 & 0.11 & @xmath216 & 6 + j1238@xmath421946 & 16170 & 5.28 & 0.22275 & 258.6 & 0.17 & @xmath227 & 0.88 & @xmath242 & + j1436 + 5010 & 16550 & 6.69 & 0.04580 & 347.4 & 0.24 & @xmath213 & 0.60 & @xmath243 & 4,9 + j1443 + 1509 & 8810 & 6.32 & 0.19053 & 306.7 & 0.17 & @xmath244 & 1.15 & @xmath245 & 1 + j1630 + 4233 & 14670 & 7.05 & 0.02766 & 295.9 & 0.30 & @xmath246 & 0.37 & @xmath247 & 8 + j1741 + 6526 & 9790 & 5.19 & 0.06111 & 508.0 & 0.16 & @xmath248 & 1.55 & @xmath249 & 1 + j1840 + 6423 & 9140 & 6.16 & 0.19130 & 272.0 & 0.17 & @xmath227 & 0.88 & @xmath250 & 1 + j2103@xmath350027 & 10000 & 5.49 & 0.20308 & 281.0 & 0.17 & @xmath251 & 0.99 & @xmath252 & 0 + j2119@xmath350018 & 10360 & 5.36 & 0.08677 & 383.0 & 0.17 & @xmath253 & 1.04 & @xmath254 & 2 + j2132@xmath420754 & 13700 & 6.00 & 0.25056 & 297.3 & 0.17 & @xmath255 & 1.33 & @xmath256 & + j2338@xmath352052 & 16630 & 6.87 & 0.07644 & 133.4 & 0.27 & @xmath257 & 0.18 & @xmath258 & + nltt11748 & 8690 & 6.54 & 0.23503 & 273.4 & 0.18 & 0.76 & & 7.2 & 5,10,11 + j0022 + 0031 & 17890 & 7.38 & 0.49135 & 80.8 & 0.38 & @xmath259 & 0.26 & & 6 + j0152 + 0749 & 10840 & 5.80 & 0.32288 & 217.0 & 0.17 & @xmath260 & 0.78 & & 1 + j0730 + 1703 & 11080 & 6.36 & 0.69770 & 122.8 & 0.17 & @xmath261 & 0.41 & & 0 + j0755@xmath424800 & 19890 & 7.46 & 0.54627 & 194.5 & 0.42 & @xmath262 & 1.18 & & + j0802@xmath350955 & 16910 & 6.42 & 0.54687 & 176.5 & 0.20 & @xmath260 & 0.76 & & + j0811@xmath420225 & 13990 & 5.79 & 0.82194 & 220.7 & 0.17 & @xmath263 & 1.71 & & + j0815@xmath422309 & 21470 & 5.78 & 1.07357 & 131.7 & 0.17 & @xmath225 & 0.63 & & + j0840@xmath421527 & 13810 & 5.04 & 0.52155 & 84.8 & 0.17 & @xmath257 & 0.19 & & + j0845 + 1624 & 17750 & 7.42 & 0.75599 & 62.2 & 0.40 & @xmath211 & 0.22 & & 0 + j0900 + 0234 & 8220 & 5.78 & & @xmath264 & 0.16 & & & & 1 + j0917 + 4638 & 11850 & 5.55 & 0.31642 & 148.8 & 0.17 & @xmath265 & 0.36 & & 12 + j1046@xmath350153 & 14880 & 7.37 & 0.39539 & 80.8 & 0.37 & @xmath211 & 0.23 & & + j1104@xmath420918 & 16710 & 7.61 & 0.55319 & 142.1 & 0.46 & @xmath266 & 0.70 & & + j1151@xmath425858 & 15400 & 6.09 & 0.66902 & 175.7 & 0.17 & @xmath267 & 0.84 & & + j1157@xmath420546 & 12100 & 5.05 & 0.56500 & 158.3 & 0.17 & @xmath268 & 0.59 & & + j1422 + 4352 & 12690 & 5.91 & 0.37930 & 176.0 & 0.17 & @xmath269 & 0.55 & & 2 + j1439 + 1002 & 14340 & 6.20 & 0.43741 & 174.0 & 0.18 & @xmath213 & 0.62 & & 2 + j1448 + 1342 & 12580 & 6.91 & & @xmath270 & 0.25 & & & & 2 + j1512 + 2615 & 12130 & 6.62 & 0.59999 & 115.0 & 0.20 & @xmath265 & 0.36 & & 2 + j1518 + 0658 & 9810 & 6.66 & 0.60935 & 172.0 & 0.20 & @xmath271 & 0.78 & & 1 + j1538@xmath420252 & 11560 & 5.97 & 0.41915 & 227.6 & 0.17 & @xmath272 & 1.06 & & + j1557@xmath422823 & 12550 & 7.76 & 0.40741 & 131.2 & 0.49 & @xmath273 & 0.54 & & + j1625 + 3632 & 23570 & 6.12 & 0.23238 & 58.4 & 0.20 & @xmath274 & 0.08 & & 6 + j1630 + 2712 & 11200 & 5.95 & 0.27646 & 218.0 & 0.17 & @xmath275 & 0.70 & & 2 + j2252@xmath350056 & 19450 & 7.00 & & @xmath276 & 0.31 & & & & 2 + j2345@xmath350102 & 33130 & 7.20 & & @xmath277 & 0.42 & & & & 2 + lp400@xmath3522 & 11170 & 6.35 & 1.01016 & 119.9 & 0.19 & @xmath269 & 0.52 & & 13,14 73 natexlab#1#1 , h. , allende prieto , c. , an , d. , et al . 2011 , , 193 , 29 , c. , beers , t. c. , wilhelm , r. , et al . 2006 , , 636 , 804 , c. , sivarani , t. , beers , t. c. , et al . 2008 , , 136 , 2070 , c. g. , van kerkwijk , m. h. , koester , d. , & verbunt , f. 2006 , , 456 , 295 , p. , wesemael , f. , & beauchamp , a. 1995 , , 107 , 1047 , l. , shen , k. j. , weinberg , n. n. , & nelemans , g. 2007 , , 662 , l95 , j. m. , kilic , m. , brown , w. r. , & kenyon , s. j. 2011 , , 729 , 2 , w. r. , geller , m. j. , & kenyon , s. j. 2009 , , 690 , 1639 . 2012 , , 751 , 55 , w. r. , geller , m. j. , kenyon , s. j. , & kurtz , m. j. 2005 , , 622 , l33 . 2006 , , 640 , l35 . 2006 , , 647 , 303 , w. r. , geller , m. j. , kenyon , s. j. , kurtz , m. j. , & bromley , b. c. 2007 , , 660 , 311 . 2007 , , 671 , 1708 , w. r. , kilic , m. , allende prieto , c. , & kenyon , s. j. 2010 , , 723 , 1072 . 2011 , , 411 , l31 . 2012 , , 744 , 142 , w. r. , kilic , m. , hermes , j. j. , et al . 2011 , , 737 , l23 , g. , ferraro , f. r. , possenti , a. , & damico , n. 2006 , , 641 , l129 , a. h. , romero , a. d. , althaus , l. g. , & hermes , j. j. 2012 , , 547 , a96 , m. , rosswog , s. , guillochon , j. , & ramirez - ruiz , e. 2011 , , 737 , 89 . 2012 , , 422 , 2417 , b. , rood , r. t. , & oconnell , r. w. 1993 , , 419 , 596 , d. , cheimets , p. , caldwell , n. , & geary , j. 1998 , , 110 , 79 , s. , schaffenroth , v. , hirsch , h. , et al . 2012 , in asp conf . ser , , vol . 452 , fifth meeting on hot subdwarf stars and related objects , ed . d. kilkenny , c. s. jeffery , & c. koen , 129 , a. , bergeron , p. , & ruiz , m. t. 2011 , , 743 , 138 , l. , bertelli , g. , bressan , a. , et al . 2002 , , 391 , 195 , l. , grebel , e. k. , odenkirchen , m. , & chiosi , c. 2004 , , 422 , 205 , u. 2009 , , 47 , 211 , u. , edelmann , h. , lisker , t. , & napiwotzki , r. 2003 , , 411 , l477 , j. j. , kilic , m. , brown , w. r. , montgomery , m. h. , & winget , d. e. 2012 , , 749 , 42 , j. j. , kilic , m. , brown , w. r. , et al . 2012 , , 757 , l21 , j. j. , montgomery , m. h. , winget , d. e. , brown , w. r. , kilic , m. , & kenyon , s. j. 2012 , , 750 , l28 , j. j. , montgomery , m. h. , winget , d. e. , et al . 2013 , , accepted , d. l. , bhalerao , v. b. , van kerkwijk , m. h. , et al . 2013 , , accepted , a. , vennes , s. , & vaccaro , t. r. 2010 , , 516 , l7 , s. j. & garcia , m. r. 1986 , , 91 , 125 , m. , allende prieto , c. , brown , w. r. , & koester , d. 2007 , , 660 , 1451 , m. , allende prieto , c. , brown , w. r. , et al . 2010 , , 721 , l158 , m. , brown , w. r. , allende prieto , c. , kenyon , s. j. , & panei , j. a. 2010 , , 716 , 122 , m. , brown , w. r. , allende prieto , c. , pinsonneault , m. , & kenyon , s. 2007 , , 664 , 1088 , m. , brown , w. r. , allende prieto , c. , swift , b. , kenyon , s. j. , liebert , j. , & ageros , m. a. 2009 , , 695 , l92 , m. , brown , w. r. , allende prieto , c. , et al . 2011 , , 727 , 3 . 2012 , , 751 , 141 , m. , brown , w. r. , hermes , j. j. , et al . 2011 , , 418 , l157 , m. , brown , w. r. , kenyon , s. j. , et al . 2011 , , 413 , l101 , d. 2008 , arxiv:0812.0482 , m. j. & mink , d. j. 1998 , , 110 , 934 , r. l. 1993 , synthe spectrum synthesis programs and line data ( kurucz cd - rom ; cambridge , ma : smithsonian astrophysical observatory ) , j. , bergeron , p. , & holberg , j. b. 2005 , , 156 , 47 , t. r. , dhillon , v. s. , & duck , s. r. 1995 , , 275 , 828 , t. r. , nelemans , g. , & steeghs , d. 2004 , , 350 , 113 , p. , strobel , k. , barnes , j. v. , & anderson , e. 1988 , , 328 , 315 , g. , richer , j. , & richard , o. 2008 , , 675 , 1223 , f. , badenes , c. , thompson , s. e. , & lupton , r. 2009 , , 707 , l51 , k. 1982 , , 253 , 798 , b. 1971 , acta astron . , 21 , 1 , j. a. , althaus , l. g. , chen , x. , & han , z. 2007 , , 382 , 779 , c. , pelletier , c. , fontaine , g. , & michaud , g. 1986 , , 61 , 197 , s. , gnsicke , b. t. , brady , s. , et al . 2012 , , 419 , 817 , a. , nebot gmez - morn , a. , schreiber , m. r. , et al . 2012 , , 419 , 806 , g. d. , weymann , r. j. , & foltz , c. b. 1989 , , 101 , 713 , k. j. & bildsten , l. 2009 , , 699 , 1365 , r. , stensen , r. h. , bloemen , s. , et al . 2012 , , 424 , 1752 , s. a. , fink , m. , kromer , m. , et al . 2012 , , 420 , 3003 , j. d. r. , kaplan , d. l. , shporer , a. , bildsten , l. , & howell , s. b. 2010 , , 716 , l146 , t. m. , langer , n. , & kramer , m. 2012 , , 425 , 1601 , p .- e . & bergeron , p. 2009 , , 696 , 1755 , p .- e . , bergeron , p. , kalirai , j. s. , & gianninas , a. 2010 , , 712 , 1345 , v. , fontaine , g. , brassard , p. , & dupret , m .- a . 2013 , , 762 , 57 , s. , kawka , a. , vaccaro , t. r. , & silvestri , n. m. 2009 , , 507 , 1613 , s. , thorstensen , j. r. , kawka , a. , et al . 2011 , , 737 , l16 , s. e. & weaver , t. a. 1994 , , 423 , 371 table [ tab : dat ] presents our radial velocity measurements . the table columns include object name , heliocentric julian date ( based on utc ) , heliocentric radial velocity ( uncorrected for the wd gravitational redshift ) , and velocity error .

there is a vast literature on epidemics on complex networks ( see @xcite for a review ) . most of the epidemic models consider a transmission mechanism which is independent of the local condition faced by the agents concerned . however , if there is a factor of coordination or persuasion involved , relative considerations tend to be important in understanding whether some new belief or behavior is adopted @xcite . to fix ideas , it may be useful to think of the diffusion process as modeling the adoption of a new technology . in this case , when an agent is confronted with the possibility of adoption , her decision depends on the persuasion effort exerted on her by each of her neighbors in the social network . more formally , those neighborhood effects can be captured as follows : the probability that an agent adopts the new technology when @xmath0 out of her @xmath1 neighbors have already adopted can be modeled by a threshold function : this probability is one if @xmath2 and it is zero otherwise . for @xmath3 , this would correspond to a local majority rule . more generally , some simple models for the diffusion of a new behavior have been proposed in term of a basic underlying model of individual decision - making : as individuals make decisions based on the choices of their neighbors , a particular pattern of behavior can begin to spread across the links of the network @xcite , @xcite . in this work , we analyze the diffusion in the large population limit when the underlying graph is a random network with given vertex degrees . there is now a large literature on the study of complex networks across different disciplines such as physics @xcite , mathematics @xcite , sociology @xcite , networking @xcite or economics @xcite . there is also a large literature on local interaction and adoption externalities @xcite , @xcite , @xcite , @xcite . similarly to these works , we study a game where players payoffs depend on the actions taken by their neighbors in the network but not on the specific identities of these neighbors . our most general framework ( presented in section [ sec : mod ] ) allows to deal with threshold games of complements , i.e. a player has an increasing incentive to take an action as more neighbors take the action . the main contribution of our paper is a model which allows us to study rigorously semi - anonymous threshold games of complements with local interactions on a complex network . we use a very simple dynamics for the evolution of the play : each time , agents choose the strategy with best payoff , given the current behavior of their neighbors . stochastic versions of best response dynamics have been studied extensively as a simple model for the emergence of conventions @xcite , @xcite . a key finding in this line of work is that local interaction may drive the system towards a particular equilibrium in which all players take the same action . most of the literature following @xcite , @xcite is concerned with stochastic versions of best response dynamics on fixed networks and a simple condition known as risk dominance @xcite determines whether an innovation introduced in the network will eventually become widespread . in this case , bounds on the convergence time have been computed @xcite , @xcite as a function of the structure of the interaction network . in this paper , we focus on properties of deterministic ( myopic ) best response dynamics . in such a framework , as shown in @xcite , coordination on the risk dominant strategy depends on the network structure , the dynamics of the game and the initial configuration . in particular , this coordination might fail and equilibria with co - existent strategies can exist . this was shown in @xcite already in the simple case where the underlying network is the two - dimensional lattice . while the results in @xcite depend very much on the details of the particular graph considered , @xcite derives general results that depend on local properties of the graphs . more formally , agents are nodes of a fixed infinite graph and revise their strategies synchronously . there is an edge @xmath4 if agents @xmath5 and @xmath6 can interact with each other . each node has a choice between two possible behaviors labeled @xmath7 and @xmath8 . on each edge @xmath4 , there is an incentive for @xmath5 and @xmath6 to have their behaviors match , which is modeled as the following coordination game parametrized by a real number @xmath9 : if @xmath5 and @xmath6 choose @xmath7 ( resp . @xmath8 ) , they each receive a payoff of @xmath10 ( resp . @xmath11 ) ; if they choose opposite strategies , then they receive a payoff of @xmath12 . then the total payoff of a player is the sum of the payoffs with each of her neighbors . if the degree of node @xmath5 is @xmath13 and @xmath14 is her number of neighbors playing @xmath8 , then the payoff to @xmath5 from choosing @xmath7 is @xmath15 while the payoff from choosing @xmath8 is @xmath16 . hence the best response strategy for agent @xmath5 is to adopt @xmath8 if @xmath17 and @xmath7 if @xmath18 . a number of qualitative insights can be derived from this simple diffusion model @xcite . clearly , a network where all nodes play @xmath7 is an equilibrium of the game as is the state where all nodes play @xmath8 . consider now a network where all nodes initially play @xmath7 and a small number of nodes are forced to adopt strategy @xmath8 ( constituting the seed ) . other nodes in the network apply best - response updates , then these nodes will be repeatedly applying the following rule : switch to @xmath8 if enough of your neighbors have already adopted @xmath8 . there can be a cascading sequence of nodes switching to @xmath8 such that a network - wide equilibrium is reached in the limit . in particular , contagion is said to occur if action @xmath8 can spread from a finite set of players to the whole population . it is shown in @xcite that there is a contagion threshold such that contagion occurs if and only if the parameter @xmath10 is less than the contagion threshold . in particular the contagion threshold is always at most @xmath19 so that @xmath8 needs to be the risk - dominant action in order to spread . however this condition is not sufficient and @xcite gives a number of characterizations of the contagion threshold and also shows that low contagion threshold implies the existence of equilibria where both actions are played . most of the results on this model are restricted to deterministic infinite graphs and the results in @xcite depend on local properties of graphs that need to be checked node - by - node . in networks exhibiting heterogeneity among the nodes , this approach will be limited in its application . in this work , we analyze the diffusion in the large population limit when the underlying graph is a random network @xmath20 with @xmath21 vertices and where @xmath22 is a given degree sequence ( see section [ sec : conf ] for a detailed definition ) . although random graphs are not considered to be highly realistic models of most real - world networks , they are often used as first approximation and are a natural first choice for a sparse interaction network in the absence of any known geometry of the problem @xcite . our analysis yields several insights into how the diffusion propagates and we believe that they could be used in real - world situations . as an example , we show some implications of our results for the design of optimal advertising strategies . our results extend the previous analysis of global cascades made in @xcite using a threshold model . they differ greatly from the study of standard epidemics models used for the analysis of the spread of viruses @xcite where an entity begin as susceptible and may become infected and infectious through contacts with her neighbors with a given probability . already in the simple coordination game presented above , we show that connectivity ( i.e. the average number of neighbors of a typical agent ) plays an ambiguous role : while it allows the diffusion to spread , when the network is highly connected , the diffusion is also limited by high - degree nodes , which are very stable . these nodes require a lot of their neighbors to switch to @xmath8 in order to play @xmath8 themselves . in the case of a sparse random network of interacting agents , we compute the contagion threshold in section [ sec : cont ] and relate it to the important notion of _ pivotal players_. we also show the existence of ( continuous and discontinuous ) phase transitions . in section [ sec : eq ] , we analyze the possible equilibria of the game for low values of @xmath10 . in particular , we give conditions under which an equilibrium with coexistence of large ( i.e. containing a positive fraction of the total population ) connected sets of players @xmath7 and @xmath8 is possible . in section [ sec : ad ] , we also compute the minimal size of a seed of new adopters in order to trigger a global cascade in the case where these adopters can only be sampled without any information on the graph . we show that this minimal size has a non - trivial behavior as a function of the connectivity . in section [ sec : lmf ] , we give a heuristic argument allowing to recover the technical results which gives some intuition behind our formulas . our results also explain why social networks can display a great stability in the presence of continual small shocks that are as large as the shocks that ultimately generate a global cascade . cascades can therefore be regarded as a specific manifestation of the robust yet fragile nature of many complex systems @xcite : a system may appear stable for long periods of time and withstand many external shocks ( robustness ) , then suddenly and apparently inexplicably exhibit a large scale cascade ( fragility ) . our paper is divided into two parts . the first part is contained in the next section . we apply our technical results to the particular case of the coordination game presented above . we describe our main findings and provide heuristics and intuitions for them . these results are direct consequences of our main theorems stated and proved in the second part of the paper . this second part starts with section [ sec : mod ] , where we present in details our most general model of diffusion ( where initial activation and thresholds can vary among the population and also be random ) . this general model allows to recover and extend results in the random graphs literature and is used in @xcite for the analysis of the impact of clustering . we also state our main technical results : theorem [ th : epi ] and theorem [ prop : cascade ] . their proofs can be found in sections [ sec : thepi ] and [ sec : casc ] respectively . section [ sec : conc ] contains our concluding remarks . * probability asymptotics : * in this paper , we consider sequences of ( random ) graphs and asymptotics as the number of vertices @xmath21 tends to infinity . for notational simplicity we will usually not show the dependency on @xmath21 explicitly . all unspecified limits and other asymptotics statement are for @xmath23 . for example , w.h.p . ( with high probability ) means with probability tending to 1 as @xmath24 and @xmath25 means convergence in probability as @xmath26 . similarly , we use @xmath27 , @xmath28 and @xmath29 in a standard way , see @xcite . for example , if @xmath30 is a parameter of the random graph , @xmath31 means that @xmath32 as @xmath23 for every @xmath33 , equivalently @xmath34 , or for every @xmath33 , @xmath35 w.h.p . in this section , we first present the model for the underlying graph . this is the only part of this section required to understand our main technical results in section [ sec : mod ] . the remaining parts of this section present and discuss applications and special cases exploring the significance of the technical results derived in the sequel . we consider a set @xmath36=\{1,\dots , n\}$ ] of agents interacting over a social network . let @xmath37 be a sequence of non - negative integers such that @xmath38 is even . for notational simplicity we will usually not show the dependency on @xmath21 explicitly . this sequence is the degree sequence of the graph : agent @xmath39 $ ] has degree @xmath13 , i.e. has @xmath13 neighbors . we define a random multigraph ( allowing for self - loop and multiple links ) with given degree sequence @xmath40 , denoted by @xmath41 by the configuration model @xcite : take a set of @xmath13 half - edges for each vertex @xmath5 and combine the half - edges into pairs by a uniformly random matching of the set of all half - edges . conditioned on the multigraph @xmath41 being a simple graph , we obtain a uniformly distributed random graph with the given degree sequence , which we denote by @xmath42 @xcite . we will let @xmath23 and assume that we are given @xmath43 satisfying the following regularity conditions , see @xcite : [ cond ] for each @xmath21 , @xmath44 is a sequence of non - negative integers such that @xmath38 is even and , for some probability distribution @xmath45 independent of @xmath21 , * @xmath46 for every @xmath47 as @xmath48 ; * @xmath49 ; * @xmath50 } d_i^2=o(n)$ ] . in words , we assume that the empirical distribution of the degree sequence converges to a fixed probability distribution @xmath51 with a finite mean @xmath52 . condition [ cond ] ( iii ) ensures that @xmath51 has a finite second moment and is technically required . it implies that the empirical mean of the degrees converges to @xmath52 . in some cases , we will need the following additional condition which implies that the asymptotic distribution of the degrees @xmath51 has a finite third moment : [ cond2 ] for each @xmath21 , @xmath44 is a sequence of non - negative integers such that @xmath50 } d_i^3=o(n)$ ] . the results of this work can be applied to some other random graphs models too by conditioning on the vertex degrees ( see @xcite ) . for example , for the erds - rnyi graph @xmath53 with @xmath54 , our results will hold with the limiting probability distribution @xmath51 being a poisson distribution with mean @xmath52 . an interesting perspective is to understand how different network structures are more or less hospitable to cascades . going back to the diffusion model described in the introduction , we see that the lower @xmath10 is , the easiest the diffusion spreads . in @xcite , the contagion threshold of a connected infinite network ( called the cascade capacity in @xcite ) is defined as the maximum threshold @xmath55 at which a finite set of initial adopters can cause a complete cascade , i.e. the resulting cascade of adoptions of @xmath8 eventually causes every node to switch from @xmath7 to @xmath8 . there are two possible models to consider depending whether the initial adopters changing from @xmath7 to @xmath8 apply or not best - response update . it is shown in @xcite that the same contagion threshold arises in both models . in this section , we consider a model where the initial adopters are forced to play @xmath8 forever . in this case , the diffusion is monotone and the number of nodes playing @xmath8 is non - decreasing . we say that this case corresponds to the _ permanent adoption model _ : a player playing @xmath8 will never play @xmath7 again . we will discuss a model where the initial adopters also apply best - response update in section [ sec : eq ] . to illustrate the notion of contagion threshold , consider the infinite network depicted on figure [ fig : counter34 ] . it is easy to see that the contagion threshold for such a network is @xmath56 , the same as for the @xmath57-dimensional lattice . note that in the network of figure [ fig : counter34 ] the density of nodes with degree @xmath58 is @xmath59 and the density of nodes of degree @xmath60 is @xmath61 , whereas in the @xmath57-dimensional lattice , all vertices have degree @xmath60 . we will see in the sequel that the sequence of random graphs @xmath20 with asymptotic degree distribution such that @xmath62 and @xmath63 has a contagion threshold of @xmath64.to be done and with density of nodes with degree @xmath58 equals to @xmath59 and degree @xmath60 equals to @xmath61.,width=188 ] we now compute the contagion threshold for a sequence of random networks . since a random network is finite and not necessarily connected , we first need to adapt the definition of contagion threshold to our context . we start by recalling a basic result on the size of the largest component of a random graph with a given degree sequence @xcite , namely a sufficient condition for the existence of a _ giant component_. consider the random graph @xmath20 satisfying conditions [ cond ] with asymptotic degree distribution @xmath45 . if @xmath65 then the size of the largest component of @xmath20 is @xmath66 . in this case , we say that there exists a _ giant component_. for a graph @xmath67 and a parameter @xmath10 , we consider the largest connected component of the induced subgraph in which we keep only vertices of degree strictly less than @xmath68 . we call the vertices in this component _ pivotal players _ : if only one pivotal player switches from @xmath7 to @xmath8 then the whole set of pivotal players will eventually switch to @xmath8 in the permanent adoption model . for a player @xmath69 , we denote by @xmath70 the final number of players @xmath8 in the permanent adoption model with parameter @xmath10 , when the initial state consists of only @xmath71 playing @xmath8 , all other players playing @xmath7 . informally , we say that @xmath70 is the size of the cascade induced by player @xmath71 . [ prop : cap ] consider the random graph @xmath20 satisfying conditions [ cond ] and [ cond2 ] with asymptotic degree distribution @xmath45 such that @xmath72 . we define @xmath55 by : @xmath73 let @xmath74 be the set of pivotal players in @xmath20 . * for @xmath75 , there are constants @xmath76 such that w.h.p . @xmath77 and for any @xmath78 , @xmath79 . * for @xmath80 , for an uniformly chosen player @xmath71 , we have @xmath81 . the same result holds if @xmath82 players are chosen uniformly at random . note that proposition [ prop : cap ] is valid only for graphs with a giant component ( for graphs with all components @xmath83 , we have trivially that @xmath81 for all values of @xmath10 ) . we can rewrite ( [ eq : cap ] ) as follows : let @xmath84 be a random variable with distribution @xmath51 , i.e. @xmath85 , then @xmath86>{{\mathbb{e}}}[d]\right\}.\end{aligned}\ ] ] we can restate proposition [ prop : cap ] as follows : let @xmath87 be the size of a cascade induced by switching a random player from @xmath7 to @xmath8 . proposition [ prop : cap ] implies that for @xmath80 , we have @xmath88 as @xmath23 for every @xmath33 , whereas for @xmath75 there exists constants @xmath89 depending only on @xmath10 and the parameters of the graph such that for every @xmath33 , @xmath90 . informally , we will say that global cascades ( i.e. reaching a positive fraction of the population ) occur when @xmath75 and do not occur when @xmath80 . we call @xmath55 defined by ( [ eq : cap ] ) the contagion threshold for the sequence of random networks with degree distribution @xmath91 . this result is in accordance with the heuristic result of @xcite ( see in particular the cascade condition eq . 5 in @xcite ) . proposition [ prop : cap ] follows from theorem [ prop : cascade ] below which also gives estimates for the parameters @xmath92 and @xmath93 when @xmath75 . under additional technical conditions , theorem [ prop : cascade ] shows that @xmath94 in case ( i ) . note that the condition @xmath95 implies that @xmath96 so that for @xmath97 , we have , @xmath98 . hence we have the following elementary corollary : [ cor1/3 ] for any random graph @xmath20 satisfying conditions [ cond ] and [ cond2 ] with asymptotic degree distribution @xmath45 such that @xmath72 , we have @xmath99 . recall that for general graphs , it is shown in @xcite that @xmath100 . note also that our bound in corollary [ cor1/3 ] is tight : for a @xmath0-regular network with @xmath101 chosen uniformly at random ( i.e. @xmath20 such that @xmath102 for all @xmath39 $ ] ) , proposition [ prop : cap ] implies that @xmath103 corresponding to the contagion threshold of a @xmath0-regular tree , see @xcite . for @xmath101 , a @xmath0-regular random graph is connected w.h.p , so that for any @xmath104 , an initial adopter will cause a complete cascade ( i.e. reaching all players ) . in particular , in this case , the only possible equilibria of the game are all players playing @xmath7 or all players playing @xmath8 . we will come back to the analysis of equilibria of the game in section [ sec : eq ] . given a degree distribution @xmath51 , proposition [ prop : cap ] gives the typical value of @xmath105 under the uniform distribution on graphs with the prescribed degree distribution . a question of interest is to find among all graphs with a given degree sequence , which ones achieve the lowest or highest value of the contagion threshold . indeed , the contagion threshold of a graph can be seen as a proxy for its susceptibility to cascades : the higher the contagion threshold is , the more susceptible the graph is to cascading behavior . clearly if @xmath106 is the maximal degree in the graph , we have @xmath107 . in the case of regular graphs , we saw that the typical contagion threshold is equal to this lower bound : @xmath108 . note however that there are examples of regular graphs in @xcite with a strictly higher threshold ( see figure 3 there : a @xmath109-regular graph with contagion threshold @xmath110 ) . hence among all @xmath0-regular graphs , picking a graph at random will give w.h.p . a graph which is the least susceptible to cascades among those @xmath0-regular graphs . however it is not always the case that picking a graph at random among all graphs with a given degree sequence will result in the least susceptible one . there are situations where the typical value of the contagion threshold exceeds the lowest possible value : this is the case in the example presented above in figure [ fig : counter34 ] for which the contagion threshold is @xmath111 whereas the sequence of random graphs @xmath20 with asymptotic degree distribution such that @xmath62 and @xmath63 has a contagion threshold of @xmath64 . the recent work @xcite has a similar focus . for erds - rnyi random graphs and for power law graphs ( dashed curve ) as a function of the average degree @xmath52 . right : @xmath112 for erds - rnyi random graphs as a function of @xmath52 in the range @xmath113$].,title="fig:",width=188 ] for erds - rnyi random graphs and for power law graphs ( dashed curve ) as a function of the average degree @xmath52 . right : @xmath112 for erds - rnyi random graphs as a function of @xmath52 in the range @xmath113$].,title="fig:",width=151 ] we now consider some examples . first the case of erds - rnyi random graphs @xmath53 where each of the @xmath114 edges is present with probability @xmath115 for a fixed parameter @xmath52 . in this case , we can apply our results with @xmath116 for all @xmath47 , i.e. the poisson distribution with mean @xmath52 . as shown in figure [ fig : qc ] ( left ) for the case of erds - rnyi random graphs @xmath53 , @xmath55 is well below @xmath117 , indeed we have @xmath118 for any value of @xmath52 . as shown in figure [ fig : qc ] ( right ) , we see that @xmath55 is a non - decreasing function of @xmath52 the average degree in the graph for @xmath119 . clearly on figure [ fig : qc ] ( left ) , @xmath55 is a non - increasing function of @xmath52 , for @xmath120 . the second curve in figure [ fig : qc ] corresponds to the contagion threshold for a scale - free random network whose degree distribution @xmath121 ( with @xmath122 ) is parametrized by the decay parameter @xmath123 . we see that in this case we have @xmath124 . in other words , in an erds - rnyi random graph , in order to have a global cascade , the parameter @xmath10 must be such that any node with no more than four neighbors must be able to adopt @xmath8 even if it has a single adopting neighbor . in the case of the scale free random network considered , the parameter @xmath10 must be much lower and any node with no more than nine neighbors must be able to adopt @xmath8 with a single adopting neighbor . this simply reflects the intuitive idea that for widespread diffusion to occur there must be a sufficient high frequency of nodes that are certain to propagate the adoption . we also observe that in both cases , for @xmath10 sufficiently low , there are two critical values for the parameter @xmath52 , @xmath125 such that a global cascade for a fixed @xmath10 is only possible for @xmath126 . the heuristic reason for these two thresholds is that a cascade can be prematurely stopped at high - degree nodes . for erds - rnyi random graphs , when @xmath127 , there exists a giant component , i.e. a connected component containing a positive fraction of the nodes . the high - degree nodes are quite infrequent so that the diffusion should spread easily . however , for @xmath52 close to one , the diffusion does not branch much and progresses along a very thin tree , `` almost a line '' , so that its progression is stopped as soon as it encounters a high - degree node . due to the variability of the poisson distribution , this happens before the diffusion becomes too big for @xmath128 . nevertheless the condition @xmath129 is not sufficient for a global cascade . global diffusion also requires that the network not be too highly connected . this is reflected by the existence of the second threshold @xmath130 where a further transition occurs , now in the opposite direction . for @xmath131 , the diffusion will not reach a positive fraction of the population . the intuition here is clear : the frequency of high - degree nodes is so large that diffusion can not avoid them and typically stops there since it is unlikely that a high enough fraction of their many neighbors eventually adopts . following @xcite , we say that these nodes are locally stable . of the cascade ( in percent of the total population ) for erds - rnyi random graphs as a function of @xmath52 the average degree for a fixed @xmath132 . the lower curve gives the asymptotic fraction of pivotal players @xmath133.,width=170 ] the proof of our theorem [ prop : cascade ] makes previous intuition rigorous for a more general model of diffusion and gives also more insights on the nature of the phase transitions . we describe it now . the lower curve in figure [ fig : cascade ] represents the number of pivotal players in an erds - rnyi random graph as a function of @xmath52 the average connectivity for @xmath134 : hence we keep only the largest connected component of an erds - rnyi random graph where we removed all vertices of degree greater than @xmath135 . by the same heuristic argument as above , we expect two phase transitions for the size of the set of pivotal players . in the proof of theorem [ prop : cascade ] , we show that it is indeed the case and moreover that the phase transitions occur at the same values @xmath136 and @xmath130 as can be seen on figure [ fig : cascade ] where the normalized size ( i.e. fraction ) @xmath133 of the set of pivotal players is positive only for @xmath137 . hence a cascade is possible if and only if there is a giant component of pivotal players . note also that both phase transitions for the pivotal players are continuous , in the sense that the function @xmath138 is continuous . this is not the case for the second phase transition for the normalized size of the cascade given by @xmath139 in proposition [ prop : cap ] : the function @xmath140 is continuous in @xmath136 but not in @xmath130 as depicted on figure [ fig : cascade ] . this has important consequences : around @xmath136 the propagation of cascades is limited by the connectivity of the network as in standard epidemic models . but around @xmath130 , the propagation of cascades is not limited by the connectivity but by the high - degree nodes which are locally stable . of the cascade ( in percent of the total population ) for erds - rnyi random graphs as a function of the frequency of the cascade.,width=170 ] to better understand this second phase transition , consider a dynamical model where at each step a player is chosen at random and switches to @xmath8 . once the corresponding cascade is triggered , we come back to the initial state where every node play @xmath7 before going to the next step . then for @xmath52 less than but very close to @xmath130 , most cascades die out before spreading very far . however , a set of pivotal players still exists , so very rarely a cascade will be triggered by such a player in which case the high connectivity of the network ensures that it will be extremely large . as @xmath52 approaches @xmath130 from below , global cascades become larger but increasingly rare until they disappear implying a discontinuous phase transition . for low values of @xmath52 , the cascades are infrequent and small . as @xmath52 increases , their frequencies and sizes also increase until a point where the cascade reaches almost all vertices of the giant component of the graph . then as @xmath52 increases , their sizes remain almost constant but their frequencies are decreasing . we considered so far the permanent adoption model : the only possible transitions are from playing @xmath7 to @xmath8 . there is another possible model in which the initial adopters playing @xmath8 also apply best - response update . we call this model the _ non - monotonic _ model ( by opposition to the permanent adoption model ) . in this model , if the dynamic converges , the final state will be an equilibrium of the game . an equilibrium of the game is a fixed point of the best response dynamics . for example , the states in which all players play @xmath7 or all players play @xmath8 are trivial equilibria of the game . note that the permanent adoption model does not necessarily yield to an equilibrium of the game as the initial seed does not apply best response dynamics . to illustrate the differences between the two models consider the graph of figure [ fig : contrex ] for a value of @xmath141 : if the circled player switches to @xmath8 , the whole network will eventually switch to @xmath8 in the permanent adoption model , whereas the dynamic for the non - monotonic model will oscillate between the state where only the circled player plays @xmath8 and the state where only his two neighbors of degree two play @xmath8 ( agents still revise their strategies synchronously ) . ( each triangle represents a long line of players).,width=75 ] clearly if a player induces a global cascade for the non - monotonic model , it will also induce a global cascade in the permanent adoption model . as illustrated by the case of figure [ fig : contrex ] , the converse is not true in general . hence , a priori , a pivotal player as defined in previous section might not induce a global cascade in the non - monotonic model . in @xcite , it is shown that if one can find a finite set of initial adopters causing a complete cascade for the permanent adoption model , it is also possible to find another , possibly larger but still finite , set of initial adopters leading to a complete cascade for the non - monotonic model . hence the contagion threshold as defined in @xcite for infinite graphs is the same for both models . in our case , we see that if we switch from @xmath7 to @xmath8 two pivotal players who are neighbors , then the whole set of pivotal players will eventually switch to @xmath8 in the non - monotonic model . hence in our analysis on random networks , we also say that the contagion threshold is the same in both models . moreover , both models will have exactly the same dynamics if started with the set of pivotal players playing @xmath8 and all other players playing @xmath7 . in particular , it shows that the dynamic converges and reaches an equilibrium of the game . hence we have the following corollary : [ prop : equil ] consider the random graph @xmath20 satisfying conditions [ cond ] and [ cond2 ] with asymptotic degree distribution @xmath45 . for @xmath75 , there exists w.h.p . an equilibrium of the game in which the number of players @xmath8 is more than @xmath142 ( defined in proposition [ prop : cap ] ) and it can be obtained from the trivial equilibrium with all players playing @xmath7 by switching only two neighboring pivotal players . hence for @xmath75 , the trivial equilibrium all @xmath7 , is rather weak since two pivotal players can induce a global cascade and there are @xmath66 such players so that switching two neighbors at random will after a finite number of trials ( i.e. not increasing with @xmath21 ) induce such a global cascade . we call this equilibrium the pivotal equilibrium . figure [ fig : mgeo ] shows the average number of trials required . it goes to infinity at both extremes @xmath136 and @xmath130 . moreover , we see that for most values of @xmath52 inside this interval , the average number of trials is less than @xmath57 . if @xmath80 , then by definition if there are pivotal players , their number must be @xmath83 . indeed , in the case of @xmath0-regular graphs , there are no pivotal players for @xmath143 . hence , it is either impossible or very difficult ( by sampling ) to find a set of players with cardinality bounded in @xmath21 leading to a global cascade since in all cases , their number is @xmath83 . as a function of @xmath52.,width=188 ] we now show that equilibria with _ co - existent _ strategies exist . here by co - existent , we mean that a connected fraction of players @xmath7 and @xmath8 are present in the equilibrium . more formally , we call a _ giant component _ : a subset of vertices containing a positive fraction ( as @xmath21 tends to infinity ) of the total size of the graph such that the corresponding induced graph is connected . we ask if there exists equilibrium with a giant component of players @xmath7 and @xmath8 ? in the case @xmath75 , for values of @xmath52 close to @xmath130 , the global cascade reaches almost all nodes of the giant component leaving only very small of connected component of players @xmath7 . however for @xmath52 close to @xmath136 , this is not the case and co - existence is possible as shown by the following proposition : [ prop : inactiveer ] in an erds - rnyi random graph @xmath53 , for @xmath75 , there exists @xmath144 $ ] such that : * for @xmath145 , in the pivotal equilibrium , there is coexistence of a giant component of players @xmath7 and a giant component of players @xmath8 . * for @xmath146 , in the pivotal equilibrium , there is no giant component of players @xmath7 , although there might be a positive fraction of players @xmath7 . . the lower ( green ) curve is the fraction of players @xmath8 in the pivotal equilibrium as a function of @xmath52 ( for @xmath147 ) . the last ( brown ) curve is the size of the giant component of players @xmath7 in the pivotal equilibrium as a function of @xmath52.,width=188 ] figure [ fig : inactive ] illustrates this proposition in the case of erds - rnyi random graphs @xmath53 . the difference between the upper ( red ) and lower ( green ) curve is exactly the fractions of players @xmath7 in the pivotal equilibrium while the brown curve represent the size ( in percentage of the total population ) of the largest connected component of players @xmath7 in the pivotal equilibrium . this curve reaches zero exactly at @xmath148 . hence for @xmath146 , we see that there is still a positive fraction of the population playing @xmath7 , but the set of players @xmath7 is divided in small ( i.e. of size @xmath82 ) connected components , like islands of players @xmath7 in a sea of players @xmath8 . note also that for erds - rnyi random graphs , the value of @xmath148 is close to @xmath136 . proposition [ prop : inactiveer ] follows from the following proposition whose proof is given in section [ sec : inactive ] . we first introduce some notations : for @xmath84 a random variable , we define for @xmath149 , @xmath150 the thinning of @xmath84 obtained by taking @xmath84 points and then randomly and independently keeping each of them with probability @xmath151 , i.e. given @xmath152 , @xmath153 . we define @xmath154 ( corresponding to the function @xmath155 defined in previous section for @xmath156 and @xmath157 ) , @xmath158 and @xmath159 [ prop : inactive ] consider the random graph @xmath20 satisfying conditions [ cond ] and [ cond2 ] with asymptotic degree distribution @xmath45 . let @xmath84 be a random variable with distribution @xmath51 . if @xmath75 , then @xmath160 . assume moreover that there exists @xmath33 such that @xmath161 for @xmath162 . there is coexistence of a giant component of players @xmath7 and a giant component of players @xmath8 in the pivotal equilibrium if @xmath163 > 0.\end{aligned}\ ] ] we consider now scenarios where @xmath164 $ ] and the initial set of adopters grows linearly with the total population @xmath21 . more precisely , consider now a firm advertising to a group of consumers , who share product information among themselves : potential buyers are not aware of the existence of the product and the firm undertakes costly informative advertising . the firm chooses the fraction of individuals who receive advertisements . individuals are located in a social network modeled by a random network @xmath20 with given vertex degrees as in previous section . however contrary to most work on viral marketing @xcite , @xcite , we assume that the advertiser has limited knowledge about the network : the firm only knows the proportions of individuals having different degrees in the social network . one possibility for the firm is to sample individuals randomly and to decide the costly advertising for this individual based on her degree ( i.e. her number of neighbors ) . the action of the firm is then encoded in a vector @xmath165 , where @xmath166 represents the fraction of individuals with degree @xmath1 which are directly targeted by the firm . these individuals will constitute the seed ( in a permanent adoption model ) and we call them the early adopters . note that the case @xmath167 for all @xmath1 corresponds to a case where the firm samples individuals uniformly . this might be one possibility if it is unable to observe their degrees . in order to optimize its strategy , the firm needs to compute the expected payoff of its marketing strategy as a function of @xmath168 . our results allows to estimate this function in terms of @xmath168 and the degree distribution in the social network . we assume that a buyer might buy either if she receives advertisement from the firm or if she receives information via word of mouth communication @xcite . more precisely , following @xcite , we consider the following general model for the diffusion of information : a buyer obtains information as soon as one of her neighbors buys the product but she decides to buy the product when @xmath169 where @xmath1 is her number of neighbors ( i.e. degree ) and @xmath170 the number of neighbors having the product , @xmath171 is a binomial random variable with parameters @xmath170 and @xmath172 $ ] and @xmath173 is a general random variable ( depending on @xmath1 ) . in words , @xmath174 is the probability that a particular neighbor does influence the possible buyer . this possible buyer does actually buy when the number of influential neighbors having bought the product exceeds a threshold @xmath173 . thus , the thresholds @xmath173 represent the different propensity of nodes to buy the new product when their neighbors do . the fact that these are possibly randomly selected is intended to model our lack of knowledge of their values and a possibly heterogeneous population . note that for @xmath175 and @xmath172 $ ] , our model of diffusion is exactly a contact process with probability of contagion between neighbors equals to @xmath174 . this model is also called the si ( susceptible - infected ) model in mathematical epidemiology @xcite . also for @xmath157 and @xmath176 , we recover the model of @xcite described previously where players @xmath7 are non - buyers and players @xmath8 are buyers . we now give the asymptotic for the final number of buyers for the case @xmath172 $ ] and @xmath177 , with @xmath178 ( the general case with a random threshold is given in theorem [ th : epi ] ) . we first need to introduce some notations . for integers @xmath179 and @xmath180 let @xmath181 denote the binomial probabilities @xmath182 . given a distribution @xmath183 , we define the functions : @xmath184 where @xmath185 . we define @xmath186:\ : g(z;\balpha,\boldp,\pi)= 0\right\}.\end{aligned}\ ] ] we refer to section [ sec : lmf ] for an intuition behind the definitions of the functions @xmath187 and @xmath188 in terms of a branching process approximation of the local structure of the graph . [ prop : ad ] consider the random graph @xmath189 for a sequence @xmath190 satisfying condition [ cond ] . if the strategy of the firm is given by @xmath168 , then the final number of buyers is given by @xmath191 provided @xmath192 , or @xmath193 $ ] , and further @xmath194 for any @xmath195 in some interval @xmath196 . for three different values of @xmath197 ( @xmath198 and @xmath199 ) . right : function @xmath200 ( the green function is @xmath201 , the size of the initial seed).,title="fig:",width=188 ] for three different values of @xmath197 ( @xmath198 and @xmath199 ) . right : function @xmath200 ( the green function is @xmath201 , the size of the initial seed).,title="fig:",width=188 ] to illustrate this result , we consider the simple case of erds - rnyi random graphs @xmath53 with @xmath157 , @xmath167 for all @xmath1 , and @xmath202 where @xmath112 is the contagion threshold for this network ( defined in previous section ) . for simplicity , we omit to write explicitly the dependence in @xmath174 and replace the dependence in @xmath51 by the parameter @xmath52 . as we will see , for some values of @xmath52 there is a phase transition in @xmath203 . for a certain value @xmath204 , we have : if @xmath205 , the size of the diffusion is rather small whereas it becomes very large for @xmath206 . clearly the advertising firm has a high interest in reaching the critical mass of @xmath207 early adopters . this phase transition is reminiscent of the one described in previous section . recall that we are now in a setting where @xmath202 so that a global cascade triggered by a single ( or a small number of ) individual(s ) uniformly sampled is not possible . hence , the firm will have to advertise to a positive fraction of individuals constituting the seed for the diffusion . the intuition is the following : if the seed is too small , then each early adopter starts a small cascade which is stopped by high - degree nodes . when the seed reaches the critical mass , then the cascades coalesce and are able to overcome barriers constituted by high - degree nodes so that a large fraction of the nodes in the giant component of the graph adopt . then increasing the size of the seed has very little effect since most of the time , the new early adopters ( chosen at random by the firm ) are already reached by the global diffusion . our theorem [ th : epi ] makes this heuristic rigorous for the general model of diffusion described above ( with random thresholds ) . we now give some more insights by exploring different scenarios for different values of @xmath10 . . the green ( upper ) curve is the size of the final set of players @xmath8 for @xmath208 given by @xmath209 as a function of @xmath52 . the brown ( lower ) curve corresponds to the minimal size of the seed for a global cascade @xmath210 . the red curve corresponds to the size of a global cascade when @xmath211 similar as in figure [ fig : cascade].,width=188 ] consider first a case where @xmath212 , then thanks to the results of previous section , we know that there exists @xmath213 such that for any @xmath214 , a global cascade is possible with positive probability if only one random player switches to @xmath8 . in particular , if a fraction @xmath203 of individuals uniformly sampled are playing @xmath7 , then for any @xmath215 , such a global cascade will occur with high probability . for @xmath131 , we know that a single player can not trigger a global cascade . more precisely , if players are chosen at random without any information on the underlying network , any set of initial adopters with size @xmath82 can not trigger a global cascade , as shown by theorem [ prop : cascade ] . however the final size of the set of players playing @xmath8 is a discontinuous function of @xmath203 the size of the initial seed . if @xmath204 is defined as the point at which this function is discontinuous , we have : for @xmath205 , the final set of players @xmath8 will be only slightly larger than @xmath203 but if @xmath206 the final set of players @xmath8 will be very large . hence we will say that there is a global cascade when @xmath206 and that there is no global cascade when @xmath205 . as shown in figure [ fig : qpetit ] , for @xmath212 , we have @xmath216 for @xmath217 $ ] and @xmath218 for @xmath131 . in figure [ fig : qpetit ] , we also see that our definition of global cascade when @xmath131 is consistent with our previous definition since the function @xmath219 giving the size of the final set of players @xmath8 when the seed has normalized size @xmath204 is continuous in @xmath52 and agrees with the previous curve for the size of the final set of players @xmath8 when @xmath220 and with only one early adopter . . minimal size of the seed for a global cascade @xmath210 as a function of @xmath52.,width=188 ] we now consider the case where @xmath221 . in this case , thanks to the results of previous section , we know that for any value of @xmath52 , a single initial player @xmath8 ( sampled uniformly ) can not trigger a global cascade . but our definition of @xmath204 still makes sense and we now have @xmath218 for all @xmath52 . figure [ fig : alphac ] gives a plot of the function @xmath222 . we see that again there are two regimes associated with the low / high connectivity of the graph . for low values of @xmath52 , the function @xmath223 is non - increasing in @xmath52 . this situation corresponds to the intuition that is correct for standard epidemic models according to which an increase in the connectivity makes the diffusion easier and hence the size of the critical initial seed will decrease . , low connectivity regime . size of the final set of players @xmath8 , @xmath224 as a function of the size of the initial seed @xmath203 for different values of @xmath225.,width=188 ] figure [ fig : va1 ] shows the size @xmath226 of the final set of players @xmath8 as a function of the size of the initial seed @xmath203 for various small values of @xmath52 . we see that for the smallest values of @xmath52 , there is no discontinuity for the function @xmath227 . in this case @xmath204 is not defined . we also see that there is a natural monotonicity in @xmath52 : as the connectivity increases , the diffusion also increases . however figure [ fig : alphac ] shows that for @xmath228 , the function @xmath222 becomes non - decreasing in @xmath52 . hence even if the connectivity increases , in order to trigger a global cascade the size of the initial seed has to increase too . the intuition for this phenomenon is now clear : in order to trigger a global cascade , the seed has to overcome the local stability of high - degree nodes which now dominates the effect of the increase in connectivity . , high connectivity regime . size of the final set of players @xmath8 , @xmath224 as a function of the size of the initial seed @xmath203 for different values of @xmath229 ( red , green , brown respectively).,width=188 ] figure [ fig : va2 ] shows the size @xmath226 of the final set of players @xmath8 as a function of the size of the initial seed @xmath203 for various value of @xmath52 . here we see that connectivity hurts the start of the diffusion : for small value of @xmath203 , an increase in @xmath52 results in a lower size for the final set of players @xmath8 ! however when @xmath203 reaches the critical value @xmath204 , then a global cascade occurs and its size is an increasing function of @xmath52 . in other words , high connectivity inhibits the global cascade but once it occurs , it facilitates its spread . in this subsection , we say that a player @xmath7 is inactive and a player @xmath8 is active . we describe an approximation to the local structure of the graph and present a heuristic argument which leads easily to a prediction for the asymptotic probability of being active . this heuristic allows us to recover the formulas for the functions @xmath230 . results from this section are not needed for the rigorous proofs given in sections [ sec : thepi ] and [ sec : casc ] and are only presented to give some intuition about our technical formulas . this branching process approximation is standard in the random graphs literature @xcite and is related to the generating function approach in physics @xcite . it is called local mean field ( lmf ) in @xcite and in the particular case of an epidemics with independent contagion on each edge , the lmf approximation was turned into a rigorous argument . for our model of diffusion with threshold , this seems unlikely to be straightforward and our proof will not use the lmf approximation . the lmf model is characterized by its connectivity distribution @xmath51 defined in condition [ cond ] . we now construct a tree - indexed process . let @xmath231 be a galton - watson branching process @xcite with a root which has offspring distribution @xmath232 and all other nodes have offspring distribution @xmath233 given by @xmath234 for all @xmath235 . recall that @xmath236 corresponds to the probability that an edge points to a node with degree @xmath0 , see @xcite . the tree @xmath231 describes the local structure of the graph @xmath237 ( as @xmath21 tends to infinity ) : the exploration of the successive neighborhoods of a given vertex is approximated by the branching process @xmath231 as long as the exploration is local ( typically restricted to a finite neighborhood independent of @xmath21 ) . we denote by @xmath238 the root of the tree and for a node @xmath5 , we denote by @xmath239 the generation of @xmath5 , i.e. the length of the minimal path from @xmath238 to @xmath5 . also we denote @xmath240 if @xmath5 belongs to the children of @xmath6 , i.e. @xmath241 and @xmath6 is on the minimal path from @xmath238 to @xmath5 . for an edge @xmath4 with @xmath240 , we denote by @xmath242 the sub - tree of @xmath231 with root @xmath5 obtained by the deletion of edge @xmath4 from @xmath231 . we now consider the diffusion model described in section [ sec : ad ] ( with threshold for a vertex of degree @xmath1 given by @xmath176 and probability for a neighbor to be influential given by @xmath172 $ ] ) on the tree @xmath231 , where the initial set of active nodes is given by a vector @xmath243}$ ] : if @xmath244 then vertex @xmath5 is in the initial set of active vertices , otherwise @xmath245 . in our model the @xmath246 s are independent bernoulli random variables with parameter @xmath247 where @xmath13 is the degree of node @xmath5 in the tree . thanks to the tree structure , it is possible to compute the probability of being active recursively as follows : for any node @xmath248 , let @xmath249 if node @xmath5 is active on the sub - graph @xmath250 with initial set of active nodes given by the restriction of @xmath251 to individuals in @xmath242 and with individual @xmath6 held fix in the inactive state . then for any node @xmath252 of degree @xmath13 , @xmath5 becomes active if the number of active influential children exceeds @xmath253 . hence , we get @xmath254 where the @xmath255 s are independent bernoulli random variables with parameter @xmath174 . then the state of the root is given by @xmath256 in order to compute the distribution of @xmath257 , we first solve the recursive distributional equation ( rde ) associated to the @xmath258 s : thanks to the tree structure , the random variables @xmath259 in ( [ eq : y ] ) are i.i.d . and have the same distribution as @xmath258 . hence their distribution solve the rde given by @xmath260 where for a given @xmath1 , the random variable @xmath261 is bernoulli with parameter @xmath166 , @xmath262 s are independent bernoulli with parameter @xmath174 , @xmath263 has distribution @xmath233 , @xmath264 and the @xmath258 s are i.i.d . copies ( with unknown distribution ) . to solve the rde ( [ eq : rde ] ) , we need to compute only the mean of the bernoulli random variable @xmath264 . hence taking expectation in ( [ eq : rde ] ) directly gives a fixed - point equation for this mean and the following lemma follows ( its proof is deferred to appendix [ sec : techlem ] ) : [ lem : lmf ] the solutions of the rde ( [ eq : rde ] ) are bernoulli random variable with mean @xmath151 , where @xmath151 is any solution of the fixed point equation : @xmath265 . for any such solution , the random variable @xmath257 defined in ( [ eq : x ] ) is a bernoulli random variable with mean @xmath266 . by the change of variable @xmath267 , we see that lemma [ lem : lmf ] is consistent with proposition [ prop : ad ] and allow to recover the functions @xmath230 . however note that , the rde ( [ eq : rde ] ) has in general several solutions one being the trivial one with @xmath268 corresponding to the trivial equilibrium where every node are active . the fact that this rde has several solutions is a major difficulty in order to turn this lmf approximation into a rigorous argument . finally , the crucial point allowing previous computation is the fact that in recursion ( [ eq : y ] ) the @xmath258 can be computed `` bottom - up '' so that the @xmath258 s of a given generation ( from the root ) are independent . the @xmath258 s in ( [ eq : x ] ) encode the information that @xmath5 is activated by a node in the subtree of @xmath231 `` below '' @xmath5 ( and not by the root ) . if one considers a node in the original graph and runs a directed contagion model on a local neighborhood of this node where only directed contagion toward this node are allowed , then the state of the graph seen from this node is well approximated by the @xmath258 s . we first present the model for the diffusion process and then our main results for the spread of the diffusion . in this section , we describe the diffusion for any given finite graph @xmath237 with vertex set @xmath36 $ ] . we still denote by @xmath13 the degree of node @xmath39 $ ] . from now on , a vertex @xmath5 is either active or inactive . in our model , the initial set of active nodes @xmath269 ( the seed ) will remain active during the whole process of the diffusion ( results for the non - monotonic model are derived in section [ sec : inactive ] ) . we will consider * single activation * where the seed contains only one vertex and * degree based activation * where the vertex @xmath5 is in the seed with probability @xmath247 , where @xmath13 is the degree of the vertex independently of the others . in other words , each node @xmath5 draws independently of each other a bernoulli random variable @xmath246 with parameter @xmath247 and is considered as initially active if @xmath244 and not initially active otherwise . in the case of degree based activation , we denote @xmath165 the parameters of this activation . in particular , if @xmath270 for all @xmath1 , then a fraction @xmath203 chosen uniformly at random among the population is activated before the diffusion takes place . * symmetric threshold model : * we first present the symmetric threshold model which generalizes the bootstrap percolation @xcite : a node becomes active when a certain threshold fraction of neighbors are already active . we allow the threshold fraction to be a random variable with distribution depending on the degree of the node and such that thresholds are independent among nodes . formally , we define for each node @xmath5 a sequence of random variables in @xmath271 denoted by @xmath272 . the threshold associated to node @xmath5 is @xmath273 where @xmath13 is the degree of node @xmath5 . we assume that for any two vertices @xmath5 and @xmath6 , the sequences @xmath272 and @xmath274 are independent and have the same law as a sequence denoted by @xmath275 . for @xmath276 , we denote @xmath277 the probability distribution of the threshold for a node of degree @xmath278 . for example in the model of @xcite described in the introduction , we take @xmath279 so that @xmath280 . we will use the notation @xmath281 and @xmath282 denotes the distribution of thresholds . now the progressive dynamic of the diffusion on the finite graph @xmath237 operates as follows : some set of nodes @xmath269 starts out being active ; all other nodes are inactive . time operates in discrete steps @xmath283 . at a given time @xmath284 , any inactive node @xmath5 becomes active if its number of active neighbors is at least @xmath285 . this in turn may cause other nodes to become active . it is easy to see that the final set of active nodes ( after @xmath21 time steps if the network is of size @xmath21 ) only depends on the initial set @xmath269 ( and not on the order of the activations ) and can be obtained as follows : set @xmath286 for all @xmath5 . then as long as there exists @xmath5 such that @xmath287 , set @xmath288 , where @xmath289 means that @xmath5 and @xmath6 share an edge in @xmath237 . when this algorithm finishes , the final state of node @xmath5 is represented by @xmath290 : @xmath288 if node @xmath5 is active and @xmath291 otherwise . note that we allow for the possibility @xmath292 in which case , node @xmath5 is never activated unless it belongs to the initial set @xmath269 . note also that the condition @xmath293 is actually not restrictive . if we wish to consider a model where @xmath294 is possible , we just need to modify the initial seed @xmath269 so as to put node @xmath5 in @xmath269 if @xmath294 . hence this model is equivalent to ours if we increase @xmath247 accordingly . * percolated threshold model : * this model depends on a parameter @xmath172 $ ] and a distribution of random thresholds @xmath295 given by @xmath296 as described above . given any graph @xmath237 and initial set @xmath269 , we now proceed in two phases . * bond percolation : randomly delete each edge with probability @xmath297 independently of all other edges . we denote the resulting random graph by @xmath298 ; * apply the symmetric threshold model with thresholds @xmath173 : set @xmath286 and then as long as there is @xmath5 such that @xmath299 , set @xmath288 , where @xmath300 means that @xmath5 and @xmath6 share an edge in @xmath298 and @xmath13 is the degree of node @xmath5 in the original graph @xmath237 . clearly if @xmath157 , this is exactly the symmetric threshold model . if in addition @xmath301 , then this model is known as bootstrap percolation @xcite . on the other hand if @xmath302 and @xmath175 for any @xmath1 , then this is the contact process with probability of activation @xmath174 on each edge . note that the percolated threshold model is not equivalent to the symmetric threshold model on the ( bond ) percolated graph since threshold depends on the degree in the original graph ( and not in the percolated graph ) . recall that for integers @xmath179 and @xmath180 , @xmath181 denotes the binomial probabilities @xmath182 . for a graph @xmath237 , let @xmath303 and @xmath304 denote the numbers of vertices and edges in @xmath237 respectively . in this section , we assume that the followings are given : a sequence of independent thresholds @xmath305 drawn according to a distribution @xmath306 , a parameter @xmath172 $ ] and an activation set drawn according to the distribution @xmath168 . the subgraph of @xmath189 induced by the activated ( resp . inactive ) nodes at the end of the diffusion is denoted by @xmath307 ( resp . @xmath308 ) . for @xmath309 , we denote by @xmath310 the number of vertices in @xmath308 with degree @xmath278 in @xmath237 and @xmath0 in @xmath308 , i.e. the number of vertices with degree @xmath278 in @xmath237 which are not activated and with @xmath0 neighbors which are not activated either . we denote by @xmath311 the number of activated vertices of degree @xmath278 in @xmath237 ( and with possibly lower degree in @xmath307 ) . we define the functions : @xmath312 where @xmath185 . we define @xmath313:\ : g(z;\boldp,\boldt,\balpha , \pi)= 0\right\}.\end{aligned}\ ] ] we refer to appendix [ sec : max ] for a justification of the use of the @xmath314 in ( [ eq : max ] ) . [ th : epi ] consider the random graph @xmath189 ( or @xmath41 ) for a sequence @xmath43 satisfying condition [ cond ] . we are given with a sequence of independent thresholds @xmath305 drawn according to a distribution @xmath306 , a parameter @xmath172 $ ] and an activation set drawn according to the distribution @xmath168 . for the percolated threshold diffusion on the graph @xmath189 , we have : if @xmath315 , or if @xmath316 $ ] , and further @xmath317 for any @xmath195 in some interval @xmath196 , then @xmath318 if we condition the induced graph @xmath319 of inactive nodes in @xmath41 on its degree sequence @xmath320 and let @xmath321 be the number of its vertices , then @xmath319 has the distribution of @xmath322 . the proof of this theorem is given in section [ sec : thepi ] . note that proposition [ prop : ad ] follows easily . in this section , we look at the diffusion with one ( or a small number @xmath82 of ) initial active node(s ) in @xmath189 . for @xmath323 $ ] , let @xmath324 ( resp . @xmath325 ) be the subgraph induced by the final active nodes with initial active node @xmath326 ( resp . initial active nodes @xmath327 ) . we also define @xmath328 as the subgraph induced by the inactive nodes with initial active node @xmath326 . the set of vertices of @xmath324 and @xmath328 is a partition of the vertices of the original graph . to ease notation , we denote @xmath329 , @xmath330 and @xmath331 . we define @xmath332 we also define @xmath333 and @xmath334 . we call the following condition the cascade condition : @xmath335 which can be rewritten as @xmath336>{{\mathbb{e}}}[d]$ ] where @xmath84 is a random variable with distribution @xmath51 . we denote by @xmath337 the largest connected component of the graph @xmath338 on which we apply a bond percolation with parameter @xmath174 ( i.e. we remove each edge independently with probability @xmath297 ) and then apply a site percolation by removing all vertices with @xmath339 . the vertices of the connected graph @xmath340 are called pivotal vertices : for any @xmath341 , we have @xmath342 . [ prop : cascade ] consider the random graph @xmath189 ( or @xmath41 ) for a sequence @xmath43 satisfying conditions [ cond ] and [ cond2 ] . we are given with a sequence of independent thresholds @xmath305 drawn according to a distribution @xmath306 and a parameter @xmath172 $ ] . * if the cascade condition ( [ eq : casc ] ) is satisfied , then @xmath343 moreover , for any @xmath341 , we have w.h.p . @xmath344 where @xmath345 is defined by ( [ eq : defxi ] ) . moreover if @xmath346 or @xmath345 is such that there exists @xmath33 with @xmath347 for @xmath162 , then we have for any @xmath341 : @xmath348 if we condition the induced graph @xmath349 of inactive nodes in @xmath41 on its degree sequence @xmath350 and let @xmath351 be the number of its vertices , then @xmath349 has the distribution of @xmath352 . * if @xmath353 , then for any @xmath354 , @xmath355 . a proof of this theorem is given in section [ sec : casc ] . we end this section with some remarks : if the distribution of @xmath173 does not depend on @xmath1 , then the cascade condition becomes with @xmath84 a random variable with distribution @xmath51 : @xmath356}{{{\mathbb{e}}}[d(d-1)]}.\end{aligned}\ ] ] in particular , if @xmath357 ( i.e. the diffusion is a simple exploration process ) , then we find the well - known condition for the existence of a giant component. this corresponds to existing results in the literature see in particular theorem 3.9 in @xcite which extend the standard result of @xcite . more generally , in the case @xmath172 $ ] and @xmath357 ( corresponding to the contact process ) , a simple computation shows that @xmath358 where @xmath359 $ ] is the generating function of the asymptotic degree distribution . applying theorems [ th : epi ] and [ prop : cascade ] allow to obtain results for the contact process . similarly , the bootstrap percolation has been studied in random regular graphs @xcite and random graphs with given vertex degrees @xcite . the bootstrap percolation corresponds to the particular case of the percolated threshold model with @xmath157 and @xmath360 and our theorems [ th : epi ] and [ prop : cascade ] allow to recover results for the size of the diffusion . finally , the case where @xmath176 and @xmath157 implies directly proposition [ prop : cap ] . it is well - known that it is often simpler to study the random multigraph @xmath41 with given vertex sequence @xmath43 defined in section [ sec : conf ] . we consider asymptotics as the number of vertices tends to infinity and thus assume throughout the paper that we are given , for each @xmath21 , a sequence @xmath361 with @xmath362 even . we may obtain @xmath189 by conditioning the multigraph @xmath41 on being a simple graph . by @xcite , we know that the condition @xmath363 implies @xmath364 . in this case , many results transfer immediately from @xmath41 to @xmath189 , for example , every result of the type @xmath365 for some events @xmath366 , and thus every result saying that some parameter converges in probability to some non - random value . this includes every results in the present paper . henceforth , we will in this paper study the random multigraph @xmath41 and in a last step ( left to the reader ) transfer the results to @xmath189 by conditioning . we run the dynamic of the diffusion of section [ sec : stm ] on a general graph @xmath41 in order to compute the final size of the diffusion in a similar way as in @xcite . the main point here consists in coupling the construction of the graph with the dynamic of the diffusion . the main difference with previous analysis consists in adding to each vertex a threshold with distribution @xmath306 independently form the graph @xmath41 as described in section [ sec : stm ] . the proof of theorem [ th : epi ] follows then easily see section [ sec : spread ] . in order to prove theorem [ prop : cascade ] , we use the same idea of coupling ( in a similar spirit as in @xcite for the analysis of the giant component ) but we have to deal with an additional difficulty due to the following lack of symmetry : if @xmath324 is the final set of the diffusion with only @xmath326 as initial active node , then for any @xmath367 , we do not have in general @xmath368 . we take care of this difficulty in section [ sec : casc ] . in the next section , we present a preliminary lemma that will be used in the proofs . a pure death process with rate 1 is a process that starts with some number of balls whose lifetime are i.i.d . rate 1 exponentials . each time a ball dies , it is removed . now consider @xmath21 bins with independent rate @xmath369 death processes . to each bin , we attach a couple @xmath370 where @xmath13 is the number of balls at time 0 in the bin and @xmath273 is the threshold corresponding to the bin . we now modify the death process as follows : all balls are initially white . for any living ball , when it dies , with probability @xmath297 and independently of everything else , we color it green ( an leave it in he bin ) , otherwise we remove it from the bin . let @xmath371 and @xmath372 denote the number of white and green balls respectively in bin @xmath6 at time @xmath284 , where @xmath373 and @xmath374 . let @xmath375 be the number of bins that have @xmath278 balls at time @xmath12 and @xmath0 white balls , @xmath5 green balls at time @xmath284 and threshold @xmath376 , i.e. @xmath377 . in what follows we suppress the superscripts to lighten the notation . the following lemma is an extension of lemma 4.4 in @xcite : [ lem : death ] consider the @xmath21 independent processes defined above and assume that the sequence @xmath43 satisfies condition [ cond ] where @xmath45 can be a defective probability distribution : @xmath378 . we assume that the thresholds @xmath379}$ ] are independent in @xmath5 and have distribution @xmath282 as described in section [ sec : stm ] . then , with the above notation , as @xmath23 , @xmath380 let @xmath381 . in particular @xmath382 . first fix integers @xmath383 and @xmath6 with @xmath384 . consider the @xmath385 bins that start with @xmath278 balls and with threshold @xmath376 . for @xmath386 , let @xmath387 be the time the @xmath6th ball is removed or recolored in the @xmath170th such bin . then @xmath388 . moreover , for the @xmath170th bin , we have @xmath389 . multiplying by @xmath390 and using glivenko - cantelli theorem ( see e.g. proposition 4.24 in @xcite ) , we have @xmath391 since @xmath392 ( by condition [ cond](i ) and the law of large numbers ) , we see that @xmath393 the law of @xmath394 given @xmath395 and @xmath396 is a binomial distribution with parameters @xmath397 and @xmath297 . hence by the law of large numbers , we have @xmath398 hence each term in ( [ eq : limsup ] ) tends to @xmath12 in probability . the same holds for any finite partial sum . let @xmath33 and @xmath269 be such that @xmath399 . by condition [ cond](iii ) , we have @xmath400 . hence , also @xmath401 . so that for sufficiently large @xmath21 , we get @xmath402 and @xmath403 and the lemma follows . our proof of theorem [ th : epi ] is an adaptation of the coupling argument in @xcite . we start by analyzing the symmetric threshold model . we can view the algorithm of section [ sec : stm ] as follows : start with the graph @xmath237 and remove vertices from @xmath269 , i.e. the initially active nodes . as a result , if vertex @xmath5 has not been removed , its degree has been lowered . we denote by @xmath404 the degree of @xmath5 in the evolving graph . then iteratively remove vertices @xmath5 such that @xmath405 . all removed vertices at the end of this procedure are active and all vertices left are inactive . it is easily seen that we obtain the same result by removing edges where one endpoint satisfy @xmath405 , until no such edge remains , and finally removing all isolated vertices , which correspond to active nodes . regard each edge as consisting of two half - edges , each half - edge having one endpoint . we introduce types of vertices . we set the type of vertices in the seed @xmath269 to @xmath8 . say that a vertex ( not in @xmath269 ) is of type @xmath7 if @xmath406 and of type @xmath8 otherwise . in particular at the beginning of the process , all vertices not in @xmath269 are of type @xmath7 since @xmath407 and all vertices in @xmath269 are by definition of type @xmath8 . as the algorithm evolves , @xmath408 decreases so that some type @xmath7 vertices become of type @xmath8 during the execution of the algorithm . similarly , say that a half - edge is of type @xmath7 or @xmath8 when its endpoint is . as long as there is any half - edge of type @xmath8 , choose one such half - edge uniformly at random and remove the edge it belongs to . this may change the other endpoint from @xmath7 to @xmath8 ( by decreasing @xmath409 ) and thus create new half - edges of type @xmath8 . when there are no half - edges of type @xmath8 left , we stop . then the final set of active nodes is the set of vertices of type @xmath8 ( which are all isolated ) . as in @xcite , we regard vertices as bins and half - edges as balls . at each step , we remove first one random ball from the set of balls in @xmath8-bins and then a random ball without restriction . we stop when there are no non - empty @xmath8-bins . we thus alternately remove a random @xmath8-ball and a random ball . we may just as well say that we first remove a random @xmath8-ball . we then remove balls in pairs , first a random ball and then a random @xmath8-ball , and stop with the random ball , leaving no @xmath8-ball to remove . we change the description a little by introducing colors . initially all balls are white , and we begin again by removing one random @xmath8-ball . subsequently , in each deletion step we first remove a random white ball and then recolor a random white @xmath8-ball red ; this is repeated until no more white @xmath8-balls remain . we now run this deletion process in continuous time such that , if there are @xmath6 white balls remaining , then we wait an exponential time with mean @xmath410 until the next pair of deletions . in other words , we make deletions at rate @xmath6 . this means that each white ball is deleted with rate 1 and that , when we delete a white ball , we also color a random white @xmath8-ball red . let @xmath411 and @xmath412 denote the numbers of white @xmath7-balls and white @xmath8-balls at time @xmath284 , respectively , and @xmath413 denotes the number of @xmath7-bins at time @xmath284 . since red balls are ignored , we may make a final change of rules , and say that all balls are removed at rate @xmath369 and that , when a white ball is removed , a random white @xmath8-ball is colored red ; we stop when we should recolor a white @xmath8-ball but there is no such ball . let @xmath414 be the stopping time of this process . first consider the white balls only . there are no white @xmath8-balls left at @xmath414 , so @xmath415 has reached zero . however , let us consider the last deletion and recoloring step as completed by redefining @xmath416 ; we then see that @xmath414 is characterized by @xmath416 and @xmath417 for @xmath418 . moreover , the @xmath7-balls left at @xmath414 ( which are all white ) are exactly the half - edges in the induced subgraph @xmath308 of inactive nodes . hence , the number of edges in this subgraph is @xmath419 , while the number of nodes not activated is @xmath420 . if we consider only the total number @xmath421 of white balls in the bins , ignoring the types , the process is as follows : each ball dies at rate @xmath369 and upon its death another ball is also sacrificed . the process @xmath421 is a death process with rate @xmath57 ( up to time @xmath414 ) . consequently , by lemma 4.3 of @xcite ( or lemma [ lem : death ] above ) , we have @xmath422 since condition [ cond ] ( iii ) implies @xmath423 . now if we consider the final version of the process restricted to @xmath7-bins , it corresponds exactly to the death process studied in section [ sec : lem ] above with @xmath157 . we need only to compute the initial condition for this process . for a degree based activation , each vertex of degree @xmath278 is activated ( i.e. the corresponding bin becomes a @xmath8-bin ) with probability @xmath424 . hence by the law of large numbers , the number of @xmath7-bins with initially @xmath278 balls and threshold @xmath376 is @xmath425 . with the notation of lemma [ lem : death ] , we have @xmath426 with the defective probability distribution @xmath427 . hence by lemma [ lem : death ] we get ( recall that @xmath157 here ) : @xmath428 it is then easy to finish the proof as in @xcite for this model ( the complete argument is presented below for the more general percolated threshold model ) . in particular , it ends the proof of theorem [ th : epi ] for the case @xmath157 . we now consider the percolated threshold model with @xmath429 . we modify the process as follows : for any white @xmath7-ball when it dies , with probability @xmath297 , we color it green instead of removing it . a bin is of type @xmath7 if @xmath430 , where @xmath0 is the number of withe balls in the bin , @xmath5 the number of green balls ( which did not transmit the diffusion ) and @xmath278 and @xmath376 are the initial degree and threshold . let @xmath411 be the number of white @xmath7-balls . by lemma [ lem : death ] , we now have : @xmath431 in particular , we have thanks to lemma [ lem : equiv ] ( in appendix [ sec : techlem ] ) for @xmath432 , @xmath433 by looking at white balls ( without taking types in consideration ) , we see that equation ( [ eq : ab ] ) is still valid . hence , we have @xmath434 for simplicity we write @xmath435 for @xmath436 . assume now that @xmath437 is a constant independent of @xmath21 with @xmath438 so that @xmath439 and @xmath440 . hence , we have @xmath441 for @xmath442 $ ] and thus @xmath443 for @xmath444 . we can find some @xmath445 such that @xmath446 for @xmath444 . but @xmath416 , so if @xmath447 then @xmath448 and from ( [ eq : b ] ) , we have @xmath449 . in case @xmath315 , we may take any finite @xmath450 and find @xmath451 and ( [ eq : limtau ] ) with @xmath452 , yields that @xmath453 in case @xmath454 , by the hypothesis we can find @xmath455 such that @xmath456 . if @xmath457 , then @xmath458 and thus @xmath459 . hence by ( [ eq : b ] ) , we have @xmath460 . since we can choose @xmath450 and @xmath461 arbitrarily close to @xmath462 , we have @xmath463 and ( [ eq : limtau ] ) with @xmath464 yields that @xmath465 note that @xmath466 and @xmath467 . hence we proved theorem [ th : epi ] for @xmath468 and @xmath469 . the results for @xmath470 and @xmath471 follows from the same argument , once we note that @xmath472 finally , the statement concerning the distribution of the induced subgraph @xmath319 follows from the fact that this subgraph has not been explored when previous algorithm stops . we start this section with some simple calculations . we define for @xmath473 $ ] , @xmath474 for @xmath475 and @xmath476 , we have @xmath477 so that for @xmath473 $ ] , we have @xmath478 . hence by condition [ cond2 ] , @xmath479 is differentiable on @xmath480 $ ] and we have @xmath481 in particular , we have @xmath482 so that we have @xmath483 . note also that @xmath484 and @xmath485 . consider now the case ( ii ) where @xmath486 , so that @xmath487 . the proof for an upper bound on @xmath488 follows easily from theorem [ th : epi ] . take a parameter @xmath489 with @xmath490 for all @xmath1 . clearly the final set of active nodes @xmath491 will be greater than for any seed with size @xmath83 . now when @xmath203 goes to zero , the fact that @xmath492 ensures that @xmath493 in theorem [ th : epi ] so that @xmath494 . hence point ( ii ) in theorem [ prop : cascade ] follows . we now concentrate on the case where the cascade condition holds . in particular we have @xmath495 so that @xmath345 defined in ( [ eq : defxi ] ) by @xmath496 is strictly less than one and we have @xmath497 also as soon as there exists @xmath33 such that @xmath347 for @xmath498 , we can use the same argument as above . since , we have in this case @xmath499 as @xmath500 , it gives an upper bound that matches with the statement ( i ) of theorem [ prop : cascade ] . in order to prove a lower bound , we follow the general approach of @xcite . we modify the algorithm defined in section [ sec : spread ] as follows : the initial set @xmath269 now contains only one vertex , say @xmath501 . when there is no half - edge ( or ball ) of type @xmath8 , we say that we make an exception and we select a vertex ( or a bin ) of type @xmath7 uniformly at random among all vertices of type @xmath7 . we declare its white half - edges of type @xmath8 and remove its other half - edges , i.e. remove the green balls contained in the corresponding bin if there are any . exceptions are done instantaneously . for any set of nodes @xmath502 , let @xmath503 be the subgraph induced by the final active nodes with initial active nodes @xmath502 . if @xmath504 , then clearly when the algorithm has exhausted the half - edges of type @xmath8 , it removed the subgraph @xmath505 from the graph and all edges with one endpoint in @xmath505 . then an exception is made by selecting a vertex say @xmath506 in @xmath507 . similarly when the algorithm exhausted the half - edges of type @xmath8 , it removed the subgraph @xmath508 and all edges with one endpoint in this set of vertices . more generally , if @xmath170 exceptions are made consisting of selecting nodes @xmath509 , then before the @xmath510-th exception is made ( or at termination of the algorithm if there is no more exception made ) , the algorithm removed the subgraph @xmath503 and all edges with one endpoint in this set of vertices . we use the same notation as in section [ sec : spread ] . in particular , we still have : @xmath511 we now ignore the effect of the exceptions by letting @xmath512 be the number of white @xmath7 balls if no exceptions were made , i.e. assuming @xmath513 for all @xmath284 . if @xmath514 is the maximum degree of @xmath515 , then we have @xmath516 $ ] . by condition [ cond ] ( iii ) , @xmath517 , and thus @xmath518 . hence we can apply results of previous section : @xmath519 we now prove that : @xmath520 the fact that @xmath521 is clear . furthermore , @xmath522 increases only when exceptions are made . if an exception is made at time @xmath284 , then the process @xmath8 reached zero and a vertex with @xmath6 white half - edges is selected so that @xmath523 . hence we have @xmath524 between exceptions , if @xmath7 decreases by one then so does @xmath525 , hence @xmath522 does not increase . consequently if @xmath278 was the last time before @xmath284 that an exception was performed , then @xmath526 and ( [ eq : bd ] ) follows . let @xmath527 , then we have @xmath528 equation ( [ eq : bd ] ) can be written as @xmath529 we first assume that @xmath345 given by ( [ eq : defxi ] ) is such that @xmath530 and there exists @xmath33 such that @xmath347 and hence @xmath531 for @xmath162 . let @xmath532 . then by ( [ eq : pos ] ) , we have @xmath533 for @xmath534 so that @xmath535 and hence by ( [ eq : tb ] ) , @xmath536 by condition [ cond ] ( iii ) , @xmath537 . consequently , ( [ eq : bd1 ] ) yields @xmath538 and thus by ( [ eq : tb ] ) @xmath539 by assumption , there exists @xmath33 sufficiently small for @xmath540 . since @xmath533 on the interval @xmath541 $ ] , ( [ eq : b2 ] ) implies that w.h.p . @xmath412 remains positive on @xmath541 $ ] , and thus no exception is made during this interval . on the other hand , @xmath540 and ( [ eq : tb ] ) implies @xmath542 , while @xmath543 . thus with @xmath544 , w.h.p . @xmath545 while ( [ eq : tbtau ] ) yields @xmath546 w.h.p . consequently , w.h.p . @xmath547 and an exception is performed between @xmath414 and @xmath548 . let @xmath549 be the last time an exception was performed before @xmath550 and let @xmath551 be the next time it is performed . we have shown that for any @xmath33 , w.h.p . @xmath552 and @xmath553 . [ lm : int ] let @xmath554 and @xmath555 be two random times when an exception is performed , with @xmath556 , and assume that @xmath557 and @xmath558 , where @xmath559 . if @xmath560 is the number of vertices removed by the algorithm by time @xmath561 , then @xmath562 in particular , if @xmath563 , then @xmath564 . by definition , we have : @xmath565 since @xmath566 and @xmath567 is continuous , @xmath568 , and ( [ eq : tb ] ) and ( [ eq : bd1 ] ) imply , in analogy with ( [ eq : tbinf ] ) and ( [ eq : tbtau ] ) , @xmath569 and @xmath570 let @xmath571 be the number of bins if no exceptions were made . then we clearly have @xmath572 since each time an exception is made @xmath573 increases by one while @xmath522 increases by more than one . hence ( [ eq : lev ] ) follows from results in previous section . let @xmath574 ( resp . @xmath575 ) be the subgraph induced by the vertices removed by the algorithm between @xmath12 and @xmath549 ( resp . @xmath551 ) . by lemma [ lm : int ] , we have @xmath576 hence informally , the exception made at time @xmath549 triggers a large cascade . now consider the case @xmath346 . note in particular that we have @xmath577 and since @xmath578 , we also have @xmath579 . then with the same argument as above , we have that @xmath412 remains positive on @xmath580 and thus no exception is made after a last exception made at time @xmath549 with @xmath581 . hence ( [ eq : c ] ) and ( [ eq : c ] ) follow with @xmath575 being the whole graph ( except possibly @xmath83 vertices ) , @xmath579 . we now finish the proof of theorem [ prop : cascade ] in the case where @xmath346 or there exists @xmath33 such that @xmath347 for @xmath582 . first by @xcite theorems 3.5 and 3.9 , the cascade condition implies that @xmath583 . the result @xmath584 could be derived from @xcite . we give an alternative proof for this result at the end of this section . for now , we denote @xmath585 . coming back to the diffusion process analyzed above , we clearly have @xmath586 . we now prove that @xmath587 . this is clear in the case @xmath346 . we now concentrate on the case @xmath530 . first , let @xmath588 be the first time after @xmath551 when an exception is made . since @xmath522 increases by at most @xmath589 each time an exception is made , we obtain from ( [ eq : ta - a ] ) : @xmath590 hence similarly as in ( [ eq : tau ] ) , we have for every @xmath33 , w.h.p . @xmath591 . since also @xmath592 , it follows that @xmath593 . if @xmath594 is the subgraph removed by the algorithm between @xmath551 and @xmath588 , then lemma [ lm : int ] implies that @xmath595 . assume now that @xmath596 , then with probability at least @xmath597 , the vertex chosen for the exception at @xmath551 belongs to @xmath340 and then we have @xmath598 , in contradiction with @xmath595 . hence we have @xmath599 and then @xmath587 as claimed . we clearly have for any @xmath341 : @xmath600 hence we only need to prove that @xmath601 . to see this , attach to each vertex @xmath5 a random variable @xmath602 uniformly distributed over @xmath480 $ ] . each time an exception has to be made , pick the vertex among the remaining ones with minimal @xmath602 so that we do not change the algorithm described at the beginning of the section . from the analysis above , we see that all exceptions made before @xmath549 are vertices not in @xmath340 and the exception made at time @xmath549 belongs to @xmath340 . now consider the graph @xmath603 obtained from the original graph where @xmath574 has been removed but all other variables are the same as in the original graph . since @xmath604 , this graph satisfies conditions [ cond ] and [ cond2 ] . hence previous analysis applies and we have in addition that the first exception made by the algorithm belongs to @xmath340 since a pivotal vertex in @xmath237 is also pivotal in @xmath603 . hence the subgraph of @xmath603 removed between times @xmath605 and @xmath606 by the algorithm is exactly @xmath607 in @xmath603 . since @xmath603 is a subgraph of @xmath237 , we have @xmath608 in the original graph . and the first claim in ( i ) follows from ( [ eq : c ] ) applied to the graph @xmath603 . the second claim in ( i ) follows exactly as in the proof of theorem [ th : epi ] given in previous section . now consider the case where @xmath530 but for any @xmath33 , there exists @xmath162 such that @xmath609 and hence @xmath610 . the idea to get a lower bound is to let @xmath174 vary . since @xmath611 , for any @xmath612 , we see that by a standard coupling argument , for any given initial seed all active nodes in the model with @xmath613 will also be active in the model with @xmath174 . hence the model with @xmath613 provides a lower bound for the number of active nodes in the model with @xmath174 . now consider the function @xmath614 as a function of @xmath174 . we have @xmath615\end{aligned}\ ] ] since @xmath345 is a local minimum of @xmath616 and @xmath155 is differentiable as a function of @xmath195 , we have @xmath617 hence we have @xmath618 . in particular for @xmath619 , we have @xmath620 . let @xmath621 , then we have @xmath622 for any @xmath623 and @xmath624 for @xmath625 . moreover , we have @xmath626 as @xmath627 and previous argument is valid for the model with @xmath613 as close as desired from @xmath174 showing that @xmath628 is a lower bound for the fraction of active nodes in the model with @xmath174 . we finish this proof by computing the asymptotic for the size of @xmath340 using our previous analysis but for a modified threshold as done in @xcite . we add a bar for quantities associated to this new model . namely , consider a modification of the original diffusion with threshold @xmath629 . in words , a node @xmath5 becomes active if one of its neighbor is active and @xmath630 in the original diffusion model . clearly the nodes that become active in this model need to have only one active neighbor in the original diffusion model with parameter @xmath631 . we denote by @xmath632 the subgraph of final active nodes with initial active node @xmath326 . note that our algorithm is an exploration process of the components of the graph on which we apply a bond percolation with parameter @xmath174 and a site percolation by removing all vertices with @xmath339 . in particular , the largest component explored is exactly @xmath340 . note that the computations made before are valid with the following functions : @xmath633 hence we have @xmath634,\ : \bg(z)=0 \}$ ] . note that if we denote @xmath635 , then we have @xmath636 in particular , @xmath637 is concave on @xmath638 $ ] and strictly concave unless @xmath639 for @xmath640 . note also that @xmath641 , so that under the cascade condition @xmath642 and @xmath637 is strictly concave . hence , we have @xmath643 and @xmath644 for @xmath645 and if @xmath530 , then @xmath646 for @xmath647 . in particular , previous analysis allows to conclude that @xmath584 . we start with the proof of proposition [ prop : inactive ] . by theorem [ prop : cascade ] , when the cascade condition holds , the set of active vertices contain the set of pivotal vertices and hence has a giant component . we denote @xmath648 the induced subgraph of inactive vertices , in the pivotal equilibrium ( i.e. in the final state when all pivotal nodes are initially active ) . by theorem [ prop : cascade ] , we have for @xmath649 , the number of vertices in @xmath648 with degree @xmath0 in @xmath648 : @xmath650 we denote @xmath651 . thanks to the result on the distribution of @xmath648 , there is a giant component of inactive vertices if @xmath652 which can be rewritten as in proposition [ prop : inactive ] . we now prove proposition [ prop : inactiveer ] . we assume that @xmath653 , @xmath157 and @xmath654 . the function @xmath655 is increasing for @xmath656 and then decreasing . we assume that @xmath10 is fixed such that the cascade condition holds . then @xmath657 . then @xmath658 while @xmath659 . we denote @xmath660 then , there is a giant component of inactive vertices if @xmath661 . both functions are non - increasing in @xmath137 and are intersecting only once in @xmath662 . this paper analyzes the spread of new behaviors or technologies in large social networks . our analysis is motivated by the two qualitative features of global cascades in social and economics systems : they occur rarely , but are large when they do . in our simplest model , agents play a local interaction binary game where the underlying social network is modeled by a sparse random graph . considering the deterministic best response dynamics , we compute the contagion threshold for this model . we find that when the social network is sufficiently sparse , the contagion is limited by the low connectivity of the network ; when it is sufficiently dense , the contagion is limited by the stability of the high - degree nodes . this phenomenon explains why contagion is possible only in a given range of the global connectivity . most importantly , we identify the set of agents able to trigger a large cascade : the pivotal players , i.e. the largest component of players requiring a single neighbor to change strategy in order to follow the change . the notion of pivotal player is crucial in the context of random graphs . cascades occur only when pivotal players represent a positive fraction of the population and in this case , any cascade will be triggered by such a pivotal player . we found that these pivotal players exist only if the connectivity of the network is in a given range . at both ends of this range ( i.e. for low and high - connectivity ) , the number of pivotal players is low . however in the high - connectivity case , we found that the system displays a robust - yet - fragile quality : while pivotal players are very rare , they trigger very large cascades . this feature makes global contagions exceptionally hard to anticipate . we also analyze possible equilibria of the game and in particular , we find conditions for the existence of equilibria with co - existent conventions . motivated by social advertising , we also consider cases where the number of pivotal players is a negligible fraction of the population . in this case , contagion is still possible if the set of initial adopters is sufficiently large . we compute the final size of the contagion as a function of the fraction of the initial adopters . we find that the low and high - connectivity cases still have different features : in the first case , the global connectivity helps the spread of the contagion while in the second case , high connectivity inhibits the global contagion but once it occurs , it facilitates its spread . finally , we analyze a general percolated threshold model for the diffusion allowing to give different weights to the ( anonymous ) neighbors . our general analysis gives explicit formulas for the spread of the diffusion in terms of the initial condition , the degree sequence of the random graph , and the distribution of the thresholds . it is hoped that this analysis will stimulate applications of our results to other practical cases . one such possibility is the study of financial networks as started in @xcite and @xcite . amini , h. ( 2010 ) . bootstrap percolation and diffusion in random graphs with given vertex degrees . _ 17_(1 ) , research paper 25 , 20 . athreya , k. b. and p. e. ney ( 1972 ) . . new york : springer - verlag . die grundlehren der mathematischen wissenschaften , band 196 . bailey , n. t. j. ( 1975 ) . ( second ed . ) . hafner press [ macmillan publishing co. , inc . ] new york . balogh , j. and b. g. pittel ( 2007 ) . bootstrap percolation on the random regular graph . _ 30_(1 - 2 ) , 257286 . battiston , s. , d. gatti , m. gallegati , b. greenwald , and j. stiglitz ( 2009 ) . blume , l. , d. easley , j. kleinberg , r. kleinberg , and e. tardos ( 2011 ) . in _ focs 2011_. blume , l. e. ( 1993 ) . the statistical mechanics of strategic interaction . _ 5_(3 ) , 387424 . blume , l. e. ( 1995 ) . the statistical mechanics of best - response strategy revision . _ 11_(2 ) , 111145 . evolutionary game theory in biology and economics . bollobs , b. ( 2001 ) . ( second ed . ) , volume 73 of _ cambridge studies in advanced mathematics_. cambridge : cambridge university press . carlson , j. and j. doyle ( 2002 ) . . 99_(suppl 1 ) , 2538 . coupechoux , e. and m. lelarge ( 2011 ) . . netgcoop 2011_. durrett , r. ( 2007 ) . . cambridge : cambridge university press . easley , d. and j. kleinberg ( 2010 ) . . cambridge university press . ellison , g. ( 1993 ) . learning , local interaction , and coordination . _ 61_(5 ) , 10471071 . ellison , g. and d. fudenberg ( 1995 ) . word - of - mouth communication and social learning . _ 110_(1 ) , 93125 . gai , p. and s. kapadia ( 2010 ) . . _ 466_(2120 ) , 2401 . galeotti , a. and s. goyal ( 2009 ) . influencing the influencers : a theory of strategic diffusion . _ 40_(3 ) , 509532 . galeotti , a. , s. goyal , m. o. jackson , and f. vega - redondo ( 2010 ) . network games . _ 77_(1 ) , 218244 . harsanyi , j. c. and r. selten ( 1988 ) . , volume 1 of _ mit press books_. the mit press . jackson , m. o. ( 2008 ) . . princeton , nj , usa : princeton university press . jackson , m. o. and l. yariv ( 2007 ) . diffusion of behavior and equilibrium properties in network games . _ 97_(2 ) . janson , s. ( 2009a ) . on percolation in random graphs with given vertex degrees . _ 14 _ , 86118 . janson , s. ( 2009b ) . the probability that a random multigraph is simple . _ 18_(1 - 2 ) , 205225 . janson , s. and m. j. luczak ( 2007 ) . a simple solution to the @xmath170-core problem . _ 30_(1 - 2 ) , 5062 . janson , s. and m. j. luczak ( 2009 ) . a new approach to the giant component problem . _ 34_(2 ) , 197216 . janson , s. , t. uczak , and a. rucinski ( 2000 ) . . wiley - interscience series in discrete mathematics and optimization . wiley - interscience , new york . kallenberg , o. ( 2002 ) . ( second ed . ) . probability and its applications ( new york ) . new york : springer - verlag . kandori , m. , g. j. mailath , and r. rob ( 1993 ) . learning , mutation , and long run equilibria in games . _ 61_(1 ) , 2956 . kempe , d. , j. kleinberg , and e. tardos ( 2003 ) . maximizing the spread of influence through a social network . in _ proceedings of the ninth acm sigkdd international conference on knowledge discovery and data mining _ , kdd 03 , new york , ny , usa , pp . lelarge , m. ( 2008 ) . diffusion of innovations on random networks : understanding the chasm . in c. h. papadimitriou and s. zhang ( eds . ) , _ wine _ , volume 5385 of _ lecture notes in computer science _ , pp . springer . lelarge , m. ( 2009 ) . efficient control of epidemics over random networks . in j. r. douceur , a. g. greenberg , t. bonald , and j. nieh ( eds . ) , _ sigmetrics / performance _ , pp . lelarge , m. and j. bolot ( 2008 ) . network externalities and the deployment of security features and protocols in the internet . in z. liu , v. misra , and p. j. shenoy ( eds . ) , _ sigmetrics _ , pp . lpez - pintado , d. ( 2008 ) . diffusion in complex social networks . _ 62_(2 ) , 573590 . molloy , m. and b. reed ( 1995 ) . a critical point for random graphs with a given degree sequence . _ 6_(2 - 3 ) , 161179 . montanari , a. and a. saberi ( 2010 ) . . 107_(47 ) , 2019620201 . morris , s. ( 2000 ) . contagion . _ 67_(1 ) , 5778 . newman , m. e. j. ( 2003 ) . the structure and function of complex networks . _ 45_(2 ) , 167256 ( electronic ) . richardson , m. and p. domingos ( 2002 ) . mining knowledge - sharing sites for viral marketing . in _ kdd 02 : proceedings of the eighth acm sigkdd international conference on knowledge discovery and data mining _ , new york , ny , usa , pp . 6170 . acm . vega - redondo , f. ( 2007 ) . , volume 44 of _ econometric society monographs_. cambridge : cambridge university press . watts , d. j. ( 2002 ) . a simple model of global cascades on random networks . _ 99_(9 ) , 57665771 ( electronic ) . young , h. p. ( 1993 ) . the evolution of conventions . _ 61_(1 ) , 5784 . young , h. p. ( 2000 ) . the diffusion of innovations in social networks . economics working paper archive 437 , the johns hopkins university , department of economics . young , h. p. ( 2009 ) . innovation diffusion in heterogeneous populations : contagion , social influence , and social learning . _ 99_(5 ) , 18991924 . this follows from the following observations for @xmath666 : @xmath667 where the @xmath668 s and @xmath258 s are independent bernoulli random variables with parameter @xmath174 and @xmath669 respectively . now we also have : @xmath670 which corresponds exactly to ( [ eq : tec ] ) . the relation is clearly true for @xmath676 . we now assume that @xmath666 and compute the following expression : @xmath677,\end{aligned}\ ] ] where the @xmath668 s and @xmath258 s are independent bernoulli random variables with parameter @xmath174 and @xmath151 respectively and @xmath678 is distributed according to @xmath679 and independent form the @xmath262 s and @xmath258 s . this expression is natural when one consider the branching process approximation of the diffusion described in @xcite section 3.3 . we have : @xmath680 note that @xmath681 for @xmath475 , so that @xmath682 where in the last equality we used the fact that for @xmath683 , we have @xmath684 . now we have @xmath688 so that @xmath689 , where we denote explicitly the dependence in @xmath174 for @xmath690 . the monotonicity of @xmath567 in @xmath174 can be easily seen form ( [ eq : monot ] ) since @xmath691 is strictly decreasing in @xmath174 . we now justify the use of the @xmath314 in ( [ eq : max ] ) . let @xmath261 be a bernoulli random variable with parameter @xmath166 . let for @xmath149 , @xmath150 be the thinning of @xmath84 obtained by taking @xmath84 points and then randomly and independently keeping each of them with probability @xmath151 . thus given @xmath152 , @xmath153 . with these notations , we have @xmath693\\ h_1(z;\boldp,\boldt,\balpha , \pi)&= & { { \mathbb{p}}}\left(\sigma(d)=0,\ : d_{1-\pi+\pi z}\geq d - k(d)\right),\end{aligned}\ ] ] so that both @xmath692 and @xmath694 are non - decreasing in @xmath195 and non - increasing in @xmath174 . note that if @xmath156 , then @xmath695 so that @xmath696 and @xmath697 . we now consider the case @xmath698 , so that there exists @xmath699 such that @xmath700 . in this case , we have @xmath701 . the statement then follows from the fact that the only possible jumps for @xmath702 are downwards . more precisely , let @xmath703 , g(z;\boldp,\boldt,\balpha , \pi)=0\}$ ] . since the function @xmath692 is non - decreasing in @xmath195 , its set of discontinuity points is denumerable say @xmath704 and @xmath692 admits a left and right limit at any point @xmath195 denoted by @xmath705 and @xmath706 respectively . if @xmath707 , then we have @xmath708 . in particular , we have @xmath709 for any @xmath710 which contradicts the fact that @xmath711 . hence the functions @xmath692 and @xmath155 are continuous at @xmath188 and the @xmath712 is attained and can be replaced by a @xmath314 . taking expectation in ( [ eq : rde ] ) , we get with @xmath713 $ ] , @xmath714 note that @xmath681 for @xmath475 , so that @xmath715 hence , we get @xmath716 and the first part of the lemma follows . because we have ( see @xcite ) : @xmath717 taking expectation in ( [ eq : x ] ) gives : @xmath718 = 1-\sum_{s;j\geq s-\ell}(1-\alpha_s)p_st_{s\ell}b_{sj}(1-x\pi)$ ] , and the second part of the lemma follows .

universality is the amazing phenomenon when the critical behavior arising in different physical systems is characterized by a common set of universal quantities . all systems belonging to a certain universality class share the same critical exponents , universal amplitude ratios and universal scaling functions . typically , universality classes are determined by spatial dimension , symmetry and the range of interaction potentials . at times , two systems belonging to one universality class appear to be entirely disparate at first sight . in this paper we discuss such a remarkable case , _ viz_. that of driven surface diffusion under detailed balance conditions and crystalline solids near their shear instability . we show that these respectively two and three dimensional systems are described by a common field - theoretic hamiltonian . we calculate the scaling behavior of this universality class using field - theoretic renormalization group ( rg ) methods . surface growth problems constitute an important class of generic nonequilibrium phenomena @xcite . particles are deposited on a surface and can diffuse around on it . in ballistic deposition processes such as molecular beam epitaxy ( mbe ) , desorption and bulk defect formation can often be neglected . after an initial transient , a steady state is established which is characterized by time - independent macroscopic properties , provided that a suitable reference frame is chosen . if one is primarily interested in universal , large - scale , long - time characteristics , the dynamical evolution can often be formulated in terms of a langevin equation which then can be analyzed using the methods of renormalized field theory . several years ago , schmittmann , pruessner and one of us ( h.k.j . ) @xcite , denoted in the following as spj , extended a model by marsili , maritan , toigo , and banavar ( mmtb ) @xcite , to describe ballistic deposition processes with oblique particle incidence . in broad terms , their approach can be described as follows : it is assumed that there is no shadowing , i.e. , that there are no overhangs so that the growing surface can be described in terms of a height variable @xmath0 ( so - called monge representation ) . furthermore , it is assumed that particle desorption and bulk defect formation can be neglected , so that the ( deterministic ) surface relaxation processes are particle - number conserving . then , the time rate of change of the surface - hight is formulated in an idealized continuum description @xcite in terms of the divergence of a current . since the absolute hight of the surface is irrelevant , these currents are constructed from the slope of the height @xmath1 and its derivatives . next , shot noise modeling the random deposition of the particles is incorporated into the growth equation by adding a non - conserving stochastic term which dominates the conserving diffusional noise . since the oblique particle beam selects a preferred ( longitudinal ) direction in the substrate plane , the resulting langevin equation is necessarily anisotropic , and the primary relaxation mechanism is driven surface diffusion . the interplay of interatomic interactions and kinetic effects , such as ehrlich - schwoebel barriers , generates an anisotropic effective surface tension which can become very small or even negative . due to the anisotropy , these mechanisms lead to a phase space with four different regimes , and the potentially scale - invariant behaviors in these regimes were studied by spj . the most interesting case turned out to have a transversal instability which leads to the generation of longitudinal ripples via a continuous phase transition in contrast to a longitudinal instability which leads to the generation of transversal ripples by a ( fluctuation induced ) first order transition . such ripple structures are commonly found in sandy deserts where they are generated under the influence of a steady wind as the driving force , see fig . [ fig : ripples ] . it is worth noting that mmtb and spj formulated their model under the assumption of an invariance with respect to tilts of the surface . however , in a real setup , the orientation of the basic surface relative to the incident particle beam is generally fixed , and hence it is useful to generalize their model to go beyond the limiting tilt - invariant case . here , in the present work , we do _ not _ assume tilt invariance . we will see as we move along that the resulting growth equation is in detailed balance when an appropriate co - moving coordinate system is used and a certain model parameter vanish . note that we use the term detailed balance not in a microscopic sense @xcite . even if the microscopic laws lack detailed balance , coarse graining and the following neglect of irrelevant terms can lead to detailed balance in the asymptotic semi - macroscopic equation of motion for a driven system @xcite . under the assumption that this is the case , we derive a quasi - static hamiltonian @xmath2 which describes the stationary distribution , and whose basic ingredients are derivatives of the surface height field @xmath3 . the elastic free energy of crystalline solids with hexagonal or tetragonal ( including cubic ) symmetry shows an instability against shear when the elastic constant @xmath4 goes to zero . the crystallographic unit cell undergoes an elastic deformation with a certain linear combination of components of the strain tensor as an order parameter . in the case that no third order invariants of the order parameter are possible ( or vanish by chance in the cubic case ) , the transition to the structurally distorted crystal is continuous , and the order parameter can be reduced to the spatial derivatives of the displacement field @xmath5 along the crystallographic symmetry axis @xcite . augmenting the stretching energy with stabilizing bending terms and then reducing it to its parts that are relevant in the rg sense , we set up a generalized elastic energy that is of the same form as the aforementioned quasi - static hamiltonian @xmath2 with the displacement field @xmath6 taking on the role of the height field @xmath7 . note , however , there is one important distinction between the two systems . whereas in the driven diffusion problem , both the longitudinal component @xmath8 and the transversal components @xmath9 lie in the plane of the substrate , i.e. , in the plane orthogonal to the height @xmath1 , the coordinate @xmath8 in the elastic problem lies in the direction of the crystal axis and is therefore parallel to the displacement @xmath5 . hence , the physical dimension of @xmath9 is @xmath10 in driven surface diffusion and @xmath11 in the elastic transition . the outline of the present paper is as follows : in sec . [ sec : model ] we set up field theoretic models for driven surface diffusion with detailed balance and elastic phase transitions , respectively , through augmenting and modifying the existing models for these systems . we unify the two field theoretic models by formulating a hamiltonian that commonly describes both systems . in sec . [ sec : rpt ] we present our renormalized perturbation theory . we sketch our diagrammatic calculation , discuss the resulting renormalization group and its flow . in sec . [ sec : scalingandexponents ] , we derive the scaling forms for several experimentally measurable quantities . for driven surface diffusion , we calculate critical exponents , and for elastic transitions we calculate logarithmic corrections to the leading mean - field behavior . in sec [ sec : concludingremarks ] we give concluding remarks . appendix [ app : calcdiagrams1loop ] contains some details on our calculation of feynman diagrams . appendix [ app : addren ] discusses some intricacies related to the fact that the calculation of the specific heat for the elastic transitions requires additive renormalizations . here , we set up our model for driven surface diffusion following the work of mmtb @xcite and spj . as a starting point , let us review the langevin equation presented by spj . in the to - interpretation , this langevin equation describes the hight @xmath0 of a @xmath12-dimensional growing hyper - surface as @xmath13 the incident particle beam has components @xmath14 and @xmath15 normal and parallel to the substrate plane , respectively . as discussed in detail by mmtb and spj , the oblique particle incidence induces a fundamental anisotropy into the system which breaks the full rotational symmetry within the @xmath12-dimensional space of the substrate . to account for this anisotropy , it is useful to choose a coordinate system in which a particular axis , say the @xmath8-axis , is aligned with @xmath15 and then to write the spatial coordinates as @xmath16 . the two uniform growth terms , the first two terms on the right hand side of eq . ( [ genlang ] ) , can be removed by a galilei transformation @xmath17 . in the following , we will work in this co - moving frame . @xmath18 is the surface current , and we model its longitudinal and transversal components following spj by @xmath19 note that under the assumptions made by mmtb and spj , @xmath20 . as we will discuss in more detail below , we will work under different assumptions here . the langevin noise @xmath21 accounts for the randomness in the particle deposition . because this shot noise adds particles to the surface , it is commonly modeled as non - conserving noise with correlations of the form @xmath22 in principle , conserved diffusional noise could also be incorporated into the model . however , it turns out to be irrelevant compared to the non - conserved noise . the character of the surface produced by the process depends on the values of the critical parameters @xmath23 and @xmath24 . our model leads to qualitatively the same phase diagram in the space spanned by @xmath23 and @xmath24 as the spj model , see fig . 1 of ref . @xcite . this phase diagram contains , in particular , a continuous transition from isotropic , edwards - wilkinson @xcite type behavior to a state characterized by longitudinal ripples for @xmath25 and @xmath26 . in the following , we will focus on this transition . the other transitions in the phase diagram which includes a fluctuation induced discontinuous transition to a surface with transversal ripples for @xmath27 and @xmath28 and a multi critical point for @xmath27 and @xmath29 are set aside for future work . now , we review the symmetries and the physical contents of the langevin equation ( [ genlang ] ) with its ingredients ( [ parcurr ] ) to ( [ shot - noise ] ) in some more depths . the work by mmtp and spj assumed that the physics of the system is invariant under an infinitesimal tilt of the surface by an angle @xmath30 , @xmath31 . the assumption of this invariance enforces the relation @xmath20 mentioned above @xcite . however , this tilt - invariance is non - physical since the orientation of the substrate plane is fixed by the normal component @xmath14 of the incident current . in the following , we will refrain from imposing this condition . stability demands positive @xmath32 and @xmath33 as well as @xmath34 . at the transition for @xmath25 and @xmath26 that we are interested in , @xmath32 and @xmath35 are irrelevant . the positive parameters @xmath33 and @xmath23 can be set equal to @xmath10 and a dimensionless positive parameter @xmath36 , respectively , via a simple rescaling of the coordinates @xmath16 and @xmath37 . note that after this transformation , the coordinates , the time and the `` temperature '' scale as @xmath38 , @xmath39 , @xmath40 and @xmath41 , respectively , where @xmath42 is an inverse external length scale appropriate to discuss the infrared limit that will be used in the renormalization program to follow . it thus follows from the langevin equations , eqs . ( [ genlang ] ) and ( [ shot - noise ] ) , that the shot noise @xmath21 and the height @xmath1 scale as @xmath43 and @xmath44 . furthermore , the coupling constants scale as @xmath45 and @xmath46 which shows that @xmath47 is irrelevant and that the upper critical dimension is @xmath48 . returning to the tilt - invariance and the fact that we do not assume it in the present work , we note that there is a possible contribution to @xmath49 of the form @xmath50 that was omitted by mmtb and spj on grounds of tilt - invariance , and we have to consider it here . power counting reveals that @xmath51 , i.e. , this term is relevant . even if we omitted it at the onset , it would be generated by the rg . here , we include it in our model at this stage . we are interested here in driven surface diffusion with detailed balance so that the statistical properties of the steady state can be derived from a free energyor hamiltonian @xmath52 . for this to hold true , we have to be able to express the components of the surface current through functional derivatives of @xmath52 . because the absolute height of the surface should not matter , i.e. , @xmath53 with a constant @xmath54 should be a symmetry , the functional variation of @xmath52 with respect to @xmath1 should be of the form @xmath55 which defines the surface current @xmath18 up to divergence - free additive part . it is easy to see that this is possible if we equalize the coupling constants @xmath56 , and in the following we will do so . then the hamiltonian is given by@xmath57 for further details and background on detailed balance in the context of driven surface diffusion , we refer to ref . @xcite . there is another symmetry present that we have not mentioned yet , viz . the so - called up - down symmetry . this symmetry , which physically means that mounds and valleys have the same form , translates for mbe with oblique incidence to a form - invariance under the simultaneous inversion @xmath58 and @xmath59 . this type of up - down symmetry was assumed by mmtb and spj , and we also assume it here . however , even though it is frequently assumed in mbe theory , up - down symmetry has to be taken with a grain of salt because the particle beam sets a preferred direction in space . therefore , the up - down symmetry is strictly speaking more appropriate for interface than for surface models . we keep this symmetry here firstly because we are primarily interested in the connection between and the mutual properties of driven surface diffusion and elastic phase transitions and secondly because the up - down symmetry simplifies the model . in a sense , this simplified model can be viewed as a leading order approximation for mbe with oblique incidence that can be improved later by including terms that break the up - down symmetry . in fact , we plan to study the effects of breaking the up - down symmetry in the future . note that this symmetry has been discussed in the literature @xcite as a prerequisite for having a free energyin this kind of system . however , we can add a further coupling proportional to @xmath60 to the hamiltonian ( [ hamsurf ] ) which explicitly destroys up - down symmetry . only in the one - dimensional case this coupling is a total gradient and therefore vanishes after integration over the full one - dimensional space . the langevin equation ( [ genlang ] ) now takes the form @xmath61 in the comoving - frame introduced above . to generate correlation and response functions of this stochastic process , it lends itself to reformulate the langevin equation as a dynamical response functional @xmath62 so that averages can be derived from path integrals with the statistical weight @xmath63 . via introducing a response field @xmath64 , the functional @xmath62 ( in to - discretization ) that we extract with standard methods @xcite from eqs . ( [ speclang ] ) and ( [ shot - noise ] ) reads @xmath65=\int d^{d}x\int dt\ \lambda \big\{\tilde{h}\big(\lambda^{-1}\partial_{t}h+\frac{\delta\mathcal{h}_{\text{surf}}[h]}{\delta h}\big)-\tilde{h}^{2}\big\}\ , , \label{def_j}\ ] ] and shows that our driven surface diffusion model obeys detailed balance @xcite . the fluctuation - dissipation theorem ( fdt ) holds , and all stationary expectation values can be calculated from path - integrals with the statistical weight @xmath66 . it is a well established fact that in all dynamical field theories that are invariant under a field shift of the type @xmath53 with a constant @xmath54 , the dynamical exponent is given by @xmath67 , where @xmath68 is the anomalous scaling exponent of the order parameter field @xmath1 . as it should , the same applies here . this can be verified in detail by noting that in the full dynamical field theory based on the response functional ( [ def_j ] ) one does not need to introduce renormalization factors for the @xmath69 and @xmath70 terms when using minimal renormalization , i.e. , dimensional regularization in conjunction with minimal subtraction . hence , we can calculate all of the critical exponents of driven surface diffusion with detailed balance from @xmath52 including the dynamical exponent . an elastic phase transitions is a structural phase transition in which the crystallographic unit cell undergoes an elastic deformation when a normal mode becomes soft in response to the vanishing of an elastic constant or combination of elastic constants . in this section , we essentially follow the early work of cowley @xcite , and folk , iro , and schwabl ( fis ) @xcite . for background information on elastic phase transitions , we refer to these original papers as well as reviews by bruce and cowley @xcite , rao and rao @xcite , and schwabl and tuber @xcite . in general , the order parameter for such a transition is a certain component of the lagrange strain tensor or a linear combination of its components . depending on the symmetry of a crystal , its lagrange elastic energy density can be unpleasantly complicated even at harmonic order due to the presence of a host of different elastic constants . here , we restrict ourselves to crystals with tetragonal or higher symmetry , i.e. , crystals whose harmonic lagrange energy density can be obtained from the tetragonal case by imposing relations between certain elastic constants . as we will see below , crystals with lower symmetry such as orthorhombic or monoclinic symmetry have simple mean field behavior at their respective elastic transition and are hence less interesting from the standpoint of critical phenomena than the higher - symmetry crystals we focus on here . when generalized to @xmath12 spatial dimensions , the lagrange energy density of tetragonal solids to second order in the strains reads @xmath71 where @xmath72 , and @xmath73 for a physical solid . the lagrange strain tensor ( see , e.g. , refs . @xcite ) is defined via the derivatives of the displacement field @xmath74 with respect to the reference space coordinates of the undeformed solid , @xmath75 with @xmath76 . if we specialize the elastic constants @xmath77 to @xmath78 , @xmath79 , @xmath80 or @xmath81 , eq . ( [ fren ] ) reduces to the harmonic energy density of a cubic or hexagonal crystal , respectively . if all these conditions hold , eq . ( [ fren ] ) describes an isotropic solid . moreover , the hexagonal case with the additional relation @xmath82 reduces to the isotropic case after a suitable rescaling of the coordinates . because of crystal anisotropy , sound velocities depend on the direction of propagation , and modes become soft only in certain selected directions , so - called soft sectors , which may be one- or two - dimensional in 3@xmath12 crystals . the type of the soft modes depends on which elastic constant or combination of elastic constants vanishes . in the following , we focus on the shear instability that occurs when @xmath4 goes to zero . the accompanying elastic phase transition is continuous if there are no third order couplings of the corresponding order parameter . we assume that this is the case . the upper critical dimension of the phase transition is @xmath83 , where @xmath84 is the dimension of the soft mode sector in fourier space , as was shown by fis . this implies that , from the standpoint of critical phenomena , @xmath85 is the most interesting case since otherwise the upper critical dimension is lower then the physical dimension @xmath86 , and the transition is simply be described by mean - field - theory . in the remainder , we focus on @xmath85 . in this case , the @xmath87-dimensional soft sector is orthogonal to the polarization of the corresponding critical sound waves , i.e. , to the direction of the symmetry - axis in fourier space . following fis , we radically reduce the displacement field to the critical displacement alone@xmath88 this step is justified if there are no third order couplings in the critical strain . such couplings can be ruled out on symmetry grounds for the crystals we focus on except for the cubic case . for cubic crystals , we assume that third order couplings are absent for some given reason , e.g. , chance . otherwise , the elastic transition becomes discontinuous . we will comment on this issue and possible couplings to the non - critical displacements @xmath89 further below . to figure out which parts of the elastic energy are relevant in the sense of the rg , it is useful to rank the different terms according to the order at which they contribute in the long - wavelength limit . to this end , let us introduce the smallness parameter @xmath90 , and set the derivatives to@xmath91 with the reduction ( [ reduz ] ) , the strain tensor becomes@xmath92 and we obtain the free energy density ( [ fren])@xmath93 where we have set @xmath94 and @xmath95 . now we see that it is appropriate to rescale the vanishing elastic constant @xmath4 by@xmath96 and reduce @xmath97 to the leading terms proportional to @xmath98:@xmath99 of course , the anisotropic last term in @xmath100 with hypercubic symmetry vanishes in the hexagonal case . at this stage , a comment regarding higher than harmonic terms in the initial lagrange energy density ( [ fren ] ) is warranted . as long as we ignore the uncritical displacement @xmath101 , the addition to the right hand side of eq . ( [ fren ] ) of third and fourth order terms in the strains leads to exactly the same form of @xmath102 as given in eq . ( [ redelaen ] ) . the only effect of the higher order strain terms is to modify the constants decorating the individual terms in eq . ( [ redelaen ] ) , and hence their omission is justified . however , if we included the noncritical displacements which scale as @xmath103 , then already the harmonic strain terms in the lagrange elastic energy density would lead to couplings of the form @xmath104 , @xmath105 and @xmath106 which contribute for @xmath107 at the same order in @xmath90 as the terms retained in eq . ( [ redelaen ] ) . thus far , we only included stretching in our discussion of elastic energy densities . however , since we are interested in the limit @xmath108 where there is an elastic instability , bending terms will be important as they stabilize the crystal . the bending terms that are relevant here are @xmath109 with @xmath110 in the hexagonal case . adding stretching and bending contributions , we obtain the model elastic energy or hamiltonian @xmath111 which is , up to the anisotropic cubic terms which are absent in the hexagonal case , of the same form as the hamiltonian ( [ hamsurf ] ) for driven surface diffusion . this is the hamiltonian that we will study , after suitable rescalings and renamings , in our renormalized perturbation theory . form this point on , we will disregard the hyper - cubic terms , i.e. , the last term in the stretching energy density ( [ redelaen ] ) and the term proportional to @xmath112 in the bending energy density ( [ redbenden ] ) . as mentioned above , crystals with cubic symmetry are expected to have anomalous scaling behavior and are hence certainly interesting , however , the presence of the hyper - cubic terms makes the diagrammatic perturbation calculation significantly more difficult . though we have done this calculation , we feel that a useful discussion of the cubic case goes beyond the scope of the present paper . we will cover elastic transitions in cubic crystals in a future paper @xcite . before we move on , let us make some comments on the relation of our hamiltonian @xmath113 to other , prominent model elastic energies that have been discussed in the literature . first , for @xmath114 ( and @xmath115 ) , @xmath113 reduces to one of the two models studied by fis ( their model ii ) . we will refer to this model as the fis model . secondly , in the limiting case of a fully isotropic crystal with @xmath116 , @xmath117 , @xmath118 , we have@xmath119^{2}\ , , \label{isotr}\ ] ] with@xmath120=\big\{\frac{c_{44}}{c_{11}}+\bigl(\partial_{\vert}u\bigr)+\frac{1}{2}\bigl(\nabla_{\bot}u\bigr)^{2}\big\}\ , , \label{invar}\ ] ] up to a total derivative . @xmath121 $ ] is invariant under a type of a galilean transformation , @xmath122 with infinitesimal transversal @xmath123 . when augmented with bending the bending term proportional to @xmath124 , this isotropic elastic energy density leads to a hamiltonian @xmath125 that is equivalent to that considered by grinstein und pelcovits ( gp ) in their seminal work on the anomalous elasticity of smectic liquid crystals @xcite . we will refer to this model as the gp model . jointly , we will refer to the fis and gp models as the special models . to further emphasize that both driven surface diffusion with detailed balance and elastic transition are described by the same field theoretic hamiltonian , we rescale the coordinates and @xmath5 in eq . ( [ hel ] ) . after renaming the order parameter field and the coupling constants in the rescaled version of eq . ( [ hel ] ) as well as renaming the order parameter field @xmath1 in eq . ( [ hamsurf ] ) , both our model systems are described by the hamiltonian @xmath126 where @xmath127 is the common order parameter field . next , let us discuss the symmetry transformations of this hamiltonian and the combinations of its parameters that are invariant under the respective transformation . knowing these combination will come in handy later on when we analyze the rg flow . this hamiltonian is invariant under the longitudinal scale transformation @xmath128 the following combinations of coupling constants are invariant under this transformation : @xmath129 a further symmetry of the hamiltonian @xmath2 is the @xmath8-dependent field shift @xmath130 the galilean transformation with transversal @xmath123 @xmath131 is a symmetry if only and only if the conditions @xmath132 hold . of course , if @xmath133 , the hamiltonian is invariant under the inversion @xmath134 . in this section , we outline our renormalized perturbation theory . for background on the methods that we use , we refer to the textbooks @xcite . the feyman diagrams of our perturbation theory consist of the gaussian propagator @xmath135 and two vertices , see fig . [ fig : statelem ] . the @xmath136-term in the denominator of the propagator will be omitted in the following because an expansion of any diagram in powers of this term will inevitably lead to unnecessary contributions that are convergent in the ultraviolet ( uv ) . there is a 3-leg vertex @xmath137 \label{3-vert}\ ] ] and a 4-leg vertex @xmath138\ , . \label{4-vertex}\ ] ] we use dimensional regularization about @xmath139 dimensions in conjunction with minimal subtraction , i.e. , we use minimal renormalization . our renormalization scheme reads @xmath140 where @xmath141 we choose the additive renormalization @xmath142 so that @xmath143 . the longitudinal tilt - transformation ( [ longtilt ] ) leads to the exact relation@xmath144 note that formally @xmath145 in minimal renormalization . in general , i.e. , beyond minimal renormalization , @xmath146 is a positive or negative constant . in a momentum shell rg , for example , it depends on the momentum cut - off and on the coupling constants . we calculate the 1-loop contributions to the superficially uv - divergent vertex functions @xmath147 , @xmath148 and @xmath149 . some details of this calculation can be found in appendix [ app : calcdiagrams1loop ] . for the counter - terms that render their respective vertex function finite at 1-loop order , we obtain @xmath150 form these counter - terms , we extract the 1-loop renomalization factors @xmath151 note that if the system is invariant under the galilean transformation ( [ galtraf ] ) which implies @xmath152 , the conditions ( [ galsym ] ) lead to the following two _ exact _ relations between the renormalization factors : @xmath153 now , we set up our rg equations and discuss the rg flow . to simplify the involved equations , we use in the following the shorthands @xmath154 for the renormalized versions of the invariant combination of coupling constants given in eq . ( [ invcoup ] ) . in terms of @xmath155 and @xmath156 , the rg @xmath157-functions are defined as @xmath158 where @xmath159 and @xmath160 denote the @xmath161-residues of the respective renormalization factors . to 1-loop order , we obtain @xmath162 inspecting the relations ( [ galbez ] ) , in particular for @xmath163 , we figure that it is useful to define @xmath164 this quantity has a perturbation expansion in @xmath155 and @xmath156 which is to lowest order given by @xmath165 from the rg @xmath157-functions , we get the wilson and gell - mann low functions @xmath166 note that for @xmath163 , which corresponds to the gp model where the ward identities ( [ galbez ] ) do hold , we have @xmath167 to arbitrary order in perturbation theory . our 1-loop results are consistent with these relations and hence satisfy an important check . having these rg functions , we can write down the rg differential operator . applying @xmath168 to cumulants ( green functions ) amounts to to rg equation ( rge ) @xmath169 with the rg differential operator @xmath170 now , we are in position to discuss the fixed points of the rg . after setting @xmath171 , the flow of @xmath155 and @xmath156 is given by the ordinary differential equations @xmath172 with @xmath173 and @xmath174 . these fixed points , which are approached asymptotically for @xmath107 , follow from the zeros of the gell - mann low functions , @xmath175 with these functions given to @xmath10-loop order by @xmath176 in addition to the unstable trivial fixed point @xmath178 , there are three more fixed points for @xmath177 . ( 1 ) the galilean - invariant fixed point @xmath179 that corresponds to the gp model , @xmath163 . we will refer to this fixed point as the gp fixed point . ( 2 ) the inversion - invariant fixed point @xmath180 that corresponds to the fis model . we will refer to this fixed point as the fis fixed point . ( 3 ) the new fixed point @xmath181 we will see below that this is the only stable fixed point . next , we discuss the stability of the fixed points which is determined by the stability matrix whose components are given by the partial derivatives with respect to @xmath155 and @xmath156 , respectively , @xmath182 its eigenvalues ( the wegner correction exponents ) and the corresponding right eigenvectors which indicate the flow near the respective fixed points are as follows : ( 1 ) for the gp fixed point , we have @xmath183 with @xmath184 and @xmath185 with @xmath186 . thus , this fixed point is stable only in one direction . ( 2 ) for the fis fixed point , we have @xmath187 with @xmath188 and @xmath189 with @xmath190 . hence , this fixed point is also stable in only one direction . ( 3 ) for the new fixed point , we have @xmath191 with @xmath192 and @xmath193 with @xmath194 . therefore , the new fixed point is stable in any direction . the picture of the rg flow that emerges from the three invariant lines @xmath195 , @xmath163 , and @xmath196 , as well as from the fixed points and their respective stability is sketched in fig . ( [ fig : flussbild ] ) and can be described as follows : the domain of attraction of the stable fixed point is given by the positive quadrant @xmath197 , @xmath198 , that is bounded by the lines @xmath163 ( gp model ) and @xmath195 ( fis model ) . the unstable fixed points that correspond to the gp and fis models , respectively , are located on these lines . for negative @xmath156 , the rg flow runs off to infinity . this suggests that we are dealing with a fluctuation - induced first order transition in the case @xmath199 . of course , the trivial fixed point @xmath178 is stable for @xmath200 . thus , for dimensions @xmath201 , we find mean - field behavior . right at the critical dimension , the mean - field power laws have logarithmic corrections . this case is of particular importance here because for the elastic phase transitions the upper critical dimension and the physical dimension are the same . in the next section , we will discuss these logarithmic corrections in detail . as a prelude , we study here the asymptotic flow of @xmath155 and @xmath156 to zero for @xmath202 . for @xmath202 , the flow equations ( [ flow ] ) in conjunction with gell - mann low functions ( [ beta1l ] ) lead to the differential equations@xmath203 solving these equations for @xmath204 , we obtain@xmath205 the amplitudes are given in general by @xmath206 the special models lead to@xmath207 in this section , we first study the scaling behavior of the cumulants of our theory . then , we extract from this general result the critical behavior of several experimentally relevant quantities for driven surface diffusion and elastic transitions , respectively . now , we determine scaling properties of the cumulants @xmath208 . besides the order - parameter - field @xmath209 and its spatial derivatives , we consider the energy - field @xmath210 . this field is a composite field in the language of renormalization theory with its own scaling behavior . this scaling behavior can be studied via so - called insertions that are generated by variations @xmath211 of the cumulants with respect to the local temperature@xmath212 . in other words , the objects of our study here are the expectation values @xmath213^{n}[e]^{m}\rangle ^{(cum)}=\big[-\frac{\delta}{\delta\tau}\big]^{m}g_{n,0 } \ , , \label{kum}\ ] ] which we have written here in a somewhat compressed yet evident form . to extract the scaling properties of the green functions , we first note that simple dimensional analysis gives @xmath214 for the leading , mean - field behavior . next , we solve the rge for the green functions to get the anomalous contributions to scaling . for @xmath215 , this rge reads @xmath216 as usual , this partial differential equation can be solved by the method of characteristics . the characteristics read @xmath217 , @xmath218 and @xmath219 , where the @xmath220-factors are the solutions of the ordinary differential equations@xmath221 collecting the solutions of these equations , the flowing couplings @xmath222 that solve eq . ( [ flow ] ) and the mean - field scaling as given in eq . ( [ kumscal ] ) , we obtain the scalings forms @xmath223 this result is correct as long as @xmath215 . in the case @xmath224 , @xmath225 or @xmath11 one has to account for additive renormalizations which become essential for the specific heat for @xmath226 . the additive renormalizations and its consequences are the theme of appendix [ app : addren ] . we apply the scalings ( [ kumscal ] ) to the result ( [ g - fin ] ) and obtain@xmath227^{m-1}\ , , \label{enkorr}\ ] ] with @xmath228 given by eq . ( [ i - l ] ) . in experiments on surface growth , one often measures the roughness of a surface and the associated roughness exponents . these exponents are defined through the height - height correlation function@xmath229 for surface diffusion , the physical dimension is @xmath230 . hence , we are interested in the scaling behavior of the height - height correlation and its roughening properties for @xmath177 . the asymptotic solutions of the rg - equations ( [ ampldiff ] ) are@xmath231 where @xmath232 , @xmath233 , and @xmath234 are the values of the rg - functions at the fixed point @xmath235 , @xmath236 . for the stable fixed point as well as for the gp fixed point , we have @xmath237 . hence we get from the definitions ( [ rgfu ] ) the exact relation@xmath238 the scaling behavior of the height - height correlation function ( [ hh - corr ] ) can be deduced from the general result ( [ genscal ] ) : @xmath239 where we have defined the anisotropy and the correlation length exponents@xmath240 which are related ( with exception of the fis - model ) by@xmath241 due to the presence of strong anisotropy , we have to dealing with two different roughness exponents @xmath242 and @xmath243 defined by @xmath244 the two exponents @xmath245 are read off immediately . in the dimensional - expansion about three surface - dimensions , we get@xmath246 to first order in @xmath161 . as mentioned above , the physical and the upper critical dimensions coincide for elastic transitions . therefore we have @xmath226 , and we are interested in logarithmic corrections . the typical measurable quantities are the susceptibility @xmath247 of the critical modes @xmath248 , and the critical part of the specific heat @xmath249 that , as the second line eq . ( [ defspez ] ) indicates , can be calculated from the energy correlation function . here , we need the asymptotic solutions of the rg - equations ( [ ampldiff ] ) for the amplitudes using the assymtotic flow ( [ asympt ] ) . we get@xmath250 with the constants @xmath251 and @xmath252 given in eqs . ( [ ampl ] ) and ( [ specampl ] ) . from the general scaling properties of the greens - functions ( [ genscal ] ) and bysetting@xmath253 ( we have set the scale so that @xmath254 to simplify the conclusions ) , we obtain @xmath255 asymptotically . the correction exponent @xmath256 reads @xmath257 whereas the special models lead to@xmath258 next , using the scaling result for the energy correlations ( [ enkorr ] ) we get @xmath259 with the correction exponent @xmath260 . we recognize that the term originating from the additive renormalization overwhelms the first one asymptotically . the correction exponent reads @xmath261 the special models lead to@xmath262 note that our results for the correction exponents for the susceptibility and the specific heat of the fis model are in full agreement with those obtained by fis originally . hence , our calculation passes an another consistency check . in summary , we have studied driven surface diffusion at the phase transition from isotropic , edwards - wilkinson type behavior to a state characterized by longitudinal ripples and continuous elastic phase transitions . we have shown that the two transitions belong under appropriate conditions to the same universality class . we derived the field theoretic hamiltonian for this universality class which has an upper critical dimension of @xmath263 . the dimension @xmath12 plays different roles in the two systems . in driven surface diffusion , @xmath12 is the dimension of the substrate whereas for the elastic phase transition , @xmath12 is the full dimension of space . hence , physically , @xmath230 in the former problem and @xmath73 in the latter . thus , the critical behavior in the surface diffusion problem is described by power laws with anomalous scaling exponents whereas that of the elastic phase transitions under consideration is described by power laws with mean - field exponents decorated by logarithmic corrections . we calculated these anomalous exponents and logarithmic corrections , respectively , for several experimentally relevant quantities . as far as elastic phase transitions were concerned , we restricted ourselves to crystals with tetragonal or higher symmetry , i.e. , crystals whose harmonic lagrange energy density can be obtained from the tetragonal case by imposing relations between certain elastic constants because crystals with lower symmetry such as orthorhombic or monoclinic symmetry have simple mean field behavior . in addition to that , we have left aside crystals with cubic symmetry in the transversal subspace . these do have anomalous scaling behavior , however , the diagrammatic perturbation calculation gets significantly more involved when cubic terms are present and is beyond the scope or the present paper . we will present our results for the cubic case in a future publication @xcite . our model hamiltonian differs from the spj model in that we do not assume tilt invariance and that we do assume detailed balance , at least on a semi - macroscopic scale . our result for the anisotropy exponent @xmath264 is smaller and those for the roughness exponents are larger than the corresponding 1-loop results produced by the spj model . hence , in comparison , our model leads to a less anisotropic but slightly rougher surface . its restriction to detailed balance makes our model a special case of a more general model . our results , moreover , show that our model corresponds to a rg fixed point of this more general model . whether or not this fixed point is the stable fixed point of the more general model , we can not judge at this point , and we will address this question in future work . this question is also important from an experimental standpoint . if our fixed point turns out being stable , this would mean that the asymptotic behavior in a typical experiment on driven surface diffusion with oblique incidence would show detailed balance as characterized by our fixed point . our model hamiltonian differs from the fis model in that it includes a relevant third order coupling that does not destroy the continuous character of the transition . the presence of this term modifies the exponents of the logarithmic corrections to the susceptibility and the specific heat . for the former , the difference is quantitative , for the latter it is even qualitative . the fis model predicts that the specific heat diverges as the transition is approached , whereas we find that it vanishes . as the fis model , our model entirely neglects the noncritical transversal components of the elastic displacement field . if the order parameter field corresponded directly to a component of the strain tensor , one could simply integrate out the noncritical strain components . this would produce additional negative contributions to the positive ( on grounds of stability ) fourth order coupling constant of our model . if the additional negative contributions exceed the positive contribution in magnitude , the transition becomes first order . however , here we have to integrate out the transversal displacement components and not entire components of the strain tensor . the transversal components of the displacement are contained together with its critical components in different components of the strain tensor , and it is not clear how to integrate them out . we will return to the question of the influence of the noncritical displacements in future work . from a conceptual standpoint , it is worthwhile to reflect on the different roles played by reference and target space symmetries in elastic phase transitions and related problems @xcite . invariance under rigid rotations in target space , unless broken by external fields , is a symmetry that applies likewise to any crystal , or for that matter to any elastic system that can be described by lagrange elasticity theory , including smectic liquid crystals , liquid crystal elastomers etc . reference space symmetry , on the other hand , is determined by the crystallographic point group of a given elastic system . it has been an established approach for setting up field theoretic models for elastic systems for about three decades to employ gradient expansion and then to retain only those terms that are deemed relevant based on dimensional analysis . examples where this approach has been used include , to name a few , the original work on nematic liquid crystals by gp , isotropic @xcite and anisotropic @xcite tethered membranes , sliding columnar phases @xcite , nematic elastomers @xcite and their membranes @xcite and so on , and we also used this approach here . for systems with comparatively high reference space symmetry , vestiges of the target space rotational invariance remain in the final field theoretic hamiltonian . the nonlinear part of the strain @xmath265 in the gp model , for example , ensures that the target space symmetry is realized at least linearly . the situation is similar in the other examples just mentioned . for the elastic transition studied in the present paper , with exception of the limiting isotropic case that coincides with the gp model , the reference space symmetry is comparatively low , and there are no remains of the target space rotational invariance in the final hamiltonian . it transpires that the extend to which the latter symmetry survives when setting up field theories depends on the former and that , in a sense , reference space symmetry is superior to target space symmetry . it is desirable to get a deeper understanding of the interplay of the two symmetries in field theories , and we see here an interesting opportunity for future work . this work was supported in part ( os ) by the national science foundation ( dmr1104707 and mrsec dmr-1120901 ) . os thanks t.c . lubensky for spurring his interest in elastic phase transitions and useful discussions . here , we present a few cornerstones of our calculation of feynman diagrams . the 1-loop diagrams of our theory are depicted in figs . [ fig:2-leg ] to [ fig:4-leg ] . a typical integral encountered in calculating these diagrams reads @xmath266 where @xmath267 . in all of our 1-loop integrals , we can set the external longitudinal momentum to zero . via setting @xmath268 , the integration over the internal longitudinal momentum can then be performed using the identity @xmath269^{2}(2c)^{2n-1}}\ , . \label{hilfe}\ ] ] we obtain @xmath270 the remaining integrations are simplified by switching integration variables to @xmath271 and @xmath272 and defining @xmath273 : @xmath274 using dimensional regularization of the integral with @xmath139 , we finally obtain @xmath275 where @xmath276 . all other feyman integrals at 1-loop order can be calculated by similar means . we get:@xmath277@xmath278@xmath279 with @xmath280.@xmath281@xmath282 the @xmath11-leg diagrams are shown in fig . ( [ fig:2-leg ] ) . with the results above we obtain their singular parts that constitute the counter - terms@xmath283@xmath284 and@xmath285 our calculation of the @xmath86-leg diagrams ( fig . ( [ fig:3-leg ] ) ) yields @xmath286 and@xmath287 after symmetrization of the external leg moments . from the @xmath288-leg diagrams , fig . ( [ fig:4-leg ] ) we obtain@xmath289@xmath290 and@xmath291 again after symmetrization . as pointed out in sec . [ sec : scalingandexponents ] , the green functions @xmath292 and @xmath293 can not be renormalized entirely by multiplicative renormalization alone . rather , they require extra additive renormalization which leads to the inhomogeneous rges ( @xmath294 ) @xmath295^{m-1}\ , , \label{rge02}\ ] ] where @xmath296 is a function of the invariant coupling constants @xmath155 and @xmath156 only . to lowest order , the additive renormalization arises from the uv - divergence of the diagram ( 4a ) in fig . ( [ fig:4-leg ] ) with removed external legs and coupling constants . this divergence is cancelled by the additive counter - term@xmath297 which leads after applying the operator @xmath298 to @xmath299 at lowest order . to proceed towards its solution , let s discus the form of the rge ( [ rge02 ] ) in some more detail . this rge is a partial differential equation that describes the dependence of @xmath300 upon the variables @xmath301 . as far as the rge is concerned , the spatial coordinate plays solely the role of a constant parameter . the general form of the rge is a quasi - linear partial differential equation ( pde ) , see _ e.g. _ @xcite@xmath302 to find its solution , we consider @xmath303 as the solution of the implicit equation @xmath304 and write the pde as@xmath305 this homogeneous pde can be solved by employing the characteristics @xmath306 , @xmath307 determined as usual by the ordinary differential equations@xmath308 with @xmath309 , and initial conditions @xmath310 , @xmath311 . since @xmath312 remaines constant along the characteristic flow , the solution is@xmath313 and @xmath307 is of the same functional form as that of @xmath314 as a function of @xmath315 , @xmath316 in our case , the characteristics @xmath317 , @xmath218 , @xmath318 , @xmath319 , @xmath320 are defined by the differential equations ( [ floweq ] ) and ( [ ampldiff ] ) , and their asymptotic solutions in the case @xmath226 , @xmath107 are given in equations ( [ asympt ] ) and ( [ ampllog ] ) . it remains to solve the differential equation for @xmath307@xmath321 this equation has the solution @xmath322 where @xmath228 is the integral@xmath323 since @xmath324 and @xmath325 , we finally get @xmath326^{m-1 } \label{g - fin}\ ] ] from eq . ( [ u - l ] ) . janssen , z. phys . b * 23 * , 377 ( 1976 ) ; r. bausch , h.k . janssen , and h. wagner , z. phys . b * 24 * , 113 ( 1976 ) ; h.k . janssen , in _ dynamical critical phenomena and related topics _ ( _ lecture notes in physics _ , vol . 104 ) , edited by c.p . enz , ( springer , heidelberg , 1979 ) .

positronium ( ps ) , the bound state of electron and positron , was extensively investigated since its discovery @xcite contributing to the development of bound state qed ( see e.g. @xcite for a review on the current status of this field ) . furthermore , this system provided stringent limits on possible deviation from the standard model that could indicate new physics @xcite ( see also @xcite for a comprehensive review of former experiments ) . moreover , in the field of materials science ps found various applications due to its unique properties ( see e.g. @xcite for modern reviews on this subject ) . one recent example is the characterization via positron annihilation lifetime spectroscopy ( pals ) of low - k dielectrics that are potential candidates for the next generation of integrated circuits @xcite . the motivation of the work presented in this paper is to understand if mesoporous silica with an interconnected pore network could be used for producing a high fraction of positronium at low temperatures . this would open the door for a new generation of experiments in fundamental research . cold positronium could be used to improve the precision of spectroscopic studies of ps @xcite or to perform the first spectroscopy of the ps@xmath2 molecule @xcite . furthermore , it could provide an alternative to the methods that are used for anti - hydrogen formation @xcite-@xcite . as it was suggested sometime ago @xcite-@xcite , anti - hydrogen could be formed using charge - exchange of ps with anti - protons . this was demonstrated for the charge conjugate reaction @xcite and in the atrap collaboration resonant charge exchange collisions of positrons with rydberg cs atoms were used to form rydberg ps that via charge exchange with the anti - protons produced anti - hydrogen in rydberg states @xcite . recently , two experiments @xcite were proposed to perform an anti - gravity test using this process . in both experiments , one of the main issues is that a high fraction of ps at low temperatures should be available . another interesting application would be the possibility to perform an experiment in order to confirm the interpretation of the recent dama / libra annual modulation signal @xcite as generated by mirror type dark matter @xcite . a step further would be to achieve bose - einstein condensation of ps @xcite . this would allow the exploration for the first time of the effects of the collective properties of a matter - antimatter system . in mesoporous silica , ps is produced by injecting positrons into the film and the distribution of the implantation depth follows a makhovian profile @xcite . in the following we solely consider the long lived triplet spin state ( called ortho - positronium with 142 ns lifetime , o - ps ) because the singlet spin state ( called para - positronium , p - ps ) has a very short lifetime of 125 ps and can be considered as annihilating in the target . the o - ps ( for simplicity we will refer to it as ps ) that diffuses into the pores loses its kinetic energy via scattering . if the pores are interconnected , the ps has a probability to tunnel from one pore to another . a fraction of the ps reaches the film surface and exits into vacuum . a classical model of thermalization process was developed by nagashima et al . their calculations reproduce very well the behavior for sio@xmath2 aerogel with pore sizes of about 100 nm . however , a classical approach is not expected to give reliable predictions for ps confined in few nm pores because quantum mechanical effects become relevant . as a matter of fact , in this regime the de broglie thermal wavelength of ps is comparable with the size of the pores . recently , mariazzi et al . @xcite considered phonon scattering to reproduce the thermalization process of ps in a box ( closed porosity ) showing that the minimal energy is not that of the lowest accessible level because the momentum phonons can exchange is fixed . in the same paper , they pointed out that this is not the case in rectangular channels because in one direction the side of the potential well tends to infinity ( z - axis ) . therefore , the magnitudes of the @xmath3 and @xmath4 momentum are quantized but the @xmath5 tends to be a continuum and the minimal energy that ps can reach is given by the ground state in the x - y components . thus measuring the ps emission energy provides a method to distinguish between the two different pores architectures ( see section [ sec : discussion ] ) . in previous studies of ps emission in vacuum using time of flight ( tof ) , many interesting effects , such as the emission from the surface of different materials @xcite-@xcite were investigated . recently , this technique was applied to study mesoporous and hybrid silica films @xcite-@xcite , to evaluate the continuity barrier @xcite and the effect of thermalization for pore surfaces decorated with different groups @xcite . however , the influence of the temperature on ps emission in vacuum from mesoporous thin films was never studied in detail . to our knowledge , only a work of mills et al . @xcite with fumed silica revealed a dependence with the temperature of the ps emission in vacuum . the fraction of ps in the low energy tail was estimated but no quantitative estimate on the value of the minimal emission energy was obtained . very recently cassidy et al . @xcite measured the emission energy of ps from mesoporous films using doppler spectroscopy . this is a different technique from that used in the present study . results consistent with the one presented in this paper ( though not sample temperature dependencies ) have been obtained . in this paper , we study the ps yield and emission energy in vacuum as a function of the target temperature for two different kinds of mesoporous thin films with the same tetraethoxysilane ( teos ) mineral source for the silica network skeleton precursor : ctacl - teos and f127-teos . the density of the c sample is approximately 1.2 g/@xmath6 and of the f sample is 1.5 g/@xmath6 . both samples were spin - coated on glass similar to the ones we measured in @xcite . the c samples are prepared via a sol - gel process using cetyl trimethyl ammonium chloride ( ctacl ) cationic surfactants as the organic pore generator ( porogen ) agent @xcite . a pure aqueous method is used . the ctacl / teos molar ratio for the films prepared is 0.22 . after deposition , the ctacl - teos / glass samples are treated at 130 @xmath7c and stored in air . the f samples use non - ionic pluronic f-127 triblock copolymer ( eo@xmath8po@xmath9eo@xmath8 ) as surfactant and were prepared in the same way as described in @xcite . both samples were calcinated for 15 minutes at 450 @xmath7c in air immediately before the e@xmath10 measurements . the recorded x - ray diffraction patterns indicate no symmetry in the pore organization . the ethz slow positron beam used for these measurements is described in greater detail in @xcite . the positrons flux is 25000 e@xmath10/s . the slow positron beam is stopped in the sio@xmath2 target . the positrons can either form positronium , i.e. o - ps or p - ps , or annihilate into @xmath11 s . the detection with an mcp ( micro - channel plate ) of the secondary electrons ( se ) emitted when the positrons hit the target serves for tagging the positronium formation . the se leave the target accelerated to 1 - 11 kev by the same voltage applied to the target relative to the grounded transport tube that is used to implant the positrons in the positronium converter . the se are then transported by a magnetic field in the backward direction , as shown in figure [ traj ] . the electrons move along the magnetic field line in spirals and are deflected to the mcp region by the @xmath12 filter . the tagging efficiency varies from 70% to 30% in the energy range 1 - 10 kev . the samples are mounted on a cryocooler head to allow the possibility of varying their temperature in the range of 50 - 400 k. the start time t@xmath13 for the detectors is triggered by the mcp detecting the se emitted when the positrons hit the target . in both pals and tof detectors the stop is given by one ( or more ) annihilation photons depositing some energy in the calorimeter ( ecal ) . both ecal are composed of bgo crystals with hexagonal shape , 61 mm external diameter and 200 mm length . the time resolution of the system mcp - ecal was measured to be around 5 ns ( fwhm ) . the typical energy resolution of the crystals is about 25 - 30@xmath14 ( fwhm ) . the ecal for the tof is placed behind a lead slit at a distance z from the target that can be varied ( as sketched in figure [ traj ] ) . the detector is screened by four half cylinders of lead surrounding the beam pipe . the thickness of the shielding is 70 mm , the width of the slit is set to 5 mm and its position with respect to the target can be adjusted . for this measurement in order to maximize the signal to background ratio the center of the slit was placed at a distance of 18 mm from the target . the main contribution to the background is given by photons coming from direct positrons and p - ps annihilations in the target ( so called prompt peak ) . some of those photons can be detected in the crystals after compton scattering in the lead shield . the determination of the slit position with respect to the target was done by scanning the slit position in 0.1 mm steps and recording the maximum of the 511 kev annihilation peak . the scans were performed 30 min after the temperature was set on the sample . this technique was very important in order to correct the slit position as a function of the temperature . in fact , at 50 k the contraction of the cryocooler head was measured to be @xmath15 mm in agreement with the prediction of finite elements calculations performed with the comsol package @xcite . the pals detector was designed to have a large acceptance to provide a uniform efficiency for detecting the ps emitted in vacuum @xcite . in order to design the detectors and interpret the data , simulations served as a powerful tool . in the monte carlo ( mc ) simulations of our setup , the 3d eb - fields were calculated with the comsol multi - physics program and the positron / electron trajectories in the beam were simulated with geant4 . the simulation of the photon detection in the apparatus was based on the same package @xcite . new classes were written in order to simulate the o - ps production , propagation in the beam pipe , reflection on pipe walls that were assumed to be knudsen - like ( ps is reflected isotropically ) , pick - off effect or decay ( ps is not a standard particle in geant4 ) . the events for the o - ps@xmath16 process were generated taking into account the decay matrix element @xcite . the geometries of the beam transport pipe , photon detector , positron tagging system and its material were coded into simulations . the results were cross - checked with our experimental measurements for both photon detection @xcite and particle transport in the eb - fields @xcite . the measurements were taken in a clean vacuum of @xmath17 mbar . to avoid water contamination of the film during the cooling down , the target was kept at room temperature for an hour using a heater before lowering its temperature . the cooling cycles were repeated and the data confirmed the reproducibility of the results . the signals from the bgo s photomultipliers are split to record both energy and timing with a charge - to - digital converter ( qdc caen v792 ) and a time - to - digital converter ( tdc caen v775 ) . a cut on the energy deposited in the bgo between @xmath18 kev was applied to optimize the signal to background ratio suppressing compton scattering events in the collimator from direct and p - ps annihilations in the target . in this section we present the results we obtained from the pals measurements . the spectra are analyzed using the lt9 @xcite program . the lifetimes of the decay components and its fractions are resolved by fitting the pals spectra . the program finds 3 exponentials convoluted with a 5 ns fwhm resolution function of the spectrometer . the shortest exponential ( less than 4 ns ) is originated by direct positron and p - ps annihilations . it is disregarded because we are interested only in contribution of o - ps . we define as ( @xmath19 , @xmath20 ) , and ( @xmath21 , @xmath22 ) the intensities and the lifetimes of the two longer exponentials . to determine the yield @xmath23 of ps emitted in vacuum and the lifetime @xmath24 in the pores of the film we used a model of ps escaping in vacuum @xcite . according to this model the lifetime @xmath24 in the pores of the film is defined as : @xmath25 the yield @xmath23 of ps emitted in vacuum was calculated according to : @xmath26 where @xmath27 is the escape rate of free ps into vacuum . we extracted the films thicknesses for the two samples by fitting the total ps yield as a function of the positron implantation energy to the makhovian profile . these thicknesses are reported in table [ tbl1 ] in which we also present the values of @xmath24 and @xmath28 calculated with the expressions above . we use the results of the fits of the pals spectra at 6 kev for the c sample and 10 kev for the f sample at 50 k and 300 k. these implantation energies will serve as a reference for the rest of the paper . we choose those values in order to maximize the amount of thermalized ps . at these energies the majority of the positrons are still implanted within the films ( no significant drop of the total ps yield @xmath29 is observed , see fig . [ palsdata ] ) and the emission energy is in the constant region ( see next section ) . with @xmath24 one can calculate the effective pore size @xmath30 in the films applying the gidley et al . @xcite extension of the tau - eldrup model @xcite-@xcite ( we will call it hereafter rte model ) . as one can see , at lower temperature the lifetime in the film increases as predicted by the rte model . this can be understood in the following way : the overlap of the ps wave function with the volume contained within a distance @xmath31 nm @xcite from the walls for which the annihilation rate is assumed to increase is less for ps confined in the pores occupying the ground state than for ps in excited states . since at lower temperatures the population of the ground state is higher the pick - off rate decreases . however , the measured lifetimes are lower than what is expected by the calculations using the rte model . for the c sample the discrepancy is less than 10% but for the f sample the measured value is 30% lower suggesting that the pick - off rate increases . deviations from the rte model were already observed in previous measurements using the 2 to 3@xmath32 ratio technique @xcite and in a measurement using pals @xcite . different factors could be responsible for that as pointed out in those papers and deserves further studies . since the parameter @xmath33 of the rte model was calibrated at room temperature we use the lifetimes at 300 k to determine the effective pore size that ps experiences . we report the results at the end of the paper in table [ tbl3 ] for both rectangular channels and cubic boxes . interestingly , we found that the yield of ps emitted in vacuum can be considered as independent of the temperature of the target ( see fig . [ palsdata ] ) . ) and total yield of ps ( @xmath34 ) for f ( upper plot ) and c samples ( lower plot ) at room temperature and 50 k.__,scaledwidth=50.0% ] this may seem to contradict the fact that the lifetime in the films increases with the temperature . however , our measurements indicate that the escape rate in vacuum is smaller at lower temperature ( see table [ tbl1 ] ) , explaining the observation that the yield is constant . a possible explanation of this effect can be found considering that ps tunnels from one pore to the other . in this case , the tunneling probability will decrease with temperature explaining our measurements . .[tbl1 ] film thickness ( z ) , ps lifetime in the pores @xmath35 and escape rate @xmath28 at 50 k and 300 k. [ cols="^,^,^,^,^,^",options="header " , ] figure [ tof_fd ] shows the acquired tof spectra at different implantation energies for the f sample . for every run the time spectra of each crystal are calibrated finding the position of the peak arising from the annihilations in the target that defines t@xmath13 . first we extract the mean energy of ps emission in vacuum applying the analysis method proposed in @xcite as follows . the background due to the annihilations in the target is subtracted by fitting the measured time spectra with the resolution function of our detector determined using targets ( aluminum and kapton ) in which the ps formation is negligible . the tof spectra are corrected for the ps decays and the time spent in front of the detector with the factor @xmath36 . the maximum of the peak distribution defines the mean ps emission energy in the direction perpendicular to the film surface . let us note that , neglecting the reflection of ps in the beam pipe , with the time - of - flight method one measures only the mean of the energy component perpendicular to the collimator that we define as the z - axis ( we define it as @xmath37 ) . the triangles in fig . [ opsevsv ] , represent it as a function of the implantation energy . as a function of the implantation voltage at a target temperature of 300 k. the triangles represents the energy extracted with the maximum analysis method ( ma ) , the squares are after the subtraction of the non - thermalized part ( maximum analysis method after correction of the spectra , ma - cs).__,title="fig:",scaledwidth=50.0% ] as a function of the implantation voltage at a target temperature of 300 k. the triangles represents the energy extracted with the maximum analysis method ( ma ) , the squares are after the subtraction of the non - thermalized part ( maximum analysis method after correction of the spectra , ma - cs).__,title="fig:",scaledwidth=50.0% ] one can see that starting from 3 kev for the c sample ( this was confirmed by other measurements on similar samples @xcite ) and 4 kev for the f sample the value of the energy emitted in vacuum tends to be constant . clearly , the emission energy of ps calculated in this way is not the minimal energy due to the confinement in the pores because one has different contributions given by the convolution of the emission energy with the implantation profile . in order to isolate the thermalized part from the tof spectra , one has to subtract the contributions of non - thermalized ps components ( see fig . [ components ] ) . to estimate these non - thermalized contributions , we suppose that the shape of their distribution , nt(t ) , is represented by the tof spectrum obtained at a low implantation energy . to select the implantation energy ( from the ones we had measured ) and the scaling factor of nt(t ) for the subtraction , we relied on the mc . we used the values for which the best fit between the mc and the spectra obtained after the subtraction of nt(t ) was achieved ( for more details see @xcite ) . we found that the best fits were obtained for 2 kev with l@xmath38=200 nm in the case of the c sample , and 3 kev and l@xmath39=350 nm for the f sample . for a given implantation energy e@xmath40 , nt(t ) is scaled down by the fraction of positrons implanted at depths smaller than l@xmath38 and l@xmath39 . this was determined by using a makhovian profile . the results of this analysis are shown as the squares in fig . [ opsevsv ] . a parabolic fit is used to determine the position of the maximum . the statistical error of the fit is typically @xmath419 ns for the f and @xmath416 ns for the c sample . the uncertainty on the determination of the slit position of @xmath42 mm results in a systematic error of the order of @xmath43 mev in the determination of the ps mean emission energy @xmath37 ( for implantation energies above 3 kev ) . the subtraction procedure described above introduces a systematic error that we estimated analyzing the data using different values of l@xmath44 nm and l@xmath45 nm . the estimated error is @xmath46 mev for the c and @xmath47 mev for the f sample . thus the combined statistical and systematic error is at a level of @xmath412.9 mev for the f and @xmath413.0 mev for the c sample . the results for the c and the f samples at room temperature and at 50 k are shown in figs . [ opsevsv_rt]-[opsevsv_lt ] . for implantation energies higher than 4 kev for the c and 5 kev for the f sample the values @xmath37 of the mean emission energy are constant . in the c sample , @xmath37 reaches its constant value at lower implantation voltages because the pore size is smaller than in the f sample ( see fig . [ opsevsv_rt ] ) . those values are higher than the thermal energy that ps will have if it would thermalize at the temperature of the film . as expected in the presence of confinement in the pores , the mean emission energy is higher for the c sample with pore sizes smaller than the f sample . as a function of the positron implantation energy for c and f samples at 300 k.__,scaledwidth=50.0% ] for implantation energies higher than 2 kev at 50 k and 300 k for the c sample . lower plot : ps mean energy as a function of the implantation energies higher than 4 kev at 50 k and 300 k for the f sample.__,scaledwidth=50.0% ] the tof technique measures @xmath37 , the mean energy of the ps atoms in the z - direction . to find the mean emission energy of ps in vacuum , @xmath37 should be multiplied by a factor @xmath48 that takes into account the angular distribution . assuming that the ps is emitted mono - energetically and isotropically from the surface , one can calculate that @xmath48 is equal 2 . in this estimation , the reflection of ps in the beam pipe and the detector acceptance , i.e. the fact that a fraction of events decaying before or after the collimator aperture are detected , are not taken into account . therefore , to determine @xmath48 considering these effects we used the mc simulation we described in section [ sec : mc ] . as shown in the upper plot of fig . [ mc ] , a satisfactory agreement between the data and the mc ( adding the spectra at 2 kev that takes into account the non - thermalized ps ) is achieved . we attribute the difference between 40 and 100 ns to the approximation used in the subtraction method where the contribution of the non - thermalized ps is underestimated since only a spectrum of a defined energy is used for this correction . the fact that the ps is emitted with an angular spread is clearly supported by the data . as one can see in the lower plot of fig . [ mc ] for ps emitted with no angular spread the data are not reproduced . the physical interpretation is that in the films we studied the pores have no organization thus they are expected to be randomly aligned . the value of @xmath48 estimated with the mc is 1.7 . this is consistent with the expectation of the analytical result and the values reported in previous experiments @xcite . a detailed scan shows that the mean ps energy ( @xmath49 ) decreases with the sample temperature down to a minimum level ( see fig . [ opsenergyvst ] ) . for the c sample this value is basically constant ( 73@xmath415 mev ) in the range of temperature in which we performed our measurement . this can be understood by the fact that in this sample the confinement energy is much higher than the thermal energy at room temperature ( @xmath50 mev ) thus almost all ps is in the ground state . for the f sample there is a weak dependence on the temperature . due to the bigger pore size compared to the c sample , the energy of the ground state is only twice the thermal energy at room temperature . therefore , the probability to find the ps occupying an excited state is higher . as expected , this probability decreases with the temperature thus the minimal energy reaches its constant value of 48@xmath415 mev . the time that ps spends in the films before being emitted in vacuum was not considered in our determination of the emission energy . the measurements presented in fig . [ opsenergyvst ] are in a regime in which a classical approach is not expected to give reliable results . some theoretical work to develop a full quantum mechanical picture of the emission process is required to address this problem ( as pointed out in @xcite as well ) . to understand the behavior of the value of the minimal energy as a function of the film temperature we present a simple model of ps in thermodynamic equilibrium at a temperature t in rectangular boxes . the expectation value @xmath51 of the hamiltonian operator for ps confined in 1 dimensional infinite well in contact with a reservoir at a temperature @xmath52 is given by : @xmath53 where z is the partition function defined as @xmath54 where @xmath30 is the dimension of the well , @xmath55 is the ps mass , @xmath56 is the principal quantum number and @xmath57 and @xmath58 are the planck and the boltzmann constants . to calculate the mean value @xmath59 of the energy for the 3d case we can use @xmath60 where to calculate @xmath61 and @xmath62 one can substitute the pore side length @xmath30 in eq . [ eq : z ] with @xmath63 and @xmath64 . for ps confined in rectangular pores , we thus obtain : @xmath65 for the case of a cubic box one can set @xmath66 . to compare the prediction of the pore size that one can extract from the tof measurements with the pals results we fit the data using eq . [ eq : h ] . to construct the function used for the fit we kept only the first 50 terms of the sum . this is very conservative , the probability for ps to occupy a state higher than n@xmath6710 for the kind of target and the temperatures we used in this study is already negligible . the solid lines in fig . [ opsenergyvst ] represent the fit to the data where the pore side length ( @xmath30,@xmath63,@xmath64 ) are left as free parameters . the fits were repeated assuming cubic box pores and the results are shown as the dashed lines in fig . [ opsenergyvst ] . we used migrad from the minuit package @xcite as a minimization procedure . as one can see , the fit to the data suggests that the pores of both samples are better modeled as rectangular pores than cubic boxes . as proposed in @xcite , to compare the values obtained from the fit with the ones extracted from the pals measurements , one has to add twice the parameter @xmath33 ( see previous section ) . the pore side lengths obtained in this way are reported in table [ tbl3 ] . in particular , the fit supports the idea that the pores are better modeled by rectangular channels . the pore side length in one direction ( @xmath64 ) obtained from the fit is much longer than the substrate thickness . since , the same values for the side length @xmath30 and @xmath63 are obtained fitting the data assuming rectangular channels instead of rectangular boxes we do not report this value in table [ tbl3 ] as it has no physical meaning . with the pore side lengths @xmath30,@xmath63,@xmath64 left as free parameters . the dashed lines were obtained fitting with eq . [ eq : h ] with a single side length free @xmath68 ( cubic box pores).__,scaledwidth=50.0% ] ccc + & tof & pals + @xmath69 [ nm ] & 3.3@xmath410.1 & 4.2@xmath410.5 + @xmath70 [ nm ] & ( 2.7@xmath410.8 , 2.7@xmath410.8 ) & ( 3.3@xmath410.5 , 3.3@xmath410.5 ) + @xmath71 [ mev ] & 73@xmath415 + + & tof & pals + @xmath69 [ nm ] & 4.1@xmath410.1 & 6.4@xmath410.5 + @xmath70 [ nm ] & ( 2.9@xmath410.8 , 3.6@xmath411.8 ) & ( 4.3@xmath410.5 , 5.3@xmath410.5 ) + @xmath71 [ mev ] & 48@xmath415 + according to the quantum mechanical model for ps thermalization of mariazzi et al . @xcite the index @xmath56 in the sum of eq . [ eq : z ] could differ from 1 for cubic pores if the level separation of two close ps energy levels is higher than the maximum momentum that a single phonon can exchange . we performed fits with different @xmath72 but the results did not improve , supporting the idea that the pores are not very well modeled by cubic boxes . to summarize , for the c sample with the tof data the best fit was obtained for pores modeled as square channels of 2.7@xmath410.8 nm side length while the lifetime method gives 3.3@xmath410.5 nm . the best fit to the data for the f sample was obtained for rectangular channels with a cross section of 2.9@xmath410.8 nm by 3.6@xmath411.8 nm that has to be compared with 4.3@xmath410.5 by 5.3@xmath410.5 of the pals measurement . this pore size was extracted by applying the rte model to reproduce the measured lifetime of 74 ns assuming the ratio @xmath73 to be the same as in the tof results . both values obtained with the tof measurement are systematically lower than the pals results . nevertheless , considering the approximation used in our model , the assumptions that the pores can be treated as rectangular channels and the uncertainty in the determination of the pore size we conclude that our results ( summarized in table [ tbl3 ] ) are in reasonable agreement with expectations . in this paper , we show that the yield of ps emitted in vacuum measured with the pals technique is independent of the temperature of the mesoporous thin films . the lifetime in the films increases with a decrease in temperature but the measured values are lower than expected by the rte model . this suggests that another source of pick - off which depends on the temperature should be invoked . further studies are needed to investigate the origin of this effect . the escape rate in vacuum decreases explaining the observation that the yield of ps emitted in vacuum is the same at 50 and 300 k. furthermore , we show that due to quantum mechanical confinement in the pores the ps emission energy into vacuum has a minimal value . the minimal energy is higher for the sample with smaller pores and this constant value is reached at a lower positron implantation energy . our results are in fair agreement with a model of ps confined in rectangular channels in thermal equilibrium with the sample . the measured minimal energy for the c sample was found to be 73@xmath415 mev while for the f sample it is 48@xmath415 mev . these experimental results provide a solid ground to understand how to produce ps at lower temperature and could serve to develop a more realistic model to interpret the data . this work was supported by the swiss national foundation , the eth zurich and the cnpq ( brazil ) . p.c . would like to thank c. lenz cesar , p. lotti and b. barbiellini for the very useful discussions . m. deutsch , phys . 82 , 455 ( 1951 ) . s. karshenboim , int . j. mod . phys . a 19 , 3879 ( 2004 ) . a. badertscher , p. crivelli , w. fetscher , u. gendotti , s. gninenko , v. postoev , a. rubbia , v. samoylenko , d. sillou , phys . rev . d 75 , 032004 ( 2007 ) [ arxiv : hep - ex/0609059 ] . a. rubbia , int . phys . a 19 , 3961 ( 2004 ) [ arxiv : hep - ph/0402151 ] . s.w.h . eijt , b. barbiellini , a.j . houtepen , d. vanmaekelbergh , p.e . mijnarends , a. bansil , phys . ( c ) 4 , 3883 ( 2007 ) , arxiv:0705.1113v1 . d. w. gidley , h. g. peng , r. s. vallery annual review of materials research 36 , 49 ( 2006 ) . d. w. gidley , w. e. frieze , t. l. dull , a. f. yee , e. t. ryan , and h. m. ho , phys . rev . b 60 , r5157 ( 1999 ) . d. w. gidley , w. e. frieze , t. l. dull , j. sun , a. f. yee , c. v. nguyen , and d. y. yoon , appl . 76 , 1282 ( 2000 ) . s. chu , a. p. mills , j. l. hall , phys . 52 , 1689 ( 1984 ) . m. s. fee , a. p. mills , s. chu , e. d. shaw , k. danzmann , r. j. chichester and d. m. zuckerman , phys . 70 , 1397 ( 1993 ) . d. b. cassidy , h. w. k. tom and a. p. mills , aip conf . 1037 , 66 ( 2008 ) . m. amoretti , et al . [ athena collaboration ] , nature ( london ) 419 , 456 ( 2002 ) . g. gabrielse , n. bowden , p. oxley , a. speck , c. h. storry , j. n. tan , m. wessels , d. grzonka , w. oelert , g. schepers , t. sefzick , j. walz , h. pittner , t. w. hnsch , and e. a. hessels , phys . 89 , 213401 ( 2002 ) . g. b. andresen , [ alpha collaboration ] , phys . b 685 , 141 ( 2010 ) . j. w. humberston , m. charlton , f. m. jacobsen and b. i. deutch , j. phys . b 20 , l25 ( 1987 ) . m. charlton , phys . a 143 , 143 ( 1990 ) . b. i. deutch , m. charlton , m. h. holzscheiter , p. hvelplund , l. v. jorgensen , h. knudsen , g. laricchia , j. p. merrison and m. r. poulsen , hyperfine interaction 76 , 153 ( 1993 ) . j. p. merrison , h. bluhme , j. chevallier , b. i. deutch , p. hvelplund , l. v. jorgensen , h. knudsen , and m. r. poulsen , m. charlton , phys . 78 , 2728 ( 1996 ) . c. h. storry , a. speck , d. le sage , n. guise , g. gabrielse , d. grzonka , w. oelert , g. schepers , t. sefzick , h. pittner , m. herrmann , j. walz , t. w. hnsch , d. comeau , e. a. hessels , phys . 93 , 263401 ( 2004 ) . a. kellerbauer et al . [ aegis collaboration ] , nucl . instr . and meth . b 266 , 351 ( 2008 ) . p. perez , l. liszkay , m. f. barthe , j .- boilot , m. carty , c. corbel , p. crivelli , a. curtoni , o. delferrierre , p. desgardin , m. etienne , u. gendotti , m. hassan , a. mohri , t. ohdaira , j .- m . rey , a. rubbia , h. saitoh , y. sauce , r. suzuki , a. walcarius , and y. yamazaki , aip conf . 1037 , 35 ( 2008 ) . r. bernabei et al . [ dama collaboration ] , eur . j. c 56 , 333 ( 2008 ) . [ arxiv:0804.2741 [ astro - ph ] ] . r. foot , phys . d 78 , 043529 ( 2008 ) [ arxiv:0804.4518 [ hep - ph ] ] . a. badertscher , a. belov , p. crivelli , s. n. gninenko , j. p. peigneux , a. rubbia and d. sillou , int . j. mod . phys . a 19 , 3833 ( 2004 ) . p. m. platzmann and a. p. mills jr . , phys b 49 , 454 ( 1994 ) . s. valkealahti and r. nieminen , appl . phys . a 32 , 95 ( 1983 ) . m. tuomisaari , r. h. howell , and t. mcmullen , phys . b 40 , 2060 ( 1989 ) . y. nagashima et al . , phys . rev . b 58 , 12676 ( 1998 ) . a. p. mills , jr . , e. d. shaw , r. j. chichester , and d. m. zuckerman , phys . rev . b 40 , 2045 ( 1989 ) . r. s. yu , t. ohdaira , r. suzuki , k. ito , k. hirata , k. sato , y. kobayashi , and j. xu , appl . 83 , 4966 ( 2003 ) . k. ito , r. s. yu , k. sato , k. hirata , y. kobayashi , t. kurihara , m. egami , h.arao , a. nakashima , and m. komatsu , j. appl . 98 , 094307 ( 2005 ) . c. he , m. muramatsu , n. oshima , t. ohdaira , a. kinomura , and r. suzuki , phys . lett . a 355 , 73 ( 2006 ) . m. p. petkov , m. h. weber , k. g. lynn , and k. p. rodbell , appl . 79 , 3884 ( 2001 ) . r. suzuki , t. ohdaira , y. kobayashi , k. ito , r. s. yu , y. shioya , h. ichikawa , h. hosomi , k. ishikiriyama , h. shirataki , s. matsuno , and j. xu , mater . forum 445 - 446 , 224 ( 2004 ) . j. xu , a. p. mills , jr . , and c. case , appl . 87 , 54105 ( 2005 ) . h. k. m. tanaka , t. kurihara , and a. p. mills , jr . b 72 , 193408 ( 2005 ) . c. he , t. ohdaira , m. muramatsu , n. oshima , a. kinomura , r. suzuki , t. oka and y. kobayashi phys . b 75 , 195404 ( 2007 ) . d. b. cassidy , p. crivelli , t. h. hisakado , l. liszkay , v. e. meligne , p. perez , h. w. k. tom , and a. p. mills jr . a 81 , 012715 ( 2010 ) . l. liszkay , c. corbel , p. perez , p. desgardin , m .- f . barthe , t. ohdaira , r. suzuki , p. crivelli , u. gendotti , a. rubbia , m. etienne , and a. walcarius , appl . lett . , 92 , 063114 ( 2008 ) . y. cohen , k. landskron , n. tetretault , s. fournier - bidoz , b. hatton , g. a. ozin , adv . 15 , 593 ( 2005 ) . w. yantasee , y. lin , x. li , g. e. fryxell , t. s. zemanian and v. v. viswanathan , analyst , 128 , 899 ( 2003 ) . n. alberola , t. anthonioz , a. badertscher , c. bas , a.s . belov , p. crivelli , s.n . gninenko , n.a . golubev , m.m . kirsanov , a. rubbia and d. sillou , nucl . instr . method a 560 , 224 ( 2006 ) . p. crivelli , u. gendotti , l. liszkay , p. perez , a. rubbia , mat . forum 607 , 251 ( 2009 ) . agostinelli et al [ geant4 collaboration ] , nucl . instr . method a 506 , 250 - 303 ( 2003 ) . a. ore and j. l. powell , phys . 75 , 1696 ( 1949 ) . a. badertscher , p. crivelli , m. felcini , s.n . gninenko , n.a . goloubev , p. nedelec , j.p . peigneux , v.postoev , a. rubbia and d. sillou phys . p. crivelli , a new search for invisible decays of ortho - positronium " , phd thesis , no.16617 , eth zurich , switzerland ( 2006 ) . l. liszkay , c. corbel , l. raboin , j .- boilot , p. perez , a. brunet - bruneau , p. crivelli , u. gendotti , a. rubbia , t. ohdaira and r. suzuki , appl . 95 , 124103 ( 2009 ) . j. kansy , nucl . instrum . meth . a 374 , 235 ( 1996 ) . l. dull , w. e. frieze , d. w. gidley , j. sun , and a. f. yee , j. phys . b 105 , 4657 ( 2001 ) . s. j. tau , chem . 56 , 5499 , ( 1972 ) . m. eldrup , d. lightbody amd j. n. sherwood , chem . phys . lett . 63 , 51 , ( 1981 ) . t. goworek , k. ciesielski , b. jasinska , and j. wawryszczuk , chem . 272 , 91 ( 1997 ) . k. ito , h. nakanishi and y. ujihira j. phys . b 103 , 4555 - 4558 ( 1999 ) . s. mariazzi , l. toniutti , n. patel , r.s . brusa , applied surface science 255 , 191 - 193 ( 2008 ) . c. g. fischer , m. h. weber , c. l. wang , s. p. mcneil , and k. g.lynn , phys . rev . b 71 , 180102r ( 2005 ) . s. thraenert , e. m. hassan , d. enke , d. fuerst and r. krause - rehberg , phys . status solidi c 4 , 3819 ( 2007 ) . u. gendotti , ph.d thesis unpublished , eth zurich , switzerland ( 2010 ) . f. james , m. roos , minuit , function minimization and error analysis , cern d506 ( long writeup ) .

recently , the proximity effect between a superconductor ( @xmath0 ) and a ferromagnet ( @xmath1 ) has attracted much attention because , due to advances in materials growth and fabrication techniques @xcite , well controlled structures , in which it can occur became available . such @xmath1/@xmath0 hybrid structures are important from the point of view of their intrinsic scientific interest , as they allow the study of the interplay between ferromagnetism and superconductivity @xcite as well as of device applications in such areas of technology as magnetoelectronics @xcite or quantum computing @xcite . in the present context the proximity , on one hand , means the leakage of superconductivity into the ferromagnet , and on the other , the spin polarization of the superconductor . in the normal metal - superconductor systems the proximity effect has been studied over 30 years , and is now well understood @xcite to be governed by the andreev reflection @xcite processes , in which an electron at the interface is reflected as a hole while a cooper pair is transfered into superconductor . this mechanism makes the movement of electrons and holes in the normal metal highly coherent . if the normal metal is replaced by a ferromagnet a number of new phenomena may occur . for instance one may expect a version of the @xmath2 state , predicted in sixties for a bulk superconductor in a strong spin exchange field , by fulde and ferrell @xcite and larkin and ovchinikov @xcite , to be realized . in the bulk , as it is well known , the exchange field tends to polarize the conduction electrons and holes . in particular if these electrons are cooper paired , naively one would expect that either the exchange field is too week to break the pairs , or it leads to the first order phase transition to the normal state . however , as was demonstrated in ref . @xcite , for certain values of the exchange field a new superconducting depairing state is realized through first order phase transition from the @xmath3 state and it transforms continuously , by a second order phase transition , into the normal state as the strength of the exchange field is increased . this @xmath2 state has a spatially dependent order parameter corresponding to the nonzero center of mass motion of the cooper pairs . another novel feature of this state is a current flowing in the ground state . the unpaired electrons tend to congregate at one portion of the fermi surface so a quasiparticle current is produced . in order to satisfy the bloch theorem : no current in the ground state , a supercurrent , generated by the nonzero value of the pairing momentum , flows in opposite direction , and the total current is zero . interestingly , similar oscillations of the pairing amplitude ( cooper pair density ) have been predicted @xcite-@xcite in a ferromagnet proximity coupled to a superconductor . it turns out that these oscillations are responsible for the oscillatory dependence of the critical temperature on the thickness of the ferromagnetic slab @xcite . this effect as well as the corresponding oscillations of the density of states at the fermi level @xcite , have been observed experimentally @xcite . in this paper we investigate further the ramifications of these interesting phenomena . one of these is the formation of so called @xmath4 junction state @xcite which is also a fingerprint of the oscillatory behaviour of the pairing amplitude . the @xmath4-state effect has been extensively studied in connection with the high-@xmath5 superconductors , where due to spatially inhomogeneity of the order parameter , the cooper pair wave function can change sign at the interface and this leads to the formation of a zero - energy mid - gap state . remarkably , this zero - energy state is unstable to the occurrence of a spontaneous current flowing parallel to and within a coherence length of the interface ( see eg . @xcite and references therein ) . in what follows we report on our finding similar spontaneous currents but of different origin in a @xmath0/@xmath1/@xmath6 trilayer . we also find zero energy bound states due to the finite size of the @xmath1 slab . in fact even for a @xmath7/@xmath0 heterostructure , in agreement with @xcite , we have found states within the @xmath0 gap . although the energy of these states depend on the thickness of the normal metal , they never reach zero , namely @xmath8 . interestingly , when the normal metal is replaced by a ferromagnet , these bound states split , and for certain values of the exchange field , they cross the fermi level @xcite . as we shall show , this circumstance leads to a current flow . clearly , the spontaneous current is strictly related to the zero energy states , as in the case of the high - @xmath5 superconductors , but the origin of these states is completely different . as mentioned above , there are two competing parts to the total current in the @xmath2 state . one is related to the unpaired electron movement and the other is a supercurrent . so one may expect , that in the case of the @xmath1/@xmath0 heterostructure , these parts could be spatially unbalanced creating regions of net current flow . indeed this is what we have discovered . a brief report of our main results has already been published @xcite . here we wish to present a more detailed and systematic study of the zero energy andreev bound states and the corresponding spontaneous current in the @xmath1/@xmath0 heterostructures . the paper is organized as follows : in sec . [ sec2 ] the simple model which allows for self - consistent description of the @xmath1/@xmath0 heterostructure is introduced . we also derive equations which have to be solved self - consistently . in particular , these include the maxwell s equation ( ampere s law ) for the coupling of the current to the magnetic field . then we describe some technical details concerning the principal layer technique , which allows us to describe a semi - infinite superconductor . in sec . [ sec3 ] the nature of the andreev bound states in the ferromagnet is discussed . the spontaneously generated current and corresponding magnetic field in the ground state are studied in the sec . [ sec4 ] . in sec . [ sec5 ] we provide some suggestions of how one might observe the spontaneous current and its polarization experimentally . finally , we conclude in sec . to study spontaneous currents and their polarization in the ferromagnet - superconductor heterostructure we have adopted a single orbital , nearest neighbour hopping hubbard model with negative @xmath9 on the @xmath0 side and zero @xmath9 otherwise . additionally we have assumed the site energies @xmath10 to be exchange split in the ferromagnet and spin independent in the superconductor . moreover , the simplest geometry allowing for a current flow is the @xmath11 system , shown in the fig.[fig1 ] , where the magnetic field in one direction , the vector potential and the current in another and the spatial modulation in a third orthogonal direction are explicitly indicated . in short , our model hamiltonian is given in the form : @xmath12 c^+_{i\sigma } c_{j\sigma } + \sum_{i\sigma } \frac{u_i}{2 } \hat n_{i\sigma } \hat n_{i-\sigma } \label{hamiltonian}\end{aligned}\ ] ] where , in the presence of a vector potential @xmath13 , the hopping integral is given by @xmath14 for nearest neighbour lattice sites , whose positions are @xmath15 and @xmath16 , and zero otherwise . the site energies @xmath10 are @xmath17 on the superconducting side and equal to @xmath18 on the ferromagnetic side , @xmath19 is the chemical potential , the on - site interaction @xmath20 is @xmath21 in the superconductor and zero elsewhere , @xmath22 , ( @xmath23 ) are the usual electron creation ( annihilation ) operators and @xmath24 is the electron number operator . note that the above description of the electrons with charges @xmath25 includes a coupling to a magnetic field @xmath26 , which is necessary for calculating the effects of current on the electronic states . in what follows we shall work within spin - polarized - hartree - fock - gorkov ( @xmath27 ) approximation , which means that we have approximated the interaction term in hamiltonian ( [ hamiltonian ] ) by the usual mean field theory mapping : @xmath28 . for the magnetic field we have chosen the landau gauge where @xmath29 and hence @xmath30 . furthermore , we have assumed that the effective @xmath27 hamiltonian is periodic in the direction parallel to the interface and therefore we work in @xmath31 space in the @xmath32 direction but in real space in the @xmath33-direction ( see fig.[fig1 ] ) . labelling the planes ( lines ) in fig.[fig1 ] by integers @xmath34 and @xmath35 at each @xmath36 point of the brillouin zone we shall solve the following @xmath27 layer index ( @xmath34 ) matrix equation for the retarded green s function matrix @xmath37 : @xmath38 where @xmath39 is the pauli unit matrix and @xmath40 is of the form : @xmath41 with @xmath42 . our system consists of infinite number of layers ( @xmath34 ) , because of fact that we have a semi - infinite superconductor . the @xmath1 region spreads from @xmath43 to @xmath17 , while superconductor is defined for @xmath44 ( see fig . [ fig1 ] ) . in order to describe semi - infinite system , we need infinite range matrix with block elements @xmath45 ( @xmath34 , @xmath35 going from @xmath46 through @xmath17 to @xmath47 ) . in short , strictly speaking , we have to invert the infinite range matrix : @xmath48 to render the problem traceable , it is useful to define the surface green s function ( @xmath49 ) @xcite as @xmath50^{-1}\}_{n , n } \label{sgfdef}\end{aligned}\ ] ] @xmath51 being semi - infinite submatrix for @xmath52 ( lower right corner of the matrix in eq.([hinf ] ) ) . the physical meaning of the @xmath49 is that it represents a square diagonal subblock corresponding to the semi - infinite bulk shown with a shaded region in the fig . [ fig1 ] . now we can assume , that electronic properties in the @xmath0 differ from that in the bulk only over the finite distance from the interface ( @xmath53 ) . @xmath54 can correspond to the distance of a few @xmath0 coherence lengths . for @xmath55 the electronic properties are that of the bulk superconductor . in other words , we assume that the system for @xmath55 is homogeneous . namely , for @xmath55 we have : @xmath56 as a consequence the surface green s function ( [ sgfdef ] ) is also @xmath34-independent and can be obtained from the bulk @xmath57 @xcite : @xmath58 where @xmath59 is the fourier transform of the green s function of the homogeneous superconductor @xmath60 , namely @xmath61 . clearly , the green s function @xmath60 for the case where the vector potential is equal to zero is given by the usual result : @xmath62 where @xmath63 is the single electron energy spectrum . the homogeneity of the system for @xmath55 leads to the conclusion , that the inversion of a semi - infinite block tridiagonal matrix ( [ hinf ] ) can be reduced to the inversion of the finite tridiagonal matrix : @xmath64 in the above equation the effect of the semi - infinite homogeneous system is contained in the embedding potential @xmath65 , which is related to the surface green s function through : @xmath66 evidently , by increasing @xmath54 until @xmath67 in the surface region does not change would constitute an exact albeit numerical solution of the eq . ( [ hfg ] ) . all calculations reported here were performed for @xmath68 . however , we have checked that larger @xmath54 has no influence on the results . the calculation of the various physical quantities , like particle concentration @xmath69 , spin polarization @xmath70 , order parameter @xmath71 , etc . , needs the evaluation of integrals of the product of the green s function and fermi distribution function as in @xmath72 . while working on the real energy axis , we need a large number of integration points to get good accuracy , unless there are no singularities in the density of states . however , in our calculations , there is a van hove singularity in the @xmath11 @xmath73 as well as @xmath3-like singularities . in fact we need approximately @xmath74 - @xmath75 points to get good accuracy , which is very time - consuming in self - consistent calculations . the situation is similar for summation over the matsubara energies @xmath76 because of the poles of the fermi function . there is an infinite number of poles , and infinite sum does not always converge well , so numerical treatment needs a large number of data points . to overcome these difficulties we follow ref . @xcite and approximate the fermi distribution function by : @xmath77 to get the correct behaviour near the fermi energy ( @xmath78 ) we choose @xmath79 . one can check , that the discrepancy between the approximate function and the true one , starts to play role for energies lower than @xmath80 . if we choose @xmath81 , where @xmath82 is the half of the bandwidth , the approximation works very well . for example , for the temperature @xmath83 it is sufficient to consider only @xmath84 points to get very good accuracy . the poles of the approximate function lie on a circle given by @xmath85 the solution of the above equation gives the poles in the form : @xmath86 in the limit @xmath87 , the poles ( [ polesenergy ] ) move to the matsubara energies , and the number of poles , @xmath88 , becomes infinite . the calculation of any physical quantity becomes : @xmath89 where @xmath90 is given by eq.([polesenergy ] ) . as usual in problems of interfaces , a large number of equations has to be solved self - consistently . in our case the parameters , to be determined self - consistently at each layer @xmath34 , are : the total electron density @xmath69 , the spin polarization @xmath70 , the @xmath0 order parameter @xmath71 , the current parallel to the interface @xmath91 and the vector potential @xmath92 . most of these quantities can be expressed in terms of the corresponding layer - diagonal elements of the green s functions @xmath93 ( see eq.([hhfg ] ) ) . the exception is the vector potential , for which we need an additional ( maxwell s ) equation . these relations are ( see eq.([oaverage ] ) ) : @xmath94 @xmath95 @xmath96 where @xmath97 is the @xmath98-th element of the matrix green s function , which is the solution of the @xmath27 equation ( eq . ( [ hfg ] ) ) . the current , for spin up electrons , in the @xmath32-direction can be calculated from the relation : @xmath99 which follows from the continuity equation for the charge ( @xmath100_- = - \sum_j j_{ij}$ ] ) . this current will give rise to a vector potential @xmath101 which will have to be used to update @xmath102 in eq.([hfg ] ) at the end of each self - consistency cycle . we calculated this new vector potential by solving numerically ampere s law , @xmath103 , which for the lattice problem at hand , is @xmath104 in all our calculations the eqs.([n])-([maxwell ] ) have been solved in each iteration step until self - consistency has been achieved . the numerical calculations have been performed at temperature @xmath83 for @xmath84 energy points , @xmath105 - @xmath36 points and @xmath106 layers ( @xmath107 ferromagnetic and @xmath107 superconducting ) . typically self - consistency , at the level of @xmath108 on all densities in eqs.([n])-([maxwell ] ) , has been achieved after @xmath109 - @xmath110 iterations . since we have determined the ferromagnetic ( [ m ] ) and superconducting ( [ delta ] ) order parameters on both sides of the interface fully self - consistently , we were able to study both @xmath0 and @xmath1 proximity effects . this means , that we were able to describe situations in which ferromagnetism and superconductivity coexist near the interface . firstly , we want to discuss the magnetic proximity effect , namely the entering of the spin polarization into the superconductor . the typical behaviour of the spin polarization ( @xmath111 ) is plotted in the fig . [ fig2 ] as a function of the layer index @xmath34 for a number of exchange splittings @xmath112 . we can see that the spin polarization exponentially decays over the distance of the @xmath0 coherence lengths @xmath113 , which in the present case is @xmath114 in units of the lattice constant @xmath115 . we have checked this by explicit calculations of @xmath113 for various values of the superconducting energy gap . so we can conclude that effect of proximity of the ferromagnet on the superconductor is very similar to case of the meissner effect , where an external magnetic field is excluded from the sample . note however that while the magnetic field is excluded on the spatial scale of the penetration depth the effective exchange field drops to zero within the distance of the coherence length @xmath113 in the superconductor . more interesting is the behaviour of the spin polarization on the ferromagnetic side of the heterostructure . as it is seen from fig . [ fig2 ] , it oscillates around its bulk value with period depending on the exchange splitting @xmath112 but not on the superconducting gap . however we were not able to fit the numerical points to any simple relation between parameters of the model . the best fit we have is @xmath116 . whilst this means that the period of the oscillations is a linear function of the exchange splitting the above relation is valid only for the thickness of the @xmath1 slab @xmath117 layers . for different thicknesses we have different relations and different numbers of period within the slab . for example , for @xmath118 , @xmath70 shows one period less for the corresponding exchange splittings @xmath112 . we now turn to the superconducting proximity effect , namely the leakage of the @xmath0 properties into the non - superconducting region . in our case , although @xmath119 on the ferromagnetic side , due to the fact that for @xmath120 , @xmath121 , the pairing amplitude @xmath122 can , in general , be non - zero and it usually is . the proximity effect is well understood in the case of non - magnetic metal - superconductor interface ( @xmath123/@xmath0 ) @xcite in terms of the andreev reflections @xcite . the andreev process describes a situation in which an electron impinging onto the @xmath123/@xmath0 interface is reflected back as a hole ( with opposite spin ) and a cooper pair is created in the @xmath0 . moreover , the movement of the electrons and holes in the @xmath123 close to the interface ( over the distance of the @xmath0 coherence length @xmath113 ) is highly correlated . the pairing amplitude , in this case , enters the normal metal and decays exponentially over the @xmath113 . when the normal metal is replaced by a ferromagnet one might naively expect that the proximity effect will be suppressed due to the pair breaking effect of the exchange field @xcite . this suggests that the pairing amplitude @xmath124 will decay over a distance much shorter than the superconducting coherence length . remarkably , as we can see in the fig . [ fig3 ] , this is not the case . instead , we can observe a very long range proximity effect with an oscillating pairing amplitude . this effect has been first noted by a. i. buzdin _ _ @xcite and have been studied in several recent publications @xcite . according to them , it can be attributed to a @xmath2 - like phenomena @xcite in @xmath1/@xmath0 heterostructures . the period of the oscillations of the @xmath124 depends on the exchange splitting @xmath112 ( see fig . [ fig3 ] ) and , similarly to the oscillations of @xmath70 , does not depend on the superconducting energy gap . clearly this behaviour is different from @xmath123/@xmath0 case , where the proximity effect depends on the @xmath0 coherence length @xmath113 . here , it depends only on the properties of the ferromagnet . moreover , the numerically obtained behaviour of the pairing amplitude @xmath124 is consistent with the analytical formula @xmath125 with @xmath126 being ferromagnetic coherence length . reassuringly this result is fully consistent with those of ref . the comparison our numerical results with the formula ( [ chi_analit ] ) for the exchange splitting @xmath127 , which gives @xmath128 is depicted in the inset of the fig . in fact we had to change @xmath129 by @xmath130 , in order to get better fit . so the solid line in the inset of the fig . [ fig3 ] corresponds to the formula @xmath131 with @xmath132 . it turns out that as exchange splitting becomes smaller the fitting to the analytical formula ( [ chi_analit ] ) becomes better . the small discrepancy in the coherence length may come from the fact , that we have used bulk value of the exchange field while calculating the coherence length . however in our system , the spin polarization oscillates around its bulk value , and if we calculate the average value of the spin polarization @xmath133 , we get a little bit smaller value than the bulk one . so , we conclude , that the effective exchange splitting @xmath134 also has to be smaller than its bulk value . indeed , when we used @xmath134 to calculate the coherence length , the agreement is improved . another reason may be the approximate form of the formula for the coherence length in the lattice case @xmath135 . the proximity effect manifests itself not only in the pairing amplitude and the spin polarization characteristics but also in the density of states ( @xmath73 ) . the layer resolved @xmath73 is defined as @xmath136 with @xmath137 being the @xmath98-th element of the matrix green s function - solution of the eq . ( [ hfg ] ) . we shall also use the density of states integrated over the total sample : @xmath138 . at this point we want to stress that , when we calculate thermodynamic quantities ( eq.([n])-([current ] ) ) we work on the complex energy plane , but calculating the @xmath73 we solve eq . [ hfg ] on the real energy axis only once . in fact we have added a small imaginary part ( @xmath139 ) for numerical purposes . the quantities : @xmath69 , @xmath70 , @xmath71 , @xmath92 as determined by the previous self - consistent procedure allow us to find the green function on the real energy axis in only one step . we also check if the new quantities like @xmath69 , @xmath70 , etc . agree with the old ones . for all such calculations the agreement has been found to be excellent . while the general solution of eq.([hfg ] ) allows for a current flow , in this section we show the results for a solution for which the current has been constrained to be zero . to appreciate the effect of spin polarization in the @xmath1 slab , let us recall again the case of no exchange splitting . if the non - magnetic metal ( @xmath123 ) is in the contact with superconductor ( @xmath0 ) , the density of states on the @xmath123 side shows features of the @xmath0 @xmath73 , i.e. the pseudo - gap opens up around the fermi energy in the @xmath73 . however it never turns into a true gap , namely the @xmath123 @xmath73 always possess low energy excitations . moreover , the gap in the @xmath0 @xmath73 also starts to fill up , due to the proximity of the normal metal . so in this case we deal with gapless superconductivity close to interface of the @xmath123/@xmath0 heterostructure . compared to the above summary , the ferromagnet - superconductor interface is enriched with a number of new features . for instance , the @xmath0 @xmath73 splits , if the exchange field is small . for strong enough exchange field , @xmath0 gap can be suppressed , indicating transition to the non-@xmath0 ( normal ) phase . an example of the layer resolved density of states for spin up ( solid line ) and down ( dashed ) is shown in the fig . [ fig4 ] . evidently there is a pseudogap in the @xmath1 @xmath73 , indicating that superconducting correlations are present in the @xmath1 layers . this effect is particularly strong for the layer index @xmath140 and @xmath141 . on the other hand , the proximity of the ferromagnet leads to existence of the low energy states in the @xmath0 @xmath73 . this effect is also observed in @xmath123/@xmath0 heterostructure but here we also notice differences in the spin up and spin down @xmath73 on the @xmath0 layers close to the interface . these are clear manifestations of the @xmath1 proximity effect . the most interesting physics of the heterostructures we consider , is the formation of andreev bound states in the @xmath123 or @xmath1 layers . they are particle in a box like states which can be associated with the semiclassical orbits bouncing back and forth between the @xmath0 and @xmath6 regions , as depicted in the fig . [ fig5 ] . as is well known in the case of @xmath123 layers the reflections at the @xmath0 are andreev ones , while those at the @xmath6 are normal reflections @xcite . the arrows in the fig . [ fig5 ] point along the momenta of the particle segment of the orbit . its @xmath32 component is @xmath36 which labels the state . clearly , the collection of states labelled in this way forms an andreev band . for the normal metal - superconductor heterostructure , andreev bound states ( bands ) are symmetrically placed with respect to the fermi energy . of course the position of these states changes with the thickness of the sample . as we increase the size of the sample , the andreev bound states approach the fermi energy , but they will never reach it . in other words , there is no possibility of crossing the fermi level . somewhat surprisingly , the situation is quite different for the @xmath1/@xmath0 structure @xcite . in this case the bound state energies can cross the fermi level . in short they can become the zero - energy mid - gap states @xcite . this situation is illustrated , for @xmath142 , in the fig . [ fig6 ] , where the bound state energies are plotted against the exchange splitting . if we take into account states for different @xmath36 ( andreev band ) , the situation is more complicated as in general the states with different symmetry of the wave function can mix , and so we are not able to separate andreev bands so easily . the example of such andreev bands rather than states is shown in the fig . [ fig7 ] , where the density of states of the spin up electrons is plotted for a few values of the exchange splitting . as we shall report presently the fact that the andreev bound states can cross the fermi level has remarkable consequences : it leads to the generation of the spontaneous current in the ground state . in fact the situation is rather similar to that of exotic superconductors ( with non-@xmath143-wave order parameter ) , where the surface @xmath73 shows the zero - energy state , leading to the spontaneously generated current . however the origin of this state is quite different , as there it comes from the symmetry related sign change of the order parameter at the surface @xcite . here , in the @xmath1/@xmath0 heterostructure , the zero - energy state comes from the oscillatory behaviour of the pairing amplitude ( @xmath144 ) in @xmath1 , which changes its phase by @xmath4 each time crosses zero . nevertheless it leads to the same effect : a spontaneously generated current . the most remarkable feature of our calculations is that the solution of the @xmath27 equations frequently converges to a solution with the finite current @xmath145 even though the external vector potential is zero . the generation of the spontaneous current is strictly related to the crossing of the andreev bound states through the fermi energy . in the fig . [ fig7 ] , we have plotted the @xmath73 for different values of the exchange splitting for calculations where the current was allowed to develop in the @xmath32 - direction . for the @xmath146 we observe that spontaneous current is generated in the ground state . we have checked that the solution with the current flowing is the true ground state , as it has energy lower than solution without the current . physically , in the presence of the current , the zero energy states split because of the momentum dependence of a @xmath147 contribution to the quasiparticle energies where the velocity vector is defined by the current @xmath148 . the example of such a doppler splitting in the @xmath73 is shown in fig . [ fig8 ] for the exchange field @xmath146 . we can see that , besides the zero energy state , the other andreev states also are split . it turns out that the splitting of these states is only weakly sensitive to the exchange field @xmath112 , but it strongly depends on the thickness of the ferromagnetic slab and it decreases as the thickness increases . interestingly , we have found a simple relation between the splitting of the zero energy state @xmath149 and the vector potential @xmath92 . this is given by : @xmath150 where @xmath151 is the value of the vector potential @xmath92 averaged over the @xmath1 side of the system only , i.e. @xmath152 . corresponding to the splitting of the zero energy states , when the current flows , we find differences in the @xmath73 at the fermi level . in fact , the difference is quite small . the main behaviour remains the same . the comparison of the density of states at the fermi energy for both cases ( @xmath153 and @xmath154 ) is shown in the fig . [ fig9 ] . the @xmath73 for the @xmath123/@xmath0 structure ( dashed line with circles ) is also drawn . as we can see , the main features remain more or less the same , in particular the period of the oscillations . similarly like the spin polarization ( fig . [ fig2 ] ) and pairing amplitude ( fig . [ fig3 ] ) , the density of states at the fermi level shows decreasing period of the oscillations as the exchange splitting increases . as we already discussed , the zero energy state is responsible for the generation of the spontaneous current in the system . the typical example of such a current , flowing parallel to the @xmath1/@xmath0 interface , ( @xmath155 ) is shown in the fig . [ fig10 ] for a few values of the exchange splitting . evidently , the behaviour of the current , as a function of the layer index , is very similar to the density of states at the fermi level . for example , if we compare the dashed curve with the triangles in fig . [ fig10 ] to the dashed or solid line with triangles in fig . [ fig9 ] , we readily notice the similarity in the oscillating nature of the current and @xmath73 respectively . another important issue is the distribution of the current through the whole trilayer structure . we find that it flows mostly in the positive @xmath32 direction on ferromagnetic side and in the negative direction in the superconductor . notably the total current , integrated over the whole sample , is equal to zero within numerical accuracy . this is as it should be for the true ground state and found to be in the @xmath2 state , where the current associated with the unpaired electrons is balanced by the supercurrent flowing in the opposite direction . while the total current in the system is equal to zero , the current integrated over @xmath1 side only ( @xmath156 ) has a finite value , almost independent of the exchange splitting . but that current is very sensitive to the thickness of the @xmath1 slab , and for example for @xmath157 @xmath158 ( @xmath159 per layer ) , while for @xmath160 we have @xmath161 ( @xmath162 per layer ) . this suggests that effect associated with the spontaneous current is very important in the samples of the small thicknesses and it seems to play minor role if the size of @xmath1 part of the system is large . obviously , the above current distribution should lead to the generation of the magnetic flux through the sample . in fact we have found that again , as the current , the magnetic flux bounded by the layers only weakly depends on the exchange splitting but it does change with the distance from the interface . for @xmath157 the magnetic flux @xmath163 is nearly equal to the half of the flux quantum @xmath164 , @xmath165 . this gives a flux per plaquette , associated with the layers @xmath34 and @xmath166 , @xmath167 . on the other hand , for @xmath160 we have @xmath168 and @xmath169 respectively . the typical layer dependence of the spontaneous magnetic flux @xmath170 ( magnetic flux through a plaquette ) is shown in the fig . [ fig11 ] . we can see that while there are small differences in the behaviour on the @xmath1 and @xmath0 sides of the sample , related to the different current distributions ( see the fig . [ fig10 ] ) , the total magnetic flux @xmath171 is roughly the same and more or less independent of the exchange splitting . as we have recalled above , in the bulk @xmath2 state the total current vanishes . moreover , the quasiparticle current @xmath172 is exactly equal to the supercurrent @xmath173 ( related to the @xmath174 in eq . ( [ hfg ] ) ) with opposite sign . this means that in a bulk layered superconductor , these two currents cancel each other layer by layer , i.e. @xmath175 . evidently this is the case because the exchange field and superconducting pairing potential are present in the whole system . by contrast , in the case of the @xmath1/@xmath0 heterostructure , the exchange field and the pairing potential are spatially separated . in this case @xmath176 is not equal to @xmath177 within each layer , but it is when integrated over the whole sample , i. e. @xmath178 thus the lack of exact cancellation of the current layer by layer leads to a finite current on both sides of the interface , but zero in the whole sample . an example of such behaviour is shown in fig . [ fig12 ] . we see , that the quasiparticle part of the total current ( solid line ) is almost in anti - phase to the supercurrent ( dashed line ) . moreover , on the superconducting side , the current is carried mainly by quasiparticles . this is rather surprising since usually one expects current to be related to the cooper pairs . however due to the proximity of the ferromagnet , there are also normal particles in the superconductor near the interface . these certainly contribute to the total current , like in the bulk @xmath2 state . in our system , only the cooper pairs close to the interface can feel the exchange field of the ferromagnet and have moving center of mass . note that magnetization goes to zero very quickly as we go away from the interface ( see fig . [ fig2 ] ) . clearly this makes the supercurrent equal to zero except within a few @xmath0 layers . interestingly we also found that whenever the pairing amplitude @xmath124 changes sign , the local quasiparticle current as well as the local supercurrent becomes large . in other words , the maxima of the @xmath176 and @xmath179 correspond to the zeros of the pairing amplitude @xmath124 . as we have mentioned above , the ground state does not carry current for all values of the exchange splitting . we have shown , that whenever there is a current flowing , there are the andreev bound states crossing the fermi level . also we have found another quantity which is related to the current flowing . namely , the pairing amplitude at the @xmath1/@xmath6 interface ( i.e. @xmath180 - see fig . [ fig1 ] ) . if @xmath181 changes its sign , the spontaneous current is generated . this happens exactly for the same values of the exchange splitting , for which a band of the andreev bound states crosses the fermi energy . moreover , the presence or absence of the current has a dramatic effect on the pairing amplitude @xmath181 . this can be seen in the fig . [ fig13 ] , where @xmath181 is plotted as a function of the dimensionless parameter @xmath182 . as implied by eq.[delta ] , @xmath124 is the real part of a complex order parameter which has an amplitude @xmath183 and a phase @xmath184 . clearly , fig . [ fig13 ] can be red as @xmath185 flipping from @xmath186 where @xmath181 has the same sign as @xmath144 in @xmath0 , to @xmath187 , where @xmath181 has the opposite sign . evidently , the pairing amplitude is pinned to zero by the spontaneous current in the system . introducing the dimensionless quantity @xmath188 we can say , that system is in a @xmath4-like phase ( for odd @xmath188 ) or in the @xmath17-phase ( even @xmath188 ) , in close analogy to studies on the @xmath4-effect in the @xmath0/@xmath1/@xmath0 structure by chtchelkatchev et al . @xcite . in ref . @xcite , we have shown a similar picture for the thickness of the @xmath1 slab equal to @xmath189 layers . here we show the corresponding results for @xmath190 layers . to be precise in the present case @xmath188 has slightly different value with prefactor @xmath191 ( not @xmath192 as it was for @xmath189 layers case ) . this discrepancy may come from the fact , that effective exchange splitting is little bit smaller than its bulk value ( see discussion in the paragraph [ subsec3.a ] ) . as our system consists of ferromagnet and superconductor , we can expect the spontaneous current to be polarized . however , up to now , we have presented results for the special case of the particle - hole symmetry ( @xmath193 or @xmath194 ) and we found that the spontaneous current was , in fact , not spin - polarized . the reason is as follows . within linear response theory , the total current can be divided into two parts : a diamagnetic one - giving the response of the bulk density , and a paramagnetic one , which is proportional to the density of states at the fermi energy , as it comes from the deformation of the wave function at the fermi surface @xcite . now if we note , that for the particle - hole symmetric case ( @xmath193 ) , the spin up @xmath195 and spin down @xmath196 @xmath73 at the fermi energy are equal , we are not surprised that the polarization of the current is equal to zero . as we go away from the half filling , the difference between spin up and spin down @xmath73 ( @xmath197 starts to play a role as it leads to the polarization of the current . a typical example of the polarization of the current @xmath198 is shown in the fig . [ fig14 ] for concentration of electrons @xmath199 and @xmath200 . to make the two calculations comparable we have adjusted the parameters so the value of the @xmath0 order parameter would be the same . this was achieved by shifting of the coupling constant @xmath201 to @xmath202 for @xmath203 , @xmath204 for @xmath205 as well as chemical potential to @xmath206 and @xmath207 respectively . the total polarization of the current @xmath208 is of order @xmath209 and does not depend much on the concentration of electrons @xmath210 . moreover , we have found a very interesting property of this current , namely , the sign of the polarization of the current depends on whether @xmath211 or @xmath212 . this means , that for @xmath213 we get the same results as for @xmath214 , except of the sign of the polarization of the current . this effect can be explained very easily if we recall that polarization of the current is proportional to the difference in the @xmath73 at the fermi energy , and one can check , that @xmath215 is an antisymmetric function under the shift @xmath216 , i.e. @xmath217 . clearly , the important issue raised by the above results is `` how to observe experimentally the predicted current ? '' but first of all let s estimate how big this current is . if we assume a bandwidth equal to @xmath218 , the average current per layer is of order of tenths of @xmath219 . in general the polarization of the ferromagnet can be large , however it can not be equal to @xmath220 . there must be a small number of the minority electrons . in other words it does nt matter how big the ratio @xmath221 is , we should always observe the current . the next point is the thickness of the @xmath1 region . namely , the current is bigger for a thinner sample , but even for thickness of order of several hundreds of @xmath129 , we expect the current to be easily detectable . clearly , the presence of the current should manifest itself in a number of experiments . for example , since there is a net magnetic flux associated with the current , it could be detected by @xmath222 experiments . another possibility is to measure the conductance across the @xmath1/@xmath0 junction @xcite . in such experiments the splitting of the zero energy state in the @xmath73 is expected to be observed and , as indicated by eq.[splitting ] the energy distance between the split peaks measures the current flowing . finally one can also use some local probes like @xmath223 techniques to measure the layer resolved @xmath73 . confirming that the current is polarized is more troublesome . as we have shown , to get polarized current the system has to be away from the @xmath193 point . in our case , to avoid a fermi energy mismatch , both the @xmath1 and @xmath0 are moved from @xmath193 . fortunately , this is only necessary for the ferromagnet . the superconductor can be in a particle - hole symmetric state . in short , realistic estimations of the spin polarization of the spontaneous current require more material specific calculations . in conclusion , we have studied properties of the insulator ( or vacuum ) - ferromagnet - superconductor trilayer . we have shown , that such a structure supports andreev bound states forming andreev bands , whose position can be tuned by thickness of the sample or exchange splitting . moreover , when a band crosses the fermi energy , a spontaneous current ( and magnetic flux ) is generated in the ground state . we have found a relation between the pairing amplitude in the ferromagnet , which has oscillatory behaviour , and the andreev bound states . namely , when the andreev band crosses the fermi energy the pairing amplitude at the surface opposite to the @xmath1/@xmath0 interface changes sign , and as long as current flows it is pinned to zero . the polarization of the current strongly depends on the band filling and is related to the difference in the spin up and spin down density of states at the fermi level . we also have discussed the question of the supercurrent vs quasiparticle current and gave some experimental clues on how to observe that current . the presented results can be , and we have argued that they are , attributed to the @xmath2 effect influenced andreev bound states and can be verified experimentally . p. m. tedrow , r. meservey , phys . rep . * 238 * , 173 ( 1994 ) ; v. t. petrashov , v. n. antonov , s. v. maximov , r. s. shaikhaidarov , jetp lett . * 59 * , 551 ( 1994 ) ; v. t. petrashov , i. a. sosnin , i. cox , a. parsons , c. troadec , j. low temp * 118 * , 689 ( 2000 ) ; h. -u . habermeier , g. cristiani , r. k. kremer , o. lebedev , g. van tendeloo , physica * c364 - 365 * , 298 ( 2001 ) . n. f. berk , j. r. schrieffer , phys . * 17 * , 433 ( 1966 ) ; c. pfleiderer , m . uhlarz , s. m. hayden , r. vollmer , h. v. lohneysen , n. r. bernhoeft , g. g. lonzarich , nature * 412 * , 58 ( 2001 ) ; d. aoki , a. huxley , e. ressouche , d. braithwaite , j. flouquet , j. -p . brison , e. lhotel , c. paulsen , nature * 413 * , 613 ( 2001 ) . g. e. w. bauer , yu . v. nazarov , d. huertas - hernando , a. brataas , k. xia , p. j. kelly , materials sci . eng . * b84 * , 31 ( 2001 ) ; s. oh , d. youm , m. r. beasley , appl . . lett . * 71 * , 2376 ( 1997 ) ; l. r. tagirov , phys . rev . lett . * 83 * , 2058 ( 1999 ) . h. k. wong , b. y. jin , h. q. yang , j. b. ketterson , j. e. hilliard , j. low temp . phys . * 63 * , 307 ( 1986 ) ; j. s. jiang , d davidovi , d. h. reich , c. l. chien , phys . lett . * 74 * , 314 ( 1995 ) ; v. mercaldo , c. attanasio , c. coccorese , l. maritato , s. l. prischepa , m. salvato , phys . rev . * b53 * , 14 040 ( 1996 ) ; th . mhge , n. n. garifyanov , yu . v. goryunov , g. g. khaliullin , l. r. tagirov , k. westerholt , i. a. garifullin , h. zabel , phys . * 77 * , 1857 ( 1996 ) ; c. l. chien , d. h. reich , j. magn . magn . mater . * 200 * , 83 ( 1999 ) . v. proki , a. i. buzdin , l. dobrosavljevi - gruji , phys . * b59 * , 587 ( 1999 ) ; a. kadigrobov , r. i. shekhter , m. jonson , z. g. ivanov , phys . rev . * b60 * , 14 593 ( 1999 ) ; l. dobrosavljevi - gruji , r. ziki , z. radovi , physica * c331 * , 254 ( 2000 ) ; m. zareyan , w. belzig , yu . v. nazarov , phys . lett . * 86 * , 308 ( 2001 ) .

galaxy clusters are the largest known gravitationally bound systems in the universe . the detailed analysis of the mass distribution of clusters is thus important in the process of understanding the large scale structure , and the nature of dark matter . the three main methods of measuring galaxy cluster masses : virial masses from velocity dispersions of cluster galaxies , x - ray imaging and spectroscopy of the intra - cluster medium ( icm ) emission , and the gravitational lensing of background galaxies , have been found in recent years to be in disagreement for some clusters . generally , the x - ray estimates are in good agreement with gravitational lensing for clusters with a high concentration of central x - ray emission ( the so - called `` cooling flow '' clusters ) but seemingly in disagreement for other , less centrally peaked objects @xcite . to obtain the estimate of the total mass ( including that due to dark matter ) of a galaxy cluster from its x - ray emission commonly assumed to be from optically thin , hot plasma that subtends the space between galaxies it is necessary to make the assumption of hydrostatic equilibrium . this is appropriate of course only for clusters that have had time to relax into equilibrium and have not experienced any recent merger events . generally , it is assumed that clusters with circular isophotes meet this criterion . cold dark matter ( cdm ) hierarchal clustering is the leading theory describing the formation of large scale structure quite well . in particular , the numerical simulations such as @xcite ( nfw ) successfully reproduce the observed dark matter halo profiles , which appear be independent of the halo mass , initial power spectrum of fluctuations , and cosmological model . however , observations often disagree with the numerical models . for instance , one disagreement regards the rotation curves of dwarf elliptical galaxies which appear to be the result of a constant - density core whereas numerical simulations predict cuspy dark matter halo profiles @xcite . in addition , observations show fewer milky way satellites than predicted by the models @xcite . clearly , to understand the nature of galaxy clusters and the dark matter they consist of , it is important to measure the matter distribution in clusters via all available means . fortunately , there are two superb x - ray observatories , chandra and xmm - newton , featuring excellent angular resolution and exceptional effective area coupled with good spectral resolution , and those are well suited for detailed analysis of the x - ray emitting gas of galaxy clusters abell 1689 is a cluster showing a large discrepancy among various mass determinations , and we chose it for a detailed study . it is a rich cluster , @xmath5 , without a pronounced cooling flow but with an approximately circular surface brightness distribution suggesting a relatively relaxed structure . the large mass , @xmath6 @xcite , and apparent symmetric distribution of abell 1689 make it a suitable cluster for gravitational lensing measurements as well as x - ray measurements . however , the type of clusters that are believed to be the most relaxed have a cool central component of enhanced surface brightness . the absence of such a component in abell 1689 suggests that the cluster is not fully relaxed . also , the galaxy content of the cluster is unusually spiral rich for a cluster with high spherical symmetry and richness , with a galaxy type ratio e : s0 : sp of approximately 22 : 22 : 28 plus 25 unidentified galaxies @xcite . @xcite measure a velocity dispersion of @xmath7 for 66 cluster members , unusually high for a cluster of this temperature . positions and redshifts from @xcite for all cluster members with red magnitude @xmath8 and within @xmath9 of the brightest central galaxy are shown in fig . [ figoptical ] together with logarithmic x - ray intensity contours from xmm mos 1 . the gravitational lensing estimate from 6000 blue arcs and arclets calibrated by giant arcs at the einstein radius of abell 1689 gives a best fit power - law exponent of @xmath10 for the projected density profile from @xmath11 to @xmath12 @xcite . ( unless otherwise stated , we assume an einstein - desitter ( eds ) cosmology with @xmath13 and @xmath14 . ) this is steeper than the profile of an isothermal sphere @xmath15 . the strong lensing analysis of two giant arcs directly gives @xmath16 @xcite . the mass profile derived from the deficit of lensed red galaxies behind the cluster due to magnification and deflection of background galaxies suggests a projected mass profile of @xmath17 for @xmath18 , close to that of an isothermal sphere @xmath19 . the mass interior to @xmath20 from this method is @xmath21 @xcite . measurement of the distortion of the luminosity function due to gravitational lens magnification of background galaxies gives @xmath22 @xcite . finally , the weak gravitational shear of galaxies in a eso / mpg wide field imager @xmath23 image gives a mass profile with best fit nfw profile with @xmath24 and @xmath25 or a best fit sis with @xmath26 @xcite . there is a good indication from the optical data that the cluster consists of substructures . @xcite suggest a strong lensing model with two clumps in order to reproduce the positions of the brightest arcs . a larger mass clump @xmath27 is centered on the brightest cluster galaxy while a smaller clump @xmath28 is located @xmath29 northeast of the main clump . they arrive at a mass a factor 2 - 2.5 lower for their x - ray estimate compared to their gravitational lensing model . @xcite identify two distinct substructures centered on redshifts @xmath30 and @xmath31 using positional and redshift data of cluster galaxies from @xcite , providing further evidence that the cluster is not relaxed . these clumps are also aligned in the southwest northeast direction but the locations do not agree with the ones of @xcite . both structures are found to have @xmath32 yielding virial masses several times smaller than those derived from lensing and x - ray estimates . can the x - ray observations provide any evidence for a substructure in abell 1689 , and what are the implications on the mass inferred from the x - ray data ? in an attempt to answer this , we analyze xmm - newton epic pn and epic mos data to measure the mass profile of a1689 and to investigate the spatial structure of the cluster . 2 contains the details of the observations and data reduction with the xmm - newton as well as with summary of the rosat , asca and chandra data ; 3 covers the methods of spectral fitting of the xmm data and presents the analysis of cluster asymmetry ; 4 derives the mass and the slope of the mass distribution in the core ; and 5 presents the inferences about the structure of the cluster inferred from the spatial analysis . the paper concludes with the summary in 6 . abell 1689 was observed with xmm - newton for 40 ks on december 24th 2001 during revolution 374 . for imaging spectroscopy we use data from the european photon imaging camera ( epic ) detectors mos1 , mos2 and pn . both mos cameras were operating in the full frame mode whereas pn was using the extended full frame mode . the extended full frame mode for pn is appropriate for studying diffuse sources since it has lower time resolution and so it is less sensitive to contamination from photons being detected during readout of the ccds . these events ( so called out - of - time events ) show up as streaks across the x - ray image and are especially bothersome when the goal of an observation is spatially resolved spectroscopy of diffuse sources . all cameras used the thin filter during the observation . epic background is comprised mainly of three components . the external particle background consists primarily of soft protons @xmath33 being funneled through the mirrors and causing a time variable flaring signal in the detector . the internal particle background is mainly due to high energy particles interacting with the detector material and causing a roughly flat spectrum with flourescent emission - lines characteristic to the detector material . this component varies over the detector surface . the third source of background is the cosmic x - ray background ( cxb ) which is roughly constant in time but varies over the sky . for all data reduction we use the software and calibration data implemented in xmm science analysis software ( sas ) 5.4.1 . to exclude the events contaminated by proton flares , we produce light curves in the @xmath34kev band where the true x - ray signal is low . we screen the data using a constraint on the total count rate of less than 1.5 ct s@xmath35 for mos and 1.1 ct s@xmath35 for pn in this band , leaving an effective exposure time of 37 ks for mos and 29 ks for pn . this screening corresponds to a limit on the count rate of approximately @xmath36 above the quiescent period in the @xmath37kev band . the effective mirror collecting area of xmm - newton is not constant across the field of view : it decreases with increasing off - axis angle and this decrease is energy dependent . this results in a position dependent decrease in the fraction of detected events and when doing imaging spectroscopy for extended sources , we need to account for this effect . by generating an ancillary response file ( arf ) for each source spectrum region , using xmm sas 5.4.1 command ` arfgen ` we calculate an average effective area for each region considered by us ( see below ) . in order to correctly account for particle - induced and cosmic x - ray background , it would be optimal to extract a background spectrum from the same detector region collected at the same time as the source spectrum . this is of course impossible , and the background can be taken from a source - free region of the detector ( other than the target , but near it ) , or can be estimated using blank sky data . we adopt the former method , noting that the background is not entirely constant over the field of view ; however , it can be assumed to be approximately constant with the exception of the fluorescent cu line at @xmath38 in pn which is the strongest contaminant emission line . using this assumption , we can effectively subtract the particle background by extracting a spectrum in a source free region in the same exposure . for pn data in the @xmath39 range is excluded due to the strong spatial dependence of the cu internal emission . using an arf generated for a source region on a background subtracted spectrum will not take into account that the spectrum used as background was extracted from a different region where the vignetting was higher , due to the larger off - axis distance of the background region . this will have the effect that the cxb component in our source spectrum will be under - subtracted and the net spectrum will contain some remnant of the cxb . the particle induced background however is not vignetted and therefore should leave no remnant . to estimate the cxb component in our exposure we take the events outside of the field - of - view to be the particle background . another spectrum is then extracted from a large source free area , located away from the cluster . the particle background is normalized and removed from this spectrum . the resulting spectrum is fitted to a broken power law model to determine the cxb flux . the incorrect vignetting correction of the cxb is found to cause an at most 2 % over - estimation of the flux which is the case in the outermost annulus in our analysis ( see below ) . however since the vignetting is energy - dependent , the incorrect vignetting may cause a small ( @xmath40 kev ) shift in temperature for the outermost region . the overall uncertainty in the background is estimated to be at most a few percent . we compare our background subtraction method with the method of using xmm standard blank sky data compiled by @xcite ; we find that both methods give similar fit parameters . however , given the difficulty of normalizing the cxb , the different high - energy leftovers from proton - flare subtraction and the different internal particle background which occurs when using background data from a different exposure , we chose to use in - field ( rather than blank - sky ) background . we find this is more robust in keeping the overall shape of the source spectrum uncontaminated . in all our analysis we use a background region of a circular annulus with inner and outer radii of @xmath41 and @xmath42 respectively . in fact , recent work by @xcite suggests that the in - field background method is probably more accurate . the clusters in the above paper , however , have smaller apparent angular size which makes the background method more reliable . in analyzing the vignetting corrected , background subtracted radial count - rate profile , we fit it to a conventional beta model , @xmath43 , where @xmath44 is the source surface brightness at radius @xmath45 . the fit gives @xmath46 and @xmath47 with @xmath48 using data out to @xmath49 . clearly this model is not a very good fit ; we show it here only for completeness and comparison with previous work , and note that it is not used in the subsequent analysis . the bad fit above results from the fact that the cluster emission is more peaked in the core than the best fit beta model . to obtain a general idea of the properties of the cluster and compare this with previous results , we extracted spectra for the central @xmath50 @xmath51 region , centered on the x - ray centroid at @xmath52 @xmath53 . this radius corresponds to @xmath54 or @xmath55 . background spectra were extracted from source - free regions from the same exposure . for xmm pn we use single and double pixel events only whereas for xmm mos we also use triple and quadruple pixel events . for spectral fitting we use the xspec @xcite software package . we fit the data in the @xmath56 range using the mekal @xcite model for the optically thin plasma and galactic absorption . with the absorption fixed at the galactic value , @xmath57 @xmath58 @xcite , and assuming a redshift @xmath59 we arrive at a temperature of @xmath60 and a metal abundance of @xmath61 solar for both mos cameras with @xmath62 . from the pn camera we get @xmath63 and @xmath64 solar with @xmath65 . the best fit models with residuals can be seen in fig . [ mospn ] for mos ( left ) and pn ( right ) . the pn temperature is in disagreement with mos data and the reason for this effect can be seen from the residuals below 1 kev for pn ( fig . [ mospn ] ( right ) ) . fitting the pn data above 1 kev we get @xmath66 and @xmath67 solar with @xmath68 , in better agreement with mos . the temperature from mos is in agreement with that found by asca , @xmath69 @xcite . this observation does not suffer from pile - up , nor is the low energy spectrum sensitive to background subtraction . the background uncertainties below 1 kev for this pn spectrum are less than 1 % whereas the pn soft excess is sometimes higher than 10 % . the excess is certainly not background related . it is possible that the pn low - energy discrepancy can be due to incorrect treatment of charge collection at lower energies ( s. snowden , priv . comm . ) . to resolve the discrepancy between mos and pn we decided to compare with asca gis / sis , rosat pspc and chandra acis - i data . besides the discrepancy regarding the softest x - ray band for the abell 1689 data between the mos and pn detectors aboard xmm - newton , the spectral fits to the chandra data presented by @xcite imply a higher absorbing column , @xmath70 @xmath58 than the galactic value of @xmath71 @xmath58 @xcite . since such excess absorption is not commonly detected in x - ray data for clusters , this requires further investigation . to determine if there is indeed any additional component of absorption beyond that attributable to the galactic column and assess the reliability of the softest energy band of the pn vs. mos data we used the most sensitive soft ( @xmath72 kev ) x - ray data for this cluster obtained prior to the xmm - newton observations , collected by the rosat pspc . the rosat observation conducted during july 18 - 24 , 1992 ( available from heasarc ) yielded 13.5 ks of good data . we extracted the rosat pspc counts from a region @xmath73 in radius , centered on the nominal center of x - ray emission . for background , we selected a source - free region of the same pspc image . using these data over the nominal energy range 0.15 - 2.1 kev with the standard pspc response matrix applicable to the observation epoch , we performed a spectral fit to a simple , single - temperature mekal model with soft x - ray absorption due to gas with solar abundances , using the xspec package as above . in the fit we use metal abundances of 0.27 solar obtained in the xmm mos fit above . while the limited bandpass of rosat pspc precludes an accurate determination of the temperature ( the best fit value is @xmath74 kev , 90% confidence regions quoted ) , the pspc data provide a good measure of the absorbing column : the best value is @xmath75 @xmath58 , certainly consistent with the galactic value . we note that the difference between the temperature inferred from the pspc fit and that obtained from the xmm - newton data as above is a result of the limited bandpass of the pspc , located much below the peak of the energy distribution of the cluster photons . the measurement of the absorbing column , however , clearly indicates that the column inferred from the chandra observation by @xcite is not correct , and might be due to instrumental effects . we conclude that the absorbing column is consistent with the galactic value . to obtain further constraints on the absorbing column , we also used the asca gis and sis data together with the pspc data for an independent constraint on the continuum radiation in the fitting process . we performed standard extraction of data from all asca detectors , also from a region from a region @xmath73 in radius for the source , and a source - free region of the same image for background . we performed a spectral fit simultaneously to data from the pspc and four asca detectors . to account for possible flux calibration differences , we let the normalization among all the different detectors run free . we used the energy range of @xmath76 and @xmath77 for the asca gis and sis cameras respectively . since asca detectors ( and in particular , the siss ) often return spectral fits with excess intrinsic absorption @xcite we also let the absorbing column be fitted independently for the gis , sis and pspc detectors . temperature and metal abundances were tied together for all datasets in the fit . the optical redshift @xmath0 was used . the joint fit of rosat and asca data gives us the best fit temperature of @xmath78 kev and abundances of @xmath79 solar , in agreement with the values quoted by @xcite , and the absorption ( for the pspc data ) @xmath80 @xmath58 , in agreement with the galactic hi 21 cm data . we shall use this value for the absorption in the subsequent analysis . this best fit rosat / asca model was compared with the data from the same region in the xmm - newton cameras giving an unfitted reduced @xmath81 of 1.39 for pn and 1.06 for both mos cameras combined . the ratio of these spectra to the rosat / asca model can be seen in fig . [ mospndisc ] . from this result we conclude that mos low energy response is more consistent with previous data and subsequently , we choose to ignore all pn data below @xmath82 . re - fitting the xmm data from the above region using @xmath83 for mos data and @xmath84 for pn leaving the absorbing column as a free parameter yields @xmath85 @xmath58 , a temperature of @xmath86 and metallicity of @xmath87 solar . while this fitted value of absorption is formally inconsistent with the galactic and rosat - inferred values , this is a relatively small difference , which might be due to the slightly imperfect calibration of the mos detectors or the assumption of isothermality made by us for this fit ( since @xmath88 and @xmath89 are correlated in the fitting procedure ) . we note that using the rosat value for absorption will give us a somewhat lower measure on the temperature ( see below ) . finally , we reduced the chandra data for abell 1689 using the most recent release of data reduction software ; specifically , due to the degradation of the chandra acis - i low - energy response correction for the charge transfer inefficiency ( cti ) is necessary , and we applied this to the chandra data . as of chandra data analysis software package ciao ver . 2.3 this correction can be applied in the standard event processing . it is also necessary to account for the acis excess low - energy absorption due to hydrocarbon contamination . we used the ` acisabs ` code for the correction to the auxiliary response function . the event grades used in the acis analysis were grade=0,2,3,4 and 6 . the cluster was observed with the chandra acis - i detector array at two separate occasions for @xmath90ks each on 2000 - 04 - 15 and 2001 - 01 - 07 . spectra were extracted for the central @xmath50 region centered on the x - ray centroid at @xmath52 @xmath53 following the ciao 2.3 science threads for extended sources . [ chandra ] shows the ratio of the combined chandra data to the best fit model determined from rosat and asca above . the left spectrum is before and the right after the ` acisabs ` correction . fitting the corrected spectrum to an absorbed mekal model in the energy range 0.3 - 7.0 kev yields absorption of @xmath91 @xmath58 , a temperature of @xmath92 kev and abundances of @xmath93 solar using the optical redshift z=0.183 . the discrepant absorption found by @xcite ( using data 0.7 - 9.0 kev ) is apparently corrected for by ` acisabs ` and the value is in agreement galactic absorption . however there is still a large discrepancy in temperature , which also can be seen from the high energy ends of fig . [ chandra ] . part of this effect can possibly be attributed to the high energy particle background but more likely to uncorrected instrumental effects . repeating all above steps for single pixel events only ( grade=0 ) in order to achieve higher spectral accuracy gives us a best fit temperature of @xmath94 kev . we can not account for the differences between the two chandra data sets ( using grade=0 vs. grade=0,2,3,4 and 6 events ) nor between the results of the chandra and xmm spectral fits . we note here that the photon statistics resulting from the xmm observation is superior to that in the chandra data , and since our analysis does not require the superior angular resolution of the chandra mirror , we limit the analysis below to the xmm data . to obtain a radial profile of cluster gas temperature , abundance , and density , we first make the assumption that the cluster is spherical and that above properties are only functions of radius . for this , we divide the image of the cluster into 11 concentric annuli out to @xmath95 @xmath96 centered on the x - ray centroid . for each annulus , we extract spectra from all epic cameras , and we set the inner and outer radii of each region by requiring that each annulus contains at least 9000 counts per each mos camera and 13000 for pn . this allows us to derive a reliable estimate of temperature in each region . point sources with intensities greater than @xmath97 over the average are excluded . the outer radii of the annuli are as follows : @xmath98 , @xmath99 , @xmath100 , @xmath101 , @xmath102 , @xmath103 , @xmath104 , @xmath105 , @xmath106 , @xmath107 and @xmath108 . average cluster properties were determined using all annular spectra simultaneously out to @xmath109 ( @xmath110 , @xmath111 ) using the same spectral fitting procedures as in section [ fitting ] . we use the optical redshift of @xmath59 @xcite and the line of sight absorption of @xmath112 as measured from rosat data , and also consistent with the galactic column density ( section [ rosasc ] ) . the mean temperature of the cluster is found to be @xmath113 and the mean abundance @xmath114 solar . we note that leaving the redshift as a free parameter gives a best fit redshift @xmath115 ( 90% confidence range ) which is considerably less than measured via optical observations . considering mos and pn data separately gives @xmath116 and @xmath117 respectively . in order to take into account the three - dimensional nature of the cluster we consider the spectrum from each of the 11 annuli to be a superposition of spectra from a number of concentric spherical shells intersected by the same annulus . the spherical shells have the same spherical radii as the projected radii of the annuli . we assume that each shell has a constant temperature , gas density and abundance . the volume for each annulus / shell intersection is calculated to determine how large a fraction of emission from each shell should be attributed to each annulus . assigning a spectral model to each spherical shell we can simultaneously fit the properties of all spherical shells . in practice we use the method of @xcite where we have a matrix of @xmath118 mekal models with absorption in xspec . in the process of fitting the data in xspec , each datagroup ( consisting of three data files : pn , mos1 and mos2 spectra for each annulus ) is fitted using the same set of models . for the central annulus ( the one with zero inner radius so it s actually a circle ) , which will intersect all 11 spherical shells , we will need to apply 11 mekal models to represent these . this means that we have to apply 11 mekal models to each datagroup ( annulus ) with absorption where each model represents the properties of one spherical shell . the normalization of each shell model is set to be the ratio of the volume that shell occupies in the cylinder that is the annulus / shell intersection to the volume it occupies in the central annulus . abundance and temperature are tied together for the models representing the same spherical shell . of course not all annuli intersect every shell , and for those shells not intersected by the annulus to which they are attributed , the model normalization will be zero . the matrix of mekal models is thus triangular and can be fitted directly to the spectra we have extracted from annular regions in the data . this approach allows for all data to be fitted simultaneously , and we do not have the problem with error propagation which occurs when starting to fit the outermost annulus and propagating inward subtracting contributions from each previous shell . the metallicity profile ( fig . [ abun ] ) from the deprojection shows signs of increasing abundance toward the center of the cluster . in temperature [ temps ] ) we find an apparent decrease for large radii ( @xmath119 ) , an effect that has been seen in analysis of other clusters with xmm ( see e.g. @xcite ) . gas dynamic simulations of the formation of galaxy clusters also show a decline of temperature at large radii @xcite . we do not find a significant cooling in the cluster core with the highest temperature ( @xmath120 ) near the core radius : in fact , we will show in section [ xyana ] that the temperature profile is not symmetric around the cluster center . we note that for completeness , we also performed the above analysis with the best - fit value of the redshift inferred from the x - ray data alone , and while the exact values of temperature and elemental abundances are slightly different , about 0.2 kev lower for temperature , the general trends in the radial runs of the parameters are the same . the limited point spread function ( psf ) of the xmm mirrors is a potential problem especially for the annuli located close to the center since those are not much larger than the psf fwhm of @xmath121 . some of the flux incident on the central ( circular ) region will be distributed over the outer annuli and vice versa . this flux redistribution will have the effect of smoothing out the measured temperature profile since all annuli will have some flux that originally belong in other annuli . this effect has been studied in detail by e.g. @xcite who find that correcting for the psf redistribution gives a profile that is consistent with an uncorrected profile . abell 1689 has a temperature profile without large temperature variations and no large central flux concentration . we conclude that the effect of flux redistribution in our case will be small and we do not attempt to correct for this . the psf is also weakly energy dependent , and to quantify its possible effect on the observed temperature profile , we calculate the energy dependent flux loss from the central annulus and the effect on the central temperature . the difference in flux loss between various energy bands ( ranging from 1.5 to 7.5 kev ) for the on - axis psf is @xmath122 . we find that for a cluster with an assumed temperature of 9 kev , this could give an error of the central temperature by at the most @xmath123 . we note that this is a maximum difference since in practice , the flux gained from outer annuli could somewhat reduce this effect by working in the opposite manner . the luminosity of the cluster in the @xmath124 band , calculated from the best fit model above ( with @xmath59 ) gives @xmath125 for an eds ( @xmath126 , @xmath127 ) universe or @xmath128 for a cdm ( @xmath129 , @xmath130 ) universe . this corresponds to a bolometric luminosity of @xmath131 or @xmath132 . all above values should include 10% as the absolute calibration error of xmm . from chandra analysis , @xcite find @xmath133 whereas @xcite find @xmath134 from asca data , both in agreement with our results . this does not provide any new information regarding the location of abell 1689 in the luminosity - temperature @xcite relation and it is still in a close agreement with the trend suggested by other clusters . llll x - ray ( rosat ) & @xmath135 & @xmath136 & 1 + x - ray ( chandra ) & @xmath137 & @xmath138 & 2 + lensing & @xmath139 & @xmath140 & 1 + optical & @xmath141 & @xmath142 & 3 + x - ray ( xmm ) & @xmath52 & @xmath53 & 4 + with the good quality xmm data , it is possible to verify the result of @xcite that there is no discrepancy between the optical and x - ray centers of the cluster . we determine the center of x - ray emission for abell 1689 using ximage command ` centroid ` . we also include a measurement of the lensing center from @xcite and the x - ray measurement from rosat by @xcite ( table [ centers ] ) . we find that all values agree within @xmath143 except the rosat estimate , the offset of which we attribute to uncertainties in rosat hri astrometry . this apparently perfect agreement among x - ray , lensing , and optical centers leads us to conclude that the icm density peak and the central dominant galaxy is probably located at the bottom of the dark matter potential well . still , the apparent non - uniformity of the icm radial temperature distribution as well as the offset of optical and x - ray redshifts prompted us to analyze the spatial structure of the cluster . in fig . [ btemps ] we show the spatial temperature distribution . spectra were extracted in rectangular regions and fitted using the same method as in section [ fitting ] . the temperature in the figure scales linearly from 6 kev ( black ) to 10 kev ( white ) . the errors on the temperature of the inner 16 regions are @xmath144 , while the errors on the outer 8 regions are @xmath145 . we clearly see an asymmetry in the temperature around the cluster center with an overall increase toward the northeast . to check the consistency of these results and to increase our accuracy , we re - group the data in larger spatial regions and perform a fit using the same above procedure . we first fit the data keeping the redshift frozen at the optical value . temperatures derived from this fit are shown in fig . [ urdl ] ( left ) . thereafter , we leave the redshift as a free parameter and re - fit the data . the fitted redshifts for these regions are shown in fig . [ urdl ] ( right ) . all errors in fig . [ urdl ] are 90% confidence limits . in the temperature map , we see a clear discrepancy between the northern and southern part of the cluster with a hint of a temperature gradient in the southwest northeast direction . the redshift map reveals a high redshift structure to the east at @xmath146 separated from the rest of the cluster at @xmath147 . analyzing the mos and pn data separately for this high redshift region gives a broad minimum in @xmath81 at @xmath148 for mos whereas pn shows several minima in the @xmath149 range . this region may indicate a subcluster falling inward away from the observer . this is further supported by the optical data , indicating that there are also high - redshift giant elliptical galaxies in this region ( fig . [ figoptical ] ) : it is interesting to speculate if this is actually the remains of a cluster core ? especially intriguing is the apparent coincidence between the smaller subcluster as suggested by strong lensing @xcite and our high redshift gas region approximately @xmath150 northeast of the main cluster . another possibility is that the redshift variation is due to large bulk motions of the intra - cluster gas . it has been shown in gas dynamic simulations that clusters with apparently relaxed x - ray profiles can have complex gas - velocity fields and be far from relaxed @xcite . this kind of motion could give rise to non - thermal emission from shocks etc . in our analysis we can not distinguish between the two possibilities of bulk motion and subclustering . this measurement of non - uniform redshift distribution of the x - ray emitting gas is important , and to verify if it could be due to instrumental effects , we investigated this in more detail . the data in the regions above are from different ccd chips in the pn data whereas for mos all data are from the same chip . since the pn camera provides about half of the data , we want to verify that there are no gain shifts between the ccds , which , if present , could easily cause such an effect . most of the cluster emission is on the ccd chips 4 and 7 . to test for any possible gain offset , we extracted the data from each of the chips individually to verify the position of the internal fluorescent @xmath151 line , with the dominant component at @xmath152 . the spectra in the range @xmath39 are fitted to a gaussian profile : both datasets yield essentially the same peak position of @xmath153 . this corresponds to a possible artificial redshift offset of maximum @xmath154 . to explain the difference in the offset measured by us as an instrumental effect , we would have to have an offset ( say at the fe k line at 6.7 kev ) of 62 ev , and not on the order of 1 ev , as inferred from the cu k instrumental line . hence we conclude that there is no gain shift that could alter our redshift measurements between pn ccds 4 and 7 . according to xmm calibration documentation @xcite the magnitude of calibration errors for pn & mos should be no larger than 10 ev . if we assume that the cluster is spherical with a smooth static gravitational potential and that the intra - cluster medium is a pressure - supported plasma , we can employ the hydrostatic equilibrium equation . the circular x - ray isophotes ( fig . [ figoptical ] ) generally indicate that a cluster is in dynamical equilibrium . the hydrostatic equation can be written as @xcite : @xmath155 where @xmath156 is the enclosed total gravitating mass enclosed within a sphere of a radius @xmath45 , @xmath157 and @xmath158 are temperature and density of the icm at @xmath45 , @xmath159 is the mean particle weight and @xmath160 is the proton mass . using the temperature and normalizations from the spectral deprojection fitting we calculate the total gravitating mass . errors are treated by error propagation . the mass distribution is fitted to a singular isothermal sphere ( sis ) @xmath161 where @xmath162 is the 1-dimensional velocity dispersion , which is used here for a comparison with previous mass estimates . the predicted density profile from cdm hierarchal clustering according to @xcite for dark matter halos is @xmath163 where @xmath164 is the critical energy density at halo redshift @xmath165 and @xmath166 is characteristic density defined by @xmath167 } \label{eqnfw2}\ ] ] where @xmath168 is the concentration of the halo defined as the ratio of the virial radius @xmath169 to @xmath170 , which in turn is a characteristic radius in the nfw model . the critical density at redshift @xmath165 for a flat ( @xmath171 ) universe is @xmath172 \label{rhocrit}\ ] ] where @xmath173 is the hubble constant , @xmath174 and @xmath175 are the current contributions of matter and vacuum energy respectively to the energy density of the universe . more recent numerical studies suggest a steeper core slope and a sharper turn - over from small to large radii @xcite . both models can be generalized as @xmath176^{(\beta - \gamma)/\alpha } } \label{eqgeneral}\ ] ] @xcite , where @xmath177 and @xmath178 characterize the density slope at small and large radii respectively whereas @xmath179 determines the sharpness in the turn - over . most studies agree on @xmath180 but the value of @xmath177 is still being debated . the nfw and moore profiles fit into the parameter space @xmath181 as @xmath182 and @xmath183 respectively . we choose to fit our data to the nfw model ( eq . [ eqnfw1 ] ) which , when integrated over @xmath45 , yields @xmath184 \label{massnfw}\ ] ] where @xmath185 . we find that the data give the best fit for the nfw model with @xmath1 and @xmath186 whereas the sis fit gives @xmath187 . the total mass data and models are shown in fig . [ mass3d ] together with the mass of the x - ray emitting gas @xmath188 . model parameters are summarized in table [ bestfit ] . for a cosmology with @xmath129 and @xmath130 , the best fit nfw model changes to @xmath189 and @xmath190 . another cosmologically important quantity for clusters is the fraction of the total mass that is in the x - ray emitting gas , @xmath191 . @xcite analyze data from the chandra observatory for 10 dynamically relaxed clusters between @xmath192 and @xmath193 and measure an average redshift independent @xmath194 at the @xmath195 radius ( where the total mass density is 2500 times the critical density at the redshift of the cluster ) . the cosmology where this is valid is a flat @xmath196 cosmology with @xmath197 assuming @xmath198 and @xmath199 . the error on @xmath200 above is the rms dispersion of the @xcite sample which is comparable to the individual errors on @xmath200 . with our best fit nfw model in the above cosmology we find @xmath201 for abell 1689 at the @xmath195 radius . this is lower than all the clusters in the @xcite sample and significantly lower than the mean . we find that in our estimate @xmath200 has not converged to a constant and this might help explain part of the discrepancy . however , it may be the case that for many clusters @xmath200 does not converge until well beyond @xmath195 . cccc @xmath202 & @xmath203 & @xmath204 & 7.64 ( 8) + @xmath205 & @xmath203 & @xmath206 & 7.64 ( 8) + sis & @xmath203 & @xmath207 & 61.4 ( 9 ) + pow & @xmath208 & @xmath209 & + pow & @xmath210 & @xmath211 & + comparing the mass and temperature at @xmath195 of abell 1689 to the m - t relation derived for a set of relaxed clusters @xcite shows a low mass for the temperature of abell 1689 . the m - t relation predicts @xmath212 for a @xmath213 cluster where @xmath214 is the hubble constant at the redshift of the cluster . for abell 1689 we find @xmath215 , significantly lower . the above values were derived assuming a flat @xmath196 cosmology with @xmath129 , @xmath130 and @xmath216 . the unusually low mass may be due to the fact that the mass of abell 1689 seems to increase steadily beyond @xmath195 . however , we note that a lower mass ( than would be predicted by the m - t relation ) is not an uncommon feature for unrelaxed clusters ( cf . @xcite ) . for completeness , we note that the calculated total mass includes the icm and galaxy mass contributions as well as the dark matter . the proper way of fitting the nfw model would be to subtract these contributions prior to performing the fit . the nfw model for the actual dark matter halo is not used in this paper ; we only calculate the total mass profile . the mass data were fitted to a simple power law ( @xmath217 ) in the ranges @xmath218 and @xmath219 . we find the best fit of the slope of the matter profile to to be @xmath220 and @xmath221 for small and large radii , respectively . this corresponds to total mass density slopes ( @xmath222 ) of @xmath223 and @xmath224 , in good agreement with what is expected from numerical simulations of cdm hierarchal clustering . we note here that @xcite have measured the density profile of abell 1689 for @xmath225 using chandra data and found @xmath226 . we do not observe a flattening of the core density profile but find a slope close to the core to be somewhere in between the preferred moore and nfw profiles . thus we do not claim to be able to detect nor dismiss any kind of modification to standard cold dark matter . nonetheless , the x - ray data imply an upper limit on the self - interaction of dark matter ; see , e.g. , the discussion in @xcite . such comparisons are more meaningful for cooling - flow clusters which are presumably more relaxed objects . to compare our results with those obtained from gravitational lensing , we reprojected our derived mass distribution into a two dimensional distribution by summing up the contributions from each shell along the line - of - sight . of course this method assumes that the outermost shell is the absolute limit of the cluster , and we recognize that this will not be entirely accurate . therefore we also include our projected best - fit nfw model which has an analytic expression . reprojected mass and nfw model are shown in fig . [ mass2d ] where pluses with error bars are the reprojected mass , the solid line is our nfw model , the triangle is the strong lensing result @xcite , the asterisks are lensing magnification results measured by the distortion of the background galaxy luminosity function @xcite , the dot - dashed lines are lensing magnification results measured by the deficit of red background galaxies @xcite , and the dashed line is the best fit nfw model from weak gravitational shear analysis @xcite . the comparison of our results with measurements from gravitational lensing is summarized in table [ oldres ] . the x - ray mass appears to be in good agreement with that derived from weak gravitational shear but can not be reconciled with the strong lensing data nor with the data from gravitational lens magnification . this discrepancy in lensing was noted earlier by @xcite as well as @xcite who can not find an agreement using any realistic corrections . the high velocity dispersion measured by @xcite and the apparent grouping of galaxies along the line of sight @xcite may indicate that this cluster is not as regular as we expect from its smooth circular surface intensity . the two component model of @xcite also strongly suggests non - uniformity . this might explain the discrepancies between mass estimates from gravitational lensing and from x - rays . a possibility could be that this is a cluster undergoing a major merger , where a sub - clump close to or along the line - of - sight is being stripped from gas or just has very low x - ray luminosity . cccc @xmath227 & @xmath228 & 1 & strong lensing + @xmath229 & @xmath228 & 4 & x - ray + @xmath230 & @xmath231 & 2 & weak lensing magnification + @xmath232 & @xmath233 & 3 & weak lensing magnification + @xmath234 & @xmath233 & 4 & x - ray + to illustrate the possible explanation that a merger might be taking place in this cluster and motivate why the x - ray derived mass should be lower than expected we employ a simple model . we assume that we have two perfectly spherical clusters aligned exactly along the line of sight having identical mass ( @xmath235 ) and gas density ( @xmath236 ) . since the thermal bremsstrahlung emissivity scales as @xmath237 we will measure a surface intensity @xmath238 and from this infer a gas density @xmath239 . in this example however we have @xmath240 and hence we will conclude that @xmath241 . this means that we underestimate the total mass of the x - ray emitting gas by a factor @xmath242 . from the hydrostatic equation ( eq . [ mass ] ) we can write @xmath243 . from the gradient of the measured surface brightness @xmath244 we infer a gradient on the gas density . in our scenario it is true that @xmath245 which yields @xmath246 . for the total mass we will measure @xmath247 . the actual mass of the cluster pair is of course @xmath248 which is twice what we measure . we have underestimated the total gravitating mass by a factor @xmath249 . this is in agreement what we find in our comparison with gravitational lensing derived masses which would be able to measure the total mass accurately in this example . while this exact scenario is not very probable it shows that a close configuration of clusters will underestimate the x - ray mass , maybe by a factor as large as 2 . in the above example the gas mass fraction would actually be overestimated which is the opposite of what we find . however if a merger was in fact taking place the hydrostatic equation would probably be quite inaccurate estimation of the mass . there would be other sources of pressure and probably the cluster would not be in equilibrium : examples here might be magnetic fields , or additional pressure support from non - thermal particles , likely to be accelerated in shocks that arise during a merger . also since we do not detect two separate peaks in the surface intensity map , it is much more likely we have a lower density companion cluster or one merging irregular system . there is a good evidence suggesting that abell 1689 is undergoing a merger . we find that the x - ray measurement yields the value of mass that is low in comparison to gravitational lensing and the m - t relation for relaxed clusters , both by about a factor 2 . we have shown that this may be the result of grouping or elongation along the line of sight . the cluster has a low gas mass fraction compared to other clusters , which is perhaps the result of large scale gas motion . the velocity dispersion of the member galaxies is very high and there are hints of subgroupings in the redshift space . the high number of spiral galaxies is unusual for a rich cluster and suggests a lower density region , where the spiral fraction would be higher , perhaps in front of or behind the main cluster . a state of complex dynamics is supported by the non - uniform temperature distribution and maybe most importantly by the variation in redshift of the x - ray emitting gas across the cluster . finally , recent measurements of the sunyaev - zeldovich effect by @xcite indicate that the inferred optical depth of the comptonizing gas might actually be higher than one would infer from the simple spherically symmetric model obeying hydrostatic equilibrium , and a clumpy or elongated structure ( aligned along the line of sight ) would alleviate the discrepancy . clearly , this cluster deserves further detailed studies ; it is a very interesting potential target for the astro - e2 mission , where the x - ray calorimeter will provide an unprecedented spectral resolution , capable of clearly confirming the complex redshift structure of the intra - cluster gas , hinted by the xmm data . if the cluster is indeed undergoing a merger , this might provide an environment where at least some fraction of the gas is accelerated ( for instance , via shocks ) to form a non - thermal distribution of particles , which might produce radiation detectable in radio ( via synchrotron emission ) , in hard x - rays ( by comptonizing the cosmic microwave background ) , and in gamma rays , potentially detectable by the future mission glast . the superior effective area of the xmm - newton telescope has allowed us to perform a detailed analysis of abell 1689 . comparing with the data from rosat , asca and chandra we verified that the data are consistent with earlier observations . importantly , there is no indication of any additional absorbing component in the xmm data , attributing the low - energy excess absorption found by @xcite in the chandra data to uncorrected instrumental effects . the now available ` acisabs ` code successfully corrects for these effects . we confirm earlier findings that there is a large discrepancy , of a factor 2 or more , between mass estimates from gravitational lensing and x - ray derived mass using an unparameterized deprojection technique . our analysis indicates that this discrepancy is true for the central part of the cluster , but also might be the case for the entire observable cluster ; our finding is in contrast to @xcite , who conclude that at large radii , there is no disagreement between the x - ray and lensing masses . although the x - ray determined mass appears to be discrepant from the values determined from most lensing techniques , it seems to be in good agreement with that derived from weak gravitational shear . we compare the gas mass of the cluster with the total mass , and find that for abell 1689 , @xmath191 is @xmath250 , significantly less than @xmath251 , the value derived by @xcite for 10 dynamically relaxed clusters . our calculation of the asymmetric temperature distribution of the cluster provides further evidence that this cluster is not in a relaxed state : the lower than expected gas mass fraction is yet another piece of evidence . we also present the first measurement of asymmetries in redshift for different regions of the icm in the cluster determined from the x - ray data alone . a similar analysis was done by @xcite who claim a 3-@xmath252 detection of bulk motion from asca observations of the centaurus cluster . we argue that the redshift variation detected here might be either due to line - of - sight clustering or possibly due to large bulk motions of the gas . even though this cluster is clearly not as relaxed as might be expected from its apparent spherical form , we find the slope of the total mass in the central region to be in good agreement with what is expected from numerical simulations of structure formation . the density slope for @xmath253 is @xmath254 . while our current understanding of the structure and dynamics of galaxy clusters is insufficient to put limits on the self - interaction of dark matter , better data and more accurate simulations appear promising for the future . we would like to thank yasushi ikebe and steve snowden for kindly helping in answering questions regarding xmm - newton data reduction and calibration issues , and james chiang for the suggestions regarding spectral deprojection . we are grateful to the xmm sas software team for supplying the data reduction software . k. andersson would also like to thank per carlson and tsuneyoshi kamae for making this research possible . x - ray data used in this work was extracted from nasa s heasarc archives . we acknowledge the use of the digitized sky survey , produced at the space telescope science institute under us government grant nag w-2166 , based on the original photographic data obtained by the uk schmidt telescope . this work was partially supported by the alice & knut wallenberg foundation via the wallenberg research link , by nasa chandra funds administered via sao as go1 - 2113x , and by department of energy contract de - ac03 - 76sf00515 to the stanford linear accelerator center . we are grateful to the anonymous referee for many constructive suggestions and comments .

the perron - frobenius theorem @xcite asserts that a real square matrix with strictly positive entries has a ` spectral gap ' , i.e. the matrix has a positive simple eigenvalue and all other eigenvalues are strictly smaller in modulus . more generally , let @xmath4 be a bounded linear operator acting upon a real or complex banach space and of spectral radius @xmath5 . we say that @xmath4 has a spectral gap if ( 1 ) it has a simple isolated eigenvalue @xmath6 the modulus of which equals @xmath5 and ( 2 ) the remaining part of the spectrum is contained in a disk centered at zero and of radius strictly smaller than @xmath5 . jentzsch generalized in @xcite the perron - frobenius theorem to integral operators with a strictly positive continuous kernel . the proof uses the schauder - tychonoff theorem to produce a dual eigenvector and then a contraction on the kernel of this eigenvector to get a spectral gap . kren - rutman ( * ? ? ? * theorem 6.3 ) ( see also @xcite and @xcite ) gave an abstract setting of this result by considering a punctured real closed cone mapped to its interior by a compact operator . compactness of the operator essentially reduces the problem to finite dimensions . birkhoff , in a seminal paper @xcite , developed a more elementary and intuitive ( at least in our opinion ) perron - frobenius ` theory ' by considering the projective contraction of a cone equipped with its associated hilbert metric . birkhoff noted that this projective metric satisfies a _ contraction principle _ , i.e. any linear map preserving the cone is a contraction for the metric and the contraction is strict and uniform if the image of the cone has finite projective diameter . all these results , or rather their proofs , make use of the ` lattice'-structure induced by a real cone on a real banach space ( see @xcite and also @xcite ) . on the other hand , from complex analysis we know that the poincar metric on the unit disk , @xmath7 , and the induced metric on a hyperbolic riemann surface enjoy properties similar to the hilbert metric , in particular a _ contraction principle _ with respect to conformal maps . more precisely , if @xmath8 is a conformal map between hyperbolic riemann surfaces then its conformal derivative never exceeds one . the map is a strict contraction unless it is a bijection ( see e.g. ( * ? ? ? * chapter i.4 : theorems 4.1 and 4.2 ) ) . by considering analytic images of complex discs kobayashi @xcite ( see also @xcite ) constructed a hyperbolic metric on complex ( hyperbolic ) manifolds , a tool with many applications also in infinite dimensions ( see e.g. ( * ? ? ? * appendix d ) ) . given a real cone contraction , perturbation theory allows on abstract grounds to consider ` small ' complex perturbations but uniform estimates are usually hard to obtain . uniform complex estimates are needed e.g. when proving local limit theorems and refined large deviation theorems for markov additive processes ( see @xcite and references therein ) and also for studying the regularity of characteristic exponents for time - dependent and/or random dynamical systems ( see e.g. @xcite ) . it is desirable to obtain a description of a projective contraction and , in particular , a spectral gap condition for complex operators without the above - mentioned restrictions . we describe in the following one way to accomplish this goal . + in section [ sec c cones ] we introduce families of @xmath9-invariant cones in complex banach spaces and a theory for the projective contraction of such cones . the central idea is simple , namely to use the poincar metric as a ` gauge ' on 2-dimensional affine sections of a complex cone . at first sight , this looks like the kobayashi construction . a crucial difference , however , is that we only consider disk images in 2-dimensional subspaces . also we do not take infimum over chains ( so as to obtain a triangular inequality , see appendix [ a proj metric ] ) . this adapts well to the study of linear operators and makes computations much easier than for the general kobayashi metric . lemma [ projective gauge ] shows that this gauge is indeed projective . the contraction principle for the poincar metric translates into a contraction principle for the gauge and , under additional regularity assumptions , developed in section [ sec complex ] , into a projective contraction , and finally a spectral gap , with respect to the banach space norm . in sections [ section real case ] and [ sec canonical complexification ] we consider real cones and define their _ canonical complexification_. for example , @xmath10 is the canonical complexification of the standard real cone , @xmath11 . we show that our complex cone contraction yields a genuine extension of the birkhoff cone contraction : a real birkhoff cone is _ isometrically _ embedded into its canonical complexification . it enjoys here the same contraction properties with respect to linear operators . we obtain in section [ c dominated ] then one of our main results : when a complex operator is _ dominated _ by a sufficiently regular real cone - contraction ( assumption [ main assumption ] ) then ( theorem [ c domin sp gap ] ) the complex operator has a spectral gap . it is of interest to note that the conditions on the complex operator are expressed in terms of a real cone and sometimes easy to verify . sections [ sec applications]-[sec rpf ] thus presents a selection of complex analogues of well - known real cone contraction theorems : a perron - frobenius theorem for complex matrices ( as stated at the end of the abstract ) , jentzsch s theorem for complex integral operators , a kren - rutman theorem for compact and quasi - compact complex operators and a ruelle - perron - frobenius theorem for complex transfer operators . acknowledgements : i am grateful to a douady for a suggestion in the proof of lemma [ lemma contraction ] and to the anonymous referee for several valuable suggestions and corrections . let @xmath12 denote the riemann sphere . when @xmath13 is an open connected subset avoiding at least three points one says that the set is _ hyperbolic_. we write @xmath14 for the corresponding hyperbolic metric . we refer to ( * ? ? ? * chapter i.4 ) or ( * ? ? ? * chapter 2 ) for the properties of the hyperbolic metric which we use in the present paper . as normalization we use @xmath15 on the unit disk @xmath16 and the metric @xmath14 on @xmath17 induced by a riemann mapping @xmath18 . one then has : @xmath19 let @xmath20 be a complex topological vector space . we denote by @xmath21 the complex subspace generated by two vectors @xmath22 and @xmath23 in @xmath20 . 1 . we say that a subset @xmath24 is a * closed complex cone * if it is closed in @xmath20 , @xmath9-invariant ( i.e. @xmath25 ) and contains at least one complex line we say that the closed complex cone @xmath26 is * proper * if it contains no complex planes , i.e. if @xmath22 and @xmath23 are independent vectors then @xmath27 . throughout this paper we will simply refer to a proper closed complex cone as a * @xmath28-cone*. let @xmath26 be a @xmath9-cone . given a pair of non - zero vectors , @xmath29 , we consider the subcone : @xmath30 . we wish to construct a ` projective distance ' between the complex lines @xmath31 and @xmath32 within this subcone . we do this by considering the affine plane through @xmath33 and @xmath34 , choosing coordinates ( the choice to some extend being arbitrary ) as follows : @xmath35 with the convention that @xmath36 iff @xmath37 . the interior of this `` slice '' is denoted @xmath38 ( for the spherical topology on @xmath39 ) . we note that when @xmath22 and @xmath23 are linearly independent , continuity of the canonical mapping @xmath40 implies that @xmath41 is a closed subset of @xmath39 . as the cone is proper , @xmath42 is a strict subset so that @xmath43 is open and non - empty , whence contains ( more than ) 3 points . if , in addition , @xmath44 is connected it is a hyperbolic riemann surface ( ( * ? ? ? * theorem i.3.1 ) ) . given a @xmath9-cone , we define the * gauge * , @xmath45 $ ] , between two points @xmath46 as follows : when two vectors are co - linear we set @xmath47 . if they are linearly independent and @xmath48 and @xmath49 belongs to the same connected component @xmath17 of @xmath38 we set : @xmath50 in all remaining cases , we set @xmath51 . when @xmath52 is a ( sub-)cone of the @xmath9-cone @xmath26 we write @xmath53 $ ] for the projective * ` diameter ' * of @xmath54 in @xmath26 . we call it a diameter even though the gauge need not verify the triangular inequality , whence need not be a metric ( see appendix [ a proj metric ] for more on this issue ) . let @xmath26 be a @xmath9-cone . the gauge on the cone is symmetric and projective , i.e. for @xmath46 and @xmath55 : @xmath56 proof : for @xmath57 we write @xmath58 with @xmath59 then @xmath60 extends to a conformal bijection @xmath61 ( a mbius transformation of @xmath39 ) preserving @xmath48 and @xmath49 . the hyperbolic metric is invariant under such transformations so indeed @xmath62 ( but both could be infinite ) . similarly , the map @xmath63 yields a conformal bijection between the domains @xmath64 and @xmath65 , interchanging @xmath48 and @xmath49 and the symmetry follows . + let @xmath66 be a complex linear map between topological vector spaces and let @xmath67 and @xmath68 be @xmath9-cones for which @xmath69 . then the map , @xmath70 is a contraction . if the image has finite diameter , i.e. @xmath71 , then the contraction is strict and uniform . more precisely , there is @xmath72 ( depending on @xmath73 only ) for which @xmath74 proof : let @xmath75 and set @xmath76 and @xmath77 for which we have @xmath78 suppose that @xmath79 are linearly independent and that @xmath80 and @xmath81 are hyperbolic ( if not , @xmath82 vanishes and we are through ) . since shrinking a domain increases hyperbolic distances , it follows that @xmath83 ( although both could be infinite ) . suppose now that @xmath84 . then @xmath48 and @xmath49 belong to the same connected component , @xmath54 , of @xmath85 . we may suppose that @xmath48 and @xmath49 also belong to the same connected component , @xmath17 , of @xmath38 ( or else @xmath86 and we are through ) . our assumptions imply that @xmath87 is a strict inclusion and that @xmath88 . choose @xmath89 and @xcite pick @xmath90 for which @xmath91 ( this is possible as the inclusion @xmath87 is strict and the diameter of @xmath17 did not exceed @xmath73 ) . the inclusion @xmath92 is non - expanding and the inclusion @xmath93 is a contraction which has conformal derivative uniformly smaller than some @xmath72 on the punctured @xmath73-neighborhood , @xmath94 , of p ( see remark [ remark punctured disk ] ) . in particular , the composed map ( see figure [ fig inclus ] ) @xmath95 has conformal derivative smaller than @xmath96 at @xmath97 . as @xmath89 was arbitrary this is true at any point along a geodesic joining @xmath48 and @xmath49 in @xmath17 so that @xmath98 [ remark punctured disk ] an explicit bound may be given using the expression @xmath99 for the metric on the punctured disk at @xmath100 ( see e.g. ( * ? ? ? * example 2.8 ) ) . denoting , @xmath101 , we obtain the bound , @xmath102 . often , however , it is possible to improve this bound . for example , suppose that @xmath17 is contractible in @xmath54 ( e.g. if @xmath103 is simply connected ) and that @xmath17 is contained in a hyperbolic ball of radius @xmath104 . lifting to the universal cover we may assume that @xmath105 and that @xmath106 with @xmath107 . the inclusion @xmath108 has conformal derivative @xmath109 for @xmath110 . we may thus use @xmath111 for the contraction constant . recall that for a real birkhoff cone @xcite one may take @xmath112 ( an open interval in @xmath113 of diameter @xmath73 is a ball of radius @xmath114 in @xmath113 ) . let @xmath115 be a complex banach space and let @xmath116 be a @xmath9-cone ( definition [ def complex cone ] ) . we denote by @xmath117 the dual of @xmath115 and we write @xmath118 for the canonical duality @xmath119 . we will consider a bounded linear operator @xmath120 which preserves @xmath121 and is a strict and uniform contraction with respect to our gauge on @xmath26 . we seek conditions that assure : ( 1 ) the presence of an invariant complex line ( existence of an eigenvector of non - zero eigenvalue ) and ( 2 ) a spectral gap . in short , an invariant line appears when the cone is not too ` wide ' and the spectral gap when , in addition , the cone is not too ` thin ' . let @xmath116 be a @xmath9-cone in a complex banach space ( in section [ section real case ] we will use the very same definition for a real cone in a real banach space ) . when @xmath122 is a non - zero functional , bounded on the vector space generated by @xmath26 , we define the aperture of @xmath26 relative to @xmath123 : @xmath124.\ ] ] we define the aperture of @xmath26 to be : @xmath125 $ ] . 1 . we call @xmath26 _ inner regular _ if it has non - empty interior in @xmath115 . + we say that @xmath26 is _ @xmath126-inner regular _ ( with @xmath127 ) if there are @xmath128 and @xmath129 so that @xmath130 ( when @xmath131 the cone is inner regular ) . we say that @xmath26 is _ outer regular _ if @xmath132 . + we say that @xmath26 has _ @xmath133-bounded sectional aperture _ ( with @xmath134 ) iff for every pair @xmath135 , the sub - cone @xmath136 is of @xmath133-bounded aperture , i.e.there is a non - zero linear functional , @xmath137 , such that @xmath138 3 . we say that @xmath26 is _ regular _ iff the cone is inner and outer regular . + when a cone is of bounded aperture then the cone has a bounded global transverse section not containing the origin . this is often a too strong requirement . for example , in @xmath139-spaces this is usually ok but not in @xmath140 with @xmath141 unless we are in finite dimensions . being inner regular means containing an open ball and this typically fails in @xmath140 for @xmath142 , again with the exemption of the finite dimensional case . the notions of bounded sectional aperture and @xmath126-inner regularity , respectively , are more flexible and may circumvent the two above - mentioned restrictions . we illustrate this in example [ ex real jentzsch ] and theorem [ complex jentzsch ] . it is necessary to create a passage between the cone - gauge and the banach space norm . the regularity properties defined above will enable us to do so through the following two lemmas : let @xmath26 be a @xmath9-cone and let @xmath143 , @xmath144 . suppose that there is @xmath128 such that @xmath145 for all @xmath146 for which @xmath147 . then 1 . @xmath148 2 . if @xmath123 is a linear functional on @xmath149 , which never vanishes on the punctured subcone , @xmath150 , then also : @xmath151 . 3 . if @xmath26 is of @xmath133-bounded sectional aperture then : @xmath152 . proof : let @xmath147 . using ( [ slice ] ) and the scale - invariance of the cone we see that + @xmath153 . our hypothesis then implies that @xmath154 contains a disc of radius @xmath155 , centered at 1 . shrinking a domain increases hyperbolic distances , whence @xmath156 if @xmath123 is non - zero on the punctured subcone , then @xmath157 for all @xmath147 and this implies the second claim . for the last assertion let @xmath123 be as in ( [ m functional ] ) with @xmath158 . possibly after multiplying @xmath22 and @xmath159 with complex phases we may assume that @xmath160 . then @xmath161 . let @xmath26 be a closed complex cone of @xmath133-bounded sectional aperture . then @xmath26 is proper , whence a @xmath9-cone ( definition [ def complex cone ] ) . if @xmath162 and @xmath163 is a functional associated to the subcone @xmath30 as in ( [ m functional ] ) then : @xmath164 proof : we normalize the functional so that @xmath158 . then @xmath165 for all @xmath166 . denote @xmath167 and @xmath168 and consider , as a function of @xmath169 , the point @xmath170 . when @xmath171 the properties of @xmath123 show that @xmath172 and therefore , @xmath173 setting @xmath174 $ ] we see that @xmath175 . the radius @xmath176 is bounded iff @xmath22 and @xmath23 are independent so the cone is proper . enlarging a domain decreases hyperbolic distances so @xmath177 therefore , @xmath178 , and the stated bound follows . + let @xmath26 be a @xmath9-cone of @xmath133-bounded sectional aperture . let @xmath120 be a strict cone - contraction , i.e.@xmath179 with @xmath180 . then @xmath26 contains a unique @xmath181-invariant complex line , @xmath182 . proof : let @xmath183 and set @xmath184 . we will construct a cauchy - sequence @xmath185 recursively . given @xmath186 , @xmath187 choose , as in definition [ def regular cone ] ( 2 ) , a functional @xmath188 normalized so that @xmath189 , associated to the subcone @xmath190 . set @xmath191 ( for which we have the bound @xmath192 ) and define the next element in our recursion : @xmath193 using lemma [ bounded sectional aperture ] and then lemma [ lemma contraction ] ( with a contraction constant @xmath194 ) we obtain for @xmath187 : @xmath195 as @xmath196 and @xmath197 we get : @xmath198 noting that @xmath199 , the first inequality implies : @xmath200 the sequence , @xmath201 , is therefore cauchy , whence has a limit , @xmath202 , @xmath203 . the limit belongs to @xmath26 because the cone was assumed closed . writing @xmath204 and using the second inequality in ( [ en bounds ] ) as well as ( [ e bounds ] ) and @xmath205 we obtain @xmath206 so also the limit @xmath207 exists . but @xmath208 shows that @xmath209 which implies that @xmath210 , whence that @xmath211 is a @xmath181-invariant complex line . suppose that also @xmath212 ( with @xmath213 ) is @xmath181-invariant . then @xmath214 and this implies @xmath215 so the two vectors must be linearly dependent . thus , @xmath182 is unique . + let @xmath120 and let @xmath26 be a @xmath9-cone of @xmath133-bounded sectional aperture which is _ @xmath126-inner regular _ for some @xmath127 . suppose that @xmath181 is a strict cone - contraction , i.e. @xmath179 with @xmath180 . then @xmath181 has a spectral gap . proof : by the previous theorem @xmath181 has a unique eigenvector in the cone , @xmath216 , with a non - zero complex eigenvalue , @xmath6 . in order to simplify the notation we replace @xmath181 by @xmath217 and assume thus that there is @xmath218 , @xmath203 for which @xmath219 . a slight complication is that @xmath220 need not be in the interior of @xmath26 , or even worse , the interior of @xmath26 may be empty . @xmath126-inner regularity ( definition [ def regular cone ] ( 1 ) ) allows us to proceed as follows : let @xmath221 , @xmath222 , @xmath127 and @xmath128 be such that @xmath130 . we write @xmath223 , @xmath224 for the iterates of @xmath225 . by taking limits in equations ( [ e bounds ] ) and ( [ l bounds ] ) we see that the sequences , @xmath226 and @xmath227 , constructed in the preceding theorem verify : @xmath228 then @xmath229 , so for all @xmath187 : @xmath230 we also get that @xmath231 now let @xmath144 . by our choice of @xmath225 when @xmath232 then @xmath233 . by lemma [ lemma inner regular ] , @xmath234 applying the contraction in lemma [ lemma contraction ] , we get @xmath235 in order to get a norm - estimate out of this we pick a sequence , @xmath236 , @xmath189 ( as in equation ( [ m functional ] ) ) , this time associated to the subcones , @xmath237 . by lemma [ bounded sectional aperture ] , @xmath238 develop the left hand side in @xmath239 , multiply by @xmath240 ( which is bounded by @xmath241 and retain the linear term to obtain @xmath242 valid for any @xmath144 . let us write @xmath243 for the coefficient to @xmath244 . since @xmath245 whenever @xmath246 and @xmath247 we also have @xmath248 for such @xmath239-values . the second half of lemma [ lemma inner regular ] then shows that @xmath249 uniformly in @xmath0 . using the identity @xmath250 . and the three bounds ( [ x ineq ] - [ alpha ineq ] ) we obtain for @xmath187 : @xmath251 therefore , @xmath252 exists . the limit is necessarily proportional to @xmath220 ( since @xmath253 and @xmath254 ) and because of ( [ alpha ineq ] ) and ( [ x - bound ] ) we also have @xmath255 . then , @xmath256 so that @xmath257 linearity of @xmath181 implies that the mapping @xmath258 must be linear , and as a linear functional it is bounded in norm by @xmath259 . finally , this time returning to the unnormalized operator , we have shown that @xmath260 with @xmath261 . it follows that @xmath6 is a simple eigenvalue of @xmath181 corresponding to the eigenprojection , @xmath262 and that the remainder has spectral radius not exceeding @xmath263 . + let @xmath115 be a complex banach space and consider @xmath264 , @xmath265 with @xmath266 . we write @xmath267 for the associated one dimensional projection . for @xmath268 we set @xmath269 then @xmath270 and @xmath271 so that @xmath272 is a regular @xmath9-cone . furthermore , if @xmath273 a calculation shows that @xmath274 . we have the following characterization of the spectral gap property : a bounded linear operator , @xmath120 , has a spectral gap iff it is a strict contraction of a regular @xmath9-cone . proof : one direction is the content of theorem [ spectral gap ] ( since a regular cone in particular is of uniformly bounded sectional aperture ) . for the other direction one uses the spectral gap projection @xmath275 to construct an adapted norm ( equivalent to @xmath276 ) : @xmath277 for some fixed choice of @xmath278 . using this norm to define the cone family in ( [ c standard ] ) , it is not difficult to see that @xmath181 is a strict and uniform contraction of @xmath272 , @xmath279 . let @xmath280 denote a real banach space . recall that a subset @xmath281 is called a ( real ) proper closed convex cone if it is closed and convex and if @xmath282 we note that convexity is a useful property that _ a fortiori _ is lost when dealing with complex cones . in the following , we will refer to a real proper closed convex cone as an _ @xmath113-cone_. we assume throughout that such a cone is non - trivial , i.e. not reduced to a point . given an @xmath113-cone one associates a projective ( hilbert ) metric for which we here give two equivalent definitions ( for details we refer to @xcite ) . the first , originally given by hilbert , uses cross - ratios and is very similar to our complex cone gauge : let @xmath283 denote the extended real line ( topologically a circle ) . for @xmath284 , we write @xmath285 with the convention that @xmath286 iff @xmath287 . properness of the cone implies that @xmath288 iff @xmath22 and @xmath23 are co - linear . in that case we set their distance to zero . otherwise , @xmath289 is a closed ( generalized ) segment @xmath290\subset \hatrr$ ] containing the segment @xmath291 $ ] , see figure [ fig cone embed ] in section [ sec canonical complexification ] . the logarithm of the cross - ratio of @xmath292 , @xmath293 then yields the hilbert projective distance between @xmath22 and @xmath23 . birkhoff @xcite found an equivalent definition of this distance : for @xmath284 , one defines @xmath294 \mylabel{birkhoff beta}\ ] ] in terms of which : @xmath295 . \mylabel{birkhoff def}\ ] ] a simple geometric argument shows that indeed the two definitions are equivalent . given a linear functional , @xmath296 , the image of the cone , @xmath297 , equals either @xmath298 , @xmath299 , @xmath300 or @xmath113 . one defines the dual cone as @xmath301 and using mazur s theorem , cf . e.g. @xcite , one sees that the @xmath113-cone itself may be recovered from : @xmath302 given an @xmath113-cone @xmath303 we use definition [ def aper one ] ( replacing @xmath9 by @xmath113 , complex by real ) to define the aperture of @xmath303 . it is given as the infimum of @xmath133-values for which there exists a linear functional @xmath123 satisfying ( see figure [ real cone figure ] ) @xmath304 the aperture , @xmath305 $ ] , of an @xmath113-cone , @xmath281 , is determined by @xmath306 proof : let @xmath307 and note that @xmath308 belongs to @xmath309 , the convex hull of cone - elements of norm one . the reciprocal of the right hand side in ( [ real aper ] ) therefore equals @xmath310 $ ] . suppose that @xmath128 . then @xmath311 and @xmath312 are disjoint convex subsets . the vector difference , @xmath313 , is open , convex and does not contain the origin , whence ( * ? ? ? * lemma 2.2 , p.89 ) there is @xmath314 whose kernel does not intersect @xmath315 . we may normalize @xmath316 so that @xmath317 then @xmath318 and therefore @xmath319 . to get the converse inequality let @xmath123 be positive and verify ( [ r aper ] ) . then @xmath320 + let @xmath281 be a @xmath321-dimensional @xmath113-cone of @xmath133-bounded sectional aperture . then @xmath303 itself is of @xmath322-bounded aperture . proof : let @xmath323 . by a theorem of caratheodory , a point in the convex hull , @xmath324 , is _ a fortiori _ in the convex hull of @xmath325 points in @xmath326 ( see e.g. @xcite ) . if @xmath327 , we may even write it as a limit of convex combinations of @xmath321 points in @xmath326 now , apply this to the set @xmath4 in the proof of the previous lemma . in the formula , ( [ real aper ] ) it thus suffices to consider @xmath321 cone - elements which we may order decreasingly according to their norm , @xmath328 . using lemma [ l real aper ] with @xmath329 , the k - bounded sectional aperture implies that @xmath330 thus @xmath331 and in view of lemma [ l real aper ] , we see that @xmath303 is of @xmath322 bounded aperture . + let @xmath332 be an @xmath113-cone . then @xmath303 is outer regular . proof : as in the previous lemma it suffices to look at the supremum in ( [ real aper ] ) over @xmath321-tuples . the set @xmath333 is compact and @xmath334 is continuous and non - vanishing on @xmath4 , whence has a minimum , @xmath128 . it follows that @xmath335 . + in the literature an @xmath113-cone @xmath303 is said to be norm - directed ( with a constant @xmath336 ) if @xmath337 , @xmath338 . for an @xmath113-cone our notion of uniformly bounded sectional aperture is equivalent ( up to a small unavoidable loss in constants ) to that of being norm - directed . to see this note that if @xmath303 is of @xmath133-bounded sectional aperture and @xmath316 verifies ( [ l - inequality ] ) then @xmath339 : @xmath340 which shows that @xmath303 is @xmath133-norm - directed . conversely , if @xmath303 is @xmath133-norm - directed then @xmath341 and lemma [ l real aper ] shows that @xmath303 is of @xmath342-bounded sectional aperture . for example , @xmath343 is 1-norm directed and of 1-bounded aperture , whereas @xmath344 is 1-norm directed , of 2-bounded sectional aperture but only of @xmath321 bounded aperture . lemma [ lemma inner regular ] ( 3 ) is a complex cone - analogue of being norm - directed . let @xmath345 and let @xmath281 be an @xmath113-cone which is @xmath346-inner regular ( for some @xmath127 ) and @xmath133-norm - directed . suppose that @xmath4 is a strict cone - contraction , i.e.@xmath347 with @xmath348 . then @xmath4 has a spectral gap . more precisely , there is @xmath349 and a one dimensional projection @xmath275 for which @xmath350 has spectral radius not greater than @xmath351 . proof : see @xcite in the case of @xmath303 being inner regular and @xmath133-norm directed . when @xmath303 is assumed only @xmath346-inner regular with @xmath352 one may either adapt the proof of birkhoff ( easy ) or use remark [ c contains r ] below . let @xmath303 be an @xmath113-cone in @xmath353 , @xmath354 and suppose that @xmath345 verifies @xmath355 . then @xmath4 has a spectral gap . proof : implicitly it is assumed that @xmath303 has non - empty interior . lemma [ r outer reg ] shows that @xmath303 is outer regular , in particular , norm - directed . local compactness of @xmath353 implies that @xmath356 so the corollary follows from birkhoff s theorem . + let @xmath281 be an @xmath113-cone . it is standard to write @xmath357 for the induced partial ordering of @xmath358 . for @xmath359 , we also write @xmath360 . the following dominated cone contraction theorem is trivial in the context of an @xmath113-cone contraction . in section [ c dominated ] we show that a similar ( non - trivial ) result holds in the complex case . let @xmath361 be contractions of the @xmath113-cone @xmath303 . suppose that there are constants @xmath362 for which @xmath363 . then @xmath364 proof : given @xmath365 , suppose that @xmath366 are such that @xmath367 , @xmath368 . then also @xmath369 and @xmath370 so that @xmath371 and the claim follows by birkhoff s characterization ( [ birkhoff def ] ) . + + 1 . let @xmath372 and suppose that @xmath373 for all indices . setting @xmath374 we see that @xmath375 which is of zero projective diameter in @xmath11 . by theorem [ real domin gap ] we recover the standard result : @xmath376 in particular , if @xmath377 and @xmath378 denote the leading eigenvalue and the second largest eigenvalue ( in absolute value ) , respectively , then @xmath379 . the standard perron - frobenius theorem generalizes easily to integral operators , cf . jentzsch s theorem @xcite and the generalization given by birkhoff in @xcite ) . we present a somewhat different generalization : let @xmath380 be a measure space and let @xmath381 , @xmath382 . let @xmath383 ( @xmath384 , a.e . ) and @xmath385 ( @xmath386 , a.e . ) with @xmath387 $ ] being the conjugated exponent so that @xmath388 . let @xmath389 be a @xmath390-measurable map . we suppose there are constants @xmath391 so that for @xmath392-almost all @xmath393 : @xmath394 let @xmath345 be the integral operator defined by @xmath395 . then @xmath4 has a spectral gap ( again with a contraction rate given by @xmath396 ) . + proof : we write @xmath397 for the cone of positive @xmath140-functions ( @xmath398 , a.e . ) and compare the operator @xmath4 with the one - dimensional projection @xmath399 . our assumption , @xmath400 shows that @xmath401 and that @xmath402 . thus , @xmath403 . + in general , @xmath303 need not be inner regular ( unless @xmath404 ) . it is , however , @xmath346-inner regular ( with @xmath405 ) for any value of @xmath406 . one has : @xmath407 so @xmath408 when @xmath409 . to see that @xmath303 is of uniformly bounded sectional aperture let @xmath410 be of unit norm in @xmath411 , @xmath412 and pick @xmath413 with @xmath414 ( the case @xmath404 , @xmath415 should be treated separately ; we leave this to the reader ) and @xmath416 . the functional @xmath417 then verifies @xmath418 for all @xmath419 . thus @xmath303 is of 2-bounded sectional aperture . a complex cone yields a genuine extension / generalization of the cone contraction described by birkhoff @xcite . more precisely , we will show that any birkhoff cone may be isometrically embedded in a complex cone , enjoying qualitatively the same contraction properties . let @xmath280 be a banach space over the reals . a complexification @xmath420 of @xmath280 is a complex banach space , equipped with a bounded anti - linear complex involution , @xmath421 , @xmath422 , @xmath423 , @xmath424 , @xmath425 , @xmath426 , for which @xmath427 is the real part . then @xmath428 is a direct sum . [ note that this is not the same as regarding @xmath420 as a real banach space . for example , @xmath429 is a complexification of @xmath430 for any @xmath431-norm , @xmath406 , while the real dimension of @xmath429 is @xmath432 . usually @xmath433 will be an isometry on @xmath420 in which case the canonical projections , @xmath434 and @xmath435 , have norm one . we note that any real banach space , @xmath436 , admits a complexification , @xmath428 as follows : we adopt the obvious rules for multiplying by complex numbers , set @xmath437 and introduce a norm e.g. using real functionals , @xmath438 the latter norm is equivalent ( within a factor of 2 ) to any other norm on @xmath420 having as real part the given space @xmath436 . for the rest of this section @xmath280 will denote the real part of a complex banach space @xmath420 . a real linear functional , @xmath439 , extends to a complex linear functional by setting @xmath440 for @xmath441 . given an @xmath113-cone @xmath281 we define its _ canonical complexification _ : @xmath442 we have the following * polarization * identity : @xmath443 proof : let @xmath444 . our defining condition ( [ complexification ] ) means that @xmath445 ( assume here it is non - zero ) must have an argument that vary within a @xmath446 angle as @xmath447 varies . normalizing appropriately , we may write @xmath448 with @xmath449 and @xmath450 . if we set @xmath451 with @xmath358 then @xmath452 for all @xmath447 . hence , @xmath453 for all such functionals and by ( [ dual cone char ] ) this is equivalent to @xmath454 . if @xmath455 ( or @xmath456 ) then we may write @xmath457 ( or @xmath458 ) so we may always assume @xmath459 . + let @xmath303 be an @xmath113-cone of @xmath133-bounded aperture . then its canonical complexification , @xmath460 , is of @xmath461-bounded aperture . proof : let @xmath314 satisfy @xmath462 and extend @xmath316 to a complex linear functional . when @xmath463 we use polarization , lemma [ c cone char ] , to write @xmath464 with @xmath465 . then @xmath466 , from which @xmath467 and @xmath468 so that @xmath469 . as @xmath470 we also have @xmath471 . therefore , @xmath472 let @xmath303 be an @xmath113-cone . if @xmath303 is ( 1 ) inner regular/ ( 2 ) outer regular/ ( 3 ) of bounded sectional aperture then so is its canonical complexification . proof : ( 1 ) when @xmath303 contains a ball @xmath473 one verifies that @xmath460 contains the ball @xmath474 . + ( 2 ) as shown in lemma [ c aperture ] , if @xmath475 is of @xmath133-bounded aperture then @xmath476 is of @xmath461-bounded aperture . + ( 3 ) let @xmath477 and write @xmath478 for the sub - cone generated by these two elements . we also write @xmath479 and @xmath480 which is an at most 4 and at least 1-dimensional @xmath113-subcone of @xmath303 . now , if @xmath481 then @xmath482 with @xmath483 and clearly also @xmath484 . but then @xmath485 so that also @xmath486 , with @xmath476 being the complexification of @xmath487 . by lemma [ cr bd aperture ] , @xmath476 is of @xmath488 bounded aperture so by lemma [ c aperture ] , @xmath476 and therefore also @xmath489 are of @xmath490 bounded ( complex ) aperture . + let @xmath303 be an @xmath113-cone and let @xmath460 denote its canonical complexification ( [ complexification ] ) . then @xmath460 is a @xmath9-cone ( definition [ def complex cone ] ) . writing @xmath491 for our projective gauge on the complex cone , the natural inclusion , @xmath492 is an isometric embedding . proof : the set @xmath460 is clearly @xmath9-invariant . consider independent vectors , @xmath365 . by lemma [ r outer reg ] any finite dimensional subcone of @xmath303 is outer regular , so in particular of uniformly bounded sectional aperture . our previous lemma shows that the corresponding complex cone is of bounded sectional ( complex ) aperture . but then @xmath460 must be proper by lemma [ lemma inner regular](3 ) . regarding the embedding we may normalize the points so that @xmath493 $ ] is a bounded segment in @xmath113 . define the ` boundary ' points , @xmath494 and @xmath495 . for any @xmath496 the point @xmath497 is outside the closed convex cone @xmath303 . by mazur s theorem , @xcite , we may separate this point from @xmath303 by a functional @xmath498 . for any @xmath496 we may then find @xmath499 for which @xmath500 then @xmath501 only if @xmath502 for any @xmath503 whence only if @xmath504 conversely , when @xmath505 and @xmath506 then @xmath507 so this condition is also sufficient . we thus have : @xmath508 . therefore @xmath509 is a generalized disc , symmetric under complex conjugation for which @xmath510 ( see figure [ fig cone embed ] ) . in this situation we know explicit formulae for both ( [ hilbert metric ] ) the real and ( [ c metric ] ) the complex hyperbolic metrics in terms of cross - ratios so we get @xmath511 . + for @xmath187 the set @xmath512 is a regular @xmath9-cone . the inclusion , @xmath513 is an isometric embedding . proof : let @xmath514 , @xmath515 denote the canonical coordinate projections then @xmath516 and @xmath517 . thus , @xmath518 is the canonical complexification of the standard real cone @xmath11 . see figure [ fig can complex ] for an illustration of @xmath519 . + below we shall need the following complex polarization let @xmath520 be at a distance @xmath521 . we may then find @xmath522 so that the vector @xmath523 , or equivalently : @xmath524 verifies : @xmath525 in particular , @xmath526 and @xmath527 proof : possibly after replacing @xmath528 by @xmath529 for a suitable @xmath530 we may assume that @xmath531 and @xmath532 . then @xmath533 is a real segment @xmath534 $ ] for which @xmath535t_1,t_2 [ \supset [ -1,1]$ ] ( see figure [ fig geom ] ) . now write @xmath536 , @xmath537-\pi/2,\pi/2[$ ] . since @xmath538 we get in the @xmath539-coordinate ; @xmath540 & = & \ { \theta \in i : x - \tan \theta \ ; y \in \ccr\ } \nonumber \\ & = & \ { \theta \in i : \cos \theta \ ; x - \sin \theta \ ; y \in \ccr\ } \nonumber \\ & = & \ { \theta \in i : { { \rm re}}\ ; e^{i\theta } ( x+iy ) \in \ccr\ } \label{thetaxy}\end{aligned}\ ] ] in @xmath539-coordinates the projective distance between @xmath541 and @xmath542 is then given by @xmath543 if we do a complex rotation , @xmath544 with @xmath545-\frac{\pi}2 + \theta_1 , \frac{\pi}{2 } - \theta_2[$ ] then the last expression in ( [ thetaxy ] ) shows that @xmath546 $ ] . now , the derivative of @xmath547 equals @xmath548 so the minimal distance between @xmath549 and @xmath550 is obtained for @xmath551 . this corresponds to a symmetric configuration in which @xmath552 $ ] with @xmath553 and @xmath554 or equivalently , @xmath555 since @xmath556 whenever @xmath557 we obtain the first claim . the second is a consequence of @xmath558 for all @xmath559 . let @xmath560 be at a distance @xmath561 . through a positive real rescaling , e.g. replacing @xmath562 by @xmath563 for a suitable @xmath564 , we may assure that @xmath565 proof : from the birkhoff characterization ( [ birkhoff def ] ) of the projective distance we may rescale , say @xmath562 , to obtain @xmath566 and @xmath567 . then @xmath568 and @xmath569 . from this we get : @xmath570 and similarly with @xmath562 and @xmath571 interchanged . rearranging terms the claim follows . + let @xmath572 be an inclusion of @xmath113-cones and denote by @xmath573 the inclusion of the corresponding complexified cones . let @xmath574 $ ] and @xmath575 $ ] be the projective diameters of the respective inclusions . then @xmath576 is finite iff @xmath577 is finite . proof : from the embedding in theorem [ thm embedding ] we see that @xmath578 which implies one direction . to see the converse , suppose that @xmath579 and let @xmath580 . possibly after rotating the polarization of @xmath581 and @xmath582 , we may by lemma [ complex polarization ] assume that @xmath583 and @xmath584 with @xmath585 . eventually applying a real rescaling of e.g. @xmath581 ( which does not change its polarization ) , we may by lemma [ real rescaling ] assure that in addition : @xmath586 let us write @xmath587 , @xmath425 and similarly for @xmath588 and @xmath589 . in order to prove our claim it suffices to find a fixed open neighborhood @xmath590 of the segment @xmath291\subset \cc$ ] , depending on @xmath591 but not upon @xmath581 and @xmath582 , such that @xmath592 for every @xmath593 . let @xmath594 . then @xmath595 and we get ( with @xmath596 ) ( a ) : @xmath597 as well as ( b ) : @xmath598 . we also obtain the estimates ( c ) : @xmath599 and ( d ) : @xmath600 . let us write @xmath601 with @xmath594 and @xmath602 . using the expansion @xmath603 and inserting the estimates ( a)-(d ) we obtain @xmath604 & \geq & ( 1-\eta^2 ) \la \ell_1 , x_t\ra \la \ell_2,x_t \ra -\\ & & |z| \eta \sqrt{2 } \sqrt{1+\eta^2 } \left ( \la \ell_1,x_t \ra \la \ell_2,x_1+x_2\ra + \la \ell_1,x_1+x_2 \ra \la \ell_2,x_t\ra\right ) \\ & & -|z|^2 2\eta^2 \la \ell_1,x_1+x_2 \ra \la \ell_2,x_1+x_2\ra\\[2 mm ] & \geq & \la \ell_1 , x_t\ra \la \ell_2,x_t \ra \left(2 - \left(\sqrt{1+\eta^2 } + |z| \frac{\eta \sqrt{2 } } { 1-\eta |t| } \right)^2 \right).\end{aligned}\ ] ] this remains positive when @xmath605 is sufficiently small ( recall that @xmath606 ) . the set of such @xmath601-values thus defines an open connected neighborhood , @xmath607 of the segment @xmath291\subset \cc$ ] . since enlarging a domain decreases hyperbolic distances , we conclude that @xmath608 . + suppose that @xmath181 is a real bounded linear operator , that @xmath303 is a @xmath126-inner regular ( some @xmath127 ) , norm - directed cone and that the real operator @xmath181 maps @xmath303 into a subcone of finite projective diameter in @xmath303 . by birkhoff s theorem , @xcite , the operator has a spectral gap . in view of remark [ remark norm - directed ] and the properties of the canonical complexification shown above , the same conclusion follows when considering the complexified operator acting on the canonically complexified cone . our complex cone contraction thus contains the real contraction as a special case ( but , of course , with a more complicated proof ) . a real operator @xmath275 which contracts a real cone @xmath303 contracts _ a fortiori _ the corresponding complexified cone @xmath460 ( easy ) . it is then natural to ask if this complex contraction may be preserved when adding an imaginary part to the operator . many of our applications below are cast over this idea and has lead us to state an abstract assumption for the action upon @xmath303 and a corresponding complex contraction theorem for complexified cones : let @xmath609 be a contraction of an @xmath113-cone @xmath303 . let @xmath610 be an operator acting upon the corresponding complex banach space . we say that @xmath611 is dominated by @xmath275 with constants @xmath612 provided that for all @xmath613 and @xmath614 : @xmath615 { { \rm re}}\la \ell , mx\ra \la \ell,\barm x \ra & \leq & \beta \la \ell , px\ra \la \ell , p x \ra \mylabel{domin b}\\[2 mm ] | { { \rm im}}\la \ell_1,mx_1\ra \la \ell_2,\barm x_2 \ra | & \leq & \gamma \la \ell_1,px_1\ra \la \ell_2,p x_2\ra \mylabel{domin c } \end{aligned}\ ] ] + we say that @xmath303 is @xmath616-inner regular ( @xmath127 ) if there are @xmath617 and @xmath618 so that @xmath619 the above conditions are @xmath299-invariant and also stable when taking convex combinations . it thus suffices to verify that these conditions hold for subsets , @xmath620 and @xmath621 which are generating for the cone and the dual cone , respectively , i.e. for which : @xmath622 note also that when @xmath303 is inner regular the last condition ( [ aux condition ] ) is automatically satisfied ( choose for @xmath225 an interior point in @xmath303 ) . when @xmath623 an operator @xmath611 verifying the above assumption is essentially real . possibly after multiplication with a complex phase the operator maps @xmath303 into @xmath303 itself . the above condition then reduces to the real cone - dominated condition of theorem [ real domin gap ] . our goal is here to show that the conclusion of that theorem also applies when @xmath611 is allowed to have a non - trivial imaginary part . it turns out that the allowed ` amount ' of imaginary part depends on the rate of contraction of @xmath275 . let @xmath281 be a proper convex cone and let @xmath624 be a strict cone - contraction , i.e. @xmath625 . write @xmath460 for the canonical complexification of @xmath303 . suppose that @xmath610 is @xmath275-dominated ( assumption [ main assumption ] ) with constants that satisfy : @xmath626 then @xmath627 and @xmath628 . if , in addition , @xmath303 is @xmath616-inner regular ( @xmath127 ) and of uniformly bounded sectional aperture then @xmath611 has a spectral gap . proof : let @xmath444 and @xmath629 . we write @xmath630 . the first step is to establish the following inequality ( which , in particular , implies that @xmath631 ) : @xmath632 we will use polarization twice to achieve this . first , write @xmath633 with @xmath634 and @xmath459 . then @xmath635 + i \left[\la \ell_1,m y\ra \la \ell_2,\barm x\ra - \la \ell_1,m x\ra \la \ell_2,\barm y\ra \right]\\ & = & \frac12 \left [ \la \ell_1 , m(x\!+\!y)\ra \la \ell_2 , \barm(x\!+\!y)\ra \!+\ ! \la \ell_1 , m(x\!-\!y)\ra \la \ell_2,\barm ( x\!-\!y)\ra \right ] + \\ & & \frac{i}{2}\left [ \la \ell_1,m(x\!+\!y)\ra \la \ell_2,\barm(x\!-\!y)\ra \!-\ ! \la \ell_1 , m(x\!-\!y)\ra \la \ell_2,\barm(x\!+\!y)\ra \right ] \\ & \equiv & \frac12 [ a ] + \frac{i}2 [ b ] . \end{aligned}\ ] ] since @xmath636 we may use inequality ( [ domin a ] ) of our assumption to deduce : @xmath637 for the second term we have by ( [ domin c ] ) @xmath638 where for the last inequality we used schwarz inequality . from these two estimates we get : @xmath639 we note that ( [ mu mu ] ) is here independent of the choice of polarization . since @xmath640 we see that the elements @xmath641 and @xmath642 are at a projective distance not exceeding @xmath643 . we may then use lemma [ complex polarization ] to rotate the polarization again and write @xmath644 where @xmath645 for all @xmath498 . but then @xmath646 we also obtain @xmath647 so that @xmath648 together with ( [ mu mu ] ) this establishes ( [ eq ell u ] ) . in order to obtain an estimate for @xmath649 we also need the following inequality : @xmath650 this follows by setting @xmath651 in the expression for @xmath4 and @xmath652 above and using the upper bounds ( [ domin b ] ) and ( [ domin c ] ) of our assumption : @xmath653 and the bound @xmath654 as before . + consider @xmath655 . using the polarization identity , proposition [ equiv c ] , we may assume that @xmath656 with @xmath657 so that @xmath658 . then also @xmath659 and with the same bounds for @xmath660 . through a real rescaling , lemma [ real rescaling ] , we may also assume that @xmath661 . we also write @xmath662 with @xmath601 , @xmath663 ( and similarly for @xmath588 and @xmath589 ) . by the choice of polarization @xmath664 so that @xmath665 , i.e. belongs to the complex cone for all @xmath663 . we want to show that when @xmath605 is small enough the same is true for @xmath666 . first note that @xmath667 . applying the inequality ( [ gamma ineq ] ) we deduce that @xmath668 and @xmath669 . using ( [ eq ell u ] ) on @xmath670 and the expansion @xmath671 , we obtain the inequality @xmath672 & \geq & { { \rm re}}\ , \la \ell_1 , mu_t\ra\ , \la \ell_2 , \overline { mu_t}\ra\ \nonumber \\ & & - |z| \ \left(\ \strut |\la \ell_1,m u_t\ra | \ |\la \ell_2,m ( u_1-u_2)\ra | \ ; + \ ; |\la \ell_2,m u_t\ra | \ |\la \ell_1,m ( u_1-u_2)\ra | \ \right ) \nonumber \\ & & - |z|^2 \ |\la \ell_1,m ( u_1-u_2)\ra | \ |\la \ell_1,m ( u_1-u_2)\ra | \nonumber \\[2 mm ] & \geq & \la \ell_1 , p x_t\ra \ , \la \ell_2 p x_t \ra \ \times \left(\frac{\alpha}{\cosh({\delta_p}/{2})}-\gamma + 2 ( \beta+\gamma ) - 2(\beta+\gamma ) \left ( 1 + \frac{|z|}{1-\eta|t| } \right)^2 \right ) . \nonumber \mylabel{c diam}\end{aligned}\ ] ] the set of @xmath673 values , @xmath594 for which the latter quantity is non - negative contains an open neighborhood , @xmath674 of the segment @xmath291\subset \cc$ ] . it follows that @xmath675 for the last assertion note that the condition ( [ aux condition ] ) implies that @xmath676 which remains non - negative when @xmath677 . thus @xmath460 is @xmath616-inner regular . by proposition [ c regular cone ] , @xmath460 is also of uniformly bounded sectional aperture so the conclusion follows from our spectral gap theorem , theorem [ spectral gap ] . the most striking application is also the simplest . a complex perron - frobenius theorem : let @xmath1 and suppose there is @xmath2 for which @xmath678 for all indices . then @xmath4 has a spectral gap . proof : the cone @xmath679 is regular in @xmath430 . by corollary [ corollary cplus ] , @xmath680 is regular in @xmath430 . we will compare @xmath611 with the constant matrix @xmath374 with respect to the real cone @xmath11 . as in example [ ex real pf](1 ) , @xmath402 . the canonical basis and its dual generates the cone and its dual , respectively , cf . remark [ remark generating ] . the constants from assumption [ main assumption ] then become ( sups and infs over all indices ) ( a ) @xmath681 , ( b ) @xmath682 and ( c ) @xmath683 . our spectral gap condition of theorem [ c domin sp gap ] simply reads @xmath684 and by finiteness of @xmath0 this is equivalent to the stated assumptions on @xmath4 . + in the following , denote by @xmath685 the essential oscillation of a real valued function @xmath220 on a measured space . theorem [ thm pf complex ] may ( almost ) be viewed as a special case of the following complex version of a result of jentzsch @xcite : let @xmath380 be a measure space and let @xmath686 , with @xmath687 . let @xmath688 , @xmath384 a.e . and @xmath689 , @xmath386 a.e . with @xmath690 so that @xmath691 , cf . example [ ex real jentzsch](2 ) . given @xmath692 we define the integral operator , @xmath693 : @xmath694 set @xmath695 and @xmath696 . suppose that @xmath697 and that @xmath698 . then @xmath699 has a spectral gap . proof : as in example [ ex real jentzsch](2 ) we consider the @xmath113-cone @xmath700 and we compare with @xmath701 . we have that @xmath702 and @xmath402 . we obtain the following estimate for the constants @xmath703 since @xmath704 , @xmath705 and @xmath706 , the cone @xmath303 is @xmath699-inner regular . as shown in example [ ex real jentzsch ] ( 2 ) the real cone has bounded sectional aperture so by proposition [ c regular cone ] ( 3 ) , the same is true for the complexified cone . the spectral gap condition of theorem [ c domin sp gap ] then translates into the stated condition on @xmath539 and @xmath707 . let @xmath115 be a complex banach space . we denote by @xmath708 denote the set of complex planes in @xmath115 , i.e. subsets of the form @xmath709 with @xmath22 and @xmath23 independent vectors in @xmath115 . if we write @xmath710 for the unit sphere in @xmath115 then @xmath711 defines a metric on @xmath708 . in the following let us fix a norm on @xmath429 . the choice may affect the constants below but is otherwise immaterial . the space @xmath712 is then a sequentially compact metric space . let @xmath713 be a @xmath9-cone and let @xmath714 . suppose there is @xmath715 , @xmath128 such that @xmath716 . then @xmath717 has at most @xmath718 bounded aperture . proof : let @xmath719 be a linear functional with @xmath720 and @xmath721 . choose @xmath722 for which @xmath723 . if @xmath724 then @xmath725 so that @xmath726 and therefore , @xmath727 the 2-dimensional space @xmath326 is spanned by @xmath159 and @xmath22 so @xmath728 . + let @xmath729 be a @xmath9-cone . then there is @xmath730 so that @xmath54 is of @xmath133-bounded sectional aperture . proof : suppose that this is not the case . then we may find a sequence @xmath731 of planes for which the aperture @xmath732 diverges . taking a subsequence we may assume that @xmath731 converges in @xmath708 to a plane @xmath326 . as @xmath54 is proper , @xmath733 is a strict subset of @xmath326 . thus there is @xmath734 . but @xmath54 is closed in @xmath429 so there is @xmath128 so that @xmath735 is disjoint from @xmath54 as well . given another complex plane , @xmath736 , we may find @xmath737 for which @xmath738 . when @xmath326 and @xmath736 are close enough , @xmath739 and @xmath740 is also disjoint from @xmath54 . by our previous lemma , @xmath741 is of aperture not exceeding @xmath742 . but this contradicts the divergence of @xmath743 as @xmath744 . + let @xmath729 be a @xmath9-cone and let @xmath745 be a closed complex sub - cone with @xmath746 . then there is @xmath747 such that for @xmath748 : @xmath749 proof : we denote by @xmath750 the canonical projection to complex projective space . we consider @xmath751 as a metric space with the metric @xmath752 as in ( [ proj metric ] ) . the projected image , @xmath753 , is compact in the open set @xmath754 so there is @xmath755 for which the @xmath756-neighborhood of @xmath753 is contained in @xmath757 . let @xmath748 be linearly independent and suppose that @xmath758 . let @xmath759 be the complex plane containing @xmath22 and @xmath23 . denote by @xmath760 the connected set in @xmath761 containing @xmath22 and @xmath23 . let @xmath762 , @xmath763 be an @xmath764-maximally separated set in @xmath760 . thus , the balls @xmath765 , @xmath763 are all disjoint and @xmath766 . the cardinality of @xmath433 is bounded by a constant depending on @xmath756 only . then @xmath767 , @xmath763 so by lemma [ xp lower bound](2 ) each @xmath768 , @xmath763 is of radius not greater than @xmath769 for the @xmath770-metric . also @xmath771 contains @xmath760 which is connected . it follows that @xmath772 does not exceed @xmath773 . which is bounded by a constant depending on @xmath756 only . + let @xmath729 be a closed subset which is @xmath9-invariant and contains no complex planes ( in terms of definition [ def complex cone ] , @xmath54 is a @xmath9-cone ) . suppose that @xmath774 is a linear map for which @xmath775 . then @xmath4 has a spectral gap . proof : we write @xmath776 for the image of @xmath54 and use the notation and constants from the two previous lemmas . first note that for @xmath777 , @xmath778 to see this note that when @xmath779 then @xmath780 so by lemma [ wv cone ] , @xmath781 . if @xmath782 denotes the connected component of @xmath783 containing @xmath22 and @xmath23 then also @xmath784 . by lemma [ lemma contraction ] , @xmath785 . iterating this argument we see that @xmath786 , @xmath187 . by lemma [ unif k aperture ] , @xmath54 is of @xmath133-bounded sectional aperture , so lemma [ xp lower bound](1 ) assures that @xmath787 . so that @xmath789 . now let @xmath790 , @xmath763 be an @xmath764-maximally separated set in @xmath489 . setting @xmath791 with @xmath763 we see that @xmath792 , @xmath793 . it follows that there is a map , @xmath794 so that @xmath795 , @xmath793 . since @xmath433 is of finite cardinality , @xmath796 must have a cycle . thus , there are @xmath797 and @xmath788 for which @xmath798 . the cone @xmath799 is regular ( easy ) and of bounded diameter in @xmath800 so @xmath801 has a spectral gap and therefore also @xmath4 . + when the operator is sufficiently regular one may weaken the assumptions on the contraction and the outer regularity of the cone . this is illustrated by the following complex version of a theorem of kren and rutman ( * ? ? ? * theorem 6.3 ) : let @xmath802 be a @xmath9-cone in the banach space @xmath420 . let @xmath803 be a quasi - compact operator or a compact operator of strictly positive spectral radius and suppose that @xmath4 verifies @xmath804 then @xmath4 has a spectral gap . proof : let @xmath275 be the spectral projection associated with eigenvalues on the spectral radius circle , @xmath805 . by hypothesis @xmath806 is finite dimensional and we may find @xmath634 such that @xmath807 we claim that @xmath808 is non - empty : let @xmath809 and define @xmath810 . suppose first that @xmath811 . then @xmath812 so that the distance between @xmath186 and @xmath806 tends to zero . since @xmath813 is locally compact and @xmath186 is bounded we may extract a convergent sub - sequence @xmath814 . suppose instead that @xmath815 then @xmath816 so there is @xmath128 for which @xmath817 . we may then replace @xmath22 by @xmath818 where @xmath819 , @xmath820 and we are back in the first case . thus @xmath821 . now , @xmath822 the latter for the topology in @xmath806 . in particular , @xmath823 is non - empty so @xmath824 is an inner regular @xmath9-cone in a finite dimensional space and @xmath825 . we may then apply the finite dimensional contraction theorem , theorem [ thm finite dim ] , to @xmath826 . it follows that @xmath827 , whence also our original operator @xmath4 has a spectral gap . + in the real cone version ( replacing @xmath9 by @xmath113 ) of theorem [ complex krein rutman ] it is not necessary to assume that the spectral radius of @xmath4 is strictly positive . this forms part of the conclusion . to see this pick @xmath809 of norm one . then @xmath816 so there is @xmath349 for which @xmath828 . therefore , @xmath829 and then also @xmath830 by the properties of an @xmath113-cone . more generally , @xmath831 . as @xmath832 it follows that @xmath833{|a^nx| } \geq \lambda > 0 .\ ] ] the fact that this conclusion is non - trivial is illustrated e.g. by the operator , @xmath834 , @xmath835 , which is compact when acting upon @xmath836)$ ] . it contracts ( though not strictly ) the cone of positive elements but has spectral radius zero . in the complex setup , if one assumes that @xmath26 is of @xmath133-bounded sectional aperture then strict positivity of @xmath5 also comes for free : suppose that @xmath837 , @xmath838 and @xmath839 , @xmath128 . then @xmath840 , @xmath841 and also @xmath842 for such @xmath6-values . by lemma [ lemma inner regular ] we see that @xmath843 from which @xmath844 . the ruelle - perron - frobenius theorem , @xcite ( see also @xcite ) , ensures a spectral gap for certain classes of real , positive operators with applications in statistical mechanics and dynamical systems . ferrero and schmitt @xcite used birkhoff s theorem on cone contraction to give a conceptually new proof of the ruelle - perron - frobenius theorem . see also @xcite and @xcite for further applications in dynamical systems . we present here a generalization to a complex setup . let @xmath845 be a metric space of finite diameter , @xmath846 . when @xmath847 ( or @xmath9 ) we denote by @xmath848 $ ] the associated lipschitz constant and by @xmath849 the supremum . then @xmath850 ( and similarly for @xmath420 ) is a banach algebra . let @xmath851 and let @xmath852 be an unramified covering map of @xmath853 which is uniformly expanding . for simplicity , we will take it to be of finite degree ( it is an instructive exercise to extend theorem [ thm rpf ] below to maps of countable degree ) . more precisely , we assume that there is @xmath854 and a finite index set @xmath433 so that for every couple @xmath855 we have a pairing @xmath856 of the pre - images , @xmath857 and @xmath858 , for which @xmath859 , @xmath860 . + fix an element @xmath861 and define for @xmath862 ( or @xmath705 ) : @xmath863 the norm of @xmath699 when acting upon @xmath864 ( in the uniform norm ) is given by @xmath865 and a straight - forward calculation shows that @xmath866 with @xmath867 . + [ thm rpf ] denote @xmath868 , @xmath869 and @xmath870 . suppose that @xmath871 then @xmath872 has a spectral gap . proof : we will compare @xmath699 with the real operator @xmath873 . for @xmath279 the set , @xmath874 defines a proper convex cone in @xmath280 which in addition is regular . inner regularity : let @xmath875 , @xmath876 and @xmath877 . then @xmath878 provided @xmath879 . whence @xmath880 . outer regularity : pick @xmath881 and set @xmath882 . for @xmath883 we have @xmath884 so that @xmath885 , and this shows outer regularity . let @xmath886 and @xmath887 . as in ( [ birkhoff beta ] ) let @xmath888 . a calculation using the defining properties of the cone - family yields : @xmath889 and we get the following bound for the diameter @xmath890 , cf . ( [ birkhoff def ] ) : @xmath891 the injection @xmath892 is thus a uniform contraction for the respective projective metrics . given @xmath883 and using the pairing @xmath893 we get for the operator @xmath894 : @xmath895 this implies that @xmath896 with @xmath897 . if we choose @xmath898 then @xmath275 becomes a strict cone contraction of the regular cone @xmath899 . we also get the estimate ( to obtain an a priori estimate for the contraction one may here try to optimize for the value of @xmath900 ) : @xmath901 by theorem [ thm birkhoff ] , @xmath609 has a spectral gap ( see @xcite and also @xcite ) . + returning to the complex operator , @xmath699 , let us fix @xmath855 and the corresponding pairing of pre - images @xmath893 as described above . let @xmath902 and write @xmath903 with @xmath904 in order to compare with the real operator , we define complex numbers @xmath905 , @xmath906 , through the relation @xmath907 equivalently ( when the denominator is non - zero ) : @xmath908 we may apply lemma [ exp estimate ] below with the bounds @xmath909 and @xmath910 to deduce that @xmath911 given @xmath912 we have : @xmath913 the two last factors are real and non - negative ( because @xmath914 ) and the complex pre - factor belongs to the set @xmath915 summing over all indices we therefore obtain @xmath916 in which @xmath315 is an average of numbers in @xmath4 whence belongs to @xmath917 , the convex hull of @xmath4 . when @xmath918 we conclude that the bounds in assumption [ main assumption ] are verified for the constants @xmath919 , @xmath920 and @xmath921 . the spectral gap condition in theorem [ c domin sp gap ] then reads as follows : @xmath922 now , in order to get a more tractable and explicit formula we make the following ( not optimal ) choice for @xmath900 : @xmath923 then @xmath924 so that @xmath925 . using ( [ delta p estimate ] ) we obtain @xmath926 one also checks that @xmath927 implies that @xmath928 so we may replace ( [ sp gap cond ] ) by the stronger condition @xmath929 finally inserting @xmath930 we obtain the claimed condition which is thus sufficient for a spectral gap . + in the literature , one often includes a statement on gibbs measures as well . if we let @xmath931 denote the leading spectral projection of @xmath873 , then positivity of @xmath275 implies that the ` state ' @xmath932 is uniformly bounded with respect to @xmath849 . by continuity , @xmath933 extends to a linear functional on @xmath934 . if , in addition , we assume @xmath853 compact , then by riesz , this functional defines a borel probability measure @xmath935 on @xmath853 . the measure is invariant and strongly mixing for @xmath936 . it is known as a gibbs measure for @xmath936 and the weight @xmath937 . this part of the theorem , however , needs the partial ordering induced by the cone of positive continuous functions and does not extend to a complex setup ( in general , it is even false there ) . in the proof we made use of the following complex estimate : let @xmath938 be such that @xmath939 and define @xmath940 through @xmath941 then @xmath942 proof : writing @xmath943 and @xmath944 we have : @xmath945 taking real and imaginary parts , @xmath946 and @xmath947 , we get @xmath948 . also @xmath949 so that @xmath950 . let @xmath115 be a complex banach space . given non - zero elements @xmath951 we write @xmath952 iff @xmath953 . let @xmath954 denote the quotient map and write @xmath955=\cc^*x$ ] for the equivalence class of @xmath956 . we equip the quotient space @xmath957 with the following metric @xmath958,[y ] ) = \dist_h ( \cc x \cap s , \cc y \cap s ) = \inf \left\ { \left\| \frac{\mu x}{\|\mu x\|}\!-\ ! \frac{\nu y}{\|\nu y\| } \right\| : \mu,\nu\in \cc^*\right\ } , \ \ x , y\in x^ * \label{proj metric}\ ] ] in which @xmath959 is the hausdorff distance between non - empty sets and @xmath960 is the unit - sphere . 1 . let @xmath116 be a @xmath9-cone of @xmath133-bounded sectional aperture . then for all @xmath162 : @xmath958,[y])\leq 2 k d_{\calc}(x , y).\ ] ] 2 . let @xmath956 , @xmath128 and set @xmath961,r)$ ] . then for all @xmath962 @xmath963,[y])}{r - d_{\pi(x^*)}([x],[y])}.\ ] ] proof : using the inequality @xmath964 we obtain from lemma [ bounded sectional aperture ] : @xmath965 and the first conclusion follows . for the second claim , normalize so that @xmath966,[y])=\|x - y\|<r$ ] and @xmath967 . let @xmath968 . by ( [ xy inequality ] ) , @xmath969 which remains smaller than @xmath970 when @xmath971 $ ] . then @xmath972 + given any two points @xmath46 we may follow kobayashi @xcite and define a projective pseudo - distance between @xmath22 and @xmath23 through : @xmath973 since @xmath974 is a ( projective ) metric , the previous lemma implies that in other words , this new distance does not degenerate when taking the inf over finite chains , so distinct complex lines in @xmath26 have a non - zero @xmath977-distance . this is conceptually very nice , but , in our context , not particularly useful . the reason is that even if @xmath120 maps @xmath782 into a subset of finite diameter in @xmath782 for the metric @xmath978 this does not seem to imply a uniform contraction of @xmath181 , i.e. no spectral gap . we leave a further study of this metric to the interested reader . + i am indepted to a. douady for the elegant argument in the proof of lemma [ lemma contraction ] . p. ferrero and b. schmitt , ruelle s perron - frobenius theorem and projective metrics , _ coll . soc , j bollyai _ * 27 * ( 1979 ) .

josephson junction arrays ( jja ) have been widely studied , both theoretically and experimentally@xcite . they show a number of phenomena , vortex physics@xcite , quantum phase transitions . they are also a paradigm ; a large variety of physical systems can be modelled using josephson junctions arrays , for example , one - dimensional thin superconducting wires@xcite and granular superconductors . nowadays , they are used to realize protected qbits@xcite or for metrological applications , for establishing current standards through frequency measurements of shapiro steps@xcite . here , we shall restrict ourselves to the 1-dimensional case . furthermore , we shall neglect dissipation , which is crucial in some quantum wires . no coupling to a dissipative environment is considered . for normal regular junction chains , the essential physics is the competition between josephson coupling and local coulomb interactions . in one dimension , true long range order can not exist , the correlations of the superconducting order parameter decay at best as a power law . when local coulomb interactions are increased , the power law is modified , and eventually a berezinskii - kosterlitz - thouless ( bkt ) transition occurs . beyond the transition , when coulomb interaction dominates , the decay of correlations becomes exponential . this is due to the unbinding of phase slips , whose existence have been shown directly in recent experiments @xcite . topological defects have always played a central role , from bkt transitions in magnets , roughening of interfaces , disordered luttinger liquids and some stripe models for high-@xmath2 superconductors . we would like in this paper to study the role of long range coulomb interactions , hereafter denoted as lri , which have not been taken into account so far in one - dimensional jj chains . one motivation is that , in one - dimensional clean superconducting wires , the long range part of the coulomb interactions is poorly screened . for example , the plasma mode frequency goes to zero for small frequency@xcite . this has motivated some recent studies@xcite . clearly , in these systems , fluctuations of the phase of the order parameter have to be taken into account . the hamiltonian looks very much like the one of a one - dimensional josephson array , with a long range coulomb term in addition . long range interactions , hereafter denoted by lri , were studied a long time ago at the spin - wave level for electronic systems@xcite . they also appear naturally in modelling underdoped high-@xmath2 superconductors , where phase fluctuations of the order parameter are important@xcite . influence of lri have also been studied recently@xcite theoretically in two and three dimensions . our goal is to obtain the phase correlator of the superconducting order parameter as a function of time and space . this quantity governs numerous physical properties . in this paper , we shall give only one application . namely , we calculate the average andreev current from a normal metal to a jj chain . because lri are normally screened in fabricated jj chains , the model discussed here is not meant to be a realistic model for a jj chain but can be relevant to thin 1-dimensional superconducting wires . however , calculations developped in this paper can be adapted to take into account the screening of the coulomb interaction between islands and make predictions about andreev current in real jj chains . after the introduction , the body of the paper is organized in four parts . in part two , we first describe the model of jj chain we use , and its connection to high-@xmath2 materials . it is convenient to go to imaginary time . when the josephson coupling dominates over the coulomb repulsion , if phase slips are neglected , the modes are spin waves . the long range interactions cause the decay of correlations to change from those of a luttinger liquid to a more pronounced decay@xcite . in the next section , we take into account phase slips . the hamiltonian decouples into a part describing spin - waves ( long wavelength excitations ) and another part describing phase slips ( rapid variations of the phase ) . in the last sections , we discuss the interaction between phase slips . for distances much smaller than a characteristic distance , they interact logarithmically . beyond this distance , the interaction behaves only as the square root of the logarithm of the distance@xcite . in part three , we examine how the kt real space renormalization group ( rg ) analysis is modified by lri . there , as soon as lri are present , strictly speaking , no quasi long range order is possible anymore . however , the initial long range interactions can be very weak . if we iterate the rg flow , it will take some time for lri to alter the behavior of the usual kt flow . the lri will show up only after a certain crossover `` time '' , which we calculate and corresponds to some temperature @xmath3 . on the other hand , rg has to be stopped at some cutoff `` time '' , @xmath4 . therefore , in order to see the influence of lri , the temperature must be lower than @xmath3 . the dependence of @xmath3 on the values of parameters such as the coulomb energy , josephson coupling and strength of lri is explicited . part four is devoted to the calculation of the phase correlation function . analytical continuation is required to obtain the correlator in real time . this part is rather technical ; some aspects of the calculations are relegated to appendices . the main result is the expression of the phase correlator , which differs from the usual ornstein - zernike expression . in the last part , we discuss one physical situation . the andreev current between a metal and a jj chain , as a function of applied voltage is calculated . we compare the curves obtained in three cases . the first case corresponds to only on - site coulomb interaction . in the second case , a coulomb interaction between nearest - neighbor islands is added . this model is often used in the literature . finally , in the third case , lri are added . for relatively low voltages , the andreev current remains always smaller with lri . we focuse on the case where , without lri , we would be in the disordered phase . there is a threshold voltage @xmath1 below which no andreev current can flow . this remains true in the presence of lri . however , the current versus voltage ( @xmath0 ) curves are different from the case without lri . without lri , the andreev current starts as @xmath5 , whereas with lri , it starts to increase faster than any power law . this can be viewed as a signature of lri and could be tested in experiments . in the conclusion , we discuss two possible applications , one for realistic fabricated jj chains and the other for one - dimensional superconducting wires . for jj chains , the interaction between island charges is screened but the methods used in this paper can be adapted to make at least qualitative predictions . for 1d wires , owing to the strong dependence of a phase slip fugacity with temperature and wire width , and also for other reasons detailed below , it is not obvious that lri can be detected experimentally . maybe they could for very long wires . we would like to study the influence of poorly screened coulomb interactions in 1 dimension . one possibility , which has been widely used some time ago , mainly in dimensions two and three , would be to take the hamiltonian for a bcs superconductor @xcite and to introduce hubbard - stratonovich transformations in the particle - hole and particle - particle channels in order to make the usual gap @xmath7 appear . the amplitude fluctuations are integrated out , so are the phase of the density fluctuations , leaving us with a phase action only . this procedure has been used in ref . , except that they stayed at the rpa level and did not consider topological excitations for the phase @xmath8 of the order parameter . another way would be to start with a one - dimensional array of josephson junctions , modelled by an ideal josephson junction ( jj ) and a capacity @xmath9 in parallel . in addition , on each island , the capacity to the ground has @xmath10 to be taken into account . cm the hamiltonian is @xmath11 with @xmath12 and the interaction potential @xmath13 and @xmath14 is the capacitance matrix . the last term of eq . ( 1 ) represents lri , @xmath15 is a constant and @xmath16 a capacity . usually , only the capacity of one junction @xmath9 and the capacity from an island to the ground @xmath10 are taken into account , the capacity matrix @xmath17 has the form @xmath18 if @xmath19 , @xmath20 if @xmath21 , and @xmath22 otherwise . it is reasonable to approximate the interaction potential@xcite by the form @xmath23 with @xmath24 . the screening length is of the order @xmath25 , with @xmath26 the interjunction spacing . typically , for one - dimensional fabricated chains@xcite @xmath27 but still @xmath28 is of order @xmath29 to @xmath30 , not more . lri thus do not appear to be completely realistic for fabricated jj chains , but for thin superconducting wires , the coulomb interaction is badly screened@xcite and can be longer range . we shall thus take a hamiltonian of the form @xmath31 with @xmath32 typically an effective capacitance to the ground . we can also see this hamiltonian as a 1d analog of the following hamiltonian for superconductors with fluctuations of the phase of the order parameter @xmath33 in this case , @xmath34 is the superfluid density in 3d . @xmath35 is the charge density , @xmath36 will denote the phase of the order parameter . its conjugate variable is @xmath37 , which is proportional to the charge @xmath38 . when phase slips are neglected , the procedure to obtain an effective action for @xmath36 is standard and we just recall it here for completeness . writing @xmath39 with @xmath40 integers , @xmath41 and @xmath40 are conjugate variables , @xmath42 , @xmath43 with @xmath44 the charging energy for nearest neighbor . following ref . , @xmath45 denoting the temperature and @xmath46 , the partition function is calculated in imaginary time , by discretizing in time steps of size @xmath47 . the system without lri would be equivalent to a usual xy model with isotropic coupling . in the following , @xmath48 will no longer be operators but numbers ; index @xmath49 from 1 to @xmath50 denotes the position in real space and the second index @xmath51 from @xmath52 to @xmath53 the position on the imaginary time axis . making @xmath54 , the partition function reads @xmath55 where the sum is on all configurations @xmath56 . now , in the spin - wave approximation , the cosines are expanded and the lagrangean @xmath57 and action @xmath58 take the form @xmath59 applying a fourier transform and integrating out the @xmath60 field , @xmath61 where @xmath62 is the fourier transform of the coulomb potential , regularized@xcite at very small @xmath38 . @xmath63 is in fact @xmath64 , and @xmath65 is the fourier transform in space of @xmath48 . rescaling the imaginary time and counting it in units of @xmath47 , and setting @xmath54 gives the effective action on a discrete space - time lattice @xmath66 with @xmath67 and @xmath68 the fourier transform in space of @xmath48 defined as @xmath69 for @xmath38 multiples of @xmath70 . now , taking into account the periodic character of @xmath36 , we can take the effective action @xmath71 having derived an effective action , we first turn to the case where the periodic character of the phase can be neglected . the approximation is good for large @xmath72 . cooper pairs feel very little coulomb - blockade . the fugacity of phase slips goes essentially@xcite as @xmath73 . phase slips are neglected and a phase correlator derived . owing to lri , the system is no more a luttinger liquid . correlations ressemble some correlators appearing in one - dimensional quantum wires with low electron densities@xcite , sometimes dubbed `` wigner crystals '' . in this section , we allow only smooth variations of @xmath36 . it is then possible to expand the cosine term . this gives immediately a spin - wave hamiltonian . we check here the known spin wave dispersion relations . @xmath74 with @xmath75 for small @xmath38 . setting @xmath76 , this results in @xmath77 and thus @xmath78 where @xmath46 and @xmath79 are the bosonic matsubara frequencies . for distances @xmath80 and times @xmath81 not too long , so that the long range term does not influence the behavior , the correlations are those of a luttinger liquid @xmath82 , where @xmath83 is a short range cutoff ( corresponding to finite bandwidth ) , whereas for very large times@xcite , @xmath84 if @xmath81 is counted in units of @xmath85 . the correlation function of @xmath86 for spin - waves read @xmath87 with @xmath88 . we would like to calculate the full correlation function , not only in the spin wave approximation , but including phase - slips . this turns out to be more intricate , we will neglect all factors that vary slower than any power law , such as , for example , logarithmic corrections or terms as @xmath89 with @xmath90 a constant . within this approximation , we have the spin - wave correlator @xmath91 now , we take into account phase slips , that is vortices for the @xmath6 model in 1 space and 1 time dimension . the procedure to split the action into two parts , one due to spin - waves and the other to vortices is standard without lri , however we recall it here with some details to point out the influence of lri . we use the model of villain and transform the exponential of the cosine term in ( [ effectiveaction ] ) as the sum over @xmath92 integers but with a quadratic action@xcite , involving @xmath36 and @xmath92 . the space direction will be denoted as @xmath80 and the imaginary time direction as @xmath93 . the long range interaction occurs only in the @xmath80 direction , for bonds directed in the @xmath93 direction . let us denote by @xmath94 as a consequence , the coulomb term , i.e. the second term on the r.h.s of eq . ( [ effectiveaction ] ) , has to be modified from @xmath95 to @xmath96 where @xmath97 denotes the ( @xmath98 times ) integer attached to the bond which goes from site @xmath99 to site @xmath100 , and pertaining to the @xmath93 direction . the @xmath101 are simply the fourier transforms of the @xmath102 . the lattice spacing will be denoted @xmath26 , and @xmath103 ( @xmath104 ) will be the unit vector in the @xmath80 ( @xmath93 ) direction . the part not involving the @xmath105 s is of the type @xmath106 so that finally the total action is of the form @xmath107 where @xmath108 and @xmath109 , @xmath110 where @xmath103 and @xmath104 are the unit lattice vectors . @xmath111 is an infinitesimally small positive quantity . from now on , we will denote @xmath112 so that @xmath113 has components @xmath114 . the only difference with the normal xy model is that the coupling constant in the @xmath93 direction has a dependence in @xmath115 for small @xmath115 ( long wavelengths ) . setting @xmath116 and @xmath117 , @xmath118 diagonalizing this action yields the decomposition into a continuous field @xmath119 and a discrete field @xmath120 . the action must be of the form @xmath121 where @xmath83 and @xmath122 take two values , @xmath80 and @xmath93 . the quantities @xmath123 and @xmath124 are functions of @xmath113 . the linear transformation is given explicitly by @xmath125 and @xmath126 if @xmath127 and zero otherwise . next , @xmath128 is rewritten in terms of vortices defined as @xmath129 which yields the vortex hamiltonian @xmath130 the apparent divergence at @xmath131 is in fact absent because of charge neutrality @xmath132 . going back to real space , denoting @xmath133 , the total number of sites : @xmath134 enables to rewrite the vortex hamiltonian as @xmath135 with , taking into account values of @xmath136 , @xmath137 and @xmath138 , @xmath139 first , the self - interaction @xmath140 has to be withdrawn , but it is not so crucial because of charge neutrality ( @xmath141 ) . second , the vortex hamiltonian as it stands is incomplete . the vortex fugacity term , of the form @xmath142 is missing , with @xmath143 being the fugacity of a vortex . in fact , one should start with a generalized villain model @xcite , introduced by jos , kadanoff , kirkpatrick and nelson , hereafter denoted by jkkn . as noted in ref . , such a term is naturally generated by real space ( migdal - kadanoff ) renormalization . it is therefore necessary to reinstall this term . next , we look at the shape of the interaction between vortices . although it is known that , with lri , the vortex - vortex coupling for large separations @xcite ( much larger than the lattice constant ) goes only as in @xmath144 , we need to derive a quantitative expression . the behavior is dominated by the behavior of the integral for small @xmath145 . for theses small wave - vectors , @xmath146 and @xmath147 , @xmath26 being the lattice constant . we can take @xmath148 and @xmath149 . @xmath150 the cosine term will be close to @xmath52 only for @xmath115 much lower than @xmath151 , and @xmath152 much lower than @xmath153 . to extract the main behavior for large distances @xmath154 , we can replace it by the integral @xmath155 where @xmath156 is a wave - vector much larger than @xmath157 , but still sufficiently small that the expansion of @xmath158 and @xmath159 for small @xmath160 holds . the integral reads @xmath161 as @xmath154 gets large , for @xmath162 different from @xmath163 and @xmath164 , @xmath165 goes to infinity , so the first @xmath166 term in the bracket is close to @xmath164 whereas the second @xmath166 term remains much smaller than @xmath52 , provided @xmath167 is not too small . we can thus safely replace the term into brackets by @xmath164 . the integral on @xmath115 is then readily performed and is in @xmath168 . the limit @xmath169 is singular . it is better to discuss limiting cases . case 1 : @xmath170 @xmath171 , this corresponds to distances where the decrease of the correlations due to long range unscreened coulomb interactions still has not taken place . we have then @xmath172 case 2 : @xmath170 @xmath173 , this corresponds to distances where the decrease of the correlations due to long range unscreened coulomb interactions has taken place . we have then @xmath174 thus , at very large distances , there is no longer sufficient interaction between vortices to hold them tied together , whatever the short range term is @xcite . mathematically the bkt transition is expected to disappear . however , we need to put this on a more quantitative basis . furthermore , in real systems , the coulomb interactions remain unscreened only up to a certain screening length @xmath175 , which can be large but is not infinite . also , the magnitude of the lri coulomb term may be much smaller than the one of the local coulomb term . in order to determine the phase diagram and the phase correlations , we need to implement a real space renormalization group scheme analog to the one developped in refs . . care has to be taken since lri will eventually dominate under the rg group . the last term in eq . ( 2 ) is more relevant than the second one . however , under the rg fow , it will take some time for the last term ( lri ) to overcome the effect of the local one ( i.e. the second term in eq . the situation is reminiscent of lri in one - dimensional quantum wires with very low electron densities ; there are important differences however . there will be remnants of the bkt transition . to proceed , we perform a rg in real space closely following kosterlitzoriginal paper @xcite . a quick look at the situation is that , owing to the coulomb potential , at very large distances @xmath176 , the interaction between vortices goes like @xmath177 ( instead of @xmath178 in the usual case ) , and is too weak to overcome the entropy term for placing a free vortex , which goes as @xmath178 . nevertheless , the last terms in eqs . ( [ hjja],[hsupra1d ] ) can be small compared to the second ones , even if they are less relevant in rg sense , their initial value can be much smaller . thus , it makes sense to treat them as a small perturbation , at least initially . this leads us to investigate how the usual kt rg flow is modified . here , the partition function has a form which is slightly different from the usual kt flow . the gas is neutral , without electric field . @xmath92 will be the number of `` @xmath179 '' charges . @xmath180 where @xmath181 is a coupling constant , @xmath182 and @xmath183 is a function that behaves as @xmath52 for not too large @xmath184 but decays as @xmath185 for very large @xmath186 . a model @xmath183 can be @xmath187 @xmath81 is the short time cutoff and @xmath188 the fugacity . in the case where the long range term is much smaller than the short range one ( at the beginning of the renormalization procedure only ) , we can however follow the same steps as kosterlitz . the details are shown in appendix a. we express the rg equation , using the variables @xmath189 now , with the long range interaction , this becomes @xmath190 with @xmath191 . note that @xmath192 can be much smaller than one . now , because @xmath192 is non - zero , the line @xmath193 is no longer a line of ( stable or unstable ) fixed points . the flow is no longer autonomous , as soon as @xmath194 . we expect a deformation of the usual flow . in particular , the line @xmath193 which was a line of fixed points in regular kt flow is no longer a true line of fixed points . cm if the lri were absent , we could be either in the case where phase correlations are decaying exponentially ( disordered phase ) or in the `` ordered '' phase , where correlations decay as a power law . in the first case , adding the long range interactions will just make the phase correlation length smaller . the more interesting case is when we start from the ordered phase . we thus examine quantitatively how switching lri drives us into the disordered phase and , in the next section , we shall examine what the phase correlations look like . in the usual case , ( without lri ) , the flow is autonomous ( i.e. does not depend explicitly on time ) and trajectories are of the form @xmath195 , where @xmath196 is the starting point . setting @xmath197 , they are hyperbolae in the @xmath198 plane . in the `` ordered '' phase , the flow on the invariant curves is : @xmath199 which integrates to @xmath200 @xmath201 is the initial `` time '' which can be set to zero . integration to obtain @xmath202 is also straightforward . the important thing is that , as @xmath203 becomes very large , @xmath204 , which is related to the fugacity , tends to zero exponentially , while @xmath202 tends to a strictly positive value @xmath205 , see fig . 2 . in our case , ( with lri ) , we can see what happens by performing the following approximation : - if @xmath206 is much smaller than @xmath204 , we neglect it , - when , it becomes of the order of @xmath204 , we do the opposite and neglect @xmath204 in the r. h. s. of the @xmath207 expression , eq . ( [ firstflowequation ] ) . thus , there are at least two stages . in the first stage , we have the usual kt flow whereas in the second stage the long range term comes into play . we shall see that there are three stages actually . we now describe the first stage . starting from a definite value of @xmath204 , by the regular kt flow , @xmath204 will decrease . suppose that @xmath204 has decreased by a substantial amount , such that @xmath204 will be much smaller than @xmath208 . this enables to replace @xmath209 by @xmath210 . inverting the relation yields @xmath211 with @xmath212 . ( @xmath213 is not necessarily much larger than one ) . the physics is that @xmath208 is proportional to the initial distance to the separatrix . it is proportional to @xmath214 in the thermal x - y model . now , we have to stop when @xmath204 reaches the value @xmath215 , because we are no longer allowed to neglect the term @xmath216 in eq . ( [ firstflowequation ] ) . this happens when @xmath217 to get the main feeling , we can show that the typical value of @xmath203 for which this happens is @xmath218 , such that @xmath219 . this corresponds to a typical imaginary time @xmath220 , or corresponding lengthscale @xmath221 such that @xmath222 neglecting @xmath223 contributions , @xmath224 , the corresponding value of @xmath204 is @xmath225 @xmath224 goes to zero as @xmath192 goes to zero . now , we will enter the second stage . we now describe this second stage , where the second term in eq . ( [ firstflowequation ] ) becomes larger than @xmath204 . then , for @xmath203 larger than @xmath226 , we integrate the rg equations , neglecting the @xmath204 term with respect to @xmath206 . @xmath227 and , using the other flow equation to obtain @xmath204 , @xmath228 then , @xmath202 will decrease and @xmath204 will continue to decrease , till @xmath202 passes the value @xmath163 . @xmath204 will be minimum for @xmath229 and then will start to increase . in fact , it increases faster than exponentially , so it will quickly become larger than @xmath230 , so that our approximation will fail . the value of @xmath231 for which this happens can be estimated , but anyway @xmath232 , corresponding to a length @xmath233 . then , there is a third stage , where @xmath204 becomes larger than @xmath234 . then , we can again neglect @xmath235 in front of @xmath204 . we are then in the third regime which has analogies with the kt flow but on the disordered part of the transition . this lasts till @xmath204 has reached an appreciable value such that one is no longer in the vicinity of the critical point . let @xmath236 be the corresponding value of @xmath203 , and @xmath237 . @xmath238 will be also of order @xmath239 . since @xmath208 is proportional to @xmath240 , where @xmath2 is the transition temperature without lri , it is tempting to conclude that , since @xmath192 goes as @xmath241 , the correlation length with lri @xmath242 behaves as @xmath243 with @xmath244 an unimportant constant of order one . it has the same form as the kt correlation length in the disordered phase ( without lri ) but @xmath245 replaces the usual@xcite constant @xmath246 . of course , @xmath242 tends to infinity as @xmath192 tends to zero . the general shape of the `` flow diagram '' is best seen on fig . 2 . renormalization has to be stopped when @xmath81 becomes equal to @xmath247 . equating @xmath242 to @xmath248 yields a temperature @xmath3 given by @xmath249 with @xmath2 the critical temperature in the absence of lri . if experiments are performed at a temperature larger than @xmath3 , it will not be possible to see the influence of lri . having the general shape of the `` flow - diagram '' , we now turn to the determination of the phase correlation function . this section is a little technical . just as the hamiltonian decouples into a vortex part , the phase correlation function also do . estimation of the spin - wave part is straightforward , but the estimation of the vortex part is more intricate . in section a , we first determine the contribution of one vortex to the phase correlator . this involves a vortex influence function @xmath250 . comparison between the usual case and the case with lri is emphasized . then , in section b , average must be carried out over the possible configurations of vortices , using the boltzmann factor with the vortex hamiltonian . a coulomb gas formulation is necessary . distinction must be done between smooth variations of the vortex influence function @xmath250 , studied in section c , and abrupt variations ( section d ) ; there are thus two contributions . summing both contributions gives the vortex part of the phase corrrelator . adding the spin - wave part yields the total phase correlator in imaginary time . in section f , analytical continuation is performed to obtain it in real time . the phase correlator in real time is necessary for the calculation of the transport properties of jj chains . for the determination of the correlation function , we use the method of jkkn . in the vortex hamitonian , the core energy has to added . such a term naturally arises from their migdal - kadanoff renormalization group . therefore , we must use the generalized villain lodel . the corresponding vortex hamiltonian is @xmath251 with @xmath93 the charge fugacity . the @xmath252 are integer charges living on the dual lattice of the original lattice . they represent vortices . the correlation function to be computed is @xmath253 . written in coulomb gas language , this correlation function disentangles into a spin - wave part and a vortex part . we closely follow the method @xcite of pelcovits . the first thing to do is to determine what is the influence of a vortex sitting at @xmath254 on the vortex part of the correlation function @xmath255 . the total correlation function is @xmath256 following jkkn , @xmath257 is written as a product of nearest neighbor correlation functions , along a path @xmath258 that connects @xmath259 to @xmath260 on the real lattice . this is for the total correlation function ( vortices plus spin - waves ) . after ( gaussian ) integration of the spin - wave degrees of freedom , @xmath261 the vortex part , @xmath255 is given by @xmath262 with @xmath263 where @xmath254 are all the positions on the dual lattice , the function @xmath264 ( @xmath265 ) is zero everywhere except on sites of the dual lattice that lie immediately on the left ( right ) of the path chosen joining @xmath259 to @xmath266 . @xmath267 is simply @xmath268 , see eq . ( [ vrhorhoprime ] ) . the brackets mean thermodynamical average with respect to @xmath269 . this results directly from gaussian integration , it can be understood as follows : the contribution of a vortex sitting at location @xmath254 , and vorticity @xmath270 is @xmath271 . the function @xmath272 depends on @xmath259 and @xmath266 . it has also another physical interpretation . it is the phase generated at site @xmath254 by a vortex - antivortex pair , where one vortex has @xmath273 and is sitting on @xmath259 and the antivortex ( @xmath274 ) is sitting on @xmath266 . the determination of @xmath272 has been done in the literature without the long range coulomb interaction ( lri ) . this function will be called @xmath275 , and the potential energy between vortices will be denoted as @xmath276 . we explain how the lri affect the results . following jkkn , without lri , @xmath277 where @xmath278 . 3 . \7 . cm for not too large distances , the interaction between vortices is still logarithmic and we just recall the known results for completeness . for simplicity , we take a path joining @xmath259 and @xmath266 to be a straight line ( see fig . 3 ) . working in cartesian coordinates , and taking the origin at @xmath259 , the @xmath80 axis , along the line going from @xmath259 to @xmath266 , the coordinates of @xmath254 , @xmath259 and @xmath260 will be @xmath279 , @xmath280 and @xmath281 respectively . the current integration variable @xmath282 on the dual lattice immediately above the cut will be of coordinate @xmath283 . @xmath284 which gives @xmath285 with @xmath286 , if @xmath287 . now , consider the lri , at large distances @xmath288 the function @xmath275 at large distances must be replaced now by @xmath272 . @xmath289 integrating by parts , to extract the main contribution at large distances , yields the result @xmath290 with @xmath291 a quantity which is smaller by a factor @xmath292 than the main one . thus , @xmath272 behaves@xcite as @xmath293 for large @xmath254 . from eq . ( [ vortexpartofg ] ) , the average of @xmath294 has to be performed with the full vortex hamiltonian ( including core energy ) . the boltzmann weight of a particular configuration @xmath270 is given by @xmath295 with @xmath296 and thus @xmath297 a high temperature expansion of @xmath255 is needed . applying the poisson formula @xmath298 to the function @xmath299 and integrating over the continuous auxiliary field @xmath300 enables to write it as @xmath301 with @xmath302 where the sum is over all configurations of @xmath303 . @xmath304 is the correlation function calculated with weight @xmath305 but as though the @xmath306 were continous variables . @xmath307 is the same as @xmath308 except that @xmath272 is absent . however , because a switch from the @xmath270 variables to the @xmath303 variables ( through the poisson formula ) has been made , now a low temperature expansion is needed . a high @xmath45 expansion in the @xmath270 variables transcribes into a low @xmath45 expansion in the @xmath303 variables . in the absence of lri , @xmath309 has been calculated in refs . . even with the lri , the correlation @xmath304 remains short range and can be approximated by nearest - neighbor only . substituting the exact @xmath304 by a nearest neighbor interaction results in @xmath310 following refs . , a low temperature expansion is performed , keeping only the essential diagrams . to simplify further , only the non - overhanging self - avoiding - walks ( saw s ) joining @xmath254 to @xmath266 are kept which results in @xmath311 where @xmath312 means integration over all the plane except on the saw . @xmath313 is the number of non - overhanging saw s of @xmath92 steps joining @xmath259 and @xmath266 . the calculation of the correlation function splits into the integration on the smooth variations of the field @xmath272 and on the abrupt discontinuities caused by the different diagrams occuring in the high @xmath45 expansion of @xmath308 . in our case , the situation looks a little more complicated because @xmath272 does not satisfy the laplace equation . however , this causes no trouble because the volume contributions coming from @xmath314 , where @xmath315 is the laplacian , will turn out to be very small . the main contributions come from the regions near the path @xmath258 going from @xmath259 to @xmath266 . more precisely , the weight @xmath316 is to be evaluated , where the prime means integrations everywhere in the plane except on the diagrams . in the usual case , integration is made by parts @xmath317 @xmath318 denotes the shaded aera in fig . 4 , @xmath319 is a contour which encloses the cut ( from @xmath259 to @xmath266 ) at a small distance @xmath320 and @xmath321 is a circle of radius @xmath322 where @xmath322 will eventually be allowed to go to @xmath323 . @xmath324 is the unit normal to the graph . on fig . 4 , it is simply the normal to the line joining @xmath259 to @xmath266 . the first term @xmath325 vanishes since @xmath326 . the second term @xmath327 will tend to zero as @xmath322 goes to infinity . the last term @xmath328 has to be calculated using eq . ( [ uder ] ) . care has to be taken that @xmath329 behaves only as @xmath330 for sufficiently large distances and this form of @xmath331 should not be used for small distances . cm with this in mind , taking @xmath332 and thus @xmath333 , @xmath334 the factor @xmath335 comes from the fact that integration above and below the cut @xmath336 has to be performed ( see fig . 4 ) . we take the integration path a finite amount @xmath320 above ( or below ) the cut to avoid spurious divergencies ( see above ) . then @xmath320 will be unimportant . the angle difference in the difference of the two @xmath166 functions eq . ( [ t3 ] ) will correspond to the change in angle , taken to be @xmath337 . dropping the spurious @xmath338 terms yields@xcite exactly @xmath339 . now , we turn to the case of lri . the analog of the terms @xmath340 , @xmath341 to @xmath342 will be denoted as @xmath343 and @xmath344 has to be replaced by @xmath345 . calculations are tedious and relegated to the appendices . it turns out that @xmath346 and @xmath347 are completely negligible whereas @xmath348 , with @xmath349 a constant . even after taking the exponential of this will only lead to logarithmic factors for the correlation functions and we thus omit this factor . here , the part due to abrupt variations of the field @xmath272 , is the same as without lri sum on saw joining @xmath259 and @xmath260 and the result is analogous to the usual case , namely , the contribution , called @xmath350 reads : @xmath351 @xmath352 is the plasma frequency @xmath353 . being the result of a high - t expansion , this expression corresponds to short correlation length @xmath354 . it is probably not valid for very large @xmath354 . in the limit @xmath355 , the usual divergence close to the bkt transition must be retrieved . the spin - wave part , the vortex part coming from smooth variations of @xmath272 and the vortex part coming from abrupt variations of @xmath272 just need to be multiplied . thus we have , with @xmath356 , @xmath357 with @xmath349 a constant . contrary to the usual case , the part of the vortex contribution corresponding to smooth variations of @xmath358 no longer cancels the spin - wave contribution . essentially @xmath359 where we systematically dropped all contributions going slower than any power law . we see that , contrary to what happens in the usual case , the total correlator is not of the ornstein - zernike type . however , in the limit @xmath360 , the ornstein - zernike form is retrieved . it is now necessary to obtain the phase correlator in real time , and then in real frequency . we need to perform an analytical continuation @xmath361 . we do not attempt to get the short time behavior . this would correspond to high voltages . at long times , @xmath362 where @xmath363 means `` proportional to '' . normalization of @xmath364 would require more information on the shape of @xmath365 for shorter times . this yields in real frequencies @xmath366 , for small @xmath366 s ( with respect to @xmath352 ) @xmath367 where @xmath368 is the heaviside function . for larger @xmath366 , this form is not valid . one would need the short time behavior of @xmath365 , which we do not have explicitly . however , on those short timescales , the long range part of the interaction will play a minor role . therefore , in order to perform calculations , a reasonable assumption is that , as without lri , @xmath369 is negligible@xcite for @xmath366 larger than @xmath370 , with @xmath371 . enforcing the normalization @xmath372 , we thus have the approximate form @xmath373 with @xmath374 , and the function @xmath375 is defined by @xmath376 . having obtained the phase correlator in real time , we then use it for the calculation of the andreev current . a normal metal is placed on the left of the josephson junction chain and biased by a voltage @xmath377 , below the gap @xmath378 , see fig . 1 . the average andreev current flowing from the metal into the superconductor can be readily expressed as a function of the distribution of phase modes @xmath369 . the case @xmath379 corresponds to a usual bcs superconductor , and btk results@xcite are retrieved . the case where @xmath369 is given by the luttinger liquid theory or usual 2d @xmath6 model has been studied in refs . . the contact between the metal and the jj chain can range from tunnel to perfect contact . following the keldysh formalism , and performing a nonperturbative expansion , to all orders in the tunneling amplitude , one can describe all situations@xcite . without phase fluctuations , the procedure has been shown to be exactly equivalent to solving the bogolubov - de gennes ( bdg ) equations . with phase fluctuations , it is only an approximation . the average current@xcite reads : @xmath380 with @xmath381 @xmath382 being the bandwidth and @xmath383 a hopping parameter . for tunnel contacts , @xmath384 . a perfect contact corresponds to @xmath385 exactly . see ref . for details . we would like to compare two cases , with and without lri . for jj chains , lri are usually not taken into account . one exception are the models of jj chain considered in ref . , where long range interactions in the capacitance matrix were examined . these models yield physical results which are esentially the same as the usual xy model , except in the case where capacitance of the islands to the ground , @xmath386 is strictly zero . only in this case does the bkt transition disappear and the chain is always in the disordered phase , because of the proliferation of unbound vortices . in most experimental situations , it does not seem to be realistic to assume that @xmath386 is really completely negligible with respect to nearest - neighbor islands capacitive coupling ( or even with respect to longer range inter - island capacitive couplings ) . one exception may be the case of the experiments of ref . . by contrast , the model studied in this paper always keeps a non - zero capacitance to the ground and still always gives unbound vortices . now , we would like to compare the results of our calculation with a model commonly used in the literature . in this latter model , a nearest - neighbor interaction bewteen island charges is included , and there is no lri . we discuss qualitatively this case and compare the predictions from such a model with our case . a term of the form @xmath387 must be added to the hamiltonian , see eq.([hjja ] ) . qualitatively , this will add some charge fluctuations and should decrease the phase correlations . in a more quantitative way , this will change for example the correlator of @xmath36 in fourier space eq . ( [ thetaqw ] ) to @xmath388 where @xmath26 is the spacing between josephson junctions . calculations are straightforward and the main changes with respect to the case with just a local coulomb term are the following . in the quasi long range phase , for frequencies much smaller than @xmath352 , the plasma frequency is now @xmath389 , and thus the constant @xmath390 has to be changed accordingly to @xmath391 . this no longer holds for frequencies of order @xmath352 . the critical point will be shifted . defining @xmath392 , the usual bkt transition which occurred at @xmath393 will now take place at @xmath394 . in the disordered phase , but still no too far from criticality , the correlation length will be depressed . however , the phase correlator will retain approximately its ornstein - zernike form , eq . ( [ oznormal ] ) , with different values of @xmath352 and @xmath354 . the important thing is that the pre - exponential factor is still in @xmath395 . as an example , we take a perfectly transmitting interface and a situation where the josephson chain would be insulating ( disordered phase ) even in absence of lri . the presence of lri will modify the correlation length @xmath354 but also alter the form of the correlator @xmath369 . then , @xmath396 is a constant and the current assumes the form @xmath397 with @xmath398 @xmath399 remains zero for voltages smaller than @xmath400 , just as without lri . the new thing is that , for voltages slightly larger than @xmath1 , @xmath399 no longer behaves as @xmath5 . we introduce the dimensionless quantity @xmath401 and the corresponding voltage @xmath402 . @xmath403 is a crossover value of the voltage , not a sharp threshold as @xmath1 . for voltages @xmath377 , such that @xmath404 , then , @xmath405 and @xmath399 behaves as @xmath406 with @xmath407 a constant . thus , the current starts to increase faster than any power law of @xmath408 . if , on the other hand , @xmath377 is larger than @xmath403 , then , @xmath399 increases again as @xmath5 . for even larger voltages , properties of the phase correlator at short time scales are needed . we expect that , the lri will not be relevant on these scales and that the linear behavior of the @xmath0 curves in the usual case will be retrieved . this is illustrated in fig . 5 , where we plot @xmath399 in units of @xmath409 versus @xmath410 , for @xmath411 . since our calculations describe only moderate lri , for the lri parameter @xmath15 , we take @xmath412 which results in @xmath413 . for the inverse correlation `` length '' @xmath414 , we take@xcite it to be @xmath415 . this serves just as a rough estimate since the exponential behavior of the correlation length applies only ( above and ) close to the usual bkt transition point . it leads to @xmath416 , to be compared to @xmath417 . we see that the andreev current is strongly depressed by lri . we also compare with the model without lri but with nearest neighbor coulomb interactions , mentioned above . we have taken @xmath418 . this causes the voltage threshold @xmath1 to be increased to a slightly higher value @xmath419 , but the current still starts as @xmath420 ; see fig . ( 5 ) . to be more quantitative , we make a logarithmic plot , on fig . 6 , of @xmath399 versus @xmath421 . whereas the curve without lri , for @xmath422 has a slope @xmath423 , this is not the case with lri . the solid curve catches only with the curve without lri , for @xmath424 of the order of @xmath425 to @xmath52 , which is slightly larger than of order @xmath426 . for the model with nearest - neighbor interaction , we plot instead @xmath399 versus @xmath427 . the curve ( dash - dotted line ) is still very close to a straight line of slope @xmath423 . in the case where the josephson chain would be in the ( quasi ) ordered phase , @xmath428 , without lri , lri impose a correlation length @xmath429 of order @xmath430 and thus a threshold voltage @xmath431 . a study of the influence of unscreened long range coulomb interaction was carried out for one - dimensional array of josephson junctions . this can also have implications for one - dimensional thin superconducting wires . we focussed on the case where the initial term describing lri was much smaller than josephson coupling and local coulomb blockade term . deriving a hamiltonian describing the interaction of phase slips , we examined how the kt flow is modified by lri . implications on andreev reflexion between a normal metal and a superconductor was examined and lri introduce a new voltage scale . the analytical form of the @xmath0 characteristics was derived for any value of the transparency of the normal - superconductor interface , ( from tunnel contact to perfect interface ) . if we start from a situation where the charging energy dominates the josephson coupling , corresponding to the jj chain being insulating , the introduction of lri changes the behaviour above the threshold voltage @xmath432 from a @xmath433 behavior to a sharper increase , faster than any power - law of @xmath434 . this lasts up to a crossover voltage @xmath403 , whose value depends on the strength of the long range coulomb interactions . if , on the contrary , we start from a situation where the josephson coupling dominates without lri ( quasi ordered phase ) , then lri will always generate a threshold voltage @xmath1 below which no andreev current can flow . this threshold @xmath1 goes as @xmath435 where @xmath15 is proportional to the strength of the lri and @xmath436 is the distance to the critical point in the absence of lri , with @xmath437 and @xmath438 . with lri , above @xmath1 , the current starts to increase faster than any power law . the form of the phase correlator obtained in this paper could be used for calculation of other physical quantities relevant to jj chains or to one - dimensional superconducting thin quantum wires . such calculations are not directly relevant for high-@xmath2 superconductors@xcite , because of their layered structure and their strong anisotropy in the direction transverse to the planes , whereas the former calculation remained striclty one - dimensional . for designed jj chains , the inverse capacitance is generally cut exponentially with a screening length of the order of @xmath439 , with usually @xmath440 , ( @xmath26 is the spacing between neighboring junctions ) , @xcite . the number @xmath441 of junctions@xcite can be up to @xmath442 , so that @xmath443 . thus , the model considered here should not be construed as a model for fabricated jj chains but rather an attempt to include lri , which do exist in isolated thin wires . however , we can use the above calculations , replacing the badly screened coulomb potential by the exponentially screened potential @xmath444 , see eq . ( [ uij ] ) . this leads qualitatively to an interaction between vortices @xmath445 which has the following properties . for distances @xmath446 much smaller than @xmath447 , @xmath448 , with @xmath449 a constant of the order of ten the strength of the interaction between vortices decays very rapidly , ( much faster than for the badly screened coulomb interaction ) . this lasts till @xmath446 becomes of the order of @xmath447 , where @xmath445 stops to decrease and behaves as @xmath450 , for large distances , i.e. @xmath451 , theoretically , the rg group is started at coarse graining scales smaller than @xmath447 and as the rg flow proceeds , the interaction between vortices gets smaller . this may tend to drive towards the disordered region . this is not unexpected , given the large amount of coulomb repulsion provided by the charge coupling of the islands which are less than @xmath452 apart . when rg proceeds further and the coarse - graining length becomes larger than @xmath447 , then , the interaction @xmath268 bewteen vortices ceases to decay and becomes logarithmic with distance . according to the value of the coupling reached at that moment , we can flow either to the disorded region or to a quasi - ordered fixed point . the screening of the coulomb interaction renders quasi long range - ordering more favorable . in particular , contrary to the badly screened interaction case , phase correlations do not always decay exponentially at the end . let us suppose first that @xmath453 is sufficiently large so that we end in the ( quasi)-ordered phase , @xmath369 will decay as a power law for small @xmath366 , with some exponent @xmath83 , for @xmath366 smaller than @xmath454 . for @xmath366 larger than @xmath454 , the decay of @xmath369 corresponds to a larger effective vortex interaction , and phase correlations are enhanced , leading to a larger @xmath369 . as we go to large distances ( or longer times or smaller voltages ) , vortex interactions get reduced until the distance @xmath447 is reached . this is due large coulomb repulsion at shorter lengthscales . this has experimental consequences on the shape of @xmath455 vs. @xmath377 curves . frequencies smaller than @xmath454 correspond to voltages smaller than @xmath456 . for these voltages , @xmath0 curves will start as power law with some exponent , normally as @xmath457 with @xmath458 , where @xmath459 and @xmath460 are renormalized values of @xmath461 and @xmath453 . when @xmath377 becomes larger than @xmath456 , the current is enhanced and increases faster than @xmath457 . this remains true for voltages still smaller than @xmath456 . now , let us suppose that @xmath453 is not so large and one ends in the disordered phase . there is a threshold voltage but beyond @xmath432 , the current starts to increase as @xmath433 in the usual theory . inclusion of screened interaction leads to an increase which is faster than @xmath433 , which could also be tested experimentally . the model with badly screened interactions may also have applications for very long ( but not unrealistically long ) 1d superconductor thin wires . more precisely in thin wires ( typically @xmath342 to @xmath462 @xmath463 diameter ) , the interaction between phase slips has the same structure as the hamiltonian ( 1 ) , but the physical origin is somewhat different @xcite . first , in concrete situations , it is not the two - dimensional btk transition which is seen but rather a one - dimensional schmid transition@xcite . we briefly recall the reasons below . in experiments with very homogenous thin wires in the difusive limit , of diameter 3 to 10 nm , such as the ones used in refs . the maximum length @xmath57 is about @xmath464 . operating temperatures are between @xmath465 and a few kelvin . if the length were infinite , renormalization had to be stopped when coarse graining length @xmath81 on the imaginary time axis reaches @xmath247 . however , in practise , two - dimensional renormalization has to be stopped when it reaches @xmath466 where @xmath467 is the mooj - schn velocity@xcite . @xmath468 is equivalent of the spin - wave velocity , where @xmath469 is the capacity per unit length and @xmath470 the diameter of the wire . @xmath467 is of the order @xmath471 , so that for a @xmath472 long wire , @xmath473 is of the order @xmath474 and for temperatures lower than @xmath475 , renormalization has to be stopped first on the @xmath80 axis , when the coarse graining length reaches the size of the wire . we end up with charges in 1 dimension ( as for a single josephson junction ) which interact logarithmically as @xmath476 , with @xmath477 a renormalized coupling . then , renormalization proceeds solely on the imaginary time direction . depending on the value of @xmath477 , fugacity of vortices renormalizes to zero or not , resulting in the usual schmid transition . the spatial structure on the x axis has been completely lost and the long - range interaction do not play any role , except that they changed slightly the values of @xmath477 . thus , in this case , lri do not seem to play an important role . thus , in order to see the influence of lri , longer wires are needed . for longer wires ( but not totally unrealistic , a few microns long ) , we can expect temperatures larger than @xmath478 , so that renormalization has to be cut on the imaginary time axis first . then , lri will drive the resistance @xmath176 to a finite value . if @xmath176 versus @xmath45 curves are plotted , no curve should plunge towards zero resistance , unless @xmath45 becomes of the order of @xmath478 , which is about @xmath479 for @xmath480 . there are difficulties to see lri . one could be that stray capacities ( unavoidable in experiments ) can also cut the long - range term in the capacitance for small @xmath38 ( in real space ) so that @xmath481 is no longer of the form @xmath482 for small @xmath160 . this depends on the charateristics of the wires . another problem is that the action for a single phase slip , which controls its bare fugacity , is strongly dependent on temperature and on the diameter of the wire . this dependence seems to be more important than the interaction bewteen phase - slips in determining the shape of the curves showing the resistance @xmath176 versus @xmath45 , for different wire thicknesses @xcite . long range interactions are originally not screened in isolated one - dimensional wires . however , the presence of an environment could alter the screening . situations where a plate or a wire is placed in the vicinity of a superconducting thin wire are now experimentally@xcite and theoretically studied @xcite . this aims at providing mechanisms of dissipation so as to damp the phase fluctuations . though we did not consider in this paper coupling to an external environment , our calculation may be viewed as a first step towards understanding the role of lri , including , both spin waves and phase slips ( solitons ) . in this appendix , we derive the rg flow in the presence of long - range interactions . we define @xmath483 as @xmath484 with @xmath485 integration has to be done over @xmath486 in the whole plane except the two circles of radii @xmath81 around @xmath487 and @xmath488 . setting @xmath489 and denoting by @xmath188 the vortex fugacity , the partition function @xmath307 reads @xmath490 where @xmath491 is the whole plane except aera inside the circle of radii @xmath81 around the @xmath492 , @xmath493 . let us denote by @xmath494 where @xmath449 is the whole plane except the two circles of radii @xmath81 around @xmath487 and @xmath488 . given the value of @xmath495 , see eq . ( [ fcaleq ] ) , we have @xmath496 for @xmath497 . denote by @xmath498 where @xmath58 is the aera . collecting all the terms , @xmath499 using @xmath500 , for small @xmath80 , @xmath501 if @xmath507 were a constant , we would retrieve the usual kt flow . however , since we are looking at larger and larger lengthscales , the effective coupling will not only decrease because of rg but also because of its @xmath508 dependence . the smallest @xmath176 contributing to @xmath307 is @xmath81 . thus , adding a term proportional to @xmath509 will take this into account . now , the partition function has to be rearranged so as to look as its original form , with renormalized parameters and an overall multiplicative factor , setting @xmath510 , we obtain @xmath511 switching to the usual variables @xmath512 and linearizing around @xmath229 , the flow reads @xmath513 with @xmath514 . here , we give details on the contributions to the vortex part of the correlation function @xmath255 coming from smooth variations of @xmath515 . the first term to evaluate is @xmath516 where the double integral extends over the whole shaded surface in fig . @xmath315 is the laplacian operator ( with respect to variable @xmath254 ) . in this calculation , @xmath260 and @xmath259 are fixed , and , without restriction , @xmath259 is taken to be zero for simplicity . if @xmath176 is the modulus of @xmath254 , since @xmath345 decreases as @xmath517 for large @xmath176 , it turns out that @xmath314 behaves , as @xmath518 . thus , @xmath519 behaves as @xmath520 . when surface integration is performed , the integral tends to a constant , for @xmath521 because of the @xmath522 factor . when exponentiated this will give a constant . since , we shall neglect all terms that depend on @xmath523 , weaker than any power law , we shall neglect this factor . the second term is a line integral over the outer contour @xmath321 on fig . 4 , @xmath524 where @xmath324 is the normal to the contour @xmath321 which is here chosen as a circle of radius @xmath322 . for all @xmath254 s belonging to @xmath321 , @xmath176 are equal to @xmath322 and @xmath345 are of order @xmath525 , the leading terms in @xmath526 are of order @xmath527 . thus @xmath327 is at most of order @xmath528 and vanishes as @xmath322 goes to infinity . the third term to evaluate is a line integral over the inner contour @xmath319 . 4 . @xmath529 a calculation completely analogous to the case without lri shows that the leading terms are of order @xmath530 , with @xmath523 the modulus of @xmath531 . this behaves , for large @xmath523 as @xmath532 . v. l. berezinskii , zh . * 59 * , 907 ( 1970 ) , @xmath533 sov . jetp * 32 * , 493 ( 1971 ) @xmath534 ; j. m kosterlitz and d. j. thouless , j. phys . c * 6 * , 1181 ( 1973 ) . j. m. kosterlitz , j. phys . c * 7 * , 1046 ( 1974 ) .

the study of resonances in scattering systems with chaotic classical dynamics is an active field of research . it is at the same time driven by physicists being interested in the quantum properties of open chaotic systems as well as by mathematicians being interested in geometric analysis and partial differential equations ( pde ) on non - compact manifolds . while in scattering theory , resonances are poles of the scattering matrix and correspond to quasi - stationary quantum states with purely in- or outgoing boundary conditions , from the pde point of view resonances are often seen as the spectral invariants defined as the poles of a meromorphic continuation of the resolvent operator . aside from some very special exceptions ( e.g. the walsh quantized baker map @xcite ) the position of the resonances of such complex systems can not be determined exactly . it has , however , been a very successful approach to gain coarser information on general properties of the resonance distribution in terms of the underlying classical dynamics and geometry , respectively . for example results on spectral gaps @xcite predict resonance free regions and the fractal weyl law @xcite gives an upper bound on the growth of the number of the so - called `` long - living '' resonances in the high frequency limit . despite the large progress that has been made in obtaining rigorous results on the distribution of the resonances , there are still many open questions ( we refer to @xcite for an overview over recent results and open questions ) . for example the upper bound on the exponential growth of the resonance density which has been proven in the fractal weyl law @xcite is only conjectured to be sharp and there exists also a conjecture on the improvement of the spectral gap estimates @xcite . it has thus been a very important approach to study the resonance structure numerically and by physical experiments in order to test the present conjectures and to show that the results on resonance distributions can be observed in real physical systems . there have been established two paradigmatic scattering systems for these kind of tests : the @xmath0-disk systems and schottky surfaces . a @xmath0-disk system describes the scattering of one particle at @xmath0-circular obstacles in the euclidean plane ( see ) . it has been introduced by gaspard - rice @xcite and cvitanovi - eckhardt @xcite in physics and by ikawa @xcite in mathematics . in physics especially the completely symmetric 3-disk system , where 3 disks of the same size are arranged on a regular triangle , enjoys great popularity . it is already complex enough to have a classical dynamic with a non - trivial , fractal repeller and is still simple enough to be treated efficiently by numerics . there exist very efficient approaches to calculate the quantum resonances based on a direct scattering matrix ansatz @xcite as well as by a semiclassical approach using zeta functions @xcite . consequently it has been among the first systems to numerically observe the fractal weyl law @xcite . further importance arises from its experimental realizability @xcite and in recent microwave experiments the spectral gap @xcite and first hints of the fractal weyl law @xcite have been found . from a mathematical point of view the disadvantage of the @xmath0-disk system is that many of the numerical algorithms are not based on mathematical rigorous results . for example the conjecture that the zeros of the semiclassical zeta function are related to the quantum resonances , which has been thoroughly tested in physics @xcite , is mathematically open ( cf . * section 1.6 - 1.8 ) for recent results into this direction ) . this is the reason for the popularity of schottky surfaces as an explicit model system in mathematics . schottky surfaces are two - dimensional non - compact surfaces with constant negative curvature which are convex co - compact ( see for a sketch of a 3-funneled schottky surface ) . the constant negative curvature allows to prove that the zeros of the selberg zeta function are related to the resonances of the laplacian @xcite and not only approximately as conjectured for the 3-disk system in physics . furthermore jenkinson - pollicott @xcite derived , analogously to the cycle expansion of the 3-disk system , efficient formulas to numerically compute the zeros of the selberg zeta functions . the big drawback of these surfaces from the physical point of view is that there is no known experimental realization ( see @xcite for attempts to realize the wave equation on negatively curved surfaces with water waves ) . as the curvature of these surfaces is , however , strictly negative , the geodesic flow is hyperbolic and these surfaces are an interesting model for open chaotic systems . additionally the study of the resonances on constant negative curvature surfaces and the relation to their geometry is a mathematically interesting question itself ( see e.g. @xcite ) thus we will call these resonances _ geometric resonances_. the most simple example of schottky surfaces which already has a non - trivial fractal repeller are the 3-funneled surfaces as shown in the upper part of . for these surfaces borthwick recently presented a thorough numerical study of their resonances structure @xcite . one of the most surprising features of this resonance structure is an observation which has before already been made for the 3-disk system @xcite : even though the underlying classical dynamic is completely hyperbolic , i.e. strongly chaotic , the distribution of the resonances is highly ordered and the resonances form a striking chain structure in the complex plane ( see ) . the 3-disk system and the schottky surfaces are , however , not the only systems to embody these chain structures . very similar chains have for example also been observed in higher dimensional systems , as the 4-sphere billiard @xcite , and even in microdisk resonators which are important models for physical applications such as microdisk lasers @xcite . furthermore the phenomenon of resonance chains is not only restricted to resonances of the laplace operator and quantum resonances , respectively . also the spectrum of the classical ruelle resonances which describe the decay of correlations in classical dynamical systems has been reported to reveal these clear resonance chains @xcite . ( upper right ) and the quantum 3-disk system with @xmath1 ( lower ) . note : the resonances for the schottky surfaces are so dense that they appear as continuous lines in the plot above , however , they are discrete ( see ) . , scaledwidth=80.0% ] though these chains are eye - catching , little is known about these structures . for the completely symmetric 3-disk system and the completely symmetric 3-funneled schottky surface , the distance between the resonances on one chain has been observed to be related to the length of the shortest periodic orbit @xcite . additionally the oscillation length of the chains has been found to be related to the length difference between two fundamental periodic orbits . recently a numerical study of the 3-disk system @xcite showed that the resonance chains are not only a visual effect , but that the resonances on one chain share common physical properties . it was , however , shown that the resonance chains are not caused by scarring of the scattering states along individual classical orbits , but are related to the interatction of several classical orbits . two fundamental questions are , however , still open and will be addressed in this article . first of all the resonance chains , are a visual impression , i.e. our brain groups the resonances to different chains and we have the impression that they are strung along different continuous lines . in this article we will give a system intrinsic definition of these continuous lines . secondly , resonance chains have been observed in many , quite different systems . we will give a unifying condition on the classical length spectrum that is able to explain and predict the existence of resonance chains . the article is organized as follows : in section [ sec:3-disk ] we treat the resonance chains of the symmetric 3-disk system . after a general introduction of 3-disk systems ( section [ sec : intro_3-disk ] ) we will numerically illustrate that the resonance chains of the symmetric 3-disk system are created by fast rotating eigenvalues of a suitable transfer operator on the canonical poincar section . this idea applies for the semiclassical as well as for the classical resonances ( section [ sec : rot_res ] ) . then we give an introduction to schottky surfaces ( section [ sec : intro_schottky ] ) and show in section [ sec : gen_zeta ] that a straightforward application of this idea on the resonance spectrum of completely symmetric schottky surfaces fails to explain the resonance chains . however , the ideas can be abstracted . this leads us to the introduction of a generalized zeta function and the generalized spectrum which are shown to be the right objects to describe and understand the resonance chains . these objects immediately also allow to deal with non - symmetric schottky surfaces and 3-disk systems and also gives a condition for the existence of resonance chains ( section [ sec : existence_of_chains ] ) . in section [ sec : num_test ] we will finally test this condition in different cases and use its predictive power to identify systems incorporating different coexistent chain structures . as this article addresses a physical as well as a mathematical readership we decided to always state clearly which results and relations are known to be true by mathematically rigorous proofs and which results are achieved by physical arguments or experiments . this will help to keep an overview which parts of the argumentation are based on rigorous arguments and where further mathematical work remains to be done . this article itself does not claim to derive any rigorous statements either . instead our aim is to gain a universal understanding of the appearance of the resonance chains and to identify the mathematical quantities , which are crucial for their understanding , via numerical investigations . as the question , whether a chain structure is visible or not , is mathematically not well posed but depends on the observer of the resonance structure , our condition can not be a _ if and only if _ condition either . the condition established in section [ sec : gen_zeta ] is thus an approximate condition and our hypothesis is that the better this condition is fulfilled the clearer the chains will be visible . this hypothesis is tested and supported by various numerical results presented in section [ sec : num_test ] . a rigorouse proof of the existence of resonance will be given in @xcite , where the ideas and insights presented in this article are an essential ingredient . a 3-disk system describes one free particle in an euclidean 2-dimensional space that scatters at three non - overlapping hard disks @xmath2 and @xmath3 . for further use we will call the union of the disks by @xmath4 . classically its trajectories are given by straight lines and hard wall reflections at the disk boundaries @xmath5 ( see left panel of ) . its quantum dynamics is described by the laplacian @xmath6 with dirichlet boundary conditions at the disk boundaries and its quantum resonances can rigorously be defined as the poles of the resolvent @xmath7 this resolvent , as a map between compactly supported @xmath8 functions and locally @xmath8-integrable functions , is known to extend meromorphically to the @xmath9 where @xmath10 is any line connecting zero with infinity @xcite . these poles do indeed coincide with the physical scattering resonances i.e. if @xmath11 is a pole with multiplicity @xmath12 , then there exist @xmath12 solutions of @xmath13 with dirichlet and purely outgoing boundary conditions . the latter condition means that the solution @xmath14 in polar coordinates @xmath15 can be written asymptotically in the limit @xmath16 as @xcite @xmath17 contrary purely incoming boundary conditions would correspond to @xmath18 here @xmath19 are constant coefficients and for fixed @xmath10 the asymptotic @xmath20 dependence equals the asymptotic dependence of a free particle state with angular momentum @xmath10 . an important class of 3-disk systems are those which are completely symmetric . in those systems all three disks have the same size and the disks are arranged on an equilateral triangle ( see left part of ) . historically the side length of the triangle is denoted by @xmath21 and the disk radius by @xmath22 @xcite . these systems are then completely characterized , up to a scaling , by the ratio @xmath23 . as the system is highly symmetric and possesses the symmetry group @xmath3 of the equilateral triangle often the symmetry reduced system is studied in physics . this system consists of an open billiard representing the fundamental domain of the 3-disk system ( see light shaded region in the left part of ) . as already mentioned before the classical time evolution of a particle between two disk reflections is just a linear propagation interspersed with hard wall reflections . the trajectory of a particle therefore does not depend on the absolute value of its momentum and we can always scale it to @xmath24 . it is furthermore convenient to get rid of the parts of linear propagation where no complex dynamic takes place and consider only the poincar section of disk reflections . a particle at the boundary of a disk @xmath25 of radius @xmath26 whose momentum is scaled to one , is completely described by its birkhoff coordinates @xmath27 . they are defined as the arc length @xmath28 $ ] on the disk boundary as well as the projection of the normalized incoming momentum to the unit tangent vector at the disk boundary @xmath29 $ ] ( see ) . the complete space @xmath30 of the poincar section consists of three copies of @xmath31\times[-1,1]$ ] . as the system is open the dynamics are not defined on all @xmath30 , because in many cases a particle with birkhoff coordinates @xmath27 will just escape to infinity after the reflection and will not hit any of the other disks in the future . we therefore introduce the forward trapped set @xmath32 of order @xmath0 as the set of all points in @xmath30 which will hit at least @xmath0 disks on their future trajectories . analogously we define the backward trapped set @xmath33 for the time inverted dynamics . the discrete time dynamics on the poincar section is then a diffeomorphism in the figure ) and by the projection of the unit momentum vector on the unit tangent vector . , scaledwidth=70.0% ] @xmath34 note that this diffeomorphism only leaves invariant the trapped set @xmath35 which is known to be a fractal subset of @xmath30 if the 3 disks are sufficiently far away from each other @xcite . using this discrete time dynamics we can define for any smooth function @xmath36 the transfer operator @xmath37 by setting @xmath38 which is an operator with integration kernel @xmath39 in the theory of dynamical systems this function @xmath40 is commonly called the _ potential function _ even if @xmath40 does not at all play the role of a potential in the sense of classical mechanics . with this transfer operator we can now introduce the zeta function . if we suppose that @xmath41 can be extended to a banach space @xmath42 such that @xmath41 is trace class then we can study the dynamical zeta function associated to the potential @xmath40 which is defined by the fredholm determinant @xmath43 this function is then known to be analytic in @xmath44 and the discrete spectrum @xmath45 of the transfer operator is related to the zeros of the dynamical zeta function by @xmath46 by the explicit form ( [ eq : kernel_transfer_op ] ) of the kernel @xmath47 we can on the other hand formally define the so - called flat trace @xmath48 where @xmath49 is the jacobian of @xmath50 , @xmath51 the iterated product and @xmath52 the set of all fixed points of @xmath50 . by the flat trace we can define in complete analogy to ( [ eq : dynamical_zeta_fredholm ] ) another zeta function by @xmath53 by a standard calculation for zeta functions ( see [ app : product_form ] ) this expression can be brought to the form @xmath54 here @xmath55 is the set of all _ primitive periodic orbits _ of the map @xmath56 , i.e. the set of all periodic orbits @xmath57 for which there is no @xmath58 which also fulfills @xmath59 . furthermore @xmath60 is the discrete time length of this orbit , @xmath61 is the iterated product of the potential @xmath40 along the primitive orbit @xmath62 and @xmath63 is the eigenvalue of @xmath64 whose absolute value is larger than one . it is important to note that the expressions ( [ eq : dynamical_zeta_flat ] ) and ( [ eq : dynamical_zeta_product_form ] ) only converge absolutely for @xmath65 sufficiently small . furthermore ( [ eq : dynamical_zeta_flat ] ) directly implies that the zeta function has no zeros in the region of absolute convergence . equations ( [ eq : dynamical_zeta_flat ] ) and ( [ eq : dynamical_zeta_product_form ] ) are thus of no direct use for a calculation of the zeros of the dynamical zeta function . if one knows , however , that @xmath66 can be extended analytically beyond the region of absolute convergence , its taylor expansion in @xmath44 around zero @xmath67 will automatically converge absolutely on any disk contained in the region of analytic continuation . this trick has been introduced in physics by cvitanovi and eckhardt @xcite under the name cycle expansion . from a mathematical point of view , the biggest challenge is to prove the analytic continuation of the zeta function defined by ( [ eq : dynamical_zeta_flat ] ) and to show that its zeros correspond to the spectrum of @xmath41 . this is for example possible if one finds function spaces such that @xmath41 is trace class and that its trace in this space equals the flat trace . then one obtains by the analyticity of the fredholm determinant automatically an analytic continuation of the zeta function . such results are known for several classes of systems ( for example for real analytic expanding or hyperbolic maps @xcite ) but to our knowledge not in any setting in which the 3-disk system fits . physicists usually skip this difficulty and test for concrete examples in which regions the cycle expansion numerically converges . it is then assumed that this region coincides with the region of analytic continuation . furthermore it is usually assumed that the zeros of the zeta function defined by the flat trace ( [ eq : dynamical_zeta_flat ] ) are eigenvalues of the transfer operator . as there is not yet a satisfying mathematical theory in the case of the 3-disk system we will also make these assumptions from now on . for the 3-disk system two particular choices for the potential are of special interest as they lead to the classical zeta function and to the semiclassical or gutzwiller - voros zeta function , respectively . the return time or return length of a particle at the disk boundary with birkhoff coordinates @xmath68 is defined as the length of the trajectory until the next disk reflection . this defines a smooth bounded function @xmath69 and we first can consider for @xmath70 the analytic family of potentials @xmath71 plugging this potential into ( [ eq : dynamical_zeta_product_form ] ) we obtain a dynamical zeta function depending on two variables @xmath72 @xmath73 where @xmath74 is the birkhoff sum along this orbit defined by @xmath75 for @xmath76 being an arbitrary point on the orbit @xmath62 . it is related to the selberg - smale zeta function by @xcite @xmath77 this zeta function is of special interest because its zeros , the so - called ruelle resonances , determine the decay of correlations of the classical dynamical system ( at least for systems where the theory of ruelle resonances is established , for the 3-disk system this is again conjectured to be true , c.f . this is why these zeta functions have been studied numerically by gaspard and ramirez @xcite for the symmetric 3-disk system where they observed the resonance chains presented in . using the ideas of @xcite we can also express the gutzwiller - voros zeta function by transfer operators . we therefore define the function @xmath78 on @xmath79 to be the restriction of the jacobian @xmath80 to the unstable direction which is simply given by the multiplication with a nonzero number . if @xmath76 is a fixed point of an iteration of the poincar map @xmath81 , then the iterated product of @xmath78 will be exactly the unstable eigenvalue @xmath82 of @xmath83 . with this function we define @xmath84 as well as the corresponding fredholm determinants @xmath85 we can then study the quotient @xmath86 for the last equality we used that the poincar map is orientation inverting and consequently @xmath87 which implies that @xmath88 the final expression is related to the so - called gutzwiller - voros zeta function @xcite by @xmath89 it has been numerically thoroughly checked @xcite that the zeros of @xmath90 coincide to a high precision to the quantum resonance which have numerically been obtained by a quantum scattering approach . it is conjectured that the zeros of the gutzwiller - voros zeta function converge towards the quantum resonances in the semiclassical limit . this relation @xcite is , however , based on the gutzwiller - trace formula for the quantum propagator beyond the ehrenfest time , and from the mathematical point of view the relation to the quantum spectrum is completely open . for practical purpose this relation between the zeros of the zeta function and the quantum resonances has , however , a huge advantage because especially in the regime of high energies the calculations of the zeros of @xmath90 via the cycle expansion is much faster than the calculation based on the quantum scattering matrix and it has been exploited in many numerical works on the quantum spectrum @xcite . we now want to understand the resonance chains for the symmetric 3-disk system via the zeta functions introduced above . the zeta functions allow to handle the resonance chains for classical ruelle resonances and quantum resonances by the same approach , even though the physical nature of these resonances is completely different . in both cases one has a zeta function @xmath91 which depends on two complex variables and the resonances are defined by ( see ( [ eq : gv_transfer_equivalence ] ) ) . in order to emphasize the uniform approach to quantum and classical resonances as well as the geometric resonances on schottky surfaces , we will , however , from now on consider the quantum resonances to be parametrized by @xmath92 which simply leads to a clockwise rotation by @xmath93 of the quantum resonances in the plots . ] @xmath94 note that for the quantum case we use the hypothesis that the semiclassical and quantum resonances coincide . we will , however , only use this correspondence for the symmetric 3-disk system in an energy regime , where this correspondence is known to hold with a very high precision @xcite . beside the set of resonances which we have defined to be the zeros in @xmath95 of @xmath91 for fixed @xmath96 it seems also natural to consider for a fixed @xmath97 the zero set in @xmath44 . in order to have a direct physical interpretation of this set we consider the inverses of these zeros and define for a fixed @xmath97 @xmath98 using the fredholm determinant representation of the classical zeta function @xmath99 one directly identifies @xmath100 which is the spectrum of @xmath101 . for the semiclassical resonances this identification is slightly more subtle as @xmath102 has been defined as the quotient of two zeta functions . if we assume , however , that @xmath103 is analytic in a neighborhood of @xmath104 , then the gutzwiller - voros zeta function can only vanish if the nominator of the quotient vanishes , i.e. if @xmath105 and @xmath106 . note that the analyticity assumption of @xmath103 is not very strong and in the @xmath95-range which we will consider in the sequel it is always fulfilled ( see [ app : top_pres ] ) summarizing , we have the following two objects in the classical as well as in the semiclassical or quantum case : on the one hand we have the resonance spectrum which equals the zeros of the zeta function in @xmath95 for @xmath96 and on the other hand we have for any @xmath97 the discrete set @xmath107 which can be interpreted as the eigenvalues of a transfer operator @xmath108 which is parametrized by a complex number @xmath95 . obviously @xmath95 is a resonance if and only if the operator @xmath109 has @xmath24 as an eigenvalue i.e. @xmath110 the simple but important observation is now the following : if we start with a resonance @xmath111 of multiplicity one and continuously vary the parameter @xmath95 then we will be able to trace the eigenvalue which was equal to @xmath24 for @xmath112 and which continuously depends on @xmath95 . we will call this eigenvalue @xmath113 . a closer look in the concrete form of our transfer operators @xmath108 allows us to predict how the eigenvalues will approximately depend on @xmath95 . in the quantum case for @xmath114 as well as in the classical case with @xmath115 we have @xmath116 where @xmath117 is simply the multiplication operator . in the case of the symmetric 3-disk system the return time is a smooth function which becomes more and more homogeneous the bigger @xmath23 becomes . assuming as a first order approximation that @xmath118 then @xmath119 if this approximation was exactly true and if we moved with @xmath95 parallel to the imaginary axis , then the eigenvalue @xmath113 would simply rotate around zero on the unit circle . note that by the definition of the resonance set , every @xmath95 value for which @xmath113 crosses the value point 1 on the unit circle is itself a resonance . by this argument a single eigenvalue of @xmath120 would create a whole chain of resonances , parallel to the imaginary axis.the resonances would be equally distributed along this chain with a distance of @xmath121 . as the return time is in reality not exactly constant one would expect that slight variations deform the chains and also lead to a distance variation between the extremal values of the return time . it can be checked numerically that it is exactly this mechanism of a rotating eigenvalue of the operator @xmath108 which creates the resonance chains . in this is visualized for the quantum resonances of the symmetric 3-disk system with @xmath1 . on the left we can see a plot of the quantum resonances @xmath122 in the complex plane . ] . on the right the spectrum of the transfer operator @xmath123 is presented . the blue crosses indicate the eigenvalues with @xmath124 for an @xmath95-value which equals the resonance @xmath125 . as expected we find an eigenvalue which is equal to @xmath24 . there is also another eigenvalue plotted which corresponds to the second chain which is nearby the chain to which @xmath126 belongs . there are further eigenvalues of the transfer operator near zero which we did not plot as the convergence of the zeta function becomes more and more complicated the smaller the absolute values of @xmath44 . the colored points show how the eigenvalue at @xmath24 moves , if the @xmath95 values varies between two resonances @xmath126 and @xmath127 . as suggested by the heuristic arguments above the eigenvalue moves along the unit circle in a clockwise orientation and creates the next resonances on the chain . exactly the same can be observed for the classical ruelle resonances of the 3-disk system with @xmath1 ( see ) . . the inset shows the chain structure of the resonances on a large domain in the complex plane . the main plot shows a strong zoom into one chain , such that only few individual resonances are in the plot region ( blue crosses ) . the region of the zoom is indicated in the inset by a red rectangle . the colored dots between the two resonances @xmath125 and @xmath127 represent a discrete interpolation between the two resonances . the interpolation follows the line suggested by the resonance chains . right : spectrum of the transfer operator on the poincar section . the blue crosses represent the spectrum of @xmath128 with @xmath124 . by the colored dots we follow the evolution of the eigenvalue starting at 1 for @xmath126 while @xmath95 varies as indicated by the colored points in the left panel . the evolution of the second eigenvalue is not plotted for more clarity . ] . the inset shows the chain structure of the resonances on a large domain in the complex plane . the main plot shows a strong zoom into one chain , such that only few individual resonances are in the plot region ( blue crosses ) . the region of the zoom is indicated in the inset by a red rectangle . the colored dots between the two resonances @xmath129 and @xmath130 represent a discrete interpolation between the two resonances following the line suggested by the resonance chains . right : spectrum of the transfer operator on the poincar section . the blue crosses represent the spectrum of @xmath128 with @xmath124 . by the colored dots we follow the evolution of the eigenvalue starting at 1 for @xmath126 . ] and @xmath127 now do not follow the line which is suggested by the resonance chains but a deformed path . as a consequence the eigenvalue of @xmath128 starting at 1 for @xmath126 now turns one time around zero on a path which is not on the unit circle ( black , dotted ) . ] these numerical results do , however , not only confirm that the hypothesis that each resonance chain is created by a rotating @xmath108-eigenvalue is true . they also indicate what the continuous mathematical object is that the resonances are distributed on . note that taking two neighboring resonances of one chain there are multiple paths to connect these two resonances . while each path will effectively lead to one rotation of @xmath113 around zero , there is only one such path , for which @xmath113 stays on the unit circle . as it can be seen in such a rotation on the unit circle is obtained if the two resonances are connected along the path which is suggested by the structure of the resonance chains . if one , however , takes a different path between the two resonances , then the corresponding eigenvalue of @xmath108 will turn on a deformed circle around zero ( see for an arbitrary path between the two resonances ) . we can thus define @xmath131 and from the numerics above we conclude that the continuous object on which the resonances are distributed is given by @xmath132 where @xmath133 is the projection on the @xmath95-component . we have thus seen in this section that the resonance chains of the classical and the quantum resonances are generated by the same mechanism . as in both cases the resonances can be seen as the zero set in @xmath95 of a zeta function , which can be represented as a fredholm determinant of a @xmath95-dependent family of operators . the resonance chains are generated by single eigenvalues of @xmath108 which circle around zero while @xmath95 moves along the chain . furthermore this approach allows a system - intrinsic definition of a continuous object that coincides with the visually suggested curves on which the resonances are strung together . in the next section we will see how this idea can be extended to the resonance chains on schottky surfaces and examine in detail the question in which systems resonance chains will occur and in which not . schottky surfaces are a special class of infinite volume riemann surfaces . in this section we will give a brief introduction to its most important properties . for a detailed introduction we refer to @xcite . the easiest way to define schottky surfaces is as a quotient of the upper half plane by a so - called schottky group . the upper half plane @xmath134 together with the riemannian metric @xmath135 has a constant negative gauss curvature @xmath136 and is a standard model of hyperbolic geometry . the group @xmath137 of real invertible matrices with determinant one acts on @xmath138 via moebius transformations @xmath139 in fact for every @xmath140 the moebius transformation is an orientation preserving isometry and all orientation preserving isometries of the upper half - plane can be expressed as an @xmath137-action . a schottky group is then a discrete subgroup @xmath141 which is constructed in the following way : let @xmath142 be disks with centers on the real line and mutually disjoint closure . then for every @xmath143 there exists an element @xmath144 that maps the boundary @xmath145 to the boundary of @xmath146 and the interior of @xmath25 to the exterior of @xmath147 . from the disjointness of the disks follows that the elements @xmath148 are the generators of a free discrete subgroup @xmath149 and such groups are called _ schottky groups_. the quotient @xmath150 is then again a surface with constant negative curvature and is called _ schottky surface_. the 3-funneled surfaces , which we are interested in , are from the dynamical point of view the simplest nontrivial example and they are constructed from four disks arranged as in . the surface can be thought as the set @xmath151 glued together along the two dotted blue half circles as well as along the two dashed red half circles ( see ) . it is known that such 3-funneled schottky surfaces are uniquely determined by the three lengths @xmath152 of the geodesics @xmath153 that wind once around one of the 3-funnels ( see upper part of ) . furthermore for any triple of positive numbers @xmath152 there exists a 3-funneled schottky surface which has fundamental geodesics of this length and we will call this surface @xmath154 . . the dotted blue and dashed red lines indicate the lines along which the fundamental domain is glued together in order to obtain the surfaces . lower part : configuration of 4 disks that give rise to the construction of the schottky group for a 3-funneled surface . the upper half plane without the disks represents a fundamental domain and the surface can be obtained by gluing together the red dashed lines with the blue dotted lines . in black the three fundamental closed geodesics @xmath155 from the upper part of the figure are shown . while @xmath156 and @xmath157 are only represented by one arc each , the geodesic @xmath158 appears as two arcs in the fundamental domain.,title="fig:",scaledwidth=70.0% ] . the dotted blue and dashed red lines indicate the lines along which the fundamental domain is glued together in order to obtain the surfaces . lower part : configuration of 4 disks that give rise to the construction of the schottky group for a 3-funneled surface . the upper half plane without the disks represents a fundamental domain and the surface can be obtained by gluing together the red dashed lines with the blue dotted lines . in black the three fundamental closed geodesics @xmath155 from the upper part of the figure are shown . while @xmath156 and @xmath157 are only represented by one arc each , the geodesic @xmath158 appears as two arcs in the fundamental domain.,title="fig : " ] in order to define the geometric resonances on a schottky surface @xmath159 we consider the positive laplacian which is defined by the metric and which we call @xmath160 and we define the resolvent @xmath161 by mazzeo - melrose @xcite and guillope - zworski @xcite we know that the resolvent admits a meromorphic continuation and we define the resonance set @xmath162 again as the set of its poles . note that the resolvent s parametrization is different from the parametrization via @xmath163 in the 3-disk case , which leads to the fact that the resonance chains do not oscillate along the real axis but along the imaginary axis . this parametrization is , however , very useful as it admits the following exact relation between the resonances of the laplacian and a zeta function : let @xmath164 be the set of all primitive closed geodesics on @xmath159 , i.e. those closed geodesics that can not be obtained by repeting a shorter closed geodesic . if for @xmath165 , @xmath166 denotes the length of the periodic geodesic , then the selberg zeta function is defined as @xmath167 this product is known to converge uniformly on any compact set with @xmath168 and for @xmath159 being a schottky surface the selberg zeta function is known to extend analytically to the whole complex plane @xmath169 and there is the following relation between its zeros and the resonances of @xmath160 : if @xmath170 then @xmath95 is either a topological zero with @xmath171 or a resonance i.e. @xmath172 @xcite . for schottky surfaces we can express the selberg zeta function as a fredholm determinant of a transfer operator . we therefore define the _ bowen - series map _ of a schottky group @xmath173 by @xmath174 here we used the convention that for @xmath175 , @xmath176 . we can then define a family of transfer operators parametrized by @xmath97 via its action @xmath177 where @xmath178 is the complex derivative of the bowen - series map and @xmath179 . these transfer operators are known to be trace class on the space of holomorphic @xmath8-functions on @xmath180 ( see @xcite for the original proof in slightly different functions spaces or ( * ? ? ? * lemma 15.6 ) ) and one therefore knows that its dynamical zeta function @xmath181 is analytic . we will give a short sketch of the idea how to relate the dynamical zeta function to the selberg zeta function . for more details we refer to ( * ? ? ? * section 15.3 ) . one first can proof that the trace of @xmath108 equals its flat trace and obtains @xmath182 as in the case of the 3-disk system ( cf . equation ( [ eq : dynamical_zeta_product_form ] ) and [ app : product_form ] ) one can transform the dynamical zeta function into a product over prime orbits of the bowen - series map @xmath183 . in a second step one can prove that the set of primitive periodic geodesics is in one - to - one correspondence to the primitive periodic orbits of the bowen - series maps , i.e. for each @xmath165 there exists a unique @xmath184 and a periodic orbit @xmath185 and the length of the geodesic is related to the stability of the fixed point by @xmath186 if we denote this @xmath0 as the _ bowen - series order _ @xmath187 of the geodesic @xmath188 , then we obtain , combining all arguments @xmath189 in particular this yields @xmath190 and we deduce that @xmath191 implies that @xmath108 has an eigenvalue equal to 1 . the previous discussion on the relation between the spectrum of @xmath108 and the resonances of the laplacian , together with the observations from section [ sec : rot_res ] that for 3-disk systems rotating eigenvalues of a transfer operator create the resonance chains would suggest that the resonance chains for the 3-funneled schottky surfaces are generated by rotating eigenvalues of the bowen - series transfer operator @xmath108 . however , this idea fails completely to explain the resonance chains for the 3-funneled symmetric surfaces : in we have plotted for @xmath192 the spectrum of @xmath108 for different @xmath95-values interpolating between two resonances @xmath126 and @xmath193 . . the inset shows the chain structure of the resonances on a large domain in the complex plane . the main plot shows a strong zoom into one chain , such that the individual resonances become visible ( blue crosses ) . the region of the zoom is indicated in the inset by a red rectangle . the colored dots between the two resonances @xmath194 and @xmath195 represent a discrete interpolation between the two resonances , following the line suggested by the resonance chains . right : spectrum of the bowen - series transfer operator with @xmath196 . the blue crosses represent the spectrum of @xmath128 . by the colored dots we follow the evolution of two particular eigenvalues , as the @xmath95-values evolve as indicated by the colored points in the left panel . we only follow the evolution of the eigenvalue starting at 1 for @xmath126 and the one ending at 1 for @xmath193 . the evolution of the other eigenvalues is not shown for more clarity . ] we observe that the eigenvalues start to turn while we move along the chain . however , the eigenvalue which was equal to @xmath24 for @xmath197 is drawn towards zero and when we approach @xmath198 it is another eigenvalue of @xmath108 which has been coming out from the interior , that creates the next resonance . in order to understand why a direct application of the idea of rotating eigenvalues to the spectrum of the bowen - series transfer operator @xmath108 fails , we recall that for the 3-disk system it has been an important argument that the return time of the poincar section is very homogeneous . this assumption is , however , not at all fulfilled for the standard bowen - series map , which corresponds morally to a poincar section indicated by the blue and red lines in . the return time or return length respectively is thus very inhomogeneous : while on @xmath192 for the closed geodesics @xmath156 and @xmath157 the return length is equal to @xmath199 , it equals only @xmath200 for the geodesic @xmath158 as this geodesic crosses the poincar section twice . in order to fix this issue we could try to construct another transfer operator which still has the property that its dynamical zeta function is related to the selberg zeta function but which has a homogeneous return time for the completely symmetric 3-funneled schottky surface . such a construction is in principle possible ( see @xcite ) , this solution would then , however , be restricted to completely symmetric schottky surfaces . non - symmetric surfaces , for which resonance chains have also been observed , would then need a construction of more and more complicated transfer operators . we will thus choose a different approach and work directly with the zeta functions which turns out to be much more flexible . looking at equation ( [ eq : dyn_zeta_bs ] ) the choice of the poincar section appears only in the order function @xmath201 . this order function has been defined as the length of the prime @xmath183-orbit which is associated to the primitive geodesic @xmath188 and it is exactly given by the number of times the geodesic passes one of the cut lines in . instead of constructing another poincar section we can also directly modify the zeta function and we can introduce for each map @xmath202 the _ generalized zeta function _ @xmath203 of course for a general order function @xmath204 it is not directly clear where this product converges and whether it has an analytic extension . for all order functions that will appear in the sequel , one can , however , show , that such an analytic extension exists and we refer to @xcite for a rigorous proof . the important observation is now that changing the order function only changes the values of the dynamical zeta functions for @xmath205 and the important relation @xmath206 is still true . we can therefore define in analogy to the spectrum of the transfer operator ( [ eq : sigma_s_3-disk ] ) , the _ generalized spectrum _ @xmath207 from ( [ eq : gen_zeta - selberg_zeta ] ) we conclude that also for this generalized spectrum we have the important relation @xmath208 thus instead of tracing the eigenvalues of a transfer operator we can follow the generalized spectral values while interpolating between different resonances on one chain . the order function which would correspond to a poincar section of the symmetric surface @xmath192 with a homogeneous return time counts for each closed geodesic how often it winds around one of the funnels . in we traced the generalized spectrum while we interpolate between the same two resonances as in and indeed we observe that the generalized spectral value which equals @xmath24 for @xmath197 starts to turn around zero and creates all the other resonances on this chains by passing through @xmath96 . additionally we see in that if the @xmath95-values follow the path which is suggested by the resonance chains , then the generalized spectral value stays on the unit circle . the continuous lines on which the resonances lie can thus be understood by the real analytic variety @xmath209 precisely they are again given by @xmath210 , the projection of this variety on the @xmath95-component . with @xmath196 for an order function that counts the number of windings around the funnels . the blue crosses represent the spectrum @xmath211 . by the colored dots we follow the evolution of the spectral value which equals 1 for @xmath197 . the evolution of the other spectral values is again not plotted . ] this approach of working with a generalized zeta function immediately allows us to understand resonance chains for non - symmetric surfaces , as well . these chains have first been observed by borthwick , for example for the surface @xmath212 ( cf . * figure 7 ) ) . a homogeneous return time would here require to cut one turn around the first funnel into 12 pieces , one turn around the second funnel into 13 and a turn around the third funnel into 14 pieces . we therefore define for @xmath213 the order function @xmath214 where @xmath215 counts the windings of the geodesic around the @xmath216-th funnel . figure [ fig : rot_res_12_13_14 ] shows the generalized spectrum for this order function and it perfectly explains the resonance chains . again the spectral value turns once around the unit circle and creates the next resonance on the chains when the @xmath95 value moves along the chains . note that the two other spectral values in @xmath211 ( blue crosses in the left part of ) seem also to lie on the unit circle , which is not the case . their absolute values are @xmath217 and @xmath218 and they belong to nearby chains , which are visible in the left part of . there are also further spectral values with absolute value slightly smaller then one . we however decided to not plot them in for more clarity and as a reliable calculation of the generalized spectrum becomes more and more complicated for small absolute values of @xmath44 as the order function @xmath204 appears with high exponents ( cf . [ app : numeric ] ) . . the inset shows the chain structure of the resonances on a large domain in the complex plane . the main plot shows a zoom into one chain , such that the individual resonances become better distinguishable ( blue crosses ) . for orientation the region of the zoom is indicated in the inset by a red rectangle . the colored dots between the two resonances @xmath219 and @xmath220 represent a discrete interpolation between the two resonances following the line suggested by the resonance chains . right : generalized spectrum for the order function as defined in ( [ eq : funnel_order_func ] ) with @xmath221 for spectral values with @xmath222 . the blue crosses represent the spectrum @xmath223 . by the colored dots we follow the evolution of one spectral value as the @xmath95-values evolve as indicated by the colored points in the right panel . we only follow the evolution of the spectral value starting at 1 for @xmath197 for more clarity . ] from the examples in the previous section we have learned that the resonance chains for symmetric as well as for asymmetric schottky surfaces are best understood in terms of the generalized zeta function for a suitable choice of an order function . up to now we have only presented two examples in ( and ) together with a good choice of an order function but we have not yet clearly worked out under which circumstances it is possible to construct a `` good '' order function . we will see that this question automatically leads us to a condition on the observability of resonance chains . the primitive length spectrum , which contains the lengths of all primitive closed geodesics , repeated by multiplicity , can be seen for @xmath192 in and for @xmath212 in . we see that the length spectra of both surfaces form a clear clustering on multiples of a base length @xmath224 . . the thin light lines indicate the integer multiples of the base length @xmath199 and one observes that the lengths form clear clusters around these values . this clustering is however not perfect as can be seen in the inset where a histogram with a much smaller binsize resolves the peak at a length around 60 . ] . the thin light lines indicate the integer multiples of the base length @xmath24 and one observes that the lengths form clear clusters around these values . again this clustering is not perfect as can be seen in the inset where a histogram with a much smaller binsize resolves the peak at a length around 51 . ] the big difference between these two cases is that the common base length to all clusters is @xmath225 for @xmath192 and @xmath226 for @xmath212 . analyzing the two order functions which we have used in these examples , we observe that they exactly respect this clustering : all geodesics of one cluster are mapped to the same integer which is given by the multiple of the base length @xmath224 on which the cluster is located . all geodesics that are necessary to be taken into account for the calculation of the resonances fulfill very well the approximation @xmath227 if this condition is exactly fulfilled , then this will lead , similar to the case of a constant return time in section [ sec : rot_res ] , to the existence of straight resonance chains . in fact under this assumption the generalized zeta function could be written as @xmath228 and if @xmath229 , i.e. @xmath230 then for all @xmath231 also @xmath232 . this condition , however , directly implies that @xmath233 for each @xmath234 , so all resonances would equidistribute on straight lines with a distance of @xmath235 . note that for the hyperbolic cylinder ( [ eq : order_func_condition ] ) is exactly fulfilled and the resonance spectrum is known to be ordered along straight lines ( * ? ? ? * section 5.1 ) . for non - elementary surfaces ( [ eq : order_func_condition ] ) can , however , only hold approximately , and one thus expects that these straight chains start to bend and that the spacing of the resonances along the chains becomes more irregular . we therefore can formulate the following hypothesis : _ if the primitive length spectrum of an infinite volume riemannian surface ( or an open euclidean billiard ) forms clusters at integer multiples of a base length , i.e. if there is a length @xmath224 and an order function @xmath236 such that ( [ eq : order_func_condition ] ) holds , then one will observe resonance chains . these chains will be described by the projection @xmath237 of the analytic variety @xmath238 to the @xmath95-component and the resonances will be approximately spaced by a distance of @xmath235 on these lines . the worse the approximation ( [ eq : order_func_condition ] ) is fulfilled , the stronger the chains will bend and the more irregular the resonance distribution on the chains will become . _ of course ( [ eq : order_func_condition ] ) can not hold on the whole length spectrum , but recall that for the calculation of the resonances in a bounded subset of a complex plane , one only needs finitely many geodesics . we thus suppose that the chains will be observable in this bounded domain , if ( [ eq : order_func_condition ] ) is fulfilled on the geodesics which are necessary for their calculation . we already saw several examples in section [ sec : rot_res ] and [ sec : gen_zeta ] which support our hypothesis on the existence of resonance chains . we will now test its validity against some more examples . first of all our condition should be able to explain the non - observability of resonance chains in some systems which are structurally similar to systems which show resonance chains . for example one observes that symmetric 3-disk billiards only show resonance chains if the disks are sufficiently far away from each other . the resonance spectrum of the symmetric 3-disk billiard with @xmath239 shows no chain structure contrary to the more open case with @xmath1 ( see ) . this is easily understood by looking at the primitive length spectrum . while for @xmath1 one sees a clear clustering of the lengths , the length spectrum for @xmath239 equidistributes very quickly and it is impossible to choose an order function which fulfills ( [ eq : order_func_condition ] ) . ( upper plot ) and @xmath1 ( lower plot ) . while for @xmath1 one observes a clear clustering of the lengths and an obvious chain structure in the resonances , the lengths equidistribute very quickly and their are no visible chains in the resonance spectrum.,title="fig : " ] ( upper plot ) and @xmath1 ( lower plot ) . while for @xmath1 one observes a clear clustering of the lengths and an obvious chain structure in the resonances , the lengths equidistribute very quickly and their are no visible chains in the resonance spectrum.,title="fig : " ] our hypothesis also explains why symmetric 4-disk systems show in general no chain structure ( see e.g. fig . 13 in @xcite for the ruelle resonances of a 4-disk system ) whereas the 3-dimensional 4-sphere billiard shows them : if 4 disks are distributed on a square , the lengths between two neighboring disks and the length between diagonal disks are in general not multiples of each others but form two incommensurable base lengths , so it will not be possible to construct a good order function . for the 4-sphere billiard where the spheres are placed on a regular tetrahedron , the distance between an arbitrary pair of spheres is equal and , if the spheres are sufficiently far away from each other , one expects a clustering of the length spectrum on multiples of this base length . exactly in this case of sufficiently open 4-sphere billiards the chain structure has been observed ( cf . figure 5.2 - 5.9 in @xcite ) . . the thin light lines indicate the integer multiples of the base length @xmath226 and red dashed lines the multiples of a second base length @xmath225 . one can observe two types of clustering , one very coarse clustering around the multiples of @xmath225 and a much finer clustering around the multiples of @xmath226 . ] we will end this section by using the predictive power of our hypothesis in order to identify a new interesting system which shows the coexistence of two chains . as we only demand ( [ eq : order_func_condition ] ) to hold approximately there should be examples for which there are different compatible base lengths and order functions , e.g. the surface @xmath240 . as can be seen in the length spectrum can be either interpreted as forming clusters at multiples of the base length @xmath225 or at multiples of the base length @xmath226 . the hypothesis thus predicts two different chain structures , one where the resonances are spaced by a distance of @xmath241 and another one with a larger spacing of @xmath242 . as for @xmath225 the approximation ( [ eq : order_func_condition ] ) holds much worse then for @xmath226 our hypothesis predicts that the resonance chains belonging to @xmath225 are much more twisted then the chains belonging to @xmath226 . as figures [ fig : rot_res_12_12_13 ] and [ fig : rot_res_12_12_13_long ] show , all these predictions are observed in the resonance structure . choosing the order function ( [ eq : funnel_order_func ] ) with @xmath243 which corresponds to the base length @xmath225 we observe strongly twisted resonance chains as indicated by the red circles in the left part of . the generalized spectrum @xmath244 of this order function creates this kind of chains by the rotating spectral values and if one connects neighboring resonances on this chain by the continuous line which is suggested by the chain structure , the generalized spectral value moves along the unit circle . this verifies that the chains are described by @xmath210 . choosing , however , the order function ( [ eq : funnel_order_func ] ) with @xmath245 which is the corresponding choice for the base length @xmath226 we obtain a generalized zeta function which describes a second structure of resonance chains , on which the resonance distances are much larger and which are much straighter ( see ) . in the complex plane are marked by crosses . the red circles highlight the resonances on the chain , which belongs to the base length @xmath225 . the colored dots between the two resonances @xmath246 and @xmath247 represent a discrete interpolation between the two resonances following the line suggested by the resonance chains . right : generalized spectrum for the order function as defined in ( [ eq : funnel_order_func ] ) with @xmath243 . the blue crosses represent the spectrum @xmath223 for @xmath196 . by the colored dots we follow the evolution of the spectral value which equals one while interpolating the @xmath95-values between @xmath126 and @xmath193 as indicated by the colored points on the left . ] in the complex plane are marked by crosses . the red circle highlight the resonances on the chain , which belongs to the base length @xmath226 . the colored dots between the two resonances @xmath246 and @xmath248 represent a discrete interpolation between the two resonances following the line suggested by the resonance chains . right : generalized spectrum for the order function as defined in ( [ eq : funnel_order_func ] ) with @xmath243 . the blue crosses represent the spectrum @xmath223 for @xmath249 . by the colored dots we follow the evolution of the spectral value which equals one while interpolating the @xmath95-values between @xmath126 and @xmath193 as indicated by the colored points on the left . ] in this article we presented a unifying approach to the chain structure of quantum resonances , classical ruelle resonances and geometric resonances . we showed at the example of 3-disk systems and schottky surfaces that the resonance chains can be understood in all three cases by means of a generalized zeta function @xmath250 which depends on the choice of an order function @xmath251 on the set of primitive closed classical orbits @xmath55 . the central property of this generalized zeta function is that independent of the choice of the order function the resonances are given by the zeros of zeta function @xmath252 . we showed that if the order function is chosen such that @xmath253 for an arbitrary base length @xmath224 then the second complex variable allows to interpolate between the resonances of one chain . furthermore we demonstrated that the continuous lines where the resonances are found on are given by the projection onto the @xmath95-component of the real analytic variety @xmath254 the existence of an order function that fulfills the condition ( [ eq : of_cond_concl ] ) does of course depend on the structure of the length spectrum of the classical system . this led us to the hypothesis that the existence of resonance chains is directly linked to a clustering of the classical length spectrum that allows us to choose an order function according to ( [ eq : of_cond_concl ] ) . we finally validated this hypothesis by presenting several examples of systems with and without resonance chains . in all cases the resonance chains were found to be linked to a clustering of the length spectrum . furthermore we showed that this new understanding allows to construct scattering systems with a customized resonance chain structure . this control of the resonance chains could be of importance in microdisk cavities where similar resonance chains have already been observed @xcite and where a precise control of the resonance position in the complex plane is of great interest for their application as microdisk lasers . additionally this understanding is crucial for fundamental questions in quantum chaos . in chaotic systems the clear clusters in the length spectrum will only exist for short lengths . for larger lengths the clusters become broader and broader and will finally overlap . however , we required the condition ( [ eq : of_cond_concl ] ) only to hold in the length range which is necessary for the calculation of the resonances in the studied frequency regime . in other words for the existence of resonance chains in a given frequency range , condition ( [ eq : of_cond_concl ] ) only has to hold for the lengths which can be resolved by the planck cells at this frequency . if one still observes resonance chains , then one has to keep in mind that one is not yet at high enough frequencies to resolve the length spectrum in a regime , where the clusters dissolve . this could be especially important for the test of results that are obtained by arguing with phase cancellation such as the conjectures on the improved spectral gaps @xcite finally the study of resonance chains opens a variety of mathematical questions . we numerically observed in this article that the chain structure becomes the clearer the better condition ( [ eq : of_cond_concl ] ) is fulfilled . by studying not just a single system but a whole family of scattering systems , where ( [ eq : of_cond_concl ] ) is increasingly well fulfilled in a certain limit one can then try to prove the existence of resonance chains asymptotically in this limit . furthermore one can try to give precise , simple formulas for the location of these chains . for symmetric 3-disk systems this limit would correspond to large @xmath23-values and for schottky surfaces to large funnel widths . for the latter case , first results into this direction will be presented in @xcite . we are grateful to ulrich kuhl for many interesting stimulating discussions and detailed feedback concerning this work . furthermore we thank bruno eckhardt , david borthwick , stphane nonnenmacher , pablo ramacher and gabriel rivire for discussions and comments . this work was supported by the project anr 2009 - 12 methchaos . s.b . acknowledges financial support by dfg via the project for760 and t.w . by the german national academic foundation in this appendix we will recall the derivation of the product form ( [ eq : dynamical_zeta_product_form ] ) from the definition of the dynamical zeta function as it has been defined in ( [ eq : dynamical_zeta_flat ] ) by the flat trace @xmath255 for each fixed point @xmath81 the jacobian is a hyperbolic , symplectic @xmath256 matrix and it thus has two real eigenvalues @xmath82 and @xmath257 with @xmath258 and the determinant can be written as @xmath259 where for the last equality we used that a positive integer @xmath163 can be written as @xmath260 different sums of two positive integers @xmath261 and @xmath95 . as a next step we can treat the double sum @xmath262 which is simply the sum over all fixed points of @xmath50 for arbitrary length @xmath56 . first we note that the iterated product of the potential function @xmath263 as well as the instable eigenvalue @xmath82 do not depend on the choice of the fixed point in a fixed orbit @xmath264 . additionally we see that if an orbit of length @xmath265 is a @xmath12-times iterate of a primitive orbit of length @xmath60 , then the orbit contains @xmath60 fixed points and as well the iterated product as the instable eigenvalue are the @xmath12-th power of the values from the primitive orbit @xmath266 this allows us to write the double sum @xmath262 as a double sum over all primitive periodic orbits and their repetitions @xmath267 and we obtain @xmath268^{k+1}&\nonumber \\ & = \prod_{k\geq0}\prod_{p\in p}\left[1-z^{n_p}\frac{v_{n_p}(x)}{|\lambda_{n_p}|\lambda_{n_p}^{k}}\right]^{k+1}&\nonumber\end{aligned}\ ] ] where we used the taylor series of @xmath269 in the last equality . in this appendix we will shortly recall how the resonance spectrum on schottky surfaces is usually calculated ( see @xcite for more details ) and what modifications have to be performed for the calculation of the generalized zeta function and the generalized spectrum . usually the resonance spectrum is obtained by taylor expanding ( [ eq : bs_dyn_zeta_fixpoint_formula ] ) in @xmath44 around zero @xmath270 the taylor coefficients are then explicitly given by ( * ? ? ? * proposition 8) @xmath271 where @xmath272 are all @xmath261-partitions of @xmath163 , i.e. all @xmath261-tuples of integers whose sum equals @xmath163 . the numerical task in order to calculate these coefficients then consists in calculating the stabilities @xmath273 of sufficiently many fixed points . for a 3-funneled schottky surface @xmath154 this can efficiently be done using the generators of the corresponding schottky group and a symbolic dynamic . for @xmath154 the generators can be written as @xmath274 where the parameter @xmath22 is chosen such that @xmath275 and as usual one writes @xmath276 and @xmath277 . if we then define the set of words of length @xmath0 by @xmath278 then we can define for each @xmath279 the hyperbolic isometry @xmath280 one then can easily show from the definition of the bowen - series maps ( * ? ? ? * section 15.2 ) that for each @xmath281 there is exactly one @xmath282 and @xmath283 which allows an efficient calculation of ( [ eq : b_k_bowen_series ] ) and thus of @xmath284 . for the calculation of @xmath250 one analogously has to taylor - expand the zeta function in @xmath44 around zero @xmath285 then one has to calculate the new taylor coefficients @xmath286 . as we want to use the word coding to calculate the lengths of the geodesics or the stabilities @xmath273 , respectively , we first have to transfer the order function @xmath287 to an order function on @xmath288 . this transformation can be obtained via the correspondence between closed geodesics on @xmath154 and fixed points of @xmath183 ( * ? ? ? * proposition 15.5 ) . using this identification one calculates that the order function as defined in ( [ eq : funnel_order_func ] ) is then given by @xmath289 where @xmath290 the bowen - series order function @xmath291 can be translated much more easily . from its definition in section [ sec : intro_schottky ] it immediately follows , that @xmath292 for all @xmath293 . using the order function on the words as well as the identification of fixed points and words , equation ( [ eq : bs_dyn_zeta_fixpoint_formula ] ) can be written as @xmath294 starting with ( [ eq : gen_zeta ] ) and reversing the argumentation of [ app : product_form ] the generalized zeta function @xmath250 can be written as @xmath295 an analogous calculation to those in ( * ? ? ? * proposition 8) then gives us the formula for the taylor coefficients @xmath296 the stabilities @xmath297 can be calculated by the generators @xmath298 of the schottky group as described above and in order to simplify the combinatorial task of building together the taylor coefficients @xmath286 from these stabilities we can use the same recurrence trick as presented in @xcite . we therefore write @xmath299 and @xmath300 we can then use the recurrence relations @xmath301 the only difference in calculating the generalized zeta function compared to the dynamical zeta function of the bowen - series maps thus consists in the modified formula for the functions @xmath302 . note that as @xmath250 is analytic the taylor coefficients @xmath286 decay superexponentially in @xmath163 and theoretically one can compute @xmath250 for arbitrary @xmath303 . practically , however , the convergence becomes worse and worse the smaller @xmath304 and the bigger @xmath65 become . especially for order functions ( [ eq : funnel_order_func ] ) with larger values of @xmath305 as they appeared for example for the surfaces @xmath212 the convergence depends very strongly on @xmath65 . this implies that a reliable calculation of @xmath306 is only possible for @xmath44 values which are slightly greater then @xmath24 . in terms of the generalized spectrum this implies that the numerical calculation is only possible for @xmath44 slightly smaller then @xmath24 . as the interesting region for the understanding of the resonance chains is , however , the unit circle , this is no severe problem for the numerical investigations presented in this article . the topological pressure of a 3-disk system is a function @xmath307 of a positive parameter @xmath308 that describes the convergence behavior of certain zeta functions . writing @xmath309 the topological pressure of the 3-disk system can be defined @xcite to be the real number @xmath307 such that @xmath310 is absolutely convergent for @xmath311 . this directly implies that @xmath312 has no zeros for @xmath313 and that @xmath314 has no zeros for @xmath315 . for a symmetric 3-disk system with @xmath1 this value of the topological pressure is given by @xmath316 consequently all zeros of the gutzwiller - voros zeta function @xmath317 with @xmath318 are automatically zeros of the fredholm determinant @xmath319 and can thus be interpreted by the spectrum of the transfer operator @xmath123 . note that all the numerical investigations presented in this article are within this @xmath95-range .

in recent years , the goal of providing high speed wireless data services has generated a great amount of interest among the research community . recent information theoretic results have demonstrated that the ability of a system to support a high link quality and higher data rates in the presence of rayleigh fading improves significantly with the use of multiple transmit and receive antennas @xcite . a very important aspect is always the availability of analytical expressions to describe the stochastic nature of the channel under consideration as given in @xcite for the mimo channel . this offers an opportunity to obtain , e.g. , closed - form analytical formulas for the ergodic capacity or the outage mutual information of such mimo channels . e.g. , in @xcite , the probability density function ( pdf ) of the random mutual information for independent identically distributed ( i.i.d . ) mimo channels was derived in the form of the inverse laplace transform and a gaussian approximation of the pdf was presented . in @xcite , the impact of mimo channel rank deficiency and spatial fading correlation on the mutual information was analyzed . furthermore , for correlated channels , the optimal transmit strategy and the impact of correlation on the outage probability were derived in @xcite . there has been considerable work on a variety of new codes and modulation signals , called space - time codes , in order to approach the huge capacity of such mimo channels . the performance criteria of space - time codes were derived in @xcite . one scheme of particular interest is the alamouti scheme @xcite for two transmit antennas . later on , @xcite proposed more general schemes referred to as orthogonal space - time block codes ( ostbc ) with the same properties as the alamouti scheme like , e.g. , a remarkably simple maximum - likelihood decoding algorithm . interestingly , the combination of ostbc with a mimo antenna system can be represented equivalently as a single - input - single - output ( siso ) system , where the channel gain is equal to the frobenius norm of the actual mimo channel . the performance of orthogonal space - time block codes @xcite with respect to mutual information was analyzed ( among others ) for the uncorrelated rayleigh fading case in @xcite and for the more general case with different correlation scenarios and line of sight ( los ) components in @xcite . ostbc exploit multiple antennas at both the transmitter and receiver in order to obtain transmit and receive diversity and therefore increase the reliability of the system . with the knowledge of the stochastic nature of the resulting equivalent channel due to the employment of ostbc in a mimo system , the loss in mutual information of ostbc in subject to transmission rate , number of receive antennas and channel rank was quantified in @xcite , whereas in @xcite a comparison of ostbc with a system applying beam - forming was presented . unfortunately , the alamouti space - time code for two transmit and one receive antennas is the only ostbc , which , to the best of our knowledge , achieves the maximum possible mutual information of a mimo system @xcite , since we can not construct an ostbc with transmission rate equal one for more than two transmit antennas @xcite . therefore , @xcite designed a quasi - orthogonal space - time block code ( qstbc ) with transmission rate one for four and eight transmit antennas . by properly choosing the signal constellations as done in @xcite , it is possible to improve the ber performance with ml - detection for the codes given in @xcite . the ber performance of qstbc with suboptimal detectors has been analyzed in @xcite . the performance of qstbc with respect to outage mutual information ( omi ) for the special case of one receive antenna and four or eight transmit antennas was analyzed via simulations in @xcite and @xcite and it was shown , that the qstbc are capable to achieve a significant portion of the mimo - omi . furthermore , it was shown in @xcite , that qstbc in conjunction with optimal ( nonlinear ) and suboptimal ( linear ) detectors provide a tradeoff between performance and complexity . the key achievements of this paper are as follows * we generalize the results in @xcite to @xmath3 transmit and an arbitrary number of receive antennas . * we show , that due to the employment of qstbc the eigenvalues of the resulting equivalent channel are pairwise independent and identical ( i.i.d ) noncentral chi - square distributed with @xmath2 degrees of freedom ( @xmath4 ) with noncentrality parameter @xmath5 . furthermore , we show that the eigenvectors of the equivalent channel are independent of each channel realization , i.e. they are constant . in other words , we first show that the combination of qstbc with a mimo system results also in a equivalent channel similar to ostbc and then fully characterize the stochastic nature of this equivalent channel . based on these important insights , we are able to provide the following results * an analytical lower bound for the outage probability achieved with qstbc , which is tight for low signal - to - noise - ratios ( snr ) values and also for increasing number of receive antennas ; * an upper bound on the outage probability , which is tight for high snr values . for the case of four transmit and an arbitrary number of receive antennas we derive analytical expressions for this bound * finally , we exploit the special structure of the qstbc and apply a new transmit strategy , which decouples the signals transmitted from different antennas in order to detect the symbols separately as in the case of ostbc . the performance of this linear detector is equivalent to the non - linear maximum - likelihood ( ml)-detector in @xcite . the remainder of this paper is organized as follows . in section [ seq : system_model ] , we introduce the system model and establish the notation . the design of qstbc for @xmath3 transmit antennas is shown in section [ sec : diversity_analy ] . the complete characterization of the equivalent channel model and other key achievements of this paper are described in section [ sec : preprocessing ] . as an possible application of the results in this section , we present the analysis of the outage probability achieved with qstbc in section [ sec : perfanaly ] , followed by some simulations and concluding remarks in section [ seq : simulations ] and [ seq : conclusion ] . we consider a system with @xmath0 transmit and @xmath1 receive antennas . our system model is defined by @xmath6 where @xmath7 denotes the ( @xmath8 ) transmit matrix , @xmath9 $ ] the ( @xmath10 ) receive matrix , @xmath11 $ ] the ( @xmath12 ) channel matrix , and @xmath13 $ ] the complex ( @xmath10 ) white gaussian noise ( awgn ) matrix , respectively . an entry @xmath14 of @xmath15 ( @xmath16 ) denotes the complex noise at the @xmath17th receiver for a given time instant @xmath18 . the real and imaginary parts of @xmath19 are independent and @xmath20(0,@xmath21 ) distributed . an entry of the channel matrix is represented by @xmath22 and describes the complex gain of the channel between the @xmath23th transmit ( @xmath24 ) and the @xmath17th receive ( @xmath16 ) antenna , where the real and imaginary parts of the channel gains are independent and normal distributed random variables and @xmath25 is @xmath26 distributed , where @xmath27 is the channel mean or ricean component . the channel matrix is assumed to be constant for a block of @xmath28 symbols and changes independently from block to block . the average power of the symbols transmitted from each antenna is normalized to one , so that the average power of the received signal at each receive antenna is @xmath29 and the signal - to - noise ratio ( snr ) is @xmath30 . it is further assumed that the transmitter has no csi and the receiver has perfect csi . a space - time block code is defined by its transmit matrix @xmath7 , which is a function of the vector @xmath31^t$ ] . the rate @xmath32 of a space - time block code is defined as @xmath33 . in this paper , we focus on rate one qstbc with length @xmath34 , therefore @xmath35 . now , let us split the vector @xmath36 into two vectors , @xmath37 and @xmath38 , for reasons that will be clear later on . the elements of @xmath36 with odd index @xmath23 are collected in @xmath37 and with even index in @xmath38 , respectively . both parts of @xmath36 are given as @xmath39= \mathbf{\gamma}\mathbf{s}^- , \mathbf{x}_{\mathrm{even}}= \mathbf{\gamma}\left[\begin{array}{c } s_{\nicefrac{n_t}{2}+1 } \\ \vdots \\ s_{n_t } \\ \end{array}\right]= \mathbf{\gamma}\mathbf{s}^+,\ ] ] with @xmath40 , where @xmath41 denotes a complex modulation signal set with unit average power , e.g. @xmath42-psk . furthermore , @xmath43 is a unitary matrix . more details on @xmath44 and its effect on the detection scheme will be discussed in section [ sec : trans_strat ] . starting with the well known alamouti scheme @xcite for @xmath45 transmit antennas as a @xmath46\;,\ ] ] the generalization of the transmit matrix for the qstbc with @xmath0 ( @xmath47 ) is done in the following recursive way @xmath48\nonumber\;,\ ] ] where @xmath49 and the diagonal @xmath50 matrix @xmath51 is given by @xmath52 . for the case of @xmath53 transmit antennas we have @xmath54\;.\ ] ] in this work , we use the alamouti scheme as the basis in order to construct the rate one qstbc . however , it is also possible to construct qstbc with rates lower than one based on other ostbc @xcite . in the following section , we perform channel - matched filtering as the first stage of preprocessing at the receiver in order to obtain the equivalent channel model , followed by the decoupling of the system model in two parts . afterwards , we analyze the eigenvalues and the eigenvectors of the resulting equivalent channel , leading to important insights of the properties of qstbc . noise pre - whitening as the second stage of preprocessing at the receiver is considered in section [ sec : noise pre - whitening ] . first of all , we briefly review the usual mimo fading channel without any coordinated coding and the impact of ostbc on the mimo channel in order to provide a better insight into the properties of qstbc . in this case , after channel matched filtering to ( [ eq : system ] ) , we have @xmath55 where @xmath56 is the singular value decomposition ( svd ) of @xmath57 , where the unitary matrices @xmath58,@xmath59 contain the eigenvectors of @xmath57 . the joint density function of the eigenvalues @xmath60 of @xmath61 in @xmath62 in the rayleigh fading case @xmath63 is given as @xcite @xmath64 where @xmath65 is a normalizing factor , @xmath66 and @xmath67 . it is obvious , that the eigenvalues are not independent of each other and it is well known that the matrix of eigenvectors @xmath59 depend on the actual channel realization . in case of ostbc , the following holds for the transmit matrix @xmath68 starting with ( [ eq : system ] ) , after some manipulations and channel matched filtering one arrives at @xmath69 where @xmath70.\ ] ] since there is no interaction between the elements of @xmath36 , the equation above can be decomposed into @xmath71 parts . the resulting equivalent channel for each element of @xmath36 of the ostbc is then a single - input - single - output ( siso ) channel given as @xmath72 which is equal to the frobenius norm of the actual mimo channel matrix @xmath57 . in case of the rate one qstbc discussed in this paper , the actual mimo channel is also transformed into a equivalent channel given as @xmath73 . differently from the ostbc the equivalent channel of qstbc is still a mimo channel , however with very interesting properties like constant eigenvectors and i.i.d . eigenvalues following a noncentral @xmath4-distribution as derived in the following . after rearranging and complex - conjugating some rows of @xmath74 the system equation in ( [ eq : system ] ) can be rewritten as @xmath75 where @xmath76^t$ ] and @xmath77 is given as @xmath78.\ ] ] thus it appears that the @xmath79 , with @xmath80 , are obtained recursively , where the recursion starts with @xmath45 , @xmath81\;.\ ] ] in order to perform channel - matched filtering we multiply @xmath82 from left to ( [ eq : system_h_vorne ] ) to get @xmath83 where the noise vector @xmath84 is spatially colored and @xmath85 is given as @xmath86,\ ] ] where @xmath87 and @xmath88 . an important property of the qstbc the system in ( [ eq : system_zerlegt ] ) can be decoupled into two parts due to the special structure of @xmath85 as described in the following . the decoupling comes from the fact that for qstbc , it holds that @xcite @xmath89 where @xmath90^t=[x_1,0,x_3,0,\dots , x_{n_t-1},0]^t$ ] and @xmath91^t$ ] . the property in ( [ eq : zerlegprobqstbc ] ) is very crucial , because this enables a simple maximum - likelihood decoding algorithm . assuming perfect channel estimation is available , the receiver computes the following decision metric over all possible transmit matrices and decides in favor of the transmit matrix that minimizes the following decision metric based on ( [ eq : system ] ) : @xmath92 after some manipulations , we arrive at @xmath93 where @xmath94 is the trace function . @xmath95 and @xmath96 are given as @xmath97 respectively . the above decision metric can be decomposed into two parts , one of which @xmath98 is only a function of @xmath99 , and the other one @xmath100 is only a function of @xmath101 . thus the minimization of ( [ eq : maximum_liklihood ] ) is equivalent to minimizing these two parts separately . note that , due to the processing at the receiver , the property in ( [ eq : zerlegprobqstbc ] ) is projected on the channel matrix @xmath85 in ( [ eq : system_zerlegt ] ) . the decoupled parts depend either on @xmath37 or @xmath38 given in ( [ eq : x_oddandx_evendef ] ) . thus , it is now possible to write an decomposed system model for each part based on ( [ eq : system_zerlegt ] ) . the decomposed system model for the part with @xmath37 ( and similarly for @xmath38 ) can be rewritten as @xmath102 for illustration , we present two examples for the case of @xmath53 and @xmath103 transmit antennas . ( @xmath53 transmit antennas ) in this case , @xmath104 in ( [ eq : neuer_kanal ] ) is given as @xmath105 \nonumber \;.\ ] ] and @xmath106 appears in ( [ eq : system_zerlegt ] ) as @xmath107\left[\begin{array}{c } x_1 \\ x_2 \\ x_3 \\ x_4 \\ \end{array}\right]+ \mathbf{n}''\;,\ ] ] where @xmath108 and @xmath109 are given as @xmath110 respectively . from ( [ eq : system_zerlegbar_bsp4mt ] ) , it is now directly obvious , that the system equation can be decoupled into two parts , which then can be considered separately . for the case considered in this example , the decomposed system model for @xmath37 ( and similarly for @xmath38 , cf . ( [ eq : system_model_s1s3_colored_noise ] ) ) can be written as @xmath111 + \widetilde{\mathbf{n}}\;,\ ] ] with a non - orthogonal @xmath112 \;.\ ] ] ( @xmath103 transmit antennas ) the same procedure applied here results in a @xmath113 given as @xmath114\;,\ ] ] where @xmath115 the general case of arbitrary @xmath0 and very important insights about the eigenvalue decomposition , the eigenvalues themselves and the eigenvectors of the equivalent channel @xmath73 , which are crucial and necessary for further analysis , e.g. , the derivations of the lower and upper bound , are provided in the following section . in order to completely characterize the equivalent channel with respect to the eigenvalues and eigenvectors , we first look at the properties of matrices with certain structures and then show that the equivalent channel matrices fulfill this special structure . let the matrix @xmath116 , where @xmath117 , has the following recursive definition @xmath118\;.\ ] ] similarly @xmath119\;,\ ] ] where the recursion starts with @xmath120 \;\textrm { and } \ ; \mathbf{n}_2(\alpha_3,\alpha_4)=\left[% \begin{array}{cc } i\alpha_3 & \alpha_4 \\ -\alpha_4 & i\alpha_3 \\ \end{array}% \right]\;.\ ] ] the matrices @xmath121 and @xmath122 have the following eigenvalue decompositions @xmath123 where @xmath124\ ] ] and @xmath125 = \phantom{i}\left[% \begin{array}{cc } \alpha_1+\alpha_2 & 0 \\ 0 & \alpha_1-\alpha_2 \\ \end{array}% \right]\ ] ] @xmath126=i\left[% \begin{array}{cc } \alpha_3-\alpha_4 & 0 \\ 0 & \alpha_3+\alpha_4 \\ \end{array}% \right]\ ] ] immediately the following question follows : is their any structure how to derive the eigenvalues of the matrices of higher @xmath127 , i.e. , if the eigenvalues of @xmath128 are given , how can we compute the eigenvalues of @xmath129 . ( note that , if the eigenvalues of @xmath128 are given it is straightforward to derive the eigenvalues of @xmath130 ) . in order to answer this question we are able to state the following lemma , where the arguments of @xmath116 and @xmath131 are omitted . [ lem : fixedeigenv ] let @xmath132 be as given in ( [ eq : structurea_n]),([eq : structureb_n ] ) , then @xmath132 with @xmath133,@xmath134 have the following eigenvalue decomposition : @xmath135 where @xmath136 @xmath137\mathbf{\pi}_n^h\ ; , \end{aligned}\ ] ] @xmath138\mathbf{\pi}_n^h\ ; , \end{aligned}\ ] ] and @xmath139_{ij}=\delta\left[2j-1-i\right]+\delta\left[2(j-\frac{n}{2})-i\right]\\ ] ] with @xmath140 $ ] denoting the delta function , giving @xmath141=1 $ ] for @xmath142 and @xmath141=0 $ ] for @xmath143 , and @xmath144_{ij}$ ] denotes the @xmath145-element of the @xmath146 permutation matrix @xmath147 . the proof is given in appendix [ sec : svd_vonhtildeallg ] . + it is important to realize that @xmath148 and @xmath149 are constructed with different arguments . the from @xmath85 in ( [ eq : channel_vorzerleg ] ) in even and odd block structure re - sorted equivalent channel matrix @xmath73 in ( [ eq : system_model_s1s3_colored_noise ] ) has exactly the same structure as @xmath116 . therefore , lemma [ lem : fixedeigenv ] can be directly applied to @xmath73 . to emphasis the usefullness of the resulting property of the qstbc , we are able to state the following theorem . [ theo : consteigenvec ] the left and right eigenvectors of the equivalent channel in ( [ eq : system_model_s1s3_colored_noise ] ) of qstbc , which fulfill the recursive construction rule of ( [ eq : neuer_kanal ] ) are given by eq . ( [ eq : eigenvecvn ] ) and therefore constant for any arbitrary channel realization . another important aspect of lemma [ lem : fixedeigenv ] is the fact that the eigenvalues in @xmath148 can be obtained simply by adding the eigenvalues of @xmath150 and @xmath151 in an appropriate manner as done in ( [ eq : building_eigvalues ] ) ( cf . appendix [ sec : svd_vonhtildeallg ] ) , which will be used in the following analysis of the qstbc . [ lem : eigenwertemitmatra ] let @xmath152 , @xmath153 , @xmath154 be as in lemma [ lem : fixedeigenv ] . the eigenvalues of the equivalent channel matrix @xmath73 of the qstbc are given by the recursive equations ( [ eq : eigenvalsn ] ) . let @xmath155 and @xmath156 , where @xmath153 , @xmath157 are the eigenvalues of @xmath158 . then for any @xmath29 and @xmath1 , the eigenvalues @xmath159 of @xmath160 are obtained as follows @xmath161 where the matrices @xmath162 with @xmath163 and @xmath0,@xmath164 are given as @xmath165 , \mathbf{a}_{n_t}^{j+1}=\frac{1}{2}\left[\begin{array}{rr } \mathbf{a}_{\frac{n_t}{2}}^{j ' } & \mathbf{b}_{\frac{n_t}{2}}^{j ' } \\ -\mathbf{b}_{\frac{n_t}{2}}^{j ' } & \mathbf{a}_{\frac{n_t}{2}}^{j ' } \\ \end{array}\right ] % \frac{1}{2}\left(\mathbf{i}_2\otimes\mathbf{a}_{\frac{n_t}{2}}^j \mp i\left[\begin{array}{rr } % 0 & 1 \\ % -1 & 0 \\ % \end{array}\right]\otimes\mathbf{b}_{\frac{n_t}{2}}^{j}\right)\ ] ] with @xmath166 and @xmath167 . the proof is given in appendix [ sec : proof_eigenwertemitmatra ] . [ theo : muirhead ] if @xmath25 is @xmath26 and @xmath168 is an @xmath169 matrix then @xmath170 has a noncentral @xmath171 distribution if and only if @xmath168 is idempotent ( @xmath172 ) , in which case the degrees of freedom is @xmath173 ( where @xmath174 and @xmath175 denote the rank and trace of @xmath168 , respectively ) and @xmath176 . [ lem : matrahermitidempotent ] the matrices @xmath162 are hermitian and idempotent . the proof is given in appendix [ sec : proof_matrahermitidempotent ] . + from lemma [ lem : eigenwertemitmatra ] and [ lem : matrahermitidempotent ] , it is now possible to prove the following theorem . [ lem : unabh_eigenwerte ] let @xmath177 defined as in lemma [ lem : eigenwertemitmatra ] be the diagonal eigenvalue matrix of the equivalent channel matrix of an qstbc , which fulfill the recursive construction rule of ( [ eq : neuer_kanal ] ) . then for any @xmath29 and @xmath1 , the eigenvalues @xmath159 of @xmath160 are pairwise independent and identical noncentral chi - square distributed with @xmath2 degrees of freedom . it is important to realize that @xmath178 and @xmath179 . recall from section [ sec : preprocessing ] , that we have decoupled the system into two independent parts . the derivation above holds therefore for both parts , i.e. , each eigenvalue appears twice , once for each part . the proof of theorem [ lem : unabh_eigenwerte ] is given in appendix [ sec : proof_unabh_eigenwerte ] . since @xmath180 in ( [ eq : system_model_s1s3_colored_noise ] ) is colored noise , the next step is to perform pre - whitening . with the knowledge of the theorems [ theo : consteigenvec ] and [ lem : unabh_eigenwerte ] it is easy to compute the pre - whitening filter @xmath181 at the receiver now . to this end , we need just the eigenvalue decomposition of @xmath73 given as @xmath182 with @xmath155 . therefore the pre - whitening filter is given as @xmath183 . by multiplying @xmath181 from the left to ( [ eq : system_model_s1s3_colored_noise ] ) we arrive at @xmath184 where the entries of @xmath185 are mutually i.i.d . gaussian processes again . in the case of @xmath53 transmit antennas @xmath186 in ( [ eq : system_model_nach_whiting ] ) is given as @xmath187\;,\ ] ] and @xmath188 from theorems [ theo : consteigenvec ] and [ lem : unabh_eigenwerte ] , it is now possible to determine @xmath44 adequately , resulting in a attractive system equation , which allows a very simple but effective ml - decoding . to emphasize this property we formulate the following corollary . [ cor : linmldetec ] by choosing the matrix @xmath44 in ( [ eq : x_oddandx_evendef ] ) as @xmath189 , ( [ eq : system_model_nach_whiting ] ) can be rewritten as @xmath190 at this point , the elements of @xmath191 ( and also @xmath192 ) are completely decoupled , since they experience no interference from each other . thus , a linear ml - detector is able to detect the symbols ( or elements ) transmitted from the antennas separately . the matrix @xmath186 in ( [ eq : system_model_nach_whiting ] ) can be decomposed as @xmath193 where @xmath194 is a diagonal matrix containing the singular values of @xmath186 . since @xmath195 is constant for all channel realizations , we can set @xmath196 without any knowledge of the current channel realization . using ( [ eq : zerlegung_h ] ) in ( [ eq : system_model_nach_whiting ] ) results in ( [ eq : systemwithd ] ) . that concludes the proof . for @xmath53 transmit antennas , @xmath197 and @xmath198 are given as @xmath199 , \mathbf{d}_2=\sqrt{2}\left[\begin{array}{rr } \tilde{\mu}_4 ^ 1 & 0 \\ 0 & \tilde{\mu}_4 ^ 2 \\ \end{array}\right]\;.\ ] ] for @xmath103 transmit antennas , we have the following @xmath200 @xmath201\ ] ] and @xmath202 with @xmath203 based on these new insights , we provide some performance analysis in this section , where we focus on the case of rayleigh fading @xmath63 . the mutual information of a mimo system with @xmath29 transmit and @xmath1 receive antennas with no csi at the transmitter and perfect csi at the receiver by using the optimal transmit strategy is given as @xcite @xmath205 the portion of the mutual information achieved with qstbc is @xmath206 the outage probability @xmath207 achievable with qstbc is defined as the probability that @xmath208 is smaller than a certain rate @xmath32 , i.e. @xmath209\;.\ ] ] unfortunately , the exact analysis of @xmath207 is not available . therefore , we provide a lower and upper bound in the following . the outage probability @xmath207 is lower bounded by @xmath210 by using the arithmetic mean - geometric mean inequality , i.e. @xmath211 with equality if and only if @xmath212 we obtain an upper bound for @xmath208 ( and therefore a lower bound on @xmath207 ) given as @xmath213 where @xmath214 for the general case of arbitrary @xmath29 similar to the cases @xmath53 and @xmath103 . the lower bound on the outage probability @xmath207 can be written as @xmath215\geq\mathrm{pr}[i_q^u < r]= \mathrm{pr}\left[\alpha_1<\frac{n_t}{\rho}\left(2^r-1\right)\right]\;. \nonumber\ ] ] since @xmath108 is chi - square distributed random variable with @xmath216 degrees of freedom , @xmath207 is given as @xcite ) in ( [ eq : poutlowerbound ] ) . that concludes the proof . [ lem : tightness_lowerbound ] the lower bound in ( [ eq : poutlowerbound ] ) gets tight for low snr values or when @xmath1 increases . the proof is given in appendix [ sec : proof_tighnesslowbound ] . using the positive definiteness of @xmath177 , ( [ eq : mutualinfoqstbcdet ] ) can be lower bounded as @xmath217 thus , the upper bound on @xmath207 is given as @xmath218 \leq \mathrm{pr}[i_q^l < r]=p\big[\det\left(\mathbf{d}_{\frac{n_t}{2}}\mathbf{d}_{\frac{n_t}{2}}\right)<\underbrace{\left(\frac{n_t}{\rho}\right)^{\nicefrac{n_t}{2}}\left(2^{\frac{rn_t}{2}}-1\right)}_{\tilde{r}}\big ] \;.\nonumber\end{aligned}\ ] ] for the special case of @xmath53 transmit antennas , @xmath207 is given as @xmath219\;.\ ] ] since @xmath220 is a product of two chi - square distributed random variables , both with @xmath221 degrees of freedom , @xmath207 is given by ( * ? ? ? * , eq.9.9.34 ) @xmath222 where @xmath223 is the psi function ( * ? ? ? * , eq.8.360 ) and @xmath224 is the gamma function @xcite . note that for high snr a useful and simple approximation of the outage probability can be obtained by retaining only the first term ( i.e. @xmath225 ) of the series within the upper bound . some simulation results of the performance of qstbc and their interpretation are presented in the following section . in fig . [ fig : capacity_nr1 ] , the omi of qstbc @xmath208 with our new transmit strategy and our linear detector is compared with the nonlinear ml - detector and zf - detector in @xcite . additionally , the omi of a mimo system with @xmath53 and @xmath226 is depicted . from the fig . , we observe that our new transmit strategy outperforms the zf - detector of @xcite and achieve the same portion of mutual information as the non - linear ml - detector presented in @xcite . and @xmath227 . ] in fig . [ fig : capacity_nr2 ] , the performance of qstbc in terms of omi with @xmath53 and @xmath103 antennas is depicted for @xmath228 . , @xmath103 transmit and @xmath228 receive antennas . ] for @xmath226 , the performance with @xmath103 is similar to the case of @xmath53 transmit antennas ( depicted in fig . [ fig : capacity_nr2 ] ) , i.e. we achieve a significant fraction of the omi . however , by increasing the number of receive antennas , we observe in fig . [ fig : capacity_nr2 ] , that the performance of qstbc with @xmath103 as well as with @xmath53 is dramatically reduced in terms of achievable omi . in fig . [ fig : outage ] , @xmath207 of qstbc with @xmath53 transmit and @xmath226 to @xmath229 and @xmath230 receive antennas is depicted . from the fig . we observe that our lower bound on the performance of qstbc with respect to @xmath207 gets tight for increasing number of receive antennas . even the upper bound performs very well and shows to be useful . transmit and different size of receive antennas @xmath1 , rate=4 . ] the performance of qstbc with respect to @xmath207 is depicted in fig . [ fig : outage_nt8 ] for @xmath103 transmit antennas . similarly to the case of @xmath53 transmit antennas , the lower bound gets tight by using many antennas at the receiver side . transmit and different size of receive antennas @xmath1 , rate=4 . ] in this paper , we generalized the qstbc in @xcite to @xmath3 transmit and an arbitrary number of receive antennas . our most important results are given in the theorems [ theo : consteigenvec ] and [ lem : unabh_eigenwerte ] in section [ sec : eigvalandeigvecequichannel ] and reveal the useful properties of the resulting equivalent channel for qstbc . in more details , theorem [ theo : consteigenvec ] state that the eigenvectors of the equivalent channel are independent of each channel realization , i.e. they are fixed . this property can be directly exploited for preprocessing , which allows a very efficient linear ml detection ( corollary [ cor : linmldetec ] ) . in more details , we developed a new transmit strategy , which allows to apply a linear ml - detector as in the case of ostbc , such that the symbols from different antennas can be detected separately . the performance of our linear detector in terms of mutual information is equal to the nonlinear ml - detector in @xcite . the theorem [ lem : unabh_eigenwerte ] offers insights in the statistics of qstbc . it is proved that the eigenvalues of the equivalent channel are i.i.d . noncentral @xmath4 . these insights were used to derive upper and lower bounds on the outage probability with qstbc . it was shown , both analytically and via simulations , that the lower bound gets tight for increasing number of receive antennas and also in the low snr - regime . from simulation results , we observed that the qstbc approaches the outage mutual information only in the case of @xmath226 receive antenna . by increasing the number of receive antennas , the loss in terms of mutual information increases unbounded with the snr . in the following , the arguments of @xmath231 and @xmath149 are omitted occasionally in order to increase the readability of the paper . the proof is done by the principle of induction . we start with the initial case @xmath232 . from ( [ eq : structurea_n ] ) , it follows @xmath233\stackrel{(\ref{eq : svd_a_2b_2})}{= } \left[% \begin{array}{rr } \mathbf{v}_2\mathbf{s}_2\mathbf{v}_2^h & \mathbf{v}_2\mathbf{t}_2\mathbf{v}_2^h \\ -\mathbf{v}_2\mathbf{t}_2\mathbf{v}_2^h & \mathbf{v}_2\mathbf{s}_2\mathbf{v}_2^h \\ \end{array}% \right ] \nonumber \\ = & \left[% \begin{array}{cc } \mathbf{v}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2 \\ \end{array}% \right]\left[% \begin{array}{cc } \mathbf{s}_2 & \mathbf{t}_2 \\ -\mathbf{t}_2 & \mathbf{s}_2 \\ \end{array}% \right]\left[% \begin{array}{cc } \mathbf{v}_2^h & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2^h \\ \end{array}% \right ] \nonumber\end{aligned}\ ] ] @xmath234\mathbf{\pi}_4 \left[% \begin{array}{cccc } \mu_2 ^ 1 & \nu_2 ^ 1 & \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } \\ -\nu_2 ^ 1 & \mu_2 ^ 1 & & \\ \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } & \mu_2 ^ 2 & \nu_2 ^ 2 \\ & & -\nu_2 ^ 2 & \mu_2 ^ 2 \\ \end{array}% \right]\mathbf{\pi}_4^h \left[% \begin{array}{cc } \mathbf{v}_2^h & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2^h \\ \end{array}% \right ] \nonumber\\ = & \underbrace{\left[% \begin{array}{cc } \mathbf{v}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2 \\ \end{array}% \right]\mathbf{\pi}_4 \left[% \begin{array}{cc } \mathbf{v}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2 \\ \end{array}% \right]}_{\mathbf{v}_4 } \mathbf{s}_4 \underbrace{\left[% \begin{array}{cc } \mathbf{v}_2^h & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2^h \\ \end{array}% \right]\mathbf{\pi}_4^h \left[% \begin{array}{cc } \mathbf{v}_2^h & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_2^h \\ \end{array}% \right]}_{\mathbf{v}_4^h}\ ; , \nonumber\end{aligned}\ ] ] with @xmath235 given as @xmath236\;,\ ] ] @xmath237 , where @xmath238 which is equivalent to @xmath239\mathbf{\pi}_4^h\ ] ] the same procedure applied to @xmath240 results in a @xmath241 given as @xmath242 with @xmath243 which is equivalent to @xmath244\mathbf{\pi}_4^h\;.\ ] ] now assume that the following hypothesis holds for @xmath245 , i.e. @xmath246 then the following inductive step is true @xmath247\stackrel{(\ref{eq : induct_hypothesis})}{= } \left[% \begin{array}{rr } \mathbf{v}_\frac{k}{2}\mathbf{s}_\frac{k}{2}\mathbf{v}_\frac{k}{2}^h & \mathbf{v}_\frac{k}{2}\mathbf{t}_\frac{k}{2}\mathbf{v}_\frac{k}{2}^h \\ -\mathbf{v}_\frac{k}{2}\mathbf{t}_\frac{k}{2}\mathbf{v}_\frac{k}{2}^h & \mathbf{v}_\frac{k}{2}\mathbf{s}_\frac{k}{2}\mathbf{v}_\frac{k}{2}^h \\ \end{array}% \right]\\ = & \left[% \begin{array}{cc } \mathbf{v}_\frac{k}{2 } & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_\frac{k}{2 } \\ \end{array}% \right]\left[% \begin{array}{cc } \mathbf{s}_\frac{k}{2 } & \mathbf{t}_\frac{k}{2 } \\ -\mathbf{t}_\frac{k}{2 } & \mathbf{s}_\frac{k}{2 } \\ \end{array}% \right]\left[% \begin{array}{cc } \mathbf{v}_\frac{k}{2}^h & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_\frac{k}{2}^h \\ \end{array}% \right]\\ = & \left[% \begin{array}{cc } \mathbf{v}_\frac{k}{2 } & \mathbf{0 } \\ \mathbf{0 } & \mathbf{v}_\frac{k}{2 } \\ \end{array}% \right]\mathbf{\pi}_k \left[% \begin{array}{cccc } \mathbf{q}_{\frac{k}{2}}^1 & & \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } \\ & \mathbf{q}_{\frac{k}{2}}^2 & & \\ \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } & \ddots & \\ & & & \mathbf{q}_{\frac{k}{2}}^{\frac{k}{2 } } \\ \end{array}% \right]\mathbf{\pi}_k^h \left[% \begin{array}{cc } \mathbf{v}_\frac{k}{2}^h & \mathbf{0 } \\ \mathbf{0}\mathbf { } & \mathbf{v}_\frac{k}{2}^h \\ \end{array}% \right ] , \nonumber\end{aligned}\ ] ] @xmath248 where @xmath249\ ] ] and @xmath250 with @xmath251 which is equivalent to @xmath252\mathbf{\pi}_k^h.\ ] ] the same procedure applied to @xmath253 results in the same @xmath200 and @xmath254 with @xmath255 which is again equivalent to @xmath256\mathbf{\pi}_k^h.\ ] ] since the initial case of @xmath257 is true and the inductive step is true , the statement in ( [ eq : svd_statement ] ) is true for all @xmath258 . that concludes the proof . the proof is done by the principle of induction . the outline of the proof is as follows . for the initial case of @xmath53 , we need the eigenvalues @xmath259 and @xmath260 for @xmath45 , i.e. the alamouti scheme , as indicated in ( [ eq : eigvalnt4aussummuandnu ] ) . the first step is therefore to construct the eigenvalues for @xmath53 with the eigenvalues of @xmath45 . using ( [ eq : building_eigvalues ] ) and ( [ eq : build_eigvalnu ] ) , we observe that the eigenvalues for @xmath53 can be also obtained with the eigenvalues for @xmath103 , which is the second step revealing an important instruction of constructing eigenvalues @xmath261,@xmath262 from @xmath263,@xmath264 . it follows the hypothesis and the inductive step concluding the proof . now , we start with the well known alamouti scheme . by applying the alamouti scheme ( @xmath45 ) , @xmath160 as well as @xmath37 and @xmath38 are only scalars . thus , the only eigenvalue of @xmath265 of the decomposed system model for the part with @xmath37 ( and similar for @xmath38 ) is given as @xmath266 where @xmath267^t$ ] and @xmath268 similarly , @xmath269 we are now able to start with the initial case of @xmath53 . the first eigenvalue of the qstbc for the part with @xmath37 ( and similar for @xmath38 ) is given as in ( [ eq : building_eigvalues ] ) ( with @xmath156 and @xmath270 ) @xmath271 @xmath272\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow 2,i } \\ \mathbf{h}_{3\rightarrow 4,i } \\ \end{array}% \right]- i\left[% \begin{array}{cccc } -h_{3i}^ * & h_{4i}^ * & h_{1i}^ * & -h_{2i}^ * \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow 2,i } \\ \mathbf{h}_{3\rightarrow 4,i } \\ \end{array}% \right]\right)\nonumber\end{aligned}\ ] ] @xmath273\mathbf{h}_{1\rightarrow 4,i}-i\left[% \begin{array}{cccc } h_{1i}^ * & -h_{2i}^ * & -h_{3i}^ * & h_{4i}^*\\ \end{array}% \right]\frac{1}{2}\left[% \begin{array}{cc } \mathbf{0 } & \mathbf{a}_2 \\ \mathbf{a}_2 & \mathbf{0 } \\ \end{array}% \right]\mathbf{h}_{1\rightarrow 4,i } \label{eq : expansnt2tont4}\\ & = \sum_{i=1}^{n_r}\mathbf{h}_{1\rightarrow 4,i}^h\frac{1}{2}\left[% \begin{array}{cc } \mathbf{a}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{a}_2 \\ \end{array}% \right]\mathbf{h}_{1\rightarrow 4,i}-i\mathbf{h}_{1\rightarrow 4,i}^h\frac{1}{2}\left[% \begin{array}{cc } \mathbf{\theta}_2 & \mathbf{0 } \\ \mathbf{0 } & -\mathbf{\theta}_2 \\ \end{array}% \right]\left[% \begin{array}{cc } \mathbf{0 } & \mathbf{a}_2 \\ \mathbf{a}_2 & \mathbf{0 } \\ \end{array}% \right ] \mathbf{h}_{1\rightarrow 4,i}\nonumber\\ & = \sum_{i=1}^{n_r}\mathbf{h}_{1\rightarrow 4,i}^h\frac{1}{2}\left(\left[% \begin{array}{cc } \mathbf{a}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{a}_2 \\ \end{array}% \right]-\left[% \begin{array}{cc } \mathbf{0 } & i\mathbf{\theta}_2\mathbf{a}_2 \\ -i\mathbf{\theta}_2\mathbf{a}_2 & \mathbf{0 } \\ \end{array}% \right]\right)\mathbf{h}_{1\rightarrow 4,i } = \mathbf{h}_{1\rightarrow 4,i}^h\mathbf{a}_4 ^ 1\mathbf{h}_{1\rightarrow 4,i}\;,\label{eq : eigvalnt4fromalpha}\end{aligned}\ ] ] where @xmath274\;.\ ] ] in an analogous manner , we get @xmath275 given as @xmath276 on the other hand , with ( [ eq : building_eigvalues ] ) we have @xmath277 \mathbf{h}_{1\rightarrow 8,i}\right)\nonumber\\ & = \sum_{i=1}^{n_r}\frac{1}{4}\mathbf{h}_{1\rightarrow 8,i}^h \mathbf{h}_{1\rightarrow 8,i}-i\mathbf{h}_{1\rightarrow 8,i}^h\frac{1}{4}\left[% \begin{array}{cccc } \mathbf{\theta}_2 & & \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } \\ & -\mathbf{\theta}_2 & & \\ \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } & \mathbf{\theta}_2 & \\ & & & -\mathbf{\theta}_2 \\ \end{array}% \right]\left[% \begin{array}{cccc } \mathbf{0 } & \mathbf{a}_2 & \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } \\ \mathbf{a}_2 & \mathbf{0 } & & \\ \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } & \mathbf{0 } & \mathbf{a}_2 \\ & & \mathbf{a}_2 & \mathbf{0 } \\ \end{array}% \right ] \mathbf{h}_{1\rightarrow 8,i}\nonumber \end{aligned}\ ] ] @xmath278 -i\left[% \begin{array}{cccc } \mathbf{0 } & \mathbf{\theta}_2\mathbf{a}_2 & \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } \\ -\mathbf{\theta}_2\mathbf{a}_2 & \mathbf{0 } & & \\ \multicolumn{2}{c}{\multirow{2}{*}{{\mbox{\huge $ 0 $ } } } } & \mathbf{0 } & \mathbf{\theta}_2\mathbf{a}_2 \\ & & -\mathbf{\theta}_2\mathbf{a}_2 & \mathbf{0 } \\ \end{array}% \right]\right ) \mathbf{h}_{1\rightarrow 8,i}\nonumber \end{aligned}\ ] ] @xmath279\right ) \mathbf{h}_{1\rightarrow 8,i}\nonumber\\ & = \sum_{i=1}^{n_r}\mathbf{h}_{1\rightarrow 8,i}^h\left(\frac{1}{2 } \left[% \begin{array}{cc } \mathbf{a}_4 ^ 1 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{a}_4 ^ 1 \\ \end{array}% \right]\right ) \mathbf{h}_{1\rightarrow 8,i}=\sum_{i=1}^{n_r}\mathbf{h}_{1\rightarrow 8,i}^h\frac{1}{2}\left ( \mathbf{i}_2 \otimes \mathbf{a}_4 ^ 1 \right ) \mathbf{h}_{1\rightarrow 8,i}\label{eq : eigvalnt4kronprod } \end{aligned}\ ] ] @xmath280\left[% \begin{array}{cc } \mathbf{a}_4 ^ 1 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{a}_4 ^ 1 \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow 4,i } \\ \mathbf{h}_{5\rightarrow 8,i } \\ 4,i}^h\mathbf{a}_4 ^ 1\mathbf{h}_{1\rightarrow 4,i}+\frac{1}{2}\mathbf{h}_{5\rightarrow 8,i}^h\mathbf{a}_4 ^ 1\mathbf{h}_{5\rightarrow 8,i}\label{eq : eigvalnt4fromnt8}\end{aligned}\ ] ] in an analogous manner , we get @xmath275 given as @xmath281\left[% \begin{array}{cc } \mathbf{0 } & i\mathbf{\theta}_k\mathbf{a}_4 ^ 1 \\ -i\mathbf{\theta}_k\mathbf{a}_4 ^ 1 & \mathbf{0 } \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow 4,i } \\ \mathbf{h}_{5\rightarrow 8,i } \\ \end{array}% \right]\ ] ] since the eigenvalues @xmath282 and @xmath283 can be obtained very easily in a similar way , we omit the derivations here . comparing ( [ eq : eigvalnt4fromnt8 ] ) with ( [ eq : eigvalnt4fromalpha ] ) shows that in order to get the eigenvalues of @xmath103 , the eigenvalues of @xmath53 have to be expanded by using the kronecker product of @xmath284 and @xmath285 , divided by @xmath286 in order to incorporate the channel entries @xmath287 as given in ( [ eq : eigvalnt4kronprod ] ) . actually , this can also be observed in the expansion from @xmath45 to @xmath53 by comparing ( [ eq : eigenvalalam ] ) with the first addend in ( [ eq : expansnt2tont4 ] ) . now assume that the following hypothesis holds @xmath288\mathbf{h}_{1\rightarrow 2k , i}\label{eq : eigvalntsubtoh1toh2k}\;,\end{aligned}\ ] ] similarly @xmath289\mathbf{h}_{1\rightarrow 2k , i}\label{eq : eigvalnuntsubtoh1toh2k}\;,\end{aligned}\ ] ] then the following inductive step is true @xmath290\frac{1}{2}\left[% \begin{array}{cc } \mathbf{a}_k^{j ' } & \mathbf{0 } \\ \mathbf{0 } & \mathbf{a}_k^{j ' } \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow k , i } \\ \mathbf{h}_{k+1\rightarrow 2k , i } \\ \end{array}% \right]\\ \phantom{= } & \mp [ \mathbf{h}_{1\rightarrow k , i}^h \mathbf{h}_{k+1\rightarrow 2k , i}^h]\frac{1}{2}\left[% \begin{array}{cc } \mathbf{0 } & i\theta_k\mathbf{a}_k^{j ' } \\ -i\theta_k\mathbf{a}_k^{j ' } & \mathbf{0 } \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow k , i } \\ \mathbf{h}_{k+1\rightarrow 2k , i } \\ \end{array}% \right]\nonumber\end{aligned}\ ] ] @xmath291\frac{1}{2}\left[% \begin{array}{cc } \mathbf{a}_k^{j ' } & \mp i\theta_k\mathbf{a}_k^{j ' } \\ \pm i\theta_k\mathbf{a}_k^{j ' } & \mathbf{a}_k^{j ' } \\ \end{array}% \right]\left[% \begin{array}{c } \mathbf{h}_{1\rightarrow k , i } \\ \mathbf{h}_{k+1\rightarrow 2k , i } \\ \end{array}% \right ] \nonumber\end{aligned}\ ] ] @xmath292\mathbf{h}_{1\rightarrow 2k , i}\;,\end{aligned}\ ] ] with @xmath293 and @xmath167 . that concludes the proof with [ eq : def_matrizenant ] , @xmath294 and @xmath295 , @xmath296 and @xmath167 are given as @xmath297\;.\ ] ] thus , @xmath298 are only hermitian , if @xmath299 is hermitian . since @xmath300 is hermitian , it follows that @xmath162 , @xmath0,@xmath296 are hermitian . similarly , @xmath301\;.\ ] ] thus , @xmath298 are only idempotent , if @xmath299 is idempotent . since @xmath300 is idempotent , it follows that @xmath162 , @xmath0,@xmath296 are idempotent . that concludes the proof . we first prove the independency of the eigenvalues . since the matrices @xmath302 are hermitian and idempotent , we are able to rewrite ( [ eq : eigenval_sum ] ) as @xmath303 independency between the eigenvalues in the gaussian case is given if and only if the eigenvalues are uncorrelated , i.e. @xmath304=0\ ; \text { $ \forall j , j\neq k$}\ ; , \nonumber\ ] ] which is fulfilled if @xmath305 by applying the eigenvalue decomposition to ( [ eq : unabh_eigenwerte ] ) , one has to distinguish between the case , where the eigenvalues are given as @xmath306 and @xmath307 with @xmath167 , @xmath308 and @xmath309 . note that the entries on the @xmath310th - diagonals of the matrices @xmath311 , where @xmath312 ( @xmath142 represents the main diagonal , @xmath313 above the main diagonal , and @xmath314 below the main diagonal ) , are equal to zero . due to the special structure of the matrices @xmath311 , it follows that @xmath315 is orthogonal to @xmath316 and @xmath315 is orthogonal to @xmath317 , with @xmath318 . thus , the eigenvalues in ( [ eq : eigenvaluesi ] ) and ( [ eq : eigenvaluesii ] ) are zero and therefore the eigenvalues in ( [ eq : eigenval_sum ] ) are independent . the probability density function ( pdf ) of the eigenvalues @xmath319 can be obtained from ( [ eq : eigenval_sum ] ) as follows . the rank @xmath320 of @xmath321 is @xmath322 . furthermore , @xmath323 \nonumber\\ & = & \mathrm{rk}((\mathbf{i}+\mathbf{\theta}_{n_t})\mathbf{a}_{2^{n-1}}^{j'})+\mathrm{rk}((\mathbf{i}-\mathbf{\theta}_{n_t})\mathbf{a}_{2^{n-1}}^{j ' } ) = \mathrm{rk}(\mathbf{a}_{2^{n-1}}^{j'})\ ; , \nonumber\end{aligned}\ ] ] where @xmath58 contains the eigenvectors of @xmath324 . since @xmath325 , the matrices @xmath326 have all rank @xmath322 , and thus the following holds @xmath327\;,\ ] ] where @xmath59 is a unitary matrix . with ( [ eq : svd_a_idem ] ) , the pdfs are given as @xmath328\right ) = p\left(\mathrm{tr}\left[\sum\limits_{i=1}^{n_r}\mathbf{\bar{h}}_i^h\left[% \begin{array}{cc } \mathbf{i}_2 & \mathbf{0 } \\ \mathbf{0 } & \mathbf{0 } \\ \end{array}% \right]\mathbf{\bar{h}}_i\right]\right)\;,\nonumber\end{aligned}\ ] ] which is the sum of squares of @xmath329 independent complex normal distributed variables , i.e. a noncentral chi - square distribution with @xmath2 degrees of freedom . that concludes the proof . the inequality ( [ eq : mutualinfoupperbound ] ) is tight only , if the ratio of two eigenvalues , i.e. @xmath330 , for all @xmath331 . from this it follows that it has to be shown that the following holds @xmath332 since the eigenvalues are chi - square distributed with each @xmath2 degrees of freedom , the ratio of the eigenvalues is distributed as follows @xmath333 which is the well - known f distribution ( * ? ? ? * , eq.9.9.35 ) . therefore , when @xmath1 goes infinity , the f distribution is given as @xmath334 where @xmath335 is the delta distribution . it follows that the lower bound gets tight for increasing @xmath1 . the lower bound is also tight for low snr values , which is obvious after rewriting ( [ eq : mutualinfoqstbcdet ] ) as follows @xmath336 furthermore , @xmath337 for small snr . as an example , @xmath338 for @xmath53 . therefore , as the snr gets smaller , the lower bound gets tighter . that concludes the proof .

are concurrent computing machines equivalent to turing machines ? this question , which amounts to confront two fundamental notions like time and computation may be treated in a purely mathematical framework . practical consequences however can not be independent of concrete realizations , that is concrete machines performing actual computations in physical time . this remark may seem curious , if one aims at showing theorems , which can not depend on the physical properties of time or machines . but , even in a mathematical treatment of concurrent computation , one needs a representation of time . usually , time is modelized as a real parameter , shared by all parts of the computation . unfortunately , such a representation does not correspond to the observable time that can be obtained from physical systems like clocks , neither to the reference time that is defined by metrology , nor to the operational time that occurs in practical realizations of logical circuits . without questioning the validity of demonstrated theorems , difficulties may emerge when trying to find practical applications . the distinction just made between abstract and concrete machines raises related questions . when a machine @xmath0 can be simulated by a program @xmath1 running on another machine , how can one identify the concrete machine @xmath0 and the program @xmath1 ? and in case a machine can not be simulated on another one , indicating some greater expressive power , is the latter due to computation or to some fundamental physical property ? before entering such questions , one must first consider the sequential machines that are presently realized . these machines which we get from a constructor , which are made of matter , which transform electric energy into heat , which we communicate with through a keyboard and a screen , why do we need them ? let us note that all the features that make them concrete are inconveniences : we would prefer them lighter , smaller , less power consuming and less dissipating . it would be ideal to make all these parameters equal to zero . in fact , we need them for their logical function , their ability to compute . but then , since this function is mathematically known , modelized and even simulated , where is the need for a concrete machine , whose features are mainly inconveniences ? the natural answer is that these machines compute faster than humans can do , with just a pencil and paper . and yet , pencil and paper are already rudimentary elements of a concrete machine , using physical objects to memorize different steps of computations . the interest in concrete machines comes from their intimate relation with physical time . if computing machines go faster than humans , then one must be confident in their action , as in most cases one is unable to check their output . indeed , in very specific cases , one can verify the correctness of a result in much less time than is needed to obtain it . this is the case for instance of the prime factorization of integers . but few concrete applications have this property . in most cases , one can not check the result in much less time than the computation itself . if the result is important , and no other way is available to obtain it , then one must be confident in the machine . what can support such a confidence ? necessarily reasoning , founded on correct functioning of the machine at a given time on some particular computations , generalized to other times and other computations . a computing machine can not be tested for all computations it can do , at any time . even for a finite machine , the number of possible computations increases exponentially with the memory size , and a memory of one hundred bits already allows a number of configurations that can not be tested in less time than the age of the universe . to establish a reasoning leading to confidence , one must : - check that each elementary component effectively realizes the function it has been designed for ( physical validation ) . - prove in a deductive way that the particular composition of these elementary components building the machine effectively leads to the global function used ( logical validation ) . the first condition is ensured by choices in implementation design and by tests made by the constructor . the second condition is obtained from a mathematical representation of the machine and from the logics of computation . these two steps of validation require good representations of all components at the physical level , and of the global machine at the logical level . if the confidence one can put in a machine relies on good modelizations of both its physical and logical functioning , how could such a machine perform more than it has been designed for , more than our present theories can modelize ? even if the existence of a new type of calculus , still unknown today , can be envisaged , with machines performing this new type of calculus , how could one build such machines without having for them good modelizations ? in such a case , one could not ensure the two validation steps , and one could not say that these machines operate correctly neither be confident in their ouput . one consequence is that realistic models are necessary , both of the computational structure and of the physical implementation of logical operators . modelization is made easier when logical and physical constraints can be separated . this is the reason for developing sequential machines or synchronous concurrent machines . in that case , the logical validation of the machine can be made , whilst ignoring the implementation characteristics of its components . the latter will finally and mainly limit the performance of the machine through the value of the clock frequency . the machine can equivalently be simulated on another concrete machine with identical clock frequency , at the expense of slower performances . however , in the case of asynchronous concurrent machines , logical and physical constraints are more involved . although machines built with asynchronous circuits are less widely used , much effort has been devoted to their understanding and modelization @xcite . in fact , they may even appear as an unavoidable evolution of computing machines . on one hand , clock timed circuits are reaching limits where clock signal distribution consumes too many resources and progress in performances approaches saturation point . on the other hand , asynchronous circuits constitute the most general class of circuits , and thus allow one to express in the most general way the questions raised by the implementation of computation on physical systems , and the solutions that may be brought . in this article , we shall be concerned with the relation of concrete computing machines with physical time . after recalling the general characteristic properties of physical time and of computing machines which are presently realized , we shall compare the solutions provided by synchronous and asynchronous systems to the implementation of logical operations . we shall show that they give different implementations of causal relations , reflecting in that way different causal structures for space - time . the notion of time may be seen to follow from two necessities . from a logical point of view , time can be considered as the concept which allows one to make a distinction between two different types of propositions : general and universal propositions ( like mathematical ones ) which are eternal , and particular propositions which are related to changing reality ( like those describing physical systems ) . moreover , time is also rendered necessary by the formulation of physics : time is the concept which allows one to give a formal expression to movement , and hence to the laws of physics . properties of time are in fact imposed by the functions that this notion must fulfill . from the logical side , time allows one to conceive a same object by characterizing it by its different states , these states being asssociated with the object at different times . a time parameter can then be used not only to index the different states characterizing a same object , but also to organize the states of different objects into classes of simultaneity . the relation of order that can be introduced on the time parameter allows one to define a relation of logical causality between the state transitions affecting different objects . however , in order to be realized physically , for instance on real machines , the causal relation between states can not be independent of the real motions affecting physical systems . in particular , the simultaneity classes defined with the help of the time parameter must coincide with those that are associated with real events occuring in physical space , hence with the physical time . the notion of physical time is intimately related to the laws of physics . after having remarked that pendulum oscillations are isochronous , galileo galilei could give a mathematical representation of motion induced by free fall , by relating the undergone distance to the elapsed time , the latter being understood as a universal reference for all motions . the existence of such a reference is made possible by the existence of physical laws governing all movements , and in particular by the existence of regular movements like inertial motions . this introduction of time leaves an important conventional part in the definition of a time reference , even if a natural choice is provided by motions which appear as most regular , like the earth motion around the sun . this leads in fact to distinguish two types of time . thus , leibniz @xcite , relying on logical arguments , could consider that space and time are mere relations between objects or events , which are fixed by an observer in a conventional way , thus building subjective space and time . still , one is also bound to admit the existence of objective space and time , as the only way to understand how physical laws governing displacements of objects and time ordering of events can be formulated in a universal way , independently of the observer . the formulation of the universal law of gravitation led newton @xcite to fix the role played by time in physical laws , and to endow it with the mathematical representation that we still use nowadays : that of a real parameter which all physical quantities depend on . in fact , newton introduced two different notions of time , which he distinguished both in their conception and in their usage . the first one , which he called `` absolute and mathematical '' , allowed him to write mathematical equations for the laws of mechanics and gravitation . the second notion , which he called `` common and sensible '' , allowed him to relate the motions of different physical systems , including clocks . even if newton privileged the first notion , which he considered as representing absolute space and time , seeing clocks as systems to be improved in order to make them as close as possible to ideal space and time , he nevertheless made two distinct uses of these notions . the first one , which identifies with the curvilinear coordinate on the planet s trajectory , he used as a mathematical tool to deal with infinitesimals of different orders . the second one , which is the physical time as can be defined by kepler s area law , he used as a measure of inertial motions , which he compared planetary motions with . the theory of relativity @xcite has led to question the a priori and absolute character of physical space and time . according to relativity , the notion of time relies on clocks , the date of an event being defined by coincidence of this event with a top delivered by a clock located at the same place . but in order to be defined in whole space , the notion of time also relies on the exchange of light signals , which are necessary to compare and synchronize the indications of remote clocks . the universal and finite velocity of light propagation then leads to a definition of time simultaneity which depends on the observer s motion . in other words , time simultaneity is not given a priori but results from a construction , or clock synchronization . by exchanging light signals , on which time references provided by clocks are encoded , one can compare these references and synchronize clocks . then , time allows one to construct space . by comparing the light signals received from several remote clocks , one can , by quadrangulation , determine positions both in time and space . this relativistic definition of time and space is rendered necessary as soon as a high precision must be attained . this is the case for instance when corrections linked to the finite velocity of light , or relativistic effects , must be taken into account @xcite . hence , this relativistic definition is the one used in physics for high precision space - time measurements @xcite , and in metrology to define time and space standards @xcite and to construct the space - time reference systems required by physics @xcite . it is also the one used in modern practical positioning systems at the surface of the earth , like gps @xcite . finally , as it will appear in the following , it is also the notion of time which is implicitly used by asynchronous communicating and computing systems @xcite . the consequences of the theory of relativity on our conception of space and time have been remarkably discussed at the logical level by russell @xcite . our representation in terms of permanent material structures located in space and evolving according to a unique external time , must be replaced by that in terms of events which are located both in space and time . this conception of space - time not only affects the formulation of modern theories in a fundamental way @xcite , but also underlies present applications in physics and metrology @xcite . when refering to physical time , simultaneity classes can not be defined a priori any more , and rely on a physical implementation by means of propagating light signals . this constructive character of time has important consequences on the functioning of devices which rely on the physical exchange of information . causal relations between events can not be derived by simple comparison with an external , a priori given , parameter . for systems which are unlocalized in space , like communicating processors , this means that the time order relation of occuring events , even if it can be defined unambiguously at the local level of each processor , nevertheless requires a more complete representation to be defined over the whole system in a consistent way @xcite . in the following , we shall analyse how the functioning of actual devices depends on the causal structure of physical space - time . to discuss the intimate relation between time and computation , one must first recall some general principles which underly the physical implementation of computing systems , and which are applied in concrete machines realized with present technologies . * 3.1 implementation of logical operations * in cmos ( complementary mos ) technology , logical gates are implemented using two electrical networks @xmath2 and @xmath3 , as represented on figure 1 . @xmath4 describe input channels , and @xmath5 the output channel . @xmath2 and @xmath3 are built with electrical switches which are combined in series / parallel networks , thus allowing to implement the logics of propositions @xcite . ( 0,0 ) # 1#2#3#4#5 @font ( 1530,2808)(856,-2635 ) ( 2341,-1051)(0,0)[lb ] ( 1711,-826)(0,0)[lb ] ( 1036,-1816)(0,0)[lb ] ( 856,-1276)(0,0)[lb ] ( 2386,-1276)(0,0)[lb ] ( 1711,-1816)(0,0)[lb ] ( 1396,-151)(0,0)[lb ] ( 1036 , 29)(0,0)[lb ] ( 1036,-2401)(0,0)[lb ] ( 1396,-2581)(0,0)[lb ] ( 1036,-781)(0,0)[lb ] fig . 1 . a logical gate . each switch is implemented with a transistor . a function implemented by @xmath6 will be said to be true , for a particular set of values of @xmath4 , if and only if there is a path that connects extreme connections of @xmath6 . for instance , for the network represented by figure 1 : - when @xmath2 is _ true _ , output @xmath5 is forced to value _ true_. - when @xmath3 is _ true _ , output @xmath5 is forced to value _ false_. the possibility that ( @xmath7 ) for some values of @xmath4 must be excluded , otherwise current could flow both through @xmath2 and @xmath3 , resulting in a short circuit between voltage sources . in the following , we shall always impose that ( @xmath8 ) is verified for all configurations of variables @xmath4 ( @xmath9 , @xmath10 and @xmath11 define as usual negation , logical disjunction and logical conjunction ) . ( 0,0 ) # 1#2#3#4#5 @font ( 3117,2025)(1126,-2401 ) ( 2476,-1141)(0,0)[lb ] ( 1801,-2401)(0,0)[lb ] ( 1306,-1861)(0,0)[lb ] ( 4006,-781)(0,0)[lb ] ( 1306,-1501)(0,0)[lb ] ( 1126,-826)(0,0)[lb ] ( 1666,-826)(0,0)[lb ] ( 3151,-646)(0,0)[lb ] ( 1531,-511)(0,0)[lb ] ( 2026,-511)(0,0)[lb ] ( 3151,-1051)(0,0)[lb ] fig . 2 . nand gate . two cases must be considered . 1- @xmath2 and @xmath3 are always opposite , for all values of variables @xmath4 , ( @xmath12 ) . this case implements propositions of classical logic ( complementary mos ) . the simple example of the _ nand _ gate is represented on figure 2 ( a bubble represents negation ) : @xmath13 2- @xmath2 and @xmath3 can be simultaneously _ false _ , i.e. ( @xmath14 ) can be _ true_. in such configurations , the output @xmath5 is not connected to any voltage source . then , because of electrical capacities , @xmath5 memorizes its previous value . this allows one to realize memories , like the _ latch _ represented in figure 3 : @xmath15 ( 0,0 ) # 1#2#3#4#5 @font ( 1620,1179)(631,-2761 ) ( 2251,-1951)(0,0)[lb ] ( 946,-1951)(0,0)[lb ] ( 946,-2536)(0,0)[lb ] ( 1981,-1726)(0,0)[lb ] ( 721,-2761)(0,0)[lb ] ( 1576,-2131)(0,0)[lb ] ( 631,-1726)(0,0)[lb ] fig . 3 . transparent latch . in practice , in order to have memorization last for long enough , and quite generally for all memories , one must ensure memory stability by using some feed - back , by means of a looped amplifier . this feed - back can be permanent ( static ) or recurrent ( dynamic logic ) . one possibility of electrical feed - back is shown in figure 4 , where two looped amplifiers have been added on output @xmath5 , one of them being weak , in the sense that it can not create any serious short circuit when it conflicts with any of the two networks @xmath2 and @xmath3 . ( 0,0 ) # 1#2#3#4#5 @font ( 1992,2385)(1126,-2401 ) ( 1801,-151)(0,0)[lb ] ( 1126,-781)(0,0)[lb ] ( 1711,-871)(0,0)[lb ] ( 1711,-1816)(0,0)[lb ] ( 1801,-2401)(0,0)[lb ] ( 2566,-826)(0,0)[lb ] ( 1126,-1816)(0,0)[lb ] ( 2971,-1141)(0,0)[lb ] fig . 4 . electrical feed - back . another way to realize a stable memory is to implement a static feed - back on a logical gate corresponding to the first case . then , the quoted _ latch _ can be realized with a looped multiplexer ( _ mux _ ) , as shown in figure 5 : @xmath16 ( 0,0 ) # 1#2#3#4#5 @font ( 1632,1497)(991,-2491 ) ( 2611,-1501)(0,0)[lb ] ( 991,-1681)(0,0)[lb ] ( 991,-2491)(0,0)[lb ] ( 1711,-1186)(0,0)[lb ] fig . transparent latch with multiplexer . this exhibits a very general difficulty which characterizes looped systems : variable @xmath5 appears on both sides of its defining equation ( [ latchdef ] ) . this equation does not mean that an equality must be realized , for instance with electric voltages , but that the following assignment must be realized : @xmath17 @xmath18 in other words , two values of the variable @xmath5 must be distinguished , which correspond to successive times : @xmath19 ( after ) and @xmath20 ( before ) are linked by equation ( [ latche ] ) . it is required that the two values @xmath19 and @xmath20 do not interfere , and that variable @xmath5 change from @xmath20 to @xmath19 . the assignment represented by equation ( [ latcha ] ) expresses a causality requirement that must be implemented in order to realize computations . in the particular case of the _ latch _ just described , operation may only cause problem in case @xmath21 is falling . indeed , in other cases : - when @xmath21 is low , @xmath5 is memorized - when @xmath21 is high , @xmath5 copies @xmath22 - when @xmath21 rises , @xmath5 begins to copy @xmath22 so that the circuit operates correctly in these three cases . however , if @xmath21 falls down while @xmath22 changes , @xmath5 will hesitate between two values of @xmath22 . the whole circuit may enter a metastable state which is invalid ( electric voltage will stay in metastable balance at an intermediate level ) and which may last for an unbounded time . it may leave this state for any of the two possible values of @xmath5 , and this in an undeterministic way , which may not be eventually acceptable for the type of computation envisaged . let us note that the circuit of figure 4 shows the same defects , for it involves a feed - back , although this may be less apparent when treated at the electrical level . although chosen here as an example , the _ latch _ shows properties which are encountered quite generally in looped devices . this brief discussion shows that the correct operation of a circuit can not be analysed without taking its environment into account , in particular the time ordering relations of input and output signals . this is entailed by the existence of loops and must be dealt with quite generally , for computing machines are naturally looped systems . in all cases , time constraints must be implemented in order to ensure the causality relations which are necessary for computation . these constraints will take very different forms , according to the type of implementation chosen , whether by means of synchronous or asynchronous systems . before discussing separately these two classes of systems , we shall first recall one important property they share , as it is also imposed by implementation of complex computations , the property of modularity . * * there exist many various ways to organize electronic components into logical circuits , in order to realize machines performing computations . usually and quite generally , one defines complex circuits as hierarchies built with elementary circuits called primitives . this method , imposed by practical considerations , indeed hints at a logical necessity : one must be able to design and realize with the same rigour circuits of increasing complexity . more precisely , one must insure that circuits implementing logical functions of high complexity level behave as they should , and one must obtain this confidence in a rather short time . because of the exponential increase of the number of configurations to check , this requirement implies that a direct physical test of the circuit s behaviour soon becomes impossible when the complexity of the logical function increases . this aim can then only be attained with the help of modular implementations , by taking advantage both of their composite logical structure , and of the logical simplicity of chosen primitives @xcite . proofs relying on known properties of composition of primitives may be developed , which allow one to deduce the correct functioning of a whole modular complex from that of its constituent primitives . then , a test of the whole complex reduces to that of some of its constituents , which are logically simple . although efficient , such strategy may not reveal itself so straightforward . according to the type of physical implementation chosen for the primitives , problems may appear which prevent the systematic development of complex circuits operating correctly , and which do not occur when the choice of primitives is modified , or when pecular constraints are put on their composition . then , there results that logical and physical aspects of modular implementations must be analysed concurrently . time appears in computing systems very early , already in the definition of the electronic circuits which implement logical functions . most circuits which are known and used are synchronous . synchronous circuits may be defined as automata whose transitions between successive states are triggered by pulses delivered by a global clock . an alternative class of circuits is provided by asynchronous circuits . in this section , we introduce the strategies followed by these two main classes of circuits for making the implementation of computation effective , and , in particular , for dealing with problems of computation interferences . * 4.1 synchronous circuits * vlsi circuits which are produced nowadays are highly concurrent devices ( a microprocessor can contain up to @xmath23 transistors , i.e. @xmath24 logical gates ) , and yet most of them can be modelized as non concurrent devices and can be considered as single finite automata . this property comes from their synchronous character , which means that all operations on internal memories are simultaneously activated by a single pulse of a global clock shared by the whole circuit . synchronous implementations use a global clock to avoid the stability problem which has been discussed in previous section . more precisely , with same notations , all _ latches _ are systematically operated in such a way to ensure that logical variables @xmath22 be stable when variables @xmath21 are falling . there exist many different types of memories , but all present the same problem , reflecting the time character of logical assignment . for the sake of simplicity , we shall only discuss the case of the previous _ latch _ , and consider it as a generic example . a synchronous device using two items of this _ latch _ for each register bit ( master - slave _ flip - flop _ ) is sketched on figure 6 . the enabling signals @xmath25 and @xmath26 are mutually excluded in time , and are derived systematically from a common clock . the output is fed back under the form of input variables @xmath27 into a combinatorial operator . if the clock period is larger than the feedback time , then variables @xmath28 are always stable when @xmath25 is falling and the _ latches _ act as required , i.e. , they make the iteration of the combinatorial function effective . the global state @xmath29 is encoded by the state of all memory bits , and can only change at the arrival of a clock pulse . regarding specification and design , synchronous circuits may be considered as modular composites , where primitives , and other modules as well , are finite automata of the type described by figure 6 . the register encodes the state of the circuit , while the combinatorial operator represents the implementation of its transition function . all registers are activated by a single clock pulse . by connecting several automata of this type , one obtains another automaton of the same type , only with a larger memory and a more complex combinatorial function . in such a representation , and from a logical point of view , neither time nor space are involved . one only needs to consider the successive logical steps associated with successive clock periods . the logical time of a computation reduces to a mere integer , which one only relates to physical time by multiplying it by the mean clock period . in other words , time is discretized . ( 0,0 ) # 1#2#3#4#5 @font ( 3072,5916)(946,-6010 ) ( 1711,-3031)(0,0)[lb ] ( 2926,-2401)(0,0)[lb ] ( 3331,-1411)(0,0)[lb ] ( 946,-1951)(0,0)[lb ] ( 2746,-1411)(0,0)[lb ] ( 3241,-2176)(0,0)[lb ] ( 2746,-646)(0,0)[lb ] ( 2881,-2176)(0,0)[lb ] ( 1666,-1276)(0,0)[lb ] ( 1846,-1411)(0,0)[lb ] ( 3511,-1816)(0,0)[lb ] ( 1621,-5281)(0,0)[lb ] ( 1621,-4921)(0,0)[lb ] ( 3376,-5551)(0,0)[lb ] ( 1936,-5956)(0,0)[lb ] ( 3376,-3391)(0,0)[lb ] ( 1261,-3391)(0,0)[lb ] ( 1936,-3616)(0,0)[lb ] ( 2746,-3616)(0,0)[lb ] ( 2116,-4201)(0,0)[lb ] ( 2926,-4201)(0,0)[lb ] fig synchronous circuit . the only physical constraint one must impose is that the clock frequency be smaller than the limit value necessary for all internal propagations to be performed in less time than the clock period . space neither plays a role in the logical function . the whole circuit may be considered as local , i.e. propagation times need only to be taken into account when circuits are connected on large distances , that is when propagation times are large when compared to the clock period , as is the case when computers are connected . implementation on a silicon chip must take into account and control all propagation times within a circuit , so that to insure that all inputs become stable before the end of each clock period . the clock signal must be implemented so that it arrives simultaneously at all _ latches _ , that is with negligible delays when compared to the clock period . clearly , such properties can only be checked once the whole circuit has been specified . such type of circuit can not be implemented incrementally , i.e. by implementing a part without any knowledge on its connections with other parts and on the clock frequency of the whole circuit . in other words , all parts must be local at every scale , from primitives to the whole circuit . the synchronous approach is however questioned in present vlsi designs . this arises mainly because of difficulties which are encountered when distributing simultaneously a same clock signal to millions of _ latches _ , over several @xmath30 , and at a frequency of the order of a gigaherthz . clock distribution results in using an important part of the chip surface and in producing an important part of the overall dissipation . * 4.2 asynchronous circuits * asynchronous circuits may be defined in opposition to synchronous circuits , by the requirement of not using a global clock . but , rather than being complementary , the class they build includes synchronous circuits . usual models represent asynchronous circuits as general devices which are distributed and communicate along connecting channels , as shown on figure 7 . the activity of such circuits is not ruled by the pulses of a global clock , but proceeds with communications distributed between many concurrent parts . these devices can be simple logical gates ( a few transistors ) or , at the opposite , complex processors . communications can be realized through a single wire or through a complex network . clearly , concurrency can not be ignored any more . indeed , one can no more define a logical state which would be associated with the global circuit at a definite time . for each device can change its state following a communication , without being synchronized with most other devices . the notion of computing step itself must be revised , as it relies on a total ordering of all logical events . ( 0,0 ) # 1#2#3#4#5 @font ( 2319,1374)(1024,-2008 ) ( 2071,-916)(0,0)[lb ] ( 1081,-1411)(0,0)[lb ] ( 2971,-1141)(0,0)[lb ] ( 2206,-1861)(0,0)[lb ] fig . asynchronous circuit . problems of computation interference may then arise at two different levels . at lowest level , functioning of a single component may be endangered by computing interferences within the component itself , due to internal loops and instabilities of internal variables . at highest level , composition of asynchronous circuits may induce computation interferences due to exchanges between one module and its environment , the latter sending signals which conflict with the module operation . we shall discuss the second case only and shall assume that primitives may be defined which are free from internal computing interferences ( see for instance @xcite ) . we first briefly describe those that are most frequently used in asynchronous circuits . in some examples , logical functions are defined in a way which does not distinguish between rising and falling edges . these undistinguished transitions are called events , and the logical function operates on these events . but the events are still transitions between different levels ( or boolean variables ) , so that each primitive can be considered in both ways , either as an operation on boolean variables or as ( another ) operation on events . a most frequently used primitive is the _ join _ element , or muller s _ c - element_. it has two inputs @xmath31 and @xmath32 and one output @xmath5 , and its logical function can be described in the following way : - if @xmath33 , then @xmath34 - if inputs become different from one another , then @xmath5 keeps its previous value . then , ouput @xmath5 only changes after both inputs @xmath31 and @xmath32 have changed . this allows a rendez - vous to be realized between levels ( wait until two inputs acquire the same value ) or between events ( wait until two inputs have received the same number of rising or falling edges , after proper initialization ) . the primitive _ c - element _ can be realized using an electrical feed - back analogous to that of figure 4 . along this line , a frequent realization is represented on figure 8 . ( 0,0 ) # 1#2#3#4#5 @font ( 2352,2004)(451,-1738 ) ( 2656,-646)(0,0)[lb ] ( 1711,-1681)(0,0)[lb ] ( 1711 , 74)(0,0)[lb ] ( 2161,-331)(0,0)[lb ] ( 451,-466)(0,0)[lb ] ( 451,-826)(0,0)[lb ] fig . 8 . simplified c - element . another primitive is the _ toggle _ , represented in figure 9 , which possesses one input @xmath22 and two ouputs @xmath35 and @xmath36 . successive events on the input are alternatively sent to outputs @xmath35 and @xmath36 . the first event after initialization is sent to the marked output @xmath35 . the _ or _ operation between events , also called _ merge _ , can be realized with a classical _ exclusive _ _ or _ gate ( _ xor _ between levels ) . the _ sequencer _ , represented in figure 9 , possesses three inputs @xmath31 , @xmath32 and @xmath37 and two outputs @xmath35 and @xmath36 . its role is to grant a given resource to one of two different processes which can make requests on inputs @xmath31 and @xmath32 . when an event is received on @xmath37 , a granting event is produced either on @xmath35 or @xmath36 according to an existing request respectively on @xmath31 or @xmath32 . when two requests are present , the _ sequencer _ arbitrates between the two , and thus introduces some part of indeterminism . the _ sequencer _ may take an unbounded time to arbitrate , but it is required to realize the mutual exclusion of the two grant signals . ( 0,0 ) # 1#2#3#4#5 @font ( 1947,3435)(901,-3661 ) ( 1891,-1276)(0,0)[lb ] ( 2476,-3661)(0,0)[lb ] ( 2791,-2131)(0,0)[lb ] ( 2791,-1231)(0,0)[lb ] ( 2746,-2806)(0,0)[lb ] ( 2791,-1456)(0,0)[lb ] ( 2746,-3076)(0,0)[lb ] ( 1936,-1951)(0,0)[lb ] ( 1936,-2176)(0,0)[lb ] ( 1936,-2761)(0,0)[lb ] ( 1936,-3031)(0,0)[lb ] ( 2341,-691)(0,0)[lb ] ( 1891,-691)(0,0)[lb ] ( 1891,-421)(0,0)[lb ] ( 901,-691)(0,0)[lb ] ( 901,-1411)(0,0)[lb ] ( 901,-2221)(0,0)[lb ] ( 901,-2986)(0,0)[lb ] ( 2791,-601)(0,0)[lb ] fig . 9 . some asynchronous primitives . in this section , we discuss the composition of asynchronous circuits , and some solutions which have been brought to the problem of computing interferences . compositions of asynchronous circuits correspond to distributed systems , where different parts communicate in a way which is not regularized by a global clock . then various and arbitrary time delays affect successive transitions at the input of a module . inputs may then conflict with the correct operation of the module itself . the approaches followed in circuit design to deal with computing interferences fall into two main classes . one practical approach to timing problems consists in working directly on the physical implementation , by keeping track of all the delays occuring in the logical circuit , together with all the constraints which must be satisfied by these delays in order to make the whole circuit operate correctly . then , programs are developed to find and optimize solutions in a systematic way @xcite . although practically very efficient , this strategy rapidly attains such a complexity that it becomes very difficult to distinguish fundamental issues from practical choices . in the other class of approaches , one attempts to separate as much as possible the logical issues related with timing from their physical manifestations , that is mainly from the values of time delays . this has led to different studies , focussing either on the determination of a best choice of logical primitives , satisfying criteria like speed - insensitivity or delay - insensitivity @xcite , or on a more restrictive definition of modular composition , like delay - insensitive compositions @xcite . in the following , we shall only briefly discuss approaches related with delay - insensitivity , and focus on the fundamental relation they tend to exhibit between the occurence of computing interferences and the causal structure of physical space - time . in order to make the analysis easier to follow , we shall introduce a graphic representation of the communications occuring between modules of a composition ( see figure 11 ) . these graphs are analogous to those that can be used in relativistic physics to represent the space - time evolution of localized physical systems , together with the light signals they exchange . as discussed in a previous section , an essential feature is the absence of an a priori given global and common time . only a local time ordering can be made between the successive events occuring on each module , reflecting the causal relations which can be made locally . although time is represented as the vertical axis , this only indicates the direction for increasing time on each module . different modules are displayed on the horizontal axis , which roughly corresponds to space . each module is then represented by a vertical line , indicating the causal succession of the local events occuring at its inputs or outputs . communications are then represented by inclined arrows leaving a module ( output ) to reach another module ( input ) . although they may vary , the slopes of theses arrows must always be greater than a strictly positive lower bound , which corresponds to light velocity . varying slopes indicate that varying speed and delays affect communications between modules . in the following , we shall denote by `` event '' each arrow corresponding to a communication , and shall call `` point '' the intersection of this event with the time evolution of a module ( following in that way the notation introduced by russell in his discussion of the causal structure of relativistic space - time @xcite ) . the logical specification of each module is translated into causal relations between the points which represent the occurence of events on the module . these local constraints may be given a precise expression using a formal language well suited to represent time ordered event structures @xcite . as propagation delays play an essential part , ordering constraints will be most conveniently visualized on graphic representations , which allow the analysis of global causal relations within distributed systems . * 5.1 delay - insensitivity * in order to discuss the role of delays in computation interferences , let us first analyse the illustrating example of the _ q - element _ @xcite , which is represented in figure 10 . the formal expression describing the logical function of the _ q - element _ can be written in a language which is derived from csp ( communicating sequential processes ) @xcite : @xmath38 ; y_o\uparrow ; [ y_i ] ; u\uparrow ; [ u ] ; y_o\downarrow ; [ \lnot y_i ] ; x_o\uparrow ; [ \lnot x_i ] ; u\downarrow ; [ \lnot u ] ; x_o\downarrow \right]\ ] ] each variable between brackets , which precedes a transition , represents a logical variable which must be true before the circuit can execute the transition which follows ( @xmath39 denotes time succession , and @xmath40 arbitrary repetition of the expresssion in brackets ) . thus , the circuit waits for @xmath4 to be true , then emits a rising edge on output @xmath41 , etc ... this logical function can be implemented as a composition of a _ c - element _ with two _ and _ gates , as represented in figure 10 . output @xmath42 of the _ c - element _ is followed by a fork , which relates @xmath42 to one input of each of the _ and _ gates . two other forks also dispatch the event produced by the environment @xmath4 ( resp . @xmath43 ) on two inputs denoted by @xmath31 and @xmath32 ( resp . @xmath44 and @xmath45 ) . ( 0,0 ) # 1#2#3#4#5 @font ( 2712,1872)(1216,-2773 ) ( 2251,-2581)(0,0)[lb ] ( 2476,-1996)(0,0)[lb ] ( 3826,-1186)(0,0)[lb ] ( 3016,-1096)(0,0)[lb ] ( 2251,-1591)(0,0)[lb ] ( 3061,-1546)(0,0)[lb ] ( 3736,-2581)(0,0)[lb ] ( 1216,-1096)(0,0)[lb ] ( 1216,-2491)(0,0)[lb ] ( 1621,-2626)(0,0)[lb ] ( 3421,-1321)(0,0)[lb ] ( 2026,-2266)(0,0)[lb ] ( 2926,-2221)(0,0)[lb ] fig . q - element . the logical operation of the circuit may be represented using a space - time graph , as in figure 11 . left and right parts of the circuit environment are respectively represented as @xmath46 and @xmath47 . the series of points corresponding to the definition of the logical function of each module can be followed on each vertical line . situations which correspond to rendez - vous , i.e. intervals where a primitive is waiting for the arrival of two events in any order , have been represented by a thick line . this is systematically the case for the _ c - element _ , but also for the _ and _ gates , when they are waiting for their two inputs to be true . pairs of points which can not occur in reverse order without ruining computation , have been signalized by dashed lines . the two cases involve the internal variable @xmath42 and one event , @xmath48 or @xmath49 produced by the environment ( @xmath47 or @xmath46 ) . event @xmath48 must reach the _ and _ gate b before event @xmath50 , recalling that the latter has been produced by the arrival of event @xmath51 on the _ c - element_. then , the fork which dispatches both events @xmath48 and @xmath51 plays a crucial role in determining the order of points on _ and _ gate b. a few remarks are in order . concurrent computing is well illustrated by figure 11 . different computations proceed along paths involving vertical and propagation lines , each representing a causally ordered series of operations . causal order makes only sense either within each vertical line , where it is associated with the logical function of the module , or within propagating lines , where it connects the output of one module to the input of another module . but no a priori total order exists between all points of the graph . this is illustrated by the independence of computation on the order of some pairs of points . for instance , two events belonging to different branches of the fork on variable @xmath42 at the output of the _ c - element _ may have arbitrary relative order . ( 0,0 ) # 1#2#3#4#5 @font ( 2757,5487)(1384,-6418 ) ( 2071,-5416)(0,0)[lb ] ( 1891,-6181)(0,0)[lb ] ( 2971,-6181)(0,0)[lb ] ( 4051,-6181)(0,0)[lb ] ( 1441,-1096)(0,0)[lb ] ( 2431,-6181)(0,0)[lb ] ( 3511,-6181)(0,0)[lb ] ( 2026,-5956)(0,0)[lb ] ( 3241,-5956)(0,0)[lb ] ( 3151,-5506)(0,0)[lb ] ( 3736,-5281)(0,0)[lb ] ( 4141,-5056)(0,0)[lb ] ( 2566,-4966)(0,0)[lb ] ( 3151,-4381)(0,0)[lb ] ( 2656,-4381)(0,0)[lb ] ( 3736,-4291)(0,0)[lb ] ( 2566,-3796)(0,0)[lb ] ( 1621,-3211)(0,0)[lb ] ( 3106,-2491)(0,0)[lb ] ( 2656,-2491)(0,0)[lb ] ( 2026,-2356)(0,0)[lb ] ( 1621,-2041)(0,0)[lb ] ( 1621,-5731)(0,0)[lb ] ( 2116,-2896)(0,0)[lb ] ( 3106,-3031)(0,0)[lb ] ( 2026,-3481)(0,0)[lb ] ( 3196,-3526)(0,0)[lb ] ( 4141,-3976)(0,0)[lb ] ( 3691,-4786)(0,0)[lb ] fig . 11 . space - time graph of the q - element . imposing a total ordering would amount to implement a global time , by means of clock distribution for instance , which would allow one to draw horizontal lines on the graph of figure 11 . but such condition is too restrictive , as computation only relies on causal relations imposed by vertical and propagation lines . the remaining freedom in the ordering of events , as the one related to the fork at the output of the _ c - element _ , is necessary for optimizing the circuit performance . for a definite implementation , event ordering will depend on the relative spatial localization of modules , so that the remaining freedom may be used to find an optimal arrangement of modules on the chip . the property of delay - insensitivity @xcite is easily seen on the graph . it corresponds to the independence of causal ordering of computation on delays occuring in responses of modules or in propagations of signals , i.e. on vertical or horizontal displacements of the modules . such property is made possible by using primitives which wait for the arrival of events at their input before producing other events at their output . but this condition appears to be unsufficient . in that respect , it is instructive to compare the two kinds of forks used by the previous composition implementing the _ q - element_. no constraint affects the events produced by the fork at the output of the _ c - element _ ( thick lines in figure 11 ) . however , forks dispatching the events produced by the environment must be implemented in such a way to respect the causal order of the events which they generate and which finally arrive at the same _ and _ gate ( dashed lines in figure 11 ) . such forks , which are called _ isochronic forks _ @xcite , must be isolated and given a special treatment at the implementation level , in order to satisfy the delay constraints which are necessary for preserving causal ordering . the property of delay - insensitivity ( di ) has been introduced and much developed as a simple condition one can impose on primitives and logical circuits , with the aim to design in a systematical way asynchronous circuits of arbitrary complexity , without having to take time scales into account . one approach consists in defining di circuits as compositions of stable primitives devoid of internal loops ( only electrical loops being used for memorization ) @xcite . a primitive is defined to be stable , by imposing that an input which changes the output can not change before the output has been established . it can then be shown that only compositions of _ c - elements _ can be di according to this definition . but , it can also be shown that compositions using _ c - elements _ ( and generalized _ c - elements _ with more inputs ) exclusively , strongly limit the type of allowed computations , excluding most circuits of interest @xcite . the _ isochronic fork _ may then be advocated as a weakest compromise to delay insensitivity . adding the _ isochronic fork _ and using this extended class of elements , called quasi delay - insensitive ( qdi ) , complex and efficient asynchronous circuits have been realized @xcite . however , as illustrated by the example of the _ q - element _ , _ isochronic forks _ need to be identified at the logical level and their implementation must be given a special treatment , which may reveal itself intricate for very complex circuits . * 5.2 delay insensitive composition * another approach for avoiding computation interferences @xcite , consists in defining a less restrictive set of di primitives , together with a notion of di composition of these primitives . circuits are represented in a formal language , called trace theory , similar to the one used in equation ( [ qelement ] ) , with further syntax rules on logical operations . computing interferences are avoided by imposing structural constraints under the form of simple rules . let us first recall definitions and some properties of trace structures @xcite . trace structures are defined as triples @xmath52 , where @xmath53 and @xmath54 are finite sets of symbols , respectively the input alphabet and the output alphabet , and @xmath55 is the set of traces , which is a subset of @xmath56 , the set of all finite - length sequences of symbols taken in the union set @xmath57 . trace structures are traditionally denoted by capital letters , while lower case letters @xmath58 , @xmath59 , @xmath60 denote symbols and @xmath61 , @xmath62 traces . the following short notations are also frequently used : & & a ? < a , , a > + & & b ! operations of concatenation , union , repetition , prefix - closure , projection and weaving are defined on trace structures : [ operations ] & & r;s + & & r|s + & & * [ r ] + & & * pref*r + & & ra < * i*r a , * o*r a , ta| t r > + & & r||s where , for convenience , notation @xmath63 has been introduced for the total alphabet of @xmath64 , where @xmath65 denotes the projection of trace @xmath62 on alphabet @xmath66 and @xmath67 is the set of all finite - length concatenations of traces in @xmath55 ( symbols @xmath68 , @xmath69 , @xmath70 , @xmath71 and @xmath72 denote as usual , existence , universality , set belonging , set intersection and set union ) . the @xmath73 operator constructs prefix - closed structures , while the projection operator hides internal symbols ; finally , the weave operator expresses instantaneous synchronization . a circuit is specified by a prefix - closed , non empty , trace structure @xmath64 with @xmath74 . the trace structure representing the environment of a circuit with trace structure @xmath64 is the reflection of the latter , and may also be given a compact notation : = < * o*r , * i*r , * t*r > a trace structure @xmath64 may be physically implemented by letting each symbol @xmath58 in the alphabets @xmath53 and @xmath54 correspond to a channel , and each occurence of this symbol in a trace of @xmath55 correspond to an event , i.e. a high or low transition , on the corresponding channel . symbols in @xmath53 or @xmath54 describe communication actions that are respectively produced by the environment or ( exclusive or ) by the circuit . in order to be able to ignore transmission delays while avoiding transmission and computing interferences , the following rules may be imposed @xcite . & & * r_0 * sr , ar s a a r + & & * r_1 * s , t r , ( a , br ) ( a , br ) + & & s a b t r s b a t r + & & * r_2 * s , t r , ( ar br ) ( ar br ) + & & ( s a b r s b a r ) ( s a b t r s b a t r ) rule @xmath75 excludes two consecutive transitions on the same wire , and hence transmission interferences that may result . rule @xmath76 expresses independence of computation on the order of signals travelling in the same direction , as this order may depend on suffered delays . the _ c - element _ is easily seen to satisfy this rule . however , the _ and _ gate only complies with the rule when it is waiting for a rising edge on its two inputs , and does not in all other cases . thus the _ and _ gate , and also the _ or _ gate are excluded by this rule , although the _ toggle _ , the _ merge _ and the _ sequencer _ are compatible . rule @xmath77 expresses the same property for signals travelling in opposite directions , in case their order does not change the result locally . note that due to the necessary symmetric treatment of a circuit and its environment , all rules are symmetric in the exchange of input and output symbols . one must exclude the possibility for a symbol of one type to disable a symbol of another type ( symbol @xmath58 is said to disable symbol @xmath59 in trace structure @xmath64 , if there is a trace @xmath61 with @xmath78 ) . such exclusion is necessary to prevent an admissible input symbol to get disabled by an output signal , depending on the delay the former has suffered on its way to the circuit ( by symmetry the same property must also hold for the environment and output signals ) . depending on the level of exclusion , this property leads to define three classes , with rules @xmath79 , @xmath80 and @xmath81 : & & * r_3^ * s , a= br + & & ( s a r s b r ) s a b r & & * r_3^ * s , a= br ar br + & & ( s a r s b r ) s a b r & & * r_3^ * s , ( ar br ) ( ar br ) + & & ( s a r s b r ) s a b r these rules successively allow for more decision possibility . rule @xmath79 does not permit data transmission and is called synchronization class ( an example is provided by the _ c - element _ ) . rule @xmath80 allows for two inputs to disable each other and is called data communication class . with rule @xmath81 , a circuit may choose between two output symbols and belongs to the arbitration class . finally , rule @xmath77 appears on specific examples to be too restrictive @xcite . an alternative and more generally efficient rule is provided by : & & * r_2^ * s , t r , ( a , cr br ) ( a , cr br ) + & & ( s a b t cr s b a tr ) s b a t c r this rule , which is conveniently expressed on a space - time graph , as shown in figure 12 , concerns three events @xmath58 , @xmath59 , @xmath60 connecting one module m and its environment @xmath82 . it stipulates that , if two time orders are allowed for the occurences of two events of different types ( i.e. one input and one output ) @xmath58 and @xmath59 , then if the event @xmath60 , of the same type as @xmath58 , is a consequence of the order `` @xmath58 then @xmath59 '' , it should also be a consequence of the other order `` @xmath59 then @xmath58 '' . this rule imposes that if an order on events is differently seen by a module and its environment , due to propagation time delays , then this order should have no consequence on the logical behavior of the module . as illustrated by figure 12 , this rule only affects the case on the left part of the figure , that is , only the case when propagation can change the order of events . ( 0,0 ) # 1#2#3#4#5 @font ( 2104,1932)(1159,-2368 ) ( 1216,-601)(0,0)[lb ] ( 1576,-2311)(0,0)[lb ] ( 2071,-2311)(0,0)[lb ] ( 2701,-2311)(0,0)[lb ] ( 3196,-2311)(0,0)[lb ] ( 1756,-1141)(0,0)[lb ] ( 2881,-1051)(0,0)[lb ] ( 1756,-1951)(0,0)[lb ] ( 1756,-1501)(0,0)[lb ] ( 2881,-1906)(0,0)[lb ] ( 2881,-1546)(0,0)[lb ] fig . space - time graph for rule @xmath83 . the set of di components is given by trace structures , defined according to definitions 1 and 2 , which satisfy the weakest form of the rules , i.e. @xmath84 , @xmath85 , @xmath86 and @xmath87 @xcite . a set of di primitive components for asynchronous circuits can thus be obtained with the following list of specifications in terms of trace structures ( see table 1 ) . .di primitive components [ cols="^,^",options="header " , ] [ table1 ] the _ wire _ corresponds to a component which waits for an event to occur on its input , then sends an event on its output , and repeats this sequence indefinitely . the inverted wire ( _ iwire _ ) behaves similarly , but begins by sending an event on its output . the _ fork _ duplicates one input . as can be seen from definitions ( [ operations ] ) , weaving not only consists in putting in parallel , but also in synchronizing common output symbols . in the particular case of two _ wires _ with a common output , weaving leads to the _ c - element_. the other components correspond to the primitive circuits which have been previously introduced ( see figure 9 ) . the objective is to realize circuits corresponding to given complex specifications by combining simple di primitive circuits . this aim may be attained by making use of operations such as decomposition and substitution , together with two theorems setting the conditions for performing these operations . a component @xmath84 is said to be decomposed into components @xmath88 ( for @xmath89 if the following conditions are satisfied . letting @xmath90 [ decomposition ] & & ( i ) _ 0i < n ( * o*s_i ) = _ 0i < n ( * i*s_i ) + & & ( ii ) * o*s_i s_j = , 0i , j < n i= j + & & ( iii ) t , x , i ( 0i < n ) + & & tt xs_i txs_i s_i txt + & & ( iv)*t*ts_0 = * t*s_0 conditions in ( [ decomposition ] ) respectively describe a closed network ( each input is connected to an output and conversely ( i ) ) , absence of output interferences ( two outputs can not be connected ( ii ) ) , absence of computing interferences ( any event produced by a component is compatible with the behavior of the component which receives it ( iii ) ) and correct behavior at circuit boundary ( network behaves as prescribed ( iv ) ) . decomposition will be denoted by @xmath91 . let us now state two useful theorems ( proofs may be obtained in @xcite ) . for components @xmath84 , @xmath85 , @xmath92 , @xmath93 and @xmath94 & & r_0 ( r_1 , s ) s ( r_2 , r_3 ) + & & r_0 ( r_1 , r_2 , r_3)holds if ( * a*r_0 r_1 ) ( * a*r_2 r_3 ) = * a*s the latter condition stipulates that internal symbols of @xmath94 , i.e. symbols in @xmath95 , where @xmath96 means set deletion , should not appear in @xmath97 . it can be realized by an appropriate renaming of internal symbols of @xmath94 . [ theorem]separation theorem for components @xmath98 and @xmath99 ( @xmath100 ) & & r_0 ( r_i)_1i < n s_0 ( s_i)_1i < n + & & r_0||s_0 ( r_i||s_i)_1i < n holds if & & ( _ 1i < n(*a*r_i ) \*a*r_0 ) ( _ 1i < n(*a*s_i ) \*a*s_0 ) = [ sep1 ] + & & , 1i= j < n + & & ( * o*r_i s_i ) ( * o*r_j s_j ) = + & & ( * o*r_i s_i ) ( * o*|r_0 ) = [ sep2 ] condition ( [ sep1 ] ) stipulates that the internal symbols of the decompositions of @xmath84 and @xmath101 are disjoint ( this condition may be satisfied by renaming some of these symbols ) , and conditions ( [ sep2 ] ) stipulate that the outputs of any two components @xmath102 and @xmath103 are also disjoint when the components are different ( these conditions may also be satisfied by reordering the components ) . with the help of these two theorems , the previously defined di primitives may be combined to give modular compositions which are delay insensitive , hence circuits where computing interferences can not be introduced by delay modifications only . we briefly describe an example of a circuit which can be obtained with such a composition of di primitives , the _ token - ring _ interface @xcite . the _ token - ring _ interface is a device allowing to connect several machines , which must share a common resource ( like a memory , or a bus ) . one item @xmath104 of this device will be associated with each machine @xmath105 , all items being identical and realizing the same function , as shown by figure 13 . item @xmath104 of this device is connected to two environments , the machine @xmath105 on top of the figure , and , at bottom , the ring @xmath64 where a token circulates . the arrival of the token at @xmath104 corresponds to an event on @xmath59 , its departure to an event on @xmath106 . the machine @xmath105 can make a request under the form of an event on @xmath107 . @xmath104 grants the resource to the machine @xmath105 by an event on @xmath108 . the machine @xmath105 signals the end of its use of the resource by an event on @xmath109 , which is acknowledged by @xmath104 under the form of an event on @xmath110 . initially , the _ token - ring _ interface is specified by the following trace structure : & & * [ a1?;p1!;a0?;p0 ! ] + & & || * pref**[b?;(q!|p1!;a0?;q ! ) ] this specification results from weaving two trace structures which respectively describe the communications of the token - ring interface with the machine @xmath105 and with the ring @xmath64 . the two trace stuctures interact through their common dependence on two events @xmath111 and @xmath112 . each trace structure may be decomposed into primitive elements . substitution and separation theorems may be applied , finally leading to a possible decomposition , as shown by figure 13 in a graphic way : ( & & * [ a1?;p1 ! ] + & & || * pref**[rq1?;q1 ! ] + & & || * pref**[b?;(q1!|p1!)],@xmath113,\ ] ] @xmath114,\ ] ] @xmath115,\ ] ] * [ ( q1?|q0?);q ! ) ] ) [ token - dec ] the first component is a _ sequencer _ ( see table 1 ) , necessary for synchronizing the output @xmath111 shared by the two trace structures defining the _ token - ring_. the _ sequencer _ also arbitrates between corresponding inputs . other components describe an _ iwire _ , two _ wires _ and a _ merge_. although they do not appear explicitly in decomposition ( [ token - dec ] ) , two forks appear in figure 13 , as a consequence of double occurences of @xmath116 and @xmath117 in ( [ token - dec ] ) . ( 0,0 ) # 1#2#3#4#5 @font ( 2814,2953)(1564,-4169 ) ( 3646,-2536)(0,0)[lb ] ( 4186,-1726)(0,0)[lb ] ( 3241,-2311)(0,0)[lb ] ( 2341,-2311)(0,0)[lb ] ( 3151,-2536)(0,0)[lb ] ( 3556,-1726)(0,0)[lb ] ( 2161,-2356)(0,0)[lb ] ( 3691,-2986)(0,0)[lb ] ( 2341,-2896)(0,0)[lb ] ( 2521,-3886)(0,0)[lb ] ( 3961,-3886)(0,0)[lb ] ( 2116,-1726)(0,0)[lb ] ( 1621,-3481)(0,0)[lb ] ( 2116,-2581)(0,0)[lb ] ( 2611,-2761)(0,0)[lb ] ( 1576,-1726)(0,0)[lb ] ( 2611,-3301)(0,0)[lb ] ( 2926,-4111)(0,0)[lb ] ( 2386,-1411)(0,0)[lb ] fig . 13 . token - ring interface . the di property of this implementation can be visualized on a space - time graph , as in figure 14 . two cases have been represented in figure 14 . in the first case , the request @xmath107 done by the machine @xmath118 is not granted , the token being sent back to the ring . when the token arrives a second time , the resource is granted to the machine @xmath105 which was waiting . this illustrates the undeterministic behavior of the module @xmath119 , which depends on arbitration performed by _ sequencer _ b. the figure also shows that the two forks , that on @xmath120 ( output of b ) and that defined by d can not create computation interferences , so that no particular constraints are necessary . this results from the function of _ sequencer _ b , which is not perturbed whatever the order of the events on its inputs . _ sequencer _ b waits for an event on @xmath59 to make a decision , and then arbitrates between the different requests it has received . as shown by the example of the _ token - ring _ interface , di primitives and di decomposition may be used to generate modular compositions which are delay insensitive , and , as shown with the help of space - time graphs , that remain free of computing interferences . delay - insensitivity appears as a simple criterion for escaping problems raised by computing interferences in a purely logically way , i.e. without recourse to a detailed analysis of the physical implementation of a circuit . the di criterion allows one to treat asynchronous circuits efficiently , like in the case of synchronous circuits , by allowing to represent them formally ( in terms of trace structures ) . although revealing a genuinely different underlying structure , the causal constraints on asynchronous circuits , as exhibited by space - time graphs , can nonetheless be embedded in a simple set of formal rules which limit the definition and composition of di circuits . in general , these rules allow di circuits to be decomposed into a number of di primitive components which increases linearly with the length of the circuit specification @xcite . ( 0,0 ) # 1#2#3#4#5 @font ( 3082,5532)(1306,-6418 ) ( 2071,-4606)(0,0)[lb ] ( 3421,-2086)(0,0)[lb ] ( 2296,-2401)(0,0)[lb ] ( 4051,-2761)(0,0)[lb ] ( 2296,-3436)(0,0)[lb ] ( 1756,-4381)(0,0)[lb ] ( 3421,-5281)(0,0)[lb ] ( 1756,-5281)(0,0)[lb ] ( 1621,-6181)(0,0)[lb ] ( 2161,-6181)(0,0)[lb ] ( 3241,-6181)(0,0)[lb ] ( 3781,-6181)(0,0)[lb ] ( 4321,-6181)(0,0)[lb ] ( 2656,-6181)(0,0)[lb ] ( 1486,-1051)(0,0)[lb ] ( 2926,-2761)(0,0)[lb ] ( 3466,-2806)(0,0)[lb ] ( 2971,-3391)(0,0)[lb ] ( 3421,-3931)(0,0)[lb ] ( 4006,-4561)(0,0)[lb ] ( 1891,-4741)(0,0)[lb ] ( 2341,-5731)(0,0)[lb ] ( 3376,-4561)(0,0)[lb ] ( 1606,-3706)(0,0)[lb ] fig . 14 . space - time graph of the token - ring interface . without giving definite answers to the problems raised in the introduction , we have nethertheless tried to provide some hints on the essential role played by physical time in computation . the necessary reference to physical time in physical implementations of logical circuits forces one to give an explicit treatment of computation interferences . these arise as obstructions when trying to make the causality underlying logical circuits coincide with the physical causality of their implementations . for synchronous circuits , these may be avoided by ruling the whole circuit with a single clock , which thus provides a global reference to a newtonian time . in general however , circuits must be considered as asynchronous and physical space - time as relativistic . in the latter , not all points are causally related , but only those such that one point falls within the light cone issued from the other . in that respect , asynchronous circuits and relativistic space - time share the same founding point of view . points derive from events and not the converse , propagating events being treated as primary entities and not as successions of points . the distinction between two classes of points can also be seen in a simple way : two points are causally related if and only if there exists a path between them using vertical or propagation lines ( in different spatial directions , but in the same time direction ) ; on another hand , points defined on two different events originating from the same point are not causally related @xcite . similarly for a concurrent computation , each computing path connects points which are causally related . avoiding computing interferences corresponds to impose that different computing paths respect a same time ordering , but only for pairs of causally related points . remedies to computing interferences in asynchronous circuits consist in recognizing paths which may conflict with a module specification , and in eventually delaying these paths , so that to respect a prescribed time ordering . this can be done either physically , at the implementation level by introducing explicit time delays , or at the logical level , by imposing specification rules which prevent the occurence of such conflicts . the latter solution , by imposing delay insensitivity both on circuits specification and decomposition in a consistent way , has the advantage of providing a purely logical characterization of the causal constraints . di circuits then build a class which may be seen as intermediate between synchronous circuits and general asynchronous circuits . they share with the former the possibility to be completely characterized by formal expressions and rules . but they rely on the same causal structure as the latter . synchronous circuits rely on time simultaneity classes , and thus on a causal structure which is typical of newtonian space - time . asynchronous circuits , on another hand , rely on a consistent treatment of propagation delays and time ordering , hence on a causal structure which characterizes relativistic space - time . delay - insensitivity provides an interesting transition between local properties , like those defining sequential processors , and global ones , like those exhibited by distributed systems . but di circuits hardly exhaust the computation potentialities brought by the introduction of asynchronous circuits . the critical consequences of delay sensitivity rather suggest to consider a further alternative when attempting to classify the different types of computations , i.e. those performed by synchronous , by di asynchronous and by ds ( delay - sensitive ) asynchronous circuits . similarly , in the same way as asynchronous computing machines may not always allow simulation by synchronous computing machines , one may infer that physical processes and physical laws , which intrinsically obey relativistic causality , may be simulated by synchronous machines in particular cases only . this hints at another advantage of computations based on asynchronous circuits , i.e. the ability to simulate in a universal way real physical processes . chakraborty , s. , yun , k.y . and dill , d.l . ( 1988 ) practical timing analysis of asynchronous circuits using time separation of events . _ proceedings of the ieee custom integrated circuits conference _ , 455458 . ieee computer society press ( may ) . davis , a. and nowick , s.m . ( 1995 ) asynchronous circuit design : motivation , background and methods . in g. birtwistle and a. davis ( editors ) , _ asynchronous digital circuit design , workshop in computing _ , 149 . bcs - springer . furber , s. ( 1995 ) computing without clocks : micropipelining the arm processor . in g. birtwistle and a. davis ( editors ) , _ asynchronous digital circuit design , workshop in computing _ , 211262 . bcs - springer . jaekel , m.t . and reynaud , s. ( 1998 ) quantum observables associated with einstein localisation . in e. zavattini , d. bakalov and c. rizzo ( editors ) , _ frontier tests of qed and physics of the vacuum ( sandansky , 1998 ) _ , 389404 . heron press . mallon , w.c . , udding , j.t . and verhoeff , t. ( 1999 ) analysis and applications of the xdi model . in async99 , proc . international symposium on advanced research in asynchronous circuits and systems _ , 231242 . ieee computer society press . martin , a.j . ( 1990b ) the limitations to delay - insensitivity in asynchronous circuits . in w. j. dally ( editor ) _ advanced research in vlsi : proceedings of the sixth mit conference _ , 263278 . mit press , cambridge , ma . martin , a.j . , lines , a. , manohar , r. , nystrm , m. , penzes , p. , southworth , r. , cummings , u. and tak - kwan - lee ( 1997 ) the design of an asynchronous mips r3000 microprocessor . in _ proceedings of the seventeenth conference on advanced research in vlsi _ , 164181 . ieee computer society press . matherat , p. and jaekel , m.t . ( 1996 ) dissipation logique des implmentations dautomates - dissipation du calcul . _ technique et science informatiques _ * 15*(8 ) , 10791104 . english translation at the url http://www.arxiv.org/abs/quant-ph/9805018 . mazurkiewicz , a. ( 1989 ) basic notions of trace theory . in j. w. de bakker , w. p. de roever , and g. rozenberg ( editors ) _ linear time , branching time , and partial order in logics and models for concurrency . lncs 354 _ , 285 . springer - verlag . school / workshop , noordwijkerhout , the nederlands , may - june 1988 . stevens , k. , ginosar , r. and rotem , s. ( 1999 ) relative timing . in _ async99 , proc . international symposium on advanced research in asynchronous circuits and systems _ , 208218 . ieee computer society press . winskel , g. ( 1989 ) an introduction to event structures . in j. w. de bakker , w. p. de roever , and g. rozenberg ( editors ) _ linear time , branching time , and partial order in logics and models for concurrency . lncs 354 _ , 364397 . springer - verlag . school / workshop , noordwijkerhout , the nederlands , may - june 1988 .

cataclysmic variables ( cvs ) are close binary systems consisting of a low - mass main - sequence ( ms ) star that fills its roche lobe and a white dwarf . main - sequence star ( mass - donating star , usually referred to as star 2 or secondary ) loses matter through the vicinity of the inner lagrangian point @xmath0 . white dwarf ( accretor star , referred to as star 1 or primary ) accretes at least a fraction of this matter via accretion disk or via accretion columns in the polar zones in the case when white dwarf has a strong magnetic field . physics and evolution of cvs as well as evolution of similar to them low - mass x - ray binaries were investigated starting from the late sixties ( see , e.g. , [ [ rjw82][fed97 ] ] and references therein ) . these studies were motivated by the recognition of the fact that evolution of cvs is determined by losses of orbital angular momentum due to radiation of gravitational waves [ [ kmg62],[p67 ] ] and magnetically coupled stellar wind [ [ sch62][vz81 ] ] . a number of studies investigated the influence of the loss of angular momentum associated with the material outflow from the system on the cv evolution ( see , e.g. , [ [ yu73][kk95 ] ] ) . however , in the absence of gas dynamical simulations of the mass transfer in the binaries , these processes were considered under the parametric approach . recent three - dimensional ( 3d ) gas dynamical simulations of the structure of gaseous flows in semi - detached binaries proved an important role of a circumbinary envelope ( see , e.g. [ [ bbkc98b],[bbkc98c ] ] ) . these calculations also show that during mass exchange a significant fraction of matter leaves the system . the present work is mainly devoted to the numerical investigation of the evolution of cvs using the data of 3d gas dynamical calculations on the losses of mass and angular momentum from close binaries . the main attention is paid to the stability of mass exchange against runaway mass loss and its dependence on the donor to accretor mass ratio @xmath1 . it is known that the loss of matter by the donor star results in the violation of its hydrostatic and thermal equilibrium . hydrostatic equilibrium is restored adiabatically , i.e. in dynamical time scale , while thermal equilibrium is restored in thermal time scale . this transition to a new equilibrium state leads to a change of donor radius . the sign of radius variation depends on the convective and radiative stability of outer envelope of the star . for stars with masses @xmath2 with deep convective envelope as well as for white dwarfs mass loss results in increase of the radius , while for stars with radiative envelopes it leads to the shrinkage of the star . mass exchange in close binary is unstable when in the course of evolution mass - losing star tends to overfill the roche lobe . it can occur when radius of the donor @xmath3 increases faster ( or decreases slower ) than the effective radius of roche lobe @xmath4 . is determined as the radius of a sphere with a volume equal to the volume of the roche lobe . ] the case when the radius of the donor increases while the radius of the roche lobe decreases is also possible . thus , the question of stability of mass exchange against runaway is determined by the balance of derivatives @xmath5 and @xmath6 . it means that unstable mass exchange is possible even for stars which contract when losing matter . the study of conditions of stable mass exchange is motivated by two related problems : i ) it is necessary to explain why @xmath7 10% of cvs with well - determined masses of components have combination of donor mass and mass ratio of components which is `` forbidden '' in the evolutionary models based on conventional assumptions on the conservation of angular momentum of a binary ; ii ) for population synthesis studies of cvs it is necessary to distinguish progenitors of cvs among all binaries containing a white dwarf and a low - mass companion ( i.e. , specify in which binaries of this type stable mass exchange is possible ) . a problem similar to the last one appears in the studies of low - mass x - ray binaries . the paper is organized as follows . in section 2 we describe major factors affecting the evolution of cvs and conditions of stable mass exchange in binaries . the reasons which motivated us to rule out the conservative approximation for mass exchange are considered in section 3 . sections 4 and 5 describe some results of 3d gas dynamical simulations of the flow structure in semi - detached binaries and introduce the model for description of angular momentum losses due to non - conservative mass exchange . results of evolutionary calculations for conservative and `` non - conservative '' models are compared in section 6 . criteria of stable mass exchange for `` non - conservative '' model with mass and angular momentum loss are determined in section 7 . section 8 summarizes the main results of the work . in the context of the present work the term `` stable mass exchange '' means dynamical stability . in other words , in the course of mass exchange the donor star can be out of thermal equilibrium and mass transfer rate may exceed the rate corresponding to the kelvin time scale ( @xmath8 , where radius @xmath9 , luminosity @xmath10 , and mass of the star @xmath11 are in solar units and mass transfer rate @xmath12 is in @xmath13 ) . let us define unstable mass exchange as the situation when the radius of the donor changes faster than the effective radius of roche lobe : @xmath14 . using the derivatives of the radii w.r.t . mass of the donor we can rewrite the condition of the stability of mass transfer as : @xmath15 the effective radius of roche lobe can be estimated , for instance , using interpolation formula given by eggleton [ [ eggl83 ] ] : @xmath16 where @xmath17 is the semimajor axis of the orbit . for @xmath18 the approximation suggested by paczyski [ [ pacyn71 ] ] is more convenient : @xmath19 derivative of the roche lobe radius can be written as @xmath20 it is seen from ( [ rrp ] ) and ( [ drra ] ) that the variation of effective radius of roche lobe is defined by the change of donor and accretor masses @xmath21 and @xmath22 as well as by the change in the separation @xmath17 , and , hence , it is influenced by the loss of mass and angular momentum from the system . as a rule , only systemic angular momentum is considered as the orbital momentum , i.e. the sum of angular momenta of two stars , which are considered as point masses . momenta of the spin of the components , momentum of the accretion disc ( if it s present ) , as well as momentum of gaseous flows inside the system as a rule are not included into the total momentum . deviation of the momentum of the donor from that of the point mass object ( which can be significant , see [ [ kuz95 ] ] ) usually is not taken into account as well . such a simplified approach is compelled , since it is very difficult to include all the abovementioned factors into calculations . having in mind all these simplifications , one gets for the orbital momentum of a binary system with the circular orbit @xmath23 where @xmath24 is the gravitational constant . according to the widely accepted models , the evolution of cataclysmic binaries is determined by the loss of angular momentum from the system via gravitational waves radiation ( gwr ) and/or magnetic stellar wind ( msw ) of the donor as well as the mass exchange between components . in the standard models of evolution it s accepted that the mass exchange among components does not change the systemic angular momentum , and influences it indirectly through a possible loss of mass from the system and angular momentum ablation by leaving matter ( term with subscript loss in eq . [ dj ] below ) . therefore , within conventional models , the equation for the change of orbital momentum can be written as : @xmath25 let us consider the terms of expression ( [ dj ] ) : + _ 1 ) loss of the systemic angular momentum due to gwr _ + the change of the systemic orbital momentum as a result of gwr is given by a formula ( see , e.g. , [ [ ll62 ] ] ) : @xmath26 where @xmath27 is the speed of light . intensity of gwr is characterized by a very strong dependence on the separation of components and respectively on the orbital period . this process is essential for short - period systems with @xmath28 , since only then the time scale of angular momentum loss becomes shorter than the hubble time . + _ 2 ) loss of the systemic angular momentum due to msw of the donor _ + a mechanism for the loss of the angular momentum of the system by means of the msw was suggested by schatzman [ [ sch62 ] ] and by mestel [ [ mes68 ] ] . if the star has a convective envelope and , subsequently , a surface magnetic field , its own axial rotation is braked by the magnetic stellar wind , and the angular momentum loss rate can be essential even for small mass loss rate . the subsequent synchronization of the spin of the donor and orbital rotation caused by tidal interaction between components results in the loss of the orbital momentum of the system and reduction of the separation of components @xmath17 . to take this effect into account , verbunt and zwaan [ [ vz81 ] ] extrapolated the data on rotation of f - type stars [ [ skum72 ] ] to the k and m type components of cvs and suggested a widely used semi - empirical formula which allows to determine the temporal behavior of the orbital momentum : @xmath29 were @xmath30 is the radius of gyration of the donor star ; @xmath31 is a numerical factor determined from the comparison of theoretical calculations and observational data . well - defined reduction of magnetic field for stars with masses less than @xmath32 permits to suggest that vanishing of a radiative core of a star when its mass decreases to this threshold leads to an abrupt `` switch - off '' of the so - called `` @xmath33''-dynamo mechanism that is responsible for the generation of stellar magnetic fields [ [ sr83 ] ] . at the instant of cessation ( or sharp reduction ) of msw the time scale of angular momentum loss [ defined by eq . ( [ djmsw ] ) ] is shorter than the thermal time scale of a donor with @xmath34 . thus , the latter is out of thermal equilibrium and its radius is larger than the radius of thermally equilibrium star of the same mass.,[pol96 ] ] . ] when the action of msw and corresponding loss of angular momentum stop , the rate of reduction of the semimajor axis declines , the donor shrinks down to the equilibrium radius and loses the contact with roche lobe [ [ sr83 ] ] . meanwhile , the system continues to lose angular momentum due to gwr , components continue to approach each other and after some time the optical star fills its roche lobe again . further evolution of the system is determined by the loss of angular momentum via gwr . cessation of msw after the star becomes completely convective is widely adopted hypothesis explaining satisfactorily the so - called orbital period gap in cvs . the value of coefficient @xmath31 in eq . ( [ djmsw ] ) can be determined by comparison of theoretical width of period gap and observational one . in the present study we adopted @xmath35 in accordance to [ [ ft94 ] ] . + _ 3 ) loss of the systemic angular momentum due to mass loss _ + mass loss from the system is mainly considered as a parameter fine tuning of which permits to get an agreement with observational minimal period of cvs ( e.g. [ [ rjw82 ] ] ) . more definite assumptions are made only for specific cases when examining evolution of the systems , in which : i ) the rate of accretion on white dwarf is limited by the hydrogen burning rate ( @xmath36@xmath13 ) , and the mass loss rate by the donor is much higher than this limit , but does not exceed eddington limit for the dwarf ( that is close to @xmath37@xmath13 , see , e.g. , [ [ hkn96],[yl98 ] ] ) , the mass excess being lost by means of stellar wind [ [ kh94 ] ] ; ii ) mass is lost due to outbursts occurring on white dwarf ( see , e.g. , [ [ yu73 ] , [ yuetal96][skr98 ] ] ) . in both cases it is usually assumed that the specific angular momentum of the matter leaving the system is equal to the specific momentum of the accretor . to describe the loss of mass and momentum out of the system , it is convenient to introduce parameters describing the degree of non - conservativity of the evolution respective to the mass [ [ ty71 ] ] : @xmath38 and to the angular momentum : sometimes specific angular momentum lost from the system is measured in units of @xmath39 and used instead of @xmath40 : @xmath41 ] @xmath42 where @xmath43 . with these parameters , the expression for the loss of momentum by the matter leaving the system gets the following form : @xmath44 in the calculations of the cataclysmic binary evolution an equation for the variation of the semiaxis of the orbit is used directly , instead of the equation for the variation of momentum over time . let us differentiate the expression ( [ j ] ) , substitute the result into equation ( [ dj ] ) , and use the relation @xmath45 . then the equation for the variation of orbital semiaxis can be written as : @xmath46 here , index ` exch ' corresponds to the change of @xmath17 resulting from the mass transfer between components . note that while the term @xmath47 is absent in the equation ( [ dj ] ) , the appropriate term @xmath48 is present in the equation ( [ da ] ) , being a natural consequence of the dependence of the orbital momentum both on semimajor axis of the orbit and masses of components . when mass exchange is conservative , the evolution of the orbital separation is determined by the assumption that the mass transfer does not change the orbital momentum of the system . in the course of mass exchange the matter that initially had specific angular momentum of the donor is transferred onto accretor and finally gets the specific momentum of the accretor . the assumption of conservation of the orbital momentum implies that if the mass of the donor is lower than the mass of accretor , the excess momentum of accreted gas should transform into the orbital one , and the mass exchange acts in the direction of increasing the orbital semiaxis . if the donor is more massive than the accretor , the lacking momentum of the accreted gas should be taken from the orbital momentum , therefore , the mass exchange acts in the direction of shrinking of the binary system . the numerical investigation of the evolution of cataclysmic binaries consists of simultaneous calculations of the evolution of the donor and of the variation of the orbital separation in time . let us consider the processes that determine the evolution of the orbital semiaxis . using parameters @xmath49 and @xmath40 , we can write formulae for the terms of equation ( [ da ] ) ( here masses and distances are given in solar units and time is given in years ) : @xmath50 @xmath51 @xmath52 @xmath53 it is seen from eqs ( [ daex])([damsw ] ) that changing of @xmath4 is mainly determined by the value of @xmath54 . on the other hand , changing of @xmath3 depends on @xmath21 and @xmath55 . thus the most convenient way for studying of limits of stable mass exchange is analysis of `` ratio of donor - star mass to accretor mass @xmath54 donor - star mass @xmath21 '' diagram . in a number of cases one may simplify expression ( [ drra ] ) and obtain analytical formulae for the derivative of the roche lobe radius . in particular , when the angular momentum changes due to gwr and msw can be neglected , one may deduce from ( [ daex ] ) and ( [ daloss ] ) the expression for the change of separation @xmath17 [ the first term in the formula ( [ drra ] ) ] as a function of mass ratio @xmath54 and the parameters of the loss of matter and momentum @xmath49 and @xmath40 : @xmath56 using the expression ( [ rrp ] ) for @xmath4 we can express the second term of ( [ drra ] ) as : @xmath57 using ( [ dabpsi ] ) and ( [ dabpsi1 ] ) , for the case of totally conservative mass exchange ( @xmath58 ) , we obtain : @xmath59 from the same equations , for the so - called ` jeans mode ' of mass loss , when the mass exchange in the system does not occur , but the donor loses its mass due to stellar wind and matter leaves the system carrying away specific angular momentum of the donor ( @xmath60 , @xmath61 ) , we obtain : @xmath62 for the case of stellar wind from the accretor when all matter lost by the donor is transferred onto accretor but then leaves the system carrying away specific angular moment of the accretor ( formally , one should take here @xmath60 and @xmath63 ) we obtain : @xmath64 the comparison of the rates of variation of the radius of the donor and effective radius of the roche lobe ( [ drra ] ) determines the state of stability / instability of mass exchange in any phase of evolution . it is possible to determine the boundaries of stable mass exchange region by two different methods : * it is possible to calculate derivatives of the radius of the donor w.r.t . its mass directly for different @xmath55 and to compare it with derivatives of the effective radius of roche lobe depending on @xmath22 , @xmath21 , and @xmath17 . this method was used , for instance , in [ [ tfy82 ] ] and it permits to restrict a region of @xmath65 diagram where mass exchange is dynamically stable and also to detect the regions where mass exchange should occur in different time scales : thermal , nuclear , or in the time scale of angular momentum loss . * alternatively , a requirement that the rate of accretion should not exceed the eddington limit for the dwarf can be accepted as a criterion of stable mass exchange . this assumption seems to be more reasonable since evolutionary calculations often suggest very high rate of mass loss during very initial moments of time after overfilling the roche lobe , while after that @xmath12 falls off rapidly . in fact , this approach involves a restriction on dynamical stability of mass transfer but permits loss of matter in thermal time scale . figure [ ris1 ] depicts the boundary of the stable mass exchange region in conventional conservative model of the evolution of cv . boundary 1 was found in [ [ tfy82 ] ] , and boundary 2 was found in [ [ kool92 ] ] . these boundaries coincide well for of donors of low mass but differ a little for large ones . unfortunately , in [ [ kool92 ] ] the procedure of the calculation of the boundary is nt described in detail . we may assume that this discrepancy is due to the differences in the codes , in a different treatment of possible weakening of donor msw at masses greater than @xmath66 , and different techniques for computation of stellar radii derivatives in this region . figure [ ris1 ] also shows cvs with known masses of components from the catalog [ [ cat97 ] ] . the total number of binaries in the diagram is 80 , and 11 of them are in the region of unstable mass exchange . for 8 of these 11 systems uncertainties of parameters are known , they are shown in fig . [ ris1 ] as error bars . it is seen that the location of stars in a region of unstable mass exchange can not be explained by uncertainties in masses of components . in fig . [ ris1]a we plot cvs with the best determined masses [ [ smith_dhillon ] ] , parameters for a number of systems being different compared to the data from [ [ cat97 ] ] . nevertheless , even for well - determined systems the problem of `` improper '' location of cvs remains 3 out of 22 systems lie in the `` forbidden '' region . to explain this disagreement we suggest a model of cvs evolution which involves loss of mass and angular momentum from the system . this model is based on 3d gas dynamical simulations of the flow structure in close binary systems . the above - stated analysis proves that the rate of change of the distance between components @xmath17 in evolutionary calculations depends heavily on two parameters : the degree of non - conservativity of mass exchange w.r.t . the matter @xmath49 and w.r.t . to the angular momentum @xmath40 . these parameters can be determined only by means of 3d gas dynamical simulations of mass transfer in binary systems . these investigations were made in [ [ bbkc98b ] ] for a binary system with mass ratio of components @xmath67 , and in [ [ bbkc98c ] ] for a binary with @xmath68 . besides that , we have specially made a numerical simulation of a binary system with @xmath69 to study the influence of parameter @xmath54 on the flow structure . analysis of the obtained results proves that in the binary systems considered by us for different values of parameter @xmath54 mass transfer is non - conservative and the non conservativity degree w.r.t . the matter is @xmath70 . simulations of flows in the 3d numerical models prove also that the matter leaves the system with a significant angular momentum . the non conservativity degree of mass exchange w.r.t . the matter can be easily determined directly from gas dynamical model , while the determination of non conservativity degree w.r.t . the angular momentum is a sophisticated problem . let us consider the equations which determines the transport of angular momentum in the system . we should consider navier - stokes equations for a correct analysis because viscosity plays a significant role in the redistribution of angular momentum . from the steady - state gas dynamical equations in a rotating frame we can deduce the equation for the angular momentum transfer @xmath71 where , as usually , @xmath72 is the velocity vector , @xmath73 is the density , @xmath74 is the roche potential , @xmath75 is the orbital angular velocity , @xmath76 is the stress tensor , @xmath77 is the radius - vector of the center of masses of the system , and angular momentum @xmath78 ( in a laboratory , i.e. inertial or observer s frame ) is defined by expression : @xmath79 let us write down the equation for angular momentum transfer in a divergent form : @xmath80 from this equation , we can obtain the integral law of change of angular momentum as follows : @xmath81 the term @xmath82 determines the change of the angular momentum due to non central nature of the force field , defined by the roche potential . by definition , the value @xmath83 is the specific flux of angular momentum @xmath78 . as a result , for the steady - state case we obtain : @xmath84 where @xmath85 and @xmath86 are the boundaries of the donor and the accretor , respectively , and @xmath87 is the outer boundary of the calculation domain . by analogy to the flux of matter @xmath88 one may estimate the flux of angular momentum from the system as @xmath89 and to use it in the expression ( [ eqpsi ] ) , determining the non conservativity parameter w.r.t . the angular momentum . note that numerical simulations were conducted for eulerian equations for in - viscid gas and not for navier - stokes equations . respectively , to calculate integral ( [ error ] ) we use isotropic term describing gas dynamical pressure in the expression for the specific flux of angular momentum : @xmath90 . is kronecker s symbol : @xmath91 ] this substitution is correct because on the outer border where the integral in ( [ error ] ) is estimated , viscosity does not play a significant role . using expression ( [ error ] ) in the gas dynamical calculations we obtained @xmath92 corresponding to @xmath93 ( see definition of @xmath94 in the footnote on page ) for the binary system with @xmath67 [ [ bbkc98b ] ] , and @xmath95 that corresponds to @xmath96 [ [ bbkc98c ] ] for a binary with @xmath97 the last value is in good agreement with the results of two - dimensional calculations of binary system with equal masses of components [ [ sawada84 ] ] where @xmath98 was obtained . however , the use of these estimates for @xmath40 in evolutionary calculations proves that this rate of an angular moment loss is so great that in the majority of binary systems the mass exchange rapidly becomes a runaway process and the rate of mass loss by the donor increases without any limit . apparently , the application of formula ( [ error ] ) to the estimation of the angular momentum loss in the evolutionary calculations is not quite correct . since conjunction of gas dynamical and evolutionary models is not completely consistent . the considered gas dynamical model does not take into account the change in the position of stars in time and , hence , the change in the angular momentum of the gas due to the non - central nature of the force field is not compensated by respective change in the orbital angular momentum . in a more general gas dynamical model , in which the change in the position of stars is taken into account , one may correctly estimate the change of the systemic angular momentum in the form of angular momentum flux through the external boundary for the `` donor + accretor + gas '' system . however , to take into account such gas dynamical calculations appears rather complicated . such estimates are possible only for a specific stage in the life of a binary system . moreover , thus obtained gas dynamical results can not be used directly in the standard evolutionary models that ignore the presence of circumbinary envelope in the system . therefore , a simplified model for estimation of the angular momentum loss developed on the basis of gas dynamical calculations will be a promising approach in the nearest future . gas dynamical simulations of mass transfer in semi - detached binaries confirm that outflow of matter from the donor occurs through a small vicinity of the inner lagrangian point @xmath0 . in this case specific angular momentum of lost matter ( in corotating frame ) can be estimated as @xmath99 , where @xmath100 . respectively , the flow of angular momentum from the donor is equal to @xmath101 due to non - conservativity of mass exchange , only a fraction @xmath49 of the matter outflowing through @xmath0 is accreted . specific angular momentum of accreted matter is equal to the angular momentum of accretor ( here we neglect the finite radius of accretor and/or residual angular momentum of accreted matter , and this approximation appears to be reasonable for cvs ) . it means that accreted matter has zero angular momentum in the rotating frame , and , hence , does not accelerate the rotation of accretor . thus the question on the efficiency of the angular momentum transfer from the spin rotation of accretor to the orbital rotation of the binary can be ignored . to summarize , we can describe the flow of accreted angular momentum as @xmath102 these expressions for the angular momentum flows do not contradict the general formula for specific flow of angular momentum [ see ( [ potok ] ) ] . indeed , viscosity does not play a significant role neither in the vicinity of the surface of the donor nor in the vicinity of the accretor where the matter has already lost its angular momentum and the flow has radial direction in the rotating frame . it means , that in these cases the stress tensor @xmath76 is reduced to the isotropic pressure as well , and does not deposit into the integral of specific flow of angular momentum . our assumptions permit to write the expression for the flow of angular momentum lost from the system in the following form : @xmath103 where @xmath104 is a parameter defining the fraction of the circumbinary envelope angular momentum that is lost with the matter leaving the system . then @xmath105 is the fraction of the circumbinary envelope angular momentum that returns back into the system via tidal interaction . henceforth we adopt for the parameter @xmath104 a value of 1 that corresponds to the case when the angular momentum of the circumbinary envelope is totally lost with the matter leaving the system . in this case the specific angular momentum of the matter leaving the system is equal to @xmath106 or in the units of systemic specific angular momentum ( @xmath107 ) : @xmath108 where @xmath109 is dimensionless distance from the center of the donor to @xmath0 . we would like to stress that formula ( [ xxxx ] ) was derived under the assumption of aligned synchronous rotation of components of the binary system . it is known that the loss of angular momentum due to outflow of the matter from the vicinity of @xmath0 can result in the development of a misalignment of the vector of spin rotation of the donor and the vector of the orbital rotation of the binary [ [ matese83 ] ] . the flow structure in the binaries with misaligned asynchronous rotation was studied by the authors in [ [ bbkc98d ] ] . having in mind the simplified character of evolutionary models we do nt consider the effect of the misalignment in the present work . the graph of @xmath110 is shown in fig . [ ris2 ] for various values of non - conservativity degree w.r.t . the matter @xmath49 . the values @xmath111 and @xmath112 are also shown in the figure by dashed lines . in the region @xmath113 each unit of the matter leaving the system carries away an angular moment that is higher than the mean specific angular momentum of the system . this results in diminishing of systemic specific angular momentum : @xmath114 in the region @xmath115 each unit of the matter leaving the system carries away an angular moment that is lower than the mean specific angular momentum of the system , and , hence , @xmath116 figure [ ris2 ] shows that the value of @xmath49 becomes negative in a certain range of @xmath54 for the majority of cases of non - conservative mass exchange ( except @xmath117 ) . this means that the escape of the matter to infinity causes redistribution of the angular momentum in the `` donor + accretor + gas '' system . this results in the growth of the total angular momentum of the binary : @xmath118 . it is evident that such a `` pumping '' of angular momentum into the system can stabilize the mass exchange in cv . indeed , consideration of orbit semiaxis change due to conservative mass exchange shows that the binary system widens for @xmath119 only . for the `` non - conservative '' model the range of @xmath54 corresponding to the widening of binary due to mass and angular momentum losses is larger for @xmath120 it happens for @xmath121 . note that @xmath122 dependence has a curios feature [ see fig . [ ris2 ] and formula ( [ xxxx ] ) ] all curves @xmath110 for different values of @xmath49 pass through the same point . this point corresponds to the case when the center of mass of the binary system is located exactly equidistantly between @xmath0 and accretor , and in this case @xmath40 does not depend on @xmath49 . fortunately , this specific feature does not influence the analysis of solution . three - dimensional gas dynamical simulation of mass transfer in cvs confirms conclusion that the outflow of matter from the system can lead both to decrease and to increase of the total angular momentum of a binary . the typical flow structures [ see fig . [ ris3]a , b ] for binaries with @xmath123 and @xmath124 demonstrate that in an inertial frame gas can have different directions of rotation ( in both figures orbital rotation corresponds to the counter - clockwise direction ) . respectively , the change of angular momentum has opposite signs in the obtained solutions for @xmath67 the systemic angular momentum decreased , and for @xmath69 it increased . it seems reasonable to explain the different types of gas dynamical solutions using the following qualitative speculation : * the flow of the gas leaving the system consists of the matter of the circumbinary envelope that has an angular momentum large enough to overcome the gravitational attraction of the accretor . initial velocity of this matter ( in rotational frame ) has the same direction as the orbital movement of the accretor . * velocity of the gas leaving the system changes under the action of the gravitational attraction of both components of the binary system , centrifugal force , coriolis force , and pressure gradient . the change of azimuthal velocity is determined mainly by the coriolis force and gravitation forces , since centrifugal force is axially - symmetrical and deviation of pressure gradient from the axial symmetry is small as well . action of coriolis force results in deflecting of the stream in the direction opposite to that of the orbital movement . consequently , this permits to subdivide the whole flow leaving the system through the vicinity of @xmath125 into two parts : the first part retains the direction of original movement , and the direction of movement of the second one is changed . * it is evident that in an axially - symmetrical gravitational field the action of coriolis force can not produce a solution when the direction of gas movement in the inertial frame is opposite to the orbital one . the situation is changed drastically for the case of non - axially - symmetrical gravitational field of a binary system . in this case gravitational attraction of the donor can accelerate the gas of the stream and result in the solution with opposite direction of movement in the inertial frame . * it follows from eq . ( [ xxxx ] ) and fig . [ ris3 ] that there is a range of values of @xmath54 where gas moves in the direction opposite to the orbital one in the inertial frame . these values of @xmath54 correspond to the cases when attraction of the donor is large enough to accelerate the gas properly ( left end of the range ) , but not too large to produce accretion of gas on the surface of the donor ( right end of the range ) . the values of @xmath54 out of this range correspond to the case when the flow directed along the orbital movement prevails . appearance of an additional range of values of @xmath54 where the `` non - conservative '' mass exchange is accompanied by increase of systemic angular momentum results in expansion of the stable mass exchange region . to illustrate this fact let us consider a graph that is often used in the qualitative studies of the stability of mass transfer in cv . this is a @xmath126 graph with curves corresponding to the analytical expressions for derivatives of effective radius of the roche lobe w.r.t . the mass of the donor @xmath127 , and straight lines corresponding to the values of derivatives of the radius of the star w.r.t . to its mass @xmath128 , the latter being known from the investigations of the internal structure of star ( see , e.g. , [ [ tfy82 ] ] ) . mass exchange is stable for the range of @xmath54 where curve @xmath127 passes below the straight line @xmath128 for the donor of corresponding type . this graph is presented in fig . it shows the derivative of effective radius of roche lobe @xmath129 as a function of @xmath54 for a scenario in which the loss of angular momentum is defined by formula ( [ xxxx ] ) for various values of the parameter @xmath49 . dependencies @xmath130 for the scenario of the completely conservative mass exchange and for the case when the outflowing matter carries away specific angular momentum of the donor or specific angular momentum of accretor are shown as well . the same fig . [ ris4 ] also shows the values of derivative @xmath131 for completely convective and degenerate stars ( @xmath132 , for thermally equilibrium stars of ms with @xmath133 ( @xmath134 , and for subgiants with degenerate low - mass helium cores ( @xmath135 . the analysis of fig . [ ris4 ] proves that the outflow of matter and loss of angular momentum from the system in accordance to law ( [ xxxx ] ) stabilize mass exchange in a wider range of @xmath54 compared to the conventional models of conservative mass exchange . it is also worth noting that for conservative scenario of mass exchange its time scale for stars with masses @xmath136 is determined by the time scale of the loss of angular momentum or the nuclear time scale for @xmath137 [ [ tfy82 ] ] , while for above - considered `` non - conservative '' model this limit on @xmath54 is sufficiently higher . we would like to stress also that in the case under consideration there exists a region of unstable mass exchange where value of @xmath54 is very small . the reason of instability [ i.e. the violation of criteria ( [ dr2r ] ) ] is fast increase of the ratio of the lost angular momentum to the mean specific momentum of binary system [ see eq . ( [ dabpsi ] ) and ( [ xxxx ] ) ] . this circumstance can lead to the destruction of a donor of very low mass . is also known . it may be caused by low efficiency of interaction between accretion disk and orbital movement ( see , e.g. , [ [ rsh83],[hp84 ] ] ) . ] thus , cvs that begin their evolution in a stable mode finish it by a catastrophic disruption of the donor when @xmath54 becomes lower than a certain threshold . to take into account results of 3d gas dynamical simulations of the mass transfer in cvs and to incorporate proper estimates of the loss of angular momentum we investigated the evolution of cvs on the basis of various assumptions on its conservativity . evolution of the donor was calculated using a modified version of the code for numerical modeling of low - mass stars which was used in the previous works of the authors ( see , e.g. , [ [ ft94],[fed97],[fy84][tf89 ] ] ) . the code uses opacity tables of huebner _ et al . _ [ [ hmma77 ] ] with addition of data of alexander _ et al . _ [ [ ajr83 ] ] for low - temperature region . equation of state was adopted from fontaine _ et al . _ [ [ fgh77 ] ] with modifications suggested by denisenkov [ [ den89 ] ] . rates of nuclear reaction were adopted from [ [ hfcz83],[cfhz85 ] ] . all runs were carried out under assumption that the donor fills its roche lobe immediately after arriving on zams . masses of the donors were adopted in the range from @xmath138 to @xmath139 and for various values of @xmath54 . chemical composition was adopted as follows : @xmath140 , @xmath141 , @xmath142 . for the mixing length parameter a value of @xmath143 was adopted . we used the method of calculation of the donor mass loss rate for known stellar radius and effective radius of roche lobe suggested by kolb & ritter [ [ kr90 ] ] . based on the results of gas dynamical calculations , we assumed that the fraction of matter accreted by white dwarf is @xmath120 , and that angular momentum carried away by the matter leaving the system is defined by expression ( [ xxxx ] ) . we did nt include into the model mass and angular momentum loss due to outbursts . let us consider evolution of cv with parameters corresponding to the stable mass exchange in conservative model . in fig . [ ris5 ] ( `` orbital period logarithm of donor - star mass loss rate '' diagram ) two evolutionary tracks are shown . one of the tracks is calculated in the conservative approximation ( track 1 ) , and another one in the `` non - conservative '' approximation ( track 2 ) . initial masses of both donor and accretor were equal to @xmath144 . it is seen that these two tracks are almost identical . during the early phase evolution is driven by angular momentum loss due to the msw of the donor . this results in decrease of the orbital semiaxis and orbital period in the course of evolution . decrease of the mass of the donor as well as its radius lead to decrease of @xmath145 [ see eq . ( [ damsw ] ) ] , and consequently to decrease of mass loss rate by the donor @xmath55 . for the `` non - conservative '' track , the value of @xmath55 is remarkably lower than for the `` conservative '' one , because in this phase the value of @xmath40 is negative ( see fig . [ ris2 ] ) and net loss of orbital angular momentum in the `` non - conservative '' model is lower than in the conservative one . the situation is inverted in later phases of evolution as @xmath54 reaches the values corresponding to large angular momentum loss . after the star becomes completely convective , msw from the donor stops . mass of the star at this instant is determined by the deviation from the thermal equilibrium . single ms stars become completely convective at @xmath146 , while mass - losing components of binaries do at lower masses @xmath147 ( under adopted assumption on chemical composition and opacity of stellar matter ) . in the `` conservative '' track mixing occurs at @xmath148 ( when the orbital period is equal to @xmath149 ) , and in the `` non - conservative '' track at @xmath150 ( @xmath151 ) . while in reality in this case the evolution of binary has to be terminated by a thermonuclear explosion of accretor . the continuation of calculation for @xmath152 can be advocated by the spin - up of accretor due to accretion because in this case the critical mass can remarkable exceed @xmath153 . ] thus , values of donor mass and orbital period differ only a little for the `` conservative '' track and for the `` non - conservative '' one . the rate of angular momentum loss decreases abruptly and contraction of binary systems becomes slower after the donor msw is stopped . as a result donor contracts ceases to lose mass . since at the moment of cessation of mass loss the radius of the donor is larger than the radius of thermally equilibrium ms star , the star contracts to the equilibrium radius . as a result the ratio of stellar radius to the effective radius of roche lobe decreases even more . in this phase the evolution of binary system is determined by gwr only . during this detached phase the binary can not be identified as cv , and certain range of the orbital periods becomes devoid of stars . this explains the origin of so called period gap of cvs . observed period gap of cvs lies between @xmath154 and @xmath155 [ [ cat97 ] ] . the boundaries of the `` theoretical '' period gap depend on adopted rate of angular momentum loss due to msw and on input physics parameters for evolutionary code that determine theoretical radii of stars . we used a code that gives somewhat overestimated value for the lower boundary . gap boundaries for the `` conservative '' track are equal to @xmath156 , and for the `` non - conservative '' one they are equal to @xmath157 . this discrepancy is related to the different rate of angular momentum loss just prior to gap , and with different masses of accretors after termination of semi - detached phase of evolution as well . note that the period gap is characterized not by total absence of cvs but rather by deficiency of cvs . the presence of cvs in the period gap can be explained as follows : i ) for cvs with masses of donors equal to @xmath158 mass exchange begins immediately in gap ; ii ) in the case when the donor fills its roche lobe not on zams but later , when hydrogen is almost burn out , then as mass decreases down to @xmath32 complete mixing does not occur due to presence a helium core in the star ( see , e.g. , [ [ tfey87],[fe89 ] ] ) . though , it must be admitted that the number of such systems is very small . the phase of evolution after period gap is characterized by lower rate of mass loss by the donor , because evolution now is driven by gwr . in this phase the rate of angular momentum loss from the system is sufficiently lower compared to the donor msw driven phase of evolution . another feature of this phase of evolution is existence of a minimum period of binary . short - period cutoff is related to the onset of significant degeneracy of the material of the donor and consequently to the change of mass - radius dependence law [ [ f71],[p81 ] ] : for small masses of donors @xmath159 their radii increase as mass decreases and the exponent of the mass - radius law tends to @xmath160 . minimal period corresponds to the value of exponent equal to @xmath161 . ) for the effective radius of roche lobe @xmath162 , condition @xmath163 , third kepler s law @xmath164 , and definition of @xmath128 ( [ dr2r ] ) . as a result we have : @xmath165 @xmath166 @xmath167 and , finally @xmath168 this conclusion is independent of the assumption of conservativity or non - conservativity of mass exchange . ] for `` conservative '' track the minimum period is equal to 75 minutes and for `` non - conservative '' one it is equal to 81 minutes . the difference in the values of @xmath169 is related to the difference in the total mass of the system and to the difference in the deviations of the radii of donor stars from thermal equilibrium values . the latter deviations are caused by different rates of mass loss . after passing the minimum the orbital period begins to increase , and in this phase the rate of mass loss by the donor quickly decreases . it is seen from fig . [ ris5 ] that difference between `` conservative '' and `` non - conservative '' tracks becomes larger in this stage . this can be naturally explained by increase of the rate of angular momentum loss with decrease of @xmath54 in the `` non - conservative '' case . `` conservative '' and `` non - conservative '' tracks have appreciably different values of mass transfer rates after passing the minimum of orbital period . this fact is of importance since the probability of discovery of a cv in a given range of @xmath170 is proportional to @xmath171 where @xmath172 is a positive constant @xmath173 for visual magnitude limited sample [ [ kkr98 ] ] . after passing the minimum of the orbital period ( i.e. in period range @xmath174 ) the mean value of the probability of discovery of cv for the `` non - conservative '' track is 1.2 times larger than for the `` conservative '' one . thus , despite a substantial difference in the rate of mass exchange , the probability of discovery of cv differs only a little in two cases under consideration . this is due to faster crossing of the corresponding period range in the `` non - conservative '' case . nevertheless , predicted difference in @xmath55 for `` conservative '' and `` non - conservative '' cases could be reflected in the distribution of cvs over types of variability since the character of variability of a cv depends on the rate of mass exchange . figure [ ris5 ] also shows the evolutionary tracks for stars that would be dynamically unstable within the conservative model . initial masses of the components were equal to @xmath144 and @xmath175 . of course , this combination of masses is rather extreme but we adopted these parameters for the sake of illustration . track 3 was calculated under the conventional assumptions on the loss of angular momentum due to gwr and msw only . this is a typical case of unstable mass exchange with unlimited growth of mass exchange rate and fast convergence of the time scale of mass loss to the dynamical one , i.e. runaway . strictly speaking this run should be considered as purely illustrative because in this case a common envelope has to form that increases the rate of angular momentum loss even more ( it was not taken into account in calculations ) . track 4 was calculated within the model in which the binary loses matter and angular momentum in accordance with eq . ( [ xxxx ] ) . it resembles the tracks for conventional case of stable evolution of cv and differs slightly from track 2 by a higher rate of mass loss in the initial phase of evolution . the latter is caused by some increase of @xmath176 at large @xmath54 [ see eq . ( [ damsw ] ) ] . it results in more strong deviations of the donor from thermal equilibrium and corresponding change of the upper border of period gap : it is equal to @xmath177 for track 4 , and @xmath178 for track 2 . the minimum of period for track 4 is equal to 76 minutes , i.e. it is nearly the same as for track 1 but slightly less than for track 2 . boundaries of stable mass exchange region were determined as follows : several sets of evolutionary tracks were calculated for systems with predefined masses of donors and values of @xmath54 that were varied with step 0.1 . the value of @xmath54 was considered as critical if for a given evolutionary track the mass accretion rate did nt exceed eddington s limit and if for the next evolutionary track ( with @xmath54 higher by 0.1 ) it did . note that consequent increasing of @xmath54 leads to unlimited growth of mass loss rate ( at values of @xmath54 greater by 0.2 0.3 than critical ) . note also , that results of such procedure are in good agreement with technique based on comparison of derivatives of stellar radius and effective roche lobe radius . the new boundary of the stable mass exchange region is shown in fig . [ ris1new ] ( under assumption of `` non - conservative '' evolution ) . it is well seen that the new region of stable mass exchange contains all observed cvs with estimated masses of components . thus , within the `` non - conservative '' model the problem of location of observed systems outside the region of stable mass exchange does not arise . on the other hand , in the @xmath65 diagram all observed cvs are concentrated near @xmath179 , while `` non - conservative '' model supposes sufficiently higher values of @xmath54 . to investigate this phenomenon , we have calculated a set of `` non - conservative '' tracks with initial donor masses from @xmath180 up to @xmath181 with step @xmath138 and with such values of @xmath54 that tracks start in the vicinity of the new boundary of stable mass exchange region . these tracks are shown in fig . [ ris7 ] ( @xmath65 diagram ) . based on calculations of these tracks , it is possible to get isochrones that are also shown in fig . [ ris7 ] for the moments of time equal to @xmath182 , @xmath183 , @xmath184 , @xmath185 , @xmath186 , @xmath187 years ( time is measured from the onset of mass exchange between components ) . years and @xmath185 years for tracks with initial donor mass @xmath188 merge , since this interval of time binaries spend in the period gap . ] it is this `` extreme '' set of tracks that was chosen since these systems spent the majority of time between an old boundary of the stable mass exchange region and the new one . it is clear , that systems with smaller values of @xmath54 ( tracks of which begin to the left from new boundary ) spent there less time . thus , cvs with parameters that are close to the `` extreme '' ones have most chances to be found out in this region . with respect to the location of isochrones in fig . [ ris7 ] all tracks can be subdivided into two distinct groups : i ) tracks for systems with massive donors ( from @xmath189 up to @xmath139 ) , for which the evolution prior to the period gap is defined mainly by the loss of systemic angular momentum due to the donor msw ; ii ) tracks for systems of low - mass donors ( from @xmath138 up to @xmath190 ) , for which this stage is absent , because their donors are initially completely convective . for the tracks of the first group a high ( @xmath191@xmath13 ) rate of mass loss by the donor prior the period gap is typical , and , hence , donor mass and mass ratio decrease rapidly . this results in a fast evolution in the @xmath65 diagram . as a result , the part of the diagram near the new boundary is crossed very quickly . this explains the absence of observed cvs near new boundary . for the tracks of the second group ( dropping a short initial phase with large rate of mass exchange ) the evolution of binary system is characterized by low and nearly constant mass loss rate ( @xmath192@xmath13 ) for a long time until period minimum . consequently , the mass of the donor and mass ratio of components change slowly . hence , if beginning the evolution near the new boundary of stable mass exchange region , the system should remain there long enough . the absence of cvs in this region can be explained as follows : masses of dwarfs in zero - age cvs can not be higher than @xmath193 that is minimum mass of helium core for a star leaving ms . results of three - dimensional gas dynamical simulations of the flow structure in semi - detached binaries show that during mass exchange a significant fraction of the matter leaves the system . on the other hand , a number of observed cvs have a combination of donor mass and mass ratio of components forbidden in the evolutionary models based on conventional assumptions on the loss of angular momentum of binaries , since in these cases mass exchange should be unstable . this contradiction can be resolved in the framework of the model suggested here . this model takes into account losses of mass and angular momentum from the system according to the results of gas dynamical simulations . it is shown that approximation for outflowing angular momentum as the difference of specific angular momentum of the donor matter in @xmath0 and specific angular momentum of accreted matter ( being multiplied by appropriate mass flows ) permits to explain observations satisfactorily . note , that the new `` non - conservative '' model practically does not change such parameters of evolutionary tracks as limits of period gap and minimal period . thus , the main conclusion of our study is : if we use a `` non - conservative '' approximation for cv evolution , we obtain a new boundary of the region of stable mass exchange in @xmath194 plane which explains the distribution of observed cvs better than conventional models . this work was supported by the russian foundation for basic research grants 99 - 02 - 17619 and 99 - 02 - 16037 and russian federation presidential grant 99 - 15 - 96022 . 1 . [ rjw82 ] _ rappaport , s. , joss , p.c . , & webbink , r.f . _ // astrophys . 2 . [ rvj83 ] _ rappaport , s. , verbunt , f. , & joss , p.c . _ // astrophys . [ it84 ] _ iben , i.,jr . , & tutukov , a.v . _ // astrophys . [ p84 ] _ patterson , j. _ // [ w85 ] _ webbink , r.f . _ // interacting binary stars / eds w.h.g . lewin , j. van paradijs , e.p.j . van den heuvel . cambridge : cambridge univ . press , 1985 . [ kolb93 ] _ kolb , u. _ // astron . astrophys . p.149 7 . [ ft94 ] _ fedorova , a.v . , & tutukov , a.v . _ // astron . zhurn . p.431 ( astron . reports v.38 . p.377 ) . 8 . [ rit95 ] _ ritter , h. _ // evolutionary processes in close binary stars / eds r.a.m.j . wijers , m.b . davies , c.a . tout . dordrecht : kluwer , 1992 . [ pol96 ] _ politano , m. _ // p.338 . 10 . [ fed97 ] _ fedorova , a.v . _ // binary stars / ed . moscow : cosmoinform , 1997 . p.179 ( in russian ) . 11 . [ kmg62 ] _ kraft , r.p . , mathews , j. , greenstein , j.l . _ // p.312 . 12 . [ p67 ] _ paczyski , b. _ // acta astron . . p.267 . 13 . [ sch62 ] _ schatzman , e. _ // ann . p.18 . 14 . [ mes68 ] _ mestel , l. _ // mon . not . r. astron . soc . p.359 . 15 . [ egg76 ] _ eggleton , p.p . _ // structure and evolution of close binary stars / eds p.p . eggleton , s. mitton , j. whelan . dordrecht : reidel , 1976 . p.209 . 16 . [ vz81 ] _ verbunt , f. , & zwaan , c. _ // astron . astrophys . [ yu73 ] _ yungelson , l.r . _ // nauchn . inform . p.71 ( in russian ) . 18 . [ kieboom81 ] _ kieboom , k.h . , & verbunt , f. _ // astron . astrophys . p.l11 . 19 . [ kk95 ] _ king , a. , & kolb , u. _ // astrophys . p.330 . 20 . [ bbkc98b ] _ bisikalo , d.v . , boyarchuk , a.a . , chechetkin , v.m . , kuznetsov , o.a . , & molteni , d. _ // mon . not . r. astron . p.39 . 21 . [ bbkc98c ] _ bisikalo , d.v . , boyarchuk , a.a . , kuznetsov , o.a . , & chechetkin , v.m . _ // astron . zhurn . 1998 . v.75 . p.706 ( astron . reports v.42 . p.621 ; preprint astro - ph/9806013 ) . 22 . [ eggl83 ] _ eggleton , p.p . _ // astrophys . v.268 . p.368 . [ pacyn71 ] _ paczyski , b. _ // astron . astrophys . p.183 . 24 . [ kuz95 ] _ kuznetsov , o.a . _ // astron . p.508 ( astron . reports v.39 . p.450 ) 25 . [ ll62 ] _ landau , l.d . , & lifshitz , e.m . _ the classical theory of fields . 1971 , oxford : pergamon press 26 . [ skum72]_skumanich a. _ // p.565 . 27 . [ sr83 ] _ spruit , h.c . , & ritter , h. _ // astron . 1983 , v.124 . p.267 . 28 . [ hkn96 ] _ hachisu , i. , kato , m. , & nomoto , k. _ // [ yl98 ] _ yungelson , l. , & livio , m. _ // astrophys . j. 1998 , v.497 . p.168 . 30 . [ kh94 ] _ kato , m. , & hachisu , i. _ // astrophys . p.802 . 31 . [ yuetal96 ] _ yungelson , l. , livio , m. , truran , j. , tutukov , a. , & fedorova , a. _ // . p.890 . 32 . [ lgr91 ] _ livio , m. , govarie , a. , & ritter , h. _ // astron . astrophys . v.246 . p.84 [ skr98 ] _ schenker , k. , kolb , u. , & ritter , h. _ // mon . not . r. astron . soc . p.633 . 34 . [ ty71 ] _ tutukov , a.v . , & yungelson , l.r . // nauchn . inform . p.88 ( in russian ) . 35 . [ tfy82 ] _ tutukov , a.v . , fedorova , a.v . , & yungelson , l.r . pisma astron . zhurn . p.365 ( sov . p.198 ) . 36 . [ kool92 ] _ de kool , m. _ // astron . v.261 . p.188 [ cat97 ] _ ritter , h. , & kolb , u. _ // astrophys . suppl . p.83 . 38 . [ smith_dhillon ] _ smith , d.a . , & dhillon , v.s . // mon . not . r. astron . soc . v.301 . p.767 . [ sawada84 ] _ sawada , k. , hachisu , i. , & matsuda , t. _ // mon . not . r. astron . soc . [ matese83 ] _ matese , j.j . , & whitmire , d.p . // astrophys . v.266 . p.776 [ bbkc98d ] _ bisikalo , d.v . , boyarchuk , a.a . , kuznetsov , o.a . , & chechetkin , v.m . _ // astron . zhurn . 1999 . v.76 . p.270 ( astron . reports v.43 . p.229 ; preprint astro - ph/9812484 ) . 42 . [ rsh83 ] _ ruderman , m.a . , & shaham , j. _ // nature . 1983 . v.304 . p.425 . 43 . [ hp84 ] _ hut , p. , & paczyski , b. _ // astrophys j. 1984 . v.284 . p.675 [ fy84 ] _ fedorova , a.v . , & yungelson , l.r . _ // astrophys . space sci . p.125 . 45 . [ tfey85 ] _ tutukov , a.v . , fedorova , a.v . , ergma , e.v . , & yungelson , l.r . pisma astron . zhurn . p.123 ( sov . p.52 ) . 46 . [ tfey87 ] _ tutukov , a.v . , fedorova , a.v . , ergma , e.v . , & yungelson , l.r . _ // pisma astron . zhurn . p.780 ( sov p.328 ) . 47 . [ fe89 ] _ fedorova , a.v . , & ergma , e.v . // astrophys . space sci . [ tf89 ] _ tutukov , a.v . , & fedorova , a.v . _ // p.1172 ( sov p.606 ) . 49 . [ hmma77 ] _ huebner , w.f . , merts , a.l . , magee , n.h . , & argo , m.f . astrophysical opacity library : los alamos sci . lab . rep . la-6760-m . los alamos , 1977 . 50 . [ ajr83 ] _ alexander , d.r . , johnson , h.r . , & rypma , r.l . _ // astrophys . v.272 . p.773 [ fgh77 ] _ fontaine , g. , graboske , h.c . , & van horn , h.m . // astrophys . v.35 . p.293 . 52 . [ den89 ] _ denisenkov , p.a . _ // nauchn . p.145 ( in russian ) . 53 . [ hfcz83 ] _ harris , m.j . , fowler , w.a . , caughlan , g.r . , & zimmerman , b.a . astron . astrophys . p.165 . 54 . [ cfhz85 ] _ caughlan , g.r . , fowler , w.a . , harris , m.j . , & zimmerman , b.a . _ // atom . data and nucl . data tabl . p.197 . 55 . [ kr90 ] _ kolb , u. , & ritter , h. _ // astron . astrophys . v.236 . p.385 [ f71 ] _ faulkner , j. _ // 1971 v.170 . p.l99 . [ p81 ] _ paczyski , b. _ // acta astron . [ kkr98 ] _ kolb , u. , king , a.r . , & ritter , h. _ // mon . not . r. astron . v.298 . p.l29 .

the key principle governing the onset of superconductivity is the spontaneous breaking of the u(1 ) gauge symmetry via a non - zero vacuum expectation value of a charged field @xcite . an immediate consequence of this is the generation of a photon mass and , for type ii superconductors , the formation of vortices which confine magnetic flux to narrow tubes as revealed by the ginzburg - landau theory@xcite . lattice studies of dual superconductivity in su(n ) gauge theories seek to exploit this connection in establishing the underlying principle governing color confinement . on a four dimensional lattice , the effective euclidean lattice higgs theory is the appropriate generalization of the ginzburg - landau theory . the higgs field is a 0-form living on the sites and the gauge field is a 1-form living on the links . classical solutions to this theory exhibit the connection between the non - zero vacuum expectation value of the higgs field and vortex formation . in u(1 ) lattice pure gauge theory ( no higgs field ) , this same connection is seen to be present , not in the defining variables , but rather in the dual variables . more specifically : 1 . a field with non - zero magnetic monopole charge , @xmath0 , has been constructed@xcite . it is a composite 4-form living on hypercubes constructed from gauge fields . there are also monopole current 3-forms . + on the dual lattice this monopole operator is a 0-form living on dual sites . the monopole currents are 1-forms living on dual links . these currents either form closed loops or end at monopole operators . + the monopole operator has a non - zero vacuum expectation value in the dual superconducting phase , @xmath1 , thereby signaling the spontaneous breaking of the u(1 ) gauge symmetry . dual abrikosov vortices have been seen in simulations@xcite . they are identified by the signature relationship between the electric field and the curl of the monopole current in the transverse profile of the vortex . the dual coherence length , @xmath2 measures the characteristic distance from a dual - normal - superconducting boundary over which the dual - superconductivity turns on . the dual london penetration length , @xmath3 measures the attenuation length of an external field penetrating the dual - superconductor . the dual photon mass @xmath4 and the dual higgs mass @xmath5 . a signal @xmath1 without the consequent signal of a dual photon mass does not imply confinement . an observation of a dual photon mass , i.e. vortex formation , without @xmath1 does not reveal the underlying principle governing the phenomenon . the lattice higgs theory , treated as an effective theory , i.e. limited to classical solutions , and considered in the dual sense , provides a model for interpreting simulations of the pure gauge theory that can reveal these important connections . the link of these considerations to confinement in non - abelian gauge theory is through the abelian projection@xcite . one first fixes the gauge while preserving u(1 ) gauge invariance . the non - abelian gauge fields can be parametrized in terms of a u(1 ) gauge field and charged coset fields . the working hypothesis is that operators constructed from the u(1 ) gauge field alone , i.e. abelian plaquettes , abelian wilson loops , abelian polyakov lines and monopole currents , will exhibit the correct large distance correlations relevant for confinement . but there is as yet no definitive way to choose the gauge which defines the abelian projection and hence no unique way to define abelian links and coset fields from the su(2 ) links . we are seeking a judicious choice of dynamical variables defined by a particular abelian projection which can account for confinement via an effective theory involving these dynamical variables . this should be our first goal . if this is solved satisfactorily , then we can investigate how the picture changes if we go to an alternative set of variables a different abelian projection since the phenomenon we are describing is of course gauge invariant . we do not expect that the same mechanism would describe confinement in two different abelian projections . there is a very nice illustration of this point in a paper by chernodub , polikarpov and veselov@xcite . they compare two abelian projections . * the first is the maximal abelian gauge@xcite which is the most widely studied and perhaps the most promising candidate . monopoles are the magnetic charge carriers of the persistent currents . * second they exhibit an abelian projection in which confinement is due to objects other than monopoles . they choose the `` minimal abelian projection '' and show that confinement is due to topological objects which are denoted `` minopoles '' . here we have two projections , two sets of dynamical variables , and two different descriptions of confinement , both viable candidate mechanisms . our goal in this paper is to test the connection between vortex formation and a non - zero vacuum expectation of the monopole field in the same abelian projection . there is an intimate connection between these two results in a higgs theory and hence it is a strong test of the idea for the dual theory . the two obvious candidate projections are the polyakov gauge and the maximal abelian gauge . polyakov gauge : : : del debbio , di giacomo , paffuti and pieri@xcite have constructed a monopole field operator that shows a very strong signal with a sharp discontinuity in the vacuum expectation value of the monopole operator as a function of temperature in the neighborhood of @xmath6 , the deconfining temperature . their calculation is manifestly gauge invariant but a polyakov line was used in the definition the monopole operator . in this sense we associate their calculation with the polyakov gauge as we will explain further below . we look at vortex formation in the present paper . maximal abelian gauge : : : the vortices are well established@xcite . the static potential constructed from abelian links gives as definitive a signal of confinement as the gauge invariant static potential as found by suzuki et.al.@xcite , stack et.al . * ; * ? ? ? * ) and bali et . bali et.al . find the abelian string tension calculation gives @xmath7 times the full string tension for @xmath8 . whether this approaches @xmath9 in the continuum limit remains to be seen . + the calculation of del debbio et.al@xcite of the monopole field operator as a practical matter is not adaptable to this gauge . it would require hundreds of gauge fixing sweeps of the whole lattice in order to accept or reject a single link update . chernodub , polikarpov and veselov@xcite have more recently calculated the constraint effective potential as a function of the monopole field in this gauge and found a symmetry breaking minimum . however they reported a problem of obtaining statistics and instead calculated an approximation to the effective potential . in this paper we seek to establish vortex formation in the polyakov gauge . more recently nakamura et al . @xcite studied an alternative monopole operator defined in terms of the variables which occur in the monopole form of the action . we have not yet addressed the issue of establishing vortex formation in that framework . we point out some technical difficulties in implementing the polyakov abelian projection . we use abelian polyakov lines to represent static sources . however they have an anomalous behavior which we discuss in sec.iv . after an appropriate modification of the sources we find vortices . however they are suppressed relative to the analogous calculation in the maximal abelian gauge . our conclusion is that the polyakov abelian projection is not likely to give a quantitative description of confinement within the the confines of this particular scenario . the fact that there are vortices in conjunction with a non zero vacuum expectation value of a monopole operator confirms the connection we sought . however that is not enough to get agreement between the string tensions as this calculation shows . the charged coset fields are thought not to contribute to long range physics , but they are capable of partially screening the sources . this consideration does not arise in the u(1 ) theory because there is no charged dynamical field and this effect appears to be minimized in the maximal abelian gauge . we conclude in this paper that the screening due to these charged fields is quite large , and that is why the vortices are suppressed . the signal of this suppression can be seen most easily in measurements of @xmath10@xmath11 . measuring this on a site coincident with the static source shows a very large attenuation in the polyakov projection compared to the maximal abelian projection . a recent measurement of a gauge invariant definition of @xmath10@xmath11 by skala et . al.@xcite for su(3 ) which uses the polyakov line to define the adjoint field exhibited no screening . to unify the discussion of various choices of the abelian projection in sec . ii we stress the introduction of a composite field @xmath12 constructed from the the gauge configuration @xmath13 . in sec . [ sec:3 ] we review the gauge invariant definition of field strength in order to connect our calculation to the monopole operator calculation of del debbio et.al.@xcite in sec . [ sec:4 ] we confront problems with the polyakov gauge as mentioned above . we must represent static sources with polyakov lines in order to study the theory at finite temperature . there is an inherent ambiguity on the lattice of ordering the eigenvalues of a diagonalized matrix at each site . the anomalous behavior is intimately connected with this freedom . we do not find a completely satisfactory solution , nevertheless we are led to an acceptable definition of sources . in sec . [ sec:5 ] we present the results of our simulation . the abelian projection requires a gauge condition that breaks the su(2 ) gauge invariance but preserves u(1 ) gauge invariance . this can be accomplished through the following two - step construction : 1 . consider a path ordered product of links , i.e. a wilson line , which begins and ends at a particular site @xmath14 . this defines an su(2 ) group element , @xmath15 , e.g. , an open plaquette or polyakov line at the site @xmath14 ( no final trace at @xmath14 ) . this in turn defines a composite field , @xmath16 . @xmath17 . \label{e2.1}\end{aligned}\ ] ] this composite field transforms under the adjoint representation of su(2 ) , @xmath18 we require two further generalizations to cover the adjoint field definitions used in this paper . ( i ) in the su(2 ) fundamental representation , a sum of group elements is , up to a normalization , also a group element . hence we can also construct an adjoint field from a sum of the above defined wilson lines . ( ii ) further , the gauge invariant path ordered products of links can include the adjoint field itself , @xmath19 , at sites along the path . this leads to a self consistent definition of an adjoint field as , e.g. , the case of the maximal abelian gauge . 2 . after constructing @xmath16 , we then perform a gauge transformation at each @xmath14 that fixes the adjoint field @xmath20 in the @xmath21 direction @xmath22 this gauge fixing construction is invariant under subsequent u(1 ) gauge transformations which rotate in the @xmath23 plane , leaving the @xmath21 direction invariant . the ambiguity of rotating @xmath24 into the @xmath25 or @xmath26 @xmath21-direction is equivalent to the choice of ordering the eigenvalues of @xmath27 . there are @xmath28 variants on gauge fixing corresponding to choosing @xmath29 or @xmath30 in the 3-direction at each of the n sites . we address this issue in sec . the polyakov gauge is defined by the above construction in which @xmath15 is taken to be an open polyakov line , @xmath31 beginning and ending at @xmath14 . @xmath32 . \label{e2.1a}\end{aligned}\ ] ] in section iv we argue there that we should choose @xmath33 at all sites , or equivalently @xmath34 at all sites . the maximal abelian gauge@xcite is conventionally defined as a configuration , @xmath13 , which maximizes @xmath35 $ ] , @xmath36 = \sum_{x , \mu } \frac{1}{2}tr\biggl[\sigma_3 u_{\mu}(x)\sigma_3 u_{\mu}^{\dagger}(x)\biggr ] , \label{e2.16}\end{aligned}\ ] ] over the set of gauge equivalent configurations @xmath13 . the continuum limit of this condition is @xmath37 = 0.\end{aligned}\ ] ] we wish to cast this into the language of the composite adjoint field . to do this let us consider a gauge fixing sweep in which we propose a gauge transformation on the links @xmath38 to test for an increase in @xmath39 . then @xmath39 becomes @xmath40 & = & \sum_{x , \mu } \frac{1}{2}tr\biggl[\sigma_3 u_{\mu}'(x)\sigma_3 u_{\mu}'^{\dagger}(x)\biggr ] , \nonumber \\ & = & \sum_{x , \mu } \frac{1}{2}tr\biggl[\sigma_3 \biggl(g(x ) u_{\mu}(x ) g^{\dagger}(x + \mu ) \biggr)\sigma_3 \biggl(g(x + \mu ) u_{\mu}^{\dagger}(x ) g^{\dagger}(x ) \biggl)\biggr ] , \nonumber \\ & = & \sum_{x , \mu } \frac{1}{2}tr\biggl[\biggl(g^{\dagger}(x)\sigma_3 g(x)\biggr ) u_{\mu}(x ) \biggl(g^{\dagger}(x + \mu ) \sigma_3 g(x + \mu)\biggr ) u_{\mu}^{\dagger}(x ) \biggr ] . \label{e2.16a}\end{aligned}\ ] ] hence a proposed gauge transformation of @xmath41 is equivalent to proposing a change in the component of @xmath42 from @xmath43 at each site to @xmath44 , while leaving @xmath41 unchanged . let us reformulate this procedure slightly by defining a gauge invariant quantity @xmath45 similar to @xmath39 but in which @xmath43 is promoted to a variable @xmath44 transforming as eqn.([2.101 ] ) . @xmath46 = \sum_{x , \mu } \frac{1}{2}tr\biggl[\phi(x)u_{\mu}(x)\phi(x+\mu ) u_{\mu}^{\dagger}(x)\biggr ] . \label{e2.15}\end{aligned}\ ] ] we search for the maximum by proposing @xmath47 , while holding @xmath13 constant , to test for an increase in @xmath45 . having found the maximum , if we then perform a guage transformation on both @xmath48 and @xmath41 , which rotates @xmath49 , or @xmath34 for all @xmath14 , then @xmath50 . but @xmath45 is gauge invariant by construction and hence in this last step , the value of @xmath45 does not change , and the transformed links are the solution to the maximization problem , eqn.([e2.16 ] ) . this construction divides the maximal abelian gauge fixing procedure into two logically independent steps described above , unifying maximal abelian gauge fixing with other gauge fixing schemes . in order to find the set @xmath51 corresponding to the maximum of @xmath45 , consider variations of @xmath52 about the stationary point under an infinitesimal transformation of @xmath16 , while leaving @xmath13 constant : @xmath53 the condition that @xmath45 is stationary under these variations is @xmath54 \phi(x)\biggr ) = 0 , \;\;\;\ ; \forall x , \label{e2.18}\end{aligned}\ ] ] where @xmath55 is defined @xmath56 the notation @xmath57 means that the quantity is parallel transported to the point @xmath14 as indicated in fig . the sum is over the nearest neighbor sites to the point @xmath14 . writing @xmath58 , the stationary condition , eqn . ( [ e2.18 ] ) , gives : @xmath59 since @xmath60 is arbitrary , @xmath12 must be either parallel or antiparallel to @xmath61 at each site . therefore @xmath62 where @xmath63 . let us now examine the second derivatives of @xmath52 about the stationary point . define @xmath64 - { \cal s}[\phi(x ) ] . \label{e5.25}\end{aligned}\ ] ] then , using the stationary condition , eqn . ( [ e2.2 ] ) , and keeping second order terms @xmath65 . \label{e5.3}\end{aligned}\ ] ] this can be cast into the form @xmath66 this shows that the second derivative with respect to the variables @xmath67 is @xmath68 a negative number . if we choose @xmath69 for all @xmath14 , then @xmath45 is a local maximum under these infinitesimal variations . this defines the composite adjoint field @xmath70 , i.e. step ( 1 ) in the process . in step ( 2 ) we perform a gauge transformation which takes @xmath71 into the 3-direction . if we choose a gauge transformation that brings @xmath72 at all sites , then @xmath45 is cast into the form @xmath39 . this is the transformation to the the maximal abelian gauge , and justifies our notation : @xmath70 . equivalently , we can take @xmath73 at all sites and reach the same conclusion . consider a solution corresponding to @xmath74 . if we then set @xmath75 at one site , @xmath76 , and search anew for a solution , we will then obtain one in which @xmath77 is given by the normalized average of @xmath77 at the neighboring sites , parallel transported to @xmath14 , except that @xmath78 is antiparallel to the average of its neighbors . equation([e5.4 ] ) , shows that the new solution is no longer a local maximum of @xmath45 and hence it is not associated with the maximal abelian gauge . we will return to this issue in a discussion of the minimal abelian gauge in sec . iv . in order to find a stationary solution one can start with an arbitrary configuration @xmath51 and then iterate . to update the site @xmath14 , one calculates @xmath79 , eqn.([e2.19 ] ) , based on the current neighboring @xmath48 s . then set @xmath80_{normalized } \;\ ; \rightarrow \;\ ; \widehat{\phi}(x),\end{aligned}\ ] ] where @xmath81 is the over - relaxation parameter . for @xmath82 we go directly to the stationary point for that site based on the current values of the neighboring @xmath77 s . however by employing over - relaxation with @xmath83 , the overall search rate for a self consistent solution is greatly enhanced . we tune this parameter to optimize performance . this procedure requires an iterative search for a solution but it does not require iterative gauge transformations . after finding a solution , a single gauge transformation brings the gauge field into the maximal abelian gauge . this method of obtaining the maximal abelian gauge involves fewer multiplications than directly maximizing @xmath39 through gauge transformations and hence is a little faster . the maximal abelian gauge condition has multiple solutions , known as gribov copies@xcite for a given field configuration . the particular solution depends on the initial configuration and on the update algorithm . however bali et . al.@xcite have studied these multiple solutions and have offered an algorithm that limits this ambiguity . this gribov ambiguity does not seem to have a large effect on the measured values of abelian projected observables . in our discussion of the maximal abelian gauge , we should understand that the above procedure is very unlikely to find the global maximum since there is no direct procedure to achieve that . hence it must be understood that in general we are only at a local maximum , not the global maximum , of @xmath39 . the lattice georgi - glashow model@xcite has an elementary adjoint higgs field living on the sites coupled to the gauge field living on the links . this allows one to define a gauge invariant field strength . the lattice definition of this is given in fig . this is the construction used to identify the ordinary magnetic field due to monopoles in non - abelian gauge theory@xcite . the classical continuum limit of these operators gives : @xmath84 \biggr)\biggr ] \nonumber\\ & & + \frac{1}{2 } a^2 tr\biggl[\phi \biggl(\partial_{\mu } \phi - ig[a_{\mu},\phi]\biggr ) \biggl(\partial_{\nu } \phi - ig[a_{\nu},\phi]\biggr)\biggr ] , \label{e3.1}\end{aligned}\ ] ] where @xmath85 is defined @xmath86 , @xmath87 is the lattice spacing and @xmath88 is the gauge coupling . this manifestly gauge invariant form can be cast into the alternative form : @xmath89 , \label{e3.2}\end{aligned}\ ] ] where we have used @xmath90 , ( @xmath91 ) , and @xmath92 . we can implement a gauge transformation that brings @xmath93 for all @xmath14 . then the last term of eqn . ( [ e3.2 ] ) vanishes . the first term gives a conventional abelian field strength , @xmath94 whether @xmath77 denotes an elementary higgs field , or a composite adjoint field as we discussed in sec . ii , this definition of field strength , eqn . ( [ e3.1 ] ) , is a manifestly gauge invariant quantity . however this is perhaps misleading in the composite adjoint field case because the definition of the field is identified with a specific gauge choice . del debbio et al . @xcite employed @xmath95 in defining the abelian field strength . this was needed in order to construct the monopole field operator . the vacuum expectation value of this field gave a very clear signal for dual superconductivity . although their construction is manifestly gauge invariant , it is identified with the polyakov gauge in the sense of this section . it is this connection that prompted us to study vortex formation in the polyakov gauge to probe the connection between vortices and a non - zero vacuum expectation value . the abelian projection consists of ( i ) choosing a gauge that preserves a u(1 ) gauge invariance , and ( ii ) identifying abelian links associated with that residual invariance@xcite . in sec . ii we discussed the first step of choosing the gauge except for the ambiguity of transforming @xmath77 into the @xmath96 or @xmath97-direction at each site . we will argue here that the @xmath98-direction is the preferred choice for all sites or @xmath97 for all sites . however this leads to an unsettling result for the polyakov gauge as we point out in part b below . hence we are led to examine the issue more broadly to check out alternatives in the following parts of this section . there are @xmath99 choices , where @xmath100 is the number of sites . the choice of rotating @xmath77 into the + 3-direction at every site is the prescription one would anticipate in order to get a smooth classical limit in which the changes in the gauge field are small between one site and its neighbors . but we feel alternatives should be presented for the record . the reader can skip to part h , the conclusion for this section , without losing continuity . we have summarized the key results of this section in table [ t1 ] . for the purpose of defining abelian links , consider the parametrization@xcite : @xmath101 the abelian link is defined through the factorization of eqn . ( [ e4.1 ] ) @xmath102 where @xmath103 . the abelian link is given by @xmath104 . the complex coset field is given by @xmath105 . under a u(1 ) gauge transformation , @xmath106 @xmath107 the coset vector fields live on the sites and are doubly charged . the ambiguity of rotating @xmath77 to the @xmath108 direction , @xmath109 , is the same as the ambiguity of ordering the eigenvalues of this @xmath110 diagonal matrix . the order of the eigenvalues for a particular link can be interchanged via an su(2 ) gauge transformation @xmath111 at the two neighboring sites corresponding to the link in question . translation invariance suggests that whatever rule we adopt for ordering the eigenvalues , we should use the same rule at every site . let us choose @xmath113 , denoting this as the @xmath112 prescription . in the polyakov gauge if we choose @xmath114 at every site , then there is a difficulty . the su(2 ) polyakov line , eqn . ( [ e2.1 ] ) , can be written @xmath115 . \label{4.31}\end{aligned}\ ] ] the @xmath116 element of this is the u(1 ) polyakov line @xmath117 . since @xmath118 is distributed in the domain @xmath119 $ ] the imaginary part of @xmath31 can not be negative . as a consequence , the average of @xmath31 will have a non - zero imaginary part due to this kinematical effect . this is anomalous . polyakov lines represent static sources . if we are to get the ` expected ' distribution in which @xmath120 is distributed symmetrically in the domain @xmath121 $ ] we would need a rule that allows the choice @xmath122 with a statistical weight equal to the @xmath114 case . if we choose @xmath73 at all sites , we reach the same conclusion except that @xmath123 can not be positive . we can examine the u(1 ) model for guidance . we expect abelian projected su(2 ) to resemble the u(1 ) theory . consider the limit in which the su(2 ) theory goes over to the u(1 ) theory . in su(2 ) the links lie in the group manifold @xmath124 . in u(1 ) they lie in a submanifold @xmath125 . consider adding a term to the action that is invariant under u(1 ) but which breaks su(2 ) . further choose it to bias the links into a peaked distribution centered on @xmath125 . then as the width of the peak approaches zero , the su(2 ) theory goes over to the u(1 ) theory . in u(1 ) , the polyakov line @xmath126 takes values for @xmath118 distributed in the domain @xmath121 $ ] . positive and negative values of @xmath118 are equally probable . to try to resolve the discrepancy between the u(1 ) theory and abelian projected su(2 ) in the @xmath112 prescription , let us consider a prescription in which we choose @xmath112 or @xmath26 depending on the value of @xmath12 . for example , we can correlate the signs @xmath112 or @xmath26 with the sign of @xmath129_3 $ ] . in other words , if @xmath12 is in the upper hemisphere then choose @xmath112 . if it is in the lower hemisphere then choose @xmath26 . we denote this the @xmath127 prescription . the phase of the abelian polyakov line will now be distributed in the expected way , @xmath121 $ ] . this is perhaps the prescription most closely associated with the limiting procedure in the previous section . this prescription is equivalent to choosing @xmath112 or @xmath26 randomly . the reason is that a gauge transformation can give an arbitrary orientation to @xmath77 . if after each update , we follow it with a random gauge transformation , ( an update followed by a gauge transformation is indistinguishable from an update alone ) then the assignments of @xmath112 or @xmath26 at each site will be random . the su(2 ) transformations do not respect the homotopy classes of u(1 ) . we now apply the @xmath127 prescription to the abelian projection in the polyakov gauge and point out a flaw in this choice . let @xmath130 denote the open polyakov line at site @xmath14 , and @xmath131 the open polyakov line at the neighboring site in the @xmath132 direction . then @xmath133 where @xmath41 is the link in the @xmath134 direction connecting the two sites . now go to the polyakov gauge . suppose we choose @xmath114 at both sites . then @xmath135 and the solution to eqn . ( [ e4.41 ] ) gives @xmath136 showing that the links in the @xmath134 direction are diagonal . contrast this with the case in which we choose @xmath114 at one site and @xmath122 at the neighboring site . then @xmath137 and the solution to eqn . ( [ e4.41 ] ) gives @xmath138 this second case does not give an acceptable abelian projection since the @xmath134 abelian link is undefined , i.e. the @xmath116 element of the link matrix is zero . we can still define a @xmath127 prescription appropriately modified for the polyakov gauge in which the spatial sites have independent random @xmath127 choices , but that the choice is independent of the values of @xmath134 . however this construction has a further problem . consider the correlator shown in fig . [ f2.5 ] . _ the correlation of the field strength with a product of two abelian polyakov lines representing a static quark - antiquark pair vanishes . _ this correlation is given by @xmath139 the @xmath127 signs , arising from the prescription of a random gauge transformation , @xmath111 , have the effect of flipping the sign of the charge of the two sources . with a statistical weight equal for each of the four signs , the field averages to zero . the correlation also vanishes if one only considers random @xmath140 gauge transformation at the sites contiguous with the plaquette . consider the parametrization of a link in eqn . ( [ e4.1 ] ) . we adopt this parametrization for a link in which the @xmath112 prescription is applied at both ends . the su(2 ) link and the corresponding abelian link are @xmath141 different prescriptions will change the su(2 ) links and accordingly the choice of the abelian angles . let @xmath142 be the gauge transformation changing the orientation of the adjoint field . ( this gauge transformation is defined up to a @xmath143 transformation . ) the link will transform according to one of the following cases : @xmath144 where the @xmath145 notation refers to the prescription at each end of the link . after a random gauge transformation the @xmath146 of the abelian plaquette angle will assume one of the forms shown in fig . we are working in the polyakov gauge hence there are only four possibilities since the adjoint field must be constant in @xmath134 . all the above configurations have the same statistical weight and for each configuration there is always another configuration differing only by the sign . hence the field strength in the presence of two sources vanishes in the @xmath127 prescription according to this argument also . this argument generalizes to any wilson loop since the parametrization could apply to a wilson line instead of a single link . further we can replace the polyakov line source by another wilson loop and then we obtain @xmath147 for the @xmath127 prescription if the two wilson loops do not share any spatial sites . to get further support for the @xmath112 prescription consider the abelian projected theory in the polyakov gauge but instead of calculating the field strength from the abelian plaquette let us instead represent the field strength by the manifestly gauge invariant form , fig . [ f2.0 ] . in the polyakov gauge , @xmath12 is independent of @xmath134 and as a consequence , the second term in fig . [ f2.0 ] vanishes . consider a static source represented by an abelian polyakov line at point @xmath14 , and the field strength at point @xmath148 as shown in fig . [ f3.0](a ) . let us define variables for the ( open ) plaquette in fig . [ f3.0](a ) by @xmath149 then the correlator is given by @xmath150 where the last equality is due to the choice of the polyakov gauge . measuring the field strength with the abelian plaquette , fig . [ f3.0](b ) , gives @xmath151 the field strength in eqn . ( [ e4.62 ] ) is manifestly gauge invariant , and therefore the prescription at the site @xmath148 has no bearing on the definition of the correlator . if we use the abelian plaquette , eqn . ( [ e4.63 ] ) , instead , then the value depends on the prescription . however the @xmath112 prescription has the same classical limit for small lattice spacing @xmath87 as the classical limit of the gauge invariant form , eqn.([e3.3 ] ) . this follows from the discussion in sec . [ sec:4.d ] . in fig . [ f2.6 ] in the cases in which the four points of the plaquette have the same prescription , the abelian plaquette has the standard classical limit . in the other cases , the plaquette involves angles arising from the off - diagonal elements of the su(2 ) link . for completeness we compare this to the connected definition of field strength at point @xmath148 as shown in fig . [ f3.0](c ) . in our notation the correlator is given by @xmath152 where @xmath153 means that the variables are parallel transported from @xmath14 to @xmath148 in the same sense as applied to eqn.([e2.19 ] ) . we consider only lattices which allow an even - odd checkerboard assignment to the sites . consider two solutions of eqn . ( [ e2.2 ] ) , corresponding to different values of @xmath154 . @xmath155 for the case in which the lhs of eqn . ( [ e2.2 ] ) involves only odd ( even ) sites then the rhs involves only even ( odd ) sites , and vice versa . hence @xmath156 in other words , the @xmath157 solution is identical to the @xmath158 solution except that it is flipped at the even ( odd ) sites . this is in contrast to the example in sec . iib where a change in @xmath154 led to a new solution which has no simple relation to the old solution . we further note that each term in @xmath159 $ ] , eqn.([e2.15 ] ) , involves @xmath12 at an even site and @xmath12 at an odd site . therefore the effect of the flipping is to change the sign of @xmath45 @xmath160 = -{\cal s}[\widehat{\phi}_{ma } ] \label{e4.641 } \end{aligned}\ ] ] since @xmath161 maximizes @xmath45 , it follows that @xmath162 minimizes @xmath45 . this is corroborated by the stability analysis , eqn.([e5.4 ] ) . the adjoint field @xmath162 defines the minimal abelian projection introduced by chernodub , polikarpov and veselov@xcite . they showed that confinement is due to objects other than monopoles , which they denoted as minopoles . the arguments in this section lead us to adopt the @xmath112 prescription for abelian projected su(2 ) in the polyakov gauge . this prescription is not completely satisfactory . it does not go over smoothly to a u(1 ) theory as one might expect . this is indicated by the behavior of the phase angle of the abelian projected polyakov line . it is bounded artificially in the interval @xmath119 $ ] giving the expectation value of the polyakov line an anomalous imaginary part . the adjoint field defined for the maximal abelian gauge is a construct that does relate directly to an observable in this sense and does not seem to cause a similar difficulty . further we must rethink how to represent static sources . the distribution of the abelian polyakov phase angle in the domain @xmath119 $ ] is weighted by the haar measure , @xmath163 which is peaked at @xmath164 . this means , e.g. that the average of two polyakov lines @xmath165 will have a non - vanishing real part arising from the same kinematic effect . our solution to this is to define the quark - antiquark source as the connected part of the product of the two polyakov loops , @xmath166 in this section we study the electric flux and the magnetic monopole current distributions around static sources in the polyakov gauge with the @xmath112 prescription and the definition of abelian static sources as in sec . [ sec:4.g ] . moreover we make a comparison between this projection and the maximal abelian projection in which the main properties of dual vortices have already been studied . our simulations were performed on a @xmath167 lattice with the standard form of the wilson action . the calculation follows the method given in @xcite . we used two abelian polyakov lines correlated along the 3-direction as static sources in the finite temperature theory . the longitudinal electric field between two charges separated by a distance @xmath168 in the 3-direction is given by : @xmath169 where @xmath170 is the angle of the abelian plaquette in the @xmath171 plane , @xmath172 is the angle of the product of the two abelian polyakov lines and @xmath173 is the electric charge . magnetic monopole currents can be identified by the dual maxwell equations that in the continuum are : @xmath174 on the lattice one can choose @xcite : @xmath175 we measured the magnetic current distribution around static sources : @xmath176 where the curl of the magnetic current is constructed as in @xcite , and @xmath177 is the magnetic charge . we made our simulations at @xmath178 and @xmath179 . as we chose the temporal extent of the lattice equal to 4 , the first value of @xmath180 corresponds to the confined phase while the second corresponds to the unconfined phase . the results for the polyakov gauge are shown in fig . [ empoly ] for @xmath181 . for @xmath178 we see a clear signal for both the electric field and the curl of the magnetic currents on the axis of the two charges . as we move in the transverse direction @xmath182 , we observe a rapid fall off of the electric field that always vanishes after a few lattice spacings and a behavior of the monopole current consistent with the predictions of the ginzburg - landau theory . as the distance @xmath168 of the static charges increases , the values of the electric field decreases as expected . in the unconfined phase ( @xmath179 ) we see a remarkably different behavior of the curl of the magnetic currents that is much smaller than in the other phase and at @xmath183 is consistent with zero everywhere . moreover the electric field approaches zero less rapidly although not clearly evident from the figure . a similar scenario has been found for the maximal abelian gauge@xcite . in fig . [ emmag ] we show the results in this gauge for the case @xmath183 . we used the procedure described in sec . [ sec:2 ] to define the abelian projection . although both abelian projections produce results consistent with the ginzburg - landau theory , the dual superconductivity parameters are quite different for the two projections . we are not able to perform a quantitative analysis in order to evaluate the coherence length @xmath2 and the london penetration length @xmath3 . nevertheless while the data for the maximal abelian gauge support a non - vanishing value for the coherence length , there is no evidence for @xmath2 being different from zero in the polyakov gauge . moreover a comparison between the two different projections shows that in the polyakov gauge the peak values of the electric field and of the curl of the magnetic currents are more than an order of magnitude smaller than the correspondent values in the maximal abelian gauge . suzuki et . al.@xcite reported similarly that the string tension was suppressed in the polyakov gauge . to understand these differences we remember that after abelian projection we are left not only with the abelian links but also with the coset vector fields . the dynamics of these doubly charged fields is clearly dependent on the choice of the abelian projection . to show the different role of the coset fields in the two projections , we measure the divergence of the electric field giving the spatial distribution of the electric charge . the divergence at site @xmath14 is defined as @xmath184 so that all the six plaquettes have in common the link starting at @xmath14 and extending in the time direction . in fig . [ gauss ] we show the results for two polyakov lines separated by three lattice spacings . the figure clearly shows that the effective charge of the static source is much smaller in the polyakov gauge and that the coset fields in the two prescriptions respond in opposite ways to the presence of an electric charge with the fields in the polyakov gauge shielding the static charge . skala et al . have studied @xmath185 in su(3 ) in a gauge invariant formulation and conclude there is no screening @xcite . using the u(1 ) theory as a guide , we explore the connection between ( i ) the non - vanishing of a vacuum expectation value of the monopole field and ( ii ) vortex formation . item ( i ) provides the underlying principle , item ( ii ) gives direct evidence of a photon mass and confinement . in the abelian projection of su(2 ) @xmath186 u(1 ) there is an additional consideration . after identifying the u(1 ) gauge field , there remain dynamical charged vector coset fields . they can screen the static sources and change the electric flux , affecting the string tension . from prior work we know that , e.g. , the minimal abelian projection and the maximal abelian projection account for the physics of confinement through very different mechanisms . we anticipated that the monopoles in the polyakov projection would give a picture very similar to the maximal abelian gauge . however we found that the coset fields greatly suppress the static sources . it could be that we are much farther from the continuum limit than we thought . but that seems unlikely since other quantities are close to scaling values . our definitions of field strength , monopoles , static sources etc . have some leeway , but all should give the same continuum limit . or it could be that the dynamical variables arising from the polyakov abelian projection do not adequately separate the short distance and long distance physics . this leaves the maximal abelian gauge as the prime candidate to define an effective theory of confinement in this scenario . we would like to thank adriano di giacomo and tsuneo suzuki for discussions . work was partially supported by the department of energy under grant no . de - fg05 - 91er40617 . giuseppe di cecio was partially supported by a university of pisa postdoctoral fellowship . 99 s. weinberg , vol ii , cambridge university press , cambridge 1996 , see chapter 21.6 for a thorough treatment of this viewpoint . for a review and further references see e.g. r.w . haymaker , _ `` dual abrikosov vortices in u(1 ) and su(2 ) lattice gauge theories '' _ , proceedings of the international school of physics `` enrico fermi '' , course cxxx , a. di giacomo and d. diakonov ( eds . ) ios press , amsterdam 1996 , hep - lat 9510035 . for a review of monopoles and confinement , see e.g. t. suzuki , nucl.phys . * b * proc . suppl . * 30 * , 176 ( 1993 ) . v. l. ginzburg and l. d. landau , zh . ekxperim , i theor . fiz , * 20 * , 1064 ( 1950 ) . a. a. abrikosov , sov . jetp * 32 * , 1442 ( 1957 ) . m. tinkham , introduction to superconductivity , ( mcgraw - hill , new york , 1975 ) j. frlich and p. a. marchetti , euro . * 2 * , 933 ( 1986 ) ; commun . phys . * 112 * , 343 ( 1987 ) ; * 116 * , 127 ( 1988 ) ; * 121 * , 177 ( 1989 ) ; lett . phys * 16 * , 347 ( 1988 ) . l. polley and u .- j . wiese , nucl . * b356 * , 621 ( 1991 ) . m. i. polikarpov , l. polley and u .- j . weise , phys . b253 * , 212 ( 1991 ) . l. del debbio , a. di giacomo and g. paffuti , phys . lett . * b349 * , 513 , ( 1995 ) ; nucl . phys b ( proc . suppl . ) * 42 * , 231 ( 1995 ) ; v. singh , r.w . haymaker , and d.a . browne , phys . rev . * d47 * , 1715 ( 1993 ) . haymaker , v. singh , d. browne and j. wosiek ; proc . of workshop on qcd vacuum structure and its applications , the american university of paris , june 1 - 5 , 1992 , p184 , world scientific 1993 . g. t hooft , nucl . phys . * b 190 * , 455 ( 1981 ) . chernodub , m.i . polikarpov and a.i . veselov , phys . lett . * b 342 * , 303 ( 1995 ) . a.s . kronfeld , g. schierholz and u.j . wiese , nucl . * b 293 * , 461 ( 1987 ) ; a.s . kronfeld , m.l . laursen , g. schierholz and u.j . wiese , phys . * b 198 * , 516 ( 1987 ) ; t. suzuki and i. yotsuyanagi , phys . rev . * d 42 * , 4257 ( 1990 ) . l. del debbio , a. di giacomo , g. paffuti and p. pieri , phys . lett . * b 355 * , 255 ( 1995 ) . v. singh , d.a . browne , and r.w . haymaker , nucl . b ( proc . suppl . ) * 30 * , 658 ( 1993 ) ; phys . b306 * , 115 ( 1993 ) . y. matsubara , s. ejiri , and t. suzuki , nucl . b34 * ( proc . suppl . ) 176 ( 1994 ) . y. peng and r.w . haymaker , phys . rev . * d 52 * 3030 , ( 1995 ) . p. cea and l. cosmai , nucl . * b * proc . suppl . * 30 * , 572 ( 1993 ) ; phys . rev . * d 52 * , 5152 ( 1995 ) . h. shiba and t. suzuki , phys . * b 333 * , 461 ( 1994 ) . j. stack and r. wensley , nucl . * d 22 * , 597 ( 1992 ) . j. stack , s. neiman and r. wensley , phys . rev . * d 50 * , 3399 ( 1994 ) g. s. bali , v. bornyakov , m. mller - preussker and k. schilling , phys . rev . * d 54 * 2863 ( 1996 ) . chernodub , m.i . polikarpov and a.i . veselov , nucl . . suppl . * 49 * , 307 , ( 1996 ) . n. nakamura , v. bornyakov , s. ejiri , s. kitahara , y. matsubara and t. suzuki , lattice-96 conference abstract , st . louis , june 1996 e - print 9608004 . p. skala , m. faber and m. zach , ahrenshoop symp . 301 - 306 , 1995 hep - lat/9603009 . v.n . gribov , nucl . phys . * b 139 * , 1 ( 1978 ) . h. georgi and s.l . glashow , phys . * 28 * , 1494 ( 1972 ) . g. t hooft , nucl . phys . * b 79 * , 276 ( 1974 ) . m. zach , m. faber , w. kainz and p. skala , phys . * b 358 * , 325 ( 1995 ) . t. suzuki , s. ilyar , y. matsubara , t. okude and k. yotsuji , phys . * b 347 * , 375 ( 1995 ) ; _ erratum - ibid . _ * b 351 * , 603 ( 1995 ) . .summary of gauge fixing . the problem in defining time - like abelian links can be remedied by choosing either @xmath112 " or @xmath26 " at all sites along a particular polyakov line . ) [ cols="<,^ , < , < " , ]

wyner and ziv @xcite , we study a setting where a sequence generated by a source is to be described succinctly to a reconstructor ( `` decoder '' ) with access to some side information . wyner and ziv showed that , although the side information is not available at the describing terminal ( `` encoder '' ) , it can be beneficial in improving the trade - off between the rate of description and the reconstruction distortion . they fully characterized this trade - off for memoryless sources with single - letter distortion measures . unlike the case without side information since the side information is used in the reconstruction process , and since the side information is not available at the describing terminal the describing terminal can not tell how the source sequence it observes will be reconstructed . in some settings , this is unacceptable . steinberg @xcite therefore studied the common - reconstruction problem where an additional restriction is imposed that the reconstruction sequence be computable with probability nearly one at the describing terminal . this greatly limits the extent by which the reconstruction can depend on the side information . more generally , there is a tension between the degree by which the reconstructing terminal utilizes the side information and the precision with which the describing terminal can compute the reconstruction sequence . it is this tension that we study in this paper . ( 7.7,3.8)(-4.5,-2.4 ) ( -5,-0.75)(1,0)1 ( -4.8,-0.5)@xmath0 ( 2.5,1)(0,-1)1 ( 2.7,0.5)@xmath1 ( -1.5,-0.5)@xmath2 ( -2,-0.75)(1,0)3.5 ( 2.5,-1.5)(0,-1)1 ( 2.75,-2.25)@xmath3 ( -3,-1.5)(0,-1)1 ( -2.8,-2.25)@xmath4 ( -0.3,-3.5)@xmath5 \leq \dd$ ] ( -6,-3.5)@xmath6 \leq \de;$ ] ( -3.75,-0.85)encoder ( 1.8,-0.85)decoder ( -4.0,-1.5)(0,1)1.5 ( -4.0,0)(1,0)2 ( -2,-1.5)(0,1)1.5 ( -2,-1.5)(-1,0)2 ( 1.5,-1.5)(0,1)1.5 ( 1.5,0)(1,0)2 ( 3.5,-1.5)(0,1)1.5 ( 3.5,-1.5)(-1,0)2 to quantify this tension , we require that the describing terminal generate an estimate of the sequence that will be produced at the reconstructing terminal ( figure [ fig : problem - diagram ] ) . we then study the distortions that can be simultaneously achieved at the describing terminal ( the encoder distortion ) and at the reconstructing terminal ( the decoder distortion ) as a function of the description rate . if the encoder s distortion function is the hamming distance and if the allowed distortion is zero , then our problem reduces in essence to steinberg s common - reconstruction problem . in section [ sec : related ] ahead . ] and if the allowed encoder distortion is infinite , our problem reduces to wyner and ziv s problem . we can thus view our problem as a generalization of the wyner - ziv problem and steinberg s common reconstruction problem . for discrete memoryless sources and finite single - letter distortion functions , we provide a single - letter characterization of the trade - off between the description rate and the distortions at the encoder and decoder sides . we also calculate this trade - off for a memoryless gaussian source and squared - error distortion functions . finally , in section [ sec : kdistortions ] , we generalize the results to account for more than two constraints and to allow each distortion function to depend on three arguments : the source symbol , the encoder s reconstruction symbol , and the decoder s reconstruction symbol . steinberg s work was also extended in other ways . kittichokechai , oechtering , and skoglund @xcite determined the rate - distortion function under a common - reconstruction constraint for a modified wyner - ziv setup where the encoder can influence the decoder s side information via an action - generator . timo , grant , and kramer @xcite , @xcite and ahmadi , tandon , simeone , and poor @xcite , @xcite derived the rate - distortions function under a common - reconstruction constraint for two special cases of the heegard - berger / kaspi problem ( the wyner - ziv problem with two decoders ) : @xcite , @xcite for physically degraded side informations , and @xcite , @xcite for complementary side informations . ahmadi , tandon , simeone , and poor @xcite , @xcite also presented the rates - distortions function under a common - reconstruction constraint for a cascade source - coding problem when the side informations are physically degraded . finally , already in @xcite , steinberg studied the implications of the common - reconstruction constraint on the simultaneous transmission of data and state and on joint source - channel coding for the degraded broadcast channel . the paper is organized as follows . in the rest of this section we introduce our notation . in section [ sec : setup ] we treat discrete sources and general distortions , and in section [ sec : gauss ] gaussian sources with quadratic distortions . in section [ sec : kdistortions ] we revisit discrete sources but this time with more and more general distortion constraints . random variables are denoted by upper - case letters and their realizations by lower - case letters . vectors are denoted by bold - face letters : random vectors by upper - case bold - face letters , and deterministic vectors by lower - case bold - face letters . sets and events are denoted by calligraphic letters , i.e. , @xmath7 . an @xmath8-tuple @xmath9 is denoted @xmath10 , and the @xmath8-fold cartesian product of the set @xmath7 is denoted @xmath11 . the convex hull of a set @xmath7 is denoted by @xmath12 . to indicate that the random variables @xmath13 and @xmath14 and conditionally independent given @xmath15 we write @xmath16 the transpose of a vector @xmath17 is denoted by @xmath18 ; its euclidean norm by @xmath19 ; and the euclidean inner product between the vectors @xmath17 and @xmath20 by @xmath21 . the set of real numbers is denoted @xmath22 and its @xmath23-fold cartesian product @xmath24 . the nonnegative reals are denoted @xmath25 , and the positive reals @xmath26 . the respective @xmath23-fold cartesean products are denoted @xmath27 and @xmath28 . we use @xmath29 to denote the indicator function : @xmath30 is equal to one if the statement is true and is equal to zero if it is false . throughout the paper @xmath31 denotes base-2 logarithm , and @xmath32 . the abbreviation iid stands for independently and identically distributed . our setting is illustrated in figure [ fig : problem - diagram ] and is specified by a tuple @xmath33 where @xmath34 are finite sets , @xmath35 is a probability distribution on @xmath36 ; @xmath37 and @xmath38 are nonnegative functions @xmath39 and @xmath40 and @xmath41 are nonnegative real numbers . the sets @xmath42 , @xmath43 , and @xmath44 model the source , side information , and reconstruction alphabets . a source sequence @xmath45 is observed at the encoder ( but not at the decoder ) and a side - information sequence @xmath46 at the decoder ( but not at the encoder ) . the sequence of pairs @xmath47 is assumed to be drawn iid according to the joint law @xmath35 . the encoder describes the source sequence @xmath48 to the decoder by an index rcl[eq : enc ] & = & ( x^n ) where @xmath49 is the encoding function and @xmath50 based on the index @xmath51 and its side information @xmath52 , the decoder forms a reconstruction sequence rcl[eq : decd ] _ ^n & = & ( , y^n ) where @xmath53 is the decoder s reconstruction function . the encoder s estimate of the decoder s reconstruction sequence is rcl[eq : dece ] _ e^n & = & ( x^n ) for some @xmath54 the goal of the communication is that the decoder s reconstruction @xmath3 matches the source sequence @xmath0 up to a distortion no larger than @xmath40 and the encoder s estimate @xmath4 matches the decoder s reconstruction @xmath3 up to a distortion no larger than @xmath41 . the distortions are measured by the bounded , nonnegative , single - letter distortion functions @xmath37 and @xmath38 . we say that a nonnegative triple @xmath55 is _ achievable _ if for every @xmath56 and sufficiently large @xmath8 there exists a message set of size @xmath57 as above such that the _ decoder - side reconstruction constraint _ @xmath58 \leq \dd + \epsilon\ ] ] and the _ encoder - side reconstruction constraint _ @xmath59 \leq \de+\epsilon\ ] ] are both met . our problem is not very interesting if the distortion constraints can not be met even when the source sequence is revealed losslessly to the reconstructor . consequently , we shall make the following assumption throughout : [ assumption ] the distortion functions @xmath60 and @xmath61 are such that for each @xmath62 there exist @xmath63 satisfying @xmath64 and @xmath65 . as we shall see , this assumption ensures that the triple @xmath66 is achievable whenever @xmath67 . we are interested in finding the smallest rate @xmath68 such that a given distortion pair @xmath69 is achievable . for given @xmath70 , let @xmath71 denote the set of rates @xmath72 such that the tuple @xmath73 is achievable : @xmath74 notice that by the assumption above , the set @xmath75 contains all rates @xmath67 and is thus nonempty . we can now define _ rate - distortions function _ as @xmath76 where the minimum exists because the set @xmath77 is nonempty , closed , and bounded from below by 0 . wyner and ziv s classic lossy source - coding problem with side information @xcite is similar to our problem except that wyner and ziv do not impose the encoder - side reconstruction constraint . informally , our problem thus reduces to the wyner - ziv problem if we set @xmath41 to infinity . wyner and ziv s result can be summarized as follows : [ thm : wz ] the rate - distortion function @xmath78 in the wyner - ziv setup is given by @xmath79 where @xmath80 , and where the minimization is over all functions @xmath81 and discrete random variable @xmath82 for which : @xmath82 takes values in an auxiliary alphabet @xmath83 of size at most @xmath84 ; @xmath85 forms a markov chain ; and rcl & & . since imposing the encoder - side reconstruction constraint can not increase the set of achievable rates , @xmath86 equality holds whenever the encoder - side reconstruction constraint does not pinch . for example , when @xmath87 ; @xmath88 ; and @xmath89 indeed , in this case the encoder can set @xmath90 to be @xmath91 . this results in being identical to and thus superfluous . steinberg s setup in @xcite is obtained from ours by replacing the encoder - side distortion constraint by the more stringent perfect - reconstruction constraint @xmath92 [ thm : steinberg ] the rate - distortion function @xmath93 in steinberg s setup is given by @xmath94 where the minimization is over all @xmath95 taking value in @xmath44 and satisfying @xmath96 and @xmath97 [ rem : steinberg ] constraint is equivalent to the block - distortion constraint @xmath98 thus , when in our setup @xmath99 is the hamming distortion and @xmath100 , then steinberg s setup differs from ours only in that is a block - distortion constraint whereas is an average - per - symbol distortion constraint . to describe the rate - distortions function for the setup of section [ subsec : setup ] , we introduce the function @xmath101 . the expression for @xmath101 in is similar to the expression for @xmath78 in except that in the expression for @xmath101 we have the additional constraint ; see ahead . given the joint law @xmath35 of the source and side information , and given the distortion functions @xmath102 , this function is defined as @xmath103 where the minimization is over all discrete random variables @xmath82 taking value in some finite auxiliary alphabet @xmath83 and forming the markov chain @xmath104 and over the functions @xmath105 and @xmath106 satisfying rcl & & [ eq : distdr0 ] + & & . [ eq : dister0 ] note that , thanks to assumption [ assumption ] , the feasible set in is not empty : we can choose @xmath82 as @xmath107 and @xmath108 , @xmath109 as the functions whose existence is guaranteed by the assumption . this choice demonstrates that @xmath110 using the convex cover method @xcite it can be shown that : allowing for sets @xmath83 of cardinality greater than @xmath111 does not decrease the value of the optimization problem . a consequence of this remark is that the minimum in is achieved : indeed , we may choose @xmath112 as the set latexmath:[$\{1 , \ldots , functions @xmath108 , @xmath109 , and the problem is reduced to minimizing a continuous function over a compact set . the key properties of @xmath101 are summarized in the following proposition : [ prop : key_properties ] the function @xmath114 is bounded from above by @xmath115 and is nondecreasing in the distortions @xmath116 moreover , it is convex and continuous . see appendix [ app : key_properties ] . our main result can be now stated as : [ thm : finite ] the rate - distortions function for the setup in section [ subsec : setup ] is equal to @xmath101 @xmath117 the coding scheme that establishes achievability is a variation on the coding scheme of wyner and ziv @xcite and is thus only sketched . its analysis is omitted . fix @xmath118 satisfying and , and fix also a blocklength @xmath8 and some ( small ) @xmath56 . let @xmath119 be a random blocklength-@xmath8 codebook with @xmath120 bins , each containing approximately @xmath121 codewords with the total number of codewords thus being @xmath122 . generate the codewords independently with the components of each codeword being drawn iid @xmath123 . number the bins @xmath124 through @xmath125 . upon observing the source sequence @xmath0 , the encoder seeks a codeword @xmath126 in @xmath127 that is jointly typical with @xmath0 . if successful , it sends the number of the bin containing @xmath126 as the message @xmath2 . it also produces the reconstruction sequence @xmath128 by applying the function @xmath129 componentwise to @xmath126 and @xmath0 . the decoder seeks a codeword @xmath130 in bin @xmath2 that is jointly typical with its side - information @xmath1 and applies the reconstruction function @xmath131 componentwise to @xmath130 and @xmath1 to produce @xmath132 . the converse is proved in subsection [ sec : th1 ] . though not identical , steinberg s setup is very similar to our setup when @xmath133 is the hamming distortion and @xmath134 is zero ( remark [ rem : steinberg ] ) . it is therefore not surprising that , as the following corollary shows , the two setups lead to identical rates : [ cor : steinberg ] let @xmath135 be arbitrary , and let @xmath99 be the hamming distortion function @xmath136 then @xmath137 see appendix [ app : rdcr ] . our results can be extended to a scenario where the encoder observes not only the source sequence @xmath138 but also some sequence @xmath139 which is correlated with the decoder s side - information sequence @xmath140 . this additional sequence @xmath139 makes it easier for the encoder to estimate the decoder s reconstruction sequence and thus allows the decoder to rely more heavily on its side information @xmath140 . to see how this seemingly more general scenario reduces to our scenario assume that @xmath141 are iid random triples of law @xmath142 and that @xmath143 takes value in the finite set @xmath144 consider now a new iid source @xmath145 taking value in the set @xmath146 according to the law @xmath147 with @xmath148 . the encoder now observes the source sequence @xmath145 only and no additional sequences . the decoder side information is still @xmath140 , and the joint law of @xmath149 is @xmath142 . finally define the new decoder distortion function @xmath150 as @xmath151 i.e. , the distortion function @xmath152 does not depend on the @xmath153-component . solving the original scenario for this new source and new decoder distortion function is equivalent to solving the seemingly more general problem we described . to establish the converse , we show that if a triple @xmath154 is achievable , then for every @xmath155 @xmath156 since @xmath101 is continuous ( proposition [ prop : key_properties ] ) , and since @xmath157 can be arbitrarily small , this implies that @xmath158 whenever @xmath55 is achievable , and consequently that @xmath159 . the first part of our proof identifying the auxiliary random variable @xmath160 and the function @xmath161 is similar to the proof of the wyner - ziv result @xcite . for a given blocklength-@xmath8 code @xmath162 , @xmath163 , @xmath164 satisfying , we have rcl + & & h ( ) [ eq:25 ] + & & i(x^n;|y^n ) + & & _ i = 1^n i(x_i;|y^n , x^i-1 ) + & = & _ i = 1^n h(x_i|y^n , x^i-1 ) - h(x_i|,y^n , x^i-1 ) + & & _ i = 1^n h(x_i|y_i ) - h(x_i|,y^n , x^i-1 ) + & & _ i = 1^n h(x_i|y_i ) - h(x_i|,y^n ) + & & _ i = 1^n h(x_i|y_i ) - h(x_i|z_i , y_i ) + & = & _ i = 1^n i(x_i;z_i|y_i ) + & & _ i = 1^n h(z_i|y_i ) - h(z_i|x_i ) + & = & _ i = 1^n i(x_i;z_i ) - i(y_i;z_i ) , [ equ : cv - part1-last ] where follows by ; follows because conditioning can not increase entropy and because @xmath165 ; follows from the chain rule for mutual information ; follows because the pair @xmath166 is independent of the tuple @xmath167 ; follows from the fact that conditioning can not increase entropy ; follows by defining @xmath168 and follows because with the definition above @xmath169 denote by @xmath170 the function that maps @xmath171 to the @xmath172-th component of the @xmath8-tuple @xmath173 , and denote by @xmath174 the function that maps @xmath0 to the @xmath172-th component of the @xmath8-tuple @xmath175 . since there is a one - to - one correspondence between the pairs @xmath176 and @xmath177 , we can define a function @xmath178 that maps @xmath176 to @xmath179 @xmath180 we now define @xmath181 where @xmath182 $ ] is with respect to @xmath183 . by definitions and , @xmath184 where @xmath182 $ ] is with respect to @xmath185 . we next turn to the encoder - side distortion . we will show that there exists a deterministic function @xmath186 that achieves a distortion no larger than @xmath187 , where @xmath187 is the distortion achieved by @xmath188 , namely , @xmath189 to this end , we express @xmath187 as rcl + & = & + & = & _ x^n , z_i + & = & _ x^n , z_i , [ equ : cv - exp - de - xnminusi ] where @xmath190 . for every @xmath191 , we define @xmath192 ( or for short @xmath193 ) as : is not unique , @xmath194 is defined as the first in lexicographical order . ] @xmath195{\de \bigl ( \deci(y_i , z_i),\ercni(x_i,{{x}_{\!\backslash \!i } } ) \bigr ) } \end{gathered}\ ] ] or in any other way that guarantees @xmath196 { \de\bigl(\deci(y_i , z_i),\ercni(x_i,{{x}_{\ ! \backslash\!i}})\bigr ) } \geq \\ \bige[y_i|x_i = x_i,{{x}_{\ ! \backslash\!i}}={{x}_{\!\backslash \!i}^*},z_i = z_i]{\de\bigl(\deci(y_i , z_i),\ercni(x_i,{{x}_{\!\backslash \!i}^*})\bigr)}.\end{gathered}\ ] ] we can now define the function @xmath197 as [ block : amos_gym ] rcl & & + ( x_i , z_i ) & & ( x_i , x_\i^*(x_i , z_i ) ) . for every @xmath198 , we have rcl + & & + & & + & & , [ equ : cv - cond - exp - de - erci ] where follows from the definition of @xmath193 ; follows because @xmath199 and follows from the definition of @xmath197 . it now follows from and that @xmath200{\de\bigl(\deci(y_i , z_i),\erci(x_i , z_i)\bigr ) } \leq \dei.\ ] ] continuing from ( [ equ : cv - part1-last ] ) we thus obtain rcl n(r+ ) & & _ i = 1^n i(x_i;z_i ) - i(y_i;z_i ) + & & _ i = 1^n ( , d_,i ) [ equ : cv - part2-first ] + & & n _ i = 1^n ( , d_,i ) + & & n ( _ i = 1^n , _ i = 1^n d_,i ) + & & n [ eq : amos300 ] where follows from the definition of @xmath101 and from , , and ; follows by multiplying by @xmath124 ; follows from the convexity of @xmath101 ( proposition [ prop : key_properties ] ) ; and follows from the monotonicity of @xmath101 ( proposition [ prop : key_properties ] ) and the fact that @xmath201 and @xmath202 . this establishes and thus concludes the proof of the converse . we next consider the case where the source , side information , and reconstruction alphabets @xmath34 are the reals @xmath203 ; the distortion functions @xmath60 and @xmath61 are quadratic rcl d_(x , _ ) & = & ( x-_)^2,[eq : dd ] + d _ ( _ , _ ) & = & ( _ -_)^2;[eq : de ] and the source and side - information pair @xmath204 is a centered bivariate gaussian , where @xmath107 is of variance @xmath205 @xmath206 and @xmath207 for some centered gaussian @xmath208 that is independent of @xmath107 and that is of variance @xmath209 and where @xmath210 is a nonzero constant . is zero , because in this case the side information is independent of the source and is thus irrelevant . ] the rate - distortions function depends on @xmath210 only through the ratio @xmath211 , because the receiver can premultiply its side information by @xmath212 without affecting the rate - distortions function . in the following we thus assume that @xmath213 , i.e. , @xmath214 we denote the rate - distortions function for this setup by @xmath215 . when @xmath216 is zero the problem is not interesting , because in this case the source sequence is determined by the side information , and @xmath215 is thus zero for all nonnegative values of @xmath217 and @xmath218 . we shall henceforth thus assume @xmath219 in this case , no finite rate can allow @xmath217 to be zero ( even if we ignore the encoder - side reconstruction constraint ) . thus , we shall also assume @xmath220 as we have seen in section [ sec : related ] , the wyner - ziv setup is obtained from ours if the encoder - side reconstruction constraint is omitted , and steinberg s common reconstruction setup is obtained if is replaced by . for a gaussian source and quadratic distortion measures , steinberg s common reconstruction rate - distortion function is @xcite @xmath221 the wyner - ziv rate - distortion function is @xcite @xmath222 this is the rate - distortion function even if the side information is revealed not only to the decoder but also to the encoder . [ thm : gaussian ] for a gaussian source and quadratic distortion measures , the rate - distortions function @xmath215 can be expressed as follows : if @xmath223 , then @xmath224 if @xmath225 , then lcl [ eq : ropga2 ] r^g(d _ , d _ ) & = & ^+ ( ) . the direct part is proved in section [ sec : achgauss ] and the converse in section [ sec : convgauss ] . if @xmath100 , then our rate - distortions function @xmath226 coincides with steinberg s common - reconstruction rate - distortion function @xmath227 of : @xmath228 [ rem : wzg ] if @xmath229 and @xmath134 are such that @xmath230 or @xmath231 then @xmath232 coincides with wyner and ziv s rate - distortion function @xmath233 in . thus , if or holds , then relaxing constraint and/or revealing the side information also to the encoder does not decrease the rate - distortions function . in the two cases that we shall describe in and ahead , no encoding is necessary because the encoder and the decoder can produce sufficiently good reconstructions @xmath128 and @xmath132 based solely on their observed sequences @xmath0 and @xmath1 . in these cases @xmath232 is thus zero . 1 . if + [ eq : d3 ] @xmath234 and @xmath235 + then the encoder and decoder can produce the sequences + rcl _ ^n & = & x^n + _ ^n & = & y^n + which satisfy the distortion constraints . 2 . if + [ eq : d4 ] @xmath236 and @xmath237 + then the encoder and decoder can produce the sequences + rcl[eq : xe1 ] _ ^n & = & x^n + _ ^n & = & y^n[eq : xd1 ] + which satisfy the distortion constraints . the achievability of theorem [ thm : gaussian ] in the remaining cases will be established using the following proposition with a judicious choice of the parameters . [ prop : gach ] for the setup in section [ sec : gausssetup ] of a gaussian source and quadratic distortion measures , the tuple @xmath238 is achievable whenever rcl r [ equ : gaussian - ac - diff - i - log ] for some parameters @xmath239 and @xmath240 satisfying [ block : amos_lunch ] @xmath241 and @xmath242 thus , @xmath243 where the minimization is over all @xmath239 and @xmath240 satisfying . see appendix [ sec : gscheme ] . we can now prove the achievability part of theorem [ thm : gaussian ] for the remaining cases . 1 . if + [ eq : d1 ] @xmath234 and @xmath244 + then the choice + [ eq : c1 ] @xmath245 ( which is positive by ) and @xmath246 + satisfies because @xmath247 and @xmath248 moreover , for this choice , + rcl[eq : rates2 ] + & = & . + thus , by and by proposition [ prop : gach ] , we conclude that when @xmath217 and @xmath218 satisfy , @xmath249 2 . if + [ eq : d2 ] @xmath250 and @xmath251 + then we consider the choice + [ eq : c2 ] @xmath252 + to see that the rhs of is positive note that implies that the denominator is positive , and implies that the numerator is positive because @xmath253 ( since @xmath254 is smaller than one , the lhs of implies that @xmath255 . this , and the fact that the lhs of also implies that @xmath256 demonstrates that the lhs of also implies that @xmath257 . ) + this choice satisfies because @xmath258 and @xmath259 moreover , for this choice , + rcl [ eq : requal ] + & = & . + thus , by and by proposition [ prop : gach ] , we conclude that when holds , @xmath260 the expressions in proposition [ prop : gach ] and their relation to become more transparent when we define [ eq : choice ] rcl z&=&a(x+w ) + _ & = & by+z + _ & = & bx+z for @xmath261 , @xmath262 , and @xmath263 a centered gaussian of positive variance @xmath264 independent of the pair @xmath204 . with these definitions [ eq : equivgpar ] rcl i(x;z|y ) & = & + & = & ( 1-a - b)^2 _ x^2 + a^2 _ w^2 + b^2 _ u^2 + & = & b^2 _ u^2 . since @xmath265 for all choices of the parameters @xmath261 , @xmath262 , @xmath266 , we can also rewrite as : @xmath267 where the minimum is over all @xmath268 that are of the form in and satisfy the distortion constraints rcl & & , + & & . if @xmath269 then the converse follows by relaxing the constraint ; see remark [ rem : wzg ] . we thus focus on the case where @xmath270 we define the function @xmath271 like @xmath272 except that its first argument ( @xmath217 ) is strictly positive ; the minimum is replaced by an infimum ; and the size of the auxiliary alphabet @xmath83 can be unbounded . thus , @xmath273 where the infimum is over all choices may take value , and we must restrict the functions @xmath131 and @xmath129 to be measurable . in the converse @xmath82 will correspond to the tuple @xmath274 , and we can therefore restrict @xmath82 here to be the space where such tuples take value . ] of the random variable @xmath82 and functions @xmath275 satisfying [ eq : amos600 ] rcl & & , [ equ : gaussian - cv - max - h - x - dec - constraint ] + & & , [ equ : gaussian - cv - max - h - x - enc - constraint ] + z & x & y , [ eq : markov ] where rcl _ & & ( y , z),[eq : phi ] + _ & & ( x , z).[eq : psi ] in analogy to proposition [ prop : key_properties ] we have : [ lem : gaus_mon_convex ] over @xmath276 the function @xmath277 is finite ; monotonic in each of its arguments ; and convex . the function is bounded by the rate - distortion function of the gaussian source without side information . the proof of monotonicity is identical to the proof of monotonicity in proposition [ prop : key_properties ] . the proof of convexity is also very similar ; only a minor change is needed to account for the fact that , _ prima facie _ , the infimum need not be achieved . the following lemma provides an explicit expression for @xmath277 when holds . [ lemma : amos_aldi100 ] if @xmath278 and @xmath279 satisfy , then @xmath280 we first prove @xmath281 to this end , we present a choice for @xmath82 , @xmath282 , @xmath283 that satisfies the constraints and is such that the objective function @xmath284 in evaluates to the rhs of . our choice depends on whether @xmath285 or @xmath286 in the first case the rhs of evaluates to 0 , whereas in the second case it is positive . when @xmath217 and @xmath218 satisfy , a suitable choice is as in and in the proof of the direct part @xmath287 when @xmath217 and @xmath218 satisfy , a suitable choice is as in and in the direct part rcl z&=&a(x+w),_=bx+z , _ = by+z , where @xmath263 is a centered gaussian of variance @xmath288 and independent of the pair @xmath204 and where @xmath289 and @xmath290 . that this choice has the desired properties follows by and . having established , we now complete the proof of the lemma by proving the reverse inequality @xmath291 since rates are nonnegative , it suffices to prove @xmath292 where @xmath293 has been replaced by @xmath294 . since the joint law of @xmath204 is fixed and is a bivariate gaussian law rcl i(x;z|y ) & = & h(x|y ) - h(x|y , z ) + & = & ( 2 e ) - h(x|y , z ) . [ equ : gaussian - cv - i - as - diff - log ] consequently , is equivalent to @xmath295 where @xmath296 is defined as @xmath297 under the same constraints that define @xmath277 in . to prove we first note that , since @xmath282 is a deterministic function of @xmath298 , rcl h(x|y , z ) & = & h(x - _ |y , z , _ ) + & = & h(x - _ |x - _ + u , z , _ ) + & & h(x - _ |x - _ + u ) [ equ : gaussian - cv - h - x - upper - bounded - by - h - xxd - cond ] [ equ : gaussian - cv - h - x - upper - bounded - by - log - var ] where in the second line we recalled that @xmath299 , and where the last line follows because conditioning can not increase differential entropy . the markov condition @xmath300 and the fact that @xmath299 imply that @xmath301 this , combined with the assumption that @xmath208 is independent of @xmath107 , implies that @xmath208 is independent of @xmath302 . and since @xmath283 is a function of @xmath302 , @xmath303 this independence implies that @xmath208 is independent of @xmath304 . this latter independence and the fact that @xmath305 can be expressed as @xmath306 implies that @xmath307 from , , the fact that the variance of a random variable can not exceed its second moment , and the fact that the magnitude of a correlation coefficient can not exceed @xmath124 , it follows that @xmath308 where @xmath309 is defined as @xmath310 subject to the relaxed constraints [ eq : amos650 ] @xmath311 we now proceed to study @xmath309 . define @xmath312 so @xmath313 subject to [ eq : aamos650 ] @xmath314 by the conditional max - entropy theorem @xcite , the supremum in is achieved when @xmath315 are jointly gaussian , as we henceforth assume . as we next argue , the lemma s hypothesis that holds implies that the choice of @xmath13 as @xmath316 is not in the feasible set . indeed , with this choice @xmath317 is equal to @xmath318 , which violates because and imply @xmath319 we thus assume in the following that @xmath13 is jointly gaussian with @xmath208 and that @xmath320 . consequently , rcl + & = & ( 2 e ( _ u^2 - ) ) [ eq : diffentropy ] + & = & ( 2 e ) where @xmath321 and @xmath322 . we can thus rewrite the optimization problem in as @xmath323 subject to @xmath324 ( we have to add the last constraint because the magnitude of a correlation coefficient can not exceed one . ) for fixed @xmath325 , the objective function in is monotonically increasing in @xmath326 ( see also ) , and so is the rhs of constraint . therefore , it is optimal to choose in @xmath327 substituting this choice in and yields @xmath328 subject to and @xmath329 notice that , whenever holds , the rhs of is upper - bounded by the square of @xmath330 . consequently , @xmath331 since the rhs of implies , @xmath332 and constraint is redundant . we therefore ignore constraint and study the maximization in subject to only . to this end , we compute the derivative of the objective function in with respect to @xmath325 : rcl + & = & .[eq : dev ] by , the derivative in is negative for all feasible @xmath325 . hence , the objective function in is decreasing on the ( symmetric ) interval of interest , and it is optimal to choose @xmath333 the optimality of this choice allows us to evaluate @xmath309 via and hence to upper - bound @xmath296 via . this yields the desired bound , which establishes the lemma . using lemma [ lem : gaus_mon_convex ] and lemma [ lemma : amos_aldi100 ] we can follow the steps of the proof in section [ sec : th1 ] of the converse part of theorem [ thm : finite ] . the remaining technicality is continuity . continuity in the interior , i.e. , on @xmath334 follows from convexity . it thus only remains to establish continuity when @xmath335 , holds , and @xmath134 is zero . this can be done by inspecting . so far we have only studied settings with two distortion functions , one of which the decoder - side distortion function @xmath336depends on the source symbol and the decoder s reconstruction , and the other the encoder - side distortion function @xmath337depends on the decoder s and the encoder s reconstruction symbols . in this section we extend our setting to allow for more than two distortion functions and to allow for distortions that depend on all three symbols : the source symbol @xmath338 , the decoder s reconstruction symbol @xmath339 , and the encoder s reconstruction symbol @xmath340 . we shall also allow the reconstruction alphabets to differ . but all alphabets are assumed finite . the new setup differs from the setup in section [ sec : setup ] in two ways . * the encoder - side reconstruction @xmath128 and the decoder - side reconstruction @xmath132 take value in the finite alphabets @xmath341 and @xmath342 which can be different . * there are @xmath343 ( possibly larger than @xmath344 ) distortion constraints specified by the @xmath343 distortion functions @xmath345 and the corresponding @xmath343 maximal - allowed distortions @xmath346 ( all of which are assumed to be nonnegative ) . we say that the tuple @xmath347 is _ achievable _ if for every @xmath56 and sufficiently large @xmath8 there exist a message set @xmath348 of size @xmath349 and functions [ eq : funk ] rcl f^(n ) & & ^n + ^(n ) & & ^n _ ^n + ^(n ) & & ^n ^n _ such that the message @xmath350 and the reconstruction sequences @xmath351 and @xmath352 satisfy : rcl [ eq : distck ] _ i=1^n d_k+ , k\{1 , , k}. + in analogy to assumption [ assumption ] , we shall assume : [ ass : extended ] to each @xmath353 corresponds some @xmath354 and some @xmath355 satisfying @xmath356 we seek the smallest rate @xmath68 for which the tuple @xmath357 is achievable . this is defined as follows . given a maximal - allowed - distortion tuple @xmath358 , let @xmath359 assumption [ ass : extended ] implies that the set @xmath360 contains all rates exceeding @xmath115 and is thus nonempty . the rate - distortions function @xmath361 can now be defined as @xmath362 where the minimum exists because the region @xmath363 is nonempty , closed , and bounded from below by 0 . to describe the rate - distortions function for the extended setup of section [ sec : problem_extended ] , we next introduce the function @xmath364 . given the joint law @xmath35 of the source and side information , and given the distortion functions @xmath365 , this function is defined as @xmath366 where the minimization is over all discrete auxiliary random variables @xmath82 and @xmath208 satisfying @xmath367 and over all functions @xmath368 and @xmath369 that simultaneously satisfy the @xmath343 distortion constraints rcl d_k , k\{1 , , k}. + [ eq : disterk ] the following proposition provides cardinality bounds on the support sets of the auxiliary random variables . [ prop : amoscardbound ] the minimum defining @xmath364 is not increased if we restrict the cardinality of the support set @xmath83 of @xmath82 to @xmath370 and the cardinality of the support set @xmath371 of @xmath208 to @xmath372 the cardinality bound on @xmath83 can be justified using the convex cover method @xcite . the cardinality bound on @xmath371 is proved in appendix [ app : amos ] . the cardinality bound on @xmath371 can be strengthened : @xmath373 need not exceed the number of distortion constraints in that depend on @xmath90 . the latter number equals @xmath124 in the original setup of section [ sec : setup ] thus allowing us to recover theorem [ thm : finite ] . [ prop : key_propertiesk ] the function @xmath374 is bounded from above by @xmath115 ; it is nondecreasing in the distortions @xmath375 and it is convex and continuous . the proof is similar to the proof of proposition [ prop : key_properties ] in appendix [ app : key_properties ] and is omitted . [ thm : extension ] the rate - distortions function for the setup in section [ sec : problem_extended ] is equal to @xmath376 : @xmath377 the achievability , i.e. , that @xmath378 can be proved using a scheme that is similar to the one that was sketched in the proof of theorem [ thm : finite ] . the only difference is that , to produce the reconstruction sequence @xmath128 , the encoder applies the function @xmath129 component - wise to the tuple @xmath379 , where , conditional on @xmath380 , the components of the sequence @xmath381 are generated independently according to the conditional law @xmath382 . the analysis of this scheme is omitted . we next prove the converse , i.e. , that @xmath383 fix some positive @xmath157 , a blocklength @xmath8 , and a rate @xmath68 . let @xmath348 be a message set of size @xmath384 , and let @xmath162 , @xmath385 , and @xmath386 be encoding and reconstruction functions as in that satisfy the @xmath343 distortion constraints in . for every @xmath387 , define @xmath388 in @xmath389 and define @xmath390 as @xmath391 notice that for every @xmath387 @xmath392 also , following the steps in , we can conclude that @xmath393 we further define as in section [ sec : th1]@xmath170 to be the function that maps @xmath171 to the @xmath172-th symbol of @xmath173 and @xmath174 to be the function that maps @xmath0 to the @xmath172-th symbol of @xmath175 . then , the symbol @xmath394 can be written as @xmath395 and @xmath396 can be written as @xmath397 for some functions @xmath161 and @xmath398 with arguments in the respective domains . we finally define for each @xmath399 and @xmath387 @xmath400 where @xmath182 $ ] is with respect to @xmath183 . notice that @xmath401 because the chosen encoding and reconstruction functions @xmath162 , @xmath385 , and @xmath386 satisfy . moreover , by definitions , @xmath402 where @xmath182 $ ] is with respect to @xmath403 . combining and with the definition of @xmath404 , we obtain rcl n(r+ ) & & _ i = 1^n i(x_i;z_i ) - i(y_i;z_i ) + & & _ i = 1^n _ ext(d_1,i , , d_k , i ) [ eq : rext1 ] + & & n _ ext ( _ i = 1^n d_1,i , , _ i = 1^n d_k , i ) + & & n _ ext ( d_1 + , , d_k+),[eq : fina ] where the last two inequalities follow by the convexity and the monotonicity of @xmath404 and by . by the continuity of @xmath404 and because @xmath56 and the blocklength @xmath8 are arbitrary , the converse follows immediately from . we acknowledge helpful discussions with prof . g. kramer . when @xmath99 is the hamming distortion and @xmath100 , our average - per - symbol distortion constraint is less stringent than the block - distortion constraint in steinberg s setup ( remark [ rem : steinberg ] ) . consequently , @xmath405 it remains to prove the reverse inequality . let @xmath82 , @xmath131 , and @xmath129 be minimizers of @xmath406 , so @xmath407 @xmath408 \leq \dd\ ] ] @xmath409 @xmath410 to prove the reverse inequality we shall upper - bound @xmath93 by showing that @xmath411 is feasible in the minimization that defines it . from the definition of @xmath95 and from , it follows that @xmath95 is computable ( almost surely ) from @xmath302 . this combines with to establish that @xmath412 and , _ a fortiori _ , that [ block : amos_ayef30 ] @xmath413 and by and , @xmath414 \leq \dd.\ ] ] it follows from that @xmath95 is feasible in the minimization defining @xmath93 and thus rcl r_cr(d _ ) & & i(x;)- i(y ; ) + & = & i(x;|y ) [ eq : amos_ayef30b ] + & & i(x;z|y ) [ eq : amos_ayef30c ] + & = & i(x;y)-i(x;z ) [ eq : amos_ayef30d ] + & = & r(d_,0 ) [ eq : amos_ayef30e ] where follows from ; where follows , by the ( conditional ) data processing inequality , from @xmath415 ( which holds by ) ; where follows from ; and follows from . inequalities and establish the corollary . that @xmath101 is bounded by @xmath115 is just a restatement of . monotonicity holds because the feasible set in the minimization defining @xmath101 is enlarged ( or is unaltered ) when @xmath40 and/or @xmath41 are increased . as to the convexity , let @xmath416 and @xmath417 be the random variables and functions that achieve the minima in the definitions of @xmath418 and @xmath419 . let @xmath420 be independent of @xmath421 . define @xmath422 and the functions @xmath423 @xmath424 then @xmath425 @xmath426}\quad \\ & = & \lambda e[d_{\mathrm{d}}(x,\phi^{(1)}(y , z^{(1 ) } ) ) ] \nonumber \\ & & + ( 1-\lambda ) e[d_{\mathrm{d}}(x,\phi^{(2)}(y , z^{(2 ) } ) ) ] \\ & \leq & \lambda d_{\mathrm{d}}^{(1 ) } + ( 1-\lambda ) d_{\mathrm{d}}^{(2 ) } ; \label{equ : cx - dd - leq - lambda}\end{aligned}\ ] ] and @xmath427 } \quad \\ & = & \lambda e[d_{\mathrm{e}}(\phi^{(1)}(y , z^{(1)}),\psi^{(1)}(x , z^{(1 ) } ) ) ] \nonumber \\ & & + ( 1-\lambda ) e[d_{\mathrm{e}}(\phi^{(2)}(y , z^{(2)}),\psi^{(2)}(x , z^{(2)}))]\ieeeeqnarraynumspace \\ & \leq & \lambda d_{\mathrm{e}}^{(1 ) } + ( 1-\lambda ) d_{\mathrm{e}}^{(2 ) } ; \label{equ : cx - de - leq - lambda}\end{aligned}\ ] ] so @xmath428 are feasible for the distortions @xmath429 consequently , rcl + & & i(x;z ) - i(y;z ) + & = & h(x ) - h(x|z ) - h(y ) + h(y|z ) + & = & h(x ) - h(x|z^(q),q ) - h(y ) + h(y|z^(q),q ) + & = & h(x ) - h(x|z^(1 ) ) - ( 1- ) h(x|z^(2 ) ) + & & - h(y ) + h(y|z^(1 ) ) + ( 1- ) h(y|z^(2 ) ) + & = & ( i(x;z^(1 ) ) - i(y;z^(1 ) ) ) + & & + ( 1- ) ( i(x;z^(2 ) ) - i(y;z^(2 ) ) ) . [ equ : cx - i - eq - lambda ] + & = & ( d_^(1),d_^(1 ) ) + ( 1- ) ( d_^(2),d_^(2 ) ) . to conclude the proof it remains to prove that @xmath101 is continuous on @xmath430 . ( continuity on @xmath431 is a consequence of the convexity , but we also claim continuity in the closed set @xmath431 . ) since @xmath430 is locally simplicial ( as can be verified by the definition in ( * ? ? ? * section 10 , p. 84 ) or using ( * ? ? ? * theorem 20.5 , p 184 ) ) , the convexity of @xmath101 on @xmath430 implies its upper - semicontinuity relative to @xmath430 . it thus remains to prove lower - semicontinuity relative to @xmath430 . that is , we need to show that @xmath432 implies that there is a subsequence @xmath433 such that @xmath434 let @xmath435 , @xmath436 , @xmath437 achieve @xmath438 with @xmath439 . since there are only a finite number of functions from @xmath440 to @xmath44 and only a finite number of functions from @xmath441 to @xmath44 , we can choose a subsequence @xmath433 along which : the mappings @xmath442 do not depend on @xmath443 and can be thus denoted @xmath108 ; the mappings @xmath444 do not depend on @xmath443 and can be thus denoted @xmath109 ; and the conditional laws @xmath445 converge to some conditional law that we denote @xmath446 . by the continuity of mutual information , @xmath447 converges to @xmath448 evaluated with respect to @xmath449 , and @xmath450 can not exceed this value because @xmath446 , @xmath109 , and @xmath108 are in the feasible set defining it . we present and analyze a scheme that achieves the rate - distortions tuples in proposition [ prop : gach ] . before describing the scheme , we introduce some notation and lemmas on @xmath8-dimensional spheres . an _ @xmath8-sphere of radius @xmath451 centered at @xmath452 _ is the set of all vectors @xmath453 satisfying @xmath454 when the center of the sphere @xmath455 is the origin @xmath456 , we call it a _ centered _ sphere , and when the radius of the sphere is @xmath124 , we call it a _ unit _ sphere . we denote the angle between two nonzero vectors @xmath457 by @xmath458 . its cosine is @xmath459 given a nonzero vector @xmath460 on an @xmath8-sphere @xmath461 , the _ spherical cap of half - angle @xmath462 centered at @xmath460 _ is the set of all vectors @xmath463 on @xmath461 satisfying @xmath464 the surface area of such a spherical cap does not depend on the vector @xmath465 but only on the dimension @xmath8 , the radius of the sphere @xmath466 , and the angle @xmath462 . if the radius @xmath467 , we denote this surface area by @xmath468 . we say that a random @xmath8-vector is uniformly distributed over an @xmath8-sphere , if it is drawn according to a uniform probability measure over the surface of this sphere . the proofs of the following four lemmas are based on results in @xcite and omitted . [ lem : cnuniform ] let @xmath469 be uniformly distributed over the centered unit @xmath8-sphere , and let @xmath470 be a deterministic unit - length vector in @xmath471 . then , @xmath472 [ lem : cnlim ] for @xmath473 : @xmath474 [ lem : ft ] let @xmath475 $ ] be such that the limit @xmath476 exists and @xmath477 . then , @xmath478 [ lem : lim0 ] for @xmath479 @xmath480 whereas for @xmath481 @xmath482 our scheme has parameters @xmath483 that must satisfy conditions and , which we repeat for convenience here : @xmath484 to describe and analyze the scheme we use vector notation . let @xmath485 denote the @xmath8-dimensional column - vector that results when the source symbols are stacked on top of each other @xmath486 likewise define the side - information vector @xmath487 and the reconstruction vectors @xmath488 , and @xmath489 . let rcl _ z^2 & & a^2 ( _ w^2 + _ x^2 ) , + r & & ( ) , + r & & ( ) . draw @xmath490 independent random @xmath8-vectors @xmath491 uniformly over the centered @xmath8-sphere of radius @xmath492 . assign these vectors to @xmath493 bins : the first @xmath494 are assigned to bin @xmath124 , the following @xmath494 vectors are assigned to bin @xmath344 , etc . more specifically , if @xmath495 denotes the set of vectors assigned to bin @xmath496 , then @xmath497 for @xmath498 and @xmath499 the codebook @xmath500 . given the source sequence @xmath501 , the encoder looks for the codeword @xmath502 that is closest to having the `` correct '' angle with @xmath503 : @xmath504 the encoder then sends @xmath505 , where @xmath506 denotes the index of the bin containing @xmath507 . it also produces the reconstruction sequence @xmath508 . given @xmath505 and the side - information vector @xmath509 , the decoder chooses @xmath510 and produces the reconstruction sequence @xmath511 . with probability 1 the @xmath512s in and are unique . we fix @xmath56 sufficiently small such that @xmath513 and define the following four events : 1 . @xmath514 `` the source and side information are atypical '' , i.e. , + rcl & |^2 - _ x^2| > _ x^2 + & |^2 - _ y^2| > _ y^2 + & | ( , ) - _ xy| > _ xy[eq : rxy ] + where @xmath515 denotes the correlation coefficient between @xmath107 and @xmath516 : + rcl _ xy=. 2 . @xmath517 `` no codeword has a good angle with the source sequence '' , i.e. , + rcl| ( , ) - | > . @xmath518 `` the chosen codeword @xmath519 does not have the correct angle with the side - information sequence '' , i.e. , + rcl| ( , ) - | > + [ eq : dec1 ] 4 . @xmath520 `` the decoder does not find the correct codeword '' , i.e. , @xmath521 also , we define the event @xmath522 [ lem : lime ] @xmath523 we note rcl[eq : sume ] [ ] & & [ _ ] + [ _ | _ ^c ] + [ _ 1 | _ ^c _ ^c ] + & & + [ _ 2 | _ ^c _ ^c ] . in the following we show that each term on the rhs of tends to zero as the blocklength @xmath8 tends to infinity . the first limit @xmath524 = 0\ ] ] follows directly from the weak law of large numbers . the second limit @xmath525=0\ ] ] can be shown following the same steps as in the proof of limit ( 134 ) in @xcite . the third limit @xmath526=0\ ] ] is proved as follows . we have @xmath527 where @xmath528 and @xmath529 denote the components of @xmath530 and @xmath531 that are orthogonal to @xmath485 : @xmath532 and @xmath533 let @xmath534 satisfy @xmath535\ ] ] and let @xmath463 and @xmath536 be vectors in @xmath471 satisfying rcl & |^2 - _ x^2| _ x^2 + & |^2 - _ y^2| _ y^2 _ y^2 + & | ( , ) - _ xy| _ xy . [ eq : rhoxy ] then , conditional on events @xmath537 by and , we have [ eq : product ] rcl ( , ) ( , ^ * ) & & ( 1 + ) _ xy(1 + ) + & & ( 1 + 3 ) and rcl ( , ) ( , ^ * ) & & ( 1- ) _ xy(1- ) + & & ( 1 - 3 ) , + where inequalities @xmath538 follow because @xmath539 and because @xmath540 . moreover , conditional on the events in , the vector @xmath529 is uniformly distributed over a centered @xmath541-dimensional sphere of radius @xmath542 , and thus limit on top of the next page follows by lemmas [ lem : cnuniform ] and [ lem : lim0 ] . we can combine limit and inequalities to obtain the limit on top of the next page . @xmath543 = 1\ ] ] @xmath544 = 1\ ] ] ' '' '' if in we take the expectation with respect to @xmath545 , and @xmath546 ( but keep the conditioning on events @xmath547 and @xmath548 ) , we obtain the desired third limit . we finally prove the fourth limit @xmath549=0.\ ] ] to this end , we define event @xmath550 as @xmath551 recalling the decoding rule in and the definition of event @xmath552 in , we see that when @xmath553 and @xmath550 occur simultaneously , then by condition the decoder finds the correct codeword @xmath554 . therefore , rcl[eq : sumu ] & & 1 - , and thus and the limit @xmath555 establish . we now prove . for each @xmath556 , we index the vectors in the @xmath557-th bin from @xmath124 to @xmath558 and we shall refer to the @xmath559-th vector in this @xmath557-th bin by @xmath560 . let @xmath561 be the index of @xmath519 , i.e. , @xmath562 . by the symmetry of the code construction and the encoding rule , the probability @xmath563 does not depend on the values @xmath557 and @xmath559 . we therefore , assume in the following that @xmath564 and @xmath565 . if we additionally condition on @xmath566 and on @xmath567 , the vectors @xmath568 ( i.e. , the vectors in bin @xmath124 that are not @xmath519 ) are independent and uniformly distributed over the centered @xmath8-sphere of radius @xmath569 without the spherical cap of half - angle @xmath570 centered at @xmath463 . thus , @xmath571 is an upper bound on the conditional density of the normalized vectors @xmath572 on the centered unit @xmath8-sphere . applying lemma [ lem : cnuniform ] , we therefore obtain inequality shown on top of the next page . rcl[eq : in1 ] + & = & 1 - _ k=2^|(1)| ( 1 - ) + & & < 1 - ( 1 - ) ^|(1)|-1 [ equ : gaussian - ac - er - other - closer - cond - pr - others - close ] + & & 1 - ( 1- ) ^2^n ( r - r- ) [ equ : gaussian - ac - er - other - closer - pr - others - close - bound ] ' '' '' we note that for any @xmath573 $ ] @xmath574 and hence the mapping @xmath575 is decreasing in @xmath576 . therefore , since @xmath577 we further obtain . if now we take the expectation with respect to @xmath578 , @xmath2 , and @xmath561 ( but keep the conditioning on @xmath579 and @xmath580 ) , results in @xmath581 \nonumber\\ & < 1 - \left(1-\frac{2 \ , c_n\left(\arccos\sqrt{1 - 2^{-2 ( r ' - r-\frac{\delta}{2})}}\right)}{c_n(\pi ) } \right)^{2^{n ( r ' - r-\delta)}}\!\!. \label{equ : gaussian - ac - er - other - closer - pr - others - close - bound2 } \end{aligned}\ ] ] the desired limit ( [ equ : gaussian - ac - er - other - closer - pr - others - close ] ) follows by and by lemma [ lem : ft ] . in fact , applying lemma [ lem : ft ] to @xmath582 and to the function @xmath583 we obtain that the right - hand side of tends to 1 as @xmath8 tends to infinity because rcl[equ : gaussian - ac - er - other - closer - lim - log - cn ] _ 1&&- _ n ( ) + & = & r - r-/2 + & > & _ 2 . here , the equality holds by lemma [ lem : cnlim ] and because the factor 2 in the logarithm does not change the limit , and the inequality holds by and because @xmath584 . this concludes the proof of limit ( [ equ : gaussian - ac - er - other - closer - pr - others - close ] ) and thus of the fourth limit . combining finally with and establishes the proof of the lemma . we can now bound the expected distortions of our scheme . we have rcl[equ : gaussian - ac - exp - dd ] & = & + & & + , and rcl[equ : gaussian - ac - exp - de ] & = & [ ^c ] + & & + [ ] . the decoder - side distortion satisfies rcl d_^(n ) ( , _ ) & = & - ^ * - b ^2 + & & ^2 + ^ * ^2 + ^2,[eq : ddrough ] where the inequality holds by the cauchy - schwarz inequality and because an arithmetic mean of two nonnegative numbers can not be smaller than it s geometric mean . therefore , rcl + & & + & = & + & & - + & & 3(_x^2+_z^2 + b^2 ( _ x^2+_u^2 ) ) + & & - 3 ( _ x^2(1- ) + _ z^2 + b^2 ( _ x^2+_u^2)(1- ) ) + & & 3(_x^2+_z^2 + b^2 ( _ x^2+_u^2 ) ) ( 1- ( 1- ) ) .[eq : dde ] in the event @xmath585 , we can derive a bound on the decoder - side distortion @xmath586 that is tighter than : @xmath587 where the first inequality follows from the definition of the event @xmath585 , the second by throwing away some negative @xmath157-terms , and the third from condition and because @xmath588 . since @xmath589 \leq 1 $ ] , we thus have : rcl[eq : ddec ] + & & d_d + ( _ x^2 + b^2 ( _ x^2 + _ u^2 ) + 4 a _ x^2 + 8 b _ x^2 + 18 ab _ x^2 ) . + combining , , and , we obtain rcl + & & d_+ 3(_x^2 + _ z^2 + b^2_y^2)(1- ( 1 + ) ) + & & + ( _ x^2 + b^2 ( _ x^2 + _ u^2 ) + 4 a _ x^2 + 8 b _ x^2 + 18 ab _ + [ eq : concdd ] similarly , we have for the encoder - side distortion : rcl d_^(n ) ( , _ ) & = & -b ^2 + & & ^2 + b^2 ^2 , [ eq : derough ] and thus , rcl + & & + & & - + & & 2(b^2 ( _ x^2+_u^2)+ b^2 _ x^2 ) ( 1- ( 1- ) ) .[eq : dee ] moreover , in the event @xmath585 we can derive a bound on the encoder - side distortion @xmath590 that is tighter than : @xmath591 where the last inequality follows by assumption and because @xmath588 . since @xmath592\leq 1 $ ] , we thus have rcl[eq : deec ] [ ^c ] & & d _ + b^2 ( 9_x^2 + _ u^2 ) . combining finally , , and , we obtain rcl + & & d _ + 2(b^2 _ y^2 + b^2 _ x^2 ) ( 1- ( 1- ) ) + & & + b^2 ( 9_x^2 + _ [ eq : concde ] recall that the rate of our scheme is smaller than @xmath593 and that @xmath594 can be chosen arbitrarily close to 0 . therefore , from , , and lemma [ lem : lime ] we conclude that when @xmath595 and @xmath596 satisfy and , then our scheme can achieve the triple @xmath597 this establishes proposition [ prop : gach ] . to prove the cardinality bound on @xmath371 , we shall need the following variation on carathodory s theorem . let @xmath599 be a compact subset of @xmath24 , and let @xmath463 be a boundary point of its convex hull @xmath600 . since @xmath463 is in the convex hull of @xmath599 , it follows from carathodory s theorem that there exist @xmath601 or fewer points @xmath602 and positive coefficients summing to @xmath124 @xmath603 such that @xmath604 we shall show that , in fact , of these @xmath443 points , we can find @xmath23 or fewer points whose convex combination is @xmath463 . so @xmath609 we shall next show that the points @xmath610 are in @xmath605 . to that end we note that by @xmath611 where the second equality holds because the @xmath612 s sum to @xmath124 . since the @xmath612 s are all positive , it follows from that all the terms on the rhs are nonnegative . since they sum to zero , they must all be zero . and since the @xmath612 s are positive , we conclude that @xmath613 and the vectors @xmath614 are all in @xmath605 . the vector @xmath463 can thus be written as a convex combination of the @xmath443 vectors in @xmath615 in @xmath605 . since @xmath605 is @xmath616-dimensional , it follows from carathodory s theorem that @xmath463 is in fact a convex combination of @xmath23 or fewer of the vectors @xmath615 . let the discrete random variables @xmath208 and @xmath82 over the alphabets @xmath371 and @xmath112 , the function @xmath617 , and the function @xmath618 satisfy and . we shall exhibit a random variable @xmath619 over the alphabet @xmath620 and a function @xmath621 satisfying @xmath622 and the @xmath343 distortion constraints @xmath623 since the markov conditions and imply @xmath624 this will allow us to replace @xmath208 and @xmath129 with @xmath619 and @xmath625 and thus conclude the proof . to describe @xmath619 and @xmath625 , we need some definitions . for each pair @xmath626 and each @xmath399 , define @xmath627 } \nonumber \\ & = \e { d_k \bigl ( x , \fdc(y , z),\psi(x , z , u ) \bigr ) } , \label{eq : dkxz}\end{aligned}\ ] ] where the expectation is , by , with respect to @xmath628 . define also the vector - valued function where the expectation is with respect to @xmath629 . let @xmath630 denote the image of @xmath631 : @xmath632 by definitions @xmath633 and , consequently , there exists a point @xmath634 on the boundary of conv@xmath635 with @xmath636 since @xmath630 is compact ( it contains at most @xmath637 points because @xmath638 depends on @xmath639 only via @xmath640 ) , lemma [ lem : card ] implies that @xmath641 can be written as a convex combination of @xmath343 or fewer points in @xmath630 : @xmath642 where @xmath643 and the coefficients @xmath644 $ ] sum to 1 . let @xmath645 be preimages of @xmath646 so @xmath647 we can now define the function @xmath625 as mapping every pair @xmath648 and every @xmath649 to @xmath650 and we define the random variable @xmath619 to be conditionally independent of @xmath516 given @xmath302 with the conditional law @xmath651 the markov condition thus holds by definition . moreover , , , and combine to prove that @xmath619 and @xmath625 also satisfy the @xmath343 distortion constraints in : denoting the @xmath559-th component of the vector @xmath652 by @xmath653 , for @xmath654 , k. kittichokechai , t. j. oechtering , and m. skoglund , `` source coding with common reconstruction and action - dependent side information , '' in _ proceedings 2010 ieee information theory workshop _ , dublin , ireland , 30 aug.3 sep . 2010 , pp . r. timo , a. grant , and g. kramer , `` rate distortion functions for source coding with complementary side information , '' in _ proceedings ieee international symposium on information theory _ , st . petersburg , russia , 31 july5 aug . 2011 . b. ahmadi , r. tandon , o. simeone , and h. v. poor , `` on the heegard - berger problem with common reconstruction constraints , '' in _ proceedings ieee international symposium on information theory _ , mit , cambridge , ma , 2012 . b. ahmadi , t. r. , s. o. , and p. h. v. , `` heegard - berger and cascade source coding problems with common reconstruction constraints , '' _ submitted to ieee transactions on information theory _ [ online ] . available : http://arxiv:1112.1762v4

self - organized criticality may be the underlying cause of power laws in a wide range of natural and man - made phenomena @xcite . systems exhibiting self - organized criticality naturally approach a critical state , without any intrinsic time or distance scales . the critical point is reached without any fine - tuning of parameters . this is unlike most critical points seen in physics for example , the ising model is only critical at a single , very specific , temperature . this lack of fine - tuning is essential if we are to understand power laws in nature , where no fine - tuning is possible . since the concept of self - organized criticality was first introduced by bak , tang , and wiesenfeld in 1987 , a number of models have been developed to investigate this phenomenon @xcite . however , the original model , the two - dimensional , isotropic , abelian sandpile model ( asm ) , is still one of the simplest and most interesting of the models . the asm is simple and robust , which are necessary features for any model of self - organized criticality . while natural phenomena are quite complex , any model that seeks to explain the ubiquity of power laws in nature must , paradoxically , be very simple ; if we are to have a robust model for the generation of power laws , we must neither have finely - tuned parameters , nor finely - tuned rules . the asm is defined on a lattice of sites , and is described by a toppling matrix @xmath1 , whose dimension is equal to the number of sites in the sandpile . the sandpile evolves stochastically . in each time step , a grain of sand is added to a random site . then , sites are checked for stability . if the number of grains at a site @xmath2 is greater than @xmath3 , then the site @xmath2 is unstable , and topples , losing @xmath4 grains , while every other site @xmath5 gains -@xmath6 grains . ( generally , @xmath7 is zero except when @xmath5 neighbors @xmath2 . ) typically , models are conservative , which means that each toppling in the bulk conserves the total number of grains ( @xmath8 ) . only for topplings along the boundary , where grains can fall off the edge , can the total number of grains change . we continue toppling unstable sites until no sites are unstable . then , we begin a new time step , and again add a grain to a random site . the asm is surprisingly tractable @xcite . we only briefly cover some of the essential points here for comprehensive reviews , see @xcite . after a large number of time steps , the asm reaches a well - defined distribution of states . of the stable height configurations , some are transient , and occur with probability zero after a long amount of time . all other states are recurrent , and occur with equal probability . dhar showed that the total number of recurrent states is just @xmath9 @xcite . this is also equal to the number of spanning trees that can be drawn on the lattice , showing a connection between the sandpile and spanning tree problems @xcite . these statements hold for all asms , which define a large class of models . now , we specialize to the two - dimensional , conservative , isotropic asm , which is defined on a two - dimensional square lattice , where each site has a maximum height of four , and where upon toppling at any site , one grain is sent to each of the site s four neighbors . furthermore , we work in the limit where the lattice is infinite . the two - dimensional , isotropic , spanning tree problem is equivalent to the central charge -2 logarithmic conformal field theory ( @xmath10 lcft ) @xcite , which has the simple gaussian action @xmath11 , where @xmath12 and @xmath13 are complex grassman variables . the @xmath10 lcft is described in @xcite . while the two - dimensional , conservative , isotropic asm is just one of many possible asms , it is the original , standard , model @xcite , and it is reasonable to simply refer to it as `` the asm , '' which we do for the remainder of this paper . calculations of correlation functions , using methods to be described in the next section , have confirmed that there is a relationship between the asm and @xmath10 lcft . two - point correlation functions of unit height variables decay as @xmath14 in the bulk @xcite , as do all two - point height correlations along open and closed boundaries @xcite ; these correlations can be understood as equivalent to correlations of lcft operators . furthermore , calculations of certain three - point correlation functions of heights along closed boundaries , and all multipoint correlations of heights along open boundaries , have allowed lcft field identifications for heights along boundaries @xcite . the asm is not robust to all perturbations . if we relax the constraint that the model be conservative , and instead allow grains to be lost in any bulk toppling ( i.e. allow dissipation ) , correlations decay exponentially , and we are taken off the critical point @xcite . the condition of conservation can be considered a `` natural '' one , rather than one requiring `` fine - tuning . '' deeper probes of the the conformal structure can be obtained by looking at correlations both on and off the critical point . mahieu and ruelle calculated a number of off - critical correlation functions of certain height configurations , known as weakly allowed clusters ( wacs ) , and used their correlation functions to propose field identifications for the 14 simplest wacs @xcite . they found that their correlation functions could be explained by assuming that all 14 wacs took the form @xmath15 the coefficients @xmath16 , @xmath17 , @xmath18 , @xmath19 , and @xmath20 vary from wac to wac . @xmath20 is the probability for the cluster at the critical point , and @xmath21 is the mass , a measure of how far the model is from the critical point . the correlation functions that they used were mostly two - point functions along horizontal or diagonal axes , as well as some three - point and four - point functions for the two simplest wacs . while these calculations provide strong evidence for the identification of the asm with the @xmath10 lcft , and the field identification in eq . ( [ eq : lcft.rep ] ) , the fact that only specific correlations were considered limits the range of the identification . it would be surprising if new orientations of correlation functions , or new wacs , were found to be inconsistent with eq . ( [ eq : lcft.rep ] ) ; but the calculations in @xcite do not rule this possibility out . more importantly , since each correlation function in @xcite required a new and separate calculation , it is hard to understand , mathematically , why these results occured . while their end results showed that certain correlations of wacs in the asm are equal to correlations of eq . ( [ eq : lcft.rep ] ) in the lcft , it was not mathematically transparent as to why this should be . nor was it clear why , or if , the same coefficients would appear in other properties , such as off - boundary correlations , or correlations with defects . here , we calculate all correlation functions of all local bond modifications ( lbms ) , for arbitrary numbers and types of lbms at arbitrary positions ( far from one another ) ; our calculations confirm that all lbms should receive the field identification in eq . ( [ eq : lcft.rep ] ) . by lbms , we mean any set of local changes in the sandpile toppling rules . for the asm at the critical point , we will assume that the lbms are conservative ( do not create or destroy grains ) . since all wacs can be calculated by lbms , our results automatically include all correlations of wacs . while the wacs are the most important types of lbms , and the easiest to find probabilities of , in numerical simulations , we generally discuss our results in terms of lbms , to emphasize the generality of our results . we give closed form expressions for @xmath16 , @xmath17 , @xmath18 , and @xmath22 and describe methods that allow rapid calculation of these coefficients . while a computer is needed for the calculation of specific @xmath16 , @xmath17 , and @xmath18 coefficients , the general calculations can be done by hand . by showing how calculation for all lbms can be done at once , we make the mathematical structure clearer . for example , we can quickly see why the coefficients @xmath16 , @xmath17 , and @xmath18 appear in other properties . we illustrate this by looking at off - boundary lbm probabilities , and correlations with bond defects ( either in the bulk , or along a boundary ) . interestingly , we find that some bond defects are represented by a lcft operator with a scaling dimension of four ; this is , to the best of our knowledge , the first dimension four operator found in the asm . while our calculations have been done in both the normal asm , and the asm with dissipation , we focus our discussion on the simpler analysis at the critical point , and only discuss the more complicated massive correlations in the last section , and in the appendices . the methods used in this paper are not powerful enough to calculate probabilities and correlations for any height configurations . even the calculation of the probability for a site to have height two requires much more complicated methods @xcite , and the correlation function of two height two variables remains unknown . this is because the condition for a site to have height two involves a nonlocal condition . as already stated , the most important lbms are those used to calculate properties of wacs @xcite . wacs are related to forbidden subconfigurations ( fscs ) . an fsc is a height configuration over a subset of sites @xmath23 , such that for every site @xmath24 , the number of neighbors of @xmath2 in @xmath23 is greater than or equal to the height at @xmath2 . fscs are important because asm height configurations are recurrent if and only if they have no fscs @xcite . a wac is a height configuration that contains no fscs , but becomes a fsc if any height in the wac is decreased by one . three wacs are shown in the left side of figure [ fig : wacs ] . wacs are analytically tractable because it turns out that the number of sandpiles with a particular wac is equal to the number of recurrent states in a sandpile with modified toppling rules @xcite . there are actually several different ways to modify the toppling rules to obtain the wac probability . the simplest is to , for each connected piece of the wac , remove all but one of the bonds connecting it to the rest of the lattice the modified lattices corresponding to the wacs are shown in the right side of figure [ fig : wacs ] . ( see @xcite for a discussion of other ways in which the sandpile can be modified to obtain the wac probabilities . ) in these modified sandpiles , grains of sand can not flow along the removed bonds ; to continue to conserve the number of grains during each toppling , the condition for instability must be decreased at the sites at the end of the removed bond . these changes result in a new toppling matrix @xmath25 . as already stated , the number of recurrent states in the asm is @xmath9 . the number of recurrent states that have the wac is given by @xmath26 . ( we discuss this equivalence further in section [ sec : wac.lbm.issues ] . ) so the bulk probability for the wac is given by @xmath27 where we have defined @xmath28 and @xmath29 . @xmath30 is the well - studied lattice green function ( at the critical point ) ; exact expressions are known for the green function between nearby sites , and asymptotic expressions for the green function between distant sites @xcite . while @xmath1 , @xmath31 , and @xmath30 all have large dimensions ( equal to the number of sites ) , @xmath32 is zero outside of a finite collection of sites . when the bond between @xmath2 and @xmath5 is removed , @xmath33 and @xmath34 are both increased by 1 , while @xmath35 and @xmath36 are both decreased by 1 . for example , for the unit height probability , we have @xmath37 here @xmath2 is the site fixed at height one , while @xmath38 , @xmath39 , and @xmath40 are the three sites that @xmath2 has been disconnected from . for any wac , the fact that @xmath32 is finite - dimensional means that the height probability can be found by calculating a simple , finite - dimensional , matrix determinant . all wacs thus correspond to lbms . however , many lbms do not correspond to wacs . lbms are simply any sandpile modifications that can be modeled with a @xmath32 matrix that is conservative ( every row and column sums to zero ) and symmetric . our analysis gives all correlations of lbms , which thus automatically gives all correlations of wacs . for an @xmath41-point correlation function of lbms , we can still use this method . the only difference is that the removed bonds are located in @xmath41 distant clusters ; this is illustrated in figure [ fig : wac.correlation ] . removal of bonds in this fashion will give @xmath42 and @xmath43 block matrix structures . for example , for a 3-point function , we will have @xmath44 @xmath45 is the modification to the toppling matrix for the set of bonds removed about the @xmath46 lbm . @xmath47 is the green function matrix between sites of the @xmath46 lbm , and its elements are @xmath48 . @xmath49 , @xmath50 is the green function matrix between the sites of the @xmath46 and @xmath51 lbms , and its elements are given by the bulk green function , @xmath52 where @xmath53 is the euler - mascheroni constant @xcite . we work in the limit where the lbms are all very far from each other we assume that any two of the @xmath41 lbms are the same order - of - magnitude , @xmath54 , apart . since the green function diverges as @xmath55 with increasing @xmath56 , calculation of @xmath57 initially looks very difficult . however , every row of every @xmath45 sums to zero this follows from the manner in which we constructed @xmath45 , and reflects the fact that grains of sand are still conserved in each toppling in the bulk of the modified sandpile . this implies that parts of @xmath49 that depend only on the column index make no contribution to @xmath58 , and thus no contribution to the correlation function , @xmath57 . so we only care about differences ( discrete derivatives ) of green functions between columns of @xmath49 , and the elements of @xmath49 are effectively @xmath59 , rather than @xmath60 . for lbms , every @xmath45 is symmetric , so every column of every @xmath45 sums to zero . using the matrix identity @xmath61 , this in turn means that the parts of @xmath49 that depend only on the row index make no contribution to the probability . this is , in effect , like taking another discrete derivative of the green function , so that the elements of @xmath49 are effectively @xmath62 . to make this concrete , suppose that the local origin of the @xmath46 lbm is located at @xmath63 , and the local origin of the @xmath51 lbm is located at @xmath64 . the @xmath46 lbm covers a set of sites at locations @xmath65 , relative to @xmath63 , and the @xmath51 lbm consists of a series of sites at locations @xmath66 , relative to @xmath67 . ( @xmath68 , @xmath69 , @xmath70 , and @xmath71 are all @xmath48 . ) then , the elements of @xmath49 all have the form @xmath72 . the last two paragraphs show that we only need the parts of @xmath49 that depend on _ both _ the row _ and _ column indices . that is , we only need the parts of the green function that depend on _ both _ @xmath65 _ and _ @xmath66 , and can drop all other terms . expanding eq . ( [ eq : bulk.green ] ) in powers of @xmath73 , we find that the lowest - order term not dropped is @xmath74 the green function can thus be treated as @xmath62 for correlations of lbms . for more general local arrow diagrams , such as those that appear in the calculations involving the height two variable @xcite , the @xmath42 matrices are not symmetric , and we can no longer drop the parts of @xmath49 that depend only on the row index . to get the connected @xmath41-point function from @xmath57 , we need to pick at least one element off the block diagonal in every block row of @xmath30 , and in every block column of @xmath30 . this means , at the minimum , picking @xmath41 elements off the block diagonal of @xmath30 , resulting in a leading - order contribution to the correlation function of @xmath75this is the universal part of the correlation function . mahieu and ruelle showed that for two - point functions , the constraint of picking only two elements off the block diagonal allows the correlation function to be written as @xcite @xmath76 similarly , they found that the leading - order contribution to the three - point probability is @xmath77 more generally , for an @xmath41-point correlation , if only @xmath41 terms are picked off the block diagonal , then the connected correlation function is given by @xmath78 \sum_s { { \rm trace}}\left\ { \prod_{x=1}^n \left [ { { { \mathbb i}}\over{{\mathbb i}+{\bf b}_{u_{s(x ) } } { \bf g}_{u_{s(x)}u_{s(x ) } } } } { \bf b}_{u_{s(x ) } } { \bf g}_{u_{s(x)}u_{s(x+1 ) } } \right ] \right\ } \ , \label{eq : general.trace.formula}\end{aligned}\ ] ] where @xmath79 is summed over all one - to - one mappings from @xmath80 to @xmath80 , and we identify @xmath81 with @xmath82 . mahieu and ruelle wrote eqs . ( [ eq : mr.trace.2]-[eq : mr.trace.3 ] ) in different , but equivalent , forms . @xmath83 is the bulk probability of the @xmath46 lbm . ( note that the two trace terms of eq . ( [ eq : mr.trace.3 ] ) are actually equal . we have written the three - point function in this form to make clear how the form generalizes for @xmath41-point functions . ) we can rewrite eq . ( [ eq : green.r2.expansion ] ) as @xmath84 @xmath85 is the column vector of the horizontal positions of the sites of the @xmath46 lbm , relative to the @xmath46 local origin ( i.e. , the elements of @xmath85 are @xmath48 ) . @xmath86 is the corresponding vector of vertical positions . @xmath85 and @xmath86 are both length @xmath87 , where @xmath87 is the number of sites needed to represent the @xmath46 lbm with the methods of the previous section ( e.g. @xmath88 for the unit height variable ) . we insert eq . ( [ eq : g.decompose ] ) into eq . ( [ eq : general.trace.formula ] ) . for each of the @xmath41 @xmath49 s , we can pick any of the four matrices of eq . ( [ eq : g.decompose ] ) , resulting in @xmath89 terms . in each of these @xmath89 terms , each @xmath49 has been replaced with the product of a column vector and a row vector . using the cyclicity of the trace to move one row vector at the end of the trace to the start of the trace , we see that each matrix @xmath90 is bracketed by a row vector to its left , and a column vector to its right , producing a @xmath91 matrix . so each of the @xmath89 terms is the product of @xmath41 numbers . we can represent the decisions as to which terms of eq . ( [ eq : g.decompose ] ) to pick by representing @xmath49 with a @xmath92 matrix , @xmath93 . the possible ways to bracket @xmath90 can be represented with a @xmath92 matrix , @xmath94 . we have @xmath95 we have defined @xmath96 then , the correlation function of @xmath41 lbms is given by @xmath97 where the other trace terms are derived by permutations of @xmath98 , as in eq . ( [ eq : general.trace.formula ] ) . we can compare this to correlation functions of fields in the @xmath10 lcft . mahieu and ruelle proposed that the wacs are represented , at the critical point , by @xmath99 ( the `` c '' term in eq . ( [ eq : lcft.rep ] ) only appears off the critical point . ) we can compute connected @xmath41-point correlations of these fields in the @xmath10 lcft . we use the formulation of the @xmath10 lcft where the action is @xmath100 where we do nt integrate over zero modes in expectation values . since the theory is gaussian , to calculate correlation functions we simply need to take wick contractions . the relevant nonzero ones are @xmath101 each term of eq . ( [ eq : lcft.rep.critical ] ) has one @xmath12 , and one @xmath13 . the only difference between terms is whether the derivative on the @xmath12 is holomorphic or antiholomorhic , and whether the derivative on the @xmath13 is holomorphic or antiholomorphic . we can use a @xmath92 matrix to represent the choice of which terms of @xmath102 are picked : @xmath103 the contractions of eqs . ( [ eq : crit.contraction.1]-[eq : crit.contraction.2 ] ) can then be represented with the matrix @xmath104 the contribution to the correlation function where the @xmath13 of the 1st lbm contracts with the @xmath12 of the 2nd lbm , the @xmath13 from the 2nd lbm contracts with the @xmath12 of the 3rd lbm , and so on , is @xmath105 other contractions give other permutations , just as in eq . ( [ eq : npt.trace.form ] ) . finally , @xmath106 differs from @xmath107 only by a matrix rotation , which will not affect the trace , if we make the following identifications : @xmath108 so the traces in eq . ( [ eq : npt.trace.form ] ) and eq . ( [ eq : lcft.critical.trace ] ) are equal , and all lbms are indeed represented by the field in eq . ( [ eq : lcft.rep.critical ] ) . these formulas for the coefficients have the appropriate transformation properties under @xmath109 rotations , and x and y reflections . ( technically , the overall sign of eq . ( [ eq : a ] ) is still undetermined at this point , since all correlation functions have even numbers of @xmath16 s . to determine the signs of the @xmath16 s we need to look at at least one massive correlation function . we can do this by consulting the massive three - point function of the unit height variable in @xcite , or more broadly , by looking at the general massive correlations in section [ sec : dissipation.general ] . ) @xmath16 , @xmath17 , and @xmath18 , can be calculated on a computer with eqs . ( [ eq : a]-[eq : b2 ] ) and eqs . ( [ eq : cu.kk]-[eq : cu.ll ] ) . evaluating eqs . ( [ eq : cu.kk]-[eq : cu.ll ] ) , as written , requires taking a matrix inverse , which can be computationally time - consuming for larger lbms . the calculation can be made substantially faster with the following matrix identity , which we state without proof : @xmath110 this identity holds for any vectors @xmath111 and @xmath112 , and c - number @xmath113 . it allows us to compute the @xmath114 s in eqs . ( [ eq : cu.kk]-[eq : cu.ll ] ) as matrix determinants , which is faster than computing matrix inverses . furthermore , we note that , in general , certain combinations of rows ( and columns ) of @xmath45 will sum to zero , which means that we can perform a matrix rotation to reduce the size of the matrix determinant . with these methods , computation of @xmath16 , @xmath17 , and @xmath18 for the ten simplest wacs takes roughly one hour , using mathematica on a computer with a 1.2 ghz processor . the results agree with those found in @xcite . comparison with two of larger wacs , which they label @xmath115 and @xmath116 , requires a more detailed discussion of the mapping between wacs and lbms , which is done in the next section . we illustrate the mapping between larger weakly allowed clusters and local bond modifications with the sandpile modification shown in figure [ fig : s10.lattice.modification ] . in this modified sandpile , a five - site cluster is separated from the rest of the sandpile , except by a single bond . the number of states in the modified sandpile of figure [ fig : s10.lattice.modification ] is equal to the number of states of the unmodified asm where decreasing the five - site cluster s leftmost site ( which we call @xmath2 ) from 2 to 1 makes the five - site cluster a fsc , _ and _ does not produce any larger fscs @xcite . the condition that the fsc produced be maximal is necessary for this equivalence , although this condition was not explicitly stated in @xcite . if changing the height of @xmath2 from 2 to 1 makes the five - site cluster a fsc , the original height configuration ( before this change ) must have been one of the four configurations shown in figures [ fig : s10.mr ] and [ fig : s10.actual.wacs ] . of these four configurations , the one in figure [ fig : s10.mr ] is not a wac , while the three in figure [ fig : s10.actual.wacs ] are . absent other conditions , the configuration in figure [ fig : s10.mr ] does not have the same probability as the configurations in figure [ fig : s10.actual.wacs ] . the configuration in figure [ fig : s10.mr ] is more likely , as an allowed configuration always stays allowed when a site height is increased . however , for some configurations of heights outside the five - site cluster , decreasing one of the three height two sites other than the one at @xmath2 , to height one , will create an fsc outside the five - site cluster , so that figure [ fig : s10.mr ] is allowed , but figure [ fig : s10.actual.wacs ] is not . however , if we impose the condition that taking the height of @xmath2 from two to one should produce no fsc larger than the five - site cluster , then all four configurations in figure [ fig : s10.mr ] and [ fig : s10.actual.wacs ] are equally probable . with this condition , decreasing one of the height two sites in figure [ fig : s10.mr ] can not possibly create an fsc outside the five - site cluster , since then the union of this fsc with the five sites would be a larger fsc upon decreasing the height of @xmath2 . for a wac , it can be shown that if any height is decreased , not only does the wac become a fsc , but it is not contained in any larger fsc . therefore , for the three height configurations in figure [ fig : s10.actual.wacs ] , the condition that the five - site fsc generated is maximal is automatic . ( however , for the configuration in figure [ fig : s10.mr ] it is not . ) therefore , the probability associated with the modified sandpile in figure [ fig : s10.lattice.modification ] is four times the probability of any of the three wacs in figure [ fig : s10.actual.wacs ] . it is also four times the probability of the height configuration in figure [ fig : s10.mr ] , if we impose on this configuration the condition that decreasing the height of @xmath2 should produce no fsc larger than the five - site cluster ( although clearly this is not as physically interesting ) . in @xcite , the configurations that they labelled @xmath115 and @xmath116 were not wacs . once they are modified to be wacs with the same shape , our values for @xmath16 , @xmath17 , and @xmath18 , obtained with the methods of the previous section , agree with theirs . although we chose a specific sandpile modification , the discussion is easily generalized . generally , consider sandpile modifications similar to those in figure [ fig : s10.lattice.modification ] , which separate a cluster of sites from the rest of the sandpile , except for one linking site . there will be @xmath117 possible height configurations in the cluster that become fscs if the height of the linking site is reduced . some of these will be wacs and some will not . the probability associated with the lattice modification will be @xmath117 times the probability for any individual wac . ( it will also be @xmath117 times the probability for any of the other height configurations , given the condition that decreasing the height of the link site should produce no fsc larger than the cluster in question . ) these issues did not need to be discussed for the simpler wacs , such as those in figure [ fig : wacs ] ( or the other wacs of @xcite ) . for each of the sandpile modifications in the right side of figure [ fig : wacs ] , only one corresponding wac height configuration is possible ( @xmath118 ) . these methods allow us to quickly determine the effects of a number of defects or boundary conditions on all lbms , and see that each time , we obtain the same coefficients . in many cases , a defect or boundary changes the green function matrix in a manner such that the change factorizes , taking the form @xmath119 , for some vectors @xmath111 and @xmath112 . ( [ eq : if.g.factorizes ] ) then shows that the effects of the change can be written as a linear combination of the @xmath114 s , and thus as a linear combination of @xmath16 , @xmath17 , and @xmath18 . for example , consider the probability for an lbm located a distance @xmath120 from a boundary ( open or closed ) . let the boundary be at @xmath121 , and @xmath122 be the coordinate along the boundary . then the green function is modified to @xcite @xmath123 placing the local origin of the lbm at ( 0,y ) , the green function between points @xmath65 and @xmath66 , relative to this local origin , where @xmath68 , @xmath70 , @xmath69 and @xmath71 are all @xmath48 , is @xmath124 in the @xmath125 , the top sign is for closed boundaries , and the bottom sign is for open boundaries . as with eq . ( [ eq : green.r2.expansion ] ) , we have only kept terms that depend on both @xmath65 and @xmath66 ( otherwise , there would be terms of @xmath126 and @xmath127 ) . then , using eq . ( [ eq : if.g.factorizes ] ) , keeping only terms of @xmath128 , and using the definitions in eqs . ( [ eq : a]-[eq : b2 ] ) and eqs . ( [ eq : cu.kk]-[eq : cu.ll ] ) , we immediately see that the probability for the lbm is @xmath129 this agrees with results for the unit height variable found in @xcite . we have also used the methods described here to investigate bond defects in the asm . we change the toppling matrix from the defect - free matrix @xmath130 to @xmath131 if @xmath132 and @xmath133 are adjacent sites , and @xmath134 , then this corresponds to removing the bond between @xmath132 and @xmath133 . if @xmath135 , this corresponds to adding a bond between @xmath132 and @xmath133 . the green function is the inverse of the toppling matrix , and the effects of this perturbation can be calculated by summing a geometric series . the result is @xmath136 where @xmath137 the correction to the green function factorizes , just as in eq . ( [ eq : if.g.factorizes ] ) , and the corrections to the lbm probabilities again depend on the same coefficients . if the two ends of the bond defect are at @xmath63 and @xmath138 , where @xmath139 and @xmath140 are both @xmath48 , then from eqs . ( [ eq : cu.kk]-[eq : cu.ll ] ) , eqs . ( [ eq : a]-[eq : b2 ] ) , and eq . ( [ eq : if.g.factorizes ] ) , the probability for a lbm of type @xmath141 at @xmath142 , for @xmath143 , is @xmath144 this is consistent with identifying the bond defect with @xmath145 it was shown in @xcite that any local arrow diagram along an open boundary is represented in the lcft by the operator @xmath146 where @xmath147 is the vector of distances perpendicular to the boundary . the arguments there worked for _ any _ local arrow diagram ( not just those corresponding to lbms ) , and similarly to our arguments in section [ sec : wac.npoint ] , used the fact that the green function along open boundaries falls off as @xmath148 . along closed boundaries ( and in the bulk ) , the green function grows as @xmath149 , so the situation is more complicated , and not all operators are proportional to @xmath150 . ( see , for example , the operators for the height two and three variables along closed boundaries , given in @xcite . ) however , we can now derive an expression similar to eq . ( [ eq : open.general.field ] ) for lbms along closed boundaries . the green function for two sites on a closed boundary , and @xmath151 apart , was found in @xcite , using eq . ( [ eq : green.closed ] ) , to be @xmath152 using the recursion relation @xmath153 on this equation , we can extend eq . ( [ eq : green.closed.expansion.1 ] ) for points @xmath48 from the boundary : @xmath154 ( @xmath155 , @xmath156 , @xmath157 , and @xmath158 are all @xmath48 . ) the green function diverges as @xmath149 , but if we are calculating correlations of lbms along closed boundaries , we can use the arguments of section [ sec : wac.npoint ] to see that we only care about the parts of the green function matrix that depend on both the row and the column indices . the part of eq . ( [ eq : green.closed.expansion.2 ] ) that depends on both @xmath159 and @xmath160 is @xmath161 using logic identical to that in the bulk case , we can use the @xmath162 part , to derive field identifications for lbms along closed boundaries : @xmath163 @xmath164 is the vector of position coordinates parallel to the boundary . we now introduce a bond defect of strength @xmath165 , along an open or closed boundary , between sites @xmath166 and @xmath167 , that are @xmath48 apart . these bond defects can be analyzed as in the previous section . along an open boundary , the bond defect is represented by @xmath168 if @xmath169i.e . if the bond defect has a vertical component . on the other hand , if the bond defect is purely horizontal , it is represented by @xmath170 a purely horizontal bond along an open boundary is represented by a dimension four operator . for closed boundaries , the bond defect is represented by @xmath171 if the bond has a horizontal component . if the bond is purely vertical , it is represented by @xmath172 along closed boundaries , it is the purely vertical bonds that have dimension four . we have verified these field identifications of bond defects with more general calculations , involving multiple fields and multiple bond defects . this required generalizing eq . ( [ eq : green.bond ] ) for multiple bond defects . however , the generalization is straightforward , and not particularly instructive , so is not shown here . we now consider the addition of dissipation . as explained in the introduction , this takes the asm off the critical point , as shown by both numerical simulations , and an exact analysis @xcite . the toppling matrix becomes @xmath173 now , with each toppling in the bulk of the sandpile , @xmath174 grains of sand are lost . if we set @xmath175 , we get back the original , critical asm . it should be noted that the interpretation of the continuous @xmath176 limit is potentially problematic . the modification to @xmath1 described in eq . ( [ eq : delta.dissip ] ) is only sensible for integer @xmath174 . it can be extended to rational @xmath174 @xcite . however , for @xmath174 rational , but not an integer , the interpretation of the sandpile modifications ( i.e. the @xmath32 matrix ) associated with a wac is changed , so that what is meant by taking a limit of infinitesimally small , rational @xmath174 is unclear . however , presumably the fact that the massive results are good for all integers @xmath177 can justify an analytic continuation to @xmath175 . regardless , we are certainly able to formally expand all correlation functions in taylor series about @xmath175 , which is what we do here . @xmath174 defines an effective mass @xmath21 for the sandpile , where @xmath178 @xcite . @xmath179 can be thought of as the lattice spacing . in looking at off - critical correlations of lbms , we are interested in correlation functions where the number of lattice spacings between any two of the lbms is @xmath180 . taking the @xmath181 limit then defines the way in which we simulataneously take @xmath176 and distances between lbms to infinity . for the off - critical sandpile , as discussed in @xcite , we can use the same methods as before to calculate correlations , with two modifications . first , we need to use a different @xmath32 matrix than before . previously , we required that our sandpile modifications be conservative , which meant that the each row and each column of @xmath32 summed up to zero . however , for the massive sandpile we will often want to consider nonconservative @xmath32 s . ( for the unit height variable , the sandpile modification in eq . ( [ eq : b.unit.height ] ) will no longer restrict the height of @xmath2 to 1 , but rather to any height from 1 to @xmath182 . ) we are most interested in lbms associated with wacs . if we want to force the heights to the heights of the wac , the @xmath32 matrix must be changed to @xmath183 @xmath184 is the @xmath32 matrix that would be used for this wac for the non - dissipative asm ( e.g. eq . ( [ eq : b.unit.height ] ) ) , and @xmath185 by `` in the wac height configuration , '' we mean in the set of sites whose heights are fixed , and not in one of the bordering sites needed to to form the @xmath32 matrix ( e.g. , for the unit height configuration , only one site is `` in the wac height configuration '' ) . more generally , for other lbms , @xmath186 is the nonconservative part of the @xmath32 matrix . second , we need to use a new green function between lattice sites . as always , the green function is given by the inverse of the toppling matrix . in appendix [ sec : massive.green ] , we calculate the green function in the limit @xmath176 , when the distance between two sites scales as @xmath187 , and find that it approaches @xmath188 , where @xmath189 is the modified bessel function of the second kind . the asymptotic expansion in @xmath187 is then @xmath190 we have not calculated the @xmath191 functions , because they turn out to not affect the universal parts of any correlation functions . in principle , they can contain bounded functions of @xmath187 , such as @xmath192 , but we have not explicitly indicated this @xmath174-dependence , since it does not affect our analysis . ( mahieu and ruelle found , for @xmath193 and @xmath194 , eq . ( [ eq : g.k0.expansion ] ) , and the specific forms of @xmath195 and @xmath196 @xcite . they found that @xmath197 for these angles , so it is possible that @xmath197 for all @xmath198 , although we have not investigated this . ) we decompose @xmath49 as a sum of four @xmath199 matrices : @xmath200 @xmath201 is a matrix in which every element is identical . @xmath202 and @xmath203 are matrices in which the elements depend only on the row index , or only on the column index . parts of @xmath43 which can not be written in these forms go into @xmath204 . all four of these matrices are taylor series in @xmath205 , whose @xmath48 terms depends only on the taylor expansion of @xmath188 , and whose higher order terms in @xmath205 depend on @xmath195 , @xmath196 , etc . . . for example every element of @xmath201 is @xmath206 . the elements of @xmath202 and @xmath203 depend on only the coordinates in the @xmath46 lbm , or on only the coordinates in the @xmath51 lbm , and thus require one derivative ( finite difference ) of the green function . the elements of @xmath204 require two or more derivatives of the green function . for correlations at the critical point , we saw that @xmath201 , @xmath202 , and @xmath203 could all be ignored in calculating lbm correlations . however , the arguments there relied on the fact that every row and every column of every @xmath45 summed to zero . that no longer holds here , and we thus need to reconsider which terms of the green function we need to keep . we expand the correlation functions in powers of @xmath174 , and look for the lowest , nonzero , power of @xmath174 . as with critical correlations , to prove the validity of eq . ( [ eq : general.trace.formula ] ) , we need to show that the lowest - order term of the @xmath41-point function only comes from the parts of @xmath57 with @xmath41 terms off the block diagonal . however , the proof is much harder in this case ; we sketch the proof in appendix [ sec : massive.trace.formula ] . similarly to eq . ( [ eq : green.r2.expansion ] ) , we need to taylor expand @xmath207 then , just as in eq . ( [ eq : g.decompose ] ) , we can write @xmath49 as a sum of terms , each of which is the product of a length @xmath87 column vector and a length @xmath208 row vector . defining @xmath209 to be the vector of length @xmath87 , all of whose entries are 1 , we have @xmath210 in the ellipses , we have dropped not only terms of @xmath211 and higher , but all terms with @xmath195 or @xmath196 . the terms with @xmath195 or @xmath196 are not necessarily higher - order in @xmath174 than the terms shown . @xmath195 contributes terms of @xmath212 to @xmath213 , and terms of @xmath214 to @xmath215 , @xmath216 , @xmath217 , and @xmath218 . @xmath196 contributes terms of @xmath214 to @xmath213 . similarly to eq . ( [ eq : crit.n.matrix ] ) , we represent @xmath49 with a @xmath219 matrix , @xmath220 , where each element of @xmath220 represents a different choice of row vector and column vector : @xmath221 to save space , we have here abbreviated @xmath222 and @xmath223 . now , while some of the terms depending on @xmath195 and @xmath196 are @xmath212 and @xmath214 , they are higher order terms in @xmath201 , @xmath202 , @xmath203 , and @xmath204 , and thus are higher order in the specific matrix elements of @xmath220 that they contribute to . we will later see that this justifies dropping them . just as in section [ sec : wac.npoint ] , when each off - diagonal green function matrix is replaced by the product of a column vector and a row vector , each @xmath224 is bracketed by a row vector to its left and a column vector to its right , producing a @xmath91 matrix . we thus get a matrix , @xmath225 , similar to the @xmath94 of eq . ( [ eq : crit.m.matrix ] ) . @xmath225 is @xmath219 , rather than @xmath92 , because the vectors bracketing @xmath224 to the left and to the right can now be @xmath85 , @xmath86 , or @xmath209 . in principle , when calculating @xmath225 , @xmath45 and @xmath47 should be the matrices for the massive sandpile . however , we only need most elements of @xmath225 in the limit @xmath176 ( this will be justified shortly ) . in this limit , we replace @xmath45 with @xmath226 , and the elements of @xmath47 with the normal , well - known , critical green function . thus , to lowest - order , the elements of @xmath94 in @xmath225 are unchanged : @xmath227 for the new entries of @xmath225 , it is not hard to show that @xmath228 where @xmath229 is the probability for the lbm , and c is defined as the number of sites in the wac height configuration ( as defined below eq . ( [ eq : b.nc ] ) ) . @xmath230 is the only element of @xmath225 where we need the @xmath214 term . mahieu and ruelle defined @xmath19 as the coefficient in eq . ( [ eq : lcft.rep ] ) , and observed that for the 14 wacs that they considered , @xmath19 always turned out to always be equal to the number of sites in the wac height configuration @xcite . we will see that our @xmath19 is the same as their @xmath19 , proving that their observation holds for all wacs . we now have @xmath231 the @xmath41-point correlation is given by @xmath232 it is easy to now verify that the higher - order terms in @xmath174 in @xmath225 and @xmath220 that we dropped in eq . ( [ eq : n.dissip ] ) and eq . ( [ eq : m.dissip ] ) indeed give contributions of @xmath233 or higher to the @xmath41-point correlation , justifying the approximations used . we now have all correlations of lbms in the asm with dissipation . we can show that the correlations we have found are the same as the correlations of the field in eq . ( [ eq : lcft.rep ] ) . mahieu and ruelle proposed that the appropriate massive extension of the lcft in eq . ( [ eq : lcft.action ] ) is @xcite @xmath234 this theory is still gaussian , with correlation functions @xmath235 other two - point correlations can then be obtained by taking holomorphic or antiholomorphic derivatives . then , just as in section [ sec : wac.npoint ] , since the theory is gaussian , we can write the @xmath41-point correlation of eq . ( [ eq : lcft.rep ] ) exactly . the result is formally identical to eq . ( [ eq : lcft.critical.trace ] ) , where the analogues of @xmath236 and @xmath237 in eqs . ( [ eq : f.critical]-[eq : h.critical ] ) are now @xmath219 ( since the @xmath12 and @xmath13 fields in eq . ( [ eq : lcft.rep ] ) may have holomorphic derivatives , antiholomorphic derivatives , or no derivatives at all ) . using the identifications in eq . ( [ eq : a]-[eq : b2 ] ) , and @xmath178 , the results agree with the correlation function in eq . ( [ eq : massive.npt.trace.form ] ) ( up to a proportionality factor , @xmath238 ) . the green function for the asm with dissipation is the inverse of the toppling matrix @xmath1 , given in eq . ( [ eq : massive.delta ] ) . in the limit of an infinite lattice , this can be found by fourier transform @xcite : @xmath239 we are interested in this integral either for @xmath240 and @xmath41 both @xmath48 , or for @xmath240 and @xmath41 both proportional to @xmath187 , in the limit @xmath176 . we saw in section [ sec : dissipation.general ] , that for @xmath240 and @xmath41 both @xmath48 , we only needed @xmath241 , in the limit @xmath176 . but this is just the standard , dissipation - free , lattice green function , whose properties can be looked up in standard references for example , @xcite . so we only need to discuss the case where @xmath240 and @xmath41 are both proportional to @xmath187 : @xmath242 in the limit @xmath176 , the exponential oscillates infinitely rapidly , so when multiplied by any function smooth in the limit @xmath176 , gives an integral of zero . so no error is introduced by changing the region of integration to a small disc of radius @xmath243 about @xmath244 . similarly , because @xmath245 is smooth over this region , we can replace the first term with the second one . changing to polar coordinates , the integral is @xmath246 with a change of variables , the integral becomes @xmath247 where @xmath189 is the modified bessel function of the second kind . to find the leading - order critical limit , we take @xmath248 ( while @xmath249 is still large ) , and use @xmath250 , reproducing the first term of eq . ( [ eq : bulk.green ] ) . in this appendix , we sketch the proof that the trace formula in eq . ( [ eq : general.trace.formula ] ) is valid for all @xmath41-point correlations of lbms off the critical point . this requires showing that in the determinant @xmath57 , if we only want terms up to @xmath251 , we never need to pick more than @xmath41 terms off the block diagonal of @xmath252 . in other words , when we calculate the determinant with where @xmath254 is summed over all permutations of @xmath255 , we never have more than @xmath41 of the @xmath256 s from the off - diagonal blocks . the off - diagonal blocks of @xmath252 all have the form @xmath257 , @xmath50 , where @xmath45 and @xmath49 can be written with eq . ( [ eq : b.dissip ] ) and eq . ( [ eq : g.dissip.general ] ) . since every row of @xmath258 sums to zero , @xmath259 , and @xmath260 . the off - diagonal block can thus be written ( we need the @xmath211 terms for the calculation of the correlation functions in section [ sec : dissipation.general ] , but do not need their explicit form for this proof . ) since there are terms of @xmath212 in the off - diagonal blocks , naively , to get the @xmath251 contribution to the correlation function , we would need parts of the determinant with up to @xmath262 terms off the block diagonal . so we need to explain why the terms with more than @xmath41 terms off the block diagonal in fact have all their contributions to the @xmath251 part of the correlation function cancel ( as well as why all the terms with lower powers of @xmath174 cancel ) . we define a `` row matrix '' to be a matrix in which the entries depend only on the row index when we consider contributions to @xmath57 , we consider , for each contributing matrix element off the block diagonal , whether it is from @xmath263 , @xmath264 , @xmath265 , or from the elements of @xmath211 or higher . we use several matrix theorems ; the first two are general , not referring specifically to @xmath32 or @xmath30 . it is not hard to show by `` zero contribution , '' we mean that while specific terms in eq . ( [ eq : determinant.definition ] ) may be nonzero , when all such terms are considered , they cancel . theorem [ thm:1 ] is easy to prove : if @xmath268 is the element from the @xmath266 block , and @xmath269 is the element from the @xmath267 block , we get a cancelling contribution from @xmath270 and @xmath271 . it is somewhat harder to show the following theorem , which we state without proof : suppose the @xmath272 block ( @xmath50 ) is a row matrix , every column of the @xmath273 block sums to one , and for every @xmath274 , every column of the @xmath275 block sums to zero . then the determinant has zero contribution from the matrix elements of the @xmath272 block ; in other words , the determinant is unchanged if every element of the @xmath272 block is set to zero . [ thm:2 ] we note that @xmath263 and @xmath265 are both row matrices , and that every column of @xmath263 sums to zero , as does every column of @xmath264 . then , repeatedly applying theorems [ thm:1 ] and [ thm:2 ] allows us to prove the following : in the determinant @xmath57 , suppose the ( @xmath276 ) block ( @xmath277 ) has a contributing matrix element from @xmath278 . then , in the terms that produce a nonzero contribution to the determinant , there is an ordered sequence of distinct block indices , @xmath279 , @xmath280 , such that for all @xmath281 , the @xmath282 block has a term from @xmath283 . furthermore , either + 1 ) the @xmath284 block has a term of @xmath214 or higher . + or + 2 ) there is a term of order @xmath211 or higher in an off - diagonal block , @xmath285 . + a different @xmath286 block @xmath287 with a matrix element from @xmath288 will produce an ordered sequence of distinct block indices , @xmath289 , with no elements in common with @xmath290 . theorem 3 shows that each term in the determinant of type @xmath292 can be associated with a distinct term of type @xmath293 or @xmath294 , so that @xmath295 . the number of powers of @xmath174 from this contribution to the determinant is then _ at least _ so if we only want @xmath251 contributions to the correlation function we should never have more than @xmath41 terms off the block diagonal . furthermore , we want a connected correlation function , so we should always have exactly @xmath41 terms off the block diagonal . this concludes the proof that eq . ( [ eq : general.trace.formula ] ) is valid for off - critical @xmath41-point correlations . to get strict equality in eq . ( [ eq : order.inequality ] ) , we need @xmath297 , @xmath298 , and @xmath299 . furthermore , the terms of type @xmath293 should be exactly proportional to @xmath300 . the fact that @xmath298 means that in the diagonal blocks , @xmath301 , we can set @xmath175 at the start of our calculations , as already seen by other means in section [ sec : dissipation.general ] .

during the last few years , a two - dimensional hexagonal lattice of carbon atoms known as graphene has attracted widespread attention owing to its unusual physical properties and promising applications @xcite . unlike conventional condensed matter systems , the quasiparticles in graphene are massless or very light particles which at low energies obey the relativistic dirac equation , where the speed of light @xmath0 is replaced with the fermi velocity @xmath1 ( see the monograph @xcite and review papers @xcite ) . this makes the electronic properties of graphene highly nontrivial , and , specifically , results in the existence @xcite of a universal ( up to small nonlocal corrections ) frequency - independent conductivity @xmath2 in the limit of zero temperature . the electrical conductivity of graphene was investigated in many theoretical papers using the current - current correlation functions in the random phase approximation , the two - dimensional drude model , the quasiclassical approach , the kubo formula , and boltzmann s transport theory ( see , e.g. , refs . @xcite ) . at first , a few distinct values for the universal conductivity @xmath2 have been obtained by different authors ( see refs . @xcite for a review ) depending on the order of limiting transitions in the used theoretical approach . at a later time , a consensus was achieved on the value @xmath3 @xcite . however , up to now there is no complete agreement between the results of the different approximate and phenomenological approaches to the conductivity of graphene at nonzero temperature , in the presence of nonzero mass - gap parameter and chemical potential . in doing so , the experimental results @xcite are not of sufficient precision for a clear discrimination between the various theoretical predictions . quantum electrodynamics at nonzero temperature provides a regular way for describing the response of a physical system to the electromagnetic field by means of polarization tensor . for graphene with an arbitrary mass - gap parameter @xmath4 an explicit expression for the polarization tensor in ( 2 + 1)-dimensional space - time was found in ref . @xcite at zero temperature . the temperature correction to the polarization tensor of graphene was derived in ref . @xcite at the pure imaginary matsubara frequencies . the results of refs . @xcite have been used to calculate the casimir and casimir - polder forces @xcite in many real physical systems incorporating graphene sheets @xcite ( previously such calculations were performed by means of the density - density correlation functions of graphene , the spatially nonlocal dielectric permittivities , phenomenological drude model , etc . the polarization tensor of graphene was also used @xcite to compare with theory the experimental data on measuring the casimir interaction in graphene systems @xcite , and very good agreement was obtained . another representation for the polarization tensor of graphene valid over the entire plane of complex frequency was derived in ref . it was applied to investigate the origin of large thermal effect in the casimir interaction between two graphene sheets @xcite and to calculate the reflectivity properties of gapless graphene @xcite . in ref . @xcite the same polarization tensor allowed calculation of the reflectivity properties of graphene - coated substrates . the derivation of the polarization tensor valid along the real frequency axis made it possible to investigate the electrical conductivity of graphene on the basis of first principles of quantum electrodynamics at nonzero temperature . note that in ref . @xcite explicit expressions for the polarization tensor at real frequencies were obtained only for a pure ( pristine ) graphene having a zero mass - gap parameter , whereas for a gapped graphene the respective expressions were presented in an implicit form . because of this , as the starting point , the conductivity of pure graphene was investigated @xcite . in doing so , a few results obtained previously using different theoretical approaches have been reproduced , and several novel analytic asymptotic expressions and numerical results ( especially for the imaginary part of conductivity ) have been obtained . in this paper we apply the formalism of the polarization tensor to systematically investigate the electrical conductivity of gapped graphene on the solid foundation of quantum electrodynamics . for this purpose , an explicit form of the results of ref . @xcite related to the case of nonzero mass - gap parameter is employed , which was found in ref . it is known that quasiparticles in graphene may acquire some small mass @xmath5 under the influence of impurities , electron - electron interactions , and in the presence of substrates @xcite . the conductivity of gapped graphene was studied in ref . @xcite using the two - band model and in ref . @xcite using the static polarization function . at zero temperature some analytic expressions for the conductivity of gapped graphene were obtained in refs . @xcite , whereas ref . @xcite contains the final results for both real and imaginary parts of conductivity in the local approximation . we reproduce the zero - temperature analytic results of ref . @xcite within the formalism of the polarization tensor and generalize them taking into account the effects of nonlocality . specifically , it is shown that at zero temperature the real part of the conductivity of gapped graphene is equal to zero in the frequency region from zero to the characteristic frequency , which is slightly larger than @xmath6 due to nonlocal effects . with increasing frequency the real part of conductivity decreases monotonously from @xmath7 to @xmath2 . according to our results , nonzero temperature has a dramatic impact on the conductivity of gapped graphene . thus , the real part of conductivity for frequencies larger than the characteristic frequency may be either decreasing or increasing function depending on the relationship between the mass - gap parameter and the temperature . the imaginary part of the conductivity of gapped graphene varies from plus to minus infinity when frequency increases from zero to the characteristic frequency ( at zero temperature it varies from zero to minus infinity in the same frequency region ) . we obtain the analytic asymptotic expressions and perform numerical computations for both real and imaginary parts of the conductivity of gapped graphene at nonzero temperature . the paper is organized as follows . in sec . ii , we present explicit expressions for the polarization tensor of gapped graphene along the real frequency axis . in sec . iii , both real and imaginary parts of the conductivity of gapped graphene are derived at zero temperature . sections iv and v investigate the impact of nonzero temperature on real and imaginary parts of the conductivity of gapped graphene , respectively . our conclusions and discussion are contained in sec . vi . here , we present analytic results for the polarization tensor of gapped graphene along the real frequency axis derived in refs . we also demonstrate how real and imaginary parts of the conductivity of graphene are expressed via imaginary and real parts of the polarization tensor , respectively . both the longitudinal ( along the graphene surface ) and transverse ( normal to the graphene surface ) conductivities are considered . the description of graphene using the polarization tensor in the one - loop approximation and the density - density correlation functions in the random phase approximation are equivalent @xcite . the former method is , however , somewhat advantageous because it relies on well understood formalism of quantum field theory at nonzero temperature . this allowed complete derivation of the polarization tensor of gapped graphene valid over the entire plane of complex frequency @xcite . in ref . @xcite a more convenient form of this tensor especially along the real frequency axis has been obtained for a gapped graphene . as usual , we notate the polarization tensor as @xmath8 , where @xmath9 , @xmath10 is the frequency of an electromagnetic wave , @xmath11 is the magnitude of the wave vector component parallel to graphene , and @xmath12 is the temperature . we assume real photons on a mass - shell , so that the inequality @xmath13 is satisfied . it has been shown @xcite that only the two components of the polarization tensor of graphene are independent . it is conventional to use @xmath14 and @xmath15 as the two independent quantities . for our purposes , however , it is more convenient to use the following combination @xmath16 as the second quantity . we present @xmath14 and @xmath17 as the sums of zero - temperature contributions and temperature corrections @xmath18 the thermal corrections @xmath19 and @xmath20 vanish in the limit of zero temperature . the analytic expressions for the polarization tensor of gapped graphene at @xmath21 were derived in ref . @xcite . they result in @xmath22 where @xmath23 is the fine structure constant , the quantity @xmath24 is defined as @xmath25 and the function @xmath26 along the real frequency axis is given by @xcite @xmath27 { \rm arctanh}\frac{\hbar\omega\eta}{\delta } , & \hbar\omega<\frac{\delta}{\eta } , \\[3 mm ] \frac{2\delta}{c}-\frac{2\hbar\omega\eta}{c}\left[1 + \left(\frac{\delta}{\hbar\omega\eta}\right)^2\right ] \left({\rm arctanh}\frac{\delta}{\hbar\omega\eta}+ i\frac{\pi}{2}\right ) , & \hbar\omega\geq\frac{\delta}{\eta}. \end{array } \right . \label{eq5}\ ] ] here , the energy gap @xmath28 and @xmath5 is the mass of a quasiparticle ( according to some estimations @xcite @xmath29 may achieve 0.1ev ) . as is seen from eqs . ( [ eq3])([eq5 ] ) , under the condition @xmath30 the quantities @xmath31 and @xmath32 are real . if the condition @xmath33 is satisfied , they have both real and imaginary parts . we now continue by considering the thermal corrections in eq . ( [ eq2 ] ) . the imaginary parts of both @xmath19 and @xmath20 are also different from zero only under the condition @xmath33 @xcite . the explicit forms are the following @xcite : @xmath34 & & \label{eq6 } \\ & & { \rm im}\delta_t\pi(\omega , k , t)= \frac{8\alpha\hbar\eta\omega}{v_f^2}\int_{u^{(-)}}^{u^{(+)}}\!\!\ ! \frac{du}{e^{\beta u}+1 } \frac{(2cu-\omega)^2+v_f^2k^2[1-a(\omega , k ) ] } { \sqrt{v_f^2k^2a(\omega , k)-(2cu-\omega)^2 } } , \nonumber\end{aligned}\ ] ] where @xmath35 and the integration limits are @xmath36 . \label{eq8}\ ] ] the above expressions ( [ eq2])([eq13 ] ) provide an exact representation for the polarization tensor of graphene at arbitrary frequency , wave vector , temperature and mass - gap parameter in the application region of the dirac model . this opens up possibility for a comprehensive investigation of the conductivity of gapped graphene on the basis of first principles of quantum electrodynamics . the point is that the longitudinal and transverse conductivities of graphene are directly connected with the respective density - density correlation functions @xcite . the latter are expressed via the components of the polarization tensor @xcite . as a result , the longitudinal and transverse conductivities of graphene take the form @xcite @xmath46 in the next sections , using these equations , the conductivities of gapped graphene are treated at both zero and nonzero temperature . at @xmath21 the polarization tensor of gapped graphene is characterized by the quantities @xmath31 and @xmath32 which are presented in eqs . ( [ eq3])([eq5 ] ) . we consider separately the real and imaginary parts of the conductivity . according to eq . ( [ eq14 ] ) , the real parts of graphene conductivities at zero temperature are given by @xmath47 from eqs . ( [ eq3])([eq5 ] ) one obtains @xmath48 \theta(\hbar\omega\eta-\delta ) , \nonumber \\[-2 mm ] & & \label{eq16 } \\ & & { \rm im}\pi^{(0)}(\omega , k)= -\frac{\alpha\pi\hbar k^2\omega\eta}{c}\left[1 + \left(\frac{\delta}{\hbar\omega\eta}\right)^{\!\!2}\right ] \theta(\hbar\omega\eta-\delta ) , \nonumber\end{aligned}\ ] ] where @xmath49 is the step function . substituting eq . ( [ eq16 ] ) in eq . ( [ eq15 ] ) and taking into account that @xmath23 , one arrives at @xmath50 \theta(\hbar\omega\eta-\delta ) , \label{eq17 } \\ & & { \rm re}\sigma_{\bot}^{(0)}(\omega , k)={\sigma_0}{\eta } \left[1 + \left(\frac{\delta}{\hbar\omega\eta}\right)^{\!\!2}\right ] \theta(\hbar\omega\eta-\delta ) , \nonumber\end{aligned}\ ] ] where the universal conductivity of graphene is @xmath51 as is seen from eq . ( [ eq17 ] ) , in the frequency region satisfying the condition @xmath30 the real parts of the conductivities of gapped graphene are equal to zero . note that the vanishing of the conductivity of gapped graphene at low frequencies was noted in many papers using various formalisms ( see , for instance , refs . @xcite ) . taking into account that @xmath13 , one obtains @xmath52 using eq . ( [ eq4 ] ) , we expand eq . ( [ eq17 ] ) in powers of the small parameter ( [ eq19 ] ) and in the lowest order find @xmath53\right\ } \theta(\hbar\omega\eta-\delta ) , \nonumber\\[-2 mm ] & & \label{eq20 } \\ & & { \rm re}\sigma_{\bot}^{(0)}(\omega , k)={\sigma_0}\left\ { 1+\left(\frac{\delta}{\hbar\omega}\right)^{\!\!2}- \frac{1}{2}\left(\frac{v_fk}{\omega}\right)^{\!\!2}\left [ 1-\left(\frac{\delta}{\hbar\omega}\right)^{\!\!2 } \right]\right\ } \theta(\hbar\omega\eta-\delta ) . \nonumber\end{aligned}\ ] ] as is seen from eq . ( [ eq20 ] ) , @xmath54 and @xmath55 differ only in the first perturbation order in the small parameter ( [ eq19 ] ) and , thus , can be considered as equal for all practical purposes . they are normalized to @xmath2 and plotted in fig . [ fg1 ] as functions of @xmath56 . note that in this and all the following figures computational results for the longitudinal and transverse conductivities are indistinguishable in the used scales . because of this , they are always presented by a single line . in the region @xmath57 the conductivities of graphene are the monotonous functions decreasing from @xmath7 to @xmath2 when the frequency increases from @xmath6 to infinity . in the local limit , where the results are independent on the wave vector , eq . ( [ eq20 ] ) leads to @xmath58 note that in our case the same results are obtained from eq . ( [ eq20 ] ) taken at @xmath59 . they coincide with the results presented in ref . @xcite obtained in the framework of another approach . using eq . ( [ eq14 ] ) , for the imaginary part of conductivity of graphene at zero temperature we obtain @xmath60 for the real part of the polarization tensor from eqs . ( [ eq3])([eq5 ] ) one finds @xmath61 where @xmath62 \left\ { \begin{array}{ll } { \rm arctanh}\frac{\hbar\omega\eta}{\delta } , & \hbar\omega\eta<\delta , \\ { \rm arctanh}\frac{\delta}{\hbar\omega\eta } , & \hbar\omega\eta\geq\delta . \end{array } \right . \label{eq24}\ ] ] using the identity @xcite @xmath63 eq . ( [ eq24 ] ) can be rewritten in the form @xmath64 \ln\left|\frac{\delta+\hbar\omega\eta}{\delta-\hbar\omega\eta } \right| . \label{eq26}\ ] ] substituting eqs . ( [ eq23 ] ) and ( [ eq26 ] ) in eq . ( [ eq22 ] ) , one arrives at @xmath65 , \nonumber\\[-2 mm ] \label{eq27 } \\ & & { \rm im}\sigma_{\bot}^{(0)}(\omega , k)=\frac{\sigma_0\eta}{\pi } \left[\frac{2\delta}{\hbar\omega\eta}- \frac{(\hbar\omega\eta)^2+\delta^2}{(\hbar\omega\eta)^2 } \ln\left|\frac{\delta+\hbar\omega\eta}{\delta-\hbar\omega\eta } \right|\right ] . \nonumber\end{aligned}\ ] ] from eq . ( [ eq27 ] ) it is seen that the nonlocal effects play equally small role in the imaginary parts of conductivity of graphene as in its real part . in the local limit we have @xmath66 and eq . ( [ eq27 ] ) reduces to @xmath67 . \label{eq28}\ ] ] this coincides with the result presented in ref . @xcite . in fig . [ fg2 ] we plot the imaginary part of the conductivity of graphene normalized to @xmath2 at @xmath21 as a function of the ratio of @xmath68 to the width of the gap @xmath4 . as is seen in fig . [ fg2 ] , for any fixed @xmath4 the imaginary part of conductivity decreases monotonously from zero to minus infinity when frequency increases from 0 to @xmath6 and then increases to zero with further increase of frequency . we start from the real parts of thermal corrections to the conductivities of graphene . according to eqs . ( [ eq2 ] ) and ( [ eq14 ] ) , they are given by @xmath69 recall that the quantities @xmath70 and @xmath71 ( and , thus , @xmath72 and @xmath73 ) are equal to zero under the condition @xmath30 . under the opposite condition @xmath33 they are given by eq . ( [ eq6 ] ) . it is convenient to introduce another integration variable in eq . ( [ eq6 ] ) @xmath74 where @xmath75 is defined in eq . ( [ eq7 ] ) . then eq . ( [ eq6 ] ) takes the form @xmath76 , \nonumber \\ & & \label{eq31 } \\[-2 mm ] & & { \rm im}\delta_t\pi(\omega , k , t)= 4\alpha\hbar\omega\eta\frac{k^2}{c } \int_{-1}^{1}\frac{d\tau}{\exp\left ( \frac{\hbar\omega}{2k_bt}+\gamma\tau\right)+1 } \left[\frac{1}{\sqrt{1-\tau^2}}-a(\omega , k)\sqrt{1-\tau^2 } \right ] . \nonumber\end{aligned}\ ] ] here , the parameter @xmath77 is defined as @xmath78 it is equal to zero in the local limit . expanding the exponent - containing fraction in the first formula of eq . ( [ eq31 ] ) in a series and calculating the integrals @xcite one obtains @xmath79 \nonumber \\ & & \label{eq33 } \\[-2 mm ] & & ~~ = -4\alpha\hbar\frac{ck^2}{\omega\eta } \sum_{n=1}^{\infty}(-1)^{n-1}e^{-\frac{n\hbar\omega}{2k_bt } } \left\{\left[1-a(\omega , k)\right ] { i}_0(n\gamma)+ \frac{a(\omega , k)}{n\gamma}{i}_1(n\gamma)\right\ } , \nonumber\end{aligned}\ ] ] where @xmath80 are the bessel functions of an imaginary argument . substituting eqs . ( [ eq33 ] ) and ( [ eq34 ] ) with added @xmath82-functions in eq . ( [ eq29 ] ) , we obtain the exact expressions for the real parts of thermal corrections to longitudinal and transverse conductivities of gapped graphene at nonzero temperature @xmath83 { i}_0(n\gamma)+ \frac{a(\omega , k)}{n\gamma } { i}_1(n\gamma)\right\ } , \nonumber\\ & & \label{eq35 } \\[-2 mm ] & & { \rm re}\delta_t\sigma_{\bot}(\omega , k , t)=-{4\sigma_0}{\eta } \theta(\hbar\omega\eta-\delta ) \sum_{n=1}^{\infty}(-1)^{n-1}e^{-\frac{n\hbar\omega}{2k_bt } } \left [ { i}_0(n\gamma)- \frac{a(\omega , k)}{n\gamma } { i}_1(n\gamma)\right ] . \nonumber\end{aligned}\ ] ] as is seen from eq . ( [ eq38 ] ) , the real parts of thermal corrections to the conductivity of gapped graphene are mostly determined by the local contributions @xmath87 the nonlocal contributions in eq . ( [ eq38 ] ) become dominant only at very low temperature when the complete thermal correction is exponentially small . by adding eqs . ( [ eq21 ] ) and ( [ eq39 ] ) , one arrives at the final expression for the real part of conductivity of gapped graphene at nonzero temperature in the local approximation @xmath88 in fig . [ fg3](a ) we plot the real part of the normalized to @xmath2 longitudinal and transverse conductivities of graphene with the gap parameter @xmath89ev as a function of frequency ( measured in ev ) at different temperatures . the lines from top to bottom correspond to @xmath90 , 100 , and 300k , respectively ( note that @xmath91k corresponds to @xmath92ev ) . to gain a better understanding , the region of very low frequencies is shown in fig . [ fg3](b ) on an enlarged scale . according to fig . [ fg3](a , b ) , the real part of conductivity decreases with increasing @xmath12 . this is because the thermal correction to conductivity presented in eq . ( [ eq39 ] ) is negative . as is seen in fig . [ fg3](a , b ) , the real part of the conductivity of graphene increases with increasing frequency and goes to @xmath2 . this is , however , not a universal property , but is determined by chosen ( small ) value of the gap parameter . to illustrate the latter statement , in fig . [ fg4 ] we plot the real parts of the conductivities of graphene with the larger gap parameters equal to ( a ) @xmath93ev and ( b ) @xmath94ev as the functions of frequency . the lines from top to bottom are plotted for @xmath90 , 100 , and 300k , respectively . as is seen in fig . [ fg4](a ) , at @xmath95 and 100k the real part of conductivity increases , but at @xmath96k it decreases with increasing frequency . figure [ fg4](b ) illustrates what is happening with an increase of the gap parameter to @xmath94ev . in this case the real part of conductivity at @xmath90 and 100k decreases with increasing frequency , whereas at @xmath91k it becomes nonmonotonous due to an interplay of two factors in eq . ( [ eq40 ] ) . with further increase of the gap parameter ( up to @xmath97ev or larger ) the real part of conductivity becomes monotonously decreasing at all considered temperatures , whereas the values of @xmath98 at @xmath90 and 100k become indistinguishable . the imaginary parts of thermal corrections to the conductivity of graphene are obtained from eqs . ( [ eq2 ] ) and ( [ eq14 ] ) @xmath99 at first we consider the frequency region @xmath30 . here , the real parts of the temperature corrections to the polarization tensor are presented by eq . ( [ eq9 ] ) . thus , the exact expressions for @xmath100 and @xmath101 in this region are given by eqs . ( [ eq9 ] ) and ( [ eq41 ] ) . the total imaginary part of the conductivity is given by @xmath102 where @xmath103 is defined in eq . ( [ eq27 ] ) . we expand eq . ( [ eq9 ] ) up to the first order in a small parameter ( [ eq19 ] ) and find @xmath104 , \nonumber \\[-2 mm ] & & \label{eq43 } \\ & & { \rm re}\delta_t\pi(\omega , k , t)=8\alpha\hbar k^2 \int_{\delta/(2\hbar c)}^{\infty}\frac{du}{e^{\beta u}+1 } \left[1+\frac{\delta^2+(\hbar\omega)^2}{4(\hbar cu)^2- ( \hbar\omega)^2}\right ] . \nonumber\end{aligned}\ ] ] substituting these equations in eq . ( [ eq41 ] ) , one arrives at equal results for @xmath100 and @xmath101 in the local approximation @xmath105 . \label{eq44}\ ] ] similar to sec . iv , it can be shown that the nonlocal corrections to @xmath106 are negligibly small . the total imaginary part of the conductivity of graphene in the local approximation is given by @xmath107 where @xmath103 and @xmath108 are given by eqs . ( [ eq28 ] ) and ( [ eq44 ] ) , respectively . note that under the condition @xmath109 we can obtain simple asymptotic expression for @xmath108 in eq . ( [ eq44 ] ) . for this purpose we take into account that the main contribution to the integral is given by @xmath110 and that the condition @xmath109 in the frequency region under consideration leads to @xmath111 as well . then , one can neglect by the second term in square brackets of eq . ( [ eq44 ] ) as compared to unity and , after the integration , arrive at @xmath112 as is seen from eqs . ( [ eq28 ] ) and ( [ eq44 ] ) , the imaginary part of the conductivity of gapped graphene is negative , whereas the thermal correction to it is positive . this means that the total imaginary part of the conductivity may change its sign at some frequency . numerical computations confirm this conclusion . we have computed @xmath113 normalized to @xmath2 as a function of @xmath10 in the region @xmath114 at room temperature @xmath91k in two ways : by eqs . ( [ eq9 ] ) , ( [ eq27 ] ) , ( [ eq41 ] ) , and ( [ eq42 ] ) , i.e. , using the exact polarization tensor with several selected values of @xmath11 , and by eqs . ( [ eq28 ] ) , ( [ eq44 ] ) , and ( [ eq45 ] ) , i.e. , in the local approximation , with nearly coincident results . in fig . [ fg5](a ) the normalized imaginary part of conductivity of graphene with the gap parameter @xmath115ev is shown as a function of frequency in the region @xmath114 ( the latter is marked by the vertical dashed line ) . as is seen in fig . [ fg5](a ) , @xmath113 varies from infinity at zero frequency to minus infinity when @xmath116 . it has a root at some @xmath68 close to @xmath4 . for better vizualization , in fig . [ fg5](b ) we show an immediate vicinity of the frequency @xmath117 on an enlarged scale . the root of @xmath113 is @xmath118ev . the behavior of @xmath108 at @xmath119 is determined by the thermal correction @xmath108 , whereas the behavior in the limiting case @xmath116 is caused by the imaginary part of conductivity at zero temperature @xmath103 ( compare with fig . [ fg2 ] ) . for comparison purposes , in fig . [ fg6 ] the computational results for @xmath113 at @xmath91k , @xmath114 are shown also for a larger gap @xmath97ev . as is seen in fig . [ fg6 ] , in this case the root of @xmath113 , @xmath120ev , is much more separated from the border frequency @xmath121 than in fig . [ fg5 ] ( see below for the discussion of the frequency region @xmath57 ) . we are coming now to the imaginary parts of the conductivities of graphene @xmath113 under the condition @xmath57 . the imaginary parts of the thermal corrections , @xmath108 , are again given by eq . ( [ eq41 ] ) where , however , the quantities @xmath38 and @xmath39 are expressed by eqs . ( [ eq10])([eq13 ] ) . these expressions , together with the contribution at zero temperature in eq . ( [ eq27 ] ) , are used below in numerical computations . it is possible also to obtain the asymptotic expressions valid under certain additional conditions . thus , according to ref . @xcite , at relatively high frequencies satisfying the condition @xmath122 one obtains @xmath123 where @xmath124 is the riemann zeta function . substituting these equations in eq . ( [ eq41 ] ) , we have @xmath125 if @xmath126 and @xmath127 the following asymptotic expressions are valid @xcite @xmath128 then from eq . ( [ eq41 ] ) one arrives at @xmath129 the complete imaginary part of the conductivity of gapped graphene under respective conditions is obtained by summing eqs . ( [ eq48 ] ) and ( [ eq50 ] ) with eq . ( [ eq28 ] ) . now we present the results of numerical computations using the exact formulas ( [ eq10])([eq13 ] ) , ( [ eq27 ] ) , and ( [ eq41 ] ) . in figs . [ fg5](a ) and [ fg5](b ) the computational results for @xmath113 normalized to @xmath2 are presented in the frequency region @xmath130ev at room temperature @xmath91k . in fig . [ fg6 ] similar results are shown for graphene with the gap parameter @xmath97ev . as is seen in figs . [ fg5](a , b ) and [ fg6 ] , in the frequency region @xmath131 the imaginary part of @xmath132 varies from minus infinity to zero with increasing frequency , i.e. , qualitatively in the same way as at zero temperature . by comparing figs . [ fg5](a ) and [ fg6 ] it is seen that for a larger gap parameter @xmath113 faster approaches to zero with increasing frequency . it can be seen that fig . [ fg6 ] in the frequency region @xmath131 is in a rather good agreement with the asymptotic expression ( [ eq50 ] ) added to the zero - temperature contribution ( [ eq28 ] ) . this is in accordance with the application conditions of eq . ( [ eq50 ] ) . in the foregoing , we have investigated the conductivity of gapped graphene described by the dirac model on the basis of first principles of quantum electrodynamics . in doing so , no input parameter has been used . both cases of zero and nonzero temperature have been considered . the longitudinal and transverse conductivities of graphene were expressed via the components of the polarization tensor in ( 2 + 1)-dimensional space - time valid over the entire plane of complex frequency @xcite . this is a microscopic description in the sense that the polarization tensor eventually represents the response of individual electronic excitations to the electromagnetic field . at zero temperature , we have confirmed the analytic expressions ( [ eq21 ] ) and ( [ eq28 ] ) for both real and imaginary parts of the conductivity of graphene obtained earlier in the local approximation @xcite . we have also generalized these expressions in eqs . ( [ eq17 ] ) and ( [ eq27 ] ) taking into account nonlocal effects . the exact expressions for the real parts of thermal corrections to the real parts of conductivities of gapped graphene at zero temperature are defined in eq . ( [ eq35 ] ) . the small role of nonlocal effects is demonstrated in eq . ( [ eq38 ] ) . finally , simple analytic expression for the real part of the total conductivity of gapped graphene at any temperature is given by eq . ( [ eq40 ] ) . according to our results , the real part of conductivity is equal to zero for @xmath114 . in the frequency region @xmath57 the real part of conductivity takes some value @xmath133 belonging to the interval @xmath134 at @xmath121 and goes to @xmath2 with increasing frequency . the value @xmath133 depends on the values of the temperature and gap parameter . for the imaginary parts of thermal corrections to the imaginary parts of conductivities of gapped graphene at zero temperature , the exact expressions are obtained in eqs . ( [ eq41 ] ) and ( [ eq9 ] ) in the frequency region @xmath30 . in the local approximation , the simple eq . ( [ eq44 ] ) is obtained , as well as the asymptotic expression ( [ eq46 ] ) . the most important impact of temperature on the imaginary part of conductivity of gapped graphene is that it goes to infinity when the frequency vanishes ( at @xmath21 the imaginary part of conductivity goes to zero with vanishing frequency ) . in the frequency region @xmath33 the exact expressions for the imaginary parts of thermal corrections are given by eqs . ( [ eq41 ] ) and ( [ eq10])([eq13 ] ) . the simple asymptotic expressions in this case are given in eqs . ( [ eq48 ] ) and ( [ eq50 ] ) . in all cases , the results of numerical computations performed using the exact formulas are in a very good agreement with that using the local approximation and with the asymptotic expansions in their areas of application . in the limiting case @xmath135 all the results of this paper smoothly transform into the previously obtained ones for pure graphene @xcite . finally , we conclude that the description of graphene by means of the polarization tensor calculated using the first principles of quantum electrodynamics at nonzero temperature in ( 2 + 1)-dimensional space - time turns out to be very fruitful not only in the casimir effect , but also for a better understanding of more conventional physical phenomena , such as the reflectivity and conductivity properties of graphene . in this respect it would be interesting to apply the results of the recent ref . @xcite for theoretical description of electrical conductivity of graphene with nonzero chemical potential using the formalism of the polarization tensor . the work of v.m.m . was partially supported by the russian government program of competitive growth of kazan federal university . 99 m. i. katsnelson , _ graphene : carbon in two dimensions _ ( cambridge university press , cambridge , 2012 ) . a. h. castro neto , f. guinea , n. m. r. peres , k. s. novoselov , and a. k. geim , rev . mod . phys . * 81 * , 109 ( 2009 ) . n. m. r. peres , rev . phys . * 82 * , 2673 ( 2010 ) . s. das sarma , s. adam , e. h. hwang , and e. rossi , rev . mod . phys . * 83 * , 407 ( 2011 ) . a. w. w. ludwig , m. p. a. fisher , r. shankar , and g. grinstein , phys . b * 50 * , 7526 ( 1994 ) . t. ando , y. zheng , and h. suzuura , j. phys . . jpn . * 71 * , 1318 ( 2002 ) . v. p. gusynin and s. g. sharapov , phys . b * 73 * , 245411 ( 2006 ) . m. i. katsnelson , eur . j. b * 51 * , 157 ( 2006 ) . k. ziegler , phys . 97 * , 266802 ( 2006 ) . k. ziegler , phys . b * 75 * , 233407 ( 2007 ) . v. p. gusynin , s. g. sharapov , and j. p. carbotte , phys . 98 * , 157402 ( 2007 ) . m. trushin and j. schliemann , phys . * 99 * , 216602 ( 2007 ) . l. a. falkovsky and a. a. varlamov , eur . j. b * 56 * , 281 ( 2007 ) . l. a. falkovsky and s. s. pershoguba , phys . b * 76 * , 153410 ( 2007 ) . m. auslender and m. i. katsnelson , phys . b * 76 * , 235425 ( 2007 ) . v. p. gusynin and s. g. sharapov , j. phys . : condens . matter * 19 * , 026222 ( 2007 ) . gusynin , s. g. sharapov , and j. p. carbotte , int . j. mod b * 21 * , 4611 ( 2007 ) . t. stauber , n. m. r. peres , and a. k. geim , phys . b * 78 * , 085432 ( 2008 ) . n. m. r. peres and t. stauber , int . j. mod . b * 22 * , 2529 ( 2008 ) . t. g. pedersen , a .- jauho , and k. pedersen , phys . b * 79 * , 113406 ( 2009 ) . m. lewkowicz and b. rosenstein , phys . * 102 * , 106802 ( 2009 ) . j. j. palacios , phys . b * 82 * , 165439 ( 2010 ) . l. moriconi and d. niemeyer , phys . b * 84 * , 193401 ( 2011 ) . p. v. buividovich , e. v. luschevskaya , o. v. pavlovsky , m. i. polikarpov , and m. v. ulybyshev , phys . b * 86 * , 045107 ( 2012 ) . . b ' acsi and a. virosztek , phys . b * 87 * , 125425 ( 2013 ) . c. a. dartora and g. g. cabrera , phys . rev . b * 87 * , 165416 ( 2013 ) . t. stauber , j. phys . : condens . matter * 26 * , 123201 ( 2014 ) . t. louvet , p. delplace , a. a. fedorenko , and d. carpentier , phys . b * 92 * , 155116 ( 2015 ) . d. k. patel , a. c. sharma , and s. s. z. ashraf , phys . status solidi * 252 * , 282 ( 2015 ) . m. merano , phys . a * 93 * , 013832 ( 2016 ) . tan , y. zhang , k. bolotin , y. zhao , s. adam , e. h. hwang , s. das sarma , h. l. stormer , and p. kim , phys . lett . * 99 * , 246803 ( 2007 ) . r. r. nair , p. blake , a. n. grigorenko , k. s. novoselov , t. j. booth , t. stauber , n. m. r. peres , and a. k. geim , science * 320 * , 1308 ( 2008 ) . z. li , e. henriksen , z. jiang , z. hao , m. martin , p. kim , h. stormer , and d. basov , nature phys . * 4 * , 532 ( 2008 ) . k. f. mak , m. y. sfeir , y. wu , c. h. lui , j. a. misewich , and t. f. heinz , phys . * 101 * , 196405 ( 2008 ) . j. horng , c .- f . chen , b. geng _ et al . b * 83 * , 165113 ( 2011 ) . m. bordag , i. v. fialkovsky , d. m. gitman , and d. v. vassilevich , phys . b * 80 * , 245406 ( 2009 ) . i. v. fialkovsky , v. n. marachevsky , and d. v. vassilevich , phys . rev . b * 84 * , 035446 ( 2011 ) . g. l. klimchitskaya , u. mohideen , and v. m. mostepanenko , rev . phys . * 81 * , 1827 ( 2009 ) . m. bordag , g. l. klimchitskaya , u. mohideen , and v. m. mostepanenko , _ advances in the casimir effect _ ( oxford university press , oxford , 2015 ) . m. bordag , g. l. klimchitskaya , and v. m. mostepanenko , phys . b * 86 * , 165429 ( 2012 ) . m. chaichian , g. l. klimchitskaya , v. m. mostepanenko , and a. tureanu , phys . a * 86 * , 012515 ( 2012 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . b * 87 * , 075439 ( 2013 ) . g. l. klimchitskaya , u. mohideen , and v. m. mostepanenko , phys . rev b * 89 * , 115419 ( 2014 ) . g. l. klimchitskaya , v. m. mostepanenko , and bo e. sernelius , phys . b * 89 * , 125407 ( 2014 ) . b. arora , h. kaur , and b. k. sahoo , j. phys . b * 47 * , 155002 ( 2014 ) . k. kaur , j. kaur , b. arora , and b. k. sahoo , phys . b * 90 * , 245405 ( 2014 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . b * 91 * , 045412 ( 2015 ) . g. gmez - santos , phys . b * 80 * , 245424 ( 2009 ) . d. drosdoff and l. m. woods , phys . b * 82 * , 155459 ( 2010 ) . d. drosdoff and l. m. woods , phys . a * 84 * , 062501 ( 2011 ) . bo e. sernelius , europhys . lett . * 95 * , 57003 ( 2011 ) . bo e. sernelius , phys . b * 85 * , 195427 ( 2012 ) . a. d. phan , l. m. woods , d. drosdoff , i. v. bondarev , and n. a. viet , appl . phys . lett . * 101 * , 113118 ( 2012 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . rev a * 89 * , 052512 ( 2014 ) . a. a. banishev , h. wen , j. xu , r. k. kawakami , g. l. klimchitskaya , v. m. mostepanenko , and u. mohideen , phys . b * 87 * , 205433 ( 2013 ) . m. bordag , g. l. klimchitskaya , v. m. mostepanenko , and v. m. petrov , phys . d * 91 * , 045037 ( 2015 ) ; * 93 * , 089907(e ) ( 2016 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . rev b * 91 * , 174501 ( 2015 ) . g. l. klimchitskaya , int . a * 31 * , 1641026 ( 2016 ) . g. l. klimchitskaya , c. c. korikov , and v. m. petrov , phys . b * 92 * , 125419 ( 2015 ) ; * 93 * , 159906(e ) ( 2016 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . rev b * 93 * , 245419 ( 2016 ) . g. l. klimchitskaya and v. m. mostepanenko , phys . rev a * 93 * , 052106 ( 2016 ) . k. pyatkovsky , j. phys . : condens . matter * 21 * , 025506 ( 2009 ) . v. p. gusynin , s. g. sharapov , and j. p. carbotte , new j. phys . * 11 * , 095013 ( 2009 ) . s. a. jafari , j. phys . : condens . matter * 24 * , 205802 ( 2012 ) . bo e. sernelius , j. phys . : condens . matter * 27 * , 214017 ( 2015 ) . i. s. gradshtein and i. m. ryzhik , _ table of integrals , series and products _ ( academic press , new york , 1980 ) . m. bordag , i. fialkovsky , and d. vassilevich , phys . b * 93 * , 075414 ( 2016 ) . the real part of the conductivity of gapped graphene at zero temperature normalized to @xmath2 is shown by the solid line as a function of frequency divided by the width of the gap . the dashed line illustrates the condition @xmath121 . ] the imaginary part of the conductivity of gapped graphene at zero temperature normalized to @xmath2 is shown by the solid line as a function of frequency divided by the width of the gap . the dashed line illustrates the condition @xmath121 . ] ( a ) the normalized to @xmath2 real part of the conductivity of graphene with the gap parameter @xmath89ev is shown as a function of frequency by the three lines plotted from top to bottom at temperature @xmath90 , 100 , and 300k , respectively . the dashed line illustrates the condition @xmath121 . ( b ) the same is shown on an enlarged scale at low frequencies . ] the normalized to @xmath2 real part of the conductivity of graphene with the gap parameter ( a ) @xmath93ev and ( b ) @xmath94ev is shown as a function of frequency by the three lines plotted from top to bottom at temperature @xmath90 , 100 , and 300k , respectively . the dashed line illustrates the condition @xmath121 . ] ( a ) the normalized to @xmath2 imaginary part of the conductivity of graphene with the gap parameter @xmath115ev is shown as a function of frequency by the solid line plotted at temperature @xmath91k . the dashed line illustrates the condition @xmath121 . ( b ) the same is shown on an enlarged scale in the vicinity of @xmath121 . ] the normalized to @xmath2 imaginary part of the conductivity of graphene with the gap parameter @xmath97ev is shown as a function of frequency by the solid line plotted at temperature @xmath91k . the dashed line illustrates the condition @xmath121 . ]

the emergence of classical dynamical chaos from more profound quantum mechanics is one of the most intriguing problems in physics @xcite . dynamical chaos in classical mechanics is a special kind of random motion in dynamical systems without any noise and random parameters that is characterized by sensitive dependence on initial conditions in a bounded phase space . the current consensus is that _ isolated bounded quantum systems _ do not show sensitive dependence on initial conditions in the same way as classical systems because their evolution is unitary . so the question is : what is the fundamental mechanism of arising classical chaos from quantum mechanics ? the purpose of this lecture is to show how classical hamiltonian chaos with sensitive dependence on initial conditions , positive lyapunov exponents , fractal properties of underlying phase space , and anomalous statistical characteristics may arise from quantum electrodynamics of a single atom strongly interacting with a quantized mode of the electromagnetic radiation field in a high - quality cavity . the study of the fundamental atom - photon interaction constitutes the rapidly growing field of cavity quantum electrodynamics ( for a review see @xcite ) . the experimental state of the art has reached in this field the stage where the transition from classical to quantum dynamical regimes can now be probed directly . atoms and photons , confined in a high - quality cavity , are ideal objects to study quantum - classical correspondence and quantum chaos . real quantum systems are not isolated . they interact with their environment and in attempts to measure their states with classical measuring devices which should , by virtue of their purpose , be in unstable states . so we are dealing with quantum - classical hybrids . it is not a simple question which degrees of freedom should be considered in a given physical situation as quantum ones and which ones as classical . let us consider an excited atom in a single - mode high - quality cavity whose frequency is close to the frequency of one of the atomic electrodipole transitions . the atom emits a photon in the cavity mode and goes to a lower - lying state . under the conditions of the strong atom - field coupling , the photon may be reemitted and reabsorbed by the atom many times . sooner or later , the atom will emit a photon in one of the other modes of the electromagnetic field that are not sustained by the cavity ( for example , transverse modes in an open fabry - perot cavity ) and the photon will be lost in a surrounding , the process known as spontaneous emission . in real experiments , the cavity relaxation due to the losses in the cavity walls can not be neglected . it is usually modeled by a weak coupling of the selected cavity mode with a bath of harmonic oscillators ( spanning a wide frequency range ) in thermal equilibrium at a given temperature . sooner or later , the atom will relax to a lower - lying state that is not resonant with the cavity mode . it seems to be reasonable to treat two near - resonant atomic levels , strongly coupled to a cavity mode , as a quantum dynamical system weakly coupled to the other electromagnetic modes , the other atomic levels and the cavity walls . an infinite number of the respective degrees of freedom , which are coupled to the quantum dynamical system s degrees of freedom but are not affected by them , forms an inexhaustible external reservoir ( that may be treated classically ) usually called the environment " . it is a kind of the external coupling which is inevitably present in reality . there is another kind of coupling called the dynamical coupling . for example , if the average number of photons in the cavity mode is sufficiently large , one may treat the field as a classical object dynamically coupled to quantized atoms ( the semiclassical approximation ) and take into account the feedback effect of the atoms on the radiation field . the semiclassical approximation breaks down when we deal with single cold atoms and photons in a cavity . in this situation we should take into account the translational ( external ) atomic degree of freedom . when a cold atom emits and absorbs photons its momentum and position may vary significally due to the recoil effect . the external atomic degree of freedom , which may be treated classically if the values of the atomic momentum are greater than the photon momentum , is dynamically coupled to the internal atomic and field degrees of freedom which are treated quantum mechanically . we show in this lecture that , even in the absence of any other interaction with the environment , the interaction of the quantum atom - field system with the external atomic degree of freedom provides , under appropriate conditions , the emergence of classical chaos from quantum electrodynamics . the discovery that a single - mode laser , a symbol of coherence and stability , may exhibit deterministic chaos is especially important not only because lasers are one of the main instruments in physics but lasers , as well , provide almost ideal systems to test general ideas in quantum mechanics and statistical physics . from the standpoint of nonlinear dynamics , laser is an open dissipative system that transforms an external excitation into a coherent output in the presence of loss . some manifestations of strange attractors and dissipative chaos have been observed with different types of lasers ( for a review , see @xcite ) . a collection of identical two - level atoms , interacting with a single - mode electromagnetic field , provides the simplest model for laser dynamics . because of a large number of atoms ( and photons ) , laser dynamics can be adequately described in the so - called semiclassical approximation , where one treats atoms as two - level objects , which may be described by the bloch equations for two components of the collective electrodipole polarisation and collective atomic population inversion , interacting with a field mode to be governed by the classical maxwell equations for the field strengths with the right hands depending on the atomic polarisation . with a single - mode homogeneously broadened laser , operating in resonance with the gain center , these five equations can be reduced to three real - valued equations for slowly varying amplitudes which have been shown to be equivalent to the well - known lorentz model for fluid convection @xcite . practically in the same time , ideas of dynamical chaos have been explored with fundamental models of the matter - radiation interaction , comprising of a collection of two - level atoms _ interacting with their own radiation field _ in a perfect single - mode cavity without any loss and external excitation . it has been shown theoretically and numerically @xcite that the semiclassical maxwell bloch equations , following from the dicke hamiltonian @xcite , may demonstrate _ hamiltonian semiclassical chaos _ if one goes beyond the so - called rotating - wave approximation , i. e. if one takes into account energy non - conserving terms in the hamiltonian ( for details , see the beginning of the next section ) . this mechanism of arising hamiltonian semiclassical chaos is rather weak under realistic assumptions . on the other hand , rotating - wave approximation suppresses chaos with atoms at rest due to existence of an additional integral of motion , the conserving interaction energy . it has been shown in a series of our papers @xcite that hamiltonian chaos may arise within the rotating - wave approximation under conditions of a modulation of the atom - field interaction . in a natural way it occurs when atoms move through a cavity in a direction along which the cavity sustains a standing - wave field that is periodic in space @xcite . when moving with a constant velocity , atoms `` see '' the field whose strength changes in time periodically . it breaks down the interaction - energy integral and may cause hamiltonian semiclassical chaos of a homoclinic type @xcite . there are another ways of modulations to be considered in @xcite . the modulation of the detuning between the atomic transition frequency and the frequency of the field mode has been shown to produce parametric instability and hamiltonian semiclassical chaos @xcite . structural hamiltonian chaos @xcite may arise as a result of a harmonic modulation of cavity length which causes the respective oscillations of the nodes of a standing - wave . generally speaking , the model of the atom - field interaction should involve not only the internal atomic and field degrees of freedom but also the center - of - mass motion of the atom . when emitting and absorbing photons , atoms not only change their internal states but their position and momentum are changed as well due the photon recoil effect . this effect may be neglected if one deals with rydberg atoms interacting with a microwave field in a cavity ( as it has been done in @xcite ) or with thermal usual atoms interacting with a visible light . it has been theoretically and numerically shown in @xcite that hamiltonian semiclassical chaos may arise with cold atoms with recoil in a standing - wave microcavity . in trying to describe adequately dynamics of single atoms and photons in a high - quality cavity , one should go beyond the semiclassical approximation and treats the atom - photon interaction on a quantum ground . the fully quantum model of the interaction between a single two - level atom ( without a recoil ) with a single - mode quantized field in an ideal cavity is known as the jaynes - cummings model @xcite . it describes the atom - field system as a quantum - electrodynamical object whose evolution in time is ( quasi)periodic . in schrdinger picture , it may be described by an infinite set of linear ordinary differential equations for the probability amplitudes to find the atom in the ground / excited state and the field in the state with @xmath0 photons , where @xmath0 runs from zero to infinity . when adopting the semiclassical approximation , we reduce , by hook or by crook , this infinite set to a small number of equations for atomic and field variables . in fact , we decouple the quantum atom - field system into an atom and field parts which may exchange excitations with each other . in doing so , we get automatically products of the atom and field variables in the equations of motion . the respective maxwell bloch equations may , under appropriate conditions , produce chaos in the classical sense of sensitive dependence on initial conditions in the reduced classical phase space spanned by the atomic and field expectation values . it resembles the procedure of deriving the famous lorenz equations from an infinite hierarchy of mode equations in fluid convection by reducing it to three main modes only . the reduction of an infinite set of linear equations , that are not chaotic in the classical sense , to a finite set of nonlinear equations , that may be chaotic , makes an impression that it is only a mathematical trick . it seems that it is not only a useful trick enabling us to handle with the equations of motion but the reduction is a model of processes that may occur in nature . coherence loss caused by inevitable interaction with environment ( decoherence ) breaks quantum unitarity suppressing some quantum properties of motion and revealing its classical properties . such a situation can be modeled with the use of a quantum - classical hybrid , a system with quantum and classical degrees of freedom dynamically coupled to each other . which part of the whole system under consideration is quantized and which one is treated classically depends on the physical situation . if we have a large number of atoms ( or photons ) we may adopt with a good accuracy the semiclassical approximation treating the cavity field as a classical one . when considering a single atom with a recoil in a high - quality microcavity it is necessary to quantize the internal electronic atomic and field degrees of freedom but the translational atomic motion may be treated classically if the values of the atomic momentum are greater than the photon momentum . the interaction between matter and radiation , commonly evidenced by spontaneous emission and induced emission and absorption of photons by atoms , is one of the most fundamental dynamical interactions in nature . in free space , an excited atomic state decays irreversibly because an infinity of vacuum electromagnetic states is available to the radiated photons which disappears in the free - space mode continuum . the situation may be changes if the mode structure and the density of modes are modified by placing the radiating atom into a high - quality cavity . a cavity - induced enhancement and inhibition of spontaneous emission has been observed in a number of experiments in different ranges of the electromagnetic field , from microwaves to visible light ( for a review see @xcite ) . in a near - resonant ideal cavity with a single - mode frequency closed to the atomic transitional frequency , the radiation emitted by the atom is reflected at the cavity walls and is reabsorbed by the atom many times before it dissipates . in particularly , spontaneous emission becomes reversible . in cavity quantum electrodynamics near - resonant interaction of a single atom with a single - mode cavity field is commonly modeled by the hamiltonian @xmath1 where @xmath2 represents the free motionless two - level atom with @xmath3 being the energy separation between the electronic states , the ground @xmath4 and excited @xmath5 ones . the pauli spin operator acts on the states as @xmath6 the hamiltonian @xmath7 represents the mode of the isolated field . the photon annihilation @xmath8 and creation @xmath9 operators with the commutation rule @xmath10 = 1 $ ] act on the photon number states as follows : @xmath11 the electric dipole interaction between the atom and the field mode is described by the operator @xmath12 where @xmath13 is the atom - field coupling constant at the atomic position @xmath14 in the one - dimensional cavity , which is known under the names `` vacuum or single - photon rabi frequency '' . the pauli spin operators @xmath15 with the commutation rules @xmath16=\mp 2\hat \sigma_\pm$ ] , @xmath17=\hat \sigma_z$ ] describe transitions between the ground and excited electronic states @xmath18 the first form of the hamiltonian ( [ 6 ] ) follows from the quantization of the classical electric dipole atom - field interaction by replacing the atomic dipole moment by the operator @xmath19 and the classical electric field by the operator @xmath20 . the products @xmath21 and @xmath22 describe the transition of the atom from the ground ( excited ) to the excited ( ground ) state and simultaneous annihilating ( creating ) a photon in the field mode , respectively . they are usual energy - conserving processes . the product @xmath23 describes the transition of the atom from the ground to the excited state and simultaneous creating a photon in the mode . the reverse process of simultaneous atom and field de - excitation is described by the term @xmath24 . the last two processes are not energy - conserving and may be neglected while we deal with a single atom . omitting the energy - nonconserving terms defines the jaynes - cummings model @xcite which was proposed originally in 1963 as a purely theoretical tool for studying fundamentals of the interaction between matter and radiation . it seemed to be far from reality because neither the atom nor the cavity mode can hardly be perfectly isolated from environment with those - days technique . exciting progress in experimental techniques has drastically changed the situation . experiments , using very high - quality microwave cavities with @xmath25 and optical microcavities with @xmath26 , have now achieved the exceptional circumstance of strong coupling between atoms and a cavity field with the strength of the coupling exceeding the atomic and cavity decays that provides manipulations with single atoms and photons and experimental proving of some predictions of the jaynes - cummings model @xcite . the art of experimentalists and theoretical efforts have opened exciting perspectives for realizing gedanken experiments on fundamentals of quantum mechanics and implementing quantum communications and computing . as we will show in this lecture , an atom , interacting with a radiation field in a high-@xmath27 cavity , provides an example of strongly - coupled microscopic nonlinear system whose dynamics may be very complicated and even chaotic . the jaynes - cummings hamiltonian has the form @xmath28 where @xmath29 is an amplitude value of the vacuum rabi frequency and @xmath30 is a shape function of the cavity mode . let us expand the atom - field state vector over the electronic atomic states and fock ( or photon - number ) field states @xmath31 @xmath32 where @xmath33 are probability amplitudes to find the atom in the excited ( ground ) state and @xmath0 photons in the field mode , respectively . substitution of eqs.([8 ] ) and ( [ 9 ] ) into the schrdinger equation @xmath34 results in an infinite - dimensional set of coupled linear equations for the complex - valued probability amplitudes @xmath35,&\\ \dot b_{n+1}&=\phantom{- } i\left[\delta_b b_{n+1}^ * + \sqrt{n+1}\,f(x ) a_n^ * \right],&\\ \end{aligned } \label{11}\ ] ] where dot denotes differentiation with respect to dimensionless time @xmath36 , @xmath37 , and @xmath38/\omega_0 $ ] . it is convenient to define new real - valued combinations of the probability amplitudes @xmath39 for which we get from ( [ 11 ] ) the quantum bloch - like equations of motion @xmath40 where @xmath41 is the dimensionless detuning between the atomic transition , @xmath42 , and the field - mode , @xmath43 , frequencies . for each specified photon number @xmath0 , the quantity @xmath44 is conserved . two global integrals of motion @xmath45 reflect conservation of the total energy and of the total probability , respectively . eqs.([13 ] ) can be easily solved for each @xmath0 in terms of trigonometric functions ( we assume that @xmath30 does not depend on time ) . the general solution of the schrdinger equation is the sum of the solutions of an infinite set of the independent bloch - like equations . the exact general solution for one of the measured quantities , the atomic population inversion @xmath46 is the following : @xmath47 where @xmath48 is the @xmath0-photon raby frequency , @xmath49 , @xmath50 and @xmath51 are the respective initial values which are determined by the initial atomic and field states . as it is seen from ( [ 16 ] ) , the solution with initially sharply defined atomic and field energy states ( @xmath4 or @xmath5 and @xmath31 ) describes a periodic exchange of one quantum of energy between the atom and the field . the fock state @xmath31 with a specified number of photons in the mode is an exotic field state . in general , a pure quantized field state is an infinite superposition of the photon - number states @xmath52 where @xmath53 is the probability for observing @xmath0 photons . the most classical of single - mode quantum states is a coherent state @xmath54 whose photon probabilities follow a poisson distribution with the mean number of photons @xmath55 and the root mean square spread @xmath56 . evolution of the population inversion of an atom , @xmath57 , in the field , initially prepared to be in a coherent state , contains all the rabi frequencies @xmath58 . the incommensurability of rabi frequencies for different @xmath0 inevitably washes out the periodicity of population transfer resulting in a collapse of the excited - state occupation probability . since the frequencies @xmath58 form a * ( a ) * + * ( b ) * + discrete set , the signal reappears after a time ( see fig . [ fig1]a computed with @xmath59 , @xmath60 and with the initial conditions ( [ 33 ] ) ) . the collapse - revival phenomenon , a direct demonstration of quantum nature of the radiation field , has been demonstrated experimentally @xcite . let us consider as well another pure field state @xmath61 whose photon probabilities follow a bose - einstein distribution @xmath62 collapses and revivals also occur with this field state ( see fig . [ fig1]b computed with @xmath59 , @xmath60 and with the initial conditions ( [ 34 ] ) ) , but the collapse time is much shorter as compared with a coherent state because the bose - einstein spread in the photon number is far larger than for a coherent field with the same mean number of photons . [ fig1 ] demonstrates that the simple jaynes - cummings model with @xmath59 may produce very complicated time evolution . however , the respective oscillations , of course , are not chaotic in the classical sense of exponential sensitivity to small changes in initial conditions . it is clear from the very structure of the equations of motion ( [ 13 ] ) which are infinite - dimensional but linear . in classical mechanics the phase space is continuous and states of a classical system can be arbitrarily close to each other providing a possibility of exponential diverging of initially close trajectories in a bounded region of the phase space . in quantum mechanics there is no notion of a trajectory , and the quantum phase space is not continuous due to the heisenberg uncertainty principle . evolution of _ an isolated quantum system _ is unitary and no dynamical chaos in the classical sense is possible with such systems . the notion `` quantum chaos '' just refers to the behavior of quantum systems whose classical counterparts behave chaotically @xcite . real quantum systems are not isolated , they interact with their environment and , under attempts to measure their states , with classical devices which , by virtue of their purpose , should be in unstable states . loosing of coherence , due to inevitable interaction with environment ( decoherence ) , breaks down quantum unitarity suppressing quantum properties of motion and manifesting classical ones . a quantum rotor with periodic kicks provides an illustrative example of suppressing quantum dynamical localization when one takes into account a thermal `` bath '' in the respective equations of motion @xcite . in the next section we will elaborate these ideas with an extended version of the strongly coupled atom - field system considered above . in the process of emitting and absorbing photons of the cavity - field mode , the atom not only changes its internal electronic states but its external translational state is changed as well due to the photon recoil effect . in this section we consider a single two - level atom with mass @xmath63 moving in an ideal cavity which sustains a single standing - wave mode along the axis @xmath14 with the wave vector @xmath64 and the shape function @xmath65 . the respective hamiltonian is the following extension of the jaynes - cummings hamiltonian ( [ 8 ] ) @xmath66 where the momentum @xmath67 and position @xmath68 operators satisfy the standard commutation relation @xmath69=i\hbar$ ] . operators , belonging to different degrees of freedom , commute with each other at the same time moment . we have now three degrees of freedom , the internal atomic and the field ones , and the external ( or translational ) atomic degree of freedom . the first two degrees of freedom are treated as fully quantum ones in the schrdinger picture . in fact , there are an infinite number of quantum degrees of freedom ( see eq . ( [ 13 ] ) ) which are entangled . the external degree of freedom will be treated as the classical one that may be justified by large values of the atomic momentum as compared with the photon momentum @xmath70 . the hamilton equations of motion for the classical external degree of freedom is easily found from the hamiltonian ( [ 22 ] ) @xmath71 where @xmath72 denotes an expectation value of the corresponding operator over a quantum state @xmath73 of the electronic - field hamiltonian . using the normalizations @xmath74 , @xmath75 , and @xmath76 , we get from ( [ 23 ] ) @xmath77 where @xmath78 is the normalized recoil frequency which characterizes the average change in kinetic energy of the atom , @xmath79 , in the process of emission and absorption of a photon . after computing the expectation value in ( [ 24 ] ) with the state vector ( [ 9 ] ) , we obtain our basic hamilton - schrdinger equations @xmath80 which describe the atom - field quantum - classical hybrid with the quantum degrees of freedom dynamically coupled to the classical degree of freedom . this infinite set of _ nonlinear ordinary differential equations _ possesses an infinite number of the integrals of motion , the total energy integral @xmath81 the bloch - like integral for each @xmath0 @xmath82 and the global integral @xmath83 the infinite - dimensional nonlinear dynamical system ( [ 25 ] ) is a quantum generalization of the five - dimensional semiclassical set of the equations of motion for the same problem that was derived in refs @xcite with a classical field . the later one was shown @xcite to be chaotic with positive values of the maximal lyapunov exponent in some ranges of the system s control parameters . it is rather difficult to compute this quantitative indicator of dynamical chaos with an infinite dimensional set of ode s . however , we have found another signatures of hamiltonian chaos with eqs . ( [ 25 ] ) , the main of which are fractals which will be demonstrated in the next section . * ( a ) * + * ( b ) * + let us compare the rabi - oscillation signals with the atom - field quantum - classical hybrid ( [ 25 ] ) and with the fully quantum jaynes - cummings model without recoil ( [ 13 ] ) . in fig . [ fig2]a and b the population inversion @xmath57 ( see eq . ( [ 15 ] ) ) for an initially excited atom interacting in resonance with the field that is initially in a coherent state ( [ 19 ] ) and in a bose - einstein state ( [ 20 ] ) , respectively , is shown . this figure should be compared with fig . [ fig1 ] where the rabi oscillations with the jaynes - cummings model have been computed for the same initial conditions and the field states with @xmath84 but with a motionless atom . more pronounced collapses and revivals occur with the quantum - classical hybrid as compared with the fully quantum model without recoil . in order to understand peculiarities of the rabi oscillations in fig . [ fig2 ] look at fig . [ fig7]a demonstrating the scheme of gedanken experiments . an atom starts at @xmath85 and moves to the right with the initial momentum @xmath86 . it reaches the first node of the standing wave , where the coupling coefficient with the field mode is zero at time moment @xmath87 in dimensionless units ( we choose @xmath88 in our computer simulations ) . for both the field states , one can see the respective slowing down of the rabi oscillations at the moments of time @xmath89 , when the atom transverses the @xmath90-th node . the pronounced peaks of revivals of the rabi oscillations occur at the moments @xmath91 , when the atom transverses the @xmath92-th antinode of the standing wave where its coupling with the field is maximal . in spite of a complicated character of the rabi oscillations , shown in fig . [ fig2 ] , they are regular because at exact resonance , @xmath93 , all the values of the real - valued amplitudes @xmath94 are conserved during the evolution . with initially excited atom we have @xmath95 for each @xmath0 . it immediately follows from eqs.([25 ] ) that the atom moves with a constant velocity through the cavity . it means that @xmath14 is a linear function of time , and the hamilton - schrdinger equations([25 ] ) reduce to the periodically modulated linear bloch - like equations ( [ 13 ] ) with periodic solutions . the exact solution for the atomic population inversion can be easily found @xmath96 it is a periodic function ( shown in fig . [ fig2 ] ) with the period to be equal to @xmath97 and the maxima at @xmath98 @xmath99 . out off resonance , @xmath100 , the quantum - classical hybrid may demonstrate chaos . to diagnose chaos it is instructive to compute the maximal lyapunov exponent @xmath101 whose values depend on initial conditions , on the detuning @xmath102 , the mean number of photons in the mode , and on the recoil frequency @xmath103 . if the quantized field is initially prepared in the fock state @xmath31 with exactly @xmath0 quanta in the mode , the infinite - dimensional set ( [ 25 ] ) reduces to 8 equations @xmath104 with initial conditions @xmath105 describing the atom - field system initially prepared in a superposition state @xmath106 the so - called topographic @xmath101-maps , showing by color modulation values of @xmath101 in the ranges of values of two control parameters with the third one to be fixed , have been computed with the set ( [ 29 ] ) . in fig . [ fig3 ] we show the @xmath101-map with initial zero atomic population , @xmath107 , @xmath108 , in the ranges of the detuning @xmath102 and the recoil frequency @xmath103 with the fixed value of the initial number of photons @xmath109 . at exact resonance , the set ( [ 29 ] ) is integrable and @xmath110 at @xmath93 . in the range @xmath111 , that corresponds to realistic values of the recoil frequency , the maximal lyapunov exponent may be positive and one may expect chaotic atomic motion in the respective ranges of @xmath103 and @xmath102 . another @xmath101-map with initially excited atom , @xmath112 , @xmath113 , demonstrates in fig . [ fig4 ] the values of @xmath101 in dependence on @xmath103 and @xmath0 at @xmath114 in double logarithmic scale . the magnitude of the maximal lyapunov exponent grows , in average , with increasing the initial number of photons in the cavity mode . * ( a ) * + * ( b ) * computing the values of @xmath101 is difficult with the cavity field initially prepared in a superposition state ( [ 18 ] ) ( including coherent and bose - einstein states ) that generates an infinite number of the equations of motion ( [ 25 ] ) . we have found another signatures of chaos . in fig . [ fig5 ] an atomic trajectory , @xmath115 , and oscillations of the atomic momentum , @xmath116 , are shown to illustrate manifestations of chaos in the set ( [ 25 ] ) with the initially coherent field and the excited atom with @xmath117 , @xmath118 , @xmath88 , and @xmath119 . a typical weakly chaotic atomic trajectory represents a kind of a random walking with small oscillations of the atom in potential wells , long ballistic flights of the lvy type with almost constant velocity and erratic turnbacks . just the presence of lvy flights fluenses strongly the statistical properties of the system generating power - like statistical laws that have been extensively studied in the semiclassical approximation in refs.@xcite . * ( a ) * + * ( b ) * fig . [ fig6 ] presents poincar sections of motion in the set ( [ 25 ] ) with five different values of the initial atomic momentum to be projected on the plane of the atomic external variables @xmath120 . [ fig6]a demonstrates an example of chaotic poincar section @xmath121 whereas fig . [ fig7]b presents a regular motion under the same conditions except for the detuning @xmath114 . a feasible scheme for detecting manifestation of chaos with hot two - level rydberg atoms moving in a high - q microwave cavity has been proposed in @xcite . the same idea could be realized with cold usual atoms in a high - q microcavity . consider a 2d - geometry of a gedanken experiment with a monokinetic atomic beam propagating almost perpendicularly to the cavity axis @xmath14 . in a reference frame moving with a constant velocity in the perpendicular direction , there remains only the transverse atomic motion along the axis @xmath14 . one measures atomic population inversion after passing the interaction zone . before injecting atoms in the cavity , it is necessary to prepare all the atoms in the same electronic state , say , in the excited state , with the help of a @xmath122-pulse of the laser radiation . it may be done with _ only a finite accuracy _ , say , equal to @xmath123 for the initial population inversion @xmath124 . the values of the population inversion @xmath125 are measured with detectors at a fixed time moment . if we would work with the values of the control parameters corresponding to the regular atom - field dynamics , we would expect to have a regular curve @xmath125@xmath124 . in the chaotic regime , the atomic inversion at the output can be predicted ( within a certain confidence interval @xmath126 ) for a time not exceeding the so - called predictability horizon @xmath127 which depends weakly on @xmath128 and @xmath126 . since the maximal confidence interval lies in the range @xmath129 and @xmath101 may reach the values of the order of @xmath130 ( see @xmath101-maps ) , the predictability horizon in accordance with the formula ( [ predtime ] ) can be very short : with @xmath101 = 0.5 the predictability horizon @xmath131 may be of the order of 10 in units reciprocal of the vacuum rabi frequency @xmath29 that corresponds to @xmath132 s with the realistic value of @xmath133 rad@xmath134 s@xmath135 . on its initial values @xmath124 with @xmath118 ( a ) at @xmath136 and ( b ) at @xmath137 with the inset showing this dependence at the exact atom - field resonance , @xmath93.,scaledwidth=45.0% ] in the regular regime , the inevitable errors in preparing @xmath128 produce the output errors @xmath138 of the same order . in the chaotic regime , the initial uncertainty increases exponentially resulting in a complete uncertainty of the detected population inversion in a reasonable time . it is demonstrated in fig . [ fig7 ] , where we plot the dependence of the values of @xmath139 at @xmath136 ( fig . [ fig7]a ) and @xmath137 ( fig . [ fig7]b ) on the values of @xmath140 in the chaotic regime with an initially fock field at @xmath118 and @xmath141 . simulation shows that an initial error @xmath142 in preparing the atomic electronic state leads to complete uncertainty @xmath143 in a rather short time . to see the difference , it is desirable to carry out a control experiment at the exact resonance ( @xmath93 ) when the atomic motion is fully regular with any initial values . the dependence @xmath125@xmath124 with @xmath93 is demonstrated in the inset in fig . [ fig7]b with all the other control parameters and initial values being the same . in this section , we treat the atom - photon interaction in a high-@xmath27 cavity as a chaotic scattering problem . * ( a ) * + * ( b ) * let us consider the scheme of scattering of atoms by the standing wave shown in fig . atoms , one by one , are placed at the point @xmath85 with different initial values of the momentum @xmath144 along the cavity axis . for simplicity , we suppose that they have no momentum in the other directions ( 1d - geometry ) . we compute the time the atoms need to reach one of the detectors placed at the cavity mirrors . the dependence of this exit time @xmath145 on the initial atomic momentum @xmath144 is studied under the other initial conditions and parameters being the same . to avoid complications that are not essential to the main theme of this section , we consider the cavity with only two standing - wave lengths . in this subsection the field is supposed to be initially prepared in a fock state . before injecting into a cavity , atoms are suppose to be prepared in the superposition state with @xmath146 , @xmath147 , i. e. in the state with zero population inversion @xmath148 . at exact resonance ( @xmath93 ) with @xmath149 , the optical potential @xmath150 is equal to zero , and the analytical expression for the dependence in question can be easily found to be the following : @xmath151 atoms simply fly through the cavity in one direction with their initial constant velocity and are registered by one of the detectors . out off resonance ( @xmath152 ) , the atomic motion has been numerically found in preceding section to be chaotic with positive values of the maximal lyapunov exponent in the following ranges of the values of the control parameters : the detuning @xmath153 and the recoil frequency @xmath154 . * ( a ) * + * ( b ) * + * ( c ) * fig . [ fig9 ] shows the function @xmath155 with the normalized detuning @xmath118 , the recoil frequency @xmath156 , and the initial number of cavity photons @xmath157 . the exit - time function demonstrates an intermittency of smooth curves and complicated structures that can not be resolved in principle , no matter how large the magnification factor . [ fig9]b shows magnification of the function for the small interval @xmath158 . further magnification in the range @xmath159 shown in fig . [ fig9]c reveals a beautiful self - similar structure . some structures in fig . [ fig9]a that look like fractal are not , in fact , unresolvable and self - similar . magnification of the structure in the range @xmath160 demonstrates quite a smooth function without unresolvable substructures and with only two singular points on the borders of the respective momentum interval . beating in all the structures of the atomic fractal in fig . [ fig9 ] should be attributed to the structure of the hamilton - schrdinger equations ( [ 29 ] ) which describe two atom - field oscillators with slightly different frequencies . the exit time @xmath145 , corresponding to both smooth and unresolved @xmath144 intervals , increases in average with increasing the magnification factor . it follows that there exist atoms never reaching the detectors in spite of the fact that they have no obvious energy restrictions to leave the cavity . tiny interplay between chaotic external and internal dynamics prevents these atoms from leaving the cavity . the similar phenomenon in hamiltonian systems is known as _ dynamical trapping _ @xcite . different kinds of atomic trajectories , which are computed with the system ( [ 29 ] ) , are shown in fig . a trajectory with the number @xmath161 transverses the central node of the standing - wave , before being detected , @xmath161 times and is called @xmath161-th trajectory . there are also special separatrix - like @xmath162-trajectories following which atoms in infinite time reach the stationary points @xmath163 ( @xmath164 ) , @xmath165 , transversing @xmath161 times the central node . these points are the anti - nodes of the standing wave where the force acting on atoms is zero . a detuned atom can asymptotically reach one of the stationary points after transversing the central node @xmath161 times . the trajectory with number 1 , showing in fig . [ fig8]b , is close to a separatrix - like @xmath166-trajectory . the smooth @xmath144 intervals in the first - order structure in fig . [ fig9]a correspond to atoms transversing once the central node and reaching the right detector . the unresolved singular points in the first - order structure with @xmath167 at the border between the smooth and unresolved @xmath144 intervals are generated by the @xmath166-trajectories . analogously , the smooth @xmath144 intervals in the second - order structure in fig . [ fig9]b correspond to the 2-nd order trajectories with singular points between them corresponding to the @xmath168-trajectories and so on . there are two different mechanisms of generation of infinite exit times , namely , dynamical trapping with infinite oscillations ( @xmath169 ) in a cavity and the separatrix - like motion ( @xmath170 ) . the set of all initial momenta generating the separatrix - like trajectories is a countable fractal . each point in the set can be specified as a vector in a hilbert space with @xmath161 integer nonzero components . one is able to prescribe to any unresolved interval of a @xmath161-th order structure a set with @xmath161 integers , where the first integer is a number of a second - order structure to which trajectory under consideration belongs in the first - order structure , the second integer is a number of a third - order structure in the second - order structure mentioned above , and so on . such a number set is analogous to a directory tree address : `` @xmath171a subdirectory of the root directory@xmath172/@xmath171a subdirectory of the 2-nd level@xmath172/@xmath171a subdirectory of the 3-rd level@xmath172/ ... '' . unlike the separatrix fractal , the set of all initial atomic momenta leading to dynamically trapped atoms with @xmath169 seems to be uncountable . * ( a ) * + * ( b ) * nonlinear fractal dynamics implies specific statistical properties of chaotic motion in hamiltonian systems ( for a review see @xcite ) . even a finite - dimensional atom - field system s phase space has a complicated topology whose simplified image is given by projections of poincar sections and fractal scattering functions . tiny interplay between all the degrees of freedom is responsible for dynamical trapping of atoms even in a very short microcavity and for anomalous statistical properties of the fundamental atom - photon interaction . the probability distribution of exit times @xmath173 for @xmath174 events with initial atomic momenta in the range @xmath175 is shown in fig . [ fig10]a in double logarithmic scale . it is close to a poissonian distribution with comparatively short exit times up to @xmath176 and demonstrates an algebraic decay with the characteristic exponent @xmath177 at its tail ( @xmath178 ) . in the range of initial atomic momenta @xmath179 , the respective pdf for @xmath180 events is computed under the same other conditions to be a poissonian - like . * ( a ) * + * ( b ) * + * ( c ) * following to the scheme in fig . [ fig8]a , let us consider in this section scattering of atoms by the standing wave field initially prepared in the coherent state ( [ 19 ] ) . before injecting into a cavity , atoms supposed to be prepared in the excited state @xmath5 with @xmath113 . in this case the initial conditions are @xmath181 fig . [ fig11 ] demonstrates the respective exit - time function @xmath155 with @xmath182 , @xmath88 , and @xmath84 . a coherent state is not a sharply defined state , as a fock one , but a superposition of an infinite number of fock states . we used a truncated basis of @xmath183 fock states for the cavity mode in our simulation . resolution of one of the unresolved structures in fig . [ fig11]a is shown in fig . [ fig11]b . it resembles the respective fragment of the fock fractal in fig . further magnification of the function in fig . [ fig11]b , which is shown in fig . [ fig11]c , reveals again a self - similar structure with smooth and singular zones . we collect an exit time statistics with @xmath184 events by counting atoms with initial momenta in the range @xmath185 reaching the detectors . the plot of the respective histogram of exit times , shown in fig . [ fig12 ] , demonstrates a few local maxima . the corresponding times of exit of atoms can be approximately estimated with the help of the formula ( [ 33 ] ) for resonant atoms . the first maximum around @xmath186 corresponds , mainly , to atoms which , being initially directed to the right with comparatively low momenta @xmath187 , turn back before reaching the central node and are registered by the left detector . it follows from ( [ 33 ] ) that for such atoms @xmath188 at @xmath189 . however , the atoms , to be injected initially with the momentum @xmath190 and registered by the right detector , may contribute to the first maximum as well because , after transversing the central node , they can be accelerated and gain the values of the momentum @xmath191 up to @xmath192 . the second local maximum around @xmath193 corresponds , mainly , to atoms with @xmath194 which transverse the central node and are registered by the right detector for the time @xmath195 . the other local maxima of the pdf are not so pronounced as the first ones , they are formed by the atoms transversing the central node a few times . the pdf in fig . [ fig12 ] demonstrates an exponential decay at the tail up to the exit times @xmath196 . in conclusion of this section we present in fig . 13 the fractal computed with the field initially prepared in a bose - einstein state ( [ 20 ] ) and atoms supposed to be prepared in the excited state @xmath5 with @xmath113 . the initial conditions are @xmath197 the values of the control parameters are the same as for the coherent fractal . fig . 13 demonstrates again a self - similar structure of the exit - time distribution . we want to stress that the hamilton - schrdinger sets of equations generating the fock and the other fractals differ strongly in the number of equations . nevertheless , the respective @xmath155 functions are rather similar in their main features . the atomic fractals we have found are generated by dynamical chaos in the atom - photon interaction but not by noise . * ( a ) * + * ( b ) * + * ( c ) * + [ fig13 ] we have studied in the strong - coupling regime the fundamental interaction between a two - level atom with recoil and a quantized radiation field in a single - mode cavity in the mixed quantum - classical formalism modelling quantum evolution of the electronic - field purely quantum system to be disturbed by translational atomic motion . we have managed to derive an infinite set of the hamilton - schrdinger nonlinear ode s for real valued electronic - field probability amplitudes and expectation values of the atomic center - of - mass position and momentum . these equations of motion are able to demonstrate known features of the atom - field quantum evolution in the strong - coupling limit such as collapses and revivals of the atomic population inversion modulated by atomic motion through the nodes and antinodes of the standing wave . we have shown that even in the absence of any other interaction with environment the hamiltion - schrdinger dynamical system provides the emergence of classical hamiltonian dynamical chaos from cavity quantum electrodynamics . we have investigated the atom - photon nonlinear dynamics with fock , coherent and bose - einstein quantum states of the initial cavity field . positive values of the maximal lyapunov exponent @xmath101 have been found with reasonable values of the control parameters , the detuning of the atom - field resonance , @xmath102 , the atomic recoil frequency , @xmath103 , and the initial mean number of photons . exponential sensitivity to initial conditions may manifest itself in the chaotic dependence of the output values of the atomic population inversion on its input values that may , in principle , be measured in real experiments . new manifestations of hamiltonian chaos in the atom - photon interaction , we have found numerically in the dependence of the atomic exit times on the initial momentum , are atomic fractals . they were classifield by the names of the respective quantum field states , i.e. , the fock , coherent and bose - einstein fractals . we have found anomalous statistical properties of the chaotic atom - photon interaction : lvy atomic flights through a cavity and a power law at the tail of the exit - time distribution function in the fock field . power - law tails should appear ( at least , in some ranges of the control parameters ) in the pdf s with coherent and bose - einstein fields , as well , but their detection would require a longer computation time than we have used . i am grateful to leonid konkov and michael uleysky for preparing the figures . this work was supported by the russian foundation for basic research under grant no . 020217796 , by the program mathematical methods in nonlinear dynamics " of the russian academy of sciences , and by the program for basic research of the far eastern division of the russian academy of sciences . chirikov , phys . 52 ( 1979 ) 263 . zaslavsky , phys . 80 ( 1981 ) 157 . f. haake , _ quantum signatures of chaos _ , springer , berlin ( 1991 ) . gutzwiller , _ chaos in classical and quantum mechanics _ , springer , new york ( 1990 ) . , ed . by p.r . berman , academic press , new york ( 1994 ) . h. walther , phys . rep . 219 ( 1992 ) 263 . j. m. raimond , m. brune , s. haroche , rev . 73 ( 2001 ) 565 . w. vogel , d .- welsch , s. wallentowitz , _ quantum optics : an introduction _ , wiley - vch , berlin ( 2001 ) . ackerhalt , p.w . milonni , m .- shih , phys . 128 ( 1985 ) 205 . harrison , contemp . 29 ( 1988 ) 341 . khanin , _ foundations of laser dynamics _ , nauka , moscow ( 1999 ) [ in russian ] . h. haken , phys . lett . a 53 ( 1975 ) 77 . belobrov , g.m . zaslavskii , g.kh . tartakovskii , zh . 71 ( 1976 ) 1799 [ sov . phys . jetp 44 ( 1976 ) 945 ] . dicke , phys . 93 ( 1954 ) 493 . prants , l.e . konkov , pisma zh . 65 ( 1997 ) 801 [ jetp lett . 65 ( 1997 ) 833 ] . prants , l.e . konkov , phys . lett . a 225 ( 1997 ) 33 . prants , l.e . konkov , zh . eksp . teor . ( 1999 ) 740 [ jetp 88 ( 1999 ) 406 ] . prants , l.e . konkov , i.l . kirilyuk , phys . e 60 ( 1999 ) 335 . prants , l.e . konkov , phys . e 61 ( 2000 ) 3632 . ioussoupov , l.e . konkov , s.v . prants , physica d 155 ( 2001 ) 311 . s.v . prants , l.e . konkov , pisma zh . 73 ( 2001 ) 200 [ jetp lett . 73 ( 2001 ) 180 ] . prants , v.yu . sirotkin , phys . a 64 ( 2001 ) 033412 . jaynes , f.w . cummings , proc . ieee 51 ( 1963 ) 89 . raimond , s. haroche , in : _ confined electrons and photons _ , ed . by e. burstein and c. weisbuch , plenum press , new york , 1995 . g. rempe , h. walther , n. klein , phys . 58 ( 1987 ) 353 . prants , pisma zh . eksp . ( 2002 ) 777 [ jetp letters 75 ( 2002 ) 651 ] . prants , m. edelman , g.m . zaslavsky , phys . e 66 ( 2002 ) 046222 . argonov , s.v . prants , zh . ( 2003 ) 946 . [ jetp 96 ( 2003 ) 832 ] . zaslavsky , _ physics of chaos in hamiltonian systems _ , academic press , oxford , 1998 .

during the past two decades much attention has been devoted to the study of disordered bulk metals with electron - electron interactions @xcite from the one hand and arrays of normal metallic grains connected by tunnel junctions @xcite from the other . the coulomb interaction plays a very important role in both types of system and many non - trivial physical effects occur due to it . in disordered metals , _ e.g. _ , the coulomb interaction results in quantum corrections to conductivity and reduces the density of states , a phenomenon known as zero bias anomaly . qualitatively similar effects were predicted @xcite and studied experimentally @xcite for arrays of tunnel junctions in the normal state , where a sufficiently strong coulomb interaction can suppress the conductivity ( coulomb blockade of tunneling ) . in spite of the similarities in the physical behavior of disordered metals and normal arrays of tunnel junctions , completely different approaches are used for the theoretical description of these respective systems . as concerns disordered metals with interaction , either diagrammatic methods @xcite or non - linear @xmath0-models @xcite are employed . the starting point of these theories is a model of a weakly disordered bulk metal . on the other hand , for the description of coulomb blockade effects in a junction array or a granular material , a completely different approach based on the theory of ambegaokar , eckern and schn ( aes ) @xcite is commonly used . the free energy functional proposed by aes does not contain any disorder : the finite conductance obtained within this model is due to tunneling of electrons between grains and can be incorporated into the theory through a term describing the so - called quantum dissipation , first introduced phenomenologically in ref . . although both approaches have been introduced almost 20 years ago , only a few indirect attempts have been made to reconcile them . we mention the work by nazarov @xcite and levitov and shytov @xcite that treat the diffusive zero bias anomaly and the coulomb blockade of tunneling on the same footing . more recently , nazarov @xcite showed how coulomb blockade phenomena survive in diffusive systems . indeed , the properties of a granular metal without electron - electron interactions are qualitatively similar to those of a disordered bulk metal ( see e.g. ref . ) and it would be natural to expect that this qualitative similarity persists even in the presence of interactions . in fact , one may conjecture that bulk disordered metals from the one hand , and granular metals or arrays of tunnel junctions from the other , can be described within one unifying scheme . in order to develop such a scheme , we consider in this paper a granular normal metal with coulomb interaction between electrons . the consideration is simplified by taking into account only the long range part of the interaction that leads to coulomb blockade effects . we assume that the macroscopic dimensionless conductance @xmath1 of the granular metal is large , which enables us to develop a perturbative theory with @xmath2 as the small parameter . within this theory , physical quantities of interest can be calculated at arbitrary temperatures . the diagrammatic analysis is supplemented by the derivation of a @xmath3-model that can be considerably simplified provided one does not take into account weak localization effects . we demonstrate below that granular metals and networks of tunnel junctions can always be described using the standard techniques developed in the theory of disordered metals ( expansions in cooperons and diffusons , @xmath3- model calculations ) . with the help of these techniques one can in principle calculate any physical quantity without using the notion of quantum dissipation . disorder is inevitably present in the system even if electrons move ballistically within the grains . in the latter case any small irregularity in the shape of the grains causes the electron motion within the grains to be chaotic and this assumption is sufficient for our theory to be applicable . an integrable shape of the grains does not seem to be realistic , and even if it happened for isolated grains , tunneling from grain to grain would add an additional chaoticity . these ideas about intrinsic disorder in granular metals have been used in previous studies @xcite . we find that , at not too low a temperature , the aes free energy functional gives correct results . however in the limit @xmath4 aes theory is no longer applicable . the region of validity of the aes free energy functional is determined by the inequality @xmath5 , where @xmath6 is the mean level spacing , @xmath7 is the density of states at the fermi surface and @xmath8 is the volume of a single grain . we conclude in particular that the aes notion of quantum dissipation at @xmath9 can only be applied to tunnel junctions connecting infinitely large conductors . if the volume of the conductors is finite the aes picture can be applied at finite temperatures only . apparently , the notion of quantum dissipation should be treated with care . a qualitative discussion of the problem `` dissipation versus dynamical screening '' in the case of superconducting grains can be found in a recent lecture by altshuler@xcite . the remainder of the paper is organized as follows . in sec . [ choice ] we formulate the model and write physical quantities in terms of functional integrals over anticommuting fields . we decouple the coulomb interaction by an integration over auxiliary fields . in sec . [ aes ] we consider the aes action and show that it does not correspond to the physics of normal metals with a screened coulomb interaction . in sec . [ metal ] we develop a diagrammatic technique . we show what kind of diagrams correspond to the aes action and demonstrate that , at low temperatures , additional contributions arise . summing up these additional diagrams we arrive at expressions corresponding to the dynamically screened coulomb interaction of a normal metal . [ sigma ] contains a derivation of a new action for a granular metal with electron - electron interaction which is applicable at arbitrary temperatures . the results are compared with the results obtained diagrammatically and with those obtained on the basis of the aes action . our results are summarized in the conclusion . we consider an array of normal metallic grains coupled to each other . the dimensionality of the array may be arbitrary . the grains are assumed to contain imperfections : there can be impurities inside them as well as on their surface . this implies that the electron motion in the grains is chaotic . the mean level spacing in a single grain is equal to @xmath6 . however , the shapes of the grains need not considerably differ from each other and we assume for simplicity that the grains are arranged in a cubic lattice . the electrons can hop from grain to grain ; moreover , they interact with each other . we are interested only in the long range part of the coulomb interaction and write it in a simplified form describing charging of the grains . in such a formulation the spin of the electrons is not important and can be taken into account at the end of the calculations when writing proper densities . under these assumptions the electron motion can be conveniently described by a functional integral of the type @xmath10\right ) { \cal d}\psi , \ ] ] with an action @xmath11 $ ] that can be written in the form @xmath12=s_{{\rm g}}[\psi ] + s_{{\rm t}}[\psi ] + s_{{\rm c}}[\psi ] . \label{a1}\ ] ] the action @xmath13 $ ] in eq . ( [ a1 ] ) describes the electron motion within the grains and we write it in the form @xmath14=\sum_{i}\int_{0}^{\beta } drd\tau \psi _ { i}^{\ast } ( \tau , r)\left ( \partial _ { \tau } -\mu + \hat{h}_{i}\right ) \psi _ { i}(\tau , r ) , \label{a2}\ ] ] where @xmath15 is a fermion field in grain @xmath16 at imaginary time @xmath17 and coordinate @xmath18 ( @xmath19 is its complex conjugate ) , @xmath20 is the chemical potential , and @xmath21 is the inverse temperature . the fermion fields @xmath22 satisfy the condition @xmath23 the operator @xmath24 in eq . ( [ a2 ] ) includes scattering by impurities and can be written as @xmath25 where @xmath26 is the impurity potential . we assume that it is gaussian distributed with the correlation @xmath27 the action @xmath28 $ ] stands for tunneling between the grains , @xmath29=\sum_{ij}\int_{0}^{\beta } d\tau \psi _ { i}^{\ast } ( \tau ) t_{ij}\psi _ { j}(\tau ) , \label{a4}\ ] ] and @xmath30 $ ] describes the charging of the grains @xmath31=\frac{e^{2}}{2}\sum_{ij}\int_{0}^{\beta } d\tau n_{i}(\tau ) c_{ij}^{-1}n_{j}(\tau ) , \label{a5}\ ] ] where @xmath32 is the density field in grain @xmath16 , @xmath33 is the electron charge , and @xmath34 is the capacitance matrix . ( [ a1 ] - [ a5 ] ) describe the model completely and one can start explicit calculations . due to the coulomb interaction , the action @xmath35 $ ] is not quadratic in @xmath36 and the integration over @xmath36 can not be performed immediately . a convenient way to proceed in such a case is to decouple the term @xmath37 $ ] by a gaussian integration over auxiliary fields . this transformation can be written as follows : @xmath38}={\cal n}\int { \cal d } ve^{-s_{2}[v]}\int { \cal d}(\psi ^{\ast } , \psi ) e^{-s_{1}[\psi , v ] } , \label{a6}\ ] ] where @xmath39 is a normalization factor and @xmath40 is the decoupling field . the action @xmath41 $ ] in eq . ( [ a6 ] ) has the form @xmath42=\frac{1}{2}\int_{0}^{\beta } d\tau \sum_{ij}v_{i}(\tau ) c_{ij}v_{j}(\tau ) , \label{a7}\ ] ] which shows that the variable @xmath43 has the meaning of a voltage on grain @xmath16 at the time @xmath17 . the new effective action @xmath44 $ ] reads @xmath45=s_{{\rm g}}[\psi , v]+s_{{\rm t}}[\psi ] , \ ] ] with @xmath46=\sum_{i}\int_{0}^{\beta } d\tau \psi _ { i}^{\ast } ( \tau ) \left ( \partial _ { \tau } -\mu + \hat{h}_{i}+iev_{i}(\tau ) \right ) \psi _ { i}(\tau ) . \label{a8}\ ] ] one more transformation can be performed exactly . following the procedure of ref . where a single grain was considered , we represent @xmath43 in the form @xmath47 where @xmath48 is the static part and @xmath49 . the replacement @xmath50 with @xmath51 does not violate the condition given by eq . ( [ a51 ] ) and we can remove the variable @xmath52 from the action @xmath53 $ ] , eq . ( [ a8 ] ) . however , this variable appears in the action @xmath54 $ ] as an additional phase of the field @xmath55 . as concerns the static part @xmath48 , its fluctuations can be neglected even in a single isolated grain , provided the temperature is high @xmath56 @xcite . this restriction is not necessary for the system of the coupled grains in the limit of large conductance @xmath57 considered here . using eqs . ( [ a1]- [ a10 ] ) we finally reduce the calculation of physical quantities to the computation of a functional integral of the form @xmath58-s_{1}[\psi , \phi ] -s_{2}[\phi ] \right ) { \cal d } \psi { \cal d}\phi , \label{a11}\ ] ] where @xmath59=\sum_{i}\int_{0}^{\beta } d\tau \psi _ { i}^{\ast } ( \tau ) \left ( \partial _ { \tau } -\mu + \hat{h}_{i}\right ) \psi _ { i}(\tau ) , \label{a12}\ ] ] @xmath60=\sum_{ij}\int_{0}^{\beta } d\tau t_{ij}\psi _ { i}^{\ast } ( \tau ) \psi _ { j}(\tau ) \exp \left ( i\phi _ { ij}\left ( \tau \right ) \right ) , \label{a13}\ ] ] @xmath61=\int_{0}^{\beta } d\tau \sum_{ij}\frac{c_{ij}}{2e^{2}}\frac { d\phi _ { i}(\tau ) } { d\tau } \frac{d\phi _ { j}(\tau ) } { d\tau } , \label{a14}\ ] ] and @xmath62 . eqs . ( [ a11]-[a14 ] ) completely specify the model that will be studied in the subsequent sections . disorder is still present in the operators @xmath63 , eq . ( [ a3 ] ) , but the mean free path within the grains is assumed to be large , @xmath64 , where @xmath65 is the fermi momentum . in this limit , the macroscopic conductivity is determined mainly by the tunneling conductance @xmath1 , @xmath66 where @xmath67 is the tunneling resistance . a functional integral formulation was used in ref . to treat the quantum dynamics of a josephson junction . the action derived in that work ( aes action ) was used later in the context of a normal tunnel junction @xcite , as well as in connection with arrays of tunnel junctions in a number of papers @xcite . the aes action can be obtained from eqs . ( [ a11]- [ a14 ] ) . first , one should average the green function @xmath68 corresponding to the action @xmath69 $ ] , eq . ( [ a12 ] ) , over impurities in the first born approximation , which gives ( in fourier representation ) @xmath70 where @xmath71 is defined in eq . ( [ a30 ] ) and @xmath72 . after that one should make a cumulant expansion in the term @xmath73 $ ] , eq . ( [ a13 ] ) . keeping only the second order of the expansion and assuming for simplicity that the capacitance matrix @xmath74 is diagonal one arrives at the aes action @xmath75 $ ] @xcite , @xmath76=\frac{1}{4e_{c}}\sum\limits_{i}\int\limits_{0}^{\beta } d\tau \left ( \frac{d\phi _ { i}}{d\tau } \right ) ^{2}+\frac{g_{t}}{2 } \sum\limits_{i , j}\int\limits_{0}^{\beta } d\tau d\tau ^{\prime } \alpha ( \tau -\tau ^{\prime } ) \left ( 1-\cos \left ( \phi _ { ij}(\tau ) -\phi _ { ij}(\tau ^{\prime } ) \right ) \right ) , \label{s}\ ] ] where @xmath77 in eq . ( [ s ] ) @xmath78 is the charging energy @xmath79 . the second term in eq . ( [ s ] ) contains the sum over neighboring grains and @xmath80 is the tunneling energy between them . in the limit of small phase fluctuations we can expand the second term in eq . ( [ s ] ) with respect to @xmath81 and obtain @xmath82=\frac{1}{4e_{c}}\sum\limits_{i}\int\limits_{0}^{\beta } d\tau \left ( \frac{d\phi _ { i}\left ( \tau \right ) } { d\tau } \right ) ^{2}+ \frac{g_{t}}{2 } \sum\limits_{i , j}\int\limits_{0}^{\beta } d\tau d\tau ^{\prime } \alpha ( \tau -\tau ^{\prime } ) \left ( \phi _ { ij}(\tau ) -\phi _ { ij}(\tau ^{\prime } ) \right ) ^{2}. \label{s2}\ ] ] using the periodicity of all functions in @xmath83 we write the fourier expansion as @xmath84 where @xmath85 are matsubara frequencies . hence we obtain for the action @xmath82=\frac{1}{4e_{c}}\sum\limits_{i}t\sum\limits_{\omega _ { n}}\omega _ { n}^{2}\mid \phi _ { i}(\omega _ { n})\mid ^{2}+\frac{g_{t}}{\pi } \sum\limits_{i , j}t\sum\limits_{i\omega _ { n}}\mid \omega _ { n}\mid \mid \phi _ { ij}(i\omega _ { n})\mid ^{2}. \label{s3}\ ] ] the second term in eq . ( [ s3 ] ) describing the tunneling is linear in frequency and keeps its form down to @xmath9 . therefore , it was attributed to quantum dissipation @xcite . performing a fourier transformation in space @xmath86 we rewrite the action @xmath87 $ ] as @xmath82=\frac{1}{2}\sum\limits_{{\bf k}}t\sum\limits_{\omega _ { n}}\left ( \frac{\omega _ { n}^{2}}{2e_{c}}+\frac{4g_{t}}{\pi } |\omega _ { n}|\sum_{a}(1-\cos { \bf k}{\bf d}_{a})\right ) |\phi _ { \omega _ { n } } ( { \bf k})|^{2 } \label{smallfluct } , \ ] ] where @xmath88 , @xmath89 , @xmath90 , @xmath91 @xmath92 are vectors connecting the centers of the neighboring grains . the phase - phase correlation function @xmath93 can be immediately calculated using eq . ( [ smallfluct ] ) and we obtain @xmath94 suppose that one assumes , following refs . , that eq . ( [ s ] ) is applicable to a system of tunnel junctions down to @xmath9 . this would imply in particular that eq . ( [ cor ] ) is correct for @xmath95 , too . however , the derivative @xmath96 is proportional to the voltage on the grains which in turn is linearly related to the charge density . multiplying eq . ( [ cor ] ) by @xmath97 , one thus obtains the density - density correlation function @xmath98 . let us consider limit of small frequencies and wave vectors @xmath99 . in this limit , the network should correspond to a bulk normal metal , where one should have the propagator corresponding to a screened coulomb interaction . however , the limit @xmath100 , @xmath101 in eq . ( [ cor ] ) does not reproduce the screened coulomb interaction . how to resolve this discrepancy ? in the next section we carry out an expansion in @xmath102 and demonstrate that keeping only the second order , which led us to eq . ( [ s ] ) is not sufficient and that , in the limit of small frequencies or temperatures , one should sum up an infinite class of diagrams . in section [ sigma ] we will then show how these processes can be included into an effective action formulation structurally similar to the aes approach . in this section we present a detailed derivation of the density - density correlation function @xmath103 summing an infinite series of diagrams . for the computation we use the standard diagram technique @xcite with a modification suggested in refs . to include tunneling between the grains . within this technique one should make an expansion in the tunneling @xmath102 , the phases @xmath104 and in the impurity potential @xmath26 . below we denote the electron green functions by solid lines , the phases @xmath105 by wavy lines and the tunneling elements @xmath102 by crossed circles . impurity propagators corresponding to eq . ( [ a3 ] ) are denoted by dashed lines . before turning to the quantitative discussion , let us briefly outline the skeleton of the analysis . to lowest order in the tunneling matrix elements , the polarization operator of the system is represented by the process depicted in fig . [ grainver ] . this diagram describes the ( absolute square of ) amplitude for tunneling from one grain into the next and the subsequent relaxation inside the target grain ( represented through the green function ) . the quantitative evaluation of this process leads to the coupling constant @xmath106 of the aes approach . on this level , coherence effects associated to mulitiple tunneling and/or impurity scattering are completely neglected ; the tunneled electron forgets about its phase memory implying that the dissipative picture of the aes approach obtains . indeed , the two green functions depicted in fig . [ grainver ] live in different grains which means that no phase correlation is possible . how can this picture change in principle ? including one more order in the tunneling , leads to corrections of the structure shown in fig . [ hightun ] . what makes these processes qualitatively different from those discussed above is that they include the possibility of _ phase coherent _ multiple impurity scattering : the two green functions labeled @xmath16 or @xmath107 scatter off the same impurities and/or chaotic potential where the associated scattering phases may cancel due to the fact that one of the green functions is advanced , the other retarded . the net two - particle modes emanating from such processes , diffusons in the context of disordered systems , are represented through the hatched regions of fig . [ diffuson ] , right . the key question now is , are such processes relevant or not ? a crude estimate can be given as follows : it is known that for low temperatures ( for a precise definition of low , see below ) , each diffuson mode diverges as @xmath108 . the fact that we are considering processes of higher order in the tunneling amplitude introduces one more power of @xmath106 . thus , for @xmath109 , higher order processes become as important as the first order contribution . for lower temperatures , these processes _ must _ be taken into account to obtain a physically correct picture . in the following we will put this discussion onto a firm basis and discuss how the coherent tunneling series can be summed in a controlled way . before starting the computation for a granular system we want to mention an important difference between the diagrams for the current - current and density - density correlation functions . calculating the classical conductivity for a granular system in the lowest order in the tunneling we need not renormalize the current vertices by impurities . to understand this fact let us recall a well known result for bulk metals . the diagram for the drude conductivity is shown in fig . [ vertex ] . the second diagram in fig . [ vertex ] , renormalizing the current vertex , gives an additional integral over momentum that yields zero , @xmath110 in eq . ( [ b1 ] ) , the green function @xmath111 is determined by eq . ( [ a15 ] ) . diagrams with many impurity lines yield zero as well and the current vertices are not renormalized . the same is true for granular metals , which can be seen after a proper replacement of the current @xmath112 as concerns the polarization operator or the density - density correlation function , the situation is different . in this case , instead of vector vertices we have scalar ones and the renormalization due to impurities can be important . expanding the functional integral , eq . ( [ a11 ] ) in the tunneling term @xmath113 , eq . ( [ a13 ] ) , we obtain to second order in @xmath102 and in @xmath105 the diagrams represented in fig . [ grainver ] . the two electron lines relate to different grains . as the impurities in the different grains are not correlated , impurity lines connecting two green functions are do not exist ( although the diagram represents a polarization loop . ) the second and third diagrams in fig . [ grainver ] are equal to each other and have an opposite sign with respect to the first diagram . the analytical result for the sum of the three diagrams reads @xmath114 where the dimensionless tunneling conductance @xmath1 is given by eq . ( [ sg ] ) . we see from eq . ( [ b3 ] ) that , although the green functions @xmath115 determined by eq . ( [ a15 ] ) contain @xmath71 , all information about the disorder drops out of the polarization loop . to calculate the phase - phase correlation function one should notice that the bare propagator corresponding to @xmath116 , eq . ( [ a14 ] ) , has the form @xmath117 and diverges at @xmath118 . therefore , one should sum up the infinite geometrical series represented in fig . [ series ] . = 1.2 cm the summation can be easily performed and we arrive at the propagator @xmath119 , eq . ( [ cor ] ) . summarizing , we conclude that , in order to reproduce the results obtained in the quadratic approximation of the aes action , eqs . ( [ s2 ] , [ s3 ] , [ smallfluct ] ) , one should sum up the diagrams of fig . [ series ] . on the other hand , the form of the propagator @xmath120 does not correspond to the propagator of a normal metal with a screened coulomb interaction ( see _ e.g. _ ) , which signals that something is missing in this analysis . to resolve this paradox one should consider diagrams of higher order in tunneling , which complicates the calculations . first of all , let us note that in order to consider higher order diagrams one should specify tunneling more explicitly . we assume that the radius of the area at which the grains contact each other much exceeds atomic distances ( large number of conducting channels ) . at the same time , the potential barrier between the grains can be arbitrarily large , which can lead to an arbitrary tunneling conductance . such a situation can be conveniently modeled by random tunneling matrix elements @xmath121 correlated as @xmath122 where @xmath121 are written for the same contact . correlations of the matrix elements of different contacts are assumed to be zero . to simplify notations we do not write subscripts numerating the contacts . taking into account eq . ( [ average ] ) is indeed very important . in order to illustrate this fact we consider two diagrams represented in fig . [ hightun]a . we see immediately the difference between them : the first diagram is not averaged over @xmath121 and can contains four different momenta . after averaging over @xmath121 one obtains the second diagram containing only three different momenta . as it has been mentioned , we will expand not only in the tunneling amplitude @xmath123 but also in the phases @xmath124 . performing the expansion in the phases and denoting them by wavy lines we have three different possibilities as depicted in fig . [ hightun]b . now we can start averaging over impurities . as usual @xcite , we average first the green functions and obtain eq . ( [ a15 ] ) for the average green function @xmath115 . then , we can consider averages with impurity lines connecting the green functions on opposite sides of the diagrams in fig . [ hightun]b . diagrams with non - intersecting impurity lines are most important and we recognize the well - known diffusons . in the present paper we consider the limit of sufficiently small grains , such that all relevant energies like temperature @xmath125 , frequency @xmath100 , tunneling energy @xmath80 , etc . are much smaller than the thouless energy @xmath126,where @xmath127 is the diffusion coefficient of a single grain and @xmath128 is the radius of the grain . in this limit the spectrum of the diffuson in a single grain is zero - dimensional ( no dependence on momenta inside the grain ) . coupling between the grains leads to a dependence on quasi - momentum @xmath129 . diagrams that should be summed in order to give the complete form of the diffuson are represented in fig . [ diffuson]a . now let us present some details of the calculation of the diagrams in fig . [ hightun ] . the different classes of diagrams contributing to the polarization operator of the granular metals before impurity averaging are shown in fig . [ hightun]b . before impurity averaging the sum over matsubara frequencies for the first diagram in fig . [ hightun]b is @xmath130 writing this expression we used eq . ( [ average ] ) . after impurity averaging of the first diagram in fig . [ hightun]b , we obtain the diagram represented in fig . [ diffuson]a . now the disorder averaging can be done and we obtain , in a standard way , the average of the product of two green functions for a granular metal , @xmath131 . here @xmath133 being the diffuson for a single isolated grain . the equation for the complete diffuson @xcite is shown in fig . [ diffuson]b . the self - energy @xmath134 in eq . ( [ dyson ] ) is the sum of the three diagrams in fig . [ diffuson]c . the second and third diagrams in fig . [ diffuson]c are equal to each other and have the opposite sign with respect to the first diagram . solving eq . ( [ dyson ] ) we obtain @xmath135 the calculation of the second diagram in fig . [ hightun]b is analogous to the calculation of the first one . the third diagram in fig . [ hightun]b is independent ofthe frequency @xmath136 and yields a constant . the diagram in fig . [ diffuson]a contains only one diffuson . more complicated diagrams can be drawn and one of them is depicted in fig . [ dif ] . = 2.5 cm however , all such many - diffuson contributions can be neglected provided the tunneling conductance @xmath1 , eq . ( [ sg ] ) , is large . a direct calculation of this diagram shows that it involves the small parameter @xmath137 . this is similar to expansions in the diffusion modes for a bulk metal where the inverse conductivity is the expansion parameter . the condition @xmath57 means that we are far from the anderson metal - insulator transition . the result of averaging and summation of the diagrams represented in fig . [ hightun ] can be written as @xmath138 we note here that the functions @xmath139 and @xmath140 have a different sign . comparing the function @xmath141 , eq . ( [ newdif ] ) , with the quantum dissipation part @xmath142 , eq . ( [ b3 ] ) , represented in fig . [ grainver ] we see immediately that the contribution @xmath141 , eq . ( [ newdif ] ) , can be neglected only if the temperature @xmath125 is sufficiently high @xmath143 ( we recall that the static component of the phase @xmath124 has been neglected and therefore @xmath144 in eq . ( [ newdif ] ) . so , the notion of quantum dissipation is applicable essentially at non - zero temperatures only for a granular material . however , it can be used for a contact connecting two bulk metals with an infinite volume . at low temperatures @xmath145 , there is no reason to neglect the contribution @xmath146 , eq . ( [ newdif ] ) , and its presence changes completely the form of the propagator . adding the functions @xmath147 , eq . ( [ b3 ] ) , and @xmath148 , eq . ( [ newdif ] ) , we obtain the total polarization @xmath149 @xmath150 proper diagrams contributing to this function are represented in fig . [ complete ] . = 2.5 cm now , we can write the final expression for the phase correlation function @xmath151 that should be written instead of @xmath120 , eq . ( [ cor ] ) . using the bare propagator @xmath152 , eq . ( [ b4 ] ) , and the self - energy part @xmath153 , eq . ( [ total ] ) , we write the phase propagator @xmath154 in the form @xmath155 using the relation between the phase @xmath124 and the voltage @xmath8 , eq . ( [ a10 ] ) , we derive an equation for the effective coulomb interaction @xmath156 on the other hand , the dynamically screened coulomb interaction in a disordered bulk metal can be written in the form @xmath157 where @xmath158 is the bare coulomb potential . we see that eqs . ( [ f1 ] , [ f2 ] ) written in the continuum limit correspond to eq . ( [ f3 ] ) . thus , we conclude that , when calculating physical quantities , the main contribution is due to small @xmath99 only in the limit @xmath159 . provided this inequality is satisfied , the properties of the tunnel junction array are equivalent to the properties of a weakly disordered bulk metal with coulomb interaction . in the opposite limit @xmath160 , one can describe the system in the `` quantum dissipation '' approximation . calculations were performed in this section using a diagrammatic expansion in the lowest order in the field @xmath124 . on the other hand , the aes free energy functional , eq . ( [ s ] ) , is written for arbitrary values of @xmath124 . can one generalize eq . ( [ f2 ] ) to arbitrary values of @xmath124 ? we will try to do this in the next section , deriving a simplified version of a replica @xmath3-model and using a non - trivial saddle point . in the preceding sections we analyzed effects of the coulomb interaction on the electron motion in granular materials using a diagrammatic expansion . this approach works well when fluctuations of the phase @xmath124 or , in other words , fluctuations of the voltages on the grains are small . however , one can imagine situations when the fluctuations are large . for example , one can try to consider non - perturbative excitations like instantons @xcite corresponding to an integer charge transfer between grains . in principle , the aes action , eq . ( [ s ] ) , can describe such excitations very well but , as we have seen previously , it can only be used at sufficiently high temperatures @xmath160 . so , our task now is to derive an action that will be applicable at lower temperatures . the derivation at temperatures @xmath145 is more complicated because now we should explicitly take into account disorder , which was not necessary for the derivation of the aes action ( strictly speaking , disorder determines the mean free time @xmath161 in the green functions used for derivation of the aes functional but this is a trivial contribution ) . on the other hand , the problem is not as complicated as the problem of anderson localization in the presence of interactions considered in refs . . in our diagrammatic expansions we considered the limit of large conductances @xmath57 and neglected all weak localization corrections . nevertheless , we start the derivation of the action using the replica @xmath3-model approach for interacting systems suggested by finkelstein @xcite . necessary simplifications will be made in the @xmath3-model . although there are difficulties in using the replica @xmath3-models for non - perturbative calculations , our goal is more modest and the final results obtained in the limit @xmath57 neglecting all weak localization corrections will not depend on replica indices at all . the model under consideration has been formulated in section [ choice ] . the derivation of the @xmath3-model can be carried out starting from eqs . ( [ a1]-[a5 ] ) . instead of the fields @xmath36 and @xmath8 one should write fields @xmath162 and @xmath163 carrying the replica index @xmath164 upon calculating physical quantities for an arbitrary @xmath18 , one should put @xmath165 . a proper @xmath3-model for bulk systems with interaction has been derived in refs . . a generalization to granular metals can been made without difficulties following ref . , where a @xmath3-model was written for a granular metal without interaction . although one can proceed along the lines of refs . , _ i.e. _ expanding the action obtained after integration over @xmath36 in the field @xmath8 , a more economic way is to use the substitution given by eqs . ( [ a90 ] , [ a10 ] ) , as well as the subsequent eqs . ( [ a11]-[a14 ] ) . of course , this is already an approximation because we neglect the static component of the phase @xmath124 . however , for our purposes this is not an essential restriction because we do not want to consider effects of localization . moreover , working in the limit @xmath166 or @xmath57 allows us to ignore the static component of @xmath124 . the derivation of the @xmath3-model starting from eqs . ( [ a11]-[a14 ] ) is practically the same as the one presented in ref . . as usual , one decouples the `` effective interaction '' of the type @xmath167 that appears after averaging over impurities by a gaussian integration over matrices @xmath168 . these matrices contain as elements both the replica indices and matsubara frequencies ( or two imaginary times @xmath17 and @xmath169 ) . neglecting the tunneling term , eq . ( [ a13 ] ) , one would obtain a zero - dimensional @xmath3-model ( we assume that all relevant energies are smaller than the thouless energy @xmath170 of a single grain ) . the relevant tunneling term in the @xmath3-model is obtained by a cumulant expansion in the tunneling . the only difference with respect to the model without interaction of ref . is the presence of the phases @xmath124 in eq . ( [ a13 ] ) , which leads to additional phase factors in the term describing the coupling between the grains . the final result for the free energy @xmath171 $ ] can be written as @xmath172=f_{2}[\phi ] + f_{{\rm \omega } } + f_{{\rm t}}[q,\phi ] , \label{full5 } \\ & & \qquad f_{2}[\phi ] = \frac{1}{2e^2}\sum_{i , j}\int { \rm tr}c_{ij}\frac{\partial \phi _ { i}\left ( \tau \right ) } { \partial \tau } \frac{\partial \phi _ { j}\left ( \tau \right ) } { \partial \tau } d\tau , \label{full6 } \\ & & \qquad f_{\omega } [ q]=\frac{\pi } { \delta } \sum_{i}\int { \rm tr } ( \hat{\omega}_{\tau } q_{i}\left ( \tau , \tau \right ) ) d\tau , \label{full7 } \\ & & \qquad f_{{\rm t}}[q,\phi ] = -\frac{g_{t}}{4}\sum_{i , j } \int { \rm tr}\left ( e^{i\phi _ { ij}(\tau ) } q_{i}\left ( \tau , \tau ^{\prime } \right ) e^{-i\phi _ { ij}(\tau ^{\prime } ) } q_{j}\left ( \tau ^{\prime } , \tau \right ) \right ) d\tau d\tau ^{\prime } , \label{full8}\end{aligned}\ ] ] where @xmath62 and the symbol tr implies a trace over both replica indices . the field @xmath124 is diagonal in the replica indices , while the matrix @xmath173 is a @xmath174 matrix with the constraints @xmath175 the action of the operator @xmath176 in eq . ( [ full7 ] ) on an arbitrary function @xmath177 is given by the relation @xmath178 in frequency representation this operator is equal to a vector of fermionic matsubara frequencies . the matrix @xmath173 , satisfying the constraints , eq . ( [ e1 ] ) , can be conveniently parametrized as @xmath179 where @xmath180 and @xmath181 are unitary matrices @xmath182 the matrix @xmath183 in frequency representation has the form @xmath184 the limiting cases of the free energy @xmath171 $ ] , eqs . ( [ full5]-[e2 ] ) are rather simple . without electron - electron interaction the phase difference between neighboring grains equals to zero @xmath185 . such an action for granular metals corresponds to that derived within the supersymmetry scheme by one of the authors @xcite . the second limiting case is achieved at sufficiently high temperatures . in this limit disorder is not important and the matrix @xmath168 does not fluctuate , being equal to the value @xmath183 , eq . ( [ e4 ] ) . inserting @xmath186 into eqs . ( [ full5]-[full8 ] ) we reproduce immediately the aes action , eq . ( [ s ] ) ( in fact we obtain the @xmath18 times replicated aes functional but only one replica field is important for us ) . at not too high a temperature , eqs . ( [ e3]-[e5 ] ) lead to a non - trivial behavior even if the disorder is weak and all effects related to weak localization can be neglected . the limit of weak disorder enables us to neglect fluctuations of @xmath168 at given @xmath124 and take its value from a saddle - point equation . we thus reduce the computation of the functional integral with the free energy functional to the solution of a saddle - point equation , substituting the solution to the free energy @xmath171 $ ] , eqs . ( [ full5]-[full8 ] ) , calculating a functional integral with the reduced free energy @xmath187 $ ] @xmath188=f[\bar{q},\phi ] \label{e6}\ ] ] where @xmath189 is the solution of the saddle - point equation for a given @xmath124 . minimizing the functional @xmath171 $ ] , eqs . ( [ full5]-[full8 ] ) , with the constraints , eqs . ( [ e1 ] ) , we obtain the following equation @xmath190d\tau ^{\prime \prime } + { \frac{\pi } { \delta } } [ \bar{q}_{i}\left ( \tau , \tau \right ) , \hat{\omega})]=0 , \label{e7}\ ] ] where the summation over @xmath107 is performed over the nearest neighbors of @xmath16 . in the commutator @xmath191 $ ] one should exchange properly the times when changing the order of the functions . ( [ e7 ] ) represents an integral equation that enables us to find , in principle , the solution @xmath192 for arbitrary values of the parameter @xmath193 . in general , this equation can be solved only numerically . however , for @xmath194 drastic simplifications arise . for large intergranular coupling , the phase @xmath195 fluctuates only weakly from grain to grain which means that @xmath196 is small . under these conditions , a solution of the diagonal form @xmath197 can be sought for . as the solution is assumed to have a diagonal form , all replica indices decouple and the equations can be solved for each replica separately . therefore we can drop the replica indices and consider eq . ( [ e7 ] ) as an equation for a single function @xmath192 for a given function @xmath198 . in the high temperature limit @xmath160 the second term in eq . ( [ e7 ] ) is much larger than the first one and we arrive at the solution @xmath199 which leads to the aes free energy functional , eq . ( [ s ] ) . in the opposite limit of low temperatures @xmath159 eq . ( [ e7 ] ) can be simplified using the assumption that @xmath192 varies slowly in space . then , one may expand the exponentials in @xmath200 , which leads to the equation @xmath201 + \frac{i\pi } { \delta } [ \frac{\partial } { \partial \tau } , k_{i}]=0 . \label{e10}\ ] ] eq . ( [ e10 ] ) is presented without writing the integration over imaginary times @xmath17 explicitly , but of course this integration is implied . transforming this equation to frequency and momentum space we obtain @xmath202 which gives the solution @xmath203 where @xmath204 and @xmath205 are the fourier transforms of the functions @xmath192 and @xmath43 respectively . substituting this ansatz for @xmath206 back into the action we obtain @xmath207 = \frac{1}{4e_c } \sum_n \sum_k v_{k , n}\left(1 + 4e_c \frac{(2/\pi)g_t\sum_a(1-\cos k_ad)}{|\omega_n|+(2/\pi)g_t\delta \sum_a(1-\cos k_ad)}\right)v_{-k ,- n}\ ] ] where @xmath208 is the charging energy . in the limit of small characteristic momentum , @xmath209 , we recover the diffusively screened coulomb interaction in a disordered evironment . from here one can proceed to the calculation of physical observables such as the conductance , the density of states or others . to this end one should introduce configurations @xmath210 fluctuating around the mean field configuration discussed above . next one would follow the standard algorithm of doing calculations within the @xmath211-model approach : ( i ) express the quantity of interest in terms of @xmath168 , ( ii ) compute the functional average over the effective action @xmath212=f[v , t]$ ] as good as is possible . let us outline how the connection between the enlarged aes - type formulation , ( [ full5 ] ) - ( [ full8 ] ) and the finkelstein approach for interacting disordered media can be made explicit . to do so , we subject our @xmath168 s to a gauge transformation , @xmath213 as a result ( i ) the hopping part of the action , @xmath214 becomes @xmath195-independent . however ( ii ) in the frequency part , @xmath215 . gaussian integration over @xmath216 then produces an effective action @xmath212 $ ] which , after taking a continuum limit , is identified as the action of the finkelstein approach . in this paper we constructed a theoretical framework to describe electronic transport in arrays of tunnel junctions or granular metals , in the limit of large inter - granule conductance , @xmath106 . both disorder and the electron - electron interaction were taken into account . the prime objective of this enterprise was to unify two large , and seemingly non - overlapping theories of interacting metallic compounds : the aes approach , focusing on charging phenomena in individual grains and their impact on large scale transport behaviour , and the finkelstein theory of disordered interacting metals with its emphasis on the interplay interaction / disorder . the key to the reconciliation of these two approaches was to observe that for low enough temperatures , @xmath217 , the effective action underlying the aes theory becomes incomplete . the physical reason is that for low temperatures the electrons and holes participating in the tunneling processes between individual grains maintain their quantum phase memory for a long time , largely in excess of the average tunneling time . this means that the particle tunneling is not only accompanied by charging and dissipation ( as in the aes approach ) , but also by other , more long ranged physical processes . specifically , we identified the quantum dissipation contained in the aes action as the high temperature limit of the long ranged screened electron - electron interaction , an observation first made in@xcite . we re - emphasize that these phenomena were not bound to the presence of a significant disorder concentration ; indeed , none of the results discussed above , displayed dependence on some disorder concentration. the only thing that mattered was chaoticity of the electron motion on scales set by the phase coherence length , a condition that is practically always met in real life systems . technically , two different routes for including these processes into the theory were proposed . one consisted of the perturbative summation of diagram classes associated to multiple chaotic scattering . the other , based on an effective action formulation , indentified what is missing in the aes action . remarkably , this second approach readily led to a unification of the aes approach and the finkelstein nonlinear @xmath211-model for weakly disorderd interacting metals . with the benefit of hindsight , this fact is easy to understand : it is a well known fact in mesoscopic physics , that the long processes resulting from multiple scattering can be interpreted as a certain type of goldstone modes . within an effective action approach , these modes must be described by an own degree of freedom , the @xmath168-matrices of the nonlinear @xmath211-model . thus it is no surprise that we obtained an action of the type ( [ full5 ] - [ full8 ] ) , involving @xmath168-fields and the coulomb phase fields of the aes approach , as an effective description of the low energy phase of the system . for large temperatures , the fluctuations of the @xmath168-matrices became inessential and the aes action was retrieved . in contrast , the continuum limit of this action was identified as the finkelstein model and the @xmath168-matrices smoothly integrated into the aes approach , without leading to conceptual complications . let us make some remarks on potential experimental ramifications of our findings . it has been predicted in refs . that two - dimensional normal arrays of tunnel junctions should undergo a kosterlitz - thouless - berezinskii ( ktb ) phase transition at a temperature @xmath218 of the order of the charging energy @xmath208 . at low temperatures , @xmath219 , the array should be in an insulating state ( coulomb blockade ) , whereas for @xmath220 the array is conducting . in particular , at @xmath221 , the conductivity @xmath211 of the array is predicted to increase with temperature according to a square - root cusp dependence , @xmath222 , where @xmath223 is a constant of order unity . the experiments @xcite mentioned in the introduction searched for this ktb transition . in view of our results , the applicability of the ktb scenario should be governed by the parameter @xmath224 . the experiments of refs . were done on arrays with relatively large grains , such that @xmath225 . indeed , the results obtained in refs . were in agreement with the ktb scenario . however , ref . found thermally activated behavior of @xmath226 , rather than the predicted square root cusp dependence on temperature . on the other hand , the experiment by yamada _ et al . _ @xcite was done on arrays consisting of relatively small cu grains with a size of about @xmath227 , such that @xmath228 . clearly this is beyond the range of applicability of refs . and the ktb scenario should be treated with care in this case . remarkably , ref . concludes good agreement of the results with the theory @xcite . given these discrepancies , we conclude that a careful analysis of the temperature - dependent conductivity of junction arrays in the framework of the theory presented in this paper would be of interest . the authors thank a. andreev , r. fazio and a. tschersich for helpful discussions . a support of the graduiertenkolleg 384 and the sonderforschungsbereich 237 is greatly appreciated .

left - right symmetric model ( lrsm ) @xcite has been one of the most widely studied beyond standard model ( bsm ) scenarios in the last few decades . the model not only explains the sub - ev neutrino mass @xcite naturally through seesaw mechanisms @xcite but also gives rise to low energy parity violations though spontaneous breaking of a parity preserving symmetry at high scale that can also be incorporated within grand unified theory groups like @xmath1 naturally . the minimal version of this model give rise to tiny neutrino masses through a combination of type i @xcite and type ii @xcite seesaw mechanisms . due to the structure of type i seesaw term , one usually has heavy right handed neutrinos in such a model , in order to generate sub - ev neutrino mass from type i seesaw . in tev scale lrsm , one can have right handed neutrino mass @xmath2 tev ( say ) which can give sub - ev light neutrinos from type i seesaw if the dirac neutrino yukawa couplings should be fine tuned to @xmath3 . similar fine - tuning is also involved in the type ii seesaw term as we discuss in the next section . such fine - tunings become more severe if we wish to keep one or two of the right handed neutrinos at very light scale , say between ev to kev . such light sterile neutrinos could have very interesting implications for low energy neutrino oscillation experiments @xcite , neutrinoless double beta decay @xmath4 @xcite , dark matter @xcite and cosmology in general @xcite . the lrsm also received significant attention after the recent hints from the large hadron collider ( lhc ) about the possibility of tev scale new physics : the cms @xmath5 excess @xcite , the atlas diboson excess @xcite and more recently , the 750 gev diphoton excess @xcite . the first two excesses around 2 tev can be explained within a version of lrsm where the discrete left - right symmetry ( to be introduced below ) gets broken at a high energy scale @xcite whereas the gauge symmetry of lrsm remains unbroken all the way down to the tev scale . the recent 750 gev diphoton excess can however , be explained within lrsm framework only when it is extended with additional vector like fermions @xcite or with fields having high @xmath6 dimensions the vector like fermions , as discussed in these works , can serve three purposes : ( i ) to explain the large diphoton cross section at 750 gev , ( ii ) to provide masses to all the fermions of the standard model ( sm ) through a universal seesaw mechanism @xcite and ( ii ) to assist in gauge coupling unification @xcite . this model can also have interesting implications for cosmology @xcite and @xmath7 @xcite . in the present work , we study a class of left - right symmetric models to see the possibility of having light sterile neutrinos at a scale much below the scale of left - right symmetry . typically , the right handed neutrinos in lrsm have masses around the scale of left - right symmetry . thus , in tev scale lrsm , one usually have right handed neutrinos in the gev to tev range . if we fine tune the yukawa couplings to be as small as electron yukawa , then we can get right handed neutrinos of the order of tens of mev . we therefore study some non - minimal versions of lrsm where one can have light right handed neutrinos in the ev - kev regime even with order one yukawa couplings . starting with the lrsm having universal seesaw for all fermions , we explore a few other possibilities : ( i ) light neutrinos as dirac fermions , ( ii ) both left and right handed neutrinos acquire majorana mass terms through radiative type ii seesaw and ( iii ) left handed neutrinos acquire masses through radiative type i seesaw . in all of these possibilities , we consider additional discrete symmetries to stabilise the particles going inside the loops so that the lightest neutral particle among them can be a cold dark matter ( cdm ) candidate . this follows the basic idea of scotogenic models first proposed by @xcite . apart from the stable cold dark matter candidate , these models can also have a long lived kev scale right handed neutrino that can be a warm dark matter ( wdm ) candidate . such kev scale sterile neutrino dark matter within minimal lrsm was discussed by @xcite and a radiative neutrino mass model with both cold and warm dark matter components have been proposed recently by @xcite . we also discuss the possible implications of these scenarios in @xmath7 , collider experiments and cosmology . this article is organised as follows . in section [ sec1 ] , we briefly discuss the minimal lrsm and then discuss the lrsm with universal seesaw in section [ sec2 ] . in section [ sec3 ] we discuss different possible versions of scotogenic lrsm where neutrinos acquire masses at one loop level with dark matter particles going inside the loops . in section [ sec4 ] , we discuss the possibilities of light sterile or right handed neutrinos in these models followed by their implications in @xmath7 and colliders in section [ sec5 ] and [ sec6 ] respectively . we then comment upon cosmological implications of these scenarios in section [ sec7 ] and finally conclude in section [ sec8 ] . left - right symmetric model @xcite is one of the best motivated bsm frameworks where the gauge symmetry of the electroweak theory is extended to @xmath8 . the right handed fermions are doublets under @xmath9 similar to the way left handed fermions transform as doublets under @xmath10 . the requirement of an anomaly free @xmath11 makes the presence of right handed neutrinos a necessity rather than a choice . to allow dirac yukawa couplings between @xmath12 doublet fermions , the higgs field has to transform as a bidoublet under @xmath12 . in order to break the gauge symmetry of the model to that of the sm spontaneously , scalar triplet fields with non - zero @xmath11 charges are introduced , which also give majorana masses to the left and right handed neutrinos . the fermion content of the minimal lrsm is @xmath13 @xmath14 similarly , the higgs content of the minimal lrsm is @xmath15 @xmath16 where the numbers in brackets denote the transformations of the fields under the gauge group @xmath17 . during the spontaneous symmetry breaking of lrsm gauge group down to the sm gauge group , the neutral component of the higgs triplet @xmath18 acquires a non - zero vacuum expectation value ( vev ) after which the neutral components of higgs bidoublet @xmath19 acquire non - zero vev s to break the sm gauge symmetry into the @xmath20 of electromagnetism . this symmetry breaking chain can be denoted as : @xmath21 denoting the vev of the two neutral components of the bidoublet as @xmath22 and that of triplets @xmath23 as @xmath24 and considering @xmath25 , @xmath26 and @xmath27 , the approximate expressions for gauge boson masses after symmetry breaking can be written as @xmath28 @xmath29 where @xmath30 is the weinberg angle . the relevant yukawa couplings for fermion masses can be written as @xmath31 where @xmath32 . in the above yukawa lagrangian , the indices @xmath33 correspond to the three generations of fermions . the majorana yukawa couplings @xmath34 is same for both left and right handed neutrinos because of the in built left - right symmetry @xmath35 . these couplings @xmath34 give rise to the majorana mass terms of both left handed and right handed neutrinos after the triplet higgs fields @xmath36 acquire non - zero vev . although it is the @xmath18 field which gets a vev at high scale breaking the left - right symmetry , the subsequent electroweak symmetry breaking induces a non - zero vev to the left handed counterpart . the induced vev for the left - handed triplet @xmath37 can be shown for generic lrsm to be @xmath38 with @xmath39 gev being the weak boson mass such that @xmath40 in general @xmath41 is a function of various couplings in the scalar potential of generic lrsm . using the results from deshpande et al . , @xcite , @xmath41 is given by @xmath42 where @xmath43 are dimensionless parameters of the scalar potential . without any fine tuning @xmath41 is expected to be of the order unity ( @xmath44 ) . however , for tev scale type i+ii seesaw , @xmath41 has to be fine - tuned as we discuss later . the type ii seesaw formula for light neutrino masses can now be expressed as @xmath45 if we consider the first term on the right hand side of the above expression , one can make an estimate of neutrino mass for tev scale lrsm . considering @xmath46 tev , the type ii seesaw term will be of the order of light neutrino mass @xmath47 ev if @xmath48 where @xmath49 is the right handed neutrino mass . thus , for tev scale right handed neutrino masses , the dimensionless parameter @xmath41 is fine tuned to the level of @xmath50 in order to get correct order of neutrino masses . this will involve unnatural fine tuning of the scalar potential parameters appearing in the expression for @xmath41 given in . similar but slightly less fine tuning is involved in the type i seesaw term for tev scale @xmath51 . the dirac yukawa couplings should be fine tuned to around @xmath52 in order to get light neutrino mass of order @xmath53 ev . in minimal lrsm with higgs bidoublet and higgs triplets discussed above , the light neutrino masses arise at tree level . if the triplets are replaced by higgs doublets , then additional fermions are required for seesaw mechanisms . in the absence of them , neutrinos have only a dirac mass term due to the bidoublet , which typically is of the order of charged lepton masses . thus , the tree level dirac mass term of neutrinos can be prevented only in the absence of higgs bidoublet . indeed , lrsm without higgs bidoublet has found some attention in the literature @xcite . this model was studied later in the context of cosmology @xcite and neutrinoless double beta decay @xcite . very recently this model was also studied in the context of 750 gev di - photon excess at lhc @xcite by several authors @xcite . the model has additional vector like fermions in order to generate all the fermion masses from universal seesaw . the fermion content of the lrsm with universal seesaw is the extension of the mlrsm fermion content by the following vector like fermions @xmath54 @xmath55 per generation of quarks and leptons . instead of higgs bidoublet , this model has a pair of higgs doublets @xmath56 the non - zero vev of the neutral component of @xmath57 also breaks the @xmath58 symmetry of the model into @xmath59 of the standard model . the left handed higgs doublet can acquire a non - zero vev @xmath60 at a lower energy to induce electroweak symmetry breaking . however , the left - right symmetry of the theory forces one to have the same vev for both @xmath61 and @xmath57 that is , @xmath62 which is unacceptable from phenomenological point of view . to decouple these two symmetry breaking scales , the extra singlet scalar @xmath63 is introduced into the model . this field is odd under the discrete left - right symmetry and hence couple to the two scalar doublets with a opposite sign . after this singlet acquires a non - zero vev at high scale , this generates a difference between the effective mass squared of @xmath61 and @xmath57 which ultimately decouples the symmetry breaking scales . due to the absence of the usual bidoublet , the left and right handed fermion doublets of the mslrm can not directly couple to each other . however , they can couple to the scalar fields @xmath64 via the additional vector like fermions . @xmath65 after integrating out the heavy fermions , the charged fermions of the standard model develop yukawa couplings to the scalar doublet @xmath61 as follows @xmath66 where @xmath67 is the vev of the neutral component of @xmath57 . the apparent seesaw then can explain the observed mass hierarchies among the three generations of fermions . the neutrino mass arises in a more complicated seesaw due to additional majorana mass terms for @xmath68 shown above . as shown in @xcite , the neutrino mass matrix in the basis @xmath69 can have three independent terms @xmath70 @xmath71 @xmath72 after integrating out the heavy neutral fermions @xmath73 . this mass matrix can further be diagonalised to get the type i seesaw formula for neutrino mass which however , is suppressed by @xmath74 compared to @xmath75 . thus , the left and right handed neutrinos have majorana mass terms which are related by @xmath76 if we have a tev scale left - right symmetry breaking say @xmath77 tev , then the above relation says @xmath78 . thus for @xmath79 ev , the right handed neutrino masses are @xmath80 ev . such light right handed neutrinos will have very interesting implications in cosmology , collider as well as neutrinoless double beta decay experiments , when the lr gauge boson masses are around the tev corner . one can however , keep them hidden in neutrinoless double beta decay experiments by tuning the @xmath81 element of the right handed neutrino mass matrix to zero . on the other hand , direct search experiments like the lhc search for heavy neutrinos and @xmath82 bosons at 8 tev centre of mass energy already excludes a wide range of heavy neutrino masses @xcite . however , as can be seen from the @xmath83 exclusion plot in @xcite , the experimental sensitivity is not so good for the low mass regime of right handed neutrinos . also , a right handed neutrino in the ev - kev regime will not give rise to the same sign dilepton plus two jets signatures which are being looked at by lhc experiments to put bounds on @xmath83 parameter space @xcite . such light right handed neutrinos will also induce non - unitarity to leptonic mixing matrix of the order @xmath84 which will again constrain @xmath85 to a small number for tev scale @xmath67 . if we want to keep the right handed neutrino mass at kev scale , the left - right symmetry breaking scale has to be @xmath86 tev , beyond the reach of lhc . since the tev scale lrsm with universal seesaw for all fermions generate three ultra light right handed neutrinos , which may be in conflict with experiments from colliders to cosmology , one can find suitable alternative to this . one such immediate remedy is to consider universal seesaw for charged fermions and radiative masses for neutrinos . one such possibility was discussed very recently in @xcite where neutrino mass was generated at two - loop level along with a bonus of having a long lived kev neutrino dark matter . here we consider another possibility of having radiative neutrino mass where the particles running inside the loop can be stable dark matter candidates . we consider two possibilities where light neutrinos can be either of dirac type or of majorana type . let us consider a version of lrsm with universal seesaw for all the charged fermions but zero tree level masses for neutrinos . this corresponds to the model discussed above but without the neutral fermion pairs @xmath73 . instead of that , consider the addition of following particles into the model with their transformations under the lrsm gauge group as well as an additional symmetry forbidding tree level neutrino mass and guaranteeing the stability of dark matter . these new particle sector consists of scalar doublets @xmath87 and fermion singlets @xmath88 . the additional symmetry can either be discrete @xmath89 or a continuous one like @xmath90 . we show one simple example of @xmath91 in this subsection . the fields taking part in radiative neutrino mass generation are shown in table [ tab : data1 ] . since the higgs doublets @xmath64 have trivial transformations under the additional @xmath91 , the usual mass terms for charged fermions are allowed . all other particles in lrsm with universal seesaw for charged fermions do not transform under this additional symmetry . such a structure of new particles and symmetry will give rise to the one - loop dirac neutrino mass shown in figure [ fig1 ] . one loop dirac neutrino mass in a scotogenic fashion was first proposed in @xcite within a model where the sm gauge symmetry was extended by @xmath11 and additional @xmath0 discrete symmetries . .particle content for radiative dirac neutrino masses [ cols="^,^,^",options="header " , ] due to the presence of several new vector , scalar and fermionic particles in lrsm , one can have various new physics contribution to neutrinoless double beta decay @xmath92 . the relevant feynman diagrams and the calculation details can be found in several earlier works @xcite and references therein . apart from the sm contribution mediated by light neutrinos and @xmath93 bosons , there can be contributions from ( a ) right handed neutrino and @xmath82 , ( b ) mixing between heavy and light neutrinos ( c ) @xmath94 mixing and ( d ) triplet scalars in type ii seesaw scenarios . here we consider only the right handed neutrino with @xmath82 contribution as new physics and ignore the active - sterile , @xmath95 mixing and triplet scalar contributions . the amplitude of the light neutrino contribution is given by @xmath96 where @xmath97 mev is the average momentum exchange of the process , @xmath98 are the masses of three light neutrinos and @xmath99 is the standard pontecorvo - maki - nakagawa - sakata ( pmns ) leptonic mixing matrix . the contribution from the heavy neutrino and @xmath82 exchange can be written in the limit @xmath100 as @xmath101 where @xmath102 is the heavy neutrino mixing matrix and @xmath103 are the masses of heavy right handed neutrinos . thus the total contribution to @xmath92 process is @xmath104 where @xmath105 are effective neutrino masses corresponding to light neutrino and heavy neutrino exchanges respectively . here , @xmath106 is the nuclear matrix element . now , for lrsm with universal seesaw , the light and heavy neutrino mass matrices are related by a factor @xmath107 where @xmath108 gev . similarly in the radiative type ii seesaw model , they are related by a factor @xmath109 which is also true for tree level type ii seesaw model . since the heavy and light neutrino mass matrix is similar upto a factor , we can consider @xmath110 in the above expressions . also the mass eigenvalues will be related by the same numerical factor @xmath111 . thus , the new physics contribution can be written as @xmath112 using the standard parameterisation of pmns mixing matrix in terms of three angles and three cp phases ( one dirac and two majorana phases ) , one can write down the light neutrino contribution as @xmath113 where @xmath114 , @xmath115 being the mixing angles and @xmath116 are majorana cp phases . fixing the lightest right handed neutrino mass to be 1 gev and hence @xmath117 , we calculate the heavy neutrino contribution to the effective neutrino mass . the contribution is found to be minimal for majorana cp phases @xmath118 as shown in figure [ fig3 ] . it can be seen from the figure that such a scenario with @xmath119 gev , @xmath120 tev and normal hierarchy of light neutrinos is ruled out from existing experimental constraint . for inverted hierarchy also , the region for @xmath121 ev is ruled out whereas the @xmath122 ev region contribution almost coincides with the experimental upper bound . choosing different values of majorana phases rules out even more parameter space . similarly , we calculate the right handed neutrino contribution for masses @xmath123 mev . in such low mass regime , the right handed neutrino contribution has a similar expression as standard light neutrino contribution upto a scaling factor of gauge boson mass ratio . the right handed neutrino contribution is found to be minimal for majorana cp phases @xmath124 as seen from figure [ fig31 ] . it can be seen that inverted hierarchy of light neutrino spectrum is already ruled out for such a case with @xmath120 tev . the normal hierarchical scenario survives only in a narrow region of parameter space around @xmath125 ev . after showing the allowed parameter space for a specific choice of @xmath126 , we show the allowed regions of @xmath127 and @xmath128 parameter space in figure [ fig4 ] and [ fig41 ] , by incorporating the experimental upper bound on @xmath7 amplitude . as seen from figure [ fig41 ] , one can still have kev scale right handed neutrino along with a tev scale @xmath82 after incorporating experimental constraints on @xmath7 lifetime . depending on the origin of light neutrino masses , there can be different collider signatures for the new physics sector responsible for it . if neutrinos are dirac fermions such that @xmath129 form a dirac fermion with ev scale dirac mass , then one can not have on - shell production of same sign dilepton plus dijet signatures mediated by @xmath82 bosons . such a process was proposed @xcite to be a clean signature of right handed gauge bosons with heavy majorana neutrinos . in these scenarios @xmath82 boson can decay into @xmath130 final states in a way similar to @xmath93 decay into @xmath131 . also , due to the presence of scalar doublet dark matter candidate , @xmath82 can also have decay into dijet plus missing energy through @xmath132 or @xmath133 . here @xmath134 in final states can give rise to missing energy , if it is the lightest neutral particle stabilised by the remnant @xmath0 symmetry . however , if neutrinos are majorana fermions and heavy neutrinos are in the gev - tev range with tev scale @xmath82 , they can be constrained by lhc data from the search of dilepton plus dijet . for even lower masses of right handed neutrinos say , in the kev regime , lhc sensitivity may not be enough and one has to go for alternate probes like neutrinoless double beta decay , discussed above . apart from heavy neutrino mediated dilepton plus dijet final states , one can also have other interesting collider signatures due to the presence of additional scalars . the scalar sector of mlrsm and some of its minimal variants have been extensively studied by several groups @xcite . in the model with radiative type ii seesaw , one can have light doubly charged scalars similar to mlrsm . however , their collider signatures can be entirely different . the present lhc lower bounds on doubly charged scalar masses ( 465 - 550 gev for @xmath135 , 370 - 435 gev for @xmath136 ) @xcite are derived by considering @xmath137 branching ratio into same sign lepton pairs . in the radiative type ii seesaw model discussed above , the doubly charged scalars do not have tree level coupling to the leptons and hence their decay into lepton pairs will be suppressed . thus , the above bounds are not applicable and one can have lighter doubly charged boson as well . in the model discussed here , one can have @xmath138 or @xmath139 with the @xmath93 boson finally decaying into two jets or leptons . considering leptonic final states only , both the decay modes can contribute to same sign dilepton plus missing energy signatures . similar conclusions can be made for right handed doubly charged scalar @xmath140 as well . in the model with radiative type i seesaw , the doubly charged scalars have tree level coupling to leptons . however , the neutral component of the left handed triplet does not acquire any vev . therefore the decay channel @xmath138 is absent . however @xmath36 still couples to scalar doublets @xmath87 and hence both @xmath141 and @xmath139 are possible in such a scenario . same conclusion can be made for right handed triplet scalar as well . cosmological implications for sterile neutrinos with masses from ev scale to gev - tev scale are well described in the recent review @xcite . here we focus only on the ev - kev range as the above analysis was primarily focussed on generating sterile neutrino masses in this range . the left - right symmetric model with light sterile neutrinos in ev to kev scale along with a cold dark matter candidate as discussed above can have interesting implications for cosmology . if one or more of the right handed neutrinos have masses in the ev range , then they can contribute to the total radiation content of the universe . such extra relativistic degrees of freedom may affect the big bang nucleosynthesis ( bbn ) predictions as well as cause changes in the cosmic microwave background ( cmb ) spectrum . the latest cosmology data from planck experiment @xcite constrains the the effective relativistic degrees of freedom @xmath142 . in the model where the light neutrinos are dirac , the right handed components contribute extra degrees of freedom to @xmath143 . as shown in @xcite , this can evade the bbn bound if the right handed neutrinos decouple much earlier that is @xmath144 . the @xmath145 contribution to @xmath143 gets diluted due to the decrease in effective relativistic degrees of freedom @xmath146 as the universe cools down from @xmath147 to @xmath148 : @xmath149 since @xmath150 for the relativistic degrees of freedom in sm at @xmath151 mev , the right handed neutrinos can evade the planck bound if @xmath152 . this implies that the right handed neutrinos should decouple before the qcd phase transition temperature 200 - 400 mev . the decoupling temperature of right handed neutrinos can be calculated by following the same procedure for left handed neutrinos and replacing the @xmath93 mass with @xmath82 . in terms of @xmath148 , it can be written as @xmath153 demanding @xmath154 mev and taking @xmath155 mev , we can arrive at the bound on @xmath126 as @xmath156 where we have considered @xmath157 , which is the number of relativistic degrees of freedom just above the qcd phase transition temperature . the bound on @xmath82 mass from cosmology in this scenario is stronger than the direct search limits discussed earlier . similar bounds will also be applicable in those models where active neutrinos are majorana but one or two of the sterile or right handed neutrinos have masses in the ev scale . even if the light right handed neutrinos decouple before the qcd temperature , they can again be produced at lower temperatures through active - sterile oscillations . this can again be in tension with the planck bound on the relativistic degrees of freedom . in the scenarios discussed above , such a situation can occur for the model with particle content shown in table [ tab : data2 ] . for the particle content in table [ tab : data2 ] , both active - active , sterile - sterile and active - sterile mass terms can appear at one loop level . similarly , in the lrsm with universal seesaw also , one can have sizeable active - sterile neutrino mixing by suitably choosing @xmath158 in equation . due to large active - sterile mixing , flavour oscillations will produce the sterile neutrinos at low temperatures making such a scenario highly disfavoured from cosmology . additional new physics is required to suppress such late production of sterile neutrinos through flavour oscillations @xcite . the dark matter phenomenology can be very rich if we have a kev scale sterile neutrino along with a stable dark matter candidate with electroweak scale mass or above . as discussed before , such a possibility exists in the radiative type ii seesaw model if we want to keep the right handed gauge bosons near the tev corner . in lrsm with universal seesaw , the symmetry breaking scale has to pushed high in order to have kev right handed neutrinos whereas in the radiative type i seesaw model , unnatural fine tuning is required to generate kev scale right handed neutrino masses , similar to the mlrsm . as discussed in @xcite such kev sterile neutrinos with gauge interactions , remain in thermal equilibrium in the early universe and remain like a thermal relic after their freeze - out . typically , for tev scale masses of these additional gauge bosons , these sterile neutrinos decouple at a low temperature , leading to overproduction . such an overproduction can be avoided by late time entropy release from the decay of a heavier particle into the sm particles @xcite after the freeze - out of kev sterile neutrino . using these prescriptions , the authors of @xcite showed that the right handed gauge bosons have to be heavier than 10 tev for correct abundance of kev sterile neutrino dark matter in mlrsm . later , a more careful study @xcite found a tiny window near @xmath82 mass of 5 tev for which a kev sterile neutrino dark matter of mass @xmath159 kev can give the correct relic abundance . similar analysis can also be performed within the models discussed in this work . similarly , the cold dark matter analysis can also be performed in the usual way of calculating relic abundance after freeze - out . analysis of scalar doublet dark matter in lrsm was recently done by @xcite . however , due to the mixed dark matter scenario , a careful detailed calculation is required , as the late time entropy release to suppress the wdm abundance will also affect the abundance of cdm . such a mixed dark matter model can be very interesting from astrophysical structure point of view , as it may provide a way to solve the small scale structure problem @xcite . this can also be interesting from dark matter experiments point of view as the cold component can provide signatures at gamma ray experiments @xcite whereas the warm component with mass around a few kev can be responsible for the anomalies at x - ray telescopes @xcite . we have presented a class of left - right symmetric models without the conventional higgs bidoublet that allows the possibility of having light sterile neutrinos along with a cold dark matter candidate stabilised by additional discrete symmetries . the charged leptons acquire masses through a universal seesaw mechanism due to the presence of additional vector like fermions . the neutrinos can either acquire masses through the same universal seesaw or through one loop radiative corrections , if suitable particles are added with non - trivial transformations under additional symmetries . the same additional symmetry can also stabilise one of the particles going inside the loop in neutrino mass diagram , resulting in a stable dark matter candidate . depending on the additional particles and symmetries , the light neutrinos can be either dirac or majorana . in case of dirac light neutrinos , the dark sector is composed of only cold dark matter component . however , in case of majorana light neutrinos , the one or more of the right handed neutrinos can have mass in the ev - kev range without any unnatural fine tuning of the yukawa couplings . a kev sterile neutrino can be a warm dark matter candidate giving rise to a mixed dark matter scenario along with the cold component . such light right handed neutrinos can have interesting implications for neutrinoless double beta decay and collider experiments , constraining the right handed gauge boson masses . cosmology constraints on the number of relativistic degrees of freedom also constraints the right handed gauge boson masses to be heavier than a few tev so that the right handed neutrinos get decoupled before the qcd phase transition temperature . a detailed calculation of the relic abundance of such mixed dark matter scenario is worth investigating and we leave it for a future work . the author would like to express a special thanks to the mainz institute for theoretical physics ( mitp ) and the munich institute for astro and particle physics ( miapp ) for hospitality and support while some part of this work was completed . 99 j. c. pati and a. salam , phys . * d10 * , 275 ( 1974 ) ; r. n. mohapatra and j. c. pati , phys . rev . * d11 * , 2558 ( 1975 ) ; g. senjanovic and r. n. mohapatra , phys . * d12 * , 1502 ( 1975 ) ; r. n. mohapatra and r. e. marshak , phys . lett . * 44 * , 1316 ( 1980 ) ; j. f. gunion , j. grifols , a. mendez , b. kayser and f. i. olness , phys . rev . * d40 * , 1546 ( 1989 ) . n. g. deshpande , j. f. gunion , b. kayser and f. i. olness , phys . * d44 * , 837 ( 1991 ) . s. fukuda et al . ( super - kamiokande ) , phys . rev . lett . * 86 * , 5656 ( 2001 ) , hep - ex/0103033 ; q. r. ahmad et al . ( sno ) , phys . lett . * 89 * , 011301 ( 2002 ) , nucl - ex/0204008 ; phys . 89 * , 011302 ( 2002 ) , nucl - ex/0204009 ; j. n. bahcall and c. pena - garay , new j. phys . * 6 * , 63 ( 2004 ) , hep - ph/0404061 ; k. nakamura et al . , . * g37 * , 075021 ( 2010 ) . k. abe et al . [ t2k collaboration ] , phys . rev . lett . * 107 * , 041801 ( 2011 ) , [ arxiv:1106.2822 [ hep - ex ] ] . y. abe et al . , phys . lett . * 108 * , 131801 ( 2012 ) , [ arxiv:1112.6353 [ hep - ex ] ] . f. p. an et al . [ daya - bay collaboration ] , phys . lett . * 108 * , 171803 ( 2012 ) , [ arxiv:1203.1669 [ hep - ex ] ] . j. k. ahn et al . [ reno collaboration ] , phys . lett . * 108 * , 191802 ( 2012 ) , [ arxiv:1204.0626][hep - ex ] ] . p. adamson et al . ( minos ) , phys.rev.lett . * 110 * , 171801 ( 2013 ) p. minkowski , phys . lett . * b67 * , 421 ( 1977 ) ; m. gell - mann , p. ramond , and r. slansky ( 1980 ) , print-80 - 0576 ( cern ) ; t. yanagida ( 1979 ) , in proceedings of the workshop on the baryon number of the universe and unified theories , tsukuba , japan , 13 - 14 feb 1979 ; r. n. mohapatra and g. senjanovic , phys . rev . lett * 44 * , 912 ( 1980 ) ; j. schechter and j. w. f. valle , phys . rev . * d22 * , 2227 ( 1980 ) . r. n. mohapatra and g. senjanovic , phys . rev . * d23 * , 165 ( 1981 ) . g. lazarides , q. shafi and c wetterich , nucl . phys . * b181 * , 287 ( 1981 ) ; c. wetterich , nucl . phys . * b187 * , 343 ( 1981 ) ; j. schechter and j. w. f. valle , phys . rev . * d25 * , 774 ( 1982 ) ; b. brahmachari and r. n. mohapatra , phys . rev . * d58 * , 015001 ( 1998 ) ; r. n. mohapatra , nucl . . suppl . * 138 * , 257 ( 2005 ) ; s. antusch and s. f. king , phys . lett . * b597 * , ( 2 ) , 199 ( 2004 ) . r. foot , h. lew , x. g. he and g. c. joshi , z. phys . * c44 * , 441 ( 1989 ) . c. athanassopoulos et al . [ lsnd ] , phys . lett . * 77 * , 3082 ( 1996 ) ; a. aguilar et al . , phys . rev . * d64 * , 112007 ( 2001 ) . g. mention et al . , phys . rev . * d83 * , 073006 ( 2011 ) . m. a. acero , c. giunti and m. laveder , phys . rev . * d78 * , 073009 ( 2008 ) ; c. giunti and m. laveder , phys . rev . * c83 * , 065504 ( 2011 ) . a. gando et . al . , [ kamland - zen collaboration ] , phys . * 110 * , 062502 ( 2013 ) . a. gando et al . , [ kamland - zen collaboration ] , arxiv:1605.02889 [ hep - ex ] . m. agostini et . , [ gerda collaboration ] , phys . lett . * 111 * , 122503 ( 2013 ) . r. adhikari et al . , arxiv:1602.04816 . g. hinshaw et al . , arxiv:1212.5226 . a. r. ade et al . , [ planck collaboration ] , arxiv:1502.01589 . v. khachatryan et al . [ cms collaboration ] , eur . j. * c74 * , 3149 ( 2014 ) . g. aad et al . [ atlas collaboration ] , jhep * 1512 * , 055 ( 2015 ) . `` atlas and cms physics results from run 2 '' , talks by j. olsen and m. kado , cern , december 15 , 2015 . the atlas collaboration , atlas - conf-2015 - 081 . cms collaboration [ cms collaboration ] , collisions at 13tev , cms - pas - exo-15 - 004 . d. chang , r. n. mohapatra and m. k. parida , phys . * 52 * , 1072 ( 1984 ) . f. f. deppisch , c. hati , s. patra , p. pritimita and u. sarkar , phys . b757 * , 223 ( 2016 ) . d. borah , s. patra and s. sahoo , int . j. mod a31 * , 1650097 ( 2016 ) . b. brahmachari , e. ma and u. sarkar , phys . * 91 * , 011801 ( 2003 ) . a. davidson and k. c. wali , phys . lett . * 59 * , 393 ( 1987 ) ; k. s. babu and r. n. mohapatra , phys . lett . * 62 * , 1079 ( 1989 ) ; p. -h . gu and m. lindner , phys . lett . * b698 * , 40 ( 2011 ) . gu , phys . rev . * d81 * , 095002 ( 2010 ) . s. patra , phys . rev . * d87 * , 015002 ( 2013 ) . e. ma , phys . rev . * d73 * , 077301 ( 2006 ) . f. bezrukov , h. hettmansperger and m. lindner , phys . rev . * d81 * , 085032 ( 2010 ) . m. nemevsek , g. senjanovic and y. zhang , jcap * 07 * , 006 ( 2012 ) . r. adhikari , d. borah and e. ma , phys . lett . * b755 * , 414 ( 2016 ) . v. khachatryan et al . , [ cms collaboration ] , eur . j. * c74 * , 3149 ( 2014 ) . t. nomura , h. okada and y. orikasa , arxiv:1602.08302 . y. farzan and e. ma , phys . rev . * d86 * , 033007 ( 2012 ) . j. kopp , p. a. n. machado , m. maltoni and t. schwetz , jhep * 1305 * , 050 ( 2013 ) . k. a. olive et al . , [ particle data group ] , chin . phys . * c38 * , 090001 ( 2014 ) . v. tello , m. nemevsek , f. nesti , g. senjanovic and f. vissani , phys . rev . lett . * 106 * , 151801 ( 2011 ) ; p. s. b. dev , s. goswami , m. mitra and w. rodejohann , phys * d88 * , 091301 ( 2013 ) ; s .- f . ge , m. lindner and s. patra , arxiv:1508.07286 ; j. chakrabortty , h. z. devi , s. goswami and s. patra , jhep * 08 * , 008 ( 2012 ) ; j. barry and w. rodejohann , jhep * 1309 * , 153 ( 2013 ) ; d. borah and a. dasgupta , jhep * 1511 * , 208 ( 2015 ) ; g. bambhaniya , p. s. b. dev , s. goswami and m. mitra , arxiv:1512.00440 ; s. p. das , f. f. deppisch , o. kittel and j. w. f. valle , phys . rev . * d86 * , 055006 ( 2012 ) ; m. k. parida and s. patra , phys . b718 * , 1407 ( 2013 ) ; r. l. awasthi , m. k. parida and s. patra , jhep * 1308 * , 122 ( 2013 ) ; w. -c . huang and j. lopez - pavon , eur . j. * c74 * , 2853 ( 2014 ) ; p. s. b. dev , s. goswami and m. mitra , phys . rev . * d91 * , 11 , 113004 ( 2015 ) ; d. borah and a. dasgupta , arxiv:1606.00378 . keung and g. senjanovic , phys . lett . * 50 * , 1427 ( 1983 ) . g. barenboim , m. gorbahn , u. nierste and m. raidal , phys . rev . * d65 * , 095003 ( 2012 ) ; j. polak and m. zralek , phys b276 * , 492 ( 1992 ) ; g. azuelos , k. benslama , j. ferland , j. phys . * g32 * , 73 ( 2006 ) ; d. w. jung and k. y. lee , phys . rev . * d78 * , 015022 ( 2008 ) ; g. bambhaniya , j. chakrabortty , j. gluza , m. kordiaczynska and r. szafron , jhep * 1405 * , 033 ( 2014 ) ; b. dutta , r. eusebi , y. gao , t. ghosh and t. kamon , phys . rev . * d90 * , 055015 ( 2014 ) ; g. bambhaniya , j. chakrabortty , j. gluza , t. jeliski and m. kordiaczynska , phys . rev . * d90 * , 095003 ( 2014 ) ; a. maiezza , m. nemevsek and f. nesti , phys . rev . lett . * 115 * , 081802 ( 2015 ) ; g. bambhaniya , j. chakrabortty , j. gluza , t. jelinski and r. szafron , phys . rev . * d92 * , 015016 ( 2015 ) ; p. s. b. dev , r. n. mohapatra and y. zhang , arxiv:1602.05947 . et al . _ [ atlas collaboration ] , jhep * 1503 * , 041 ( 2015 ) . k. a. olive , g. steigman and d. n. schramm , phys . * 43 * , 239 ( 1979 ) ; s. m. davidson and h. e. logan , phys . rev . * d80 * , 095008 ( 2009 ) . s. hannestad , r. s. hansen and t. tram , phys . lett . * 112 * , 031802 ( 2014 ) ; b. dasgupta and j. kopp , phys . * 112 * , 031803 ( 2014 ) . u. seljak , a. makarov , p. mcdonald and h. trac , phys . lett . * 97 * , 191303 ( 2006 ) . r. j. scherrer and m. s. turner , phys . rev . * d31 * , 681 ( 1985 ) . c. garcia - cely and j. heeck , jcap * 1603 * , 021 ( 2016 ) . a. harada and a. kamada , jcap * 1601 * , 031 ( 2016 ) . t. daylan , d. p. finkbeiner , d. hooper , t. linden , s. k. n. portillo , n. l. rodd and t. r. slatyer , phys . dark univ . * 12 * , 1 ( 2016 ) . e. bulbul , m . markevitch , a . foster , r . k. smith , m . loeewenstein and s. w. randall , astrophys . j. * 789 * , 13 ( 2014 ) a. boyarsky , o. ruchayskiy , d. iakubovskyi and j. franse , phys . * 113 * , 251301 ( 2014 ) . a. boyarsky , j. franse , d. iakubovskyi and o. ruchayskiy , phys . lett . * 115 * , 161301 ( 2015 ) .

recent developments involving imaging atmospheric cherenkov telescopes ( iacts ) have revolutionised the field of very high energy ( vhe ) @xmath1-ray astronomy . indeed , observational data obtained using the current generation of stereoscopic arrays have contributed over 100 new sources to the catalogue of confirmed gev - tev emitters . notwithstanding these impressive achievements , the motivation for astronomers to continue identifying and categorising additional @xmath1-ray sources and source populations remains undiminished . the initial source detections are likely to be dominated by luminous or nearby @xmath1-ray emitters , implying a requirement for subsequent observations to target progressively fainter source populations . accordingly , the development of analytical techniques which enable maximal exploitation of the available instrumental sensitivity will become increasingly important as the life - cycles of the various iact arrays unfold . this emerging philosophy is exemplified by the recent work of @xcite which employs accurate knowledge of the instrumental response to enhance the performance and reliability of an established technique for source detection originally developed by @xcite . until recently , the inherent faintness of astrophysical vhe @xmath1-ray fluxes has resulted in statistically limited signal measurements with uncertainties that are dominated by poissonian shot noise . accordingly , many traditional iact analyses have justifiably neglected the existence of previously sub - dominant systematic effects . in contrast , successful source detection at the limits of instrumental sensitivity requires deep observations and the accumulation of rich data sets with abundant event statistics . in this new regime , the relative contribution of statistical fluctuations to the overall error budget is suppressed with measurement accuracy ultimately becoming systematically limited . this work investigates methods for the appropriate statistical treatment of systematic uncertainties affecting iact observational data and their subsequent analysis . in cherenkov astronomy , the synthesis of large data sets typically involves aggregation of temporally sparse event statistics which correspond to distinct values of various observational parameters , each of which can introduce distinct , observation - specific systematic biases . accordingly , analyses that are designed to ameliorate the impact of irreducible systematic effects must incorporate sufficient flexibility to successfully model this inter - observational variability . in the subsequent discussion , the relevant statistical issues and the techniques designed to address them are illustrated using the specific example of _ stacking _ analysis . in such analyses , measurement data that correspond to multiple , independent but similar experiments are combined to enhance their scientific utility . in the context of @xmath1-ray astronomy , stacking analyses are commonly applied in situations when a target population comprises a large number of individually undetectable source candidates that are theoretically expected to exhibit at least one identical signal characteristic . in the absence of genuine @xmath1-ray emission , the superimposed observational data will be consistent with random poisson fluctuations about the mean background level . conversely , multiple faint signals produced by a genuine sample of faint @xmath1-ray sources will reinforce upon combination , yielding a significant overall excess with respect to the measured background . accordingly , stacking analyses often facilitate the derivation of average source properties for the target population as a whole , even when detection of individual @xmath1-ray signals is rendered impossible by inadequate instrumental sensitivity . the specific consideration of stacking is instructive because combined source analysis inherently segregates subsets of observational data which are likely to incorporate distinct systematic biases . moreover , stacking analyses provide a more intuitive distinction between systematic effects that undergo temporal variation , such as telescope elevation or optical efficiency , and those that vary on a target - wise basis , such as the ambient night - sky brightness of the number of @xmath1-ray sources within the telescope field - of - view . compelling and theoretically well motivated target populations for combined source analyses include x - ray binary systems @xcite , combined pulsar populations within globular clusters and the self - annihilating dark matter components of dwarf spheroidal galaxies ( e.g. * ? ? ? the atmosphere of the earth is almost completely opaque to multi - gev and tev radiation and accordingly , the operational mode of iact arrays relies upon indirect detection of astrophysical @xmath1-rays . indeed , the majority of incident vhe photons undergo electron - positron pair production interactions within the coulomb fields of stratospheric nuclei . frequently , these interactions initiate electromagnetic cascades of charged particles , which radiate cherenkov light as they descend superluminally through the dielectric medium of the atmosphere . iact arrays are designed to image the resultant distribution of cherenkov photons at ground - level , using the encoded information to reconstruct the properties of the progenitor @xmath1-ray . the set of reconstructed @xmath1-ray properties define a multi - dimensional parameter space , with suitable sub - spaces that define the event populations that are required for specific astrophysical analyses . the situation is complicated somewhat by the fact that iact arrays operate in a strongly background dominated regime and must successfully distinguish a comparatively weak @xmath1-ray signal from an almost overwhelming background of air - showers initiated by hadronic cosmic rays . fortunately , the properties of hadron- and @xmath1-ray - initiated air - showers are sufficiently disparate that an effective segregation is possible based on the characteristics of the captured cherenkov images . traditionally , a simplified elliptical parameterisation of the cherenkov light distribution was used in conjunction with lookup tables of simulated air - shower properties to separate the signal and background event populations @xcite . modern computational resources permit utilisation of the complete cherenkov images ( e.g * ? ? ? * ; * ? ? ? * ) and the application of advanced machine - learning algorithms ( e.g * ? ? ? * ) to further enhance the background rejection efficiency . despite these sophisticated filtering techniques , the sheer number of air - showers that trigger an iact array during a given observation inevitably results in a subset of hadron - originated cherenkov images which are sufficiently @xmath1-ray - like in appearance to survive the event selection procedure . quantitatively , the probability that a @xmath1-ray - like event having specific reconstructed parameters will pass background rejection is described by a multi - dimensional _ acceptance _ function . the reconstructed @xmath1-ray signal ( @xmath2 ) within a suitably defined sampling region of the event property parameter space is a superposition of the true @xmath1-ray signal ( @xmath3 ) on an irreducible background ( @xmath4 ) of spuriously classified events . @xmath5 the magnitude of the background component is typically estimated using the number of @xmath1-ray - like events ( @xmath6 ) that are reconstructed within one or more nominally off - source regions . adopting this convention , is normally re - expressed as @xmath7 where @xmath0 is a normalisation factor which compensates for any disparities in the instrumental response among the various sampling regions . fundamentally , the precise value of @xmath0 is dependent upon specific characteristics of the individual observations that contribute to the overall data set . numerous variable factors such as the target zenith and azimuthal angles , the configuration and functionality of individual telescopes within a larger array , or the ambient atmospheric conditions may modify the nominal system performance . the functional dependency of @xmath0 on the various observational and instrumental parameters is complicated and is must typically be resolved for each observation using a multi - dimensional model estimate of the prevailing system acceptance . candidate models may be synthesised by interpolating the distributions of representative @xmath1-ray - like background data sets , assembled from multiple independent observations of empty fields - of - view . accurate replication of the true system acceptance using this technique requires an extensive database of suitable observations with characteristics that sample the entire space of observational parameters with adequate resolution . for certain well - understood parameters it may be possible to derive semi - analytic parameterisations of the corresponding systematic distortions , which can improve the interpolation accuracy and reduce the required sampling density nonetheless , for complicated instruments with a large number of possible configurations , the observation time required to populate the reference database may be prohibitive . alternatively , response models which inherently incorporate the appropriate systematic offsets can be constructed on an observation - wise basis using the subset of @xmath1-ray - like background events that fall outside of the fiducial sampling regions . however , this approach is susceptible to contamination by imperfectly excluded @xmath1-ray sources and is ultimately limited by the availability of event statistics within a single field - of - view . for the current generation of iacts , residual discrepancies between the generated model and the true instrumental response reflect a superposition of imperfectly modelled systematic offsets and are typically at the few percent level . however , if observational situations arise in which both approaches for modelling the system acceptance are compromised , the resultant disparity could be much larger . for example , if the telescope configuration or the prevailing operating conditions restrict the availability of appropriate archival background data and the target field - of - view is crowded , then systematic effects could dominate statistical uncertainties when searching for faint @xmath1-ray sources . moreover , since the combined impact of multiple systematic effects is often reflected by _ non - uniform _ variations in the system acceptance , the relative extent and location of the sampling regions within the event property parameter space may also bias the estimated value of @xmath0 . most experiments tacitly acknowledge the uncertainty in @xmath0 by attaching a conservative systematic error term to their quoted results . conventional stacking analyses typically entail simple summation of the observed event counts associated with the individual targets in the population sample , with the aggregate datasets forming the basis for subsequent astrophysical inference . however , the background estimation procedures outlined in @xmath8[subsec : obs_technique ] render this traditional approach ( hereafter referred to as _ data stacking _ ) sub - optimal in the context of vhe @xmath1-ray observations . in particular , data stacking analyses are complicated somewhat by the requirement to synthesise a combined value of the normalisation parameter @xmath9 which corresponds to the stacked dataset . a viable approach uses the definition of the overall signal excess , @xmath10 to form an average over all @xmath11 targets in the sample , weighted by the individual off - source counts . @xmath12 unfortunately , this technique fails to account for any target - specific uncertainties that arise during derivation of the individual @xmath13 . even if no systematic _ biases _ are introduced , the diverse observational parameters that correspond to each individual target dataset will likely affect accuracy to which the corresponding values of @xmath13 can be determined . the resultant non - uniformity renders subsequent propagation of the individual error estimates to the value of @xmath9 somewhat problematic for the data stacking approach . indeed , it is often unclear how the individual uncertainties should be combined in a statistically consistent manner , particularly when they are characterised by asymmetric intervals or more complicated probability distributions . assignment of a uniform but conservative systematic error term to all of the @xmath13 renders the problem tractable , but inevitably results in an overall loss of sensitivity . moreover , indications exist that the shortcomings of the data stacking technique can be problematic in experimental situations . indeed , @xcite derived anomalously high combined significance results when applying a data stacking procedure to two independent target ensembles comprising 37 and 44 putative vhe @xmath1-ray sources for which @xmath14 and @xmath15 . by studying the on and off - source event count distributions , @xcite concluded that a relative target - wise dispersion @xmath16 was required to explain the derived significances in the absence of a true @xmath1-ray signal . in subsequent sections , the statistical behaviour of the data stacking approach is investigated , highlighting the limited applicability of this technique in scenarios for which @xmath0 is uncertain . in response to these inherent shortcomings , a statistically robust method for the combination of iact observations is outlined , which allows systematic uncertainties to be treated on a target - by - target basis . using this technique , observations yielding weakly constrained values of @xmath13 may be usefully and consistently combined with more reliable datasets , without diluting the scientific utility of the latter . the new method operates by defining target - specific likelihood functions which facilitate appropriate treatment of relevant systematic uncertainties . stacking is implemented by forming a product of individual target likelihood functions and using this combined likelihood to estimate the shared characteristics of the putative @xmath1-ray signals . it should be emphasised that the applicability of the method is contingent upon an implicit assumption of at least one identical signal characteristic for all targets comprising the sample . realistically , signal characteristics for which the required assumption of universality is reasonable are typically target - specific functions of the experimental observables . as discussed in the subsequent paragraphs , disparities between putative sub - populations comprising the target sample are inherently problematic in the context of data stacking . in contrast , the suggested approach intrinsically facilitates straightforward customisation of the individual target likelihoods . construction of the single target likelihood function proceeds under the assumption that the sample of @xmath1-ray - like events that are reconstructed within a nominal sampling region is drawn from a parent population with poisson distributed arrival times . accordingly , if the value of @xmath0 were precisely constrained , then would imply @xmath17 where @xmath18 and @xmath19 denotes the true mean value of the poisson distributed variables @xmath2 and @xmath6 respectively . the new parameter @xmath20 represents the unknown , true value of the derived normalisation parameter . in fact , the assumed validity of is implicit in the derivation of @xmath9 in with likely deviations from this idealised behaviour undermining the reliability of the traditional data stacking framework . conversely , the combined likelihood approach treats target - specific systematic uncertainties by modelling @xmath0 as a random variable with arbitrary distribution @xmath21 . as indicated in @xmath8[subsec : obs_technique ] , the specific functional definitions of @xmath21 are dependent upon observational parameters which are likely to vary on a target - by - target basis . the probability density function which models an abstract single - target dataset @xmath22 is simply the product of the distributions of its component data . @xmath23 accordingly , substitution for @xmath24 using a concrete observational dataset @xmath25 enables calculation of the target - specific likelihood describing the true @xmath1-ray signal . @xmath26 modelling systematically uncertain parameters as random variables with known probability density functions is a common statistical approach , with applications extending beyond the limited domain of stacking analyses ( e.g. * ? ? ? indeed , equations and are equally valid in single target iact analyses that must accommodate non - negligible systematic uncertainties on the estimated value of @xmath0 . the ultimate goal of this stacking procedure is to obtain improved constraints on the global @xmath1-ray signal characteristics . in this context , @xmath19 and @xmath20 are not of primary interest and are therefore categorised as nuisance parameters throughout the subsequent analysis . finally , the joint likelihood describing the global signal characteristics of the entire ensemble of @xmath11 targets is simply the product of the individual target likelihoods . @xmath27 accordingly , the set of values which parameterise @xmath28 is the union of the parameters of the component single - target likelihood functions and therefore incorporates @xmath29 uncertain nuisance parameters . the initial objective of a stacking analysis is typically the conclusive detection of a combined @xmath1-ray excess , with detailed investigation of specific signal characteristics assuming subsidiary importance . in such situations , the probability of detection is typically evaluated in terms of the the significance , @xmath30 with which experimental data exclude the null hypothesis of zero signal . for an abstract likelihood function @xmath31 relating the experimental data @xmath32 to multiple nuisance parameters , @xmath33 and parameters of interest @xmath34 , a popular formula for calculation of detection significance @xcite employs the likelihood ratio defined by @xcite @xmath35 where a caret denotes the maximum likelihood estimate of the accented symbol , @xmath36 denotes the value of @xmath37 which corresponds to the null hypothesis , and @xmath38 represents the conditional maximum likelihood estimate for @xmath39 given that @xmath40 . the significance is subsequently defined in terms of @xmath41 @xmath42 which reduces to @xmath43^{n_{\rm on } } + n_{\rm off}\ln\left[(1+\alpha)\frac{n_{\rm off}}{n_{\rm on } + n_{\rm off } } \right]^{n_{\rm on}}\right\}^{\frac{1}{2}}\ ] ] as reported by ( * ? ? ? * equation 17 ) and is nominally @xmath44 distributed for a typical , single - target dataset with negligible @xmath0 uncertainty i.e. @xmath22 . proponents of data stacking typically emulate the single - target approach by using equation 17 in conjunction with a pseudo - dataset @xmath45 . consequently , non - uniform observational dependencies the individual @xmath13 become entangled , which inevitably complicates statistically robust treatment of their effects . conversely , direct substitution of @xmath46 in the calculation of @xmath47 in the joint - likelihood approach ensures that the resultant expression for @xmath30 is constructed using all available information regarding target - specific systematic uncertainties . recently , @xcite proposed an extension to the method suggested by @xcite which uses well - calibrated estimates of the instrumental point - spread function to achieve enhanced sensitivity relative to the more traditional analysis . their technique ameliorates the effect of systematic variations in the telescope acceptance characteristics by segregating the observational data on the basis of prevailing instrumental and environmental parameters . results that are derived from a joint likelihood stacking analysis comprise estimates of the various experimental parameters of interest , with any uncertainty quantified using associated confidence intervals . the frequentist definition of a properly constructed @xmath48 confidence interval states that @xmath48 of such intervals , generated from a large number of independent experimental measurements of a quantity @xmath49 , will contain ( or _ cover _ ) the true value @xmath50 . confidence intervals which fulfil this criterion are described as having correct _ coverage_. construction of confidence intervals typically uses experimental data to define a test statistic which is subsequently used to identify regions of the relevant parameter space that fulfil the required statistical criteria . the _ profile likelihood _ is an appropriate choice of test statistic in situations that require the treatment of multiple nuisance parameters . the profile likelihood is defined as the ratio @xmath51 where previously encountered symbols retain their earlier definitions and @xmath52 represents the conditional maximum likelihood estimate for @xmath39 assuming a specific value of @xmath37 . conveniently , the distribution of @xmath53 converges to that of a @xmath54 random variate with @xmath55 degrees of freedom for large experimental data sets ( e.g. * ? ? ? * ) . accordingly , straightforward derivation of confidence intervals is facilitated by comparison of the derived test statistic with appropriate percentiles of the relevant @xmath54 distribution . moreover , the profile likelihood has the desirable property of being independent of @xmath39 , facilitating the derivation of robust confidence intervals , even in the presence of uncertain nuisance parameters . verification of the performance and reliability joint likelihood stacking procedure employed a computerised toy monte carlo approach . the roofit framework @xcite was used to generate multiple simulated datasets consistent with the assumed distributions of @xmath2 , @xmath6 and @xmath0 for specific values of the true signal parameters @xmath56 $ ] , @xmath57 and @xmath58 . to better understand the effect of the distribution of @xmath0 on the outcome of each stacking procedure , data were generated using three alternative parameterisations for @xmath59 . a basic test case , simulating unbiassed systematic uncertainties , employed a gaussian parameterisation @xmath60 with @xmath61 ( hereafter referred to as _ model a _ ) . two less idealised examples , adopting bifurcated gaussian functions @xmath62 with @xmath63 and @xmath64 ( hereafter referred to as _ model b _ and _ model c _ , respectively ) , allowed the effect of increasing asymmetry in the modelled @xmath0 distribution to be investigated . figure [ fig : alphadists ] shows representative distributions of @xmath0 corresponding to each of the simulated scenarios . models a and b mimic somewhat pessimistic observational situations in which the choice of off - source sampling regions is restricted , limiting event statistics and biasing the derivation of @xmath0 . accordingly , the corresponding results are particularly relevant to galactic source populations , for which source confusion and diffuse background contamination within target fields - of - view can be particularly problematic . model c represents an extreme case which is unlikely to occur in real experimental contexts , but is nonetheless included to demonstrate the efficacy of the joint likelihood technique . corresponding to model a ( _ black _ ) , model b ( _ red _ ) , and model c ( _ blue_).,scaledwidth=60.0% ] using roofit , ensembles of the generated single - target datasets were then employed in conjunction with the relevant compound probability density functions to obtain multiple realisations of the joint likelihood @xmath65 corresponding to a discrete range of target multiplicities @xmath66 $ ] . for each ensemble dataset @xmath67 , the statistical significance ( @xmath68 ) of the combined @xmath1-ray signal was evaluated using the likelihood ratio prescription of @xcite @xmath69 for comparison , a traditional data stacking analysis was also applied to the generated data . the components of the ensemble datasets used to construct each realisation of @xmath65 were combined to obtain corresponding stacked datasets , @xmath70 . subsequent statistical inference employed a likelihood function @xmath71 , which was formed by substitution of the stacked data into the single target probability density function @xmath72 with @xmath73 defined as a delta function centred on the stacked @xmath9 estimate , @xmath74 . following the convention outlined in @xmath8[sec : inf_sig_char ] , equation was evaluated using the components of @xmath75 do derive the statistical significance @xmath76 of the combined @xmath1-ray signal corresponding to the data stacking approach . for both stacking methodologies , construction of the associated profile likelihoods was accomplished using the intrinsic capabilities of roofit , adopting @xmath77 as the sole parameter of interest . confidence intervals were then defined as the ranges of @xmath77 outside which @xmath53 exceeded an appropriate percentile of the nominal @xmath78 distribution . for each distinct combination of the true signal parameters , modelled distribution of @xmath0 , and target multiplicity , the monte carlo approach outlined in @xmath8[sec : monte_carlo ] was applied to derive 5000 independent estimates of @xmath79 and @xmath30 corresponding to each stacking technique . in combination with appropriate statistical metrics , the resultant data enable a comparative evaluation of the joint likelihood and traditional data stacking approaches . an experimentally relevant diagnostic for both stacking analyses is the probability with which data corresponding to a particular value of @xmath77 will yield a significance @xmath30 , in excess of a specified threshold @xmath80 . assuming data that are distributed in accordance with the null hypothesis , this probability is called the _ significance level_. although the specific choice of @xmath80 is arbitrary , it is typically chosen in order to obtain a desired significance level and ensure a controlled rate of false - positive detections . objective comparison of the alternative stacking techniques is facilitated by a threshold which corresponds to a nominal percentile of the appropriate null distribution of @xmath30 . a related statistic is the _ power _ , which describes the probability that derived values of @xmath81 will correctly identify a genuine signal , and provides useful insights regarding the sensitivity of the corresponding analytical method . the relevant monte carlo distributions for @xmath30 , corresponding to @xmath82 , are plotted for each target multiplicity in figure [ fig : null_sigma_dists ] . assuming that the true value of @xmath20 is 0.1 , systematic uncertainties which affect the estimation of @xmath13 according to the distributions plotted in figure [ fig : alphadists ] will generate misleading results in a data stacking analysis . in many experimental situations , a small number of erroneous results is tolerable , as long as the expected frequency with which they occur can be reliably predicted . this predictability criterion is realised if the derived values consistently conform to a known statistical distribution . irrespective of @xmath21 , and for all @xmath83 $ ] , the significance estimates derived using the joint likelihood technique appear to retain the @xmath44 distribution which arises in case of zero @xmath0 uncertainty . in contrast , results that correspond to traditional data stacking exhibit substantial deviations from this nominal distribution . moreover , adoption of asymmetric @xmath0 distributions introduces a monotonic dependence on the target multiplicity , yielding systematically larger significance estimates as @xmath11 increases . consideration of the procedure used during data stacking to synthesise a combined @xmath0 estimate suggests an intuitive explanation of the observed behaviour . models b and c generate ensemble data sets in which the subset of overestimated @xmath0 values is typically predominant and has an absolute _ mean _ offset from @xmath20 which exceeds that of the remaining , underestimated @xmath13 . accordingly , the frequency with which equation yields a substantial overestimate for @xmath9 should increase with the ensemble size , which is consistent with the observed behaviour . conversely , symmetrically distributed @xmath0 uncertainties should not inherently bias the data stacking process . instead , the target - specific weighting of each @xmath13 specified in equation prevents perfect cancellation of the systematic offsets and inhibits convergence to the nominal null distribution for @xmath30 as @xmath84 . as outlined in @xmath8[sec : monte_carlo ] all monte carlo data sets used for this study assume a common @xmath57 , with the range of inter - target variability in @xmath85 restricted to the level of poisson fluctuations . in experimental contexts , the true number of off events is likely to be different for each target and the weights accorded to the corresponding @xmath13 would be more diverse . the adverse impact of symmetric systematic uncertainties would then depend critically on the joint , target - wise distribution of @xmath0-offset and @xmath85 , with specific ensembles potentially realising particularly benign or pathological scenarios . as an alternative to the absolute measure of significance defined by equation , experimental practitioners often choose to multiply @xmath30 by the sign of the corresponding @xmath1-ray excess . adopting this convention , systematic _ _ over__estimates of @xmath9 during data stacking would tend to over - generate negative significances when @xmath82 . in many situations the notion of a negative signal is not physically meaningful and such results would probably be disregarded . in general , the ability to eliminate spurious results on the basis of their physical validity may ameliorate the impact of particular systematic effects , and reduce the rate of false - positive detections . nonetheless , unfeasibly negative significances should raise concerns regarding the trustworthiness of corresponding upper limits . corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] + corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] + corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood and data stacking approaches . each histogram is representative of 5000 independent monte carlo datasets , which were generated assuming @xmath82 . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) for target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] significance thresholds corresponding to the 95th percentile of the various null distributions of @xmath30 are plotted in the left - hand column of figure [ fig : power_diffs ] . results derived using the joint - likelihood approach exhibit minimal multiplicity dependence and appear consistent with the nominal @xmath86 expected for a @xmath44 random variate . conversely , thresholds which correspond to the data stacking approach systematically exceed their joint likelihood counterparts , and become increasing discrepant for larger @xmath11 in simulations that assume asymmetric distributions of @xmath0 . accordingly , recourse to a nominal distribution of @xmath30 is not appropriate in the data stacking framework , and valid interpretation of the significance of a particular detection requires empirical calibration of distinct significance thresholds for each target multiplicity . the values of @xmath87 plotted in figure [ fig : power_diffs ] permit derivation of appropriately calibrated estimates of the statistical power at each vertex of the investigated model parameter space . the resultant probabilities are presented in figure [ fig : powers ] , while the results plotted in the right - hand column of figure [ fig : power_diffs ] reveal the corresponding differences power between the alternative stacking techniques . results pertaining to models a and b exhibit saturation at extreme values of @xmath77 and @xmath11 , caused by the limited statistics of the monte carlo datasets . results corresponding a symmetric gaussian distribution for @xmath0 reveal derived powers which are compatible to @xmath88 for all @xmath66 $ ] and @xmath56 $ ] . although traditional data stacking appears to offer marginally superior performance for this scenario , asymmetric parameterisations for the distribution of @xmath0 reverse this outcome . indeed , powers derived using the joint likelihood approach in conjunction with models b and c exhibit substantial enhancement with respect to their data stacking counterparts . moreover , powers corresponding to the data stacking technique exhibit a marked sensitivity to the degree of asymmetry in the adopted @xmath0 distribution , which appears to be substantially ameliorated by the joint likelihood approach . strong suppression of the resultant data stacking power is evident when model c is applied to the monte carlo analysis , with derived discrepancies @xmath89 separating the alternative stacking techniques at the upper extremes of the modelled parameter ranges . more significantly for practical applications , assuming the intermediate scenario represented by model b , reveals a maximum disparity of @xmath90 for the experimentally interesting region of @xmath91 as @xmath92 . in this context , the joint likelihood approach yields a two - fold enhancement in the likelihood of detecting a genuine @xmath1-ray signal . for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] + for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] + for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] for the joint likelihood ( _ blue _ ) and data stacking ( _ red _ ) approaches . _ right - hand column : _ differences between the derived powers corresponding to the joint likelihood and data stacking approaches . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . each bin is derived using analyses of 5000 independent monte carlo datasets which assume to true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=45.0% ] corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] + corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] + corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] corresponding to the 95th percentile of the null distributions of @xmath30 for the joint likelihood ( _ left - hand column _ ) and data stacking ( _ right - hand column _ ) approaches . each bin is derived using analyses of 5000 independent monte carlo datasets . the plotted results correspond to model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) assuming true signals @xmath56 $ ] and target multiplicities @xmath66 $ ] . the underlying datasets are as for figure [ fig : power_diffs],title="fig:",scaledwidth=45.0% ] for an analysis to perform reliably in an experimental context , it is important that any generated confidence intervals have correct coverage . in reality , despite sophisticated treatment of associated systematic and statistical uncertainties , residual disparities between the assumed and actual distributions of experimental data often prevent practical realisation of this criterion . moreover , deviations from correct coverage often lead to an unexpected increase in the frequency of spuriously inferred scientific conclusions . accordingly , the degree by which the derived confidence intervals deviate from their nominal coverage provides an informative comparator for the alternative stacking techniques . which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] + which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] + which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] which include the true signal value , @xmath77 . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] figure [ fig : coverage_fractions ] displays the fraction of derived 95% confidence intervals for @xmath3 which cover the true signal value , for each parameterisation of @xmath21 . complementary insight is provided by figure [ fig : coverage_residuals ] , which illustrates the corresponding deviations from correct coverage as a function of @xmath77 and the target multiplicity , @xmath11 . it is evident that confidence intervals generated by the traditional data stacking approach are generally incompatible with their stated coverage . while the derived coverage fractions exceed the nominal 95% for low signals and target multiplicities , the majority of the @xmath93 parameter space is characterised significant negative deviations . this behaviour is typically described as _ under - coverage _ , and implies excessively permissive confidence intervals with an increased propensity to generate false - positive results . the observed under - coverage is maximised for small non - zero signals and appears to be exacerbated by the addition of targets to the ensemble dataset . moreover , increased asymmetry of the modelled @xmath0 distribution has a markedly detrimental effect on the derived coverage characteristics . indeed , the maximum divergence from correct coverage increases from 4% to 25% and then to 85% for model a , model b , and model c , respectively . in the worst case , this means that only 10% of the derived intervals include the true parameter value . the inherent inability of the data stacking technique to acknowledge systematic uncertainties leads to confidence intervals which are unrealistically narrow . consequently , reported accuracy of experimental measurements will be excessively optimistic and in the worst cases could be used to infer spurious scientific conclusions . corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] + corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] + corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] corresponding to the joint likelihood ( jl ) and data stacking ( ds ) approaches . results are shown which correspond model a ( _ top row _ ) , model b ( _ middle row _ ) and model c ( _ bottom row _ ) . the underlying datasets are as for figure [ fig : power_diffs].,title="fig:",scaledwidth=45.0% ] in contrast , notable stability in the coverage fractions derived using the joint likelihood technique indicates robust handling of systematic uncertainties . confidence intervals derived under the assumption of a gaussian distribution in @xmath0 exhibit coverage which is correct to within @xmath94 throughout the studied region of the model parameter space . variation of the coverage fraction as a function of @xmath77 remains minimal when asymmetric parametrisations of @xmath21 are assumed . although under - coverage which worsens with increasing target multiplicity is evident in results that correspond to models b and c , the discrepancy with respect to correct coverage does not exceed @xmath95 . the comparative evaluation presented in @xmath8[sec : results ] reveals several limitations to the applicability of traditional data stacking in situations when @xmath0 is uncertain , many of which are addressed by the joint likelihood technique . reliable inference of signal characteristics is impaired by significant levels of under - coverage for @xmath96 in the data stacking approach . moreover , the inherent permissivity of derived confidence intervals is compounded by the inclusion of additional targets . in combination , these characteristics imply a substantially increased tendency to generate spurious results which is exacerbated as the stacked target ensemble is enlarged . in an experimental context , this behaviour would be highly undesirable and largely undermines the motivation for combined source analysis . in contrast , confidence intervals provided by the joint likelihood approach closely approximate correct coverage under the assumption of symmetric uncertainties on @xmath0 , and exhibit multiplicity - dependent under - coverage at @xmath97 levels if an asymmetric distribution is adopted . the deviations from correct coverage that are exhibited by the data stacking approach are consistent with the erroneous assumption of negligible uncertainty in the value of @xmath0 . resolution of this issue is complicated by the lack of a straightforward prescription for propagating systematic uncertainties to the calculated value of @xmath9 . conversely , the joint likelihood method provides a robust technique for the treatment of uncertain @xmath0 estimates on a target - specific basis . appropriate interpretation of the combined signal significance is also more complicated within the data stacking framework . the monte carlo null distributions for @xmath30 presented in figure [ fig : null_sigma_dists ] indicate that application of equation requires proper calibration of @xmath87 to prevent spurious rejection of the null hypothesis . the situation deteriorates when increasingly asymmetric parameterisations of the generated @xmath0 distribution are adopted . under these circumstances , nave assumption of a nominal @xmath44 distribution for @xmath30 induces a systematic overestimation of corresponding significance levels which worsens with increasing target multiplicity . in an experimental context , the implied increase in the false - positive detection rate for larger target samples is contrary to common expectation and therefore represents a serious shortcoming of the data stacking technique . in contrast , significance estimates derived for @xmath77 using the joint likelihood approach appear insensitive to the value of @xmath11 and display excellent correspondence with the expected @xmath44 distribution . notwithstanding appropriate calibration of a suitable detection threshold @xmath87 , the powers which characterise the data stacking technique indicate that asymmetry in the adopted @xmath0 distribution substantially diminishes the overall sensitivity . conversely , the powers corresponding to the joint likelihood analysis appear relatively unaffected by the choice of @xmath21 . the alternative approaches exhibit similar properties , and indeed data stacking appears to marginally outperform the joint likelihood approach , if symmetric systematic uncertainties are assumed . it should be acknowledged that the degree of adverse behaviour inherent in the data stacking technique is contingent upon the adopted values of @xmath20 and @xmath19 , as well as the assumed magnitude of the simulated systematic effects . accurate modelling of the system acceptance may ameliorate the observed pathologies , with the cumulative effect of target - specific @xmath0 uncertainties only becoming significant for target ensembles with @xmath98 . figure [ fig : delta_alpha_mag ] illustrates the combined effect on data stacking of systematic offsets at the few percent level using calibrated values of the significance threshold that correspond to the 95th percentile of simulated null distributions for @xmath30 . the monte carlo data sets assume relative target - wise uncertainties @xmath99 , for both symmetric and asymmetric parameterisations for @xmath21 . the curves correspond to simulated ensembles comprising @xmath100 target sources for which @xmath101 and @xmath102 . despite an inevitable reduction in the influence of the modelled systematic offsets , it is clear that even small target - specific biasses can reinforce to distort the combined significance for large observational datasets . moreover , even the pessimistic parameter values that define models a and b may be practically relevant , especially when the extent of off - source regions is restricted by nearby sources , or stringent event selection limits the overall count statistics . for the and data stacking approach . the illustrated results correspond to gaussian ( _ left _ ) and bifurcated gaussian ( _ right _ ) parameterisations for @xmath21 , using target - specific uncertainties @xmath103 for the symmetric models , and @xmath104 with @xmath105 for the asymmetric models . each datapoint is derived using analyses of 5000 independent monte carlo datasets , which assume ensembles of @xmath100 targets having true signals @xmath106 . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=48.0% ] for the and data stacking approach . the illustrated results correspond to gaussian ( _ left _ ) and bifurcated gaussian ( _ right _ ) parameterisations for @xmath21 , using target - specific uncertainties @xmath103 for the symmetric models , and @xmath104 with @xmath105 for the asymmetric models . each datapoint is derived using analyses of 5000 independent monte carlo datasets , which assume ensembles of @xmath100 targets having true signals @xmath106 . all monte carlo datasets were generated assuming @xmath57 and @xmath58.,title="fig:",scaledwidth=48.0% ] the joint likelihood method demonstrates good performance in a number of simple but experimentally relevant scenarios where the reliability of traditional data stacking is seriously impaired . nonetheless , it should be stated that the _ full _ potential of the joint likelihood method can only be realised if the distributions of all the systematic uncertainties affect each observational dataset are parameterisable with sufficient accuracy . admittedly , definitive derivation of the distribution of @xmath0 is unlikely to be straightforward in many situations . semi - analytic estimation of the required distributions necessitates detailed knowledge of the instrumental response as well as careful measurement of the environmental conditions at the time of observation . accordingly , imperfect or incomplete understanding of the dominant systematic biases might render this approach untenable . plausible alternative strategies include jackknife or bootstrap resampling of off - source data from randomly selected background sampling regions to derive multiple estimates for @xmath0 and construct an empirical distribution . in this case , the accuracy of the resultant parameterisation is limited by the finite number of permutations available for the resampling process . emulating the approach of @xcite and segregating data on the basis of quantifiable _ operating conditions _ would enable the use of large representative datasets to derive template distributions that pertain to finite subsets of the observational parameter space . in situations where the suggested techniques are unable to accurately reconstruct the shape of the @xmath0 distribution , a generic parameterisation could be tailored using target - specific estimates of the dispersion in derived values of the on / off normalisation . by acknowledging the existence of the associated systematic uncertainties , this first - order approximation would ameliorate under - coverage as well as deviations from the nominal @xmath44 distribution for @xmath30 , thereby reducing the rate of unexpected false - positive detections moreover , by facilitating limited adaptation of the modelled systematic uncertainties , this approach retains a key advantage of the full joint likelihood technique that can not be straightforwardly emulated by traditional data stacking . indeed , it is likely that this simplified approach might recover a significant proportion of the power which is achieved when the true distribution of @xmath0 is available . nonetheless , as demonstrated by the monte carlo results for a symmetric distribution of @xmath0 , there may be situations in which the advantages offered by the joint likelihood approach are marginal and the performance of the data stacking approach is deemed sufficient . while this study has focussed upon parameterisation of @xmath0 uncertainties , the joint likelihood technique represents a highly flexible and extensible analytical approach . indeed , related analyses are able to parameterise arbitrary properties of the instrument response , the measured @xmath1-ray signal , and even suspected source properties such as the spectral index and temporal variability ( e.g. * ? ? ? toy monte carlo simulations have been used to investigate the effect of non - negligible systematic uncertainties affecting the relative normalisation of the on and off - source event sampling regions in atmospheric cherenkov telescope observations . as a specific example , the properties of two alternative strategies for combined source analysis have been examined . situations have been identified in which the traditional _ data stacking _ approach exhibits unexpected and undesirable statistical behaviour . even when correctly calibrated null distributions for the combined significance are employed , the sensitivity of the data stacking approach appears markedly impaired if asymmetrically distributed @xmath0 uncertainties are assumed . the data stacking technique is also yields unreliable estimates for the @xmath1-ray signal parameters , with derived confidence intervals deviate from their stated coverage by a margin which widens with increasing target multiplicity . an alternative to data stacking has been outlined which combines target - specific likelihood functions that explicitly account for non - uniform systematic uncertainties . as an analytical approach , the _ joint likelihood _ technique exhibits many characteristics which surpass those of traditional data stacking . it offers a statistically robust prescription for the treatment of target - specific systematic uncertainties , and effectively addresses the shortcomings which are inherent in the data stacking framework . indeed , the joint likelihood effectively eliminates pathological behaviour inherent in data stacking , with combined source significances conforming to a single , nominal null distribution . moreover , it achieves equivalent or superior sensitivity to the data stacking technique and yields parameter confidence intervals that exhibit minimal deviation from their stated coverage . many of the insights provided by this study of combined source analyses are equally applicable to single - target analyses involving similarly abundant event statistics . indeed , the synthesis of any extensive dataset often combines observations which incorporate diverse systematic effects . this study has shown that the joint likelihood method provides a viable approach for the robust combination of data with inhomogeneous systematic uncertainties , irrespective of the celestial target coordinates . although it is a generally applicable technique , the joint likelihood analysis is likely to prove most useful in experimental situations involving limited event statistics and a restricted choice of off - source sampling regions . , m. , ajello , m. , albert , a. , atwood , w. b. , baldini , l. , ballet , j. , barbiellini , g. , bastieri , d. , bechtol , k. , bellazzini , r. , berenji , b. , blandford , r. d. , bloom , e. d. , bonamente , e. , borgland , a. w. , bregeon , j. , brigida , m. , bruel , p. , buehler , r. , burnett , t. h. , buson , s. , caliandro , g. a. , cameron , r. a. , caadas , b. , caraveo , p. a. , casandjian , j. m. , cecchi , c. , charles , e. , chekhtman , a. , chiang , j. , ciprini , s. , claus , r. , cohen - tanugi , j. , conrad , j. , cutini , s. , de angelis , a. , de palma , f. , dermer , c. d. , digel , s. w. , do couto e silva , e. , drell , p. s. , drlica - wagner , a. , falletti , l. , favuzzi , c. , fegan , s. j. , ferrara , e. c. , fukazawa , y. , funk , s. , fusco , p. , gargano , f. , gasparrini , d. , gehrels , n. , germani , s. , giglietto , n. , giordano , f. , giroletti , m. , glanzman , t. , godfrey , g. , grenier , i. a. , guiriec , s. , gustafsson , m. , hadasch , d. , hayashida , m. , hays , e. , hughes , r. e. , jeltema , t. e. , jhannesson , g. , johnson , r. p. , johnson , a. s. , kamae , t. , katagiri , h. , kataoka , j. , kndlseder , j. , kuss , m. , lande , j. , latronico , l. , lionetto , a. m. , llena garde , m. , longo , f. , loparco , f. , lott , b. , lovellette , m. n. , lubrano , p. , madejski , g. m. , mazziotta , m. n. , mcenery , j. e. , mehault , j. , michelson , p. f. , mitthumsiri , w. , mizuno , t. , monte , c. , monzani , m. e. , morselli , a. , moskalenko , i. v. , murgia , s. , naumann - godo , m. , norris , j. p. , nuss , e. , ohsugi , t. , okumura , a. , omodei , n. , orlando , e. , ormes , j. f. , ozaki , m. , paneque , d. , parent , d. , pesce - rollins , m. , pierbattista , m. , piron , f. , pivato , g. , porter , t. a. , profumo , s. , rain , s. , razzano , m. , reimer , a. , reimer , o. , ritz , s. , roth , m. , sadrozinski , h. f .- w . , sbarra , c. , scargle , j. d. , schalk , t. l. , sgr , c. , siskind , e. j. , spandre , g. , spinelli , p. , strigari , l. , suson , d. j. , tajima , h. , takahashi , h. , tanaka , t. , thayer , j. g. , thayer , j. b. , thompson , d. j. , tibaldo , l. , tinivella , m. , torres , d. f. , troja , e. , uchiyama , y. , vandenbroucke , j. , vasileiou , v. , vianello , g. , vitale , v. , waite , a. p. , wang , p. , winer , b. l. , wood , k. s. , wood , m. , yang , z. , zimmer , s. , kaplinghat , m. , martinez , g. d. , dec . constraining dark matter models from a combined analysis of milky way satellites with the fermi large area telescope . physical review letters 107 ( 24 ) , 241302 . , f. , akhperjanian , a. g. , anton , g. , barres de almeida , u. , bazer - bachi , a. r. , becherini , y. , behera , b. , bernlhr , k. , boisson , c. , bochow , a. , borrel , v. , braun , i. , brion , e. , brucker , j. , brun , p. , bhler , r. , bulik , t. , bsching , i. , boutelier , t. , chadwick , p. m. , charbonnier , a. , chaves , r. c. g. , cheesebrough , a. , chounet , l .- m . , clapson , a. c. , coignet , g. , dalton , m. , daniel , m. k. , davids , i. d. , degrange , b. , deil , c. , dickinson , h. j. , djannati - ata , a. , domainko , w. , oc . drury , l. , dubois , f. , dubus , g. , dyks , j. , dyrda , m. , egberts , k. , emmanoulopoulos , d. , espigat , p. , farnier , c. , feinstein , f. , fiasson , a. , frster , a. , fontaine , g. , fling , m. , gabici , s. , gallant , y. a. , grard , l. , giebels , b. , glicenstein , j. f. , glck , b. , goret , p. , hauser , d. , hauser , m. , heinz , s. , heinzelmann , g. , henri , g. , hermann , g. , hinton , j. a. , hoffmann , a. , hofmann , w. , holleran , m. , hoppe , s. , horns , d. , jacholkowska , a. , de jager , o. c. , jung , i. , katarzyski , k. , katz , u. , kaufmann , s. , kendziorra , e. , kerschhaggl , m. , khangulyan , d. , khlifi , b. , keogh , d. , komin , n. , kosack , k. , lamanna , g. , lenain , j .- p . , lohse , t. , marandon , v. , martin , j. m. , martineau - huynh , o. , marcowith , a. , maurin , d. , mccomb , t. j. l. , medina , m. c. , moderski , r. , moulin , e. , naumann - godo , m. , de naurois , m. , nedbal , d. , nekrassov , d. , niemiec , j. , nolan , s. j. , ohm , s. , olive , j .- f . , de oa wilhelmi , e. , orford , k. j. , ostrowski , m. , panter , m. , paz arribas , m. , pedaletti , g. , pelletier , g. , petrucci , p .- o . , pita , s. , phlhofer , g. , punch , m. , quirrenbach , a. , raubenheimer , b. c. , raue , m. , rayner , s. m. , reimer , o. , renaud , m. , rieger , f. , ripken , j. , rob , l. , rosier - lees , s. , rowell , g. , rudak , b. , rulten , c. b. , ruppel , j. , sahakian , v. , santangelo , a. , schlickeiser , r. , schck , f. m. , schrder , r. , schwanke , u. , schwarzburg , s. , schwemmer , s. , shalchi , a. , skilton , j. l. , sol , h. , spangler , d. , stawarz , . , steenkamp , r. , stegmann , c. , superina , g. , szostek , a. , tam , p. h. , tavernet , j .- terrier , r. , tibolla , o. , van eldik , c. , vasileiadis , g. , venter , c. , venter , l. , vialle , j. p. , vincent , p. , vivier , m. , vlk , h. j. , volpe , f. , wagner , s. j. , ward , m. , zdziarski , a. a. , zech , a. , may 2009 . hess upper limit on the very high energy @xmath1-ray emission from the globular cluster 47 tucanae . 499 , 273277 . , m. f. , fegan , d. j. , gibbs , k. , gorham , p. w. , hillas , a. m. , lamb , r. c. , liebing , d. f. , mackeown , p. k. , porter , n. a. , stenger , v. j. , aug . 1985 . application of imaging to the atmospheric cherenkov technique . in : jones , f. c. ( ed . ) , international cosmic ray conference . . 453456 . , abramowski , a. , acero , f. , aharonian , f. , akhperjanian , a. g. , anton , g. , barnacka , a. , barres de almeida , u. , bazer - bachi , a. r. , becherini , y. , becker , j. , behera , b. , bernlhr , k. , bochow , a. , boisson , c. , bolmont , j. , bordas , p. , borrel , v. , brucker , j. , brun , f. , brun , p. , bulik , t. , bsching , i. , carrigan , s. , casanova , s. , cerruti , m. , chadwick , p. m. , charbonnier , a. , chaves , r. c. g. , cheesebrough , a. , chounet , l .- m . , clapson , a. c. , coignet , g. , conrad , j. , dalton , m. , daniel , m. k. , davids , i. d. , degrange , b. , deil , c. , dickinson , h. j. , djannati - ata , a. , domainko , w. , drury , l. o. c. , dubois , f. , dubus , g. , dyks , j. , dyrda , m. , egberts , k. , eger , p. , espigat , p. , fallon , l. , farnier , c. , fegan , s. , feinstein , f. , fernandes , m. v. , fiasson , a. , fontaine , g. , frster , a. , fling , m. , gallant , y. a. , gast , h. , grard , l. , gerbig , d. , giebels , b. , glicenstein , j. f. , glck , b. , goret , p. , gring , d. , hague , j. d. , hampf , d. , hauser , m. , heinz , s. , heinzelmann , g. , henri , g. , hermann , g. , hinton , j. a. , hoffmann , a. , hofmann , w. , hofverberg , p. , horns , d. , jacholkowska , a. , de jager , o. c. , jahn , c. , jamrozy , m. , jung , i. , kastendieck , m. a. , katarzyski , k. , katz , u. , kaufmann , s. , keogh , d. , kerschhaggl , m. , khangulyan , d. , khlifi , b. , klochkov , d. , kluniak , w. , kneiske , t. , komin , n. , kosack , k. , kossakowski , r. , laffon , h. , lamanna , g. , lennarz , d. , lohse , t. , lopatin , a. , lu , c .- c . , marandon , v. , marcowith , a. , masbou , j. , maurin , d. , maxted , n. , mccomb , t. j. l. , medina , m. c. , mhault , j. , moderski , r. , moulin , e. , naumann , c. l. , naumann - godo , m. , de naurois , m. , nedbal , d. , nekrassov , d. , nguyen , n. , nicholas , b. , niemiec , j. , nolan , s. j. , ohm , s. , olive , j .- f . , de oa wilhelmi , e. , opitz , b. , ostrowski , m. , panter , m. , paz arribas , m. , pedaletti , g. , pelletier , g. , petrucci , p .- o . , pita , s. , phlhofer , g. , punch , m. , quirrenbach , a. , raue , m. , rayner , s. m. , reimer , a. , reimer , o. , renaud , m. , de los reyes , r. , rieger , f. , ripken , j. , rob , l. , rosier - lees , s. , rowell , g. , rudak , b. , rulten , c. b. , ruppel , j. , ryde , f. , sahakian , v. , santangelo , a. , schlickeiser , r. , schck , f. m. , schnwald , a. , schwanke , u. , schwarzburg , s. , schwemmer , s. , shalchi , a. , sikora , m. , skilton , j. l. , sol , h. , spengler , g. , stawarz , . , steenkamp , r. , stegmann , c. , stinzing , f. , sushch , i. , szostek , a. , tavernet , j .- terrier , r. , tibolla , o. , tluczykont , m. , valerius , k. , van eldik , c. , vasileiadis , g. , venter , c. , vialle , j. p. , viana , a. , vincent , p. , vivier , m. , vlk , h. j. , volpe , f. , vorobiov , s. , vorster , m. , wagner , s. j. , ward , m. , wierzcholska , a. , zajczyk , a. , zdziarski , a. a. , zech , a. , zechlin , h .- s . , h.e.s.s . collaboration , mar . constraints on dark matter annihilations towards the sculptor and carina dwarf galaxies . astroparticle physics 34 , 608616 . , s. , degrange , b. , punch , m. , barrau , a. , bazer - bachi , r. , cabot , h. , chounet , l. m. , debiais , g. , dezalay , j. p. , djannati - atai , a. , dumora , d. , espigat , p. , fabre , b. , fleury , p. , fontaine , g. , george , r. , ghesquiere , c. , goret , p. , gouiffes , c. , grenier , i. a. , iacoucci , l. , malet , i. , meynadier , c. , munz , f. , palfrey , t. a. , pare , e. , pons , y. , quebert , j. , ragan , k. , renault , c. , rivoal , m. , rob , l. , schovanek , p. , smith , d. , tavernet , j. p. , vrana , j. , oct . 1998 . a new analysis method for very high definition imaging atmospheric cherenkov telescopes as applied to the cat telescope . nuclear instruments and methods in physics research a 416 , 425437 . neyman , j. , pearson , e. s. , 1933 . on the problem of the most efficient tests of statistical hypotheses . philosophical transactions of the royal society of london . series a , containing papers of a mathematical or physical character 231 ( 694 - 706 ) , 289337 . http://rsta.royalsocietypublishing.org / content/231/694-% 706/289.short[http://rsta.royalsocietypublishing.org / content/231/694-% 706/289.short ] , t. c. , cawley , m. f. , fegan , d. j. , gibbs , k. g. , hillas , a. m. , kowk , p. w. , lamb , r. c. , lewis , d. a. , macomb , d. , porter , n. a. , reynolds , p. t. , vacanti , g. , jul . 1989 . observation of tev gamma rays from the crab nebula using the atmospheric cerenkov imaging technique . 342 , 379395 .

the lagrangian points of a gravitationally bound two - body system are a feature of newtonian gravity . unlike general relativity ( gr ) newton s gravity admits analytic solutions for the two - body problem ; furthermore looking for the positions , on the joint orbital plane , where the attraction of both bodies on a negligible mass test particle counterbalances exactly the centrifugal force , one finds five points where such a condition is fulfilled , with an orbital angular velocity coinciding with that of the two main bodies around their common center of mass . the traditional labelling of the five points is @xmath0 , @xmath1 , @xmath5 , @xmath2 , @xmath3 and the geometry of the system is as sketched in fig . [ fig . 1 ] . + three points ( @xmath6 , @xmath7 , @xmath8 ) are saddle points of the effective potential ; in other words , the equilibrium there is unstable , however in the case of the sun / earth pair the instability is very mild . the remaining two points ( @xmath9 and @xmath10 ) are real local minima so the equilibrium there is stable , though corresponding to a shallow potential well . when coming to a relativistic approach , even though we may guess that the situation is marginally or even negligibly different from the newtonian case , the existence of lagrangian points is a priori not garanteed so that the problem needs a careful discussion . hopefully the final conclusion is indeed that libration points ( lagrange points ) still exist also in a relativistic 2 plus 1 body configuration , at least in a range of masses including the sun / earth pair , even though no closed solution is available for the position of such points @xcite . the advantage of the set of the lagrangian points is that they form a configuration rigidly rotating together with the earth . this property has already been exploited many times for space missions , such as wmap , the herschel space observatory , planck ( all concluded ) and now gaia , in @xmath7 ; the deep space climate observatory , the solar and heliospheric observatory ( soho ) and lisa pathfinder , in @xmath6 . the list is not exhaustive and many more missions are planned directed again to @xmath6 or @xmath7 . it is also worth mentioning that proposals have been issued to exploit , for fundamental physics , the lagrangian points of the pair earth - moon @xcite . the stability of the positions with respect to one another and to the earth makes the lagrangian points appropriate to work as basis for a physical reference frame at the scale of the inner solar system . furthermore , considering the size of the `` graph '' having the @xmath11 s as knots , we may remark that the time of flight of electromagnetic signals going from one point to another is in the order of some 10 minutes or more ; such long time may act as a multiplier for the tiny asymmetries originated by the angular momentum of the sun and for the strain due to passing gravitational waves . the present paper will discuss the possibilities listed above , highlighting the advantages for fundamental physics experiments and for the positioning and guidance of spacecraft out of the terrestrial environment . in particular , we shall nickname lagrange the proposal of exploiting time of flight measurements along a closed path having @xmath11 points as vertices , in order to take advantage of the asymmetric propagation produced by the spin of the sun in the case of two counter - rotating electromagnetic signals . the use of one and the same loop will avoid delays due to different geometric paths for the two beams ; the cancellation of the purely geometric component of the time of flight will let the above mentioned asymmetry emerge . in section [ sec - timedelay ] we discuss the gr time delay due to the sun ; then in subsection [ sec - delay ] , we specialize the analysis to the case where emitter , central body and receiver are aligned ( indeed a special case for lagrange ) . in the calculation , the contribution of the quadrupole of the central body will be included . in section [ lt ] the analysis will concern the more general configuration with a closed contour encompassing a wide area and the purpose will be the measurement of the solar gravitomagnetic effect with an accuracy better than @xmath4 . section [ sec : galactic ] evaluates the possibilities to retrieve information about a possible galactic gravito - magnetic field , if it is there . section [ sec : rps ] presents a relativistic positioning system ( rps ) based on a set of emitters of electromagnetic signals located at the mentioned l - points . in section [ other ] some additional opportunities given by the lagrangian reference system , including gravitational waves ( gw ) detection , are listed . methodological and technological challenges posed by the lagrange project are reviewed in sect . [ sec - tech ] . a short conclusion closes the paper . a gravitational field produces a time delay and a deflection ( deviation ) on the propagation of electromagnetic waves . presently , we are interested only to describe the effects of the delays in time propagation , because the effects of the deflection on the time propagation are negligible with respect to the leading contributions . the main effect depends on the mass of the central body ( the sun ) and is fully explained in terms of the metric developed by schwarzschild in 1916 @xcite . therefore , these delays are related to the gravitoelectric field of gr @xcite . the first successful measurements were obtained by shapiro and collaborators by means of radar - echoes from earth to the planets mercury and venus @xcite . successively , anderson and collaborators @xcite repeated the measurement of the delay in the round trip time from the two spacecraft mariner 6 and 7 orbiting the sun . finally , shapiro @xcite and reasenberg @xcite obtained the most accurate results with this technique by means of a transponder placed on the surface of mars by the nasa mission viking . the agreement between the measured delay and its general relativity prediction was around 0.1% . this kind of measurements are quite important because they allow to constrain the ppn parameter @xmath12 , which measures the space curvature per unit of mass . currently , its best measurement has been obtained by the radar tracking of the cassini spacecraft during a superior conjunction with the sun along its cruise to saturn @xcite . bertotti and collaborators obtained @xmath13 . the advantage of the latter measurement relies on the doppler tracking ( not exploited before ) and the multi - frequency link ( both x - band and ka - band ) that allow for the plasma compensation of the solar corona . this delay , which is now known as the shapiro time delay , represents the first gr correction to the time propagation of an electromagnetic signal between an emitter and a receiver with respect to the time of propagation that is needed in the flat spacetime of minkowski . lagrange , with its multi - spacecraft configuration , would allow the measurement of the time delay in the propagation of the electromagnetic signals in several different geometrical configurations . at the same time , it would extend the measurement of the delay not only to the effect previously mentioned , the so - called shapiro time - delay , but also to the delay produced by the gravitomagnetic field @xcite of the sun and/or to that of the earth . for instance , referring to previous fig . [ fig . 1 ] , we can consider the propagation of light and the corresponding delay between the two equilateral lagrangian points @xmath2 and @xmath3 . in this case , the impact parameter @xmath14 , the point of closest approach to the sun , is equal to 0.5 au , i.e. comparable with the other distances , avoiding the problem connected with the plasma of the solar corona , as well as the additional delay produced by the quadrupole moment of the sun . another very interesting geometrical configuration is the one represented by the two collinear points @xmath0 and @xmath1 . the propagation of signals between these two points will allow , for the first time , a direct measurement ( in the field of the earth ) of the overall delay on their propagation , as produced by the combined action of the mass and angular momentum of the earth plus the additional delay due to its oblateness . some of the corresponding measurements with lagrange will allow to improve the current limits in gravitational physics by exploiting the present know - how and accuracy in time of flight measurements and with the present state of art in atomic clocks precision and accuracy . conversely , other effects , in order to emerge from the noise , need an improvement in the current technology of time measurements . the lagrange measurements would be based on the application of null geodesics around a spinning body in the weak field and slow motion limit ( wfsm ) of gr . in terms of metric , the kerr metric will be the reference @xcite , or , to say better , its weak field limit @xcite , with a non - diagonal component @xmath15 proportional to the intrinsic angular momentum ( spin ) @xmath16 of the central body . we consider a quasi - cartesian coordinate system at the post - newtonian level with origin in the central ( deflecting and delaying ) body . we consider the propagation in the @xmath17 plane ( coincident with the plane of the ecliptic ) and assume that the angular momentum @xmath18 of the body is along the _ z - axis _ and that this axis is also the symmetry axis of the body ( i.e we assume cylindrical , or axial , symmetry ) . in particular , we assumed a standard isotropic pn approximation @xcite where the receiver ( or observer ) has to be considered positioned along the positive _ y - axis_. under the above approximations , the line element can be written as : @xmath19 where or @xmath20 represents the time - time component of the metric , while the other terms provide spatial and mixed contributions . ] @xmath21 in the above expressions , @xmath22 represents the speed of light , @xmath23 the ( invariant ) proper time , @xmath24 the newtonian gravitational constant , @xmath16 the angular momentum of the central body , @xmath25 the distance in the reference plane , @xmath26 the kronecker symbol and , finally , @xmath27 represents the gravitational potential @xmath28 where @xmath29 , @xmath30 and @xmath31 are , respectively , the mass , radius and quadrupole moment of the body . this configuration is particularly interesting when the delays in the propagation of the electromagnetic signal are analysed in order to take care also of , besides the contributions from the gravitoelectric and gravitomagnetic fields of gr , the effect in the time propagation produced by the quadrupole moment of the central object . in the case of the propagation of electromagnetic waves we need to impose the condition of null geodesics with the further condition that we restrict to the propagation in the reference plane @xmath17 with @xmath32 constant and @xmath33 . therefore , eq . ( [ ppn - metric ] ) reduces to : @xmath34 we can now solve for the coordinate time element @xmath35 from eq . ( [ ppn - metric - luce ] ) and integrate the final expression from the emitter position at @xmath36 up to the receiver ( or observer ) position at @xmath37 ( @xmath38 and @xmath39 are positive quantities ) . for the propagation time @xmath40 we finally obtain : @xmath41 where smaller contributions to the time delay have been neglected . the first term in eq . ( [ delta_t ] ) accounts for the time propagation in the flat spacetime of special relativity . the second term represents the contribution from the gravitoelectric field of gr in the weak field approximation : it is the shapiro time delay . the third contribution arises from the gravitomagnetic field in the same approximation . the @xmath42 sign accounts for the chirality of this contribution : it is positive for a propagation of the signals in the same sense of rotation of the central mass , it is negative in the case of the opposite sense for the propagation . finally , the last term represents the contribution that arises from the oblateness of the central body . the solution provided in eq . ( [ delta_t ] ) implies that emitter , central body and receiver have the same @xmath43 and @xmath44 coordinates ( with @xmath45 ) and differ only for the @xmath46 coordinate ( which is negative for the emitter , null for the central body and positive for the receiver ) . the result obtained in eq . ( [ delta_t ] ) can be considered as a particular case of two results obtained in previous works @xcite . in fact , our result coincides with that obtained in @xcite when that work is restricted to the lensing effect which can be obtained for light propagating in their symmetry plane ( coincident with our reference plane ) with the transformations @xmath47 and @xmath48 in their expressions and with the further condition @xmath49 or @xmath50 in their final expressions for the delays due to the angular momentum and the quadrupole coefficient ( see in particular their section 2 ) ( not to be confused with the ppn parameter commonly designated by the same symbol ) and @xmath51 represent two of the euler angles that define the orientation of their symmetry plane with respect to the lens plane , while @xmath52 represents the angular position of a generic light ray over the lens plane . ] . conversely our first three terms in eq . ( [ delta_t ] ) coincide with equations ( 55 ) , ( 56 ) and ( 57 ) of @xcite with the transformation @xmath53 for their @xmath38 and the approximation @xmath54 and @xmath55 for their coordinates . by applying the measurements based on the propagation time determined with eq . ( [ delta_t ] ) to the configuration @xmath0earth@xmath1 , it will be possible ( at least in principle ) to obtain a measurement of the earth s quadrupole coefficient in a way independent from the usual space geodesy techniques based on the inter - satellite tracking by means of the two twin grace satellites @xcite and from the precise orbit determination ( pod ) of laser - ranged satellites in orbit at a relatively high altitude , as in the case of the two lageos @xcite . considering that the distance between the two lagrangian points @xmath0 and @xmath1 is @xmath56 m , and assuming an impact parameter @xmath14 of the order of the earth s radius @xmath57 m , for the propagation time of eq . ( [ delta_t ] ) we obtain : @xmath58 where the contribution of each term has been highlighted . if we consider a round trip travel for the propagation time , the smaller contribution of the gravitomagnetic field cancels out when we consider the propagation on the same side of the earth , and the quadrupole effect can be extracted after modelling the shapiro delay and the larger effect of the propagation time in the flat spacetime of minkowski . the knowledge of the oblateness of the earth is particularly important because of its long - term variations in relation to the earth s internal structure and its mass distribution . in fact , phenomena like the melting from glaciers and ice sheets as well as mass changes in the oceans and in the atmosphere are responsible for variations in the rate of the global mass redistribution with a consequent time dependency in the quadrupole coefficient characterized by annual and interannual variations @xcite . this kind of measurement can be initiated by earth , the delaying body , with all the advantages of an earth based laboratory equipped with the best time - measuring apparatus to perform the experiment . in particular , optical clocks and lattice clocks based on @xmath59-atoms have reached outstanding fractional frequency instabilities down at a level of about @xmath60 or less , with @xmath61 the integration time @xcite . for instance , with an integration time of about @xmath62 s it is possibile to reach a precision in the measurement of the quadrupole coefficient of about @xmath63 , comparable with the current best determinations from earth ( with calibrated errors ) using the lageos data , and even better with longer integration times . the lense - thirring effect ( lt ) or inertial frame dragging by a moving massive body is a feeble effect of gr , first considered by thirring @xcite and lense and thirring @xcite in 1918 , while studying the influence of rotating masses ( in particular a rotating hollow massive spherical shell ) on a test particle . lt may also be considered as a manifestation of gravito - magnetism i.e. of that typical component of the gr gravitational interaction resembling the magnetic field of moving charges . so far , lt has been verified experimentally in a limited number of cases . the gravity probe b experiment measured the precession induced by the gravitomagnetic field of the earth on four orbiting gyroscopes , with a @xmath64 accuracy @xcite . a careful analysis of the orbits of the lageos and lageos 2 satellites , monitored by laser ranging , evidenced the lt drag of the nodes of the orbits with a @xmath65 accuracy @xcite . the ongoing lares experiment ( combined with the previous data from the two lageos ) has attained a @xmath66 accuracy @xcite . with a different technology , the ginger experiment is under study and preliminary test of the technology . it is based on the use of a three - dimensional array of ring lasers to be located underground at the national gran sasso laboratories in italy @xcite . ring lasers are extremely sensitive rotation measuring devices . their operating principle is a gr evolution of the old sagnac effect @xcite ; what is measured in practice are frequency and amplitude of a beat between two stationary counter - propagating light beams in the ring . rotations , either of kinematical origin or due to the chirality of the gravitational field ( gravitomagnetic component ) , produce a right / left asymmetry of the propagation along the ring . the aim of ginger is to verify the terrestrial lt within @xmath4 or better . the use of the sun - earth lagrangian frame would allow a measurement of the solar gravitomagnetic field ( solar lt ) , exploiting the sagnac approach but resorting to time of flight measurements rather than to interference phenomena or beat tones . for our purpose we may start from the external line element of a steadily rotating body in a reference frame where the main mass is at rest and the axes do not rotate with respect to the distant sky ( to the quasars ) . as in the previous section , weak field conditions are assumed , but now , for convenience , we use polar coordinates in space . it is : @xmath67 if @xmath29 is the mass of the source , it is @xmath68 with the dimension of a length . similarly , if @xmath16 is the modulus of the angular momentum of the source , it is @xmath69 with the dimension of a squared length . the time variable is @xmath70 , again with the dimension of a length . the index @xmath71 labels the coordinates specific of the non - rotating , asymptotically flat reference frame . in the case of the sun it is : @xmath72 it is convenient to use coordinates apt for a terrestrial ( or co - orbiting with the earth ) observer . in practice we need to combine a rotation of the axes at a rate @xmath73 corresponding to the orbital motion of the earth , together with a boost at the tangential speed of the earth on its orbit @xmath74 @xcite . what holds for the earth , holds for the lagrangian points too . since we are considering free fall the orbital rotation rate is keplerian , so that : @xmath75 here @xmath30 is the radius of the orbit of the earth ( @xmath76 m ) , @xmath77 , and @xmath78 . we may now restrict our attention to the orbital plane , so that it is @xmath79 . in the new reference frame and with the new coordinates ( see the appendix for the details ) the line element is @xmath80d\tau^2 \nonumber \\ & & - \left(1 + 2\frac{m}{r}+4\frac{m^2}{r^2}\right)dr^2 \nonumber\\ & & - \left[1-r\frac{m}{r^2}-\frac{m^2}{r^2}\left(1 - 2\frac{r}{r}\right)\right]r^2d\phi^2 \label{line1 } \\ & & + 2\left[2\frac{j}{rr}-\sqrt{\frac{m}{r}}-\left(1 - 2\frac{r}{r}\right)\left(\frac{m}{r}\right)^{3/2}\right]r d\tau d\phi \nonumber\end{aligned}\ ] ] the approximation has been kept to the lowest order in @xmath81 and with reference to the numerical values holding for the sun . for short we write : @xmath82r\ ] ] the frame is non - inertial and comoving with the laboratory ; the origin remains in the center of the sun . in order to find out the time of flight of an electromagnetic signal along a given path , we may extract the time element from eq . ( [ line1 ] ) remembering that it is @xmath83 . in terms of a generic stationary axially symmetric space - time and referring to general coordinates , we find : @xmath84 in order to insure an evolution towards increasing real times , the @xmath85 sign must be chosen . then we see that the term containing the square root in the right hand side of eq . ( [ ds0 ] ) does not change sign when reversing the sense of motion along a given path , whereas the first term in the numerator does . since we are interested in the asymmetries in the propagation we consider the difference between the right- and left - handed time of flight along the same elementary section of the path ; in this way the square root cancels and the other term doubles . finally we integrate along the whole closed space trajectory and express the result in terms of the proper time of the observer . the total time of flight asymmetry turns out to be @xcite : @xmath86 casting into eq . ( [ tofd ] ) the information extracted from eq . ( [ line1 ] ) and preserving the solar weak field approximation , we get : @xmath87\right\}\allowbreak \oint\left(2\frac{j}{r^2}-\frac{\sqrt{mr}}{r}\right)rd\phi \nonumber \\ & \simeq & -4\oint\frac{j}{r}d\phi \label{tofu}\end{aligned}\ ] ] suppose now that the closed path is a polygon , whose edges are light rays . of course the corresponding null trajectories will be affected by the gravitational lensing due to the mass of the sun . however we know that the angular deviation due to the lensing effect is proportional to @xmath88 , so that its influence in the calculation of ( [ tofu ] ) is negligible . in practice we may assume the space trajectories of electromagnetic signals to be straight ; the typical equation is simple : @xmath89 the closest distance from the straight line to the center of the system is @xmath14 and the azimuth of the closest point is @xmath90 . suppose for example that a signal , propagating in the ecliptic plane , goes from position @xmath91 , with coordinates @xmath92 and @xmath93 to the arrival point @xmath94 with coordinates @xmath95 and @xmath96 . we easily work out the contribution of this stretch to the integral ( [ tofu ] ) : @xmath97 let us apply the above result to a triangular loop having @xmath2 , @xmath1 and @xmath3 at the corners . the coordinates in the plane of the ecliptic , measuring the angles from the sun - earth line , are : @xmath98 it is @xmath99 m , so that @xmath100 . the minimum distance from the @xmath101 ( or @xmath102 ) line and the center of the system is @xmath103 numerically : @xmath104 m. the angular coordinate of the minimum distance point , on one side or the other , is @xmath105 . coming to the @xmath106 line , it is @xmath107 and @xmath108 . summing up , and considering the full triangle , we have : @xmath109 eq . ( [ triangolo ] ) may be approximated to first order in @xmath110 i.e. at the @xmath111 level : @xmath112 finally , casting in numbers and dividing by @xmath22 in order to have the result in seconds , we obtain : @xmath113 the total expected time of flight asymmetry is well within the range of measurability , at least in terrestrial laboratory conditions . the challenge is to measure it in space . besides making use of sun s angular momentum as a source of a lt field for a basic science experiment , i.e. for another precise test of gr , there are other , some even truly practical , reasons for such observations . it is widely believed that the observed differential rotation of the sun triggers a near - surface layer of rotational shear , known as _ tachocline _ , where large - scale dipole magnetic fields are generated by dynamo action , ultimately leading to the 11-year solar cycle of sunspots @xcite . crucial to the possible role of the tachocline in the dynamo are its location and depth . while sun s photosphere can be directly observed and also neutrinos provide some direct information about processes in the core of our star , the tachocline is not directly observable . until now , all available information about this boundary layer between the radiative interior and the differentially rotating outer convective zone , have been collected via helioseismology observations , mainly using the solar and heliospheric orbiter ( soho ) and the solar dynamics observatory ( sdo ) probes @xcite . the estimated location of the shear layer at sun s equator is ( @xmath114 , i.e. beneath the convection zone base , and with a width of @xmath115 . using sun s density profile based on the standard solar model @xcite , the tachocline itself should contribute at the level @xmath116 to the total angular momentum of the sun , i.e. to the source of the lt field . although such precision of the solar lt field determination lies at the very edge of the expected lagrange project sensitivity , a periodic low frequency temporal variation of the lt field strength would open another window for sun interior studies . when addressing the effects of rotating massive bodies ( sun ) on a local space - time geometry in our planetary system , it is rational to consider also possible analogous effects originating from larger structures dynamics , i.e. from our galaxy or even more . the main reason is that fields associated with metric tensor components ( @xmath117 or @xmath15 as used in eq.s ( [ ppn - metric ] ) and ( [ line0 ] ) ) might mimic the effects typically associated to the presence of dark matter ( dm ) , i.e. additional centripetal acceleration ( @xmath118 ) and gravitational lensing ( effective refraction index @xmath119 ) . the gravito - magnetic field potential ( @xmath120 ) and field strength ( @xmath121 ) , rather than metric tensor components , are used for clearer analogy only . the best studied and quantified dm problem is related to the dynamic stability of dwarf and spiral galaxies . in the case of the milky way ( mw ) the distribution of the accounted for luminous mass in stars ( @xmath122 ) , nonluminous interstellar gas and dust ( @xmath123 ) , central black hole ( @xmath124 ) and central bulge ( @xmath125 ) is not compatible with the observed nearly flat rotation curves [ @xmath126 of stars and gas in the disk @xcite . the same property of rotation curves for spiral galaxies has been confirmed for star - free , edge regions , via radio emission observations of neutral hydrogen @xcite . for the mw , the mutual gravitational attraction of stars , central black hole and interstellar dust provide a significant part ( nearly all ) of the required centripetal force at small distances ( up to @xmath127 kpc ) from the center , while flat rotation curves at larger distances ( above @xmath128 kpc ) undeniably point towards some other source of centripetal force @xcite . the typical approach to address this problem is to postulate the presence of a large galactic halo , extending beyond @xmath129 kpc , consisting of massive nonluminous particles with isothermal spherical distribution . in such models , the stabilizing effect of dm is being contemplated through a static gravitational field ( i.e. the @xmath130 component of the metric tensor ) due to the mass of the invisible halo . depending on the peculiar dm model , the total mass of the mw may be entirely dominated by the dark halo and could reach values ranging from @xmath131 @xcite up to @xmath132 @xcite . using a naive but straightforward example , a lt field of @xmath133 s@xmath134 strength , would suffice to provide all necessary centripetal acceleration to account for the motion of the sun around the mw center with a tangential velocity @xmath135 km / s at @xmath136 kpc radius . according to a realistic mw mass distribution model @xcite , an additional force component is needed , to account for @xmath129 km / s of the total @xmath137 km / s orbital velocity . a local value of the lt field @xmath138 s@xmath139 would account for this additional centripetal force . similar values could be deduced for the m31 galaxy @xcite in which the rotational velocity term associated with the dm scenario provides a linearly growing contribution to the rotation curve with a slope of @xmath140 s@xmath134 . this result could also be interpreted as the influence of a homogenous lt field , perpendicular to the plane of the disk of the galaxy , with an identical intensity of @xmath141 s@xmath134 . the hypothesized lt field strengths are weaker than the current experimental possibilities @xcite but well within the lagrange project scope . on the contrary , the galactic lt field strengths , calculated from known baryonic mass - velocity data , are typically much weaker : only @xmath142 s@xmath134 for the mw at the sun distance from the center . taking into account that none of the recent experiments have been able to detect the physical nature of dm ( see overview in @xcite ) and a clear observational evidence of a strong correlation between galactic baryonic content and plateau velocity of the rotation curves ( @xmath143 ) , expressed through the tully - fisher relation ( @xmath144 ) @xcite for more than @xmath145 rotationally supported galaxies of different masses and morphologies @xcite , constitutes a justified reason to consider alternatives to standard dm scenarios and assess the possible presence of local lt fields with strength in the @xmath146 s@xmath134 range . overwhelming share of non - baryonic energy - mass density , in coincidence with observed clustering of visible matter over a larger volume in the observable universe , suggests that other potential sources of lt fields , with strengths within the interval of interest , are viable . in that respect , even a residual primordial lt field , originating from the initial singularity and the subsequent fast evolution processes ( inflation era ) can not be excluded . presumed primordial lt fields would imprint on the cmb spectrum in a similar way as dm and would influence large scale structures evolution into characteristic filaments with congregated clusters of galaxies and large voids in between @xcite . similar filaments and voids structures and other relevant analogies are commonly observed in solid state systems @xcite . filaments of vortex lines of quantum vortices in superfluid helium or magnetic flux bundles in superconducting materials are the best examples . although such simple analogies can not guarantee they would have something in common with large scale structures in the universe , the existence of forces inside and among such filaments , originating from the interaction with the bulk of the medium ( space time voids ) , looks a lot like dm and dark energy ( de ) effects . therefore local lt field search ( measurement ) might open another window into the dm and de problem . the solution adopted for global positioning on earth or in its vicinity is mainly based on the gps method and on the gps , glonass , galileo and other present or future satellites dedicated constellations . without entering into a discussion of the strengths and weaknesses of that approach , it is easily agreed that it can not be extended beyond the near terrestrial environment or , at least , that the application to space navigation is an opportunity to reconsider the whole method , especially regarding the way to account for the effects of special and general relativity . an intrinsically relativistic positioning system ( rps ) has been proposed and is described in @xcite . it is based on the local timing of at least four remote independent sources of electromagnetic pulses ; the essence of the method is graphically presented in fig . [ fig . 2 ] . successive pulses ( but they could also be periodic equal phase hypersurfaces ) cover space - time by a regular four - dimensional lattice . the world - line of an observer intersects the walls of successive cells of the lattice ; the proper time interval measured by the observer between consecutive crossings provides the basic information . counting the pulses ( after identifying the various sources ) and applying a simple linear algorithm it is possible to calculate the coordinates ( including time ) of the receiver in the fiducial reference frame @xcite . s are the null wavevectors of the signals coming from the two ( four in the full space - time ) independent sources . the @xmath147 s label the null wavefronts of the pulses . the wiggling line is the worldline of the receiver . the dots show some of the intersection points ( reception of a signal ) . the relevant quantity is the length ( i.e. proper time span ) between successive reception events.,title="fig:",width=302 ] + the dimensionless coordinates along the light cone of a source are expressed as the sum of an integer part @xmath148 ( the subscript @xmath149 labels the sources ) and of a fractional part @xmath150 . the integer is obtained just counting the successive arrivals of the pulses ; the fractional part is given by a simple linear algorithm applied to sequences of arrival times in the proper reference of the observer @xcite . projecting the @xmath151 light cone coordinates onto the axes of the fiducial reference frame finally produces the practical coordinates we are interested in . of course the sources may be orbiting satellites , but in that case you have to know with the best possible accuracy , the position of each satellite ( i.e. its real orbit ) while time passes . the situation would be far simpler if the position of the emitter were fixed in the fiducial reference frame . this possibility is implemented in nature if the signals come , for instance , from pulsars : their positions in the sky are indeed fixed or slowly moving at a well known rate ; furthermore pulsars are also very good clocks , even better , in the long term , than our atomic clocks . an exercise application of the rps , using pulsars , is presented in @xcite . the inconvenience of pulsars is that their pulses are extremely weak so that large antennas are needed and special techniques must be implemented in order to identify and extract the signal from an overwhelming noise . such troubles can be removed placing artificial `` pulsars '' in points that keep rigidly their positions in an appropriate reference system . that is indeed the case of the lagrangian points . @xmath0 , @xmath1 , @xmath2 and @xmath3 , equipped with emitters of regular pulses would form a very interesting basis for a physical reference frame co - orbiting with the earth . @xmath5 has not been considered because it is located in the opposite side of the sun with respect to the earth , so being invisible from our planet . an important feature of the system is that the distances between the reference points range between @xmath152 million km approximately ( from @xmath0 or @xmath1 to the earth ) to @xmath153 million km ( from the earth to @xmath2 or @xmath3 ) . such large values dramatically reduce the effect of the geometric dilution which renders gps ( and the other terrestrial positioning systems ) useless when extended away from our planet . of course all lagrangian points lie in a plane , but that is usually the case also for most space missions and the size of the base allows for good results even for a couple of astronomic units out of the ecliptic plane . the spacecraft carrying the emitter devices could in general not coincide with the corresponding lagrange point , but would rather orbit around the point on stable ( @xmath2 and @xmath3 ) or on weakly unstable lissajous orbits ( @xmath0 and @xmath1 ) . the final accuracy of the positioning would depend mainly on the accuracy with which the instantaneous position on the orbit is known . the other limiting factor for the final result is the quality of the clock used by the receiver , but there it would be easy to attain centimeter accuracies or even better . besides the proposals already described in the previous sections , it is worth mentioning a number of other opportunities that a lagrangian frame as the one presented in this paper could lend . among the weak gravitational effects of gr a long sought for , and now world - famous , one , are gravitational waves ( gw ) . after the recent detection of a signal from the merger of a pair of black holes @xcite the international effort to further detections has become even more intense than before . besides the terrestrial facilities , great hopes are put on the elisa space interferometers , when they will fly . in the same time , it is worth exploring various approaches to the way gw can be detected and there is one that can employ the same setup and method described for the lt measurement . in the @xmath154s of the past century sun , fejer , gustafson and byer @xcite , starting from a previous idea of weiss , proposed and analyzed a so called zero - area sagnac interferometer for the detection of gw . the zero - area sagnac interferometer organizes the path of light beams so that the contoured area be zero . in practice , one has a pair of twisted half loops and each beam travels along the two loops once clockwise , once counter - clockwise . the consequence is that the signal of the kinematic and effective rotations cancel ; if , however , the strain induced by a gw is present , the transient difference in the metrics along the two arms of the interferometer maintains its effect and is recorded as a residual time depending asymmetry in the phases of the beams . on earth the principle works but the storage time in one arm of the interferometer can hardly be longer than a ms . in the case of our lagrangian system a zero - area configuration could be achieved including earth - based stations and using for instance an @xmath155 circuit . in that case the `` storage time '' ( here the flight time ) in a half loop would be in the order of one thousand seconds . the angle between the `` arms '' ( important in order the strain of the gw to be manifested ) would be @xmath156 rather than @xmath157 , but the permanence time would be @xmath158 times longer than in a terrestrial device . using eq.s ( 2 ) or ( 3 ) of ref . @xcite ( where a continuous harmonic wave is considered ) as a rule of thumb , we see that the length of the travel time could indeed , in the case of @xmath159mhz frequency of the wave , boost the signal by a factor @xmath160 . of course an application like this deserves a consideration much more attentive than this qualitative note , but it would be worth doing once lagrange had been put into operation . once the lagrange stations would be _ in situ _ they would allow a number of additional scientific usages . 1 . provided at least one of the spacecraft be equipped with optical devices , following the apparent position of @xmath2 or @xmath3 with respect to the background stars would give an interesting means to study the gravitational lensing in the solar system . once the `` artificial pulsars '' consisting in the emitters for the lagrange rps were active , they could be used as gauges for studying real pulsars . 3 . spacecraft located in @xmath2 and @xmath3 would be an interesting opportunity to study the space environment there and in particular the already discovered @xcite and maybe not yet discovered trojan asteroids there . 4 . monitoring the movements around the l - points would allow to map the gravitational field in the corresponding regions . the time measurements proposed in the previous sections are quite challenging , from both the point of view of science and technology , especially given the fact that the measurements will not take place in the `` earth laboratory '' . consequently , in order to verify the practical possibilities to carry out the measurements of the time - of - flight between two or more lagrange points , a number of studies devoted to investigate deeply the scientific and technological aspects are necessary . in this section , we start by highlighting the main issues that should be considered and attacked in the future in order to guarantee the success of this proposal . for practical reasons , we will discuss separately at this first level the science issues from the technological ones . of course , these two fields science and technology are not separated , but indeed tightly related to each other . the scientific aspects are mainly related to the gravitational physics of a three - body problem and of its gr interpretation . no exact solution exist for this problem , neither in classical celestial mechanics ( cm ) , nor in gr . in particular , since a spacecraft is massless with respect to both the earth and the sun , and since the orbit of the earth around the sun can be approximated as circular at least in first approximation for our purposes we can report our initial considerations to the restricted three - body problem of cm . our approach will be based on the characteristic tools of perturbations theory for what concerns both gravitational and non gravitational perturbations . the n - body problem of cm will be simplified in the classical three - body problem of the sun - earth system , in the so - called restricted case . the variations with respect to the more general three - body problem and those related to the influences of a fourth - body will then be analysed perturbatively . the gr effects will be considered within the corrections of relativistic celestial mechanics , i.e. considering the gr effects in terms of disturbing potentials and accelerations to be treated perturbatively with the well known lagrange or gauss equations . we refer to the classical textbooks of celestial mechanics for details @xcite . as it was briefly mentioned earlier , the orbits of a spacecraft around the lagrangian points are not all stable , even if the lagrange points are points of equilibrium , i.e. they are points where the gradient of the potential is null . in fact , since we are in a rotating frame , the apparent forces come to play an important role . for instance , the two triangular points @xmath2 and @xmath3 , although they are points of maximum potential , are stable points because of the coriolis force , which acts as an attractive force . conversely , the three collinear points @xmath0 , @xmath1 and @xmath5 are unstable points . therefore , a spacecraft in orbit around one of the triangular points will follow a periodic orbit . conversely , a spacecraft in orbit around one of the collinear points will undergo a lissajous orbit , with the necessity of periodic manoeuvres in order to keep the satellite sufficiently close to the unstable equilibrium point . as a consequence of these very preliminary and quite general considerations we list below the first requirements that from both the scientific and conceptual point of view we need to investigate by means of dedicated studies . 1 . the solar system is of course not made of the sun - earth pair only . an analysis of the gravitational perturbations induced by the other celestial bodies is in order , both for the stability of the positions of the lagrangian points and for the influence on the signal that would be measured . 2 . all effects have to be analysed from the view point of gr in order to build a consistent interpretation frame for the data in view of the expected results . a detailed study of the effects of deviations from gr in the framework of ppn formalism is in order . an analysis is already present in @xcite and the relevant parameters taken into account there have been @xmath161 ( accounting for the presence and relevance of preferred reference frames ) and @xmath12 ( accounting for the effect of spatial curvature alone ) . the possible relevance of such and other parameters must be discussed at the scale of the inner solar system . furthermore , the possible presence of galactic or even relic , gm fields must be considered and modeled in order to provide appropriate templates for the interpretation of the results . 4 . being the positioning at the lagrange points crucial , it is of paramount importance to study the orbits of the spacecraft of the experiment around those points and the related dynamics and stability . libration movements are expected in @xmath2 and @xmath3 , where the orbits can be stable , whereas weakly unstable lissajous orbits are typical of @xmath0 and @xmath1 ( saddle points of the effective potential ) . stabilization strategies must be envisaged . one should also account for the proper motion of the lagrange points with respect to the earth , due to the eccentricity of the orbit of our planet . a model must be worked out accounting for the effects of the movements of the stations , according to point 4 , on the final evaluation of the time of flight difference along closed paths among the lagrange points . semi - analytical investigations of the main periodic orbits around the lagrangian points @xcite provide accurate approximations to model distance variations . they can also be used as seeds for numerical surveys of the unstable dynamics around @xmath0 and @xmath1 . as previously noted , the close links between scientific analysis and the corresponding technological aspects required to achieve the scientific objectives already emerge in the description of the above points . the technological aspects to consider are several and have to be all tackled in order to ensure the realization of the proposed time - of - flight measurements . in particular , the measurement of the time - of - flight difference within a closed path and with the required precision represents , among all the aspects to consider , the most challenging one . as seen in previous section [ subsec : polygon ] , the interval between the arrivals of the clock- and respectively counterclock - wise pulses is in the order of @xmath162 s ( in the best case @xmath163 s ) . in order to attain a 1% accuracy the clock ( and the measurement process ) must have a sensitivity of at least @xmath164 s. this is not a problem in terrestrial laboratories , where one can achieve results one or two orders of magnitude better than that , depending on the integration time . of course , one may consider the possibility to repeat more than once the closed path . this will facilitate the measurement of the time difference , but will impose a more stringent requirement on the clock stability . therefore , a balance is at the end necessary for these two aspects . each spacecraft will have to be equipped with a receiver / transmitter device , or transponder , in order to allow an electromagnetic signal to travel along the closed circuit formed by the lagrangian points . of course , deep space antennas ( dsa ) will play a fundamental role in all phases of the mission , from launch to the cruise phase of each spacecraft , to their orbit insertion in the proximity of a lagrangian point and , finally , during the time measurements in the operative phase of the mission ( see for instance @xcite ) . for what concerns the capability to reach the lagrangian points , this is not a huge problem , especially in the case of @xmath0 and @xmath1 , and also in the case of the most far away @xmath2 and @xmath3 , but reaching @xmath5 and placing in its surroundings a spacecraft would be quite challenging , because of the shielding effect by the sun , and this is the reason why we let it aside from the beginning . the main in - orbit station should be probably located in the spacecraft orbiting @xmath1 . this for the ease of communications with earth and its shielding of the perturbing solar radiation . we list below the first requirements that , from the technological point of view , we need to investigate carefully by means of dedicated studies . 1 . the core measurement of lagrange concerns time . the main problem is to think of atomic clocks in space , onboard an observatory located in one of the lagrange points ( such as @xmath0 or @xmath1 ) , endowed with the required accuracy and a stability sufficient for a long mission ( not less than months ) . a possible solution in order to ensure the best accuracy , and for the longest time , would be to permanently keep ( and for all configurations ) an earth laboratory as hub of our experimental polygonal network . in any case , the need for having and controlling clocks and oscillators in space can not be dispensed of completely , so that the problem must be carefully studied . 2 . the main station must be equipped also with an appropriate elaboration capacity . even though the requirement there is not much stringent : the algorithm to be used , in the case of rps , is linear . more important is the onboard data storage capacity , connected with an effective communication channel for transferring information to the earth . 3 . the technology of the transponders to be used at the other ( with respect to the main one ) stations has to be analysed carefully . each transponder is expected to work as much as possible like a mirror , with a fixed and as short as possible delay between the arrival and the bounce back of the incoming pulses . 4 . in case of the use of lagrange as a base for positioning and navigation in the solar system , each lagrange point must host an emitter of regular pulses , possibly combined with the transponder . each emitter needs to be controlled by an atomic clock ; the main constraints are not on the frequency ( ghz or even mhz repetition times for the pulses are sufficient ) , but rather on the stability over months or even years , if the rps has to be considered as a permanent equipment for navigation . 5 special attention must be paid to the design of all antennas installed onboard . all electronic equipment , at all stations , of course needs an appropriate power supply , whose main components are solar cells and storage capacity . these , and other , engineering aspects will be considered in the near future in parallel with the previous scientific requirements . after the review presented in the previous section , our conclusion can be very short . we have discussed the numerous and relevant possibilities for science , which could be implemented setting up a physical reference frame based on the lagrange points of the sun - earth system . we have concentrated our attention on the information that could be extracted from time of flight measurements of electromagnetic signals exchanged between the stations located at the l - points , christening lagrange the corresponding proposal . among others , we have analyzed in some detail two scientific and one practical usages of lagrange : accurate determination of the gravitational time delay in the propagation of the signals , with the purpose to ameliorate the evaluation of the ppn @xmath12 parameter ; measurement of the solar lense - thirring effect and possibly of a galactic gravito - magnetic field ; establishing of a relativistic positioning and navigation system in the solar system . among the interesting information that could be retrieved from the measurement we have included new constraints on the quadrupole moment , @xmath31 , both of the earth and the sun . the preliminary analysis puts these goals within the range of feasibility , posing , in the same time , important challenges for data analysis techniques and hard technologies to be deployed in space . starting from eq . ( [ line0 ] ) we apply a rotation by an angle @xmath165 , where @xmath73 is given by the first equation of the ( [ kepler ] ) . using the angular coordinate @xmath166 the line element assumes the new form : @xmath167 the next step is to apply a boost at the keplerian speed @xmath74 written in the second of the eq.s ( [ kepler ] ) . both time and the azimuth are affected ; the final coordinates expressed by the boost will obey the relations : @xmath168 for the application of interest here it is @xmath79 and all terms of the line element are developed at the order of @xmath169 with the numerical values of the sun , as cited in the text . the result is indeed eq . ( [ line1 ] ) . k. schwarzschild . ber das gravitationsfeld eines massenpunktes nach der einsteinschen theorie ( on the gravitational field of a mass point according to einstein s theory ) , sitzungsberichte der kniglich preuischen akademie der wissenschaften ( berlin ) , 1916 , seite 189 - 196 k. s. thorne , gravitomagnetism , jets in quasars , and the stanford gyroscope experiment . near zero : new frontiers of physics ; fairbank , j.d . ; deaver , jr . , b.s . ; everitt , c.w.f . ; michelson , p.f . , eds . , 1988 ,

this paper is a continuation in a series of papers ( papers i vi and viii ix ) of radial - velocity studies of close binary stars @xcite and presents data for the ninth group of ten close binary stars observed at the david dunlap observatory . for technical details and conventions , and for preliminary estimates of uncertainties , see the interim summary paper ( * ? ? ? * paper vii ) . selection of the targets is quasi - random : at a given time , we observe a few dozen close binary systems with periods shorter than one day , brighter than 10 11 magnitude and with declinations @xmath2 ; we publish the results in groups of ten systems as soon as reasonable orbital elements are obtained from measurements evenly distributed in orbital phases . whenever possible , we estimate spectral types of the program stars using our classification spectra . these are compared with the mean @xmath3 color indexes taken from the _ tycho-2 _ catalog @xcite and the photometric estimates of the spectral types using the relations published by @xcite . the observations reported in this paper have been collected between june 1997 and november 2004 . the ranges of dates for individual systems can be found in table [ tab1 ] . all systems discussed in this paper , except three : hs aqr , sw lac , and ks peg , have been observed by us for radial - velocity variations for the first time . we have derived the radial velocities in the same way as described in previous papers . see paper vii for a discussion of the broadening - function approach used in the derivation of the radial - velocity orbit parameters : the amplitudes , @xmath4 , the center - of - mass velocity , @xmath5 , and the time - of - primary - eclipse epoch , @xmath6 . the novelty of this paper is the treatment of binaries with very weak signatures of binarity in their broadening functions , resulting either from exceptionally weak mg i @xmath7 triplet lines ( v395 and , section [ v395and ] ) or from the presence of a bright third star in the system ( tu umi , section [ tuumi ] ) which limits the dynamic range in the broadening function . we utilize here an approach borrowed from image processing techniques in that we represent the broadening function information in the phase domain as a two - dimensional image where smoothing in the phase coordinate restores the temporal correlation between successive broadening functions . this paper is structured in a way similar to that of previous papers , in that most of the data for the observed binaries are in two tables consisting of the radial - velocity measurements in table [ tab1 ] and their preliminary sine - curve solutions in table [ tab2 ] . table [ tab1 ] contains data for nine stars as we have not been able to derive individual radial velocities for the difficult case of tu umi . the data in table [ tab2 ] are organized in the same manner as in previous papers . in addition to the parameters of spectroscopic orbits , the table provides information about the relation between the spectroscopically observed epoch of the primary - eclipse t@xmath8 and the recent photometric determinations in the form of the @xmath9 deviations for the number of elapsed periods @xmath10 . it also contains our new spectral classifications of the program objects . section [ sec2 ] of the paper contains brief summaries of previous studies for individual systems and comments on the new data . figure [ fig1 ] and figure [ fig4 ] show the 2-d broadening function images for v395 and and tu umi while figure [ fig2 ] and figure [ fig3 ] show the radial velocity data and solutions . figure [ fig5 ] shows the bf s for all systems ; the functions have been selected from among the best defined ones , usually around the orbital phase of 0.25 using the photometric system of phases counted from the deeper eclipse . photometric variability of the star was discovered by the hipparcos satellite @xcite ; the star was assigned rrc type in the 74th special name - list of variable stars @xcite on the basis of the hipparcos data . @xcite suggested that v395 and is a contact eclipsing system on the basis of the period - color relation . light elements were calculated by @xcite , hjd=2448500.4372 + e 0.684656 using the hipparcos photometric data and the doubled value of the original period . v395 and was listed in the @xcite with the spectral type of this object estimated as b8 ( this type is also quoted in the washington double star catalog , from now on called wds ) . and radial velocity of @xmath11 km s@xmath12 ; in the hd catalog the spectral type is a0 . our spectra of the magnesium triplet in v395 and showed an exceptionally weak , poorly defined , strongly broadened feature with depth of only 2 percent of continuum with no discernible doubling . while we expected that the feature will be weak in a nominally a0 type star ( simbad ) , the particular weakness of the line was surprising . the star is relatively bright , @xmath13 . it is a member of a visual binary ( wds 23445 + 4623 ) . the companion at a separation of 5.7 arcsec is 2.2 mag fainter than v395 and . it is visible at the position angle @xmath14 and thus is projected into our east west spectrograph slit . its presence may have produced weakening of the spectral lines by probably less than 15% . original estimates based on the very shallow and poorly defined broadening functions ( bf s ) suggested that the star will be entirely intractable and may have to be abandoned . our standard technique of measuring radial velocities requires that each bf , for each spectrum , gives a peak or a pair of peaks for radial velocity centroid determination ; such well defined peaks simply did not exist for v395 and . however , a new approach , partly based on ideas used in image processing worked relatively well : the bf s were ordered in as columns in a two - dimensional image , with y - coordinate being the heliocentric radial velocity . the x - coordinate , the phase , is obviously unevenly filled and the data can be interpolated into a regular grid of phases . at this point , we utilize the fact that the bf s are correlated in the temporal ( phase ) domain so we can use smoothing . thus , horizontal cuts through the bf - image were convolved with a gaussian kernel with the full width at half maximum ( fwhm ) of 0.025 in phase , and then re - binned into a new uniform phase system with 40 phase points . the `` broadening function image '' is shown in figure [ fig1 ] while a section at phase 0.75 is shown in figure [ fig5 ] , among the broadening functions for other stars of this group . even after processing of all available 103 spectra of v395 and , as described above , only some of the smoothed bf s were measurable for individual radial velocities . they are listed in table [ tab1 ] as for other stars , but they should be used with considerable caution . the smoothed broadening functions were originally found at 0.025 intervals in phase and heliocentric times were calculated for them using the @xmath6 prediction given above . some of the smoothed bf s were still too poor and have been eliminated from the listing . the final solution is based on the data as listed in table [ tab1 ] and required a new value of the initial epoch @xmath6 , as given in table [ tab2 ] . the star appears to be a very close pair with almost identical components , @xmath15 . the component potentially eclipsed at the primary minimum ( corresponding to @xmath6 ) is less massive ; its peak in the broadening function is slightly sharper and taller indicating higher surface temperature and lower rotation rate . the very small amplitude of light variations , @xmath16 , suggests that the eclipses will not be actually observed and that the binary is an `` ellipsoidal variable '' showing only proximity effects . our classification spectra clearly indicated a spectral type as early as b7 which is difficult to reconcile with the close orbit with the period of 0.685 days . the two spectra that we obtained were not identical and showed some peculiar variability and emission at he i @xmath17 . no spectral estimate of @xmath18 was possible due to peculiarity of the spectra . stars of the spectral type of b7 or b8 have masses around 2.8 to 3.0 @xmath19 and radii around 2.2 to 2.4 @xmath20 . if the orbit has @xmath21 then the current data would imply @xmath22 which is just about possible for a contact system . hs aqr is a detached or semi - detached binary . spectroscopic studies based on the high quality echelle spectra were presented by @xcite . this was followed by an extensive discussion of the photometric and spectroscopic data by @xcite . the above sources provide the most extensive references to properties of the system . the value of @xmath6 was taken from @xcite : @xmath23 . our observations of hs aqr give a good solution of the radial velocity orbit ( figure [ fig2 ] ) , but perhaps not of such a high quality as that of popper which was based on the echelle spectrograph spectra . however , since the main uncertainties in this field are not in random errors , but in systematic differences in measuring techniques , we give our solution here for future reference . the broadening function at phase close to 0.25 ( figure [ fig5 ] ) shows two well defined peaks for detached components , with that for the secondary components a bit broader indicating that this component is larger , if this star is in rotation - revolution synchronism . this agrees with the supposition @xcite that the secondary component is close to or fills its roche lobe . our new spectral type estimate of f6v is slightly earlier than previously discussed ( f8v and g8 - 9 ) , in a better agreement with the tycho-2 @xcite color index @xmath24 . the binary is in a visual system ( ccdm j20409 - 0036ab , ads 14147ab ) , but the companion at 1.6 arcsec separation is some 4 magnitudes fainter than hs aqr thus of no consequence for the radial velocity orbit . the recent minimum prediction by @xcite perfectly agrees with our determination of the initial epoch @xmath6 . the star is bright , @xmath25 , but was not included in the hipparcos database , so its parallax is currently unknown . v449 aur was discovered photometrically by the hipparcos mission . the primary eclipse prediction was the hip one : @xmath26 . the mean @xmath27 data of @xcite , @xmath28 and @xmath29 , suggest an early - type no later than a5/6v , while our spectral classification is a2iv and , through a comparison with the color index , implies some reddening . the binary type of the system is not clear at this moment : the broadening functions are partially blended , as in a contact binary , and somewhat difficult to measure for centroid determinations ( see figure [ fig5 ] ) . on the basis of its very shallow hipparcos light curve , but with rather well - defined eclipses , the binary was considered by @xcite to be a detached one ( ea ) ; the broadening functions would rather suggest a contact binary . however , the large widths of the peaks in the broadening functions and the resulting blending may result from rotational velocities higher than the synchronous ones . the mass ratio is not far from unity , @xmath30 , which is frequently encountered among early - type very close binaries . the scatter of the radial velocity observations of the slightly more massive component , which is eclipsed in the primary minimum , appears to be genuinely larger at phases around 0.2 0.4 ( figure [ fig2 ] ) ; we have no explanation for it . fp boo is another case of the hipparcos photometric discovery . the hipparcos period and @xmath6 serve the current observations relatively well : @xmath31 . the previous spectral type given by simbad and in the hipparcos database , a5 , does not agree with our estimate of f0v ; the latter would imply @xmath32 whereas the tycho-2 mean color index is even redder , @xmath33 , corresponding to f2v . the hipparcos photometry was analyzed by @xcite using a simplified approach . the mass ratio was estimated at @xmath34 , which is in a surprisingly good agreement with the spectroscopic result , @xmath35 , taking into account that the light curve is shallow and does not show any obvious total eclipses . ( this case is in our view an exception , a fortuitous coincidence . normally , we do not trust the so - called photometric mass ratios because so many stars in our rv program show @xmath36 ) . fp boo was somewhat faint ( @xmath37 ) for hipparcos photometry , so the light curve could be easily improved . since the velocity amplitudes are large and thus imply the orbital inclination close to 90 degrees , it is likely that more precise photometry will show total eclipses . in its small mass ratio and the spectral type of f0/2v , the system is somewhat similar to the well - known contact binary aw uma . the radial velocity orbit is shown in figure [ fig2 ] while a broadening function close to phase 0.25 is shown in figure [ fig5 ] . sw lac is one of the most frequently photometrically observed contact binaries . it is a bright ( @xmath38 ) binary with a large amplitude and an unstable light curve and thus an easy target for investigations utilizing small telescopes . several discussions addressed the matter of evolving surface spots , the most recent being by @xcite where references to previous work can be found . sw lac is a late type contact binary with the spectral type variously assigned as g3 , g5 and k0 . our new classification is g5v which predicts a color index a bit less red than the observed @xmath39 ; possibly g8v would be a good compromise . from the moment of minima predictions as well as the direct detection of faint spectral signatures @xcite , the binary appears to have at least one probably two nearby , low mass companions @xcite . the system is known to change the orbital period . we used the moment of primary minimum of @xcite : @xmath40 which ideally predicts our @xmath6 . our observations form the third currently available set of radial velocity data ( see figure [ fig2 ] ) . @xcite published the first radial velocity orbit , with @xmath41 . a similar , but not identical result was obtained soon after by @xcite who published only the mass ratio , @xmath42 , without more details . in spite of the large brightness of the star and excellent definition of the individual peaks in the broadening functions , our mass ratio , @xmath43 , carries a larger error than for most of our targets , probably because of the genuine changes in the broadening functions , which are caused by elevated stellar surface activity what is important , however , that radial velocity semi - amplitudes in our orbit ( figure [ fig2 ] ) are larger than previously observed : our total mass estimate , @xmath44 ( the primary 2.18 @xmath19 , the secondary 0.92 @xmath19 ) , is significantly larger than for the orbit of @xcite , @xmath45 . since most of systematic effects result in a reduction of the radial velocity amplitudes , our result is more trustworthy as it is based on the superior technique of the broadening functions . thus , sw lac is a relatively massive , but surprisingly cool system , a situation which was very clearly seen before in the case of another cool contact binary ah vir @xcite . as in the case of ah vir , the masses are characteristic for f - type stars while the combined luminosity is in agreement with small star sizes implied by the roche geometry of a very short period system and their low temperature ( g5v ) . indeed , for sw lac , the @xmath46 calibration @xcite predicts @xmath47 as compared with the parallax - based , observed @xmath48 , so that the geometry and photometric properties are those of a late - type , compact contact system , only the masses are too large . we have not noticed in individual broadening functions any traces of the third component detected by @xcite and apparently producing part of the time - of - minima perturbations @xcite . however , the same spectra , when averaged in the heliocentric system , clearly show a late type dwarf contributing about 1 percent to the total light at our 5184 spectral window ( caroline dangelo and marten van kerkwijk , 2005 , private communication ) . the bright star ks peg ( @xmath49 , @xmath50 ) , also called 75 peg , has been known as a radial velocity variable for a long time , since the work of @xcite . @xcite recognized the peculiar properties of this binary star , in particular the very small mass ratio of @xmath51 making the velocity amplitude of the primary component small and the secondary component invisible in their spectra . the photometric variability was observed by @xcite while a good quality light curve was provided by the hipparcos @xcite project . the light curve with amplitude of only 0.1 magnitude shows very wide minima with unusual , sharply peaked , brief light maxima . we observed ks peg in our standard way , and because the star is so bright with a particular attention to detection of the secondary component . in our series of observations , we have already detected several binaries with @xmath52 , with some even below this limit , such as sx crv at @xmath53 @xcite , so we hoped to see a signature of the secondary . the spectral type of ks peg is a1v ; we used the star of a2v as a main template , but we also used another f8v radial velocity standard to possibly capture the signature of the secondary component of later spectral type . none of these led to detection of any traces of the secondary , even when applying the technique of the `` bf - image '' smoothing described above for v395 and ( section [ v395and ] ) which permits detection of weak features correlating in the phase domain . since the spectra of such a bright star result in particularly good determinations of the broadening functions , the non - detection of the secondary is puzzling and indicates that the secondary may be a very low mass star or even a massive planet , just able to enforce rapid rotation on the visible component , but too faint to contribute light to the overall luminosity of the binary . one of the panels of figure [ fig5 ] shows comparison of broadening functions at both orbital elongations ( averages of 6 bf s at each elongation ) . rapid rotation of the primary component and its ( small ) orbital displacements are very clearly visible , but there is no sign of the secondary component . note that systematic trends in the baselines are due to the rectification process with inevitably different tracing of the continuum for the program and the standard template spectra . these trends may mask a very weak secondary component . our initial reference epoch was that of the hipparcos photometry , @xmath54 , which predicts our @xmath6 very well . in fact , with the same period as in the table , even the epoch of the @xcite is well predicted more than 16 thousand epochs earlier ( the authors in fact suggested 0.5021035 , which does not tie all observations that well ) . the remarkable constancy of the orbital period may be one of useful hints in interpretation of ks peg . ks peg is an important object because it is a bright star and thus belongs to the complete , magnitude - limited hipparcos sample of short - period binaries brighter than 7.5 magnitude @xcite . its parallax is moderately large , @xmath55 milli - arcsec , and the star is intrinsically bright , @xmath56 . the absolute magnitude agrees very well with the expected value for a main sequence a1 star , @xmath57 . the large projected rotational velocity of the primary component , @xmath58 km s@xmath12 , implies that the orbital inclination can not be far from @xmath59 degrees . assuming @xmath60 , @xmath61 , one can achieve a consistent picture , in terms of reproduction of @xmath62 and @xmath63 , with a relatively small secondary star with @xmath64 and thus @xmath65 , in a binary with the orbital inclination @xmath66 . iw per is somewhat similar to ks peg described above : it is also bright , @xmath67 , small amplitude , 0.04 mag . , photometric variable . the variability was recognized by @xcite , but most of the research concentrated on on the chemical peculiarities and am characteristics @xcite . iw per was not a lucky star in our observations . a new ccd system which was used for it failed after some time and the data could not be related to those taken before or after as our technique requires high consistency in observations of program and template stars . later on , it was realized , that the heavy dewar of the ccd system induced flexure in the spectrograph of up to 1/2 of the pixel in the detector focal plane . we decided to analyze and publish the available data for iw per when it was realized that it is a single - line spectroscopic binary , precluding any in - depth studies of this system . in the final radial velocity solution , the spectra taken at large hour angles have been given half weight to account for the spectrograph flexure . the period and the initial epoch for our observations have been taken from the hipparcos database . @xmath68 . in contrast to ks peg , iw per is not a very low amplitude radial velocity variable : the half amplitude @xmath69 km s@xmath12 indicates that the mass ratio is not very small , probably around @xmath70 and that the spectroscopically invisible component is a star with a mass only slightly smaller than solar . the visible component rotates with @xmath71 km s@xmath12 which is consistent with synchronous rotation in a moderately wide binary with the orbital period of 0.917 days . iw per has the best determined parallax among stars of th current group , @xmath72 mas . the implied luminosity , @xmath73 , the spectral type a3v and @xmath74 are mutually consistent . v592 per is another photometric discovery of hipparcos . it is a relatively bright , @xmath75 , basically unstudied contact binary . the moment of primary eclipse from hipparcos , @xmath76 serves well for prediction of @xmath6 . the star has been known to possess a visual companion at a separation of only 0.16 arcsec @xcite . data for wds 04445 + 3953 indicate that the position angle changes rather rapidly so the system is definitely a physical one . we re - discovered the companion spectroscopically : in the broadening functions , its signature is strong , because of its sharpness , but the companion is fainter than the binary with @xmath77 , or @xmath78 mag ; in the wds catalog @xmath79 mag . to determine the close binary orbit , we analyzed our observations the way as before for triple systems , by first fitting three gaussians to the whole bf , then subtracting the third star gaussian and then measuring the radial velocities of both binary components from the residual broadening functions . in spite of the presence of the visual companion , the contact binary orbit is well defined , with @xmath80 the spectroscopic result is in total disagreement with the mass ratio estimated from the light curve of @xmath81 @xcite . once again , we stress that due to the very low information content of light curves light curve solutions for partially eclipsing contact binaries have little sense and they should not be attempted without supporting spectroscopic data . our observations are spread over time in that there is a gap of two full seasons . the radial velocities of the third component were estimated from all available spectra giving the average value of @xmath82 km s@xmath12 for the first season ( @xmath83 ) and @xmath84 km s@xmath12 for the second season ( @xmath85 ) . our systematic uncertainties are at the level of @xmath86 km s@xmath12 , while measurements in the presence of the variable binary `` background '' difficult , so we can not claim that we see seasonal variation in @xmath87 ; it is basically identical to @xmath5 of the contact binary . in view of similar luminosities , the spectral type f2iv estimated by @xcite most probably applies to both stars , the contact binary v592 per and the third star in the system . this spectral type does not agree with @xmath88 estimated from the tycho-2 data @xcite which suggests f5f6 . tu umi is one of the hipparcos photometric discoveries . a description of new photometric observations conducted concurrently with the spectroscopic observations was given by @xcite ; the paper contains also references to work related to identification of the variability and of the determination of the period at twice the hipparcos original period . the previous photometric paper @xcite was based on a short time interval and could not provide a new value of the period which was estimated to only 4 significant figures . however , the hipparcos period , which we use here , together with the new photometric epoch , jd(min ) = 2,452,725,6262 @xcite , counted back to the hipparcos epoch ( additionally shifted by 1/4 of the period because the star was thought to be a pulsating one so the time of its maximum light was used ) , jd(min ) = 2,448,499.9747 , indicate a good stability of the period . tu umi is a case of a contact binary with a bright close companion . we were not aware that it had been known as a visual binary wds 14557 + 7618 with separation of 0.2 arcsec ( this is well below the 1.88 m telescope resolution in our location ) when we started the observations . the spectral lines of the close binary are very broad and , in fact , entirely invisible in individual spectra , being masked by the strong sharp spectrum of the companion . with 70 individual spectra rather evenly covering orbital phases , we attempted to analyze the data using the approach similar to that described for v395 and ( section [ v395and ] ) by ordering in phase and smoothing the broadening functions . this case is much more difficult than that of v395 and because , even after averaging , the binary features are too weak and too poorly defined for being measured ; most of the dynamic range in the bf is `` wasted '' on the sharp peak of the tertiary component ( figure [ fig5 ] ) . we attempted to remove this peak from the broadening function . first , we followed our normal routines of approximating the third star feature by a gaussian and subtracting it . this did not work as the profile is not gaussian at its base where its shape is particularly important . then we attempted to find the shape of the third - star peak by combining all broadening functions and determining their lower envelope . this still did not provide a correct shape of the tertiary peak at the base , since none of the bf s was really free of the binary signature . thus , we were unable to remove the tertiary and to determine individual radial velocities of components for tu umi ; this star does not have any data listed in table [ tab1 ] . while this is disappointing , we feel that it is doubtful if this star will ever permit a detailed parameter determination . in this situation , we present only very rough estimates from what can be derived from just eyeballed fit to the broadening functions in the 2-dimensional representation ( figure [ fig4 ] ) . in plotting the expected changes in the broadening functions as a function of phase , we assumed the initial epoch as observed simultaneously photometrically by @xcite . the star appears to be a w - type contact system with a rather large amplitude of the secondary star variation , perhaps as large as @xmath89 km s@xmath12 . an estimate for the primary is @xmath90 km s@xmath12 . these amplitudes are shown on the bf image in figure [ fig4 ] . integrations of the wide binary feature at several phases indicate that the binary is in fact only marginally fainter than the third component , @xmath91 , but the sharpness of its peak dominates in the combined spectrum of the system . the above estimate of the relative luminosity may carry a large systematic error since it is strongly dependent on how the baseline of the broadening function is drawn ; in fact , this luminosity ratio implies @xmath92 , which poorly agrees with the direct magnitude difference given in the wds catalog ( @xmath93 ) . the third component was observed to have the mean velocity of @xmath94 km s@xmath12 ( a single observation @xmath95 km s@xmath12 ) . fo vir is a totally eclipsing , detached or semi - detached binary , with the larger component close to filling the roche lobe , thus possibly an object in the pre - contact state . an extensive discussion of its properties , following the photometric study of @xcite who established its binary character , was presented by @xcite . on the basis of photometry and radial velocities of the dominant a7v component , they were able to infer many parameters of the system , in particular they predicted the mass ratio , @xmath96 , in a semi - detached system with the more massive component filling or almost filling its roche lobe . similar results were obtained by @xcite . we have been able to detect the secondary component of fo vir , but with some difficulty , because it is relatively small and faint at the observed mass ratio @xmath97 and much later spectral type than that of the a7v primary , probably early k. a somewhat indirect detection of the secondary was reported by @xcite who observed the lyman - alpha @xmath98 emission moving in anti - phase with the primary star and thus most likely related to the secondary ; their @xmath99 is in agreement with the current determination . we experienced difficulties with measuring the radial velocities of both components in the first half of the orbit ; this could be related to the previously reported strong and variable oconnell effect @xcite . in addition , the case of fo vir confronted us with what may be a limitation of our broadening functions approach : we observed relatively strong ( similar to those of the secondary component ) , initially unexplained broadening - function features moving in phase with the primary . we suspect that these `` ghosts '' may appear because of the low metallicity of the primary component @xcite and a mismatch between the spectra of program and template stars ; we simply could not locate a proper , low metallicity template among bright standard stars . many tests and experiments with different templates give us assurance that the final orbital elements are secure ones , although certain element of uncertainty still remains . the radial velocity data for fo vir may have to be re - interpreted when a better @xmath6 prediction for the duration of our observations ( 1997 2001 ) becomes available . with observations spanning four seasons , we could establish that the period must be shorter than that used by the hipparcos project ( 0.775569 ) and we used one which is only by one digit in the last place shorter than that ; however , the period may be in fact systematically changing so that a full discussion is clearly in order , especially in view of a possibly sporadic mass transfer in the system . our @xmath6 is shifted by 50 minutes from the time predicted by the hipparcos project . a re - discussion of rather few available moments of minima for 1983 1991 gave @xmath100 , which we took as a guidance for adopting a shorter period , but which nevertheless resulted in the spectroscopic @xmath6 for our program too early by 40 minutes . fo vir is a member of a visual binary wds 13298 + 0106 , but the companion is some 6.5 fainter so that its presence was of no consequence for the spectroscopic observations . this paper presents radial - velocity data and orbital solutions for the ninth group of ten close binary systems observed at the david dunlap observatory . this group contains very interesting objects : ks peg and iw per are single - line ( sb1 ) binaries without any traces of companions ; the former may have a very small - mass secondary ( @xmath51 ) . v395 and , hs aqr and v449 aur are very close , early - type binaries ; the matter of them being detached systems requires further work ; v395 and with its spectral type of b7/8 is particularly interesting because of its orbital period of only 0.685 days . fp boo , sw lac , v592 per and tu umi are contact binaries . two of them , v592 per and tu umi , have bright tertiary companions , while fp boo is a very small mass - ratio ( @xmath101 ) contact binary . the frequently observed system sw lac is more massive than estimated before ; its broadening functions show the large surface brightness and rather compact dimensions of its secondary component , both possibly indicating that it is really a semi - detached binary and not a contact system . finally , we have been able to detect the low - mass secondary in fo vir ( @xmath102 ) ; the system is most probably a semi - detached binary before establishing contact . support from the natural sciences and engineering council of canada to smr and swm and from the polish science committee ( kbn grants po3d 006 22 and p03d 003 24 ) to wo and rp is acknowledged with gratitude . the research made use of the simbad database , operated at the cds , strasbourg , france and accessible through the canadian astronomy data centre , which is operated by the herzberg institute of astrophysics , national research council of canada . this research made also use of the washington double star ( wds ) catalog maintained at the u.s . naval observatory . abt , h. a. & morrell , n. i. 1995 , , 99 , 135 adelman , s. j. 1998 , , 132 , 93 albayrak , b. , djurasevic , g. , erkapic , s. & tanriverdi , t. 2004 , , 420 , 1039 bessell , m. s. 1979 , , 91 , 589 clausen , j. v. , helt , b. e. & olsen , e. h. 2001 , , 374 , 980 duerbeck , h. w. 1997 , inf . bull . stars , 4513 duflot , m. , figon , p. & meyssonnier , n. 1995 ; , 114 , 269 grenier s. , baylac m.o . , rolland l. , burnage r. , arenou f. , briot d. , delmas f. , duflot m. , genty v. , gomez a.e . , halbwachs j .- l . , marouard m. , oblak e. , sellier a. 1999 , , 137 , 451 european space agency . 1997 . the hipparcos and tycho catalogues ( esa sp-1200)(noordwijk : esa ) ( hip ) heintz , w. d. 1990 , , 74 , 275 hendry p. d. & mochnacki , s. w. 1998 , , 504 , 978 hg , e. , fabricius , c. , makarov , v. v. , urban , s. , corbin , t. , wycoff , g. , bastian , u. , + schwekendiek , p. , & wicenec , a. 2000 , , 355l , 27 hrivnak , b. j. 1992 , bull . aas , 24 , 686 hube , d. p. & gulliver , a. f. 1985 , , 97 , 280 hube , d. p. , martin , b. e. & gulliver , a. f. 1988 , inf . bull . stars , 3151 jordi , c. , figueras , f. , torra , j. & asiain , r. 1996 , , 115 , 401 kazarovets , a. v. , samus , n. n. , durlevich , o. v. , frolov , m. s. , antipin , s. v. , kireeva , n. n. , pastukhova , e. n. 1999 , inf . bull . stars , 4659 kreiner , j. m. 2004 , acta astr . , 54 , 207 lu , w.x . , & rucinski , s. m. 1999 , , 118 , 515 lu , w. , & rucinski , s. m. 1993 , , 106 , 361 ( paper i ) lu , w. , rucinski , s. m. , & ogloza , w. 2001 , , 122 , 402 ( paper iv ) mason , b. d. , wycoff , g. l. , hartkopf , w. i. , douglass , g. g. , & worley , c. e. 2001 , , 122 , 3466 ( wds ) mochnacki , s. w. , fernie , j. d. , lyons , r. , schmidt , f. h. & gray , r. o. 1986 , , 91 , 1221 morris , s. l. 1985 , , 295 , 143 pauzen , e. & maitzen , h. m. 1998 , , 133 , 1 plaskett , j. s. , harper , w. e. , young , r. k. , plaskett , h. h. 1920 , publ . dominion astroph . obs . , 1 , 163 popper , d.m . 1996 , , 106,133 popper , d.m . 1997 , , 114 , 1195 popper , d.m . 2000 , , 119 , 2391 poretti , e. , niarchos , p. g. , mantegazza , l. , antonello , e. & conconi , p. 1987 , , 69 , 337 pribulla , t. , chochol d. , parimucha , . 1999 , contr . skalnat pleso , 29 , 111 pribulla , t. , kreiner , j. m. & tremko , j. 2003 , contr . skalnat pleso , 33 , 38 pych , w. & rucinski , s. m. 2004 , inf . stars , 5524 pych , w. , rucinski , s. m. , debond , h. , thomson , j. r. , capobianco , d. c. , blake , r. m. , ogloza , w. , stachowski , g. , rogoziecki , p. , ligeza , p. , gazeas , k. 2004 , , 127 , 1712 ( paper ix ) rucinski , s. m. 2002a , , 124 , 1746 ( paper vii ) rucinski , s. m 2002b , , 114 , 1124 rucinski , s. m. , & duerbeck , h.w . 1997 , , 109 , 1340 rucinski , s. m. , & lu , w. 1999 , , 118 , 2451 ( paper ii ) rucinski , s. m. , lu , w. , & mochnacki , s. w. 2000 , , 120 , 1133 ( paper iii ) rucinski , s. m. , lu , w. , mochnacki , s. w. , ogloza , w. , & stachowski , g. 2001 , , 122 , 1974 ( paper v ) rucinski , s. m. , lu , w. , capobianco , c. c. , mochnacki , s. w. , blake , r. m. , thomson , j. r. , ogloza , w. , & stachowski , g. 2002 , , 124 , 1738 ( paper vi ) rucinski , s. m. , capobianco , c. c. , lu , w. , debond , h. , thomson , j. r. , mochnacki , s. w. , blake , r. m. , ogloza , w. , stachowski , g. , & rogoziecki , p. 2003 , , 125 , 3258 ( paper viii ) schmidt , f. h. & fernie , j. d. 1984 , inf . stars , 2527 selam , s. o. 2004 , , 416 , 1097 shaw , j. s. , guinan , e. f. & ivester , a. h. 1991 , bull.aas , 23 , 1415 zhai , d. & lu , w. 1989 , chinese astr . astroph . , 13 , 350 2452507.6792 & 122.09 & 1.00 & -143.05 & 1.00 + 2452507.7194 & 91.43 & 1.00 & -104.07 & 1.00 + 2452507.7372 & 73.05 & 1.00 & -99.45 & 0.50 + 2452507.7543 & 60.43 & 0.50 & 0.00 & 0.00 + 2452507.7744 & 38.30 & 0.50 & 0.00 & 0.00 + 2452507.7875 & 0.00 & 0.00 & 0.00 & 0.00 + 2452509.6398 & 84.74 & 1.00 & -103.00 & 0.50 + 2452509.6551 & 95.57 & 1.00 & -116.69 & 1.00 + 2452509.6704 & 108.27 & 1.00 & -120.46 & 1.00 + 2452509.6863 & 115.38 & 1.00 & -133.96 & 1.00 + v395 and & ea & hd 222900 & @xmath103(1.51 ) & 133.35(2.55 ) & 14.20 & 53,278.0000(30 ) & 0.684656 & 0.879(25 ) + & b7/8 : & hip 117111 & & 151.72(2.06 ) & 9.09 & @xmath104 [ 6978 ] & 1.647(80 ) & + hs aqr & ea & hd 197010 & @xmath105(0.64 ) & 111.57(0.56 ) & 3.36 & 52,507.7981(12 ) & 0.710188 & 0.626(7 ) + & f6v & bd@xmath106 & & 178.24(1.46 ) & 9.94 & @xmath107 [ 10 ] & 1.795(38 ) & + v449 aur & ew(a ) & hd 41578 & @xmath108(1.22 ) & 128.27(1.45 ) & 9.72 & 52,307.0853(24 ) & 0.703648 & 0.730(15 ) + & a2/3v & hip 29108 & & 175.61(2.15 ) & 16.76 & @xmath109 [ 5410 ] & 2.051(73 ) & + fp boo & ew(a ) & bd@xmath110 & @xmath111(1.02 ) & 26.88(0.98 ) & 8.08 & 52,388.2227(36 ) & 0.6404870 & 0.106(5 ) + & f0v & hip 76970 & & 254.04(2.80 ) & 14.28 & @xmath112 [ 6070 ] & 1.475(60 ) & + sw lac & ew(w ) & hd 216598 & @xmath113(1.17 ) & 173.91(1.78 ) & 10.11 & 52,484.1080(8 ) & 0.3207158 & 0.776(12 ) + & g5v & hip 113052 & & 224.14(1.70 ) & 11.54 & @xmath114 [ -50 ] & 2.101(55 ) & + ks peg & sb1 ( eb ? ) & hd 222133 & @xmath115(0.62 ) & 25.19(0.80 ) & 5.25 & 50,735.6405(28 ) & 0.502103 & + & a1v & hip 116611 & & & & @xmath116 [ 16,115 ] & & + iw per & sb1 ( eb ? ) & hd 21912 & @xmath117(0.31 ) & 98.64(0.51 ) & 3.46 & 53,031.4192(8 ) & 0.91718 & + & ( a3v ) & hip 16591 & & & & @xmath118 4940 ] & & + v592 per & ew(a ) & hd 29911 & @xmath119(1.04 ) & 93.74(1.25 ) & 9.42 & 52,516.8641(20 ) & 0.715722 & 0.408(7 ) + & ( f2iv ) & hip 22050 & & 229.97(1.85 ) & 15.00 & @xmath120 [ 5612 ] & 2.521(72 ) & + tu umi & ew(w ) & bd@xmath121 & @xmath122(0.20 ) & 35(15 ) & 30 & 52,725.6262 & 0.377088 & 0.16(7 ) + & ( f2 ) & hip 73047 & & 220(20 ) & 40 & 0 [ 0 ] & 0.65(27 ) & + fo vir & ea / eb & hd 117362 & @xmath123(1.45 ) & 33.84(1.02 ) & 6.75 & 51,660.1839(25 ) & 0.775568 & 0.125(5 ) + & ( a7v ) & hip 65839 & & 269.91(2.95 ) & 15.87 & @xmath124 [ 4074 ] & 2.257(89 ) & +

the phenomenological interest of prompt - photon production in fixed - target experiments @xcite resides mainly in its use as a gluon probe in structure - function studies . prompt - photon production is historically our main source of information on the gluon parton density at large @xmath4 ( e.g. @xmath5 ) @xcite , a region which has very little influence on the evolution of the deep - inelastic - scattering structure functions . this same region is relevant for hadron colliders in production phenomena at very large transverse momenta , and thus its understanding is crucial in order to disentangle possible new physics signals from the qcd background . for example , a particularly interesting problem has emerged in the past few years in the production of large - transverse - energy ( ) jets at the tevatron . an excess over the qcd prediction has been reported by the cdf collaboration @xcite , for jets with @xmath6 gev . while the excess has not been confirmed by the d@xmath7 data @xcite , it is of interest to study the uncertainty in the high- tail of the jet distribution due to the gluon density systematics , to see whether there is room for deviations as large as those detected by cdf . for example , a suitable modification of the gluon density at large @xmath4 has been proposed ( cteq4hj @xcite ) , which is consistent with the excess observed by cdf . the study of the recent e706 prompt - photon data @xcite , however , suggests that a consistent fit of the large-@xmath8 ( @xmath9 ) rate is incompatible with the cteq4hj gluon density @xcite . moreover , both jet cross sections and direct - photon cross sections at high transverse energy are affected by soft - gluon effects . these effects should be understood in both cases in order to be able to claim a discrepancy with qcd predictions . in particular , these effects can be very important in the direct - photon case , since the typical values probed are much smaller than in the case of jet production at the tevatron and , therefore , the size of the running coupling @xmath10 at the relevant scales is bigger . comparisons between theory and prompt - photon experimental results have been carried out recently in refs . the recent e706 data @xcite seem to differ most from the next - to - leading order calculation , over the whole @xmath8 range . in refs . @xcite , an attempt is made to fit the e706 data by introducing an intrinsic transverse momentum of the incoming partons with @xmath11 . the precise details of how the intrinsic @xmath12 is incorporated in the calculations , however , can significantly affect the impact of these corrections , as shown by the large variations reported in ref . @xcite . the use of an intrinsic - transverse - momentum model is sometimes motivated as a way of estimating the effects of soft - gluon emission . the most prominent effect of soft - gluon emission in drell - yan pair production is the generation of the characteristic transverse - momentum spectrum of the lepton pair . this can be modeled with an appropriate intrinsic transverse momentum of the incoming partons . as a matter of fact , the formalism for soft - gluon resummation in drell - yan pair production can be shown to merge , at very small transverse momenta , into some non - perturbative intrinsic transverse momentum of the partons inside the hadron @xcite . while this approach is not unreasonable when one considers the transverse momentum of the produced _ pair _ , it can however lead to inconsistencies for the problem of _ single_-photon production . in fact , for example , it is quite clear that the photon @xmath8 spectrum at large @xmath8 explores the kinematic region of @xmath13 in the parton densities , which is certainly not the case for the transverse - momentum distribution of a drell - yan pair . thus , as of now , a method for the inclusion of non - perturbative effects in the resummed formulae for the high-@xmath8 limit of the inclusive photon cross section is not available . furthermore , in the opposite limit of small @xmath8 , it is the multiple emission of hard ( rather than soft ) gluons that leads to a sizeable perturbative broadening of the transverse momenta of the incoming partons @xcite . in this work , we consider the effect of soft - gluon resummation in prompt - photon production near the threshold limit , that is to say for @xmath14 . the theoretical evaluation of these effects , at the next - to - leading logarithmic accuracy , has been carried out independently in refs . @xcite and @xcite . we shall review in the next section the necessary formalism , using the language of ref . @xcite . in the rest of the paper , we will present its phenomenological applications , and we will thus discuss its numerical implementation as well as its impact on physical cross sections . as is well known , prompt - photon production takes place both by hard - photon emission from initial- or final - state quarks ( direct component ) , and by collinear radiation from final - state partons . this last mechanism is not fully calculable in perturbation theory and , in fact , it depends upon the photon fragmentation function . because of the large suppression of the fragmentation function at large momentum fractions @xmath15 , it is usually believed that this contribution becomes irrelevant when @xmath16 increases . contrary to common wisdom , we shall instead show that the very - large-@xmath8 behaviour of the direct and of the fragmentation production processes is the same if the incoming hadrons do not contain valence antiquarks , as in the case of @xmath17 collisions . under these circumstances , resummation should therefore be performed for the fragmentation emission too . we will show , however , that in the cases of practical interest the corrections due to the fragmentation processes are small , and we shall limit our considerations to the hard - photon part . the plan of the paper is as follows . in section [ secnotat ] we review the formalism for the resummation of threshold effects , and the main formulas valid for the specific case of prompt - photon production . there we also recall the main issues related to the inversion of the resummed expressions from mellin space back to the physical @xmath4 space . in section [ results ] we study numerically the impact of the resummation corrections . we explore the effects both at the parton and hadron level , considering kinematical configurations and distributions of phenomenological relevance for current experiments . in particular , we concentrate on the study of the size of the resummation corrections , and of the residual dependence on the choice of renormalization and factorization scales . section [ datacomp ] contains a comparison between our results and the data from some recent experiments . this does not want to be a comprehensive phenomenological study , but a preliminary analysis of the impact of our results on the comparison of theory and data . our conclusions , and the outlook for future progress , are given in section [ conclusions ] . an appendix collects some details of the resummation formulas . we consider the inclusive production of a single prompt photon in hadron collisions : [ procgamma ] h_1(p_1 ) + h_2(p_2 ) ( p ) + x . the colliding hadrons @xmath18 and @xmath19 carry momenta @xmath20 and @xmath21 , respectively . we parametrize the momenta in terms of light - cone coordinates : p^= ( p^+ , p_t , p^- ) , p^ ( p^0 p^3 ) . in their centre - of - mass frame , using massless kinematics , the momenta of the colliding hadrons have the following light - cone coordinates [ mom ] p_1^= ( 1,0,0 ) , p_2^= ( 0,0,1 ) , where @xmath22 is the centre - of - mass energy squared . the photon momentum @xmath23 is thus parametrized as [ momp ] p^= ( e^y , e_t , e^-y ) , where @xmath24 and @xmath25 are the transverse energy and the rapidity , respectively . we also introduce the customary scaling variable @xmath8 @xmath26 : [ xt ] x_t = . in the present paper we are mostly interested in the prompt - photon production cross section integrated over @xmath25 at fixed @xmath24 . according to perturbative qcd , the cross section is given by the following factorization formula [ 1pxsgamma ] & & = _ a , b _ 0 ^ 1 dx_1 f_a / h_1(x_1,_f^2 ) _ 0 ^ 1 dx_2 f_b / h_2(x_2,_f^2 ) + & & _ 0 ^ 1 dx \ { ( x - ) _ ab(x , ( ^2 ) ; e_t^2 , ^2 , _ f^2 , _ f^2 ) . + & & + . _ c _ 0 ^ 1 dz z^2 d_c/(z,_f^2 ) ( x - ) _ abc(x , ( ^2 ) ; e_t^2 , ^2 , _ f^2 ) } . where @xmath27 denotes the parton indices @xmath28 , and @xmath29 and @xmath30 are the parton densities of the colliding hadrons , evaluated at the factorization scale @xmath31 . the first and the second term in the curly bracket on the right - hand side of eq . ( [ 1pxsgamma ] ) represent the _ direct _ and the _ fragmentation _ component of the cross section , respectively . the fragmentation component involves the parton fragmentation function @xmath32 of the observed photon at the factorization scale @xmath33 , which , in general , differs from the scale @xmath31 of the parton densities . the _ rescaled _ ( @xmath34 ) . ] partonic cross sections @xmath35 and @xmath36 in eq . ( [ 1pxsgamma ] ) are computable in qcd perturbation theory as power series expansions in the running coupling @xmath37 , @xmath38 being the renormalization scale in the @xmath39 renormalization scheme : [ pxsg ] _ ab(x , ( ^2 ) ; e_t^2 , ^2 , _ f^2 ) = ( ^2 ) , [ pxsc ] _ abc(x , ( ^2 ) ; e_t^2 , ^2 , _ f^2 ) = ^2(^2 ) . note that the ratio between the direct and the fragmentation terms in eqs . ( [ pxsg ] ) and ( [ pxsc ] ) is of the order of @xmath40 , where @xmath41 is the fine structure constant . this ratio is compensated by the photon - fragmentation function @xmath42 , which ( at least formally ) is of the order of @xmath40 , so that direct and fragmentation components equally contribute to eq . ( [ 1pxsgamma ] ) . throughout the paper we always use parton densities and parton fragmentation functions as defined in the @xmath39 factorization scheme . in general , we consider different values for the renormalization and factorization scales @xmath38 , @xmath31 , @xmath33 , although we always assume that all of them are of the order of the photon transverse energy @xmath24 . the lo terms @xmath43 in eq . ( [ pxsg ] ) are due to the following parton - scattering subprocesses at the tree - level [ loproc ] q + |q g + , q + g q + , + g + . using our normalization , the two independent ( non - vanishing ) partonic cross sections for the direct component are : [ siqqgamma ] _ q|q g^(0)(x ) = e_q^2 ( 2 - x^2 ) [ siqggamma ] _ qg q^(0)(x ) = _ |qg ^(0)(x ) = e_q^2 ( 1 + ) , where @xmath44 is the quark electric charge . note that , having integrated over the photon pseudorapidity , the expressions ( [ siqqgamma ] ) and ( [ siqggamma ] ) are even functions of the photon transverse energy @xmath24 , i.e. they depend on @xmath45 rather than on @xmath4 . the nlo terms @xmath46 in eq . ( [ pxsg ] ) were first computed in ref . @xcite . the partonic contributions @xmath36 to the fragmentation component of the cross section are exactly equal to those of the single - hadron inclusive distribution . note that , unlike in the case of the direct component , _ all _ the parton - parton scattering subprocesses @xmath47 ( i.e. including @xmath48 ) contribute to the fragmentation component already at lo . the explicit calculation of @xmath36 up to nlo was performed in ref . @xcite . the behaviour of the lo and nlo perturbative contributions to the direct component of the prompt - photon cross section is shown in fig . [ fig : qalonlo ] . the lo terms @xmath49 and @xmath50 are both singular when @xmath51 : [ lobeh ] _ abd ^(0)(x ) ~ , ( x 1 ) , and they both vanish in the high - energy limit @xmath52 . the integrable singularity in eq . ( [ lobeh ] ) is a typical phase - space effect , while the vanishing behaviour at small-@xmath4 is due to the dominance of fermion ( i.e. spin @xmath53 ) exchange in the @xmath54-channel . two new dynamical features appear at nlo . near the threshold region @xmath51 , the nlo contributions are double - logarithmically enhanced , [ nlobeh ] ^(1)(x ) ~ ^(0)(x ) ^2 ( 1-x ) , ( x 1 ) , because the radiation of soft and , possibly , collinear partons is strongly inhibited by the kinematics . in the high - energy limit @xmath52 , the partonic cross sections @xmath55 approach constant values @xcite : this regge plateau follows from the fact that at nlo single - gluon ( i.e. spin 1 ) @xmath54-channel exchange affects all the partonic subprocesses . the behaviour of the partonic contributions @xmath56 in the remaining intermediate region of @xmath4 has no straightforward physical interpretation ( e.g. @xmath57 even becomes negative ) because it strongly depends on the scale - dependent corrections already subtracted in the definition of the parton densities and parton fragmentation functions . higher - order perturbative qcd corrections in the small-@xmath8 regime can systematically be computed by using the @xmath58-factorization approach @xcite , which consistently takes into account the _ perturbative _ broadening of the transverse momenta of the incoming partons . we are interested in this work in the behaviour of the qcd corrections near the partonic - threshold region @xmath51 , i.e. when the transverse energy @xmath24 of the photon approaches the partonic centre - of - mass energy @xmath59 . in this region , the singularities in eqs . ( [ lobeh ] , [ nlobeh ] ) are enhanced by double - logarithmic corrections due to soft - gluon radiation and the higher - order cross section contributions in eqs . ( [ pxsg ] , [ pxsc ] ) behave as [ hobeh ] ^(n)(x ) ~ ^(0)(x ) . resummation of these soft - gluon effects to all orders in perturbation theory can be important to improve the reliability of the qcd predictions . the resummation program of soft - gluon contributions has to be carried out @xcite in the mellin - transform space , or @xmath60-space . working in @xmath60-space , we can disentangle the soft - gluon effects in the parton densities from those in the partonic cross section and we can straightforwardly implement and factorize the kinematic constraints of energy and longitudinal - momentum conservation . the latter point is particularly relevant for soft - gluon resummation in hadron collisions @xcite . indeed , all - order soft - momentum recoil can not exactly be taken into account by directly working in @xmath4-space and the ensuing kinematics approximation leads to ( same - sign ) factorially growing coefficients . this implies @xcite that no resummed logarithmic hierarchy can consistently be defined in @xmath4-space ( the classes of leading logs @xmath61 , next - to - leading logs @xmath62 and so forth in eq . ( [ hobeh ] ) are not separately summable , because they lead to divergent and not integrable contributions at @xmath63 ) . on the contrary , no kinematics approximation ( in the soft limit ) is required in @xmath60-space and the corresponding logarithmic hierarchy of @xmath64-contributions is systematically well defined . to work in @xmath60-space , it is convenient to consider the mellin transform @xmath65 of the dimensionless hadronic distribution @xmath66 . the @xmath60-moments with respect to @xmath67 and at fixed @xmath24 are thus defined as follows : [ shn ] _ , n(e_t ) _ 0 ^ 1 dx_t^2 ( x_t^2)^n-1 e_t^3 . in @xmath60-moment space , eq . ( [ 1pxsgamma ] ) takes a simple factorized form [ 1pxsngamma ] _ , n(e_t ) & = & _ a , b f_a / h_1,n+1(_f^2 ) f_b / h_2,n+1(_f^2 ) + & & \ { _ ab , n((^2 ) ; e_t^2 , ^2 , _ f^2 ) + & + & . _ c _ abc , n((^2 ) ; e_t^2 , ^2 , _ f^2 ) d_c/ , 2n+3(_f^2 ) } , where we have introduced the customary @xmath60-moments @xmath68 and @xmath69 of the parton densities and parton fragmentation functions : [ pdm ] f_a / h , n(^2 ) & & _ 0 ^ 1 dx x^n-1 f_a / h(x,^2 ) , + [ pffm ] d_a/ , n(^2 ) & & _ 0 ^ 1 dz z^n-1 d_a/(z,^2 ) . note that the @xmath60-moments of the partonic cross sections in eq . ( [ 1pxsngamma ] ) are again defined with respect to @xmath67 : [ ndef ] _ ab , n((^2 ) ; e_t^2 , ^2 , _ f^2 ) _ 0 ^ 1 dx^2 ( x^2)^n-1 _ ab(x,(^2 ) ; e_t^2 , ^2 , _ f^2 ) . the explicit expressions of the @xmath60-moments @xmath70 , @xmath71 of the lo contributions in eqs . ( [ siqqgamma ] ) and ( [ siqggamma ] ) were obtained in ref . @xcite and are recalled in appendix [ appa ] . note also the pattern of moment indices in the various factors of eq . ( [ 1pxsngamma ] ) , that is , @xmath72 for the parton densities and @xmath73 for the parton fragmentation functions . this non - trivial pattern follows from the conservation of the longitudinal and transverse momenta . the threshold region @xmath74 corresponds to the limit @xmath75 in @xmath60-moment space . in this limit , the soft - gluon corrections ( [ hobeh ] ) to the higher - order contributions of the partonic cross sections become [ hobehn ] ^(n)_n ~ ^(0)_n . the resummation of the soft - gluon logarithmic corrections to all orders in perturbation theory has been considered in refs . @xcite and @xcite . in the following section we recall the main results . in ref . @xcite soft - gluon resummation has been performed in detail for the various partonic channels that contribute to the _ direct _ component of the prompt - photon cross section @xmath65 in eq . ( [ 1pxsngamma ] ) . we discuss first the large-@xmath60 behaviour of the partonic cross sections @xmath76 for the partonic channels @xmath77 that start to contribute at lo . these cross sections can be written as [ sigreslo ] _ ab , n = _ ab , n^(res ) , ab = q|q , qg , |qg , where @xmath78 denotes terms that contribute beyond lo and are furthermore suppressed by a relative factor @xmath79 at large @xmath60 . the logarithmically - enhanced soft - gluon corrections are included in the resummed expressions @xmath80 and can be factorized with respect to the corresponding lo cross sections @xmath81 . the _ all - order _ resummation formulae are _ q|q , n^(res)((^2 ) ; e_t^2 , ^2 , _ f^2 , _ f^2 ) & = & ( ^2 ) _ q|qg , n^(0 ) c_q|q ( ( ^2),q^2/^2;q^2/_f^2 ) + [ gammaresqq ] & & _ n+1^q|q g ( ( ^2),q^2/^2;q^2/_f^2 ) , + + _ qg , n^(res)((^2 ) ; e_t^2 , ^2 , _ f^2 , _ f^2 ) & = & ( ^2 ) _ qgq , n^(0 ) c_qg ( ( ^2),q^2/^2;q^2/_f^2 ) + [ gammaresqg ] & & _ n+1^qg q ( ( ^2),q^2/^2;q^2/_f^2 ) , + + [ gammaresqbarg ] _ |qg , n^(res)((^2 ) ; e_t^2 , ^2 , _ f^2 , _ f^2 ) & = & _ qg , n^(res)((^2 ) ; e_t^2 , ^2 , _ f^2 , _ f^2 ) , where [ etscale ] q^2 = 2 e_t^2 . the functions @xmath82 in eqs . ( [ gammaresqq ] , [ gammaresqg ] ) do not depend on @xmath60 . thus , the @xmath64-dependence of the resummed cross sections is entirely embodied by the radiative factors @xmath83 . they depend on the flavour of the qcd partons @xmath84 involved in the lo hard - scattering subprocess @xmath85 and can be expressed in an exponential form : & & _ n^ab d ( ( ^2 ) , ; ) = \ { n g_ab^(1)(b_0(^2)n ) + & & + g_ab^(2)(b_0(^2)n , q^2/^2;q^2/_f^2 ) [ deltanll ] + o((n)^k ) } , where @xmath86 is the first coefficient of the qcd @xmath87-function [ betafirst ] b_0 = . note that the functions @xmath88 , @xmath89 and so forth in the exponent do not depend separately on @xmath10 and @xmath64 . they are functions of the expansion variable @xmath90 and vanish when @xmath91 . this means that the exponentiation structure in eq . ( [ deltanll ] ) is not trivial and , in particular , that all the double logarithmic ( dl ) terms @xmath92 in eq . ( [ hobehn ] ) are taken into account by simply exponentiating the lowest - order contribution @xmath93 . the exponentiation in eq . ( [ deltanll ] ) defines an improved perturbative expansion in the threshold region . the function @xmath94 resums all the _ leading _ logarithmic ( ll ) contributions @xmath95 in the exponent , @xmath89 contains the _ next - to - leading _ logarithmic ( nll ) terms @xmath96 , and so forth . once the functions @xmath97 have been computed , we have a systematic perturbative treatment of the region of @xmath60 where @xmath98 , which is much larger than the domain @xmath99 where the fixed - order calculation in @xmath10 is reliable . the ll and nll functions @xmath88 and @xmath89 in eq . ( [ deltanll ] ) have been explicitly computed in ref . the ll functions @xmath88 are different for the @xmath100 and @xmath101 partonic channels of eqs . ( [ gammaresqq ] ) and ( [ gammaresqg ] ) but they can be expressed in terms of parton colour factors and a single ( parton - independent ) function @xmath102 : g_q|q^(1 ) ( ) & = & ( 2c_f - c_a ) h^(1 ) ( ) + c_a h^(1)(/2 ) , + [ g1fun ] g_qg^(1 ) ( ) & = & c_a h^(1 ) ( ) + c_f h^(1)(/2 ) , with [ htll ] h^(1 ) ( ) = . the explicit expressions of the nll functions @xmath89 are recalled in appendix [ appa ] . note that the ll functions @xmath88 do not depend on the factorization scale @xmath31 . this dependence starts to appear only in the nll functions @xmath89 . note also the mismatch between the moment index of the radiative factor and that of @xmath81 in eqs . ( [ gammaresqq ] , [ gammaresqg ] ) : the former depends on @xmath103 , like the parton densities in eq . ( [ 1pxsngamma ] ) . the explicit @xmath31-dependence of @xmath89 exactly matches the scale dependence of the parton densities at large values of @xmath60 . thus , when ( and only when ) nll resummation is included , we can expect @xcite better stabilization of the calculation of the cross section at large @xmath8 with respect to variations of the factorization scale @xmath31 ( see sec . [ hlresults ] ) . the functions @xmath82 in eqs . ( [ gammaresqq ] , [ gammaresqg ] ) contain all the terms that are constant in the large-@xmath60 limit . they are produced by hard _ virtual _ contributions and by subdominant ( non - logarithmic ) soft corrections to the lo hard - scattering subprocesses . these functions are computable as power series expansions in @xmath10 c_ab((^2),q^2/^2;q^2/_f^2 ) & = & 1 + _ n=1^+ ( ) ^n c_ab^(n)(q^2/^2;q^2/_f^2 ) + [ cgamma ] & = & 1 + c_ab^(1)(q^2/^2;q^2/_f^2 ) + o(^2 ) . at present , we know only the first - order constant coefficients @xmath104 and @xmath105 in eqs . ( [ cgamma ] , [ gammaresqq ] , [ gammaresqg ] ) . these coefficients can be extracted @xcite from the complete nlo analytic results of refs . their values are recalled in appendix [ appa ] . the inclusion of the @xmath60-independent function @xmath82 in the resummed formulae does not affect the shape of the cross section near threshold , but improves the soft - gluon resummation by fixing the overall ( perturbative ) normalization of the logarithmic radiative factor . we can explicitly show @xcite the theoretical improvement that is obtained by combining the nll radiative factor with the first - order coefficient @xmath106 . expanding the resummation formulae ( [ gammaresqq ] , [ gammaresqg ] ) in towers of logarithmic contributions as in eq . ( [ hobehn ] ) , we have & & _ n^(res)(;e_t^2,^2,_f^2 ) = _ n^(0 ) \ { 1 + _ n=1^ ^n } , [ sigtow ] where @xmath107 . the dominant and next - to - dominant coefficients @xmath108 and @xmath109 are controlled by evaluating the radiative factor to nll accuracy . when the nll radiative factor is supplemented with the coefficient @xmath110 , we can correctly control also the coefficients @xmath111 . in particular , we can predict @xcite the large-@xmath60 behaviour of the next - to - next - to - leading order ( nnlo ) cross sections @xmath112 in eq . ( [ pxsg ] ) up to @xmath113 . note also that the coefficients @xmath108 are scale independent and the coefficients @xmath109 depend on the sole factorization scale @xmath31 . in the tower expansion ( [ sigtow ] ) , the first terms that explicitly depend on the renormalization scale @xmath38 ( and on @xmath31 , as well ) are those controlled by @xmath111 . their dependence on @xmath38 is obtained by combining that of @xmath114 with that of the radiative factor at nll order . the inclusion of the first - order constant coefficient @xmath110 thus theoretically stabilizes the resummed partonic cross section at large @xmath8 with respect to variations of the renormalization scale . this scale dependence is numerically studied in sec . [ hlresults ] . so far we have only considered the near - threshold behaviour of the partonic cross sections @xmath115 , @xmath116 , @xmath117 in eq . ( [ sigreslo ] ) . the behaviour of other partonic channels @xmath118 that contribute to the direct component of the prompt - photon cross section was discussed in ref . it turns out that the partonic channel @xmath119 enters the resummed cross section only at next - to - next - to - leading logarithmic ( nnll ) accuracy and that all the other channels are relatively suppressed in the same way as the correction @xmath78 on the right - hand side of eq . ( [ sigreslo ] ) . since we are interested in explicitly performing soft - gluon resummation up to nll order , we can limit ourselves to considering the resummed expressions in eqs . ( [ sigreslo])([gammaresqbarg ] ) . detailed numerical studies of the resummed cross sections are presented in sec . [ results ] . however , from the analytical results reviewed in this section , we may already anticipate that soft - gluon resummation increases the perturbative qcd predictions in the large-@xmath8 region . this conclusion can be argued by a simplified treatment within the dl approximation . to dl accuracy , the exponent of the radiative factors in eq . ( [ deltanll ] ) has to be expanded to its first order in @xmath10 , and we obtain [ dlresqg ] & & \ { ^2 n } = \ { ( c_f + 2c_a ) ^2 n } > 1 , + & & + [ dlresqq ] & & \ { ^2 n } = \ { ( 4c_f - c_a ) ^2 n } > 1 , + & & + [ dlrat ] & & \ { 3(c_a - c_f ) ^2 n } > . for the sake of completeness , in the square bracket on the right - hand side of eqs . ( [ dlresqg ] ) and ( [ dlresqq ] ) we have explicitly separated the positive contributions coming from the initial - state partons and the negative contribution from the final - state recoil . from these equations we see that the resummed partonic cross sections @xmath120 and @xmath121 are both enhanced with respect to their lo approximations @xmath70 , @xmath71 . moreover , the enhancement in the @xmath101 partonic channel is larger than that in the @xmath100 channel . we refer the reader to ref . @xcite for a discussion on the physical origin of this behaviour . we can now comment on the large-@xmath24 behaviour of the fragmentation component of the prompt - photon cross section , by comparing the direct and fragmentation contributions in eq . ( [ 1pxsngamma ] ) . the partonic cross sections @xmath76 and @xmath122 have the same large-@xmath60 behaviour , but , owing to the hard ( although collinear ) emission always involved in any splitting process @xmath123 , the photon fragmentation function @xmath124 is of the order of @xmath125 . therefore , in the curly bracket on the right - hand side of eq . ( [ 1pxsngamma ] ) the fragmentation component is formally suppressed by a factor of @xmath125 with respect to the direct component . this suppression is consistent with the fact that the resummed partonic cross sections for the direct processes ( see the right - hand side of eqs . ( [ gammaresqq ] ) and ( [ gammaresqg ] ) ) turn out to be independent of the photon fragmentation scale @xmath33 . this argument shows that , in many cases , the fragmentation contributions are subdominant near threshold and , thus , they can be neglected in resummed calculations at large @xmath8 . the caveat ` in many cases ' in the above conclusion regards the fact that the argument applies to the partonic contributions in the curly bracket of eq . ( [ 1pxsngamma ] ) . in other words , the argument assumes that all the different initial - state partonic channels @xmath126 give comparable contributions to the hadronic cross section . this is not always true once the effect of the parton densities is included . a relevant exception is indeed the case of prompt - photon production in proton - nucleon collisions . owing to the low antiquark content of the colliding hadrons , the hadronic cross section is mostly due to the partonic channels @xmath127 and @xmath128 : [ xsngammapn ] _ , n(e_t ) ~ _ , n^qg(e_t ) + _ , n^qq(e_t ) , ( pn ) . as for the @xmath101 initial - state contribution @xmath129 , we can use the above argument to conclude that its direct component dominates at large @xmath24 . setting all the scales equal to @xmath24 , for the sake of simplicity , we can write : [ xsngammapnqg ] _ , n^qg(e_t ) ~ _ , n^qg ( dir)(e_t ) ~f_q , n+1(e_t^2 ) f_g , n+1(e_t^2 ) _ qg , n((e_t^2 ) ) . however , in the case of the @xmath130 initial state , the direct component enters only at nlo and , thus , the cross section is dominated by the fragmentation part and , in particular , by photon fragmentation from a final - state quark of the lo scattering subprocess @xmath131 . we can write : [ xsngammapnqq ] _ , n^qq(e_t ) ~ _ , n^qq ( frag)(e_t ) ~f_q , n+1(e_t^2 ) f_q , n+1(e_t^2 ) _ qqq , n((e_t^2 ) ) d_q/ , 2n+3(e_t^2 ) . taking the ratio of the two initial - state contributions and replacing @xmath132 and @xmath133 by their lo contributions @xmath134 in eqs . ( [ pxsg ] , [ pxsc ] ) , we obtain [ qgqqratio ] ~ d_q/ , 2n+3(e_t^2 ) . the factor @xmath135 on the right - hand side is compensated by the behaviour of the photon fragmentation function @xmath136 . in the large-@xmath60 limit , the ratio of the lo partonic contributions @xmath134 is constant and , thus , the fragmentation function produces an @xmath79-suppression factor . nonetheless , this suppression can be balanced by the parton density contribution @xmath137 since , at large @xmath4 , the gluon density is typically softer than the quark density . as a matter of fact , using the altarelli - parisi evolution equation at lo and under reasonable assumptions on the large-@xmath4 behaviour of the parton densities at the initial evolution scale , it is easy to show that we have the following asymptotic behaviour at very large values of @xmath60 and of the evolution scale : [ eq : largenfg ] f_g , n ~ f_q , n , + [ eq : largendg ] d_q/ , n ~ . combining these results with eq . ( [ qgqqratio ] ) , in the large-@xmath60 limit we therefore get [ eq : fraglargen ] . this discussion shows that , in the case of large-@xmath8 prompt - photon production in @xmath17 collisions , the contribution of the fragmentation component to the hadronic cross section can become comparable to that of the direct component . although on the right - hand side of eq . ( [ qgqqratio ] ) we have approximated the partonic contribution @xmath138 by its lo expansion , the inclusion of higher - order terms and , in particular , of resummation effects does not substantially modify the conclusion . performing soft - gluon resummation in the partonic cross sections @xmath122 of the fragmentation component , we can write an expression that is analogous to eq . ( [ sigreslo ] ) : [ sigreslof ] _ abc , n = _ abc , n^(res ) . limiting our treatment to the ll accuracy , the resummed cross section is given by @xcite [ gammaresf ] _ abc , n^(res)((^2 ) ; e_t^2 , ^2 , _ f^2 ) ^2(^2 ) _ abdc , n^(0 ) _ n+1^ab dc((^2),q^2/^2;q^2/_f^2,q^2/_f^2 ) . the radiative factor is [ deltanllf ] _ n^ab d c((^2 ) , ; , ) = \ { n g_abdc^(1)(b_0(^2)n ) + o((n)^k ) } , where the ll function @xmath139 is analogous to those in eq . ( [ g1fun ] ) and it can be expressed in terms of the colour charges @xmath140 ( @xmath141 , if the parton @xmath142 is a quark , and @xmath143 , if @xmath142 is a gluon ) of the partons involved in the lo hard - scattering subprocess : g_abdc^(1 ) ( ) = ( c_a + c_b + c_c - c_d ) h^(1 ) ( ) + c_d h^(1)(/2 ) . in particular , for the @xmath144 channel we have [ g1qqqq ] g_qqqq^(1 ) ( ) = 2 c_f h^(1 ) ( ) + c_f h^(1)(/2 ) , which is very similar to @xmath145 in eq . ( [ g1fun ] ) because @xmath146 . more precisely , since @xmath147 , @xmath148 is slightly smaller than @xmath145 as long as they are evaluated in the perturbative region @xmath149 . we can now come back to the effect of the fragmentation component in @xmath17 collisions . using the resummed partonic cross sections in eqs . ( [ gammaresqg ] , [ gammaresf ] ) rather than their lo approximations , the right - hand side of eq . ( [ qgqqratio ] ) has to be multiplied by an additional contribution , as given by the ratio of the corresponding radiative factors , namely [ deltacorr ] . because of the relation @xmath150 between the colour charges , this factor does not sizeably differ from unity , as it can be argued by its dl approximation on the right - hand side of eq . ( [ deltacorr ] ) ( see also the comment below eq . ( [ g1qqqq ] ) ) . we have thus shown that , at least at the ll level , soft - gluon resummation does not enhance the relative importance of the fragmentation component for prompt - photon production at large @xmath8 . the importance of the fragmentation component in @xmath17 collisions mainly depends on the detailed behaviour of the parton densities at large @xmath4 and on how large are the values of @xmath24 of interest . this issue , as well as the impact of the nll corrections to the ll results obtained above , require further studies that will be presented in a future work . as for the present study , we limit ourselves to perform soft - gluon resummation in the direct component and we check that the fragmentation component does not sizeably contribute to the hadronic cross section in the actual experimental configurations investigated in the paper ( see sec . [ datacomp ] ) . we use soft - gluon resummation to nll accuracy at the parton level to introduce an improved prompt - photon cross section @xmath151 as follows _ , n^(res)(e_t ) & = & _ ab = q|q , qg , |qg f_a / h_1,n+1(_f^2 ) f_b / h_2,n+1(_f^2 ) + & & + [ ppcsres ] & + & _ , n^(nlo)(e_t ) , where @xmath152 is the prompt - photon hadronic cross section at nlo , @xmath80 is given in eqs . ( [ gammaresqq])([gammaresqbarg ] ) and @xmath153 represents its perturbative truncation at order @xmath154 ( i.e. at nlo ) . thus , because of the subtraction in the square bracket on the right - hand side , eq . ( [ ppcsres ] ) exactly reproduces the nlo results and resums soft - gluon effects beyond @xmath155 to nll accuracy . in general , we evaluate @xmath80 using the nll expression ( [ deltanll ] ) of the radiative factors and including the @xmath156 contribution ( [ cgamma ] ) of the constant factors @xmath82 . this defines our nlo+nll predictions . the resummed formulae presented so far are given in @xmath60-moment space . to obtain cross sections in the physical @xmath8-space ( i.e. as functions of the centre - of - mass energy ) , one has to perform the inverse mellin transformation : [ mppp ] e_t^3 = _ c_mp - i^c_mp+i dn x_t^-2n _ , n^(res)(e_t ) . when the @xmath60-moments @xmath157 are evaluated at a fixed perturbative order in @xmath10 , they are analytic functions in a right half - plane of the complex variable @xmath60 . in this case , the constant @xmath158 that defines the integration contour in eq . ( [ mppp ] ) has to be chosen in this half - plane , that is , on the right of all the possible singularities of the @xmath60-moments . an additional complication occurs when the @xmath60-moments are computed in resummed perturbation theory . in this case , since the resummed functions @xmath159 in eq . ( [ g1fun ] ) ( as well as the nll functions @xmath160 ) are singular at @xmath161 , the soft - gluon factors @xmath162 in eq . ( [ deltanll ] ) have cut singularities that start at the branch - point @xmath163 . these singularities , which are related to the divergent behaviour of the perturbative running coupling @xmath10 near the landau pole , signal the onset of non - perturbative phenomena at very large values of @xmath60 or , equivalently , in the region very close to threshold . the issue of how to deal with the landau singularity in soft - gluon resummation formulae for hadronic collisions was discussed in detail in ref . @xcite . in the evaluation of the inverse mellin transformation ( [ mppp ] ) we thus use the _ minimal prescription _ introduced in ref . the constant @xmath158 is chosen in such a way that all singularities in the integrand are to the left of the integration contour , except for the landau singularity at @xmath164 , that should lie to the far right . this prescription is consistent @xcite with the perturbative content of the soft - gluon resummation formulae because it converges asymptotically to the perturbative series and it does not introduce ( unjustified ) power corrections of non - perturbative origin . these corrections are certainly present in physical cross - sections , but their effect is not expected to be sizeable as long as @xmath24 is sufficiently perturbative and @xmath8 is sufficiently far from the hadronic threshold . obviously , approaching the essentially non - perturbative regime @xmath165 gev , @xmath74 , a physically motivated treatment of non - perturbative effects has to be introduced . in the following sections , we limit ourselves to presenting numerical and phenomenological results that do not include any non - perturbative correction we present in this section some numerical results , to provide an illustration of the size of the effects considered and to show the improvements obtained with respect to scale variations after the inclusion of the nll corrections . we start by discussing the resummation effects at the level of partonic cross sections . the resummed partonic cross section can be obtained from eqs . ( [ ppcsres ] ) and ( [ mppp ] ) by assuming parton - density functions of the form @xmath166 , and hence @xmath167 for all complex values of @xmath60 . we consider first the @xmath168 terms in the expansion of the resummed cross section , in order to estimate to which accuracy this reproduces the exact nlo results . in fig . [ fig : qqnlo ] ( left ) we plot the function @xmath169 , defined in eq . ( [ pxsg ] ) , as a function of @xmath170 . the exact @xmath168 result @xcite is compared with three possible implementations of the resummation procedure , all equivalent at nll accuracy . the first case ( short - dashed line ) corresponds to our default resummed prediction , as given by eq . ( [ gammaresqq ] ) . in the second case ( dot - dashed line ) , we set the constant @xmath171 introduced in eq . ( [ cgamma ] ) equal to 0 . in the third case , we keep the contribution of the constant @xmath172 , but we modify it by a term suppressed by a factor of @xmath125 , in order to explore the possible effect of contributions of order @xmath125 which can not be taken into account by the soft - gluon resummation . as a constraint on the form of these corrections , we must impose that no poles appear on the positive real axis in the @xmath60 plane ( these poles would logarithmically enhance the partonic cross section when @xmath173 ) . we select a parametrization of the @xmath125 corrections that allows us to bracket the exact result at : [ eq : cqashift ] c_q |q ^(1 ) c_q |q ^(1 ) ( 1 + ) , [ eq : cqgshift ] c_q g^(1 ) c_qg^(1 ) ( 1- ) , with @xmath174 . in our applications we shall consider the two cases with @xmath175 ( namely no correction to the @xmath176 term ) and @xmath177 as a way to establish the size of subleading threshold corrections beyond the nll order . as one can see from fig . [ fig : qqnlo ] , the inclusion of the finite term @xmath178 is essential to accurately reproduce shape and normalization of the exact result not only near threshold , but below it as well . the agreement deteriorates unavoidably for @xmath179 , as , here , terms subleading in @xmath125 become important . analogous results for the @xmath101 channel are given in the right panel of fig . [ fig : qqnlo ] . note that in both cases the two choices @xmath175 and @xmath177 in eq . ( [ eq : cqgshift ] ) bracket the exact result over a large region of @xmath180 , and thus provide a good estimator of the subleading terms systematics . the choice @xmath175 , furthermore , provides a very accurate description up to values of @xmath180 of the order of 1/10 . the fully - resummed parton - level cross sections are shown in fig . [ fig : qares10 ] for the @xmath181 and @xmath101 channels ( left and right panel , respectively ) , and for @xmath182 gev ( fig . [ fig : qares100 ] collects the same results for @xmath183 gev ) . here and in the following we shall define the resummed cross sections as in eq . ( [ ppcsres ] ) , that is , we substitute their @xmath168 terms with the exact nlo result , using the same choice of renormalization and factorization scales . in this way our results are exact up to ( and including ) @xmath168 , and include the nll resummation of terms of @xmath184 and higher . we compare the fixed - order results ( dashed lines ) with the resummed results . for these we provide both the @xmath175 and @xmath177 prescriptions . note that , even at the level of resummed cross sections , the difference between the @xmath175 and @xmath177 results are rather small , in particular for the @xmath185 channel . in this section we present some results for the full hadronic cross sections . the main points we intend to highlight are : 1 . the size of the nll corrections , relative to the nlo contributions ; 2 . the scale dependence at nll order . our goal here is to explore the pure effects of resummation at higher orders . therefore we shall neglect in this section all production channels which are not improved by the resummation corrections considered in this work . this includes all processes which first appear at , such as @xmath186 and @xmath187 , as well as all contributions proportional to a parton @xmath188 fragmentation function , as discussed in the previous section . these terms will however be included , at fixed nlo , in our comparison with experimental data , performed in the next section . to be more specific , we list here the classes of diagrams included at nlo , in addition to the lo processes @xmath189 and @xmath190 ( the possible replacement of quarks with antiquarks in all these cases is understood ) : [ eq : procsin ] q g g , q|q q|q , q|q q|q , q|q gg . for the third set we only include the diagrams proportional to @xmath191 , since the part of the amplitude describing the photon emission from the final - state quarks ( which is by itself gauge invariant ) , can not be considered as a correction to any tree - level process . as a default set of parton densities we shall use the cteq4 m set described in ref . @xcite . for the purposes of the present study , no significant change is obtained if different sets are used . figures [ fig : deltaqa ] and [ fig : deltatot ] present the ratios : [ deltadef ] , , , where , for simplicity , we indicated here with @xmath192 the differential distribution @xmath193 . for each channel we present the results using both the @xmath175 ( solid lines ) and @xmath177 ( dashed lines ) prescriptions . we also show the dependence on the choice of renormalization and factorization scales , which we take equal , and varying within the set @xmath194 . in this section we shall always keep the fragmentation scale @xmath33 , necessary for the factorization of the singularities from final - state collinear photon emission , equal to . note that the size of the resummation effects is larger for the larger scales , contrary to the behaviour of the scale dependence of the nlo cross section . this suggests that the scale dependence of the resummed cross section will be reduced relative to that of the nlo results . the scale dependence of the resummed cross section ( @xmath195 contributions ) , compared to the nlo one , is given in fig . [ fig : scdepl0 ] , for the @xmath175 case . the same result for @xmath177 is given in the left panel of fig . [ fig : scdepl ] . we plot the distributions as a function of the beam energy ( @xmath196 ) for the fixed value of @xmath197 gev . different values of @xmath196 , therefore , probe different ranges of , as indicated by the upper labels on the plots . note the significant reduction in scale dependence , more marked in the @xmath175 case . to display the importance of the inclusion of the constant @xmath198 terms , we show the same scale - dependence plot with @xmath199 in the right panel of fig . [ fig : scdepl ] . while the scale sensitivity is slightly worse than in the cases with @xmath200 , there is still an important improvement over the nlo behaviour . similar results , for @xmath201 gev , are shown in figs . [ fig : scdeph0 ] and [ fig : scdeph ] . the general features of these distributions are similar to those of the plots for @xmath197 gev . small violations of -scaling can be observed between @xmath197 and @xmath202 gev , due to the evolution of the coupling constant and of parton densities . we also explored the independent renormalization- and factorization - scale dependence of our calculations . the large size of this dependence at nlo was stressed already in refs . the results , for @xmath17 collisions at @xmath203 and 10 gev , are shown in figs . [ fig : scdep5 ] and [ fig : scdep10 ] , respectively . with the exception of the renormalization - scale dependence at 5 gev , a significant improvement in the stability of the results , relative to the dependence at lo and nlo , is observed in all cases . [ cols="^,^,^,^,^,^,^,^",options="header " , ] the convergence of the higher - order corrections is displayed in table [ tab : toppert ] . the last column includes the sum of all contributions of order @xmath204 and higher , performed using the _ minimal prescription _ of eq . ( [ mppp ] ) . the fixed - order terms do not have any ambiguity due to the choice of the contour for the mellin transformation in eq . ( [ mppp ] ) . the contribution from the @xmath181 channel converges very rapidly . in the case of the @xmath101 channel the convergence is slower , in particular at the larger values of @xmath8 , but even at @xmath205 gev the size of the resummed contributions beyond order @xmath204 is only of the order of 10% of the total . this supports the validity of the _ minimal prescription _ , since the truncated resummed expansion converges to it very smoothly . we present in this section a comparison between our calculations and the results of the two most recent measurements of fixed - target direct - photon production . e706 @xcite studied photons produced in @xmath206 collisions at @xmath207 gev , covering the centre - of - mass rapidity range @xmath208 to indicate the value of photon rapidity in the hadron - hadron centre - of - mass frame . ] . the data span the region between 3.5 and 11 gev , approximately . this corresponds to @xmath209 . ua6 @xcite studied photons produced in @xmath210 collisions at @xmath211 gev , over the rapidity range @xmath212 . the region extends between 4 and 7 gev ( @xmath213 ) , approximately . what follows is not meant to be a systematic phenomenological study , but will serve as a benchmark to assess the impact of the resummation effects in realistic experimental conditions . all the calculations in this section have been done assuming @xmath214 . for recent complete studies of all available data , done using the fixed - order nl calculations , see refs . @xcite . to compare our resummed predictions with actual data , two additional things need to be done : inclusion of the @xmath125-suppressed contributions , and inclusion of realistic experimental cuts . the class of processes for which we evaluated resummation corrections in section [ secnotat ] provides the dominant contribution to the production rate in realistic experimental configurations . on the left - hand side of fig . [ fig : e706frag ] we plot the relative contribution , evaluated at nlo , of the processes with a @xmath101 , @xmath181 , @xmath215 and @xmath216 initial state for the e706 experimental configuration . as one can see , the sum of @xmath101 and @xmath181 accounts for 90% of the overall rate , independently of @xmath217 . on the right - hand side of fig . [ fig : e706frag ] we plot the rate of the direct contributions , relative to the sum of direct and fragmentation , for each given channel . in the case of the @xmath215 and @xmath216 channels we compare the absolute values of the rates , since the direct component is negative after the subtraction of the initial - state mass singularities . the comparison of the two plots in fig . [ fig : e706frag ] shows that the processes for which we are going to include resummation corrections account at nlo for a fraction of the total rate between 70 and 90% , in the range 412 gev . the situation is even better in @xmath210 collisions ( see fig . [ fig : ua6frag ] , obtained for the ua6 experimental configuration ) , where the to - be - resummed processes account for over 90% of the nlo rate . we therefore expect that the neglect of the resummation corrections to the @xmath218 and to the fragmentation processes is only a minor correction to the overall picture . for the present study , and in addition to the resummed predictions for the processes listed in eq . ( [ eq : procsin ] ) , we will therefore only include the fixed - order nl determination of these remaining components of the direct - photon production process . the fragmentation processes are evaluated using the nlo single - inclusive parton distributions from @xcite , convoluted with the grv photon fragmentation functions @xcite . we found very small sensitivity to the choice of the photon fragmentation functions . as anticipated above , the comparison of our results with actual data requires the inclusion of realistic detector acceptance cuts . the resummation formalism discussed so far allows the evaluation of the transverse - energy distributions integrated over the full range of rapidity for the observed photon . this approximation is technically correct , provided the measurement is performed within a finite range in rapidity including the value of @xmath219 in the collision centre - of - mass frame . this is because in the large- limit all production is concentrated at @xmath219 . to include the effect of experimental rapidity cuts , which usually do include the @xmath219 point , we therefore apply the following acceptance correction to our resummed cross sections : ^(res)(yy ) ^(res)(all y ) . the experimental configuration for most experiments of practical interest is wide enough that , for the relevant values of @xmath217 , the rapidity acceptance is very large . we show two examples in fig . [ fig : yaccept ] . one set of curves gives the ratio @xmath220 , evaluated at nlo for the e706 experimental configuration . the three curves correspond to different choices of scale @xmath38 , and show very small dependence on @xmath38 . we also checked that the dependence on the pdf set used is at the level of 1 - 2% . the acceptance loss is of the order of 25% for values around 4 gev , and becomes totally negligible for @xmath221 gev . the other set gives the ratio @xmath222 , evaluated at nlo for the ua6 experimental configuration . the acceptance loss is here more significant , due to the tight cut at negative rapidity . we present our prediction for the e706 data in fig . [ fig : e706comp ] has been rescaled by a 1.09 factor , to account for nuclear corrections to the @xmath206 process @xcite . ] . notice the significant reduction in scale dependence obtained when going from the fixed - order nl calculation to the resummed result . this scale reduction is particularly evident at high , where the resummation effects are more important . notice that while at low the band with the resummed prediction is all contained within the nlo uncertainty band , at high the nll result becomes larger relative to nlo for all the displayed scale choices . the plot shows a reasonable agreement between data and theory at large , indicating that no additional significant contribution is required in this region . the large disagreement between data and theory already present at nlo @xcite is still present , as no net increase is obtained from the resummation contributions . their only effect is to reduce the scale dependence . a similar picture emerges from the comparison of the theory with the @xmath210 ua6 data . this is shown in fig . [ fig : ua6comp ] . the disagreement between data and theory at low is however much less dramatic here than in the case of e706 . the extent to which these low- discrepancies can be removed by the inclusion of non - perturbative effects such as an intrinsic @xmath12 remains to be understood , as the global consistency of the different data sets is not very compelling @xcite . we presented in this paper a numerical study of the impact of resummation corrections , at the next - to - leading logarithmic level , on the transverse energy distribution of direct photons produced in hadronic collisions . we dealt with the resummation of the @xmath223 sudakov logarithms studied theoretically in refs . @xcite . as a result , this work is mostly of relevance for typical fixed - target photon production . the current prompt - photon data from the high - energy hadronic colliders cover in fact the region @xmath224 , where sudakov effects are negligible . we showed that the inclusion of higher - order sudakov corrections improves significantly the factorization and renormalization scale dependence , relative to what observed in the fixed - order nl calculations . even when the scales are varied independently , the uncertainty from scale variations is significantly reduced . as a result of the reduced scale dependence , the overall size of the resummation contributions depends significantly on the chosen scale . in general , however , the resummed cross sections for different scale choices have values contained within the nlo uncertainty band for values up to 0.5 , and exceed the upper side of the nlo band by large factors when approaches 1 . still , our resummation corrections turn out to be much smaller , at least in the range of existing data , than the effects induced at large by some implementations of intrinsic @xmath12 effects , as discussed in refs . @xcite . in these papers the effect of intrinsic-@xmath12 corrections was evaluated to be as large as factors of 3 and more , for the whole range . we believe that this is related to the absence in the intrinsic-@xmath12 models of the appropriate sudakov suppression due to the presence of the hadronic system recoiling against the photon ( represented at ll in our formalism by the negative terms in the exponents of eqs . ( [ dlresqg ] ) and ( [ dlresqq ] ) ) . in our work we did not include resummation corrections to the fragmentation processes . we proved that , in @xmath17 collisions , the large-@xmath60 behaviour of the corrections to the @xmath225 processes is formally similar to that of the corrections to the leading one , @xmath226 . these corrections are therefore not suppressed when @xmath60 increases towards larger values . whether they can be neglected or not , is therefore a pure matter of numerics . we showed that their contribution is not dominant in the regions of experimental interest , and limited ourselves to including them at the fixed next - to - leading order . as discussed in sec . [ fragcomp ] , we have no reasons to believe that the resummation corrections are any larger for the fragmentation processes than for the @xmath101 channel . a more quantitative study of these statements , and a complete phenomenological assessment of the comparison between theory and the current sets of data in view of the results presented in this paper , will be the subject of future work . in this appendix we recall ( see ref . @xcite ) the explicit expressions of the various factors that contribute to the resummed cross sections in eqs . ( [ gammaresqq ] ) and ( [ gammaresqg ] ) . we use the customary notation for the colour factors in @xmath227 qcd , namely , @xmath228 and @xmath229 . the @xmath60-moments of the lo partonic cross sections in eqs . ( [ siqqgamma ] , [ siqggamma ] ) are [ qgamman ] _ q|qg , n^(0 ) = e_q^2 ( 2+n ) , [ ggamman ] _ qgq , n^(0 ) = _ |qg , n^(0 ) = e_q^2 ( 7 + 5n ) , where @xmath230 is the euler @xmath231-function . the formulae for the nll functions @xmath89 in the exponent of the radiative factors in eq . ( [ deltanll ] ) are the following [ g2qq ] g_q|q^(2 ) ( , ; ) & = & ( 2c_f - c_a ) h^(2 ) ( ) + 2 c_a h^(2)(/2 ) + & + & 2 ( 1 - 2 ) + ( 1- ) - + & + & \ { + } , [ g2qg ] g_qg^(2 ) ( , ; ) & = & c_a h^(2 ) ( ) + 2 c_f h^(2)(/2 ) + & + & 2 ( 1 - 2 ) + ( 1- ) - + & + & \ { + } , where @xmath232 is the euler number and @xmath233 are the first two coefficients of the qcd @xmath87-function [ betass ] b_0 = , b_1 = . the auxiliary function @xmath234 that appears in eqs . ( [ g2qq ] , [ g2qg ] ) has the following expression [ htnll ] h^(2 ) ( ) = - ( 1 - 2 ) - , where the coefficient @xmath235 is given by [ kcoef ] k = c_a ( - ) - t_r n_f . the first - order coefficients @xmath104 and @xmath105 of the @xmath60-independent functions in eq . ( [ cgamma ] ) are c_q|q^(1)(q^2/^2;q^2/_f^2 ) & = & ^2 ( 2 c_f - c_a ) + - ( 2 c_f - c_a ) 2 + [ cqq1coef ] & + & k - k_q + ( 2 c_f - c_a ) + ( 2 c_f - c_a ) ^2 2 + & - & ( 2 c_f - c_f ) - b_0 , c_qg^(1)(q^2/^2;q^2/_f^2 ) & = & ^2 ( c_f + c_a ) + - ( c_f - 2 c_a ) 2 + [ cqg1coef ] & - & k_q + ( 2 c_f + 19 c_a ) + c_f ^2 2 + & - & ( ( c_f + c_a ) - c_f - b_0 ) - b_0 , where [ kcoeffq ] k_q = ( - ) c_f , and the coefficient @xmath235 is given in eq . ( [ kcoef ] ) . note that the ll functions @xmath88 are given in eq . ( [ g1fun ] ) . thus , the formulae presented in this appendix complete all the ingredients that are necessary to evaluate the resummed cross sections with nll accuracy . 99 # 1#2#3sov . # 1 * ( 19#2 ) # 3 # 1#2#3nucl . phys . * b#1 * ( 19#2 ) # 3 # 1#2#3phys . rep . * # 1 * ( 19#2 ) # 3 # 1#2#3phys . lett . * b#1 * ( 19#2 ) # 3 # 1#2#3phys . # 1 * ( 19#2 ) # 3 # 1#2#3phys . rev . * d#1 * ( 19#2 ) # 3 # 1#2#3phys . rev . * d#1 * ( 19#2 ) # 3 # 1#2#3z . phys . * c#1 * ( 19#2 ) # 3 # 1#2#3eur . j. * c#1 * ( 19#2 ) # 3 # 1#2#3comm . # 1 * ( 19#2 ) # 3 # 1#2#3#4#1 * # 2 * ( 19#3 ) # 4 # 1#2#3_ibid . _ * # 1 * ( 19#2 ) # 3 # 1 hep - ph/#1 # 1 hep - ex/#1 d. adams et al . , e704 coll . , ; + c. de marzo et al . , na24 coll . , ; + a. angelis et al . , r108 coll . , ; + a. angelis et al . , r110 coll . , ; + e. anassontzis et al . , r806 coll . , ; + t. akesson et al . , r807 coll . , ; + g. ballocchi et al . , ua6 coll . , ; + m. bonesini et al . , wa70 coll . , ; + g. alverson et al . e706 coll . , ; + j. badier et al . , ; + m. mclaughlin et al . , e629 coll . . l. apanasevich et al . , e706 collaboration , . g. ballocchi et al . , ua6 collaboration , . p. aurenche et al . , . w. vogelsang and a. vogt , . martin , r.g . roberts , w.j . stirling and r.s . thorne , , and references therein . lai et al . , cteq collaboration , preprint msu hep903100 , , and references therein . j. huston et al . , . f. abe et al . , cdf collaboration , . b. abbott et al . , d@xmath7 collaboration , , submitted to phys . j. huston et al . , , ( ) . l. apanasevich et al . , . p. aurenche et al . , preprint lapth70498 , . collins and d.e . soper , ; + j.c . collins , d.e . soper and g. sterman , . s. catani , m. ciafaloni and f. hautmann , , ; + j.c . collins and r.k . ellis and d.a . e. laenen , g. oderda and g. sterman , , ( ) . s. catani , m.l . mangano and p. nason , jhep 9807 ( 1998 ) 024 , ( ) . s. catani , m.l . mangano , p. nason and l. trentadue , , , ( ) . p. aurenche , r. baier , a. douiri , m. fontannaz and d. schiff , ; + p. aurenche , r. baier , m. fontannaz and d. schiff , . f. aversa , p. chiappetta , m. greco and j.ph . guillet , . g. sterman , . s. catani and l. trentadue , . s. catani and l. trentadue , . r. bonciani , s. catani , m.l . mangano and p. nason , , ( ) . contogouris , n. mebarki and s. papadopoulos , int . j. mod a5 * ( 1990 ) 1951 ; + a.p . contogouris and s. papadopoulos , mod . * a5 * ( 1990 ) 901 . gordon and w. vogelsang , . s. catani , g. turnock , b.r . webber and l. trentadue , ; + s. catani , l. trentadue , g. turnock and b.r . lai et al . , . m. glck , e. reya and a. vogt , , erratum .

the central idea behind the field of spintronics is to use the electron s spin degree of freedom , not its charge , for information storage and processing.@xcite since the energy required to generate magnetic fields scales unfavorably at small length scales , electrical mechanisms for the manipulation and detection of magnetic moments are crucial for successful spintronics applications . the spin transfer torque@xcite has been established as a reliable effect to manipulate the magnetization of a thin ferromagnetic layer with a spin - polarized current.@xcite whereas injection from a ferromagnet was used as a spin - polarized current source in the original realization , devices based on spin - orbit coupling have also been realized.@xcite an extreme form of spin - orbit coupling exists in the surface states of topological insulators ( tis):@xcite these states have complete spin - momentum locking , _ i.e. _ , the electron spin and its direction of motion are perfectly correlated . one consequence of the spin - momentum locking is that a surface charge current in a ti is automatically spin - polarized . recent proposals have concentrated on exploiting this effect to control the dynamics of thin magnetic layers in the proximity of the surface of a three - dimensional ti@xcite or magnets coupled to the edge of a two - dimensional ti.@xcite experiments on metallic magnetic layers in contact to a ti have reported a spin - transfer torque exceeding the values found in non - topological spin - orbit materials.@xcite in this paper , we investigate the possibility to use the spin - polarized surface currents of a topological insulator to control the magnetic moment of a molecular magnet adsorbed on the surface of a three - dimensional topological insulator or at the edge of a two - dimensional topological insulator . like a ferromagnet , a molecular magnet has degenerate magnetic ground states , separated by an energy barrier,@xcite although in the case of a molecular magnet the barrier is microscopic , not macroscopic , which leads to a finite relaxation time of the molecule s magnetic moment . of particular interest are the so - called `` single - molecule magnets'',@xcite which consist of a magnetic core with a spin @xmath0 , shielded from the environment by a ( typically ) organic ligand . for single - molecule magnets magnetic lifetimes of several years have been reported for temperatures below the `` blocking temperature '' set by the anisotropy barrier between the two magnetic ground states.@xcite single - molecule magnets have been shown to preserve their magnetic properties , including long magnetic lifetimes , when adsorbed on conducting surfaces.@xcite the mechanism by which an applied electrical current at the surface of a ti can be used to switch the orientation of a molecular magnet is best illustrated using the example of a localized spin @xmath1 coupled to the edge of a two - dimenstional topological insulator.@xcite the electronic state at the edge of a two - dimensional ti is helical , so that electrons moving in opposite directions have opposite spin.@xcite hence , a particle current to the right is polarized as `` spin up '' , see fig . [ fig : sketch ] . by spin conservation , backscattering of a right - moving electron into a left - moving state requires the localized spin to flip from down to up , so that a single backscattering event is sufficient to polarize the spin @xmath1 in the `` up '' state . if the applied current is sufficiently large , the pauli principle forbids backscattering of left - moving electrons which would be accompanied by a spin flip in the opposite direction , so that the spin @xmath1 remains in the `` up '' state as long as the current is being applied . weakly coupled to the helical edge of a two - dimensional topological insulator . backscattering of a right - moving electron is accompanied by a flip to the spin `` up '' state of the localized spin , leaving it in a fully polarized state after a single backscattering event . for a sufficiently large applied current the reverse scattering process , which would return the localized spin to the `` down '' state , is strongly suppressed.,scaledwidth=45.0% ] this simple picture needs to be refined for molecules with a higher spin and for molecular magnets adsorbed to the two - dimensional surface of a three - dimensional ti . unlike the spin @xmath1 of the example of fig . [ fig : sketch ] , a higher - spin molecule comes with its own anisotropy axis , and the argument based on spin conservation no longer applies in this simple form if the molecule s anisotropy axis is not aligned with the spin quantization axis of the ti edge state . simple spin conservation arguments can not be applied to a molecular magnet on the surface of a three - dimensional topological insulator either , because in that case there is no unique spin quantization axis , as electrons can propagate in an arbitrary direction along the surface . as we will show below , in this generic situation current - induced switching of molecular magnet on a ti surface is not perfect , although appreciable polarizations can be achieved even for randomly oriented molecules . an interesting observation is that , unlike in the example above , for a molecular magnet on the surface of a three - dimensional ti the current - induced polarization is not a monotonously increasing function of the applied current , but has a maximum at intermediate current densities . this article is organized as follows : in sec . [ sec:2dti ] we consider a molecular magnet at the edge of a two - dimensional ti , expanding the example of fig . [ fig : sketch ] to higher - spin molecules . we calculate the probability with which current - induced switching takes place , the current - induced switching rate , and the zero - current relaxation rate that arises from the exchange coupling of the molecule s spin to the ti edge . in sec . [ sec:3dti ] we consider the same questions for a molecular magnet on the surface of a three - dimensional topological insulator , which is the more likely candidate for an experimental realization . we conclude with a brief outlook in section [ sec : conclusion ] . the main text mainly focuses on molecules with integer spin @xmath2 . the case of half - integer spin , which requires a technically more demanding analysis , is discussed in the appendix . to introduce our notation and to provide a reference for further calculations , we start by considering a localized moment of spin @xmath3 exchange - coupled to the helical edge of a two - dimensional topological insulator . we assume that there is only a single localized spin coupled to the edge and that the temperature is large enough that the kondo effect can be neglected . in that case transitions between different spin states can be described using rate equations . the coupling between the edge and the magnetic moment is described by the exchange hamiltonian @xmath4 where @xmath5 labels the coordinate along the helical edge , @xmath6 is the velocity of the helical edge state , and @xmath7 is the dimensionless exchange coupling . because of spin - momentum locking of the edge states , backscattering of edge electrons from the molecular magnet involves simultaneous flips of the spin of the edge electrons and of the localized spin of the molecular magnet . we fix the spin @xmath5 axis along the direction of the spin quantization axis for the edge - state electrons , @xmath8 so that right - moving electrons at the edge have spin up and left - moving electrons have spin down . we note that the spin @xmath5 axis need not be in the same direction as the laboratory @xmath5 axis . ( in fact , for the edge of a two - dimensional topological insulator , the spin quantization axis is commonly taken to be perpendicular to the plane of the ti.@xcite ) the states @xmath9 in the helical edge are labeled by their energy @xmath10 and the propagation direction @xmath11 , where we take @xmath12 for right - moving electrons ( positive @xmath5 direction ) and @xmath13 for left - moving electrons . ( note that in this notation @xmath14 merely parameterizes the energy ; its magnitude @xmath15 equals the magnitude of the momentum , but the sign of @xmath14 is that of the energy @xmath16 , not of the momentum . ) the transition rates @xmath17 between the spin states are calculated from fermi s golden rule . the rate @xmath18 for transitions from the spin up state @xmath19 to the spin down state @xmath20 is @xmath21 \nonumber \\ & & \mbox { } \times |\langle + , \downarrow|h_{\rm ex}| -,\uparrow \rangle|^2 \delta(\varepsilon_k - \varepsilon_{k ' } ) \nonumber \\ & = & \frac{j^2}{2 \pi \hbar } \int\limits_{-\infty}^{+\infty } d \varepsilon n_{-}(\varepsilon ) ( 1 - n_{+}(\varepsilon ) ) . \label{eq : gamma1}\end{aligned}\ ] ] ( the matrix elements of @xmath22 do not depend on @xmath14 and @xmath23 , which is why we have suppressed @xmath14 , @xmath23 in our notation . ) similarly , the rate @xmath24 at which transitions from the spin down state to the spin up state take place reads @xmath25 the distribution function at temperature @xmath26 and chemical potential @xmath27 in the presence of an edge current @xmath28 is @xmath29 as one can easily verify from the relation @xmath30 $ ] . for definiteness , we will assume throughout that the particle current is to the right , @xmath31 . performing the integrations in eqs . ( [ eq : gamma1 ] ) and ( [ eq : gamma2 ] ) then gives @xmath32 where we abbreviated @xmath33 for high applied currents , @xmath34 , scattering processes in which a left - moving electron is scattered into a right - moving one and , correspondingly , the rate @xmath18 are exponentially suppressed . on the other hand , for low applied currents @xmath35 , scattering of thermally excited charge carriers dominates the spin switching rates , and the application of the current @xmath36 only gives a slight asymmetry . in the absence of coherences between the spin states @xmath37 and @xmath38 the probabilities @xmath39 and @xmath40 to find the spin a state of spin @xmath41 can be solved from the stationary solutions of the rate equation @xmath42 ( in principle , one needs a density matrix to describe possible coherent superpositions of the degenerate states @xmath37 and @xmath38 , see , _ e.g. _ , refs . and the appendix . however , such coherences do not occur in the present case , in which there is no external magnetic field and a single spin quantization axis for the conduction electrons at the ti edge and the localized spin . ) substituting the rates ( [ eq : gamma12 ] ) one immediately finds @xmath43 the same result also follows from the observation that the coupling to the conduction electrons at the ti edge effectively amounts to a zeeman shift @xmath44 between the spin up and spin down states of the molecule . we conclude that the application of a current @xmath45 leads to a complete polarization of the molecule s spin . although the application of a current through the helical edge of a two - dimensional topological insulator causes a polarization of a spin @xmath1 exchange - coupled to the edge state , the induced magnetic moment quickly disappears as soon as the current is switched off . in our model , in which the coupling to the helical edge state is the only source of relaxation , the corresponding relaxation rate @xmath46 is given by the low - current limit of eq . ( [ eq : gamma12 ] ) , @xmath47 ( a ) for a spin-@xmath48 molecule and ( b ) for a spin-@xmath49 molecule.,scaledwidth=45.0% ] longer relaxation times in the absence of an applied current requires molecules with a higher spin @xmath2 . in this case , magnetic anisotropy , @xmath50 imposes an energy barrier between states with maximal and minimal spin @xmath51 ( measured along the anisotropy axis @xmath52 ) , see fig . [ fig : sketch_spin1 ] , which , if @xmath53 is lower than the barrier energy , leads to strongly enhanced lifetimes . ( see below for a quantitative estimate for our model . ) in our analysis , we have adopted a simple easy - axis anisotropy strength @xmath54 ( which we take to have the units of frequency ) with @xmath55 , although our considerations also carry over to other forms for the anisotropy energy in which there are two different minima . ( the case @xmath56 is for easy - plane anisotropy , which does not have two separated energy minima and will not be considered here . ) we label the spin states with the quantum number @xmath57 , @xmath58 , , @xmath2 for the spin component @xmath59 along the @xmath52-direction . ( for @xmath3 , @xmath60 takes the values @xmath61 , corresponding to the notation `` @xmath62 '' and `` @xmath63 '' in the previous subsection . ) if the anisotropy axis is aligned with the spin quantization axis for the helical edge state , the analysis of the previous subsection immediately carries over , and one finds @xmath64 . \label{eq : gammass}\ ] ] for the probabilities @xmath65 one then finds @xmath66 where the prefactor is fixed by the normalization condition @xmath67 . if the temperature is much smaller than the anisotropy energy , which is a condition that we will assume throughout this article , the rate equations can be solved directly , using detailed balance , to give the zero - current switching rate @xmath68 between the two magnetic ground states @xmath69 and @xmath70 , @xmath71 if @xmath2 is integer and @xmath72 if @xmath2 is half integer . the application of a current @xmath36 slightly larger than @xmath73 leads to a quick and complete spin polarization of the molecule . the switching rate @xmath74 can be estimated as the inverse of the sum of inverse transition rates @xmath75 , which gives @xmath76 for large @xmath2 . a smaller current @xmath77 also polarizes the molecule , but since thermal activation is still needed in the switching process , the time required to reach the polarized state is still large ( though much shorter than the zero - current lifetime ) , @xmath78 if @xmath2 is integer and @xmath79 if @xmath2 is half integer . if @xmath52 is not aligned with the @xmath5 axis , the expression for the rates @xmath80 involves the matrix elements @xmath81 where @xmath82 and @xmath83 are the polar angles corresponding to the anisotropy axis @xmath84 . this gives @xmath85 \sin^2 \theta + ( 1 \mp \tau \cos \theta)^2 g[\tau { i}\mp d ( 2 s \pm 1 ) ] \right\}. \label{eq : gammaanis}\end{aligned}\ ] ] one verifies that the zero - current rates are the same as those of eqs . ( [ eq : gammass ] ) , independent of the direction @xmath52 of the anisotropy axis , so that eqs . ( [ eq : gammaswitch0even ] ) and ( [ eq : gammaswitch0odd ] ) still apply . because of the degeneracy of the states @xmath37 and @xmath38 the rate equation ( [ eq : master_time - dep ] ) is not sufficient to describe transitions between these two states and , instead , one has to use a master equation for the reduced @xmath86 density matrix that is able to account for coherent superpositions of the states @xmath37 and @xmath38 . for that reason we here restrict ourselves to the case of integer @xmath2 , for which this complication does not occur . a complete analysis and final expressions for half - integer @xmath2 are given in the appendix . we first discuss the case of an `` intermediate '' current @xmath87 , for which the transition rates ( [ eq : gammaanis ] ) can be approximated as @xmath88 the large difference between `` upstream '' transitions to higher - energy spin states and `` downstream '' transitions to lower - energy spin states ensures that the molecule will predominantly be in one of its two magnetic ground states @xmath69 and @xmath70 . the presence of the applied current breaks the symmetry between these two ground states and leads to a preferred population of the state @xmath69 if @xmath89 and @xmath70 if @xmath90 , although there is no longer a perfect polarization for generic @xmath82 , @xmath91 switching between the magnetic ground states involves thermal activation , which is why the switching rate between the magnetic ground states is exponentially long in @xmath92 , although the rate is parametricaly larger than the spontaneous relaxation rate at zero current . defining @xmath74 as the switching rate from @xmath70 to @xmath69 we have @xmath93 where @xmath68 is the zero - current relaxation rate of eq . ( [ eq : gammaswitch0even ] ) . the switching rate for the inverse process is obtained upon replacing @xmath94 . ) for @xmath95 and for angles @xmath96 , @xmath97 , @xmath98 , @xmath99 , @xmath100 , @xmath101 , @xmath102 , @xmath103 and @xmath104.,scaledwidth=45.0% ] the difference between the case @xmath96 and generic @xmath82 is most pronounced in the `` high - current '' limit @xmath105 , in which the rates @xmath80 in eq . ( [ eq : gammaanis ] ) are dominated by the terms proportional to @xmath106 $ ] . approximating @xmath107 \approx { i}$ ] for @xmath108 , we find @xmath109 for the probabilities @xmath65 this gives immediately @xmath110 for @xmath82 close to zero this still gives a narrow distribution around @xmath111 , but perfect polarization is not reached if @xmath112 independent of the magnitude of the current @xmath36 . examples of the distribution ( [ eq : ps2d ] ) are shown in fig . [ fig : ps2d ] for @xmath95 and for several representative angles @xmath82 . figure [ fig : mean2d ] shows the mean values @xmath113 of the spin component along the anisotropy axis and @xmath114 of the spin component along the current direction as a function of the angle @xmath82 for two different values of the total spin @xmath2 . except for @xmath100 , the application of a current always results in a net spin in the current direction . for a randomly oriented molecule , the average moment in the current direction approaches @xmath115 in the limit of large @xmath2 . without a perfect current - induced polarization , we should define the current - induced `` switching rate '' @xmath74 as the rate at which the steady - state probability @xmath65 is approached , starting from @xmath70 ( if @xmath89 ) . the estimate ( [ eq : gammaswitch2dhigh ] ) for the switching rate @xmath74 obtained above for @xmath52 parallel to the @xmath5 axis remains valid as a good order - of - magnitude estimate for the case of a general orientation of the anisotropy axis in the high - current regime . along the direction of the anisotropy axis ( solid curves ) and the mean spin component @xmath116 in the current direction ( dashed ) , as a function of the angle @xmath82 , for @xmath117 ( left panel ) and @xmath118 ( right panel).,scaledwidth=45.0% ] the main result of the previous section is that the current - induced polarization of a molecular magnet coupled to the edge of a two - dimensional ti is complete only if the molecule s anisotropy axis is aligned with the spin quantization axis for the ti edge state , whereas the polarization is incomplete but generally nonzero for arbitrary orientations of the molecule . the analysis was simplified by the fact that the edge of a two - dimensional ti has a unique propagation direction and , hence , a unique spin quantization axis . this is the main difference with the case of a molecule coupled to the surface of a three - dimensional ti : electrons at the surface of a three - dimensional ti can propagate in all directions in the plane of the surface , even in the presence of a large applied current . hence , there is not a unique spin quantization axis , and there will always be surface - state electrons with a spin that is not aligned with the anisotropy axis of the molecular magnet . nevertheless , as we show below , even for a molecule adsorbed on the surface of a three - dimensional ti the application of an electrical current can lead to a complete spin polarization , provided the molecule s anisotropy axis is aligned with the current direction and the magnitude of the applied current is appropriately chosen . as in the case of a two - dimensional ti , the spin polarization of the molecular magnet is not complete ( but generally nonzero ) if the molecule s magnetic anisotropy axis is not aligned with the current direction . plane . a current is applied the positive @xmath5 direction.,scaledwidth=45.0% ] to keep the notation close to that of the previous section , we take the ti surface to be the @xmath119 plane and take the direction of ( particle ) current flow to be the positive @xmath5 direction , see fig . [ fig : sketch3d ] . the electrons at the ti surface are then described by the hamiltonian @xmath120 we write the current density @xmath121 as @xmath122 where @xmath123 is the fermi wavenumber and @xmath36 has the dimension of current . for the exchange hamiltonian we take @xmath124 where , as in the previous section , a prefactor has been included to make the strength @xmath7 of the exchange interaction dimensionless . the hamiltonian @xmath125 for the magnetic anisotropy of the magnetic molecule is the same as in the previous subsection , see eq . ( [ eq : hs ] ) . the eigenstates of @xmath126 are labeled by the energy @xmath10 and by the angle @xmath127 of the propagation direction in the @xmath119 plane , such that @xmath128 corresponds to the positive @xmath5 direction . we assume that the chemical potential @xmath129 , which ensures that thermally excited carriers as well as carriers contributing to the current flow remain well away from the dirac point . in that case the distribution function in the presence of a particle current density @xmath130 is @xmath131 equation ( [ eq : n3d ] ) both follows from a solution of the boltzmann equation and maximizes the entropy under the constraint of a fixed current density @xmath132 . for the calculation of the transition rates between different spin states we need the matrix elements @xmath133 , \label{eq : matrixelements3d}\end{aligned}\ ] ] where we abbreviated @xmath134 . equation ( [ eq : matrixelements3d ] ) generalizes eq . ( [ eq : matrixelements2d ] ) to the case of a surface state . as in eq . ( [ eq : matrixelements2d ] ) , @xmath82 and @xmath83 are the polar angles marking the direction of the anisotropy axis . for the transition rates we then find , again assuming @xmath135 , @xmath136 , \label{eq : gammageneral}\end{aligned}\ ] ] where the function @xmath137 was introduced in eq . ( [ eq : g ] ) . we will now analyze these rates and the resulting probabilities @xmath65 in the regimes of low current densities , @xmath138 , intermediate current densities , @xmath139 , and high current densities , @xmath140 . ( for @xmath3 no intermediate regime exists and the high - current regime is defined as @xmath34 . ) for low current densities @xmath138 the expressions ( [ eq : gammageneral ] ) for the transition rates simplify to @xmath141 the resulting zero - current switching rate @xmath68 is given by the same expression as in the two - dimensional case , see eqs . ( [ eq : gammaswitch0even ] ) and ( [ eq : gammaswitch0odd ] ) . for intermediate current densities , @xmath142 , the exponential suppression of the transition rates to higher - energy spin states is reduced , whereas the transition rates into the lower - energy states remains approximately as in the low - current regime . for a quantitive analysis we restrict ourselves to the case of integer spin @xmath2 , referring to the appendix for a discussion of half - integer @xmath2 . in the intermediate - current regime the transition rates can be approximated as @xmath143 as in the two - dimensional case , the applied current breaks the symmetry between states with positive and negative spin @xmath60 , leading to an imbalance between the populations of the ground states @xmath69 and @xmath70 , @xmath144 the population of the excited states @xmath145 remains exponentially small in @xmath146 . the current - induced switching rate from the state @xmath70 into @xmath69 is @xmath147^s \nonumber \\ & & \mbox { } \times e^{2 \pi \hbar s { i}/k_{\rm b } t}. \label{eq : gammaswitchj3deven}\end{aligned}\ ] ] the inverse rate is given by the same expression , but with @xmath148 replaced by @xmath149 . the case @xmath150 that the molecule s anisotropy axis is aligned with the current direction requires to be considered separately . in this case the leading approximation of eq . ( [ eq : gamma3dintermediate ] ) vanishes for the `` upstream '' rate @xmath151 with @xmath152 and subleading contributions in the small parameter @xmath153 must be taken into account . one finds @xmath154 if @xmath152 . asymptotically , for @xmath34 [ but still @xmath155 , this leads to a complete spin polarization of the molecule , @xmath156 [ the probability @xmath157 vanishes @xmath158 ; all other probabilities are exponentially small in @xmath146 . ] the perfect polarization can be understood if one observes that the surface - state electrons responsible for the transitions to higher - energy spin states are predominantly electrons moving in the positive @xmath5 direction , which are then backscattered to the negative @xmath5 direction . these electrons have a unique spin quantization axis , and their scattering leads to a well - defined change of the molecule s spin state @xmath159 . the switching rate from the state @xmath70 to @xmath69 is given by eq . ( [ eq : gammaswitchj3deven ] ) with @xmath150 . the rate for the inverse process @xmath160 vanishes in the limit @xmath161 . , @xmath162 , and @xmath163 for a spin-@xmath48 molecule on the surface of a three - dimensional topological insulator to be in the corresponding spin state as a function of @xmath164 we have set @xmath165 . the inset shows the maximum value of the probability @xmath166 for a molecule with @xmath117 as a function of @xmath167.,scaledwidth=45.0% ] remarkably , the complete polarization is lost again when the current is increased further . if @xmath105 , not only electrons moving in the positive @xmath5 direction but essentially all electrons that contribute to the current @xmath121 can scatter off the localized magnetic moment and change its spin . for high current densities the transition rates are ( for arbitrary direction of the anisotropy axis ) @xmath168 where we abbreviated @xmath169 and one quickly obtains the steady - state distribution from here , @xmath170 f_+(\theta)^{s+s } f_-(\theta)^{s - s } } { f_+(\theta)^{2 s + 1 } - f_-(\theta)^{2 s + 1}}. \label{eq : ps3dhigh}\ ] ] since @xmath171 , even for a perfectly aligned molecule no perfect polarization occurs in the limit of large currents . the probability @xmath172 to find the molecule in the maximal - spin state saturates at a value slightly above @xmath173 for molecules with large spin @xmath2 . although no perfect polarization results for current densities @xmath105 , the rate at which the steady - state distribution is approached is considerably enhanced in comparison to the intermediate current - density regime , see eq . ( [ eq : gammaswitch2dhigh ] ) . as an illustration , we show the probabilities @xmath65 for @xmath174 for a molecule with @xmath117 as a function of the applied current density for @xmath175 , see fig . [ fig : spin1 ] . notice the ( faint ) maximum of @xmath176 for @xmath177 . the inset shows that the maximum value of @xmath176 approaches unity in the limit when the anisotropy energy is much larger than the temperature , @xmath178 . figure [ fig : mean3d ] shows the mean spin components @xmath179 along the molecule s anisotropy axis and @xmath116 along the current direction as a function of the angle @xmath82 for @xmath95 , for the intermediate - current regime and for the high - current regime . as in the case of a two - dimensional ti , except for @xmath100 , the application of a current always results in a net spin in the current direction . for a randomly oriented molecule , the average moment in the current direction approaches @xmath115 in the limit of large @xmath2 . along the direction of the anisotropy axis ( solid curve ) and the mean spin component @xmath116 in the current direction ( dashed ) , as a function of the angle @xmath82 , for total spin @xmath95 . left panel : intermediate - current regime @xmath142 ; right panel : high - current regime @xmath180.,scaledwidth=45.0% ] we can apply the results obtained here to a topological insulator surface with a dilute covering of molecular magnets . for a dilute covering our single - molecule analysis can be used to obtain the net polarization of the molecular layer . assuming that the molecules have randomly orientated anisotropy axis a valid assumption if the magnetic core is shielded by an approximately spherical shell , see , _ e.g. _ , ref . the net polarization is found by averaging a single molecule s average moment over the directons of the anisotropy axis @xmath52 . the calculation is straightforward in principle , starting from the steady - state distributions ( [ eq : p3deven ] ) and ( [ eq : ps3dhigh ] ) . [ see eqs . ( [ eq : p3dodd ] ) and ( [ eq : p3dhighodd ] ) for steady - state distributions for half - integer spin @xmath2 . ] instead of reporting the resulting expressions , which are rather cumbersome , we refer to fig . [ fig : average ] , which shows the resulting net polarization per molecule as a function of @xmath2 . per molecule as a function of the total spin @xmath2 , for an ensemble of randomly oriented molecular magnets on the surface of a three - dimensional topological insulator . open circles are for the intermediate - current regime @xmath181 ; filled circles are for the high - current regime @xmath182.,scaledwidth=45.0% ] the strict spin - momentum locking at the surface of a topological insulator provides an appealing mechanism to use electric currents to switch the spin state of a molecular magnet weakly coupled to the surface . whereas based on simple spin conservation arguments one expects a complete response for a molecular magnet at the edge of a two - dimensional topological insulator if the molecule s magnetic anisotropy axis is aligned appropriately with the spin quantization axis of the surface state , the situation is more complicated for molecular magnets on the surface of a three - dimensional topological insulator and/or for an arbitrary orientation of the magnetic anisotropy axis . for a molecule on the surface of a three - dimensional topological insulator , a full current - induced polarization is achieved only if the molecule s magnetic anisotropy axis is aligned with the current direction and the magnitude @xmath183 of the current density is in an `` intermediate '' range @xmath184 , with @xmath185 the molecule s anisotropy energy and @xmath2 the magnitude of its spin . yet , even for a surface covered by molecules with randomly oriented anisotropy axes the application of a surface current should result in a net magnetic moment per molecule for all current strengths . + + the minimal current density @xmath183 required to polarize the molecule and the resulting steady - state distribution of the molecule s magnetic states depends more on properties of the topological insulator surface than properties of the molecular magnet itself . the condition for an appreciable current - induced polarization is @xmath186 , _ i.e. _ , the disortion of the fermi surface by the applied current should be large in comparison to the blurring of the fermi surface by temperature . this is a condition that is easier to achieve if the chemical potential is in the vicinity of the dirac point , so that @xmath123 is small , and a relatively small current density implies a large shift of the distribution function . the role of the anisotropy energy is to set a crossover scale @xmath187 , above which current - induced transitions allow the molecule to be in all spin states , whereas for lower current densities only the magnetic ground states @xmath69 and @xmath70 are accessible . again , the crossover between these regimes happens at lower current densities if the chemical potential is closer to the dirac point . + + details of the coupling between the molecule and the surface play a crucial role when it comes to setting the switching rate . this applies both to the zero - current spontaneous switching between the degenerate magnetic ground states @xmath69 and @xmath70 and to the current - induced changes between these states . if the temperature is below the blocking temperature , appreciable switching rates can be obtained only in the `` high - current regime '' @xmath188 , where we found the switching rate to be of order @xmath189 , where @xmath7 is the dimensionless exchange coupling . an upper limit for this rate is found by setting @xmath190 , although we expect that the presence of the organic shell will typically lead to much smaller exchange couplings for single - molecule magnets on a topological insulator surface . taking a current density @xmath183 at the lower end of the high - current regime , @xmath187 , we find an upper bound for switching times in the range of @xmath191 s for a molecule with a blocking temperature of a few k , but most likely significantly longer if @xmath7 is smaller . + + our calculations have shown that in - plane currents can be used as an effective tool to switch the orientation of the spin of molecular magnets adsorbed on the surface of a three - dimensional topological insulator , even if no complete polarization is achieved for molecules without a specially aligned magnetic anisotropy axis . we hope that our observations , together with other theoretical studies highlighting the intriguing physics of molecular magnets on ti surfaces,@xcite will motivate further experimental and theoretical research in this direction . we gratefully aknowledge support from the german research foundation ( dfg ) in the framework of the collaborative research center 658 . in the case of a molecular magnet with half - integer spin @xmath2 the states @xmath37 and @xmath38 are degenerate , which means that transitions between these states can not be described by means of a rate equation , but that they need to be described with a master equation for the ( reduced ) density matrix @xmath192 the diagonal elements @xmath193 are the probabilities to find the molecule in spin state @xmath159 , @xmath194 ; the off - diagonal elements describe coherences beween these two spin states . to allow for a unified treatment of a molecular magnet coupled to the edge of a two - dimensional ti and a molecular magnet at the surface of a three - dimensional ti , we write matrix elements of the exchange hamiltonian ( [ eq : hex2d ] ) for the two - dimensional case as @xmath195 where we use the propagation angle @xmath128 for a right - moving electron ( labeled with @xmath196 in sec . [ sec:2dti ] ) and @xmath197 for a left - moving electron ( labeled with @xmath198 ) . matrix elements of the exchange hamiltonian ( [ eq : hex3d ] ) for the three - dimensional case are written @xmath199 the rates @xmath200 , see eqs . ( [ eq : gammaanis ] ) and ( [ eq : gammageneral ] ) , are then expressed as where the double integration @xmath202 should be replaced by a double sum @xmath203 for the case of a two - dimensional ti . the explicit expressions for the @xmath204 read @xmath205 where @xmath134 , see the text below eq . ( [ eq : matrixelements3d ] ) . because of the large energy differences between spin states with different values of @xmath206 and the lack of direct transitions between the degenerate spin states @xmath159 and @xmath207 for @xmath208 , the rates ( [ eq : gammassapp ] ) are sufficient to calculate the ( rate of change of the ) probabilities @xmath65 that the molecule is in spin state @xmath159 for @xmath208 , see eq . ( [ eq : master_time - dep ] ) . on the other hand , for the degenerate spin states @xmath37 and @xmath38 we must use the full density matrix @xmath209 , see eq . ( [ eq : rhodef ] ) . the equation of motion for this density matrix has the well - known lindblad form@xcite @xmath210 \label{eq : lindblad } \\ & & \mbox { } + j^2 \int \frac{d\varphi}{2 \pi } \frac{d\varphi'}{2 \pi } g[{i}(\cos \varphi'-\cos\varphi)/2 ] \left\ { m(\varphi',\varphi ) \rho m(\varphi',\varphi)^{\dagger } - \frac{1}{2 } \left [ m(\varphi',\varphi)^{\dagger } m(\varphi',\varphi ) \rho + \rho m(\varphi',\varphi)^{\dagger } m(\varphi',\varphi ) \right ] \right\ } , \nonumber\end{aligned}\ ] ] where the projection matrices @xmath211 , and @xmath212 are defined as @xmath213 one verifies that the lindblad equation ( [ eq : lindblad ] ) simplifies to the standard rate equation ( [ eq : master_time - dep ] ) if the density matrix @xmath209 is a diagonal matrix , @xmath214 . the first term in eq . ( [ eq : lindblad ] ) contains the expectation value of the exchange hamiltonian projected onto the spin states @xmath215 . it acts as an effective magnetic field driving coherent transitions between the states @xmath37 and @xmath38 . this term plays no role for states @xmath159 with @xmath208 because it is much smaller than the anisotropy field if the exchange coupling is weak , @xmath216 . however , for transitions between the states @xmath37 and @xmath38 there is no anisotropy field , and the exchange field competes with the similarly small scattering - induced transitions between the states @xmath37 and @xmath38 . with the same notation as above we have @xmath217 where , for a molecule coupled to the edge of a two - dimensional topological insulator , the integration @xmath218 should be replaced by a summation @xmath219 . performing the angular integration gives @xmath220 which is indeed what one expects for an exchange field in the @xmath5 direction , recalling that the spin states @xmath159 are defined with respect to the anisotropy axis @xmath84 . a solution of the lindblad equation proceeds by parameterizing @xmath221 where @xmath222 is the @xmath86 identity matrix and the @xmath223 are the pauli matrices in the space spanned by @xmath37 and @xmath38 . we give the equation of motion for @xmath209 for the intermediate and high current regimes , in which we may approximate @xmath224 \approx { i}(\cos \varphi-\cos\varphi')/2 $ ] if @xmath225 and @xmath224 \approx 0 $ ] otherwise ( including the case @xmath226 ) . @xmath227 , \nonumber \\ \frac{d\rho_x}{dt } & = & - \frac{j { i}}{8 } \rho_y \cos \theta + \frac{j^2 { i}}{8 } \left\{-\rho_x [ 4 \sin^2 \theta + ( 2 s+1)^2 ] + 2 ( \rho_z \cos \theta - 2 \rho_0 ) ( 2 s+1 ) \sin \theta \right\ } - \frac{1}{2 } \rho_x \sum_{\pm } \gamma_{\pm 3/2,\pm 1/2 } , \nonumber \\ \frac{d \rho_y}{dt } & = & \frac{j { i}}{16}\left [ 2 \rho_x \cos \theta + \rho_z ( 2 s+1 ) \sin \theta \right ] - \frac{j^2 { i}}{8 } \rho_y \left [ 4 \sin^2 \theta + ( 2 s+1)^2 \cos^2 \theta \right ] - \frac{1}{2 } \rho_y \sum_{\pm } \gamma_{\pm 3/2,\pm 1/2 } , \nonumber \\ \frac{d \rho_z}{dt } & = & - \frac{j { i}}{16 } ( 2 s+1 ) \rho_y \theta + \frac{j^2 { i}}{8 } ( 2 s+1 ) \left\ { ( 2 s+1)[2 \rho_0 \cos \theta - \rho_z ( \cos^2 \theta + 1 ) ] + 2 \rho_x \sin \theta \cos \theta \right\ } \nonumber \\ & & \mbox { } + \sum_{\pm } \left [ \pm \gamma_{\pm 1/2,\pm 3/2 } p_{\pm 3/2 } - \frac{1}{2 } \gamma_{\pm 3/2,\pm 1/2 } ( \rho_z \pm \rho_0)\right ] , \label{eq : lindblad2d}\end{aligned}\ ] ] in the three - dimensional case the equations of motion are @xmath227 , \nonumber \\ \frac{d\rho_x}{dt } & = & - \frac{j { i}}{8 } \rho_y \cos \theta - \frac{j^2 { i}}{3 \pi^2 } \left\ { \rho_x \left [ 4 + \sin^2 \theta + 2 ( 2 s+1)^2 \right ] + 4 \sin \theta \cos \theta \rho_z \right\ } - \frac{j^2 { i}}{4 } \rho_0 ( 2 s+1 ) \sin \theta \nonumber \\ & & \mbox { } - \frac{1}{2 } \rho_x \sum_{\pm } \gamma_{\pm 3/2,\pm 1/2 } , \nonumber \\ \frac{d \rho_y}{dt } & = & \frac{j { i}}{16 } \left [ 2 \rho_x \cos \theta + \rho_z ( 2 s+1 ) \sin \theta \right ] - \frac{j^2 { i}}{3 \pi^2 } \rho_y \left[4 + 4 \sin^2 \theta + ( 2 s+1)^2(1 + \cos^2 \theta ) \right ] \nonumber \\ & & \mbox { } - \frac{1}{2 } \rho_y \sum_{\pm } \gamma_{\pm 3/2,\pm 1/2 } , \nonumber \\ \frac{d \rho_z}{dt } & = & - \frac{j { i}}{16 } ( 2 s+1 ) \rho_y \sin \theta + \frac{j^2 { i}}{3 \pi^2 } ( 2s+1 ) \left[2 \rho_x \sin \theta \cos \theta - \rho_z ( 2 s+1 ) ( 4 - \sin^2 \theta ) \right ] + \frac{j^2 { i}}{8 } ( 2s+1)^2 \rho_0 \cos \theta \nonumber \\ & & \mbox { } + \sum_{\pm } \left [ \pm \gamma_{\pm 1/2,\pm 3/2 } p_{\pm 3/2 } - \frac{1}{2 } \gamma_{\pm 3/2,\pm 1/2 } ( \rho_z \pm \rho_0)\right ] . \label{eq : lindblad3d}\end{aligned}\ ] ] in the intermediate - current regime we may set @xmath228 see eqs . ( [ eq : gammaanis ] ) and ( [ eq : gamma3dintermediate ] ) . for the high - current regime we may approximate @xmath229 by eqs . ( [ eq : gammahigh2d ] ) and ( [ eq : gammahigh3d ] ) for the two - dimensional and three - dimensional cases , respectively . when searching for a stationary solution , the contributions from transitions to / from the spin states @xmath230 vanish in the equations of motion for @xmath231 and @xmath232 because of detailed balance . they remain , however , in the equations of motion for @xmath233 and @xmath234 . once the ratio @xmath235 has been determined from the stationary solution of the lindblad equations ( [ eq : lindblad2d ] ) or ( [ eq : lindblad3d ] ) , the remaining probabilities @xmath65 for @xmath208 can be found from detailed balance , @xmath236 and from the normalization condition @xmath237 . we now report expressions for the ratio @xmath235 for the intermediate and high current regimes @xmath238 and @xmath105 for a spin coupled to the edge of a two - dimensinal ti and to the surface of a three - dimensional ti . the expressions for the intermediate - current regime are simplified using the inequalities @xmath239 and @xmath240 , but we make no assumptions regarding the relative magnitude of these two quantities . ( note that the intermediate current regime does not exist for @xmath3 . ) the expressions for the high - current limit are simplified using the inequality @xmath240 . in the intermediate - current regime we find for the case of a molecular magnet coupled to the edge of a two - dimensional ti @xmath241 with @xmath242 this result agrees with what one obtains from naive application of the rate equations if @xmath243 , but differs otherwise because of the effect of the current - induced exchange field ( [ eq : hexexpr ] ) . note that @xmath244 . the resulting steady - state probability distribution is @xmath245 the current - induced switching rate from @xmath70 to @xmath69 can be calculated as the transition rate from @xmath38 to @xmath37 starting with the initial condition @xmath246which is the reduced density matrix after current - induced equilibration of all states with @xmath247 , including the state @xmath38 , but without including transitions from @xmath38 to @xmath37 . within a time @xmath248 , which is short in comparison to the time required to transition from to @xmath38 to @xmath37 , a quasi - steady state is reached for the reduced density matrix @xmath209 , in which the off - diagonal component @xmath234 becomes nonzero , @xmath249 the current - induced switching rate @xmath74 is then found as the rate of change of @xmath250 in this quasi - steady state , which gives @xmath251 where @xmath252 was defined in eq . ( [ eq : gplus2dintermed ] ) . the switching rate for the opposite process is given the same expression , but with the replacement @xmath94 . for the high - current limit and a two - dimensional ti we find @xmath253 with the function @xmath254 defined as @xmath255 which is different from the result one obtains from naive application of the rate equations . for @xmath3 the ratio @xmath235 is consistent with a spin polarized in the @xmath5 direction . the resulting probability distribution @xmath65 reads @xmath256 equation ( [ eq : gammaswitch2dhigh ] ) of the main text is a good order - of - magnitude estimate for the rate @xmath74 at which this steady - state distribution is approached . for a molecule on the surface of a three - dimensional ti with an applied current in the intermediate - current regime @xmath257 the stationary solution of the lindblad equation ( [ eq : lindblad3d ] ) gives the ratio @xmath253 with @xmath258 where the function @xmath259 was defined in eq . ( [ eq : fdef ] ) of the main text . this expression simplifies to the ratio @xmath260 one obtains by considering rates only in the limit @xmath261 . the corresonding steady - state distribution reads @xmath262 the switching rate from @xmath70 to @xmath69 is calculated in the same way as for the case of a two - dimensional ti , and one finds @xmath263 the switching rate for the inverse transition is again given by the replacement @xmath94 . finally , for the high - current regime in three dimensions we find that the stationary solution of the lindblad equation ( [ eq : lindblad3d ] ) has @xmath253 with @xmath264 the corresponding steady - state distribution reads @xmath265 this distribution is approached at the rate given by eq . ( [ eq : gammaswitch2dhigh ] ) of the main text . if the magnetic anisotropy axis is aligned with the current direction , the rate equation approach can be used throughout . in this case the transition rates @xmath266 and @xmath267 for the intermediate and high - current regimes read @xmath268.\end{aligned}\ ] ] in the intermediate - current regime , the steady - state spin polarization of the molecule is complete ; the probability @xmath157 vanishes @xmath269 if @xmath2 is half integer , whereas all other probabilities are exponentially small in @xmath146 . the switching rate out of the state @xmath70 is , in the intermediate - current regime , @xmath270 in the high - current regime the order - of - magnitude estimate ( [ eq : gammaswitch2dhigh ] ) for the switching rate also applies to the half - integer spin case . 30ifxundefined [ 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 http://link.aps.org/doi/10.1103/revmodphys.76.323 [ * * , ( ) ] http://dx.doi.org/10.1146/annurev-conmatphys-070909-104123 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevapplied.4.047001 [ * * , ( ) ] http://www.sciencedirect.com/science/article/pii/0304885396000625 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.54.9353 [ * * , ( ) ] http://www.sciencedirect.com/science/article/pii/s0304885307010116 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.101.036601 [ * * , ( ) ] http://scitation.aip.org/content/aip/journal/apl/97/16/10.1063/1.3502596 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.106.036601 [ * * , ( ) ] http://www.nature.com/nature/journal/v476/n7359/abs/nature10309.html [ * * , ( ) ] http://science.sciencemag.org/content/336/6081/555 [ * * , ( ) ] http://www.nature.com/nmat/journal/v12/n3/abs/nmat3522.html [ * * , ( ) ] http://www.nature.com/articles/ncomms2709 [ * * , ( ) ] http://www.nature.com/nnano/journal/v8/n8/abs/nnano.2013.145.html [ * * , ( ) ] http://www.nature.com/nmat/journal/v12/n7/abs/nmat3675.html [ * * , ( ) ] http://www.nature.com/nnano/journal/v8/n7/abs/nnano.2013.102.html [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.98.106803 [ * * , ( ) ] http://link.aps.org/doi/10.1103/revmodphys.82.3045 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.108.187201 [ * * , ( ) ] http://dx.doi.org/10.7566/jpsj.83.064710 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.109.166602 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.93.115419 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.93.125303 [ * * , ( ) ] https://arxiv.org/abs/1509.06929v1 [ * * ] http://ieeexplore.ieee.org/document/7406697/ [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.90.205403 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.93.205130 [ * * , ( ) ] http://dx.doi.org/10.1038/nature13534 [ * * , ( ) ] http://www.nature.com/nmat/journal/v13/n7/abs/nmat3973.html [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.114.257202 [ * * , ( ) ] http://www.nature.com/nmat/journal/v7/n3/abs/nmat2133.html [ * * , ( ) ] , http://www.annualreviews.org/doi/abs/10.1146/annurev-conmatphys-070909-104053 [ * * , ( ) ] https://www.cambridge.org/core/journals/mrs-bulletin/article/single-molecule-magnets/65e2461c810e8de1a2ed2bf9620d3f96 [ * * , ( ) ] http://pubs.acs.org/doi/full/10.1021/nl2006142 [ * * , ( ) ] http://www.nature.com/nmat/journal/v8/n3/abs/nmat2374.html [ * * , ( ) ] http://dx.doi.org/10.1021/nl204141z [ * * , ( ) ] http://dx.doi.org/10.1021/nl503925h [ * * , ( ) ] http://dx.doi.org/10.1002/smll.200801594 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.102.256803 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.106.236402 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.92.045430 [ * * , ( ) ] http://www.nature.com/nphys/journal/v8/n6/full/nphys2322.html [ * * , ( ) ] http://www.sciencedirect.com/science/article/pii/092145269390146w [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.57.6602 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.90.166602 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.70.195345 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevb.71.195324 [ * * , ( ) ] http://link.aps.org/doi/10.1103/physrevlett.111.167203 [ * * , ( ) ] https://arxiv.org/abs/1501.02643 [ * * ] http://www.sciencedirect.com/science/article/pii/0034487772900109 [ * * , ] http://link.springer.com/article/10.1007/bf01608499 [ * * , ]

the combinatorial datum that specifies a cluster algebra of geometric type is an extended exchange matrix @xmath2 , which consists of a skew - symmetrizable integer matrix @xmath0 , called an exchange matrix , and some additional rows called the coefficient rows of @xmath2 . certain exchange matrices @xmath0 are associated to tagged triangulations of marked surfaces , as explained in @xcite . in this case , the choice of coefficient rows amounts to choosing a set of laminations in the surface . a lamination is a collection of weighted curves in the surface . the corresponding coefficient row is a vector known as the shear coordinates of the lamination . the notion of laminations predates the notion of cluster algebras . references are found in ( * ? ? ? * sections 1213 ) . given an exchange matrix @xmath0 , we consider the category of cluster algebras of geometric type associated to @xmath0 , with maps given by certain specializations of the coefficients . we follow @xcite in broadening the category of cluster algebras of geometric type to allow infinitely many coefficient rows with not - necessarily integral entries , and then narrowing the definition of coefficient specialization , as compared with the definitions in @xcite . the category also depends on an underlying ring @xmath3 , usually @xmath4 , @xmath5 , or @xmath6 . if this category contains a universal object , the universal object is called the universal geometric cluster algebra for @xmath0 over @xmath3 . universal geometric cluster algebras over @xmath3 are known to exist ( * ? ? ? * corollary 4.7 ) when @xmath3 is a field , but there is no general theorem guaranteeing the existence of universal geometric cluster algebras over @xmath4 . however , the existence of universal geometric cluster algebras over @xmath4 is proved for various @xmath0 in @xcite , and for some additional @xmath0 in this paper , as we explain below . constructing a universal geometric cluster algebra is equivalent to constructing a universal extended exchange matrix over @xmath0 by specifying a collection of coefficient rows , or universal geometric coefficients over @xmath3 . the main theme of @xcite is to make the connection between universal geometric coefficients and the piecewise - linear geometry of a fan @xmath1 , called the mutation fan . the goal of this paper is to construct the mutation fan and universal geometric coefficients in the case of marked surfaces . both constructions involve a variant on laminations that we call quasi - laminations . we show that the fan defined naturally by quasi - laminations equals the rational part of the mutation fan if and only if the surface has a property that we call the curve separation property ( theorem [ rat fb surfaces ] ) . in most cases , the curve separation property is the following : given two incompatible allowable curves ( in the sense of quasi - laminations ) , there exists a tagged triangulation of the surface such that the shear coordinates of the two curves have strictly opposite signs in at least one entry . however , in the case of a once - punctured surface without boundary , the statement of the curve separation property must be altered slightly . ( see definition [ curve sep def ] . ) a long but elementary argument establishes the curve separation property for all marked surfaces except once - punctured surfaces without boundary ( theorem [ curve separation ] ) . we complete the construction ( corollary [ poly grow univ ] ) of universal geometric coefficients ( over @xmath4 or @xmath5 ) for a smaller family of of surfaces . specifically , we show in theorem [ null tangle theorem ] that the shear coordinates of allowable curves ( in the sense of quasi - laminations ) are universal geometric coefficients if and only if the surface has a stronger property called the null tangle property : if a finite collection of weighted curves has the weighted sum of its shear coordinates equal to zero for all tagged triangulations , then all weights are zero . ( once again , the statement of the property must be modified for once - punctured surfaces without boundary . ) we prove the null tangle property in certain cases by extending some of the arguments that establish the curve separation property . specifically , theorem [ poly grow null tangle ] establishes the null tangle property for a disk with at most @xmath7 punctures , for an annulus with at most @xmath8 puncture , and for a sphere with three boundary components and no punctures . it remains open to determine whether once - punctured surfaces without boundary have the curve separation property , and to determine which additional surfaces have the null tangle property . however , in @xcite , we establish the null tangle property ( and thus the curve separation property ) in the simplest case not covered in this paper : the once - punctured torus . this leads to an explicit construction of universal geometric coefficients over @xmath4 , @xmath5 or @xmath6 for the once - punctured torus . in this section , we briefly review the construction of cluster algebras of geometric type , in the broader sense introduced in @xcite . we describe coefficient specialization and characterize the universal objects under specialization , the universal geometric cluster algebras or cluster algebras with universal geometric coefficients . we also present some results from @xcite that are useful in constructing universal geometric coefficients for cluster algebras from surfaces . all of the constructions depend on a choice of an * _ underlying ring _ * @xmath3 . we require that @xmath3 is either the integers @xmath4 or a field containing the rationals @xmath5 as a subfield and contained as a subfield of the real numbers @xmath6 . we are most interested in the cases where @xmath3 is @xmath4 , @xmath5 , or @xmath6 . let @xmath9 be an indexing set of arbitrary cardinality . the usual definition of cluster algebras of geometric type , found in ( * ? * section 2 ) , is recovered from the definition below by taking @xmath10 and @xmath9 finite . for each @xmath11 , let @xmath12 be a formal symbol . let @xmath13 be the set of formal products of the form @xmath14 with each @xmath15 in @xmath3 . this is an abelian group with product @xmath16 and identity @xmath17 . we define an * _ auxiliary addition _ * @xmath18 in @xmath13 by @xmath19 the triple @xmath20 is called a * _ tropical semifield _ * over @xmath3 . the symbols @xmath12 are called * _ tropical variables_*. we also consider @xmath13 as a topological space . it is the product , indexed by @xmath9 , of copies of @xmath3 . we give @xmath3 the discrete topology and give @xmath13 the product topology as a product of copies of the discrete set @xmath3 . details about the product topology may be found , for example , in ( * ? ? ? * chapter 3 ) . see also ( * ? ? ? * section 3 ) . an * _ exchange matrix _ * @xmath0 is a * _ skew - symmetrizable _ * @xmath21 integer matrix . that means that there exist positive integers @xmath22 such that @xmath23 for all @xmath24 $ ] . in this paper , we will work with * _ skew - symmetric _ * exchange matrices , meaning that @xmath25 for all @xmath24 $ ] . ( the notation @xmath26 $ ] stands for the indexing set @xmath27 of @xmath0 . ) an * _ extended exchange matrix _ * is @xmath0 , augmented by an additional collection of * _ coefficient rows _ * , indexed by @xmath9 . each coefficient row is a vector in @xmath28 . if @xmath9 is infinite , then @xmath2 is not a matrix in the usual sense , but it is a collection of entries @xmath29 where @xmath30 is either in @xmath26 $ ] or in @xmath9 and @xmath31 is in @xmath26 $ ] . now set @xmath32 and let @xmath33 be a field of rational functions in @xmath34 independent variables with coefficients in @xmath35 . here @xmath35 is the group ring of the group @xmath36 , ignoring the auxiliary addition @xmath18 . a * _ ( labeled geometric ) seed _ * of rank @xmath34 is a pair @xmath37 such that @xmath38 is an @xmath34-tuple of algebraically independent elements of @xmath33 and @xmath2 is an extended exchange matrix with coefficient rows indexed by @xmath9 . the @xmath34-tuple @xmath39 is called a * _ cluster _ * and the entries of @xmath39 are called * _ cluster variables_*. the seed determines elements @xmath40 of @xmath36 called * _ coefficients_*. given a seed @xmath37 and an index @xmath41 $ ] , we define an new seed @xmath42 . the new cluster @xmath43 has @xmath44 whenever @xmath45 and @xmath46_+}\prod_{i\in i}u_i^{[b_{ik}]_+}+\prod_{i=1}^nx_i^{[-b_{ik}]_+}\prod_{i\in i}u_i^{[-b_{ik}]_+}\right),\ ] ] where @xmath47_+$ ] stands for @xmath48 . the new cluster variable @xmath49 is a rational function in @xmath39 with coefficients in @xmath50 , so that @xmath51 . the new extended exchange matrix @xmath52 is given by @xmath53_+&\mbox{otherwise , } \end{array}\right.\ ] ] where @xmath54 means @xmath55 if @xmath56 or @xmath57 if @xmath58 . since @xmath52 only depends on @xmath2 and @xmath59 , not on @xmath39 , we also write @xmath60 for @xmath52 . for any finite sequence @xmath61 of integers in @xmath26 $ ] , we use @xmath62 for @xmath63 . now let @xmath64 be the @xmath34-regular tree with an edge - labeling by integers @xmath8 through @xmath34 , with each vertex incident to exactly one edge with each label . an edge labeled @xmath59 between vertices @xmath65 and @xmath66 is denoted by @xmath67 . we fix a vertex @xmath68 and define a map @xmath69 from vertices of @xmath64 to seeds by requiring that @xmath70 and that @xmath71 whenever @xmath67 . the map @xmath69 is a * _ cluster pattern _ * and the map @xmath72 is a * _ @xmath73-pattern_*. we write @xmath74 and @xmath75 for the cluster variables and matrix entries of @xmath76 and write @xmath77 for the coefficient @xmath78 . associated to the cluster pattern is a * _ cluster algebra @xmath79 of geometric type _ * , namely the @xmath50-subalgebra of @xmath33 generated by all cluster variables in the cluster pattern . the cluster algebra is determined up to isomorphism by the extended exchange matrix @xmath2 , so we write @xmath80 for @xmath81 . for any permutation @xmath82 of @xmath26 $ ] , applying @xmath82 simultaneously to the entries of @xmath39 , to the rows of @xmath0 , to the columns of @xmath2 and to the edge - labels of @xmath64 induces a canonical isomorphism of cluster patterns and of cluster algebras . in ( * definition 4.1 ) * _ unlabeled seeds _ * are defined as equivalence classes of seeds under these simultaneous permutations . the * _ exchange graph _ * is the graph on unlabeled seeds obtained from the cluster pattern by identifying vertices that have the same _ unlabeled seed_. let @xmath32 and @xmath83 be tropical semifields over the same underlying ring @xmath3 . let @xmath37 and @xmath84 be seeds of rank @xmath34 , with coefficient rows indexed by @xmath9 and @xmath85 respectively . a * _ coefficient specialization _ * is a ring homomorphism @xmath86 such that a. the exchange matrices @xmath0 and @xmath87 coincide ; b. @xmath88 for all @xmath89 $ ] ; c. the restriction of @xmath90 to @xmath36 is a continuous @xmath3-linear map to @xmath91 with @xmath92 and @xmath93 for all @xmath89 $ ] and @xmath94 . continuity here refers to the product topology described above . the cluster algebra @xmath95 has * _ universal geometric coefficients over @xmath3 _ * if the following condition holds : given a cluster algebra @xmath96 with the same initial exchange matrix @xmath0 , there exists a unique coefficient specialization from @xmath97 to @xmath98 . the extended exchange matrix @xmath2 is then said to be * _ universal over @xmath3_*. equation leads to the definition of * _ mutation maps _ * for @xmath0 . for each @xmath41 $ ] , we define a map @xmath99 by setting @xmath100 with @xmath101 equivalently , if @xmath2 is an extended exchange matrix with exchange matrix @xmath0 and @xmath102 is a coefficient row of @xmath2 , then @xmath103 is the corresponding coefficient row of @xmath60 . more generally , the mutation maps associated to @xmath0 are defined as follows . let @xmath104 be a sequence of integers in @xmath26 $ ] , let @xmath105 and define @xmath106 for @xmath107 $ ] . then @xmath108 the map @xmath109 is continuous , piecewise linear , and invertible , with inverse @xmath110 . in particular , it is a homeomorphism from @xmath111 to @xmath111 . it also has the property that @xmath112 given a finite set @xmath113 , vectors @xmath114 in @xmath111 , and elements @xmath115 of @xmath3 , write the formal expression @xmath116 . this expression is a * _ @xmath0-coherent linear relation with coefficients in @xmath3 _ * if the equalities @xmath117 hold for every finite sequence @xmath61 of integers in @xmath26 $ ] . the symbol @xmath118 is the zero vector and @xmath119 refers to componentwise minimum . the equality for the empty sequence @xmath120 implies that a @xmath0-coherent linear relation is also a linear relation in the usual sense . let @xmath9 be some indexing set and let @xmath121 be vectors in @xmath28 . then @xmath121 is an * _ @xmath3-spanning set for @xmath0 _ * if the following condition holds : if @xmath122 , then there exists a finite subset @xmath123 and elements @xmath115 of @xmath3 such that @xmath124 is a @xmath0-coherent linear relation . also , @xmath121 is an * _ @xmath3-independent set for @xmath0 _ * if the following condition holds : if @xmath113 is a finite subset of @xmath9 and @xmath125 is a @xmath0-coherent linear relation with coefficients in @xmath3 , then @xmath126 for all @xmath127 . finally , @xmath121 is an * _ @xmath3-basis _ * for @xmath0 if it is both an @xmath3-independent set for @xmath0 and an @xmath3-spanning set for @xmath0 . the notion of a @xmath3-basis for @xmath0 is analogous to the linear - algebraic notion of a basis , with @xmath0-coherent linear relations replacing the usual linear relations . our interest in @xmath3-bases for @xmath0 arises from the following theorem , which is ( * ? ? ? * theorem 4.4 ) . [ basis univ ] let @xmath2 be an extended exchange matrix with entries in @xmath3 . then @xmath2 is universal over @xmath3 if and only if the coefficient rows of @xmath2 are an @xmath3-basis for @xmath0 . for a description of the unique coefficient specialization from @xmath80 to any other cluster algebra with the same @xmath3 and @xmath0 , see ( * ? ? ? * remark 4.5 ) . the following proposition ( * ? ? ? * proposition 4.6 ) is proved by the usual argument , using zorn s lemma , that any vector space has a ( hamel ) basis . [ basis exists ] suppose the underlying ring @xmath3 is a field . for any exchange matrix @xmath0 , there exists an @xmath3-basis for @xmath0 . given an @xmath3-spanning set @xmath128 for @xmath0 , there exists an @xmath3-basis for @xmath0 contained in @xmath128 . in light of theorem [ basis univ ] and proposition [ basis exists ] , if @xmath3 is a field , then universal geometric coefficients over @xmath3 exist for any @xmath0 . we have no general proof that @xmath4-bases exist , and indeed the second statement of proposition [ basis exists ] can fail for @xmath10 . ( see ( * ? ? ? * remark 4.9 ) . ) however , @xmath4-bases ( and thus universal geometric coefficients over @xmath4 ) are constructed for various exchange matrices @xmath0 in @xcite and later in this paper ( corollary [ poly grow univ ] ) . the following proposition is ( * ? ? ? * proposition 4.11 ) . [ no zero no piecewise ] suppose @xmath0 has no row consisting entirely of zeros . then the formal expression @xmath116 is a @xmath0-coherent linear relation if and only if condition holds for all sequences @xmath120 . a subset of @xmath111 that is closed under addition and positive scaling is a * _ convex cone_*. a * _ simplicial _ * cone is the nonnegative span of a finite set of linearly independent vectors . a * _ rational cone _ * is the nonnegative span of a finite set of rational vectors . a * _ face _ * of a convex cone @xmath129 is a convex subset @xmath130 such that any line segment @xmath131 whose interior intersects @xmath130 has @xmath132 . any face of a convex cone is itself a convex cone . a face of a closed convex cone is closed as well . a collection @xmath133 of closed convex cones is a * _ fan _ * if it has the following two properties : if @xmath134 and @xmath130 is a face of @xmath129 , then @xmath135 ; and the intersection of any two cones in @xmath133 is a face of each of the two . a * _ complete fan _ * is a fan @xmath133 such that the union of the cones in @xmath133 is the entire ambient space . a fan is * _ simplicial _ * if all of its cones are simplicial . a fan is * _ rational _ * if all of its cones are rational . if @xmath136 and @xmath137 are fans and every cone in @xmath137 is a union of cones in @xmath136 , then @xmath136 * _ refines _ * @xmath137 . as before , @xmath54 means @xmath55 if @xmath56 or @xmath57 if @xmath58 . given @xmath138 in @xmath111 , we write @xmath139 for @xmath140 . given @xmath141 , define @xmath142 to mean that @xmath143 for every sequence @xmath120 of integers in @xmath26 $ ] . a @xmath144-equivalence class is a * _ @xmath0-class_*. the closure of a @xmath0-class is a * _ @xmath0-cone_*. each @xmath0-class is a convex cone and each @xmath0-cone is a closed convex cone ( * ? ? ? * proposition 5.4 ) . the * _ mutation fan _ * for @xmath0 is the collection @xmath1 consisting of all @xmath0-cones and all faces of @xmath0-cones . ( faces of @xmath0-cones can fail to be @xmath0-cones . see ( * ? ? ? * remark 5.28 ) . ) the following results are ( * ? ? ? * proposition 5.3 ) , ( * ? ? ? * theorem 5.13 ) , and ( * ? ? ? * proposition 5.26 ) . [ linear ] every mutation map @xmath109 is linear on every @xmath0-cone . [ fan ] the collection @xmath1 is a complete fan . [ int of b - cones ] an arbitrary intersection of @xmath0-cones is a @xmath0-cone . a collection of vectors in @xmath111 is * _ sign - coherent _ * if their @xmath145 coordinates weakly agree in sign , for each @xmath41 $ ] . the following proposition is ( * ? ? ? * proposition 5.30 ) . [ contained bcone ] a set @xmath146 is contained in some @xmath0-cone if and only if the set @xmath147 is sign - coherent for every sequence @xmath120 of indices in @xmath26 $ ] . an @xmath3-basis @xmath121 for @xmath0 is * _ positive _ * if , for any vector @xmath122 , there is a @xmath0-coherent linear relation @xmath124 with each @xmath148 nonnegative . a positive @xmath3-basis seems to be the natural choice for `` the right '' @xmath3-basis for @xmath0 , and this intuition is borne out by the following results , which are ( * ? ? ? * proposition 6.2 ) and ( * ? ? ? * corollary 6.13 ) . note the appearance of @xmath6 in proposition [ pos reals fb ] , rather than @xmath3 . [ positive unique ] for any fixed @xmath3 , there is at most one positive @xmath3-basis for @xmath0 , up to scaling of each basis element by a positive unit in @xmath3 . [ pos reals fb ] if a positive @xmath6-basis for @xmath0 exists , then @xmath1 is simplicial . the basis consists of exactly one vector in each ray of @xmath1 . the following proposition is part of ( * ? ? ? * proposition 6.7 ) . [ positive cone basis ] suppose @xmath121 is an @xmath3-independent set for @xmath0 with the following property : if @xmath129 is a @xmath0-cone , then the nonnegative @xmath3-linear span of @xmath149 contains @xmath150 . then @xmath121 is a positive @xmath3-basis for @xmath0 . in this section , we give background material on cluster algebras arising from triangulated surfaces , following @xcite . [ s m def ] let @xmath151 be a surface obtained from a compact , oriented surface without boundary by deleting a finite collection of disjoint open disks . the boundary of @xmath151 is a finite collection of circles called * _ boundary components _ * , one for each disk removed . let @xmath152 be a nonempty , finite collection of points in @xmath151 called * _ marked points _ * , with at least one marked point in each boundary component . marked points in the interior of @xmath151 are called * _ punctures_*. a curve contained in the boundary of @xmath151 , connecting two points in @xmath152 , with no points of @xmath152 in the interior of the curve , is a * _ boundary segment_*. we exclude unpunctured monogons , digons , and triangles , as well as once - punctured monogons . if @xmath151 is a sphere , then we require that it have at least @xmath153 punctures . in @xcite , the surface @xmath151 is considered to be a riemann surface , but the complex structure is not needed in this paper . [ tri def ] an * _ arc _ * in @xmath154 is a curve in @xmath151 , up to isotopy relative to @xmath152 , whose endpoints are points in @xmath152 . we require that the arc not intersect itself , except that its endpoints may coincide , and that the arc is disjoint from @xmath152 and from the boundary of @xmath151 , except at its endpoints . we also exclude arcs that bound an unpunctured monogon and arcs that , together with a boundary segment connecting the endpoints , define an unpunctured digon . two arcs @xmath155 and @xmath156 are * _ incompatible _ * if they intersect in @xmath157 , and if the intersection can not be removed by ( independently ) isotopically deforming the arcs . otherwise , they are * _ compatible_*. a * _ triangulation _ * is a maximal collection of distinct pairwise compatible arcs . the exclusion of certain pairs @xmath154 in definition [ s m def ] assures that @xmath154 admits at least two triangulations . the arcs defining an triangulation divide @xmath151 into * _ triangles _ * , which have @xmath8 , @xmath7 , or @xmath158 distinct vertices and @xmath7 or @xmath158 distinct sides . ( in @xcite , the terms * _ ideal triangles _ * and * _ ideal triangulations _ * are used because the marked points lie at infinity in the complex structure . ) a * _ self - folded triangle _ * is a triangle with 2 distinct sides . ( for example , later in figure [ shear self - folded fig ] , one of the two triangles shown in each picture is self - folded . ) an arc in a triangulation is an edge of exactly two triangles in the triangulation unless it constitutes two sides of a self - folded triangle . removing the arc from the triangulation leaves a quadrilateral region , which can be completed to two triangles by either one of the diagonals of the quadrilateral . a * _ flip _ * is the operation of removing an arc from a triangulation and forming a new triangulation by inserting the other diagonal of the quadrilateral . [ tagged tri def ] a * _ tagged arc _ * in @xmath154 is a arc in @xmath151 that does not cut out a once - punctured monogon , with each end of the arc designated ( or `` tagged '' ) either * _ plain _ * or * _ notched_*. an endpoint on a boundary component must be tagged plain , and if the endpoints of an arc coincide , then they must have the same tag . two tagged arcs are * _ compatible _ * if one of the following conditions holds : the two underlying untagged arcs are distinct ( i.e. non - isotopic ) and compatible and any shared endpoints have the same tagging , or the two underlying untagged arcs are the same and their tagging agrees on one endpoint but not the other . a * _ tagged triangulation _ * is a maximal collection of distinct pairwise compatible tagged arcs . there is a map @xmath159 from ordinary arcs to tagged arcs , defined as follows : if an arc @xmath156 does not bound a once - punctured monogon , then @xmath160 is the same curve , with both endpoints tagged plain . if @xmath156 has both endpoints at @xmath161 , and bounds a once - punctured monogon with puncture @xmath162 , then @xmath160 is the arc , in the monogon , from @xmath163 to @xmath164 , tagged plain at @xmath163 and notched at @xmath164 . starting with a triangulation , and applying @xmath159 to each arc , we obtain a tagged triangulation . as an example of the map @xmath159 , consider the triangulation @xmath165 of a punctured digon shown in the left picture of figure [ tau fig ] . one arc defining @xmath165 connects a marked point on the boundary to the puncture , and the other arc ( in purple or light gray ) encloses the puncture . the right picture of the figure shows the tagged triangulation @xmath166 . [ complex and flips ] the tagged triangulations are the maximal simplices of a simplicial complex called the * _ tagged arc complex_*. this is the complex whose vertices are the tagged arcs , with a set of arcs defining a face if and only if they are pairwise compatible . all tagged triangulations of @xmath154 have the same number of tagged arcs . ( see ( * ? ? ? * proposition 2.10 ) . ) given a tagged triangulation @xmath165 and a particular tagged arc @xmath156 in @xmath165 , there is a unique tagged triangulation @xmath167 such that @xmath168 and @xmath169 . the tagged triangulations @xmath165 and @xmath167 are said to be related by a * _ tagged flip _ * , or more specifically , @xmath167 is obtained from @xmath165 by a flip at @xmath156 . an exact combinatorial description of tagged flips is found in ( * ? ? ? * remark 5.13 ) . the dual graph to the tagged arc complex is the graph whose vertices are tagged triangulations and whose edges are pairs of triangulations related by a tagged flip . this dual graph is connected ( * ? ? ? * proposition 7.10 ) , except when @xmath151 has no boundary components and exactly one puncture . in the latter case , the dual graph of the tagged arc complex has exactly two connected components , one consisting of triangulations with all arcs tagged plain , and the other consisting of triangulations with all arcs tagged notched . the two components are related by the symmetry of reversing all taggings ( plain@xmath170notched ) . given a tagged triangulation @xmath165 , there is a simple recipe ( * ? ? ? * definition 5.15 ) for writing down an exchange matrix @xmath171 , called the * _ signed adjacency matrix _ * of @xmath165 . this is an @xmath21 matrix whose rows and columns are indexed by the tagged arcs @xmath156 in @xmath165 . we do not need the details of this definition . rather , what is important for us are some observations and results of @xcite . first , @xmath171 is skew - symmetric , rather than merely skew - symmetrizable . next , @xmath171 does not have a row or column consisting of all zeros . finally , we quote very small parts of ( * ? ? ? * theorem 6.1 ) and ( * ? ? ? * corollary 6.2 ) . [ signed adjacency mut ] if @xmath165 and @xmath167 are tagged triangulations such that @xmath167 is obtained from @xmath165 by a flip at an arc @xmath156 , then @xmath172 is obtained from @xmath171 by matrix mutation at the position indexed by @xmath156 . [ cluster main ] let @xmath97 be a cluster algebra with initial exchange matrix @xmath171 and arbitrary coefficients . then the exchange graph of @xmath97 is isomorphic to the dual graph of the tagged arc complex , or , if @xmath151 has no boundary components and exactly one puncture , the connected component of the dual graph consisting of tagged triangulations with all arcs tagged plain . [ lam def ] an * _ integral ( unbounded measured ) lamination _ * is a collection of pairwise non - intersecting curves in @xmath151 , up to isotopy relative to @xmath152 . we take the curves to be distinct up to isotopy , but each curve @xmath173 appears with a positive integer weight @xmath174 . each curve must be * a closed curve , * a curve whose two endpoints are unmarked boundary points , * a curve having one endpoint an unmarked boundary point , with the other end spiraling ( clockwise or counterclockwise ) into a puncture , or * a curve spiraling into ( not necessarily distinct ) punctures at both ends . however , we exclude * curves with self - intersections , * curves that are contractible in @xmath157 , * curves that are contractible to a puncture , and * curves having two endpoints on the boundary and being contractible to a portion of the boundary containing zero or one marked points . when a curve spirals into a puncture , that puncture is a * _ spiral point _ * of the curve . a * _ rational ( unbounded measured ) lamination _ * is defined in the same way , but with positive rational weights rather than positive integer weights . [ shear def ] given a lamination @xmath175 and a tagged triangulation @xmath165 , the * _ shear coordinate vector _ * is a vector @xmath176 indexed by arcs @xmath156 in @xmath165 . the entry @xmath177 is defined as follows . we deal with each curve @xmath178 separately and set @xmath179 where @xmath174 is the weight of @xmath173 in @xmath175 and where @xmath180 is a quantity that we now define . first , everywhere there is a puncture @xmath181 incident only to notched ends of arcs , change all of these notched ends to plain ends and reverse the direction of any spirals of @xmath173 at @xmath181 . we obtain an altered tagged triangulation with no notched endpoints on arcs , except that at some points @xmath164 there are two tagged arcs , whose underlying untagged arcs are the same , one tagged plain and one tagged notched at @xmath164 . if @xmath156 is one of these arcs for some @xmath164 , then if necessary we switch notched ends with plain ends and reverse any spirals of @xmath173 at @xmath164 so that @xmath156 becomes an arc with one end notched . call the altered tagged triangulation @xmath167 and call the curve with some spirals possibly reversed @xmath182 . define @xmath180 to be @xmath183 , calculated as follows . there is a triangulation @xmath184 with @xmath185 . re - use the name @xmath156 for the arc in @xmath184 corresponding to @xmath186 . because of how we defined @xmath167 , the arc @xmath156 is contained in two distinct triangles of @xmath184 ( rather than defining two edges of the same self - folded triangle ) . choose an isotopy deformation of @xmath182 ( reusing the symbol @xmath182 for this deformation ) so as to minimize the number of intersection points of @xmath182 with @xmath156 . then @xmath183 is the sum , over each intersection of @xmath182 with @xmath156 , of a number in @xmath187 . the number is @xmath57 unless @xmath182 intersects the two triangles containing @xmath156 as shown in figure [ shear fig ] , in which case it is @xmath8 or @xmath188 as shown . in both pictures , the arc @xmath156 is the diagonal of the square and the curve @xmath182 is the vertical or horizonal line intersecting the square . figure [ shear fig ] introduces a color convention that we maintain throughout the paper . when calculating a shear coordinate at @xmath156 , we color @xmath156 purple and color the other edges red and blue so that , traveling clockwise along each triangle , we meet the colors in the order red , blue , purple . those reading the paper in grayscale may have some difficulty in distinguishing the arcs ; it may be helpful to know that the blue arcs become a darker gray than the red arcs and the purple arcs are lightest of all . we also shade the interiors of the red - blue - purple triangles with a light purple . the quadrilaterals represented in figure [ shear fig ] might have fewer than @xmath153 distinct vertices and @xmath153 distinct sides . many examples appear later on . the first example occurs in figure [ shear self - folded fig ] , which illustrates shear coordinates in the case where one of the two triangles in figure [ shear fig ] is self - folded . although a red and blue edge coincide , we show both colors for clarity , putting the colors on the appropriate sides to make the connection to figure [ shear fig ] . the following are ( part of ) ( * ? ? ? * theorem 13.5 ) and ( * ? ? ? * theorem 13.6 ) . [ mut lam ] fix a lamination @xmath175 . if @xmath165 and @xmath167 are tagged triangulations such that @xmath167 is obtained from @xmath165 by a flip at an arc @xmath156 , then @xmath189 , where @xmath190 is the mutation map at the index for the arc @xmath156 in @xmath165 . [ lam bij ] fix a tagged triangulation @xmath165 . then the map @xmath191 is a bijection between integral ( resp . rational ) unbounded measured laminations and @xmath192 ( resp . @xmath193 ) . theorem [ mut lam ] suggests the idea of constructing the mutation fan @xmath1 and a basis for @xmath0 in terms of laminations . in turns out that unbounded measured laminations are almost , but not quite , the right tool . in this section , we develop the right tool , which we call quasi - laminations . we also define the curve separation property , a property of @xmath154 that allows us to construct the rational part of the mutation fan in the context of quasi - laminations . in section [ curve separation sec ] , we establish the curve separation property for all marked surfaces except those with no boundary components and exactly one puncture . [ quasi lam def ] a curve in @xmath151 is * _ allowable _ * if it satisfies the requirements placed on curves in definition [ lam def ] and is not one of the following : * a curve that has both endpoints on the same boundary segment and that , together with the portion of the boundary between its endpoints , cuts out a once - punctured disk , and * a curve with two coinciding spiral points that cuts out a once - punctured disk . figure [ exclude curves ] illustrates these two excluded curves . two allowable curves are * _ compatible _ * if they are non - intersecting , or if they are identical except that , at one ( and only one ) end of the curve , they spiral opposite directions into the same point . figure [ allow intersect ] illustrates the allowed intersections between compatible curves . one curve is drawn black and the other is dotted and green . an * _ integral quasi - lamination _ * is a collection of pairwise compatible allowable curves with positive integer weights . a * _ rational quasi - lamination _ * is a collection of pairwise compatible allowable curves with positive rational weights . [ quasi - shear def ] let @xmath165 be a tagged triangulation , let @xmath156 be a tagged arc in @xmath165 and let @xmath175 be a quasi - lamination . then the shear coordinates @xmath194 are defined as a weighted sum of shear coordinates @xmath195 of the individual curves @xmath173 in @xmath175 , just as in definition [ shear def ] . we should think of the definition of quasi - laminations as a harmless modification of the definition of laminations . indeed , given a rational unbounded measured lamination , one can replace all curves of the type illustrated in figure [ exclude curves ] with pairs of intersecting but compatible curves of the corresponding type shown in figure [ allow intersect ] . ( if some curve @xmath173 appears twice in the result , we replace the two copies of @xmath173 with a single copy of @xmath173 whose weight is the sum of the original two weights . ) this is a bijection , preserving integrality and shear coordinates , between rational unbounded measured laminations and quasi - laminations . the following theorems are now immediate from theorems [ mut lam ] and [ lam bij ] . [ mut q - lam ] fix a quasi - lamination @xmath175 . if @xmath165 and @xmath167 are tagged triangulations such that @xmath167 is obtained from @xmath165 by a flip at an arc @xmath156 , then @xmath189 , where @xmath190 is the mutation map at the index for the arc @xmath156 in @xmath165 . [ q - lam bij ] fix a tagged triangulation @xmath165 . then the map @xmath191 is a bijection between integral ( resp . rational ) quasi - laminations and @xmath192 ( resp . @xmath193 ) . in order for us to understand the mutation fan in terms of quasi - laminations , the surface must have the following key property . [ curve sep def ] a marked surface @xmath154 has the * _ curve separation property _ * if for any incompatible allowable curves @xmath173 and @xmath196 , there exists a tagged triangulation @xmath165 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) and an arc @xmath156 in @xmath165 such that the integers @xmath180 and @xmath197 have strictly opposite signs . definition [ curve sep def ] defines the curve separation property entirely in the context of the geometry of @xmath154 . we now rephrase the property in terms of the mutation fan : curves are incompatible exactly when they are separated in the mutation fan . [ fbt compatible ] suppose @xmath154 satisfies the curve separation property , let @xmath165 be any tagged triangulation of @xmath154 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) , and @xmath198 be a set of allowable curves . then the curves in @xmath198 are pairwise compatible if and only if the vectors @xmath199 are contained in some @xmath171-cone . suppose @xmath173 and @xmath196 are not compatible . then the curve separation property says that there exists a tagged triangulation @xmath167 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) and an arc @xmath156 in @xmath167 such that the integers @xmath200 and @xmath201 have strictly opposite signs . then @xmath167 is obtained from @xmath165 by a sequence of flips , so theorem [ mut q - lam ] implies that @xmath202 and @xmath203 are obtained by applying some mutation map to @xmath195 and @xmath204 . by proposition [ contained bcone ] , @xmath195 and @xmath204 are not contained in any common @xmath171-cone . conversely , let @xmath173 and @xmath196 be compatible , allowable curves and let @xmath156 be a tagged arc in a tagged triangulation @xmath167 . we will show that @xmath200 and @xmath201 weakly agree in sign . as in definition [ shear def ] , removing tags and switching spirals on @xmath173 and @xmath196 , we pass from @xmath167 to a triangulation @xmath184 where @xmath156 is contained in two distinct triangles . let @xmath205 be the interior of the quadrilateral that is the union of these two triangles . we refer to figure [ shear fig ] , where @xmath156 is represented by a purple arc and where @xmath205 is shaded light purple . ( possibly some of the edges or vertices of @xmath205 are identified with each other . ) we see that @xmath180 and @xmath197 do not have opposite signs unless @xmath173 and @xmath196 intersect , so suppose they intersect . since @xmath173 and @xmath196 are compatible , they are identical except at one end , where they spiral opposite ways into a puncture @xmath181 . consider the connected components of @xmath206 and the connected components of @xmath207 . since @xmath173 and @xmath196 are identical except where they spiral opposite ways into @xmath181 , each connected component of @xmath206 corresponds to a connected component of @xmath207 that intersects @xmath205 in the same way , with _ one possible exception _ : if @xmath181 is a vertex of the quadrilateral , there may be one connected component @xmath182 of @xmath206 that disagrees with the corresponding component @xmath208 of @xmath207 . but this disagreement can only be slight : orienting @xmath173 towards @xmath181 , the component @xmath182 enters @xmath205 through some edge of the quadrilateral and leaves @xmath205 through some other edge . orienting @xmath196 towards @xmath181 as well , the component @xmath208 enters @xmath205 through the same edge as @xmath182 does . thus it is impossible for @xmath182 and @xmath208 to contribute of strictly opposite signs to the shear coordinates of @xmath173 and @xmath196 . if either of them ( say @xmath182 ) makes a nonzero contribution @xmath209 , then since @xmath173 is not self - intersecting , all components of @xmath206 make contributions weakly agreeing in sign with @xmath209 . therefore also all components of @xmath207 make contributions weakly agreeing in sign with @xmath209 , so @xmath200 and @xmath201 weakly agree in sign . now if @xmath198 is a set of pairwise compatible curves , then @xmath200 and @xmath201 weakly agree in sign for any pair @xmath210 . therefore , @xmath211 all weakly agree in sign for any @xmath167 . by theorem [ mut q - lam ] , applying any mutation map @xmath212 to @xmath199 produces a sign - coherent set of vectors . proposition [ contained bcone ] says that @xmath199 is contained in some @xmath171-cone . [ all arcs tagged plain ] the requirement that all arcs be tagged plain if @xmath154 has no boundary components and exactly one puncture in definition [ curve sep def ] , in proposition [ fbt compatible ] , and in what follows arises from the fact that the tagged arc complex has two connected components in this case . ( see definition [ complex and flips ] . ) the requirement is crucial in the proof of one direction of proposition [ fbt compatible ] , where it ensures that the tagged triangulations @xmath165 and @xmath167 are related by a sequence of flips . this requirement is also the reason that , in section [ curve separation sec ] , we are not able to establish the curve separation property for @xmath154 having no boundary components and exactly one puncture . we now define a rational fan @xmath213 , based on quasi - laminations , which is closely related to the mutation fan @xmath214 . specifically , @xmath213 is the rational part of @xmath214 , in a sense that we make precise below . [ rat q - lam fan def ] fix a tagged triangulation @xmath165 . let @xmath198 be a set of pairwise compatible allowable curves in the sense of definition [ quasi lam def ] . let @xmath215 be the nonnegative @xmath6-linear span of the integer vectors @xmath199 . let @xmath213 be the collection of all such cones @xmath215 . we call @xmath213 the * _ rational quasi - lamination fan _ * for @xmath165 and justify the name below in theorem [ rat fb surfaces ] . the following is a specialization of ( * ? ? ? * definition 6.9 ) . [ rat part def ] suppose @xmath133 and @xmath216 are fans such that : a. [ rat cones rat part ] @xmath216 is a rational fan . b. [ cone in cone rat part ] each cone in @xmath216 is contained in a cone of @xmath133 . c. [ max cone rat part ] for each cone @xmath129 of @xmath133 , there is a unique largest cone ( under containment ) among cones of @xmath216 contained in @xmath129 . this largest cone contains @xmath217 . then @xmath216 is called the * _ rational part _ * of @xmath133 . if the rational part of @xmath133 exists , then it is unique . [ rat fb surfaces ] let @xmath165 be any tagged triangulation of @xmath154 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) . the collection @xmath213 is a rational , simplicial fan . the fan @xmath213 is the rational part of @xmath214 if and only if @xmath154 has the curve separation property . suppose that some cone @xmath215 is not simplicial . then there is some rational point in @xmath215 that can be written as a @xmath5-linear combination of the vectors @xmath218 in two distinct ways . this point is thus obtained as the shear coordinates of two distinct rational quasi - laminations , contradicting theorem [ q - lam bij ] . let @xmath215 and @xmath219 be cones in @xmath213 . then @xmath220 contains @xmath221 . suppose @xmath222 is a rational vector in @xmath220 . then @xmath222 is the shear coordinates of some quasi - lamination supported on a set @xmath223 of compatible curves . by theorem [ q - lam bij ] , we conclude that @xmath223 is contained in both @xmath198 and @xmath224 . thus @xmath225 . since @xmath220 is a rational cone , @xmath226 , so that @xmath227 we have shown that @xmath213 is a simplicial fan , and it is rational by construction . if @xmath154 does not have the curve separation property , then theorem [ mut q - lam ] and proposition [ contained bcone ] imply that there exist incompatible allowable curves @xmath173 and @xmath196 with @xmath195 and @xmath204 contained in some @xmath0-cone @xmath129 . but @xmath195 and @xmath204 both span rays of @xmath213 not contained in a common cone of @xmath213 . thus @xmath213 fails condition of definition [ rat part def ] . now suppose @xmath154 has the curve separation property . condition of definition [ rat part def ] holds by proposition [ fbt compatible ] . we now establish condition of definition [ rat part def ] . given a @xmath171-cone @xmath129 , let @xmath198 be the set of allowable curves @xmath173 whose shear coordinates @xmath195 are contained in @xmath129 . proposition [ fbt compatible ] implies that @xmath215 is the unique largest cone among cones of @xmath213 contained in @xmath129 . it remains to show that this @xmath215 contains every rational point in @xmath129 . we argue by induction on the dimension of @xmath129 . the assertion is trivial when @xmath129 is the zero cone . suppose @xmath228 is in @xmath217 . then @xmath228 is the shear coordinates of some rational quasi - lamination by theorem [ q - lam bij ] . the shear coordinates of the individual curves in this quasi - lamination span a cone @xmath229 in @xmath213 . the cone @xmath229 is contained in some @xmath171-cone @xmath230 . if @xmath231 , then @xmath229 is contained in the largest cone @xmath232 of @xmath213 in @xmath129 . otherwise , @xmath233 , which is a @xmath171-cone by proposition [ int of b - cones ] , and whose dimension is lower than the dimension of @xmath129 . let @xmath234 be the largest cone of @xmath213 contained in @xmath235 . by induction , @xmath236 . but now @xmath234 is contained in the largest cone @xmath232 of @xmath213 in @xmath129 . in either case , @xmath228 is in @xmath215 . more generally than definition [ rat part def ] , ( * ? ? ? * definition 6.9 ) defines the @xmath3-part of a fan for an underlying ring @xmath3 , and ( * ? ? ? * proposition 6.10 ) establishes several facts about the @xmath3-part of a fan . the following corollary is obtained from ( * ? ? ? * proposition 6.10 ) by setting @xmath237 and appealing to theorem [ rat fb surfaces ] . [ r part prop ] suppose @xmath154 has the curve separation property . let @xmath165 be any tagged triangulation of @xmath154 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) . each rational @xmath171-cone is a cone in @xmath213 . the full - dimensional @xmath0-cones are exactly the full - dimensional cones in @xmath213 . the curve separation property has important consequences for finding positive @xmath4- and @xmath5-bases for @xmath171 . [ curves pos basis ] suppose @xmath154 has the curve separation property . let @xmath3 be @xmath4 or @xmath5 and fix a tagged triangulation @xmath165 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) . 1 . [ indep allowable ] if shear coordinates @xmath195 of allowable curves @xmath173 form an @xmath3-independent set for @xmath171 , then they form a positive @xmath3-basis for @xmath171 . [ basis allowable ] if a positive @xmath3-basis exists for @xmath171 , then the shear coordinates of allowable curves are a positive @xmath3-basis . to prove theorem [ curves pos basis ] , we use the following specialization of ( * ? ? ? * corollary 6.12 ) . the first part of proposition [ fb basis ray ] as stated in ( * ? ? ? * corollary 6.12 ) refers to the @xmath4-part ( as in ( * ? ? ? * definition 6.9 ) ) , not the rational part , but these two fans coincide . [ fb basis ray ] if a positive @xmath4-basis exists for @xmath0 , then the unique positive @xmath4-basis for @xmath0 consists of the smallest nonzero integer vector in each ray of the rational part of @xmath1 . if a positive @xmath5-basis exists for @xmath0 , then a collection of vectors is a positive @xmath5-basis for @xmath0 if and only if it consists of exactly one nonzero vector in each ray of the rational part of @xmath1 . for assertion , we verify the hypothesis of of proposition [ positive cone basis ] . suppose @xmath129 is a @xmath171-cone . by theorem [ rat fb surfaces ] , there is largest cone @xmath229 among cones of @xmath213 contained in @xmath129 and @xmath229 contains every point in @xmath150 . this verifies the hypothesis of proposition [ positive cone basis ] for @xmath237 . the hypothesis for @xmath10 follows by the integer case of theorem [ q - lam bij ] . assertion is an immediate consequence of theorem [ rat fb surfaces ] and proposition [ fb basis ray ] . ( theorem [ q - lam bij ] implies that the shear coordinate vector of an allowable curve is the shortest integer vector in the ray it spans . ) [ affine a2 surface ] consider an annulus with two marked points on its outer boundary and one marked point on its inner boundary . the surface admits a triangulation with three arcs , for example as shown in figure [ a2annulus ] . if @xmath165 is the triangulation shown , then @xmath171 is the matrix @xmath2380&1&1\\-1&0&1\\-1&-1&\hspace{6pt}0\\\end{bsmallmatrix*}$ ] . some allowable curves in the surface are shown ( green and dotted ) in figure [ a2annuluscurves ] . [ cols="^,^,^,^,^,^,^ " , ] all remaining allowable curves are related to the curves @xmath239 , @xmath240 , @xmath241 , and @xmath242 as we now explain . since each @xmath243 connects the inner boundary of the annulus to the outer boundary , we can modify @xmath173 by inserting spirals . specifically , for each @xmath244 , let @xmath245 be obtained from @xmath173 by tracing along @xmath173 from the inner boundary , inserting @xmath34 counterclockwise outward spirals , and then continuing along @xmath173 to the outer boundary . by convention , @xmath246 and if @xmath247 then @xmath245 is obtained from @xmath173 by inserting @xmath248 _ clockwise _ outward spirals . the curve @xmath249 is illustrated in figure [ a2annuluscurvespiral ] . keeping in mind that allowable curves are considered up to isotopy in @xmath151 relative to @xmath152 , one verifies that @xmath250 , that @xmath251 , that @xmath252 , and that @xmath253 . the set of all allowable curves is @xmath254 for @xmath255 , let @xmath256 be the shear coordinates of @xmath257 with respect to the triangulation shown in figure [ a2annulus ] . these shear coordinates are also indicated in figure [ a2annuluscurves ] . the shear coordinates of @xmath258 for @xmath259 are @xmath260 the shear coordinates of @xmath261 for n@xmath262 are @xmath263 , and similar descriptions hold for shear coordinates of @xmath264 and @xmath265 . the set of all shear coordinates of allowable curves consists of four infinite families @xmath266 for @xmath267 together with the vectors @xmath268 , @xmath269 , and @xmath270 . we will see in theorem [ curve separation ] that this surface has the curve separation property . thus theorem [ rat fb surfaces ] says that @xmath213 is the rational part of @xmath1 . in fact , @xmath213 is complete in this case , so it coincides with @xmath214 . the shear coordinates of allowable curves span the rays of @xmath1 , and the other faces of @xmath1 are spanned by sets of pairwise compatible allowable curves . the fan @xmath1 is depicted in figure [ fb a fig ] in the following way : ( 0,0)(135,-164 ) ( -69,98)@xmath271 ( -2.5,28.5)@xmath272 ( -38,8)@xmath273 ( -67,-102)@xmath274 ( 48,31)@xmath268 ( -5.5,-64.5)@xmath269 the intersection of a nonzero cone of @xmath1 with a unit sphere about the origin is a point , circular arc , or spherical triangle . the figure shows the stereographic projection of these points , arcs , and triangles . ( the vector @xmath275 $ ] projects to the origin of the plane . ) the ray for @xmath270 is not labeled but is marked as a green dot that appears as a limit of four sequences of red dots . with two exceptions , each @xmath7-dimensional face of @xmath1 is contained in exactly two @xmath158-dimensional faces of @xmath1 . the exceptions are the face spanned by @xmath268 and @xmath270 and the face spanned by @xmath269 and @xmath270 . in this section , as an aside , we further justify our interest in quasi - laminations by using results of @xcite to connect shear coordinates of allowable curves and @xmath276-vectors of cluster variables . given a tagged arc @xmath155 in @xmath154 , we define ( up to isotopy ) a curve @xmath277 . the curve @xmath277 coincides with @xmath155 except within a small ball around the endpoints of @xmath155 . within each ball around an endpoint @xmath181 , the behavior of @xmath277 is as follows : if @xmath181 is on a boundary component , then @xmath277 ends at a point @xmath278 on the same boundary component such the path along the boundary from @xmath181 to @xmath278 , keeping @xmath151 on the left , does not leave the small ball . if @xmath181 is a puncture , then @xmath277 spirals into @xmath181 , clockwise if @xmath155 is tagged plain at @xmath181 and counterclockwise if @xmath155 is tagged notched at @xmath181 . the map @xmath279 is similar to the map considered in ( * ? ? ? * definition 17.2 ) , but with the opposite orientation throughout . the following lemma is immediate . [ kappa lemma ] the map @xmath279 from tagged arcs to allowable curves is one - to - one and is surjective onto the set of allowable curves that are not closed . two arcs @xmath155 and @xmath156 are compatible if and only if the curves @xmath277 and @xmath280 are compatible . let @xmath165 be a tagged triangulation of @xmath154 . part of ( * ? ? ? * theorem 15.6 ) is a bijection between the tagged arcs in @xmath154 and the cluster variables of a cluster algebra with initial exchange matrix @xmath171 . ( if @xmath154 has exactly one puncture and no boundary components , then we assume @xmath165 has all tags plain and restrict the map to plain - tagged arcs . ) we write @xmath281 for this bijection , but we do not need the details of the map . rather , it is enough to know that tagged triangulations map to clusters . for a suitable choice of coefficients , each cluster variable @xmath282 has a @xmath276-vector of @xmath283 . this is an integer vector with entries indexed by the tagged arcs in @xmath165 , via the bijection between tagged arcs in @xmath165 and cluster variables in the initial cluster . for the definition of @xmath276-vectors , see ( * * section 6 ) . [ g surfaces ] fix a tagged triangulation @xmath165 and let @xmath155 be a tagged arc , not necessarily in @xmath165 . ( if @xmath154 has exactly one puncture and no boundary components , then take all tags on @xmath165 and on @xmath155 to be plain . ) then @xmath283 equals @xmath284 . the proof of proposition [ g surfaces ] revolves around the following weak version of ( * ? ? ? * conjecture 7.12 ) . [ gvec mu ] let @xmath285 be an edge in @xmath64 and let @xmath286 and @xmath287 be exchange matrices such that @xmath288 . for any @xmath94 and @xmath289 $ ] , the @xmath276-vectors @xmath290 and @xmath291 are related by @xmath292 , where @xmath293 is the transpose of @xmath286 . concatenating results of @xcite , as explained in ( * ? ? ? * remark 8.14 ) , we see that conjecture [ gvec mu ] is true whenever @xmath286 and @xmath287 are skew - symmetric , and in particular for exchange matrices arising from surfaces . furthermore , by skew - symmetry , we rewrite the assertion of conjecture [ gvec mu ] as @xmath294 , and then use to further rewrite the assertion as @xmath295 , for exchange matrices @xmath286 and @xmath287 related by mutation at position @xmath59 . alternately , let @xmath167 obtained from @xmath165 by a flip of the arc @xmath156 and let @xmath155 be any arc . then @xmath296 , where @xmath190 is the mutation map at the index for the arc @xmath156 in @xmath165 . on the other hand , suppose @xmath173 is an allowable curve . theorem [ mut q - lam ] says that the vector @xmath202 is @xmath297 , where @xmath190 is the mutation map at the index for the arc @xmath156 in @xmath165 . setting @xmath298 , we see that vectors @xmath299 and @xmath300 satisfy the same recurrence . when @xmath155 is an arc in @xmath165 , the vector @xmath300 has @xmath188 in the position indexed by @xmath155 and @xmath57 elsewhere . in this case , @xmath282 is the cluster variable indexed by @xmath155 in the initial cluster , so @xmath301 also has @xmath188 in the position indexed by @xmath155 and @xmath57 elsewhere . this completes the proof of proposition [ g surfaces ] . in this section , we prove the following theorem . [ curve separation ] if @xmath154 has one or more boundary components or if @xmath154 has two or more punctures , then @xmath154 has the curve separation property . the hypotheses rule out the case where @xmath154 has no boundary components and exactly one puncture . see remarks [ all arcs tagged plain ] and [ the missing case ] . as mentioned in the introduction , one additional case of the curve separation property is established in @xcite . to prove theorem [ curve separation ] , we take incompatible allowable curves @xmath173 and @xmath196 and construct a tagged triangulation @xmath165 and a tagged arc @xmath186 such that the shear coordinates @xmath180 and @xmath197 have strictly opposite signs . an allowable curve @xmath173 may intersect @xmath156 many times , but since @xmath173 may not intersect itself , all nonzero contributions to @xmath180 from these intersections have the same sign . thus to prove that @xmath180 and @xmath197 have strictly opposite signs , it is enough to find one intersection of @xmath173 with @xmath156 and one intersection of @xmath196 with @xmath156 that contribute opposite signs to @xmath180 and @xmath197 . we consider many cases . in most cases we construct , instead of a tagged triangulation @xmath165 , a triangulation @xmath184 . in these cases , we tacitly take @xmath165 to be the tagged triangulation @xmath302 . ( see definition [ tagged tri def ] . ) we avoid restating the conclusion ( that the shear coordinates @xmath180 and @xmath197 have strictly opposite signs ) in every case . in each case , part of the triangulation @xmath165 or @xmath184 is illustrated in a figure . to make the figures as clear as possible , we continue the convention on red , blue , and purple arcs and light - purple shading , established in definition [ shear def ] . the arc @xmath156 is always drawn in purple . arcs that are not used to calculate shear coordinates at @xmath156 are gray . the curve @xmath173 is solid and black , while @xmath196 , when it appears , is dotted and green . marked points are numbered as necessary . we shade areas outside of @xmath151 ( i.e. beyond the boundary ) with a dark gray . since our figures are confined to the ( topologically uncomplicated ) screen or page , we often need to indicate that certain loops are not contractible , even though on the page they appear to be contractible . to do so , we shade the inside of the loop with a light gray striped pattern . the figures show certain points and arcs as distinct when they might possibly coincide . when necessary , the proof considers the possible coincidences . in ruling out certain coincidences , it is important to remember that @xmath151 is an _ oriented _ surface . some arcs making up @xmath165 or @xmath184 are constructed by * _ following closely _ * along existing curves . this is meant as a technical term : a curve follows another closely if there are no punctures or boundary components between them . the point is to be certain that the arcs constructed do indeed bound triangles . in the figures , one might draw a curve tightly next to another curve to indicate that it follows closely . however , to aid legibility , in many cases we have straightened curves somewhat . because there are many cases , we first list and discuss the cases to make it clear that every possibility is covered . [ boundary case ] : : both endpoints of @xmath173 are on boundary segments . [ boundary spiral case ] : : @xmath173 has one endpoint on a boundary segment and one spiral point . [ closed case ] : : @xmath173 is a closed curve . by the symmetry of swapping @xmath173 and @xmath196 , in the remaining cases we are free to assume that both @xmath173 and @xmath196 have spiral points at both ends . [ spiral essential case ] : : @xmath173 and @xmath196 both have spiral points at both ends , and they have an intersection that can not be removed by changing the directions of spiral points . in the remaining cases , we can assume that the only intersections between @xmath173 and @xmath196 come at shared spiral points where their spiral directions disagree . [ spiral one not case ] : : @xmath173 and @xmath196 both have spiral points at both ends , and only intersect where they share spiral points with opposite directions . one spiral point of @xmath196 is not a spiral point of @xmath173 . again by symmetry , in the remaining case we are free to assume that the set of spiral points of @xmath173 equals the set of spiral points of @xmath196 . [ spiral equal case ] : : @xmath173 and @xmath196 both have spiral points at both ends , and only intersect where they share spiral points with opposite directions . the set of spiral points of @xmath173 equals the set of spiral points of @xmath196 . these cases exhaust the possibilities . [ the missing case ] when @xmath154 is a surface with no boundary components and exactly one puncture ( the case explicitly ruled out by hypothesis ) , to prove the curve separation property , we must construct a tagged triangulation @xmath165 and an arc @xmath156 as described above , with the additional requirement that @xmath165 has all arcs tagged plain . for such a surface @xmath154 , only [ closed case ] , [ spiral essential case ] , and [ spiral equal case ] apply . with a view towards eventually removing the extra hypothesis from the theorem , we give arguments in [ closed case ] and [ spiral equal case ] that also work for this kind of surface @xmath154 . the arguments work because when @xmath151 has no boundary components and exactly one puncture , no arc bounds a once - punctured monogon . thus when @xmath151 has no boundary components and exactly one puncture and we construct @xmath184 , we know automatically that @xmath165 has all arcs tagged plain . we are presently unable to prove [ spiral essential case ] for @xmath154 having no boundary components and exactly one puncture . we now proceed to consider each case . [ boundary case ] [ two boundary case ] we color the two boundary segments red and draw blue and purple arcs as shown in figure [ two boundary fig ] . ] the top blue arc starts at the point @xmath8 , closely follows the left red boundary segment down to @xmath173 , then closely follows @xmath173 to the right red boundary segment , and closely follows that boundary segment up to the point @xmath158 . we make the symmetric construction for the bottom blue arc . since the two red segments are distinct , the blue arcs are compatible or coincide . the purple arc is a diagonal of the red - blue - red - blue rectangle and connects points @xmath7 and @xmath158 . if the top blue arc bounds an unpunctured monogon , then points @xmath8 and @xmath158 coincide , and @xmath173 is not allowable because it is contractible to a piece of the boundary containing one marked point . if the top blue arc defines an unpunctured digon , then point @xmath158 is a neighbor to point @xmath8 on the same boundary component . in this case , we replace the top blue arc by the boundary segment ( which we again color blue ) whose endpoints are @xmath8 and @xmath158 . the symmetric argument shows that the bottom blue arc is valid or can be replaced by a boundary segment . the purple arc @xmath156 does not bound an unpunctured monogon because if points @xmath7 and @xmath158 coincide , then the two red segments coincide , contradicting the description of this case . if the purple arc defines an unpunctured digon , then one of the blue arcs bounds an unpunctured monogon , but this has already been ruled out . extend these arcs to a triangulation @xmath184 . the blue and purple arcs and the red segments define triangles in @xmath184 . if @xmath196 is a curve that intersects @xmath173 ( and if the intersection can not be avoided by independent isotopy deformations of the two curves ) , then @xmath196 must , somewhere along its extent , enter the pictured quadrilateral through a blue arc ( or originate on a blue segment ) , cross the purple arc , and exit through the opposite blue arc ( or end on the opposite blue segment ) . [ one boundary case ] the situation is illustrated in the top - left picture of figure [ one boundary fig ] . + color the boundary segment red and draw blue and purple arcs as shown in the top - right picture of figure [ one boundary fig ] . the blue arc follows closely the red segment from point @xmath8 to @xmath173 , turns left and follows @xmath173 closely without crossing it , and finally turns left and follows the red segment closely down to point @xmath7 . the purple arc @xmath156 follows the red segment closely up from point @xmath7 , crosses @xmath173 once and , the second time it meets @xmath173 , turns right and follows @xmath173 without crossing it until it meets the earlier part of the purple arc , then turns left and follows the red segment closely back down to @xmath7 without crossing itself . if the blue arc bounds an unpunctured monogon , then points @xmath8 and @xmath7 coincide and @xmath173 is not allowable because it is contractible to a piece of the boundary containing one marked point . if the blue arc defines an unpunctured digon , then points @xmath8 and @xmath7 are the only marked points on their boundary segment , and we replace the blue arc with the boundary segment connecting @xmath8 and @xmath7 that is not already red , coloring this segment blue . the purple arc does not bound an unpunctured monogon because @xmath173 is not contractible to a piece of the red boundary segment . if the purple arc defines an unpunctured digon , then the blue arc bounds an unpunctured monogon , but we have already ruled out that possibility . extend these arcs to a triangulation @xmath184 . in @xmath184 , there is a second triangle having the purple arc as an edge . if the second triangle is self - folded , then @xmath173 is an excluded curve of a type illustrated in figure [ exclude curves ] . thus the second triangle is not self - folded , and we color its other edges as in the bottom - left picture of figure [ one boundary fig ] . now suppose @xmath196 intersects @xmath173 . then @xmath196 must cross the blue arc ( or originate on the blue segment ) on the outer side of the picture and then cross either the inner red arc or the inner blue arc . if it crosses the inner blue arc , then the shear coordinates @xmath180 and @xmath197 have strictly opposite signs . otherwise , we alter the collection of arcs so that the purple arc and the inner red and blue arcs are incident to point @xmath8 instead of point @xmath7 . recoloring appropriately , we obtain the bottom - right picture of figure [ one boundary fig ] . now @xmath196 crosses the red arc on the outer side of the picture and then the inner red arc , so the two shear coordinates have strictly opposite signs . [ boundary spiral case ] suppose first that @xmath173 spirals clockwise into the puncture . color the boundary segment red and label its endpoints @xmath8 and @xmath7 as shown in figure [ boundary puncture fig ] . label the spiral point as @xmath158 . draw a blue arc from point @xmath8 , following the red segment closely to @xmath173 , following @xmath173 to point @xmath158 , going closely around point @xmath158 , following @xmath173 closely back to the red segment , and following the red segment to point @xmath7 . the blue arc does not bound an unpunctured monogon because @xmath154 is not a once - punctured monogon . if the blue arc defines an unpunctured digon , then points @xmath8 and @xmath7 are the only marked points on their boundary component and we replace the blue arc with the non - red boundary segment connecting point @xmath8 to point @xmath7 , coloring this segment blue . draw a purple arc from point @xmath7 to itself , following the red segment closely from point @xmath7 to point @xmath8 and then following the blue arc closely back to point @xmath7 . the purple arc does not bound an unpunctured monogon . it also does not define an unpunctured digon because @xmath154 is not a once - punctured monogon . now draw an arc from point @xmath7 to the puncture , without crossing the purple arc . this arc is indicated both in red and in blue in the picture to match the coloring conventions of figures [ shear fig ] and [ shear self - folded fig ] . complete these arcs to a triangulation @xmath184 . if @xmath196 is not compatible with @xmath173 , then there are two possibilities . the first possibility is that some portion of @xmath196 crosses the outer blue arc ( or originates in the outer blue segment ) then crosses the red / blue arc before crossing the outer blue arc again ( or terminating in the outer blue segment ) . in doing so , @xmath196 crosses the red - blue - red - blue quadrilateral twice , picking up a nonzero contribution to @xmath197 one of the two times . the second possibility is that @xmath173 crosses the outer blue arc ( or originates in the outer blue segment ) and then spirals counterclockwise into the puncture . in either case , we obtain the desired sign difference of shear coordinates at the purple arc @xmath156 . notice that in the case where @xmath173 and @xmath196 intersect but are compatible ( that is , when they agree except for the direction of the spiral into the puncture ) , the relevant shear coordinate for @xmath196 is zero . if @xmath173 spirals counterclockwise into the puncture , then we redraw the picture swapping the roles of points @xmath8 and @xmath7 , reversing left and right in drawing the arcs , and swapping the colors red and blue . [ closed case ] choose a direction to traverse @xmath173 . say a marked point @xmath181 is * _ left of @xmath173 _ * if , traversing @xmath173 in the chosen direction , there is a path that leaves @xmath173 , goes left , and reaches @xmath181 without crossing @xmath173 again . we define the notion * _ right of @xmath173 _ * analogously . for certain curves @xmath173 , it is possible for the same point to be both right of @xmath173 and left of @xmath173 . [ closed left right case ] label some marked point left of @xmath173 with the number @xmath8 . label some marked point right of @xmath173 with the number @xmath7 . possibly the points @xmath8 and @xmath7 coincide . we draw a blue arc that follows closely on the left some curve from point @xmath8 to the left of @xmath173 , turns left , follows @xmath173 closely on the left all around , turns left again and follows closely the curve back to point @xmath8 . we draw another blue curve in the same way on the right . these curves are shown in figure [ closed annulus fig ] . ] neither blue arc bounds an unpunctured monogon , because @xmath173 is not contractible to a puncture . if the blue arc at point @xmath8 defines an unpunctured digon , then point @xmath8 is on a boundary component and is the only marked point on that component . in this case , we replace the inner blue arc with the unique boundary segment on that component , which we then color blue . ( in this case , we should replace the light gray striped shading with dark gray shading . ) we proceed similarly if the blue arc at point @xmath7 defines an unpunctured digon . if points @xmath8 and @xmath7 coincide , then possibly the two blue curves coincide as well . if @xmath196 intersects @xmath173 , then somewhere along its extent it crosses the outer blue arc ( or originates on the outer blue segment ) and then crosses the inner blue arc ( or terminates on the inner blue segment ) . possibly @xmath196 winds around the inner blue arc a number of times before crossing it , but for simplicity , it is shown without winding in figure [ closed annulus fig ] . we draw a red arc from point @xmath8 to point @xmath7 without crossing @xmath196 as follows : the arc follows the inner blue arc closely counterclockwise to @xmath196 , then follows @xmath196 closely , without crossing it , to the outer blue arc , and finally follows closely the outer blue arc clockwise to point @xmath7 . the red and blue arcs define a quadrilateral with the red edges identified . we draw a purple arc @xmath156 that forms a diagonal of this quadrilateral . ( either diagonal works for the conclusion of the theorem , but one choice is consistent with our coloring scheme . ) we complete this set of arcs to a triangulation @xmath184 . [ closed not left right case ] without loss of generality , suppose all marked points are left of @xmath173 . since these points are not also right of @xmath173 , the surface decomposes at @xmath173 as a connected sum . in figure [ closed sum fig ] , we depict the situation with the `` left '' summand , containing marked points , outside of the curve @xmath173 and the `` right '' summand , containing no marked points , inside @xmath173 . since @xmath173 is not contractible , the right summand is topologically nontrivial . we choose a particular marked point , labeled @xmath8 , on the left and draw a blue arc closely following @xmath173 just as we drew the first blue arc in [ closed left right case ] . as in [ closed left right case ] , the blue arc does not bound an unpunctured monogon , and if it defines an unpunctured digon then we replace it with a blue boundary segment . if @xmath196 intersects @xmath173 , then @xmath196 must , somewhere along its extent , cross @xmath173 to enter the right summand , trace a nontrivial loop in the right summand , and then cross @xmath173 again to leave the right summand . in fact , @xmath196 may do this several times , but in the description below , we ignore any other pieces of @xmath196 that cross @xmath173 . the situation is depicted in the left picture of figure [ closed sum fig ] . in this picture , the entire region inside @xmath173 is shaded with light gray stripes to indicate both that the right connected summand is topologically nontrivial and that the closed path formed by part of @xmath196 and part of the blue arc is neither contractible to a point , nor isotopic to @xmath173 . now , draw a red arc and a purple arc as shown in the right picture of figure [ closed sum fig ] . the red arc follows the blue arc closely from point @xmath8 up to @xmath196 , follows @xmath196 closely without crossing it until reaching the blue arc again , and follows to blue arc back to the point @xmath8 . the purple arc follows the blue arc , crosses @xmath196 , and continues almost back to point @xmath8 before turning to follow @xmath196 closely , without immediately crossing @xmath196 again . it follows @xmath196 closely until it is close to the blue arc and then crosses @xmath196 to follow the blue arc back to point @xmath8 . the red and purple arcs do not bound unpunctured monogons or define unpunctured digons because the part of @xmath196 shown is not contractible to part of the blue arc . the red , blue , and purple arcs form a triangle , oriented consistently with the convention of figure [ shear fig ] . these arcs can be completed to a triangulation @xmath184 . in @xmath184 , there is a second triangle having the purple arc as an edge , as shown in figure [ closed sum fig 2 ] . since the point @xmath8 is the only marked point that can be reached without crossing the blue arc , all three vertices of the second triangle are at point @xmath8 . the second triangle is not self - folded , because all three of its vertices coincide . to compute @xmath180 , notice that @xmath173 intersects the red - blue - red - blue quadrilateral in at least two places . two intersections are shown in figure [ closed sum fig 2 ] . one of the two contributes @xmath57 to @xmath180 , but the other contributes @xmath8 . a similar situation arises in the calculation of @xmath197 . an explicit example of [ closed not left right case ] may be helpful . we take @xmath154 to be the double - torus with one puncture and take @xmath173 to be a curve along which @xmath151 decomposes as a connected sum of tori . this is a fairly representative example , since the right connected summand of @xmath151 in general is a nonempty connected sum of tori . specifically , let @xmath151 be the octagon with identifications indicated in the top - left picture of figure [ closed sum example fig ] . + all eight vertices of the octagon are identified , and we take this common point to be the puncture . the curves @xmath173 and @xmath196 are indicated as usual by a solid black and a dotted green line . the curve @xmath173 appears in four pieces that are identified to a closed curve ; the curve @xmath196 appears in two pieces . the top - right picture of figure [ closed sum example fig ] shows the blue , red , and purple arcs constructed first . the bottom picture shows these three arcs , straightened and completed to a triangulation of @xmath151 . [ spiral essential case ] draw a purple arc between the spiral points of @xmath173 , following @xmath173 closely except for the spirals . possibly the endpoints of the purple arc coincide , but the purple arc does not bound an unpunctured or once - punctured monogon because @xmath173 is not contractible to a spiral point and because @xmath173 is not an excluded curve in definition [ quasi lam def ] . choose a spiral point of @xmath196 and label it @xmath8 . label the spiral points of @xmath173 as @xmath7 and @xmath158 . let @xmath208 be the part of @xmath196 from the chosen spiral point to the first point where @xmath196 intersects @xmath173 , aside from any intersections that arise if point @xmath8 coincides with point @xmath7 and/or @xmath158 and spiral directions of @xmath173 and @xmath196 disagree there . draw a red arc that follows @xmath173 closely from point @xmath7 without crossing it ( except for the spiral ) to @xmath208 , then follows @xmath208 without crossing it ( except for the spiral ) to point @xmath8 . draw a blue arc from point @xmath158 to point @xmath8 closely following @xmath173 and then @xmath208 . these arcs are shown in the left picture of figure [ essential ] . until further notice , ignore the directions of spirals in the pictures . suppose the red arc bounds an unpunctured monogon . then points @xmath8 and @xmath7 coincide . furthermore , if the spiral directions of @xmath173 and @xmath196 agree at this point , then this first intersection of @xmath196 with @xmath173 can be removed by an isotopy deformation of @xmath196 . on the other hand , if spiral directions disagree at this point , then this first intersection can be removed by changing the directions of spiral points . either way , we have contradicted our construction of @xmath208 , so the red arc does not bound an unpunctured monogon . by the same argument , the blue arc also does not bound an unpunctured monogon . the red and blue arcs are not isotopic because if so , then @xmath173 is an excluded curve in definition [ quasi lam def ] . we extend these three arcs to a triangulation . in this triangulation , there is a second triangle having the purple arc as an edge . this second triangle is not self - folded , because if so , then @xmath173 is an excluded curve in definition [ quasi lam def ] . the situation is illustrated in the right picture in figure [ essential ] , still ignoring directions of spirals . possibly point @xmath153 coincides with some or all of the points @xmath8 , @xmath7 , and @xmath158 . let @xmath128 denote the red - blue - red - blue rectangle shown in the right picture of figure [ essential ] . after following @xmath208 , the curve @xmath196 exits @xmath128 either through the bottom blue arc or through the bottom red arc , possibly by spiraling into point @xmath153 . up to a global reorientation , we can assume that @xmath196 either spirals into point @xmath153 or exits through the bottom blue arc . suppose the bottom blue arc is isotopic to the top red arc . if @xmath196 spirals into point @xmath153 , then @xmath196 is an excluded curve in definition [ quasi lam def ] . if @xmath196 exits @xmath128 through the blue arc without spiraling into point @xmath153 , then the situation is as pictured in figure [ essential isotopic ] , still ignoring spiral directions . ] in this case , @xmath196 must spiral into point 2 , so it only intersects @xmath173 , if at all , by spiraling the other direction into point @xmath7 . we conclude from these contradictions that the bottom blue arc is not isotopic to the top red arc . possibly , however , the two red arcs are isotopic to each other , the two blue arcs are isotopic to each other , and/or the bottom red arc is isotopic to the top blue arc . in the triangulation containing these two triangles , there is another triangle sharing the top red arc as an edge , as shown in the left picture of figure [ essential alter ] . possibly this new triangle is self - folded , but if so , then the red arc is not one of the two edges that are folded onto each other . we flip the red arc to produce a new red - blue - purple triangle as shown in the middle picture of figure [ essential alter ] . let @xmath303 be the red - blue - red - blue rectangle consisting of the two triangles having the purple arc as an edge ( after the flip of the red edge ) . possibly the two red arcs of the original rectangle @xmath128 coincide . in this case , the bottom triangle of @xmath128 is also affected by the flip of the red arc , and so @xmath196 exits @xmath303 through its bottom blue arc , regardless of whether @xmath196 spirals into point @xmath153 or exits @xmath128 through the blue arc . if @xmath196 exits @xmath303 through the blue arc , then we make no more flips and define @xmath304 to be @xmath303 . otherwise , the two red arcs of @xmath128 do nt coincide , the bottom triangle of @xmath303 coincides with the bottom triangle of @xmath128 , and @xmath196 spirals into point @xmath153 . in this case , the bottom red arc of @xmath303 is not isotopic to the top blue arc of @xmath303 for the same reason that the top red arc of @xmath128 is not isotopic to the bottom blue arc of @xmath128 . the bottom red and blue arcs of @xmath303 are not isotopic ( i.e. the bottom triangle of @xmath303 is not self - folded ) for the same reason that the top red and blue arcs of @xmath128 are not isotopic . the two red arcs of @xmath303 may be isotopic . we flip the bottom red arc of @xmath303 , just as we flipped the top red arc of @xmath128 , to obtain a new rectangle @xmath304 , composed of two triangles sharing the purple arc as an edge , such that @xmath196 enters through the top blue arc and exits through the bottom blue arc . ( the fact that @xmath196 enters @xmath304 through the top blue arc follows from the fact that @xmath196 enters @xmath303 through the top blue arc , even in the case where the two red arcs of @xmath303 are isotopic . ) this construction of @xmath304 is illustrated in the right picture of figure [ essential alter ] . if any red or blue arc defining @xmath304 bounds a once - punctured monogon , we replace it by an arc from the vertex of the monogon to the puncture , tagged notched at the puncture . ( we have already ruled out the possibility that the purple arc @xmath156 bounds a once - punctured monogon . ) the spiral directions of spiral points of @xmath196 are now irrelevant . if the spiral direction of @xmath173 at point @xmath7 disagrees with the direction shown in figure [ essential alter ] , then we tag all arcs notched at point @xmath7 and similarly for point @xmath158 . we can do this even if points @xmath7 and @xmath158 coincide , because in that case , @xmath173 has the same spiral directions at both ends . we complete the tagged arcs to a tagged triangulation @xmath165 , change all tags to plain ( reversing spiral directions accordingly ) and apply @xmath305 to make a triangulation to calculate @xmath180 and @xmath197 as in definition [ shear def ] . this latter triangulation contains the two triangles of @xmath304 and the altered spiral directions are also as shown in the figure . [ spiral one not case ] [ spiral one not distinct case ] number the spiral points so that @xmath173 spirals into @xmath8 and @xmath7 and @xmath196 spirals into @xmath7 and @xmath158 . draw a red arc and a blue arc from @xmath8 to @xmath158 as shown in the left picture of figure [ 3 spirals fig ] . , title="fig:"],title="fig : " ] for the moment , ignore the directions of spirals in the picture . each arc follows @xmath173 and then @xmath196 closely . these two arcs are not isotopic to each other because otherwise @xmath154 is a sphere with @xmath158 punctures . also draw two gray arcs from @xmath7 to @xmath158 , following @xmath196 closely and staying between the red arc and the blue arc . tag one arc plain at point @xmath7 and tag the other arc notched at point @xmath7 . none of these arcs bounds a monogon because points @xmath8 , @xmath7 , and @xmath158 are distinct . tag the arcs at points @xmath8 and/or @xmath158 notched if the spiral directions of @xmath173 or @xmath196 disagree with figure [ 3 spirals fig ] at these points . complete to a tagged triangulation @xmath165 . if the spiral directions at point @xmath7 agree with figure [ 3 spirals fig ] , then we take @xmath156 to be the gray arc tagged notched at point @xmath7 . if the spiral directions at point @xmath7 disagree with figure [ 3 spirals fig ] , then we take @xmath156 to be the gray arc tagged plain at point @xmath7 . in either case , to calculate the shear coordinates @xmath180 and @xmath197 , we pass to a triangulation and spiral directions as in the right picture of figure [ 3 spirals fig ] . the purple arc in the right picture is @xmath156 . the other gray arc in the left picture , besides @xmath156 , becomes the arc in the right picture that we color red on one side and blue on the other side as in figure [ shear self - folded fig ] . [ spiral one not coincide case ] the curve @xmath173 has both spiral points in the same direction at the same point ( labeled @xmath8 in figure [ curve and loop ] ) . we first draw a red arc as shown in the left picture in figure [ curve and loop ] , starting at point @xmath7 , following @xmath196 closely , crossing @xmath173 twice near point @xmath8 , following @xmath173 closely back to the earlier part of the red arc , then following the red arc closely back to point @xmath7 . this red arc does not bound an unpunctured monogon because if so , then @xmath173 is contractible to the puncture @xmath8 . suppose this red arc bounds a once - punctured monogon . then the puncture is either point @xmath8 or a marked point in the region shaded striped gray in figure [ curve and loop ] . in either case , @xmath173 is an excluded curve in definition [ quasi lam def ] , so we conclude that the red arc does not bound a once - punctured monogon . we next draw a blue arc as shown , following the red arc closely from point @xmath7 to point @xmath8 and then following @xmath196 closely back to point @xmath7 . if the blue arc bounds an unpunctured monogon , then @xmath173 is an excluded curve in definition [ quasi lam def ] . if the blue arc bounds a once - punctured monogon , then replace it with an arc from point @xmath7 to the puncture , tagged notched at the puncture . now tag appropriately at points @xmath8 and @xmath7 to obtain the correct spiral directions and produce a triangulation , similarly to [ spiral one not distinct case ] . part of the triangulation is shown in the right picture of figure [ curve and loop ] . [ spiral equal case ] [ spiral equal two coincide case ] in fact , since @xmath173 and @xmath196 are incompatible , the spiral directions must disagree at both points . draw a purple arc , as in figure [ two coincide fig ] , following the two curves closely . since the purple arc has 2 distinct endpoints , it does not bound a monogon . the purple arc can be completed to some triangulation . if the purple arc constitutes two sides of a self - folded triangle in the triangulation , then we flip the third side to obtain a new triangulation in which the purple arc is an edge of two distinct triangles . these two triangles are represented in figure [ two coincide fig ] , although any of the points shown may coincide ( except points @xmath8 and @xmath7 ) , and various combinations of the arcs shown may coincide . we apply @xmath159 to the triangulation . the resulting tagged triangulation has plain tags at points @xmath8 and @xmath7 because the purple arc is not part of a self - folded triangle . we change the plain tagging at point @xmath8 and/or point @xmath7 to notched tagging if the spiral directions at these points do not agree with figure [ two coincide fig ] . the tagged triangulation with these altered taggings is @xmath165 . when shear coordinates are calculated for the purple arc @xmath156 , we pass to a triangulation and spiral directions as shown in figure [ two coincide fig ] . [ spiral equal two one not coincide case ] label the two spiral points as @xmath8 and @xmath7 . draw two gray arcs from @xmath8 to @xmath7 , following @xmath173 closely . tag these two arcs oppositely at point @xmath7 . draw a red arc following @xmath173 closely from point @xmath8 to point @xmath7 and then following @xmath196 closely from point @xmath7 back to point @xmath8 . draw a blue arc similarly on the other side of @xmath173 and @xmath196 . the red and blue arcs are not isotopic , because if so @xmath154 is a twice - punctured sphere . since @xmath173 and @xmath196 do nt coincide , even ignoring spiral directions , neither the red arc nor the blue arc bounds an unpunctured monogon . if either or both arcs bounds a once - punctured monogon , then replace it with an arc from @xmath8 to the puncture , tagged notched at the puncture . tag arcs notched at point @xmath8 if the spiral directions disagree with those shown in figure [ two spirals disagree once ] and complete to a tagged triangulation @xmath165 . if the spiral directions at point @xmath7 agree with figure [ two spirals disagree once ] , then the gray arc tagged notched at point @xmath7 becomes the purple arc in the right picture of figure [ 3 spirals fig ] and the gray arc tagged plain at @xmath7 becomes the red / blue arc . if the spiral directions at point @xmath7 disagree with figure [ two spirals disagree once ] , the two gray arcs reverse roles . the calculation of shear coordinates at the purple arc involves arcs and spiral directions as shown in the right picture of figure [ two spirals disagree once ] . [ spiral equal two two not coincide case ] label the two spiral points as @xmath8 and @xmath7 . draw a blue arc , as in the top - left picture in figure [ two spirals disagree twice ] , following @xmath173 closely . + draw a red arc , as shown , following @xmath196 closely . the red and blue arcs can be completed to some triangulation . this triangulation has two triangles containing the blue arc as an edge . ( the presence of the red edge rules out the possibility that a triangle is folded onto the blue edge . ) one of these triangles is shown in the top - right picture of figure [ two spirals disagree twice ] . because points @xmath8 and @xmath7 do nt coincide , we assume up to symmetry that points @xmath8 and @xmath158 do nt coincide . thus the purple arc does not bound a monogon . possibly points @xmath7 and @xmath158 coincide but the new red arc does not bound an unpunctured monogon because it is part of some triangulation . if the new red arc bounds a once - punctured monogon , then we replace the new red arc with a tagged arc from point @xmath7 to the puncture . for the moment , we assume that the purple arc is not isotopic to the original red arc . the case where the two are isotopic is dealt with separately below . now draw another blue arc from point @xmath158 to point @xmath7 , closely following the purple arc to point @xmath8 and then closely following the red arc to point @xmath7 . this arc is shown in the bottom - left picture of figure [ two spirals disagree twice ] in the case where points @xmath7 and @xmath158 do not coincide , and in the bottom - right picture in the case where points @xmath7 and @xmath158 coincide . in the latter case , we need to consider the possibility that the new blue arc bounds an unpunctured or once - punctured monogon . but if the new blue arc bounds an unpunctured monogon , then the original red arc is isotopic to the purple arc , a possibility that we deal with separately below . if the new blue arc bounds a once - punctured monogon , then replace it with an arc from point 2 to the puncture , tagged notched at the puncture . now tag arcs at points @xmath8 and/or @xmath7 notched if the spiral directions at these points disagree with the figure , and complete to a tagged triangulation @xmath165 . to calculate shear coordinates @xmath180 and @xmath197 , we pass to a triangulation and spiral directions as shown in one of the bottom pictures of figure [ two spirals disagree twice ] . finally , we must consider the case where the original red arc is isotopic to the purple arc . figure [ two spirals disagree twice 2 ] illustrates this case . the purple arc is omitted in favor of the original red arc . arcs are also colored differently , with the red and blue arcs in the left picture of figure [ two spirals disagree twice 2 ] having switched colors from the top - left picture of figure [ two spirals disagree twice ] . the smaller red arc from the bottom - right picture of figure [ two spirals disagree twice ] is now colored purple . we complete these three arcs to a triangulation . in this triangulation , there is a second triangle having the purple arc as an edge , as illustrated in the right picture of figure [ two spirals disagree twice 2 ] . the new blue and red arcs may be isotopic , in which case the purple arc bounds a once - punctured monogon . in this case , we replace the purple arc with an arc from point 2 to the puncture , tagged notched at the puncture . again we tag arcs at points @xmath8 and/or @xmath7 notched if the spiral directions at these points disagree with the figure , and complete to a tagged triangulation @xmath165 . when shear coordinates are calculated for the purple arc @xmath156 , we pass to a triangulation and spiral directions as shown in the right picture of figure [ two spirals disagree twice 2 ] . the white space between the red and blue arcs is not shaded striped gray because the red and blue arcs may coincide . [ spiral equal one coincide case ] draw a purple arc following @xmath173 ( and thus @xmath196 ) closely . the purple arc does not bound an unpunctured or once - punctured monogon because @xmath173 is not contractible to a puncture and is not an excluded curve in definition [ quasi lam def ] . we complete the purple arc to a triangulation @xmath184 . since the purple arc does not have two distinct endpoints , it does not constitute two edges of a self - folded triangle . thus it is contained in two distinct triangles in @xmath184 . furthermore , neither of these triangles is self - folded , because otherwise @xmath173 is an excluded curve in definition [ quasi lam def ] . the relevant part of @xmath184 is shown in figure [ one coincide fig ] . up to switching @xmath173 with @xmath196 , the spirals are as shown in the figure . [ spiral equal one not coincide case ] the common spiral point is labeled @xmath8 . draw a gray arc following @xmath173 closely and a purple arc following @xmath196 closely . + + then draw a second gray arc following @xmath173 and then @xmath196 as shown in the top picture of figure [ one spirals disagree ] . ( there are two ways to draw the second gray arc relative to the spiral directions of @xmath173 and @xmath196 . the right way is shown in the figure . ) neither the first gray arc nor the purple arc bounds an unpunctured or once - punctured monogon because because neither @xmath173 nor @xmath196 is contractible to a spiral point or an excluded curve in definition [ quasi lam def ] . the second gray arc does not bound an unpunctured monogon because @xmath173 and @xmath196 do nt coincide , even when spiral directions are ignored . for the same reason , the first gray arc and the purple arc are not isotopy equivalent . the second gray arc is not equivalent to the other two arcs because it is their product in the sense of the fundamental group and neither of the other two arcs is trivial . these three distinct arcs form a triangle and can be completed to a triangulation . in this triangulation , there is another triangle having the first gray arc as an edge . this new triangle is not self - folded because the first gray edge does not bound a once - punctured monogon . it is shown in the middle picture of figure [ one spirals disagree ] , with one of the new edges red and one gray . the point @xmath7 may coincide with the point @xmath8 . we now have two triangles sharing an edge ( the first gray edge ) . we flip this shared edge , coloring the new edge blue , as shown in the bottom picture of figure [ one spirals disagree ] . this new configuration of five arcs can also be extended to a new triangulation . the new triangulation has another triangle sharing the purple arc as an edge . as before , the new triangle is not self - folded . it is also shown in the bottom picture of figure [ one spirals disagree ] with the new edges colored red and blue . the point @xmath158 may coincide with point @xmath8 and/or point @xmath7 . the triangulation constructed has the desired properties with respect to shear coordinates at the purple arc @xmath156 . in this section , we formulate the null tangle property of a marked surface , and show that it is equivalent to the assertion that the shear coordinates @xmath195 of allowable curves @xmath173 constitute a positive @xmath3-basis for @xmath171 . we observe that the curve separation property follows from the null tangle property . the observation suggests that it might be quite difficult to establish the null tangle property in general , since the proof of the curve separation property ( theorem [ curve separation ] ) was so complex , even leaving out one family of marked surfaces . here , we prove the null tangle property for a smaller family of marked surfaces . in @xcite , the null tangle property is established for an additional marked surface : the once - punctured torus . [ weighted tangle def ] a * _ ( weighted ) tangle ( of curves ) _ * in @xmath154 is a finite collection of curves , all distinct up to isotopy and all allowable in the sense of definition [ quasi lam def ] , with no requirement of compatibility between curves , and with each curve @xmath173 having a * _ weight _ * @xmath306 . we do not require weights to be positive as we did in the definitions of laminations and quasi - laminations ( definitions [ lam def ] and [ quasi lam def ] ) . the * _ support _ * of a tangle is the set of curves having nonzero weight . an * _ straight tangle _ * is a tangle consisting of pairwise compatible curves . ( the difference between a quasi - lamination and a straight tangle is that the weights in an straight tangle are allowed to be negative . ) a tangle is * _ trivial _ * if all weights are zero . the * _ weighted union _ * of tangles @xmath307 and @xmath308 is constructed by first taking the multiset union of @xmath307 and @xmath308 . then , for each curve @xmath173 that appears in both @xmath307 and @xmath308 , we replace the two copies of @xmath173 with a single copy of @xmath173 whose weight is the sum of the original two weights . [ null tangle def ] a tangle @xmath309 has shear coordinates @xmath310 with respect to any tagged triangulation @xmath165 , given by the weighted sum of vectors @xmath195 over all of the curves @xmath311 . each @xmath195 is calculated as in definition [ shear def ] . if @xmath154 has either one or more boundary components or two or more punctures or both , then a * _ null tangle _ * in @xmath154 is a tangle @xmath309 such that @xmath310 is the zero vector for all tagged triangulations @xmath165 . if @xmath154 has no boundary components and exactly one puncture , then a * _ null tangle _ * is a tangle @xmath309 such that @xmath310 is the zero vector for all tagged triangulations @xmath165 with all tags plain . a marked surface has the * _ null tangle property _ * if every null tangle is trivial . the main results of this section are the following theorems and an immediate corollary . [ null tangle theorem ] suppose @xmath3 is @xmath4 or @xmath5 and let @xmath165 be a tagged triangulation of @xmath154 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) . the following are equivalent . a. [ null tangle theorem null ] @xmath154 has the null tangle property . b. [ null tangle theorem basis ] the shear coordinates of allowable curves form an @xmath3-basis for @xmath171 . c. [ null tangle theorem pos basis ] the shear coordinates of allowable curves form a positive @xmath3-basis for @xmath171 . [ poly grow null tangle ] if @xmath154 is a sphere having @xmath164 boundary components and @xmath181 punctures with @xmath312 , then @xmath154 has the null tangle property . the hypothesis of theorem [ poly grow null tangle ] is that @xmath154 is a disk with @xmath57 , @xmath8 , or @xmath7 punctures , an annulus with @xmath57 or @xmath8 punctures , or a sphere with three boundary components . [ poly grow univ ] if @xmath154 is a sphere having @xmath164 boundary components and @xmath181 punctures with @xmath312 , then the shear coordinates of allowable curves form a positive @xmath3-basis for @xmath0 , for @xmath10 or @xmath5 . [ poly grow remark ] comparing with ( * ? ? ? * proposition 11.2 ) , we see that the marked surfaces identified in theorem [ poly grow null tangle ] and corollary [ poly grow univ ] are precisely those whose associated cluster algebra has polynomial growth . ( see ( * ? ? ? * section 11 ) . ) given that the null tangle property is established for an additional surface in @xcite , the connection to polynomial growth is probably coincidental . [ finite null tangle remark ] the marked surfaces of theorem [ poly grow null tangle ] and corollary [ poly grow univ ] include those of finite type in the sense of ( * ? ? ? * section 1.3 ) ( the cases of an unpunctured or once - punctured disk ) . we recover a weaker version of ( * ? ? ? * proposition 17.5 ) , giving a proof entirely in terms of triangulated surfaces . [ affine null tangle remark ] the marked surfaces of theorem [ poly grow null tangle ] and corollary [ poly grow univ ] also include those of affine type in the sense of ( * ? ? ? * definition 10.9 ) ( the cases of a twice - punctured disk or unpunctured annulus ) . for exchange matrices @xmath0 of affine type , it is conjectured ( * ? ? ? * conjecture 10.10 ) that the @xmath276-vectors associated to @xmath313 , together with an additional integer vector , are a positive @xmath3-basis for @xmath0 , for any @xmath3 . the results of this section establish the marked surfaces case of ( * ? ? ? * conjecture 10.10 ) , using proposition [ g surfaces ] and the fact that there is a unique allowable closed curve in these cases . example [ affine a2 surface ] . the point of defining null tangles is the connection to @xmath0-coherent linear relations . [ null tangle is b coher rel ] a tangle @xmath309 in @xmath154 is null if and only if @xmath314 is a @xmath171-coherent linear relation for any tagged triangulation @xmath165 ( with all tags plain if @xmath154 has no boundary components and exactly one puncture ) . by theorem [ mut q - lam ] , a tangle @xmath309 is null if and only if @xmath314 is a nontrivial linear relation among the shear coordinates of allowable curves that is preserved ( in the sense of ) under mutation maps at @xmath171 . now proposition [ no zero no piecewise ] says that @xmath309 is null if and only if @xmath314 is a @xmath171-coherent linear relation . proposition [ null tangle is b coher rel ] implies the following rephrasing of the null tangle property . [ null tangle prop ] suppose @xmath3 is @xmath4 or @xmath5 and let @xmath165 be a tagged triangulation of @xmath154 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) . then @xmath154 has the null tangle property if and only if the shear coordinates of allowable curves constitute an @xmath3-independent set for @xmath171 . we now quote a version of ( * ? ? ? * proposition 4.12 ) . the proposition is stated in @xcite in terms of @xmath0-coherent linear relations , but we use proposition [ null tangle is b coher rel ] to state a specialization of the proposition in terms of tangles . [ one positive or one negative ] let @xmath309 be a tangle in @xmath154 . suppose for some tagged triangulation @xmath165 ( with all tags plain if @xmath154 has no boundary components and exactly one puncture ) , for some tagged arc @xmath156 in @xmath165 , and for some curve @xmath173 in @xmath309 that @xmath180 is strictly positive ( resp . strictly negative ) and that @xmath197 is nonpositive ( resp . nonnegative ) for every other @xmath315 . then @xmath316 . we now explain the connection between the null tangle property and the curve separation property . the * _ disorder _ * of a tangle @xmath309 is the smallest number @xmath59 such that , after deleting curves with weight zero , @xmath309 is a weighted union of @xmath59 straight tangles . then the null tangle property is the statement that the disorder of a null tangle must be @xmath57 . one approach to establishing the property would be to rule out null tangles of disorder @xmath59 for all positive @xmath59 . the curve separation property is closely related to ruling out null tangles of certain small disorders . [ no 1 2 ] null tangles of disorder @xmath8 do not exist in any marked surface . if @xmath154 has the curve separation property , then null tangles of disorder @xmath7 do not exist in @xmath154 . theorem [ rat fb surfaces ] , specifically the fact that @xmath213 is simplicial , implies that a null tangle can not have disorder @xmath8 . suppose @xmath154 has the curve separation property . if a null tangle @xmath309 has disorder @xmath7 , then it is a weighted union of two non - trivial straight tangles @xmath307 and @xmath308 , and there is a curve @xmath173 in the support of @xmath307 and curve @xmath196 in the support of @xmath308 such that @xmath173 and @xmath196 are not compatible . the curve separation property says that there is a tagged triangulation @xmath165 and an arc @xmath156 ( with all arcs tagged plain if @xmath154 has no boundary components and exactly one puncture ) such that @xmath180 and @xmath197 have strictly opposite signs . without loss of generality , take @xmath180 positive and @xmath197 negative . proposition [ fbt compatible ] says that @xmath307 is contained in a @xmath171-cone , so proposition [ contained bcone ] says that @xmath317 is nonnegative for every curve @xmath182 in @xmath307 . similarly @xmath318 is nonpositive for every curve @xmath208 in @xmath308 . since @xmath309 is a null tangle , we have @xmath319 . let @xmath167 be the tagged triangulation obtained by flipping @xmath156 in @xmath165 . since the shear coordinates @xmath317 all have weakly the same sign , theorem [ mut q - lam ] ( applied separately to each curve in @xmath307 ) implies that @xmath320 . for the same reason , @xmath321 . but @xmath171 does not have a row consisting all of zeros , so it does not have a column consisting all of zeros . thus since @xmath322 and @xmath323 have strictly opposite signs in the entry indexed by @xmath156 , looking at we conclude that @xmath324 . this contradicts the assumption that @xmath309 is a null tangle . [ null disorder 3 curve sep ] if @xmath154 does not admit null tangles of disorder @xmath7 or @xmath158 , then @xmath154 has the curve separation property . suppose the curve separation property fails . then there exist incompatible curves @xmath173 and @xmath196 such that @xmath200 and @xmath201 have weakly the same sign for all @xmath167 and @xmath325 . ( if @xmath154 has no boundary components and exactly one puncture , then instead we know that @xmath200 and @xmath201 have weakly the same sign for all @xmath167 with all arcs tagged plain . ) applying theorem [ mut q - lam ] , we see that @xmath326 and @xmath327 form a sign - coherent set for any sequence @xmath120 of indices . proposition [ contained bcone ] says that @xmath195 and @xmath204 are contained in some @xmath0-cone , so @xmath328 equals @xmath329 for any @xmath120 by proposition [ linear ] . now let @xmath175 be the unique quasi - lamination with @xmath330 . let @xmath307 be the tangle containing only the curve @xmath173 with weight @xmath8 and let @xmath308 be the tangle containing only @xmath196 with weight @xmath8 . let @xmath331 be the tangle obtained from @xmath175 by negating all weights . the weighted union of @xmath307 , @xmath308 , and @xmath331 is a null tangle of disorder @xmath332 . but @xmath333 because @xmath173 and @xmath196 are incompatible . proposition [ no 1 2 ] rules out the possibility @xmath334 , so @xmath335 is @xmath7 or @xmath158 . [ null implies sep ] if @xmath154 has the null tangle property , then it also has the curve separation property . combining results , we obtain theorem [ null tangle theorem ] . given , theorem [ curves pos basis ] , corollary [ null implies sep ] and proposition [ null tangle prop ] prove , which trivially implies , which implies by proposition [ null tangle prop ] . the relationship between the curve separation property and the null tangle property suggests that we might extend the proof of theorem [ curve separation ] to establish the null tangle property under the hypotheses of theorem [ curve separation ] . while this seems hard in general , some of the easier cases of the proof of the theorem extend to partial statements about the null tangle property , which lead to the proof of theorem [ poly grow null tangle ] . [ null tangle two boundary ] if @xmath309 is a null tangle , then no curve in the support of @xmath309 connects two distinct boundary segments . suppose @xmath173 is an allowable curve connecting two distinct boundary segments . construct a tagged triangulation @xmath167 and distinguish a ( purple ) arc @xmath325 as in [ two boundary case ] of the proof of theorem [ curve separation ] . ( see figure [ two boundary fig ] . ) we observe that @xmath173 is the unique allowable curve whose shear coordinate @xmath180 is strictly positive and apply proposition [ one positive or one negative ] to conclude that @xmath173 can not appear in the support of a null tangle . [ null tangle boundary puncture ] if @xmath309 is a null tangle , then no curve in the support of @xmath309 has one endpoint on a boundary segment and spirals into a marked point at the other end . suppose @xmath173 is an allowable curve having one endpoint on a boundary segment and a spiral point at the other end . construct @xmath167 and the purple arc @xmath325 as in [ boundary spiral case ] of the proof of theorem [ curve separation ] . ( see figure [ boundary puncture fig ] . ) again , we observe that @xmath173 is the unique allowable curve whose shear coordinate @xmath180 is strictly positive and apply proposition [ one positive or one negative ] . [ null tangle one boundary ] suppose @xmath154 is a sphere having @xmath164 boundary components and @xmath181 punctures , with @xmath312 . if @xmath309 is a null tangle , then no curve in the support of @xmath309 has both endpoints on the same boundary segment . suppose @xmath309 is a null tangle and let @xmath173 be a curve in @xmath309 with both endpoints on the same boundary segment . in particular , @xmath154 is not a disk or once - punctured disk . construct @xmath167 and the purple arc @xmath325 in the bottom - left picture of figure [ one boundary fig ] , in [ one boundary case ] of the proof of theorem [ curve separation ] . suppose @xmath196 is a curve in @xmath309 with @xmath336 . then @xmath196 has an endpoint on the same boundary segment as @xmath173 , so propositions [ null tangle two boundary ] and [ null tangle boundary puncture ] imply that @xmath196 has both endpoints on the same boundary segment as @xmath173 . if @xmath154 is a disk with @xmath7 punctures or an unpunctured annulus , then each boundary segment admits a unique allowable curve with both endpoints on that boundary segment . thus @xmath173 is the unique curve in @xmath309 with @xmath337 . we apply proposition [ one positive or one negative ] to see that @xmath173 is not in the support of @xmath309 . if @xmath154 is a sphere with three boundary components or a once - punctured annulus , then we argue in several steps , looking at the bottom - left picture of figure [ one boundary fig ] . if point @xmath158 is the puncture and the other boundary component is not contained in the gray striped area , then @xmath173 is not allowable in definition [ quasi lam def ] . if point @xmath158 is on a boundary component and no marked point off of that boundary component is contained in the gray striped area , then any curve @xmath196 in @xmath309 with @xmath197 positive is isotopic to @xmath173 . thus again @xmath173 is the unique curve in @xmath309 whose shear coordinate @xmath180 is strictly positive , so again @xmath173 is not in the support of @xmath309 . we are left with a unique choice of @xmath173 that might appear in the support of @xmath309 , namely the curve such that the area shaded striped gray contains two boundary components or a boundary component and a puncture . but then @xmath173 would be the unique curve in @xmath309 with shear coordinate @xmath180 strictly positive , so it is not in the support of @xmath309 . theorem [ curve separation ] says that the curve separation property holds , so by proposition [ no 1 2 ] , null tangles of disorder @xmath8 and @xmath7 do not exist . if @xmath154 is a disk with @xmath57 or @xmath8 punctures , then propositions [ null tangle two boundary ] and [ null tangle boundary puncture ] amount to the null tangle property . in the remaining cases , propositions [ null tangle two boundary ] , [ null tangle boundary puncture ] , and [ null tangle one boundary ] show that the support of a null tangle has no curves that involve boundary segments . if @xmath154 is an unpunctured annulus , there is one remaining allowable curve ( the closed curve ) . a null tangle has disorder at most @xmath8 , and therefore has disorder @xmath57 . if @xmath154 is a disk with @xmath7 punctures , then there are five remaining allowable curves : one closed curve and four curves with distinct spiral points . these can be arranged into two sets of pairwise compatible curves . ( group together the two curves with counterclockwise spirals at one of the punctures , group together the two curves with clockwise spirals that same puncture , and put the closed curve in either group . ) a null tangle has disorder at most @xmath7 , and therefore has disorder @xmath57 . if @xmath154 is a once - punctured annulus , then there are four remaining allowable curves : two closed curves and two curves which spiral into the puncture at both ends . a null tangle has disorder at most @xmath7 and thus has disorder @xmath57 . if @xmath154 is a sphere with three boundary components , then the only remaining allowable curves are three compatible closed curves , so a null tangle has disorder at most @xmath8 , and therefore has disorder @xmath57 . the author thanks sergey fomin for helpful converations . the author also thanks an anonymous referee for many helpful comments , and especially for pointing out that an earlier version of ( * ? ? ? * proposition 4.6 ) ( appearing here as proposition [ basis exists ] ) was incorrect .

in this paper new results from the theory of electromagnetic radiation emitted by a bunch of ultra - relativistic positrons channeling through a periodically deformed crystal ( a crystalline undulator ) are reported . we formulate the approximation for effective analytical and numerical analysis of the characteristics of the undulator radiation with account for the influence of two main parasitic effects , the positron dechanneling and the photon attenuation . the developed formalism is applied to calculate the number of the emitted photons and the brilliance of the radiation formed in crystalline undulators . in a crystalline undulator there appears , in addition to a well - known channeling radiation , the radiation of an undulator type which is due to the periodic motion of channeling particles which follow the bending of the crystallographic planes . the parameters of the undulator radiation can be easily varied by changing the energy of beam particles and the parameters of crystal bending . the feasibility of this scheme was explicitly demonstrated for in refs . . in these papers as well as in the subsequent publications @xcite the idea of this new type of radiation , the essential conditions and limitations which must be fulfilled to make possible the observation of the effect were formulated in an adequate form for the first time . a number of corresponding numerical results were presented to illustrate the developed theory . the importance of the ideas suggested and discussed in the cited papers has also been realized by other authors resulting in a significant increase of the number of publications in the field during last years @xcite but , unfortunately , often without proper citation @xcite . a detailed review of the results obtained in this newly arisen field as well as a historical survey of the development of all principal ideas and related phenomena can be found in ref . . ) which are periodically bent . the centerline of this channel ( solid line ) is described by a harmonic function @xmath0 . its period @xmath1 and amplitude @xmath2 satisfy the condition @xmath3 . the dashed curve represents the trajectory of a projectile trapped in the channel . , width=491,height=207 ] the mechanism of the photon emission by means of a crystalline undulator is illustrated in fig . [ figure1_jpg ] . the @xmath4-plane in the figure is a cross section of an initially linear crystal , and the @xmath5-axis represents the cross section of a midplane of two neighbouring non - deformed crystallographic planes ( not drawn in the figure ) spaced by the interplanar distance @xmath6 . under certain conditions the ultra - relativistic positrons will channel in the periodically bent channel . the trajectory of a particle contains two elements . firstly , there are channeling oscillations due to the action of the interplanar potential . their typical frequency @xmath7 depends on the positron energy @xmath8 and the parameters of the interplanar potential . secondly , there are oscillations related to the periodicity of the distorted midplane , - the undulator oscillations , whose frequency is @xmath9 . the spontaneous emission of photons is associated with both of these oscillations . the typical frequency of the channeling radiation is @xmath10 where @xmath11 is the relativistic lorenz factor . the undulator oscillations give rise to the photons with frequency @xmath12 where @xmath13 is the undulator parameter , @xmath14 @xcite . if strong inequality @xmath15 is met than the frequencies of the channeling radiation and the undulator radiation are also well separated , @xmath16 . in this case the characteristics of the undulator radiation are practically independent on the channeling oscillations but depend on the shape of the periodically bent midplane @xcite . there are essential features which distinguish a crystalline undulator from a conventional one based on the action of the periodic magnetic ( or electric ) field on the projectile . in the latter the beam of particles and the photon flux move in vacuum whereas in the proposed scheme they propagate through a crystalline media . therefore , to prove that the crystalline undulator is feasible , it is necessary to analyze the influence of the interaction of both beams with the crystal constituents . on the basis of such analysis one can formulate the conditions which must be met and define the ranges of parameters ( which include @xmath8 , @xmath2 , @xmath1 and also the crystal length @xmath17 and the photon energy @xmath18 ) within which all the criteria are fulfilled . in full this analysis was carried out very recently and the feasibility of the crystalline undulator was demonstrated in an adequate form for the first time in refs . and in refs . . for further referencing let us briefly mention the conditions which must be met in a crystalline undulator . a stable planar channeling of an ultra - relativistic positron in a periodically bent crystal occurs if the maximum centrifugal force , @xmath19 , is less than the maximal force due to the interplanar field , @xmath20 . notating the ratio @xmath21 as @xmath22 one formulates this condition as follows @xcite : @xmath23 there are two essentially different regimes of the radiation formation in a periodically bent crystals they are defined by the magnitude of the ratio @xmath24 . in the case of low amplitudes , @xmath25 , the characteristic frequencies of the channeling radiation and the undulator radiation become compatible @xmath26 . this results in the loss of the monochromaticity of the radiation , since the channeling radiation is essentially non - monochromatic due to noticeable deviations of the interplanar potential from a harmonic form . additionally , in this case the intensity of undulator radiation is small compared with that of the channeling radiation @xcite . on the contrary , in the limit @xmath27 not only the characteristic frequencies are well separated , @xmath28 , but also the undulator radiation intensity is higher than the intensity of the channeling radiation @xcite . as a result , if one is only interested in the spectral distribution of the undulator radiation , one may disregard the channeling oscillations and assume that the projectile moves along the centerline of the bent channel . therefore , the criterion which is imposed on the relative magnitudes of @xmath6 , @xmath2 and @xmath1 is as follows @xmath29 the second inequality ensures that the crystal is deformed elastically , and its structure and symmetry are not affected by the deformation . the term undulator implies that the number of undulator periods , @xmath30 , is large . only in this limit does the radiation formed during the passage of a bunch of relativistic particles through a periodic system bear the features of undulator radiation ( narrow , well - separated peaks in spectral - angular distribution ) rather than those of synchrotron radiation . hence , the following strong inequality , which entangles the period @xmath1 and the length of a crystal @xmath17 must be met in the crystalline undulator @xcite : @xmath31 the coherence of the radiation , emitted in the crystalline undulator , takes place if the energy of the channeling particle does not change noticeably during with the penetration distance . for ultra - relativistic projectiles the main source of energy losses are the radiative losses . therefore , it is important to establish the range of energies for which the parameters of undulator radiation formed in a perfect periodic crystalline structure are stable . in ref . a comprehensive quantitative analysis of the radiative loss of energy , @xmath32 , due to the channeling and the undulator radiation was carried out . it was established that the relative radiative losses @xmath33 become large if the initial energy of the positron bunch is @xmath34 gev . for lower energies of positrons @xmath35 the radiative losses are small , @xmath36 . as was pointed out refs . two phenomena , the dechanneling effect and the photon attenuation , lead to severe limitation on the length of a crystalline undulator . if the dechanneling effect is neglected , one may unrestrictedly increase the intensity of the undulator radiation by considering larger @xmath30-values . in reality , random scattering of the channeling particle by the electrons and nuclei of the crystal leads to a gradual increase of the particle energy associated with the transverse oscillations in the channel . as a result , the transverse energy at some distance from the entrance point exceeds the depth of the interplanar potential well , and the particle leaves the channel . the mean penetration distance covered by a channeling particle is called the dechanneling length . for given crystal and projectile the dechanneling length @xmath37 depends on the energy @xmath8 and on the parameter @xmath22 ( see ( [ condition.1 ] ) ) . to calculate the dechanneling length one can either apply the diffusion theory to describe the multiple scattering @xcite or carry out a computer simulation of the scattering process of the projectile from the crystal constituents @xcite . alternatively , to estimate @xmath38 one can use the approximate formulae @xcite . for an ultra - relativistic positron the dechanneling length in straight channels ( i.e. @xmath39 ) in various crystals lies within the interval @xmath40 @xcite , i.e. does not exceed several millimeters at gev energies of a positron . for a periodically bent channel the dechanneling length decreases as @xmath22 grows following , approximately , the law @xmath41 @xcite . the propagation of photons emitted in a crystalline undulator is strongly influenced by the atomic and the nuclear photoeffects , the coherent and incoherent scattering on electrons and nuclei , the electron - positron pair production . all these processes lead to the decrease in the intensity of the photon flux as it propagates through the crystal . a quantitative parameter , which accounts for all these effects and defines the scale within which the intensity of a photon flux decreases by a factor of @xmath42 , is called the attenuation length , @xmath43 . it is related to the mass attenuation coefficient @xmath44 as @xmath45 @xcite . the coefficients @xmath44 are tabulated for all elements and for a wide range of photon frequencies @xcite . the magnitude of @xmath46 depends on @xmath47 and on the type of the constituent atoms . for low-@xmath48 crystals ( e.g. , diamond ) the magnitude of @xmath46 exceeds that for a tungsten crystal taken for the same @xmath47 by a factor of @xmath49 . in the case of a diamond crystal the value of @xmath46 varies from @xmath50 cm at @xmath51 kev up to several cm for @xmath52 ev . the simplest way to account for the dechanneling and the attenuation is to consider the case when the crystal length satisfies the condition @xmath53 $ ] , and to assume that within the chosen @xmath17 scale neither the number of channeled particles nor the flux of emitted photons do decrease . such approach was utilized in most of the papers devoted to the crystalline undulator problem . more consistent treatment of the dechanneling process and of its influence on the parameters of the undulator radiation was carried out in refs . , where a simple analytic expression for the spectral - angular distribution was derived which contains , as a parameter , the dechanneling length . in the present work we make another step in developing the theory of the crystalline undulator . the following problems are solved and discussed below in the paper . * we propose the scheme for accurate quantitative treatment of the radiation from a crystalline undulator in presence of the dechanneling and the photon attenuation ( sect . [ crystallineundulator ] ) . as a result , we evaluate an analytic expression for the spectral - angular distribution and the number of emitted photons which contains , as parameters , three quantities @xmath17 , @xmath38 , and @xmath54 . * we demonstrate that for given type of the crystal and crystallographic plane , and for given values of @xmath8 , @xmath2 , @xmath1 and @xmath47 there exists an optimal length of the crystal which ensures the largest number of the emitted photons . * using ( a ) and ( b ) we carry out the calculation of the number of emitted photons and the brilliance of the crystalline undulator radiation ( sect . [ numericalresults ] ) . the calculations , which account for the conditions ( [ condition.1])-([condition.4 ] ) , are performed for several crystals and by using the parameters of positron bunches used in modern colliders @xcite . prior to the discussion of the radiation formed in a crystalline undulator , in sect . [ idealundulator ] we briefly summarize the results from the general theory of undulator radiation ( see , e.g. refs . ) . we use the term ` ideal undulator ' to indicate that the propagation of positrons and photons occurs in vacuum . the spectral - angular distribution of the energy @xmath55 emitted by an ultra - relativistic projectile in a planar undulator can be written in the following form @xmath56 here @xmath57 and @xmath58 are the emission angles with respect to the undulator axis , @xmath59 is the solid angle of the emission . the function @xmath60 , which does not depend on the undulator length , is given by @xmath61 \right ) , \qquad m=0,1\ , . \label{dspectral.3}\end{aligned}\ ] ] here @xmath62 , @xmath63 , @xmath13 is the undulator parameter and the parameter @xmath64 is given by @xmath65 the factor @xmath66 on the right - hand side of ( [ dspectral.1 ] ) is defined as follows @xmath67 where @xmath68 and @xmath69 is a positive integer such that @xmath70 . for @xmath71 the function @xmath72 has a sharp and powerful maximum in the point @xmath73 , where @xmath74 . the width of the peak @xmath75 is equal to @xmath76 . this behaviour of @xmath72 results in a peculiar form of the spectral - angular distribution of undulator radiation which clearly distinguishes it from other types of electromagnetic radiation formed by a charge moving in external fields . namely , for each value of the emission angle @xmath77 the spectral distribution consists of a set of narrow and equally spaced peaks ( harmonics ) . the peak intensity is proportional to @xmath78 . this factor reflects the constructive interference of radiation emitted from each of the undulator periods and is typical for any system which contains @xmath30 coherent emitters . the values @xmath79 of the harmonics frequencies follow from the condition that parameter @xmath64 becomes an integer ( this corresponds to @xmath73 ) . in particular , in the case of the forward emission , @xmath80 , the harmonics frequencies are defined from the relation @xmath81 for @xmath80 the integrals ( [ dspectral.3 ] ) can be evaluated analytically and the spectral - angular distribution calculated for @xmath82 ( for @xmath83 ) acquires the form @xcite : @xmath84 ^ 2\ , , \label{numberphotonscollection.5a}\end{aligned}\ ] ] where @xmath85 and @xmath86 is the bessel function . the finite width of the central peak of @xmath72 defines the emission cone @xmath87 and the bandwidth @xmath88 of the @xmath69th harmonic . using @xmath89 and accounting for ( [ centralline.1 ] ) one derives @xmath90 formulae ( [ numberphotonscollection.5a])-([numberphotonscollection.3 ] ) allow one to calculate the number of photons @xmath91 of energy @xmath92 $ ] emitted by a beam particle within the cone @xmath87 : @xmath93 where @xmath94 ^ 2 $ ] . let us introduce two other quantities which characterize the radiation formed in an undulator and are closely related to the number of the emitted photons , but also take into account the properties of the beam of ultra - relativistic particles . these quantities are the _ flux _ and the _ brilliance _ ( see , e.g. ref . ) . the flux @xmath95 describes the number of photons per second of the @xmath69th harmonic emitted in the cone @xmath87 and in a given bandwidth . a quantitative definition of this quantity , measured in @xmath96 ( the abbreviation bw stands for the bandwidth @xmath97 ) , is given by the following formula @xcite : @xmath98 } \ , , \label{b&b.4}\end{aligned}\ ] ] where @xmath99 is the electric current of the beam . in the latter expression @xmath99 is measured in amperes . the general definition of brilliance of the photon source of a finite size is given in terms of the number of photons of energy @xmath18 emitted in the cone @xmath100 per unit time interval , unit source area , unit solid angle and per bandwidth @xcite . to calculate this quantity is it necessary to know the beam sizes @xmath101 , @xmath102 and angular divergencies @xmath103 , @xmath104 in two perpendicular directions , as well as the divergency angle of the radiation and the size of the photon beam . the brilliance of undulator radiation can be related to the flux @xmath95 as follows @xcite : @xmath105 here @xmath106 are the total emittance of the photon source in the @xmath107 and @xmath108 directions , with @xmath109 being the angular width of the @xmath69th harmonic and @xmath110 is the apparent source size taken in the diffraction limit @xcite . to obtain brilliance in the units @xmath111 the quantities @xmath112 and @xmath113 must be measured in millimeters whereas the angular variables @xmath114 and @xmath115 - in milliradians . in an ideal undulator the beam of particles and the emitted photons propagate in vacuum . in a crystalline undulator , due to the interactions with crystal atoms , the particles can dechannel , and thus be lost for further motion through the undulator . additionally , the photons emitted inside the crystal can be absorbed or scattered while making their way out from the crystal . therefore , it is necessary to account for the processes of dechanneling and photon attenuation . in what follows we carry out the qualitative analysis of the influence of these two processes on the characteristics of the radiation formed in a crystalline undulator . let the crystal length , the amplitude and period of bending and the energy @xmath8 satisfy the conditions ( [ condition.1])-([condition.4 ] ) . a positron , which enters the crystal at small incident angle with respect to the curved crystallographic plane , penetrates through the crystal following the bending of its channel . however , due to random scattering by the electrons and nuclei of the crystal the energy of the transverse oscillations of the positron in the channel increases , and finally the particle leaves the channel , becoming lost for the crystalline undulator . although the rigorous treatment of the dechanneling process can not be implemented by analytical means only , it is possible to develop a model approach based on the assumption that the probability @xmath116 for a particle to penetrate at a distance @xmath5 along the undulator axis ( @xmath117 $ ] ) can be described by the exponential decay law @xmath118 in intermediate formulae when referring to the dechanneling length we omit its arguments @xmath8 and @xmath22 . with the effect of dechanneling taken into account the spectral - angular distribution of the radiated energy per one particle can be written as follows : @xmath119 the first term is the contribution to from all the processes in which the particle dechannels somewhere inside the crystal . to calculate this term one notices that the quantity @xmath120 defines the probability of a particle to channel through the distance @xmath5 and then dechannel within the interval @xmath121 . such a particle emits the radiation which corresponds to the undulator of the length @xmath5 and the number of periods @xmath122 . therefore @xmath123 where @xmath124 denotes the spectral - angular distribution from the undulator of the length @xmath5 . the superscript ` ( att ) ' indicates that to calculate this quantity one has to account for the photon attenuation . the second term on the right - hand side of ( [ dechannelinginundulator.6a ] ) is due the process when the projectile channels through the whole length @xmath17 . its probability is given by the factor @xmath125 . therefore one can write @xmath126 if the photon attenuation is neglected , then to calculate ( [ dechannelinginundulator.5 ] ) one uses ( [ dspectral.1 ] ) instead of @xmath127 . the integral in ( [ dechannelinginundulator.4 ] ) is also evaluated with the help of ( [ dspectral.1 ] ) where one substitutes @xmath30 with @xmath122 . such approach was applied ref . with the only difference that in the cited paper to calculate ( [ dechannelinginundulator.4 ] ) and ( [ dechannelinginundulator.5 ] ) we used the discrete probabilities instead of the continuous distribution function ( [ dechannelinginundulator.1 ] ) . the use of the latter implies that the dechanneling effect is small over the scale of one undulator period and , therefore , @xmath128 . now let us turn to the derivation of the quantity @xmath129 which is the spectral - angular distribution of radiation formed in the undulator of the length @xmath130 in presence of the attenuation . in the intermediate formulae we assume that the ratio @xmath131 is an integer number which corresponds to the number of periods in this undulator . in the final formula this limitation will be omitted . throughout the text the notations @xmath17 and @xmath30 are reserved for the length of the crystal and the number of undulator periods within @xmath17 . as mentioned above , if one neglects the photon attenuation effect , the distribution @xmath129 is described by ( [ dspectral.1 ] ) where one substitutes @xmath30 with @xmath132 . the only quantity in ( [ dspectral.1 ] ) which depends on the number of undulator periods is the factor @xmath133 defined in ( [ idealundulator.2a ] ) . this factor appears in the formula for spectral - angular distribution as a result of squaring the modulus of a coherent sum of the amplitudes of electromagnetic waves emitted from spatially different but similar parts of the undulator . in more detail , @xmath133 is given by @xmath134 . the argument @xmath135 ( @xmath136 is the wavenumber ) stands for the phase of the electromagnetic wave emitted within the @xmath137th period of the undulator and detected at some distant point @xmath138 from the undulator . it is assumed that the quantities @xmath17 , @xmath5 and @xmath138 satisfy the relations : @xmath139 . th period of the undulator of the length @xmath5 , can be absorbed ( or scattered ) in the part of crystal of thickness @xmath140 on its way to a distant detection point @xmath138 ( @xmath141).,width=377,height=207 ] in a crystalline undulator a photon emitted within the @xmath137th period in the direction of the point @xmath138 can be absorbed within the distance @xmath140 while propagating through the crystal , see fig . [ fordn.eps.fig ] . to account for this possibility one can assume that the wavenumber becomes complex , @xmath142 . the quantity @xmath143 defines the attenuation length @xmath144 within which the photon flux is reduced by a factor of @xmath42 . for a complex @xmath145 the phase factor @xmath146 , which in an ideal undulator is the same for all periods @xmath147 , is replaced with @xmath148 , and a proper expression for @xmath149 is @xmath150 the spectral - angular distribution of radiation in presence of the photon attenuation acquires the form @xmath151 in the limit @xmath152 ( i.e. , when there is no attenuation ) the factor @xmath153 becomes equal to @xmath133 from ( [ idealundulator.2a ] ) , and the right - hand side of ( [ attenuationinundulator.8 ] ) reduces to that of eq . ( [ dspectral.1 ] ) . to derive the explicit expression for the spectral - angular distribution of the radiated energy from a crystalline undulator one uses ( [ attenuationinundulator.8 ] ) in ( [ dechannelinginundulator.6a])([dechannelinginundulator.5 ] ) . let us note here , that although the expression ( [ attenuationinundulator.3 ] ) was obtained for the case when the ratio @xmath122 is an integer , its use in the integral from ( [ dechannelinginundulator.4 ] ) can be justified by the above mentioned conditions , that the undulator period @xmath1 is small compared to the dechanneling length @xmath154 , the attenuation length , @xmath155 , and the length of crystal , @xmath17 . hence , the relative error , which appears when one uses ( [ attenuationinundulator.8 ] ) in ( [ dechannelinginundulator.4 ] ) , is small , being of the order of magnitude @xmath156 . carrying out the integration one represents the total spectral - angular distribution ( [ dechannelinginundulator.6a ] ) of radiation formed in the crystalline undulator in the form similar to ( [ dspectral.1 ] ) @xmath157 where the function @xmath60 , defined by ( [ dspectral.2a ] ) , does not depend on @xmath17 , @xmath154 and @xmath155 . these parameters enter the factor @xmath158 which is given by the expression : @xmath159\ , , \label{dechannelingandattenuation.2}\end{aligned}\ ] ] where the following notations are used : @xmath160 despite a cumbersome form of the right - hand side of ( [ dechannelingandattenuation.2 ] ) its main features can be easily understood . firstly , we notice that if the dechanneling is neglected , @xmath161 ( or @xmath162 ) , the function @xmath158 reproduces @xmath163 from ( [ attenuationinundulator.3 ] ) . in another limit @xmath164 ( i.e. , no attenuation and dechanneling ) eq . ( [ dechannelingandattenuation.2 ] ) reduces to the definition of the factor @xmath72 which characterizes the ideal undulator . in the case when only the attenuation effect is neglected the limit of @xmath158 can also be easily evaluated . in either of these cases the main maximum of @xmath158 is located in the point @xmath73 , i.e. when the parameter @xmath64 reduces to an integer , and , therefore , the harmonics frequencies are still defined by ( [ centralline.1 ] ) . the maximum value @xmath165 can be presented as follows : @xmath166 , \label{dechannelingandattenuation.5}\end{aligned}\ ] ] where the quantity @xmath167 stands for the number of undulator periods within @xmath154 , and the ratio @xmath168 does not depend on the crystal length @xmath17 . the width of the central peak @xmath169 , which in the case of an ideal undulator equals to @xmath76 , is increased due to the photon attenuation and the dechanneling . formally , the additional widths are due to the factors @xmath170 and @xmath171 which enter ( [ dechannelingandattenuation.2 ] ) . the widths associated with these factors are , respectively , @xmath172 and @xmath173 . thus , the total width of the peak is : @xmath174 the additional widths lead to the enlargement of the solid angle @xmath87 of the emission cone in the forward direction . in accordance with ( [ dechannelingandattenuation.6 ] ) one derives @xmath175 the formulae for the number of photons @xmath176 emitted in the cone @xmath87 as well the corresponding flux of radiation @xmath177 one derives similarly to how it was done in sect . [ idealundulator ] for an ideal undulator . the result is : @xmath178 } \ , . \label{cucharacteristics.4b}\end{aligned}\ ] ] the difference between these equations and formulae ( [ numberphotonscollection.1 ] ) and ( [ b&b.4 ] ) is that the number of undulator periods @xmath30 , met in the latter , is substituted with the effective number of periods , @xmath179 , which is defined as follows : @xmath180 \sqrt { 1 + { { \kappa}}_{\rm d}^2\,{(x-1)^2 + 1 \over 4\pi^2 } } \ , . \label{cucharacteristics.6}\end{aligned}\ ] ] we use these equations in the subsequent section to define the optimal length of a crystalline undulator . in the case of an ideal undulator one can , in principle , increase infinitely the length of the undulator . this will result in the increase of the number of photons , the photon flux , and the brilliance since they are proportional to the number of periods . the limitations on the values of @xmath17 and @xmath30 are mainly of a technological nature . the situation is different for a crystalline undulator , where the number of channeling particles and the number of photons which can emerge from the crystal decrease with the growth of @xmath17 . it is seen from ( [ cucharacteristics.6 ] ) that if @xmath181 then the parameters @xmath182 and @xmath183 also become infinitely large , and the effective number of periods goes to zero leading to @xmath184 . this is quite natural result , since in the limit @xmath185 all particles leave the channeling mode and , thus , do not undulate in the most part of the crystal , whereas all emitted photons are absorbed inside the crystal if @xmath186 . another formal ( and physically trivial ) fact , which follows from ( [ cucharacteristics.5 ] ) and ( [ cucharacteristics.6 ] ) , is that @xmath187 also for a zero - length undulator , when @xmath188 . vanishing of a positively defined quantity @xmath179 at two extreme boundaries suggests that there exists the length @xmath189 for which the effective number of periods ( taken for fixed values of @xmath155 , @xmath154 and @xmath1 ) attains the maximum . to define the value of @xmath189 or , what is equivalent , the quantity @xmath190 , one carries out the derivative of @xmath191 with respect to @xmath192 and equalizes it to zero . the analysis of the resulting equation shows that for each value of @xmath193 there is only one root @xmath194 . hence , the equation defines , in an inexplicit form , a single - valued function @xmath190 which ensures the maximum of @xmath179 for given @xmath155 , @xmath154 and @xmath1 . it is important to note that the crystal length enters eqs . ( [ cucharacteristics.4a])-([cucharacteristics.4b ] ) only via the ratio @xmath192 . all other quantities , met in these formulae as well as in ( [ cucharacteristics.5 ] ) and ( [ cucharacteristics.6 ] ) , are independent on the length of the crystal . therefore , the quantity @xmath189 ensures the highest values of @xmath176 and @xmath177 for the radiation formed in the crystalline undulator . in this sense @xmath189 can be called the _ optimal length _ of the undulator which corresponds to a given value of @xmath107 . and @xmath195 on @xmath196.,width=340,height=340 ] the dependences of @xmath190 and of the ratio @xmath197 on @xmath107 are presented in fig . [ optimalundnom.eps.fig ] . for a given crystalline structure , the dechanneling length @xmath154 is uniquely defined by the energy @xmath8 and the parameters of bending @xmath2 and @xmath1 . on the other hand , the attenuation length @xmath155 is the function of @xmath47 . therefore , fixing @xmath8 , @xmath2 , @xmath1 and @xmath47 one calculates @xmath196 and , then , using the dashed curve in the figure finds the optimal length of the crystalline undulator @xmath189 which accounts for the dechanneling effect and the photon attenuation . simultaneously , from the solid curve one finds the effective number of the undulator periods @xmath198 which defines the number of emitted photons , the flux and the brilliance of radiation . from ( [ cucharacteristics.4a ] ) follows , that to find the number of photons @xmath176 emitted in a crystalline one has to calculate two factors . the factor @xmath199 ( see ( [ numberphotonscollection.1 ] ) ) depends on the harmonic number @xmath69 and on the undulator parameter @xmath13 which , in turn , is defined by the values of @xmath8 , @xmath2 and @xmath1 through the relation @xmath200 . the second factor , @xmath179 depends on @xmath17 , @xmath1 , @xmath37 and @xmath201 . it was explained in sect . [ optimallength ] that once the quantities @xmath1 , @xmath38 and @xmath46 are known the length of the crystal can be fixed by the condition @xmath202 which results in the maximum values of @xmath179 and @xmath177 with respect to @xmath17 . the numerical data presented below in this section was obtained for the optimal length of undulator . therefore , to calculate @xmath176 one fixes , in addition to the crystallographic plane , the values of @xmath69 , @xmath8 , @xmath2 and @xmath1 ( the three latter are subject to the conditions ( [ condition.1])([condition.4 ] ) ) which uniquely define the quantities @xmath13 , @xmath22 , @xmath79 , @xmath46 . however , there is some uncertainty with respect to the magnitude of the dechanneling length . this uncertainty is not intrinsic to the case of a periodically bent crystal but rather reflects the stochastic nature of the interaction of a channeling particle with crystal constituents . as mentioned above to calculate the dechanneling length one can apply the diffusion theory to describe the multiple scattering or carry out numerical simulations of the scattering process . alternatively , one can use model - dependent analytic expressions for @xmath38 . in the present paper we utilize the approach , presented in ref . , and approximate @xmath38 with @xmath203 here where @xmath204 cm is the electron classical radius , @xmath205 mev is the electron rest energy , @xmath206 is the thomas - fermi radius of the crystal atom . the parameter @xmath22 is defined by eq . ( [ condition.1 ] ) . the quantity @xmath207 stands for the dechanneling length of a positron in a straight crystal@xcite . the quantity @xmath208 , with @xmath99 denoting the ( average ) ionization potential of the crystal atom , is the coulomb logarithm characterizing the ionization losses of an ultra - relativistic particle in amorphous media . for a quick estimation of @xmath207 ( in cm ) one can re - write the right - hand side of the second equation from ( [ dechannnelingattenuation.3 ] ) as @xmath209 , with @xmath206 and @xmath6 measured in and @xmath8 in gev . the values of @xmath206 and @xmath6 , are presented in table [ crystals.table ] . . parameters @xmath6 , @xmath206 and @xmath210 for different crystals and channels . [ cols="<,<,<,<,<",options="header " , ] the results of our calculations are presented in figs . [ s_si_ge_w.eps.fig ] and [ bril_c_si_ge_w.eps.fig ] . the choice of the crystals was motivated by the fact that c , si , ge and w crystals are frequently used in channeling experiments ( see , e.g. , ref . ) . an additional reason is that for a given photon frequency the magnitude of @xmath46 rapidly decreases with the growth of atomic number of the constituent atoms . therefore , by comparing the results obtained for different crystals one can investigate the influence of the photon attenuation on the formation of the radiation in a crystalline undulator . ) per a bandwidth @xmath211 and per a positron as a function of the ratio @xmath24 calculated for the positron energies in various colliders ( see table [ particledatagroup2004.table1 ] ) as indicated . graphs ( b ) : the corresponding values of the fundamental harmonic energy ( see eq . ( [ centralline.1 ] ) with @xmath212 ) . see also explanations in the text . each vertical pair of the graphs ( a ) and ( b ) correspond to the positron channeling in the particular periodically bent channel as indicated in the graphs ( a ) . the legend refers to all graphs in the figure.,title="fig:",width=548,height=396 ] + ) per a bandwidth @xmath211 and per a positron as a function of the ratio @xmath24 calculated for the positron energies in various colliders ( see table [ particledatagroup2004.table1 ] ) as indicated . graphs ( b ) : the corresponding values of the fundamental harmonic energy ( see eq . ( [ centralline.1 ] ) with @xmath212 ) . see also explanations in the text . each vertical pair of the graphs ( a ) and ( b ) correspond to the positron channeling in the particular periodically bent channel as indicated in the graphs ( a ) . the legend refers to all graphs in the figure.,title="fig:",width=548,height=396 ] graphs ` ( a ) ' in fig . [ s_si_ge_w.eps.fig ] present the dependence of the _ maximal _ number of emitted photons of the first harmonic ( @xmath212 ) per bandwidth @xmath211 and per a positron versus the ratio @xmath24 . the curves were calculated for the positron energies indicated in table [ particledatagroup2004.table1 ] ( for bepc - ii and pep - ii colliders we used the values @xmath213 gev and @xmath214 gev , respectively ) . for each crystal and for each @xmath8 value the dependences @xmath215 were obtained as follows . there are two independent variables , @xmath1 and @xmath2 , which , ( for fixed crystal , energy and harmonic number @xmath69 ) define all other quantities on the right - hand side of ( [ cucharacteristics.4a ] ) . for practical purposes it is more convenient to chose the ratio @xmath216 and the parameter @xmath217 ( see eq . ( [ condition.1 ] ) ) as the independent variables . then , for each pair @xmath218 one finds @xmath1 , @xmath219 , @xmath220 , the dechanneling length @xmath38 and the number of periods @xmath221 , the fundamental harmonic frequency @xmath222 ( see eq . ( [ centralline.1 ] ) ) and the attenuation length @xmath223 , and the value of @xmath198 which corresponds to the optimal undulator length calculated for @xmath224 ( see ( [ cucharacteristics.5])-([cucharacteristics.6 ] ) and fig . [ optimalundnom.eps.fig ] and sect . [ optimallength ] ) . as a result , one finds the magnitude of @xmath225 . finally , scanning through all @xmath218 values one determines the highest possible value of the number of photons per bw , @xmath215 , as a function of @xmath24 . having done this one also finds the dependence @xmath226 ( graphs ` ( b ) ' in fig . [ s_si_ge_w.eps.fig ] ) as well all other characteristics of the undulator as functions of @xmath24 . let us briefly discuss the behaviour of obtained dependences . firstly , as it is seen from the graphs ( a ) , for a fixed amplitude @xmath2 the quantity @xmath215 is an increasing function of a positron energy @xmath8 . this feature becomes clear if one analyzes the @xmath8 dependence of the product @xmath227 which defines the number of emitted photons ( see eqs . ( [ cucharacteristics.4a ] ) and ( [ cucharacteristics.5 ] ) ) . all three factors are increasing functions of energy ( although it is not too obvious for @xmath228 ) . another feature of the curves @xmath215 is that they are decreasing function of @xmath24 in the region @xmath229 . to a great extent this is a consequence of the photon attenuation in the crystal . indeed , as the ratio @xmath24 increases the undulator period @xmath1 increases too , in order to maintain the inequality @xmath230 ( see eq . ( [ condition.1 ] ) ) . larger values of @xmath1 results in lowering of the emitted photon energy ( see eq . ( [ centralline.1 ] ) and the graphs ( b ) in fig . [ s_si_ge_w.eps.fig ] ) and , consequently , to the decrease of the attenuation length , @xmath46 . this , in turn , leads to the increase of the ratio @xmath196 which defines the magnitude of @xmath228 . this factor , as it is seen from fig . [ optimalundnom.eps.fig ] , rapidly falls off for @xmath231 , and this feature manifests itself in the dependence of @xmath215 on @xmath24 . in the case of crystals consisting of heavy atoms the dependence acquires additional features , which are due to the fact that the ionization potentials , @xmath232 , of the inner atomic subshells of such atoms lie within the energy range @xmath233 kev . the photons with the energy just above the threshold are absorbed much more efficiently than those with the lower energies . as a result , the dependence of @xmath46 in the vicinity of the threshold becomes a saw - like . for @xmath234 the attenuation length noticeably ( up to the order of magnitude ) exceeds @xmath46 for @xmath235 . this effect results in the irregularities of the dependence @xmath215 on @xmath24 , which in figs . [ optimalundnom.eps.fig ] are mostly pronounced for diamond and tungsten crystals . in the opposite limit , when @xmath25 the number of the emitted photons goes to zero . this tendency , which is seen explicitly for all the curves ( but the cers - c one ) in the case of c and si crystals , is also clear and is due to the fact that the case @xmath236 corresponds to the linear crystal , i.e. the absence of the crystalline undulator . in fig . [ bril_c_si_ge_w.eps.fig ] we present the peak brilliance of the crystalline undulators based on different crystals ( as indicated ) and calculated using the parameters of the positron beams from table [ particledatagroup2004.table1 ] . the data refer to the emission in the first and the third harmonics in the forward direction . it is seen that in contrast to the number of the emitted photons which is the same , by the order of magnitude , for all colliders , the magnitudes of the peak brilliance for different beams differ by orders of magnitude . to the largest extent this is due to the quality of the beam , which includes , apart from the beam current @xmath99 , its size and angular divergency , see . ( [ b&b.4 ] ) and ( [ b&b.7 ] ) . for all crystals and over the whole range of photon energies the product @xmath237 of the photon source emittances is the smallest for the kekb collider ( labeled as ` 5 ' in the graphs in fig . [ bril_c_si_ge_w.eps.fig ] ) . as a result , this beam , which does not lead to the highest values of @xmath215 , ensures the largest peak brilliance of the crystalline undulator radiation . the peak brilliance for the kekb positron beam is on the level of @xmath238 for the photon energies within @xmath239 mev range . these values can be compared with the peak brilliance of the light sources of the third generation @xcite . the peak brilliance on the level @xmath240 in the 100 kev range of photon energies by means of the undulators based on the action of magnetic field is planned to be achieved within several projects @xcite . the data from fig . [ bril_c_si_ge_w.eps.fig ] ) demonstrate that it is feasible to produce the radiation of the same level of brilliance but for much higher energies by means of crystalline undulators . , the dashed curves refer to @xmath241 . in each graph the enumerated sets of the solid and the dashed curves correspond to the parameters of the positron beams in different colliders ( see table [ particledatagroup2004.table1 ] ) . 1 : da@xmath242ne , 2 : vepp-2000 , 3 : bepc - ii , 4 : pep - ii , 5 : kekb , 6 : cers - c.,title="fig:",width=566,height=377 ] + , the dashed curves refer to @xmath241 . in each graph the enumerated sets of the solid and the dashed curves correspond to the parameters of the positron beams in different colliders ( see table [ particledatagroup2004.table1 ] ) . 1 : da@xmath242ne , 2 : vepp-2000 , 3 : bepc - ii , 4 : pep - ii , 5 : kekb , 6 : cers - c.,title="fig:",width=566,height=377 ] theoretical investigations show that it is entirely realistic to use a crystalline undulator for generating spontaneous radiation in a wide range of photon energies . the parameters of such an undulator , being subject to the restrictions mentioned in sect . [ introduction ] , can be easily tuned by varying the parameters of the bending , the positron energy and by choosing different channels . the large range of energies available in modern colliders together with the wide range preparation of periodically bent crystalline structures allow one to generate the crystalline undulator radiation with energies from hundreds of kev up to tens of mev region . the brilliance of the undulator radiation within this energy range is comparable to that of conventional light sources of the third generation but for much lower photon energies . the experimental efforts are needed for the verification of numerous theoretical predictions . such efforts will certainly make this field of endeavor even more fascinating than as it is already and will possibly lead to the practical development of a new type of tunable and monochromatic radiation sources . 99 a. v. korol , a. v. solovyov , w. greiner , _ j. phys . g : nucl . part . * 24 * , pp . l45-l53 , 1998 . a. v. korol , a. v. solovyov , w. greiner , _ int . j. mod * 8 * , pp . 49 - 100 , 1999 . a. v. korol , a. v. solovyov , w. greiner , _ int . * 9 * , pp . 77 - 105 , 2000 . w. krause , a. v. korol , a. v. solovyov , w. greiner , _ j. phys . g : nucl . part . * 26 * , pp . l87-l95 , 2000 . a. v. korol , a. v. solovyov , w. greiner , _ j. phys . g : nucl . part . * 27 * , pp . 95 - 125 , 2001 . w. greiner , a. v. korol , a. v. solovyov , _ aip conf . _ , * 576 * , pp.17 - 20 , 2001 . w. krause , a. v. korol , a. v. solovyov , w. greiner , _ nucl . methods _ a * 475 * , pp . 441 - 444 , 2001 . a. v. korol , w. krause , a. v. solovyov , w. greiner , _ nucl . methods _ a * 483 * , pp . 455 - 460 , 2002 . r. o. avakian , et . al . , _ jetp lett . _ * 68 * , pp . 467 - 471 , 1998 . u. mikkelsen , e. uggerhj , _ nucl . b * 160 * , pp . 435 - 439 , 2000 . r. o. avakian , et . al . , _ nucl . instrum . methods _ b * 173 * , pp . 112 - 120 , 2001 . r. o. avakian , et . al . , _ nucl . methods _ a * 492 * , pp . 11 - 13 , 2002 . r. o. avakian , et . al . , _ nucl . methods _ a * 508 * 496 - 499 , 2003 . s. bellucci , s. bini , v. m. biryukov , et . al . , _ phys . * 90 * , 034801 , 2003 . s. bellucci , s. bini , s. giannini , et . al . , _ phys . rev . beams _ * 7 * , 023501 , 2004 . n. a. korhmazyan , n. n. korhmazyan , n. e. babadjanyan , _ tech . phys . _ * 49*(4 ) , pp . 499 - 501 , 2004 . r. o. avakian , et . al . , _ nucl . instrum . methods _ b * 227 * , pp . 104 - 109 , 2005 . h. kudo , _ nucl . instrum . methods _ b * 189 * 609 - 614 , 1981 . j. a. ellison , s. t. picraux , _ phys . lett . _ a * 83 * , p. 271 v. m. biruykov , yu . a. chesnokov , v. i. kotov , _ crystal channeling and its application at high - energy accelerators _ , springer , berlin , 1996 . j. h. hubbel , s. m. seltzer , `` tables of x - ray mass attenuation coefficients '' , nistir 5632 - web version 1.02 , http://physics.nist.gov/physrefdata/xraymasscoef/cover.html . s. eidelman , et . al . , ` review of particle physics ' , _ phys . lett . _ * b 592 * , pp . 1 - 1109 , 2004 . d. f. alferov , yu . a. bashmakov , p. a. cherenkov , _ sov . uspekhi _ * 32 * , p. 200 kim , ` characteristics of synchrotron radiation ' , in _ aip conference proceedings _ * 189 * , pp . 565 - 632 , 1989 . p. rullhusen , x. artru , p. dhez , _ novel radiation sources using relativistic electrons _ , world scientific , singapore , 1998 . kim , ` characteristics of synchrotron radiation ' , in _ x - ray data booklet _ , lawrence berkeley laboratory , berkley , ca , ch . 4 - 1 , 1986 . kim , _ nucl . methods _ * 246 * , pp . 71 - 76 , 1986 . m. altarelli , a. salam , _ europhysicsnews _ * 35 * , no . 2 , pp.47 - 50 , 2004 . `` the european x - ray laser project xfel '' , http://xfel.desy.de/content/e169/index_eng.html k. balewski , et.al . `` petra iii : a third generation synchrotron radiation source at desy . '' technical design report '' , http://tesla.desy.de/new_pages/tdr_cd/start.html .

at long last , the experimental exploration of the higgs sector of the electroweak theory has begun . the discovery of a scalar boson with a mass of approximately @xmath8 gev at the large hadron collider ( lhc ) is a spectacular triumph of the standard model ( sm ) @xcite . the hints of a larger than expected branching fraction for @xmath9 , although they have recently weakened , could suggest possible extensions of the scalar sector @xcite or of other sectors . ( for some vector - like fermion extension examples , see refs . @xcite . ) certainly , detailed analyses of these extensions are worthwhile and might provide guidance to experimenters searching for physics beyond the sm . the simplest and most studied such extension is the two higgs doublet model ( 2hdm ) , recently reviewed in ref . a potential problem for such models are tree - level flavor changing neutral currents ( fcnc ) which will occur whenever fermions of a given charge couple to more than one higgs multiplet @xcite . while it is possible to ameliorate the constraints by assumptions about the flavor - changing couplings @xcite , the usual solution to the problem is eliminating the tree - level fcnc altogether by using a discrete @xmath1 symmetry . depending on the transformation properties of the right - handed fermions under the @xmath1 , several different models can be obtained @xcite . the most familiar two models are the type i model , in which all of the fermions couple to a single higgs multiplet , and the type ii model , in which the down - type quarks and charged leptons couple to one higgs multiplet and the up - type quarks couple to the other . a @xmath1 symmetry can be promoted to a gauged @xmath0 symmetry as discussed in refs . @xcite . in ref . @xcite , the sm fermions do not have charges of the new @xmath0 gauge symmetry and a type i 2hdm model results . ( see ref . @xcite for another example in the fourth generation context . ) in ref . @xcite , the sm fermions have charges under the @xmath0 gauge symmetry , resulting in various types of 2hdms . as emphasized in ref . @xcite , one of the interesting features in this kind of model is the absence of the pseudoscalar @xmath10 . it is a common effect @xcite of a spontaneously broken @xmath0 gauge symmetry that the pseudoscalar is eaten to become the longitudinal component of the new gauge boson @xmath2 . of course , one could restore the pseudoscalar by adding a complex higgs singlet to the model . in our paper , we will closely follow the scenario of ref . @xcite . in this model , all sm fermions are neutral under the new abelian gauge group @xmath11 , yet the @xmath2 coupling to the sm fermions can be mediated by mixing with the sm neutral gauge bosons . the @xmath2 can communicate with the sm fermions only through kinetic and/or mass mixing with the sm neutral gauge bosons . the fact that the @xmath2 does not directly interact with fermions eliminates the numerous constraints on @xmath2 masses , and allows the @xmath2 to be extremely light , with mass allowed below @xmath12 . ( for a review on a complementary heavy @xmath2 , see ref . @xcite . ) this model in ref . @xcite was not originally proposed in the context of 2hdms . it was first proposed as a generalization of the so - called dark photon model whose coupling is of the same form as the photon coupling . while the two models both adopt kinetic mixing of the @xmath13 and the @xmath11 , the difference comes from the fact that the @xmath2 gets its mass only from a higgs singlet in the dark photon model while it gets the mass from a higgs doublet ( and also from a higgs singlet if a higgs singlet exists ) in the dark @xmath3 model . as a result , the dark @xmath3 can couple to the weak neutral current as well . when the @xmath2 is very light , it can explain @xcite the astrophysical anomalies such as the @xmath14 gamma - ray from the galactic center observed by the integral satellite @xcite or the positron excess observed by atic @xcite and pamela experiments @xcite , depending on the property of the dark matter . it can also explain the @xmath15 deviation of the measured muon @xmath4 from the sm prediction @xcite . ( for some discussions about non - abelian dark gauge sectors , see refs . see also ref . @xcite for a discussion on @xmath16-@xmath17 mixing . ) it is remarkable that a very light gauge boson , introduced to provide a dark matter explanation to some astrophysical anomalies , can also provide an explanation for the absence of tree - level fcnc in a 2hdm . it leads to a very different phenomenology from the ordinary 2hdms . some of the implications for the dark @xmath3 boson as well as the @xmath5 sm - like higgs state , @xmath6 , of this model have been discussed in refs . @xcite . in this paper , we will mainly discuss the phenomenology of the other higgs bosons of the model , and refer to this model as the `` dark 2hdm '' to emphasize that our study of this model focuses on the comparison to the ordinary 2hdms . we particularly focus on the scenario that the other higgs , @xmath7 , is lighter than the sm - like higgs , @xmath6 , of @xmath5 , which was a subject of the study in ref . @xcite for the ordinary 2hdms . we will see that various constraints force one to add a singlet to the model . the decay of the @xmath7 to @xmath18 will provide a dramatic signature that could be on the verge of discovery at the lhc , and the decay of the charged higgs to @xmath19 would also lead to striking signatures . in sec . [ sec : model ] , we describe the model and the particle masses and couplings . in sec . [ sec : constraints ] , we consider the constraints from lep and the lhc on the neutral higgs bosons , and in sec . [ sec : lighthiggs ] study the remarkable signature of the light higgs and the implications for lhc experiments looking for multi - lepton jets . in sec . [ sec : chargedhiggs ] , we discuss the properties of the charged higgs boson and in sec . [ sec : conclusion ] , we present our conclusions . a discussion of the properties of the @xmath2 boson is in the appendix . in this section , we describe the dark 2hdm proposed by davoudiasl , lee , and marciano @xcite . the scalar sector is composed of two doublets and a singlet under the @xmath20 . as we will see , the model with just two doublets and no singlet is not compatible with various constraints . including a singlet can avoid this issue . the gauge group is @xmath21 , and the sm fermions are assumed to be neutral under the extra @xmath11 . the gauge part of the kinetic lagrangian includes kinetic mixing between the @xmath13 and @xmath11 @xcite and is given by _ gauge = - b _ b^ + b _ z^ - z _ z^ [ eq : kineticmixing ] where @xmath22 and @xmath23 . the hat notation indicates the states before diagonalization . as discussed in various places , for examples see refs . @xcite , one can redefine the fields to remove the kinetic mixing term parametrized by @xmath24 , which is experimentally constrained to be small . when there is one higgs doublet and the @xmath2 gets the mass from a higgs singlet , this leads to an induced coupling of the @xmath2 to the electromagnetic current _ = -e j^_em z_(j^_em q_f |f ^f ) and the induced coupling of the @xmath2 to the weak neutral current is negligibly small because of the suppression by high orders of @xmath24 . this gauge boson is widely called the `` dark photon '' , since fermions couple to the @xmath2 with a very small effective electric charge ( @xmath25 ) for a given electric charge ( @xmath26 ) of the fermion @xmath27 . the model in ref . @xcite added an additional higgs doublet charged by the @xmath11 which introduces mixing in the @xmath3-@xmath2 mass matrix . the @xmath3-@xmath2 mass mixing , parametrized by @xmath28 , is also taken to be small . the value of @xmath28 will be determined by parameters in the scalar potential . because of this mixing , the @xmath2 couples to both electromagnetic current and weak neutral current as _ = -(e j^_em + _ z g_z j^_nc ) z_(j^_nc |f ^f - [ t_3f ] |f ^_5 f ) with @xmath29 , @xmath30 and the weak mixing angle @xmath31 . the @xmath2 with more general coupling ( compared to the dark photon ) in this model was named the `` dark @xmath3 '' to emphasize that it couples to the weak neutral current . in particular , when the @xmath32 limit is taken , the @xmath2 couples only to the weak neutral current as the sm @xmath3 boson does with a suppressed coupling ( @xmath33 ) . the definition of @xmath28 and constraints on @xmath34 , @xmath24 , @xmath28 can be found in the appendix . we use a similar but slightly different notation from ref . was chosen to couple to the sm fermions rather than @xmath35 , with definitions of @xmath36 , @xmath37 . this leads to different typical values of @xmath38 and @xmath39 . care must be taken in defining the heavy higgs vs. the light higgs , which also amounts to a definition of @xmath40 . we provide a comparison of notations in the following . reference @xcite : @xmath41 couples to the sm fermions , which requires typically small @xmath38 ( as @xmath42 ) . @xmath43 corresponds to the heavy higgs @xmath44 sm - like higgs limit . this paper : @xmath35 couples to the sm fermions , which requires typically large @xmath38 ( as @xmath45 ) . @xmath46 corresponds to the heavy higgs @xmath47 sm - like higgs limit . ( @xmath43 corresponds to the light higgs @xmath47 sm - like higgs limit . ) the higgs to gauge boson couplings in eqs . - can be read for ref . @xcite by @xmath48 , @xmath49 ( or @xmath50 ) , and overall sign flips for @xmath51 , @xmath52 , @xmath53 . ] in conventional 2hdms , one defines @xmath35 to be the doublet that couples to the top quark ( or , in the type i 2hdm , to all fermions ) . following this notation , we assume the @xmath11 charges @xmath54 = q'[\phi_s ] = 1 $ ] , @xmath55 = 0 $ ] with the hypercharges @xmath56 = y[\phi_2 ] = 1/2 $ ] , @xmath57 = 0 $ ] . the @xmath11 is spontaneously broken when @xmath41 or @xmath58 gets a vacuum expectation value ( vev ) . the scalar part of the kinetic lagrangian is given by _ & = & | d__1 |^2 + & = & l _ + l _ + where d__i = ( _ + i g y[_i ] b_+ i g t_3[_i ] w_3 + i g_z q[_i ] z^0 _ ) _ i and two doublets @xmath59 with vevs @xmath60 ( @xmath61 ) , and a singlet @xmath62 with a vev @xmath63 . the @xmath64 limit corresponds to the doublets only case . the @xmath3 and @xmath2 masses , with @xmath65 and @xmath36 , @xmath66 are given by _ = m_z^0 ^ 2 z^0 z^0 - ^2 z^0 z^0 + m_z^0 ^ 2 z^0 z^0 + where m_z^0 ^ 2 & = & g_z^2 v^2 , + m_z^0 ^ 2 & = & g_z^2 ( v^2 ^2 + v_s^2 ) + g_z g v^2 ^2 + ( ) ^2 g^2 v^2 , [ eq : mzprime0 ] + ^2 & = & g_z g_z v^2 ^2 + g_z g v^2 . this leads to the @xmath3-@xmath2 mixing angle ( @xmath67 ) as 2= ( ccccc z + z + ) = ( ccccc & - + & + ) ( ccccc z^0 + z^0 + ) [ eq : massmixing ] where @xmath3 and @xmath2 are the mass eigenstates . the @xmath3-@xmath2 mixing angle is constrained to be very small ( @xmath68 few @xmath69 ) by precision @xmath3 pole measurement at lep @xcite . we are primarily interested in a very light @xmath2 [ @xmath70 . we will work in the @xmath71 limit throughout this paper , which requires @xmath72 as @xmath73 is very small . in this limit we have m_z^2 & & m_z^0 ^ 2 = g_z^2 v^2 , [ eq : mz ] + m_z^2 & & m_z^0 ^ 2 - = g_z^2 ( v^2 ^2 ^ 2 + v_s^2 ) = g_z^2 v^2 ( 1 - ^2 ^ 2_d ) , [ eq : mzprime ] + & & = ^2 + _ w . [ eq : xi ] the @xmath2 approaches the massless limit as @xmath74 or @xmath75 . with the @xmath76 notation defined in the appendix , we can use @xmath77 ( the doublets only limit corresponds to @xmath78 ) , and have m_z . for simplicity , we assume no mixing between the doublets and singlet and allow mixing only between the doublets . the pure doublets case can be reached if we take @xmath79 ( corresponding to @xmath80 ) in eqs . and , and it would not change the following couplings in eqs . - as well as the approximations in eqs . - . the relevant couplings of vector bosons to heavy higgs ( @xmath6 ) and light higgs ( @xmath7 ) , with no mixing between the doublets and singlet , are _ & = & _ hzz hzz + c_hzz hzz+ _ hzz hzz + & + & _ hzz hzz + c_hzz hzz+ _ hzz hzz + where _ hzz & = & c_hzz^ ( ^2 + ^2 ) [ eq : chzz ] + c_hzz & = & c_hzz^ ( . + ) + c_hzz & = & c_hzz^ ( ^2 + ^2 ) and _ hzz & = & c_hzz^ ( ^2 - ^2 ) + c_hzz & = & c_hzz^ ( . + & & . - ) + c_hzz & = & c_hzz^ ( ^2 - ^2 ) [ eq : chzz ] with the sm higgs-@xmath3-@xmath3 coupling @xmath81 . note that , unlike ref . @xcite , we include couplings with a general mixing angle @xmath40 . since @xmath82 , @xmath83 , we can get approximations , using eq . , _ hzz & & c_hzz^ ( - ) [ eq : chzzapprox ] + c_hzz & & - c_hzz^ ( 2 g_z / g_z ) ( - ) [ eq : chzzapprox ] + c_hzz & & c_hzz^ ( 2 g_z / g_z)^2 ( ^3 + ^3 ) [ eq : chzzapprox ] and _ hzz & & c_hzz^ ( - ) [ eq : chzzapprox ] + c_hzz & & c_hzz^ ( 2 g_z / g_z ) ( - ) [ eq : chzzapprox ] + c_hzz & & c_hzz^ ( 2 g_z / g_z)^2 ( ^3- ^3 ) [ eq : chzzapprox ] giving the expected 2hdm couplings of the two neutral higgs bosons to the @xmath3 in eqs . and . in the @xmath84 limit , we can reproduce the relation shown in ref . @xcite , i.e. @xmath85 . the scalar potential is v & = & v_1 + v_2 + v_1 & = & m_11 ^ 2 _ 1^_1 + m_22 ^ 2 _ 2^_2 + ( _ 1^_1 ) ^2 + ( _ 2^_2 ) ^2 + _ 3 ( _ 1^_1 ) ( _ 2^_2 ) + _ 4 ( _ 1^_2 ) ( _ 2^_1 ) + v_2 & = & m_33 ^ 2 _ s^_s + ( _ s^_s ) ^2 with all coefficients real . interestingly , the terms @xmath86 and @xmath87 $ ] whose coefficients are generally complex are forbidden by the @xmath11 gauge symmetry . we assume the terms @xmath88 are negligible ( no mixing between the doublets and singlet ) . the charged higgs mass is given by m_h^^2 = - v^2 which requires @xmath89 . the higgs singlet mass is given by m_s^2 = _ 6 v_s^2 and the neutral doublet higgs mass - squared matrix is given by m_^2 = ( ccccc _ 1 v_1 ^ 2 & ( _ 3 + _ 4 ) v_1 v_2 + ( _ 3 + _ 4 ) v_1 v_2 & _ 2 v_2 ^ 2 + ) . [ eq : higgsmatrix ] the mass eigenstates of the doublets are @xmath6 and @xmath7 ( with @xmath90 ) with m_h^2 = ( _ 1 v_1 ^ 2 + _ 2 v_2 ^ 2 + ) + m_h^2 = ( _ 1 v_1 ^ 2 + _ 2 v_2 ^ 2 - ) and the @xmath6-@xmath7 mixing angle is given by 2= ( ccccc h + h + ) = ( ccccc & + - & + ) ( ccccc _ 1 + _ 2 + ) . since only @xmath35 couples to the sm fermions in this model , the @xmath91 limit provides the heavier higgs ( @xmath6 ) as the sm - like higgs ( @xmath92 , @xmath93 up to a sign ) , which is the case of interest in the subsequent sections in this paper . the other limit , @xmath94 , would have provided the light higgs ( @xmath7 ) as the sm - like higgs ( @xmath95 , @xmath96 up to a sign ) . note that the sm - like limit for the @xmath6 is achieved for @xmath97 instead of @xmath98 which would be true when the light higgs @xmath7 is the sm - like one . the relative coupling of the higgses to the sm fermions and @xmath16 boson in the dark 2hdm is the same as the type i 2hdm as following . htt , hbb , h : , & & hww : ( - ) [ eq : hcoupling ] + htt , hbb , h : , & & hww : ( - ) [ eq : hcoupling ] the relative coupling of @xmath99 ( @xmath100 ) is the same as that of the @xmath101 ( @xmath102 ) to a good approximation since the @xmath3 mixing is very small [ eq . , eq . ] . + ( a ) ( b ) in this section , we consider various constraints / predictions on the neutral higgs bosons in the dark 2hdm . they include ( i ) lep limits on associated production of the light higgs ( @xmath103 ) , ( ii ) lep limits on the marginal @xmath3 width for @xmath104 decay , ( iii ) lhc measured branching ratios in various modes ( @xmath105 , @xmath106 , @xmath107 , @xmath108 ) including the most precisely measured @xmath9 . we will focus on the case in which the mass of the light higgs is @xmath109 , i.e. roughly , @xmath110 range . if it is lighter than @xmath111 , then the fact that @xmath112 decays do not dominate restricts @xmath113 to be infinitesimally close to @xmath114 @xcite . if it is heavier than @xmath115 , then the @xmath116 channel opens up although our results may not change qualitatively . a sm higgs of less than @xmath117 gev was ruled out long ago by lep experiments looking for associated production with a @xmath3 . in 2hdms , however , the @xmath102 and @xmath100 ( neglecting small @xmath3 mixing ) couplings are suppressed by @xmath118 , possibly making the @xmath7 quite gaugephobic . lep published upper bounds on this factor as a function of the higgs mass , assuming the higgs would decay into @xmath119 quarks or @xmath120 s @xcite . however , as we will discuss in sec . [ sec : lighthiggs ] , @xmath7 can mainly decay into @xmath18 in the dark 2hdm . since the lep searches did not cover fermions from a very light @xmath2 [ @xmath121 , one must look at lep bounds on invisible higgs decays . these bounds at @xmath122 cl from aleph @xcite , delphi @xcite , l3 @xcite , opal @xcite collaborations are given in fig . [ fig : lep ] ( a ) . there are also updated aleph results , which covers the higgs mass from @xmath123 @xcite , but this is beyond our range of interest . for simplicity , we will take only l3 data , which is roughly close to average values of the lep data in the @xmath124 range we consider in this paper ( @xmath110 ) . for a given @xmath7 mass , this bound is given as a band in @xmath125 plane in fig . [ fig : lep ] ( b ) . the apparent asymmetry between the @xmath126 and @xmath127 regions originates from @xmath128 , which makes the region near @xmath127 ( or @xmath129 ) preferable . as the @xmath7 mass increases from around @xmath130 , the bands quickly converge as the lep excluded region vanishes [ fig . [ fig : lep ] ( a ) ] . this is illustrated as a sudden departure of @xmath131 curve from the @xmath132 curve in fig . [ fig : lep ] ( b ) . the @xmath7 can be produced directly from @xmath3 decays into an @xmath7 and a @xmath2 , if kinematically allowed . it will appear as a contribution to the invisible @xmath3 width . for a sufficiently light @xmath2 , using eq . , we have ( z h z ) & & ( c_hzz)^2 ( 1 - ) ^3 + & & ( ) ^2 ^2 ( - ) ( 1 - ) ^3 . as was discussed in ref . @xcite , the boosted @xmath2 shows the divergent nature as @xmath133 due to the enhancement from the longitudinal component of the @xmath2 . the limit on the new physics contribution to the undetected width of the @xmath3 , which can be obtained by subtracting the sm prediction from the measured width , is @xmath134 ( 95% cl ) from lep data @xcite . for @xmath7 masses below the @xmath3 mass , we can see from fig . [ fig : lep ] ( a ) that @xmath135 is always greater than @xmath136 for l3 . the lower bound on @xmath124 is given as a function of @xmath137 in fig . [ fig : mhbound ] for two extreme values of @xmath135 . in the doublets only case ( @xmath138 ) , we would have had @xmath139 . since the @xmath124 and @xmath140 would then be of similar size , @xmath141 and @xmath142 would be also of similar size , unless some of the @xmath143 values in the higgs mass - squared matrix in eq . are larger than @xmath144 , violating perturbativity . the requirement of perturbativity then gives a limit of @xmath145 ( or @xmath146 ) . as discussed in the appendix , experimental constraints provide a tight bound on @xmath76 in the doublets only case which conflicts with this limit . with a presence of the higgs singlet , @xmath147 can be lower than @xmath148 , and the @xmath124 bound gets relaxed , as fig . [ fig : mhbound ] shows . for @xmath149 , there is essentially no bound on @xmath124 from the lep invisible decay width . thus , we find that the doublets only case is unacceptable in the scenario we consider . we see that for sufficiently small @xmath147 , one can have @xmath150 , leading to the possibility of @xmath151 . as shown in ref . @xcite , this decay would unacceptably dominate @xmath6 decays unless @xmath113 was infinitesimally close to @xmath148 . we will not consider this possibility in this paper . the discovery of a higgs boson , @xmath6 , at @xmath8 gev with roughly sm branching ratios will imply stringent constraints on additional decays of the @xmath6 . in particular , the new decays @xmath152 and @xmath153 should be sufficiently small . for a sufficiently light @xmath2 , using eqs . and , we have ( h z z ) ( ) ^2 ^2 ( - ) and ( h z z ) ( ) ^4 ( ) ^2 . in fig . [ fig : darkdecays ] , we show the region in which their partial decay widths are within @xmath154 of the sm total decay widths ( @xmath155 for the @xmath5 higgs ) , illustrating the sensitive parameter - space for each decay for a few values of @xmath156 , @xmath157 , @xmath158 . in the region between the colored bands , the partial decay widths are more than @xmath154 of the sm total decay width . a very light @xmath2 at the lhc could be found by a narrow resonance search after appropriate cuts as studied in refs . constraints on the model from these decay modes need a dedicated lhc data analysis . figure [ fig : moreconstraints ] ( a ) shows the parameter - space where the total decay width of the heavy higgs @xmath6 is close to the sm width ( within @xmath159 ) . we have used a modified version of the code of ref . @xcite for some of the plots in this section . it shows that for a relatively large @xmath147 , the @xmath160 would be quite different from the sm prediction except for a very narrow region of parameter - space . + ( a ) ( b ) + ( a ) ( b ) the scenario that the heavy higgs is the sm - like higgs in ordinary 2hdms were discussed in ref . we follow their parametrization to measure the event rate ratio compared to the sm predictions . for the @xmath9 , _ h ( ) = ( ) ^2 with @xmath161 being our calculated value , instead of the precise value of the sm @xcite . similarly , @xmath162 , etc . are defined for other decay modes . the prefactor comes from the relative yukawa coupling of the @xmath6 and sm quarks [ eq . ] , which is relevant for the dominant gluon fusion @xmath163 through quark loops . as noted in ref . @xcite , the number of events ( production cross section times branching ratio ) of the @xmath6 into many modes can vary up a factor of two , given currently available data . figure [ fig : moreconstraints ] ( b ) shows the parameter - space in the dark 2hdm where the branching ratios of the @xmath164 , @xmath165 , @xmath107 and @xmath108 are consistent with the sm prediction up to a factor of two ( @xmath166 ) . in the region between the bands , at least one of these event rate ratios is more than a factor of two different from the sm prediction . the @xmath167 mode is the decay mode that drove the recent discovery of the @xmath5 higgs boson at the lhc . we plot the allowed region corresponding to @xmath168 in fig . [ fig : diphoton ] . the exact @xmath169 is not achievable with given choices of @xmath137 ( @xmath170 , @xmath157 , @xmath158 ) , and requires much smaller values to reach this limit . in the ordinary type i model , it is well known that one can not substantially exceed @xmath169 . one can expand the allowed parameter - space considerably by going to smaller @xmath137 values . we also note the parameter - space of the diphoton constraint in fig . [ fig : diphoton ] is covered , except for a very little portion at the edge , by the parameter - space of less stringent other modes in fig . [ fig : moreconstraints ] ( b ) . the effect of the charged higgs that can contribute to diphoton signals may depend on the parameters of the scalar potential , and it is ignored ( which is valid when the charged higgs is heavy enough ) . we will consider relatively light charged higgs scenario later in this paper though , which can change the diphoton rate . it is worth mentioning that @xmath171 may not be a decisive constraint as additional vector - like leptons can increase the diphoton decay rate although they can also alter other decays @xcite . + ( a ) ( b ) the most interesting decay of the @xmath7 would be @xmath172 . ( for simplicity , we consider only the @xmath173 case to avoid @xmath116 decay . ) for a sufficiently light @xmath2 , using eq . , we have ( h z z ) ( ) ^4 ( ) ^2 . we can compare this to the typically dominant decay channel , @xmath164 , ( h b|b ) ( ) ^2 which gives = ( ) ^2 ( ) ^2 . [ eq : lighthiggsratio ] figure [ fig : lighttodark ] shows the parameter region for @xmath174 and @xmath175 in which @xmath172 dominates the light higgs decay with @xmath176 , @xmath177 of the total @xmath7 decay . the @xmath18 mode can dominate the light higgs decay in a similar manner in which the higgs decay to the weak vector bosons would dominate in the sm if the higgs mass were sufficiently heavy . this originates from the enhancement from the longitudinal polarization of vector bosons when they are boosted . in accordance with expectation from eq . , [ fig : lighttodark ] shows that the @xmath172 dominates in a larger region of parameter - space with larger @xmath124 and larger @xmath147 . especially , in the @xmath178 and sufficiently large @xmath38 region , the @xmath7 decay is almost entirely into @xmath18 as indicated by @xmath177 dashed curves . + ( a ) ( b ) we note that for @xmath179 case , in which the lighter one would be the sm - like higgs , the heavier one would dominantly decay into @xmath180 . for some discussions of a higgs decaying into @xmath18 in different contexts , see refs . @xcite . the @xmath2 will then decay into fermions with a partial decay width , when fermion masses are neglected @xcite . ( z f |f ) _ z^2 g_z^2 ( g^2_v f + g_a f^2 ) m_z , where @xmath181 and @xmath182 . @xmath183 for quarks ( leptons ) . the branching ratio of @xmath2 into the charged leptons ( which we call @xmath184 and @xmath185 ) depends on @xmath186 as well as nontrivial hadronic decays . it was shown in ref . @xcite that the individual lepton branching ratios for dark photons ( i.e. a model with only kinetic mixing @xmath24 for the interaction ) vary , roughly , between @xmath154 and @xmath187 over most of the mass range . this branching ratio would change if there are other light hidden sector particles that @xmath2 can decay into . also , our dark @xmath3 ( i.e. a model with both kinetic mixing and @xmath3-@xmath2 mass mixing parametrized by @xmath28 ) has different branching ratios in general . we assume no displaced vertex and take @xmath188 in the following sketchy analysis . the light higgs might be detectable by a similar resonance search for the heavier higgs decaying into @xmath2 in sec . [ sec:125bounds ] although a feasibility study might be necessary . very recently , atlas has looked for prompt `` lepton - jets '' at @xmath189 @xcite . a lepton - jet is a final state consisting of collimated muons or electrons . in one of their analyses , with results consistent with the sm , they look for pairs of lepton - jets , each with two or more muons ( the rest of the event is ignored ) . although they do consider a specific hidden valley model as an example , their results will apply to our case . they only consider @xmath2 masses of @xmath190 and @xmath191 , and the results are relatively insensitive to the choice of masses . the resulting @xmath192 cl upper bound on the cross section times branching ratio is @xmath193 for @xmath194 and @xmath195 for a @xmath196 . for our ballpark estimate , we will take the bound to be @xmath197 without specifying the @xmath2 mass . the sm cross section for higgs production through gluon fusion is about @xmath198 for the higgs mass of @xmath199 at @xmath189 @xcite . in our case , the rate is multiplied by the factor @xmath200 of eq . . as shown in fig . [ fig : lighttodark ] , the @xmath7 can decay into @xmath18 dominantly in most of the parameter - space of interest . if we ignore the @xmath38 and @xmath113 dependence and take @xmath201 , the light higgs @xmath7 will appear as two muon - jets in the atlas analysis @xcite . taking @xmath202 , the cross section times branching ratio for @xmath203 is bounded as ( 37 pb ) ( ) ^2 x_^2 20 fb . for relatively large @xmath38 , this gives @xmath204 for a choice of @xmath205 . as @xmath38 approaches @xmath148 , this changes to @xmath206 . as the @xmath7 mass varies , these numbers change slightly , but it is clear that one is forced into a fairly small area of parameter - space . although these values of @xmath113 might seem fine - tuned , one should note that it varies from @xmath207 to @xmath208 in terms of @xmath209 as @xmath113 varies from @xmath210 to @xmath211 . given the level of our rather sketchy estimate , we do not take these bounds too seriously , but it is indicative that a sophisticated analysis can potentially shrink the allowed region considerably ( unless the effect was discovered ) . , with the upper ( lower ) line corresponding to @xmath213 . the dotted lines are similar for the decay @xmath214 . the solid line is for @xmath215 , which is either virtual or real , assuming @xmath202 and @xmath216 . the dot - dashed line is for @xmath217 with the @xmath5 sm - like higgs . the dominance of @xmath218 decay mode is persistent for all @xmath219 values allowed by the lep over the mass range of interest : @xmath110 . , scaledwidth=32.5% ] we have noted that the dark 2hdm is unusual in that there is no pseudoscalar higgs boson ( it becomes the longitudinal component of the @xmath2 ) associated with the doublets . however , it certainly does have a charged scalar , @xmath212 , which is orthogonal to the longitudinal component of the @xmath220 . the phenomenology of the charged higgs in the ordinary type i model was recently discussed in ref . @xcite . in the type i model as well as in this dark 2hdm , the coupling of the charged higgs to fermions is suppressed by a factor of @xmath38 . ( thus @xmath221 decay data , which pushes the charged higgs mass above @xmath222 in the type ii model , is not stringent in the type i model and in fact gives no bound for @xmath223 @xcite . ) direct bounds on the charged higgs mass from lep @xmath224 searches give @xmath225 assuming the charged higgs decays only into @xmath226 and @xmath227 @xcite . if a charged higgs is light enough , one can look for it in top quark decays , but as atlas data shows , the lower bound on @xmath38 in the type i model is fairly weak , varying from @xmath148 to @xmath207 in the mass range of @xmath228 when the dominant decay mode is @xmath227 @xcite . for charged higgs masses above about @xmath229 , its primary decay mode is generally given by @xmath230 , which is very difficult to detect . for some discussion of relatively light charged higgs ( @xmath231 ) in ordinary 2hdms , see ref . @xcite . in the type i model , one can also look for the decays @xmath232 where @xmath233 is either @xmath7 , @xmath6 or @xmath10 . in the dark 2hdm considered in this paper , however , there is a new and dramatic signature . we have seen that the primary decay mode of the @xmath7 in most of the relevant parameter - space of our interest is into @xmath18 . in this case , one can have @xmath234 . much of the time , the @xmath2 will decay into lepton pairs . this will certainly be the dominant decay if the @xmath235 decay is kinematically a two - body decay , and would still be dominant even if the decay is kinematically three - body , since the competing decay will be suppressed by two powers of the @xmath120 mass . the widths for the two and three - body decays of @xmath235 can be found in refs . the two - body decay width is ( h^h w^ ) = ^3/2(m_h^^2,m_h^2,m_w^2 ) [ eq : hhw ] with @xmath236 . it decreases with @xmath124 and increases with @xmath135 . similarly , the decay width into the heavy higgs ( @xmath5 ) is given by ( h^h w^ ) = ^3/2(m_h^^2,m_h^2,m_w^2 ) , and the decay width into @xmath237 is given by ( h^t |b ) ( 1 - ) ^2 when @xmath238 is ignored and @xmath239 is taken , and a similar expression is given for @xmath227 mode . figure [ fig : chargedhiggs ] shows the partial decay widths based on the above formulas for @xmath240 , @xmath241 , @xmath242 , @xmath243 assuming that @xmath202 and @xmath244 . ( the @xmath245 has a smaller contribution than @xmath246 as in the ordinary type i model because of the small mass at the higgs mass scale as the rg running is faster with color . ) for @xmath247 region , the three - body decay @xmath248 ( and similarly for @xmath6 ) is shown in the plot . we see that the decay into @xmath249 completely dominates for the entire range of @xmath212 masses in the plot ( @xmath250 ) . we checked that , for all our interested range of @xmath110 and @xmath251 that we discussed in sec . [ sec : constraints ] , the @xmath235 decay mode keeps dominating over the other @xmath212 decay modes for the mass range of @xmath252 . while the decay rate of the charged higgs is the same as in ordinary type i model , the difference comes from how the @xmath7 can decay , leading to more dramatic signatures . the production cross section for charged higgs bosons , for fairly large @xmath38 , is dominated by drell - yan pair production . each of these charged higgs can mainly decay into @xmath253 , giving four @xmath2s in the relevant parameter - space . this may be looked for with lepton - jet searches @xcite in a similar fashion we discussed in the previous section for @xmath254 channel , with appropriate cuts and selections . note that the lepton - jet searches only require at least two muon - jets . with four @xmath2s and a @xmath255 branching ratio for @xmath256 , a substantial fraction of charged higgs pairs will give a signal . one can estimate that the current atlas bound of 20 fb is already covering a region of parameter - space , but it is clear that a much more detailed analysis is needed . given the unique signature of the model , with four @xmath2s and two @xmath16 s , a more targeted search could cover much more of the parameter - space . in the dark 2hdm , there is no tree - level vertex @xmath257 for the same reason that there is no @xmath258 vertex at tree level . namely , if one goes into a basis in which only one higgs doublet gets a vev , then the charged higgs is entirely in the other doublet . as a result , it can have no vev - dependent vertices . since @xmath20 is broken , this result will break down at one - loop , and thus @xmath259 and @xmath260 can occur at one - loop . the effect of the goldstone boson equivalence theorem , with a boosted @xmath2 , would enhance this loop - suppressed decay potentially at meaningful level . we discussed , in this paper , only the case of @xmath261 . it will be interesting and important to study how low the charged higgs mass can be in this model while satisfying all the experimental and theoretical constraints , in view of the fact that most of the lower bounds on the @xmath262 were obtained based on the typical @xmath237 , @xmath245 and @xmath227 modes . it is also noteworthy that other variants of the dark 2hdm [ for example , with some of the sm fermions carrying nonzero @xmath11 charges ] would give the similar @xmath218 dominance since eq . is valid for all types of 2hdms . although experiments at the lhc have recently found the sm - like higgs boson with a mass of @xmath5 , it is still important to search for other higgs bosons , which occur in most extensions of the standard model . in this paper , we considered a 2hdm with a new @xmath0 gauge symmetry under which the sm particles are not charged ( the dark 2hdm ) , and discussed the physics of the other higgs bosons , namely , the non - sm - like neutral higgs and the charged higgs scalar . the additional higgs doublet of this dark 2hdm is charged by the @xmath0 gauge symmetry , and this @xmath0 plays the role of the @xmath1 parity of the ordinary 2hdms , thus forbidding tree - level fcnc . kinetic mixing between the @xmath0 gauge groups will generally occur , as will mixing in the @xmath3-@xmath2 mass matrix . the possibility of a very light @xmath2 gauge boson has attracted increasing interest of late . such a light @xmath2 has been a subject of active experimental searches including the fixed target experiments at jlab in virginia and at mainz in germany . there are also searches using the decays from mesons at kloe , babar , and belle experiments . the physics connecting the higgs to a heavy @xmath2 ( for example , the @xmath2 decays into the higgs boson @xcite ) has been extensively studied , but connection of the higgs to a very light @xmath2 has not been . the light @xmath2 allows a very interesting scenario as the various higgs boson decays can involve a light gauge boson which can decay into leptons with @xmath263 branching ratio along with possible enhancement from the longitudinal polarization . interestingly , in this model , the most important predictions / constraints on higgs properties come from the properties of the @xmath2 . the @xmath2 is assumed to have a mass of @xmath264 or less . such a light mass is the region most high energy collider analyses discard to avoid large sm backgrounds . we considered the case in which the @xmath5 sm - like higgs is the heavier scalar , and studied the light higgs and charged higgs in the dark 2hdm . after describing the complete model , we studied constraints arising from lep bounds for invisible higgs decays and width of the @xmath3 as well as the @xmath265 sm - like higgs state seen at the lhc . we found that a model with just two higgs doublets is excluded by the lep invisible higgs decay bounds combined with the precision electroweak physics , rare @xmath221 decay , etc . this issue can be resolved by introducing an additional higgs singlet . we then considered the production and detection of the light higgs boson ( @xmath7 ) whose dominant decay is into @xmath18 . this leads to a remarkable and unusual signature . the @xmath2 will decay with an appreciable branching fraction into a lepton pair , which ( since the @xmath2 is so light ) will form a collimated lepton jet . a recent atlas experiment looked for events with two or more such `` muon - jets '' , and their results show that significant constraints on the model would be possible if a sophisticated analysis with larger statistics were to follow . the atlas muon - jets experiment also constrains the charged higgs boson , which decays predominantly into @xmath249 ( either as a two - body or a three - body decay ) . the charged higgs in the dark 2hdm can be very light compared to those in the ordinary 2hdms . in this paper , we limited ourselves to the case in which the @xmath5 sm - like higgs is the heavier neutral higgs , no mixing is present between the higgs doublets and a singlet , and we focused on the region in which the light higgs mass is in the range of @xmath109 . relaxing these limits will allow more studies . acknowledgments : this work was supported in part by the u.s . doe under grant no . de - ac05 - 06or23177 ( jlab ) and in part by the nsf under grant no . phy-1068008 ( w&m ) . we thank p. ferreira and r. santos for useful discussions about the code that was used , after modification , to generate some numerical results , and atlas experimenters a. hass , e. strauss , and g. watts for discussions about their muon - jet analysis . hl thanks h. davoudiasl , i. lewis , and w. marciano for many helpful discussions during the dark @xmath3 projects . in the @xmath71 limit , as shown in sec . [ sec : model ] , the @xmath3-@xmath2 mass - squared matrix is given by m_zz^2 = . we take the parametrization @xmath266 which helps separate the @xmath2 interactions with @xmath267 and @xmath268 . the @xmath24 comes from the kinetic mixing term , and the @xmath28 originates from the @xmath3-@xmath2 mass mixing from the higgs ( @xmath41 ) that is charged under both @xmath269 and @xmath11 . taking @xmath270 which is realized for @xmath271 . the @xmath3-@xmath2 mass mixing parameter @xmath28 is further parametrized by _ z with , from eqs . - , & & this would have been @xmath272 in the doublets only case ( @xmath80 ) . the approximation @xmath273 would be valid in the limit where both @xmath38 and @xmath39 are large , and partly because of this reason , we use @xmath147 as our input instead of @xmath76 in the numerical analysis of this paper , which helps in estimating the higgs singlet contribution . ignoring the higher order terms for small @xmath24 and @xmath28 parameters ( and their combination ) , we get _ & = & - e j^_em a_- g_z j^_nc z^0 _ + & & - e j^_em ( a_+ z^0 _ ) - g_z j^_nc ( z^0_- _ w z^0 _ ) [ eq : kinetic ] + & & - e j^_em ( a_+ z _ ) - g_z j^_nc ( z_+ _ z z _ ) [ eq : mass ] where eq . is obtained after field redefinition to remove kinetic mixing term [ eq . ] at leading order and eq . is after the @xmath3-@xmath2 mass - squared matrix diagonalization [ eq . ] . we can see that , because of cancellation , there is no net @xmath2 coupling to the weak neutral current induced by the kinetic mixing @xmath274 . for a very light @xmath2 , the kinetic mixing @xmath24 is constrained by various experiments including the electron beam dump , electron anomalous magnetic moment , and narrow resonance searches . in particular , there are active searches using fixed target experiments at jlab ( in virginia ) and at mainz ( in germany ) as well as the searches using the heavy meson decays at kloe , babar , belle experiments . very roughly , for the @xmath275 range , the experimental bounds of @xmath276 are present . with some combination of @xmath24 and @xmath34 ( roughly , @xmath277 and @xmath278 ) , the @xmath2 can explain the @xmath15 deviation of the muon @xmath4 . ( see ref . @xcite for details . ) the @xmath24 basically parametrizes the vector coupling of the @xmath2 as it dominates the coupling to the electromagnetic current . a good summary of the constraints and sensitivities from various experiments in the @xmath279 parameter - space can be found in ref . @xcite . the @xmath2 interaction to the weak neutral current is enabled by the @xmath3-@xmath2 mass mixing @xmath28 , and it controls the axial coupling of the @xmath2 as it dominates the coupling to the weak neutral current . the axial couplings expand the phenomenology of the light @xmath2 into more areas including the low energy parity violation and the enhancement of the production for the boosted @xmath2 . they were studied in refs . they include low - energy parity violation , rare @xmath280 decays , and rare @xmath221 decays . also , bounds from the higgs decay exists as the presence of the @xmath281 and @xmath282 modes should be still consistent with the lhc data which is consistent with the @xmath5 sm higgs property . typically , @xmath283 , depending on various conditions , is expected to satisfy all the constraints . there are some caveats and dependencies on other parameters about this bounds , and we will take @xmath284 as a firm upper bound that we should satisfy in this paper . this is quite a small quantity that can not be achieved with the pure doublets case as @xmath285 can not be too small ( as discussed in sec . [ sec:3b ] ) , which is an important reason that this model needs a higgs singlet . 99 g. aad _ et al . _ [ atlas collaboration ] , phys . lett . b * 716 * , 1 ( 2012 ) [ arxiv:1207.7214 [ hep - ex ] ] . s. chatrchyan _ et al . _ [ cms collaboration ] , phys . b * 716 * , 30 ( 2012 ) [ arxiv:1207.7235 [ hep - ex ] ] . for example , see p. m. ferreira , r. santos , m. sher and j. p. silva , phys . d * 85 * , 077703 ( 2012 ) [ arxiv:1112.3277 [ hep - ph ] ] . s. dawson and e. furlan , phys . d * 86 * , 015021 ( 2012 ) [ arxiv:1205.4733 [ hep - ph ] ] . m. carena , i. low and c. e. m. wagner , jhep * 1208 * , 060 ( 2012 ) [ arxiv:1206.1082 [ hep - ph ] ] . n. bonne and g. moreau , phys . b * 717 * , 409 ( 2012 ) [ arxiv:1206.3360 [ hep - ph ] ] . h. an , t. liu and l. -t . wang , phys . rev . d * 86 * , 075030 ( 2012 ) [ arxiv:1207.2473 [ hep - ph ] ] . a. joglekar , p. schwaller and c. e. m. wagner , jhep * 1212 * , 064 ( 2012 ) [ arxiv:1207.4235 [ hep - ph ] ] . n. arkani - hamed , k. blum , r. t. dagnolo and j. fan , jhep * 1301 * , 149 ( 2013 ) [ arxiv:1207.4482 [ hep - ph ] ] . l. g. almeida , e. bertuzzo , p. a. n. machado and r. z. funchal , jhep * 1211 * , 085 ( 2012 ) [ arxiv:1207.5254 [ hep - ph ] ] . j. kearney , a. pierce and n. weiner , phys . d * 86 * , 113005 ( 2012 ) [ arxiv:1207.7062 [ hep - ph ] ] . m. a. ajaib , i. gogoladze and q. shafi , phys . d * 86 * , 095028 ( 2012 ) [ arxiv:1207.7068 [ hep - ph ] ] . h. davoudiasl , h .- s . lee and w. j. marciano , phys . d * 86 * , 095009 ( 2012 ) [ arxiv:1208.2973 [ hep - ph ] ] . s. dawson , e. furlan and i. lewis , phys . d * 87 * , 014007 ( 2013 ) [ arxiv:1210.6663 [ hep - ph ] ] . h. davoudiasl , i. lewis and e. ponton , phys . d * 87 * , 093001 ( 2013 ) [ arxiv:1211.3449 [ hep - ph ] ] . g. c. branco , p. m. ferreira , l. lavoura , m. n. rebelo , m. sher and j. p. silva , phys . * 516 * , 1 ( 2012 ) [ arxiv:1106.0034 [ hep - ph ] ] . s. l. glashow and s. weinberg , phys . d * 15 * ( 1977 ) 1958 . e. a. paschos , phys . d * 15 * ( 1977 ) 1966 . t. p. cheng and m. sher , phys . d * 35 * ( 1987 ) 3484 . a. pich and p. tuzon , phys . d * 80 * , 091702 ( 2009 ) [ arxiv:0908.1554 [ hep - ph ] ] . h. davoudiasl , h .- s . lee and w. j. marciano , phys . d * 85 * , 115019 ( 2012 ) [ arxiv:1203.2947 [ hep - ph ] ] . p. ko , y. omura and c. yu , phys . b * 717 * , 202 ( 2012 ) [ arxiv:1204.4588 [ hep - ph ] ] . lee and a. soni , phys . lett . * 110 * , 021802 ( 2013 ) [ arxiv:1206.6110 [ hep - ph ] ] . for example , see v. barger , p. langacker , h .- s . lee and g. shaughnessy , phys . d * 73 * , 115010 ( 2006 ) [ hep - ph/0603247 ] . p. langacker , rev . phys . * 81 * , 1199 ( 2009 ) [ arxiv:0801.1345 [ hep - ph ] ] . c. boehm , d. hooper , j. silk , m. casse and j. paul , phys . lett . * 92 * , 101301 ( 2004 ) [ astro - ph/0309686 ] . c. boehm , p. fayet and j. silk , phys . d * 69 * , 101302 ( 2004 ) [ hep - ph/0311143 ] . n. arkani - hamed , d. p. finkbeiner , t. r. slatyer and n. weiner , phys . d * 79 * , 015014 ( 2009 ) [ arxiv:0810.0713 [ hep - ph ] ] . p. jean , j. knoedlseder , v. lonjou , m. allain , j .- roques , g. k. skinner , b. j. teegarden and g. vedrenne _ et al . _ , astron . astrophys . * 407 * , l55 ( 2003 ) [ astro - ph/0309484 ] . j. chang , j. h. adams , h. s. ahn , g. l. bashindzhagyan , m. christl , o. ganel , t. g. guzik and j. isbert _ et al . _ , nature * 456 * , 362 ( 2008 ) . o. adriani _ et al . _ [ pamela collaboration ] , nature * 458 * , 607 ( 2009 ) [ arxiv:0810.4995 [ astro - ph ] ] . g. w. bennett _ et al . _ [ muon g-2 collaboration ] , phys . d * 73 * , 072003 ( 2006 ) [ hep - ex/0602035 ] . j. beringer _ [ particle data group collaboration ] , phys . d * 86 * , 010001 ( 2012 ) . m. baumgart , c. cheung , j. t. ruderman , l. -t . wang and i. yavin , jhep * 0904 * , 014 ( 2009 ) [ arxiv:0901.0283 [ hep - ph ] ] . f. chen , j. m. cline and a. r. frey , phys . d * 80 * , 083516 ( 2009 ) [ arxiv:0907.4746 [ hep - ph ] ] . h. zhang , c. s. li , q .- h . cao and z. li , phys . d * 82 * , 075003 ( 2010 ) [ arxiv:0910.2831 [ hep - ph ] ] . c. d. carone , phys . b * 721 * , 118 ( 2013 ) [ arxiv:1301.2027 [ hep - ph ] ] . t. alanne , s. di chiara and k. tuominen , arxiv:1303.3615 [ hep - ph ] . h. davoudiasl , h .- s . lee and w. j. marciano , phys . lett . * 109 * , 031802 ( 2012 ) [ arxiv:1205.2709 [ hep - ph ] ] . h. davoudiasl , h .- s . lee , i. lewis and w. j. marciano , arxiv:1304.4935 [ hep - ph ] . m. ferreira , r. santos , m. sher and j. p. silva , phys . d * 85 * , 035020 ( 2012 ) [ arxiv:1201.0019 [ hep - ph ] ] . b. holdom , phys . b * 166 * , 196 ( 1986 ) . s. gopalakrishna , s. jung and j. d. wells , phys . d * 78 * , 055002 ( 2008 ) [ arxiv:0801.3456 [ hep - ph ] ] . r. barate _ [ lep working group for higgs boson searches and aleph and delphi and l3 and opal collaborations ] , phys . b * 565 * , 61 ( 2003 ) [ hep - ex/0306033 ] . s. schael _ et al . _ [ aleph and delphi and l3 and opal and lep working group for higgs boson searches collaborations ] , eur . phys . j. c * 47 * , 547 ( 2006 ) [ hep - ex/0602042 ] . m. ferreira and d. r. t. jones , jhep * 0908 * , 069 ( 2009 ) [ arxiv:0903.2856 [ hep - ph ] ] . r. barate _ [ aleph collaboration ] , phys . b * 466 * , 50 ( 1999 ) . j. abdallah _ et al . _ [ delphi collaboration ] , eur . j. c * 32 * , 475 ( 2004 ) [ hep - ex/0401022 ] . p. achard _ et al . _ [ l3 collaboration ] , phys . b * 609 * , 35 ( 2005 ) [ hep - ex/0501033 ] . g. abbiendi _ et al . _ [ opal collaboration ] , phys . b * 682 * , 381 ( 2010 ) [ arxiv:0707.0373 [ hep - ex ] ] . a. heister _ et al . _ [ aleph collaboration ] , phys . b * 526 * , 191 ( 2002 ) [ hep - ex/0201014 ] . m. s. carena , a. de gouvea , a. freitas and m. schmitt , phys . d * 68 * , 113007 ( 2003 ) [ hep - ph/0308053 ] . a. denner , s. heinemeyer , i. puljak , d. rebuzzi and m. spira , eur . j. c * 71 * , 1753 ( 2011 ) [ arxiv:1107.5909 [ hep - ph ] ] . o. lebedev , h. m. lee and y. mambrini , phys . b * 707 * , 570 ( 2012 ) [ arxiv:1111.4482 [ hep - ph ] ] . s. thomas , `` dark forces : searches for new forces at the gev - scale , '' slac national accelerator lab , 2009 ( unpublished ) . m. jung , a. pich and p. tuzon , jhep * 1011 * , 003 ( 2010 ) [ arxiv:1006.0470 [ hep - ph ] ] . [ lep higgs working group for higgs boson searches and aleph and delphi and l3 and opal collaborations ] , hep - ex/0107031 . g. aad _ et al . _ [ atlas collaboration ] , jhep * 1206 * , 039 ( 2012 ) [ arxiv:1204.2760 [ hep - ex ] ] . m. aoki , r. guedes , s. kanemura , s. moretti , r. santos and k. yagyu , phys . d * 84 * , 055028 ( 2011 ) [ arxiv:1104.3178 [ hep - ph ] ] . a. djouadi , j. kalinowski and p. m. zerwas , z. phys . c * 70 * , 435 ( 1996 ) [ hep - ph/9511342 ] . a. g. akeroyd , nucl . b * 544 * , 557 ( 1999 ) [ hep - ph/9806337 ] . k. agashe , h. davoudiasl , s. gopalakrishna , t. han , g .- y . huang , g. perez , z .- g . si and a. soni , phys . d * 76 * , 115015 ( 2007 ) [ arxiv:0709.0007 [ hep - ph ] ] . v. barger , p. langacker and h .- s . lee , phys . lett . * 103 * , 251802 ( 2009 ) [ arxiv:0909.2641 [ hep - ph ] ] . r. d. mckeown , aip conf . proc . * 1423 * , 289 ( 2012 ) [ arxiv:1109.4855 [ hep - ex ] ] .

quantum correlations have a central role in quantum information processing , in quantum foundations , as well as in the physics of strongly correlated systems @xcite . on one hand , quantum correlations , and in particular entanglement , are a resource that allows one to go beyond what classically possible in many scenarios , from communication tasks , to ( measurement - based ) quantum computing , to quantum cryptography . on the other hand , the non - classicality of quantum correlations be it in the form of non - locality , steering , entanglement , or discord is one of the most distinctive traits of quantum mechanics , and challenges our understanding of quantum mechanics itself . the interplay between local and global properties of quantum states is a key aspect in the study of quantum correlations , both from a fundamental perspective and an applicative one . for example , we may want to certify the presence of multipartite entanglement in large systems without the often inaccessible knowledge of the global state , using instead only the information that comes from reduced states . on the other hand , in condensed - matter physics , because of the typically local , two - body interactions , relevant properties are dictated by the interplay between the allowed reduced states and global correlations , giving raise to phenomena like frustration @xcite . the general study of the relations between the properties of the reduced states and the properties of the global state is known as the quantum marginal problem , which has seen a growing interest in the past years also for the reasons above @xcite . in this work we study what can be inferred about the quality of the correlations of the global state given information about the two - body reduced states , aiming at answering the question : _ what correlations need to be present globally to explain what we see locally ? _ in @xcite a characterization of multipartite entanglement in terms of even just single - party reduced states ( actually , single - party spectra ) was given , but under the assumption of dealing with a pure or quasi - pure global state . in @xcite the possibility of dealing with global mixed states is taken into account , and examples are given where two - qubit separable states are only compatible with global entanglement , intended in the sense of lack of total separability ( see section [ sec : definitions ] for definitions ) . in @xcite examples are also given where genuine multipartite entanglement a much stronger notion of global entanglement can still be deduced from the properties of the two - body reduced states , but only when these reduced states exhibit bipartite entanglement themselves . in this work we present several results that complement and generalize those of @xcite . we offer a brief summary of our findings in table [ table ] . firstly , we provide examples of triples of bipartite reduced states in the simplest case , two - qubit states that , albeit separable , are only compatible with genuine tripartite entanglement ( lower - right corner of table [ table ] ) . as far as we know , this is the strongest `` gap '' known between the entanglement properties of the marginals and of the global state . secondly , we address the issue of relating the general quantumness of correlations @xcite of reduced states to the quantum correlations of the global state . we find that strictly classical reduced states may still be compatible only with global entanglement ( upper - left corner of table [ table ] ) . thirdly , we find that , at least for qubits , the strict classicality of the two - body correlations makes it impossible to certify genuine tripartite entanglement based on the knowledge of the reductions : strictly classical two - qubit reduced states are always compatible with a global state that is not genuine tripartite entangled ( upper - right corner of table [ table ] ) . the rest of the paper is organized as follows . in section [ sec : definitions ] we define the relevant notions of correlations and classicality , and of compatibility of two - body reduced states in tripartite systems . in section [ sec : classicalcomp ] we study the relation between the classicality of reductions and their compatibility . in section [ sec : genuine ] we prove that unentangled reduced states may only be compatible with genuine multipartite entanglement at the level of the global state . finally , we conclude in section [ sec : conclusions ] . l l|>m1.5cm|>m3.5cm| & & + & & entangled & genuinely multipartite + entangled & fully + classical & & ( qubits ) & separable & @xcite & we begin by formally defining qualitatively different types of correlations . [ def : separable ] any tripartite mixed state can be written as a mixture of an ensemble of pure states as @xmath0 . we say that @xmath1 is : * _ fully separable _ , if we can take each @xmath2 to be fully factorized , e.g. @xmath3 ; * _ biseparable _ , if we can take each @xmath2 to be unentangled in at least one partition , e.g. , @xmath4 , @xmath5 , or @xmath6 ; * _ genuinely multipartite entangled _ , if for any ensemble there is at least one @xmath7 with @xmath8 that is not factorized with respect to any bipartition , i.e. , if @xmath1 is not biseparable ; * _ fully classical _ , if we can take each @xmath2 to be of the form @xmath9 , with @xmath10 orthonormal basis on @xmath11 , @xmath12 , and @xmath13 , respectively , so that , overall , @xmath14 . bipartite full classicality and separability are defined similarly : @xmath15 is _ fully classical _ if @xmath16 , for @xmath17 orthonormal bases , and _ separable _ if @xmath18 . a bipartite state is _ entangled _ if it is not separable . the notions of full separability and biseparability are redundant for bipartite states . the set of fully classical states , @xmath19 , is a subset of the set of fully separable states , @xmath20 , which in turn is a subset of the set of biseparable states , @xmath21 ; the set of genuinely multipartite entangled states , @xmath22 is the is the complement of @xmath21 in the space of all states , @xmath23 ( see figure [ fig : sets ] ) . all the mentioned sets apart from @xmath19 and @xmath22 are convex . a biseparable state may be either separable or entangled in any bipartition , but it is by definition the convex sum of three states that are each separable in one of the bipartitions @xmath24 , @xmath25 , and @xmath26 . we now move to define formally the notion of compatibility for reduced states . [ def : compatible ] given a triple of three two - party states @xmath27 , its _ compatibility set _ is defined as @xmath28 . any compatibility set is a convex set @xcite , and the property of being part of a given compatibility set defines an equivalence relation . we find it useful to denote by @xmath29 the compatibility set associated with the reduced states of @xmath30 , i.e. , the set of all states that have the same reductions as @xmath30 . a triple of two - party states @xmath31 is said to be _ compatible _ ( so that we refer to the triple as _ triple of reductions _ ) if @xmath32 , i.e. , if there is at least one global state with those reductions . the following definition links the compatibility of reduced states to the correlation properties of global states . [ def : compatible2 ] we say that the reductions @xmath31 are _ incompatible with a set @xmath33 _ ( or with the defining correlation property of @xmath33 ) if @xmath34 . we say that ( a compatible ) @xmath31 is only compatible with genuine multipartite entanglement if it is incompatible with @xmath21 . figure [ fig : sets ] illustrates the problem of deciding whether certain bipartite reductions necessarily require the presence of global correlations of a certain kind . we can always expand a generic tripartite state as @xmath35 where @xmath36 is the reduced states of party @xmath37 and @xmath38 is the normalized identity operator on the hilbert space @xmath39 . the bipartite correlation matrices @xmath40 can be defined via @xmath41 , and satisfy @xmath42 = { \tr}_l [ \chi_{kl}]=0 $ ] . it is worth noting that when a bipartite marginal state @xmath43 is fully classical , then @xmath44 = 0 $ ] , and @xmath45 is also diagonal in the same product basis as @xmath46 . the tripartite correlation matrix @xmath47 , which for a fixed @xmath1 can be defined via , satisfies @xmath48 = 0 $ ] , for all @xmath49 . for compatible reductions @xmath50 , the compatibility set @xmath51 is spanned by choosing the tripartite correlation matrix @xmath47 so that the resulting operator in is a physical state , i.e. , positive semidefinite . on the other hand , to determine whether a triple of bipartite states is compatible , we first check the basic necessary condition that the single party marginals be the same , i.e. , @xmath52 = \tr_k[\rho_{ik}]$ ] for all @xmath53 . next , we have to search for a tripartite correlation matrix , @xmath47 , such that eq . is physical . if no such @xmath47 exists , the given states are not compatible . consider the marginals from the well - known greenberger - horne - zeilinger ( ghz ) state @xmath54 : @xmath55 . these are fully classical marginals coming from a genuinely tripartite entangled state . however , these marginals are also compatible with @xmath56 , which is fully - classical . in this section we will first provide an example where fully classical two - body reduced states are not compatible with a global fully classical state , and actually require the presence of entanglement . on the other hand , we will prove that , in the case of three qubits , the fully classical two - body reductions are always compatible with a global states that is not genuine multipartite entangled . we will now derive conditions to ensure that some fully - classical marginals can not be compatible with any global fully - classical state . we start with the following lemma . [ le : commutatorlemma ] suppose three states @xmath57 , @xmath58 , and @xmath59 are fully classical . consider the commutators @xmath60 = \left [ \rho_{ij } \ox \openone_k , \ ; \rho_{ik } \ox \openone_j \right]$ ] , where the second equality is due to the assumed classicality , i.e. , @xmath61 = 0 $ ] . then : a. all commutators @xmath62 vanish if and only if there are orthonormal basis @xmath63 , @xmath64 , @xmath65 , such that @xmath66 , @xmath67 , and @xmath68 . b. if some commutator @xmath62 does not vanish , then : a. at least one @xmath69 is degenerate ( i.e. , at least two eigenvalues of some @xmath69 are the same ) ; b. there does not exist a tripartite fully classical state that is compatible with @xmath57 , @xmath59 , and @xmath58 . \(i ) the `` if '' part is trivial . let us prove the `` only if '' part . by hypothesis we may assume @xmath70 with the orthonormal basis @xmath71 on @xmath11 , @xmath72 on @xmath12 , @xmath73 on @xmath13 , and @xmath74 , @xmath75 ( similarly for @xmath76 , etc ) . then @xmath77 implies @xmath78 \label{eq : commute } \notag\\ & \quad=\left [ ( \r_{ij}-\r_i\ox\r_j ) \ox \openone_k , \ ; ( \r_{ik}-\r_i\ox\r_k ) \ox \openone_j \right ] \notag\\ & \quad=\left [ \r_{ij } \ox \openone_k , \ ; \r_{ik } \ox \openone_j \right]=0,\end{aligned}\ ] ] with @xmath79 . by setting @xmath80 in , we have @xmath81=0 $ ] , @xmath82 . thus the states @xmath83 are simultaneously diagonalizable in the orthonormal basis @xmath84 . so we may replace the basis @xmath63 and @xmath85 in and by @xmath84 . this replacement may result in the change of @xmath86 and @xmath87 . since there is no confusion , we still use them in and . next by setting @xmath88 in , we have @xmath89=0 $ ] , @xmath82 . thus the states @xmath90 are simultaneously diagonalizable in the orthonormal basis @xmath91 . so we may replace the basis @xmath64 and @xmath92 in and by @xmath91 . this replacement may result in the change of @xmath93 . since there is no confusion , we still use them in . third we set @xmath94 in and repeat the above argument to show that the basis @xmath65 and @xmath95 in and can be replaced by the orthonormal basis @xmath96 . so the assertion follows . \(ii ) suppose either condition ( a ) or ( b ) is violated . we have that either @xmath97 , @xmath98 , and @xmath99 are all non - degenerate respectively in the orthonormal basis @xmath63 , @xmath64 , @xmath65 ( violation of ( a ) ) , or that there exists a tripartite fully classical state @xmath100 that is compatible with @xmath57 , @xmath59 , and @xmath58 ( violation of ( b ) ) . in either case we have @xmath101 , @xmath102 and @xmath67 . so ( i ) implies that all commutators @xmath103 vanish , and we have a contradiction . an immediate consequence of lemma [ le : commutatorlemma ] is the following . [ thm : commuteimpliesclassical ] let @xmath57 , @xmath59 , and @xmath58 be three compatible bipartite fully classical states , such that \(i ) they all commute ( all commutators @xmath103 of lemma [ le : commutatorlemma ] vanish ) , or \(ii ) all three one - body reduced states @xmath97 , @xmath104 and @xmath99 are non - degenerate . then @xmath57 , @xmath59 , and @xmath58 are compatible with a fully classical tripartite state . the fact that all three single - system reductions are not degenerate implies , by lemma [ le : commutatorlemma].(iia ) , that all the commutators @xmath103 defined in lemma [ le : commutatorlemma ] vanish . by lemma [ le : commutatorlemma].(i ) we have that @xmath97 , @xmath98 , and @xmath99 , are diagonal in the orthonormal bases @xmath63 , @xmath105 , and @xmath106 , respectively , in which @xmath57 , @xmath59 , and @xmath58 are explicitly classical . most importantly , we have @xmath107 with @xmath108 and @xmath109 . let @xmath1 be any tripartite state with which the three two - body reductions are compatible . then also the fully classical tripartite state @xmath110 has bipartite reduced density matrices @xmath57 , @xmath59 , and @xmath58 . given theorem [ thm : commuteimpliesclassical ] , in order to construct an example where @xmath57 , @xmath58 , and @xmath59 are all fully classical but not compatible with any fully - classical state , we have first of all to construct an example where @xmath57 , @xmath58 , and @xmath59 are classical but do not commute with each other . for this , we will need the following lemma . we recall that for a bipartite state @xmath111 acting on the hilbert space @xmath112 , the partial transpose computed in the standard orthonormal basis @xmath113 of system @xmath114 , is defined by @xmath115 . one can similarly define the partial transpose @xmath116 on the system @xmath117 . [ le : classicalmain ] consider three classical - classical two - qubit states @xmath118 where @xmath119 , and any one of @xmath63 , @xmath64 , @xmath65 is a real and orthonormal basis in @xmath120 . let @xmath121 then : a. if @xmath122 then @xmath1 is separable with respect to to the partition @xmath24 , @xmath25 and @xmath26 . b. @xmath57 , @xmath59 and @xmath58 are compatible if and only if they are compatible with the biseparable state @xmath1 in . c. suppose @xmath57 , @xmath59 and @xmath58 are compatible they are compatible with a fully separable state if and only if @xmath1 is fully separable . \(i ) one may directly verify that the state is invariant under partial transposition with respect to any system , i.e. , @xmath123 for @xmath124 . since @xmath122 , the assertion follows from ( * ? ? ? * theorem 2 ) . \(ii ) the `` if '' part is trivial ; let us prove the `` only if '' part . suppose the bipartite marginals @xmath57 , @xmath59 and @xmath58 are compatible with a tripartite state @xmath125 . since @xmath57 , @xmath59 , and @xmath58 are real , they are also compatible with the real state @xmath126 , so we can assume that @xmath127 is real without loss of generality . by eq . ( 1 ) , there is a hermitian matrix @xmath47 such that @xmath128 since in our case both @xmath129 and @xmath130 are real , also @xmath131 is real . it follows from eqs . , and the fact that @xmath63 , @xmath64 , @xmath65 are real and orthonormal bases , that @xmath57 , @xmath59 and @xmath58 are invariant under the local unitary @xmath132 . so they are compatible with the state @xmath133 where we have used that , from , @xmath134 we will now argue that , for a real @xmath131 , @xmath135 so that proves that , for a physical state @xmath130 , @xmath129 is also physical , as it corresponds to the convex combination of physical states . the starting point in proving is to observe that every three - qubit correlations matrix @xmath131 is by definition the linear combination of traceless pauli matrices , i.e. @xmath136 since @xmath131 is hermitian , all coefficients @xmath137 are real . moreover , for a real ( and hence symmetric ) @xmath131 only terms with an even number of @xmath138 are present in the expansion , because @xmath139 , while @xmath140 and @xmath141 are symmetric . on the other hand , @xmath142 for @xmath143 , while , obviously , @xmath144 . since each non - zero term in the expansion of @xmath131 contains an even number of @xmath145 s , it will change sign after conjugation by @xmath146 , i.e. , @xmath147 as argued , this implies @xmath122 , with biseparability following from ( i ) . \(iii ) the `` if '' part follows from ( ii ) , let us prove the `` only if '' part . suppose @xmath57 , @xmath59 and @xmath58 are compatible with a fully separable state @xmath125 . from , @xmath1 is the convex sum of a few fully separable states . so the assertion follows . this completes the proof . we are now ready to present our example where @xmath57 , @xmath58 , and @xmath59 are all fully classical but not compatible with any fully - classical state . [ eg : commutatorex ] consider the three - qubit state @xmath148 where @xmath149 , with @xmath150 are the pauli matrices and @xmath151 ( for @xmath152 outside of this interval the matrix @xmath153 is not positive semi - definite ) , @xmath154 . it is not hard to see that each of the bipartite marginals states is fully classical but with respect to different bases . a quick method to verify this assertion is using theorem 1 in @xcite . moreover the reductions of @xmath153 do not commute with each other . part ( ii ) of lemma [ le : commutatorlemma ] implies that there is no tripartite fully classical state that is compatible with these bipartite marginals : @xmath155 . surprisingly , we can find a value of @xmath152 for which the fully classical marginals in fact require some global entanglement . consider the state @xmath156 . the range of @xmath157 does not contain any fully factorized pure state , hence it can not be fully separable . now , up to a local unitary , the classical bipartite marginals @xmath158 , @xmath159 , @xmath160 , and the entangled state @xmath157 can be written as - , respectively . it follows from lemma [ le : classicalmain].(iii ) that @xmath158 , @xmath159 , and @xmath160 can not be compatible with any fully separable state . in appendix [ sec : appuni ] , which focuses on the uniqueness of global states with fixed two - body reductions , we provide also an example of classical reductions compatible with a _ unique _ global state that is not fully classical , although fully separable . although , as we have just seen , compatible classical marginals may require global quantum correlations or even entanglement , it turns out that for the case of _ three qubits _ they will never require the global state to be genuinely multipartite entangled . to see this , will need an additional lemma . [ le : nongenuine ] let @xmath63 , @xmath64 be an orthonormal basis on @xmath11 and @xmath12 , respectively . suppose the bipartite marginals @xmath57 , @xmath58 and @xmath59 are compatible . they are compatible to a non - genuinely entangled tripartite state when one of the following conditions is satisfied : ( i ) @xmath161 and @xmath162 . ( ii ) @xmath163 . suppose @xmath57 , @xmath58 and @xmath59 are compatible with @xmath1 . if hypothesis ( i ) is satisfied , then they are also compatible with the state @xmath164 , which is biseparable . on the other hand if hypothesis ( ii ) is satisfied , then it follows from @xcite that there is a quantum channel @xmath165 on @xmath13 such that @xmath166 , for @xmath167 . so we obtain @xmath168 and @xmath169 . then @xmath57 , @xmath58 and @xmath59 are compatible with the fully separable state @xmath170 . this completes the proof . now we are ready to give the proof of the following . [ thm : classical ] any three compatible classical - classical two - qubit states @xmath57 , @xmath59 and @xmath58 are compatible with a tripartite biseparable state . suppose some fully classical bipartite marginals @xmath57 , @xmath59 and @xmath58 are only compatible with genuinely entangled states @xmath1 . then the one - party reduced density operators @xmath171 have to be the maximally mixed states , @xmath172 . indeed , if , without loss of generality in the argument , @xmath173 is non - degenerate , the two - party reduced states would also be compatible with a global state given by the a locally ( on @xmath114 ) dephased version of @xmath1 , which would be separable in @xmath24 , leading to a contradiction . thus , up to local unitaries , we may assume @xmath174 where @xmath175 and @xmath176 . since @xmath177 , we have @xmath178 . so we obtain @xmath179 $ ] . since @xmath1 is genuinely entangled , the cases @xmath180 and @xmath181 are excluded by lemma [ le : nongenuine].(i ) and [ le : nongenuine].(ii ) , respectively . so we obtain @xmath57 as eq . . by similar arguments and performing suitable diagonal local unitary gates on systems @xmath182 , the classical - classical two - qubit states @xmath59 and @xmath58 can be simplified to the forms eqs . and , respectively . meantime , @xmath57 is unchanged . since @xmath57 , @xmath59 and @xmath58 are compatible , it follows from lemma [ le : classicalmain].(ii ) that they are compatible with a biseparable state . it gives us a contradiction . so there are no compatible bipartite marginals that are only compatible with genuine entangled states . this completes the proof . we have seen that the condition of classicality of marginals is strong enough to exclude the need for global genuine multipartite entanglement . we will now construct non - classical separable marginals that are only compatible with global genuine multipartite entanglement , but first we need to establish some more definitions and notation . we set @xmath183 , @xmath184 and @xmath185 . we denote @xmath186 , @xmath187 the rank and range of any square matrix @xmath188 , respectively . a quantum state is a positive semidefinite linear operator @xmath189 with @xmath190 . we say @xmath1 is a @xmath191 state by meaning that the reduced density operators satisfy @xmath192 , @xmath193 , @xmath194 . the ranks of the reduced density operators of @xmath1 are invariant when we perform an invertible local operator ( ilo ) on @xmath1 . that is , let @xmath195 such that @xmath196 . then @xmath197 , @xmath198 and @xmath199 where @xmath200 . we also denote @xmath201 as the vector whose components are the complex conjugate of those of @xmath202 . so @xmath202 is real when @xmath203 . evidently @xmath204 , where @xmath205 , we recall , denotes partial transposition . we call the integer pair @xmath206 the _ birank _ of @xmath111 , and the two integers may be different . for such examples of two - qubit and qubit - qutrit separable states , we refer the readers to ( * ? ? ? * table i , ii ) . furthermore , we say that @xmath111 is a ppt [ npt ] state if @xmath207 [ @xmath208 has at least one negative eigenvalue ] . evidently , a separable state must be ppt . the converse is true only if @xmath209 @xcite . we say a bipartite state @xmath57 is _ @xmath114-finite _ when for any subspace @xmath210 , @xmath211 and any state @xmath212 , it holds that @xmath213 . in other words , @xmath57 is not @xmath114-finite when @xmath214 contains a 2-dimensional subspace spanned by @xmath215 with some linearly independent states @xmath216 . so if @xmath57 is not @xmath114-finite , there must be infinitely many product states in @xmath214 . besides these notions and notation , we will need also the following lemma . [ le : symmetricab ] suppose the bipartite marginals @xmath57 , @xmath59 and @xmath58 are compatible with @xmath1 , and @xmath57 is @xmath114-finite and @xmath117-finite . then @xmath1 is either separable with respect to the partition @xmath217 , or genuinely multipartite entangled . suppose @xmath1 is biseparable , so @xmath218 , with @xmath219 separable in the @xmath24 partition ( similarly for @xmath220 and @xmath221 . we argue that @xmath222 is fully separable , and a similar argument will apply to @xmath223 . let @xmath224 . since @xmath57 is @xmath117-finite and @xmath225 , any @xmath226 must be a product state . so @xmath222 is fully separable . similarly one can show that @xmath223 is also fully separable , so @xmath1 is separable with respect to @xmath217 . we are now in the position to prove the following . [ thm : main ] suppose the triple @xmath227 is compatible . then , @xmath228 if all the following conditions are met : ( i ) for any @xmath229 , the state @xmath230 is @xmath231-finite and @xmath232-finite ; ( ii ) @xmath59 has birank @xmath233 ; ( iii ) @xmath57 has birank @xmath234 , @xmath235 . suppose @xmath57 , @xmath59 and @xmath58 are compatible with a biseparable state @xmath1 . by hypothesis ( i ) and lemma [ le : symmetricab ] , we obtain that @xmath1 is separable with respect to the partition @xmath217 . let @xmath236 where @xmath8 , the @xmath237 on @xmath13 are pairwise linearly independent . note that the product subspace @xmath238 , @xmath239 . this fact and ( i ) imply that @xmath240 , @xmath239 . by similar arguments we have @xmath241 . we may assume @xmath242 . hence @xmath243 . by ( ii ) we have @xmath244 where @xmath245 . without loss of generality , we may assume that the states @xmath246 , @xmath247 span @xmath248 . we choose a suitable ilo @xmath249 such that @xmath250 , @xmath251 and @xmath252 is real . by performing @xmath249 on the state @xmath1 , we have @xmath253 and @xmath254 for any @xmath231 . since the operation @xmath249 does not change the rank of quantum states , it follows from ( ii ) that @xmath255 has birank @xmath256 . recall that the @xmath237 are pairwise linearly independent . since ( i ) is not changed under ilos , it follows from ( i ) that @xmath252 is not parallel to any state @xmath257otherwise @xmath258 would not be @xmath117-finite . since @xmath252 is real , the two @xmath259-dimensional subspaces @xmath260 and @xmath261 are equal and spanned by @xmath262 and @xmath263 . so the states @xmath264 , for any @xmath265 . then ( i ) implies that these @xmath266 are real up to an overall phase . so @xmath267 . it implies @xmath268 which is a contradiction with ( iii ) . therefore @xmath57 , @xmath59 and @xmath58 are not compatible to any non - genuinely entangled state . this completes the proof . we will now make use of theorem [ thm : main ] and offer an example of separable marginals only compatible with genuinely entangled tripartite states . [ ex : main ] consider the family of rank-@xmath259 states on @xmath269 given by @xmath270 , with @xmath271 @xmath272 , @xmath273 , and @xmath274 . it is easy to see that the only biseparable pure state in @xmath275 is @xmath276 . the bipartite reduced density operators of @xmath277 are @xmath278 with @xmath279 and @xmath280 . the three two - qubit marginals @xmath281 , @xmath255 and @xmath282 are positive under partial transposition ( ppt ) , so they are separable @xcite . hence , @xmath57 , @xmath59 and @xmath58 are separable too ; they also evidently satisfy condition ( i ) of theorem [ thm : main ] . furthermore , @xmath59 has birank @xmath283 , while @xmath284 , and @xmath57 has birank @xmath285 . so also conditions ( ii ) and ( iii ) of theorem [ thm : main ] are satisfied , and we conclude that @xmath57 , @xmath59 and @xmath58 are only compatible with genuinely tripartite entangled states . the example shows that for any fixed local dimension @xmath286 , there exist triples of two - qudit separable states that are only compatible with genuine multipartite entanglement . the `` core '' of our construction is the genuine multipartite entangled three - qubit state @xmath287 of eq . . it turns out that @xmath287 is actually the only state compatible with its reductions . the proof of this is given in the appendix . it is worth comparing this with the results of @xcite . there it was proven that for almost all pure entangled states of three qubits @xmath288 it holds @xmath289 , with the exception of states of the generalized ghz form @xmath290 ( up to local unitary transformations ) , which satisfy , e.g. , @xmath291 . interestingly , the only three - qubit pure states that have separable reduction are of the generalized ghz form @xcite . this implies that any three - qubit state @xmath129 such that ( i ) its reductions are separable and ( ii ) @xmath292 , must be mixed . since the state @xmath287 has rank two , we can think of it as the simplest possible example that satisfies ( i ) and ( ii ) , with the additional property of being uniquely determined by its reductions . we generalize example [ ex : main ] in several ways , all presented in the appendix . while we showed that there exist genuine multipartite states whose compatibility set contains only genuine multipartite states , it is natural to ask how common this phenomenon is , i.e. , whether such states have finite volume in the set of all states . this is important also from the point of view of the potential realization of such states in the lab , which can never be perfect . we answer this question in the affirmative . we introduce a parameter of compatibility of a tripartite state @xmath30 with @xmath293 as @xmath294 , where we have used the hilbert - schmidt norm @xmath295 in its arguments and the fact that @xmath296 is positive and vanishes if an only if @xmath297 . ] . we further define @xmath298 . we have @xmath299 for any triple @xmath300 such that @xmath301 , even if the triple of reduced states is compatible , as in example [ ex : main ] . finally , given a tripartite state @xmath287 , we define @xmath302 and @xmath303 . now , consider a genuinely entangled multipartite state @xmath304 with separable reductions such that @xmath305 , and the convex combination of @xmath304 with an arbitrary fully separable state @xmath306 : @xmath307 for @xmath308 . since the set of biseparable states is closed , there exists @xmath309 such that @xmath310 is genuine multipartite entangled for all @xmath306 and all @xmath311 . since @xmath306 is fully separable , so are the two - party reduced states of @xmath312 . furthermore , since @xmath313 is continuous , there exists @xmath314 such that @xmath315 for all @xmath306 and all @xmath316 . for any local finite dimensions , the set of fully separable states has non - zero volume among all states , because there exists a ball of fully separable states around the maximally mixed state @xcite . thus , the argument above proves that also the set of tripartite states whose two - party marginals are separable but only compatible with genuine multipartite entanglement has non - zero volume . we analyzed the relation between the character of correlations of tripartite states and the ones exhibited by their bipartite reductions , i.e. , a version of the quantum marginal problem that focuses on the compatibility of bipartite reductions with certain global properties . we constructed examples where separable reductions are only compatible with genuine multipartite entanglement . up to our knowledge , this is the `` largest '' known separation between the character of correlations of bipartite reductions and what can be inferred about the quality of correlations of the global state , based only on the knowledge of the reductions . on the other hand , at least for qubits we were able to prove that compatible reductions that are fully classical can always originate from a biseparable global state . nonetheless , bipartite reductions that are fully classical may still require the presence of some entanglement in the global state . our results show that the relation between global and `` local '' correlations is far from trivial . notably , the notion of fully classical correlations is strong enough to `` break '' the need for genuine multipartite entanglement , but not the potential need for global entanglement altogether . an interesting open question is whether compatible classical - classical marginals in high dimension are always compatible with a biseparable tripartite state . another question is how to quantitatively bound the certifiable genuine multipartite entanglement in terms of the non - classicality of the two - body reductions , at least in the three - qubit case . the latter problem is reminiscent of the case of entanglement distribution , where the non - classicality of correlations rather than the entanglement present between a quantum carrier and distant labs constitutes a bound on the entanglement that can be generated between the labs by exchanging the carrier @xcite . we are grateful to a. acn and o. ghne for useful discussions . lc started this work when at the institute for quantum computing & department of pure mathematics , university of waterloo , and was mainly supported by mitacs and nserc ; lc s research was funded in part by the singapore national research foundation under nrf award nrf - nrff2013 - 01 . og is grateful for the support of the austrian science fund ( fwf ) and marie curie actions ( erwin schrdinger stipendium j3312-n27 ) . km was supported by the john templeton foundation , national research foundation and the ministry of education ( singapore ) during the completion of this work . mp acknowledges support from nserc , cifar , and ontario centres of excellence . we first prove that @xmath287 in eq . in the paper is the only state compatible with its reductions , a fact of interest in its own . [ le : entunique ] for @xmath287 in eq . it holds @xmath317 . suppose @xmath318 has the same reductions as @xmath287 , i.e. , @xmath319 . we can always write its spectral decomposition as @xmath320 , where @xmath321 , and @xmath322 , @xmath239 . we have @xmath323 . it follows from eq . ( 5 ) in the paper that @xmath324 . so any four states of @xmath325 are linearly dependent . using the freedom in the choice of the pure - state ensemble representation of a mixed states @xcite , we can choose a suitable linear combination of @xmath226 , @xmath326 , such that it is equal to @xmath327 . so the state can be written as @xmath328 where @xmath329 . by applying this procedure to another four states @xmath330 with @xmath331 respectively , we can realize @xmath332 . by relabeling the states , we can write @xmath333 with @xmath334 on @xmath335 . we have @xmath336 and @xmath337 by eq . . since @xmath281 is x - finite for @xmath338 , we have @xmath339 . by the similar argument we can show @xmath340 . so we have @xmath341 . let @xmath342 where @xmath343 and @xmath344 . by eq . we have @xmath345 , @xmath346 . by eq . , we have @xmath347 and @xmath348 . these equations imply for @xmath343 , we have @xmath349 the coefficients of @xmath350 in both @xmath281 and @xmath282 are @xmath351 , so @xmath352 . by replacing @xmath353 by a suitable linear combination of them , we may assume @xmath354 . so the tripartite state can be rewritten as @xmath355 , where @xmath356 since @xmath324 , we have @xmath357 . by eq . , we have @xmath358 . so @xmath359 . by eq . again , we have @xmath360 , @xmath361 , and @xmath362 . now we see @xmath363 in eq . . this completes the proof . we further derive ( and later use in appendix [ sec : appgen ] ) the following lemma . [ le : unique ] suppose @xmath57 , @xmath59 , @xmath58 are only compatible with a tripartite state @xmath1 , and @xmath281 , @xmath255 , @xmath282 are compatible with another tripartite state @xmath277 . if @xmath364 , then @xmath277 is the only state with which @xmath281 , @xmath255 , @xmath282 are compatible . suppose @xmath365 are compatible with another state @xmath366 . since @xmath367 , we may find a small enough @xmath368 and a tripartite state @xmath369 , such that @xmath370 . so the bipartite reductions @xmath371 are compatible with the state @xmath372 , which is different from @xmath1 . it gives us a contradiction . we conclude this section presenting separable marginals that are only compatible with a unique quantum correlated ( unentangled ) state . [ pp : sepunique ] the separable states @xmath373 , where @xmath374 , are only compatible with the separable state @xmath375 . we will use the following observation in the proof and it is easy to verify . for any @xmath376 , there are only two product states @xmath377 , they also span the space @xmath378 . that is , any state in @xmath378 is the linear combination of @xmath379 and @xmath380 . it is clear that @xmath381 are compatible with @xmath1 . suppose they are compatible with another three - qubit state @xmath382 . by applying the observation to system @xmath383 we have @xmath384 with some complex numbers @xmath385 . by applying the observation to system @xmath182 we have @xmath386 , and hence @xmath387 . so we may assume @xmath388 . as a result , the range of the state @xmath277 is spanned by the product states @xmath389 . by simple algebra one can see that the only feasible @xmath277 compatible to @xmath390 and @xmath58 is the convex sum of @xmath391 and @xmath392 . by using the condition @xmath393 we obtain @xmath394 . this completes the proof . we provide here some further examples of states with separable reductions that are only compatible with genuine multipartite entanglement , also making use of proposition [ le : entunique ] and lemma [ le : unique ] . note that @xmath395 for @xmath287 in eq . . it follows from lemma [ le : unique ] that for any @xmath396 , the separable states @xmath397 , @xmath398 , and @xmath399 are uniquely compatible with the state @xmath400 . so we have generated a family of separable bipartite marginals which are uniquely compatible with a genuinely entangled state , extending example [ ex : main ] . we now generalize example [ ex : main ] to a different family of states that satisfy the conditions in theorem [ thm : main ] . let @xmath277 be as in eq . , and the product state @xmath401 . such product state always exists because @xmath402 , and there is a product state in any 2-dimensional two - qubit subspace . for example , we can choose @xmath403 . we have the following corollary now . [ cr : mixture ] let @xmath404 , @xmath405 , @xmath273 , @xmath406 . for @xmath407 suppose the product states @xmath408 where @xmath409 is real and @xmath255 is the reduced density operator of the state @xmath277 in eq . . the three reduced density operators of the three - qubit state @xmath410 are only compatible with genuinely entangled states . it is sufficient to show that the three reduced density operators @xmath411 , @xmath412 , and @xmath413 satisfy the three conditions ( i ) , ( ii ) , and ( iii ) in theorem [ thm : main ] . recall that @xmath277 satisfies these conditions . since @xmath414 and @xmath415 , we have @xmath416 for any @xmath376 . so condition ( i ) is satisfied . next , the same argument shows that @xmath411 has birank @xmath417 , which is exactly condition ( iii ) . third , the hypothesis @xmath418 and @xmath409 is real imply that the birank of @xmath413 is @xmath419 . so condition ( ii ) is also satisfied . k. binder and a. p. young , rev . phys . * 58 * , 801 ( 1986 ) ; h. t. diep ( ed . ) , _ frustrated spin systems _ ( world scientific , singapore , 2005 ) ; c. m. dawson and m. a. nielsen , phys . a * 69 * , 052316 ( 2004 )

the determination of scattering amplitudes between pairs of mesons is an interesting problem in strong qcd . it is also a complicated problem , because both @xmath14 annihilation to @xmath15-channel resonances and `` nonresonant '' scattering are important effects , and it is often difficult to separate the various contributions . however by specializing to annihilation - free channels such as i=2 @xmath0 and @xmath16 , i=3/2 k@xmath17 , kn and nn , one may study nonresonant scattering in relative isolation . the determination of resonance parameters , reaction mechanisms , and many other aspects of hadron physics are complicated by the presence of nonresonant scattering , which is treated as an ( often poorly understood ) initial - state and final - state rescattering effect . developing an accurate description of nonresonant scattering would help clarify many other aspects of hadron physics . a further interesting possibility is that sufficiently attractive nonresonant scattering may lead to weakly bound hadron - hadron or multihadron states , as does happen in nuclei and hypernuclei . we may also find a rich spectrum of meson - meson bound states , the study of which will extend nuclear physics into the largely unexplored field of mesonic nuclei " or molecules " @xcite . an understanding of psps , psv and other meson - meson scattering amplitudes is also important for the interpretation of non - qcd processes such as nonleptonic weak decays , since these show evidence of important hadronic final state interactions . the @xmath18i=1/2 rule is a well known example . similarly , a recent study of d and d@xmath19 decays to k@xmath20 @xcite found that the dalitz plots are dominated by two - meson isobars , including @xmath21 , k@xmath22 and k@xmath23 , and complex relative amplitudes are required to describe the d@xmath24 dalitz plot . without final state interactions one would expect relatively real couplings to these final states . one finds a surprising variety of approaches to strong hadron - hadron scattering in the literature . there are many studies using effective hadronic lagrangians , such as the chiral perturbation theory " description of the psps sector . although this method is convenient because it uses perturbative qft techniques , it is incomplete in that it takes effective lagrangian vertex strengths from the data ; one should be able to calculate these hadronic couplings directly from quark - gluon forces . second , there are studies that model the low energy hadron - hadron scattering mechanism , which include the apparently dissimilar meson exchange and quark - gluon descriptions of hadronic forces . meson exchange models are again attractive for their simplicity , since they use perturbative qft techniques to determine scattering amplitudes . this approach has been elaborated in greatest detail in models of the nn force @xcite , in which a large number of meson exchanges is assumed . with this large parameter space a good description of this interaction is possible , although there is a concern that one may be parametrizing other scattering mechanisms in addition to @xmath25-channel meson exchange . alternatively , one may calculate hadron - hadron forces directly from the fundamental quark - gluon interaction , using quark model hadron wavefunctions . this approach has also seen its most detailed development in studies of the nn interaction @xcite , and is most successful in describing the short - ranged repulsive core . maltman and isgur @xcite also found a physically reasonable intermediate ranged attraction from a color van der waals effect in the quark - gluon approach , which is _ not _ equivalent to the usual @xmath0 or @xmath26 meson exchange explanation of this force . the quark description of hadron - hadron interactions is complicated by the combinatorics of matrix elements between quark bound states , but has the advantage that it can easily be extended to a wide range of spin and flavor channels through a simple change of the external hadron wavefunctions . a third promising approach is to infer hadron scattering amplitudes from lgt . to date lgt has seen little application to scattering problems because of the difficulty of treating systems that are not in their ground states . estimates of the i=0 and i=2 @xmath0 scattering lengths have been obtained by exploiting a theoretical relation to finite - size effects @xcite , and more recently very interesting results for nuclear physics potentials in the @xmath27 system were reported @xcite . in future it may be possible to improve hadron scattering models through comparisons with similar `` lgt data '' . in this paper we are concerned with the derivation of meson - meson scattering amplitudes from quark - gluon forces . we derive meson - meson scattering amplitudes at lowest order in the quark - quark interaction , which leads to a quark interchange model described by `` quark born diagrams '' @xcite . since the quark - quark interaction is already well established from hadron spectroscopy , our predictions have little parameter freedom . in previous work we and others ( usually assuming oge hyperfine dominance ) have shown that this approach gives a reasonably accurate description of s - wave scattering in a wide range of channels without @xmath14 annihilation , including i=2 @xmath0 @xcite , i=3/2 k@xmath17@xcite , i=0,1 kn@xcite , i=0,1 bb@xcite ( compared to lgt data ) , and the nn repulsive core @xcite . this approach has also been applied to @xmath28 @xcite and other reactions relevant to heavy ion collisions , where the experimental low energy cross sections are as yet unclear . the principal new contribution of this paper is a detailed analytical derivation of the meson - meson scattering amplitudes that follow from the complete quark - quark interaction , including color coulomb , linear scalar confinement , oge spin - spin , oge spin - orbit , oge tensor and linear spin - orbit forces . as a future application of these results , one might hope to clarify the relationship between meson exchange and quark interchange models by a detailed comparison of the spin dependence of hadron - hadron scattering amplitudes , which we expect to be sensitive to the details of the scattering mechanism . here we consider both pseudoscalar - pseudoscalar ( psps ) and pseudoscalar - vector ( psv ) scattering . the former is a `` standard benchmark '' for meson scattering models , because i=2 @xmath0 low energy scattering has no @xmath15-channel resonances and has been the subject of many experimental phase shift analyses . although we find reasonable agreement with s - wave i=2 @xmath0 scattering , this channel has no spin degree of freedom , and so can not be used to test the characteristic spin dependences predicted by the quark model s oge and linear scalar confinement forces . we find in contrast that psv is an excellent theoretical laboratory for the study of spin - dependent forces , as it can accommodate both meson - meson spin - orbit and tensor interactions . the spin - dependent forces at the meson - meson level are closely related to the corresponding terms in the quark - quark interaction in our approach . although the study of psv scattering is essentially a theoretical exercise at present , these phase shifts are accessible experimentally , for example through measurement of the relative s and d final state phases in @xmath29 . thus it should be possible to measure psv phase shifts from resonance decays to multiamplitude psv final states . before we proceed to our detailed results , we note that some work has already appeared on meson - meson scattering in psv systems . numerical results for many light s - wave psv meson channels were previously reported by swanson @xcite using a similar quark model approach that incorporated oge spin - spin and linear confinement forces . theoretical results for psv scattering ( @xmath16 in particular ) in a meson exchange model were published by janssen _ _ @xcite and bckmann _ et al . _ @xcite , assuming @xmath17 , vector and @xmath30 exchange . since the @xmath31 , @xmath32 and @xmath33 vertex strengths are relatively well established , it was possible to evaluate these scattering amplitudes numerically . these papers did not consider the exotic i=2 channel , so a direct comparison with our quark model psv results is not possible at present . we approximate the full hadron - hadron scattering amplitude by a single ( born - order ) matrix element of the quark - quark interaction hamiltonian @xmath34 . since @xmath34 is @xmath35 in color , one must then have quark line rearrangement to have a nonvanishing overlap with two color - singlet mesons in the final state . in @xmath36 scattering there are four independent born - order diagrams , which we label according to which pair of constituents interacted ; these are transfer@xmath37 " ( t1 ) , transfer@xmath38 " ( t2 ) , capture@xmath37 " ( c1 ) and capture@xmath38 " ( c2 ) , which are shown in fig.1 . in the special case of identical quarks _ and _ identical antiquarks , which is relevant here , there is a second set of four symmetrizing " diagrams t1@xmath39 c2@xmath40 , which are identical to t1@xmath41c2 except that the quark lines are interchanged rather than the antiquark lines . the hadron - hadron t - matrix element @xmath42 for each diagram can conveniently be written as an overlap integral of the meson wavefunctions times the underlying quark @xmath42 . these overlap integrals ( specializing ref.@xcite to the case of equal quark and antiquark masses ) are @xmath43 the quark @xmath42 has momentum arguments @xmath44 , which are defined in fig.2 . in this paper we will evaluate these overlap integrals with standard gaussian quark model wavefunctions and the quark @xmath42 for the complete set of oge color coulomb , linear scalar , oge spin - spin , oge and linear scalar confinement spin - orbit and oge tensor interactions . these interactions are given in app.a . we specialize the general problem of psps scattering without @xmath14 annihilation to i=2 @xmath0 because many experiments have published phase shift analyses of this channel . the other @xmath0 channels have large @xmath15-channel @xmath14 annihilation contributions . the full i=2 @xmath0 born - order t - matrix element is determined by adding the individual contributions of app.b , with psps spin matrix elements given in app.c , part c2 . there are also flavor and color factors for each diagram and an overall signature " phase of @xmath45 , and a second set of symmetrizing " diagrams for identical quarks and identical antiquarks , as discussed in detail in ref.@xcite . on summing these contributions we find @xmath46 where @xmath47 is an abbreviation for the confluent hypergeometric ( kummer ) function @xmath48 . the three separate expressions above are the oge spin - spin , color coulomb and linear confinement contributions respectively . the @xmath49 terms come from the transfer diagrams , and the remaining , isotropic , terms come from the capture diagrams . the spin matrix elements of the spin - orbit and tensor terms vanish identically in the psps channel . since @xmath50 and @xmath51 , one can equivalently write this amplitude as a function of the cm momentum and scattering angle using @xmath52 , where @xmath53 . the bose symmetry required for this @xmath0 scattering amplitude is evident . we may derive the elastic born - order i=2 @xmath0 phase shifts from eq.([pipit ] ) , using the relation between phase shifts and the t - matrix given in app.d , especially eq.(d17 ) , and the integrals in app.g . the result we find for the s - wave is @xmath54 where we have introduced @xmath55 . the total born - order s - wave phase shift is the sum of these three contributions . this s - wave phase shift is shown in fig.3 with our standard quark model parameter set @xmath56 , @xmath57 gev , @xmath58 gev and @xmath59 gev@xmath60 . we also use @xmath61 gev throughout . this confirms that the color coulomb and linear confinement interactions make relatively small contributions to the i=2 @xmath0 s - wave at moderate energies . the weakly repulsive linear confining interaction in i=2 @xmath0 near threshold was previously found numerically by swanson@xcite . one might be concerned about the approximation of using sho wavefunctions , especially at higher energy scales where there should be strong short - distance components in the pion wavefunction due to the attractive spin - spin hyperfine interaction . to test the sensitivity to sho wavefunctions we evaluated the i=2 @xmath0 scattering amplitudes and phase shifts numerically using coulomb plus linear plus hyperfine @xmath14 wavefunctions and monte carlo integration of the real space integrals corresponding to the t - matrix integrals ( [ t1def]-[c2def ] ) . as usual this requires a `` smearing '' of the contact hyperfine term , @xmath62 , to allow solution of the schrdinger equation with an attractive delta - function interaction . in the literature the inverse smearing length is typically taken to be @xmath63 gev . ( a calculation of i=2 @xmath0 scattering with this interaction and @xmath64 gev was reported previously by swanson @xcite . ) with our standard light - quark parameter set @xmath56 , @xmath65 gev@xmath60 and @xmath66 gev , we found that fitting the @xmath67 splitting required a value of @xmath68 gev . to illustrate the dependence of the scattering amplitude on this parameter , in fig.4 we show the i=2 @xmath0 s - wave that follows from our standard quark model set @xmath69 , with @xmath70 and @xmath71 gev . clearly the predicted phase shift is rather similar to the sho result of fig.3 , although the effect of short - distance peaking in the @xmath17 wavefunction is evident above @xmath72 gev . we also show most of the higher statistics experimental results for the i=2 @xmath0 s - wave phase shift in figs.3 and 4 . the references shown are colton _ et al._@xcite , durusoy _ et al._@xcite ( ope extrapolation , solid ; ope + dp form factor , open , slightly displaced in @xmath73 for visibiity ) , hoogland _ et al._@xcite ( extrapolation b ) , and losty _ et al._@xcite . et al._@xcite found a wide range of results from three different off - shell extrapolations , so we simply quote their fitted scattering length below . clearly there is already reasonable agreement with the experimental s - wave phase shift at lower energies without fitting the quark model parameters . the model predicts a rather dramatic extremum in this phase shift near @xmath74 gev , which is unfortunately beyond the limiting invariant mass of most of the experimental studies . there are some measurements of this phase shift at higher invariant mass with lower accuracy due to durusoy _ et al._@xcite , which are also shown in the figure . the durusoy _ _ results support our predicted extremum near @xmath74 gev ; indeed , their phase shift above @xmath74 gev appears to fall even more rapidly than we predict . we have investigated optimal parameter fits of the s - wave phase shift formula eq.(6 ) to the data , but we find that these are rather unstable because the color coulomb and linear confinement contributions are small and are qualitatively similar functions . in any case the durusoy _ data and fig.4 show that the hyperfine smearing distance @xmath26 is an important parameter , and this will not be well determined until accurate phase shift data becomes available at higher invariant mass . an accurate measurement of i=2 @xmath0 scattering amplitudes near and above @xmath74 gev would clearly be very useful as a test of this and other models of meson - meson scattering . the i=2 @xmath0 scattering length is defined by @xmath75 . the results we find from eq.(6 ) are @xmath76 and their numerical values with our standard quark model parameters set are @xmath77 } & { $ s\cdot s$ } \cr -0.007 \ { \rm [ fm ] } & { cou . } \cr -0.017 \ { \rm [ fm ] } & { lin . } \cr -0.109 \ { \rm [ fm ] } & { \rm total}. } \ ] ] the coulomb and linear contributions were independently checked by monte carlo integration of the corresponding real - space overlap integrals . the relative sizes of these numerical contributions _ a posteriori _ justify the approximation of neglecting the color coulomb and linear terms in i=2 @xmath0 scattering . the i=2 @xmath0 scattering length has been calculated previously using many other theoretical approaches . a summary of some of these predictions is given below . ( we use a current value of @xmath78 mev in weinberg s pcac formula @xmath79 . ) @xmath80 } & { lgt \cite{lgt_pipi } } \cr -0.052 \ { \rm [ fm ] } & { meson exchange \cite{mesex , fs } } \cr -0.053 \ { \rm [ fm ] } & { roy eqs.\cite{ana } } \cr -0.063 \ { \rm [ fm ] } & { pcac \cite{wein}}. } \ ] ] although @xmath81 and other effective lagrangian parameters are normally taken from experiment , these parameters are of course calculable from quark - gluon forces . as an example , our result for @xmath82 yields the following expression for @xmath81 , @xmath83 the dominant contribution comes from the @xmath84 oge @xmath85 term . experimental determinations of the scattering length have yielded results which are larger than theoretical expectations ; @xmath86 } & { losty { \it et al.}\cite{losty } } \cr -0.24(2 ) , -0.22{+0.03\atop -0.04 } \ { \rm [ fm ] } & { prukop { \it et al . } \cite{prukop } } \ . } \ ] ] we speculate that this discrepancy is due to the use of a simple @xmath87 effective range formula in extrapolation . the difficulty of extrapolating experimental phase shifts to threshold has been stressed by morgan and pennington @xcite . we advocate the use of a generalized specific heat plot " of low energy phase shifts for this purpose @xcite . this plot takes into account the threshold behavior seen in eq.(6 ) , @xmath88 where @xmath89 is a relatively slowly varying function of @xmath90 . thus the threshold behavior is approximately proportional to @xmath91 rather than just @xmath92 , and since @xmath93 is quite small this leads to rapid variation near threshold and makes the linear-@xmath92 extrapolation inaccurate . we suggest removal of all this dependence by displaying @xmath94 versus @xmath1 . the intercept in this plot is the scattering length , and the slope at intercept implies the effective range . this generalized specific heat plot is shown in fig.5 for i=2 @xmath0 scattering . an extrapolation of the moderate-@xmath92 data can now be seen to be much closer to the theoretical scattering lengths . the small-@xmath92 dependence of @xmath95 was calculated by donoghue @xcite in a chiral effective lagrangian , which gave the weinberg result at @xmath96 and an @xmath97 correction factor of @xmath98 . this is just the correction due to an overall factor of @xmath99 , so this predicts a zero slope in @xmath1 for @xmath94 at threshold . the jlich meson - exchange model @xcite , which is dominated by @xmath25-channel @xmath100 exchange in this process , also predicts rapid variation in @xmath95 near threshold . the prediction of this model for @xmath101 @xcite , shown in fig.5 , is rather similar to our quark model result . low energy `` phase shift equivalent '' gaussian i=2 @xmath0 potentials , derived using the method of mott and massey @xcite as described in app.e , are given below . we quote separate gaussians for the transfer and capture contributions from each of the three interactions , spin - spin contact , color coulomb and linear hyperfine . however their predicted phase shift decays more slowly at large momentum , probably due to the use of power law form factors in their vertices . @xmath102 in ref.@xcite we derived i=2 @xmath0 potentials for the spin - spin contact interaction using the locality expansion " method of ref.@xcite . this gave an identical result for the spin - spin transfer potential , because this amplitude ( before bose symmetrization ) is a function of @xmath25 only . however for the capture diagrams the mott - massey approach used here gives a different potential , since it is constrained to reproduce the @xmath103 series expansion of the phase shift in eq.(6 ) , but the local approximation is not . the two capture potentials reproduce the scattering length , but the local approximation gives an incorrect effective range . the low energy mott - massey i=2 @xmath0 potential is shown in fig.6 for our standard quark model parameters @xmath56 , @xmath65 gev@xmath60 and @xmath66 gev . the spin - spin hyperfine contribution is dominant over the range shown . the higher partial waves ( l @xmath105 ) may be evaluated similarly . according to eq.([pipit ] ) , these receive contributions only from the transfer diagrams . the born - order d - wave phase shift with sho wavefunctions is given by @xmath106 these three expressions are numerically rather small , and their phases are such that they approximately cancel ; at @xmath74 gev they are respectively @xmath107 , @xmath108 and @xmath109 . to see this more clearly , the leading @xmath110 behavior predicted by eq.(14 ) is @xmath111 and the three dimensionless combinations @xmath112 , @xmath113 and @xmath114 are comparable in size . we have also evaluated this d - wave phase shift using coulomb plus linear plus hyperfine wavefunctions . the result is shown in fig.4 , and is numerically similar to the sho d - wave , eq.(14 ) . in comparison the experimental d - waves reported by durusoy _ _ @xcite and hoogland _ et al . _ @xcite are @xmath115 at @xmath74 gev ( see fig.4 ) . ( losty _ et al _ @xcite report a rather larger but inconsistent low - energy d - wave . ) this is clearly larger than our prediction , although the rather slow variation of the durusoy _ et al . _ and hoogland _ d - waves with @xmath116 may indicate a problem with the measurements ; the expected threshold behavior of @xmath117 is much more rapid than the observed energy dependence . unfortunately the dispersion relations represented by the roy equations have technical difficulties with determining d- and higher waves @xcite . they do however lead to predictions of a _ positive _ d - wave close to threshold , which is not evident in the data . the d - wave may well have important meson exchange contributions , since this type of model can accommodate the reported experimental phase shift @xcite . for simplicitly we will initially quote results only for i=2 @xmath16 . the other isospin channels are simply related by flavor factors , which we will discuss subsequently . we assume identical spatial wavefunctions , so only the @xmath100 spin degree of freedom and difference in phase space distinguish this case from @xmath0 . summing the individual contributions in app.b with the appropriate flavor and color factors and the @xmath45 signature phase , and using the psv spin matrix elements of app.c part c3 , we find for the i=2 @xmath16 t - matrix @xmath118 @xmath119 @xmath120 @xmath121 @xmath122 \hfill\ ] ] @xmath123 \hfill\ ] ] @xmath124 \hfill\ ] ] @xmath125 + \bigg [ \vec s_\rho \cdot \vec c\;\vec s_\rho \cdot \vec c - { 2\over 3}\vec c^{\ , 2 } \bigg ] \bigg ] \hfill \ .\ ] ] @xmath126 the individual contributions in this result are respectively oge spin - spin ; oge color coulomb ; linear confinement ; oge spin - orbit ; linear scalar confinement spin - orbit ; oge tensor ( transfer diagrams ) ; and oge tensor ( capture diagrams ) . in all these we list transfer followed by capture contributions . @xmath127 is the @xmath100 meson spin vector , @xmath128 and @xmath129 are the initial and final @xmath17 momenta , @xmath130 and @xmath131 are the initial and final @xmath100 momenta , and @xmath132 as in @xmath0 . since this result was derived in the cm frame , @xmath133 . this @xmath42 evidently describes @xmath16 spin - orbit and tensor interactions , in addition to spin - independent scattering . it is interesting that there is a one - to - one mapping between the quark - quark spin - orbit and tensor interactions and these @xmath16 spin - orbit and tensor terms . this simple result need not be true in general ; a given spin - dependent interaction at the quark level may give rise to a different type of hadron - hadron interaction . as an example , a mapping of a tensor nucleon - nucleon force into a nucleon - nucleus spin - orbit interaction was discussed by stancu , brink and flocard.@xcite to evaluate phase shifts and inelasticities it is convenient to calculate the matrix element of our @xmath16 t - matrix eq.([pirhot ] ) between general @xmath134 states , which gives the reduced matrix element @xmath135 @xmath136 as discussed in app.d . this is a straightforward exercise , although integrals of special functions and a careful angular analysis of the spin - orbit and tensor terms are required ; the details are discussed in apps.g and h. this matrix element is diagonal in @xmath137 except for the tensor interaction , which has both diagonal and off - diagonal ( transfer ) and fully off - diagonal ( capture ) contributions . the @xmath137-diagonal results , again showing transfer diagram contributions followed by capture , are @xmath138 and the off - diagonal tensor matrix elements are 0.5 cm @xmath139^{1/2 } \over ( 2j+1)\hskip0.3 cm } \ ; x \bigg ( { \cal f}^{(j-1)}_{{5\over 2},{7\over 2}}(x ) + 2 { \cal f}^{(j)}_{{5\over 2},{7\over 2}}(x ) + { \cal f}^{(j+1)}_{{5\over 2},{7\over 2}}(x ) \bigg)\ ; \ ; e^{-2x } \\ & & \nonumber \\ & & t^j_{l'\ne l}|^{\it capture}_{oge\ t } = { \pi^2 \alpha_s \over m^2 } \ ; \delta_{j1 } \ ; \big ( \delta_{l2 } \delta_{l'0 } + \delta_{l0 } \delta_{l'2 } \big ) \ ; { 2^{19/2 } \over 3^{11/2}\cdot 5}\ ; x\ ; f_{{5\over 2},{7\over 2}}(2x/3 ) \ ; e^{-2x } \ .\end{aligned}\ ] ] 0.5 cm in these formulas @xmath140 is a modified spherical bessel function , the tensor @xmath141 matrix element between @xmath142 @xmath16 states is @xmath143 and the integral @xmath144 is evaluated in app.g . in s - wave to s - wave scattering the spin - orbit and tensor @xmath16 t - matrix contributions vanish , and we are left with color coulomb , linear and spin - spin contributions , just as in the i=2 @xmath0 case . the i=2 @xmath16 s - wave phase shifts that result from these interactions , again using eq.(d17 ) , are @xmath145 where @xmath146 , and again @xmath55 . in fig.6 we show these individual components and the total s - wave phase shift with our standard quark model parameter set and meson masses ( used throughout ) of @xmath147 gev and @xmath148 gev . the forces considered here evidently lead to strong repulsion in the i=2 @xmath16 channel . the spin - orbit and tensor terms in eqs.(21 - 23 ) all contribute to @xmath150 @xmath16 scattering , and there is also an odd-@xmath137 , @xmath151-independent term due to the oge spin - spin interaction in eq.(18 ) , which is not symmetric under @xmath152 . the color coulomb and linear confinement spin - independent terms , eqs.(19,20 ) , contribute only to even @xmath137 . adding the various diagonal matrix elements of eqs.(18 - 23 ) and using eq.(d17 ) gives phase shifts for each @xmath153l@xmath154 partial wave . in fig.8 we show results for all p - wave channels and for j = l@xmath155 in d- and f - wave . note that there is a large , inverted spin - orbit force in the p - wave , so the @xmath153p@xmath156 phase shift is widely separated from @xmath153p@xmath38 , and has an even larger maximum phase shift than the s - wave . the higher - l channels show decreasing phase shifts with increasing l , as expected for short - ranged quark - gluon forces . the relative importance of the individual contributions to the spin - dependent force is of considerable interest . in fig.9 we show the various spin - dependent contributions to the i=2 @xmath153p@xmath38 @xmath16 phase shift . the largest contribution arises from oge spin - orbit , in particular from the transfer diagrams . the oge and confinement spin - orbit capture diagrams give smaller contributions of the same sign . finally , the confinement spin - orbit transfer diagrams have a sign opposite to all these and reduces the total spin - orbit force somewhat . this dominance of the psv spin - orbit by oge is an interesting result , especially since mukhopadhyay and pirner @xcite found the opposite result in kn . in that system they concluded that confinement , not oge , makes the largest contribution to the spin - orbit force . the oge tensor in i=2 @xmath153p@xmath38 @xmath16 is weakly repulsive ; it makes a much larger contribution to @xmath153p@xmath37 and @xmath153p@xmath156 , where the tensor matrix element is respectively @xmath157 and @xmath158 times as large . the oge tensor is evident in fig.8 , in the departure of the ratio ( @xmath153p@xmath159p@xmath37):(@xmath153p@xmath160p@xmath156 ) from the pure spin - orbit value of 2:1 at higher energies . there is also an off - diagonal coupling due to the oge tensor terms , given by eqs.(24,25 ) , but we have neglected this in calculating phase shifts because we find that it is numerically a small effect . the largest coupling at low energies is @xmath161s@xmath162d@xmath37 , which leads to an inelasticity of only @xmath163 by @xmath164 gev ( calculated using eqs.(d18-d20 ) ) . we may determine low energy gaussian equivalent @xmath16 potentials from the phase shifts , as discussed in app.e . the most interesting potential phenomenologically is the spin - orbit one , since the origin of the spin - orbit interaction in the nn system is a long - standing and still poorly understood problem . in particular we derived gaussian potentials corresponding to the p - wave phase shifts due to the oge and linear scalar confinement spin - orbit interactions , using eq.(e4 ) of app.e . the results for the transfer and capture contributions to these potentials are @xmath165 the oge , linear and total spin - orbit potentials for the @xmath153p@xmath38 wave of i=2 @xmath16 are shown in fig.10 for our standard parameter set . the largest contribution to the @xmath16 spin - orbit force comes from oge transfer diagrams ; the linear confinement spin - orbit from the transfer diagrams is about half as large and opposite in sign , and the two capture diagram contributions are much smaller . since the confinement capture and transfer diagrams have opposite signs , the net result is dominance of the psv spin - orbit by the oge contribution . 0.5 cm . j@xmath3 exotic states in psv . [ cols="^,^,^,^,^,^,^,^,^,^ " , ] [ table2 ] inspection of the tables shows that the largest exotic scattering amplitude should be in the i=2 @xmath16 @xmath167 p - wave . the elastic phase shift in this channel is the @xmath153p@xmath156 curve in fig.8 . the large negative phase shift shows that this is a strongly repulsive channel ; the maximum phase shift is predicted to be a quite large @xmath168 at @xmath169 gev , which exceeds even the s - wave phase shift maximum . the largest _ attractive _ exotic phase shift we have found in psv is the i=0 partner , which is @xmath170 of i=2 , giving a maximum phase shift of @xmath171 at the same mass . we do not find sufficient attraction to form a meson - meson `` molecular '' bound state in any of these nonstrange j@xmath3-exotic psv channels . the @xmath172 channels are relatively weak because only the @xmath173 part of the @xmath172 contributes to these diagonal scattering amplitudes ; the @xmath174 component leads to open - strange final states ( k@xmath175 for example ) after quark line interchange . regarding candidate exotic resonances , there have been speculations that the determination of the mass and width of the exotic candidate @xmath4 may have been compromised by inelastic rescattering effects @xcite , analogous to the `` deck effect '' proposed as a nonresonant explanation of the @xmath176 . for example , crossing the @xmath177 threshold at @xmath178 gev in the process @xmath179 might mimic resonant phase motion if this process has a rapidly varying inelasticity . we can test this and other nonresonant possibilities by calculating the elementary @xmath180 scattering amplitudes using our quark model approach . some important results follow from simple flavor factors . note in particular that the nonresonant scattering amplitude @xmath181 vanishes in any i=1 channel , including the @xmath182 exotic one . this is a general result whenever the quark line diagram of fig.1 dominates ; clearly a pair of oppositely charged , nonstrange @xmath14 mesons @xmath183 can not scatter into another charged pair @xmath184 under quark interchange . a comparison with isospin matrix elements shows that this implies that scattering of any two @xmath14 isovectors in i=1 vanishes . this isospin selection rule eliminates two subprocesses discussed by donnachie and page @xcite as deck effect backgrounds that might shift a higher - mass exotic resonance to an apparent @xmath4 , @xmath185 and @xmath186 . independent of any scattering model , one should note that the coupling @xmath187 is probably small because of the strong ves experimental limit ( reported by v.dorofeev @xcite ) of @xmath188 since @xmath189 is a large mode ( @xmath190 @xcite ) , if @xmath187 rescattering were important we would also expect to observe a large @xmath191 branching fraction . we also expect the final background process suggested by donnachie and page ( @xmath192 ) to be small , because the direct time ordering @xmath193 vanishes in p - wave in this model . finally , the rescattering process they propose , @xmath194 , does not vanish in the quark interchange model , although the required @xmath195 and suppressed @xmath172 flavor factor may nonetheless make this a relatively weak amplitude . a calculation of this and related scattering amplitudes is planned for a future publication . although there is little experimental information about psv interactions at present , these phase shifts actually are experimentally accessible in existing data , for example as relative fsi phases in the d and s amplitudes in @xmath196 . these are usually , and incorrectly , taken to be relatively real amplitudes . the relative phase including the fsi is @xmath197 @xcite , and is observable for example as a reduction in the strength of the sd cross term in the @xmath198 angular distribution by @xmath199 . since this method requires individual measurements of the s@xmath60 , d@xmath60 and sd cross term in the angular distribution , it should be applicable to cases such as @xmath200 and @xmath201 where s and d are of comparable magnitude @xcite . the @xmath202 phases we predict at these masses ( which are calculated from + 1/2 times the @xmath153s@xmath37 and @xmath153d@xmath37 i=2 @xmath16 phases in figs.7,8 ) are @xmath203 and @xmath204 . this proposed technique is similar to that used in k@xmath205 decays @xcite , in which the low energy i=0 @xmath0 s - wave phase shift is actually observed as the difference between the i=0 s - wave and i=1 p - wave @xmath0 fsi phases . in this paper we have derived meson - meson scattering amplitudes , including spin - dependent forces , from a calculation of the born - order matrix element of the quark - quark interaction between two - meson states . since @xmath14 annihilation is not included in the model , it describes scattering that does not involve coupling to @xmath15-channel resonances . this includes for example i=2 and the nonresonant backgrounds in all channels , including exotic j@xmath3 . we considered the cases of psps and psv scattering , and derived the scattering amplitudes in all @xmath206 channels for these cases . the parameters of the model were previously fixed by quark model studies of hadron spectroscopy . where possible we have compared the results to experiment . in i=2 @xmath0 ( the best studied psps case ) the results were shown to be in reasonable agreement with experiment in s - wave scattering , and an extremum predicted near @xmath74 gev is supported by the data . rapid variation of @xmath95 is predicted near threshold , which may reconcile theoretical expectations of a small scattering length with larger reported experimental values based on extrapolation in @xmath1 . the experimental d - wave , although quite small , is clearly larger than the model predicts . the psv system is a convenient theoretical laboratory for studying spin - dependent forces , since it can accommodate both spin - orbit and tensor interactions , and is simpler than kn or nn . we derived analytical results for these spin - dependent psv interactions ( t - matrices and phase shifts ) given sho wavefunctions and the standard spin - dependent quark model forces . the quark - quark spin - orbit and tensor forces map directly into spin - orbit and tensor psv interactions . we find that the oge spin - orbit force in the psv system is quite large in p - wave , and so is expected to be large in many other hadron - hadron systems as well . there is no psv phase shift data at present . we noted however that psv phase shifts actually can be measured in multiamplitude resonance decays to psv final states , so it should be possible to test theoretical predictions for psv scattering amplitudes in future experimental studies . our predictions for scattering in j@xmath3-exotic channels are of current interest because the reported exotics might be complicated by large and rapidly varying nonresonant inelasticities . one speculation is that the @xmath4 parameters might be strongly affected by the opening of inelastic couplings to the @xmath177 channel . in our model ( and in any @xmath14 constituent interchange model ) several of these nonresonant processes can be rejected as significant complications because of vanishing flavor factors . in future we plan to extend our calculations to other exotic meson - meson channels , such as s+p , to test whether strong attractive interactions are predicted that might support multiquark exotics " such as s+p molecules . we also plan to apply the current approach to the study of spin - dependent interactions in other hadronic systems , including kn , nn and light hadron + charmonium systems . we are grateful to our colleagues d.blaschke , n.cason , h.g.dosch , s.krewald , e.klempt , n.isgur , c.michael , m.r.pennington , b.pick , h .- j.pirner , g.rpke , f.sassen , j.speth , fl.stancu and c .- y.wong for useful discussions of various issues relating to meson - meson scattering . this research was supported in part by the doe division of nuclear physics , at ornl , managed by ut - battelle , llc , for the us department of energy under contract no . de - ac05 - 00or22725 . es acknowledges support from the doe under grant de - fg02 - 96er40944 and doe contract de - ac05 - 84er40150 under which the southeastern universities research association operates the thomas jefferson national accelerator facility . tb acknowledges additional support from the deutsche forschungsgemeinschaft dfg at the university of bonn and the forschungszentrum jlich under contract bo 56/153 - 1 . the various contributions to the quark - quark @xmath42 ( with color factors of @xmath207 removed ) are + @xmath208 & { \rm oge spin - spin}\cr + { 4\pi \alpha_s \over \vec q^{\ , 2 } } & { \rm oge color coulomb}\cr + { 6\pi b \over \vec q^{\ , 4 } } & { \rm linear conft . } \cr + { 4i\pi \alpha_s \over m^2\vec q^{\ , 2}}\ ; \bigg [ \vec s_1\cdot \big ( \vec q\times ( { \vec p_1\over 2}-\vec p_2)\big ) + \vec s_2\cdot \big(\vec q\times ( \vec p_1-{\vec p_2\over 2 } ) \big ) \bigg ] & { \rm oge spin - orbit}\cr - { 3i\pi b \over m^2\vec q^{\ , 4}}\ ; \bigg [ \vec s_1\cdot ( \vec q\times \vec p_1 ) - \vec s_2\cdot ( \vec q\times \vec p_2 ) \bigg ] & { \rm linear spin - orbit } \cr + { 4\pi \alpha_s \over m^2\vec q^{\ ; 2}}\ ; \bigg [ \vec s_1 \cdot \vec q \ ; \vec s_2 \cdot \vec q -{1\over 3}\ ; \vec q^{\ ; 2}\ , \vec s_1\cdot \vec s_2 \bigg ] & { \rm oge tensor.}\cr } \label{quarkt}\ ] ] the standard @xmath14 quark model gaussian wavefunction is given by @xmath209 where in general @xmath210 and for our special case of equal quark and antiquark masses @xmath211 in this appendix we give the explicit meson - meson t - matrix elements that follow from the overlap integrals eqs.([t1def]-[c2def ] ) with gaussian wavefunctions and the various quark t - matrix elements . the oge spin - spin hyperfine , color coulomb and linear confinement results were derived previously @xcite . for completeness we quote the formulas here , as well as giving the new spin - orbit and tensor results . the multiplicative diagram - dependent color and flavor factors and the signature phase ( which is @xmath45 for these meson - meson scattering diagrams ) are not included in the results given below . these formulas abbreviate the confluent hypergeometric function @xmath48 as @xmath47 , and @xmath212 . these simple contact matrix elements were evaluated previously , for example in ref . @xcite ( in an equivalent form , but incorporating color factors and the signature phase , as eqs.(71 - 73 ) of that reference ) . the results are @xmath213 \\ & & { t_{fi}}^{\rm ( t2 ) } = { t_{fi}}^{\rm ( t1)}(\vec c \rightarrow -\vec c ) \\ \rule{0mm}{1 cm } & & { t_{fi}}^{\rm ( c1 ) } = -\ ; { 2 ^ 6 \over 3^{5/2}}\ ; { \pi\alpha_s \over m^2}\ ; e^{-\vec a^{\ , 2}/3\beta^2 } \bigg [ \vec s_1\cdot \vec s_2 \bigg ] \\ & & { t_{fi}}^{\rm ( c2 ) } = { t_{fi}}^{\rm ( c1 ) } \ . \end{aligned}\ ] ] the contribution of the oge color coulomb interaction to the meson - meson t - matrix follows from the evaluation of the integrals eqs.([t1def]-[c2def ] ) with the second quark - quark @xmath42 in eq.([quarkt ] ) . the results are @xmath214 the linear confinement integrals were carried out in coordinate space , since the fourier transform of the linear potential is singular . the results are @xmath215 one may also obtain these results using the momentum space integrals eqs.([t1def]-[c2def ] ) , but the @xmath216 quark - quark @xmath42 in eq.([quarkt ] ) must include a long - distance regularization in the intermediate stages of the integration . the final result is well defined due to the gaussian damping provided by the wavefunctions . we have checked both the linear and color coulomb @xmath42 results by comparing the expressions eqs.([b5]-[b12 ] ) with monte carlo evaluations of the corresponding real - space overlap integrals . the four oge spin - orbit overlap integrals can be evaluated similarly using the fourth quark - quark @xmath42 in eq.([quarkt ] ) , which gives @xmath217 \\ & & { t_{fi}}^{\rm ( t2 ) } = { t_{fi}}^{\rm ( t1)}(\vec c \rightarrow -\vec c ) \\ \rule{0mm}{1 cm } & & { t_{fi}}^{\rm ( c1 ) } = + { 4 \over 3^{5/2 } } \ ; { \pi\alpha_s\over m^2\beta^2}\ ; { \rm f}_{{3\over 2},{5\over 2}}(\vec a^{\ , 2 } /6\beta^2 ) \ ; e^{-\vec a^{\ , 2}/2\beta^2}\ ; \bigg [ ( \vec s_1 - \vec s_2)\cdot i(\vec a \times \vec c ) \bigg ] \\ & & { t_{fi}}^{\rm ( c2 ) } = -\ ; { t_{fi}}^{\rm ( c1 ) } \ . \end{aligned}\ ] ] the matrix elements eqs.([t1def]-[c2def ] ) , of the scalar confinement spin - orbit interaction in eq.([quarkt ] ) are @xmath218 \\ & & { t_{fi}}^{\rm ( t2 ) } = { t_{fi}}^{\rm ( t1)}(\vec c \rightarrow -\vec c ) \\ \rule{0mm}{1 cm } & & { t_{fi}}^{\rm ( c1 ) } = + { 1 \over 2\cdot 3^{1/2 } } \ ; { \pi b\over m^2\beta^4}\ ; { \rm f}_{{1\over 2},{5\over 2}}(\vec a^{\ , 2 } /6\beta^2 ) \ ; e^{-\vec a^{\ , 2}/2\beta^2}\ ; \bigg [ ( \vec s_1 - \vec s_2)\cdot i ( \vec a \times \vec c ) \bigg ] \\ & & { t_{fi}}^{\rm ( c2 ) } = -\ ; { t_{fi}}^{\rm ( c1 ) } \ . \end{aligned}\ ] ] finally , for the oge tensor @xmath42 ( the last entry in eq.([quarkt ] ) we find @xmath219 \\ & & { t_{fi}}^{\rm ( t2 ) } = { t_{fi}}^{\rm ( t1)}(\vec c \rightarrow -\vec c ) \\ & & { t_{fi}}^{\rm ( c1 ) } = + \ ; { 2 ^ 5 \over 3^{5/2}\cdot 5}\ ; { \pi\alpha_s\over m^2\beta^2}\ ; { \rm f}_{{5\over 2},{7\over 2}}(\vec a^{\ , 2 } /6\beta^2)\ ; e^{-\vec a^{\ , 2}/2\beta^2}\ ; \bigg [ \vec s_1 \cdot \vec a\ ; \vec s_2 \cdot \vec a -{1\over 3}\ , \vec a^{\ , 2}\ ; \vec s_1 \cdot \vec s_2 \bigg ] \label{b23 } \\ & & { t_{fi}}^{\rm ( c2 ) } = { t_{fi}}^{\rm ( c1 ) } \label{b24 } \ . \end{aligned}\ ] ] the tensor matrix elements in the capture diagrams , eqs.([b23],[b24 ] ) , are the only cases in which we have found a post - prior discrepancy in these psps and psv scattering amplitudes ; the post forms of these matrix elements involve a tensor in @xmath129 rather than @xmath128 , @xmath220 \ .\ ] ] these capture tensor terms vanish in the psps channel . they do make a small , off - diagonal contribution to psv scattering , _ albeit _ only in the @xmath161s@xmath221d@xmath37 amplitude . in these scattering amplitude calculations the matrix elements of spin - dependent quark interactions ( the spin - spin , spin - orbit and tensor forces ) involve matrix elements of linear and bilinear quark spin operators . since the quark spins are not directly observed , it is useful to replace them by the spins of the external hadrons . this appendix gives the ( diagram dependent ) mapping from quark spins to hadron spins in the psps and psv cases considered in this paper . the spin matrix elements we require are @xmath222 , @xmath223 , @xmath224 and @xmath225 between general initial and final psps and psv spin states . our convention for the diagrams ( fig.1 ) is that mesons a and c are always ps ( _ e.g. _ @xmath17 ) , and b and d are ps or v ( _ e.g. _ @xmath17 or @xmath100 ) . first , in psps scattering there are no external meson spins , so the quark spin matrix elements are proportional to geometrical tensors such as @xmath226 . the matrix elements by diagram are @xmath227 note that the spin - orbit and tensor terms vanish identically in psps scattering ; this follows from applying eqs.(c2,c3 ) to eq.(a1 ) . in psv scattering the vector ( _ e.g. _ @xmath100 ) meson spin @xmath228 provides an additional degree of freedom , and the linear and quadratic quark spin matrix elements can be expressed in terms of the @xmath100 spin matrix elements @xmath229 and @xmath230 . the mapping of quark to meson spins is @xmath231 since total angular momentum is conserved , the t - matrix is block diagonal in a total angular momentum basis , and can be written as @xmath232 in the special case of spinless scattering these basis states are eigenstates of orbital angular momentum @xmath233 so the t - matrix is given by @xmath234 between momentum eigenstates , @xmath235 are related by @xmath236 and hence @xmath237 where @xmath238 . a more familiar quantum - mechanical result follows from fixing the incident direction @xmath239 in eq.([tp ] ) and integrating over final angles with a @xmath240 weight , which gives @xmath241 since we define @xmath242 by @xmath243 it is related to the lorentz invariant @xmath180 scattering amplitude @xmath244 defined by the pdg@xcite ( their eq.(35.8 ) ) by @xmath245 and hence to the cm scattering amplitude @xmath246 ( their eq.(35.48 ) ) by @xmath247 ( here @xmath248 is the cm momentum of any particle , @xmath249 and the relation between a diagonal partial wave amplitude @xmath250 and the phase shift @xmath251 allow us to determine @xmath252 from @xmath42 . for purely elastic scattering , and assuming small phase shifts so that @xmath253 , eqs.([tp]-[al ] ) give @xmath254 this relation was used previously to determine for example k@xmath17 @xcite and @xmath137-diagonal kn @xcite elastic phase shifts . in the case of elastic scattering of identical bosons , such as i=2 @xmath0 , there is an additional factor of two for identical particles @xcite , so the relation between the born - order phase shift and the t - matrix element becomes @xmath255 since the angular integral in eq.([deltal1 ] ) is proportional to the amplitude @xmath256 , we may also write the born - order elastic phase shift ( for distinguishable particles ) directly in terms of the t - matrix , @xmath257 this formula has a straightforward generalization to the case of external hadrons with spin , which we use to evaluate @xmath16 phase shifts and inelasticities . in the case of an @xmath137- and @xmath15-diagonal interaction this is @xmath258 this is adequate for diagonal forces such as our @xmath16 spin - orbit interactions . the tensor force however is not @xmath137-diagonal , so for this case we must introduce a more general parametrization . since the tensor interaction couples @xmath259 channel pairs which have the same @xmath151 but differ by @xmath260 , it leads to @xmath261 @xmath262-matrices . these can be parametrized as @xmath263 \ .\ ] ] in our calculation , both of the phase shifts and the inelasticity @xmath264 are @xmath265 , so to this order we can relate these linearly to the matrix elements of @xmath34 . the born order phase shift formula ( d17 ) remains valid for both channels , and the inelasticity is @xmath266 the overall phase of @xmath267 is dependent on the state normalizations , but the familiar @xmath268 is unique . potentials provide a very useful representation of hadron - hadron interactions . they have a clear intuitive meaning , and can easily be used in the nonrelativistic schrdinger equation in searches for bound states or in coupled channel problems . unfortunately , one may define hadron - hadron potentials in many different ways . ideally they should reproduce phase shifts or t - matrix elements , at least in the low energy limit . the assumption of a unique , purely local potential is in general overly restrictive , as it leads to a scattering amplitude that is a function of @xmath25 only . in general we find @xmath180 scattering amplitudes that depend on both @xmath15 and @xmath25 . one approach to this problem is to introduce nonlocal gradient " corrections to the potential @xcite , which can be expressed for example as @xmath269 terms ; this approach leads to the familiar breit - fermi hamiltonian for one - photon and one - gluon exchange , and was used to define hadron - hadron potentials in our previous work @xcite . alternatively one may project the scattering amplitude onto a given angular channel @xmath137 so that only @xmath15 dependence remains , and find a local @xmath137-wave potential that describes the scattering in that channel . this definition of potentials was discussed by mott and massey @xcite , and is equivalent to the definition we shall use here . this approach was previously used by swanson @xcite to define meson - meson potentials from scattering amplitudes . the quantum mechanical relation between the phase shift @xmath270 and the radial wavefunction @xmath271 in potential scattering of a mass @xmath272 particle ( which becomes the reduced mass below ) in first born approximation is @xmath273 since our t - matrix elements implicitly determine the elastic scattering phase shifts , for example the @xmath274 @xmath0 s - wave in eq.(6 ) , we can invert this formula for each @xmath137 to determine the corresponding @xmath137-wave local potential @xmath275 . in practice we find that our phase shifts are sufficiently hard " at high energies to require singular potentials ; this is presumably an artifact of our approximations , such as assuming a contact spin - spin interaction . for this reason we do not completely invert the phase shift relation ( e1 ) , and instead simply fit an assumed gaussian form @xmath276 to our theoretical _ low energy _ phase shift . the two gaussian parameters are determined from the @xmath277 and @xmath103 terms in the expansion of the phase shift near threshold , which in the s - wave case are equated to @xmath278 the generalization to higher @xmath137 , also using eqs.(e1 ) , is @xmath279 in determining the gaussian potentials that correspond to our derived phase shifts such as eqs.(6,14,28 ) , we set the external factors of @xmath99 and @xmath280 equal to @xmath281 and @xmath282 before expanding in @xmath248 . this corresponds to using nonrelativistic phase space and nonrelativistic external hadron line normalizations in our t - matrix calculations , which we assume is the appropriate choice for the derivation of a nonrelativistic equivalent potential . the post - prior discrepancy " is a familiar problem in rearrangement collisions ; the diagrams of fig.1 treat the scattering as due to an interaction between the initial hadrons a and b ( prior " ) , but we could equally well have written the interaction between the two final hadrons c and d ( post " ) . the post t - matrices may be obtained from the prior ones by exchanging the initial and final mesons and transforming the momenta @xmath283 , @xmath284 and @xmath285 . thus for example the post c1 t - matrix is @xmath286 @xmath287 one may show that the post and prior results for the scattering amplitude are equal provided that the external wavefunctions are eigenfunctions of the hamiltonian @xcite . swanson @xcite shows an example of convergence of post and prior results for meson - meson scattering amplitudes derived from quark born diagrams as the external wavefunctions approach exact hamiltonian eigenstates . of course the gaussian wavefunctions we use to derive our analytical results are only approximations to the eigenfunctions of the full oge plus linear hamiltonian , so in general we can expect to find a post - prior discrepancy . in this study we actually find such a discrepancy only in part of the tensor interaction in psv scattering , for which we take the mean of the two results , @xmath288 it is useful in the partial wave decomposition of scattering amplitudes to expand functions of the sum and difference momentum transfers @xmath290 ( here @xmath291 ) in spherical harmonics , @xmath292 @xmath293 this expansion may be inverted to determine the coefficient functions @xmath294 , @xmath295 @xmath296 many of the scattering amplitudes derived in this paper are proportional to confluent hypergeometric functions in @xmath297 or @xmath298 , and their partial wave decomposition requires the integral of a legendre polynomial times a shifted confluent hypergeometric function . this integral is given by ( abbreviating @xmath299f@xmath300 as @xmath301 ) @xmath302 where @xmath303 and the pochhammer symbol of negative index is @xmath304 in numerical evaluations it is often useful to transform the confluent hypergeometric function in eq.([teds_excellent_int ] ) to a more rapidly converging negative - argument form , using @xmath305 . as noted in app.d , when evaluating phase shifts and inelasticities it is useful to determine t - matrix elements between @xmath134 states . in the psv system there are spin - orbit and tensor contributions to the t - matrix , and determination of the @xmath206-basis matrix elements of these terms is a complicated problem in angular analysis . here we show how these matrix elements may be evaluated . where @xmath307 . ( additional dependence on the rotational scalar @xmath308 is a trivial modification of this angular decomposition . ) to proceed , we expand @xmath309 in spherical harmonics , as in eq.(g1 ) ; @xmath310 and introduce spherical components for the spin and momentum vectors , @xmath311 @xmath312 and the usual state vector expansion , @xmath313 with these substitutions one may determine the @xmath16 @xmath314 matrix elements ( analogous to the spinless matrix element @xmath315 of eq.(d7 ) ) for the spin - orbit term ( h1 ) . the result involves a sum over a product of six clebsch - gordon coefficients , and can be written as the product of two wigner @xmath316 symbols and two @xmath317 symbols , the constraints of the @xmath316 and @xmath320 symbols force this matrix element to be diagonal in @xmath259 , and imply that the only radial components of @xmath321 in eq.([fexpand ] ) that contribute are @xmath322 . substitution of the explicit @xmath316 and @xmath317 symbols gives our final result for @xmath323 matrix elements of spin - orbit ( h1 ) type , @xmath324 \over 2 ( 2l+1 ) } \ , { \vec a\ , } ^2 \bigg ( f_{l-1 } - \ , f_{l+1 } \bigg ) \ .\ ] ] this result has the overall @xmath325 dependence that one would expect from a spin - orbit force . and @xmath327 \ .\ ] ] both tensor matrix elements have @xmath328 contributions , unlike the other interactions we have considered . the general results in terms of wigner @xmath329 and @xmath330 symbols are @xmath331 @xmath332 and r.machleidt , k.holinde and c.elster , phys . 149 , 1 ( 1987 ) . a recent variant of this approach derives nn scattering amplitudes from meson exchange at the quark level ; see d.bartz and fl.stancu , hep - ph/0006012 . there is an extensive literature on the derivation of the nn interaction from quark - gluon forces in the quark model , beginning with d.a.liberman , phys . d16 , 1542 ( 1977 ) . most studies have employed resonating group methods ( see for example c.s.warke and r.shanker , phys . c21 , 2643 ( 1980 ) ; m.oka and k.yazaki , phys . 90b , 41 ( 1980 ) ; j.e.t.ribeiro , z.phys.c5 , 27 ( 1980 ) ) and variational methods ( for example k.maltman and n.isgur , phys . d29 , 952 ( 1984 ) ) . the born - order quark - gluon approach described here was applied to nn by t.barnes , s.capstick , m.d.kovarik and e.s.swanson , phys . c48 , 539 ( 1993 ) , and in a closely related approach by d.hadjimichef , g.krein , s.szpigel and j.s.daveiga , phys . b367 , 317 ( 1996 ) . m.lscher , comm . 104 , 177 ( 1986 ) ; 105 , 153 ( 1986 ) ; nucl . b354 , 531 ( 1991 ) ; s.r.sharpe , r.gupta and g.kilcup , nucl . b383 , 309 ( 1992 ) ; r.gupta , a.patel and s.r.sharpe , phys . d48 , 388 ( 1993 ) ; m.fukugita _ et al . 71 , 2387 ( 1993 ) ; phys . d52 , 3003 ( 1995 ) ; s.aoki _ et al . _ , hep - lat/9911025 . ( fig.5 shows the scattering length quoted by aoki _ et al._. ) t.barnes and e.s.swanson , phys . d46 , 131 ( 1992 ) ; see also g .- q.zhao , x.g.jing and j-c.su , phys . rev . d58 , 117503 - 1 ( 1998 ) ; d.hadjimichef , g.krein , s.szpigel and j.s.daveiga , phys . b367 , 317 ( 1996 ) ; ann . 268 , 105 ( 1998 ) ; b.masud , j.paton , a.m.green and g.q.liu , nucl . a528 , 477 ( 1991 ) ; b.masud , phys . d50 , 6783 ( 1994 ) . t.barnes , n.black , d.j.dean and e.s.swanson , phys . c60 , 045202 ( 1999 ) . c version has an incorrect figure ( a duplicate of fig.4 was used as fig.5 ) , and minus signs were deleted from the lower @xmath337-axis in both figures . the original , archived version nucl - th/9902068 has the correct figures . d.blaschke and g.rpke , phys . b299 , 332 ( 1993 ) ; k.martins , d.blaschke and e.quack , phys . c51 , 2723 ( 1995 ) ; k.martins , prog . nucl . phys . 36 , 409 ( 1996 ) ; d.blaschke , g.burau and yu.l.kalinovsky , nucl - th/0006071 . for @xmath338 see v.dorofeev ( ves collaboration ) , in proceedings of whs99 , hep - ex/9905002 ; g.s.adams _ et al . ( e852 collaboration ) , phys . 81 , 5760 ( 1998 ) . for @xmath4 see d.r.thompson _ et al . _ ( e852 collaboration ) , phys . 79 , 1630 ( 1997 ) ; a.abele _ et al . _ ( crystal barrel collaboration ) , phys . b423 , 175 ( 1998 ) ; s.u.chung _ et al . _ ( e852 collaboration ) , phys . d60 , 092001 ( 1999 ) . n.isgur , r.kokoski and j.paton , phys . 54 , 869 ( 1985 ) ; t.barnes , f.e.close and e.s.swanson , phys . rev . d52 , 5242 ( 1995 ) ; c.bernard _ et al . _ ( milc collaboration ) , phys . d56 , 7039 ( 1997 ) ; p.lacock _ et al . _ ( ukqcd collaboration ) , phys . b401 , 308 ( 1997 ) .

there is increasing evidence that supermassive black holes ( smbhs ) exist at the center of all galaxies , and that they are the sources which power active galactic nuclei and quasars @xcite . for example , vlbi observations of the keplerian disk around an object in ngc4258 indicate that the central object has a mass @xmath7 . also , large numbers of observations are provided by the hubble space telescope suggesting that smbhs exist in galaxies such as m31 ( @xmath8 ) , m87 ( @xmath9 ) and our own galaxy ( @xmath10 ) ( see for example , * ? ? ? * for a brief overview ) . although evidence of the existence of smbhs is compelling , the actual formation process of these objects is still uncertain @xcite . several different scenarios have been proposed , some based on stellar dynamics , others on gas hydrodynamics , and still others which combine the processes . at present , there is no definitive observation as yet which confirms or rules out any one of these scenarios . here we discuss the collapse of a supermassive star ( sms ) as one scenario of formation of an smbh . this subject is also interesting from a viewpoint of general relativity , since the onset of radial instability at the mass shedding limit takes place at @xmath11 ( @xmath1 is the angular momentum , @xmath2 is the total gravitational energy ) in the uniformly rotating sms @xcite , which is also considered as a critical threshold between formation of a stationary black hole ( bh ) and of a naked singularity . the bh uniqueness theorem says that a complete gravitational collapse of a body always results in a bh rather than a naked singularity , and that the final state of the bh should go into a stationary kerr bh ( e.g. , * ? ? ? however there is a candidate to except this cosmic censor . the collapse of a prolate spheroid with large semimajor axis leads to spindle singularities without an apparent horizon @xcite . therefore , a collapse of a star with the critical value @xmath12 may show us an interesting phenomenon in general relativity . what happens when the star of @xmath13 collapses ? what is the final fate of a collapsing star of @xmath13 ? numerical simulations can clearly show the answers to these questions . the gravitational collapse of a star of @xmath13 has been investigated most of all in axisymmetric spacetime . the pioneering study in this field was made by @xcite . he set up a differentially rotating star , adding radial velocity to induce the collapse . he found that the criterion of formation of a bh apparent horizon is @xmath14 . the following study has been done by @xcite . they set up a uniformly rotating @xmath15 polytropic star and deplete 99 % of the whole pressure to induce the collapse . they found that the criterion of bh formation is @xmath16 , and when the star exceeds the above value , flattened disk formed ( @xmath17 for the case of 99 % pressure deplete ) . note that @xmath18 represents the critical kerr parameter that the collapse proceeds to a bh . finally @xcite performed a collapse of a differentially rotating @xmath15 polytropic star . he also depleted the pressure of the star to induce the collapse . he found that when @xmath19 is less than 0.5 , a bh is formed when the rest mass is larger than the maximum mass of the pressure depleted @xmath1 constant sequence . when @xmath19 is slightly less than @xmath20 , a bh formed when the rest mass is sufficiently larger than the maximum allowed mass on the pressure depleted @xmath1 constant sequence . he also found , by comparing his results from polytropic evolution and from @xmath21-law evolution of the hydrodynamics , that shock heating prevents the prompt collapse to a bh under the condition of @xmath13 . at the same time we have submitted our paper , several groups have also investigated the collapse of a relativistic star of @xmath13 in full general relativity . @xcite investigated the onset of radial instability of uniformly rotating stars in 2d and find that the criterion of @xmath19 is @xmath22 in a very soft polytropic equation of state ( @xmath23 @xmath24 ; @xmath25 is a polytropic index ) . @xcite investigated the collapse of a differentially rotating @xmath15 polytropic star in 3d by depleting the pressure and find that the criterion of bh formation is @xmath26 , and when the star exceeds the above value , the collapsing star forms a torus which fragments into nonaxisymmetric clumps . @xcite studied the collapsing star of @xmath13 in 2d with a polytropic index of @xmath15 and @xmath27 and find that the criterion for bh formation is determined by the central @xmath28 ( @xmath29 ; @xmath30 is the specific angular momentum and @xmath31 is the cylindrical mass ) at the initial state of the star . the purpose of this paper is the following threefold . the first is to verify the nature of the gravitational collapse of a differentially rotating radially unstable star from the viewpoint of cosmic censor . if we collapse a star of @xmath32 , we expect that the star can not directly form a bh of @xmath32 because of cosmic censor . therefore it is important to find the main cause to prevent bh formation of @xmath33 . from the previous computational results , shock heating @xcite and core bounce @xcite are the dominant phenomena to prevent a star from forming a bh . however , all of the previous calculations have set up violent initial data sets , adding radial velocity or depleting pressure , to induce the collapse , it may cause abrupt transport of the energy . we therefore set up a mild , natural situation , that is a collapse of differentially rotating stars from the onset of radial instability , to focus on a graduate transport of the energy . also , we compute the collapsing star in 3d to allow shock propagation or bar formation , if it occurs . the second is to determine the final outcome of the collapse of differentially rotating smss . two different types of rotation profile arises during the quasi - static evolution of the rotating star , that depends on the environment . one is a uniformly rotating star , maintained uniform rotation during the quasi - static evolution by sufficiently strong viscosity or strong magnetic field . the other is a differentially rotating star , impossible to retain uniform rotation in low viscosity and low magnetic field ( even the star initially takes uniform rotation ) . for the collapse of a uniformly rotating sms from the onset of radial instability , @xcite studied 3d relativistic hydrodynamic simulation in the post - newtonian gravitation and found that the collapse is coherent and that it is likely to form an smbh with no significant bar nor disk formation . followup computation has been performed in 2d hydrodynamics in full general relativity and found that approximately 10 % of the total mass can form a disk , while approximately 90 % of that should form a bh @xcite . for the differentially rotating stars , the final outcome may strongly depends on the amount and the distribution of initial angular momentum , and the nature of angular momentum distribution of smss is largely unknown . we also do not know what path does a differentially rotating sms take during the quasi - static evolution . when the amount of angular momentum of the star is sufficient , the star seems to follow in a quasi - static evolution up to the mass - shedding limit . @xcite find in the newtonian quasi - static evolution that bar formation is inevitable before reaching mass - shedding limit . when the amount of angular momentum of the star is not sufficient , the star seems to collapse at the onset of radial instability before reaching secular instability ( @xmath34 for a uniformly rotating incompressible newtonian stars @xcite ) or mass - shedding limit . the purpose of this paper is to examine the collapse of the sms from the onset of radial instability . when we consider the collapse of radially unstable differentially rotating stars , the final outcome has a possibility of differing from the collapse of uniform rotation because of the strong centrifugal force at the central core , which may prevent the prompt collapse . what is the final fate of the collapse ? does the star fragment due to the growth of the degree of differential rotation ? does the disk form during the collapse ? relativistic simulation can clearly answer to these questions . finally , it is important to probe whether the collapse of differentially rotating smss could be a promising source of gravitational waves . direct detection of gravitational waves by ground based and space based interferometers is of great importance in general relativity , in astrophysics , and in cosmology ( e.g. , * ? ? ? the catastrophic collapse is one of the promising sources of gravitational waves ( e.g. , * ? ? ? for a gravitational collapse of the star in a dynamical timescale , there are two main reasons that prevent a prompt collapse of the star , which should produce gravitational waves at that moment . one reason is core bounce and/or shock heating . suppose gravitational force is balanced to the centrifugal force ( @xmath35 ) in newtonian gravity . the total mass and the angular momentum ( @xmath36 ) are almost conserved during the collapse , assuming that gravitational radiation takes a little role during the collapse in the absence of nonaxisymmetric deformation prior to form a bh . note that @xmath37 is angular velocity , @xmath4 is the radius . we may estimate the bounce radius of the star in a global sense as @xmath38 . since the horizon radius is roughly the order of @xmath2 , core bounce might take place for the case @xmath32 , which prevents violation of cosmic censor @xcite . the other is bar formation . from the dimensional analysis , we can describe the rotational kinetic energy @xmath39 and the gravitational binding energy @xmath40 as @xmath41 and @xmath42 . we also accept the assumption that total mass and angular momentum are almost conserved during the collapse . we could estimate the radius of bar formation in terms of the ratio of the rotational kinetic energy to the gravitational binding energy as @xmath43 . since the dynamical instability for a uniformly rotating , incompressible newtonian star sets in at @xmath44 @xcite and for relativistic gravitation as @xmath45 @xcite , bar formation takes place at the radius @xmath46 for the collapse of @xmath13 . therefore , we may expect that a gravitational collapse of @xmath13 is a promising source of quasi - periodic gravitational waves . this paper is organized as follows . in [ sec : cf ] we present basic equations of our conformally flat approximation in general relativity . we demonstrate our code tests in [ sec : ctest ] . we discuss our initial data sets and our numerical results for our catastrophic collapse in [ sec : jm2 ] . in [ sec : discussion ] we summarize our findings . throughout this paper , we use geometrized units ( @xmath47 ) and adopt cartesian coordinates @xmath48 with the coordinate time @xmath49 . greek and latin indices run over @xmath50 and @xmath51 , respectively . in this section , we briefly describe the conformally flat spacetime ( e.g. , * ? ? ? * ) . we solve the fully relativistic equations of hydrodynamics , but neglect nondiagonal spatial metric components @xcite . we define the spatial projection tensor @xmath52 , where @xmath53 is the spacetime metric , @xmath54 the unit normal to a spatial hypersurface , and where @xmath55 and @xmath56 are the lapse and shift . within a first post - newtonian approximation , the spatial metric @xmath57 may always be chosen to be conformally flat @xmath58 where @xmath59 is the conformal factor ( see * ? ? ? * ; * ? ? ? the spacetime line element then reduces to @xmath60 we adopt maximal slicing , for which the trace of the extrinsic curvature @xmath61 vanishes , @xmath62 the gravitational field equations in conformally flat spacetime for the five unknowns @xmath55 , @xmath56 , and @xmath59 can then be derived conveniently from the 3 + 1 formalism ( e.g. , * ? ? ? * ) . since the spatial metric is conformally flat , the transverse part of its time derivative vanishes . the transverse part of the evolution equation of the spatial metric therefore relates the extrinsic curvature to the shift vector , @xmath63 inserting eq . ( [ eqn : evolutionmetric ] ) into the momentum constraint equation , then , yields an equation for the shift @xmath64 @xmath65 where @xmath66 is the flat space laplacian and @xmath67 is the momentum density . in the definition of @xmath68 , @xmath69 is the stress energy tensor . the conformal factor @xmath59 is determined from the hamiltonian constraint @xmath70 where @xmath71 is the mass - energy density measured by a normal observer . maximal slicing implies @xmath72 , so that the trace of the evolution equation for the intrinsic curvature yields an equation for the lapse @xmath55 @xmath73 combined with eq . ( [ eqn : hamiltonianc ] ) , where @xmath74 . therefore , conformally flat gravitational field equations for the five unknowns @xmath59 , @xmath75 , @xmath64 can be derived by eqs . ( [ eqn : momentumc ] ) - ( [ eqn : evolution ] ) . for a perfect fluid , the energy momentum tensor takes the form @xmath76 where @xmath77 is the rest - mass density , @xmath78 the specific internal energy , @xmath79 the pressure , and @xmath80 the four - velocity . we adopt a @xmath21-law equation of state in the form @xmath81 where @xmath21 is the adiabatic index which we set @xmath82 in this paper . in the absence of thermal dissipation , ( [ gammalaw1 ] ) , together with the first law of thermodynamics , implies a polytropic equation of state @xmath83 where @xmath84 is the polytropic index and @xmath85 is a constant . from @xmath86 together with the equation of state ( eq . [ [ gammalaw1 ] ] ) , we can derive the energy and euler equations according to @xmath87 where @xmath88 and @xmath89 , @xmath90 is the 3-velocity , pressure viscosity , respectively . note that we treat the matter fully relativistically ; the conformally flat approximation only enters through simplifications in the coupling to the gravitational fields . note also that we include an artificial viscosity , since core bounce might occur in our gravitational collapse and that should produce shocks . we will explain our form of an artificial viscosity in [ sec : ctest ] . as a consequence to treat shocks we also need to solve the continuity equation @xmath91 separately . we solve the matter evolution in second order accurate in space and time with the transfer scheme of @xcite . there are three key issues to solve gravitational field equations numerically . the first is to introduce symmetric tensor @xmath92 related to the extrinsic curvature @xmath61 as @xcite @xmath93 therefore , we describe the equations of conformal factor ( eq . [ [ eqn : hamiltonianc ] ] ) and lapse ( eq . [ [ eqn : evolution ] ] ) as @xmath94 note that the dependence value ( such as @xmath59 , @xmath75 ) of the source term is always lower than o(@xmath95 ) , o(@xmath96 ) , which is safe for a convergence during the iteration . , the first term of eq . ( [ eqn : cfactor ] ) behaves as @xmath97 . ] the second is to decompose the vector type elliptic equation . the equation for the shift , @xmath98 can be further simplified by introducing a vector @xmath99 and a scalar @xmath100 according to @xcite @xmath101 the shift can then be computed from @xmath102 and will automatically satisfy eq . ( [ eqn : shifteq ] ) . next , we decompose the symmetric tensor @xmath92 as ( e.g. , * ? ? ? * ) @xmath103 where @xmath104 therefore , we can rewrite the momentum constraint as @xmath105 we also decompose the vector @xmath106 with the same manner as the shift ; @xmath107 where @xmath108 and @xmath109 satisfies @xmath110 note that @xmath68 only appears in the presence of matter , which means that the source terms of eqs . ( [ eqn : wb ] ) and ( [ eqn : wchi ] ) are compact . therefore , we can compute these values quite accurately . to summarize , we have reduced einstein equations in a conformally flat spacetime to 10 elliptic equations for 10 variables ( @xmath108 , @xmath109 , @xmath59 , @xmath75 , @xmath99 , @xmath100 ) , @xmath111 these poisson - type equations are solved by imposing the following boundary condition at outer boundaries @xmath112 here we briefly explain our order of computing 10 poisson - type equations . first , we solve @xmath108 and @xmath109 since the source terms do not depend on other unknown spacetime variables . once we derive the symmetric tensor @xmath113 from @xmath108 and @xmath109 , we solve @xmath59 iteratively till the convergence . after that we solve @xmath75 iteratively till the convergence . finally , we solve @xmath99 and @xmath109 . we use modified incomplete cholesky decomposition conjugate gradient ( miccg ) method @xcite to solve elliptic equations . finally , we should maintain our grid resolution at the central core of the collapsing star . the nested grid is one of the most appropriate way to handle this problem ( see , e.g. , * ? ? ? our method is a mimic version of the nested grid to handle the collapsing star problem , which is based on the rigridding method of @xcite . for the post homologous collapse of the sms , there are two different timescales in the collapsing star ; one is the collapsing timescale at the center and the other is the one at the envelope , which is sufficiently longer than that at the center . therefore , @xcite add the grid number at the intermediate stage of the collapse in order to control both regions , center and envelope of the star . since we perform the collapse of an sms in 3d , we concentrate on the central core of the collapse due to the limitation of computational resource . in order to maintain spatial grid resolution especially at the center , we shrink the total grid to one half as the core radius of the star becomes half from the beginning of the present grid resolution . although this method changes the boundary condition when we `` zoom in '' the computational domain , it is approximately good when the difference of the gravitational mass between the end of the previous computational grid and the beginning is small . in fact , the difference of the gravitational mass between every two computational domains is less than @xmath114 , and therefore this `` zoom - in '' method should behave as a reasonable approximation for a collapse of differentially rotating smss . we monitor the gravitational mass @xmath2 and the angular momentum @xmath1 @xmath115 \psi^{5 } \right . \nonumber \\ & & \left . + \frac{1}{16 \pi } \psi^{5 } k_{ij } k^{ij } \right ] d^{3 } x , \\ j & = & - \frac{1}{2 \pi } \oint_{\infty } ( x k^{i}_{y } - y k^{i}_{x } ) \psi^{6 } ds_{i } \nonumber \\ & = & \int ( x j_{y } - y j_{x } ) \psi^{6 } d^{3 } x , \end{aligned}\ ] ] during the evolution . in all cases reported in [ sec : jm2 ] , the gravitational mass was conserved up to @xmath116% , and the angular momentum up to @xmath117% of their initial values . we also compute proper mass @xmath118 , rotational kinetic energy @xmath39 , gravitational binding energy @xmath40 in equilibrium as @xmath119 where @xmath37 the angular velocity of the star . since we use a polytropic equation of state at @xmath120 , it is convenient to rescale all quantities with respect to @xmath85 @xcite . since @xmath121 has dimensions of length , we introduce the following nondimensional variables @xmath122 henceforth , we adopt nondimensional quantities , but omit the bars for convenience ( equivalently , we set @xmath123 ) . first , we demonstrate 1d relativistic wall shock problem to check whether our code has an ability to treat shock . we set up the form of artificial pressure viscosity @xcite as @xmath124 where @xmath125 , @xmath126 is the local grid spacing and where @xmath127 is the dimensionless parameter . when evolving the matter equations we limit the stepsize @xmath128 by the following courant condition ( @xmath129 ) , where the second term represents dynamical time of a collapsing star . we choose the dimensionless parameter @xmath130 . we have tested the ability of our code to resolve shocks by performing a wall - shock problem , in which two phases of a fluid collide . in fig . [ fig : wshock ] we compare numerical results with the analytic solution for initial velocities that are similar to those found in our simulations below . with @xmath131 we find good agreement , and set this value to simulate the catastrophic collapse in [ sec : jm2 ] . next , we demonstrate a spherical dust collapse in fig [ fig : oscollapse ] . note that our conformally flat approximation retains all the nonlinear terms to maintain the exact dynamics for a spherical spacetime . we compare 1d and 3d results to check whether our 3d code has an ability to reproduce a spherical dust collapse . note that we construct our 1d code following wilson s approach @xcite described in appendix [ sec:1dfield ] . we choose the grid size of 1d computation as 5,000 , while of 3d one as ( @xmath132 ) . we find a very good agreement of the central lapse between our 1d result and our 3d conformally flat one within the error of 1% . we terminate the integration of our 3d code at @xmath133 , since the convergence of the iteration process in our elliptic solver becomes significantly slow . then , we demonstrate the collapse of a spherical star in fig . [ fig : spalp ] ( see table [ tbl : sphinitial ] for an equilibrium profile ) . since a radially unstable spherical star only collapses to form a spherical bh promptly , we should follow the collapse of a spherical star to check whether our code has an ability to follow the collapse and to find a signal of bh formation . we set up the same grid size as we did for a spherical dust collapse , covering the star with 81 grid points across the diameter . the central lapse of the star decreases monotonically , which means a collapse of a spherical star directly forms a spherical bh . although we terminate our integration around @xmath134 , the fact that 1d computation of the spherical star collapse finds an apparent horizon at @xmath135 indicates that a bh might form . c c c c @xmath136 & @xmath137 & @xmath138 & @xmath139 finally , we check the stability of a uniformly rotating star whether our code has an ability to determine the radial stability of a star . since we determine the radial stability of a differentially rotating star in [ sec : jm2 ] , we should check the sensitivity of our code to determine the critical onset of radial instability of uniformly rotating stars by comparing the results derived from the turning point method . to assess the ability of our code to distinguish stable stars from unstable ones with rotation , we consider an equilibrium sequence of uniformly rotating stars of the fixed angular momentum @xmath1 ( @xmath140 at the turning point ) . while the turning point criterion strictly identifies the onset of secular instability , the point of onset of dynamical instability nearly coincides with the secular instability point . we adopt the polytropic index of @xmath141 , which is regarded as radiation pressure dominant , sms sequence , and grid resolution as in the spherical simulations reported above . we decrease the initial pressure to induce the collapse ( @xmath142 ) . figure [ fig : roteq ] summarizes our dynamical stability analysis for the rotating sms . we conclude that with the adopted grid resolution , our code can distinguish stable rotating stars from unstable ones within 0.3% of the maximum gravitational mass . figure [ fig : rotst ] shows the evolution of the central density for stable and unstable rotating stars . here we explain our initial data sets for collapsing stars . since we are interested in a collapsing star of @xmath13 , we have three requirements to construct differentially rotating equilibrium stars . the first requirement is to construct radially unstable differentially rotating stars . the critical @xmath21 to the onset of radial instability for slow rotation and weak gravitational field is described analytically as @xcite @xmath143 where @xmath144 is the inertia momenta of the star . note that relativistic gravitation unstabilizes the star , while rotation , which produces centrifugal force , stabilizes the star . although we take into account the differential rotation in the equilibrium star , this criterion ( eq . [ [ eqn : critg ] ] ) may indicate the appropriate direction to choose a parameter sets of the initial condition of the rotating collapsing star . from eq . ( [ eqn : critg ] ) , we should at least choose soft equation of state to induce the collapse and set an @xmath145 polytropic star , an sms sequence of the star . the second requirement is to construct a star which holds @xmath26 . from the critical onset of radial instability in the equilibrium star , a uniformly rotating sms takes the maximum of @xmath146 when @xmath147 @xcite . to verify the nature of cosmic censor , we should at least go beyond @xmath32 . the main restriction to hold large @xmath148 comes from the mass shedding limit of the star . therefore , to construct a star of @xmath149 in radially unstable branch , differential rotation of the star is required in sms sequence . the last requirement is high degree of differential rotation . it is indeed one path to construct high degree of differential rotation in quasi - static evolution , and reach the onset of radial instability . to summarize , we need soft equation of state and high degree of differential rotation , and choose an @xmath145 polytropic star ( sms sequences ) and the degree of differential rotation as @xmath150 where we define the rotation profile as @xmath151 in the newtonian limit ( @xmath152 , @xmath153 ) , this rotation law can be written as @xmath154 where @xmath155 is the degree of differential rotation , @xmath156 is the cylindrical radius of the star . since @xmath155 has a dimension of length , we normalize it with a proper equatorial radius @xmath157 , ( @xmath158 ) . hereafter we choose @xmath159 to construct relatively a high degree of differential rotation . we briefly summarize our method to construct relativistic rotating equilibrium stars in appendix [ sec : rst ] . we summarize the parameters of our differentially rotating stars at initial in table [ tbl : initial ] . we slightly perturb our initial equilibrium state according to @xmath160 where @xmath161 . we install @xmath162 and @xmath163 density perturbation to provide the seed for one - armed spiral and bar formation , if the physical situation should lead to unstable growth . we adopt a grid size ( @xmath164 ) , so that the star is initially covered by 161 points across the equatorial diameter . we evolve the rotating sms up to the point at which the conformally flat approximation breaks down . c c c c c & @xmath165 & @xmath166 & @xmath167 & @xmath168 + @xmath169 & @xmath170 & @xmath170 & @xmath170 & @xmath170 + @xmath171 & @xmath172 & @xmath173 & @xmath174 & @xmath175 + @xmath176 & @xmath177 & @xmath178 & @xmath179 & @xmath180 + @xmath181 & @xmath182 & @xmath183 & @xmath184 & @xmath185 + @xmath2 & @xmath186 & @xmath187 & @xmath188 & @xmath189 + @xmath1 & @xmath190 & @xmath191 & @xmath192 & @xmath193 + @xmath194 & @xmath195 & @xmath196 & @xmath197 & @xmath198 + @xmath19 & @xmath199 & @xmath200 & @xmath201 & @xmath202 + @xmath203 & @xmath204 & @xmath205 & @xmath205 & @xmath206 + stability & unstable & unstable & stable & stable figure [ fig : rhomax ] shows the results of radial stability in 4 differentially rotating stars . since we do not have a tool to determine radial stability in differentially rotating stars from their equilibrium states ( but for determining the criterion of secular stability in rotating stars , see @xcite ) , the evolution is necessary to determine its stability . the criterion to determine the radial stability is as follows . when the central density of the star grows exponentially within a few dynamical time , we determine the star radially unstable . on the other hand , when the central density of the star oscillates around its equilibrium state , we determine the star radially stable . from this criterion , model i and ii have an exponential growth of the central density that means radially unstable , while model iii and iv have maximum at several dynamical times that means radially stable . we also show the evolution of central lapse in fig . [ fig : lapse ] . the rapid decrease of @xmath207 below @xmath208 indicates that a bh is likely to form . model ii shows that the central lapse monotonically decreases from @xmath209 to @xmath210 . this figure shows that we can follow the collapse from the regime of newtonian gravity ( @xmath211 ) to that of relativistic gravity ( @xmath212 ) . on the other hand model iii shows that the central lapse oscillates around its equilibrium state , and that also means the star is radially stable . we show the final density snapshots of the stars in the equatorial plane ( fig . [ fig : qeq ] ) and in the meridional plane ( fig . [ fig : qmd ] ) . even in the final snapshots of radially unstable stars , the collapse is almost axisymmetric . also , from the snapshots in the meridional plane , the material around the rotational axes collapses faster than the surrounding material due to the strong nonlinear gravitational field . let us now focus on the collapsing star ( model ii ) to probe the final outcome . we locally define the cylindrical rest mass @xmath31 , angular momentum @xmath30 , specific angular momentum @xmath213 , keplerian angular velocity @xmath214 as @xmath215 to indicate the transport of angular momentum and the distribution of @xmath216 . we assume the axisymmetric collapse to investigate the final outcome , since the meridional density snapshots ( fig . [ fig : qmd ] ) behave almost axisymmetry . figure [ fig : mass ] shows the concentration of mass density profile during the collapse . as the collapse goes on , the increase ratio of the cylindrical mass at the core of the star becomes high . in fact , @xmath217% of the cylindrical mass is inside the radius of @xmath218 at @xmath219 . we also show the mean radius , defined as @xmath220 , during the collapse in fig . [ fig : rm ] . note that @xmath221 is the rest mass . we find that the mean radius monotonically decreases , which means that the collapse of the central core is coherent . figure [ fig : omega ] shows the variation of angular velocity profile during the collapse . since the cylindrical rest mass density shows a coherent collapse , the degree of differential rotation significantly increases at the core during the collapse and approaches to the `` keplerian '' angular velocity defined in eq . ( [ eqn : keplarian ] ) . we show the specific angular momentum distribution during the collapse in fig . [ fig : jm ] . in a global sense , @xmath19 is conserved so that a final bh should violate cosmic censor , if it forms . however , we find that most of the matter collapses to form a bh , while not a few amount of angular momentum stays at the surface area of the star that prevents forming a bh of @xmath149 . we show the distribution of @xmath216 in fig . [ fig : jm2 ] during the collapse to verify cosmic censor . we find that mass density collapses first in the central part and angular momentum remains in the surface of the star that prevents forming a bh of @xmath149 . note that the distribution of @xmath222 is only an indicator to interpret the physical cause to prevent bh formation . although we should define the total gravitational mass locally to discuss the local distribution of @xmath19 , there is no knowing how to define it . therefore we use the local rest mass instead . we investigate the collapse of differentially rotating smss by means of hydrodynamic simulations in conformally flat approximation in general relativity . we start our collapse around the onset of radially instability at @xmath3 to the point where conformally flat approximation breaks down . we find that cosmic censor even holds for a gravitational collapse of a radially unstable differentially rotating equilibrium sms of @xmath32 . the main reason to prevent formation of a bh of @xmath32 is that quite a large amount of angular momentum stays at the surface , not core bounce nor shock propagation , bar formation in our model . note that even a thin disk near the surface of the star can hold relatively a large amount of angular momentum if the radius is large . the above conclusion is supported by @xcite in full general relativistic simulation that the criterion of bh formation is determined at the central @xmath28 ( @xmath223 ) . in fact , the central @xmath28 of model ii is @xmath224 , which satisfies their criterion . the collapse of a differentially rotating , radially unstable sms is coherent and likely leads to formation of an smbh . this situation is quite similar to the collapse of a uniformly rotating sms ( e.g. , * ? ? ? * ; * ? ? ? * ) , since @xmath225 seems still not sufficient to change the path of the collapse from the one of uniform rotation . however , this final outcome may depend on the equation of state of the star . @xcite treated the isothermal ( @xmath226 ) collapse of initially homogeneous , uniformly rotating , low entropy clouds via smooth particle hydrodynamics ( sph ) simulations . they found considerable fragmentation into dense clumps , and disk formation containing @xmath227 of the mass . they concluded that a seed bh will form at the center and that it likely will grow gradually by accretion in our model . also @xcite treated the collapse of uniformly rotating polytropic star from the critical onset in the range of @xmath228 and found that the mass of the disk is less than @xmath229 of the gravitational mass . we can not find any evidence of bar formation nor significant disk formation from the rotating collapse prior to bh formation . the phenomenon of no bar formation also comes from the fact that mass density collapses first to form a bh . in such case , @xmath194 can not scale in @xmath230 due to the growth of the degree of differential rotation , and as a fact the star of @xmath194 can not reach the dynamical instability point of @xmath231 . since the @xmath162 dynamical instability takes place when the star has a toroidal structure and soft equation of state @xcite , we can not find any evidence of toroidal structure at the time we stop our integration . the rotating sms collapse is a promising source of burst gravitational waves and of quasi - normal mode ringing waves . the characteristic frequency for burst ( @xmath232 ) , quasi - normal mode ringing ( @xmath233 ) , and the wave amplitude for burst ( @xmath234 ) , quasi - normal mode ringing ( @xmath235 ) , are @xcite @xmath236 , \\ h_{\rm burst } & \sim & 1 \times 10^{-18 } \left ( \frac{m}{10^{6 } m_{\odot } } \right ) \left ( \frac{1 { \rm g pc}}{d } \right ) \left ( \frac{m}{r } \right ) , \\ f_{\rm qnm } & \sim & 2 \times 10^{-2 } \left ( \frac{10^{6 } m_{\odot}}{m } \right ) { \rm [ hz ] } , \\ h_{\rm qnm } & \sim & 6 \times 10^{-19 } \left ( \frac{\delta e_{\rm gw}/m}{10^{-4 } } \right)^{1/2 } \left ( \frac{2 \times 10^{-2 } { \rm [ hz]}}{f_{\rm qnm } } \right)^{1/2 } \nonumber \\ & & \times \left ( \frac{m}{10^{6 } m_{\odot } } \right)^{1/2 } \left ( \frac{1 { \rm gpc}}{d } \right ) , \end{aligned}\ ] ] where @xmath237 is the distance from the observer and @xmath238 is the total radiated energy . we set @xmath239 , a characteristic mean radius during bh formation . since the main targets of lisa are gravitational radiation sources between @xmath240 and @xmath241 hz , it is possible that lisa can search for the burst and quasi - normal ringing waves accompanying rotating sms collapse and formation of an smbh . we would like to thank an anonymous referee for his / her critical reading of our manuscript and constructive suggestions . he also thanks hideki asada , silvano bonazzola , eric gourgoulhon , takashi nakamura , ken - ichi nakao , luciano rezzolla , stu shapiro , nick streigeolous , takahiro tanaka , koji ury , shin yoshida for discussion . this work has been supported in part by mext grant - in - aid for young scientists ( no . 200200927 ) , astronomical data analysis center , national astronomical observatory of japan , and by mext grant - in - aid for 21coe program at the department of physics , kyoto university . numerical computations were performed on the nec sx-5 machine in the yukawa institute for theoretical physics , kyoto university , on the vpp-800 machine in the academic center for computing and media studies , kyoto university , and on the vpp-5000 machine in the astronomical data analysis center , national astronomical observatory of japan . here we briefly explain relativistic hydrodynamics in spherically symmetric spacetime @xcite that we use in [ sec : cf ] . we can describe the line element of spherically symmetric spacetime in the isotropic coordinate as @xmath242 where @xmath55 is the lapse , @xmath243 is the shift , @xmath59 is the conformal factor . we use the maximal slicing condition @xmath244 , which derives the following relation to the components of an extrinsic curvature as @xmath245 as a consequence , we only need to consider @xmath246 to construct an extrinsic curvature . the momentum constraint derives the equation of the extrinsic curvature as @xmath247 the integration form of eq . ( [ eqn:1dmom ] ) is @xmath248 the restriction of the spatial metric to be conformally flat requires the following equation @xmath249 which gives us an appropriate boundary condition for the extrinsic curvature as @xmath250 . also the integration form of eq . ( [ eqn:1dbet ] ) is @xmath251 the hamiltonian constraint and the trace of the evolution equation guide @xmath252 = 2 \pi \alpha \psi^{5 } ( \rho_{\rm h } + 2 s ) + \frac{21}{16 } \alpha \psi^{5 } ( k^{r}_{r})^{2 } \label{eqn:1dev } , \end{aligned}\ ] ] with the boundary conditions @xmath253 at the grid edge and @xmath254 at the center ( @xmath255 ) . therefore , we can determine 4 unknown variables ( @xmath246 , @xmath59 , @xmath256 , @xmath243 ) with 4 equations ( eqs . [ [ eqn:1dintk ] ] , [ [ eqn:1dintb ] ] , [ [ eqn:1dham ] ] , [ [ eqn:1dev ] ] ) . instead of @xmath246 from a computational point of view ( see [ subsec : nt ] . ) . ] matter equations can be written as @xmath257 we have not taken into account artificial viscosity , since shock does not seem to play an important role in a spherical dust collapse and the collapse of a spherical star . in [ sec : ctest ] , we adopt our 1d relativistic hydrodynamic code to a spherical dust collapse by setting @xmath258 and @xmath259 @xmath260 equilibrium variables of them in typical polytropic index . we briefly mention a method to compute an apparent horizon in 1d @xcite . an apparent horizon is defined as an outermost marginally trapped surface whose future - directed outgoing null geodesics have zero expansion @xmath261 @xcite . the outgoing null vector @xmath262 can be expressed as @xmath263 , where @xmath264 is unit normal of the spacelike hypersurface defined as @xmath265(@xmath266 , 0 , 0 , 0 ) , @xmath267 is the out - directed spacelike vector orthogonal to the surface . the projection tensor @xmath268 and the extrinsic curvature @xmath61 are described as @xmath269 the expansion @xmath261 is then written as @xmath270 for spherically symmetric spacetime , we can choose the spacelike vector @xmath271 as @xmath272 , and then the radius @xmath273 of an apparent horizon satisfies the following equation , @xmath274 note that an event horizon always lies outside an apparent horizon , and both horizons coincide with each other when the spacetime is stationary . also an event horizon always forms before the time that an apparent horizon does @xcite . here we summarize our method to construct rotating relativistic equilibrium stars , which is based on @xcite ( see also * ? ? ? * for a historical review ) . first we focus on how to solve the relativistic euler equation . the equation can be described in axisymmetric spacetime as @xmath275 where @xmath276 is a specific enthalpy . we also assume specific type of rotation law as @xmath277 where @xmath278 is the central angular velocity , in order to integrate eq . ( [ eqn : reuler ] ) . the bernoulli s equation is driven by integrating the relativistic euler equation ( [ eqn : reuler ] ) as @xmath279 where @xmath280 , \\ k & \equiv & \int \frac{du^{t}}{u^{t } } = \ln u^{t } = - \frac{1}{2 } \ln [ \alpha^{2 } - \psi^{4 } [ ( \beta^{x } - \omega y)^{2 } + ( \beta^{y } + \omega x)^{2 } ] ] , \\ r & \equiv & \int u^{t } u_{\varphi } d\omega = -\frac{1}{2 } a^{2 } ( \omega_{c } - \omega)^{2 } , \end{aligned}\ ] ] where @xmath281 in this section . ] . note that we have already assumed polytropic equation of state @xmath282 where @xmath283 , @xmath25 is a polytropic index . therefore , we can describe the matter distribution equation by using eq . ( [ eqn : bernoulli ] ) and the rotation law ( eq . [ [ eqn : rlaw ] ] ) as @xmath284}{\sqrt{u^{t } } } - 1 \right ] \nonumber \\ & = & \frac{1}{n+1 } \left [ \frac{c \exp[a^{2}(\omega_{c } - \omega)^{2 } / 2 ] } { \sqrt{\alpha^{2 } - \psi^{4 } [ ( \beta^{x } - \omega y)^{2 } + ( \beta^{y } + \omega x)^{2 } ] } } - 1 \right ] \label{eqn : qdis } , \\ a^{2 } ( \omega_{c } - \omega ) & = & u^{t } u_{\varphi } = \frac{\psi^{4 } [ x ( \beta^{y } + \omega x ) - y ( \beta^{x } - \omega y ) ] } { \alpha^{2 } - \psi^{4 } [ ( \beta^{x } - \omega y)^{2 } + ( \beta^{y } + \omega x)^{2 } ] } . \label{eqn : rotlaw}\end{aligned}\ ] ] note that we assume conformally flat spacetime to derive the final part of the equality in eqs . ( [ eqn : qdis ] ) and ( [ eqn : rotlaw ] ) . from the computational point of view , we introduce nondimensional quantities rescaled in the equatorial proper radius of the star ( @xmath285 ) as @xmath286 we also rescale the lapse and conformal factor using the nature of scale free in newtonian gravity as @xmath287 to determine the matter distribution and rotation profile of the next iteration step , we have to determine five unknown variables ( @xmath285 , @xmath288 , @xmath278 , @xmath289 , @xmath290 ) from five equations . therefore , we evaluate eqs . ( [ eqn : qdis ] ) and ( [ eqn : rotlaw ] ) at three locations , the point of the maximum rest mass density , that of polar surface , and that of the equatorial surface . note that eq . ( [ eqn : rotlaw ] ) becomes an identity at polar surface of the star , and thus we have five equations to be solved . we use newton - rapson method @xcite to solve these equations . once we determine 5 unknown variables , we solve eq . ( [ eqn : rotlaw ] ) using newton - rapson method to determine the rotation profile . after that we can determine the matter distribution ( eq . [ [ eqn : qdis ] ] ) . next we briefly summarize the gravitational field equations normalized by proper radius as follows . @xmath291 we choose the same boundary conditions as we choose in our evolution code ( see [ sec : cf ] ) . note that we construct the star in 3d using octant symmetry to adopt our initial data to our 3d evolution code smoothly . to summarize , we first solve the gravitational field equations ( eqs . [ [ eqn : apcfbx ] ] [ [ eqn : apcfeta ] ] ) . next , we determine 5 unknown quantities ( @xmath285 , @xmath288 , @xmath278 , @xmath289 , @xmath290 ) from 5 equations ( eqs . [ [ eqn : qdis ] ] and [ [ eqn : rotlaw ] ] ) using newton - rapson method . finally we determine the matter distribution and rotation profile for the next iteration step . we continue this iteration cycle till @xmath285 converges ; i.e. @xmath285 goes bellow the error of @xmath292 . we also check the error rate of physical quantities such as @xmath2 and @xmath1 and find that they are in fact below the error rate of @xmath240 . 99 baumgarte , t. w. & shapiro , s. l. 1999 , , 526 , 941 blanchet , l. , damour , t. & schfer , g. 1989 , , 242 , 289 bonazzola , s. , frieben , j. , & gourgoulhon , e. 1998 , , 331 , 280 bowers , r. l. , & wilson , j. r. 1991 , numerical modeling in applied physics and astrophysics ( boston : jones and bartlett pub . ) , chap . 9 . centrella , j. m. , new , k. c. b. , lowe , l. l. , & brown , j. d. 2001 , , 550 , l193 chandrasekhar , s. 1965 , , 142 , 1488 . 1969 , _ ellipsoidal figures of equilibrium _ , ( new york : yale univ . press ) , 61 chandrasekhar , s. & lebovitz , n. r. 1965 , , 142 , 1488 cook , g. b. , shapiro , s. l. , & teukolsky , s. a. 1992 , , 398 , 203 . 1994 , , 422 , 227 duez , m. , shapiro , s. l. , & yo , h. -j . 2004 , , 69 , 104016 hawkings , s. w. , & ellis , g. f. r. 1973 , the large scale structure of space - time , chap . 9.2 hawley , j. f. , smarr , l. l. , & wilson , j. r. 1984 , , 55 , 211 isenberg , j. , & nester , j. 1980 , in general relativity and gravitation vol . 1 : one hundred years after the birth of albert einstein , ed . a. held ( new york : plenum press ) , 23 komatsu , h. , eriguchi , y. , & hachisu , i. 1989 , , 237 , 355 kormendy , j. 2004 , in carnegie observatories centennial symposium . 1 : coevolution of black holes and galaxies , ed . l. c. ho ( cambridge : cambridge univ . press ) , in press ( astro - ph/0306353 ) loeb , a. , & rasio , f. a. 1994 , , 432 , 52 leaver , e. w. 1985 , proc . london a , 402 , 285 murata , k. , natori , r. & karaki , y. 1990 , large - scale numerical simulation ( tokyo : iwanami ) , chap . 5.3 ( in japanese ) nakamura , t. 1981 , prog . phys . , 65 , 1876 nakamura , t. , oohara , k. , & kojima , y. 1987 , prog . , 90 , part i , sec . 2 nakamura , t. , shapiro , s. l. , & teukolsky , s. a. 1988 , , 38 , 2972 new , k. c. b. 2003 , living reviews in relativity , 6 , 2 new , k. c. b. , & shapiro , s. l. 2001 , , 548 , 439 oohara , k. , & nakamura , t. 1989 , prog . phys . , 82 , 535 press , w. h. , flannery , b. p. , teukolsky , s. a. , & vetterling , w. t. 1992 , numerical recipes in fortran ( 2nd ed . ; cambridge : cambridge univ . press ) , chap . 9.6 rees , m. j. 1998 , in black holes and relativistic stars , ed . r. m. wald ( chicago : univ . chicago press ) , 79 . 2001 , in black holes in binaries and galactic nuclei , ed . l. kaper , e. p. j. van den heurel , & p. a. woudt ( new york : springer - verlag ) , 351 ruffert , m. 1992 , , 265 , 82 saijo , m. , shibata , m. , baumgarte , t. w. , & shapiro , s. l. 2001 , , 548 , 919 saijo , m. , baumgarte , t. w. , shapiro , s. l. , & shibata , m. 2002 , , 569 , 349 saijo , m. , baumgarte , t. w. , & shapiro , s. l. 2003 , , 595 , 352 sekiguchi , y. -i . , & shibata , m. 2004 , , submitted ( gr - qc/0403036 ) shapiro , s. l. & teukolsky , s. a. 1980 , , 235 , 199 . 1983 , black holes , white dwarfs and neutron stars , ( new york : john wiley and sons ) , chap . 7 . 1985 , , 298 , 34 . 1991 , , 66 , 994 . 1992 , , 54 , 2006 shibata , m. 1997 , , 55 , 2002 . 2000 , prog . phys . , 104 , 325 . 2003 , , 595 , 992 . 2004 , , 605 , 350 shibata , m. , baumgarte , t. w. , & shapiro , s. l. 2000 , , 542 , 453 shibata , m & nakamura , t. 1995 , , 52 , 5428 shibata , m & shapiro , s. l. 2002 , , 572 , l39 stark , r. f. & piran , t. 1985 , , 55 , 891 ( erratum 56 , 97 ) . 1986 , in dynamical spacetimes and numerical relativity , ed . j. m. centrella ( cambridge : cambridge univ . press ) , 40 stergioulas , n. 2003 , living reviews in relativity , 6 , 3 thorne , k. s. 1998 , in black holes and relativistic stars , ed . r. m. wald ( chicago : univ . chicago press ) , 41 van leer , b. 1977 , j. comp . phys . , 23 , 276 wald , r. m. 1984 , general relativity ( chicago : univ . chicago press ) , chap . 12.1 wilson , j. r. 1979 , in sources of gravitational waves , ed . l. l. smarr ( cambridge : cambridge univ . press ) , 423 wilson , j. r. & mathews , g. j. 1989 , in frontiers in numerical relativity , ed . c. r. evans , l. s. finn , d. w. hobill ( cambridge : cambridge univ . press ) , 306 . 1995 , , 75 , 4161 york , j. w. , jr . 1979 , in sources of gravitational waves , ed . l. l. smarr ( cambridge : cambridge univ . press ) , 83

we consider stochastic billiards ( knudsen random walk ) in two - dimensional infinite domains ( ` tubes ' ) of the form @xmath0 where @xmath1 and @xmath2 is a monotone smooth function . generally speaking , billiards are dynamical systems describing the motion of a particle in a region with reflection rules at the boundary : such systems have been extensively studied in the mathematical and physical literature . when the reflection rule is randomized , we have a stochastic billiard process . the stochastic models have received much less attention ; invariant distributions for stochastic billiards in general ( mostly bounded ) domains were studied in @xcite . in the present paper we study the qualitative ( recurrence or transience ) and quantitative ( almost - sure bounds ) behaviour of stochastic billiards in _ unbounded _ domains in the plane . physical motivation for the billiards model comes from the dynamics of ideal gas models in the so - called knudsen regime where intermolecular interactions are neglected . the random reflection law is motivated by the fact that the particle is small and the surface off which it reflects has a complicated ( rough ) microscopic structure . the behaviour of such knudsen flows is of interest in several areas of physics , chemistry , and technology . for physical background , see for instance @xcite . for more on the motivation of the model in the present paper , see @xcite and references therein . we mention some related models at the end of this section . informally , the model that we study can be described as follows . a particle moves ballistically inside the domain @xmath3 at constant velocity ( of unit magnitude , say ) , and each time that the particle hits the boundary , it is reflected at a random angle @xmath4 to the inwards - pointing normal at the boundary curve @xmath5 , where @xmath6 does not depend on the initial direction of the particle . so , inside the domain the motion is deterministic : the stochasticity is introduced via the distribution @xmath6 of the random reflections . in this paper , we take the distribution of @xmath6 to be symmetric about @xmath7 . we consider two types of problem : ( i ) the continuous - time motion of the particle in the tube ; and ( ii ) the discrete - time embedded process obtained by observing the instances of collisions on the boundary ( in other words , unit time elapses between reflections ) . the embedded process is , from our point of view , of interest in its own right ( and behaves very differently to the continuous - time process ) , and is also a vital ingredient for the study of the continuous - time process . the asymptotic phenomena that we study for each process are also of two main types : ( i ) recurrence or transience , and existence of moments for recurrence times ; and ( ii ) almost - sure bounds for the location of the particle at time @xmath8 , as @xmath9 . what kind of regions @xmath3 are of interest ? in the case of @xmath10 , our model becomes symmetrically reflecting motion in a strip . roughly speaking for the moment , the horizontal motion of the particle is then described by a random walk with zero drift , and so the model is null - recurrent . on the other hand , if @xmath11 our tube is a wedge and , with our reflection rule , at any reflection there is a positive chance that the particle will head off to infinity and never return to a bounded set : it is transient . similarly if @xmath12 for @xmath13 , it is not difficult to show that the particle will follow nearby the deterministic trajectory which goes to infinity with linear speed . this dichotomy motivates the study of tubes with widths that grow sub - linearly , to probe precisely the transition between recurrence and transience . it is also natural to consider the case where the tube has decreasing width . the primary family of tubes to bear in mind has @xmath14 for some @xmath15 , although we do consider more general forms for the function @xmath16 . then there are two main cases : @xmath17 and @xmath18 . see figure [ tubes ] . we obtain criteria for recurrence and transience of our processes . loosely speaking , these results imply , for example , that for @xmath19 the particle always has transient dynamics , but can be recurrent for @xmath20 depending on the reflection distribution . also , for very shallow tubes such as @xmath21 , the introduction of random reflections ensures that the process is recurrent , while the corresponding deterministic evolution along the normals is transient . ( see theorem [ gen_g ] below . ) we also obtain almost - sure bounds for the horizontal position of the particle , in both discrete- and continuous - time settings . thus we have information about the asymptotic speed of the processes . for example , when @xmath14 for @xmath22 , theorem [ time ] below implies that for all large times @xmath8 the continuous - time process is at position @xmath23 , ignoring logarithmic terms . in particular , the motion is _ super - diffusive_. in the case @xmath18 , recurrence is evident . in this case we obtain a criterion for the finiteness of the mean recurrence - time ( roughly speaking , ` ergodicity ' ) . we also obtain polynomial estimates for the position of the particle ; the ergodicity has so - called ` heavy tails ' . a step in our proofs will involve relating the stochastic billiard process to an instance of the so - called lamperti problem ( after @xcite ) of a one - dimensional stochastic process with asymptotically zero mean drift . a crucial ingredient to our proofs for the stochastic billiard process is provided by some new results for general lamperti - type processes . these results ( theorems [ prop1 ] , [ lamp_low ] , [ prop6 ] ) are of independent interest , and are in some cases apparently new even for the nearest - neighbour random walk ( birth - and - death chain ) . to obtain our almost - sure bounds for lamperti - type processes , we state and prove some general results on obtaining almost - sure bounds for stochastic processes via semimartingale - type criteria . we conclude this section with some remarks upon related models . as previously mentioned , stochastic billiard models have received comparatively little attention . classical ( that is , non - stochastic ) billiard models have been extensively studied , particularly in mathematical physics , and the literature is vast ; see for instance @xcite . a common approach in the classical setting is via dynamical systems for which the reflection rule is elastic . in this setting , we mention that billiards in certain unbounded domains resembling the tubes considered here ( at least for @xmath24 ) have been studied ; see for instance @xcite and references therein . an infinite - tube billiard model with a stochastic component ( cf the @xmath25 case here ) is analyzed in @xcite . in the dynamical systems setting , studying ergodicity is typically a primary goal . in the stochastic setting , the results of the present paper demonstrate a complete range of behaviours , from positive - recurrence ( roughly speaking , ` ergodicity ' ) , through null - recurrence , to transience . in the next section we give a more formal definition of the model and state our main results . the statement of our auxiliary results on the lamperti problem is deferred to section [ lamperti ] . the rigorous formulation of the stochastic billiard model that we consider is essentially given in @xcite . we now describe the construction , which is modified slightly from that in @xcite to fit with our context . let @xmath26 , which will be large but fixed , to be specified later . for a monotone function @xmath2 , let @xmath27 while the continuous - time process is perhaps most natural to describe , it is more convenient to construct the discrete - time process first , and then to construct the continuous - time process from that . thus we define a discrete - time markov chain , informally obtained by recording the locations of the successive hits of the particle on the boundary @xmath5 . this is a markov chain with state - space @xmath28 that we denote by @xmath29 ( @xmath30 ) , where for @xmath31 we write in coordinates @xmath32 . thus when @xmath33 , @xmath34 . we call @xmath35 the _ collisions process_. informally , the state @xmath36 represents the absorbing state achieved if the particle attains a trajectory that will never intersect with @xmath5 again . figure [ tubes ] illustrates the idea of the construction . we construct @xmath35 formally as follows . suppose that @xmath37 , with @xmath38 . let @xmath39 be i.i.d . random variables with the distribution of a random angle @xmath6 where @xmath40 = 1 $ ] . for @xmath41 , given @xmath42 , we perform a step of our process as follows : * if @xmath43 , set @xmath44 ; * otherwise , @xmath45 specifies a ray @xmath46 starting at @xmath47 with angle @xmath45 to the interior normal to @xmath5 at @xmath48 ; we adopt the convention that positive values of the angle correspond to the right of the normal ; negative values correspond to the left . * if the ray @xmath46 does not intersect with @xmath49 , set @xmath50 . otherwise , let @xmath51 be the first point of intersection of the ray @xmath46 with @xmath5 . then if @xmath52 , set @xmath53 , else set @xmath54 . this defines the discrete - time process @xmath35 . the randomness is introduced through the random draws from the reflection distribution @xmath6 ; if @xmath6 is degenerate ( i.e. , equal to a constant almost surely ) , then @xmath35 is deterministic . note that the construction ensures that @xmath55 whenever it is defined . we need some modified form of ` reflection ' away from the vertical boundary at @xmath56 , such as that specified by ( iii ) . this is for technical reasons , and in particular ensures that the process does not jump directly to @xmath36 , since that case will not be of interest to us here . the particular form of this ` reflection ' given by ( iii ) is rather arbitrary ; any comparable rule will leave the behaviour unchanged . moreover , if the distribution of @xmath6 has no atom at @xmath7 , we could instead take the reflection rule to be the same on all boundaries without changing the characteristics of the process . we obtain the continuous - time stochastic process @xmath57 with state - space @xmath58 , which we call the _ stochastic billiard process _ , from the collisions process @xmath35 , essentially by interpolation , as follows . suppose @xmath59 . define the successive _ collision times _ of the particle by @xmath60 , and if @xmath61 @xmath62 where @xmath63 is euclidean distance on @xmath64 and @xmath65 ; if @xmath43 set @xmath66 . then for @xmath67 define @xmath68 so that @xmath69 for any @xmath67 . then for any @xmath70 we define @xmath71 as follows : @xmath72 this defines the process @xmath73 , and in particular @xmath35 is embedded in @xmath73 via @xmath74 , @xmath75 ( where @xmath76 ) . when @xmath77 , write @xmath78 in coordinates . a realization of the sequence @xmath79 therefore specifies the processes @xmath35 and @xmath73 via the construction just described . we now describe some further assumptions that we make on the function @xmath16 that specifies the region @xmath3 and on the distribution of the angle @xmath6 . suppose that the random variable @xmath6 for the angle of reflection satisfies for some @xmath80 @xmath81 = 1 , \text { and } \pr [ \alpha \geq x ] = \pr [ \alpha \leq -x ] \quad ( x \in \r ) , \ ] ] so that @xmath6 is bounded strictly away from @xmath82 and the distribution of @xmath6 is symmetric around @xmath7 . of special nature is the degenerate case where @xmath83=1 $ ] ; this leads to a deterministic evolution ( i.e. , reflection always occurs along the normals ) . we will shortly introduce some assumptions on the function @xmath16 defining the domain @xmath3 . if @xmath11 , @xmath3 is a right - cone , and at any reflection from @xmath5 there is positive probability ( in fact , @xmath84 \geq 1/2 $ ] under condition ( [ alpha ] ) ) that the process @xmath35 ( hence @xmath73 ) will go to @xmath36 : this is an obvious case of transience . moreover , if @xmath6 is degenerate , then it is clear that for @xmath16 strictly monotone @xmath85 if and only if @xmath86 ; so transience is possible even if @xmath87 grows arbitrarily slowly , at least in this deterministic case . on the other hand , if @xmath88 , we have that if @xmath6 is non - degenerate then @xmath89 performs a one - dimensional random walk on a half - line with zero drift and uniformly bounded jumps : so in this case @xmath35 will be null - recurrent ( loosely speaking for the moment ) . on another hand , it is at least intuitively plausible that if @xmath87 decreases sufficiently fast , @xmath35 will be positive - recurrent . in the present paper we will interpolate between these three situations ; formally we assume : * @xmath2 is monotonic , and thrice - differentiable . moreover , there exists @xmath90 for which , as @xmath91 , @xmath92 \frac{g(x)}{x } \\ g''(x ) & = [ \gamma ( \gamma-1 ) + o(1 ) ] \frac{g(x)}{x^2}\\ examples of functions satisfying ( a1 ) include @xmath93 or @xmath94 , where @xmath95 . then in particular we will be interested in the two cases where in addition to ( a1 ) either : @xmath96 or : @xmath97 given ( a1 ) , a necessary condition for ( [ h1 ] ) is that @xmath98 ; a sufficient condition for ( [ h2 ] ) is @xmath18 . the conditions on the derivatives of @xmath16 imposed by ( a1 ) are natural smoothness constraints that are necessary for our purposes . a particular case that we will be interested in is where for @xmath15 , @xmath14 for @xmath99 . then @xmath16 satisfies ( a1 ) and moreover satisfies ( [ h1 ] ) , ( [ h2 ] ) for @xmath17 , @xmath18 respectively . when @xmath100 , elementary geometry , using the fact that @xmath6 is bounded strictly away from @xmath82 , implies that @xmath35 ( and hence @xmath73 ) will almost surely never jump directly to @xmath36 , if @xmath101 is large enough ( see lemma [ cota_delta ] below ) . thus for the remainder of the paper we can take the state - space of @xmath35 to be @xmath5 and that of @xmath73 to be @xmath102 . in the next section we state our main results . then in section [ open ] we mention some open problems , and outline the remainder of the paper . recall the parameter @xmath26 . set @xmath103 here and throughout the paper we use the convention that @xmath104 . we will say that the process @xmath73 is recurrent if @xmath105=1 $ ] and transient otherwise ; if recurrent then it is positive- or null - recurrent according to whether @xmath106 < \infty$ ] or not . similarly for the process @xmath35 , but with @xmath107 instead of @xmath108 . by the construction of our processes , we have that @xmath109 thus the classification of recurrence or transience transfers directly between @xmath35 and @xmath73 . since @xmath110 , we have that the nonnegative random variable @xmath111 is bounded above by @xmath112 . thus @xmath113 < \infty$ ] , and @xmath114 = 0 $ ] if and only if @xmath115=1 $ ] . as an example , in the case where for @xmath80 , @xmath6 is uniformly distributed on the interval @xmath116 , it is the case that @xmath113 = \frac{{{\mathop{\rm tg}}}\alpha_0}{\alpha_0 } - 1 > 0 $ ] . our first result covers the recurrence / transience classification for our two processes @xmath35 and @xmath73 for the case of a growing tube . positive - recurrence is clearly ruled out in this case ( as our processes dominate the zero - drift process in a strip ) . so here the first important issue is that of transience versus null - recurrence . define @xmath117}{1 + 2\exp[{{\mathop{\rm tg}}}^2\alpha ] } , \ ] ] so that @xmath118 and @xmath119 if and only if @xmath120=1 $ ] . [ gen_g ] suppose that the random variable @xmath6 satisfies ( [ alpha ] ) and that for @xmath98 , @xmath16 satisfies ( a1 ) and ( [ h1 ] ) . then there exists @xmath121 such that for all @xmath122 and @xmath123 with @xmath124 : * @xmath35 , @xmath73 are transient if @xmath125 ; and * @xmath35 , @xmath73 are null - recurrent if @xmath126 . in particular , since @xmath127 , theorem [ gen_g ] says that for @xmath19 , the processes @xmath35 and @xmath73 are always transient , regardless of the distribution of @xmath6 . a special case of theorem [ gen_g ] is given by @xmath128 for some @xmath129 ; then ( a1 ) holds with @xmath25 and ( [ h1 ] ) holds . in this case it follows from theorem [ gen_g ] that @xmath35 is transient if @xmath115=1 $ ] , otherwise it is null - recurrent . that is , the introduction of _ randomness _ ensures that the process returns to a neighbourhood of the origin infinitely often ( with probability @xmath130 ) , whereas in the deterministic case the process is transient . the same remark applies for other functions @xmath16 that grow as @xmath131 . next we deal with the case of a narrowing tube , when @xmath16 is decreasing . now our processes are dominated by the zero - drift random walk in the strip , so transience is impossible . thus recurrence is evident , and the question of interest in this case is which moments of @xmath132 exist . the following result is for the collisions process @xmath35 . [ dim2_erg ] suppose @xmath6 satisfies ( [ alpha ] ) and that for @xmath133 , @xmath16 satisfies ( a1 ) and ( [ h2 ] ) . then there exists @xmath121 such that for all @xmath122 and @xmath134 with @xmath135 : * @xmath35 is positive - recurrent if @xmath136 $ ] ; and * @xmath35 is null - recurrent if @xmath137 $ ] . it follows from lemma [ compare](ii ) below that the continuous - time process @xmath73 is also positive - recurrent under the conditions of theorem [ dim2_erg](i ) . again , theorem [ dim2_erg ] shows that for a sub - polynomial function such as @xmath138 , a non - degenerate @xmath6 leads to null - recurrence . theorem [ dim2 ] below deals with the important special case of @xmath139 , @xmath17 or @xmath18 . in particular , when @xmath6 is non - degenerate , it covers the two critical cases omitted in theorems [ gen_g ] and [ dim2_erg ] ; they are null - recurrent . [ dim2 ] suppose @xmath6 satisfies ( [ alpha ] ) and is not degenerate , and that @xmath139 , @xmath15 . * suppose @xmath140 . then there exists @xmath121 such that for all @xmath122 and @xmath134 with @xmath135 , @xmath35 is null - recurrent . * suppose @xmath141 < 0 $ ] . then there exists @xmath121 such that for all @xmath122 and @xmath134 with @xmath135 , @xmath35 is null - recurrent . we now turn to our results on almost - sure bounds for our processes , that is , how far from the origin the particle will typically be in both the discrete- and continuous - time processes . first we give a basic statement that says that such questions are non - trivial . [ unbounded ] suppose that @xmath6 satisfies ( [ alpha ] ) . suppose that @xmath16 satisifies ( a1 ) and either ( i ) @xmath16 satisfies ( [ h1 ] ) ; or ( ii ) @xmath16 satisfies ( [ h2 ] ) and @xmath6 is non - degenerate . then for @xmath101 sufficiently large and any @xmath123 with @xmath142 , a.s . , @xmath143 if @xmath16 satisfies ( [ h2 ] ) , and @xmath144 is not large enough in this last result , then it is clear that for some distributions of @xmath6 ( with small bound @xmath145 ) the particle may get ` trapped ' in a bounded set . this case is not of interest for us here . we want to quantify the result of proposition [ unbounded ] . again we can proceed more generally , but our results are clearer if we take @xmath14 @xmath146 ) from now on . under the more general assumption ( a1 ) , we believe that our techniques still apply , with some modifications , and that the polynomial exponents in the theorems below remain valid . our first almost - sure bounds are for the collisions process in the case @xmath17 : [ time1 ] suppose @xmath6 satisfies ( [ alpha ] ) and @xmath14 with @xmath17 . suppose that @xmath147 for @xmath101 sufficiently large . for any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath149 moreover , if @xmath150 ( so that by theorem [ gen_g ] @xmath35 is transient ) , then there exists @xmath151 such that a.s . , for all but finitely many @xmath75 , @xmath152 theorem [ time1 ] shows that we have polynomial behaviour , but that on the time - scale of the collisions process @xmath35 , the speed @xmath153 can be very large for @xmath154 close to @xmath130 . the next result gives the corresponding bounds for the stochastic billiard process @xmath73 . on this time - scale , the behaviour is very different , since the particle can spend much more time between collisions . [ time ] suppose @xmath6 satisfies ( [ alpha ] ) and @xmath155 for @xmath17 . suppose that @xmath156 for @xmath101 sufficiently large . for any @xmath148 , a.s . , for all @xmath157 large enough , @xmath158 moreover , if @xmath150 ( so that by theorem [ gen_g ] @xmath73 is transient ) , then there exists @xmath151 such that a.s . , for all @xmath8 large enough , @xmath159 in particular , in the ( perhaps most interesting ) transient case , ( [ time2c ] ) shows that the asymptotic speed for the particle is zero , i.e. , @xmath160 a.s . , but the motion is super - diffusive . moreover , as @xmath161 , we approach linear growth , while as @xmath162 , we approach diffusive growth , as expected in view of the remarks just above the statement of ( a1 ) . ( [ time2b ] ) also demonstrates super - diffusive growth , even in the recurrent case . now we consider the case where @xmath139 for @xmath18 . if @xmath163 $ ] , we use the notation @xmath164}{(1 - 2\gamma ) \exp [ { { \mathop{\rm tg}}}^2 \alpha ] -\gamma } , \ ] ] so that @xmath165 . the next result deals with the collisions process @xmath35 . [ time2 ] suppose @xmath6 satisfies ( [ alpha ] ) and is non - degenerate , and @xmath14 for @xmath166 . suppose that @xmath156 for @xmath101 sufficiently large . * suppose that @xmath167 $ ] . then for any @xmath148 , a.s . , for all but finitely many @xmath75 , ( [ time1a ] ) and ( [ time1b ] ) hold . * suppose that @xmath168 $ ] . then for any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath169 now we state the corresponding result for the stochastic billiard process @xmath73 . [ time3 ] suppose @xmath6 satisfies ( [ alpha ] ) and is non - degenerate , and @xmath14 for @xmath166 . suppose that @xmath170 for @xmath101 sufficiently large . * suppose that @xmath167 $ ] . then for any @xmath148 , a.s . , for all @xmath157 large enough , ( [ time2b ] ) holds . * suppose that @xmath168 $ ] . then for any @xmath148 , a.s . , for all @xmath157 large enough , @xmath171 theorems [ time2](ii ) and [ time3](ii ) show that even in the positive - recurrent case , the behaviour is essentially polynomial ( i.e. , heavy - tailed ) in nature . moreover , the exponents depend explicitly on the reflection distribution @xmath6 , unlike in the other cases . we briefly mention some open problems for the model studied here . of interest is the weak ( distributional ) limiting behaviour of the horizontal components of @xmath35 , @xmath73 . we have from lemma [ zetalem3 ] and corollary [ cor1 ] below together with theorem 4.1 in @xcite that lamperti s invariance principle ( [ invar ] ) below will hold with @xmath172 and @xmath173 , 1)$ ] , @xmath6 non - degenerate , provided that lamperti s ` condition ( c ) ' from @xcite is satisfied . if lamperti s ` condition ( c ) ' were verified , this would then imply that for @xmath174 , 1)$ ] and @xmath6 non - degenerate , we would have the weak invariance principle as @xmath175 : @xmath176 here @xmath177 is a diffusion process on @xmath178 with kolmogorov backwards equation @xmath179 where @xmath180 ) , \textrm { and } b = 4 ( 1-\gamma)^2 \exp [ { { \mathop{\rm tg}}}^2 \alpha ] .\ ] ] a more challenging problem would be to determine whether any corresponding weak limit theory holds for the continuous - time process @xmath181 ( for @xmath182 $ ] , say ) . it may also be of interest to relax some of our assumptions , such as those on the reflection distribution . for instance , one might consider distributions for @xmath6 with support on all of @xmath183 $ ] ; there are several ` natural ' distributions relevant in this case @xcite . the techniques of the present paper require the assumption that @xmath6 be bounded away from @xmath82 . indeed , if @xmath6 can take value @xmath184 , say , with positive probability , the particle will eventually jump to @xmath36 in the case of a tube of nondecreasing width . on the other hand , if @xmath6 has distribution on @xmath183 $ ] with sufficiently light tails at the endpoints ( such that in particular @xmath114 < \infty$ ] , say ) , it may be possible to modify the techniques in the present paper to obtain similar results ; one would need to extend various results from processes with jumps that are bounded to processes with jumps satisfying higher - order moment assumptions , for example . the structure of the remainder of the paper is as follows . in section [ speeds ] we present and prove general semimartingale criteria ( of some independent interest ) on almost - sure bounds for one - dimensional stochastic processes . in section [ lamperti ] we make use of the results in section [ speeds ] , and give some results for the so - called lamperti problem , which are of interest in their own right as well as being a crucial element in the proofs of our main results . in section [ billiards ] we prove the results stated in section [ mainres ] on the processes @xmath185 . in this section we present and prove general semimartingale criteria for obtaining upper and lower almost sure bounds for discrete - time stochastic processes on the half - line . these results will provide some of our main tools for the study of processes with asymptotically zero mean drifts in section [ lamperti ] below , but for the present section we work in some generality . let @xmath186 be a filtration on a probability space @xmath187 . let @xmath188 be a discrete - time @xmath189-adapted stochastic process taking values in @xmath178 . suppose that @xmath190 = 1 $ ] for some @xmath191 . the types of process to which our criteria can be applied are quite general . for instance , due to the semimartingale nature of the results , we do not require that @xmath192 be a markov process . this fact is particularly useful if @xmath192 is a process of norms @xmath193 ; even if @xmath194 is markov , @xmath195 will not be , in general . we anticipate that the general results in this section are widely applicable . moreover , continuous - time processes may be treated via embedded discrete - time processes . one condition that we need for some of our results is that jumps of the process @xmath192 are uniformly bounded _ above_. elements of the proofs in the present section extend ideas used in @xcite and @xcite . our conditions will involve the existence of suitable functions @xmath196 such that the process @xmath197 satisfies an appropriate ` drift ' condition . our criteria will involve only first - order conditions ( i.e. expectations ) ; no variance results are required . our results will yield almost sure bounds for @xmath198 ( and hence @xmath199 ) in terms of the function @xmath196 and simple functions @xmath200 that control our bounds . the functions @xmath200 will belong to classes of eventually increasing functions defined as follows . we say function @xmath201 satisfies condition ( c1 ) and @xmath202 satisfies condition ( c2 ) if : * @xmath203 as @xmath204 , there exists @xmath205 such that @xmath206 is increasing for all @xmath207 , and @xmath208 ; * @xmath209 as @xmath204 , there exists @xmath210 such that @xmath211 is increasing for all @xmath212 , and @xmath213 . note that the summability condition in ( c2 ) is equivalent to the condition that @xmath214 for some ( hence all ) @xmath215 . throughout this section we interpret @xmath216 as @xmath217 . with this convention , condition ( c1 ) is satisfied , for example , by @xmath218 or @xmath219 , where @xmath148 , and ( c2 ) is satisfied , for example , by @xmath220 or @xmath221 , @xmath148 . the first result in this section is a submartingale criterion for an upper bound . [ thm4a ] let @xmath222 be a discrete - time @xmath189-adapted stochastic process taking values in @xmath178 . suppose that there exists @xmath223 a nondecreasing function such that , @xmath224 \geq 0 ~\textrm{a.s . } , \end{aligned}\ ] ] for all @xmath75 . also suppose that there exists @xmath225 such that for all @xmath226 @xmath227 \leq b n .\ ] ] define the nondecreasing function @xmath228 for @xmath229 by @xmath230 let @xmath231 satisfy ( c1 ) . then , a.s . , for all but finitely many @xmath75 , @xmath232 a sufficient condition for ( [ jj2 ] ) is clearly that there exists @xmath233 such that @xmath234 \leq b ' ~\textrm{a.s.},\ ] ] for all @xmath75 . we next present a variant of theorem [ thm4 ] , which relaxes the condition ( [ jj1 ] ) at the expense of this slightly stronger version of ( [ jj2 ] ) . theorem [ thm4a ] holds with conditions ( [ jj1 ] ) and ( [ jj2 ] ) replaced by the lone condition ( [ jj3 ] ) . [ thm4 ] for the proof of theorem [ thm4 ] , we need the following : [ lem13 ] let @xmath235 be an @xmath236-adapted process on @xmath178 and @xmath237 such that @xmath238 = 1 $ ] and for some @xmath225 and all @xmath75 @xmath239 \leq b ~{\rm a.s.}. \end{aligned}\ ] ] then for any @xmath240 and any @xmath226 @xmath241 \leq ( b n + z_0 ) x^{-1}.\end{aligned}\ ] ] * proof . * similarly to doob s decomposition ( see e.g. @xcite , p. 120 ) , set @xmath242 , and for @xmath243 let @xmath244 , where @xmath245 $ ] , @xmath246 , @xmath247 . then @xmath248=\exp [ z_{m+1}- z_m+a_m^{- } \mid { { \mathcal f}}_m ] = a_m+a_m^-=a_m^+\in [ 0,b]\ ] ] so that @xmath249 is a nonnegative @xmath250-submartingale with @xmath251 for all @xmath252 , and @xmath253\le w_0+bn = z_0+bn$ ] . hence by doob s submartingale inequality ( see e.g. @xcite , p. 137 ) @xmath254 \leq \pr \left [ \max_{0 \leq m \leq n } w_m \geq x \right ] \leq x^{-1 } \exp [ w_n ] \leq ( b n + z_0 ) x^{-1 } , \ ] ] as required . @xmath255 + * proof of theorems [ thm4a ] and [ thm4 ] . * first we prove theorem [ thm4a ] . since , by ( [ jj1 ] ) , @xmath256 is a nonnegative submartingale , doob s submartingale inequality implies that , for any @xmath226 , @xmath257 \leq ( a(n))^{-1 } \exp [ f(y_n ) ] \leq b n ( a(n))^{-1},\ ] ] using ( [ jj2 ] ) . also , for @xmath258 , @xmath259 = \pr \left [ f \left ( \max_{0 \leq m \leq n } y_m \right ) \geq a(n ) \right ] , \ ] ] since @xmath196 is nondecreasing . with @xmath228 defined by ( [ finv ] ) , let @xmath260 denote the event @xmath261 then since @xmath262 implies @xmath263 , we obtain from ( [ 1002dz ] ) and ( [ 1002cz ] ) that for all @xmath226 @xmath264 \leq \pr \left [ f \left ( \max_{0 \leq m \leq n } y_m \right ) \geq a(n ) \right ] \leq b n ( a(n))^{-1 } .\ ] ] now @xmath265 hence by ( [ lll1 ] ) and ( [ lll2 ] ) , along the subsequence @xmath266 for @xmath267 , ( c1 ) and the borel - cantelli lemma imply that , a.s . , the event @xmath260 occurs only finitely often , and in particular there exists @xmath268 such that for all @xmath269 @xmath270 every @xmath226 sufficiently large has @xmath271 for some @xmath272 ; then , a.s . , @xmath273 for all but finitely many @xmath274 . now since @xmath275 and @xmath228 is nondecreasing , ( [ 001x ] ) follows . to obtain theorem [ thm4 ] , in the previous argument we replace ( [ 1002dz ] ) by an application of lemma [ lem13 ] with @xmath276 and @xmath277 . @xmath255 + we now work towards obtaining a lower bound for @xmath278 . for @xmath279 let @xmath280 denote the _ first passage time _ of @xmath281 for @xmath192 , that is @xmath282 then @xmath283 is a nondecreasing sequence of stopping times for the process @xmath192 ; under the condition @xmath284 a.s . , @xmath285 a.s . for every @xmath281 . for @xmath279 and all @xmath75 , let @xmath286 ( where @xmath287 ) , the process stopped at @xmath280 ; then @xmath288 if @xmath289 and @xmath290 for @xmath291 . we have the following result , which is a ` reverse foster s criterion ' ( compare theorem 2.1.1 in @xcite ) . [ thm5a ] let @xmath222 be a discrete - time @xmath189-adapted stochastic process taking values in @xmath178 , such that for some @xmath292 @xmath293 = 1,\ ] ] for all @xmath75 . fix @xmath279 and suppose that there exists @xmath295 a nondecreasing function for which , for some @xmath296 , @xmath297 \geq \eps \1_{\ { \sigma_\ell > n \ } } ~\textrm{a.s.}\ ] ] for all @xmath75 . then @xmath298 \leq \frac{1}{\eps } f ( \ell + b ) .\ ] ] * proof . * let @xmath279 . taking expectations in ( [ oo1 ] ) we have for all @xmath299 @xmath300 - \exp [ f(y^\ell_m ) ] \geq \eps \pr [ \sigma_\ell > m ] .\ ] ] summing for @xmath252 from @xmath7 up to @xmath274 we obtain @xmath301 - f(y^\ell_0 ) \geq \eps \sum_{m=0}^{n } \pr [ \sigma_\ell > m ] .\ ] ] since @xmath302 we obtain @xmath303 = \lim_{n \to \infty } \sum_{m=0}^{n } \pr [ \sigma_\ell > m ] \leq \frac{1}{\eps } \limsup_{n \to \infty } \exp [ f(y^\ell_{n+1 } ) ] \leq \frac{1}{\eps } f ( \ell + b),\ ] ] using the fact that , by ( [ bndd ] ) , @xmath304 for all @xmath75 since @xmath196 is nondecreasing . @xmath255 + now we have the following lower bound : [ thm6b ] let @xmath222 be a discrete - time @xmath189-adapted stochastic process taking values in @xmath178 , such that condition ( [ bndd ] ) holds . suppose that there exists @xmath295 a nondecreasing function and @xmath296 for which @xmath305 \geq \eps ~\textrm{a.s.}\ ] ] for all @xmath75 . let @xmath306 satisfy ( c2 ) . for @xmath307 define the function @xmath308 by @xmath309 then , a.s . , for all but finitely many @xmath75 , @xmath310 * proof . * first note that by ( c2 ) and the fact that @xmath196 is nondecreasing , we have that @xmath311 is nondecreasing for all @xmath312 sufficiently large . fix @xmath313 . by markov s inequality , @xmath314 ] \leq ( v(\ell))^{-1 } .\ ] ] given ( [ lll ] ) and ( c2 ) , the borel - cantelli lemma implies that , a.s . , @xmath315 $ ] for only finitely many @xmath316 . moreover , given that @xmath196 satisfies ( [ oo2 ] ) , we have that ( [ oo1 ] ) holds for all @xmath279 and all @xmath75 . then lemma [ thm5a ] with ( [ bndd ] ) in this case implies that @xmath317 \leq \eps^{-1 } f(\ell+b)$ ] for all @xmath281 . thus we have that a.s . , for some @xmath318 and all @xmath269 , @xmath319 . hence with the definition of @xmath308 at ( [ fdef3 ] ) we have , a.s . , for all @xmath281 sufficiently large , @xmath320 since @xmath321 < \infty$ ] for all @xmath281 , we have that @xmath285 a.s . for all @xmath281 . moreover , the jumps bound ( [ bndd ] ) implies that , for all @xmath322 , @xmath323 a.s . , so that @xmath324 a.s .. thus , a.s . , every @xmath75 satisfies @xmath325 for some @xmath326 . then , a.s . , @xmath327 for all @xmath274 sufficiently large , since @xmath328 . then ( [ ee1xx ] ) and ( [ dddxx ] ) imply that , a.s . , for all but finitely many @xmath75 , @xmath329 since @xmath330 and @xmath331 is nondecreasing for @xmath274 sufficiently large . now taking a sequence of values for @xmath332 converging down to @xmath130 we obtain ( [ 003xx ] ) . @xmath255 in this section we give upper and lower almost - sure bounds for the so - called lamperti problem of a stochastic process on the half - line with mean drift asymptotically zero . we will need these results for our almost - sure bounds for the stochastic billiard model ; our results for the lamperti problem appear to be new in the generality given here , and in some cases even for a nearest - neighbour random walk . let @xmath333 be a filtration on a probability space @xmath334 . let @xmath335 be a discrete - time time - homogeneous stochastic process adapted to @xmath333 and taking values in an unbounded subset @xmath336 of @xmath178 . we suppose that jumps of @xmath337 are uniformly bounded , that is there exists @xmath225 such that for all @xmath75 and all @xmath338 @xmath339 = 0 , ~\textrm{a.s.}.\ ] ] under the jumps condition ( [ jumps ] ) , the jump - moment functions @xmath340 $ ] and @xmath341 $ ] given for @xmath342 by @xmath343 , \quad ( n \in \z^+)\ ] ] are well - defined ; in particular @xmath344 is bounded above . our basic assumption for this section will be ( a2 ) below . * let @xmath337 be a discrete - time stochastic process on @xmath178 satisfying ( [ jumps ] ) with @xmath345 as given by ( [ mus ] ) . for some of the results in this section we also assume that there exists @xmath346 such that , for all @xmath338 @xmath347 we will sometimes make the further assumption that @xmath348 = 1 .\ ] ] the model that we concentrate on is a particular case of the so - called lamperti problem ( see @xcite ) of a stochastic process on @xmath178 with mean drift asymptotically zero : that is @xmath349 as @xmath91 . our results do not actually assume @xmath349 , but it is in this case ( in fact , in the case @xmath350 ) that they are of most interest . as mentioned by lamperti @xcite , and as is also true for the results in this section , only the first two moments @xmath345 of the jump distribution are important : there is some form of ` invariance principle ' at work . we will be mainly concerned with the case that , from the point of view of recurrence , turns out to be critical ; that is where @xmath351 remains bounded away from zero and from infinity . in the particular case of a nearest - neighbour random walk , where @xmath337 is supported on @xmath352 , the problem reduces to that of a simple random walk with asymptotically zero perturbation : that model was studied by harris @xcite and by hodges and rosenblatt @xcite , and is amenable to special methods for so - called birth - and - death chains . thus many results are present in the literature for the nearest - neighbour case : in section [ literature ] below we briefly mention some of these results and their relation to the results given in this section . for the applications in the present paper , however , we can not use these nearest - neighbour results . thus we need to prove results in the general lamperti setting . in fact , as we will point out below , some of the results that we give in the present section seem to be new even for the nearest - neighbour situation . thus the results that we state in this section are of independent interest . as well as being of interest in their own right , stochastic processes on the half - line with mean drift asymptotically zero are important for the study of multidimensional processes by the method of lyapunov - type functions ( see e.g. @xcite ) . for example , if @xmath235 is a zero - drift process with bounded jumps on @xmath353 , @xmath354 , the process @xmath355 is supported on the half - line and has mean drift asymptotically zero . clearly , the process @xmath356 will not be nearest - neighbour ; thus the generality of results like lamperti s @xcite and those in the present section is valuable . fix @xmath357 , which will need to be large for some of our results . we say that @xmath337 is _ recurrent _ or _ transient _ according to whether the return time @xmath358 is almost surely finite or not ; in the former case we distinguish positive- and null - recurrence according to whether the return time has finite or infinite expectation . the recurrence and transience properties of @xmath337 were studied by lamperti , who proved the following result ( see theorems 3.1 and 3.2 of @xcite with theorem 2.1 of @xcite ; also theorem 3 of @xcite for a finer result ) . note that all of the conditions that we state in this section involving evaluating @xmath359 , @xmath344 only need apply for @xmath338 . [ lmpti ] @xcite suppose that ( a2 ) , ( [ ass2 ] ) , and ( [ limsup ] ) hold . * if for all @xmath360 , @xmath361 , then @xmath337 is null - recurrent for any @xmath362 . * suppose that there exists @xmath363 such that for all @xmath360 @xmath364 then @xmath337 is transient for any @xmath365 . * suppose that there exists @xmath363 such that for all @xmath360 , @xmath366 . then @xmath337 is positive - recurrent for any @xmath362 . for our almost - sure lower bounds , we impose an additional ` reflection ' condition that ensures that we can avoid getting trapped in a bounded set . condition ( a3 ) below is the most appropriate way of doing this for our applications to the stochastic billiard model , but is clearly stronger than is necessary for the results in the present section . * given @xmath367 , if a jump would take @xmath368 we replace the jump with @xmath369 instead . we now state our results on almost - sure bounds . in the general setting of the present section , the only almost - sure bound result that we could find in the literature is also due to lamperti @xcite ; see proposition [ lmpti2 ] in our discussion of the literature below . lamperti s bound is a weaker version of our result theorem [ prop1](i ) below , while theorem [ prop1](ii ) gives a complementary lower bound . [ prop1 ] suppose that assumption ( a2 ) holds . * suppose that there exists @xmath370 such that for all @xmath371 @xmath372 then for any @xmath373 , for any @xmath148 , a.s . , for all but finitely many @xmath374 , @xmath375 * suppose that ( a3 ) holds , and that there exists @xmath376 such that for all @xmath377 @xmath378 then for any @xmath379 , for any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath380 in the transient case , we prove the following lower bound that strengthens , in some sense , theorem [ prop1](ii ) in this case . [ lamp_low ] suppose that ( a2 ) and ( a3 ) hold . if ( [ mu_trans ] ) holds for some @xmath363 and all @xmath381 , then there exists @xmath151 such that , for any @xmath382 , a.s . , for all but finitely many @xmath75 , @xmath383 we could find no reference for a result like theorem [ lamp_low ] , even in the nearest - neighbour case . we prove theorem [ lamp_low ] in section [ lamplowprf ] ; it may be possible to extract bounds for the exponent @xmath384 in the logarithmic term in terms of the constants @xmath385 by keeping track of the constants in our proofs . note that the transient critical lamperti problem for which ( a2 ) and ( a3 ) hold , and @xmath386 for some @xmath387 and all @xmath312 large enough , satisfies the conditions of each of the results in theorems [ prop1 ] and [ lamp_low ] . the following discussion suggests that without taking into account finer behaviour ( such as the smallest value of @xmath363 for which ( [ mu_trans ] ) is satisfied ) , one can not expect to improve upon theorem [ lamp_low ] . let @xmath388 be the symmetric simple random walk on @xmath353 , @xmath354 , so that for @xmath389 , @xmath390 = ( 2d)^{-1}$ ] if and only if @xmath391 . then elementary calculations show that @xmath392 - \exp [ ( \| s^d_{n+1 } \| - \| s^d_n \|)^2 \mid s^d_n = x ] \to 1 - \frac{2}{d}\ ] ] as @xmath393 ; thus the process @xmath394 is in precisely the critical lamperti situation , and for @xmath395 satisfies ( [ mu_trans ] ) . for @xmath395 , a classical result of dvoretzky and erds @xcite says that for any @xmath148 , a.s . , @xmath396 for all but finitely many @xmath75 , and that this bound is sharp in that for @xmath397 the inequality ( [ delower ] ) fails infinitely often with probability @xmath130 . in particular , it is at least informally reasonable to argue that by letting @xmath398 in ( [ delower ] ) one might not expect to improve in general upon the arbitrary logarithmic factor in theorem [ lamp_low ] . now we state results on almost - sure bounds that we will need for our stochastic billiard model when @xmath139 with @xmath18 . again these results seem to be new in the generality given here . theorem [ prop6 ] below is further demonstration of the phenomenon of _ polynomial ergodicity _ for the critical lamperti problem , as also evidenced by results of @xcite in the context of stationary measures . [ prop6 ] suppose that ( a2 ) and ( [ ass2 ] ) hold . * suppose that there exists @xmath399 such that for all @xmath312 sufficiently large @xmath400 then for any @xmath373 , any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath401 * suppose that ( a3 ) holds , and that there exists @xmath402 such that for all @xmath312 sufficiently large @xmath403 then there exists @xmath404 such that for any @xmath379 , for any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath405 note that the @xmath406 case of theorem [ prop6](ii ) gives a weaker form of the lower bound in theorem [ prop1](ii ) under a somewhat weaker condition . before we prove the results stated in section [ lampres ] , we briefly discuss the existing literature related to the present section . this is done in section [ literature ] below . then we give the proofs of theorems [ prop1 ] and [ prop6 ] in section [ lampprf ] and theorem [ lamp_low ] in section [ lamplowprf ] . in the general setting of the so - called lamperti problem , when ( a2 ) holds , there seem to be few almost - sure bounds known . the next result , due to lamperti himself ( see theorems 2.1 , 2.2 , 4.2 , and 5.1 in @xcite ) , includes an almost - sure upper bound in a particular case where the conditions of ( i ) or ( ii ) in proposition [ lmpti ] hold . [ lmpti2 ] @xcite suppose that ( a2 ) holds , and that for any finite interval @xmath407 @xmath408 = 0.\ ] ] suppose that for @xmath409 @xmath410 then for any @xmath373 , for any @xmath148 , a.s . , for all but finitely many @xmath75 @xmath411 in addition , suppose that ( [ null ] ) holds uniformly in @xmath373 . then as @xmath175 the following invariance principle applies : @xmath412 where @xmath413 is a diffusion process on @xmath178 with kolmogorov backwards equation @xmath414 ( see section 3 in @xcite for details ) . thus theorem [ prop1](i ) improves upon the bound in ( [ lampbnd ] ) . note that the condition ( [ null ] ) , which does not distinguish between null - recurrence and transience , implies ( [ limsup ] ) . a special case of the lamperti problem on @xmath178 is the case where the process @xmath337 is supported on @xmath352 and only nearest - neighbour jumps are allowed . this special case has received much more attention than the general case described in section [ lampres ] above ; now we briefly describe known results in the nearest - neighbour case . when they exist , these nearest - neighbour results are sharper than the general results that we give in section [ lampres ] . thus it may be possible to sharpen the bounds in the results in section [ lampres ] . in the nearest - neighbour case , @xmath337 is sometimes known as a _ birth - and - death _ chain ( or birth - and - death random walk ) . precisely , suppose that there exists a sequence @xmath415 with @xmath416 such that for all @xmath417 and @xmath75 @xmath418 = 1 - \pr [ \eta_{n+1 } = x+1 \mid \eta_n = x ] = p_x,\ ] ] with reflection from @xmath7 governed by @xmath419 = 1 - \pr [ \eta_{n+1 } = 1 \mid \eta_n = 0 ] = p_0.\ ] ] ( in the literature , the slightly more general model where the walk is allowed to stay in the current position also appears . this introduces no new essential features however . ) such processes have been extensively studied in various contexts . they are often amenable to explicit computation ; one particularly fruitful approach is via orthogonal polynomials , dating back at least to karlin and mcgregor @xcite and subsequently employed for instance by voit @xcite . a recent reference is @xcite , to which we are indebted for bringing to our attention several of the papers cited in this section . the one - step mean drift of this walk is for @xmath417 @xmath420 = 1 -2p_x .\ ] ] we are in the lamperti situation if we assume @xmath421 . a case of particular interest is when @xmath416 satisfies @xmath422 for @xmath423 and @xmath424 . then with the notation at ( [ mus ] ) , @xmath425 and @xmath426 . in the nearest - neighbour case , partial versions of the recurrence result proposition [ lmpti ] were given in e.g. @xcite . when ( [ nnw ] ) holds , proposition [ lmpti ] implies that if @xmath427 , @xmath337 is null - recurrent , while if @xmath428 , @xmath337 is transient , positive - recurrent according to @xmath429 , @xmath430 . in the critical case @xmath431 , @xmath337 is transient if @xmath432 , null - recurrent if @xmath433 , and positive - recurrent if @xmath399 . almost - sure results are known for certain null - recurrent and transient situations . the following two propositions collect some known results of this type ; special cases were dealt with in @xcite . the next result follows from theorem 2.11 of @xcite ; it deals with the transient case with @xmath428 . ( there appears to be a typo in the statement of theorem 2.11 in @xcite . ) @xcite suppose that ( [ nnw ] ) holds for @xmath428 and @xmath434 . then for any @xmath373 , a.s . @xmath435 theorem 7.1 of lamperti @xcite gives a corresponding result for more general processes ( in the manner of section [ lampres ] ) , but with convergence in probability only . the following iterated logarithm result that applies in the @xmath436 , @xmath437 case of ( [ nnw ] ) follows from theorem 4(b ) of @xcite ( compare also theorem 1.3 of @xcite ) . @xcite suppose that there exist @xmath438 with @xmath439 such that for all @xmath312 sufficiently large @xmath440 then for any @xmath373 , a.s . @xmath441 apparently missing , for example , is the complete result in the case where ( [ nnw ] ) holds with @xmath431 and @xmath442 , for which the walk is null - recurrent ; a one - sided result is given in theorem 4(a ) of @xcite . also , there seem to be no ( sharp ) existing results in the critical positive - recurrent case where ( [ nnw ] ) holds with @xmath431 and @xmath443 . thus theorems [ prop1 ] , [ lamp_low ] , and [ prop6 ] add to known results even in the nearest - neighbour case . in order to prove theorems [ prop1 ] and [ prop6 ] , we apply the general results on almost - sure bounds for discrete - time stochastic processes given in section [ speeds ] . + * proof of theorem [ prop1 ] . * for @xmath371 and @xmath75 we have @xmath444 & = 2x \exp [ \eta_{n+1 } -\eta_n \mid \eta_n = x ] + \exp [ ( \eta_{n+1 } -\eta_n)^2 \mid \eta_n = x ] \nonumber\\ & = 2x \mu_1(x ) + \mu_2(x).\label{aa1}\end{aligned}\ ] ] under the conditions of part ( i ) of the theorem , the right - hand side of ( [ aa1 ] ) is uniformly bounded above . hence we can apply theorem [ thm4 ] , with @xmath445 , taking @xmath446 and @xmath447 . this proves part ( i ) . under the conditions of part ( ii ) of the theorem , we have that the right - hand side of ( [ aa1 ] ) is strictly positive for all @xmath377 . under condition ( a3 ) it suffices to consider the process on @xmath448 . hence we can apply theorem [ thm6b ] , with @xmath445 , taking @xmath446 and @xmath449 . this proves part ( ii ) . @xmath255 + now we prepare for the proof of theorem [ prop6 ] . we need two more lemmas , to identify suitable functions @xmath196 with which to apply the criteria of section [ speeds ] . [ lem5 ] suppose that ( a2 ) and ( [ ass2 ] ) hold . suppose that there exists @xmath450 such that ( [ c4 ] ) holds for all @xmath312 sufficiently large . then there exists @xmath370 such that for all @xmath371 @xmath451 \leq c.\ ] ] * proof . * it follows from taylor s theorem applied to the function @xmath452 that for @xmath453 as @xmath454 @xmath455 .\end{aligned}\ ] ] now conditioning on @xmath456 and setting @xmath457 , taking expectations in the last displayed equation gives @xmath458\\ & = \frac{x^{1+\kappa}}{\log ( 1+x ) } \left [ \frac{1+\kappa}{2x^2 } ( 2x \mu_1(x ) + \kappa \mu_2 ( x ) ) -\frac{1}{2x^2 \log ( 1+x ) } ( 2x \mu_1(x ) + ( 2\kappa+1 ) \mu_2(x ) + o(1 ) ) \right ] \\ & \leq \frac{x^{\kappa-1}}{(\log ( 1+x))^2 } \left [ -\frac{1}{2 } \mu_2 ( x ) + o(1 ) \right ] , \end{aligned}\ ] ] using ( [ c4 ] ) . then with ( [ ass2 ] ) this yields the result . @xmath255 [ lem6 ] suppose that ( a2 ) and ( [ ass2 ] ) hold . suppose that there exists @xmath402 such that ( [ c5 ] ) holds for all @xmath312 sufficiently large . then there exist @xmath404 , @xmath459 such that for all @xmath377 @xmath460 \geq \eps.\ ] ] * proof . * this time , it follows from taylor s theorem that for @xmath453 as @xmath454 @xmath461 .\end{aligned}\ ] ] now conditioning on @xmath456 and setting @xmath457 , taking expectations in the last displayed equation gives @xmath462\\ & = x^{1+\kappa } \left [ \frac{1+\kappa}{2x^2 } ( \log ( 1+x ) ) ( 2x \mu_1(x ) + \kappa \mu_2 ( x ) ) + \frac{1}{2x^2 } ( 2x \mu_1(x ) + ( 2\kappa+1 ) \mu_2(x ) + o(1 ) ) \right ] \\ & \geq x^{\kappa-1 } \left [ \frac{\kappa+1}{2 } \mu_2 ( x ) + o(1 ) \right ] \geq \eps > 0 , \end{aligned}\ ] ] for all @xmath312 large enough , using ( [ c5 ] ) , ( [ ass2 ] ) , and the fact that @xmath402 . @xmath255 + * proof of theorem [ prop6 ] . * by lemma [ lem5 ] , we can apply theorem [ thm4 ] with @xmath463 , @xmath447 , and @xmath445 to obtain part ( i ) of the theorem . on the other hand , by lemma [ lem6 ] we can apply theorem [ thm6b ] with @xmath464 , @xmath449 , and @xmath445 to obtain part ( ii ) of the theorem . @xmath255 suppose that ( [ mu_trans ] ) holds for some @xmath363 . our strategy for the proof of theorem [ lamp_low ] is to construct a scale on which the process @xmath337 is transient with mean drift that is positive and bounded uniformly away from @xmath7 . fix @xmath465 . denote @xmath466 and for @xmath467 define the intervals @xmath468,\ ] ] where @xmath469 is as in the jumps bound ( [ jumps ] ) ; then for @xmath470 sufficiently large , @xmath471 do not overlap . we look at the process @xmath337 at the moments at which it enters an interval @xmath472 different from the last one visited . we define the random times @xmath473 inductively as follows . set @xmath474 ; for @xmath475 , if @xmath476 , let @xmath477 . consider the embedded process @xmath478 . the conditions of theorem [ lamp_low ] imply those of theorem [ prop1](ii ) , so in particular ( [ limsup ] ) holds . this together with the jumps bound ( [ jumps ] ) implies that @xmath479 = 1\ ] ] for any @xmath467 , so that the process @xmath480 is well - defined and the random times @xmath481 are almost surely finite for each @xmath482 . denote @xmath483 , \quad(k \in \n).\ ] ] the next result says that the process @xmath480 has uniformly positive mean drift . [ scale - l ] under the conditions of theorem [ lamp_low ] , there exist @xmath465 and @xmath484 such that for all @xmath467 @xmath485 * proof . * let @xmath486 . taylor s theorem implies that as @xmath91 @xmath487 = x^{\lambda-2 } \left [ \lambda x \mu_1 ( x ) + \frac{\lambda(\lambda -1)}{2 } \mu_2 ( x ) + o ( x^{-1 } ) \right ] , \ ] ] using ( [ jumps ] ) and ( [ mus ] ) . then by ( [ mu_trans ] ) and ( [ jumps ] ) again , we have @xmath487 \leq x^{\lambda-2 } \left [ ( \lambda /2 ) ( \delta + \lambda b^2 ) + o ( x^{-1 } ) \right],\ ] ] where @xmath469 is the jumps bound in ( [ jumps ] ) . it follows that for @xmath488 @xmath487 \leq 0 , \ ] ] for all @xmath312 sufficiently large . hence for the stopping times @xmath481 it is also the case that @xmath489 \leq 0 , \ ] ] for all @xmath312 sufficiently large . then the supermartingale property ( [ superm ] ) implies that for @xmath470 large enough and all @xmath467 @xmath490 \leq [ ( 1+\beta)^r- b]^\lambda .\ ] ] also , we have that @xmath491 = & ~ p_r \exp [ \eta_{\ell_{k+1}}^\lambda \mid \eta_{\ell_k } \in i_r , \eta_{\ell_{k+1 } } \in i_{r+1 } ] \nonumber\\ & + ( 1-p_r ) \exp [ \eta_{\ell_{k+1}}^\lambda \mid \eta_{\ell_k } \in i_r , \eta_{\ell_{k+1 } } \in i_{r-1 } ] \nonumber\\ \geq & ~ p_r [ ( 1+\beta)^{r+1 } + b ] ^\lambda + ( 1-p_r ) [ ( 1+\beta)^{r-1 } + b]^\lambda .\end{aligned}\ ] ] combining ( [ bb1 ] ) and ( [ bb2 ] ) we see that one can choose @xmath465 , @xmath484 such that @xmath492 for all @xmath493 . + consider , for @xmath475 , @xmath494 then the process @xmath495 is a stochastic process on @xmath496 with nearest - neighbour transitions that tracks which interval @xmath472 the embedded process @xmath480 is in . with @xmath497 as defined by ( [ prdef ] ) , we have @xmath498 = 1- \pr [ z_{k+1 } = r-1 \mid z_k = r ] = p_r .\ ] ] let @xmath499 denote the time of the first visit of @xmath500 to @xmath467 , i.e. @xmath501 for @xmath467 , @xmath502 define @xmath503.\ ] ] in words , @xmath504 is the probability that the process @xmath500 hits @xmath493 in finite time , given that it starts at @xmath505 . the next result will enable us to show , loosely speaking , that the process @xmath500 leaves each state for good only shortly after its first visit . [ lastexit ] under the conditions of theorem [ lamp_low ] , there exists @xmath370 such that @xmath506 * proof . * to estimate the required hitting probability we introduce an auxiliary process . fix @xmath507 , which we will eventually take to be large . let @xmath467 . define the nonnegative process @xmath508 for @xmath509 by @xmath510 then we have for @xmath226 and @xmath509 @xmath511 = \exp \ { - c_0^{-1 } ( n - r ) \ } \exp [ \exp \ { - c_0^{-1 } ( z_{t+1 } - z_t ) \ } - 1 \mid z_t = n ] .\end{aligned}\ ] ] we will make use of the fact that there exists @xmath512 such that @xmath513 for all @xmath312 with @xmath514 . since @xmath500 has nearest - neighbour jumps , it follows that @xmath515 \nonumber\\ \leq & - c_0^{-1 } \exp [ z_{t+1 } - z_t \mid z_t = n ] + c_1 c_0^{-2 } \exp [ ( z_{t+1 } - z_t ) ^2 \mid z_t = n ] \nonumber\\ \leq & - c_0^{-1 } ( 2 p_n - 1 ) + c_1 c_0^{-2 } , \end{aligned}\ ] ] using ( [ zjumps ] ) and the fact that @xmath500 has nearest - neighbour jumps again . it then follows from ( [ uiop1 ] ) , ( [ uiop2 ] ) , and lemma [ scale - l ] that we can choose @xmath516 sufficiently large , not depending on @xmath493 , so that for any @xmath226 @xmath517 \leq 0 , \ ] ] for all @xmath509 . now let @xmath502 and take @xmath518 . recall the definition of the stopping time @xmath519 from ( [ kappaz ] ) . from ( [ ww ] ) we have that @xmath520 is a nonnegative supermartingale , so that @xmath521 exists a.s . , and @xmath522 \geq \exp [ w_{\kappa_r^z } ] $ ] . it follows that for any @xmath467 , @xmath502 @xmath523 = \exp \ { -c_0^{-1 } s \ } \geq \exp [ w_{\kappa_r^z } \1_{\ { \kappa_r^z < \infty\ } } ] = \gamma ( r , s ) , \ ] ] using ( [ gammadef ] ) . in particular , it follows that for some @xmath370 large enough @xmath524 completing the proof of the lemma . @xmath255 + to complete the proof of theorem [ lamp_low ] , we need to translate our results for the embedded process @xmath480 back to the underlying process @xmath337 . + * proof of theorem [ lamp_low ] . * consider the nearest - neighbour process @xmath525 on @xmath496 , as defined by ( [ zkdef ] ) . recall the definition of @xmath499 from ( [ kappaz ] ) . let @xmath526 denote the time of the _ last _ visit of @xmath500 to @xmath467 , i.e. @xmath527 similarly for the process @xmath337 set for @xmath528 @xmath529 with lemma [ lastexit ] , the borel - cantelli lemma implies that , a.s . , for only finitely many @xmath467 does @xmath500 return to @xmath493 after visiting @xmath530 . , for all but finitely many @xmath467 , @xmath531 set @xmath532 observe that by definition of the process @xmath500 we have that a.s . for some @xmath370 and all @xmath312 large enough @xmath533 by ( [ firstlast ] ) . again by the definition of @xmath500 , it now follows that a.s . for some @xmath534 @xmath535 for all @xmath312 large enough . the conditions of theorem [ lamp_low ] imply those of theorem [ prop1](ii ) . hence the lower bound in theorem [ prop1](ii ) applies , and so for any @xmath148 , a.s . , for all but finitely many @xmath75 @xmath536 it follows that for any @xmath148 , a.s . , for all but finitely many @xmath537 @xmath538 so by ( [ a1 ] ) and ( [ a2 ] ) there exist @xmath539 such that , a.s . , for all @xmath540 @xmath541 by the transience of @xmath337 , we have that a.s . @xmath542 for all but finitely many @xmath75 . hence a.s . , for all but finitely many @xmath75 we have @xmath543 using the jumps bound ( [ jumps ] ) for the final inequality . this proves the theorem . suppose that at time @xmath75 we have @xmath544 for @xmath545 , and then @xmath546 is reflected at angle @xmath6 to the normal . denote @xmath547 , the jump of the horizontal component of @xmath35 . also set @xmath548 , so @xmath549 . we now proceed to obtain estimates for @xmath550 and its moments . the next lemma gives an upper bound on @xmath550 that follows from the fact that for large enough @xmath312 our tube will be almost flat , while @xmath6 is bounded strictly away from @xmath82 . * fix @xmath80 . assuming ( a1 ) , we have that @xmath555 and hence @xmath556 as @xmath91 ; in particular we can choose @xmath101 large enough so that for all @xmath557 , @xmath558 and @xmath559 for some @xmath560 . by symmetry , it suffices to suppose that we start on the positive half of the curve , i.e. , at @xmath561 . we have that @xmath562 is bounded by @xmath563 first suppose ( [ h1 ] ) holds . then @xmath16 is nondecreasing , so @xmath564 . consider @xmath565 , represented as @xmath566 when @xmath567 in figure [ cota_delta_fig ] . the reflected ray at angle @xmath145 to the normal from @xmath561 has equation in @xmath568 given by @xmath569 take @xmath570 . as @xmath571 , we have @xmath572 for all @xmath312 large enough . consider the line @xmath573 this line intersects @xmath5 at @xmath574 and it intersects the reflected ray at angle @xmath145 to the normal from @xmath561 at @xmath575 with @xmath576 since @xmath577 for @xmath575 , the curve @xmath578 remains above the line @xmath579 for all @xmath575 . thus @xmath580 is bounded by @xmath581 as given by ( [ inter1 ] ) ; this is @xmath582 in figure [ cota_delta_fig ] . a related geometrical argument also provides the proof of proposition [ unbounded ] : + * proof of proposition [ unbounded ] . * under condition ( [ h1 ] ) , so that @xmath87 is strictly increasing , the deterministic path in the case @xmath115=1 $ ] tends to infinity . it is then clear that for any distribution for @xmath6 in this case that there exists @xmath148 for which @xmath585 > \eps\ ] ] for all @xmath312 large enough . the stated result for @xmath35 then follows in this case , and hence the result for @xmath73 also . now suppose that condition ( [ h2 ] ) holds and that @xmath6 is non - degenerate . then there exists @xmath586 for which @xmath587 > \eps_1 $ ] . under ( a1 ) , @xmath555 as @xmath91 , so we can choose @xmath101 big enough so that the angle @xmath588 to the normal satisfies @xmath589 for all @xmath552 . from ( a1 ) , we have that @xmath16 is monotone and @xmath590 . then it follows that for all @xmath312 in any bounded interval @xmath591 , ( [ updrift ] ) holds for some @xmath148 . thus with positive probability @xmath35 reaches any finite horizontal distance , and the result follows in this case also . @xmath255 + the next lemma gives crucial estimates for the first two moments of @xmath550 . these require fairly lengthy computations . [ moments_g ] suppose that @xmath6 satisfies ( [ alpha ] ) and @xmath87 satisfies ( a1 ) , and also or . then as @xmath91 @xmath592=2g'(x)g(x)(1 + 2\e [ { { \mathop{\rm tg}}}^2\alpha ] ) + o(g(x)^3/x^2 ) ; \ ] ] and @xmath593= 4 g(x)^2\exp[{{\mathop{\rm tg}}}^2\alpha]+o(g(x)^3/x).\ ] ] * proof . * first suppose that ( [ h1 ] ) holds ( the case of an increasing tube ) . then @xmath564 . take @xmath312 sufficiently large so that @xmath594 . consider the jump @xmath595 . we need to consider 3 cases . in the first case , @xmath596 ( see figure [ delta+ ] ) , @xmath597 in the second case , ( see figure [ delta ] ) , @xmath598 , @xmath599 in the third case , ( see figure [ delta- ] ) , @xmath600 , @xmath601 we start with the first of the three cases . using taylor s theorem , we can write @xmath605,\ ] ] where @xmath606 $ ] . thus , writing @xmath607 , @xmath608 where we have used the fact that @xmath609 ( recall that @xmath610 ) . analogously , in the second case , @xmath611 where @xmath612 , @xmath613 $ ] ; and in the third case @xmath614 where @xmath615 , @xmath613 $ ] . let us now estimate the first two moments of @xmath550 . for convenience write @xmath624 $ ] . for the first moment we obtain , using the symmetry of @xmath625 , @xmath626&=\int\limits_0^{\alpha_0}\delta_+ { { \mathrm d}}f(\alpha ) + \int\limits_{-\theta}^0\delta'_+ { { \mathrm d}}f(\alpha ) -\int\limits_{-\alpha_0}^{-\theta}\delta_- { { \mathrm d}}f(\alpha ) \\ & = \int\limits_0^{\theta } 2g(x ) \big [ \frac{g'(x)+{{\mathop{\rm tg}}}\alpha}{1 - 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}+ \frac{g'(x)-{{\mathop{\rm tg}}}\alpha}{1 + 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}\big]{{\mathrm d}}f(\alpha ) \\ & ~~+\int\limits_{\theta}^{\alpha_0 } 2 g(x ) \big [ \frac{g'(x)+{{\mathop{\rm tg}}}\alpha}{1 - 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}- \frac{-g'(x)+{{\mathop{\rm tg}}}\alpha}{1 + 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}\big]{{\mathrm d}}f(\alpha ) \\ & ~~+o(g''(x)g(x)^2)\\ & = \int\limits_{0}^{\alpha_0 } 2g(x ) \big [ \frac{g'(x)+{{\mathop{\rm tg}}}\alpha}{1 - 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}+ \frac{g'(x)-{{\mathop{\rm tg}}}\alpha}{1 + 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2}\big]{{\mathrm d}}f(\alpha ) \\ & ~~+o(g''(x)g(x)^2)\\ & = \int\limits_{0}^{\alpha_0 } 2g(x ) \big[\frac { 2g'(x)+4g'(x){{\mathop{\rm tg}}}^2\alpha -2g'(x)^3 } { ( 1 - 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2 ) ( 1 + 2g'(x){{\mathop{\rm tg}}}\alpha -g'(x)^2 ) } \big]{{\mathrm d}}f(\alpha)\\ & ~~+o(g''(x)g(x)^2 ) . \end{aligned}\ ] ] the denominator in the term in square brackets in the last integrand is @xmath627 . hence @xmath628 = 2g(x)g'(x)\big(1 + 2\e[{{\mathop{\rm tg}}}^2\alpha]\big ) + o(g''(x)g(x)^2+g'(x)^3g(x ) ) .\ ] ] then using ( a1 ) to bound the error terms , we obtain ( [ mu1 ] ) . for the second moment , we have , in a similar fashion , @xmath629 & = \int\limits_0^{\alpha_0}\delta_+^2 { { \mathrm d}}f(\alpha ) + \int\limits_{-\theta}^0(\delta'_+)^2 { { \mathrm d}}f(\alpha ) + \int\limits_{-\alpha_0}^{-\theta}\delta_-^2 { { \mathrm d}}f(\alpha ) \\ & = \int\limits_0^{\alpha_0}\big[2g(x)\frac{g'(x)+{{\mathop{\rm tg}}}\alpha}{1 - 2g'(x ) { { \mathop{\rm tg}}}\alpha - g'(x)^2 } + o(g''(x)g(x)^2)\big]^2{{\mathrm d}}f(\alpha ) \\ & ~~+\int\limits_{-\theta}^0 \big[2g(x)\frac{g'(x)-{{\mathop{\rm tg}}}|\alpha|}{1 + 2g'(x){{\mathop{\rm tg}}}|\alpha| -g'(x)^2}+o(g''(x)g(x)^2 ) \big]^2{{\mathrm d}}f(\alpha ) \\ & ~~+\int\limits_{-\alpha_0}^{-\theta } \big[2g(x)\frac{-g'(x)+{{\mathop{\rm tg}}}|\alpha|}{1 + 2g'(x){{\mathop{\rm tg}}}|\alpha| -g'(x)^2 } + o(g''(x)g(x)^2 ) \big]^2{{\mathrm d}}f(\alpha ) \\ & = \int\limits_0^{\alpha_0}\big[2g(x ) { { \mathop{\rm tg}}}\alpha + o ( g(x ) g'(x ) ) + o(g''(x)g(x)^2)\big]^2{{\mathrm d}}f(\alpha ) \\ & ~~+\int\limits_{-\theta}^0 \big[2g(x ) { { \mathop{\rm tg}}}\alpha + o ( g(x ) g'(x ) ) + o(g''(x)g(x)^2 ) \big]^2{{\mathrm d}}f(\alpha ) \\ & ~~+\int\limits_{-\alpha_0}^{-\theta } \big[2g(x ) ( -{{\mathop{\rm tg}}}\alpha ) + o(g(x ) g'(x ) ) + o(g''(x)g(x)^2 ) \big]^2{{\mathrm d}}f(\alpha ) \\ & = 2 \int\limits_{0}^{\alpha_0 } 4 g(x)^2 ( { { \mathop{\rm tg}}}^2\alpha ) { { \mathrm d}}f(\alpha ) + o(g'(x)g(x)^2)+ o(g''(x)g(x)^3)\\ & = 4 g(x)^2\e [ { { \mathop{\rm tg}}}^2\alpha ] + o(g'(x)g(x)^2)+ o(g''(x)g(x)^3).\end{aligned}\ ] ] then ( [ mu2 ] ) follows , again using ( a1 ) to bound the error terms . analogously to the case where ( [ h1 ] ) holds , we define @xmath566 , @xmath631 , @xmath632 . for @xmath633 , we have @xmath634 in the second case , when @xmath635 , @xmath636 in the third case , when @xmath637 , @xmath638 in the same way as before , we obtain @xmath639 @xmath640 and @xmath623 similar computations to before then yield the same expressions ( [ mu1 ] ) and ( [ mu2 ] ) . lemma [ moments_g ] is proved . + the next result will allow us to compare the recurrence times @xmath107 and @xmath108 . * first we observe that when @xmath643 is large enough ( such that @xmath644 ) , we have @xmath645 , that is , successive collisions are a.s . on different sides of the tube . thus , for @xmath646 , a.s . , @xmath647 now suppose that ( [ h1 ] ) holds . then for @xmath312 large enough , under ( [ h1 ] ) , ( [ lower ] ) implies @xmath648 . in other words , the time between collisions for the process @xmath73 is no less than that for the process @xmath35 , and part ( i ) follows . now suppose that ( [ h2 ] ) holds . by the triangle inequality , we have @xmath649 thus lemma [ cota_delta ] with ( [ h2 ] ) implies that for some @xmath650 , given @xmath651 , @xmath652 for all @xmath312 large enough . then part ( ii ) follows . @xmath255 our approach to studying the horizontal component of the collisions process @xmath35 is to consider a rescaled version of the process in such a way that we get exactly an instance of the lamperti problem . the key is to find a scale on which the process has uniformly bounded jumps . define the function @xmath653 via @xmath654 . under assumption ( a1 ) on @xmath16 , it follows that @xmath655 \frac{1}{g(x ) } , \\ h''(x ) & = - \frac{2 g'(x)}{g(x)^2 } - \frac{x g''(x)}{g(x)^2 } + \frac{2xg'(x)^2}{g(x)^3 } = [ \gamma ( \gamma -1 ) + o(1 ) ] \frac{1}{x g(x ) } , \\ h'''(x ) & = o \big ( \frac{1}{x g(x)^2 } \big ) .\end{aligned}\ ] ] [ zetalem3 ] suppose that ( a1 ) holds . suppose that @xmath6 satisfies ( [ alpha ] ) . then there exists @xmath225 such that for all @xmath75 and all @xmath658 @xmath659 = 1.\ ] ] also , for all @xmath75 , as @xmath660 @xmath661 & = \frac{2 \gamma ( 1-\gamma ) ( 1+\exp [ { { \mathop{\rm tg}}}^2 \alpha ] ) } { y } + o(y^{-1 } ) ; \text { and}\\ \label{zeta2b } m_2 ( y):= \exp[(\zeta_{n+1}-\zeta_n)^2\mid \zeta_n= y ] & = 4 ( 1-\gamma)^2 \exp[{{\mathop{\rm tg}}}^2\alpha ] + o(1).\end{aligned}\ ] ] * proof . * given @xmath662 , denote @xmath663 . if the reflection is at angle @xmath6 , we have from taylor s theorem that as @xmath91 @xmath664 using lemma [ cota_delta ] . by lemma [ cota_delta ] we have that @xmath665 , and then ( [ zeta0b ] ) is immediate . taking expectations in ( [ def_h2 ] ) and using lemma [ moments_g ] , we obtain @xmath666\\ & = 2 ( 1-\gamma + o(1 ) ) ( 1 + 2\e[{{\mathop{\rm tg}}}^2\alpha ] ) g'(x ) + ( 2 \gamma ( \gamma-1 ) + o(1 ) ) \e[{{\mathop{\rm tg}}}^2\alpha ] \frac{g(x)}{x } \\ & = \frac{2\gamma(1-\gamma ) ( 1 + \e[{{\mathop{\rm tg}}}^2\alpha ] ) g(x)}{x } + o ( g(x)/x).\end{aligned}\ ] ] this yields ( [ zeta1b ] ) . similarly , squaring both sides of ( [ def_h2 ] ) and taking expectations gives ( [ zeta2b ] ) . @xmath255 + this last result , together with proposition [ unbounded ] , immediately implies the following : [ cor1 ] suppose that @xmath16 satisfies ( a1 ) , and that @xmath6 satisfies ( [ alpha ] ) . then @xmath667 is a lamperti - type problem as discussed in section [ lamperti ] , satisfying ( a2 ) and ( a3 ) . moreover , ( [ limsup ] ) holds if @xmath16 satisfies ( [ h1 ] ) , and also , if @xmath6 is non - degenerate , if @xmath16 satisfies ( [ h2 ] ) . finally , ( [ ass2 ] ) holds provided that @xmath6 is non - degenerate . in the special case where @xmath139 , @xmath95 , so that @xmath668 , we will need a more precise version of lemma [ zetalem3 ] . this is lemma [ zetalem ] below . not only will this enable us to deal with the critical case in the recurrence classification ( theorem [ dim2 ] ) , it will also be crucial for our proofs of the almost - sure bounds carried out in section [ bounds ] . [ zetalem ] suppose that @xmath14 where @xmath95 , and that @xmath6 satisfies ( [ alpha ] ) . then for all @xmath75 , as @xmath660 @xmath669 & = \frac{2\gamma(1-\gamma)(1+\exp [ { { \mathop{\rm tg}}}^2 \alpha ] ) } { y}+o(y^{-1 } ( \log y)^{-1 } ) ; \\ \label{zeta2 } m_2(y ) : = \exp[(\zeta_{n+1}-\zeta_n)^2\mid \zeta_n = y ] & = 4(1-\gamma)^2\exp[{{\mathop{\rm tg}}}^2\alpha]+o((\log y)^{-1}).\end{aligned}\ ] ] * proof . * we can apply taylor s theorem to obtain , conditional on @xmath670 , @xmath671 then taking expectations and using ( [ mu1 ] ) and ( [ mu2 ] ) we obtain ( [ zeta1 ] ) . similarly we obtain ( [ zeta2 ] ) after squaring the last displayed expression . @xmath255 + * proof of theorem [ gen_g ] . * under the conditions of theorem [ gen_g ] , corollary [ cor1 ] holds . we apply lamperti s result proposition [ lmpti ] to the process @xmath672 described by lemma [ zetalem3 ] , noting that @xmath672 is null - recurrent , positive - recurrent , or transient exactly when @xmath35 is . from lemma [ zetalem3 ] , we have that @xmath673 + o(1 ) ) , \ ] ] and also @xmath674 ) - \exp [ { { \mathop{\rm tg}}}^2 \alpha ] + o(1 ) \right)\nonumber\\ & = 4 ( 1-\gamma ) \left ( ( \gamma - \gamma_c ) ( 1 + 2 \exp [ { { \mathop{\rm tg}}}^2 \alpha ] ) + o(1 ) \right ) , \end{aligned}\ ] ] where @xmath675 is given by ( [ gammac ] ) . for part ( i ) of the theorem , if @xmath676 we have from ( [ f1 ] ) that there exists @xmath677 such that @xmath678 for all @xmath679 sufficiently large , and hence by proposition [ lmpti](ii ) , @xmath672 is transient . for part ( ii ) of the theorem , it suffices to consider the case where @xmath6 is non - degenerate , so @xmath680 . then from ( [ f1 ] ) we have that for @xmath681 , @xmath682 for all @xmath679 sufficiently large . also , in this case @xmath683 > 0 $ ] , so we have from ( [ f2 ] ) that @xmath684 for all @xmath679 sufficiently large . hence by proposition [ lmpti](i ) , @xmath672 is null - recurrent . this proves theorem [ gen_g ] for the process @xmath35 , and the statement for the process @xmath73 follows from ( [ times2 ] ) and lemma [ compare](i ) . @xmath255 + * proof of theorem [ dim2_erg ] . * for part ( i ) , if @xmath685 < 0 $ ] , we have from ( [ f2 ] ) that @xmath686 for some @xmath363 and all @xmath679 sufficiently large . then ( noting corollary [ cor1 ] ) it follows from proposition [ lmpti](iii ) that @xmath672 and hence @xmath35 is positive - recurrent . for part ( ii ) , it suffices to suppose that @xmath6 is non - degenerate . then the @xmath133 case of ( [ f1 ] ) implies that @xmath687 for all @xmath679 large enough . on the other hand , if @xmath685 > 0 $ ] , we have from ( [ f2 ] ) that @xmath688 for all @xmath679 sufficiently large . then null - recurrence follows from proposition [ lmpti](i ) . + in order to complete the proof of theorem [ dim2 ] , we need a sharper form of lamperti s recurrence classification result presented in proposition [ lmpti ] . fine results in this direction are given in @xcite . we will only need the following consequence of theorem 3 of @xcite . the key to the proof of our almost - sure bounds for the stochastic billiard model is to apply our almost - sure bound results from section [ lamperti ] to the rescaled process @xmath672 that we studied in lemma [ zetalem ] . this will allow us to obtain theorems [ time1 ] and [ time2 ] . we will then derive the results for the continuous - time process @xmath73 , theorems [ time ] and [ time3 ] , from the corresponding results for @xmath35 . recall that for @xmath75 , @xmath690 . + * proof of theorem [ time1 ] . * the idea here is to apply theorem [ prop1 ] to the process @xmath672 . by lemma [ zetalem ] , we have that ( [ f2 ] ) holds . then since @xmath691 this implies that the conditions of theorem [ prop1](i ) and ( ii ) are satisfied for the process @xmath672 ( using corollary [ cor1 ] ) . thus for any @xmath148 , a.s . , for all but finitely many @xmath75 , @xmath692 and then ( [ time1a ] ) and ( [ time1b ] ) follow , since @xmath693 . also by lemma [ zetalem ] , we have that ( [ f1 ] ) holds . if @xmath125 , it follows ( using corollary [ cor1 ] ) that we can apply theorem [ lamp_low ] with @xmath694 to obtain that for some @xmath151 , a.s . , for all but finitely many @xmath75 , @xmath695 . then ( [ time1c ] ) follows . @xmath255 + * proof of theorem [ time2 ] . * this time we will apply theorems [ prop1 ] and [ prop6 ] . it follows from lemma [ zetalem ] that @xmath696 where @xmath697 ) } { ( 1-\gamma ) \exp [ { { \mathop{\rm tg}}}^2 \alpha ] } .\ ] ] hence for @xmath168 $ ] , @xmath443 and ( using corollary [ cor1 ] ) the conditions of parts ( i ) and ( ii ) of theorem [ prop6 ] are satisfied for @xmath672 . then part ( ii ) of the theorem follows . on the other hand , for @xmath167 $ ] , we have @xmath698 so that ( using corollary [ cor1 ] ) the conditions of parts ( i ) and ( ii ) of theorem [ prop1 ] are satisfied . this yields part ( i ) of the theorem , in the same way as in the proof of the corresponding results in theorem [ time1 ] . @xmath255 + the following lemma will enable us to derive our ` infinitely - often ' lower bounds for @xmath699 from bounds for @xmath700 . [ e0 ] suppose that @xmath95 . suppose that there exist @xmath701 with @xmath702 , such that for any @xmath148 , a.s . , for all but finitely many @xmath75 @xmath703 then for any @xmath148 , a.s . , for all @xmath8 sufficiently large , @xmath704 * proof . * recall the definition of the collision times @xmath705 from ( [ nudef ] ) . we have from the triangle inequality and lemma [ cota_delta ] that for some @xmath370 , for all @xmath475 , @xmath706 thus by the upper bound in ( [ e1 ] ) , for any @xmath148 , a.s . , for some @xmath370 and all @xmath41 , @xmath707 let @xmath148 , and for @xmath708 , set @xmath709 also recall the definition of @xmath710 from ( [ ntdef ] ) . then by ( [ d1 ] ) we have that for any @xmath148 , there exist @xmath370 and @xmath711 for which , a.s . , for all @xmath8 large enough @xmath712 hence for any @xmath148 , a.s . , for all @xmath8 sufficiently large @xmath713 now from ( [ d3 ] ) we have that , a.s . , for all @xmath8 large enough @xmath714 now applying the lower bound in ( [ e1 ] ) we obtain for any @xmath148 , a.s . , for all @xmath8 large enough @xmath715 using the definition of @xmath716 . simplifying leads to the desired result . @xmath255 + we will use lemma [ e0 ] in the proofs of theorems [ time ] and [ time3 ] below . we will apply the lemma taking either @xmath717 with @xmath95 or @xmath718 with @xmath719 . note that for @xmath133 , @xmath720 so that the hypothesis @xmath721 in lemma [ e0 ] is satisfied in either case . + * proof of theorem [ time ] . * first of all , from theorem [ time1 ] we have that ( [ time1a ] ) and ( [ time1b ] ) hold . thus we can apply the @xmath722 case of lemma [ e0 ] , which yields ( [ time2b ] ) . it remains to prove ( [ time2c ] ) . suppose that @xmath724 , so that we have transience . first we prove the lower bound in ( [ time2c ] ) . we have from ( [ d4 ] ) that a.s . , for all @xmath8 large enough , @xmath725 hence from ( [ time1c ] ) we have that for some @xmath726 , a.s . , for all @xmath8 large enough @xmath727 the final inequality using ( [ d3 ] ) and the fact that the function @xmath728 is eventually increasing in @xmath729 . now ( [ time2c ] ) follows by the definition of @xmath716 . now we prove the upper bound in ( [ time2c ] ) . from ( [ nudef ] ) and ( [ lower ] ) we have that @xmath730 . then using ( [ time1c ] ) we have that for some @xmath151 , a.s . , for all but finitely many @xmath475 , @xmath731 let @xmath732 , and for @xmath733 set @xmath734 then by ( [ d5 ] ) , we have that for @xmath384 large enough , a.s . , for all @xmath8 sufficiently large , @xmath735 now from ( [ d4 ] ) and ( [ time1a ] ) we have that for some @xmath370 , a.s . , for all @xmath8 sufficiently large , @xmath736 now using the fact that @xmath737 a.s . , and the definition of @xmath738 , the result follows . + * proof of theorem [ time3 ] . * we apply lemma [ e0 ] again . for part ( i ) , we have from part ( i ) of theorem [ time2 ] that ( [ time1a ] ) and ( [ time1b ] ) hold , so we can apply the @xmath722 case of lemma [ e0 ] to obtain ( [ time2b ] ) in this case . h. brzis , w. rosenkrantz , and b. singer , an extension of khintchine s estimate for large deviations to a class of markov chains converging to a singular diffusion , _ comm . pure appl . _ * 24 * ( 1971 ) 705726 . f. comets , s. popov , g.m . schtz , and m. vachkovskaia ( 2007 ) billiards in a general domain with random reflections . to appear in : _ archive for rational mechanics and analysis_. available at arxiv.org as math.pr/0612799 . a. dvoretzky and p. erds , some problems on random walk in space , in : proceedings of the second berkeley symposium on mathematical statistics and probability , 1950 , pp . university of california press , berkeley and los angeles , 1951 . menshikov , i.m . asymont , and r. iasnogorodskii , markov processes with asymptotically zero drifts , _ problems of information transmission _ * 31 * ( 1995 ) 248261 , translated from _ problemy peredachi informatsii _ * 31 * 6075 ( in russian ) .

as more observations accumulate , it has become increasingly clear that the field population of massive galaxy halos consists largely of the debris of accreted satellites , along the lines of searle & zinn ( 1978 ) . in the case of the milky way , the orbital stream of the sagittarius ( sgr ) dwarf spheroidal ( dsph ) is so complicated that it is easy to imagine its members being dispersed throughout much of the halo over a few gyr timescale ( e.g. , law et al . 2005 ; fellhauer et al . additional observations find coherent structures and satellites over much of the galactic halo ( e.g. , belokurov 2007a , b ) . evidence of mergers in the form of the giant stream and other structures , is clearly even more striking in m31 ( ibata et al . 2001 , 2007 ; ferguson et al . n - body models of these events show that even a single minor merger is capable of filling much of the halo with debris after only 2 gyr ( mori & rich 2008 ) . one is also struck by both the surface brightness and spatial extent of the giant stream and its shell feature , all of which are attributable to one relatively recent minor merger , that did not disrupt the disk ( fardal et al . 2006 , 2008 ; mori & rich 2008 ) . the importance of the giant stream event emphasizes that , as halo histories are stochastic , the populations of halos may be dominated by the debris of just a few events involving massive satellites certainly not the picture that one may infer from the study of their globular cluster systems . in the case of m31 , direct measurements of the halo star formation history from the main sequence turnoff revealed a complex range of age and abundance ( e.g. , brown et al . 2003 , 2006 , 2008 ) . in all of the m31 halo fields , there exist significant populations of intermediate - age stars , with metallicities ranging from 1/10 to twice solar . in addition to a contingent of stars ranging from 510 gyr , nearly all m31 halo fields show signs of a blue extension of @xmath1 gyr old stars ( e.g. brown et al . 2006 ; richardson et al . if stellar systems are being ingested then their properties should , in principle , reflect those of similar galaxies in the local group . although star formation likely ceases as a satellite is tidally disrupted , it is clear that much of the complexity can be observationally reconstructed , as has been accomplished for the sagittarius dsph galaxy ( e.g. , chou et al . 2007 ) . an efficient tracer of more recent accretion events through intermediate - age populations is the number ratio , c / m , of carbon - rich ( c ) stars to the oxygen - rich ( m ) stars . c- and m - type stars are the most luminous single stars found in intermediate - age populations , and their spectral signatures are distinctive . in complex stellar systems , the presence of carbon stars in significant numbers guarantees the presence of substantial intermediate - age populations . the primary correlation in the c / m ratio is with metallicity ( iben & renzini 1983 ; brewer et al . 1995 ) ; in more metal rich systems , the balance tilts toward the oxygen - rich m - stars . for example , the galactic bulge is dominated by m giants ( blanco et al . intermediate - age populations are carbon star dominated , with this tendency to increase at lower metallicity ( e.g. groenewegen 2002 ; mouhcine & lanon 2003 ; batinelli & demers 2005 ) . likewise , radial age and/or metallicity gradients in a given stellar system have been confirmed to be accompanied by gradients in the c / m ratio ( e.g. , cioni et al . furthermore , the small magellanic cloud is well known to contrast with the lmc in having a high c / m ratio ( blanco et al . 1984 ) . although carbon stars are known in star clusters of the magellanic clouds ( see e.g. mould & aaronson 1979 ) , no late - type , luminous carbon stars are found in galactic globular clusters despite there being a large contingent of massive , metal poor clusters . it may be inferred that the deep mixing on the agb necessary to produce a carbon star is only possible for intermediate age stars . most studies of the c / m ratio in the local group rely on imaging through narrow band filters that measure the relative strength of cn and tio bands ( e.g. , cook et al . 1986 ) . in this context , the -color index is a measure for the relative strengths of cn and tio bands and thus efficiently segregates m - stars with strong tio and c - stars with strong cn bands . quantitative analyses require additional color information ( such as v@xmath0i ) , in order to construct distinctive two - color diagrams ( tcds ) . specifically , c - stars in m31 have been recorded ( brewer et al . 1995 ; battinelli et al . 2003 ; demers & battinelli 2005 ; battinelli & demers 2005 ) . the inferred relatively low c / m ratios of the order of 0.1 in the m31 disk and the inner halo ( within 10 kpc ) are then in agreement with the metal rich character of these populations . for the present work , we have examined spectra from a large survey of the m31 halo ( koch et al . 2008 ; hereafter k08 ) , and we have synthesized the cn and tio filters for our spectra . from those we explore the c / m ratio in m31 halo fields , some of which are further than 100 kpc from the nucleus , but we do so without the completeness of a magnitude - limited photometric survey . even considering these limitations , we can statistically investigate the late - type stellar content in the andromeda halo and connect occurrences of either population to the prevailing ages and metallicities out to large radii . this paper is organized as follows : in 2 we briefly summarize our observations and the standard reductions , while 3 explains our definitions of the cn@xmath0tio pseudo - filters . the separation into c- and m - stars from these filters , possible biases , and the implications for m31 s halo populations are discussed in 4 . finally , 5 summarizes our findings . the data used in this work were collected in the course of a keck program focusing on the spectroscopic investigation of m31 s halo structures based on the kinematics and chemical analysis of red giants . since these data are identical to those presented in k08 , we refer the reader to the latter ( and references therein ) for details on the target selection , data collection and reduction for the whole project . to recapitulate , our data set covers fields on the south - east minor axis of m31 at projected distances of 9 kpc out to 160 kpc . these fields were observed using the deimos multislit spectrograph at the keckii 10 m telescope over several observing runs from 2002 through 2006 . the majority of the spectra was centered at a wavelength of 7800 , yielding a full spectral coverage of @xmath265009200 , which also comprises the prominent molecular tio and cn bands that are of interest for the present study . the photometry of our targets comes from two sources : for the outer halo fields ( r@xmath325 kpc ) the washington @xmath4 , @xmath5 and @xmath6 photometry of ostheimer ( 2003 ) was transformed into standard johnson - cousins v and i magnitudes following majewski ( 2000 ) . stars from this photometry were pre - selected to maximize the number of red giants , combining their location in the color magnitude diagrams ( cmds ) and tcds built from the washington filters so as to optimally avoid the dwarf sequence ( i.e. , blueward of the red giant branch [ rgb ] , and _ much _ brighter than the tip of the rgb [ trgb ] ; e.g. , majewski et al . 2000 ; palma et al . 2003 ) , and to cover the entire rgb so as to yield a broad , unbiased metallicity range . the photometry of the inner m31 fields ( r@xmath725 kpc ) was , on the other hand , taken from the megacam / megaprime archive ( gwyn 2008 ) and transformed to the johnson - cousins magnitudes via the latest padova stellar isochrones ( marigo et al . 2008 ) , which are available both in the chft photometric system and for johnson - cousins v and i. as above , the goal of the respective target selection was to avoid those regions of the cmd with the greatest foreground dwarf contamination ( and to exclude galaxies based on morphological criteria ; e.g. , gilbert et al . 2006 ) . with the aim of targeting bright stars to reach sufficient s / n ratios , highest priority was given to stars with @xmath8 mag . since m31 s trgb lies at approximately 20.5 mag ( e.g. , durrell et al . 2001 ) , some bright agb stars above the trgb will have been removed from the initial samples . in fact , our sample retains a number of bright ( red ) stars above the rgb tip , so we conclude that our relative completeness does not fall so rapidly . the redder stars are actually fainter in the i - band due to molecular absorption , thus permitting bolometrically bright stars to fall within the i - band selection ( see also cmds in fig . 5 ) . targets selected from either survey reach similar limiting magnitudes ( slightly fainter than 22.5 mag ) and span a comparably broad color range , so we conclude that no selection criterion has a bias set either in selecting c- or m - stars . based on stellar radial velocities , the equivalent widths of the gravity - sensitive na doublet and the stars v@xmath0i colors , we obtained an efficient separation of m31 red giants from the milky way foreground contamination in k08 . k08 reject 781 stars as being likely foreground dwarfs , while 1288 were classified as likely m31 giants and comprise our present sample . the narrow band ( @xmath9300 ) _ 77 _ and _ 81 _ filters centered on molecular bands have proven successful in separating c and m stars in local group galaxies , when coupled with temperature sensitive broad - band photometry . in particular , the red giant branch ( rgb ) bifurcates into the redder ( in ) c - stars and the m - stars with v@xmath0i redder than @xmath21.52.0 ( cook et al . 1986 ; cook & aaronson 1989 ; nowotny 2001 , 2003 ; harbeck et al . 2004 ) with similar cuts for other color indices ( e.g. , r@xmath0i ; battinelli et al . 2003 ; demers & battinelli 2005 ) . our program for k08 was originally designed to focus on spectroscopy so that no photometry in selected narrow - band filters was taken . thus we have to rely on integrating our spectra in band passes mimicking the traditional filter curves . depending on the choice of telescope and instrument , the exact filter curves will differ and therefore yield slightly different zero points and color indices . we investigated the different instrumental set ups by convolving our spectra with filter curves from those three sites that are most widely found in the literature on the late - type stellar content of local group galaxies : firstly , the cfht filters as used by , e.g. , battinelli et al . ( 2003 ) ; demers & battinelli ( 2005 ) , to name a few ; secondly , the band passes available for the wiyn telescope ( e.g. , harbeck et al . 2004 ) ; and finally , we applied gaussian filter curves ( cook & aaronson 1989 ) as defined in the asiago data base ( moro & munari 2000 ; fiorucci & munari 2003 ) . the respective transmission curves are compared in fig . 1 . while the _ 77_-filter curve , centered at 7750 , measures the tio band head in m - stars , it essentially samples the continuum in c - star spectra . the _ 81_-filter at 8100 , on the other hand , is a measure for cn absorption in the c - stars and covers the stellar continuum of the cool m - giants . this is illustrated in fig . 2 for two stars from our m31 sample that were classified as c- and m - stars based on the -color , and for a background object with essentially flat continuum in either filter . -band that contributes the only absorption within the band passes . both stars have radial velocities consistent with membership with m31 . the spectra were arbitrarily normalized at 7562 . indicated as dashed lines are the cfht - based filter curves we used for measuring the color index . ] the low signal - to - noise ( s / n ) ratios of these examples already indicates that a purely spectroscopic classification of our sample would be difficult to achieve ; therefore the main objective of this pseudo - photometric study is rather a statistical endeavor ( note that the median s / n of the entire sample in k08 is a mere 8.5 per pixel ) . in practice , we define @xmath10 where @xmath11 denotes the filter profiles that were taken from the aforementioned telescopes filter characteristics . the integration is carried out over the whole doppler corrected spectral flux , @xmath12 . the constant in eq . 1 accounts for the integral over the transmission curves and amounts to ( @xmath00.086 , 0.351 , 0.272 ) for the ( cfht , wiyn , asiago ) filtes , respectively . we note , however , that an accurate zero point is not required for this study , since we are primarily interested in an empirical identification of the different stellar populations . errors on the color were determined by monte carlo variations of the spectra accounting for the spectral noise and uncertainties in the radial velocities ( entering the measurements upon doppler shifting the spectra ) . while the velocity contribution is negligible , at a median @xmath13 of 0.0001 mag , the random error due to the s / n dominates the error budget and yields a median total random color uncertainty of 0.012 mag ( with 95.7% of the errors smaller than 0.04 mag ) . it turns out that all three different filter definitions yield very similar results and the resulting color indices are extremely well correlated . 3 shows a comparison of each narrow band measurement and despite the fact that the trends are not strictly unity , there is only little scatter of less than 0.1 mag ( 1@xmath14 ) around the linear relations . however , a major source of uncertainty is the presence of telluric absorption that falls into the filter band passes ( fig . 2 , bottom panel ) . in particular , the _ 77_-band slightly overlaps the atmospheric @xmath15-band at 7600 , which will remove flux from the stellar spectrum that matters for the filter integration ( e.g. , cook & aaronson 1989 ) . this prompted us to adopt the cfht - based filter definitions for our final analysis , as this particular system has the smallest bandwidth and thus shows the least overlap with the a - band ( fig . 2 ) . secondly , the gap between the two adjacent deimos ccds can introduce flux offsets that may affect a measure of the true color index . to estimate the bias introduced by these effects , we performed the identical integration over the filter curves for a subsample of background galaxies in the m31 sample ( see k08 ) with preferentially flat continua . to first order , these spectra contain only a flat continuum over the entire spectral range , plus the atmospheric contaminants ( emission lines that are red shifted into the filter band passes result in a strongly aberrant and can be easily discarded ) . as a result , the of the galaxies is compatible with zero as expected , with a 1@xmath14 scatter of 0.10 mag . an additional test of this error component consists of estimating the scatter in the pseudo - color of dwarf stars around @xmath16 ( see also fig . 4 ; bottom panel ) , which are expected to contain only the telluric features in the spectral region of interest . in concordance with the galaxies , the respective scatter of in the bluer dwarfs amounts to 0.10 as well . thus we add this value in quadrature to the color error to account for undesirable telluric absorptions . in fig . 4 we show the tcd of our stars in the sense of cn@xmath0tio ( ) versus v@xmath17 , both for giant and dwarf stars according to the separation in k08 . in order to select possible c - star candidates and m - giants , we draw empirical lines that frame these populations ( cf . cook et al . 1986 ; harbeck et al . 2004 , and references therein ) , but note that their exact choice remains somewhat arbitrary due to potential zero point uncertainties in the colors and our non - standard , i.e. , spectroscopic derivation of the color . m - stars were selected from those diagrams by imposing a cut at v@xmath0i=2 , which strictly selects spectral types later than m5 ( `` m5 + '' ; e.g. , demers & battinelli 2003 ) . to select c - star candidates , we used a bluer cut at v@xmath0i of 1.55 . although this means that we do not obtain a homogeneous selection of c- and m - stars in the same color space , this cut - off is justified by the spectral type assignments of nowotny et al . ( 2001 ) , according to which the c - star branch extends relatively bluewards ( as indicated by the square symbols in fig . 4 ) . moreover , the presence of stars in this selection box , which are clearly detached from the regular rgb and that are potentially identifiable through their spectra , is in agreement with our bluer c - star candidate criterion . the limits on the color , on the other hand are set to @xmath18 ( c ) and @xmath19 ( m ) , in accordance with the boundaries used by cook et al . ( 1986 ) ; battinelli et al . ( 2003 ) ; see also brewer et al . ( 1996 ) . overall , those five c - star candidates in the selection box are affected by low s / n and suffer from unsecure spectral classification . only for very few cases , smoothing of the spectra yields a weak indication of an excess in the cn band ( cf . fig . 2 ) . m - stars , on the other hand , are usually bright and clearly distinguished by their prominent tio bands , already upon visual inspection . we also show , in fig . 5 ( left panel ) , the location of the five tentative c - stars in a cmd . two of those are fainter than i@xmath2022 mag , while the remainder are comparably bright objects . none of remaining three candidates lies above the trgb of 20.52@xmath210.05 from durrell s et al . ( 2001 ) study of a halo field at 20 kpc . we note , however , that ibata et al . ( 2007 ; their fig . 45 ) estimate a spread in distance modulus of up to one magnitude due to the large depth extent of the m31 halo , which may affect conclusions drawn from stars covering a range of distances out to large radii . ] , although the two brightest ones are comparable to its magnitude . likewise , we show in the right panel of fig . 5 the 50 m - giant candidates . it is reassuring that they populate the same locus in the cmd as the c - star candidates . since we indeed detected copious m giants and therefore likely agb stars in this boundary , this shows that this region is not biased against c stars : if present in larger number , we should have been able to detect them regardless of our preselection criteria ( sect . 2 ) . 0.05 from durrell et al . small filled squares show the `` tcd outliers '' with red v@xmath0i colors at ( cn@xmath0tio)@xmath20 , while solid triangles are those stars with large positive ( cn@xmath0tio ) on the blue side of the tcd . ] overall , there are very few c - star candidates found in the respective region of the giants tcd ( top panel of fig . 4 ) . by simply counting numbers in either box , we find a c / m5 + ratio for the entire m31 halo sample of ( 10@xmath215)% , with an error based on poisson statistics . in total , there are 5 ( 50 ) objects with colors that _ formally _ imply that they are c- stars . an accurate account of the measurement errors prompted us to also run a number of monte carlo simulations , varying the sample by its uncertainties in either color . thereby we find a similar , statistical value of the c / m5 + ratio of ( 0.10@xmath210.03 ) , where 5 ( 53 ) stars on average are scattered into the c- ( m- ) star selection region . the late - type stars thus make up a fraction of ( 4@xmath212)% of our entire spectroscopic red giant sample . our measured c / m5 + ratio compares to a foreground - corrected value of 8.4% found in the _ outer disk _ of m31 ( battinelli et al . 2003 ) and those of @xmath210% by demers & battinelli ( 2005 ) and 2% to 10% out to 35 kpc found by brewer et al . note that all these latter studies identify their c - star candidates from narrow band photometry only and use slightly different cmd selection criteria ( see also sect . 4.3.3 ) . two regions in the tcd ( fig . 4 ) contain stars that do not seem to be compatible with the colors of any population within , nor outside of the selection boxes . the first group comprises approximately seven stars with large cn absorption compatible with that of c - stars ( @xmath20.6 mag ) , but with v@xmath0i of @xmath21 and therefore typical of the rgb . these stars cover 1.5 mag in magnitude towards the fainter end of the observed rgb ( fig . 5 ) . thus their respective spectra are largely governed by strong noise and , in two cases , flux offsets between the deimos ccds so that they were easily discarded . it is interesting , though , that battinelli & demers ( 2005 ) find comparable values of positive ( cn@xmath0tio ) on the blue side of their tcd from their photometry at large radii ( @xmath22 kpc ) in the m31 disk . secondly , we note a plume of the order of 12 red ( v@xmath0i@xmath23 ) stars with a @xmath24 , which thus scatter away from the well defined rgb . most of those spectra withstand a classification as late k - type to early m - giants ( by visual inspection ) , thus indicating highly erroneous v@xmath0i , although their nominal color uncertainties are not unusually large . the remainder of those spectra are largely dominated by noise . a possible reason for larger uncertainties in some stars colors may be their variability on the agb . as the luminosity amplitude of such stars is much larger in v than in the i - band ( as high as 1 mag ) , their colors can be falsified ( in the sense of a broader color scatter ) if the the v and i photometry were taken at different epochs . moreover , even if the v and i photometry are obtained nearly simultaneously , as for this work ( and k08 ) , v@xmath0i could be abnormally red if the observations were taken at minimum light . therefore , systematic variability of the reddest of these particular stars can not be ruled out as an additional error source ( see also battinelli & demers 2005 ) . all in all , these two components fall outside the ranges populated by the standard m- and c - star populations considered here and they are unlikely to affect the c / m ratio in the present analysis , given the empirical treatment of this ratio and the ( photometric ) errors . an accurate determination of the c / m star ratio from a spectroscopic survey as presented here can suffer from several effects : first , a potential residual contamination with foreground dwarf stars can artificially decrease the c / m ratio . as addressed in k08 , this component is expected to be largely reduced through a laborious treatment of v@xmath0i - color , velocity and na doublet equivalent widths as membership criteria . to this end , we separated the tcds in fig . 4 into foreground dwarfs and m31 giant candidates following the assignment of k08 . while there are no interfering milky way dwarf carbon stars to be expected ( totten et al . 2000 ) , as is in fact seen in our data ( bottom panel of fig . 4 ) , there can be a non - negligible dwarf contamination in the m - star regime . on the other hand , this issue is efficiently resolved by just the addition of the dwarf / giant information from k08 and we expect no significant residual dwarf component in the `` giant '' sample that would falsify the final c / m ratio . furthermore , brewer et al . ( 1996 ) found that all stars in their tcd selection region were spectroscopically confirmed c - stars with no other contaminating spectral types . we also note that the sequence of foreground m - giants ( bottom panel of fig . 4 ) is in full agreement with the observed galactic dwarf component ( e.g. , letarte et al . 2002 ) and the empirical colors of nowotny et al . ( 2001 ) , which reinforces the accuracy of our separation procedure in k08 . in order to investigate to what extent possible inhomogeneities in the input photometry and the mixture of several observing runs , extending over a period of @xmath25 years ( k08 ) could affect our conclusions , we separated our sample into individual tcds for each observing run , in particular those for the inner fields for these , the photometry is based on the homogeneous megacam survey only . these diagrams are shown in fig . 9 in the appendix . in either case , the expected tracks for the rgb and the m - giant populations are well reproduced and the distribution for the foreground dwarfs is homogeneous between the runs as well . furthermore , the values for the c / m ratio we find from data in different fields are consistent , albeit hampered by overall low numbers ( see also sect . 4.1 ) . this reassures us that our use of data from different epochs and observing runs does not incur any inhomogeneities in the evaluation of our c / m ratios . furthermore , when plotting tcds separately for all fields from megacam and all washington - selected ones , we find no difference ( fig . 10 in the appendix ) . finally , we resolve the issue of whether potential differences in the limiting magnitudes of both sources for the target ( pre- ) selection could bias the detectability of c- and m - stars with respect to numbers or their radial distribution . therefore , we construct tcds and the c / m ratio by imposing several magnitude cuts for all samples that obey i@xmath2621.5 , 22.0 , 22.5 mag , respectively , the general trends in the diagrams are broadly preserved and the same order of magnitude of @xmath210% in the c / m ratio is recovered . we conclude that non - homogeneities between our input sources are negligible and of no concern for the present study . finally , the targets for the spectroscopic study were deliberately pre - selected from washington photometry ( ostheimer 2003 ; palma et al . 2003 ) and the megacam data to follow the locus of the rgb and to cover a broad part of the giant branch to be unbiassed with regard to metallicity , thereby including some red stars . the main goal of k08 was to maximize the number of red giants while avoiding regions of the cmd with the greatest foreground dwarf contamination ( i.e. , blueward of the rgb , and much much brighter than the trgb ) . this means that also the contribution of cool , late - type stars to the present sample is potentially incomplete . strictly , the c / m ratio should be determined using c- and m- agb stars only , with a faint limit for the agb m - stars at m@xmath27 ( brewer et al . 1995 ) , corresponding to the trgb of m stars . therefore , the low value of c / m we find would be caused by an overestimate of the number of m5 + agb stars in our sample . although we do not have any means to directly uncover this number of `` true '' agb stars from our present m ( agb+rgb ) sample , we obtain order of magnitude estimates using two approaches . durrell et al . ( 2001 ) state that , at any given magnitude below the trgb , 22% of the stars are on the agb as estimated from evolutionary models . we can therefore assume that 11@xmath212 stars in our m - giant selection box are agb stars , which leads to a c / m5 + ratio of 0.45@xmath210.25 . on the other hand , it is possible to turn to the agb - fraction , @xmath28 , which is defined as the number ratio of luminous agb stars above the trgb ( per magnitude ) to the number of rgb stars below the tip ( armandroff et al . accounting for the above agb contribution below the trgb of 22% , we thus estimate n(rgb ) as approximately 37 . however , a major unknown in this reasoning is the actual value of @xmath28 . observations of dsphs find a wide range of agb fractions , e.g. , 314% ( lianou et al . 2010 ) , while martinez - delgado & aparicio ( 1997 ) find 820% based on cmd modeling of old populations . here we tentatively adopt 15% as a reasonably representative mean value , which predicts 5@xmath211 luminous agb stars in a sample like ours . only accounting for these would result in a higher c / m5 + ratio of 1.0@xmath210.7 . we note , however , that both methods assume that the number of c - stars is not affected by any of the above arguments and remains constant at five . to investigate the influence that incompleteness could have on our c - star counts , we determined empirical incompleteness factors using c - star tabulations from the literature . for a number of dwarf galaxies , we determine the number ratio of luminous agbs ( above the trgb ) to the fainter c - stars below the tip . this ratio gives us an order of magnitude for our underestimate of the luminous agb stars that we might miss , _ assuming our survey was 100% incomplete above the rgb tip_. harbeck et al . ( 2004 ) surveyed several dsphs satellites of m31 and find _ more _ c - star candidates below the trgb . the data of nowotny et al . ( 2003 ) for ngc 147 and ngc 185 indicate approximately a factor of two more c - stars above the trgb compared to the fainter ones , while battinelli & demers ( 2004a , b ) find of the order of three times more luminous than fainter c - stars in the same galaxies , that is , the trgb . while such an empirical incompleteness correction to our survey would yield a much higher c / m ratio of @xmath201 , it is interesting to note that our value found in the halo is in good agreement with that found in the disk by battinelli et al . ( 2003 ) . ] . if we were to combine all the ( extreme ) effects described in this section into an empirical correction to our measurements , we could reproduce values for m31 s c / m5 + ratio as high as @xmath23 . however , since the above treatment of possible incompleteness in our spectroscopic sample relies on several crude assumptions , we will retain in what follows the original value of ( 0.10@xmath210.03 ) determined in sect . 4.1 . in fig . 6 we show the radial distributions of c- and m - stars , which is clearly aggravated by the small number statistics of either component . in particular , we identified c - star candidates at distances of 1132 kpc . in those three radial bins that contain a non - zero number of carbon - stars , these correlate well with the respective m - giant populations so that the resulting c / m5 + ratio in each bin remains constant at @xmath20.080.13 . it is further interesting that , while the absolute numbers of m - stars in the outer fields ( r@xmath2010 kpc ) drop with radius and therefore qualitatively trace the overall , decreasing halo density profile ( e.g. , ibata et al . 2007 ) , the fraction of the m - giants relative to the entire ( k- and m- ) giant population is also invariant with distance , at @xmath20.04@xmath210.02 . gilbert et al . ( 2006 ) , ibata et al . ( 2007 ) , and k08 detect stars that belong to the m31 halo out to large distances , albeit in low numbers . their fields between 90 and 160 kpc , for instance , were found to contain a total of 42 red giant members ( k08 ) , 26 of which have metallicity measurements . as fig . 6 implies , there is only one interfering m - giant candidate present amongst the overall giant sample at 90 kpc , while their majority ( 2@xmath14 ) resides within 34 kpc . from these low numbers , we also conclude that any possible remaining `` contamination '' of our metallicity sample with the few cooler , metal rich m - giants did not falsify the metallicities in none of the radial bins in k08 , therefore confirming the finding of very metal poor stars and the strong radial metallicity gradient of k08 . the presence of such a gradient and the low metallicities in the m31 halo is , however , in contrast to the low total number of c - stars we detect in this study ( see discussion in the next section ) . to investigate whether the m - giants are representative of the overall m31 halo population , we show in fig . 7 velocity histograms of the entire giant sample ( as solid lines ) using the kinematic data of k08 . these have been scaled to the total number of the m - giants , which are shown as shaded histograms . in this plot , the `` halo fields '' contain all giant candidates from the 12 minor axis and three off - axis fields of k08 , while the `` stream fields '' refer to masks h13s and a3 on top of the giant stellar stream . it is intriguing how well the m - stars trace the overall velocity distribution of the entire giant sample , both for the halo _ and _ the giant stream in either case , the m - star and the giant velocities are in full agreement . in the halo fields , most of the oxygen - rich stars are found at the m31 systemic velocity ( at around @xmath0300 kms@xmath29 ) , while the stream fields display m - giant peaks at the stream components velocities of @xmath30 and @xmath31 kms@xmath29 . the same holds for the five carbon - star candidates , two of which are found in the stream fields . in particular , the c / m5 + ratio in the halo ( 3/32 ; that is , 0.09@xmath210.06 ) is indistinguishable from the value in the stream , at 2/18 or 0.11@xmath210.08 . the agreement of the low value found in the stream fields ( which were preselected from only one data source ) with the other halo values reassures us that this likely not an artifact of inhomogeneous selection criteria during the photometric target selections . on the other hand , we find a slightly higher relative number of m - giants in the stream as compared to the halo fields : while n(m)/n@xmath32 = ( 8@xmath212)% in the stream , this ratio is at ( 3@xmath210.5)% in the genuine halo fields . we have analyzed a total of 1288 spectra of giant stars in the m31 halo , with projected distances ranging from 9 to 160 kpc . using mock narrow - band filter convolutions , we have been able to plot these giants populations in the cn@xmath0tio vs. v@xmath0i diagram ; from this we find that the number ratio c / m5 + is less than 0.1 in _ all halo fields_. we detect a total of only five c - star candidates . therefore , the outer halo exhibits the low c / m ratio that typifies other fields that have been targeted in m31 ( brewer et al . 1995 ; nowotny et al . 2001 ; batinelli & demers 2005 ) . moreover , a few of the m31 dsph satellites only show null - detections or upper limits of c - star numbers ( groenewegen 2002 ; harbeck et al . in contrast , disk fields closer than 30 kpc ( batinelli & demers 2003 ) host far greater absolute numbers of carbon stars that roughly trace the disk surface brightness . if we assume the mean metallicity ( and 1@xmath14 spread ) of those fields , which harbor these candidates ( at @xmath26[fe / h]@xmath20=@xmath33 dex for the halo fields and @xmath34 dex for the stream fields ; k08 ) , this low c / m5 + ratio in the m31 halo fields falls @xmath1 order of magnitude below the well known correlation of local group galaxies ( fig . 8) , which is defined over a broad range of morphological types by systems like the smc , m33 , and ngc 205 ( groenewegen 2002 ; mouhcine & lanon 2003 ; cioni & habing 2003 ) . it is noteworthy , though , that the m31 dwarf galaxy satellites and vii and ngc 147 ( battinelli & demers 2004a ; harbeck et al . 2004 ) fall near our m31 halo and stream fields in this plot . ngc 147 is an m31 dwarf elliptical satellite , devoid of dust and gas ( grebel et al . 2003 ) , that exhibits a star formation history similar to the magellanic clouds or the galactic dsphs fornax and leo i ( harbeck et al . 2004 ) in that it is massive enough to form stars over most of its life time ( m@xmath35=15.1 ) . the m31 dsph companion and vii ( grebel & guhathakurta 1999 ) , on the other hand , is different from the remainder of the m31 dsphs in that it experienced prolonged star formation and it shows a predominant intermediate - age population of the order of 35 gyr ( harbeck et al . 2004 ) . these galaxies and our halo value also lie along lines of similar star formation histories , as indicated by the models by mouhcine & lanon ( 2003 ; black lines in fig . 8) , in particular those that are characterized by a higher dominance of the younger populations ( solid and long - dashed lines ) . the agreement of our c / m ratio in the m31 halo with those of ngc 147 and and vii is then consistent with the interpretation that such ( dsph and de ) galaxies are likely connected to some of the distinct merger features in m31 ; the structures and mean stellar metallicites of these features ( ferguson et al . 2002 ; harbeck et al . 2004 ) are further consistent with this picture . owing to the pre - selection criteria for the present spectroscopic sample , we can not exclude the possibility that we have underestimated the number of m - type agb stars ( sect . under realistic assumptions , however , we showed that the c / m5 + ratio is unlikely to be significantly larger than 1.0 , unless we missed a substantial c - star population in our sample . this value is then still lower than expected from a smc - type population and rather similar to the other m31 dwarf elliptical satellite , ngc 185 . in order to reach a c / m ratio as high as predicted by the models of mouhcine & lanon ( 2003 ) at the given halo metallicity , m31 s agb - fraction , @xmath28 , would have to be remarkably low , of the order of @xmath36% . moreover , only if our survey was 100% incomplete at the bright end could we attain smc - like c / m ratios of @xmath23 ( following our methodology in sect . 4.3.3 . ) . prior surveys of the m31 disk , extending to 40 kpc where the halo likely dominates , find compelling evidence that the c / m5 + fraction actually drops at the boundary of the disk ( demers & battinelli 2005 ) . the g1 cloud of ferguson et al . ( 2002 ) , once thought to be the likely remnant of a dwarf spheroidal satellite , exhibits a c / m5 + ratio typical of our halo fields . deep hst observations ( brown et al . 2003 , 2006 , 2007 , 2008 ; richardson et al . 2009 ) , find a widespread intermediate age population in the m31 halo , while kalirai et al . ( 2006 ) and k08 both find evidence of a metal poor outer halo in m31 . low metallicity and a strong intermediate age population would appear to be strong factors favoring a large carbon star population . indeed , this appears to be the case for the galactic dwarf spheroidals ( e.g. , azzopardi et al . 1985,1986 ; groenewegen 1999 ; demers & battinelli 2002 ) . however , the galactic halo does not appear to be comprised of the debris of such systems with major c - star populations . even if the parent systems disrupted , the carbon - rich stars should have survived in the halo field ( see ibata et al . 2001b for the sgr dwarf ) . we are left to conclude that two factors may decrease the expected numbers of carbon stars . first , the age distribution in the outer halo may be tilted toward older stars . however , in the g1 clump ( at a projected distance of 34 kpc ) , there is evidence that the field has a strong intermediate - age component ( rich et al . 2004 ; faria et al . 2007 ) and also an hst field at 35 kpc was shown to contain @xmath21/3 of stars younger than 10 gyr ( brown et al . the second explanation would argue that the halo is too metal rich to support a substantial population of c - stars . however , both kalirai et al . ( 2006 ) and k08 agree that [ fe / h]@xmath37 and lower at large distances , a value that is low enough to support carbon star production ; moreover , there is a dispersion in metallicity with a metal rich tail high enough to spawn m giants . considering that the origin of this metal rich tail must arise in mergers , the absence of a strong carbon star population in the outer halo must be considered to be noteworthy . therefore , the low c / m ratio in the m31 field is inconsistent with a presumed disk - like or dwarf galaxy origin ( with properties like those indicated in fig . 8) of those halo fields over a broad radial extent . although the total numbers of stars are low in the halo , the c / m _ ratio _ should be preserved . the cool m - giants with detectable tio , on the other hand , are found throughout the halo , further strengthening the case for a widespread population with [ fe / h]@xmath38 , even for the most distant giants . further , stars with measureable tio follow the same kinematics as is found for all giants with measurable radial velocities . both regarding all m stars and with respect to the c / m ratio ( as measured by our very small numbers ) the giant stream appears to harbor no special stellar population , in contrast to what has been found for the sgr dwarf spheroidal galaxy ( ibata et al . 2001b ) , which does have a concentration of carbon stars . with the explosive discovery of numerous streams and stellar systems in the halo , the picture of the halo population comprised nearly entirely of stellar subsystems has become the canon . yet well constructed surveys of our milky way halo ( e.g. mauron et al . 2008 ) find only a handful of luminous carbon stars . if indeed halos are dominated by debris , as can not ruled out from a dynamical standpoint , they are not similar in population to the smc , ngc 6822 , or m33 . if we appeal to the disruption of a disk - like galaxy to have created the giant stream ( e.g. , bearing on the model of fardal et al . 2008 ) , that progenitor was especially devoid of carbon stars and more closely resembled ngc 147 rather than m33 . in the broader picture , the late - type stellar content of the m31 halo must be judged to differ from that of the dwarf galaxies that one might propose to have contributed its stars . we would like to thank m .- cioni for helpful discussions and an anonymous referee for a constructive report . ak acknowledges support by an stfc postdoctoral fellowship . support was provided by nsf ( ast-0307931 , ast-0709479 ) , hst ( go-10265 , 10816 ) , and by r.m . some of the data presented herein were obtained using the w. m. keck observatory , which is operated as a scientific partnership among caltech , the university of california , and nasa . the observatory was made possible by the generous financial support of the w. m. keck foundation . this research used the facilities of the canadian astronomy data centre operated by the national research council of canada with the support of the canadian space agency . armandroff , t. e. , da costa , g. s. , caldwell , n. , & seitzer , p. 1993 , , 106 , 986 azzopardi , m. , lequeux , j. , & westerlund , b. e. 1986 , , 161 , 232 azzopardi , m. , lequeux , j. , & westerlund , b. e. 1985 , , 144 , 388 battinelli , p. , demers , s. , & letarte , b. 2003 , , 125 , 1298 battinelli , p. , & demers , s. 2004a , , 418 , 33 battinelli , p. , & demers , s. 2004b , , 416 , 111 battinelli , p. , & demers , s. 2005 , , 430 , 905 belokurov , v. , et al . 2007a , , 658 , 337 belokurov , v. , et al . 2007b , , 654 , 897 blanco , v. m. , mccarthy , m. f. , & blanco , b. m. 1984 , , 89 , 636 brewer , j. p. , richer , h. b. , & crabtree , d. r. 1995 , , 109 , 2480 brewer , j. p. , richer , h. b. , & crabtree , d. r. 1996 , , 112 , 491 brown , t. m. , ferguson , h. c. , smith , e. , kimble , r. a. , sweigart , a. v. , renzini , a. , rich , r. m. , & vandenberg , d. a. 2003 , , 592 , l17 brown , t. m. , smith , e. , ferguson , h. c. , rich , r. m. , guhathakurta , p. , renzini , a. , sweigart , a. v. , & kimble , r. a. 2006 , , 652 , 323 brown , t. m. , et al . 2007 , , 658 , l95 brown , t. m. , et al . 2008 , , 685 , l121 chou , m .- y . , et al . 2007 , , 670 , 346 cioni , m .- l. , & habing , h. j. 2003 , , 402 , 133 cioni , m .- l. , et al . 2008 , , 487 , 131 cook , k. h. , aaronson , m. , & norris , j. 1986 , , 305 , 634 cook , k. h. , & aaronson , m. 1989 , , 97 , 923 demers , s. , & battinelli , p. 2002 , , 123 , 238 demers , s. , & battinelli , p. 2005 , , 436 , 91 durrell , p. r. , harris , w. e. , & pritchet , c. j. 2001 , , 121 , 2557 fardal , m. a. , babul , a. , geehan , j. j. , & guhathakurta , p. 2006 , , 366 , 1012 fardal , m. a. , babul , a. , guhathakurta , p. , gilbert , k. m. , & dodge , c. 2008 , , 682 , l33 faria , d. , johnson , r. a. , ferguson , a. m. n. , irwin , m. j. , ibata , r. a. , johnston , k. v. , lewis , g. f. , & tanvir , n. r. 2007 , , 133 , 1275 fellhauer , m. , et al . 2006 , , 651 , 167 ferguson , a. m. n. , irwin , m. j. , ibata , r. a. , lewis , g. f. , & tanvir , n. r. 2002 , , 124 , 1452 fiorucci , m. , & munari , u. 2003 , , 401 , 781 gilbert , k. m. , et al . 2006 , , 652 , 1188 grebel , e. k. , & guhathakurta , p. 1999 , , 511 , l101 grebel , e. k. , gallagher , j. s. , iii , & harbeck , d. 2003 , , 125 , 1926 groenewegen , m. a. t. 2002 , in chemical evolution of dwarf galaxies , ringberg workshop ( arxiv : astro - ph/0208449 ) gwyn , s. d. j. 2008 , , 120 , 212 harbeck , d. , gallagher , j. s. , iii , & grebel , e. k. 2004 , , 127 , 2711 ibata , r. , irwin , m. , lewis , g. , ferguson , a. m. n. , & tanvir , n. 2001a , , 412 , 49 ibata , r. , irwin , m. , lewis , g. f. , & stolte , a. 2001b , , 547 , l133 ibata , r. , martin , n. f. , irwin , m. , chapman , s. , ferguson , a. m. n. , lewis , g. f. , & mcconnachie , a. w. 2007 , , 671 , 1591 iben , i. , jr . , & renzini , a. 1983 , , 21 , 271 kalirai , j. s. , et al . 2006 , , 648 , 389 kerschbaum , f. , nowotny , w. , olofsson , h. , & schwarz , h. e. 2004 , , 427 , 613 koch , a. , et al . 2008 , , 689 , 958 ( k08 ) law , d. r. , johnston , k. v. , & majewski , s. r. 2005 , , 619 , 807 letarte , b. , demers , s. , battinelli , p. , & kunkel , w. e. 2002 , , 123 , 832 lianou , s. , grebel , e.k . , & koch , a. 2010 , a&a , in press ( arxiv:1003.0861 ) majewski , s. r. , ostheimer , j. c. , patterson , r. j. , kunkel , w. e. , johnston , k. v. , & geisler , d. 2000 , , 119 , 760 marigo , p. , girardi , l. , bressan , a. , groenewegen , m. a. t. , silva , l. , & granato , g. l. 2008 , , 482 , 883 martinez - delgado , d. , & aparicio , a. 1997 , , 480 , l107 mauron , n. 2008 , , 482 , 151 mori , m. , & rich , r. m. 2008 , , 674 , l77 moro , d. , & munari , u. 2000 , , 147 , 361 mouhcine , m. , & lanon , a. 2003 , , 338 , 572 mould , j. , & aaronson , m. 1979 , , 232 , 421 nowotny , w. , kerschbaum , f. , schwarz , h. e. , & olofsson , h. 2001 , , 367 , 557 nowotny , w. , kerschbaum , f. , olofsson , h. , & schwarz , h. e. 2003 , , 403 , 93 ostheimer , j. c. j. 2003 , ph.d . thesis , university of virgina palma , c. , majewski , s. r. , siegel , m. h. , patterson , r. j. , ostheimer , j. c. , & link , r. 2003 , , 125 , 1352 rich , r. m. , reitzel , d. b. , guhathakurta , p. , gebhardt , k. , & ho , l. c. 2004 , , 127 , 2139 richardson , j. c. , et al . 2008 , , 135 , 1998 richardson , j. c. , et al . 2009 , , 396 , 1842 searle , l. , & zinn , r. 1978 , , 225 , 357 stanek , k. z. , & garnavich , p. m. 1998 , , 503 , l131 totten , e. j. , irwin , m. j. , & whitelock , p. a. 2000 , , 314 , 630

host - associated microbial communities are known to be of tremendous importance for host fitness , improving nutrient uptake , training the immune system , and resisting invasion by pathogens ( see , for example , brestoff & artis , 2013 ; kamada et al . , 2013 ; fredricks ( ed . ) , 2013 ) . our understanding of these communities , however , remains remarkably poor . the origin , maintenance , and importance of community diversity ( fierer et al . , 2011 ) , the factors determining community stability and resilience ( shade et al . , 2012 ) , and the mechanisms of community assembly ( costello et al . , 2012 ) are only some of the questions driving this rapidly expanding field . although most microorganisms can not be cultured in a laboratory setting , advances in genome - sequencing technology now allow organisms to be probed in their natural environments . in particular , the 16s rrna tag - sequencing approach identifies community members using fragments of dna from the hypervariable regions of the ribosomal 16s gene . the development of this technique and the decreasing cost of high - throughput sequencing have prompted a large number of tag - sequencing experiments , including such large - scale efforts as the human microbiome project or the earth microbiome project . the amount of collected data is growing exponentially . however , our ability to interpret this data still has important limitations . the _ de facto _ standard approach to 16s data analysis begins by clustering reads by sequence similarity into `` operational taxonomic units '' ( otus ) ; see fig . 1a ( quince et al . , 2009 ; kunin et al . , 2010 ; huse et al . , a variety of clustering techniques have been developed and are widely used in popular software tools or packages ( hunt et al . , 2008 ; schloss et al . , 2009 ; edgar , 2010 ; huang et al . , 2010 ; edgar et al . , 2011 ; quince et al . , 2011 ; schloss et al . , 2011 ; sul et al . , 2011 ; caporaso et al . , 2012 ; zheng et al . , 2012 ; morgan et al . , 2013 ; youngblut et al . , 2013 ) . despite significant progress in the development of such software , all clustering - based approaches suffer from a major shortcoming ( prosser et al . , 2007 ; hamady & knight , 2009 ; schloss & westcott , 2011 ) . although an otu is a useful concept for coarse - graining sequencing data , its definition is not biologically motivated , but as its name acknowledges is purely operational . sequences assigned to a particular otu are generally presumed to be close phylogenetic relatives and therefore likely to derive from ecologically similar bacterial subpopulations . however , the assumption that 16s sequence similarity is a good proxy for ecological similarity is notoriously problematic ( prosser et al . , 2007 ; preheim et al . , moreover , otu assignments are not definitive but depend on both the clustering algorithm and the random seed chosen ( schloss & westcott , 2011 ) . several approaches have been proposed to improve the resolution of 16s data analysis beyond the standard 97%-similarity otus . denoising algorithms exploit the predictable structure of certain error types to attempt to reassign or eliminate noisy reads ( huse et al . , 2010 ; quince et al . , 2011 ; rosen et al . , these algorithms are widely used for identifying low - abundance ( `` rare '' ) species against a noisy background , often with the aim of improving estimates of ecological diversity . these objectives , however , remain very challenging due to issues that no denoiser can fully address . any error model is necessarily approximate , and no denoising algorithm can deal with errors that are not adequately described by its error model ; when calling low - abundance species this issue becomes particularly problematic . an alternative approach termed distribution - based clustering ( dbc ; preheim et al . , 2013 ) aims to circumvent the limitations of conventional denoisers by using cross - sample comparisons , i.e. supplementing sequence information by ecological information ( distribution of abundance across multiple biological samples ) . however , dbc as an otu clustering algorithm also has important limitations : for low - count sequences , cross - sample comparisons necessarily become unreliable , and the execution time is prohibitively long even for moderately - sized datasets . here , we build on the above methods to address a distinct question . rather than trying to further improve the existing approaches to otu clustering and rare species identification , we combine error - model based denoising and systematic cross - sample comparisons to resolve the fine ( sub - otu ) structure of moderate - to - high abundance community members in 16s illumina data . importantly , our method does not rely on clustering similar sequences together . in this regard , our method is similar to oligotyping ( eren et al . , 2013 ) , but our approach does not require manual supervision and applies to an entire community rather than an isolated otu . using published data from a longitudinal study where the tongue community of two human individuals was sampled almost daily for several months ( caporaso et al . , 2011 ) , we demonstrate that sequence similarity is a very poor predictor of ecological similarity , which we quantify for two bacteria as the correlation of their abundance time traces ( `` dynamical similarity '' ) . thus , most clustering - based approaches would erroneously group together bacterial subpopulations of high ecological diversity for this data set . however , a comparative analysis of the tongue communities of the two individuals also shows that when a pair of 16s tags is observed in both individuals , the dynamical similarity of the pair as measured independently in the two individuals is highly correlated . this correlation falls off substantially when sequences differing by 1 nt out of 130 are compared . in other words , the exact sequence of the 16s tag carried by a bacterial subpopulation is predictive of its ecology , while even 99.2% similarity between tags of different subpopulations is generally not predictive of dynamical similarity , as defined above . our results lend support to the recent idea that even a purely 16s - based study can provide insight into functional relatedness of community members ( cf . picrust , langille et al . , 2013 ) , while also exhibiting and beginning to quantify the limitations of such methods . we demonstrate the applicability of our approach to a broad range of dataset types ( host - associated longitudinal ; environmental cross - sectional ; mock community ) , providing examples when highly similar sequences were found to exhibit ecologically significant distinctions . finally , we discuss how the single - nucleotide sub - otu resolution of our method can provide new insights into factors shaping community assembly . we used the raw data from a published long - term longitudinal sampling from four body sites ( gut : feces , right and left palm , and tongue ) of one male and one female individual ( caporaso et al . , 2011 ) . in this study , the hypervariable region v4 of the bacterial 16s rrna gene was amplified and sequenced with illumina ga - iix . for details on collection and sequencing see the original reference ( caporaso et al . , 2011 ) . quality - filtered data published with that study is available at mg - rast:4457768.3 - 4459735.3 and is sufficient to reproduce our results using provided analysis scripts ( see supplementary information ( si ) ) . however , to investigate the performance of our filtering approach at different quality filtering settings , for this work we used the demultiplexed , but not quality - filtered fastq data , kindly provided to us by the study authors . we split this data into per - sample fastq files using a custom matlab script ( mathworks , inc . ) and subjected it to minimal quality filtering using usearch v.7.0.1090 ( edgar , 2010 ) , truncating reads at phred quality score 2 ( other thresholds were also evaluated ; see fig . s4 ) , trimming to a fixed length of 130 nt and eliminating reads with ambiguous characters ( n ) . in addition , we removed reads with expected number of base call errors exceeding 1 ( maxee parameter in usearch ) . this criterion only eliminated 1% of trimmed reads . notably , our approach does not rely on assumptions about a maximum number of errors in a read . finally , to facilitate cross - sample comparisons , we compiled a library of all 1.4 m unique reads ever observed and a global table listing the abundances of each sequence across samples . this was done using dereplication capabilities of usearch and a custom perl script ( * mergeseqs.pl * ) . this script and others referenced in * bold * below are freely available at https://github.com/hepcat72/cff . finally , the abundance table was normalized to @xmath0 total reads per sample , to correct for varying sample size . read quality varied across lanes , so the number of reads after quality filtering was highest in a subset of tongue and fecal samples . in this work , we focused primarily on the tongue samples , as these come closest to probing the internal dynamics of a community living in a well - defined location on the body ; however , the analysis of fecal samples supports the same conclusions and is presented in fig . tongue samples were distributed over two lanes . the lane 6 samples from the male subject from day 65 onwards ( 314 consecutive samples covering a period of 355 days , @xmath1 reads in quality - filtered samples before normalization ) had approximately 4-fold more reads than those from the female subject and from days 1 - 64 of the male subject ( all on lane 5 ) . consequently , the analysis below uses the data from the male subject from day 65 onwards , and , for the comparative analysis of the two individuals , also the 135 samples collected from the female subject . the early samples from the male subject ( days 1 - 64 ) are only used for illustrative purposes ( fig . 3d ) . to demonstrate the broad applicability of our method we also employed other published data ( figs . s7 and s11 ) ; the data is described in the corresponding legends . clustering can be a useful strategy to coarse - grain 16s data while also reducing noise , but if sequencing noise is low enough , such coarse - graining may not be necessary . at low noise , each community member is predominantly represented by the same 16s sequence , surrounded by a cloud of low - abundance error sequences with the structure of the cloud determined by reproducible error rates . prior work has described such error clouds in the data ( quince et al . , 2009 ; edgar , 2013 ) , and the assumption that high - abundance sequences are likely to be error - free is used in several rank - based denoising and chimera - checking algorithms ( slp , perseus , uchime de novo , uparse , abundantotu ) . the treatment of reads that are very similar to high - abundance sequences is different across existing algorithms . for example , slp ( huse et al . , 2010 ) would consider any read differing by a single nucleotide from a higher - abundance sequence ( its `` direct neighbor '' in sequence space ) as an error . however , some of these reads may actually represent true community members ( preheim et al . , 2013 ) . a more nuanced treatment can accept a sequence as likely to be real if its observed abundance is highly unlikely to have arisen in error , given some assumptions about error rates . this idea is at the foundation of error - model based denoising . it was used in ampliconnoise ( quince et al . , 2011 ) , and its recent implementation in dada ( rosen et al . , 2012 ) makes dada , to our knowledge , the best denoiser currently available . however , no error model is perfect , and for all denoisers , errors not explicitly described by their model are labeled as true sequences . thus a denoising algorithm alone is insufficient for achieving sub - otu resolution : if two close sequences that would fall within a single otu are both identified as `` probably real '' , one of these could still be an error . in the context of a single sample , confidently resolving close sequences as independently real requires a different experimental technique ( faith et al . , 2013 ) or a complete , high - quality reference database of all bacteria in the sample , which in practice is available only for mock communities . it is possible to resolve this problem in the framework of standard 16s experiments through a comparison of multiple samples , either longitudinal or cross - sectional ( preheim et al . , 2013 ) . as an example , fig . 1b shows the abundances of the 10 highest - abundance direct neighbors of the overall top sequence of the tongue community , seq . # 1 , for a representative set of 100 consecutive samples . we see that three specific direct neighbors are strongly and consistently overrepresented compared to the other neighboring sequences and , more importantly , exhibit a dynamical behavior of their own ( consider , for example , the 3rd most abundant neighbor ) . this has a clear interpretation ( fig . 1c ) : these three sequences must belong to other , fairly abundant bacterial subpopulations , possibly related to seq . # 1 , but distinct and with their own dynamics . to achieve sub - otu resolution , we adopt precisely this strategy , namely a cross - sample correlation analysis of individually denoised samples . which denoiser should we use ? dada would be an excellent option ; however , its estimated execution time on the tongue dataset used here is @xmath2 sec ( see si ) . this is largely due to its exact treatment of probabilities , critically important for the processing of sequences with an abundance of just a few counts . however , for such sequences the imperfections of the error model become non - negligible and can not be controlled , since cross - sample comparisons are interpretable only for sequences with sufficient abundance . we therefore designed a new , simplified denoiser . our algorithm , described below , takes two orders of magnitude less time to execute , yet for sequences of moderate abundance considered here achieves performance equal to dada , as demonstrated using mock community data ( sup . table s2 ) . for 16s data obtained using the illumina platform , the main sources of errors are pcr substitutions , pcr chimeras , and substitution errors due to illumina base call errors . of these , the substitution errors are responsible for generating the largest number of unique sequences ( fig . s2 ; see also edgar , 2013 ) and have the most predictable structure : their rates can be estimated directly from the data . to do so , we considered the error clouds around the top 10 sequences by overall abundance ( in all tongue samples combined ) . assuming that most of these sequences are in fact errors , we determined the rates of specific one - nucleotide substitutions ( * errorrates.pl * with z - score threshold of 2 ; see si ) . these inferred rates were consistent across error clouds observed in the data ( fig . s3 ) , with the average error rate of only 0.10% per nucleotide ( sup . table 1 ; compare with quince et al . , 2011 , table 2 ) . we then used these error rates to predict the expected abundance of any given sequence if its presence were entirely due to independently generated sequencing errors of its more abundant neighbors ( the `` null model '' ; fig . s5 ; * nzeros.pl * ) . sequences whose abundance exceeded a threshold of 10 counts and the null - model prediction by at least 10-fold ( very conservative filtering parameters ) , were marked as `` candidates '' ; their presence can not be explained as an error within a substitution - only error model ( * getcandidates.pl * ) . candidate sequences include true biological 16s sequences , but also sequences that arose through a different type of error , most notably pcr chimeras . chimeric sequences were identified using uchime denovo ( edgar , 2011 ) on the pooled data from all samples . most such sequences were already eliminated by the abundance threshold requirement : if we relax the abundance threshold to 2 ( excluding singletons only ) , we find that the chimeras detected by uchime , when present in a sample , have abundance under 10 counts in 95% of cases . however , chimeras of highly abundant parents reproducibly occur at higher abundances ( haas et al , 2011 ) and are filtered at this step . candidate sequences that remained after filtering chimeras were labeled `` real '' . our highly conservative filtering criteria allow us to assume that this list contains only true biological sequences , i.e. that there are no false positives ( cf . table s2 ) , except possibly those due to some exceptionally frequent errors not described by our error model ( see si ) . this stringency comes at the expense of low - abundance false negatives ( true biological sequences labeled as `` possible noise '' ) . our strategy is to retain all sequences marked `` real '' in 2 or more samples ( out of 507 ; * getreals.pl * ) . this makes our denoiser specifically adapted to multi - sample analysis : in each sample , only high - confidence detections are identified , which is very fast , and then a liberal criterion applied across samples retains all sequences that ever generated a high - confidence detection , except sample singletons . in particular , we stress that our detection threshold of 10 counts is not equivalent to removing all sequences with abundance below 10 ; the only sequences excluded from consideration are those than never rise to 10 counts in the entire set of 504 tongue samples , or do so only once . for such sequences , the measured counts are dominated by detection and counting noise . in the interest of speed , and to ensure the robustness of reported sequence - abundance values with respect to the details of the error model , we did not attempt to remap noisy reads to their most probable source . our approach relies on the accuracy of measurement of relative abundances of true sequences . the error remapping process modifies sequence counts in a way that depends on the assumptions of the error model , distorting the relative abundance values whenever neighboring sequences are incorrectly classified as `` reals '' or `` errors '' . in contrast , discarding noisy reads leaves the relative abundances intact , as long as the probability of making zero errors is approximately constant across all sequences . this assumption is much weaker than adopting a particular error model . we estimate the zero - error probability at @xmath3% ( see si ) ; in other words , discarding noisy reads leads only to a @xmath4% loss of sequencing depth . if read remapping is desired , the analysis described below can be applied to dada denoiser output . since non - identical reads are never clustered together , ours is a single - nucleotide resolution approach . the complete workflow of cluster - free filtering is outlined in fig . s6 and detailed in the si . the code is freely available at https://github.com/hepcat72/cff . the starting point for our analysis is a global sequence abundance table listing the abundances of each unique 16s sequence across samples . we retained the 307 sequences that passed the multi - sample filtering algorithm described in methods , and thus putatively belong to bacteria present in the population at least part of the time . we denote these sequences by their overall abundance rank : seq . # 1 , # 2 , etc . in this list , 184 pairs of sequences were direct neighbors in sequence space ( hamming distance 1 ) . these pairs had 99.2% sequence similarity but were resolved by our criteria as independently present in the community . the population of bacteria sharing the exact same sequenced fragment of the 16s gene ( at 100% identity ) is the smallest taxonomic unit resolvable by 16s analysis . for notational convenience , throughout this work we call it the `` subpopulation '' identified by a sequence . in the standard approach to tag - sequencing data , it is assumed that sequence similarity of 16s hypervariable regions can be used as a proxy for phylogenetic , and therefore ecological relatedness . our new filtering method , applied to time - series data , allows us to bypass this assumption and assess ecological relatedness independently , based on the similarity of time traces , since each distinct subpopulation will respond in its own way to variation in environmental conditions ( youngblut et al . , 2013 ) , causing the abundance time traces to be more or less correlated ( or possibly anticorrelated ; see fig . 2a - c illustrates this by showing time traces ( normalized counts versus observation day ) for three examples of sequence pairs . we find that sequences differing by as little as 1 nucleotide ( 99.2% similarity ) can be ecologically distinct as evidenced by their very different time series ( fig . 2a ) ; see also vandewalle et al . , 2012 . for comparison , fig . 2b shows another pair of sequences , also with 99.2% sequence similarity but whose abundance time traces appear indistinguishable . the remarkable correlation between these two traces provides an internal control and demonstrates that the much lower correlation of traces in fig . 2a can not be explained by measurement error but reflects a true ecological difference . note that the abundances of the two sequences shown in fig . 2b are not equal , but occur with a highly stable ratio . this could reflect a stable difference in abundance of the bacteria they represent , but is more likely caused by differential amplification efficiency of these sequences by the pca primers ( turnbaugh et al . , 2010 ; klindworth et al . , 2013 ) and/or a different number of genomic 16s copies per cell ( tourova , 2003 ) . panels a - b show that sequence similarity need not imply ecological similarity . finally , fig . 2c illustrates that the converse is also true : sequences exhibiting identical time - dependence may have as little as 81% sequence identity . to quantify the generality of these examples , it is useful to define a measure of the ecological similarity of the bacterial subpopulations represented by two sequences . a natural candidate metric is the pearson correlation of the measured abundance traces . note , however , that the maximum correlation one can expect between the time traces of two sequences depends on their abundance : for low - abundance sequences poisson sampling noise becomes non - negligible and sets an upper bound on the correlation coefficient . we therefore define the `` dynamical similarity '' of two traces as the pearson correlation of their abundance , normalized by their maximum possible correlation @xmath5 , computed as the correlation of the higher - abundance time trace with a poisson - downsampled version of itself ( see si ) . for sequence distance , we use the hamming distance between sequences after pairwise alignment ( see si ) . with these definitions , we can present a 2d histogram of dynamical similarity vs. distance in sequence space for all sequence pairs constructed from the top 200 real sequences ( fig . as expected , most sequence pairs exhibit no significant dynamical similarity and are also far apart in sequence space , but a subset of closely similar sequences appears to display some degree of anticorrelation between the two measures . zooming in on this region ( fig . 2e ) makes this anticorrelation more apparent ; however , even when restricted to the subset shown in fig . 2e , the correlation coefficient remains weak ( @xmath6 ) . sequences separated by up to 6 - 7 nt ( 95% sequence similarity ) tend to be dynamically similar , the effect increasing for smaller distances , but this general trend is very loose and is not a reliable predictor of similarity for any particular pair . this result was not unexpected , and is frequently used in arguments against over - reliance on the 16s gene sequence ( see , for example , prosser et al . , 2007 ) , in favor of methods providing functional information , such as shotgun metagenomics . the novelty of fig . 2e lies in the fact that it was obtained entirely within the framework of 16s tag sequencing methodology . as explained in the previous section , 16s tags with low dynamical similarity clearly derive from distinct bacterial subpopulations , even if the sequences are themselves highly similar . we now consider pairs of sequences with highly correlated time traces such as observed in fig . 2b , c. such correlated pairs could derive from the same bacterial cells ( as multiple genomic copies of the 16s gene , or as exceptionally common pcr errors not included in our model ) . alternatively , they could derive from distinct bacterial subpopulations that either occupy the same ecological niche or engage in a strong obligate symbiosis . such pairs are thus of significant ecological interest , provided it can be shown that the sequences actually derive from different bacterial cells . in this section , we demonstrate that cross - sample correlation analysis can , in some cases , successfully make this subtle distinction between same - cell or different - cell sources . to draw this distinction , we make use of the following observation . the abundance ratio of two sequences that derive from the same bacterium is set by some sample - independent parameter ( e.g. involving differential amplification efficiency , 16s copy number , and/or pcr error rate ) ; therefore , any fluctuation in their abundance ratio is due to measurement noise , and must be uncorrelated between samples . any statistically significant time ( or location ; see si ) correlation of abundance ratio fluctuations , e.g. in consecutive ( or proximate ) samples , is therefore strong evidence that the two sequences are at least partially contributed by physically distinct subpopulations . for this approach to succeed , the dynamics of individual subpopulations must be slow enough to allow correlations between consecutive samples to be observed . we therefore began by computing , for each of the top 100 sequences , the autocorrelation function @xmath7 , defined as the correlation between abundance fluctuations in samples separated by @xmath8 time points , and normalized so that @xmath9 ( for simplicity , we treat samples as though they were equally spaced in time , which is approximately correct ; the mean separation between samples was 1.1 days ) . the environment experienced by tongue microorganisms changes frequently , and one might have expected that daily sampling would probe the space of possible community states , but provide little information about community dynamics as these would occur on a faster time scale . surprisingly , we found the time dependence of most sequences in the top 100 to have a significant autocorrelation despite the relatively low sampling rate ( fig . although conditions on the tongue make fast abundance changes possible , as evidenced by the large , rapid fluctuations in fig . 2a - c , we found the correlation time for the top 100 sequences to be surprisingly long , typically 2 - 4 days but often longer ( fig . 3b ) , sometimes exceeding a month ( fig . these multi - day autocorrelations make it plausible that for physically distinct subpopulations , the fluctuations of their abundances relative to each other could be slow enough to be detectable even if their ecology is similar . consider two sequences @xmath10 and @xmath11 whose abundance time traces are highly correlated . denote by @xmath12 , @xmath13 the two traces renormalized to the same mean for best overlap , as in fig . 2b , c , and let @xmath14 be their fractional difference in a given sample ( a quantity more robust to noise than the nave abundance ratio ) : @xmath15 if @xmath16 reflects abundances of two distinct subpopulations , then @xmath14 can be expected to exhibit an autocorrelation on par with that observed for the individual sequences . intuitively , if on day 1 , subpopulation @xmath10 is , say , 10% more abundant than @xmath11 , and the dynamics of both are slow , then @xmath10 is likely to maintain its lead on day 2 . in contrast , if the two sequences are genomic variants contained within the same bacterium , then any difference between @xmath12 and @xmath13 must be due to measurement noise , and @xmath14 will be uncorrelated between samples . we therefore introduce the _ persistence of difference _ @xmath17 as the 1-day autocorrelation coefficient of @xmath14 : @xmath18 where angular brackets denote averaging over time . @xmath17 characterizes the persistence of abundance fluctuations of two sequences relative to each other . for sequences arising from the same cells , @xmath17 must vanish . any pair of sequences exhibiting a statistically significant @xmath17 must be contributed , at least in part , by two physically distinct bacterial subpopulations . note that the absolute abundance of a sequence may change dramatically between days ( e.g. more favorable conditions can cause both subpopulations to proliferate quickly ) , but the normalization of @xmath14 makes @xmath17 insensitive to such overall correlated behavior . summarizing the above , we have the following expectation for @xmath17 : for a randomly chosen pair of sequences , with insignificant dynamical similarity , @xmath17 should be significantly non - zero ( due to the slow dynamics of the individual subpopulations ; see si ) , and form a unimodal distribution consistent with the null model of unrelated subpopulations . in contrast , pairs displaying high dynamical similarity come in two types , and the persistence of difference @xmath17 should display a bimodal distribution : pairs of sequences found within the same bacterial cell will have vanishing or insignificant @xmath17 , while pairs belonging to distinct subpopulations will likely exhibit a persistence of difference comparable with the null model prediction . this is precisely what we observe . 3c shows , for all sequence pairs constructed from the top 100 sequences , a scatter plot of their persistence of difference @xmath17 versus dynamical similarity as defined previously ( the normalized pearson correlation of their abundances ) . the mean and standard deviations of the distribution predicted by the null model ( unrelated subpopulations ) are indicated by the green ellipse , and were computed directly from the data by reversing in all pairs the time order for one of the sequences . the mean and standard deviations of the actual data are indicated by the red cross . we find , as expected , that the @xmath17 score of dynamically dissimilar sequence pairs is unimodal and consistent with the null - model prediction . in contrast , the @xmath17 score of dynamically similar pairs exhibits the predicted bimodality ( right side of the plot ) , with a subset exhibiting weak or negligible persistence of difference ( bottom right ) . as explained above , we interpret these low-@xmath17 pairs as corresponding to genomic 16s variants found within a single bacterium . letters a - c identify pairs shown on fig . note that the strong persistence of difference identifies the pair `` b '' as being contributed , at least in part , by distinct bacterial cells , despite 99.2% sequence similarity and an almost perfect correlation of abundances ( fig . conversely , the low-@xmath17 pair `` c '' ( with only 81% sequence similarity ) likely corresponds to an example of two dissimilar 16s genes contained within a single bacterium . note the enrichment of pairs with high sequence similarity among the dynamically similar pairs , as indicated by the color code ( compare with fig . remarkably , in the case of pair `` b '' , the conclusion of distinct bacterial subpopulations drawn from fig . 3c can be confirmed directly . panel d shows the time traces of this pair for days 1 - 64 ( normalization as in fig . due to the relatively poor sequencing depth in these early samples , they were not included in fig . the clear separation observed prior to day 40 provides an independent confirmation of our conclusion . we stress that these data were not used in the analysis presented in fig . 3c , but the sensitivity of the autocorrelation method was sufficient to identify these sequences as deriving from physically distinct cells based solely on the data shown in fig . the autocorrelation function of the fractional difference @xmath14 for this pair is shown in fig . we have verified that the persistence of difference for this pair does not change significantly if any window of 100 consecutive samples is used instead of the full time series ( data not shown ) . figs . 2a , s7 , s10 , and s11 provide examples of some fine features that standard otu - based methods would fail to detect , but which become accessible with cluster - free filtering . we now ask whether such cases are the exception or the rule . for a given sequence similarity threshold , we can define , for each of the most abundant sequences , its would - be otu , namely the ensemble @xmath19 of all `` real '' sequences within the chosen similarity threshold . we construct the time trace of the abundance of this otu as the sum of the abundances of all its members . we can now ask : how representative is this time trace of the true behavior of the member sequences ? let @xmath20 be the correlation coefficients between time traces of individual members and the otu itself , normalized to the maximum expected correlation as before . we define unweighted and weighted otu quality scores @xmath21 and @xmath22 as , respectively , the simple average of @xmath20 , and an average weighted by the abundance of the member : @xmath23 here @xmath24 is the number of subpopulations in the otu and @xmath25 is the average abundance of member @xmath26 . the weighted quality score @xmath22 is always larger , because the most abundant sequence dominates the sum and so is better correlated with the otu trace . thus @xmath22 tells us how representative the otu is of its most abundant member . the unweighted quality score @xmath21 tells us how diverse is the group of subpopulations lumped together into an otu . if the sequences grouped into an otu are all dynamically identical ( are poisson - resampled versions of each other at different abundances ) , both quality scores will be close to 1 . if the otu is dominated by one subpopulation , with other members dynamically different but very low in abundance , we will have @xmath27 , but @xmath28 . finally , if the otu contains several dynamically distinct subpopulations at comparable abundances , both quality scores will be low . the average quality scores for otus assembled around the top 5 sequences are presented in fig . 4 as a function of sequence similarity threshold . the relatively high weighted quality score @xmath22 means that an otu time trace is , on average , fairly representative of its most abundant member . the unweighted score @xmath21 is , however , dramatically lower , indicating that the otus group together sequences from subpopulations with high dynamical diversity . these quality scores rely on abundance time - trace correlations , which become contaminated with noise for low - abundance sequences . for the purposes of fig . 4 , to apply these definitions conservatively , we therefore restricted our attention only to high - abundance members of the otu , considering only sequences from the top 200 by overall abundance . further , our cluster - free filtering method also has finite resolution , as the sequences we analyze are only 130nt long and may derive from distinct 16s genes , implying some unresolved diversity . this limited resolution leads to an artificial inflation of otu quality scores as the similarity threshold approaches 100% . for both these reasons the true quality scores of otus are likely even lower ( see si ) . the fact that tag sequence similarity within the 16s gene is only loosely correlated with dynamical similarity ( fig . 2e ) was not unexpected ( see , for example , prosser et al . , 2007 and references therein ) . at a neutral mutation rate of order 10 - 9 per base pair per generation ( ochman , 2003 ) , an average difference of a single nucleotide out of 100 would already require divergence for millions of generations . a more precise estimate of divergence time should take into account the possibility of horizontal gene transfer , whose rate in an ecologically relevant setting is hard to assess . however , it is clear that , generically , two bacteria that differ by even 1 nt in a particular hypervariable region of the 16s gene likely diverged a long time ago . these bacteria are likely to also differ elsewhere in their 16s gene , and to carry even more significant differences in functional parts of their genome . in contrast , what if we consider two bacteria whose sequenced portions of their 16s genes are identical ? since the length of the sequenced fragment is small ( typically @xmath29 nt ) and the mutation rate is low , these bacteria could still have diverged a very long time ago ( lukjancenko et al . , 2010 ) . however , depending on circumstances , the actual time since the last common ancestor may be much shorter . for example , consider two communities that frequently exchange members . if two bacteria drawn from two such communities are 100% identical in their 16s tags , a likely explanation for this identity is a recent exchange event , in which case the entire genomes of these bacteria may be close to identical . we conclude that in the presence of strain exchange between communities , exact sequence identity and near - identity may have fundamentally different implications . the study of caporaso et al . sampled the tongue microbiota of two cohabiting individuals ( rob knight , personal communication ) , and so strain exchange is likely to be a highly significant factor ( song et al . , 2013 ) . we hypothesized , therefore , that these communities would share some non - negligible number of subpopulations at 100% sequence identity , and that these common subpopulations might have similar ecology in both communities . we began by identifying the fraction of common 16s sequences in the list of the top n for each individual ( at 100% sequence identity ) . based on our strain exchange hypothesis , we expected to find some matches , but were still surprised to find this fraction to be as high as 75% ( fig . such a high proportion of perfect matches provides strong evidence that the identical sequences found in these two communities most likely diverged from a common ancestor more recently than any pair of close , but non - identical sequences within the same community . the same conclusion is supported by the analysis of fecal samples from the two individuals ( fig . we then considered the 73 sequences that were found among the top 100 of both individuals and asked whether the behavior of these subpopulations was predominantly shaped by their presumed common origin ( causing them to be similar ) or by local adaptation ( causing them to diverge while leaving the 16s region intact ; see lukjancenko et al . , 2010 ) . to this end , for each pair of sequences @xmath30 drawn from this list , we measured their dynamical similarity independently in the two datasets ; @xmath31 for the male and @xmath32 for the female . if the effect of local adaptation were dominant , then the exactness of a match of 16s sequences would not carry much information : the ecologies and genomes would be no more similar between 100%-identical partners in the two communities than between any other sequences within the same bacterial `` species '' ( otu ) ; this scenario is implicitly assumed by taxonomy - based methods . alternatively , if the ecology were determined primarily by the shared recent ancestor , then identical 16s tag sequences in the two communities would correspond to bacterial subpopulations with almost identical genomes . in this scenario , provided local adaptation did not modify the ecology of a subpopulation significantly , @xmath31 and @xmath32 should be strongly correlated , and unlike the first scenario , this correlation would be noticeably degraded for any less than 100% sequence identity . the latter is indeed what we observe ( fig . 5b demonstrates that subpopulations identified by the exact same 16s tags in the two individuals are dynamically similar ; see also fig . s11d and s12 . to obtain fig . 5c , we constructed an `` inexact pairing '' of sequences between individuals , whereupon each sequence from the top 100 in the female individual was matched to the highest - abundance sequence from the top 100 in the male individual that differed from it by exactly 1 nucleotide , when such a match existed . this matching corresponds to 99.2% sequence identity , yet already substantially degrades the correlation between @xmath31 and @xmath32 ( fig . we conclude that 100% identity of tag sequences has qualitatively different implications from even 99.2% near - identity . in this work , we have demonstrated that cross - sample correlation analysis of denoised 16s data can be exploited to achieve sub - otu resolution . the cluster - free filtering approach we presented reliably identified up to 20 distinct subpopulations within standard 97% similarity otus , and a comparative analysis of oral communities of two cohabiting individuals demonstrates that most such subpopulations are shared between the two communities . furthermore , subpopulations identified by the exact same 16s tags in the two individuals are dynamically similar , whereas even a single nucleotide mismatch is enough to degrade this similarity . overall , our analysis shows that coarse - graining sequence data into otus is not essential for ecological applications of 16s tag sequencing methodology . our approach combines two novelties . first and foremost , we do not cluster similar sequences together . regrettably , in the literature the term `` clustering '' has multiple meanings . most denoising algorithms aim to assign erroneous reads to their most likely source , to make the abundance estimates of true sequences more accurate . the same term `` clustering '' is used both for this read remapping and for merging multiple true sequences into a single otu . however , these two practices are fundamentally different . read remapping constitutes data denoising ; as such , it is always advantageous , can be done in a principled way , and can be evaluated against an objective standard of performance . adding it to our approach would likely somewhat improve the results . in contrast , otu clustering is a form of data coarse - graining , and the optimal degree of coarse - graining is necessarily application - dependent . importantly , for some applications it may not be necessary or desirable . when studying coarse features of community composition and dynamics , e.g. comparing communities across habitats ( costello et al . , 2009 , huttenhower et al . 2012 ) coarse - graining is appropriate . for example , metrics of community comparison such as unifrac ( lozupone & knight , 2005 ) are widely used precisely because , by construction , they are not sensitive to otu sub - structure . however , when studying subtle differences between broadly similar communities , e.g. samples from similar habitats or repeated sampling of the same habitat , the sub - otu structure becomes a valuable source of insight . this is the intended application for our approach . although we focused on longitudinal illumina data , the denoising algorithm we developed does not assume short read length or low error rate and is directly applicable to a wide range of dataset types ( see examples in fig . s7 and s11 ) , provided the error structure is consistent across samples ( preheim et al . , we expect our approach to be useful for investigating the structure and dynamics of discrete community subtypes such as those observed in the vaginal community ( huttenhower et al . 2012 ) . our second novelty is to exploit the quantitative advantage offered by multi - sample ( time - course or cross - sectional ) data . since the copy number of the 16s gene carried by a bacterium is typically unknown ( tourova , 2003 ) , and the pcr amplification bias among different 16s fragments can sometimes reach orders of magnitude ( turnbaugh et al . , 2010 ; klindworth et al . , 2013 ) , the 16s data from a single sample carries very little quantitative information about community composition . in contrast , the ratios of sequence abundance are highly informative and can be measured very precisely , as demonstrated in fig . recently , time - course data collection has been gaining in popularity , as it was recognized that such experiments can offer valuable insight into community dynamics ( shade et al . , 2013 and references therein ) . however , another major advantage of such datasets , namely that changes in sequence abundance ratios can be measured much more accurately than absolute abundances , is only beginning to be explored . for us , time - series data provides a context where sub - otu resolution acquires its full power . specifically , we have shown that cross - sample comparisons enable us to decouple sequence similarity from dynamical similarity while remaining fully within the framework of 16s tag sequencing . high - quality reference databases can complement our approach to facilitate paralog identification . the basic methodology described here should also be extendable to other marker genes . the new approach described in this work is not a replacement for otu clustering ; it discards low - abundance sequences and so is unsuitable for studies of population - level alpha- or beta - diversity . however , the novel statistical and computational techniques we present allow full utilization of the quantitative information carried by sequences with a moderate - to - high abundance . this has promising applications for the study of factors affecting community assembly . as discussed above , sub - otu resolution can provide insight into the prevalence of strain exchange between communities , invasion / extinction dynamics of otu subpopulations , and the time scale of ecological divergence relative to sequence divergence . in addition , the dynamics of individual - specific subpopulations could help characterize the role of host genetics or the host immune system on shaping the community , particularly in the context of highly controlled experiments with germ - free animals . we thank w bialek , af bitbol , cp broedersz , ds fisher , r knight and members of his lab , sa levin , y meir , sw pacala , sp preheim and mj rosen for helpful discussions , g caporaso for providing the data , and r edgar for usearch support . this work was partially supported under the darpa biochronicity program , grant d12ap00025 , and national science foundation grants phy-1305525 and ccf-0939370 . 99 biblabel#1 @bibitem#1@bibitem#1- brestoff jr , artis d ( 2013 ) . commensal bacteria at the interface of host metabolism and the immune system . _ nat immunol _ * 14 * : 676 - 684 . caporaso jg , lauber cl , costello ek , berg - lyons d , gonzalez a , stombaugh j et al ( 2011 ) . moving pictures of the human microbiome . _ genome biology _ * 12 * : r50 . caporaso jg , lauber cl , walters wa , berg - lyons d , huntley j , fierer n et al ( 2012 ) . ultra - high - throughput microbial community analysis on the illumina hiseq and miseq platforms . _ isme j _ * 6 * : 1621 - 1624 . costello ek , lauber cl , hamady m , fierer n , gordon ji , knight r ( 2009 ) . bacterial community variation in human body habitats across space and time . _ science _ * 326 * : 1694 - 1697 . costello ek , stagaman k , dethlefsen l , bohannan bjm , relman da ( 2012 ) . the application of ecological theory toward an understanding of the human microbiome . _ science _ * 336 * : 1255 - 1262 . desantis tz , hugenholtz p , larsen n , rojas m , brodie el , keller k et al ( 2006 ) . greengenes , a chimera - checked 16s rrna gene database and workbench compatible with arb . _ appl environ microbiol _ * 72 * : 5069 - 5072 . edgar rc ( 2004 ) . muscle : multiple sequence alignment with high accuracy and high throughput . _ nucleic acids res . _ * 32 * : 1792 - 1797 edgar rc ( 2010 ) . search and clustering orders of magnitude faster than blast . _ bioinformatics _ * 26 * : 2460 - 2461 . edgar rc , haas bj , clemente jc , quince c , knight r ( 2011 ) . uchime improves sensitivity and speed of chimera detection . _ bioinformatics _ * 27 * : 2194 - 2200 . eren am , maignien l , sul wj , murphy lg , grim sl , morrison hg et al ( 2013 ) . oligotyping : differentiating between closely related microbial taxa using 16s rrna gene data . _ methods in ecology and evolution _ * 4 * : 11111119 . faith jj , guruge jl , charbonneau m , subramanian s , seedorf h , goodman al et al ( 2013 ) . the long - term stability of the human gut microbiota . _ science _ * 341 * : 1237439 . fierer n , lennon jt ( 2011 ) . the generation and maintenance of diversity in microbial communities . _ am j bot _ * 98 * : 439 - 448 . fredricks dn ( 2013 ) . _ the human microbiota : how microbial communities affect health and disease . _ wiley - blackwell : hoboken , n.j . haas bj , gevers d , earl am , feldgarden m , ward dv , giannoukos g et al ( 2011 ) . chimeric 16s rrna sequence formation and detection in sanger and 454-pyrosequenced pcr amplicons . _ genome res _ * 3 * : 494 - 504 . hamady m , knight r ( 2009 ) . microbial community profiling for human microbiome projects : tools , techniques , and challenges . _ genome res _ * 19 * : 1141 - 1152 . huang y , niu bf , gao y , fu lm , li wz ( 2010 ) . cd - hit suite : a web server for clustering and comparing biological sequences . _ bioinformatics _ * 26 * : 680 - 682 . hunt de , david la , gevers d , preheim sp , alm ej , polz mf ( 2008 ) . resource partitioning and sympatric differentiation among closely related bacterioplankton . _ science _ * 320 * : 1081 - 1085 . huse sm , welch dm , morrison hg , sogin ml ( 2010 ) . ironing out the wrinkles in the rare biosphere through improved otu clustering . _ environ microbiol _ * 12 * : 1889 - 1898 . huttenhower c , gevers d , knight r , abubucker s , badger jh , chinwalla at et al ( 2012 ) . structure , function and diversity of the healthy human microbiome . _ nature _ * 486 * : 207 - 214 . kamada n , chen gy , inohara n , nunez g ( 2013 ) . control of pathogens and pathobionts by the gut microbiota . _ nat immunol _ * 14 * : 685 - 690 . klindworth a , pruesse e , schweer t , peplies j , quast c , horn m et al ( 2013 ) . evaluation of general 16s ribosomal rna gene pcr primers for classical and next - generation sequencing - based diversity studies . _ nucleic acids res _ * 41*. kunin v , engelbrektson a , ochman h , hugenholtz p ( 2010 ) . wrinkles in the rare biosphere : pyrosequencing errors can lead to artificial inflation of diversity estimates . _ environ microbiol _ * 12 * : 118 - 123 . langille mgi , zaneveld j , caporaso jg , mcdonald d , knights d , reyes ja et al . ( 2013 ) predictive functional profiling of microbial communities using 16s rrna marker gene sequences . _ nature biotechnol _ * 31 * : 814821 . lozupone c , knight r ( 2005 ) . unifrac : a new phylogenetic method for comparing microbial communities . _ appl environ microbiol _ * 71 * : 8228 - 8235 . lukjancenko o , wassenaar tm , ussery dw ( 2010 ) . comparison of 61 sequenced escherichia coli genomes . _ microbial ecology _ * 60 * : 708 - 720 . morgan mj , chariton aa , hartley dm , court ln , hardy cm ( 2013 ) . improved inference of taxonomic richness from environmental dna . _ plos one _ * 8 * : e71974 . ochman h ( 2003 ) . neutral mutations and neutral substitutions in bacterial genomes . _ mol biol evol _ * 20 * : 2091 - 2096 . preheim sp , perrotta ar , martin - platero am , gupta a , alm ej ( 2013 ) . distribution - based clustering : using ecology to refine the operational taxonomic unit . _ appl environ microbiol _ * 79 * : 6593 - 6603 . prosser ji , bohannan bjm , curtis tp , ellis rj , firestone mk , freckleton rp et al ( 2007 ) . essay the role of ecological theory in microbial ecology . _ nature reviews microbiology _ * 5 * : 384 - 392 . quince c , lanzen a , curtis tp , davenport rj , hall n , head i m et al ( 2009 ) . accurate determination of microbial diversity from 454 pyrosequencing data . _ nat methods _ * 6 * : 639-u627 . quince c , lanzen a , davenport rj , turnbaugh pj ( 2011 ) . removing noise from pyrosequenced amplicons . _ bmc bioinformatics _ * 12 * : 38 . rosen mj , callahan bj , fisher ds , holmes sp ( 2012 ) . denoising pcr - amplified metagenome data . _ bmc bioinformatics _ * 13 * : 283 . schloss pd , westcott sl , ryabin t , hall jr , hartmann m , hollister eb et al ( 2009 ) . introducing mothur : open - source , platform - independent , community - supported software for describing and comparing microbial communities . _ appl environ microbiol _ * 75 * : 7537 - 7541 . schloss pd , gevers d , westcott sl ( 2011 ) . reducing the effects of pcr amplification and sequencing artifacts on 16s rrna - based studies . _ plos one _ * 6*. schloss pd , westcott sl ( 2011 ) . assessing and improving methods used in operational taxonomic unit - based approaches for 16s rrna gene sequence analysis . _ appl environ microbiol _ * 77 * : 3219 - 3226 . shade a , peter h , allison sd , baho d , berga m , buergmann h et al ( 2012 ) . fundamentals of microbial community resistance and resilience . _ frontiers in microbiology _ * 3*. shade a , caporaso jg , handelsman j , knight r , fierer n ( 2013 ) . a meta - analysis of changes in bacterial and archaeal communities with time . _ isme j _ * 7 * : 1493 - 1506 . song sj , lauber c , costello ek , lozupone ca , humphrey g , berg - lyons d et al ( 2013 ) . cohabiting family members share microbiota with one another and with their dogs . _ elife _ * 2*. sul wj , cole jr , jesus ed , wang q , farris rj , fish ja et al ( 2011 ) . bacterial community comparisons by taxonomy - supervised analysis independent of sequence alignment and clustering . _ proc natl acad sci _ * 108 * : 14637 - 14642 . tourova tp ( 2003 ) . copy number of ribosomal operons in prokaryotes and its effect on phylogenetic analyses . _ microbiology _ * 72 * : 389 - 402 . turnbaugh pj , quince c , faith jj , mchardy ac , yatsunenko t , niazi f et al ( 2010 ) . organismal , genetic , and transcriptional variation in the deeply sequenced gut microbiomes of identical twins . _ proc natl acad sci _ * 107 * : 7503 - 7508 . vandewalle jl , goetz gw , huse sm , morrison hg , sogin ml , hoffmann rg et al ( 2012 ) . acinetobacter , aeromonas and trichococcus populations dominate the microbial community within urban sewer infrastructure . _ environ microbiol _ * 14 * : 2538 - 2552 . youngblut nd , shade a , read js , mcmahon kd , whitaker rj ( 2013 ) . lineage - specific responses of microbial communities to environmental change . _ appl environ microbiol _ * 79 * : 39 - 47 . zhang z , schwartz s , wagner l , miller w ( 2000 ) . a greedy algorithm for aligning dna sequences , _ j comput biol _ * 7 * : 203 - 214 . zheng zj , kramer s , schmidt b ( 2012 ) . dysc : software for greedy clustering of 16s rrna reads . _ bioinformatics _ * 28 * : 2182 - 2183 . in this section , we illustrate the idea of error - model - based denoising ( see also the introduction in rosen et al . , 2012 ) and give a detailed description of the simple denoiser we designed for this work . we then describe the workflow of an open source software package we created to implement this denoiser , and compare its performance on mock community data with dada ( rosen et al . , 2012 ) . finally , to illustrate that our cluster - free filtering approach is not restricted to longitudinal illumina data , we provide an example of its application to a very different dataset , specifically 454 sequencing data from a cross - sectional environmental sampling performed by preheim et al . , 2013 . clustering can be a useful strategy for filtering noise by coarse - graining data . however , such coarse - graining may not be a necessity : if the noise level is low , as suggested by known estimates of pcr and sequencing error rates ( see , for example , quince et al . 2011 ) , then we can avoid clustering , since we expect each community member to be predominantly represented by the same 16s sequences . we begin by illustrating this idea using the tongue microbiome data of caporaso et al . since the tongue community is relatively stable ( costello et al . , 2009 ) , the low - noise scenario would predict that certain specific sequences should consistently dominate in each sample . alternatively , if the noise were high , then the high - abundance community members would be represented by clouds of similar reads , none of which would clearly dominate . to show that the data of caporaso et al . supports the first ( low - noise ) scenario , we identified the top 5 sequences by overall abundance . these sequences were strongly different ( fig . [ fig : rank ] , inset ) , corresponding for the most part to bacteria from different phyla : in decreasing order of abundance , these were _ neisseria sp . _ ( phylum _ proteobacteria _ , class _ betaproteobacteria _ ) , _ haemophilus sp . _ ( phylum _ proteobacteria _ , class _ gammaproteobacteria _ ) , _ fusobacterium sp . _ ( phylum _ fusobacteria _ ) , _ streptococcus sp . _ ( phylum _ firmicutes _ ) , and _ prevotella sp . _ ( phylum _ bacteroidetes _ ) . ( taxonomy assigned by a blast search ( blastn 2.2.22 , matrix @xmath33 , gap / extenstion penalty @xmath34 ) against greengenes database ; desantis et al . , 2006 . all 5 sequences had a match with 100% identity over 100% of sequence length . ) the sample - by - sample rank of these overall top 5 sequences was consistently in the top 10 . we stress that the goal of fig . [ fig : rank ] is not to characterize the temporal stability of community composition ( previously characterized , for example , in costello et al . , 2009 ) ; rather , it serves to show that the community members that correspond to these highest - abundance tags are consistently represented by the same 130nt sequence ( at 100% identity ) across all samples . in other words , despite the presence of noise in the data , 100% sequence identity is not an unreasonable criterion : the error rate is low enough that the error - free sequence dominates over the `` error cloud '' of its variants ( edgar , 2013 ) . this key observation is the foundation of the approach described in this work . to estimate the rates of substitution errors observed in data after quality filtering , we used the `` error clouds '' around the high - abundance sequences in the dataset . since all sequences were trimmed to a length of 130nt , each `` mother '' sequence has 390 direct neighbors in sequence space ( @xmath35 ) . for very high - abundance sequences such as seq . # 1 , all 390 neighbors were observed in at least one sample of the time series . the time series of their abundances , normalized to the abundance of seq . # 1 , is shown in fig . [ fig : errorcloud ] . for this figure , the neighbors were ordered by the type of substitution that differentiates them from the mother sequence , and , within these categories , by the position of the differing nucleotide along the sequence . we see that , with a few exceptions ( most notably the three neighbors also shown in fig . 1b ) , the abundance of a given neighbor is a constant fraction of the abundance of the mother sequence . this is precisely what we expect for neighbors that arise as pcr or sequencing errors of the mother sequence , and the abundance ratio is then the probability of that particular error . we see that the error rate is set primarily by the type of substitution , and does not exhibit significant dependence on the position along the sequence . for long reads , we would likely have seen an increase in error rates towards the end of the sequence , but our sequences are only 130nt long , well within the capabilities of accurate base - calling of the illumina platform . we can therefore assign probabilities to substitution errors based solely on the substitution type ( which nucleotide was replaced by which other ) , independent of the position along the read . to determine these probabilities , we first identify the neighbors that are outliers in their substitution category ; they likely correspond to true biological sequences physically present in the community , rather than sequencing errors . outlier exclusion is done based on z - scores , i.e. for each sequence we compare its raw cumulative abundance over all samples to the mean in its substitution category , and normalize by the standard deviation in the category . a strong outlier differing from the mother sequence by a nucleotide substitution at location @xmath24 will skew the error rate estimation at that location : some substitution type will appear to be unusually frequent . therefore , we exclude nucleotide locations that correspond to the strongest outliers , those for which the z - score exceeds some threshold . the remaining locations are then used to estimate the error rates : for each of these locations , we count the number of times a particular substitution occurred , as well as the number of times the nucleotide was recorded correctly . after appropriate normalization , these counts give us the probability of each type of the error . [ fig : errorrates ] shows the inferred substitution rates for error clouds around the top 50 sequences by abundance , using minimally quality - filtered data processed as described in the methods ( phred score cutoff 2 , z - score threshold 2 ) . we find that the rates of different substitutions can differ dramatically ( up to 50-fold ) , but our estimates are highly reproducible ( note the log scale on the y axis ) , with variability predictably increasing if lower - abundance error clouds are used . table s1 lists the error rates estimated from the error clouds of the top 10 sequences ( mean @xmath36 standard deviation ) . note that , in principle , this effective error probability includes both the base - call errors of the illumina sequencer and the single - nucleotide substitution errors occurring during pcr . however , the approximate symmetry between rates of a substitution and its reverse - complement partner ( e.g. @xmath37 ) , and a clear bias towards transitions as opposed to transversions , suggests that the observed substitutions are dominated by pcr errors ( compare with quince et al . , 2011 , table 2 ) . inferring error rates directly from the data offers multiple strong advantages . specifying the error rate as an external parameter ( e.g. , morgan et al . , 2013 ) necessarily requires resorting to a conservative global upper bound . different pcr conditions and different sequencing machines will have different error rates ( for example , compare fig . [ fig : qcparams ] and fig . [ fig : dbc]a ) . further , substitutions vary strongly in probability : in our case , using a single upper bound on error rates would have over - estimated the probability of certain error types by up to 50-fold , reducing our ability to resolve close sequences . in other words , measuring substitution rates directly from the data both reduces the number of algorithm parameters and improves performance . .substitution error rates per nucleotide , multiplied by 1000 , as measured from the `` error clouds '' of the top 10 sequences by abundance . error bars are standard deviations across the 10 estimates . [ cols="<,^,^,^,^,^",options="header " , ] the approach described in this work does not explicitly rely on the longitudinal nature of the sampling . most of our analysis can be readily applied to any multi - sample datasets , e.g. a cross - sectional sampling or a location series , provided samples were collected and processed in a similar way so that the error structure can be assumed to be similar . further , and despite the caveats we described , our method can be applied even to data collected using the 454 sequencing platform . to illustrate the broad applicability of our approach , we used data from a cross - sectional environmental sampling conducted by preheim et al . ( sra accession number from srp029470 ) . lake water microbiota were sampled at depths ranging from 0 m ( surface ) to 22 m with 1-meter depth intervals . the authors used this data to illustrate their sequence clustering algorithm ( dbc ) that also relies on cross - sample comparisons to distinguish between closely related otus ; for details , see the original reference ( preheim et al . , 2013 ) . they report their algorithm worked best with stringent quality filtering whereupon sequences were trimmed to just 76 nt , and any reads containing bases with phred quality scores at or below 16 were discarded . this filtering retained 7.78 m total sequences ( 120k unique ) . since our approach includes data denoising , we could use much more liberal quality score filtering and retain more data ( usearch maxee of 1 and truncating at phred quality score 2 ) . to compare runtime of our algorithm and dbc , we increased the read truncation length so as to keep the same total number of sequences . this set the quality - filtered sequence length to @xmath38 nt , 20% longer than used by the authors ( 7.98 m sequences , 300k unique ; ` tcsh run_cff_on_fastq.tcsh 91 analysisresults " preheimdata/*.fastq " ` . ) . [ fig : dbc]a shows the substitution error rates inferred from the data at both sets of quality filtering parameters . note that these rates are significantly lower than those of fig . [ fig : qcparams ] ( the scales of the two plots are identical ) , exhibit a very weak transition / transversion bias , and are more sensitive to phred score quality filtering than what we have seen with caporaso et al . data ( fig . [ fig : qcparams ] ) . this seems to indicate that the protocol used by preheim et al . generates significantly fewer pcr substitution errors . this dependence of error rates on the experimental protocol highlights the advantage of being able to estimate error rates for a given dataset directly from the data , without the need for a separate calibration . [ fig : dbc]bc provide examples of sequences differing by a single nucleotide exhibiting ecologically significant distinctions , as identified by our method in this environmental dataset ; compare with fig . 2a , fig . [ fig : birthdeath ] and fig . [ fig : feces]ab . sequence abundance is shown as a function of depth ( each sample was independently normalized to @xmath39 total quality - filtered reads per sample , to correct for varying sample size ) . the dbc method of preheim et al . is also capable of identifying otus differing by a single nucleotide ( compare fig . [ fig : dbc]bc to fig . 5b in the original reference ) ; however , our analysis achieved higher resolution by retaining longer reads and took only 13 min single - core processor time ( intel xeon cpu 2.83ghz ) , compared to 8 hours analysis time the authors report for parallelized dbc running on a cluster with 60 - 100 processes executing simultaneously . in fact , the true runtime difference is even greater , since the complexity of both algorithms scales with the number of _ unique _ sequences rather than total reads . for @xmath40 as used by the authors of dbc , our algorithm completes in only 210 seconds . we stress , however , that dbc and cluster - free filtering seek to achieve different goals and are not directly comparable . dbc is an otu clustering algorithm , whereas the goal of cluster - free filtering is to identify sub - otu structure of moderate - to - high - abundance community members . however , to our knowledge dbc is the only existing tool that exploits cross - sample comparisons to inform the interpretation of sequencing data , and the performance comparison above serves to illustrate the drastically different computational cost of the two approaches . we have described a method that employs cross - sample comparisons to achieve sub - otu resolution . the analysis presented in the main text uses data from a study with an uncommonly large number of samples ; in contrast , the previous section demonstrates that our method can be usefully applied to a dataset with only 22 datapoints . what is the minimum number of samples required by our method ? the answer is that the number of samples determines the resolution that can be achieved ; more samples will always allow higher resolution , but coarser differences can be resolved with just a few . for example , just 2 samples ( say , 0 m and 10 m ) would have been enough to resolve the two subpopulations presented on fig . [ fig : dbc]c . by contrast , the difference between depth traces of seq . # 10 and seq . # 49 ( fig . [ fig : dbc]b ) is less pronounced and more samples are required . finally , resolving the sequences in fig . 2b would not have been possible with fewer than @xmath41100 samples . for high - abundance sequences where the complex struture of noise in the counts can be neglected , this tradeoff can be formally quantified using the jensen - shannon divergence as a measure of distance between abundance distributions of two sequences across samples ; for details , see preheim et al . , 2013 . the maximum degree to which time traces of two sequences can be correlated is a function of their abundance : for low - abundance sequences the poisson sampling noise becomes non - negligible and sets an upper bound for the best achievable correlation coefficient . consequently , to define a correlation as strong or weak , any measured correlation coefficient should be compared to this abundance - dependent quantity rather than to 1 . let @xmath42 be the true abundance time trace of some bacterial strain ( in units of cells , rather than sequence counts ) . imagine that two sequences in the dataset were measuring the abundance of this exact same strain , but with different amplification efficiencies @xmath43 and @xmath44 ( let @xmath45 ) . neglecting all sources of noise other than the poisson counting noise , the abundance traces of these two sequences can be modeled by @xmath46,\ ] ] where @xmath47 $ ] denotes adding poisson noise . since poisson noise is unavoidable , the correlation coefficient between these two traces sets an upper bound for the correlation between @xmath48 and any other trace @xmath49 with the same mean abundance as @xmath50 . this maximum correlation depends on the shape of the trace @xmath42 and amplification efficiencies @xmath43 , @xmath44 , and can be expressed as follows : @xmath51=\operatorname{corr}\left(\operatorname{poiss}[\lambda_1 n(t)],\operatorname{poiss}[\lambda_2/\lambda_1 \ast \lambda_1 n(t)]\right).\ ] ] and therefore , in terms of measurable quantities only : @xmath52 \approx \operatorname{corr}\left(\operatorname{poiss}[n_1(t)],\operatorname{poiss}[\langle n_2\rangle/\langle n_1\rangle \ast n_1(t)]\right).\ ] ] here @xmath53 denotes the average abundance , and we use the higher - abundance trace of the pair as the best estimate of the shape of the true abundance @xmath42 . the maximum correlation coefficient depends on the shape of the trace @xmath48 and on the mean abundance of the trace we compare it to ; the lower the mean abundance is , the stronger the effect of poisson noise and the lower the @xmath54 . in practice , for a pair of traces @xmath48 , @xmath50 , we compute their best expected correlation as follows : 1 . take the more abundant trace @xmath48 2 . construct a renormalized trace @xmath55 3 . poisson - resample both of these 10 times : denote these @xmath56 , @xmath57 , @xmath58 . 4 . compute the 100 correlation coefficients between all pairs @xmath59 . set @xmath60=\langle c_{ij}\rangle$ ] . the shape of the function @xmath60 $ ] is illustrated on fig . [ fig : maxcorr ] . we use the blast definition , i.e. the ratio of the number of mismatches to the total number of columns after pairwise realignment , and multiply this ratio by the length of the sequence ( 130 nt ) . for close sequences that differ by a few substitution errors the alignment is trivial , and this normalization corresponds to the hamming distance between sequences , in nt . we define the autocorrelation time @xmath61 of a sequence as the time shift @xmath8 at which the autocorrelation function @xmath7 falls below the threshold of statistical significance . for reasons discussed above , the notions of strong ( significant ) or weak ( insignificant ) correlation of sequence time traces are abundance - dependent . therefore , instead of using a fixed threshold value for all sequences , we proceed in the following way . for a given sequence , we first compute its root - mean - square autocorrelation coefficient for time shifts between 70 and 100 samples : @xmath62 if we assume that all autocorrelation observed at such large time shifts is entirely due to noise , then @xmath63 provides a natural scale for statistical significance . we conservatively define the significance threshold at twice the magnitude of @xmath63 . note that @xmath63 provides an upper bound on a statistically significant correlation value . if some dynamical processes in the population are slow enough that they contribute to the autocorrelation function even at such large time shifts ( _ cf . [ fig : birthdeath ] ) , this will increase @xmath63 and cause us to underestimate the true autocorrelation time . this means that assuming @xmath63 was entirely due to noise is a safe approximation to make : if it does not hold , it can only strengthen our conclusion that the sequence abundance time traces exhibit multi - day autocorrelations . [ fig : birthdeath]ab shows examples of sequences exhibiting steady change in abundance for more than a month . in both cases , the slow - changing sequence is 99.2% similar to a very high - abundance community member and could not have been resolved by traditional otu - based methods . note the sharp jump in panel b at day 182 of the sequence representing the invading subpopulation ( red ) to an abundance value close to the equilibrium established after day 210 . it is intriguing to speculate that this trace may document spatial invasion of a subpopulation already established elsewhere on the tongue , a region accidentally sampled on day 182 . to distinguish between 16s tags coming from distinct subpopulations or from physically the same bacterial cells , we introduced a quantity we called the _ persistence of difference _ @xmath17 . for this , we first defined the fractional difference @xmath14 between two time traces renormalized to the same mean @xmath16 : @xmath15 we then defined the persistence of difference @xmath17 as the 1-day autocorrelation coefficient of @xmath14 . if @xmath10 and @xmath11 are two genomic variants contained within the same bacterium , then any difference between @xmath12 and @xmath13 must be due to measurement noise , and @xmath17 must vanish . if , however , @xmath16 reflect abundances of two distinct subpopulations , then @xmath14 can be expected to exhibit some degree of autocorrelation due to the slow dynamics observed for most individual sequences . we gave an intuitive argument for this in the main text . here , to gain some extra intuition about the null model for @xmath17 , we calculate it explicitly in the simplest case when the two traces @xmath64 are independent and can be approximated by a stationary , weakly fluctuating process : @xmath65 here @xmath66 have zero mean , unit variance and are uncorrelated . assuming @xmath67 , we can write : @xmath68 and therefore , making use of the independence assumption , @xmath69 here @xmath70 are the one - day autocorrelation coefficients of the fluctuations of the two individual sequences . the independence approximation made above is clearly not valid for the dynamics of most community members . for this reason , for the purposes of fig . 3c , the null - model prediction was constructed directly from the data , by reversing in all pairs the time order for one of the sequences prior to the calculation of @xmath17 . this removes any real correlations of the traces while preserving autocorrelation and other properties of the traces such as their fluctuation spectrum . nevertheless , the calculation above is useful as it explains why the null - model expectation for @xmath17 is non - zero when both sequences have slow internal dynamics . note that a sequence with an exceptionally long intrinsic time scale ( as shown in fig . [ fig : birthdeath]ab ) will have a large @xmath17 score when paired with any other sequence . these two sequences were therefore excluded from fig . none of the cross - sample comparison methodology described in this work is limited to time series data . the `` persistence of difference '' argument accompanying fig . 4 is no exception ; however , it does rely on two additional assumptions , namely that the composition of samples varies smoothly with some parameter labeling the samples , and that the sampling frequency is sufficiently high to allow correlations of fluctuations to be observed between consecutive samples . for the longitudinal data series of caporaso et al . this parameter was time ; for a location series one can expect community composition to vary smoothly in space , and the same argument can be applied . in other words , the use of `` persistence of difference '' @xmath17 need not be limited strictly to longitudinal datasets . however , autocorrelation - based analysis is particularly sensitive to the number of samples ( see section [ sec : numberofsamples ] ) . determining whether @xmath17 can be a useful concept for studying the spatial heterogeneity of populations requires further investigation . as described in the main text , for the purposes of fig . 4 , when calculating otu quality scores , we restricted our attention only to high - abundance members of the otu , considering only sequences from the top 200 by overall abundance . since most of the diversity is contributed by low - abundance species ( huttenhower et al . , 2012 ) , 4 underestimates the true diversity of an otu . including lower - abundance otu members makes otu quality scores drop continuously as new otu members are added ; however , it also becomes increasingly hard to separate dynamical diversity from the effects of noise . consequently , in fig . 4 we report our most conservative estimate of within - otu diversity , where we use only the highest - abundance members out of all those resolved by cluster - free filtering ( there was an average of @xmath71 resolved sequences within a 97% otu , and only @xmath72 per otu were used for fig . 4 ) . in addition , otu quality scores were calculated under the assumption that each sequence represents a separate subpopulation . sequences that in fact derive from the same bacteria ( 16s paralogs or errors not in our model ) appear in the defining equation as independent , dynamically identical subpopulations , increasing the apparent otu quality score . this is another reason why the true quality scores of otus are likely even lower than reported in fig . to confirm our conclusions from the analysis of tongue microbiome data presented in the main text , we repeated our analysis using fecal samples of the two individuals , collected in the same study ( caporaso et al . , 2011 ) . there were 374 samples , 243 from the male subject and 131 from the female , with @xmath73 reads per sample . we normalized the observed abundances to @xmath74 total reads in each sample to correct for varying sample size . as before , we labeled sequences in order of decreasing overall abundance ( pooling samples from both individuals ) : seq # 1f , seq # 2f , etc . , where `` f '' reflects that we are now dealing with fecal samples rather than the tongue . again , we find that sequences differing by as little as a single nucleotide can exhibit ecologically significant differences in their dynamics . the most striking example is that the dominating sequence in individual `` male 3 '' differs from the dominating sequence in `` female 4 '' by a single nucleotide , and virtually no cross - contamination is observed ( fig/ [ fig : otu]a ) . both these sequences map to _ bacteroides sp . _ in greengenes ( desantis et al . , 2006 ) . another example is presented in panel b. finally , we repeat the cross - individual analysis presented , for tongue samples , in fig . 5ab . we find that the two gut communities as probed by the fecal samples also share a large fraction of sequences at 100% identity . this once again supports the scenario whereby the communities exchange members with non - negligible frequency , although less so than the tongue samples . the observation of panel a is therefore unlikely to represent the effect of dispersal limitation , suggesting instead a functional difference between the representatives of _ bacteroides sp . _ established in the two individuals or a resistance to invasion . finally , we find that the dynamical similarity of shared sequences , when measured independently in the two individuals , is clearly correlated , just as it was for the tongue communities ( fig . 5b ) . with the number of shared sequences being lower for fecal samples than for tongue samples , the statistics were insufficient to compare dynamical similarity of `` intentionally mismatched '' sequences as in fig . 5c . the same analysis as in fig . 5b can be performed for shared 97% otus rather than shared sequences ( at 100% identity ) . we constructed otus using closed - reference otu picking as implemented in qiime , matching sequences at 97% sequence similarity against the greengenes database . [ fig : otu ] shows the scatter plot of the dynamical similarity between pairs of common otus , as measured independently in the two individuals , for 78 common otus ( those shared within the top 100 ) . note , however , that most otus are dominated by a single high - abundance sequence ( as evidenced by the high weighted quality score on fig . 4 ) , and most of these dominating sequences are shared across the two communities ( fig . for these reasons , the plot shown here is very similar to fig . 5b , but only because the within - otu diversity is masked by dominating subpopulations . 0.2 in -0.2 in 1em desantis tz , hugenholtz p , larsen n , rojas m , brodie el , keller k et al ( 2006 ) . greengenes , a chimera - checked 16s rrna gene database and workbench compatible with arb . _ appl environ microbiol _ * 72 * : 50695072 .

a practical figure of merit to compare fuel and battery materials is the specific energy , or energy to mass ratio @xcite . as a material constant , it measures the energy that will be harvested by using a unit mass of the material . about a decade ago , a highly non - traditional fuel , called `` phaseonium '' , which is a three level atom with two lower states in a quantum coherent superposition , was proposed to be used in a photonic carnot engine ( pce ) @xcite . phaseonium engine could work with a single heat bath and a phaseonium reservoir @xcite . this proposal stimulated much interest to quantum heat engines @xcite . it was later argued that existing resonator systems can not implement such an engine , due to high cavity losses and atomic dephasing @xcite . here , we address a fundamental question of how the specific energy of phaseonium fuel is scaled with the number of quantum coherent levels . a favorable scaling law against decoherence and dephasing could bring the phaseonium engine closer to available practical systems . we describe multilevel generalization of phaseonium fuel in fig . the block diagonal density matrix @xmath2 of an @xmath0 level atom is shown in fig . [ fig1](a ) . the excited level , denoted by ``@xmath3 '' , and the lower levels , denoted by `` @xmath4 '' are well separated from each other by an energy @xmath5 measured from the central lower level @xmath6 as shown in fig . [ fig1](b ) . the lower levels can be degenerate or non - degenerate . the diagonal elements @xmath7 , and @xmath8 , with @xmath9 , determine the level populations , while the off diagonal elements @xmath10 , with @xmath11 , indicate the coherence between the levels . coherence can be characterized by the magnitude and phase of the complex number @xmath10 . though both the amplitude and the phase of coherent superposition states can be controlled in experiments @xcite , the main control variable for the photonic carnot engine is the phase of the coherence as the amplitude is required to be small enough to keep the system only slightly out of thermal equilibrium . the complete graphs in fig . [ fig1](a ) have @xmath1 nodes and @xmath12 links , representing the atomic energy levels and the coherences between them , respectively . the simplest graph has @xmath13 nodes , which is the case of the original phaseonium proposal @xcite . the interplay between quantum coherence and energy discussed in photon carnot engine @xcite revealed that the energy content of the phaseonium with @xmath13 can be optimized at a certain phase of the coherence . we could envision as if we are considering more complex , larger , phaseonium molecules with the graphs having @xmath14 , corresponding to @xmath0 level atom phaseonium ( nlap ) . ( color online ) @xmath0 level atom phaseonium ( nlap ) fuel . a ) density matrix @xmath2 and complete graph representations of nlap . @xmath2 is @xmath0 dimensional square matrix . its coherent block can be represented by a complete graph with @xmath1 nodes and @xmath15 links . graphs were shown up to @xmath16 number of nodes . b ) nlap for non - degenerate and degenerate atoms . the excited state is denoted by @xmath3 and the lower levels are denoted by @xmath17 with @xmath18 . the upper level is well separated from the lower levels by an energy @xmath19 measured from the central lower level @xmath6.,width=326 ] we can imagine different phaseonium molecules can be possible for a given atom of unit mass and explore how the specific energy of the atom depends on the size of the phaseonium molecule characterized by @xmath1 . next to the phase of coherence , @xmath1 becomes another control parameter which could favourably contribute to the enhancement of the specific energy of the single atom quantum fuel . for @xmath20 1 , the number of coherences would scale quadratically , @xmath21 . if the quadratic coherence scaling could be translated into the energy content of the atomic fuel , we could overcome the cavity losses for implementation and boost the performance of quantum carnot engine for applications . from fundamental point of view , such scaling analysis could reveal profound difference of quantum fuel from a classical resource as such a scaling can not exist without quantum coherence . complete graphs of phaseonium molecules serve more than a simple counting of coherences . they emphasize the generality of our question we address in the present contribution . can we beat decoherence with the scaling advantage of quantum coherent resources ? ( color online ) photonic carnot engine with @xmath0 level atom phaseonium ( nlap ) fuel . photon gas in a high quality cavity of frequency @xmath5 is the working substance and the mirrors of the cavity play the role of the piston . nlap leaves the hohlraum at temperature @xmath22 and is subsequently prepared in a state with quantum coherence among its lower levels characterized by @xmath12 phase parameters @xmath23 with @xmath24 . created nlaps are repeatedly injected into the cavity at a rate @xmath25 in the quantum isothermal expansion process , where heat @xmath26 is transferred to the cavity . the cycle continues with quantum adiabatic expansion and quantum isothermal compression and is completed with a quantum adiabatic compression . an amount of heat @xmath27 is rejected into the entropy sink in the isothermal compression.,width=307 ] we have recently numerically examined a superradiant quantum otto engine @xcite which exhibits a similar scaling law of work with the number of atoms ; though the efficiency is independent of the temperature and the number of atoms . our present results are fully analytical , and for a carnot cycle which can harvest work from a single heat bath and the scaling laws with the number of coherences are both for the work and the efficiency . furthermore , specific energy of a superradiant cluster increase linearly with the number of atoms , while here for the single atom phaseonium fuel it is quadratic . quantum advantage in the charging power of quantum batteries with the number of qubits has been examined very recently @xcite . linear scaling of work with the number of qubits is reported ; while due to a quantum speed up of the operation time , the charging power scales quadratically @xcite . the preparation of phaseonium fuel and using it in pce can be compared to charging and discharging a single qubit quantum battery with multiple quantum coherences . phaseonium fuel or battery allows for quadratic scaling in harvested work , efficiency , and the specific energy with the number of quantum coherences . we examined the charging or preparation cost of the phaseonium battery and compared it with the harvested work by pce . our result verified that the second law is obeyed in our system . the operation of photonic carnot engine is described in fig . the working fluid of the engine is the photon gas in a high quality single mode cavity of frequency @xmath5 . the radiation pressure by the cavity photons applies on the cavity mirrors playing the role of the piston of the engine . the quantum fuel of the engine is an nlap . the quantum carnot cycle consists of two quantum isothermal and two quantum adiabatic processes . in the isothermal expansion , nlaps are generated and injected into the cavity at a rate @xmath25 . the interaction time @xmath28 between an nlap and the cavity field is short , @xmath29 , so that only one nlap can be present in the cavity @xcite . coherences in nlap are characterized by @xmath12 phase parameters @xmath23 , with @xmath30 . coherent superposition states in @xmath0 level atom system can be generated by stimulated raman adiabatic passage @xcite , morris - shore transformation @xcite , or quantum householder reflection techniques @xcite . thermalization of the single atom can be considered relatively fast and hence the injection rate would be limited by the time of coherence preparation . the choice of specific technique of coherence induction depends on the details of a particular implementation . if the amplitudes of the coherences are much smaller than the level populations , then nlap can be assumed in an approximate thermal equilibrium with a thermal reservoir ( hohlraum ) at a temperature @xmath22 . during the interaction , the mean number of photons , @xmath31 , and the cavity temperature increases ; while the expansion cools down the cavity when there is no atom inside . repeated injection of nlaps into the cavity maintains the cavity field at a temperature @xmath32 by transferring a total amount of heat into the cavity as @xmath26 . @xmath32 is an effective temperature defined in terms of the steady state photon number @xmath33 as @xmath34 , with @xmath35 is the boltzman constant . it can be higher than @xmath22 in the presence of coherence @xcite . the cavity volume , and hence the frequency , change negligibly , @xmath36 . the cycle continues with an adiabatic expansion where the entropy remains constant and the temperature drops as the @xmath5 changes appreciably . following step is the isothermal compression in which heat @xmath27 is transferred from cavity to a cold reservoir at a temperature @xmath37 . the cycle is completed by adiabatic compression where the temperature is raised back to @xmath32 . the net work extracted from the cycle is @xmath38 where @xmath39 and @xmath40 . the mean photon number @xmath41 and the temperature @xmath42 at the beginning of the @xmath43 stage determine the entropy @xmath44 by @xmath45 . using @xmath46 , @xmath47 , @xmath48 , and @xmath49 , we write @xmath50 . the efficiency of the engine is defined as @xmath51 . it reduces to @xmath52 . this coincides with the standard definition of thermodynamic efficiency in carnot cycle and used in the original proposal of the phaseonium fuel @xcite as well as in the arguments against its feasibility in the presence of decoherence channels @xcite . in order to present results comparable to the previous works , we calculate the efficiency as defined by these studies . to avoid any misleading impressions however , we emphasize that in practical considerations round trip efficiency can be more relevant figure of merit . the round trip efficiency of the engine should include the cost of the preparation of the quantum coherent atom ; which would ensure the validity of the second law @xcite . on the other hand , it was noted that the cost of quantum fuel can be expensive @xcite , but it is still appealing as it can be used to harvest work from a single heat reservoir . our objective here is not to discuss if such pces can be efficient enough for certain applications , but to examine if such devices , proposed in @xcite , can produce positive work in the presence of decoherence by exploiting a scaling advantage of multiple coherence resources , in contrast to the negative conclusions of earlier studies @xcite . during the adiabatic process @xmath53 does not change so that @xmath54 and @xmath55 . these relations reveal that work and efficiency of the photonic carnot engine can be calculated by determining the @xmath33 at the end of the isothermal expansion stage . this manuscript organised as follws : in sec . ii , we describe the n - level phaseonium model , and pce system dynamics . we review and discuss the analytical and numerical verification of the analytical results in sec . we also estimate the preparation cost of nlap in this section . we conclude the results in sec . in order to find the @xmath33 , we solve @xmath56 where @xmath57 . here @xmath58 is the fock number state for the cavity photons and @xmath2 is the reduced density matrix of the cavity field . the equation of motion for @xmath2 can be obtained by tracing the equation of motion of the complete system over atomic degrees of freedom @xmath59_{nn})\label{eq.ofmot},\ ] ] where @xmath60 is the hamiltonian of the arbitrary @xmath61 atom in the interaction picture relative to the cavity photons , with @xmath62 and @xmath63 . here @xmath64 , @xmath65 are the energies of atomic states @xmath66 and @xmath67 , with @xmath18 , @xmath68 is the coupling rate between the atom and the field , and @xmath69 is the photon annihilation operator . the model hamiltonian describes a situation where @xmath0 level atom is coupled to a single mode cavity in a fan shaped transition scheme . a more realistic model requires consideration of multimode cavity coupled to an atom with multiple upper and lower hyperfine levels @xcite . such models can be reduced to effective single mode cavity and multilevel atom interactions @xcite or can be directly described by generalized master equations of micromasers @xcite . atoms with fan shaped degenerate level schemes are also studied from the perspective of generating large superposition states @xcite . the central question for us here is the dependence of work and efficiency on the number of the superposed quantum states and we will only consider single upper level and a set of degenerate or non - degenerate lower levels for simplicity . analytically calculating the right hand side of the eq . ( [ eq.ofmot ] ) , we find ( see appendix for details ) @xmath70\nonumber\\ & + ( \sum_{i=1}^n k_{b_i}\rho_{b_ib_i}+\sum_{i < j } k_{ij}^{\phi_{ij}}|\rho_{b_ib_j}|)\nonumber\\ & \times [ n\rho_{nn}-(n+1)\rho_{n+1,n+1}]\},\end{aligned}\ ] ] where the coefficients @xmath71 and @xmath72 depend on the atomic relaxation rate @xmath73 , atomic dephasing rate @xmath74 , detuning parameter @xmath75 , with @xmath76 , and @xmath77 , as well as the coherence parameters @xmath23 and @xmath78 , by the relations given in the appendix . thus , we obtain the rate of change of average photon number @xmath79 where @xmath80 and @xmath81 . the equation of motion for the evolution of population elements in the density matrix coincides with the thermalization dynamics of a resonator coupled to a heat bath . accordingly , the coarse grained dynamics effectively describes sequence of nlap injected into the resonator as a mesoscopic ensemble of @xmath0 level atoms acting as a heat bath . the off diagonal elements of the density matrix or the coherences can be kept vanishingly small to describe the steady state approximately as a thermal equilibrium state . the corresponding effective temperature can be determined by the modified detailed balance condition to reach such a quasi equilibrium state in eq . ( [ eq.mot2 ] ) which gives @xmath82 the detailed balance between the thermal reservoir at @xmath22 and the photon gas in the resonator is broken but there is a modified detailed balance between the coherent atomic ensemble and the resonator photons . accordingly the resonator can reach a thermal equilibrium at a different temperature @xmath32 than @xmath22 . the steady state of the eq . ( [ rateeq . ] ) yields the average photon number as @xmath83 where @xmath84 is the average photon number in the absence of coherence . using @xmath85 , we determine the effective cavity temperature as @xmath86 by using high temperature approximations @xmath87 and @xmath88 in eq . ( [ nbarf ] ) . therefore , the efficiency of the photonic carnot engine in the case of nlap becomes @xmath89 where @xmath90 is the carnot efficiency . note that for @xmath91 , @xmath92 but @xmath93 could have a positive value for particular values of control parameters @xmath94 . in order to get further analytical results we will make some simplifying assumptions . we focus on degenerate nlap case to proceed analytically , for which @xmath95 , @xmath96 , @xmath97 , @xmath98 , @xmath99 and @xmath100 . in addition , we consider phase locked equal amplitude coherences with @xmath101 and @xmath102 . hence the coefficients in eq . ( [ eq.mot2 ] ) become @xmath103 , @xmath104 and @xmath105 , and hence eq . ( [ rateeq . ] ) reduces to @xmath106-\kappa\bar{n}_{\phi } , \label{degrateeq.}\end{aligned}\ ] ] for @xmath107 , @xmath108 , where @xmath109 , @xmath110 , @xmath111 with @xmath112 . here we introduced @xmath113 and @xmath114 , with @xmath115 , as the decoherence rate due to the dissipation in the cavity and a phenomenological decoherence factor due to atomic dephasing , respectively @xcite . while the dephasing factor is phenomenologically introduced in @xcite we provide its rigorous microscopic derivation in the appendix . steady state solution of eq . ( [ degrateeq . ] ) yields an effective temperature given by @xmath116 in the high temperature limit where @xmath117 with @xmath118 . for small coherence and decoherence terms in @xmath119 , an approximate expression can be written for the effective temperature @xmath120 this result shows that if the reduction of the magnitude of coherence due to dephasing is slower than the quadratic increase with @xmath1 , then the multilevel coherence could be used to beat the decoherence induced by the cavity dissipation . the magnitude of coherence @xmath121 is limited by the positivity requirement of the density matrix as well as the thermal equilibrium requirement of the cavity field . the former condition require @xmath122 so that @xmath123 for @xmath107 as @xmath124 for @xmath107 . accordingly one can fix the coherence amplitude @xmath121 as a constant as long as it remains smaller than @xmath125 for the range of @xmath1 values . for a realistic number of levels this is not a very restrictive condition . more severe limitation on @xmath121 is due to the quasi - equilibrium condition of the photon gas . we will take @xmath126 and consider @xmath127 in our numerical examinations . in the classical asymptotical limit of @xmath128 then @xmath129 as @xmath125 and hence the quadratic scaling reduces to a linear one for which the specific energy becomes a constant as expected for classical systems . mesoscopic systems in quantum regime are therefore necessary to exploit the quadratic scaling in the specific energy . we note that the coarse grained dynamics is designed on purpose to determine the steady state by rapid convergence using a numerically efficient dynamical equations . analytical solution of the mean photon number dynamics for the degenerate case and with @xmath130 , @xmath131 , @xmath132 gives @xmath133 where @xmath134 is the thermalization time . we will first discuss the predicted analytical state states by the coarse grained master equation in modern resonator settings then examine the exact numerical desription of the dynamics of the system to verify the analytical results in the subsequent subsections . ( color online ) extracted work ( @xmath135 ) and efficiency ( @xmath136 ) of photonic carnot engine , with @xmath0 level atom phaseonium ( nlap ) fuel , depending on the number of degenerate coherent ground state levels @xmath1 , for different decoherence factor models ( a)-(b ) @xmath137 , ( c)-(d ) @xmath138 . ( e)-(f ) @xmath139 , where @xmath140 . coherence parameter is @xmath141 and the initial thermal coherent atomic temperature is @xmath142 in units of @xmath143 . @xmath144 values are 0.15 , 0.1 , 0.05 , 0.001 for ( a)-(b ) , 0.14 , 0.12 , 0.1 , 0.08 for ( c)-(d ) and 0.012 , 0.01 , 0.008 , 0.006 for ( e)-(f ) in decreasing order for the lower to upper curves respectively . the plots are given for the circuit qed parameters in @xcite . the quantities @xmath145 , @xmath146 , @xmath147 , and @xmath148 , which are the coupling coefficient to the cavity field , atomic injection rate , cavity loss term , and atomic decay respectively , are dimensionless and scaled with the resonance frequency @xmath149 ghz . @xmath136 is dimensionless , @xmath135 is dimensionless and scaled with @xmath5.,width=326 ] in the high temperature limit ( @xmath150 , the entropy change in the isothermal expansion stage is @xmath151 and the heat input becomes @xmath152 . the work output at @xmath153 is found to be @xmath154 where @xmath155 , respectively . in superconducting circuit , microwave and optical resonators , it is estimated that @xmath156 @xcite . @xmath157 should be replaced by @xmath12 for smaller number of levels . accordingly , by using five or more level quantum phaseonium fuel , the working fluid can beat quantum decoherence to harvest positive work . in fig . [ fig : fig3 ] , we plot the work output and efficiency of the photonic carnot engine with degenerate nlap fuel , depending on the number of quantum coherent levels . we consider @xmath1 independent as well as @xmath1 dependent scaling models @xcite for the decoherence factor and take @xmath137 in fig . [ fig : fig3](a)-(b ) , @xmath138 in fig . [ fig : fig3](c)-(d ) , and @xmath139 in fig . [ fig : fig3](e)-(f ) , where @xmath140 as shown in the appendix . the plots are given for the circuit qed parameters in @xcite . we consider larger atomic dephasing rates than the typical values to demonstrate its limiting effect on @xmath135 and @xmath136 . the plots indicate that even when there is large dephasing , which can increase with @xmath1 linearly or quadratically , @xmath135 and @xmath136 can retain their quadratic power law with @xmath1 up to a critical @xmath1 . ( color online ) extracted work ( @xmath135 ) and efficiency ( @xmath136 ) of microwave resonator photonic carnot engine , with @xmath0 level atom phaseonium ( nlap ) fuel , depending on the number of degenerate coherent ground state levels @xmath1 for different decoherence factors ( a)-(b ) @xmath158 , ( c)-(d ) @xmath159 , ( e)-(f ) @xmath139 , where @xmath140 . coherence parameter is @xmath141 and the initial thermal coherent atomic temperature is @xmath142 in units of @xmath143 . @xmath144 values are 0.1 , 0.15 , 0.05 , 0.001 for ( a)-(b ) , 0.12 , 0.1 , 0.08 , 0.06 for ( c ) , 0.14 , 0.12 , 0.1 , 0.08 for ( d ) , 0.12 , 0.1 , 0.08 , 0.06 for ( e ) and 0.01 , 0.008 , 0.006 , 0.004 in decreasing order for the lower to upper curves respectively . the parameters @xmath160 , @xmath161 , @xmath162 and @xmath163 are the atom - field coupling coefficient , atomic injection rate , cavity loss term and atomic decay respectively . they are dimensionless and scaled with the typical resonance frequency is @xmath164 ghz @xcite . @xmath136 is dimensionless , @xmath135 is dimensionless and scaled with @xmath5.,width=326 ] similar results are found for the cases of optical and microwave cavities . we see both in fig . [ fig : fig4](a)-(b ) and fig . [ fig : fig5](a)-(b ) that when dephasing is independent of @xmath1 , the work output and efficiency increases quadratically with the number of coherent levels . if the dephasing rate is increasing linearly with @xmath1 as in fig . [ fig : fig4](c)-(d ) and fig . [ fig : fig5](c)-(d ) , or if it is increasing quadratically with @xmath1 as in fig . [ fig : fig4](e)-(f ) and fig . [ fig : fig5](e)-(f ) , the work output and efficiency of the photonic engine is enhanced quadratically with the number of coherent levels only up to critical @xmath1 . beyond which the work output and efficiency decays exponentially due to the dominating effect of decoherence . ( color online ) extracted work ( @xmath135 ) and efficiency ( @xmath136 ) of optical resonator photonic carnot engine , with @xmath0 level atom phaseonium ( nlap ) fuel , depending on the number of degenerate coherent ground state levels @xmath1 for different decoherence factors ( a)-(b ) @xmath158 , ( c)-(d ) @xmath159 , ( e)-(f ) @xmath165 , where @xmath140 . coherence parameter is @xmath141 and the initial thermal coherent atomic temperature is @xmath142 in units of @xmath143 . @xmath144 values are 0.15 , 0.1 , 0.05 , 0.001 for ( a)-(b ) , 0.14 , 0.12 , 0.1 , 0.08 for ( c)-(d ) and 0.01 , 0.008 , 0.006 , 0.004 for ( e)-(f ) in decreasing order for the lower to upper curves respectively . the parameters @xmath166 , @xmath167 , @xmath168 and @xmath169 are the coupling frequency to the cavity field , atomic injection rate , cavity loss term and atomic decay respectively . they are dimensionless and scaled with the typical resonance frequency is @xmath170 thz @xcite . @xmath136 is dimensionless , @xmath135 is dimensionless and scaled with @xmath5.,width=345 ] in order to perform a faithful numerical simulation of a typical set up described in the theory , we investigate the injection process in detail . we assume a regular atomic injection of rydberg atoms into a fabry - perot cavity @xcite with an atomic interaction time @xmath28 with the cavity field and an empty cavity time @xmath171 such that @xmath172 where @xmath25 is the injection rate . during the time interval @xmath28 , the hamiltonian is @xmath173 while for the time interval @xmath171 , @xmath174 ( @xmath175 and @xmath176 for degenerate ground state levels ) . we choose injection time @xmath177 where @xmath113 is the cavity decay rate and @xmath178 is the number of atoms kicking the cavity field in the photon lifetime . the time elapsed when cavity is empty @xmath171 can be related to the interaction time such that @xmath179 . thus , we can write @xmath180 . here , @xmath181 is a factor introduced to measure @xmath171 in terms of @xmath28 so that @xmath182 . we solve the master equation by numerical methods and compare the results with the developed theory . we use qutip package @xcite in python software to solve the master equation . we perform single atomic injection in two steps . first step is the atom - cavity field interaction ( during @xmath28 ) and the second one is the free cavity field evolution ( during @xmath171 ) . the master equation for the first step is written under markov and born - markov approximations as @xcite @xmath183+\gamma\sum_m^{n+1 } \mathcal{l}[l_m^{\gamma}]+\frac{\gamma_{\phi}}{2}\sum_n^n \mathcal{l}[l_n^{\phi}]\end{aligned}\ ] ] where last two terms stand for pure spontaneous emission and pure dephasing @xcite , respectively . here , @xmath184=(2x\rho x^{\dagger}-xx^{\dagger}\rho-\rho x^{\dagger}x)/2 $ ] is a liouvillian superoperator in lindblad form and @xmath185 , @xmath186 . we include an auxiliary state @xmath187 to the atomic state space to model the decay of the excited and the lower levels . presence of @xmath187 is not altering the initial phaseonium state . the auxiliary state is unpopulated and at the lower level energy . its use allows for faithful simulation of the excited state and the degenerate ground state @xmath188 ) decay equations in eq . ( [ eq15 ] ) . this decay model is already used in the original master equation developed for the two - level phaseonium engine @xcite . different decay models , for example decay of excited level to the lower levels are employed for other systems such as many atom superradiant otto engine @xcite and similar effect of beating decoherence with scaling up coherence is found . present contribution discusses the original photo - carnot engine @xcite as well as the objection to its feasibility due to dephasing , phenomenologically described by factor @xmath114 @xcite . our introduction of @xmath73 is an additional decoherence channel not included in ref . we have found that @xmath189 can be analytically expressed in terms of @xmath114 ( see appendix for details ) . the microscopical master equation approach describes both the original photo - carnot engine coarse - grained master equations with @xcite and without @xcite dephasing and generalizes them to the multilevel case ; as our mesoscopic master equation , eq . [ eq.mot2 ] , does analytically . in the second step , cavity decay @xmath190 is present during the time interval @xmath171 in accordance with the key assumptions of micromaser theory @xcite and the corresponding master equation is @xmath191+\kappa\mathcal{l}[\hat{a}].\end{aligned}\ ] ] ( color online ) time evolution of average photon number during thermalisation process of pce field while regular injection of 2lap depending on different @xmath178 parameters . @xmath178 values are 4500 , 1500 , 150 , 100 and 50 in decreasing order for the upper to lower curves respectively . coherence parameter is @xmath192 , initial field temperature @xmath193 and temperature of thermal coherent atoms is @xmath194 in units of @xmath143 . the resonant field frequency is @xmath164 ghz , cavity quality factor is @xmath195 , atom cavity field interaction time is @xmath196 @xmath197s , atom decay rate is @xmath198 hz , atom dephasing rate is @xmath199 hz and atom cavity field coupling is @xmath200 khz . time is dimensionless and scaled with @xmath5.,width=307 ] in fig . [ fig : fig6 ] , we present the thermalisation process of the cavity field depending on different @xmath178 values by depicting the photon number versus scaled time . physical parameters @xcite are given in the figure caption consistent with the rydberg atoms in a superconducting febry - perot cavity @xcite . for low values of @xmath178 which corresponds to large @xmath181 , we have zigzag like curves and for high values of @xmath178 , we have smoother lines . average photon number @xmath33 converges to the theoretical value for @xmath201 . thermalisation time is much longer than the convergence rate of coarse grained master equation and the microscopic exact method is much more costly numerically . ( color online ) comparison of time evolution of average photon number in presence of thermal and coherent atom injection with @xmath202 and @xmath203 . the horizontal dashed and dashed - dotted lines stands for the analytical @xmath33 and @xmath53 values in absence of loss mechanisms . solid and dotted lines stands for time evolution of @xmath33 and @xmath53 respectively in presence of dissipation channels . @xmath204 for both subplots corresponding to @xmath205 @xmath197s . insets magnifies the lines between @xmath206 and @xmath207 . all the remaining parameters are the same with that of fig . [ fig : fig6 ] . time is dimensionless and scaled with @xmath5.,width=307 ] in fig . [ fig : fig7](a)-(b ) , we express the effect of the number of degenerate atomic ground state levels @xmath1 of the coherent atoms against the dephasing and the cavity loss mechanisms . horizontal dotted and dotted - dashed lines are the analitical values of average photon numbers ( @xmath33 , @xmath53 ) of the no loss case for each @xmath1 . in fig . [ fig : fig7](a ) when decoherence channels are open , average photon number saturates below the analytical values for @xmath202 in accordance with the argument that 2lap phaseonium can not beat decoherence @xcite . average photon number exceeds @xmath53 for @xmath203 in fig . [ fig : fig7](b ) by keeping all the other parameters same . thus , we show that the decoherence can be beaten by using higher @xmath1 . to make the effect more visible in the figures , we take larger coherence magnitude , @xmath192 . ( color online ) comparison of numerical results ( dashed line ) with developed theory ( solid line ) of effective field temperature t@xmath208 , harvested work @xmath135 and efficiency @xmath136 respectively depending on the number of degenerate ground state levels @xmath1 . @xmath209 and corresponding @xmath210 @xmath197s . all the remaining parameters are the same that of fig . [ fig : fig6 ] . t@xmath208 , @xmath135 and @xmath136 is dimensionless and scaled with @xmath5.,width=336 ] we also compare the consistency of effective field temperature t@xmath208 , harvested work @xmath211 and efficiency @xmath212 versus @xmath1 in fig . [ fig : fig8 ] ( a ) , ( b ) , ( c ) between developed analytical and numerical results when dissipation channels open . we observe a good consistency between numerical and theoretical results in steady state . typical methods to generate quantum superposition states , such as pulse area , adiabatic passage , or stirap techniques @xcite , utilize optical pulses interacting with the atomic system to transfer an initial atomic state to a target one . the initial and target quantum states are known and hence one can easily determine the required unitary transformation between them . physical implementation of the required propagator is however a much more challenging problem than the calculation of the transformation matrix . an efficient strategy to sythesize the transformation matrix is to decompose it into a product of matrices , representing interacting steps that can be implemented by using optical pulses coupled to the atom . an arbitrary @xmath1 dimensional unitary matrix @xmath213 can be decomposed into so called @xmath1 generalized quantum householder reflection ( qhr ) matrices or @xmath214 standard qhrs and a phase gate @xcite . a generalized qhr is defined by @xmath215 where @xmath216 is the identity operator and the @xmath217 is the normalized column vector with dimension @xmath1 , same with the number of the pulses , and @xmath218 is an arbitrary phase factor . the decomposition of @xmath213 in terms of generalized qhrs can be written as @xmath219 for @xmath108 , eq . ( [ genreflct ] ) reduces to @xmath220 which is the standard qhr . the interaction represented by each householder matrix can be described by a propagator which can be determined by the morris - shore transformation @xcite . our n+1 level atom coupled to n optical pulses in a fan shaped transition scheme , or so called n - pod model , is a generic model that is used to discuss generation of arbitrary multilevel superposition states . under the morris - shore transformation , the lower levels of the atom are grouped into a single bright level coupled to an effective single pulse and n-1 dark levels uncoupled from the optical pulses . the propagator is then easily determined in this morris - shore basis . back transformation to the original basis gives the full propagator , or the generalized qhr matrix . both the number of qhr steps and the number of pulses used in each step are in the order of @xmath1 , therefore the total number of pulses to be used to generate the target state would be in the order of @xmath157 . this shows that the energetic cost of preparation of the target state scales with @xmath157 , same with the work and efficiency scaling in the corresponding photonic carnot engine . the preceding discussion is applicable to the case of mixed states as well , for which the normalized vectors of generalized qhrs are defined as @xcite @xmath221 here , @xmath222 is the @xmath43 column of @xmath213 , @xmath223^t$ ] and @xmath218 is an arbitrary phase where @xmath224 . it s shown that for an n - pod system any standard qhr @xmath225 can be realized by single pulses with an rms pulse area @xmath226 @xcite . the corresponding unitary transformation can only link the mixed states with the identical dynamical invariants . in our case we consider initial thermal states out of a hohlarum transformed to a final state with small coherences . the initial and final states would then possess different spectral decompositions so that they can not be unitarily connected . a resolution to this is suggested to exploit decoherence channels such as spontaneous emission or pure dephasing , in combination with the unitary transformation @xcite . we will not follow this route but use an alternative , which allows for a fully unitary procedure to generate desired coherences . as the coherences contribute additively , to exploit their scaling advantage we do not need an exact state but an approximate one would be sufficient . accordingly , we can simply consider an approximate approach and do not specify an exact value for the coherences . we only need to keep them small enough to ensure slightly out of thermal equilibrium final state . we illustrate our strategy for @xmath202 case , and suggest that in principle larger nlap can be generated by straightforward extension of this technique . unitarity of our procedure also makes the details of generation process immaterial for the cost estimation . the cost would be the same for other unitary equivalent processes to generate same states . initial state of the atom out of the hohlarum at @xmath22 is the thermal density matrix @xmath227 where @xmath228 inverse temperature @xmath229 and @xmath230 is the partition function , with @xmath231 being the atomic hamiltonian . taking @xmath202 and @xmath194 , we find @xmath232 . target density matrix @xmath233 is taken to be @xmath234 where the off - diagonal elements between degenerate ground state levels are taken real and much smaller than the diagonal elements that are equal to those of the @xmath235 . the initial and final density matrices have distinct dynamical invariants in their spectral decompositions and hence they can not be linked by a coherent evolution . let us assume however an approximate link such that @xmath236 . the unitary transformation @xmath237 can be determined from the matrix that diagonalize @xmath233 and found to be @xmath238 writing @xmath239 , the fidelity between @xmath233 and @xmath240 is determined by @xmath241 and found be @xmath242 . thus , @xmath243 can be approximately be used as the unitary operator linking the initial and the target density matrices . @xmath244 can be synthesized by using two standard qhrs and a phase gate as @xmath245 where@xmath246 and @xmath247 is a one - dimensional phase gate . by using eq . ( [ normalisedvect ] ) one finds the normalized column vectors to be @xmath248^t$ ] , @xmath249^t$ ] . the example for @xmath202 here illustrates the basic principles to generate arbitrary nlap . one can use @xmath1 pulses for each @xmath214 standard qhrs and a phase gate to synthesize a unitary transformation matrix , which is the diagonalization matrix of a quasi equilibrium thermal state at @xmath22 with small coherences . when the pulses applied to the actual atom out of hohlarum at @xmath22 , the final state will be approximately the same with the target state used to determine the properties of the pulses . we can now estimate the energy cost @xmath250 of generating @xmath233 using @xmath251 where @xmath157 is the total number of pulses used in the qhr technique and @xmath252 is the energy of a single pulse . @xmath252 can be determined from the pulse area @xmath253 as in @xcite . for a square pulse of duration @xmath254 and amplitude @xmath255 we have @xmath256 , where @xmath257 is the magnitude of the dipole moment @xmath258 where @xmath259 is the vacuum permittivity and @xmath260 is the speed of light . taking the pulse area @xmath226 we find @xmath261 the intensity of the pulse is given by @xmath262 . the pulse energy in a beam of radius @xmath263 can be estimated by @xmath264 . using @xmath265 and @xmath266 , where @xmath121 and @xmath267 are the wavelength of the optical field and the radial beam divergence , respectively , we find @xmath268 taking @xmath269 and @xmath270 @xcite , we find @xmath271 . total energy cost to reach steady state per cycle is @xmath272 where @xmath273 . here , @xmath274 is the number of atoms needed for thermalisation , @xmath25 is the injection rate , and @xmath275 is the time elapsed to reach the steady state . in our results we have found that the harvested work per cycle is much less than the resonance energy , @xmath276 , thus the generation cost of nlap fuel is several orders of magnitude larger than the harvested work @xmath277 , which confirms that the second law of thermodynamics obeyed in photonic carnot engine with nlap . generation cost is not included in the thermodynamic efficiency but it can be significant figure of merit in the round trip efficiency . to make such photonic carnot engines more appealing for certain applications , it is necessary to increase their round trip efficiency as well . for that aim one could consider the cases of larger compression ratios ( @xmath278 ) and high operation temperatures ( @xmath279 ) . our focus here is on the discussion if such engines can operate under decoherence . despite the negative conclusions for two level phaseonium case @xcite , we have found that larger phaseonium fuel allows for operational photonic carnot engines . the question of how to increase their round trip efficiency requires further analysis which is beyond the scope of present contribution . summarizing , we examined scaling of work and efficiency of a quantum heat engine with the number of quantum resources . specifically , we considered a photonic carnot engine with a multilevel phaseonium quantum fuel . we derived a generalized master equation for the cavity photons , which forms the working fluid of the engine , and determined the steady state photon number to calculate the work output and thermodynamic efficiency . we find that they scale quadratically with the number of quantum coherent levels @xmath1 . we examined the case of degenerate levels to get analytical results and to examine scaling laws against decoherence due to cavity dissipation and atomic dephasing . we verified our analytical results with detailed numerical methods and have shown consistency of coarse grained analytical results with the microscopical numerical approach . generation of multilevel phaseonium fuel using morris - shore transformation determined quantum householder reflection technique as well as its cost are examined . using typical parameters in modern resonator systems , such as circuit qed , our calculations reveal that decoherence due to cavity dissipation could be overcome by the multilevel quantum coherence even in the presence of large dephasing rate . if the dephasing rate increases with @xmath1 , then work and efficiency can still overcome the decoherence and retain their @xmath157 scaling up to a critical number of coherent levels . authors warmly thank n. allen , a. imamoglu and i. adagideli for illuminating discussions . authors acknowledge support from ko university and lockheed martin university research agreement . we generalize micromaser mesoscopic master equation treatment @xcite applied for three - level phaseonium engine @xcite to multi - level case . in @xmath1lap model , the hamiltonian of the whole system is @xmath280 where label @xmath35 implies injected @xmath61 arbitrary atom . we adopt the notation of ref . @xcite for the hamiltonian and relevant quantities . here , @xmath281 and in the interaction picture where @xmath283 , @xmath284 are atomic energy levels , @xmath68 is the atom - field coupling coefficient and @xmath5 is the single mode cavity frequency . here , we assume all levels coupled to the excited one with the same coefficient @xmath68 . the equation of motion of overall system is the equation of motion of the radiation field which is the working substance of the heat engine can be found by tracing out atomic part as @xmath288_{nn})\nonumber\\ + \mbox{tr\scriptsize at}\mathcal{l}_a[\rho]_{nn}+\mbox{tr\scriptsize at}\mathcal{l}_f[\rho]_{nn}\label{eq - mot}\end{aligned}\ ] ] where @xmath57 . here , @xmath289_{nn}={\langle n|}(h\rho-\rho h){|n\rangle}= \sum_{m}{\langle n|}h{|m\rangle}{\langle m|}\rho{|n\rangle}-\sum_{m}{\langle n|}\rho{|m\rangle}{\langle m|}h{|n\rangle}=\sum_{m}(h_{nm}\rho_{mn}-\rho_{nm}h_{mn})$ ] . in micromaser theory , due to the short atom - cavity interaction time , last term , cavity decay is usually ignored when the atom is inside the cavity . the second term is treated perturbatively and will be considered for zeroth order in g. here , we will keep it but assume it can be treated independently . one may write the partial trace operation over atomic degrees of freedom for a random single atom as , @xmath290_{nn}&=\sum_{\alpha}{\langle \alpha , n|}[h,\rho]{|\alpha , n\rangle}\nonumber\\ & = { \langle a n|}[h,\rho]{|a n\rangle}+\sum_{i=1}^{n}{\langle b_i n|}[h,\rho]{|b_i n\rangle}\label{ptrace}\end{aligned}\ ] ] where @xmath291 are the atomic basis as expressed at the right hand side . each term of eq . ( [ ptrace ] ) can be calculated by using the selective rules of the hamiltonian between certain levels @xmath292 and @xmath274 ; for instance inserting for the first term of eq . ( [ ptrace ] ) we have , @xmath293{|a n\rangle}&={\langle a n|}(h\rho-\rho h){|a n\rangle}\nonumber\\ & = \sum_{\alpha ' m}\{{\langle a n|}h{|\alpha ' m\rangle}{\langle \alpha ' m|}\rho{|a n\rangle}\nonumber\\ & -{\langle a n|}\rho{|\alpha ' m\rangle}{\langle \alpha ' m|}h{|a n\rangle}\}\nonumber\\ & = \sum_{\alpha ' m}\{h_{an,\alpha'm}\rho_{\alpha'm , an}-\rho_{an,\alpha'm}h_{\alpha'm , an}\}. \label{eq4}\end{aligned}\ ] ] @xmath294 is the hamiltonian eq . ( [ eq - int ] ) to be inserted into eq . ( [ eq4 ] ) which can be written as , @xmath295 where @xmath296 and @xmath297 . expressing the terms of eq . ( [ eq4 ] ) conveniently , we write @xmath298 and @xmath299 . besides , @xmath300 . substituting these terms into the first part of eq . ( [ eq4 ] ) we have @xmath301 . the second part of eq . ( [ eq4 ] ) is simply the complex conjugate . hence the first term of eq . ( [ ptrace ] ) is @xmath293{|a n\rangle}&=\hbar g e^{-i\omega t}\sqrt{n+1}\sum_{i=1}^{n}\rho_{b_i n+1,an}-c.c.\nonumber\\ \label{eq7}\end{aligned}\ ] ] the second term of eq . ( [ ptrace ] ) would be calculated by similar considerations . we can write @xmath302 { |b_i n\rangle}=\sum_{\alpha ' m}\{h_{b_i n,\alpha'm}\rho_{\alpha'm , b_i n}\}$ ] . in this case @xmath303 , @xmath304 and @xmath305 . then , @xmath302{|b_i n\rangle}=\hbar g e^{i\omega t}\sqrt{n}\rho_{an-1,b_in}-c.c.$ ] and finally the second term of eq . ( [ ptrace ] ) becomes , @xmath306{|b_in\rangle}=\hbar g e^{i\omega t}\sqrt{n}\sum_{i=1}^{n}\rho_{an-1,b_in}-c.c.\nonumber\\\label{eq8bi}\end{aligned}\ ] ] inserting these results into eq . ( [ eq - mot ] ) and after some rearrangements we have the field equation of motion , @xmath307_{nn}.\label{9eqmot}\end{aligned}\ ] ] we use @xmath308 hereafter for simplicity . here , @xmath309_{nn}&={\langle n|}\frac{\kappa}{2}(2\hat{a}\rho\hat{a}^{\dagger}-\rho\hat{a}^{\dagger}\hat{a}-\hat{a}^{\dagger}\hat{a}\rho){|n\rangle}\nonumber\\ & = \kappa\left\lbrace(n+1)\rho_{n+1,n+1}-n\rho_{nn}\right\rbrace.\label{louvil - rho}\end{aligned}\ ] ] in order to proceed the calculation , any single term in the summation of eq . ( [ 9eqmot ] ) should be calculated and inserted therein . the terms can be obtained by the integration of corresponding equation of motions by using the selecetive rules of the hamiltonian as expressed above . we evaluate atomic equation of motion @xmath310 order in @xmath68 first , we have where @xmath312 and @xmath313 . final two terms of eq . ( [ atomic-eq.mot ] ) corresponds to @xmath314 $ ] with @xmath315 and @xmath316 . here , @xmath317 and @xmath318 is taken equal to @xmath73 for simplicity . note that the atomic part of the master equation is for the case of pure dephasing @xmath319 relaxation and we follow the usual assumption of micromaser theory @xcite that cavity decay and atomic dynamics can be separately treated . when atom is inside the cavity , decay is not included . the equation of motion of the @xmath43 term of the first summation of eq . ( [ 9eqmot ] ) for a single atom is @xmath320 here , we have neglected the the matrix element @xmath321{|an\rangle}$ ] in accord with the assumptions indicated above . the equation of motion for @xmath322 which is the @xmath43 term of second summation of ( [ 9eqmot ] ) could be obtained by simply replacing @xmath323 in eq . ( [ eq10 ] ) , that is @xmath324 therefore we obtain @xmath325 and @xmath326 terms by integrating eq.s ( [ eq10 ] and [ eq.11 ] ) formally in the following form , the terms of eq.s ( [ eq12 ] and [ eq13 ] ) can be factorized to atomic and field density matrices for zeroth order solution in g. for instance , @xmath329 @xmath330 here , @xmath331 , @xmath332 which are the initial atomic density matrix elements , obey the respective atomic equations of motion @xmath333 in which the solutions are @xmath334 where @xmath335 . eq.s ( [ eq15 ] ) and ( [ eq16 ] ) imply that excited and ground state levels decay to a lower level . the off - diagonal elements of the atomic density matrix are equal to @xmath336 . the @xmath337 is assigned with the coherence preparation . thus , we can express eq.s ( [ eq12 ] and [ eq13 ] ) by using zeroth order atomic eq.s ( [ eq14]-[eq16 ] ) to find first order solutions in g , before proceeding , we replace the summation over number of injected atoms by integration over injection time as @xmath341 where @xmath25 is the injection rate . we also define @xmath342 where @xmath343 . then , @xmath344+c.c\big\}+\mathcal{l}_f[\rho]_{nn}.\label{eq20}\end{aligned}\ ] ] note that @xmath345 while @xmath346 . evaluating the integrals in ( [ eq20 ] ) over @xmath347 and @xmath348 after changing integration order as @xmath349 we have @xmath350-\big[\frac{1}{\gamma}\big(\sum_{i=1}^{n}\frac{1}{-i\delta_i+\gamma}\big)\rho_{b_i b_i}^0\nonumber\\ & + \sum_{i < j}\big(\frac{1}{(-i\delta_i+\gamma)}\frac{1}{(i\omega_{b_i b_j}+\bar{\gamma})}e^{i\phi_{ij}}+\frac{1}{(-i\delta_j+\gamma)}\frac{1}{(-i\omega_{b_i b_j}+\bar{\gamma})}e^{-i\phi_{ij}}\big)|\rho_{b_i , b_j}^{0}|\big]\nonumber\\ & \times\big[(n+1)\rho_{n+1 n+1}-n\rho_{nn}\big]+c.c.\bigg\}+\mathcal{l}_f[\rho]_{nn}\label{eq21}\end{aligned}\ ] ] @xmath351-\big[\frac{1}{\gamma}\big(\sum_{i=1}^{n}\frac{2\gamma}{\delta_i^2+\gamma^2}\big)\rho_{b_i b_i}^0\nonumber\\ & + \sum_{i < j}\big[\big(\frac{2\cos\phi_{ij}(\delta_i\omega_{b_i b_j}+\gamma\bar{\gamma})+2\sin\phi_{ij}(\omega_{b_i b_j}-\delta_i\bar{\gamma})}{(\delta_i^2+\gamma^2)(\omega_{b_i b_j}^2+\bar{\gamma}^2)}+\frac{2\cos\phi_{ij}(\gamma\bar{\gamma}-\delta_j\omega_{b_i b_j})+2\sin\phi_{ij}(\omega_{b_i b_j}+\delta_j\bar{\gamma})}{(\delta_j^2+\gamma^2)(\omega_{b_i b_j}^2+\bar{\gamma}^2)}\big)|\rho_{b_i b_j}^0|\big]\nonumber\\ & \times\big[(n+1)\rho_{n+1 n+1}-n\rho_{nn}\big]\bigg\}\label{eq22}\end{aligned}\ ] ] @xmath355 , @xmath356 and @xmath357 . note that @xmath358 has @xmath1 number of terms and @xmath359 has @xmath15 number of terms in the summation . since we seek the solutions in the steady state , we obtain the steady state photon number @xmath33 by solving @xmath360 where @xmath361 and we write @xmath362 by using previously obtained @xmath363 as @xmath364 where @xmath371 here , @xmath372 is the average photon number in the absence of atomic coherence in terms of average photon number @xmath373 which is the average photon number in the absence of atomic coherence and in the absence of cavity decay @xmath113 . finally we look at the degenerate ground state case ( @xmath374 , @xmath375 ) . in this case , @xmath53 can be simplified to @xmath376 where @xmath377 , @xmath378 for any @xmath379 . the simplified forms of other parameters are 44ifxundefined [ 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 * * , ( ) in @noop _ _ , vol . ( , ) pp . @noop * * , ( ) in @noop _ _ , vol . ( , ) pp . @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 _ _ , ed . ( , , ) @noop * * , ( ) @noop * * , ( ) @noop _ _ , ed . ( , ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( )

it is well known that if @xmath3 is a self - similar measure arising from an ifs satisfying the open set condition , then the set of local dimensions of the measure is a closed interval whose endpoints are easily computed . further , the hausdorff dimension of the set of points whose local dimension is a given value can be determined using the legendre transform of the @xmath4-spectrum of the measure . this is known as the multifractal formalism and we refer the reader to @xcite for more details . for measures that do not satisfy the open set condition , the multifractal analysis is more complicated and , in general , much more poorly understood . in @xcite , hu and lau discovered that the @xmath5-fold convolution of the classical middle - third cantor measure fails the multifractal formalism as there is an isolated point in the set of local dimensions . subsequently , in @xcite further examples of this phenomena were explored and it was shown , for example , that there is always an isolated point in the set of local dimensions of the @xmath0-fold convolution of the cantor measure associated with a cantor set with ratio of dissection @xmath6 when the integer @xmath2 . more recently , it was proven in @xcite that continuous measures satisfying a weak technical condition have the property that a suitably large convolution power admits an isolated point in its set of local dimensions . in @xcite , ngai and wang introduced the notion of finite type ( see section 2 for the definition ) . this property is stronger than the weak separation condition introduced in @xcite , but is satisfied by many self - similar measures which fail to possess the open set condition . examples include bernoulli convolutions , @xmath7 , with contraction factor @xmath8 equal to the reciprocal of a pisot number @xcite and the cantor - like measures mentioned above . building on earlier work ( c.f . , @xcite ) , feng undertook a study of equicontractive , self - similar measures of finite type in @xcite . his main results were for bernoulli convolutions . in particular , he proved that despite the failure of the open set condition , the multifractal formalism still holds for the bernoulli convolutions whose contraction factor was the reciprocal of a simple pisot number ( meaning , a pisot number whose minimal polynomial is of the form @xmath9 ) . a particularly interesting example is when the contraction factor is the golden ratio with minimal polynomial @xmath10 ( also called the golden mean ) . in this paper we study the local dimension theory of equicontractive , self - similar measures @xmath3 of finite type , whose support is a compact interval and for which the underlying probabilities are regular . we first give a simple formula for the value of the local dimension of @xmath3 at any `` periodic '' point of its support . the finite type condition leads naturally to a combinatorial notion we call a `` loop class '' . for a `` positive '' loop class we prove that the set of attainable local dimensions of the measure is a closed interval and that the set of local dimensions at periodic points in the loop class is a dense subset of this interval . these self - similar measures have a distinguished positive loop class that is called the essential class . thus the set of attainable local dimensions is the union of a closed interval together with the local dimensions at points in loop classes external to the essential class . we will say that a point is an essential point if it is in the essential class . the set of essential points has full lebesgue measure on the support of the measure and in many interesting examples the set of essential points is the interior of the support of the measure . this is the case with many bernoulli convolutions , @xmath7 , including when @xmath11 is the golden ratio ( c.f . section [ ex : other pisot ] ) , and with the @xmath0-fold convolution of the cantor measure on a cantor set with ratio of dissection @xmath1 when @xmath12 ( see section [ sec : cantor ] ) . when the essential set is the interior of the support of the measure @xmath13 then @xmath3 has no isolated point in its set of attainable local dimensions if and only if @xmath14 coincides with the local dimension of @xmath3 at an essential point . in that case , the set of attainable local dimensions of @xmath15 is a closed interval . the bernoulli convolution @xmath7 , with @xmath16 a simple pisot number has this property . however , we construct other examples of bernoulli convolutions ( with contraction factor a pisot inverse ) which do have an isolated point in their set of attainable local dimensions ( see subsection [ bcisolatedpt ] ) . as far as we are aware , these are the first examples of bernoulli convolutions known to admit an isolated point . we also construct a cantor - like measure of finite type , whose set of local dimensions consists of ( precisely ) two distinct points ( see example [ ex : isolated points ] ) . in all of these examples , the essential set is the interior of the support of the measure . the convolution square of the bernoulli convolution , @xmath7 , with @xmath16 the golden ratio , is another example of a self - similar measure to which our theory applies . it , too , has exactly one isolated point in its set of attainable local dimensions , although in this case the set of non - essential points is countably infinite ( see subsection [ ssec : golden square ] ) . the computer was used to help obtain some of these results . in principle , the techniques could be applied to other convolutions of bernoulli convolutions and other measures of finite type , however even with the simple examples given here , the problem can become computationally difficult . the paper is organized as follows : in section [ sec : term ] , we detail the structure of self - similar measures of finite type , introduce terminology and describe a number of examples that we will return to throughout the paper . the notion of transition matrices and properties of local dimensions of measures of finite type are discussed in section [ sec : trans ] . in section [ sec : loop ] we introduce the notion of loop class , essential class and periodic points . a formula is given for the local dimension at a periodic point and we prove that the essential class is always of positive type . in section [ sec : main ] we prove that the set of local dimensions at periodic points in a positive loop class is dense in the set of local dimensions at all points in the loop class . we also show that the set of local dimensions at the points of a positive loop class is a closed interval . in particular , this implies that the set of local dimensions at the essential points is a closed interval . in section [ sec : comp ] we give a detailed description of our computer algorithm by means of a worked example . we also explain our main techniques for finding bounds on sets of local dimensions and illustrate these by constructing a cantor - like measure of finite type whose local dimension is the union of two distinct points . in section [ sec : cantor ] we show that with our approach we can partially recover results from @xcite about the local dimensions of cantor - like measures of finite type . we also show that some facts about the endpoints of the interval portion of the local dimension , that are known to be true for cantor - like measures in the small " overlap case , do not hold in general . bernoulli convolutions , @xmath7 , where @xmath17 is the reciprocal of a pisot number of degree at most four , are studied in section [ sec : pisot ] and we see that two of these measures admit an isolated point . we also study the convolution square of the bernoulli convolution with the golden ratio in this final section . for the examples in this paper , we present only minimal information . a more detailed analysis of all of these examples can found as supplemental information appended to the arxiv version of the paper @xcite . consider the iterated function system ( ifs ) consisting of the contractions @xmath18 , @xmath19 , defined by @xmath20 where @xmath21 , @xmath22 and @xmath23 is an integer . by the associated self - similar set , we mean the unique , non - empty , compact set @xmath24 satisfying @xmath25 suppose @xmath26 , @xmath19 are probabilities , i.e. , @xmath27 for all @xmath28 and @xmath29 . our interest is in the self - similar measure @xmath30 associated to the family of contractions @xmath31 as above , which satisfies the identity @xmath32these measures are sometimes known as _ equicontractive _ , or _ @xmath8-equicontractive _ if we want to emphasize the contraction factor @xmath8 . they are non - atomic , probability measures whose support is the self - similar set . we put @xmath33 . given an @xmath34-tuple @xmath35 @xmath36 , we write @xmath37 for the composition @xmath38 and let @xmath39 the iterated function system , @xmath40 , is said to be of * * finite type * * _ _ _ _ if there is a finite set @xmath41 such that for each positive integer @xmath34 and any two sets of indices @xmath42 , @xmath43 @xmath36 , either @xmath44 where @xmath45 is the diameter of @xmath24 . if @xmath31 is of finite type and @xmath3 is an associated self - similar measure , we also say that @xmath3 is of finite type . it is worth noting here that the definition of finite type is independent of the choice of probabilities . in @xcite , nguyen showed that any ifs satisfying the open set condition is of finite type , but there are also examples of iterated function systems of finite type that fail the open set condition . indeed , recall that an algebraic integer greater than @xmath46 is called a _ pisot number _ if all its galois conjugates are less than @xmath46 in absolute value . examples include integers greater than @xmath46 and the golden ratio , @xmath47 . in ( * ? 2.9 ) , ngai and wang showed that if @xmath16 is a pisot number and all @xmath48 $ ] , then the measure @xmath3 satisfying ( [ ss ] ) is of finite type . this result allows us to produce many examples of measures of finite type that do not satisfy the open set condition . the case when the ifs is generated by two contractions is of particular interest . we will use the notation @xmath7 to denote the self - similar measure @xmath49 where @xmath50 for @xmath51 . [ ex : pisot ] when @xmath52 the measures , @xmath53 are known as * cantor measures * ( or uniform cantor measures ) . their support is the cantor set with ratio of dissection @xmath17 and they satisfy the open set condition . when @xmath54 these measures are called * bernoulli convolutions*. they fail to satisfy the open set condition , but are of finite type whenever @xmath16 is a pisot number . given two probability measures , @xmath55 , the convolution of @xmath3 and @xmath56 is defined as @xmath57the name bernoulli convolution comes from the fact that @xmath58where the infinite convolution is understood to converge in a weak sense . bernoulli convolutions , @xmath7 , with contraction factor @xmath17 the inverse of a pisot number , have been long studied . they have unusual properties and are of interest in fractal geometry , number theory and harmonic analysis . for example , although almost every bernoulli convolution is absolutely continuous with respect to lebesgue measure , and even has an @xmath59 density function , those with a pisot inverse as the contraction factor are not only purely singular , but their fourier transform , @xmath60 , does not even tend to zero as @xmath61 . we refer the reader to @xcite and @xcite for some of the interesting history of these measures . [ ex : convolution ] suppose @xmath3 and @xmath62 are @xmath17-equicontractive measures , say @xmath63 and @xmath64 where @xmath65 and @xmath66 . then @xmath67 is the @xmath17-equicontractive , self - similar measure satisfying @xmath68where @xmath69 and @xmath70 . it follows directly from ngai and wang s result @xcite that any @xmath0-fold convolution power of the bernoulli convolution or cantor measure , @xmath53 is of finite type when @xmath11 is pisot . [ ex : shmerkin ] another consequence of @xcite is that the ifs @xmath71where @xmath72 is an integer , is of finite type . the convex hull of the self - similar set is @xmath73 $ ] and the self - similar set is the full interval @xmath74 $ ] when @xmath75 . when @xmath76 , the open set condition is not satisfied . the @xmath0-fold convolutions of cantor measures with contraction factor @xmath1 are examples of self - similar measures associated with such an ifs . these cantor - like measures were studied in @xcite using different methods . in section [ sec : cantor ] we will see how our approach relates to some of their results . we will refer to the following conditions on a self - similar measure @xmath3 as our standard technical assumptions . 1 . the measure @xmath77 is a @xmath17-equicontractive , self - similar measure , as in ( [ ss ] ) , that is of finite type . 2 . the probabilities , @xmath78 satisfy @xmath79 ( we call these _ regular probabilities _ ) . the support of @xmath3 ( equivalently , the underlying self - similar set ) is a closed interval . by rescaling the @xmath80 appropriately , we can assume without loss of generality that this interval is @xmath73 $ ] . we remark that supp@xmath81 $ ] if and only if ( the rescaled ) @xmath82 satisfy @xmath83 , @xmath84 and @xmath85 for all @xmath86 . in this case , @xmath87 in the definition of finite type . although some of what we say is true more generally for self - similar measures of finite type , we make use of the standard technical assumptions at key points throughout the paper . the bernoulli convolutions @xmath7 and the @xmath0-fold convolutions of uniform cantor measures @xmath7 with @xmath88 are examples of measures satisfying the standard technical assumptions . ( see examples [ ex : pisot ] and [ ex : convolution ] ) . the measures of example [ ex : shmerkin ] , where @xmath75 , are also examples of such measures when regular probabilities are chosen . as we have seen , measures that are of finite type need not satisfy the open set condition . our primary interest is in this case . the finite type property is , however , stronger than the weak separation condition ( see ng for a proof ) , and the multifractal analysis of self - similar measures of finite type is somewhat more tractable because of their better structure . this structure is explained in detail in @xcite , but we will give a quick overview here . for each integer @xmath34 , let @xmath89 be the collection of elements of the set @xmath90 @xmath91 , listed in increasing order . put @xmath92:1\leq j\leq s_{n}\}\text{.}\]]elements of @xmath93 are called _ net intervals of level _ @xmath34 . for each @xmath94 , @xmath95 , there is a unique element @xmath96 which contains @xmath97 . we call @xmath98 the _ parent _ of @xmath97 and @xmath97 a _ child _ of @xmath98 . we denote the _ normalized length _ of @xmath99 $ ] by @xmath100note that by definition there is no @xmath101 with @xmath102 , nor can we have @xmath103 . furthermore , there must be some @xmath104 with @xmath105 for suitable choices of @xmath106 @xmath107 . next , we consider all @xmath101 with @xmath108 $ ] . as @xmath109 $ ] is a closed interval of length @xmath110 this is the same as the set of all @xmath111 with @xmath112 . we suppose @xmath113\}=\{a_{1},\dots , a_{k}\}\]]and assume @xmath114 . we define the _ neighbour set _ of @xmath97 as @xmath115 let @xmath96 be the parent of @xmath97 , and @xmath116 ( listed in order from left to right ) be all the net intervals of level @xmath34 which are also children of @xmath98 and have the same normalized length and neighbour set as @xmath97 . define @xmath117 to be the integer @xmath118 with @xmath119 . the _ characteristic _ _ vector of _ @xmath97 is the triple @xmath120 we also speak of the pair of characteristic vectors , @xmath121 as parent and child if @xmath122 and @xmath123 for a parent / child pair @xmath124 . the characteristic vector is important because it carries the neighbourhood information about @xmath97 . put @xmath125 if the measure is of finite type , then @xmath126 will contain only finitely many distinct characteristic vectors . suppose the net interval , @xmath127 has two children , @xmath128 and @xmath129 that differ only in the value of @xmath130 , that is , they have the same length and the same neighbourhood set . the characteristic vectors for the children of @xmath131 and @xmath132 , will be identical as they depend only on @xmath133 and @xmath134 and not on @xmath135 . for notational reasons , we find it convenient to take advantage of this when drawing the directed graph relating parents to children by suppressing @xmath136 equating these two children on the graph , and allowing multiple edges from @xmath98 to @xmath131 . we will call the characteristic vectors where we have suppressed the numbers @xmath130 the reduced characteristic vectors , and we will call the resulting graph the reduced transition graph . [ ex : golden diagram ] in ( * ? ? ? * section 4 ) and ( * ? ? ? * section 6 ) , feng studied the bernoulli convolution @xmath7 , with @xmath137the golden ratio , and found that there were seven characteristic vectors . their normalized lengths and neighbourhood sets are given by : * characteristic vector 1 : @xmath138 * characteristic vector 2 : @xmath139 * characteristic vectors 3a and 3b : @xmath140 and @xmath141 * characteristic vector 4 : @xmath142 * characteristic vector 5 : @xmath143 * characteristic vector 6 : @xmath144 notice there are only six reduced characteristic vectors ; we label the two characteristic vectors with identical length and neighbourhood set as 3a and 3b . in @xcite these were labelled as 3 and 7 . the directed graphs in figure [ fig : golden ] show the parent / children relationships . the term `` essential class '' , referred to in the figure , is defined in section 4 . [ ex : bc ] by an _ admissible path , _ @xmath145 _ _ of length _ _ @xmath146 we will mean an ordered @xmath34-tuple , @xmath147 where @xmath148 for all @xmath28 and the characteristic vector , @xmath149 is the parent of @xmath150 . by the _ symbolic expression _ of @xmath94 we mean an admissible path of length @xmath151 , denoted @xmath152=(\mathcal{c}_{0}(\delta _ { 0}),\dots,\mathcal{c}_{n}(\delta _ { n})),\ ] ] where @xmath153 and for each @xmath154 , @xmath155 . here @xmath156 is @xmath73 $ ] . feng @xcite proved that the symbolic expression uniquely determines @xmath97 . for @xmath157 $ ] , the symbolic representation for @xmath158 , denoted @xmath159,$ ] will mean the sequence @xmath160 of characteristic vectors where @xmath161 for all @xmath34 and @xmath162 is the parent of @xmath163 . we note that unless @xmath158 is an endpoint of a net interval ( in which case there are two representations of @xmath158 ) , @xmath159 $ ] is unique . the notation @xmath164 $ ] will mean the admissible path consisting of the first @xmath165 characteristic vectors of @xmath159 $ ] . we frequently write @xmath166 for the net interval of level @xmath34 containing @xmath158 . thus @xmath159 $ ] is the sequence where the first @xmath151 terms gives the symbolic representation of @xmath166 for each @xmath34 . given a probability measure @xmath3 , by the * upper local dimension * of @xmath3 _ _ _ _ at @xmath167 supp@xmath3 , we mean the number @xmath168)}{\log r}.\]]replacing the @xmath169 by @xmath170 gives the * lower local dimension * , denoted @xmath171 . if the limit exists , we call the number the * local dimension * of @xmath3 at @xmath158 and denote this by @xmath172 . multifractal analysis refers to the study of the local dimensions of measures . for a @xmath17-equicontractive measure @xmath3 , it is easy to check that @xmath173)}{n\log \varrho } \text { for } x\in \text{supp}\mu , \label{locdim}\ ] ] and similarly for the upper and lower local dimensions . our first several lemmas will enable us to show that we can replace the interval @xmath174 $ ] by @xmath166 . suppose @xmath3 satisfies the standard technical assumptions . let @xmath175\in \mathcal{f}_{n},$ ] with @xmath176 . then @xmath177\sum _ { \substack { \sigma \in \mathcal{a}^{n } \\ \varrho ^{-n}(a - s_{\sigma } ( 0))=a_{i}}}p_{\sigma } .\ ] ] this argument can basically be found in feng @xcite , but we give the details here for completeness . iterating ( [ ss ] ) @xmath34 times gives @xmath178 since @xmath3 is a non - atomic measure supported on @xmath73 $ ] , we have @xmath179 now , @xmath180 implies that @xmath181 $ ] , hence by definition of the neighbourhood set @xmath182 for some @xmath183 . thus @xmath184 , so @xmath185 = [ a - a_i \varrho^n , a - a_i + \varrho^n]$ ] . this implies that @xmath186 we observe that @xmath187 , and hence @xmath188 ) & = [ a_i \varrho^n + a - a_i \varrho^n , a_i \varrho^n + a - a_i \varrho^n + \ell_n(\delta)\varrho^n ] \\ & = [ a , a+\ell_n(\delta ) \varrho^n ] = [ a , b ] = \delta.\end{aligned}\ ] ] hence @xmath189 \\ & = \sum_{i=1}^{k } \mu [ a_i , a_i + \ell_n(\delta ) ] \sum_{\substack{\sigma \in \mathcal{a}^{n}\\ \varrho^{-n}(a - s_{\sigma}(0))=a_{i } } } p_{\sigma } \end{aligned}\ ] ] as claimed . for @xmath190card@xmath191 , put @xmath192and @xmath193 here we have chosen to normalize by multiplying by @xmath194 . this is done so that the minimal non - zero entry in the transition matrices ( defined in the next subsection ) is at least @xmath46 . [ cor]there is a constant @xmath195 such that for any @xmath34 and any @xmath196 @xmath197 the upper bound is clear from the lemma . for the lower bound we note that each @xmath198 > 0 $ ] as the support of @xmath3 is the full interval @xmath73 $ ] . the finite type condition ensures there are only finitely many choices for @xmath199 $ ] . [ lem1 ] suppose @xmath3 satisfies the standard technical assumptions . there are constants @xmath200 such that if @xmath131 , @xmath201 are two adjacent net intervals of level @xmath34 , then @xmath202 the proof is similar to that of ( * ? ? ? * lemma 2.11 ) and proceeds by induction on @xmath34 . the base case holds as there are only finitely many choices for @xmath203 when @xmath204 . now assume the result for level @xmath205 and we will verify it holds for level @xmath34 . if @xmath206 have the same parent @xmath98 , the result follows easily from the observation that @xmath207 otherwise , @xmath131 and @xmath201 are children of adjacent net intervals of level @xmath205 , @xmath208 respectively , and we can suppose @xmath131 is to the left of @xmath209 . as in @xcite , put @xmath210\text { and they share the same right endpoint\thinspace } \ } \\ d_{2 } & = \{\sigma \in \mathcal{a}^{n-1}:\widehat{\delta _ { 2}}\text { } \subseteq s_{\sigma } [ 0,1]\text { and they share the same left endpoint\thinspace } \ } \\ e_{j } & = \{\sigma \in \mathcal{a}^{n-1}\diagdown d_{j}:\widehat{\delta _ { j } } \subseteq s_{\sigma } [ 0,1]\}\text { , } j=1,2.\end{aligned}\ ] ] the definitions ensure that @xmath211 , @xmath212 and @xmath213 applying the induction assumption gives @xmath214 taking @xmath215 gives @xmath216 the other inequality is similar . note that in the proof the assumption that @xmath217 were regular probabilities was important . the following is immediate from the two previous results . there are constants @xmath218 such that if @xmath206 are adjacent net intervals of level @xmath34 , then @xmath219 together these results yield the following useful approach to computing local dimensions . [ cor1]suppose @xmath3 satisfies the standard technical assumptions . let @xmath167supp@xmath3 and @xmath166 denote a net interval of level @xmath34 containing @xmath158 . then @xmath220a similar statement holds for the ( lower ) local dimensions . since any net interval of level @xmath34 has length at most @xmath221 , the interval @xmath174 $ ] contains @xmath166 . the finite type property ensures it is contained in a union of a bounded number of @xmath34th level net intervals , say @xmath222 , where @xmath223 is adjacent to @xmath224 and for a suitable index @xmath28 , @xmath225 . thus for constants @xmath226 , @xmath227 ( independent of the choice of @xmath34 and @xmath158 ) , @xmath228)\leq \sum_{j=1}^{n}\mu ( \delta _ { n}(x_{j } ) ) \\ & \leq \sum_{j=1}^{n}p_{0}^{n}p_{n}(\delta _ { n}(x_{j}))\leq np_{0}^{n}c^{n}n^{n}p_{n}(\delta _ { n}(x)).\end{aligned}\ ] ] thus the limiting behaviour of the three expressions @xmath229)}{n\log \varrho } , \text { } \frac{\log \mu ( \delta _ { n}(x))}{n\log \varrho } \ ] ] and @xmath230 coincide . if @xmath231 $ ] , then @xmath232 and hence @xmath233 . consequently , @xmath234 . more generally , since @xmath235 is never empty and @xmath236 is the minimal probability , it follows that @xmath237 for all @xmath34 and @xmath238 . consequently , @xmath239 the results of the previous subsection show that for studying the local dimensions of these measures it will be helpful to make accurate estimates of @xmath240 . towards this , slightly modifying @xcite we define _ primitive transition matrices _ , @xmath241 @xmath242 for a net interval @xmath175 $ ] of level @xmath34 and parent @xmath243 $ ] as follows : suppose @xmath244 and @xmath245 . for @xmath246 and @xmath247 we set @xmath248if @xmath249 and there exists @xmath250 with @xmath251 and @xmath252 . this is equivalent to saying @xmath253we set @xmath254 otherwise . as @xmath126 is finite for a measure of finite type , there is an upper bound on the size of these matrices . the entries are non - negative and all non - zero entries are at least one . each column has at least one non - zero entry because @xmath255 if and only if there is some @xmath256 which `` contributes '' to it , in the sense defined above . it is also important to note that the standard technical assumption that supp@xmath81 $ ] guarantees that given @xmath250 such that @xmath257 , there exists @xmath249 such that @xmath258 $ ] . this means that each row of the matrix @xmath259 also has a non - zero entry . for @xmath260 card@xmath261 , put @xmath262the same reasoning as in ( * ? ? ? * theorem 3.3 ) shows that @xmath263thus if @xmath264=(\gamma _ { 0},\dots,\gamma _ { n})$ ] ( that is , @xmath265 and @xmath266)$ ] ) , then since @xmath267=\left [ 1\right ] $ ] we have @xmath268where by the norm of matrix @xmath269 we mean @xmath270 given an admissible path @xmath271 , we write @xmath272 and refer to such a product as a _ transition matrix_. with this notation , the results of the previous subsection can be stated as [ corlocdim ] suppose @xmath3 satisfies the standard technical assumptions . if @xmath167supp@xmath3 , then @xmath273)\right\vert } { n\log \varrho } \label{f2}\]]and similarly for the ( lower ) local dimension . [ ex : golden transition ] again , consider the bernoulli convolution , @xmath274 , with @xmath16 the golden ratio . feng @xcite showed that @xmath275 has symbolic representation @xmath276 $ ] and that @xmath277 $ ] . applying corollary [ corlocdim ] gives another proof that @xmath278 . next , we give a useful simple lemma . [ compare ] suppose @xmath3 satisfies the standard technical assumptions . let @xmath279 and @xmath280 be transition matrices . then @xmath281 and @xmath282 . we have @xmath283since all the entries of each of the matrices is nonnegative and each column of @xmath279 has an entry @xmath284 , it follows that @xmath285 the argument for the other inequality is similar , noting that each row of @xmath279 has an entry @xmath284 as a consequence of the standard technical assumptions . an important consequence of this result is that the local dimension of @xmath3 at @xmath158 depends only on the tail of the symbolic representation of @xmath158 . suppose @xmath159=(\gamma _ { 0},\gamma _ { 1},\dots)$ ] . for any @xmath165 , @xmath286and similarly for the ( lower ) local dimension . if @xmath287=(\gamma _ { 0},\beta _ { 1},\dots,\beta _ { n},\gamma _ { n},\gamma _ { n+1},\dots ) $ ] , then the ( upper or lower ) local dimensions of @xmath3 at @xmath158 and @xmath288 agree . this holds since @xmath289 by @xmath290 we mean the spectral radius of the square matrix @xmath291 the largest eigenvalue of @xmath292 in absolute value . recall that @xmath293 we will call a matrix @xmath292 _ positive _ if all its entries are strictly positive and write @xmath294 . we record here some elementary facts about positive matrices that will be useful later . [ normsprod ] suppose @xmath3 satisfies the standard technical assumptions . assume @xmath295 are transition matrices and @xmath280 is positive . 1 . then @xmath296 . [ it2:1 ] 2 . there is a constant @xmath297 such that if @xmath298 is a square matrix , then @xmath299 [ it2:2 ] 3 . suppose @xmath280 is a square matrix . there is a constant @xmath300 such that @xmath301 [ it2:3 ] to see ( [ it2:1 ] ) , let @xmath302 . as all @xmath303 a simple calculation gives @xmath304 for ( [ it2:2 ] ) , assume @xmath305 is a @xmath306 matrix . let @xmath307 . as the entries of @xmath279 are non - negative and the entries of @xmath280 are at least 1 , it is easy to see that @xmath308with the final inequality holding because the sum of the diagonal entries of @xmath298 is the sum of the eigenvalues of @xmath298 , counted by multiplicity . for ( [ it2:3 ] ) , let @xmath309 be the jordan decomposition of @xmath280 and let @xmath310 . by the perron - frobenius theory , @xmath311 is a simple root of the characteristic polynomial of @xmath280 and all other eigenvalues of @xmath280 are strictly less than @xmath311 in modulus . since all entries of @xmath280 are at least @xmath46 , it can be easily seen that @xmath312 . as @xmath313 , it is enough to prove @xmath314 where @xmath315 depends on @xmath280 , but not @xmath34 . assume @xmath280 is of size @xmath316 . since the jordan block for @xmath311 is @xmath317 , all entries of @xmath318 , other than the @xmath319 entry which is @xmath320 , are either @xmath275 or of the form @xmath321 where @xmath322 and @xmath323 is an eigenvalue of @xmath280 with @xmath324 . thus @xmath325where @xmath326 is the maximum of @xmath46 and the modulus of the eigenvalues of @xmath280 other than @xmath311 . as @xmath327 for all @xmath34 sufficiently large depending on @xmath328 , it follows that for some constant @xmath329 depending on @xmath330 ( and hence depending only on @xmath280 ) we have @xmath331 for all @xmath34 . this proves the right hand inequality . the left hand inequality is obvious . feng in @xcite also introduced the notion of an essential class for measures of finite type . here we introduce the more general definition of a loop class , of which the essential class is a special case . \(i ) a non - empty subset @xmath332 is called a * * loop class * * if whenever @xmath333 , then there are characteristic vectors @xmath334 , @xmath335 , such that @xmath336 , @xmath337 and @xmath338 is an admissible path with all @xmath339 . \(ii ) a loop class @xmath332 is called an**essential class * * if , in addition , whenever @xmath340 and @xmath341 is a child of @xmath323 , then @xmath342 . \(iii ) we call a loop class * maximal * if it is not properly contained in any other loop class . the finite type property ensures that every element in the support of @xmath15 is contained in a loop class . clearly every loop class is contained in a unique maximal loop class . feng in ( * ? ? ? * lemma 6.4 ) , proved there is always precisely one essential class . of course , the essential class is a maximal loop class . we will denote the essential class by @xmath343 and here - after speak of `` the '' essential class . if @xmath159=(\gamma _ { 0},\gamma _ { 1},\gamma _ { 2},\dots)$ ] with @xmath344 for all large @xmath28 , we will say that _ _ _ _ @xmath158 is an * essential point * ( or is * in the essential class * ) and call @xmath158 a * non - essential point * otherwise . the phrase , @xmath158 is * in the loop class * @xmath345 will have a similar meaning . an admissible path will be said to be in a given loop class if all its members are in that class . note that if @xmath159=(\gamma _ { j})$ ] is non - essential , then none of the characteristic vectors @xmath334 belong to @xmath343 . a non - essential point necessarily has its tail in some loop class external to the essential class . we remark that the essential class is dense in @xmath73 $ ] . this is because the uniqueness of the essential class ensures that every net interval contains a net subinterval of higher level whose characteristic vector is in the essential class . in fact , we show next that the set of essential points has full lebesgue measure in @xmath73 $ ] . suppose @xmath3 satisfies the standard technical assumptions . then the set of non - essential points is a subset of a closed set of lebesgue measure @xmath275 . as we already observed , every net interval contains a descendent net subinterval whose characteristic vector is in the essential class . ( we will abuse notation slightly and call such a net subinterval `` essential '' . ) the finite type property ensures we can find such a net subinterval in a bounded number of generations and that there exists some @xmath346 such that the proportion of the length of the net subinterval to the length of the original interval is @xmath347 . we now exhibit a cantor - like construction . we begin with @xmath73 $ ] . consider the first level at which there is a net subinterval that is essential . remove the interiors of all the net subintervals of this level that are essential . the resulting closed subset of @xmath73 $ ] is a finite union of closed intervals , say @xmath348 , whose lengths total at most @xmath349 . we repeat the process of removing the interiors of the essential net subintervals at the next level at which there are essential net subintervals in each of the intervals of @xmath348 . the resulting closed subset now has length at most @xmath350 . after repeating this procedure @xmath351 times one can see that the non - essential points are contained in a finite union of closed intervals , @xmath352 whose is total length is at most @xmath353 . it follows that the non - essential points are contained in the closed set @xmath354 and this set has measure @xmath275 . it is worth remarking that the construction above may leave some essential points within the cantor - like construction . as the resulting set is measure @xmath355 , the smaller set of just the non - essential points will also be measure @xmath356 . [ ex : golden essential ] from figure [ fig : golden ] one can see that the bernoulli convolution @xmath7 , with @xmath16 the golden ratio , has seven distinct loop classes : @xmath357 , @xmath358 , @xmath359 , @xmath360 , @xmath361 , @xmath362 , and @xmath363 . of these , @xmath362 , @xmath364 and the essential class , @xmath357 ( with 3 reduced elements ) are maximal . the two loop classes external to the essential class , @xmath362 and @xmath365 are associated to the two endpoints , @xmath275 and @xmath46 . these are the only two non - essential points , in other words , the set of essential points is @xmath366 . we will say the loop class @xmath367 is of * positive type * if there is an admissible path @xmath368 in @xmath367 such that @xmath369 is a positive matrix . we note that as there is a non - zero entry in each row and column of each primitive transition matrix , then any loop class @xmath367 of positive type has the property that for every @xmath370 there is an admissible path @xmath371 in @xmath367 such that @xmath372 is positive . [ ex : golden positive type ] the loop classes @xmath362 and @xmath363 of example [ ex : golden essential ] are of positive type since ( as shown in ( * ? ? ? * section 4 ) ) @xmath373 $ ] . as @xmath374 , $ ] the essential class , @xmath375 is also of positive type . however , the loop class , @xmath376 is not of positive type since all transition matrices from this loop class are of the form @xmath377 $ ] for some @xmath34 . in example [ ex:@xmath378 ) ] we give a bernoulli convolution where a maximal loop class is not of positive type . however , the essential class is always of positive type under our standard technical assumptions . [ prop : pos ] suppose @xmath3 satisfies the standard technical assumptions . then the essential class is of positive type . fix @xmath379 . we recall that feng in ( * ? ? ? * lemma 6.4 ) showed that given any positive integer @xmath380 cardinality@xmath381 , there is an admissible path @xmath382 in the essential class , going from @xmath383 to @xmath384 such that all entries of the @xmath351th row of @xmath385 are non - zero . choose such a path , @xmath386 with all the entries of row @xmath46 of @xmath387 non - zero . row @xmath388 of @xmath389 has a non - zero entry in some column , say @xmath390 and , of course , the @xmath391 entry is also non - zero . choose an essential , admissible path @xmath392 from @xmath393 to @xmath384 with all entries of row @xmath351 non - zero . matrix multiplication shows that all entries of both rows 1 and 2 of @xmath394 are non - zero . by repeated application of this reasoning we can construct an admissible path @xmath368 in @xmath395 from @xmath393 to @xmath396 such that all entries of @xmath372 are strictly positive . we call @xmath167supp@xmath3 a * periodic point * if @xmath158 has symbolic representation@xmath397=(\gamma _ { 0},\dots,\gamma _ { s},\theta ^{-},\theta ^{-},\dots)\]]where @xmath398 is an admissible cycle ( a non - trivial path with the same first and last letter ) and @xmath399 is the path with the last letter of @xmath398 deleted . we refer to @xmath398 as a period of @xmath158 . we will say that periodic @xmath158 is _ _ _ _ * positive * if the square transition matrix @xmath400 is positive . it is worth noting that @xmath398 and @xmath399 are not uniquely defined . for example , the path @xmath401 from example ex : golden essential can be decomposed as @xmath402 and @xmath403 , or as @xmath404 and @xmath405 , or as @xmath404 and @xmath406 , etc . in what follows , it will not make any significant difference as to which choice is made . observe that @xmath407)=t(\gamma _ { 0},\dots , \gamma _ { s},\delta _ { 1})(t(\theta ) ) ^{n}\]]when @xmath408 and @xmath409 is the first letter of @xmath410 . ( recall @xmath411 is the length of path @xmath399 . ) a periodic @xmath158 is in the essential class if it has a period that is a path in the essential class . there is a simple formula for the local dimensions at periodic points . [ periodic ] suppose @xmath3 satisfies the standard technical assumptions . let @xmath158 be the periodic point with period @xmath398 . then@xmath412 suppose @xmath159=(\gamma _ { 0},\dots,\gamma _ { s},\theta ^{-},\theta ^{-},\dots)$ ] and @xmath413 . let @xmath414 . according to lemma [ compare ] , there is a constant @xmath195 such that @xmath415for any @xmath416 . thus corollary [ corlocdim ] implies that @xmath417since @xmath418 @xmath419 the result follows . in @xcite and @xcite , feng showed that the set of local dimensions for the bernoulli convolution @xmath53 with @xmath16 the golden ratio , was an interval and determined its endpoints . this is not true , in general , for measures of finite type . for instance , it is known that the set of local dimensions of the @xmath0-fold convolution of uniform cantor measures with contraction factor @xmath6 for integer @xmath420 is the union of an interval and an isolated point ( see @xcite ) . feng also proved that the set of attainable local dimensions of @xmath274 was the closure of the set of local dimensions at periodic points . in this section we will prove that if a loop class , @xmath421 is of positive type , then the set of local dimensions at points in @xmath422 is a closed interval . moreover , this interval is the closure of the set of local dimensions at the periodic points in @xmath423 . these statements hold , in particular , for the essential class , @xmath424 since the essential class is a positive loop class according to proposition [ prop : pos ] . in sections [ sec : comp ] , [ sec : cantor ] and [ sec : pisot ] we give examples to illustrate that the set of local dimensions attained at points of a loop class external to the essential class can overlap , or may be disjoint from the local dimensions of the essential class . [ periodicdense ] suppose @xmath3 satisfies the standard technical assumptions . assume that @xmath367 is a loop class of positive type . the set of local dimensions of @xmath3 at positive , periodic points in the loop class @xmath367 is dense in the set of all local dimensions at points in @xmath367 . it is also dense in the set of all upper ( or lower ) local dimensions at points in @xmath367 . we remark that in this theorem ( and other results of this section ) the assumption that the loop class is of positive type may not be necessary for particular ifs and particular loops . we will prove denseness in the set of lower local dimensions at points in the loop class @xmath367 . the arguments for the ( upper ) local dimensions are the same . fix @xmath158 in @xmath345 say @xmath159=(\gamma _ { 0},\gamma _ { 1},\gamma _ { 2},\dots)$ ] with @xmath425 for all @xmath426 . choose a subsequence @xmath427 such that @xmath428as @xmath367 is finite , by passing to a further subsequence if necessary ( not renamed ) we can assume all @xmath429 . put @xmath430 let @xmath368 be an admissible path in @xmath367 going from @xmath393 to @xmath431 such that @xmath432 . ( we remark that it is important that this transition matrix is independent of the choice of @xmath351 . ) of course , then @xmath433 is an admissible cycle in @xmath367 . as @xmath372 is a positive matrix , lemmas [ compare ] and [ normsprod](2 ) imply that there are constants , @xmath434 , bounded above and bounded below away from @xmath275 such that @xmath435 consider the local dimension at the periodic point @xmath436 in @xmath423 given by @xmath437=(\gamma _ { 0},\dots,\gamma _ { m-1},\theta _ { k}^{-},\theta _ { k}^{-},\dots).\]]since @xmath438 is positive , @xmath436 is a positive point . furthermore , we have@xmath439fix @xmath440 . the choice of @xmath441 and boundedness of @xmath434 ensures we can choose @xmath351 large enough so that @xmath442@xmath443it is now easy to see that@xmath444 since the essential class is of positive type , we immediately deduce the following important fact about these measures of finite type . the set of local dimensions at essential periodic points is dense in the set of local dimensions at all essential points . here is another needed elementary fact whose proof is an exercise for the reader . [ lem : cyclicshift ] let @xmath445 be a cycle and let @xmath446 be any cyclic shift of @xmath398 . then @xmath447 [ limitperiodic ] suppose @xmath3 satisfies the standard technical assumptions . assume that the loop class , @xmath345 is of positive type and suppose @xmath448 is a sequence of positive , periodic points in @xmath422 . then there is some @xmath158 in @xmath367 such that @xmath449 assume @xmath450 has period @xmath451 in @xmath367 and that @xmath452 . we put @xmath453so @xmath454and @xmath455by passing to a subsequence , not renamed , we can assume all @xmath456where @xmath457 is a decreasing sequence tending to 0 . further , we can assume that all the even labelled paths , @xmath458 , have the same first letter , say @xmath459 , and the same last letter , @xmath460 . similarly , we can assume all the odd labelled paths have a common first letter @xmath461 and common last letter @xmath462 . since @xmath367 is of positive type , we can certainly choose two admissible paths in @xmath367 , @xmath463 going from @xmath464 to @xmath461 and @xmath465 going from @xmath462 to @xmath466 , so that @xmath467 and @xmath468 are positive . we want to inductively define a rapidly increasing subsequence @xmath469 such that @xmath397=(\eta , \underbrace{\theta _ { 1}^{-},\dots,\theta _ { 1}^{-}}% _ { k_{1}},\lambda _ { is},\underbrace{\theta _ { 2}^{-},\dots,\theta _ { 2}^{-}}% _ { k_{2}},\lambda _ { si},\underbrace{\theta _ { 3}^{-},\dots,\theta _ { 3}^{-}}% _ { k_{3}},\dots)\]]has the desired property , where @xmath368 is an admissible path beginning with @xmath470)$ ] and ending with the parent of @xmath461 . temporarily fix @xmath34 . with abuse of notation we will write the truncated product as @xmath397_{n}=\eta \prod_{i=1}^{n}(\theta _ { i}^{-})^{k_{i}}\lambda _ { i}\]]where @xmath471 if @xmath183 is even , and @xmath465 if @xmath183 is odd . define @xmath472 as the maximal @xmath473 coming from lemma [ normsprod](3 ) taken over all cyclic shifts of @xmath474 of @xmath475 . let @xmath476 be any prefix of @xmath477 and @xmath478 any prefix of @xmath479 . the elementary lemmas of the previous section imply that @xmath480_{n-1}(\theta _ { n}^{-})^{k}\lambda _ { n}^{\prime } ) \right\vert ) } { % l([x]_{n-1}(\theta _ { n}^{-})^{k}\lambda _ { n}^{\prime } ) } -\frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}\right\vert = 0 \label{2 - 1}\]]and@xmath480_{n-1}(\theta _ { n}^{-})^{k}\lambda _ { n}\theta _ { n+1}^{\prime } ) \right\vert ) } { l([x]_{n-1}(\theta _ { n}^{-})^{k}\lambda _ { n}\theta _ { n+1}^{\prime } ) } -\frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}% \right\vert = 0 . \label{2 - 2}\]]obviously , @xmath481_{n-1}(\theta _ { n}^{-})^{k}\lambda _ { n}\theta _ { n+1}^{\prime } ) } = 0 \label{2 - 3}\]]and@xmath482we pick @xmath483 such that the left hand side of these four limits are each less than @xmath457 for all choices of @xmath476 and @xmath478 . certainly this process defines an @xmath158 belonging to @xmath367 . we need to check that the upper and lower dimensions of @xmath3 at @xmath158 are correct . by construction , the limiting behaviour ( as @xmath484 ) of @xmath485_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}^{\prime } ) ) \right\vert ) } { l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}^{\prime } ) ) } \ \ \mathrm{and}\ \ \frac{\log ( \left\vert t([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ) \right\vert ) } { l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ) } \]]approach the values @xmath486 and @xmath487 in the infimum / supremum respectively . hence it remains to consider the case @xmath485_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}(\theta _ { n+1}^{-})^{p_{n+1}}\theta _ { n+1}^{\prime } ) \right\vert ) } { % l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}(\theta _ { n+1}^{-})^{p_{n+1}}\theta _ { n+1}^{\prime } ) } \]]for @xmath488 , and @xmath478 some prefix of @xmath489 . this is equivalent to @xmath485_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ( \theta _ { n+1}^{\ast -})^{p_{n+1}})\right\vert ) } { % l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ( \theta _ { n+1}^{\ast -})^{p_{n+1}})}\]]for some cyclic shift @xmath490 of @xmath475 . for ease of notation , let @xmath491_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ( \theta _ { n+1}^{\ast -})^{p_{n+1}})}$ ] . first , observe that : @xmath492_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ( \theta _ { n+1}^{\ast -})^{p_{n+1}})\right\vert ) } { l_{0 } } \\ & \leq \frac{\log ( \left\vert t([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } \right\vert ) } { l_{0}}+\frac{\log ( \left\vert t(\theta _ { n+1}^{\ast } ) ^{p_{n+1}})\right\vert ) } { l_{0 } } \\ & \leq \frac{l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ) } { l_{0}}\cdot \frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}+\frac{\log ( k_{n}sp((t(\theta _ { n+1}))^{p_{n}+1}))}{l_{0}}% + \varepsilon _ { n } \\ & \leq \frac{l_{0}-l((\theta _ { n+1}^{-})^{p_{n}+1})}{l_{0}}\cdot \frac{% \log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}\\&\hspace{10pt } + \frac{l((\theta _ { n+1}^{-})^{p_{n}+1})}{l_{0}}\cdot \frac{\log ( sp(t(\theta _ { n+1}))^{p_{n}+1})}{l((\theta _ { n+1}^{-})^{p_{n}+1})}+\frac{\log k_{n}}{% l_{0}}+\varepsilon _ { n } \\ & \leq \frac{l_{0}-l((\theta _ { n+1}^{-})^{p_{n}+1})}{l_{0}}\cdot \frac{% \log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}\\&\hspace{10pt } + \frac{l((\theta _ { n+1}^{-})^{p_{n}+1})}{l_{0}}\cdot \frac{\log ( sp(t(\theta _ { n+1})))}{% l(\theta _ { n+1}^{-})}+2\varepsilon _ { n } \\ & = ( 1-t_{n})\left ( \frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}% \right ) + t_{n}\left ( \frac{\log ( sp(t(\theta _ { n+1})))}{l(\theta _ { n+1}^{-})}% \right ) + 2\varepsilon _ { n}.%\end{aligned}\]]for @xmath493here the second inequality comes from ( [ 2 - 2 ] ) , lemma [ normsprod](3 ) and lemma [ lem : cyclicshift ] , and the final inequality from ( [ 2 - 3 ] ) . for the opposite inequality , we use the fact that @xmath494 and lemma [ normsprod](1 ) for the first inequality below , and ( [ 2 - 1 ] ) and ( [ 2 - 4 ] ) for the second and third inequalities to obtain @xmath495_{n-1}(\theta _ { n}^{-})^{k_{n}}\lambda _ { n}\theta _ { n+1}^{\prime } ( \theta _ { n+1}^{\ast -})^{p_{n+1}})\right\vert ) } { l_{0 } } \\ & \geq \frac{\log ( \left\vert t([x]_{n-1}(\theta _ { n}^{-})^{k_{n}}\right\vert ) } { l_{0}}+\frac{\log ( \left\vert t((\theta _ { n+1}^{\ast } ) ^{p_{n+1}})\right\vert ) } { l_{0 } } \\ & \geq \frac{l([x]_{n-1}(\theta _ { n}^{-})^{k_{n}})}{l_{0}}\cdot \frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}\\ & \hspace{10pt } + \frac{l((\theta _ { n+1}^{-})^{p_{n}+1})}{l_{0}}\cdot \frac{\log ( sp(t(\theta _ { n+1}))^{p_{n}+1})}{l((\theta _ { n+1}^{-})^{p_{n}+1})}-\varepsilon _ { n } \\ & \geq ( 1-t_{n})\left ( \frac{\log ( sp(t(\theta _ { n})))}{l(\theta _ { n}^{-})}% \right ) + t_{n}\left ( \frac{\log ( sp(t(\theta _ { n+1})))}{l(\theta _ { n+1}^{-})}% \right ) -2\varepsilon _ { n}%\end{aligned}\]]together these estimates prove @xmath496 lies within @xmath497 of the same convex combination of @xmath498 and @xmath499 , and hence @xmath500 and @xmath501 both belong to the interval @xmath502 $ ] . that completes the proof . combining theorems [ periodicdense ] and [ limitperiodic ] , it follows that these measures satisfy the following . the set of local dimensions at essential points coincides with the set of lower ( or upper ) local dimensions at essential points . moreover , @xmath503 the final result of this section will be to show that the set of local dimensions at points in a loop class of positive type is a closed interval . of course , in particular , this applies to the essential class . [ interval ] suppose @xmath3 satisfies the standard technical assumptions . further , suppose that the loop class @xmath367 is of positive type . assume @xmath288 and @xmath504 are periodic , positive points in @xmath423 . then the set of local dimensions of @xmath3 contains the closed interval with endpoints @xmath505 and @xmath506 . suppose @xmath288 has period @xmath507 and @xmath504 has period @xmath398 where @xmath508 and @xmath509 are positive matrices with spectral radii @xmath510 and @xmath311 , respectively . with this notation , @xmath511let @xmath512 . we want to prove that there exists a @xmath158 such that @xmath513 appealing to theorem [ limitperiodic ] , we see it will be enough to show that there is some sequence of periodic , positive points @xmath514 in the loop class @xmath345 such that @xmath515 to do this , we start by choosing admissible paths in @xmath367 , @xmath516 joining the last letter of @xmath398 to the first letter of @xmath517 and @xmath392 doing the opposite , such that @xmath518 . then for any positive integers , @xmath519 , @xmath520 is a square transition matrix . select two sequences of integers , @xmath521 , @xmath522 tending to infinity , with @xmath523we will prove that the periodic points with period @xmath524 satisfying@xmath525 work . as @xmath518 , lemmas [ compare ] and [ normsprod ] combine to imply that for all @xmath390@xmath526and@xmath527 thus @xmath528it follows that ( [ key ] ) is satisfied . combining the three theorems we deduce that the set of local dimensions at the essential points of such measures is a closed interval whose endpoints are given by the infimum and the supremum of the local dimensions at essential , periodic points . [ intend]let @xmath529 essential , positive periodic@xmath530 and @xmath531 essential and positive periodic@xmath532 then@xmath533\text{.}\ ] ] another immediate corollary is that if the set of essential points is @xmath366 , then the measure admits at most one isolated point . if the set of essential points of @xmath3 is @xmath366 , then the set of local dimensions of @xmath3 consists of a closed interval together with @xmath534 . [ ex : golden local dimension ] as observed in example [ ex : golden essential ] , this is the situation for the bernoulli convolution @xmath274 , with @xmath16 the golden ratio . further , since @xmath535 and @xmath360 is in the essential class , it follows that @xmath14 coincides with the local dimension at an essential point . consequently , the set of local dimensions of @xmath274 is equal to the closed interval @xmath502 $ ] consisting of the local dimensions at the essential points of @xmath7 . in @xcite it is shown that @xmath536 and @xmath537 . more generally , we have the following . if every maximal loop class is of positive type , then the set of local dimensions of @xmath3 is a finite union of closed intervals . these intervals can be disjoint . indeed , in example [ ex : isolated points ] we construct a measure whose set of local dimensions consists of two points ( degenerate intervals ) . given a family of contractions , @xmath538 , and probabilities @xmath26 , we have implemented an algorithm to find @xmath126 and all of the associated transition matrices . we use a modified version of @xcite to perform these calculations . we give an overview here , by means of a worked example . for this we consider the six contractions , the maps @xmath539 for @xmath540 , for @xmath541 the corresponding self - similar set is the @xmath542-fold sum of the middle - third cantor set rescaled to @xmath73 $ ] . suppose the maps have normalized probabilities @xmath543 . one can see that @xmath544\in \mathcal{f}_{0}$ ] has characteristic vector @xmath545 . we will call this characteristic vector 1 . here the first @xmath46 is the normalized length of the interval . the sequence @xmath546 is @xmath547 is the neighbourhood set . the last @xmath548 is @xmath549 . for the questions we are interested in , @xmath550 is not needed , hence we will suppress it in the future , instead allowing multiple edges between nodes in the graph . this gives a reduced characteristic vector of @xmath551 . instead of considering all intervals in @xmath552 , we only consider those that arise from new reduced characteristic vectors found to be in @xmath553 . so , initially , we look at those children of reduced characteristic vector 1 . in this case , these two things are the same , but we will see later on that this is not always the case . we first subdivide @xmath73 $ ] by considering the maps @xmath554 and and @xmath555 for all @xmath556 and @xmath557 . this partitions the interval at the points @xmath558 . consider first @xmath559 $ ] coming from this subdivision . we see that its normalized length is @xmath560 and it has neighbourhood set @xmath546 . we label @xmath561 as the reduced characteristic vector 2 . the @xmath80 that contributes to @xmath275 is @xmath562 and is associated to the normalized probability of @xmath46 . hence the transition matrix is @xmath563 $ ] . now , consider @xmath564 $ ] . this again has normalized length @xmath565 and neighbourhood set @xmath566 . we label @xmath567 as the reduced characteristic vector 3 . the @xmath275 comes from @xmath568 and the @xmath565 from @xmath569 , hence the transition matrix is @xmath570 $ ] . we continue in this fashion , noting that for @xmath571 $ ] , where @xmath572 , the characteristic vector is @xmath573 , which we label as 4 . in feng s notation , we would distinguish these as four different characteristic vectors by use of their third component , @xmath574 . we will not distinguish these , but will instead allow multiple maps from the reduced characteristic vector 1 to the reduced characteristic vector 4 . these are @xmath575 \text { } \mathrm{or}\ \left [ \begin{array}{ccc } p_{3 } & p_{2 } & p_{1}% \end{array}% \right ] \\ & \mathrm{or}\ \left [ \begin{array}{ccc } p_{4 } & p_{3 } & p_{2}% \end{array}% \right ] \mathrm{or}\ \left [ \begin{array}{ccc } p_{5 } & p_{4 } & p_{3}% \end{array}% \right ] .\end{aligned}\ ] ] continuing in this fashion , we can compute the finite set of reduced characteristic vectors obtaining * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath576 * reduced characteristic vector 3 : @xmath577 * reduced characteristic vector 4 : @xmath578 * reduced characteristic vector 5 : @xmath579 * reduced characteristic vector 6 : @xmath580 * reduced characteristic vector 7 : @xmath581 the characteristic vectors 4 and 5 comprise the essential set . , with @xmath582 for a complete list of transition matrices , see the supplementary information @xcite . given a loop class of positive type , there are two main ways we obtain good estimates on the set of local dimensions associated to this loop class . the first is to find explicit examples of periodic points within the loop class and calculate their local dimensions by determining the spectral radius of the transition matrix of the cycle . applying theorem [ periodicdense ] , this will produce an interval contained in the set of local dimensions of the loop class . to find upper and lower bounds on the set of local dimensions we use the family of pseudo - norms : @xmath583these are the sub - norm on indices @xmath227 , the total sub - norm and the total sup - norm respectively . the total sup - norm is actually a norm . here @xmath227 is a subset of the indices of the column vectors . care must be take here that the subset @xmath227 is valid for all matrices within the loop class one is considering , as different transition matrices may have different dimensions . for all matrices , @xmath584 we have @xmath585and latexmath:[\[\left\vert t\right\vert _ { \min } , \left\vert t\right\vert _ { c,\min } \leq @xmath368 is the number of columns in @xmath587 . we can thus obtain upper and lower bounds for the local dimensions at points in the loop class by calculating these pseudo - norms for all admissible products of primitive transition matrices from the loop class up to some fixed length and using formula ( [ f2 ] ) . we remark that pseudo - norms based on a subset of the indices of the row vectors would serve , as well . [ ex : isolated points ] a measure whose local dimension is two isolated points : we continue with the example described above , but now take the normalized uniform weights @xmath588 . let @xmath3 denote the associated self - similar measure . the essential class , as mentioned above , is \{4 , 5 } and its primitive transition matrices are : @xmath589 & t(4,5 ) & = \left [ \begin{array}{cc } 1 & 0 \\ \noalign{\medskip}1 & 1 \\ \noalign{\medskip}0 & 1% \end{array}% \right ] \\ t(4,4 ) & = \left [ \begin{array}{ccc } 1 & 1 & 0 \\ \noalign{\medskip}1 & 1 & 1 \\ \noalign{\medskip}0 & 0 & 1% \end{array}% \right ] & t(5,5 ) & = \left [ \begin{array}{cc } 1 & 1 \\ \noalign{\medskip}1 & 1% \end{array}% \right ] \\ t(5,4 ) & = \left [ \begin{array}{ccc } 1 & 1 & 1 \\ \noalign{\medskip}1 & 1 & 1% \end{array}% \right ] & t(5,5 ) & = \left [ \begin{array}{cc } 1 & 1 \\ \noalign{\medskip}1 & 1% \end{array}% \right ] \\ & & & \end{aligned}\ ] ] as the total column sub and sup - norms for all these matrices is @xmath388 , it follows that all @xmath351-fold products of primitive transition matrices in the essential class will have their ( usual ) norm in the interval @xmath590 $ ] . it follows that the local dimension at any point in the essential class is @xmath591 . as the local dimension at @xmath275 and @xmath46 is equal to @xmath592 , the set of local dimensions of @xmath3 consists of the two distinct points , @xmath593\}=\{1\}\cup \left\ { 1+\log 2/\log 3\right\ } .\ ] ] in example [ ex : shmerkin ] we considered the ifs @xmath594for integers @xmath595 and the self - similar cantor - like measures @xmath3 associated with this ifs and probabilities @xmath596 . this class of measures includes , for example , the @xmath0-fold convolution of the uniform cantor measure with contraction factor @xmath6 rescaled to @xmath73 $ ] . the self - similar set is the full interval @xmath73 $ ] when @xmath75 . the results in @xcite and @xcite , which extend @xcite , together imply that if @xmath597 and @xmath598 , @xmath599 , then the set of local dimensions of @xmath3 consists of a closed interval and one ( or two ) isolated points , the local dimensions at @xmath275 and @xmath46 . these facts can be recovered by our methods as well , when , in addition , @xmath600 . let @xmath3 be the self - similar measure associated with the ifs ( [ cantor ] ) with @xmath601 and probabilities satisfying @xmath602 . then the set of essential points is the open interval @xmath366 and @xmath14 is an isolated point in the set of all local dimensions . indeed , for @xmath603 , @xmath604 we note , first , that the iterates of @xmath275 at level @xmath34 ( meaning the real values @xmath605 for @xmath101 ) occur every @xmath606 , beginning at @xmath275 and ending at @xmath607 . similarly , the iterates of @xmath46 are spaced the same distance apart , but start at @xmath608 and end at @xmath46 . in the subinterval @xmath609 $ ] they alternate , except in the special case that @xmath610 mod@xmath611 , when they coincide . we will assume @xmath612 mod@xmath611 and leave the easier case for the reader . adjacent @xmath34th level net intervals contained in @xmath609 $ ] are of the form @xmath613 $ ] , @xmath614 $ ] where if @xmath615 is an iterate of @xmath275 , then @xmath616for suitable integers @xmath617 . there is a similar formula when @xmath615 is an iterate of @xmath46 . if @xmath618 for integer @xmath619 and @xmath620 , then it is easily seen that @xmath621 . thus @xmath34th level net intervals have either ( non - normalized ) length @xmath622and , in either case , have length between @xmath623 and @xmath624 . let @xmath625 and assume @xmath34 is chosen so large that @xmath626 . let @xmath627 $ ] denote the @xmath34th level net interval containing @xmath158 . the choice of @xmath34 ensures that all the numbers of the form @xmath628 for integer @xmath183 with @xmath629 are iterates of @xmath275 at level @xmath34 and hence comprise the neighbour set of @xmath630 $ ] . thus the neighbour set depends only upon whether @xmath279 is an iterate of @xmath275 or an iterate of @xmath46 . it follows that there are only two reduced characteristic vectors associated with these net intervals . furthermore , the length of these net intervals is sufficiently large to ensure that that the interior of each such interval contains at least one @xmath631-iterate of @xmath275 and one @xmath151-iterate of @xmath46 . consequently , each such @xmath34th level net interval contains both styles of net intervals of level @xmath151 ( and no other children ) . it follows that their characteristic vectors belong to the essential class and that proves the set of essential points is @xmath366 . to prove the upper bound given in the statement of the proposition for @xmath632 with @xmath633 we will show that if @xmath159=(\gamma _ { 0},\gamma _ { 1},\dots)$ ] and @xmath34 is sufficiently large , then @xmath634 where @xmath635 is the total column sub - norm introduced in the preceding section . we can assume @xmath636 and take @xmath637 . suppose @xmath638 , @xmath639 $ ] , @xmath640 $ ] , and the neighbour sets are @xmath641 and @xmath642 respectively , where @xmath643 . temporarily fix @xmath351 . by definition , @xmath644 when there is some @xmath101 and @xmath249 such that @xmath645 and @xmath646 and @xmath647 @xmath275 otherwise . of course , there is some choice of @xmath28 with a valid choice of @xmath648 . if @xmath649 we are done . so assume this @xmath650 . that means @xmath651 . the bounds on @xmath158 ensure that for some @xmath652 , @xmath653furthermore , @xmath654 this implies there is some @xmath183 such that @xmath655 ; indeed , @xmath656 . another routine calculation shows @xmath657 , thus @xmath658 . if , instead , @xmath659 , similar arguments prove @xmath660 . this completes the proof . we are not claiming that these bounds on @xmath661 for @xmath603 are sharp , merely illustrating that we can recover the property that @xmath662 is an isolated point with our approach . in ( * ? ? ? * thm . 6.1 ) , the minimum and maximum local dimensions , other than at @xmath663 , were investigated for the case of the @xmath0-fold convolution of the uniform cantor measure for small @xmath0 . we show that there exists @xmath0 outside of these ranges where these formula do not hold . [ ex : m - fold convolution ] consider the @xmath0-fold convolution of the uniform cantor measure with contraction factor @xmath1 , for integer @xmath664 . it is shown in @xcite that if @xmath665 , then @xmath666where @xmath667it was also shown there that for all even @xmath0 , @xmath668 . using the computer , we have checked these formulas for @xmath669 and @xmath670 . we have found that the right hand side of ( [ eq : min ] ) is not @xmath671 and is not the minimal local dimension for @xmath672 moreover , for @xmath673 the minimal local dimension does not occur at the point @xmath674 . in fact , the predicted value of the minimum local dimension is greater than the maximum local dimension other than for @xmath675 . see tables [ tab : min ] and [ tab : max ] . we note that in tables tab : min and [ tab : max ] , when the formula is known to hold for theoretical reasons , we put the precise value , otherwise we put a range , coming from the techniques of section [ sec : comp ] . in @xcite , there was also a formula given for the maximum local dimension other than at @xmath675 . let @xmath676 , @xmath677 and @xmath678 . put @xmath679it was shown that for @xmath680 , @xmath681 and @xmath682here @xmath26 should be understood as @xmath275 if @xmath683 . using our methods , we can show that for @xmath669 , @xmath684 , these statements continue to be true , that is , @xmath685 @xmath686 and @xmath687 is the value specified by the right hand side of formula ( [ eq : max ] ) . but for @xmath669 and @xmath688 , the predicted maximum local dimension from the right hand side of is too big . we have not been able to determine if @xmath689 is the point where the maximum dimension occurs . we refer the reader to the supplementary information @xcite where this example is worked out in full detail . .minimal local dimensions [ cols= " < , < , < , < " , ] . in this subsection we study the rescaled measure @xmath691 where @xmath7 is the bernoulli convolution with @xmath16 the golden ratio . this is the self - similar measure associated with the ifs of contractions @xmath692 , @xmath693 and @xmath694 , with corresponding ( regular ) probabilities @xmath695 . it is of finite type and has support @xmath73 $ ] . the reduced transition diagram has 40 reduced characteristic vectors . the essential class can be naturally identified with those labelled by @xmath696 . two cycles in the essential class are @xmath697 and @xmath698 . the spectral radius of @xmath699 is approximately @xmath700 , while the spectral radius of @xmath701 is approximately @xmath702 . this shows that@xmath703\subseteq \{\dim \nu _ { \rho } ( x):x\text{essential}\}.\ ] ] we have also been able to find upper and lower bounds on the local dimensions from the essential class using the method described in subsection [ algdim ] . we obtain an upper bound by taking the column sub - norm with the subset @xmath704 @xmath705 and taking admissible products of up to 20 primitive transition matrices . we obtain a lower bound by using the total column sup - norm with products of up to 10 primitive transition matrices . these calculations give@xmath706.\ ] ] there are four non - essential maximal loops , each of which is a singleton . the maximal loop classes @xmath362 and @xmath707 correspond to the two endpoints of the support , @xmath663 . the transition matrix in both cases is the @xmath708 identity matrix and the points have local dimension @xmath709 the other two maximal loop classes are @xmath710 and @xmath711 . the characteristic vector of 25 is @xmath712 . its transition matrix has the same spectral radius as @xmath713 , hence the local dimension at any point in the loop class @xmath710 coincides with the local dimension at some essential point . similar statements hold for the loop class @xmath711 . thus the set of local dimensions of @xmath714 consists of an interval and an isolated point , @xmath715 . we recall that a point @xmath158 can have a most two symbolic representations , and that this will occur only if @xmath716 . let @xmath717 be the point with symbolic representation @xmath718this point also has symbolic representation @xmath719we see that both representations of these points are external to the loop class , and hence we have a countable number of non - essential points . not all points with a symbolic representation in a loop class external to the essential set need be non - essential . to see this we observe that @xmath720 are two symbolic representations for the same point , one of which is in the essential class , and one of which is not . hence this point is an essential point . the transition matrices of the cycles @xmath721 and @xmath392 give the extreme values of spectral radii over all transition matrices of essential cycles of length up to 10 . it would be interesting to know if these give the endpoints of the interval portion of the set of local dimensions . we thank m. ng for helping with some of the calculations on the cantor - like measures . 99 c. bruggeman and k. e. hare , _ multi - fractal analysis of convolution powers of measures _ , real anal . exch . * 38*(2013 ) , 391 - 408 . c. bruggeman , k. e. hare and c. mak , _ multi - fractal spectrum of self - similar measures with overlap _ , nonlinearity * 27*(2014 ) , 227 - 256 . r. cawley and r.d . mauldin , _ multi - fractal decompositions of moran fractals _ , adv . math . * 92*(1992 ) , 196 - 236 . deng , k - s . lau and s - m . ngai , _ separation conditions for iterated function systems with overlaps _ , contemporary math . * 600*(2013 ) , 1 - 20 . k. falconer , _ techniques in fractal geometry _ , john wiley and sons , new york , 1997 . a - h . fan , k - s . lau and s - m . ngai , _ iterated function systems with overlaps _ , asian j. math . * 4*(2000 ) , 243 - 250 . feng , _ smoothness of the _ @xmath722__-spectrum of self - similar measures with overlaps _ _ , j. london math . 68*(2003 ) , 102 - 118 . feng , _ the limited rademacher functions and bernoulli convolutions asociated with pisot numbers _ , adv . in math . * 195*(2005 ) , 24 - 101 . feng , _ lyapunov exponents for products of matrices and multifractal analysis . part ii : general matrices _ , isr . j. math . * 170*(2009 ) , 355 - 394 . feng and k - s . lau , _ multi - fractal formalism for self - similar measures with weak separation condition _ , j. math . pures appl . * 92*(2009 ) , 407 - 428 . feng , k - s . lau and x - y . wang , _ some exceptional phenomena in multi - fractal formalism : part ii _ , asian j. math . * 9*(2005 ) , 473 - 488 . feng and e. olivier , _ multifractal analysis of the weak gibbs measures and phase transition - application to some bernoulli convolutions _ , ergodic theory dynamical systems * 23*(2003 ) , 1751 - 1784 . fong , k. e. hare and d. johnstone , _ multi - fractal analysis for convolutions of overlapping cantor measures _ , asian j. math . * 15*(2011 ) , 53 - 70 . k. e. hare , k. g. hare , k. r. matthews , _ local dimensions of measures of finite type supplementary information _ , preprint 2015 , arxiv . hu , _ the local dimensions of bernoulli convolution associated with the golden number _ , trans . 349*(1997 ) , 2917 - 2940 . hu and k - s . lau , _ multi - fractal structure of convolution of the cantor measure _ , adv . app . math . * 27*(2001 ) , 1 - 16 . s. lalley , @xmath311__-expansions with deleted digits for pisot numbers , _ _ trans . soc . * 349*(1997 ) , 4355 - 4365 . k. s. lau and s. m. ngai , _ multifractal measures and a weak separation condition _ , adv . math . * 141*(1999 ) , 45 - 96 . k. s. lau and s. m. ngai , @xmath722__-spectrum of the bernoulli convolution associated with the golden ratio _ _ , studia math . * 131*(1998 ) , 225 - 251 . k - s . lau and x - y . wang , _ some exceptional phenomena in multi - fractal formalism : part i _ , asian j. math * 9*(2005 ) , 275 - 294 . s. m. ngai and y. wang , _ hausdorff dimension of self - similar sets with overlaps _ , j. london math * 63*(2001 ) , 655 - 672 . n. t. nguyen , _ iterated function systems of finite type nd the weak separation property _ , proc . soc . * 130*(2001 ) , 483 - 487 . e. olivier , n. sidorov and a. thomas , _ on the gibbs properties of bernoulli convolutions related to _ @xmath311__-enumeration in multinacci bases _ _ , monatsh . math . * 145*(2005 ) , 145 - 174 . y. peres , w. schlag and b. solomyak , _ sixty years of bernoulli convolutions _ , fractal geometry and stochasics , ii , progress in probability * 46 * , birkhuser , basel , 2000 , 39 - 65 . a. porzio , _ on the regularity of the multifractal spectrum of bernoulli convolutions _ , j. stats . phys . * 91*(1998 ) , 17 - 29 . p. shmerkin , _ a modified multi - fractal formalism for a class of self - similar measures with overlap _ , asian j. math * 9*(2005 ) , 323 - 348 . b. solomyak , _ notes on bernoulli convolutions _ , fractal geometry and applications : a jubilee of benot mandelbrot , proc . sympos . pure math . , * 72 * , amer . soc . , providence , ri , 2004 , 207230 . b. testud , _ phase transitions for the multi - fractal analysis of self - similar measures _ , nonlinearity * 19*(2006 ) , 12011217 . dept . of pure mathematics , university of waterloo , waterloo , ont . , n2l 3g1 , canada . ` kehare@uwaterloo.ca ` + dept . of pure mathematics , university of waterloo , waterloo , ont . , n2l 3g1 , canada . ` kghare@uwaterloo.ca ` + dept . of pure mathematics , university of waterloo , waterloo , ont . , n2l 3g1 , canada . ` kevinmatthews12@hotmail.com ` * local dimensions of measures of finite type supplementary information * kathryn e. hare , kevin g. hare , and kevin r. matthews this is the supplementary information for the manuscript _ local dimensions of measures of finite type _ by kathryn e. hare , kevin g. hare and kevin r. matthews . this contains the details of the examples mentioned in this manuscript . by `` the probabilities are uniform '' we mean that @xmath723 . the spectral range of a loop class is @xmath724 recalling that the dimension of a periodic point with cycle @xmath725 is @xmath726 allows an easy translation from the spectral range of a loop and the set of local dimensions of a loop class . the outline of the examples are as follows . * examples 2.6 , 3.11 , 4.7 , 4.10 . * * section [ sec:1 ] is for the minimal polynomial @xmath727 with @xmath728 $ ] ( page ) . * examples from section 7 . * * section [ sec:2 ] is for the minimal polynomial @xmath729 with @xmath730 ( page ) . * * section [ sec:3 ] is for the minimal polynomial @xmath729 with @xmath731 ( page ) . * * section [ sec:4 ] is for the minimal polynomial @xmath729 with @xmath732 ( page ) . * * section [ sec:5 ] is for the minimal polynomial @xmath729 with @xmath733 ( page ) . * * section [ sec:6 ] is for the minimal polynomial @xmath729 with @xmath734 ( page ) . * * section [ sec:7 ] is for the minimal polynomial @xmath729 with @xmath735 ( page ) . * * section [ sec:8 ] is for the minimal polynomial @xmath729 with @xmath736 ( page ) . * * section [ sec:9 ] is for the minimal polynomial @xmath729 with @xmath737 ( page ) . * examples from section 8 , including table 3 and examples 8.1 and 8.2 * * section [ sec:10 ] is for the minimal polynomial @xmath738 with @xmath728 $ ] ( page ) . * * section [ sec:11 ] is for the minimal polynomial @xmath739 with @xmath728 $ ] ( page ) . * * section [ sec:12 ] is for the minimal polynomial @xmath740 with @xmath728 $ ] ( page ) . * * section [ sec:13 ] is for the minimal polynomial @xmath741 with @xmath728 $ ] ( page ) . * * section [ sec:14 ] is for the minimal polynomial @xmath742 with @xmath728 $ ] ( page ) . * * section [ sec:15 ] is for the minimal polynomial @xmath743 with @xmath728 $ ] ( page ) . * * section [ sec:16 ] is for the minimal polynomial @xmath744 with @xmath728 $ ] ( page ) . * example in section 8.2 . * * section [ sec:17 ] is for the minimal polynomial @xmath727 with @xmath745 $ ] ( page ) . * example 6.1 * * section [ sec:18 ] is for the minimal polynomial @xmath729 with @xmath730 ( page ) . * * section [ sec:19 ] is for the minimal polynomial @xmath729 with @xmath731 ( page ) . * * section [ sec:20 ] is for the minimal polynomial @xmath729 with @xmath732 ( page ) . consider @xmath746 , the root of @xmath727 and the maps @xmath747 with @xmath748 , and @xmath749 . the probabilities are uniform . the reduced transition diagram has 6 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath750 * reduced characteristic vector 3 : @xmath751 * reduced characteristic vector 4 : @xmath752 * reduced characteristic vector 5 : @xmath753 * reduced characteristic vector 6 : @xmath754 see figure [ fig : pic1 ] for the transition diagram . with @xmath728 $ ] , full set + essential class ] this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(3,5 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(5,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(5,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] & t(6,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 3 , 5 , 6 ] . the essential class is of positive type . an example is the path [ 5 , 6 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath756 $ ] . the minimum comes from the loop @xmath757 $ ] . the maximum comes from the loop @xmath758 $ ] . these points will include points of local dimension [ .9404200909 , 1.440420090 ] . the spectral range is contained in the range @xmath759 $ ] . the minimum comes from the total row sub - norm of length 10 . the maximum comes from the total row sup - norm of length 10 . these points will have local dimension contained in [ .8642520535 , 1.440420090 ] . there are 2 additional maximal loops . maximal loop class : [ 4 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 1.440420090 . maximal loop class : [ 2 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 1.440420090 . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath761 , @xmath762 , and @xmath763 . the probabilities are given by @xmath764 , @xmath765 , @xmath766 , and @xmath767 . the reduced transition diagram has 5 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath768 * reduced characteristic vector 3 : @xmath769 * reduced characteristic vector 4 : @xmath770 * reduced characteristic vector 5 : @xmath771 see figure [ fig : pic2 ] for the transition diagram . with @xmath760 $ ] , full set + essential class ] this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 3&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 3&3\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 1&3\end { array } \right ] \\ t(1,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 3&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 3&3\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&3\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 3&1\\ 0&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 3&3\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 3&3\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{c } 3\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&3\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 3 , 4 ] . the essential class is of positive type . an example is the path [ 3 , 4 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath772 $ ] . the minimum comes from the loop @xmath773 $ ] . the maximum comes from the loop @xmath774 $ ] . these points will include points of local dimension [ .892789260 , 1.133544891 ] . the spectral range is contained in the range @xmath775 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .892789260 , 1.150482354 ] . there are 2 additional maximal loops . maximal loop class : [ 5 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 1.892789260 . maximal loop class : [ 2 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 1.892789260 . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath761 , @xmath762 , and @xmath763 . the probabilities are uniform . the reduced transition diagram has 5 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath768 * reduced characteristic vector 3 : @xmath769 * reduced characteristic vector 4 : @xmath770 * reduced characteristic vector 5 : @xmath771 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 3 , 4 ] . the essential class is of positive type . an example is the path [ 3 , 4 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1123 $ ] . the minimum comes from the loop @xmath774 $ ] . the maximum comes from the loop @xmath773 $ ] . these points will include points of local dimension [ .8238416275 , 1.261859507 ] . the spectral range is contained in the range @xmath1124 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .7949160034 , 1.261859507 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath777 , @xmath778 , @xmath779 , and @xmath780 . the probabilities are given by @xmath781 , @xmath782 , @xmath766 , @xmath783 , and @xmath784 . the reduced transition diagram has 4 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath785 * reduced characteristic vector 3 : @xmath786 * reduced characteristic vector 4 : @xmath787 see figure [ fig : pic3 ] for the transition diagram . with @xmath776 $ ] , full set + essential class ] this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 4&1\end { array } \right ] \\ t(1,3 ) & = \left [ \begin { array}{cc } 6&4\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 4&6\end { array } \right ] \\ t(1,3 ) & = \left [ \begin { array}{cc } 1&4\end { array } \right ] & t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,3 ) & = \left [ \begin { array}{cc } 4&1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 6&4\end { array } \right ] & t(3,3 ) & = \left [ \begin { array}{cc } 1&0\\ 4&6\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 4&1\\ 1&4\end { array } \right ] & t(3,3 ) & = \left [ \begin { array}{cc } 6&4\\ 0&1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 4&6\end { array } \right ] & t(4,3 ) & = \left [ \begin { array}{cc } 1&4\end { array } \right ] \\ t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 3 ] . the essential class is of positive type . an example is the path [ 3 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath788 $ ] . the minimum comes from the loop @xmath774 $ ] . the maximum comes from the loop @xmath774 $ ] . these points will include points of local dimension [ .892789260 , 1.058745493 ] . the spectral range is contained in the range @xmath789 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .892789260 , 1.058745493 ] . there are 2 additional maximal loops . maximal loop class : [ 4 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 2.523719013 . maximal loop class : [ 2 ] . the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1 . these points have local dimension 2.523719013 . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath777 , @xmath778 , @xmath779 , and @xmath780 . the probabilities are uniform . the reduced transition diagram has 4 reduced characteristic vectors . the reduced characteristic vectors are : this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(3,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] \\ t(3,3 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] & t(3,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 3 ] . the essential class is of positive type . an example is the path [ 3 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1125 $ ] . the minimum comes from the loop @xmath774 $ ] . the maximum comes from the loop @xmath774 $ ] . these points will include points of local dimension [ .8340437666 , 1.464973520 ] . the spectral range is contained in the range @xmath1126 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .8340437666 , 1.464973520 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath791 , @xmath792 , @xmath793 , @xmath794 , and @xmath795 . the probabilities are given by @xmath796 , @xmath797 , @xmath798 , @xmath799 , @xmath800 , and @xmath801 . the reduced transition diagram has 7 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath576 * reduced characteristic vector 3 : @xmath577 * reduced characteristic vector 4 : @xmath578 * reduced characteristic vector 5 : @xmath579 * reduced characteristic vector 6 : @xmath580 * reduced characteristic vector 7 : @xmath802 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 5&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 10&5&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 10&5\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 10&10&5\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 10&10\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 5&10&10\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 5&10\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&5&10\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{cc } 1&5\end { array } \right ] \\ t(1,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 5&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 10&5&1\end { array } \right ] \\ t(2,5 ) & = \left [ \begin { array}{cc } 10&5\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 10&10&5\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 1&0\\ 10&10\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 5&1&0\\ 5&10&10\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 5&1\\ 5&10\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 10&5&1\\ 1&5&10\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 10&5\\ 1&5\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 10&10&5\\ 0&1&5\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cc } 1&0\\ 10&10\\ 0&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 5&1&0\\ 5&10&10\\ 0&0&1\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cc } 5&1\\ 5&10\end { array } \right ] & t(5,4 ) & = \left [ \begin { array}{ccc } 10&5&1\\ 1&5&10\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cc } 10&5\\ 1&5\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 5&1\\ 5&10\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 10&5&1\\ 1&5&10\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 10&5\\ 1&5\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 10&10&5\\ 0&1&5\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 10&10\\ 0&1\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 5&10&10\\ 0&0&1\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cc } 5&10\end { array } \right ] \\ t(7,4 ) & = \left [ \begin { array}{ccc } 1&5&10\end { array } \right ] & t(7,6 ) & = \left [ \begin { array}{cc } 1&5\end { array } \right ] \\ t(7,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 4 , 5 ] . the essential class is of positive type . an example is the path [ 5 , 4 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath803 $ ] . the minimum comes from the loop @xmath804 $ ] . the maximum comes from the loop @xmath805 $ ] . these points will include points of local dimension [ .972638047 , 1.027566600 ] . the spectral range is contained in the range @xmath806 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .972381959 , 1.031992942 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath791 , @xmath792 , @xmath793 , @xmath794 , and @xmath795 . the probabilities are uniform . the reduced transition diagram has 7 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath576 * reduced characteristic vector 3 : @xmath577 * reduced characteristic vector 4 : @xmath578 * reduced characteristic vector 5 : @xmath579 * reduced characteristic vector 6 : @xmath580 * reduced characteristic vector 7 : @xmath802 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] \\ t(2,5 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 1&1&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 1&1&0\\ 1&1&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 1&1&1\\ 1&1&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 1&1&1\\ 0&1&1\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\\ 0&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 1&1&0\\ 1&1&1\\ 0&0&1\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] & t(5,4 ) & = \left [ \begin { array}{ccc } 1&1&1\\ 1&1&1\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 1&1&1\\ 1&1&1\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 1&1\\ 1&1\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 1&1&1\\ 0&1&1\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] \\ t(6,4 ) & = \left [ \begin { array}{ccc } 1&1&1\\ 0&0&1\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(7,4 ) & = \left [ \begin { array}{ccc } 1&1&1\end { array } \right ] & t(7,6 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(7,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 4 , 5 ] . the essential class is of positive type . an example is the path [ 5 , 4 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1127 $ ] . the minimum comes from the loop @xmath826 $ ] . the maximum comes from the loop @xmath826 $ ] . these points will include points of local dimension [ .9999999996 , .9999999996 ] . the spectral range is contained in the range @xmath1128 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .9999999996 , .9999999996 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath808 , @xmath809 , @xmath810 , @xmath811 , @xmath812 , and @xmath813 . the probabilities are given by @xmath814 , @xmath815 , @xmath816 , @xmath799 , @xmath817 , @xmath818 , and @xmath819 . the reduced transition diagram has 6 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath820 * reduced characteristic vector 3 : @xmath821 * reduced characteristic vector 4 : @xmath822 * reduced characteristic vector 5 : @xmath823 * reduced characteristic vector 6 : @xmath824 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 6&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 15&6&1\end { array } \right ] & t(1,4 ) & = \left [ \begin { array}{ccc } 20&15&6\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 15&20&15\end { array } \right ] & t(1,4 ) & = \left [ \begin { array}{ccc } 6&15&20\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&6&15\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 1&6\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 6&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 15&6&1\end { array } \right ] \\ t(3,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 20&15&6\end { array } \right ] & t(3,4 ) & = \left [ \begin { array}{ccc } 6&1&0\\ 15&20&15\end { array } \right ] \\ t(3,4 ) & = \left [ \begin { array}{ccc } 15&6&1\\ 6&15&20\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 20&15&6\\ 1&6&15\end { array } \right ] \\ t(4,4 ) & = \left [ \begin { array}{ccc } 6&1&0\\ 15&20&15\\ 0&1&6\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{ccc } 15&6&1\\ 6&15&20\\ 0&0&1\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{ccc } 20&15&6\\ 1&6&15\end { array } \right ] & t(5,4 ) & = \left [ \begin { array}{ccc } 15&20&15\\ 0&1&6\end { array } \right ] \\ t(5,4 ) & = \left [ \begin { array}{ccc } 6&15&20\\ 0&0&1\end { array } \right ] & t(6,4 ) & = \left [ \begin { array}{ccc } 1&6&15\end { array } \right ] \\ t(6,5 ) & = \left [ \begin { array}{cc } 1&6\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 4 ] . the essential class is of positive type . an example is the path [ 4 , 4 , 4 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath825 $ ] . the minimum comes from the loop @xmath826 $ ] . the maximum comes from the loop @xmath826 $ ] . these points will include points of local dimension [ .976628124 , 1.014334771 ] . the spectral range is contained in the range @xmath827 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .975486976 , 1.014334771 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath829 , @xmath830 , @xmath831 , @xmath832 , @xmath833 , @xmath834 , and @xmath835 . the probabilities are given by @xmath836 , @xmath837 , @xmath838 , @xmath839 , @xmath840 , @xmath841 , @xmath842 , and @xmath843 . the reduced transition diagram has 9 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath844 * reduced characteristic vector 3 : @xmath845 * reduced characteristic vector 4 : @xmath846 * reduced characteristic vector 5 : @xmath847 * reduced characteristic vector 6 : @xmath848 * reduced characteristic vector 7 : @xmath849 * reduced characteristic vector 8 : @xmath850 * reduced characteristic vector 9 : @xmath851 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 7&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 21&7&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 35&21&7&1\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccc } 35&21&7\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 35&35&21&7\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccc } 35&35&21\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 21&35&35&21\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccc } 21&35&35\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 7&21&35&35\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccc } 7&21&35\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 1&7&21&35\end { array } \right ] \\ t(1,7 ) & = \left [ \begin { array}{ccc } 1&7&21\end { array } \right ] & t(1,8 ) & = \left [ \begin { array}{cc } 1&7\end { array } \right ] \\ t(1,9 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 7&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 21&7&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 35&21&7&1\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 35&21&7\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 7&1&0&0\\ 35&35&21&7\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccc } 7&1&0\\ 35&35&21\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 21&7&1&0\\ 21&35&35&21\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccc } 21&7&1\\ 21&35&35\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 35&21&7&1\\ 7&21&35&35\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 35&21&7\\ 7&21&35\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cccc } 7&1&0&0\\ 35&35&21&7\\ 1&7&21&35\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccc } 7&1&0\\ 35&35&21\\ 1&7&21\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cccc } 21&7&1&0\\ 21&35&35&21\\ 0&1&7&21\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccc } 21&7&1\\ 21&35&35\\ 0&1&7\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 35&21&7&1\\ 7&21&35&35\\ 0&0&1&7\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 35&21&7\\ 7&21&35\\ 0&0&1\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cccc } 7&1&0&0\\ 35&35&21&7\\ 1&7&21&35\\ 0&0&0&1\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccc } 7&1&0\\ 35&35&21\\ 1&7&21\end { array } \right ] \\ t(6,5 ) & = \left [ \begin { array}{cccc } 21&7&1&0\\ 21&35&35&21\\ 0&1&7&21\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccc } 21&7&1\\ 21&35&35\\ 0&1&7\end { array } \right ] \\ t(7,6 ) & = \left [ \begin { array}{ccc } 7&1&0\\ 35&35&21\\ 1&7&21\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cccc } 21&7&1&0\\ 21&35&35&21\\ 0&1&7&21\end { array } \right ] \\ t(7,6 ) & = \left [ \begin { array}{ccc } 21&7&1\\ 21&35&35\\ 0&1&7\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cccc } 35&21&7&1\\ 7&21&35&35\\ 0&0&1&7\end { array } \right ] \\ t(7,6 ) & = \left [ \begin { array}{ccc } 35&21&7\\ 7&21&35\\ 0&0&1\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cccc } 35&35&21&7\\ 1&7&21&35\\ 0&0&0&1\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccc } 35&35&21\\ 1&7&21\end { array } \right ] & t(8,5 ) & = \left [ \begin { array}{cccc } 21&35&35&21\\ 0&1&7&21\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccc } 21&35&35\\ 0&1&7\end { array } \right ] & t(8,5 ) & = \left [ \begin { array}{cccc } 7&21&35&35\\ 0&0&1&7\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccc } 7&21&35\\ 0&0&1\end { array } \right ] & t(8,5 ) & = \left [ \begin { array}{cccc } 1&7&21&35\\ 0&0&0&1\end { array } \right ] \\ t(9,7 ) & = \left [ \begin { array}{ccc } 1&7&21\end { array } \right ] & t(9,8 ) & = \left [ \begin { array}{cc } 1&7\end { array } \right ] \\ t(9,9 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 5 , 6 ] . the essential class is of positive type . an example is the path [ 6 , 5 , 5 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath852 $ ] . the minimum comes from the loop @xmath804 $ ] . the maximum comes from the loop @xmath853 $ ] . these points will include points of local dimension [ .993847946 , 1.006057728 ] . the spectral range is contained in the range @xmath854 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .993576302 , 1.007354690 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath856 , @xmath857 , @xmath858 , @xmath859 , @xmath860 , @xmath861 , @xmath862 , and @xmath863 . the probabilities are given by @xmath864 , @xmath865 , @xmath866 , @xmath867 , @xmath840 , @xmath868 , @xmath869 , @xmath870 , and @xmath871 . the reduced transition diagram has 8 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath872 * reduced characteristic vector 3 : @xmath873 * reduced characteristic vector 4 : @xmath874 * reduced characteristic vector 5 : @xmath875 * reduced characteristic vector 6 : @xmath876 * reduced characteristic vector 7 : @xmath877 * reduced characteristic vector 8 : @xmath878 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 8&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 28&8&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 56&28&8&1\end { array } \right ] \\ t(1,5 ) & = \left [ \begin { array}{cccc } 70&56&28&8\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 56&70&56&28\end { array } \right ] \\ t(1,5 ) & = \left [ \begin { array}{cccc } 28&56&70&56\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 8&28&56&70\end { array } \right ] \\ t(1,5 ) & = \left [ \begin { array}{cccc } 1&8&28&56\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{ccc } 1&8&28\end { array } \right ] \\ t(1,7 ) & = \left [ \begin { array}{cc } 1&8\end { array } \right ] & t(1,8 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,3 ) & = \left [ \begin { array}{cc } 8&1\end { array } \right ] \\ t(2,4 ) & = \left [ \begin { array}{ccc } 28&8&1\end { array } \right ] & t(3,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 56&28&8&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 8&1&0&0\\ 70&56&28&8\end { array } \right ] & t(3,5 ) & = \left [ \begin { array}{cccc } 28&8&1&0\\ 56&70&56&28\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 56&28&8&1\\ 28&56&70&56\end { array } \right ] & t(4,5 ) & = \left [ \begin { array}{cccc } 8&1&0&0\\ 70&56&28&8\\ 8&28&56&70\end { array } \right ] \\ t(4,5 ) & = \left [ \begin { array}{cccc } 28&8&1&0\\ 56&70&56&28\\ 1&8&28&56\end { array } \right ] & t(5,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 56&28&8&1\\ 28&56&70&56\\ 0&1&8&28\end { array } \right ] \\ t(5,5 ) & = \left [ \begin { array}{cccc } 8&1&0&0\\ 70&56&28&8\\ 8&28&56&70\\ 0&0&1&8\end { array } \right ] & t(5,5 ) & = \left [ \begin { array}{cccc } 28&8&1&0\\ 56&70&56&28\\ 1&8&28&56\\ 0&0&0&1\end { array } \right ] \\ t(6,5 ) & = \left [ \begin { array}{cccc } 56&28&8&1\\ 28&56&70&56\\ 0&1&8&28\end { array } \right ] & t(6,5 ) & = \left [ \begin { array}{cccc } 70&56&28&8\\ 8&28&56&70\\ 0&0&1&8\end { array } \right ] \\ t(6,5 ) & = \left [ \begin { array}{cccc } 56&70&56&28\\ 1&8&28&56\\ 0&0&0&1\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cccc } 28&56&70&56\\ 0&1&8&28\end { array } \right ] \\ t(7,5 ) & = \left [ \begin { array}{cccc } 8&28&56&70\\ 0&0&1&8\end { array } \right ] & t(7,5 ) & = \left [ \begin { array}{cccc } 1&8&28&56\\ 0&0&0&1\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccc } 1&8&28\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{cc } 1&8\end { array } \right ] \\ t(8,8 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 5 ] . the essential class is of positive type . an example is the path [ 5 , 5 , 5 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath879 $ ] . the minimum comes from the loop @xmath880 $ ] . the maximum comes from the loop @xmath880 $ ] . these points will include points of local dimension [ .995246196 , 1.003425326 ] . the spectral range is contained in the range @xmath881 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .994695517 , 1.003455237 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath883 , @xmath884 , @xmath885 , @xmath886 , @xmath887 , @xmath888 , @xmath889 , @xmath890 , and @xmath891 . the probabilities are given by @xmath892 , @xmath893 , @xmath894 , @xmath895 , @xmath896 , @xmath897 , @xmath898 , @xmath899 , @xmath900 , and @xmath901 . the reduced transition diagram has 11 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath902 * reduced characteristic vector 3 : @xmath903 * reduced characteristic vector 4 : @xmath904 * reduced characteristic vector 5 : @xmath905 * reduced characteristic vector 6 : @xmath906 * reduced characteristic vector 7 : @xmath907 * reduced characteristic vector 8 : @xmath908 * reduced characteristic vector 9 : @xmath909 * reduced characteristic vector 10 : @xmath910 * reduced characteristic vector 11 : @xmath911 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 9&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 36&9&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 84&36&9&1\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 126&84&36&9&1\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 126&84&36&9\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 126&126&84&36&9\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 126&126&84&36\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 84&126&126&84&36\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 84&126&126&84\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 36&84&126&126&84\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 36&84&126&126\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 9&36&84&126&126\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 9&36&84&126\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 1&9&36&84&126\end { array } \right ] & t(1,8 ) & = \left [ \begin { array}{cccc } 1&9&36&84\end { array } \right ] \\ t(1,9 ) & = \left [ \begin { array}{ccc } 1&9&36\end { array } \right ] & t(1,10 ) & = \left [ \begin { array}{cc } 1&9\end { array } \right ] \\ t(1,11 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 9&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 36&9&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 84&36&9&1\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccccc } 9&1&0&0&0\\ 126&84&36&9&1\end { array } \right ] \\ t(3,7 ) & = \left [ \begin { array}{cccc } 9&1&0&0\\ 126&84&36&9\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccccc } 36&9&1&0&0\\ 126&126&84&36&9\end { array } \right ] \\ t(3,7 ) & = \left [ \begin { array}{cccc } 36&9&1&0\\ 126&126&84&36\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 84&36&9&1&0\\ 84&126&126&84&36\end { array } \right ] \\ t(4,7 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 84&36&9&1\\ 84&126&126&84\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccccc } 9&1&0&0&0\\ 126&84&36&9&1\\ 36&84&126&126&84\end { array } \right ] \\ t(4,7 ) & = \left [ \begin { array}{cccc } 9&1&0&0\\ 126&84&36&9\\ 36&84&126&126\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccccc } 36&9&1&0&0\\ 126&126&84&36&9\\ 9&36&84&126&126\end { array } \right ] \\ t(4,7 ) & = \left [ \begin { array}{cccc } 36&9&1&0\\ 126&126&84&36\\ 9&36&84&126\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 84&36&9&1&0\\ 84&126&126&84&36\\ 1&9&36&84&126\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 84&36&9&1\\ 84&126&126&84\\ 1&9&36&84\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{ccccc } 9&1&0&0&0\\ 126&84&36&9&1\\ 36&84&126&126&84\\ 0&1&9&36&84\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{cccc } 9&1&0&0\\ 126&84&36&9\\ 36&84&126&126\\ 0&1&9&36\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{ccccc } 36&9&1&0&0\\ 126&126&84&36&9\\ 9&36&84&126&126\\ 0&0&1&9&36\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{cccc } 36&9&1&0\\ 126&126&84&36\\ 9&36&84&126\\ 0&0&1&9\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 84&36&9&1&0\\ 84&126&126&84&36\\ 1&9&36&84&126\\ 0&0&0&1&9\end { array } \right ] \\ t(6,7 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 84&36&9&1\\ 84&126&126&84\\ 1&9&36&84\\ 0&0&0&1\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccccc } 9&1&0&0&0\\ 126&84&36&9&1\\ 36&84&126&126&84\\ 0&1&9&36&84\\ 0&0&0&0&1\end { array } \right ] \\ t(7,7 ) & = \left [ \begin { array}{cccc } 9&1&0&0\\ 126&84&36&9\\ 36&84&126&126\\ 0&1&9&36\end { array } \right ] & t(7,6 ) & = \left [ \begin { array}{ccccc } 36&9&1&0&0\\ 126&126&84&36&9\\ 9&36&84&126&126\\ 0&0&1&9&36\end { array } \right ] \\ t(7,7 ) & = \left [ \begin { array}{cccc } 36&9&1&0\\ 126&126&84&36\\ 9&36&84&126\\ 0&0&1&9\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{cccc } 9&1&0&0\\ 126&84&36&9\\ 36&84&126&126\\ 0&1&9&36\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccccc } 36&9&1&0&0\\ 126&126&84&36&9\\ 9&36&84&126&126\\ 0&0&1&9&36\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{cccc } 36&9&1&0\\ 126&126&84&36\\ 9&36&84&126\\ 0&0&1&9\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccccc } 84&36&9&1&0\\ 84&126&126&84&36\\ 1&9&36&84&126\\ 0&0&0&1&9\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{cccc } 84&36&9&1\\ 84&126&126&84\\ 1&9&36&84\\ 0&0&0&1\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccccc } 126&84&36&9&1\\ 36&84&126&126&84\\ 0&1&9&36&84\\ 0&0&0&0&1\end { array } \right ] & t(9,7 ) & = \left [ \begin { array}{cccc } 126&84&36&9\\ 36&84&126&126\\ 0&1&9&36\end { array } \right ] \\ t(9,6 ) & = \left [ \begin { array}{ccccc } 126&126&84&36&9\\ 9&36&84&126&126\\ 0&0&1&9&36\end { array } \right ] & t(9,7 ) & = \left [ \begin { array}{cccc } 126&126&84&36\\ 9&36&84&126\\ 0&0&1&9\end { array } \right ] \\ t(9,6 ) & = \left [ \begin { array}{ccccc } 84&126&126&84&36\\ 1&9&36&84&126\\ 0&0&0&1&9\end { array } \right ] & t(9,7 ) & = \left [ \begin { array}{cccc } 84&126&126&84\\ 1&9&36&84\\ 0&0&0&1\end { array } \right ] \\ t(9,6 ) & = \left [ \begin { array}{ccccc } 36&84&126&126&84\\ 0&1&9&36&84\\ 0&0&0&0&1\end { array } \right ] & t(10,7 ) & = \left [ \begin { array}{cccc } 36&84&126&126\\ 0&1&9&36\end { array } \right ] \\ t(10,6 ) & = \left [ \begin { array}{ccccc } 9&36&84&126&126\\ 0&0&1&9&36\end { array } \right ] & t(10,7 ) & = \left [ \begin { array}{cccc } 9&36&84&126\\ 0&0&1&9\end { array } \right ] \\ t(10,6 ) & = \left [ \begin { array}{ccccc } 1&9&36&84&126\\ 0&0&0&1&9\end { array } \right ] & t(10,8 ) & = \left [ \begin { array}{cccc } 1&9&36&84\\ 0&0&0&1\end { array } \right ] \\ t(11,9 ) & = \left [ \begin { array}{ccc } 1&9&36\end { array } \right ] & t(11,10 ) & = \left [ \begin { array}{cc } 1&9\end { array } \right ] \\ t(11,11 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 6 , 7 ] . the essential class is of positive type . an example is the path [ 7 , 6 , 6 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath912 $ ] . the minimum comes from the loop @xmath913 $ ] . the maximum comes from the loop @xmath853 $ ] . these points will include points of local dimension [ .998657167 , 1.001330544 ] . the spectral range is contained in the range @xmath914 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .998541136 , 1.001708776 ] . consider @xmath746 , the root of @xmath729 and the maps @xmath747 with @xmath748 , @xmath916 , @xmath917 , @xmath918 , @xmath919 , @xmath920 , @xmath921 , @xmath922 , @xmath923 , @xmath924 , and @xmath925 . the probabilities are given by @xmath926 , @xmath927 , @xmath928 , @xmath929 , @xmath930 , @xmath897 , @xmath931 , @xmath932 , @xmath933 , @xmath934 , and @xmath935 . the reduced transition diagram has 10 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath936 * reduced characteristic vector 3 : @xmath937 * reduced characteristic vector 4 : @xmath938 * reduced characteristic vector 5 : @xmath939 * reduced characteristic vector 6 : @xmath940 * reduced characteristic vector 7 : @xmath941 * reduced characteristic vector 8 : @xmath942 * reduced characteristic vector 9 : @xmath943 * reduced characteristic vector 10 : @xmath944 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 10&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 45&10&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cccc } 120&45&10&1\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 210&120&45&10&1\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{ccccc } 252&210&120&45&10\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 210&252&210&120&45\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{ccccc } 120&210&252&210&120\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 45&120&210&252&210\end { array } \right ] & t(1,6 ) & = \left [ \begin { array}{ccccc } 10&45&120&210&252\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{ccccc } 1&10&45&120&210\end { array } \right ] & t(1,7 ) & = \left [ \begin { array}{cccc } 1&10&45&120\end { array } \right ] \\ t(1,8 ) & = \left [ \begin { array}{ccc } 1&10&45\end { array } \right ] & t(1,9 ) & = \left [ \begin { array}{cc } 1&10\end { array } \right ] \\ t(1,10 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 10&1\end { array } \right ] & t(2,4 ) & = \left [ \begin { array}{ccc } 45&10&1\end { array } \right ] \\ t(3,5 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 120&45&10&1\end { array } \right ] & t(3,6 ) & = \left [ \begin { array}{ccccc } 10&1&0&0&0\\ 210&120&45&10&1\end { array } \right ] \\ t(3,6 ) & = \left [ \begin { array}{ccccc } 45&10&1&0&0\\ 252&210&120&45&10\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 120&45&10&1&0\\ 210&252&210&120&45\end { array } \right ] \\ t(4,6 ) & = \left [ \begin { array}{ccccc } 10&1&0&0&0\\ 210&120&45&10&1\\ 120&210&252&210&120\end { array } \right ] & t(4,6 ) & = \left [ \begin { array}{ccccc } 45&10&1&0&0\\ 252&210&120&45&10\\ 45&120&210&252&210\end { array } \right ] \\ t(5,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 120&45&10&1&0\\ 210&252&210&120&45\\ 10&45&120&210&252\end { array } \right ] & t(5,6 ) & = \left [ \begin { array}{ccccc } 10&1&0&0&0\\ 210&120&45&10&1\\ 120&210&252&210&120\\ 1&10&45&120&210\end { array } \right ] \\ t(5,6 ) & = \left [ \begin { array}{ccccc } 45&10&1&0&0\\ 252&210&120&45&10\\ 45&120&210&252&210\\ 0&1&10&45&120\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 120&45&10&1&0\\ 210&252&210&120&45\\ 10&45&120&210&252\\ 0&0&1&10&45\end { array } \right ] \\ t(6,6 ) & = \left [ \begin { array}{ccccc } 10&1&0&0&0\\ 210&120&45&10&1\\ 120&210&252&210&120\\ 1&10&45&120&210\\ 0&0&0&1&10\end { array } \right ] & t(6,6 ) & = \left [ \begin { array}{ccccc } 45&10&1&0&0\\ 252&210&120&45&10\\ 45&120&210&252&210\\ 0&1&10&45&120\\ 0&0&0&0&1\end { array } \right ] \\ t(7,6 ) & = \left [ \begin { array}{ccccc } 120&45&10&1&0\\ 210&252&210&120&45\\ 10&45&120&210&252\\ 0&0&1&10&45\end { array } \right ] & t(7,6 ) & = \left [ \begin { array}{ccccc } 210&120&45&10&1\\ 120&210&252&210&120\\ 1&10&45&120&210\\ 0&0&0&1&10\end { array } \right ] \\ t(7,6 ) & = \left [ \begin { array}{ccccc } 252&210&120&45&10\\ 45&120&210&252&210\\ 0&1&10&45&120\\ 0&0&0&0&1\end { array } \right ] & t(8,6 ) & = \left [ \begin { array}{ccccc } 210&252&210&120&45\\ 10&45&120&210&252\\ 0&0&1&10&45\end { array } \right ] \\ t(8,6 ) & = \left [ \begin { array}{ccccc } 120&210&252&210&120\\ 1&10&45&120&210\\ 0&0&0&1&10\end { array } \right ] & t(8,6 ) & = \left [ \begin { array}{ccccc } 45&120&210&252&210\\ 0&1&10&45&120\\ 0&0&0&0&1\end { array } \right ] \\ t(9,6 ) & = \left [ \begin { array}{ccccc } 10&45&120&210&252\\ 0&0&1&10&45\end { array } \right ] & t(9,6 ) & = \left [ \begin { array}{ccccc } 1&10&45&120&210\\ 0&0&0&1&10\end { array } \right ] \\ t(9,7 ) & = \left [ \begin { array}{cccc } 1&10&45&120\\ 0&0&0&1\end { array } \right ] & t(10,8 ) & = \left [ \begin { array}{ccc } 1&10&45\end { array } \right ] \\ t(10,9 ) & = \left [ \begin { array}{cc } 1&10\end { array } \right ] & t(10,10 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 6 ] . the essential class is of positive type . an example is the path [ 6 , 6 , 6 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath945 $ ] . the minimum comes from the loop @xmath946 $ ] . the maximum comes from the loop @xmath946 $ ] . these points will include points of local dimension [ .999021586 , 1.000793713 ] . the spectral range is contained in the range @xmath947 $ ] . the minimum comes from the total row sub - norm of length 5 . the maximum comes from the total row sup - norm of length 5 . these points will have local dimension contained in [ .998839826 , 1.000814954 ] . the number of reduced characteristic vectors in the essential class is 46 . the essential class is of positive type . an example is the path [ 129 , 92 , 91 , 114 , 112 , 127 , 140 , 150 , 75 , 75 , 75 , 75 , 75 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath948 $ ] . the minimum comes from the loop @xmath949 $ ] . the maximum comes from the loop @xmath950 $ ] . these points will include points of local dimension [ .9702219294 , 1.077705433 ] . the spectral range is contained in the range @xmath951 $ ] . the minimum comes from the column sub - norm on the subset @xmath952 of length 15 . the maximum comes from the total column sup - norm of length 10 . these points will have local dimension contained in [ .8483019061 , 1.532658865 ] . * reduced characteristic vector 40 : @xmath953 * reduced characteristic vector 41 : @xmath954 * reduced characteristic vector 45 : @xmath955 * reduced characteristic vector 46 : @xmath956 * reduced characteristic vector 47 : @xmath957 * reduced characteristic vector 51 : @xmath958 * reduced characteristic vector 52 : @xmath959 * reduced characteristic vector 53 : @xmath960 * reduced characteristic vector 54 : @xmath961 * reduced characteristic vector 61 : @xmath962 * reduced characteristic vector 62 : @xmath963 * reduced characteristic vector 63 : @xmath964 * reduced characteristic vector 70 : @xmath965 * reduced characteristic vector 71 : @xmath966 * reduced characteristic vector 72 : @xmath967 * reduced characteristic vector 82 : @xmath968 * reduced characteristic vector 84 : @xmath969 * reduced characteristic vector 86 : @xmath970 * reduced characteristic vector 96 : @xmath971 * reduced characteristic vector 106 : @xmath972 * reduced characteristic vector 122 : @xmath973 * reduced characteristic vector 135 : @xmath974 * reduced characteristic vector 146 : @xmath975 this has transition matrices : @xmath976 \\ t(41,54 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 0&0&0&1&0&0&0\\ 0&0&1&0&1&0&0\\ 0&0&0&1&0&1&0\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(45,61 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\end { array } \right ] \\ t(46,62 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 0&1&0&0&1&0\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(47,63 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(51,70 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 0&1&0&1&0&0&0\\ 0&0&1&0&1&0&0\\ 0&0&0&1&0&0&0\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(52,71 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&0&1&0&0&0\\ 0&1&0&1&0&0\\ 0&0&1&0&1&0\\ 0&0&0&1&0&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(53,45 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 0&0&1&0&0&1\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\end { array } \right ] \\ t(53,46 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(54,72 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 0&0&0&1&0&0&0\\ 0&1&0&0&1&0&0\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(61,82 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 0&1&0&0&0\\ 1&0&1&0&0\\ 0&1&0&0&1\\ 0&0&0&1&0\\ 0&0&0&0&1\end { array } \right ] \\ t(62,84 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&0&1&0&0&0\\ 0&1&0&1&0&0\\ 0&0&1&0&1&0\\ 0&0&0&1&0&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(63,86 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 0&1&0&0&0\\ 1&0&0&1&0\\ 0&0&1&0&1\\ 0&0&0&1&0\\ 0&0&0&0&1\end { array } \right ] \\ t(70,96 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 0&0&1&0&0&1&0\\ 0&0&0&1&0&0&0\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(71,46 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(71,47 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 1&0&0&1&0&0\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(72,63 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(82,51 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&0&1&0&0&1&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\end { array } \right ] \\ t(82,52 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\end { array } \right ] & t(84,106 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(86,40 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&0&1&0\end { array } \right ] & t(86,41 ) & = \left [ \begin { array}{ccccccc } 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&1&0&0&1&0&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(96,61 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(106,122 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(122,51 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&0&1&0&0&1&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\end { array } \right ] \\ t(122,135 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&0&1&0&0&0\\ 1&0&0&1&0&0\\ 0&0&1&0&0&1\\ 0&0&0&1&0&0\\ 0&0&0&0&1&0\end { array } \right ] \\ t(122,41 ) & = \left [ \begin { array}{ccccccc } 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&1&0&0&1&0&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(135,146 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&0&1&0&0&0\\ 0&1&0&1&0&0\\ 0&0&1&0&1&0\\ 0&0&0&1&0&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(146,46 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 1&0&0&1&0&0&0\\ 0&0&0&1&0&0&1\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ \end{aligned}\ ] ] the maximal loop class is of positive type . an example is the path [ 82 , 52 , 71 , 46 , 62 , 84 , 106 , 122 , 135 ] . the maximal loop class is not a simple loop . this spectral range will include the interval @xmath977 $ ] . the minimum comes from the loop @xmath978 $ ] . the maximum comes from the loop @xmath979 $ ] . these points will include points of local dimension [ .9702219294 , .9918974638 ] . the spectral range is contained in the range @xmath980 $ ] . the minimum comes from the total column sub - norm of length 15 . the maximum comes from the total column sup - norm of length 15 . these points will have local dimension contained in [ .8875527238 , 1.309253535 ] . this has transition matrices : @xmath983 & t(30,22 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&1&0\\ 0&0&1&0\\ 0&0&0&1\end { array } \right ] \\ \end{aligned}\ ] ] the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1.380277569 . these points have local dimension .9702219294 . the number of reduced characteristic vectors in the essential class is 1207 . the essential class is of positive type . an example is the path [ 683 , 631 , 688 , 624 , 683 , 631 , 688 , 624 , 683 , 631 , 688 , 624 , 683 , 631 , 688 , 624 , 683 , 631 , 688 , 624 , 683 , 631 , 689 , 625 , 684 , 632 , 690 , 509 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath988 $ ] . the minimum comes from the loop @xmath989 $ ] . the maximum comes from the loop @xmath990 $ ] . these points will include points of local dimension [ .997948624 , 1.008536243 ] . the spectral range is contained in the range @xmath991 $ ] . the minimum comes from the column sub - norm on the subset @xmath992 of length 15 . the maximum comes from the total column sup - norm of length 10 . these points will have local dimension contained in [ .747923631 , 1.971972204 ] . * reduced characteristic vector 388 : @xmath993 * reduced characteristic vector 454 : @xmath994 * reduced characteristic vector 460 : @xmath995 * reduced characteristic vector 524 : @xmath996 * reduced characteristic vector 531 : @xmath997 * reduced characteristic vector 590 : @xmath998 this has transition matrices : @xmath999 \\ t(454,524 ) & = \left [ \begin { array}{ccccccccccccccccccccc } 1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1\end { array } \right ] \\ t(460,531 ) & = \left [ \begin { array}{ccccccccccccccccccccc } 1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 1&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&1&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1\end { array } \right ] \\ t(524,590 ) & = \left [ \begin { array}{ccccccccccccccccccccc } 1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&1&0&0&0&0&1&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0&1\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1\end { array } \right ] \\ t(531,454 ) & = \left [ \begin { array}{ccccccccccccccccccccc } 0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1\end { array } \right ] \\ t(590,388 ) & = \left [ \begin { array}{ccccccccccccccccccccc } 1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0&0&1\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&0&1&0\end { array } \right ] \\ \end{aligned}\ ] ] the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1.510392888 . these points have local dimension .998498570 . this has transition matrices : @xmath1001 \\ \end{aligned}\ ] ] the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1.506135679 . these points have local dimension 1.008536246 . this has transition matrices : @xmath1003 \\ \end{aligned}\ ] ] the maximal loop class is a simple loop . it s spectral radius is an isolated points of 1.506135679 . these points have local dimension 1.008536246 . consider @xmath746 , the root of @xmath740 and the maps @xmath747 with @xmath748 , and @xmath749 . the probabilities are uniform . the reduced transition diagram has 30 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath1006 * reduced characteristic vector 3 : @xmath1007 * reduced characteristic vector 4 : @xmath1008 * reduced characteristic vector 5 : @xmath1009 * reduced characteristic vector 6 : @xmath751 * reduced characteristic vector 7 : @xmath1010 * reduced characteristic vector 8 : @xmath1011 * reduced characteristic vector 9 : @xmath1012 * reduced characteristic vector 10 : @xmath1013 * reduced characteristic vector 11 : @xmath1014 * reduced characteristic vector 12 : @xmath1015 * reduced characteristic vector 13 : @xmath1016 * reduced characteristic vector 14 : @xmath1017 * reduced characteristic vector 15 : @xmath1018 * reduced characteristic vector 16 : @xmath1019 * reduced characteristic vector 17 : @xmath1020 * reduced characteristic vector 18 : @xmath1021 * reduced characteristic vector 19 : @xmath1022 * reduced characteristic vector 20 : @xmath1023 * reduced characteristic vector 21 : @xmath1024 * reduced characteristic vector 22 : @xmath1025 * reduced characteristic vector 23 : @xmath1026 * reduced characteristic vector 24 : @xmath1027 * reduced characteristic vector 25 : @xmath1028 * reduced characteristic vector 26 : @xmath1029 * reduced characteristic vector 27 : @xmath754 * reduced characteristic vector 28 : @xmath1030 * reduced characteristic vector 29 : @xmath1031 * reduced characteristic vector 30 : @xmath1032 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(3,6 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(4,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(6,8 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(7,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(7,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(8,9 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(8,10 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\end { array } \right ] & t(8,11 ) & = \left [ \begin { array}{cc } 0&1\\ 1&0\end { array } \right ] \\ t(8,12 ) & = \left [ \begin { array}{ccc } 1&0&1\\ 0&1&0\end { array } \right ] & t(8,13 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(9,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] & t(9,5 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(10,14 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\\ 0&1\end { array } \right ] & t(10,15 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\\ 0&0&1\end { array } \right ] \\ t(11,16 ) & = \left [ \begin { array}{cc } 0&1\\ 1&0\end { array } \right ] & t(12,17 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 1&0&1\\ 0&1&0\end { array } \right ] \\ t(12,18 ) & = \left [ \begin { array}{cc } 1&0\\ 1&0\\ 0&1\end { array } \right ] & t(13,7 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(13,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] & t(14,19 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(15,20 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&1&0\\ 0&0&1\end { array } \right ] & t(16,12 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\end { array } \right ] \\ t(16,21 ) & = \left [ \begin { array}{cc } 0&1\\ 1&0\end { array } \right ] & t(16,10 ) & = \left [ \begin { array}{ccc } 1&0&1\\ 0&1&0\end { array } \right ] \\ t(17,22 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&1&0\\ 0&0&1\end { array } \right ] & t(18,23 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(19,9 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(19,10 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\end { array } \right ] \\ t(20,24 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&0&1\\ 0&1&0\end { array } \right ] & t(20,25 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&0&1\\ 0&0&1&0\end { array } \right ] \\ t(20,26 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&0\\ 0&0&1\end { array } \right ] & t(21,27 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(22,28 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&0&1\\ 0&1&0\end { array } \right ] & t(22,29 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&0&1\\ 0&0&1&0\end { array } \right ] \\ t(22,30 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&0\\ 0&0&1\end { array } \right ] & t(23,12 ) & = \left [ \begin { array}{ccc } 1&0&1\\ 0&1&0\end { array } \right ] \\ t(23,13 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(24,28 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&0&1\\ 0&1&0\end { array } \right ] \\ t(25,29 ) & = \left [ \begin { array}{cccc } 1&0&0&0\\ 0&1&0&0\\ 1&0&0&1\\ 0&0&1&0\end { array } \right ] & t(25,30 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&1&0\\ 1&0&0\\ 0&0&1\end { array } \right ] \\ t(26,12 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\\ 0&1&0\end { array } \right ] & t(26,21 ) & = \left [ \begin { array}{cc } 0&1\\ 1&0\\ 0&1\end { array } \right ] \\ t(26,10 ) & = \left [ \begin { array}{ccc } 1&0&1\\ 0&1&0\\ 0&0&1\end { array } \right ] & t(27,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(28,12 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&1&0\\ 1&0&1\end { array } \right ] & t(28,21 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\\ 1&0\end { array } \right ] \\ t(28,10 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&1\\ 0&1&0\end { array } \right ] & t(29,24 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&0&1\\ 0&1&0\\ 0&0&1\end { array } \right ] \\ t(29,25 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&0&1\\ 0&0&1&0\\ 0&0&0&1\end { array } \right ] & t(30,26 ) & = \left [ \begin { array}{ccc } 0&1&0\\ 1&0&0\\ 0&0&1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 3 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 ] . the essential class is of positive type . an example is the path [ 5 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1033 $ ] . the minimum comes from the loop @xmath1034 $ ] . the maximum comes from the loop @xmath1035 $ ] . these points will include points of local dimension [ .8778224563 , 1.232482628 ] . the spectral range is contained in the range @xmath1036 $ ] . the minimum comes from the total row sub - norm of length 10 . the maximum comes from the total row sup - norm of length 10 . these points will have local dimension contained in [ .7764098609 , 1.232482628 ] . consider @xmath746 , the root of @xmath741 and the maps @xmath747 with @xmath748 , and @xmath749 . the probabilities are uniform . the reduced transition diagram has 11 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath1037 * reduced characteristic vector 3 : @xmath1038 * reduced characteristic vector 4 : @xmath1039 * reduced characteristic vector 5 : @xmath1040 * reduced characteristic vector 6 : @xmath1041 * reduced characteristic vector 7 : @xmath1042 * reduced characteristic vector 8 : @xmath1043 * reduced characteristic vector 9 : @xmath753 * reduced characteristic vector 10 : @xmath1044 * reduced characteristic vector 11 : @xmath1045 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(3,6 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(4,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(5,8 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(6,9 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(7,10 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(7,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(7,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(9,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] & t(9,11 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] \\ t(9,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] & t(10,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(11,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(11,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(11,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & \end{aligned}\ ] ] the essential class is : [ 3 , 5 , 6 , 7 , 8 , 9 , 10 , 11 ] . the essential class is of positive type . an example is the path [ 5 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1046 $ ] . the minimum comes from the loop @xmath1034 $ ] . the maximum comes from the loop @xmath1047 $ ] . these points will include points of local dimension [ .8531002132 , 1.137466951 ] . the spectral range is contained in the range @xmath1048 $ ] . the minimum comes from the total row sub - norm of length 10 . the maximum comes from the total row sup - norm of length 10 . these points will have local dimension contained in [ .7962268660 , 1.137466951 ] . the number of reduced characteristic vectors in the essential class is 535 . the essential class is of positive type . an example is the path [ 5 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1049 $ ] . the minimum comes from the loop @xmath1050 $ ] . the maximum comes from the loop @xmath1051 $ ] . these points will include points of local dimension [ .8572351366 , 1.075364982 ] . the spectral range is contained in the range @xmath1052 $ ] . the minimum comes from the total column sub - norm of length 10 . the maximum comes from the total column sup - norm of length 10 . these points will have local dimension contained in [ .8256729656 , 1.075364982 ] . the number of reduced characteristic vectors in the essential class is 187 . the essential class is of positive type . an example is the path [ 5 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1055 $ ] . the minimum comes from the loop @xmath1034 $ ] . the maximum comes from the loop @xmath1056 $ ] . these points will include points of local dimension [ .8974197439 , 1.110448830 ] . the spectral range is contained in the range @xmath1057 $ ] . the minimum comes from the total row sub - norm of length 10 . the maximum comes from the total row sup - norm of length 10 . these points will have local dimension contained in [ .7584448787 , 1.110448830 ] . consider @xmath746 , the root of @xmath744 and the maps @xmath747 with @xmath748 , and @xmath749 . the probabilities are uniform . the reduced transition diagram has 14 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath1060 * reduced characteristic vector 3 : @xmath1061 * reduced characteristic vector 4 : @xmath1062 * reduced characteristic vector 5 : @xmath1063 * reduced characteristic vector 6 : @xmath1064 * reduced characteristic vector 7 : @xmath1065 * reduced characteristic vector 8 : @xmath1066 * reduced characteristic vector 9 : @xmath1041 * reduced characteristic vector 10 : @xmath1067 * reduced characteristic vector 11 : @xmath1068 * reduced characteristic vector 12 : @xmath753 * reduced characteristic vector 13 : @xmath1069 * reduced characteristic vector 14 : @xmath1070 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(2,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(3,6 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(4,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(4,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(4,4 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(5,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(5,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(5,8 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(6,9 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] \\ t(7,10 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(7,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] \\ t(7,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(8,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(8,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(8,11 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(9,12 ) & = \left [ \begin { array}{cc } 1&0\\ 0&1\end { array } \right ] & t(10,13 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(10,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(10,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(11,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(12,3 ) & = \left [ \begin { array}{cc } 1&0\\ 1&1\end { array } \right ] \\ t(12,14 ) & = \left [ \begin { array}{c } 1\\ 1\end { array } \right ] & t(12,3 ) & = \left [ \begin { array}{cc } 1&1\\ 0&1\end { array } \right ] \\ t(13,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(14,7 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(14,3 ) & = \left [ \begin { array}{cc } 1&1\end { array } \right ] & t(14,5 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 3 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 ] . the essential class is of positive type . an example is the path [ 5 , 3 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1071 $ ] . the minimum comes from the loop @xmath1034 $ ] . the maximum comes from the loop @xmath1072 $ ] . these points will include points of local dimension [ .8449717217 , 1.056214652 ] . the spectral range is contained in the range @xmath1073 $ ] . the minimum comes from the total row sub - norm of length 10 . the maximum comes from the total row sup - norm of length 10 . these points will have local dimension contained in [ .7831871251 , 1.056214652 ] . consider @xmath746 , the root of @xmath727 and the maps @xmath747 with @xmath748 , @xmath1074 , and @xmath1075 . the probabilities are given by @xmath1076 , @xmath1077 , and @xmath1078 . the reduced transition diagram has 40 reduced characteristic vectors . the reduced characteristic vectors are : * reduced characteristic vector 1 : @xmath551 * reduced characteristic vector 2 : @xmath1079 * reduced characteristic vector 3 : @xmath1080 * reduced characteristic vector 4 : @xmath1081 * reduced characteristic vector 5 : @xmath1082 * reduced characteristic vector 6 : @xmath1083 * reduced characteristic vector 7 : @xmath1084 * reduced characteristic vector 8 : @xmath1085 * reduced characteristic vector 9 : @xmath1086 * reduced characteristic vector 10 : @xmath1087 * reduced characteristic vector 11 : @xmath1088 * reduced characteristic vector 12 : @xmath1089 * reduced characteristic vector 13 : @xmath1090 * reduced characteristic vector 14 : @xmath1091 * reduced characteristic vector 15 : @xmath1092 * reduced characteristic vector 16 : @xmath1093 * reduced characteristic vector 17 : @xmath1094 * reduced characteristic vector 18 : @xmath1095 * reduced characteristic vector 19 : @xmath1096 * reduced characteristic vector 20 : @xmath1097 * reduced characteristic vector 21 : @xmath1098 * reduced characteristic vector 22 : @xmath1099 * reduced characteristic vector 23 : @xmath1100 * reduced characteristic vector 24 : @xmath1101 * reduced characteristic vector 25 : @xmath1102 * reduced characteristic vector 26 : @xmath1103 * reduced characteristic vector 27 : @xmath1104 * reduced characteristic vector 28 : @xmath1105 * reduced characteristic vector 29 : @xmath1106 * reduced characteristic vector 30 : @xmath1107 * reduced characteristic vector 31 : @xmath1108 * reduced characteristic vector 32 : @xmath1109 * reduced characteristic vector 33 : @xmath1110 * reduced characteristic vector 34 : @xmath1111 * reduced characteristic vector 35 : @xmath1112 * reduced characteristic vector 36 : @xmath1113 * reduced characteristic vector 37 : @xmath1114 * reduced characteristic vector 38 : @xmath1115 * reduced characteristic vector 39 : @xmath1116 * reduced characteristic vector 40 : @xmath1117 this has transition matrices : @xmath755 & t(1,3 ) & = \left [ \begin { array}{cc } 2&1\end { array } \right ] \\ t(1,4 ) & = \left [ \begin { array}{ccc } 1&2&1\end { array } \right ] & t(1,5 ) & = \left [ \begin { array}{cc } 1&2\end { array } \right ] \\ t(1,6 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(2,2 ) & = \left [ \begin { array}{c } 1\end { array } \right ] \\ t(2,7 ) & = \left [ \begin { array}{cc } 2&1\end { array } \right ] & t(3,8 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&2&1\end { array } \right ] \\ t(3,9 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&2&1\end { array } \right ] & t(3,10 ) & = \left [ \begin { array}{ccccc } 2&0&1&0&0\\ 0&1&0&2&1\end { array } \right ] \\ t(4,11 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 0&2&0&1&0\\ 0&0&1&0&2\end { array } \right ] & t(4,12 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 1&0&2&0&1&0\\ 0&0&0&1&0&2\end { array } \right ] \\ t(4,13 ) & = \left [ \begin { array}{cccccc } 2&0&1&0&0&0\\ 0&1&0&2&0&1\\ 0&0&0&0&1&0\end { array } \right ] & t(4,14 ) & = \left [ \begin { array}{ccccc } 2&0&1&0&0\\ 0&1&0&2&0\\ 0&0&0&0&1\end { array } \right ] \\ t(5,15 ) & = \left [ \begin { array}{ccccc } 1&2&0&1&0\\ 0&0&1&0&2\end { array } \right ] & t(5,16 ) & = \left [ \begin { array}{cccc } 1&2&0&1\\ 0&0&1&0\end { array } \right ] \\ t(5,17 ) & = \left [ \begin { array}{ccc } 1&2&0\\ 0&0&1\end { array } \right ] & t(6,18 ) & = \left [ \begin { array}{cc } 1&2\end { array } \right ] \\ t(6,6 ) & = \left [ \begin { array}{c } 1\end { array } \right ] & t(7,8 ) & = \left [ \begin { array}{ccc } 1&0&0\\ 0&2&1\end { array } \right ] \\ t(7,9 ) & = \left [ \begin { array}{cccc } 0&1&0&0\\ 1&0&2&1\end { array } \right ] & t(8,10 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 2&0&1&0&0\\ 0&1&0&2&1\end { array } \right ] \\ t(9,11 ) & = \left [ \begin { array}{ccccc } 1&0&0&0&0\\ 0&1&0&0&0\\ 0&2&0&1&0\\ 0&0&1&0&2\end { array } \right ] & t(9,12 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 2&0&1&0&0&0\\ 1&0&2&0&1&0\\ 0&0&0&1&0&2\end { array } \right ] \\ t(10,19 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 2&0&1&0&0&0\\ 0&2&0&1&0&0\\ 0&1&0&2&0&1\\ 0&0&0&0&1&0\end { array } \right ] & t(10,20 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(10,21 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&2&0&1&0&0\\ 1&0&2&0&1&0\\ 0&0&1&0&2&0\\ 0&0&0&0&0&1\end { array } \right ] & t(11,22 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 2&0&1&0&0&0\\ 1&0&2&0&1&0\\ 0&1&0&2&0&1\\ 0&0&0&1&0&2\end { array } \right ] \\ t(12,23 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&1&0&0&0&0\\ 0&2&0&1&0&0\\ 0&1&0&2&0&1\\ 0&0&1&0&2&0\\ 0&0&0&0&1&0\end { array } \right ] \\ t(12,24 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(12,25 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 2&0&1&0&0&0\\ 1&0&2&0&1&0\\ 0&0&1&0&2&0\\ 0&0&0&1&0&2\\ 0&0&0&0&0&1\end { array } \right ] \\ t(13,19 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 2&0&1&0&0&0\\ 0&2&0&1&0&0\\ 0&1&0&2&0&1\\ 0&0&0&1&0&2\\ 0&0&0&0&1&0\end { array } \right ] \\ t(13,20 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(13,21 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 0&2&0&1&0&0\\ 1&0&2&0&1&0\\ 0&0&1&0&2&0\\ 0&0&0&0&1&0\\ 0&0&0&0&0&1\end { array } \right ] \\ t(14,22 ) & = \left [ \begin { array}{cccccc } 2&0&1&0&0&0\\ 1&0&2&0&1&0\\ 0&1&0&2&0&1\\ 0&0&0&1&0&2\\ 0&0&0&0&0&1\end { array } \right ] & t(15,23 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 0&2&0&1&0&0\\ 0&1&0&2&0&1\\ 0&0&1&0&2&0\\ 0&0&0&0&1&0\end { array } \right ] \\ t(15,24 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\end { array } \right ] & t(15,25 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 1&0&2&0&1&0\\ 0&0&1&0&2&0\\ 0&0&0&1&0&2\\ 0&0&0&0&0&1\end { array } \right ] \\ t(16,13 ) & = \left [ \begin { array}{cccccc } 2&0&1&0&0&0\\ 0&1&0&2&0&1\\ 0&0&0&1&0&2\\ 0&0&0&0&1&0\end { array } \right ] & t(16,14 ) & = \left [ \begin { array}{ccccc } 2&0&1&0&0\\ 0&1&0&2&0\\ 0&0&0&1&0\\ 0&0&0&0&1\end { array } \right ] \\ t(17,15 ) & = \left [ \begin { array}{ccccc } 1&2&0&1&0\\ 0&0&1&0&2\\ 0&0&0&0&1\end { array } \right ] & t(18,16 ) & = \left [ \begin { array}{cccc } 1&2&0&1\\ 0&0&1&0\end { array } \right ] \\ t(18,17 ) & = \left [ \begin { array}{ccc } 1&2&0\\ 0&0&1\end { array } \right ] & t(19,19 ) & = \left [ \begin { array}{cccccc } 1&0&0&0&0&0\\ 2&0&1&0&0&0\\ 0&2&0&1&0&0\\ 0&1&0&2&0&1\\ 0&0&0&1&0&2\\ 0&0&0&0&1&0\end { array } \right ] \\ t(19,20 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(20,26 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&0&1&0&0&0&0\\ 0&0&2&0&1&0&0\\ 0&1&0&2&0&1&0\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(21,27 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&0&1&0&0&0\\ 1&0&0&2&0&1&0\\ 0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(22,28 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&0&1&0&0\\ 0&1&0&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(22,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(22,30 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&0&1&0\\ 0&0&1&0&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(23,31 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0\\ 0&1&0&2&0&0&1\\ 0&0&0&1&0&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(24,32 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 0&1&0&2&0&1&0\\ 0&0&1&0&2&0&0\\ 0&0&0&0&1&0&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(25,24 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(25,25 ) & = \left [ \begin { array}{cccccc } 0&1&0&0&0&0\\ 2&0&1&0&0&0\\ 1&0&2&0&1&0\\ 0&0&1&0&2&0\\ 0&0&0&1&0&2\\ 0&0&0&0&0&1\end { array } \right ] \\ t(26,27 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 2&0&0&1&0&0&0\\ 1&0&0&2&0&1&0\\ 0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(27,28 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&0&1&0&0\\ 0&1&0&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(27,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(27,30 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&0&1&0\\ 0&0&1&0&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(28,33 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&1&0&0\\ 0&1&0&2&0&0&1\\ 0&0&0&1&0&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(28,34 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0&0\\ 0&1&0&2&0&1&0&0\\ 0&0&1&0&2&0&0&1\\ 0&0&0&0&1&0&0&2\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(29,35 ) & = \left [ \begin { array}{cccccccc } 1&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0\\ 0&0&2&0&1&0&0&0\\ 0&1&0&2&0&1&0&0\\ 0&0&1&0&2&0&1&0\\ 0&0&0&1&0&2&0&0\\ 0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&1\end { array } \right ] \\ t(30,36 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 2&0&0&1&0&0&0&0\\ 1&0&0&2&0&1&0&0\\ 0&0&1&0&2&0&1&0\\ 0&0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(30,37 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&0&1&0&0&0\\ 1&0&0&2&0&1&0\\ 0&0&1&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(31,28 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&0&1&0&0\\ 0&1&0&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(31,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(31,30 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&0&1&0\\ 0&0&1&0&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(32,31 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0\\ 0&1&0&2&0&0&1\\ 0&0&0&1&0&0&2\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(33,28 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&0&1&0&0\\ 0&1&0&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(33,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(34,38 ) & = \left [ \begin { array}{cccccccc } 1&0&0&0&0&0&0&0\\ 0&0&1&0&0&0&0&0\\ 0&0&2&0&1&0&0&0\\ 0&1&0&2&0&0&1&0\\ 0&0&0&1&0&0&2&0\\ 0&0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&1\end { array } \right ] \\ t(35,39 ) & = \left [ \begin { array}{cccccccc } 1&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0\\ 2&0&0&1&0&0&0&0\\ 1&0&0&2&0&1&0&0\\ 0&0&1&0&2&0&0&1\\ 0&0&0&0&1&0&0&2\\ 0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&1\end { array } \right ] \\ t(36,40 ) & = \left [ \begin { array}{cccccccc } 1&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0&0\\ 0&2&0&0&1&0&0&0\\ 0&1&0&0&2&0&1&0\\ 0&0&0&1&0&2&0&0\\ 0&0&0&0&0&1&0&0\\ 0&0&0&0&0&0&0&1\end { array } \right ] \\ t(37,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(37,30 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&0&1&0\\ 0&0&1&0&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(38,36 ) & = \left [ \begin { array}{cccccccc } 1&0&0&0&0&0&0&0\\ 0&1&0&0&0&0&0&0\\ 2&0&0&1&0&0&0&0\\ 1&0&0&2&0&1&0&0\\ 0&0&1&0&2&0&1&0\\ 0&0&0&1&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(38,37 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 2&0&0&1&0&0&0\\ 1&0&0&2&0&1&0\\ 0&0&1&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(39,28 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&0&1&0&0\\ 0&1&0&0&2&0&1\\ 0&0&0&1&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\end { array } \right ] \\ t(39,29 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&0&1&0&0\\ 0&0&1&0&0&2&0&1\\ 0&0&0&0&1&0&2&0\\ 0&0&0&0&0&1&0&2\\ 0&0&0&0&0&0&1&0\end { array } \right ] \\ t(39,30 ) & = \left [ \begin { array}{ccccccc } 0&1&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&0&1&0\\ 0&0&1&0&0&2&0\\ 0&0&0&0&1&0&2\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(40,33 ) & = \left [ \begin { array}{ccccccc } 1&0&0&0&0&0&0\\ 2&0&1&0&0&0&0\\ 0&2&0&1&0&0&0\\ 1&0&2&0&1&0&0\\ 0&1&0&2&0&0&1\\ 0&0&0&1&0&0&2\\ 0&0&0&0&0&1&0\\ 0&0&0&0&0&0&1\end { array } \right ] \\ t(40,34 ) & = \left [ \begin { array}{cccccccc } 0&1&0&0&0&0&0&0\\ 0&2&0&1&0&0&0&0\\ 1&0&2&0&1&0&0&0\\ 0&1&0&2&0&1&0&0\\ 0&0&1&0&2&0&0&1\\ 0&0&0&0&1&0&0&2\\ 0&0&0&0&0&0&1&0\\ 0&0&0&0&0&0&0&1\end { array } \right ] \\ \end{aligned}\ ] ] the essential class is : [ 28 , 29 , 30 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 ] . the essential class is of positive type . an example is the path [ 28 , 34 , 38 , 36 , 40 , 34 , 38 , 37 ] . the essential class is not a simple loop . this spectral range will include the interval @xmath1118 $ ] . the minimum comes from the loop @xmath1119 $ ] . the maximum comes from the loop @xmath1120 $ ] . these points will include points of local dimension [ 3.873239615 , 3.883344935 ] . the spectral range is contained in the range @xmath1121 $ ] . the minimum comes from the column sub - norm on the subset @xmath1122 of length 20 . the maximum comes from the total column sup - norm of length 10 . these points will have local dimension contained in [ 3.696560893 , 4.281748469 ] .

a multi - messenger approach is necessary to study the most energetic particles in nature . we expect the sources to produce cosmic rays ( i.e. charged particles ) , photons and neutrinos . their detection requires multiple techniques . cosmic rays are detected in ground arrays and atmospheric fluorescence telescopes ( i.e. the pierre auger observatory ) , detectors in balloons ( e.g. heat , act ) and satellites ( e.g. pamela , fermi space telescope ) and the space station ( ams ) . high energy gamma rays are studied with air cherenkov telescopes ( act , such as hess , veritas , cangaroo , magic ) , detectors in satellites ( e.g. integral and fermi ) , ground arrays and atmospheric fluorescence telescopes ( auger ) . high energy neutrinos require under ice or water kilometer cube detectors , such as icecube in antarctica or the km3net in the mediterranean see , and radio telescopes in balloons , such as anita . i will concentrate here on cosmic rays and will mention photons ( see the talk of francis halzen on neutrinos in this conference ) . cosmic rays have a non - thermal spectrum well approximated by a power law close to @xmath0 for 10 orders of magnitude in energy , from @xmath1 ev to above @xmath2 ev , and 30 orders of magnitude in the flux . the flux goes from 1 particle/(m@xmath3 sec ) at 10@xmath4 ev , to 1 particle/(m@xmath3yr ) at 10@xmath5 ev , to 1 particle/(km@xmath3yr ) at 10@xmath6 ev ( see fig . 1 ) . the non - thermal spectrum of cosmic rays indicates that their acceleration results from stochastic processes in the presence of magnetic fields , as first proposed by fermi . the diffusive acceleration of charged particles in the shock waves of supernova explosions may give origin to most cosmic rays . but the highest energies of cosmic rays are so extreme that it is not understood if a similar mechanism can accelerate these particles even in the most extreme sources , which may or may not be large enough and have sufficiently large magnetic fields . each energy range of cosmic rays addresses different physics . from 10@xmath7 ev to 10@xmath8 ev the modulation of the incoming charged particles by the solar wind can be studied . the expanding plasma generated by the sun decelerates and partially sweeps - off the lower energy galactic cosmic rays from the vicinity of the earth . this effect depends on the level of solar activity , which has an 11 year cycle . from 10@xmath8 ev to 10@xmath9 or 10@xmath10 ev , cosmic rays allow us to study the galactic sources that produce them and the propagation of charged particles within the galaxy . above this energy range and up to at least a few 10@xmath11 ev the cosmic rays must be of extragalactic origin although the characteristics of the transition is still not understood . thus the so called ultra - high energy cosmic rays ( uhecr ) , those above 10@xmath10 ev , allow us to study their extragalactic sources and the propagation of charged particles in the extragalactic medium . cosmic ray spectrum @xmath0 over 10 orders of magnitude from @xmath1 ev to above @xmath2 ev(from ref . the pamela satellite ( payload for antimatter exploration and light - nuclei astrophysics ) , consists of a magnetic spectrometer in orbit , launched in june 2006 , and has produced its first data releases in 2008 . of particular interest has been the positron fraction @xmath13 data the pamela collaboration announced first in july 2008 @xcite ( see fig . 1 ) , which shows an excess in the 10 -100 gev range with respect to what is expected from secondary cosmic rays , i.e. a fraction diminishing as the energy increases ( secondary electrons and some positrons are produced in the interactions of cosmic ray protons and heavier nuclei ) . this result confirms what was earlier called theheat excess " @xcite , also observed by ams-01 @xcite . fig . 1 shows the pamela data ( in red ) and the data of several prior experiments ( in black ) . the large dispersion in the data seen at energies below a few gev is expected , because the different experiments were done in different moments of the solar cycle . above 10 gev the solar modulation is not important and this is the range where the excess is significant . as seen in fig . 1 , the positron fraction is about 0.05 at 10 gev and grows steadily with energy to reach 0.1 - 0.2 at 100 gev . soon after the publication of the pamela positron data , in november 2008 , the atic collaboration announced a 6@xmath14 excess in the @xmath15 flux in the range 300 to 800 gev @xcite , with respect to what is expected from secondary cosmic rays . if the pamela positron fraction excess is due to an excess of positrons ( as opposed to a defect of electrons , which are much more abundant than positrons in secondary cosmic rays ) it may mean that positrons and electrons are been created by an unknown source . the excess electron flux in the pamela energy range would not be noticed until it becomes comparable to , and then larger than , the secondary electron flux as the energy increases . if this is so , atic may have measured the continuation to higher energies of the excess found by pamela . energy spectrum of @xmath15 multiplied by @xmath16 measured by hess @xcite , atic @xcite and other experiments . the double arrow shaded area indicate the approximate range of the systematic error in the hess spctrum ( from ref . @xcite ) ] energy spectrum of @xmath15 multiplied by @xmath16 measured by hess @xcite , atic @xcite and other experiments . the double arrow shaded area indicate the approximate range of the systematic error in the hess spctrum ( from ref . @xcite ) ] atic ( the advanced thin ionization calorimeter instrument ) is a balloon born calorimeter launched from the mcmurdo base in antarctica , which has flown several times . the data presented are from the 2000 - 2001 and 2002 - 2003 flights . a few days after the announcement of atic , the hess collaboration published its @xmath15 flux measurement above 600 gev , which is steeply falling with increasing energy @xcite . fig . 2 ( from ref . @xcite ) shows the data of hess , atic and other earlier experiments . the @xmath15 flux is multiplied by @xmath16 . the double arrow shows the possible shift of the hess spectrum due to @xmath17% energy scale uncertainty and the shaded area indicates the approximate range of the systematic error in the hess spectrum due to other reasons . the hess data show that , if the atic data are correct , there is a sharp cutoff of any excess above about 800 gev , which is compatible with the sharp cutoff expected from dark matter particle annihilation at the parent particle mass . since their appearance , at the moment of writing these notes the data of pamela and atic have generated more than 50 papers , at the rate of several per week , providing different explanations for their excesses . the main two explanations are the existence of an astrophysical source nearby , or dark matter particles annihilating nearby , within a volume around earth much smaller than the volume of the galaxy . electron and positrons rapidly loose energy through synchrotron and inverse compton processes , thus at the energies observed they must come from less than 1 kpc away ( for comparison , the distance between the sun and the center of the galaxy is about 8 kpc ) . known possible astrophysical sources are pulsars , of which there are some , such as geminga and vela , within a few 100 pc from earth @xcite . it is not known the mechanism through which they would produce such energetic electrons and positrons , thus the spectrum they could produce is unknown too . the maximum energy cutoff of the spectrum observed would in this case correspond to the maximum energy of emission of positrons and electrons . in the case of annihilating dark matter particles , the cutoff in energy would correspond to the mass of the parent particle . the cutoff could be very sharp if the particles annihilate directly into @xmath18 pairs , and this would be a clear indication of the dark matter origin of the signal @xcite ( see fig . if instead @xmath18 are produced after a chain of decays of annihilation products , the cutoff would be less sharp , similar to what would be expected from astrophysical sources . fig . 4 ( taken from ref . @xcite ) shows three fits to the atic data , assuming an astrophysical source , a dark matter particle annihilating directly to @xmath18 and another producing @xmath18 as secondaries of the annihilation products . in order to account for the pamela and/or atic data , the annihilation rate of most dark matter candidates needs to be enhanced by a factor , called boost factor " @xmath19 , which may reach @xmath20 in the case of many usual wimp dark matter candidates @xcite and may be much smaller , close to @xmath21 ( see fig . 4 ) , for some candidates that decay preferentially into electrons and positrons ( the atic collaboration finds that a kaluza - klein particle decaying mostly into electrons and positrons with mass 620 gev fits well their electron - positron data @xcite ) . examples of fits to the atic data assuming a pulsar nearby ( red ) , the annihilation of 800 gev dark matter particles producing @xmath18 as secondaries with @xmath22 ( blue ) and the annihilation of 620 gev kaluza - klein dark matter particle decaying directly into @xmath18 with @xmath23 ( black ) . from ref . @xcite ] several ways of enhancing the local annihilation rate have been proposed : there could be a dense dark matter clamp nearby ( however , the probability found in structure formation simulations for this to happen is about 10@xmath24 @xcite ) , which would enhance the rate , which depends on the square of the dark matter density , or the annihilation cross section itself could be enhanced . for example a sommerfeld resonance " results from the almost formation of a bound state by the annihilating particles at very small relative velocities . this effect would enhance the annihilation cross section in the dark halo ( but not in the early universe , where the particles move at larger velocities ) @xcite . whatever produces the positron excess seen by pamela should not produce an excess of antiprotons , since the antiproton to proton ratio @xmath25 was also measured by pamela @xcite and is compatible with what is expected from secondary cosmic rays . this could be partially explained because there are many more protons than electrons in the secondary cosmic rays ( pamela requires a 10@xmath26 discrimination of protons to measure positrons ) and because @xmath27 and @xmath28 reach us from a larger volume than @xmath29 and @xmath30 , from a large fraction of the total volume of the galaxy . the fermi space telescope ( ex glast - gamma - ray large area space telescope ) can also detect @xmath29 and @xmath30 ( it does discriminate between the two ) , besides photons . fermi was launched in june 2008 and is providing @xmath31-ray spectroscopic data of unprecedented quality . it has two detectors ( the glast burst monitor , gmb , and the large area telescope , lat ) which detect photons between 10 kev and 300 gev , and observe the whole sky every 3 hours . photons of even higher energy will be detected by air cherenkov telescopes ( acts ) , cangaroo iii in australia , hess in namibia , magic in las palmas and veritas in the us ( see the talk of manel martinez in this conference ) . these can also detect @xmath32 , as we already mentioned in the case of hess . the universe is totally transparent to photons below 100 gev . at higher energies photons interact with infra - red and optical backgrounds but still at energies below 10 s of tev arrive to us from cosmological distances ( at higher energies they are absorbed by the cosmic microwave background radiation ) . photons can again reach earth from cosmological distance at energies above @xmath1 gev , the range of uhecr . the photons observed by fermi reveal , therefore , the spacial distribution of their sources . in particular , if fermi observes the annihilation of dark matter particles into photons , it is expected to detect not only the center of our galaxy but also a large amount of dark matter clumps . as mentioned above , since the annihilation rate depends on the square of the density , high density regions boost the rate . a large amount of dark matter clumps are expected to remain within the dark halo of our galaxy . haloes grow hierarchically , incorporating lumps and tidal streams from earlier phases of structure formation . the best simulations of structure formation at present ( the via lactea ii @xcite and the aquarius project @xcite ) show that 10 s of subhaloes could be discovered by fermi . however , subhaloes are more effectively destroyed by tidal effects near the center of the galaxy , thus most of them are far from the sun . the chance that a random point close to the sun is lying within a clump is smaller than 10@xmath24 @xcite . dwarf galaxies and other objects which fermi might observe as a whole through dark matter annihilation into photons will have the annihilation rate increased by a boost factor @xmath19 , may be as large of 100 , due to the existence of substructure within them ( with respect to the rate they would have without dark matter lumps ) . but the concept of a boost factor can not be applied to the possible enhancement of a signal in electron and positrons as seen by pamela or atic , since at most only one or very few dark matter clump could be present within the volume these particles come from . the fermi space telescope can measure the @xmath15 flux in the range explored by atic , possibly to 1tev , with better accuracy and its results are expected to be made public soon , in the april 2009 meeting of the american physical society ( which will take place in the first week of may 2009 ) . all - particle energy spectrum of cosmic rays , the flux is multiplied by @xmath16 ( from ref . the lines represent spectra for elemental groups @xmath34 as indicated according to a particular model . the sum of all elements ( galactic ) and a presumably extragalactic component are shown as well ( see ref . @xcite for details ) . ] in fig . 1 two features of the cosmic ray spectrum are indicated : the knee " above 10@xmath35 ev and the ankle " above 10@xmath10 ev . these are shown also in fig . 5 ( taken from ref . @xcite ) in which the flux has been multiplied by a factor @xmath16 . at the knee there is a slight steepening of the spectrum , the spectral index changes from @xmath36 at low energies to about -3.1 at higher energies . second knee " feature is sometimes mentioned , at about 4@xmath37 ev , where the spectrum seems to become again steeper at higher energies with a spectral index approximately @xmath38 ( see fig . 5 ) . at the ankle the spectrum seems to flatten again to a spectral index of about @xmath39 . the ankle " is sometimes seen as just part of a dip " in the spectrum between 10@xmath10 and 10@xmath6 ev . at energies below 10@xmath40 ev the elemental composition of cosmic rays is well known because cosmic ray particles are detected directly . the composition and flux of cosmic rays at these energies are in agreement with their production in galactic supernova explosions @xcite . direct measurements above the atmosphere to energies slightly above 10@xmath40 ev are performed by various instruments , such as atic , cream , bess and tracer . at higher energies , the flux of cosmic rays becomes smaller than several particles /(m@xmath3 day ) so their detection with instruments carried in balloons or spacecrafts becomes impractical and cosmic rays are detected from the ground through the shower of secondary particles they produced in the atmosphere . these are called extensive air showers ( eas ) . arrays of detectors covering a large area , usually from a fraction to many km@xmath3 , operate for years to obtain enough statistics . because of the sparse sampling of the shower , the uncertainties in the modeling of the shower development and the large statistical fluctuations in the interactions in the atmosphere , the inferred energy and characteristics of the shower can only determine the relative contribution of groups of elements to the cosmic ray flux . for example , a heavy nucleus ( usually referred to just as iron " ) generates a shower with a relatively higher ratio of muons to electrons that a proton initiated shower of the same energy . a good understanding of the composition around the knee would help determining what is the origin of this feature , but this is difficult . the best measurements of the composition of cosmic rays in the knee region come from the kascade ( karlsruhe shower core and array detector ) experiment , which observes a steepening of the spectrum first of light elements and successively at higher energies for heavier ones @xcite ( see fig . 5 ) . the knee can be explained by the cutoff of the galactic proton spectrum due to the leakage of protons from the galaxy . the galactic magnetic fields are not strong enough to keep protons of this energy bound to the galaxy . heavier elements are bound to the galaxy up to higher energies , so their cutoff energy increases approximately as their charge @xmath34 . however , the data may also be explained if nuclei are accelerated in the supernova shock fronts to maximum energies proportional to @xmath34 @xcite . two possibilities for the crossing of galactic to extragalactic cosmic rays : 3a . ( left ) shows the crossing at the second knee " @xmath41 gev ( the ankle is part of a dip due to the interaction of extragalactic protons with the cmb ) ; 3b . ( right ) shows the crossing at the ankle , above 10@xmath42 gev . figure taken from ref . @xcite ] the interpretation of the second knee is linked to that of the ankle . the second knee is sometimes explained by the fall off of the heaviest elements ( including elements heavier than iron ) in the galactic cosmic rays @xcite ( see fig . the second knee would then be the place in the spectrum where an extragalactic component becomes dominant , as shown in fig . compilation by r. engel of the main cosmic rays data up to 2007 . the fluxes are multiplied by @xmath43 . the agasa spectrum does not show a gzk feature , which is present in the hires and 2007 auger spectra . ] in this case , the ankle is an absorption dip " in the spectrum between 10@xmath10 and 10@xmath6 ev proposed to be caused by electron - positron pair production of extragalactic protons on cosmic microwave background ( cmb ) photons over large distances @xcite . this is one of the possible explanations for the ankle . historically the ankle was interpreted as the transition from a rapidly falling galactic ( iron dominated ) component to a flatter spectrum of extragalactic origin subdominant at lower energies . both interpretations are shown in fig . 6b @xcite . the cosmic rays above 10@xmath10 ev are called ultra - high energy cosmic rays ( uhecr ) . pierre auger observed cosmic rays with energies above 10@xmath35 ev already in the 1930 s , and 80 years later their origin is still uncertain . he was first in understanding that extensive air showers are produced by high energy cosmic rays . giuseppe cocconi argued in 1956 that cosmic rays above the ankle are extragalactic in origin because they can not be confined by the galactic magnetic field . in 1962 john linsley at volcano ranch in new mexico observed the first 10@xmath11 ev shower . immediately after the discovery of the microwave background radiation by penzias and wilson in 1965 , in two separate papers in 1966 , greisen , and independently zatsepin and kuzmin @xcite pointed out that if the uhecr are of extragalactic origin there should be an absorption feature in their spectrum at about 4@xmath44 ev , the so called gzk cut - off " . in fact , if the uhecr are extragalactic their energy is degraded by inelastic collisions with background photons as they propagate . protons would loose energy mainly through photopion production in the cmb background @xmath45 , where the pion carries away @xmath46 of the original nucleon energy . the mean free path for this reaction is only @xmath47 mpc . heavier nuclei mostly loose energy through photodisintegration . for both , protons and heavy nuclei , the energy losses become important above about 4@xmath44 ev , thus they could not reach earth from a distance beyond 50 to 100 mpc . hillas plot " showing the magnetic field versus size of potential uhecr sources . acceleration of protons to 1@xmath48zev@xmath49 ev , or protons or fe nuclei to 100 eev @xmath50 ev require conditions above the respective line . ] until recently the existence of the gzk absorption feature was in question . the results of the main uhecr experiments until a few years ago , agasa and hires , were contradictory ( see fig.7 ) . the agasa experiment in japan , an array of 111 surface detectors ( scintillator counters ) distributed in 100 km@xmath3 with 1 km spacing between them , took data from 1984 to 2003 and found no gzk cutoff @xcite . this gave rise to several models to explain the agasa excess " above 10@xmath11 ev , in which the uhecr were produced locally so the cosmological energy absorption would not be important . these so called top - down " models invoked new physics to produce uhecr directly with the high energies required , as opposed to bottom - up " models in which uhecr are accelerated in astrophysical sites . in 1993 , the fly s eye experiment , consisting of a fluorescence telescope in utah , usa , detected a 3@xmath5110@xmath11 ev event . the upgrade of this experiment , the high resolution fly s eye ( hires ) , with two fluorescence telescopes took data between 1997 and 2006 and found a spectrum compatible with the gzk cutoff @xcite . the pierre auger observatory in argentina @xcite , a hybrid experiment which combines the two detection methods of agasa and hires , was designed to prove or disprove the existence of the gzk cutoff and to elucidate the mystery of the origin of the highest energy cosmic rays . it consists of both surface detectors ( water tanks ) , sd , and fluorescence detectors , fd . the spectrum of agasa showing no gzk feature , and those of hires and auger in 2007 compatible with a gzk feature , are shown in fig . top - down models were introduced not only to produce the highest energy cosmic rays locally and avoid the gzk absorption feature , but also as an alternative to acceleration models to explain the highest energy cosmic rays , which the latter models have difficulty explaining . only the most extreme sources in the universe are possible sites for the acceleration to protons or other nuclei at those energies : active galactic nuclei ( agn ) , gamma ray bursts ( grb ) , magnetars ( i.e. neutron stars with enormous magnetic fields @xmath52 g ) , huge lobes and hot spots of radio galaxies , super nova remnants ( snr ) , colliding galaxies , large shocks in galaxy clusters ( for a review see for example ref . @xcite , from which fig . 8 was taken ) . the maximum energy at which particles can be accelerated in astrophysical sources is limited by the acceleration rate and the energy loss rate due to inverse compton scattering or inelastic scattering . the hillas plot " @xcite ( see fig . 8) shows astrophysical sites for which the gyroradius @xmath53 of a particle of charge @xmath34 and the desired energy @xmath54 , in the estimated magnetic field @xmath19 of the site , is smaller than the size @xmath55 of the site . the idea is that only in sites where the particles can be confined for a while , the acceleration could take place . 8 shows a modern version of the hillas plot @xcite , including many more potential production sites than the original version of the plot ( from 1984 @xcite ) . as the figure shows , the possible sites where protons could be accelerated diminishes fast as the energy increases above 10@xmath11 ev . top - down models circumvented the need for acceleration of particles by producing them at very high energies . in super heavy dark matter ( shdm ) models , super heavy metastable particles were produced in early universe and remained at present as part of the dark matter of the universe . these particles ( with colorful names such as ` cryptons ' or ` wimpzillas ' ) could decay or annihilate @xcite into quarks and leptons etc which in turn produced the observed uhecr . these would come primordially from within the dark halo of our galaxy , thus shdm models predicted an excess of events towards the galactic center , which has not been found . in these models the maximum energy of cosmic rays was given by the parent particle mass . in other models , the parent particles instead of being metastable were instead produced by topological defects , such as cosmic strings or necklaces ( for a review see ref . @xcite ) . in the z - burst " model @xcite , ultra - high energy neutrinos from remote sources would annihilate at the z - resonance with relic background neutrinos near earth ( this model encountered problems with the large flux of ultrahigh energy neutrinos required ) . the z bosons would then decay , producing secondary protons , neutrinos and photons . the z - resonance , which is the uhecr energy cutoff , occurs when the energy of the incoming @xmath56 is @xmath57 . the spectra of the uhecr produced in top - down models are determined by the elementary particle physics of z - boson decays and of qcd fragmentation , which predict photon domination of the spectrum at high energies . 9a ( left panel ) fluxes measured by yakutsk , agasa , hires and auger , multiplied by @xmath16 . 9b ( righ panel ) the dip " has been used to calibrate the energy of the experiments by factors within their systematic errors . the spectra now coincide , except at the highest energies . see ref . @xcite for details . ] the pierre auger observatory consists of 1600 water tanks with 1.5 km spacing , distributed in a 3,000 km@xmath3 surface , the surface detector ( sd ) and four fluorescence detector sites , each with 6 telescopes that overlook the surface array , the fluorescence detector ( fd ) . the tanks detect the cherenkov light produced by the electromagnetic and muon component of the air showers . the telescopes detect the near ultra - violate light from the flourescent emission of nitrogen induced by the shower . auger started operating in 2004 and has been completed a few months ago , in june 2008 . the auger exposure now is of the order of 10@xmath58 km@xmath3 sr yr . the fd measures the longitudinal shower distribution and provides a calorimetric measurements of the shower energy , but can operate only in moonless nights , which gives it a 10% duty cycle . since the fd telescopes observe the light emitted by the air molecules and not the energy deposited by the shower particles in the air , the energy measurement inferred with the fd depends sensitively on the value adopted for the fluorescence yield , which is still not well known . this is an important source of systematic error . for example , auger and hires use values of the fluorescence yield that differ by about 10% . a better determination of the fluorescence yield is expected soon . the sd measures the transverse shower distribution and observes the sky all the time , thus it provides most of the statistics at the highest energies . about 15% of the showers observed by auger are hybrid events , observed both by the sd and the fd . the hybrid events are used to calibrate the energy of the much larger number of sd only events . the systematic uncertainly in the energy determination of auger is at present 22% . the fluxes measured by yakutsk , agasa , hires and auger are displayed in fig . 9a ( left panel ) . the multiplication of the fluxes by @xmath16 exacerbates the importance of the systematic errors in the energy calibration . fig.9b ( right panel ) shows how the spectra coincide , except at the highest energies , if the dip feature that appears in all of them is used to calibrate the energy by factors within the systematic errors of the different experiments . the energy calibration factors used to produce fig . 9b are 1.0 , 1.2 , 0.75 , 0.83 and 0.625 for hires , auger , agasa , akeno and yakutsk respectively ( see ref . @xcite for details ) . the pierre auger observatory was optimized to observe in the energy range between 10@xmath10 ev and 10@xmath11 ev , because this is the window in which cosmic rays could reach us from extragalactic sources . the extragalactic magnetic fields are not expected to be larger than 10@xmath59 g. recent realistic simulations of the expected large scale extragalactic magnetic fields , show that except in the regions close to the virgo , perseus and coma clusters the obtained magnetic fields are not larger than 3@xmath60 g @xcite ( however see ref . we can see then from eq . 1 that with @xmath61 g at energies @xmath62 ev protons would only be weakly deflected over distances of mpc . at energies above 10@xmath11 ev the gzk cutoff will considerably reduce the flux of uhecr . thus there is a window between 10@xmath10 ev and 10@xmath11 ev to possibly do extragalactic astronomy with charged particles . the main properties of uhecr auger is expected to determine are four : 1 ) the spectrum of the highest energy cosmic rays , in particular to prove or disprove the existence of a gzk cutoff ; 2 ) the composition of those high energy primaries ( are they protons , heavy nuclei , photons ? ) ; 3 ) the arrival direction distribution ( is there a large scale anisotropy , a correlation with particular sources ? ) ; 4 ) the sources . in the following we address these issues in turn . fits to the auger spectrum with different compositions and injection spectra of primary uhecr @xcite . ] fits to the auger spectrum with different compositions and injection spectra of primary uhecr @xcite . ] the about 2@xmath63 events above 2.5@xmath64 ev collected by auger up to august 2007 , with about 4 times the exposure of agasa and 2 times that of hires , have a spectrum compatible with a gzk cutoff at 4@xmath44 ev , as shown in fig . 10 @xcite . the lower panel of fig . 10 shows the fractional differences between the hires and auger data at the highest energies and a power law @xmath65 which fits the spectrum below 4@xmath44 ev . the number of events expected if this power law spectrum were to continue to energies above 4@xmath44 ev or above 1@xmath66 ev would be 167@xmath673 and 35@xmath671 respectively , while auger observed 69 and 1 respectively . this amounts to a 6@xmath14 significance of the flux suppression at the gzk energy , which is compatible with what was found by hires . the spectrum can not be used to determine the composition of uhecr , since models with pure proton primaries and mixed proton - iron primaries all produce good fits to the spectrum . fig.11 shows several fits to the spectrum of auger with proton primaries and mixed composition primaries and several primary spectral indices @xmath31 and maximum energies @xmath68 characterizing the spectrum emitted by the sources @xcite . even the extreme unrealistic assumption of having pure iron emitted at the sources provides good fits to the auger spectrum @xcite . it is also not possible either to determine from the spectrum if the observed suppression is due to the gzk effect or to the energy cut of the spectra of the cosmic rays emitted at the sources . illustration of the observation of the longitudinal shower development with the auger fd . the maximum defines @xmath69 . from ref . @xcite ] the main tool to study the mass composition of cosmic rays is the measurement of the longitudinal profile of the showers . 12 shows an illustration of the measurement of the longitudinal shower development observed with the fluorescence detectors of auger . for a given energy , heavy nuclei interact earlier in the atmosphere than protons ( nuclei interact as a collection of nucleons ) . photon initiated showers are much more penetrating in the atmosphere than hadrons , up to energies higher than 10@xmath11 ev at which photons produce a preshower " ( through interaction with the magnetic field of the earth ) before entering into the atmosphere . thus the column depth of air at which the shower profile is maximum , @xmath69 , as measured with the fd , is well correlated with the primary particle mass . this can be seen in fig . 13 where the expected @xmath69 as function of energy for photon , proton and iron initiated showers are shown , as predicted by several simulation programs . the value of the @xmath69 for protons is about 100 g @xmath70 larger than for iron . the fd detector of auger can measure @xmath69 with a systematic uncertainty of about 15 g / cm@xmath3 . 13 also shows a compilation of the earlier data on @xmath69 as function of the energy , for energies above 10@xmath40 ev . it shows a composition compatible with proton dominance at the highest energies . except for one data point of hires , the measurements reached 2@xmath44 ev . compilation of earlier measurements of the mean depth of shower maximum @xmath69 as function of energy , compared with the predictions of various simulations programs for photon , proton and iron initiated showers ( see for example ref . @xcite ) . ] auger data on @xmath69 @xcite are presented in fig . 14 together with the predictions for proton and iron initiated showers from several shower simulation programs . although with low statistics the auger data show a change of regime ( a change in slope , called elongation rate " ) at 2@xmath44 ev towards a heavier composition . at the highest energies the composition is intermediate between protons and iron , compatible with a mean mass number of about 5 . the errors shown in the figure are statistical . the systematic uncertainty is @xmath67 15 g / cm@xmath3 . the hires collaboration did not find the change in slope at 2@xmath44 ev present in the auger data , and its composition measurements seem compatible with proton dominance to the highest energies ( see fig.16 ) . mean depth of shower maximum @xmath69 as function of energy measured by auger @xcite compared with what is expected from proton and iron initiated showers . the last data point seems to indicate a heavier composition above 2@xmath44 ev . ] mean depth of shower maximum @xmath69 as function of energy measured by hires @xcite ( fig . from ref . @xcite ) compared with what is expected from proton and iron initiated showers . no change in the slope is observed at 2@xmath44 ev . ] auger has obtained stringent upper bounds on photons as uhecr @xcite . uhecr photons with energy above 10@xmath6ev are one of the key observables to distinguish top - down from bottom - up models ( see for example ref . because in the process of hadronization pions are produced more copiously than protons , and neutral pions decay into photons , all top - down models produce more photons than protons ( and no heavy nuclei ) . also gzk " photons , those produced in the decay of photo - produced @xmath72 in the gzk process must always be present at some level , although the level may be extremely low . the photon fraction could be as small at 10@xmath24 @xcite . finding gzk photons among the uhecr would help understanding the primary hadronic spectrum which produced them and the intervening extragalactic backgrounds with which they interact , the radio background and magnetic fields . the flux of gzk photons is related to that of cosmogenic " neutrinos , those neutrinos produced in the decay of the photo - produced charged pions in the gzk process . neutrinos and photons produced as secondaries of the gzk process would give complementary information since neutrinos come without absorption from the whole production volume , while photons can reach us only from less than 100 mpc . photons with energy @xmath73 ev pair - produce electrons and positrons on the radio background and can not reach earth from beyond 10 to 40 mpc ( although the photon energy - attenuation length is not well known , due to the uncertainties in the spectrum of the absorbing radio background ) . same as fig . 13 , but with several characteristic top - down models and other experimental upper bounds . see ref . @xcite for details . ] same as fig . 13 , but with several characteristic top - down models and other experimental upper bounds . see ref . @xcite for details . ] at energies below 10@xmath11 ev photons produce in the atmosphere deeper showers than protons ( due to the landau - pomeranchuk - migdal ( lpm ) effect ) . at higher energies the photons interact with the magnetic field of the earth before entering the atmosphere and produce several electrons below the lpm threshold ( what is called preshowering " ) but the statistics of auger is too low at those energies . bounds based on @xmath69 measurements using fd data were initially used to set an upper bound on photons @xcite . better bounds were obtained later @xcite by using two parameters , the shower - front radius and the rise time of the signal , that can be measured with the sd . late developing showers have smaller shower - front radius ( the shower - front flattens out as the shower develops ) . photon initiated showers have a lower muon content , more low energy particles , and thus longer rise - time of the signal . with these two parameters , the high statistics sd could be used to search for photons . auger found no events compatible with what is expected from a photon primary . as a result it has obtained stringent upper bounds shown in figs . 16 and 17 . the range of photon fractions in the integrated flux for some top - down models and for gzk photons from ref . @xcite together with the 2006 and 2007 auger and agasa - yakutzk ( ay ) @xcite upper bounds are shown in fig . . the auger bounds together with other upper bounds on the photon fraction and characteristic top - down models are shown in fig . 17 @xcite . these photon bounds provide an argument independent of possible correlation with astrophysical sources against most top - down models . if the suppression of the cosmic rays spectrum at 4@xmath44 ev is due to the gzk effect , namely to the interactions with the cmb photons , then the cosmic rays with higher energies should come from nearby sources . for example , above 6@xmath44 ev 90% of the protons should come from less than about 200 mpc away , while 50% should come from less than 90 mpc away . thus , the arrival directions of the highest energy cosmic rays should correlate with the distribution of visible matter nearby , which is very inhomogeneus . the angular resolution of auger above 1@xmath44 ev is less than one degree , while the deflection of protons or nuclei in the @xmath74 g magnetic field of our galaxy , at those energies is expected to be about 10@xmath75 . auger has , in fact , found a correlation with extragalactic astrophysical sources @xcite . auger reported a correlation between the arrival directions of uhecr with @xmath76 ev and the positions of nearby active galactic nuclei ( agn ) with redshift @xmath77 ( i.e. distance @xmath7871 mpc ) : 20 out of 27 cosmic ray events were found to correlate with at least one of 442 agn selected form a particular catalog , the veron - veron - cetty catalog ( 292 within the field of view of auger ) located within a circle of 3.2@xmath79 around the event arrival direction ( only 5.6 are expected to do so if the flux were isotropic ) . the auger collaboration found a statistically significant correlation with a smaller amount of data in may 2006 with very similar energy threshold and angular distance . fixing these parameter a - priori it studied the subsequent data until august 2007 when the original correlation was confirmed to a pre - determined confidence level in the additional data set alone . map in galactic coordinates of the positions of the agn within 71 mpc used in the correlation search of auger ( red stars ) and circles of 3.2@xmath79 centered at the arrival directions of the 27 events with @xmath76 ev observed by auger until august 2007 . the shading indicate levels of exposure @xcite . ] the map of the arrival directions and of the agn positions is shown in fig . there is a clear alignment of several of the correlating uhecr events close to the supergalactic plane , indicated in the figure with a dashed line . two of the events have arrival directions less than 3@xmath79 away from the nucleus of centaurus a , one of the closest agn , at 3.4 mpc away . cen a displays jets and radio lobes which extend over a scale of about 10@xmath79 along the supergalactic plane , and a variable compact radio nucleus . the two events mentioned correlate with the nucleus position while several lie in the vecinity of the radio lobe extension along the supergalactic plane . in general , agn with prominent radio lobes are rare and do not follow the observed spatial distribution . so far there are only three other doublet events ( i.e. the same agn associated with two uhecr ) , one of which is close to and other two which are outside the supergalactic plane . the most prominent radio galaxy within 70 mpc , m87 in the virgo cluster , does not correlate with any observed event above 5.7 @xmath44 ev . there is a lack of events in the region of the virgo cluster which seems to be present also in the data of the hires collaboration . may be large magnetic fields in the direction of virgo could explain the lack of events in that direction . the hires collaboration did not confirm a correlation of uhecr with agns @xcite ) : out of 13 events with @xmath80 only two correlated with an agn within 3.2@xmath79 ( and 3.2 events were expected just by chance ) . hires observed mostly the northern sky and auger the southern , could they be significantly different ? given the small statistics of the sample and the negative results of hires , the correlation found by auger does not prove that the sources of uhecr are agns . the sources may be any astrophysical objects with a space distribution similar to agns . are agn the sources or just tracers of the sources ? nuclei are attenuated by photodisintegration in the intergalactic medium and intermediate mass nuclei ( a @xmath81 ) would have to reach us from distances shorter than 70 mpc . heavier nuclei , close to fe have horizons comparable to that of protons but they would be deflected in the magnetic field of our galaxy by more than 3@xmath79 . thus the correlation indicates that the primary uhecr are most likely protons . however , in the case the primaries are protons , the actual sources of the uhecr may be further out than 70 mpc . we mentioned above that the gzk horizon " ( defined as the distance from earth which contains the sources that produce 90% of the protons which arrive with energies above a given threshold ) should be about 200 mpc for protons of energy above 6 @xmath44 gev , while it would be 70 mpc for about 9 @xmath44 gev . this may indicate that the energy calibration of auger should be shifted upwards by 30% or that there are accidental correlations with agns that are closer than the actual uhecr sources . the correlation with agns found by auger seems to indicate that the highest energy cosmic rays are protons , while the composition data of auger suggest a change towards a heavier composition above @xmath82 ev . which of these two pieces of information is correct ? there may also be a problem with the programs used to simulates the showers and determine if they are proton or iron like . data from lhcf are needed to improve them . clearly more data are needed to answer all these questions , and they are coming . there is an infield with extra muon counters in construction in the pierre auger observatory in argentina . also a new experiment , the telescope array ( with a low energy extension called tale ) is being built in utah ( usa ) . this auger observatory in argentina is the south part of the two - part project initially proposed for auger . there is a current proposal for an auger - north observatory in colorado ( usa ) with 4@xmath83 surface detectors occupying an area of 2@xmath63 km@xmath3 . there is also the jem - euso mission in preparation , the extreme universe space observatory that will be hosted in the japanese experimental module of the international space station , that will study very energetic cosmic ray showers from space . many advances have occurred in recent years in our understanding of the origin and propagation of cosmic rays but many fundamental open question remain . in the gev to tev energy range , the solar modulation of cosmic rays and the solar wind need to be better understood , and we need to understand if the pamela/ atic excess is in fact there and it is not an instrumental issue , and if so if it is due to a pulsar nearby or the dark matter annihilation . the fermi space telescope data expected soon will greatly help in this respect . in the 10@xmath58 - 10@xmath7 gev energy range , the cosmic ray composition , the galactic sources , the propagation of cosmic rays within the galaxy and the transition from the galactic to extra - galactic cosmic rays ( at the second knee or at the ankle ? ) need to be better understood . at energies above 10@xmath42 gev charged particle extragalactic astronomy has started ! but we need yet to identify the sources , understand the composition of the cosmic rays , the mechanisms of acceleration at the sources and the intervening backgrounds and magnetic fields . 9 aharonian f a , atoyan a and volk h j 1995 _ astron . astrophys . _ * 294 * l41 ; hooper d , blasi p and serpico pd 2008 _ jcap _ * 0901 * 025 [ arxiv:0810.1527 [ astro - ph ] ] ; yuksel h , kistler m d and stanev t 2008 arxiv:0810.2784 [ astro - ph ] ; profumo s 2008 arxiv:0812.4457 [ astro - ph ] . berezinsky v , gazizov az , grigorieva s i 2004 _ nucl . suppl . _ * 136 * 147 [ astro - ph/0410650 ] ; berezinsky v , gazizov az , grigorieva s i 2005 _ phys . _ b * 612 * 147 [ astro - ph/0502550 ] ; berezinsky v , gazizov az , grigorieva s i 2006 _ phys . _ d * 74 * 043005 berezinsky v , kachelriess m and vilenkin a 1997 _ phys . lett . _ * 79 * 4302 ; kuzmin v a and rubakov v a 1998 _ phys . atom . nucl . _ * 61 * 1028 [ _ yad . fiz . _ * 61 * 1122 ] ; birkel m and sarkar s 1998 _ astropart . phys . _ * 9 * 297 ; blasi p , dick r and kolb e w 2002 _ astropart . phys . _ * 18 * 57 ; aloisio r , berezinsky v and kachelriess m 2004 _ phys . rev . _ d * 69 * 094023 . dolag k , grasso d , springel v and tkachev i 2004 _ jetp lett . _ * 79 * , 583 [ _ pisma zh . eksp . fiz . _ * 79 * , 719 ] [ arxiv : astro - ph/0310902 ] ; dolag k , grasso d , springel v and tkachev i 2005 _ jcap _ * 0501 * 009 [ arxiv : astro - ph/0410419 ] .

the non - equilibrium green function ( negf ) method@xcite potentially has several advantages to investigate electronic transport properties of nano - scale materials such as single molecules,@xcite atomic wires,@xcite carbon based materials,@xcite and thin layers.@xcite the potential advantages are summarized by the following features of the negf method : ( i ) the source and drain contacts are treated based on the same theoretical framework as for the scattering region . @xcite ( ii ) the electronic structure of the scattering region under a finite source - drain bias voltage is self - consistently determined by combining with first principle electronic structure calculation methods such as the density functional theory ( dft ) and the hartree - fock ( hf ) method . @xcite ( iii ) many body effects in the transport properties , e.g. , electron - phonon@xcite and electron - electron interactions,@xcite might be included through self - energies without largely deviating the theoretical framework . ( iv ) its applicability to large - scale systems can be anticipated , since the negf method relies practically on the locality of basis functions in real space , resulting in computations for sparse matrices.@xcite due to those potential advantages , recently several groups have implemented the negf method coupled with the dft or hf method using atomic - type or the other local basis functions with successful applications for calculations of the electronic transport properties . @xcite however , a highly accurate and efficient implementation method must be still developed from the following two reasons : the first obvious reason is to extend the applicability of the negf method to large - scale systems . the efficient implementation might lead to more challenging applications of the negf method to very large - scale complicated systems . the majority part in the computational effort of the negf method mainly comes from the evaluation of the density matrix which is decomposed into the evaluation of green functions and numerical integrations . thus , the efficient calculation of the part is a key factor for extending the applicability to large - scale systems . nevertheless , accurate and efficient methods for evaluating density matrix within the negf method have not been fully developed , although several methods have been already proposed . @xcite to extend the applicability of the negf method to large - scale systems , a remarkably efficient method that we have recently developed@xcite will be applied for the problem in this study , and we will show that the new method is much faster than the other method.@xcite the second reason is that spurious scattering should be negligible when the negf method is extended to include the many body effects beyond the one particle picture.@xcite the spurious scattering accompanied by the inaccurate implementation might make the many body effects indistinct in the electronic transport properties . one can imagine that the spurious scattering can be easily produced in the negf method , since unlike the conventional band structure calculations negf has to be evaluated by a patch work that the self consistent field ( scf ) calculations of the source and drain leads are performed beforehand , and the calculated results are incorporated in the negf calculations through the self energy and the boundary conditions between the scattering region and leads . therefore , a careful treatment to handle the boundary conditions should be developed to avoid the spurious scattering . in this paper , to address the above two issues we present an accurate and efficient implementation of the negf method , in combination with dft using pseudo - atomic orbitals ( paos ) and pseudopotentials , using a contour integration method which is based on a continued fraction representation of the fermi - dirac function . for the accurate treatment of the boundary conditions between the scattering region and leads , we also develop a method for calculating the hartree potential by a combination of the two dimensional fast fourier transform ( fft ) and a finite difference method so that the boundary condition can be correctly reproduced . in addition , we discuss a careful treatment to construct the charge density near the boundaries . the efficiency and accuracy of our implementation are demonstrated by several numerical test calculations on convergence of the density matrix . this paper is organized as follows : in sec . ii the details of our implementation for treating the equilibrium state of the scattering region are discussed by focusing on the evaluation of the equilibrium density matrix , the treatment for constructing the charge density near the boundaries , and an efficient method for calculating the hartree potential . in sec . iii our implementation for the nonequilibrium state is described . in sec . iv we demonstrate the accuracy and efficiency of the implementation by a series of numerical calculations and several applications . in sec . v our implementation of the negf method is summarized . since most of practical aspects in the implementation of the negf method coupled with dft using the localized paos@xcite appear in the ground state calculation of the system at equilibrium , we start our discussion from the electronic structure calculation of the equilibrium ground state by using the green function method with dft . let us consider a system where one dimensional infinite cells are arranged with two dimensional periodicity as shown in fig . 1 . throughout the paper , we assume that the electronic transport along the * a*-axis is of interest , and that the two dimensional periodicity spreads over the * bc*-plane . the one dimensional infinite cell consists of the central region denoted by @xmath2 and the cells denoted by @xmath3 and @xmath4 , where @xmath5 . all the cells @xmath3 and @xmath4 , arranged semi - infinitely , contain the same number of atoms with the same structural configuration , respectively , but the cells @xmath3 and @xmath4 can be different from each other . in the equilibrium state with a common chemical potential everywhere in the system , the electronic structure of the system may be determined by dft.@xcite is an extended central region consisting of @xmath6 , @xmath7 , and @xmath8 . , width=302 ] due to the periodicity of the * bc*-plane the one - particle kohn - sham ( ks ) wave function in the system is expressed by the bloch function on the * bc*-plane using paos @xmath9 located on site @xmath10 as : @xmath11 where @xmath12 , @xmath13 , @xmath14 , and @xmath15 are indices for the spin , eigenstate , site , and basis orbital , respectively . the lattice vector @xmath16 and the bloch wave vector @xmath17 are given by @xmath18 , where @xmath19 and @xmath20 are the lattice vectors , and @xmath21 , where @xmath22 and @xmath23 are the reciprocal lattice vectors , respectively . the summation over @xmath14 and @xmath15 is considered for all the basis orbitals in the one dimensional infinite cell , which indicates no periodicity along the @xmath24-axis . considering the variation of the total energy , within the conventional dft , of the system expressed by the ks wave function eq . ( [ eq : c1 - 1 ] ) with respect to coefficients @xmath25 we obtain the following ks matrix equation : @xmath26 where @xmath27 is a column vector consisting of the coefficients @xmath28 . the hamiltonian @xmath29 and overlap matrices @xmath30 are given by @xmath31 where @xmath32 and @xmath33 are the hamiltonian and overlap matrix elements between two basis functions @xmath34 and @xmath35 , respectively . the overlap matrix arises from the non - orthogonality of the pao basis functions.@xcite now we consider an extended central region @xmath36 composed of the regions @xmath2 , @xmath37 , and @xmath38 as shown in figs . 1 ( a ) and ( b ) . the extension of the central region @xmath2 is made so that the relaxation of electronic structure around the interfaces between the leads , @xmath37 and @xmath38 , and the central region @xmath2 can be allowed . in addition , we impose two conditions : 1 . the localized basis orbitals @xmath39 in the region @xmath2 overlap with those in the regions @xmath37 and @xmath38 , but do not overlap with those in the regions @xmath40 and @xmath41 . the localized basis orbitals @xmath39 in the @xmath3 ( @xmath4 ) region has no overlap with basis orbitals in the cells beyond the nearest neighboring cells @xmath42 ( @xmath43 ) and @xmath44 ( @xmath45 ) . in our implementation the basis functions are strictly localized in real space because of the generation of basis orbitals by a confinement scheme.@xcite therefore , once the localized basis orbitals with specific cutoff radii are chosen for each region , the two conditions can be always satisfied by just adjusting the size of the unit cells for @xmath3 and @xmath4 . this is a benefit in the use of the strictly localized basis orbitals compared to other local basis orbitals such as slater- and gaussian - type orbitals . in the use of the strictly localized basis orbitals , the eigenvalue problem in the hilbert space spanned by the basis orbitals are solved without introducing any cutoff scheme . on the other hand , in the green function method the matrix elements of the hamiltonian and overlap matrices have to be truncated so as to satisfy the above two conditions in case of the other localized basis orbitals with the small but long tail . with the above two conditions ( i ) and ( ii ) , the hamiltonian matrix given by eq . ( [ eq : c1 - 3 ] ) is written by a block tridiagonal form as follows : @xmath46 where @xmath47 , @xmath48 , and @xmath49 are hamiltonian matrices of the central @xmath36 , left @xmath50 and right @xmath51 regions of which matrix size are the same as the number of basis orbitals @xmath52 , @xmath53 , and @xmath54 in the regions @xmath36 , @xmath40 , and @xmath41 , respectively . the other block components in eq . ( [ eq : c1 - 5 ] ) are the hamiltonian matrices connecting two regions among the regions , and these matrix sizes are deduced from those of the two regions . also the completely same structure is found in the overlap matrix . thus , the electronic structure of the system given by fig . 1(a ) can be obtained by solving the one dimensional block chain model , being @xmath17-dependent , given by fig . 1(b ) and the corresponding eq . ( [ eq : c1 - 2 ] ) . by noting @xmath55 and making use of the block tridiagonal form of the hamiltonian and overlap matrices , the green function of the central region @xmath36 can be written by @xmath56 with self energies @xmath57 and @xmath58 defined by @xmath59 where @xmath60 and @xmath61 are surface green functions of the left and right regions . it is worth pointing out that there is a total energy functional which can be variationally minimized with respect to charge density @xmath62 if eq . ( [ eq : c1 - 6 ] ) is self - consistently solved . the details of derivation for the functional is given in appendix a. in general , the surface green function @xmath63 is defined by @xmath64 , where @xmath65 and @xmath66 are the hamiltonian and overlap matrices for the lead regions , and the suffix , @xmath67 , is @xmath68 or @xmath69 . it is noted that due to the two conditions ( i ) and ( ii ) mentioned above the hamiltonian and overlap matrices for the lead regions can be written by a block tridiagonal form as follows : @xmath70 where the hamiltonian can be spin and @xmath17-dependent , while the indices for them are omitted for simplification of the notation , and also the index @xmath14 appearing in @xmath71 corresponds to the cell number for the lead @xmath72 or @xmath73 . it seems to be difficult to directly diagonalize the ks equation eq . ( [ eq : c1 - 2 ] ) for the one dimensional block chain model because of the infinite dimension of the matrices . however , instead by focusing on only the central region one can evaluate the green function of the central region as a rather small problem of @xmath74 in size . the effect of semi infinite regions @xmath68 and @xmath69 are included through the corresponding self energies @xmath57 and @xmath58 . in order to practically calculate the green function of the central region given by eq . ( [ eq : c1 - 6 ] ) , we introduce an approximation where the regions @xmath75 are all equivalent to each other with respect to the spatial charge distribution , the ks hamiltonian , and the relevant density matrix which are calculated in advance by adopting the system of which unit cell is @xmath40 and by using the conventional band structure calculation . the same approximation also applies for the regions @xmath76 . strictly speaking , the assumption is not correct , since the charge distribution must be affected by the interaction between the central region @xmath36 and the regions @xmath3 ( @xmath4 ) . however , if the size of the unit vector along the @xmath24-axis for the regions @xmath37 and @xmath38 in the extended central region @xmath36 is large enough , the assumption will be asymptotically correct as the unit vector becomes larger . the approximation enables us to evaluate the surface green function by the iterative method.@xcite the efficient iterative scheme can be performed by the following procedure : @xmath77 with a set of initial values @xmath78 by the iterative calculation , in most cases the inverse of @xmath79 rapidly converges at a part of the surface green function : @xmath80 where @xmath81 is the ( 1,1 ) block element of the surface green function @xmath82 . the ( 1,1 ) block element @xmath81 of @xmath83 ( or @xmath84 ) in size has all the necessary information to calculate the self - energies @xmath57 and @xmath58 , since there is no contribution from the other block elements because of the two conditions ( i ) and ( ii ) mentioned above . in practice , the convergence in the iterative calculation is very fast and the frobenius norm , defined by @xmath85 , of @xmath86 ( ev ) is obtained by typically only 7 iterations . one of practical difficulties in the implementation of the green function method is how the equilibrium density matrix is evaluated efficiently and accurately.@xcite in our implementation , the equilibrium density matrix is highly efficiently computed using the contour integration method with a special treatment of the fermi - dirac function @xmath87.@xcite if the hamiltonian and overlap matrices associated with eq . ( [ eq : c1 - 6 ] ) are @xmath17-dependent , it turns out that the spectrum function in the lehmann representation of the central green function is complex number in general as discussed in appendix b of ref . then , the density matrix @xmath88 , where one of the associated basis orbitals is in the central cell and the other is in the cell denoted by @xmath16 , is given by making use of both the retarded and advanced green functions @xmath89 and @xmath90 as @xmath91 with @xmath92 where @xmath93 is the volume of the unit cell , @xmath94 represents the integration over the first brillouin zone , @xmath95 a positive infinitesimal , and @xmath96 a chemical potential . the integration over @xmath17-space is numerically performed by using the monkhorst - pack mesh.@xcite it is also noted that the phase factor @xmath97 appears through eq . ( [ eq : c1 - 1 ] ) . if the hamiltonian and overlap matrices are @xmath17-independent , eq . ( [ eq : c1 - 22 ] ) can be simplified into a well - known formula : @xmath98 . \label{eq : c1 - 24}\end{aligned}\ ] ] in case of the @xmath17-independent problem , the simplified formula is used , since the number of the evaluation of the green function is reduced by half . for the efficient integration in eq . ( [ eq : c1 - 23 ] ) , in our implementation the fermi - dirac function is expressed by a continued fraction representation derived from a hypergeometric function@xcite so that the structure of poles can be suitable for the integration associated with the green function as follows : @xmath99 where @xmath100 with @xmath101 , @xmath102 is electronic temperature , @xmath103 and @xmath104 are complex variables . also , @xmath105 and @xmath106 are the poles and the associated residues of the continued fraction representation eq . ( [ eq : c1 - 25 ] ) , which are obtained via an eigenvalue problem derived from eq . ( [ eq : c1 - 25]).@xcite since all the @xmath105 are real numbers , the poles @xmath107 are located on the imaginary axis . one may find an interesting distribution of the poles on the complex plane that the interval between neighboring poles are uniformly located up to about 61 % of the total number of poles on the half complex plane with the same interval @xmath108 , and from then onward it increases very rapidly as the distance between the pole and the real axis increases . the structure of the poles in eq . ( [ eq : c1 - 25 ] ) allows us to efficiently evaluate eq . ( [ eq : c1 - 23 ] ) because of the asymptotic change , @xmath109 , of the green function in the faraway region of the real axis . in other words , the denser poles are allocated for the rapidly varying range of the green function , and the coarser for smoothly . in addition , there is no ambiguity in the choice of the path in the contour integration unlike the other schemes,@xcite which is one of advantages in our method . by terminating the summation of eq . ( [ eq : c1 - 25 ] ) at a finite number of poles @xmath110 , the integration of eq . ( [ eq : c1 - 23 ] ) can be performed by the contour integration where the poles in the upper and lower half planes are taken into account for the terms with plus and minus signs in eq . ( [ eq : c1 - 23 ] ) , respectively , and the explicit formula is given by @xmath111 where @xmath112 , and @xmath113 is the zeroth order moment of the green function @xmath114 . the zeroth order moment @xmath113 is easily calculated by @xmath115 which can be derived from the moment representation of the green function,@xcite where @xmath69 is a large real number and in this study @xmath116 ( ev ) is used in order to make the higher order moments negligible . although the number of poles in the summation of eq . ( [ eq : c1 - 26 ] ) required for the sufficient convergence depends on the electronic temperature @xmath102 , the fully convergent result within double precision is achieved by the use of only 100 poles in case of @xmath117 as shown later . if forces on atoms are calculated based on the conventional dft scheme using the non - orthogonal basis orbitals,@xcite the evaluation of the energy density matrix @xmath118 is needed . in appendix b , we derive the calculation scheme of the equilibrium energy density matrix @xmath119 based on the contour integration method . even though the basis functions we used are strictly localized in real space , there is the nonnegligible contribution for the charge density near the boundary between the central and lead regions from the basis functions located in the lead regions . note that any treatment for the contribution to the charge density has not been clarified in the other implementations.@xcite thus , we carefully calculate the charge density in the central region by considering three contributions : @xmath120 where the suffix , s , is @xmath68 or @xmath69 , and @xmath121 , @xmath122 , and @xmath123 are the charge densities contributed from the basis functions located in the central , the lead and central , and the lead regions , respectively . note that the summation over s is not required in eq . ( [ eq : c1 - 43 ] ) because of the conditions ( i ) and ( ii ) . each charge contribution is explicitly given by @xmath124 where * a * is the lattice vector of the unit cell for the @xmath37 or @xmath38 region along the * a*-axis . the displacement of @xmath125 ( @xmath126 ) denotes that the basis function is placed in the @xmath40 ( @xmath41 ) region in the configuration shown in fig . 1 . the charge density given by eq . ( [ eq : c1 - 44 ] ) is calculated by the equilibrium density matrix @xmath127 given by eq . ( [ eq : c1 - 22 ] ) . although eq . ( [ eq : c1 - 44 ] ) can give a finite electron density on the outside of the central cell with @xmath128 because of the overlap of basis functions , the contribution is reflected in the central cell with @xmath128 by considering the periodicity on the * bc*-plane . in the nonequilibrium case , the equilibrium density matrix is only replaced by the nonequilibrium density matrix which will be discussed in the next section . each term in the summations for the two contributions @xmath122 and @xmath123 survive only if the overlap of the associated two basis orbitals is not zero in the central region . since the original central region @xmath2 is extended by adding the @xmath37 and @xmath38 regions , it is expected that the density matrix elements @xmath129 and @xmath130 in eqs . ( [ eq : c1 - 45 ] ) and ( [ eq : c1 - 46 ] ) are close to those of the leads in the equilibrium condition . therefore , the density matrix elements of the leads calculated by the conventional band structure calculations are used for eqs . ( [ eq : c1 - 45 ] ) and ( [ eq : c1 - 46 ] ) . due to the treatment the charge densities @xmath122 and @xmath123 are independent of the scf iteration so that for the computational efficiency they can be computed on a numerical mesh and stored before the scf iteration . the factor 2 for @xmath122 in eq . ( [ eq : c1 - 43 ] ) is due to taking account of the contribution from @xmath131 , while the factor does not appear for @xmath123 , since all the paired terms are included by the double summation in eq . ( [ eq : c1 - 46 ] ) . we add a note that the same consideration has to be applied even for the calculation of the density of states ( dos ) . in this case , the contribution from off - diagonal block green functions connecting the central and lead regions should be added to the dos of the central region @xmath36 calculated by @xmath132 . the off - diagonal block green functions can be calculated from @xmath132 and the surface green functions , @xmath133 and @xmath134 , by making use of the identity @xmath55 as follows : @xmath135 where s is @xmath68 or @xmath69 . the hartree potential in the central region is calculated under the boundary condition that the hartree potential at the boundary between the central @xmath36 and @xmath136 regions is same as that of the lead , where the hartree potential in both the lead regions is calculated using the conventional band structure calculation before the calculation of the infinite @xmath137 in fig . 1(b ) using the green function . in our implementation , the hartree potential for the central region with the boundary condition is efficiently evaluated by a combination of the two dimensional fft and a finite difference method , while other schemes are used in the other implementations.@xcite the majority part of the hartree potential in our treatment is accurately calculated by considering the neutral atom potential which is the sum of the local potential of the pseudopotential and the hartree potential by the confined charge for the neutralization.@xcite the neutral atom potential depends on only the atomic structure and atomic species , and has no relation with the boundary condition . the effect of the relaxation of charge distribution on the hartree potential is taken into account by the remaining minority part of the hartree potential @xmath138 given by @xmath139 where @xmath140 is defined by the difference between the electron density @xmath141 calculated by the green function and the atomic electron density@xcite @xmath142 calculated by superposition of each atomic electron density @xmath143 at atomic site @xmath14 as follows : @xmath144 the fourier transformation of eq . ( [ eq : c1 - 35 ] ) on the yz plane , corresponding to the * bc*-plane depicted in fig . 1(a ) , yields @xmath145 where @xmath146 with integer numbers @xmath147 and @xmath148 . by approximating the second derivative in eq . ( [ eq : c1 - 37 ] ) with the simplest finite difference , @xmath149 we obtain a simultaneous linear equation @xmath150 of @xmath151 in size with a tridiagonal matrix defined by @xmath152 and a vector defined by @xmath153 where @xmath154 is the interval between neighboring points @xmath155 and @xmath156 , @xmath62 runs from @xmath157 to @xmath158 , and @xmath159 and @xmath160 are the boundary conditions . since the lattice vector of the extended central region @xmath36 along the * a*-axis is divided by @xmath161 for the discretization , the positions @xmath162 and @xmath163 are situated at the boundary , along the * a*-axis , of the unit cell of the central region @xmath36 . thus , @xmath159 and @xmath160 can be calculated from the hartree potential at the left boundaries along the * a*-axis of the left and right leads , respectively . after solving the simultaneous linear equations for all the @xmath164 points , the inverse two dimensional fft on the * bc*-plane yields the difference hartree potential @xmath138 under the boundary conditions . it is apparent that there is no ambiguity for the inclusion of the boundary conditions in the method . although the treatment can be easily extended to the higher order finite difference to the second derivative , we restrict ourselves to the simplest case in the implementation because of its sufficient accuracy . employing fft for two dimensional fourier transformation , the whole computational effort to solve the poisson equation eq . ( [ eq : c1 - 35 ] ) under the boundary conditions is estimated to be @xmath165 which is slightly superior to that of the three dimensional fft , where @xmath161 , @xmath166 , and @xmath167 are the number of meshes for the discretization along the @xmath24- , @xmath19- , and @xmath20-axes . our implementation of the negf method with dft is based on the strictly localized paos@xcite and a norm - conserving pseudopotential method.@xcite within the scheme , the calculation of the matrix elements such as the overlap and kinetic energy integrals , consisting of two center integrals , is performed using a fourier transform method,@xcite while the other matrix elements for @xmath168 and @xmath138 , which can not be decomposed into two center integrals , are evaluated by the numerical integration on the regular mesh in real space.@xcite the further details on how the elements of the hamiltonian and overlap matrices are calculated can be found in ref . @xcite . in addition to the above evaluation of the hamiltonian and overlap matrix elements , the hamiltonian matrix elements associated with the basis orbitals situated at near the boundary are treated in a special way as explained below . if the tails of two basis orbitals located on atoms in the central region @xmath36 go beyond the boundary between the central and the lead regions , the associated hamiltonian matrix element is replaced by the corresponding element in the lead region calculated by the conventional band structure calculation . the case can happen only if the two basis orbitals are located in the region @xmath37(@xmath38 ) because of the condition ( i ) so that the replacement of the hamiltonian matrix element can be always possible . the replacement is made by assuming that the potential profile near the boundary is similar to that near the boundary between the regions @xmath169 and @xmath170 , and can be justified if the size of the region @xmath171 is large enough . compared to conventional band structure calculations , it seems that the negf method tends to suffer from difficulty in obtaining the scf convergence . our observation in several cases suggests that the difficulty may come from charge sloshing along the * a*-axis during the scf iteration . the difference hartree potential @xmath138 change largely by imposition of the boundary condition even for a small variation in the charge density distribution , resulting in a serious charge sloshing along the * a*-axis . thus , we consider suppression of the charge sloshing along the * a*-axis by introducing the following weight factor @xmath172 : @xmath173 where @xmath174 is a smaller one of either @xmath175 or @xmath176 , and @xmath177 and @xmath178 are adjustable parameters , while keeping @xmath179 . the prefactor @xmath180 is given by @xmath181 with definitions : @xmath182 where @xmath183 is an adjustable parameter . noting that @xmath184 is the number of difference electron density of each layer indexed by @xmath185 , and that the coulomb potential induced by each charged layer depends linearly on the distance from the layer , one can notice that @xmath186 is proportional to the electric field at position @xmath14 . therefore , eq . ( [ eq : b1 - 1 ] ) takes charge density under a large electric field into significant account in addition to the suppression of the charge sloshing in the * bc*-plane in a sense by the kerker method.@xcite in our implementation , the weight factor given by eq . ( [ eq : b1 - 1 ] ) is combined with the kerker method@xcite and the residual minimization method in a direct inversion iterative subspace ( rmm - diis)@xcite with substantial improvement . a technical remark should also be added to avoid a local trap problem in the scf calculation . in systems having a long * a*-axis , a unphysical charge distribution , corresponding to a large charge separation in real space , tends to be obtained even after achieving the self consistency . in this case the hamiltonian of the central region @xmath36 at the first scf iteration , which are calculated via superposition of atomic charge density , is far from the self - consistently converged one , while the hamiltonian matrix used in the calculation for the self - energy is determined in a self - consistent manner beforehand . the inconsistency between the two matrices tends to produce a unphysical charge distribution at the first scf iteration . once the situation happens at the first scf iteration , in many cases the electronic structure keeps trapped during the subsequent scf iteration , which is a serious problem in practical applications . however , the local trap problem can be overcome by a simple scheme that the first few scf iterations are performed by using the conventional band scheme and then onward the solver is switched from the band scheme to the negf method . in the band structure calculation for the first few iterations , such a unphysical charge distribution does not appear due to no self - energy involved . in most cases we find that the simple scheme works well to avoid the local trap problem in the scf convergence . based on the negf theory mainly developed by schwinger@xcite and keldysh@xcite , the density matrix in the nonequilibrium state of the central region is evaluated by@xcite @xmath187 in addition to the equilibrium density matrix @xmath88 given by eq . ( [ eq : c1 - 22 ] ) , a correction term defined by @xmath188 is taken into account , where @xmath189 is defined by @xmath190 with @xmath191 and @xmath192 either the left @xmath193 or right chemical potential @xmath194 which is lower than the other is used for the calculation of the equilibrium density matrix in eq . ( [ eq : c1 - 40 ] ) . as well , in eqs . ( [ eq : c1 - 42 ] ) and ( [ eq : c1 - 52 ] ) the chemical potentials are given by the rule that @xmath195 and @xmath196 if @xmath197 , and @xmath198 and @xmath199 if @xmath200 . starting from the negf theory , the formula , eq . ( [ eq : c1 - 40 ] ) , may be derived by introducing two assumptions . the first assumption is that the occupation of the wave functions incoming from the left ( right ) region still obeys the fermi - dirac function with the left ( right ) chemical potential even in the central region . the assumption can be justified within at least the one - particle picture , since the same result can be obtained from the lippmann - schwinger equation for a non - interacting system.@xcite the second assumption is that in the central region the states in the energy regime below the lower chemical potential is in equilibrium due to the other physical obstacles such as the electron - phonon@xcite and electron - electron interactions,@xcite which are not considered explicitly in our implementation , although the definite role of those obstacles is obscure as for the occupation of the states . the second assumption allows electrons to occupy in highly localized states below the lower chemical potential through the first term of eq . ( [ eq : c1 - 40 ] ) . only the states in the energy regime between two chemical potentials are treated as in the nonequilibrium condition in which the contribution of the wave functions incoming from the lead with the higher chemical potential is taken into account to form the correction term given by eq . ( [ eq : c1 - 41 ] ) . the integrand in eq . ( [ eq : c1 - 42 ] ) is not analytic apart from the real axis , since the integrand is a function of both @xmath201 and @xmath202 . thus , one can not apply the contour integration method that we use for the equilibrium density matrix . instead , a simple rectangular quadrature scheme is applied to the integration of eq . ( [ eq : c1 - 42 ] ) on the real axis with a small imaginary part @xmath203 . since the integrand contains the difference between two fermi - dirac functions , the energy range for the integration can be effectively reduced to a narrow range that the difference is larger than a threshold , where the threshold of @xmath204 is used in this study . with the threshold and the step width of 0.01 ( ev ) , the number of meshes on the real axis is 152 for @xmath205 ( ev ) at @xmath206 . the convergence speed depends on the shape of the integrand and how large @xmath203 is employed for smearing the integrand , which will be discussed later . the source - drain bias voltage applied to the left and right leads is easily incorporated by adding a constant electric potential @xmath207 to the hartree potential in the right lead region . the effect of the bias voltage appears at three places . the first effect is that the hamiltonian matrix in the right region given by eq . ( [ eq : c1 - 9 ] ) is replaced using eq . ( [ eq : c1 - 10 ] ) as @xmath208 this can be easily confirmed by noting that the matrix elements for the constant potential @xmath207 is @xmath209 . the off - diagonal block elements @xmath210 , @xmath211 @xmath212 , and @xmath213 appearing eqs . ( [ eq : c1 - 7 ] ) and ( [ eq : c1 - 8 ] ) are also replaced in the same way . the second effect is that the chemical potential of the right lead is replaced as @xmath214 the treatment is made so that the first replacement can be regarded as just shifting the origin of energy in the right lead . the last effect is that the boundary condition in eq . ( [ eq : c1 - 39 ] ) is replaced as @xmath215 where @xmath216 is calculated by the fourier transformation on the * bc*-plane for @xmath138 at @xmath163 plus @xmath207 . since only the difference of the bias voltages applied to the left and right leads affects the result , one can consider the replacements on only the right lead at the three places as shown above . it is noted that the replacement by eq . ( [ eq : c1 - 55 ] ) corresponds to adding a linear potential @xmath217 to the hartree potential in the central region @xmath36 , where @xmath218 and @xmath219 are determined by the boundary conditions : @xmath216 and @xmath220 . in our implementation , the gate voltage @xmath221 is treated by adding an electric potential defined by @xmath222 , \label{eq : c1 - 56}\end{aligned}\ ] ] where @xmath223 is a constant value corresponding to the gate voltage , @xmath224 the center of the region @xmath2 , and @xmath225 the length of the unit vector along * a*-axis for the region @xmath2 . due to the form of eq . ( [ eq : c1 - 56 ] ) , the applied gate voltage affects mainly the region @xmath2 in the central region @xmath36 . the electric potential may resemble the potential produced by the image charges.@xcite the spin resolved transmission is evaluated by the landauer formula for the non - interacting central region @xmath36 connected with two leads : @xmath226 where @xmath227 is the spin and @xmath228-resolved transmission defined by @xmath229 \label{eq : c1 - 58}.\end{aligned}\ ] ] using the transmission formula , the current is evaluated by @xmath230 the formula can be derived by starting from a more general formula of the current for the interacting central region @xmath36 and by replacing the involved green functions with the non - interacting green functions as shown by meir and wingreen.@xcite we perform the integration in eq . ( [ eq : c1 - 59 ] ) on the real axis with a small imaginary part @xmath203 by the same way as for the nonequilibrium density matrix of eq . ( [ eq : c1 - 42 ] ) . all the calculations in this study were performed by the dft code , openmx.@xcite the paos centered on atomic sites are used as basis functions.@xcite the pao basis functions we used , generated by a confinement scheme,@xcite are specified by h5.5-s2 , c4.5-s2p2 , o5.0-s2p2d1 , fe5.0-s2p2d1 , mg5.5-s2p2 , la7.0-s3p2d1f1 , sr7.0-s3p2d1f1 , and mn6.0-s3p2d2 , where the abbreviation of basis functions , such as c4.5-s2p2 , represents that c stands for the atomic symbol , 4.5 the cutoff radius ( bohr ) in the generation by the confinement scheme , and s2p2 means the employment of two primitive orbital for each of s- and p - orbitals . norm - conserving pseudopotentials are used in a separable form with multiple projectors to replace the deep core potential into a shallow potential.@xcite also , a local density approximation ( lda ) to the exchange - correlation potential is employed.@xcite , while a generalized gradient approximation ( gga)@xcite is used only for calculations of the lamno@xmath0srmno@xmath1 superlattice . the real space grid techniques are used with the cutoff energies of 120 - 200 ry in numerical integrations and the solution of poisson equation using fft.@xcite in addition , the projector expansion method is employed in the calculation of three - center integrals for the deep neutral atom potentials.@xcite ) per atom of a linear carbon chain with a bond length of 1.4 under the zero bias voltage at electronic temperature of 300 , 600 , and 1200 k , where the regions @xmath37 , @xmath38 , and @xmath2 contain four carbon atoms , respectively . for comparison the same calculation _ ( closed contour ) _ using a closed contour method@xcite is also shown for t=600 k. the definitions of @xmath231 and @xmath232 are found in appendix a. the reference values are obtained from calculations with a large number of poles . ( b ) total dos of the carbon linear chain , calculated by the conventional band structure calculation ( solid line ) and the egf method ( dotted line ) , under the zero bias voltage at 300 k , where 160 poles are used for the integration of the equilibrium density matrix . it is hard to distinguish two lines due to the nearly equivalent results.,width=302 ] the accuracy and efficiency of the implementation are mainly determined by the evaluation of density matrix given by eq . ( [ eq : c1 - 40 ] ) which consists of two contributions : the equilibrium and non - equilibrium terms given by eqs . ( [ eq : c1 - 22 ] ) and ( [ eq : c1 - 41 ] ) , respectively . in this subsection , we discuss the convergence properties of the equilibrium and non - equilibrium terms in the density matrix as a function of numerical parameters . ) per atom of the same linear carbon chain as in fig . 2 , calculated by various imaginary parts , under a finite bias voltage of 0.5 v , where the other calculation conditions are same as for fig . the number pointed by the arrow denotes a grid spacing ( ev ) corresponding to the number of grid points . the reference values are obtained from calculations with a large number of grid points . ( b ) mulliken populations in the carbon chain . the sequential numbers 1 and 12 correspond to the most left and right hand side atoms in the central region @xmath36 , respectively.,width=302 ] since the majority part of the density matrix given by eq . ( [ eq : c1 - 40 ] ) is the equilibrium contribution , let us first discuss convergency of the equilibrium density matrix as a function of poles . the absolute error in @xmath233 of a carbon linear chain is shown in fig . 2(a ) as a function of the number of poles in order to illustrate the convergence property for the equilibrium density matrix under zero bias voltage , where @xmath233 can be regarded as a conventional expression of the total energy in dft , and the definitions of the two energy terms @xmath231 and @xmath232 with the fermi - dirac function are given in appendix a. for comparison the result calculated by a closed contour method is also shown.@xcite one can see that the accuracy of @xmath234 ev per atom is obtained using 140 , 100 , and 70 poles for the electronic temperature of 300 , 600 , and 1200 k , respectively , while about 400 energy points are needed to obtain the same accuracy using the closed contour method at 600 k.@xcite for most cases , we find that the convergence rate is similar to the case shown in fig . in general , the number of poles to achieve the accuracy of @xmath234 ev per atom must be proportional to the inverse of @xmath102 , since the interval between neighboring poles of the continued fraction given by eq . ( [ eq : c1 - 25 ] ) is scaled by @xmath235 . this fact implies that the computational effort increases as the electronic temperature decreases . however , we generally use electronic temperature from 300 to 1000 k for practical calculations , which means that the use of 100 poles is enough for practical purposes . therefore , it can be concluded that the most contribution of the density matrix can be very accurately evaluated with a small number of poles , i.e. , 100 . to demonstrate the proper treatment of the boundary between the lead and the central regions in our implementation , in fig . 2(b ) we show a comparison between the conventional band structure and the egf calculations with respect to dos of the carbon linear chain . the comparison provides a severe test to check whether the egf method is properly implemented or not . it can be confirmed that dos calculated by the egf method is nearly equivalent to that by the conventional band structure calculation , which clearly shows the proper treatment of the boundary between the lead and the central regions in our scheme . as explained before , the integration of eq . ( [ eq : c1 - 42 ] ) required for the evaluation of the nonequilibrium term in the density matrix has to be performed on the real axis with a small imaginary part because contour integration schemes may not be applied due to the non - analytic nature of the integrand . the treatment might suffer from numerical instabilities in the scf iteration , since the integrand can rapidly vary due to the existence of poles of green function located on the real axis . a remedy to avoid the numerical problem is to smear the green function by introducing a relatively large imaginary part.@xcite to investigate convergency of the nonequilibrium term in the density matrix , in fig . 3(a ) we show the absolute error in @xmath233 of the same infinite carbon chain as in fig . 2 but under a finite bias voltage of 0.5 ev as a function of the number of regular grid points used for the evaluation of the nonequilibrium term in the density matrix given by eqs . ( [ eq : c1 - 41 ] ) and ( [ eq : c1 - 42 ] ) . we also tested the gauss - legendre quadrature for the integration of the nonequilibrium term besides the integration using the regular grid , but found that the convergence rate of the gauss - legendre quadrature is rather slower than the simple scheme possibly due to the spiky structure of the integrand . thus , we have decided to use the simple scheme using the regular grid . as expected , it turns out that the number of grid points to achieve the accuracy of @xmath234 ev per atom increases as the imaginary part becomes smaller . however , the accuracy of @xmath234 ev is attainable using about 100 grid points in case of the imaginary part of 0.01 ev , while a few thousands grid points have to be used to achieve the same accuracy for the imaginary part of 0.0001 ev . although the accuracy of @xmath234 ev can be achieved by introducing the smearing scheme , however , one may consider that results can be affected by the introduction of an imaginary part . in order to find a compromise between the accuracy and efficiency , the mulliken population of the carbon linear chain under the finite bias voltage of 0.5 ev is shown in fig . we see that the use of the imaginary part of 0.01 ev gives a result comparable to that obtained by the use of 0.0001 ev . thus , the imaginary part of 0.01 ev can be a compromise between the accuracy and efficiency in this case . as a result , one can find that the energy points of 200 ( 100 and 100 for the equilibrium and nonequilibrium density matrices , respectively ) are enough to achieve the accuracy of @xmath234 ev per atom in @xmath233 at 600 k. however , it should be mentioned that a proper choice of the imaginary part must depend on the electronic structure of systems , and the careful consideration must be taken into account especially for a case that the spiky dos appears in between two chemical potentials of the leads , while for the equilibrium part of the density matrix the convergence property is insensitive to the electronic structure . we also note that the accurate evaluation of the density matrix makes the scf calculation stable even under a finite bias voltage . in fact , for the case with the bias voltage of 0.5 v the number of the scf iterations to achieve the residual norm of @xmath236 for the charge density difference is 29 which is nearly equivalent to that , 30 , for the zero bias case . since for large - scale systems it is very time - consuming to perform the scf calculation at each bias voltage , here we propose an interpolation scheme to reduce the computational cost in the calculations by the negf method . the interpolation scheme is performed in the following way : ( i ) the scf calculations are performed for a few bias voltages which are selected in the regime of the bias voltage of interest . ( ii ) when the transmission and current are calculated , a linear interpolation is made for the hamiltonian block elements , @xmath47 and @xmath237 , of the central scattering region and the right lead , and the chemical potential , @xmath238 , of the right lead by @xmath239 where the indices @xmath157 and @xmath240 in the superscript mean that the quantities are calculated or used at the corresponding bias voltages where the scf calculations are performed beforehand . note that it is also possible to perform the interpolation for @xmath17-independent hamiltonian matrix elements instead of eqs . ( [ eq : c1 - 60 ] ) and ( [ eq : c1 - 61 ] ) . in general , @xmath241 should range from 0 to 1 for the moderate interpolation . a comparison between the fully self consistent and the interpolated results is shown with respect to the current and transmission in the linear carbon chain in figs . 4(a ) and ( b ) . in this case , the scf calculations at three bias voltages of 0 , 0.5 , and 1.0 v are performed , and the results at the other bias voltages are obtained by the interpolation scheme . for comparison we also calculate the currents via the scf calculations at all the bias voltages . it is confirmed that the simple interpolation scheme gives notably accurate results for both the calculations of the current and transmission . although the proper selection of bias voltages used for the scf calculations may depend on systems , the result suggests that the simple scheme is very useful to interpolate the effect of the bias voltage while keeping the accuracy of the calculations . v. the zigzag edges are terminated by hydrogen atoms . the isosurface value of @xmath242 is used for drawing the spin density.,width=302 ] as an illustration of our implementation , we investigate transport properties of zigzag graphene nanoribbons ( zgnrs ) with different magnetic configurations . a characteristic feature in the band structure of zgnr is the appearance of flat bands around x - point near the fermi level , resulting in spin - polarization of associated states located at the zigzag edges.@xcite thus , so far several intriguing transport properties have been theoretically predicted especially for zgnrs among gnrs by focusing on the spin polarized edge states.@xcite for instance , it is found that zgnrs might exhibit an extraordinary large magnetoresistance ( mr ) effect and a spin polarized current.@xcite -@xmath207 curves for ( a ) the up - spin and ( b ) the down - spin states in the fm junction , and ( c ) the up - spin and ( d ) the down - spin states in the afm junction of 7- and 8-zgnrs.,width=302 ] here we focus on the current - bias voltage ( @xmath243-@xmath207 ) characteristic of @xmath244- and @xmath245-zgnrs with two magnetic configurations : ferromagnetic ( fm ) and antiferromagnetic ( afm ) junctions as shown in figs . 5(a ) and ( b ) , respectively , where the number , @xmath244 or @xmath245 , is the number of carbon atom in the sublattice being across zgnr along the lateral direction . the odd and even cases will also be referred to as _ asymmetric _ and _ symmetric _ , respectively . the extended central region @xmath36 consists of one sublattice and four unit cells , and contains 72 and 82 atoms for 7- and 8-zgnrs , respectively . the poles of 100 is used for the evaluation of the equilibrium density matrix with the electronic temperature of 300 k , while the nonequilibrium term in the density matrix is evaluated using the simple quadrature method with the imaginary part of 0.01 ev and the grid spacing of 0.02 ev . the geometric structures used are optimized under the periodic boundary condition until the maximum force is less than @xmath246 hartree / bohr . at each bias voltage the electronic structure of zgnr is self - consistently determined . for the fm junction of 8-zgnr under various applied source - drain bias voltages @xmath207.,width=302 ] figures 6(a ) and ( b ) show the current - voltage ( @xmath243-@xmath207 ) curves for the up- and down - spin states in the fm junctions of 7- and 8-zgnrs . it is found that the current for 7-zgnr linearly depends on the bias voltage , while the current for 8-zgnr is saturated at the bias voltage of about @xmath247 . the distinct behavior of 8-zgnr from 7-zgnr can be more definitely seen in the afm junction as shown in figs . 6(c ) and ( d ) . interestingly , 8-zgnr with the afm junction exhibits a diode behavior for the spin resolved current . only the up - spin state contributes substantially to the current in the negative bias regime . in contrary in the positive bias regime only the down - spin state contributes to the current . it is worth mentioning that the effect can also be regarded as a spin filter effect . on the other hand , the @xmath243-@xmath207 characteristics for 7-zgnr is nearly equivalent to that for the fm junction . the considerable difference between 7- and 8-zgnrs in the @xmath243-@xmath207 characteristics can be attributed to the symmetry of two wave functions , @xmath248 and @xmath249 states , around the fermi level . for 8-zgnr the wave functions of the @xmath248 and @xmath249 states are antisymmetric and symmetric with respect to the @xmath12 mirror plane which is the mid plane between two edges , respectively , while those wave functions for 7-zgnr are neither symmetric nor antisymmetric . from a detailed analysis@xcite , it can be concluded that the unique distinction in the @xmath243-@xmath207 characteristics arises from an interplay between the symmetry of wave functions and band structures of zgnrs . in addition , we find that spin moments are reduced by applying the finite bias voltage as shown in fig . since the flat bands around x - point of the minority spin state are located about 0.25 ev above the fermi level , the spin moments at the zigzag edges are largely reduced by increasing the occupation of the flat bands for the minority spin states when the bias voltage exceeds the threshold as shown in fig . 7.@xcite the details of the analysis on the unique spin diode and filter effect of zgnrs is discussed elsewhere.@xcite region , sandwiched by fe(100 ) , being the @xmath37 and @xmath38 regions . the red large and small and blue circles denote fe , o , and mg atoms , respectively . ( b ) the net charge and spin magnetic moment of a fe or mg atom belonging to each layer in the parallel magnetic configuration between the left and right leads . the position in the horizontal axis exactly corresponds to that of the layer in the above figure.,width=321 ] the applicability of our implementation to bulk systems is demonstrated by an application to a tunneling junction consisting of mgo(100 ) layers sandwiched by iron . the magneto - tunneling junction was theoretically predicted to exhibit a large tunneling - magnetoresistance ( tmr),@xcite and subsequently the tmr effect has been experimentally confirmed.@xcite we consider four mgo(100 ) layers sandwiched by iron with the ( 100 ) surface of which atomic configuration is shown in fig . 8(a ) , where the lattice constant of the * bc*-plane used is 2.866 , and they are 2.866 and 4.054 in iron and mgo regions along the * a*-axis , and the distance between the mgo and iron layers is assumed to be 2.160 . the four mgo(100 ) layers corresponds to the region @xmath2 in fig . 1(a ) , and four fe layers of the left and right hand correspond the region @xmath37 and @xmath38 , respectively . the scf calculations were performed at 1000 k under zero bias voltage using * k*-points of @xmath250 and 130 poles for the integration of the equilibrium density matrix . it is found that obtaining the scf is much harder than the case of zgnr discussed before , and that a careful and modest treatment for the charge mixing is required . were used.,width=302 ] as shown in fig . 8(b ) , the net charge of iron atoms in the interfacial layer is positive due to the coordination to an oxygen atom , and the reduction of electrons leads to the increase of the magnetic moment of iron atoms at the interfacial layer . the increase of the magnetic moment at the interfacial layer makes the distinction of the majority and minority spin states at the fermi energy clear , i.e. , the nature of the majority and minority spin states at the fermi energy can be assigned to @xmath251- and @xmath225-states , respectively . the * k*-resolved transmissions at the chemical potential for the majority and minority spin states for the parallel magnetic configuration between the left and right leads are shown in figs . 9(a ) and ( b ) , respectively . the large peak at the @xmath252 point in the majority spin state can be attributed to the @xmath251-state , while sharp peaks around four pillars come from the @xmath225-state as discussed in ref.@xcite . note that the position of the sharp peaks is rotated by 45 degrees because of the unit cell rotated by 45 degrees compared to that in ref.@xcite . the conductances , @xmath253 and @xmath254 , for the majority and minority spin states , calculated from the average transmission integrated over the first brillouin zone , are 11.99 and 2.82 @xmath255 , respectively , which implies that the tunneling junction may behave as a spin filter . the distinction in the conductance should be attributed to decay properties of states in the insulating mgo region coupled with the two states.@xcite for the antiparallel magnetic configuration the * k*-resolved transmission at the chemical potential is understood as a multiplication of the transmissions for the majority and minority spin states in the parallel configuration as shown in fig . the conductance , @xmath256 , of the antiparallel magnetic configuration is 0.34 ( @xmath257 ) which is smaller than those of the parallel case . by defining tmr@xmath258 , we obtain tmr of 2082 % , which is compared to 3700 % for a 5 layer mgo case reported in ref.@xcite . when the transmission of a system with the periodicity along the * a*-axis as well as the periodicity of the * bc*-plane is evaluated under zero bias voltage , it can be easily obtained by making use of the hamiltonian calculated by the conventional band structure calculation without employing the green function method described in the paper . this scheme enables us to explore transport properties for a wide variety of possible geometric and magnetic structures with a low computational cost , and thereby can be very useful for many materials such as supperlattice structures . once the hamiltonian and overlap matrices are obtained from the conventional band structure calculation for the periodic structure , the transmission is evaluated by eq . ( [ eq : c1 - 57 ] ) , where all the necessary information to evaluate eq . ( [ eq : c1 - 57 ] ) can be reconstructed by the result of the band structure calculation . as an example of the scheme , we calculate the conductance of a ( lamno@xmath1)/(srmno@xmath1 ) superlattice with four different magnetic structures , i.e. , ferromagnetic , a - type , g - type , and c - type antiferromagnetic configurations of mn sites.@xcite ) /(srmno@xmath1 ) superlattice with the ferromagnetic spin configuration of mn sites . the middle sized blue and small red circles denote mn and o atoms , while sr and la atoms are denoted by the mark . density of states ( dos ) of the ( lamno@xmath1)/(srmno@xmath1 ) superlattice for ( b ) ferromagnetic and ( c ) a - type antiferromagnetic configurations . , width=302 ] in recent years , it has been found experimentally that the superlattice structures consisting of lamno@xmath1 and srmno@xmath1 layers exhibit a metal - insulator transition in terms of the layer thickness.@xcite bhattacharya et al . fabricated ( lamno@xmath1)@xmath259/(srmno@xmath1)@xmath260 superlattices on a srtio@xmath1 ( 001 ) substrate , and measured the in - plane resistivity as a function of temperature.@xcite the resistivity measurement indicates that the superlattices are metallic and insulating for @xmath261 and @xmath262 , respectively . on the other hand , the ( lamno@xmath1)@xmath263/(srmno@xmath1)@xmath263 superlattices fabricated on the same substrate by nakano et al . exhibit a sample dependence in the resistivity measurement , i.e. , one of three samples is metallic and the others are insulating.@xcite they argued that the metallic behavior observed in the one sample may be attributed to a certain structural incompleteness in the superlattice structure , and that the ideal superlattice should become insulating based on their experimental results . a theoretical model calculation for the ( lamno@xmath1)@xmath259/(srmno@xmath1)@xmath260 superlattices suggests that the metal - insulator transition at @xmath264 can be explained by existence of the g - type antiferromagnetic barrier in the srmno@xmath1 layers sandwiched by the lamno@xmath1 layers with the ferromagnetic configuration.@xcite since the analysis by nakano et al . suggests that the charge transfer between the lamno@xmath1 and srmno@xmath1 layers is rather localized in the vicinity of the interface,@xcite the model should be applicable to the case of ( lamno@xmath1)@xmath263/(srmno@xmath1)@xmath263 without significantly depending on the ratio between the thicknesses of ( lamno@xmath1 ) and ( srmno@xmath1 ) layers , indicating that ( lamno@xmath1)@xmath263/(srmno@xmath1)@xmath263 is metallic . however , the naive consideration evidently contradicts the experimental result.@xcite as a first step towards comprehensive understanding of transport properties of the superlattice structures by the first principle calculations , we consider the simplest superlattice , i.e. , ( lamno@xmath265srmno@xmath1 ) . in the calculations , the in - plane lattice constant is fixed to be 3.905 which is equivalent to that of the srtio@xmath1 substrate . the out - of - plane lattice constant is assumed to be 7.735 , since those are experimentally determined to be 3.959 and 3.776 for the lamno@xmath1 and the srmno@xmath1 layers , respectively , grown on the srtio@xmath1 substrate and the average out - of - plane lattice constant for the superlattices is nearly equivalent to the average of the two values.@xcite with those lattice constants internal structural parameters are optimized for each magnetic configuration without any constraint until the maximum force is less than @xmath266 hartree / bohr . the optimized structure for the ferromagnetic configuration is shown in fig . it is found that the position of oxygen atoms is largely distorted due to the different ionic radii between la and sr atoms , showing that mn atoms are located in the center of each distorted octahedron . also , bond angles of mn - o - mn are found to be 167.4 and 161.6 ( degree ) for the in - plane and out - of - plane , respectively . the total energies relative to the ferromagnetic configuration are listed in table [ l1m1-tb ] . the calculated ground state is the ferromagnetic configuration , and the a - type antiferromagnetic configuration lies just above 5 mev per formula unit . it may be considered that the nearly degeneracy between the two configurations corresponds to the neighborhood of the boundary at @xmath267 and @xmath268 in the phase diagram for the tetragonal la@xmath269sr@xmath270mno@xmath1.@xcite the two configurations have both a metallic dos , while the ferromagnetic configuration is half - metallic as shown in figs . 10(b ) and ( c ) , reflecting the large in - plane and out - of - plane conductances as shown in table [ l1m1-tb ] . from a detailed analysis ( not shown ) of doss , we see that the electronic states at the fermi level are composed of @xmath271 orbitals of mn atoms and @xmath272 orbitals of oxygen atoms . also , the mulliken population analysis ( not shown ) suggests that the charge state of mn atoms in the superlattice is in between those in the lamno@xmath1 and srmno@xmath1 bulks . this implies that the metallic band structures are induced by partial filling of the @xmath271 bands . since the bond angle of mn - o - mn for the out - of - plane is slightly acute , therefore , this may be attributed to reduction of the conductance of the ferromagnetic state in the out - of - plane as shown in table [ l1m1-tb ] . a systematic study for the thicker cases and the effect of coulomb interaction@xcite in the @xmath271 orbitals is highly desirable , and the details will be discussed elsewhere . .total energy ( mev ) per formula unit , lamno@xmath0srmno@xmath1 , and conductance @xmath273 ( @xmath257 ) of the ( lamno@xmath265srmno@xmath1 ) supperlattice with four different magnetic configurations , i.e. , ferromagnetic ( f ) , a - type ( a ) , g - type ( g ) , and c - type ( c ) antiferromagnetic configurations of mn sites . the total energy is measured relative to that of the ferromagnetic configuration . @xmath274 is the in - plane conductance for the up spin state , and the others are construed in the similar way . for the conductance calculations * k*-points of @xmath275 were used . [ cols="^,^,^,^,^",options="header " , ] , where @xmath276 and @xmath277 are the elapsed times by a single core and a parallel calculations . the cores used in the mpi and openmp parallelizations are called _ process _ and _ thread _ , respectively . the parallel calculations were performed on a cray xt5 machine consisting of amd opteron quad core processors ( 2.3ghz ) . in the benchmark calculations , the number of processes is taken to be equivalent to that of processors . therefore , in the parallelization using 1 or 2 threads , 3 or 2 cores are idle in a quad core processor.,width=302 ] the computation in the the negf method can be parallelized in many aspects such as * k*-points , energies in the complex plane at which the green functions are evaluated , spin index , matrix multiplications , and calculation of the inverse of the matrix . here we demonstrate a good scalability of the negf in the parallel computation by a hybrid scheme using the message passing interface ( mpi ) and openmp which are used for inter - nodes and intra - node parallelization , respectively . the green function defined by eq . ( [ eq : c1 - 6 ] ) is specified by the * k*-point , energy @xmath278 , and spin index @xmath12 . since the calculation of the green function specified by each set of three indices can be independently performed without any communication among the nodes , we parallelize triple loops corresponding to the three indices using mpi . each node only has to calculate the green functions for an allocated domain of the set of indices , and partly sum up eq . ( [ eq : c1 - 26 ] ) or ( [ eq : c1 - 42 ] ) in a discretized form . after all the calculations finish , a global summation among the nodes is required to complete the calculation of eq . ( [ eq : c1 - 26 ] ) or ( [ eq : c1 - 42 ] ) , which , in most cases , is a very small fraction of the computational time even including the mpi communication among the nodes . thus , the reduction of scalability for the parallelization of the three indices is mainly due to imbalance in the allocation of the domain of the set of indices . the imbalance can happen in the case that the number of combination for the three indices and the number of processes in the mpi parallelization are relatively small and large , respectively . in addition to the three indices , one may notice that the matrix multiplications and the calculation of the matrix inverse can be parallelized , which are situated in the inner loops of the three indices . the evaluation of the central green function given by eq . ( [ eq : c1 - 6 ] ) and the surface green functions given by eq . ( [ eq : c1 - 21 ] ) , and the self energies given by eqs . ( [ eq : c1 - 7 ] ) and ( [ eq : c1 - 8 ] ) are mainly performed by the matrix multiplications and the calculation of the matrix inverse . we parallelize these two computations using openmp in one node . since the memory is shared by threads in the node , the communication of the data is not required unlike the mpi parallelization . however , the conflict in the data access to the memory can reduce the scalability in the openmp parallelization . as a whole , in our implementation the * k*-point , energy @xmath278 , and spin index @xmath12 are parallelized by mpi , and the matrix multiplications and the calculation of the matrix inverse are parallelized by openmp . in fig . 11 we show the speed - up ratio in the elapsed time for the evaluation of the density matrix of 8-zgnr under a finite bias voltage of 0.5 ev . the geometric and magnetic structures and computational conditions for 8-zgnr are same as before . the energy points of 197 ( 101 and 96 for the equilibrium and nonequilibrium terms , respectively ) are used for the evaluation of the density matrix . only the @xmath252 point is employed for the * k*-point sampling , and the spin polarized calculation is performed . thus , the combination of 394 for the three indices are parallelized by mpi . it is found that the speed - up ratio of the flat mpi parallelization , corresponding to 1 thread , reasonably scales up to 64 processes . furthermore , it can be seen that the hybrid parallelization , corresponding to 2 and 4 threads , largely improves the speed - up ratio . by fully using 64 quad core processors , corresponding to 64 processes and 4 threads , the speed - up ratio is about 140 , demonstrating the good scalability of the negf method . we have presented an efficient and accurate implementation of the negf method for electronic transport calculations in combination with dft using pseudo - atomic orbitals and pseudopotentials . in the implementation , we have developed accurate methods for the evaluation of the density matrix and the treatment of the boundary between the scattering region and the leads . a contour integration method with a continued fraction representation of the fermi - dirac function has been successfully applied for the evaluation of the equilibrium term in the density matrix , which evidently outperforms the previous method,@xcite while a simple quadrature scheme on the real axis with a small imaginary part is employed for that for the nonequilibrium term in the density matrix . it has been demonstrated by numerical calculations that the accuracy of @xmath234 ev per atom in @xmath233 is attainable using the energy points of 200 in the complex plane even under a finite bias voltage of 0.5 v at 600 k. however , the evaluation of the nonequilibrium density matrix still requires a careful treatment where the pole structure of the green functions has to be smeared by introducing a finite imaginary part . the numerical calculations suggest that the number of energy points required for the convergence can be largely reduced by introducing the imaginary part of 0.01 ev without largely changing the calculated results in a practical sense . we also note that the accurate evaluation of the density matrix provides another advantage that the scf calculations even under a finite bias voltage smoothly converge in a similar fashion as the conventional band structure calculation does . we have also developed an efficient method for calculating the hartree potential by a combination of the two dimensional fft and a finite difference method without any ambiguity in reproducing the boundary conditions . in addition , a careful evaluation of the charge density near the boundary between the scattering region and the leads is presented in order to avoid the spurious scattering accompanied by the inaccurate construction of the charge density . the proper treatment for the charge evaluation in our implementation can definitely be verified by a comparison between the conventional band structure calculation and the egf method with respect to the dos of the carbon chain . finally , we have demonstrated the applicability of our implementation by calculations of spin resolved @xmath243-@xmath207 characteristics of zgnrs , showing that the @xmath243-@xmath207 characteristics depend on the symmetry of zgnr , and that the symmetric zgnr exhibits a unique spin diode and filter effect . also , the applicability of our implementation to bulk systems is demonstrated by applications to a fe / mgo / fe tunneling junction and a lamno@xmath0srmno@xmath1 superlattice . based on the above discussions and the good parallel efficiency in the hybrid parallelization shown in the study , it is concluded that our implementation of the negf method can be applicable to challenging problems related to large - scale systems , and can be a starting point , apart from numerical spurious effects , to include many body effects beyond the one particle picture in the electronic transport . the authors would like to thank h. kondo for providing a prototype fortran code of the negf method . this work is partly supported by crest - jst , the next generation supercomputing project , nanoscience program , and nedo ( as part of the nanoelectronics project ) . let us introduce the following density functional which may define the total energy of the central region @xmath36 being a part of the extended system at equilibrium with a common chemical potential @xmath96 : @xmath279 where @xmath280 is the coulomb interaction energy between electrons and the external potential of the central region @xmath36 given by @xmath281 and @xmath282 is the hartree energy defined by @xmath283 where @xmath284 is an additional electron density which arises from the boundary condition between the central and lead regions . in fact , the additional electron density @xmath284 can be obtained by back fourier transforming eq . ( [ eq : c1 - 39 ] ) divided by @xmath285 within our treatment . the last term of eq . ( [ eq : a-1 ] ) , @xmath232 , is a self energy density functional , which may correspond to the energy contribution from the self energy due to the semi infinite leads , given by @xmath286 \label{eq : a-4}\end{aligned}\ ] ] with @xmath287 where @xmath288 , the factor of 2 is due to the spin multiplicity , and the self energy @xmath289 is the sum of the self energies arising from the left and right leads . although we neglect the spin and * k*-dependency on the formulation for the simplicity throughout the appendix , its generalization with the dependency is straightforward . the kinetic energy @xmath290 can be evaluated by eq . ( [ eq : c1 - 6 ] ) as the band energy @xmath291 minus double counting corrections as follows : @xmath292\\ \nonumber & = & e_{\rm band } - \int dr^3 n({\bf r } ) v_{\rm eff}({\bf r})\\ \nonumber & & - { \rm tr}\left [ -\frac{2}{\pi } { \rm i m } \int_{-\infty}^{\mu } de g_{c}(e^+ ) \sigma(e^+ ) \right ] \label{eq : a-6},\\\end{aligned}\ ] ] where @xmath291 is defined by @xmath293 . \label{eq : a-7}\end{aligned}\ ] ] it is noted that the last term in eq . ( [ eq : a-6 ] ) cancels the contribution from the first term in eq . ( [ eq : a-5 ] ) . the effective potential @xmath294 in the second term of eq . ( [ eq : a-6 ] ) will be defined later . also , the exchange - correlation energy @xmath295 in eq . ( [ eq : a-1 ] ) is considered to be a density functional such as local density approximations ( lda ) and generalized gradient approximations ( gga ) evaluated using electron density @xmath62 in the central region @xmath36 . we now consider the variation of the total energy @xmath231 with respect to @xmath62 . the variations of @xmath280 and @xmath282 are simply given by @xmath296 = \int dr^3 \delta n({\bf r } ) v_{\rm ext}({\bf r } ) \label{eq : a-8}\end{aligned}\ ] ] and @xmath297 = \int dr^3 \delta n({\bf r } ) \int dr'^3 \frac{n({\bf r}')+n'({\bf r } ' ) } { \vert { \bf r } - { \bf r } ' \vert}. \label{eq : a-9}\end{aligned}\ ] ] by noting the dyson equation and @xmath298 which are both derived from eq . ( [ eq : c1 - 6 ] ) , and @xmath299 , the variation of @xmath291 is given by two contributions : @xmath300 } \\ \nonumber & & = \int dr^3 \delta n({\bf r } ) { \rm tr}\left [ -\frac{2}{\pi } { \rm i m } \int_{-\infty}^{\mu } de\right.\\ \nonumber & & \qquad\qquad \left . \times e g_{c}(e^+ ) \frac{\delta h_{v } } { \delta n({\bf r } ) } g_{c}(e^+ ) s_{c } \right]\\ \nonumber & & = \int dr^3 \delta n({\bf r } ) { \rm tr}\left [ -\frac{2}{\pi } { \rm i m } \int_{-\infty}^{\mu } de \right.\\ & & \nonumber \left . \qquad\qquad \times e\left ( \frac{-\delta h_{v } } { \delta n({\bf r } ) } \right ) \frac{\partial g_{c}(e^+)}{\partial e } \right]\\ \nonumber & & \quad + \int dr^3 \delta n({\bf r } ) { \rm tr}\left [ -\frac{2}{\pi } { \rm i m } \int_{-\infty}^{\mu } de \right.\\ \nonumber & & \qquad \qquad \left . \times e g_{c}(e^+ ) \frac{\delta h_{v}}{\delta n({\bf r } ) } g_{c}(e^+ ) \frac{\partial \sigma(e^+)}{\partial e } \right].\\ \label{eq : a-10}\end{aligned}\ ] ] the trace in the first term of eq . ( [ eq : a-10 ] ) can be transformed by considering a partial integral and assuming the system to be insulating as follows : @xmath301 } \\ \nonumber & & = -{\rm tr}\left [ -\frac{2}{\pi } { \rm i m } \int_{-\infty}^{\mu } de \left ( \frac{-\delta h_{v}}{\delta n({\bf r } ) } \right ) g_{c}(e^+ ) \right]\\ & & = \int dr'^3 n({\bf r } ' ) \frac{\delta v_{\rm eff}({\bf r}')}{\delta n({\bf r})}. \label{eq : a-11}\end{aligned}\ ] ] it is also noted that the second term in eq . ( [ eq : a-10 ] ) cancels the variation of the contribution from the second term in eq . ( [ eq : a-5 ] ) . the variation of the second term in eq . ( [ eq : a-6 ] ) is easily found as @xmath302 & = & \int dr^3 \delta n({\bf r } ) v_{\rm eff}({\bf r})\\ \nonumber & + & \int dr^3 \delta n({\bf r } ) \int dr'^3 n({\bf r } ' ) \frac{\delta v_{\rm eff}({\bf r}')}{\delta n({\bf r})}.\\ \label{eq : a-12}\end{aligned}\ ] ] thus , it turns out using eqs . ( [ eq : a-8])-([eq : a-12 ] ) that the variation of the total energy @xmath231 is given by @xmath303 } \\ \nonumber & & = \int dr^3 \delta n({\bf r})\\ \nonumber & & \times \left ( -v_{\rm eff}({\bf r } ) + v_{\rm ext}({\bf r } ) + \int dr'^3 \frac{n({\bf r}')+n'({\bf r } ' ) } { \vert { \bf r } - { \bf r } ' \vert } + \frac{\delta e_{\rm xc}}{\delta n({\bf r } ) } \right).\\ \label{eq : a-13}\end{aligned}\ ] ] letting @xmath304 $ ] be zero so that the variation of the total energy @xmath231 can be always zero with respect to @xmath62 , we obtain a form of the effective potential : @xmath305 it is found that the effective potential takes the same form as in the kohn - sham method.@xcite the fact implies that the self consistent solution of the green function under the zero bias condition may correspond to the minimization of the total energy functional defined by eq . ( [ eq : a-1 ] ) , since in practice the green function defined by eq . ( [ eq : c1 - 6 ] ) is calculated using the effective potential given by eq . ( [ eq : a-14 ] ) as a consequence of combining the negf method with dft . the generalization of the functional to the metallic case with a finite temperature and the nonequilibrium state might be an important direction in the future study so that forces on atoms can be _ variationally _ calculated from the functional , since the existence of a variational functional has been recently suggested for the nonequilibrium steady state.@xcite the equilibrium energy density matrix @xmath306 , where one of the associated basis orbitals is in the central cell and the other is in the cell denoted by @xmath16 , is calculated using the contour integration method applied to the equilibrium density matrix as follow : @xmath307 with @xmath308 if the hamiltonian and overlap matrices are @xmath17-independent , as well as eq . ( [ eq : c1 - 22 ] ) , eq . ( [ eq : c1 - 28 ] ) can be simplified to : @xmath309 . \label{eq : c1 - 29}\end{aligned}\ ] ] for the general case with the @xmath17-dependent hamiltonian and overlap matrices , eq . ( [ eq : c1 - 28 ] ) is evaluated by the contour integration method with the special form of fermi - dirac function given by eq . ( [ eq : c1 - 25 ] ) as follows : @xmath310 with @xmath311 where @xmath312 is the first order moment of the green function @xmath114 . in eq . ( [ eq : c1 - 30 ] ) , a term , @xmath313 , which appears mutually for @xmath314 , is omitted , since the diverging terms cancel each other out in eq . ( [ eq : c1 - 27 ] ) . by making use of the moment representation of the green function,@xcite the following simultaneous linear equation is derived for the zero and first order moments : @xmath315 letting @xmath316 and @xmath317 be @xmath318 and @xmath319 , respectively , the zero and first order moments can be evaluated by @xmath320 where @xmath69 is a large real number so that the higher order moments can be neglected . for the nonequilibrium green function , the nonequilibrium contribution @xmath321 in the the energy density matrix @xmath322 can be calculated using the simple quadrature scheme in the same way as for the nonequilibrium term in the density matrix . in a similar fashion to the charge density , the atomic electron density @xmath142 in the central region consists of two contributions . one of them comes from atoms located in the central region @xmath36 , and the other is the contribution by the tail of atomic electron densities associated with atoms in the leads regions . since the latter is independent of the the scf iteration , it is computed on the regular mesh before the scf iteration . we used @xmath323 ev , eb=27.21 ev , and @xmath324 , being the parameters in the closed contour integration depicted in fig . 2 of ref.@xcite . the value for @xmath325 gives four enclosed matsubara poles for @xmath326 k. the path @xmath68 toward the positive direction is extended by @xmath327 from the chemical potential . the paths @xmath68 and @xmath36 are integrated using the gauss - legendre quadrature scheme , respectively . it turns out that 40 grid points for the path @xmath36 is enough to achieve the fully convergent result within double precision . so , we used 40 grid points for the path @xmath36 , and only changed the number of grid points for the path @xmath68 in the calculation for fig . 2(a ) of ref.@xcite . therefore , the number of energy points in the closed contour integration is given by 4 ( the enclosed matsubara poles ) plus 40 ( the path @xmath36 ) plus the number of grid points for the path @xmath68 .

statistical inference for time series stemming from stationary functional processes has attracted considerable interest during the last decades and progress has been made in several directions . estimation and testing procedures have been developed for a wide range of inference problems and for large classes of stationary functional processes ; see bosq ( 2000 ) , hrmann and kokoszka ( 2012 ) and horvth and kokoszka ( 2012 ) . however , the asymptotic results derived , typically depend in a complicated way on difficult to estimate , infinite dimensional characteristics of the underlying functional process . this restricts considerably the implementability of asymptotic approximations when used in practice to judge the uncertainty of estimation procedures or to calculate critical values of tests . in such situations , bootstrap methods can provide useful alternatives . bootstrap procedures for hilbert space - valued time series proposed so far in the literature , are mainly attempts to adapt , to the infinite dimensional functional set - up , of bootstrap methods that have been developed for the finite dimensional ( i.e. , mostly univariate ) time series case . politis and romano ( 1994 ) considered applications of the stationary bootstrap to functional , hilbert - valued time series and showed its validity for the sample mean for functional processes satisfying certain mixing and boundeness conditions . dehling et al . ( 2015 ) considered applications of the non - overlapping block bootstrap to u - statistics for so called near epoch dependent functional processes and sharipov et al . ( 2016 ) to change point analysis . franke and nyarigue ( 2016 ) and zhou and politis ( 2016 ) developed some theory for a residual - based bootstrap applied to a first order functional autoregressive process . notice that the transmission of other bootstrap procedures for real - valued time series to the functional set - up , like for instance of the autoregressive - sieve bootstrap , kreiss ( 1988 ) and kreiss et al . ( 2011 ) , seems to be difficult mainly due to problems associated with the estimation ( of an with sample size increasing number ) of infinite dimensional autoregressive operators . applications of bootstrap procedures to certain inference problems in functional time series analysis have been also considered in the literature . for instance , for the construction of prediction intervals , fernndez de castro et al . ( 2005 ) used an approach based on resampling pairs of functional observations by means of kernel - driven resampling probabilities . the same authors also apply a parametric , residual - based bootstrap approach using an estimated first order functional autoregression with i.i.d . resampling of appropriately defined functional residuals . for the same prediction problem , hyndman and shang ( 2009 ) applied different bootstrap approaches including bootstrapping the functional curves by randomly disturbing the forecasted scores using residuals obtained from univariate autoregressive fits . aneiros - perez et al . ( 2011 ) considered the nonparametric functional autoregressive models , while mingotti et al . ( 2015 ) the case of the integrated functional autoregressive model . apart from the lack of theoretical justification , the aforementioned bootstrap applications do not provide a general bootstrap methodology for functional time series as they are designed for and their applicability is restricted to the particular inference problem considered ; see also mcmurry and politis ( 2012 ) and shang ( 2016 ) for an overview . in this paper a general and easy to implement bootstrap procedure for functional time series is proposed which generates bootstrap replicates @xmath0 of an observed functional time series @xmath1 and is applicable to a large class of stationary functional processes . the procedure avoids the explicit estimation of process operators and exploits some basic properties of the stochastic process of fourier coefficients ( scores ) appearing in the well - known karhunen - love expansion of the functional random variables . it is in particular shown , that under quite general assumptions , the stochastic process of fourier coefficients obeys a so - called vector autoregressive representation and this representation plays a key role in developing a bootstrap procedure for the functional time series at hand . more specifically , to capture the essential driving functional parts of the underlying infinite dimensional process , the first @xmath2 functional principal components are used and the corresponding @xmath2-dimensional time series of fourier coefficients is bootstrapped using a @xmath3th order vector autoregression fitted to the vector time series of sample fourier coefficients . in this way , a @xmath2-dimensional pseudo - time series of fourier coefficients is generated which imitates the temporal dependence structure of the vector time series of sample fourier coefficients . using the ( truncated ) karhunen - love expansion , these pseudo - fourier coefficients are then transformed to functional bootstrap replicates of the main driving , principal components , of the observed functional time series . adding to these replicates an appropriately resampled functional noise , leads finally to the bootstrapped functional pseudo - time series @xmath4 . in a certain sense , our bootstrap procedure works by using a finite rank ( i.e. , @xmath2-dimensional ) approximation of the infinite dimensional structure of the underlying functional process and a @xmath3th order vector autoregressive approximation of its infinite order temporal dependence structure . to achieve consistency and to capture appropriately the entire infinite dimensional structure of the functional process , the number @xmath2 of functional principal components used as well as the order @xmath3 of the vector autoregression applied , are allowed to increase to infinity ( at some appropriate rate ) as the sample size @xmath5 increases to infinity . this double sieve property justifies the use of the term sieve bootstrap " for the bootstrap procedure proposed . we show that under quite general conditions , this bootstrap procedure succeeds in imitating correctly the entire infinite dimensional autocovariance structure of the underlying functional process . notice that apart from the problem that instead of the unknown true scores , the time series of estimated scores is used , the asymptotic analysis of our bootstrap procedure faces additional challenges which are caused by the fact that vector autoregressions of increasing order and of increasing dimension are considered and that the lower bound of the corresponding spectral density matrix approaches zero as the dimension of the vector time series of scores used , increases to infinity . we apply the proposed sieve bootstrap procedure to the problem of estimating the distribution of the functional fourier transform which is fundamental in a multitude of applications and has attracted interest in the functional time series literature ; see cerovecki and hrmann ( 2015 ) for some recent developments . a basic bootstrap central limit theorem is then established which shows validity of the functional sieve bootstrap for this important class of statistics , which includes the sample mean as a special case . using the time series of fourier coefficients in the context of functional time series analysis has been considered by many authors in a variety of applications . among others we mention hyndman and shang ( 2009 ) who , for functional autoregressive models and for the sake of prediction , used univariate autoregressions fitted to the scalar time series of scores . in the same context and more related to the approach proposed in this paper , a multivariate approach of prediction has been proposed by aue et al . ( 2014 ) which works by fitting a vector autoregressive model to the multivariate time series of scores . the paper is organized as follows . section 2 derives some basic properties and discuss the autoregressive representations of the vector process of fourier coefficients appearing in the karhunen - love expansion of the functional process . apart from being useful for bootstrap purposes , these properties are of interest on their own . the functional sieve bootstrap procedure proposed is described in section 3 where some properties of the bootstrap functional pseudo - time series are also discussed . asymptotic validity of the new bootstrap procedure for finite fourier transforms is established in section 4 . section 5 presents some numerical simulations which investigate the finite sample performance of the functional sieve bootstrap and comparisons with the performance of three popular block bootstrap methods are given . technical proofs and auxiliary lemmas are deferred to section 6 . we consider a ( functional ) stochastic process @xmath6 where for each @xmath7 ( interpreted as time ) , @xmath8 is a random element of the separable hilbert space @xmath9,\re)$ ] with parametrization @xmath10 for @xmath11 $ ] . as usual we denote by @xmath12 the inner product in @xmath13 and by @xmath14 the induced norm defined for @xmath15 as @xmath16}x(t)y(t)dt\big)^{1/2}$ ] and @xmath17 respectively . furthermore , for a matrix @xmath18 we denote by @xmath19 its frobenius norm , while for an operator @xmath20 , @xmath21 denotes its operator norm and @xmath22 its hilbert - schmidt norm , if @xmath20 is a hilbert - schmidt operator . for the underlying functional process @xmath23 it is assumed that its dependence structure satisfies the following assumption . * assumption 1 * @xmath24 is a purely non deterministic , @xmath25-@xmath26 approximable process . the general notion of @xmath27 approximability refers to stochastic process @xmath28 with @xmath29 taking values in @xmath30 , @xmath31 , and where the random element @xmath32 admits the representation @xmath33 . here the @xmath34 s are i.i.d . random elements in @xmath13 and @xmath35 some measurable function @xmath36 . if for @xmath37 an independent copy of @xmath38 , the condition @xmath39 is satisfied , where @xmath40 , then @xmath23 is called @xmath27 approximable . @xmath41 approximability is a notion of weak dependence which applies to many commonly used functional time series models , like linear functional processes , functional arch processes , etc . ; see hrmann and kokoszka ( 2010 ) for more details . let @xmath42 be the mean of @xmath23 which by stationary is independent of @xmath43 and for which we assume @xmath44 for simplicity . we denote by @xmath45 the autocovariance operator @xmath46 at lag @xmath47 defined by @xmath48 . associated with the covariance operator is the covariance function @xmath49\times[0,1 ] \rightarrow \re$ ] with @xmath50 , @xmath51 $ ] , that is , @xmath52 is an integral operator with kernel function @xmath53 . assumption 1 implies that @xmath54 and that for every @xmath55 the spectral density operator @xmath56 is well defined , continuous in @xmath57 , selfadjoint and trace class , hrmann et al . ( 2015 ) ; see also panaretos and tavakoli ( 2013 ) for similar properties under different weak dependence conditions which require summability of functional cumulants . in what follows we will strengthen somehow the assumption on the norm summability of the autocovariance operator to the following requirement . * assumption 2 * @xmath58 for some @xmath59 . furthermore , we will assume that the spectral density operator @xmath60 satisfies the following condition . * assumption 3 * for all @xmath61 $ ] , the operator @xmath62 is of full rank , i.e. , @xmath63 . for real - valued univariate processes , @xmath63 is equivalent to the condition that the spectral density is everywhere in @xmath64 $ ] strictly positive while for multivariate process to the non - singularity of the spectral density matrix for every frequency @xmath61 $ ] . notice that all eigenvalues @xmath65 , @xmath66 of @xmath67 are positive , while @xmath68 by the trace class property of @xmath60 . since @xmath69 , the positivity of @xmath60 implies that the covariance operator @xmath70 has full rank , that is , its eigenvalues @xmath71 satisfy @xmath72 for all @xmath73 and the corresponding eigenfunctions @xmath74 form a complete orthonormal basis in @xmath30 . by the symmetry and compacteness of @xmath75 , the random element @xmath32 admits the well known karhunen - love representation @xmath76 where @xmath77 , @xmath78 , are the orthonormalized eigenfunctions that correspond to the eigenvalues @xmath79 , @xmath80 , of @xmath75 . for @xmath81 , let @xmath82 , @xmath73 , and consider any subset of indices @xmath83 with @xmath84 , @xmath85 . later on , we will concentrate on the specific set @xmath86 which will be the set of the @xmath2 largest eigenvalues of the covariance operator @xmath70 . consider now the @xmath2-dimensional process @xmath87 . observe that @xmath88 is strictly stationary , purely non deterministic and has mean zero , @xmath89 . furthermore , its autocovariance matrix function @xmath90 , @xmath91 , is given by @xmath92 and satisfies by assumption 2 , @xmath93 note that the bound on the right hand side above is independent of the set @xmath94 and that although by construction it holds true that @xmath95 for @xmath96 , the random variables @xmath97 and @xmath98 may be correlated for @xmath99 . the summability property ( [ eq.gammasum ] ) implies that the @xmath2-dimensional vector process @xmath88 possesses a continuous spectral density matrix @xmath100 which is given by @xmath101 moreover , @xmath102 satisfies the following boundeness conditions . [ le.eigen ] under assumption 1 and 3 and assumption 2 with @xmath103 , the spectral density @xmath104 satisfies @xmath105$},\ ] ] where @xmath106 and @xmath107 are real numbers ( @xmath108 depends on the set @xmath94 ) , such that @xmath109 and @xmath110 is the @xmath111 unity matrix . the continuity and the boundeness properties of the spectral density matrix @xmath112 stated in lemma [ le.eigen ] , imply that the process @xmath88 obeys a so called vector autoregressive representation ; cheng and pourahmadi ( 1983 ) , see also wiener and masani ( 1957 ) and ( 1958 ) . that is , there exist an infinite sequence of @xmath113-matrices @xmath114 and a full rank @xmath2-dimensional white noise process @xmath115 , such that @xmath116 where the coefficients matrices satisfy @xmath117 and @xmath118 is a zero mean white noise innovation process , that is @xmath119 and @xmath120 , with @xmath121 if @xmath122 , @xmath123 otherwise and @xmath124 a full rank @xmath113 covariance matrix . we stress here the fact that ( [ eq.arxi ] ) does not describe a model for the process of fourier coefficients @xmath125 and should not be confused with the so - called linear , infinite order vector autoregressive ( var(@xmath126 ) ) process driven by independent , identically distributed ( i.i.d . ) in fact , representation ( [ eq.arxi ] ) is the autoregressive analogue of the well - known ( moving average ) wold representation of @xmath125 with respect to the same white noise innovation process @xmath127 . this autoregressive representation is valid for any stationary and purely non deterministic process the spectral density matrix of which is continuous and satisfies the boundness conditions ( [ eq.fbounds ] ) ; see also cheng and pourahmadi ( 1983 ) and pourahmadi ( 2001 ) for details . in contrast to the wold representation , the autoregressive representation ( [ eq.arxi ] ) seems to be more appealing for statistical purposes , since it express the vector time series of fourier coefficients @xmath128 as a function of its ( in principle ) observable past values @xmath129 , @xmath66 . in what follows we assume that the eigenvalues are in descending order , i.e. , @xmath130 and we consider the set @xmath86 of the @xmath2 largest eigenvalues of @xmath75 . the corresponding normalized eigenfunctions ( principal components ) are denoted by @xmath77 , @xmath131 and are ( up to a sign ) uniquely identified . furthermore , by parseval s identity , the quantity @xmath132 describes the proportion of the variance of @xmath32 captured by the first @xmath2 functional principal components . to simplify notation we surpass in the following the upper index @xmath133 and write simple @xmath134 for @xmath128 keeping in mind that the @xmath135th component @xmath136 of @xmath137 corresponds to the @xmath138th largest eigenvalue @xmath71 of @xmath75 , @xmath139 . furthermore , we write @xmath140 , @xmath141 , @xmath142 and @xmath143 for @xmath144 , @xmath145 , @xmath106 and @xmath146 , respectively . the basic idea of our procedure is to generate pseudo - replicates @xmath4 of the functional time series at hand by first bootstrapping the @xmath2-dimensional time series of fourier coefficients @xmath147 , @xmath148 , corresponding to the first @xmath2 principal components . this @xmath2-dimensional time series of fourier coefficients is bootstrapped using the autoregressive representation of @xmath149 discussed in section 2.2 . the generated @xmath2-dimensional pseudo - time series of fourier coefficients is then transformed to functional principal pseudo - components by means of the truncated karhunen - love expansion @xmath150 . adding to this an appropriately resampled functional noise leads to the functional pseudo - time series @xmath4 . however , since the @xmath149 s are not observed , we work with the time series of estimates scores . this idea is precisely described in the following functional sieve bootstrap algorithm . * * step 1 * : select a number @xmath151 of functional principal components and an autoregressive order @xmath152 , both finite and depending on @xmath5 . * * step 2 * : let @xmath153 be the @xmath2-dimensional time series of estimated fourier coefficients , where @xmath154 , @xmath139 are the estimated eigenvectors corresponding to the estimated eigenvalues @xmath155 of the sample covariance operator @xmath156 with @xmath157 . * * step 3 * : let @xmath158 and define functional residuals as @xmath159 , @xmath160 . * * step 4 * : fit a @xmath161th order vector autoregressive process to the @xmath2-dimensional time series @xmath162 , @xmath160 , denote by @xmath163 , the estimates of the autoregressive matrices and by @xmath164 the residuals , @xmath165 different estimators @xmath166 , @xmath167 can be used , like for instance , yule - walker estimators ; cf . brockwell and davis ( 1991 ) . * * step 5 * : generate a @xmath2-dimensional pseudo time series of scores @xmath168 , @xmath169 , as @xmath170 where @xmath171 , @xmath160 are i.i.d . random vectors having as distribution the empirical distribution of the centered residual vectors @xmath172 , @xmath173 and @xmath174 . * * step 6 * : define the pseudo - functional time series @xmath175 by @xmath176 where @xmath177 are i.i.d . random functions obtained by choosing with replacement from the set of centered functional residuals @xmath178 , @xmath160 and @xmath179 . some remarks concerning the above algorithm are in order . note that @xmath4 are functional pseudo - random variables and that the autoregressive representation of the vector time series of fourier coefficients is solely used as a vehicle to bootstrap the @xmath2 main functional principal components of the time series at hand . in fact , it is this autoregressive representation which allows the generation of pseudo - time series of fourier coefficients @xmath180 , in step 4 and step 5 , in a way that imitates the dependence structure of the sample fourier coefficients @xmath181 and which are used in the ( m - truncated ) karhunen - love representation . this together with the additive functional noise @xmath182 , lead to the new functional pseudo - observations @xmath183 . notice that the estimated eigenvectors @xmath184 may point in an opposite direction than the eigenvectors @xmath185 . in asymptotic derivations this is commonly taking care off by considering the sign corrected estimator @xmath186 , where the ( unobserved ) random variable @xmath187 is given by @xmath188 . however , since adding this sign correction will not affect the asymptotic results derived , we assume for simplicity throughout this paper , that @xmath189 , for @xmath190 . some modifications of the above basic bootstrap algorithm are possible which concern the resampling schemes used to generate the vector of pseudo - innovations @xmath171 and/or the bootstrap functional noise @xmath191 . to elaborate , and as we will see in the sequel , the applied i.i.d . resampling of the innovations @xmath171 in step 5 , suffices in order to capture the entire , infinite dimensional second order structure of the underlying functional process @xmath24 . a modification of this i.i.d . resampling scheme may be needed , however , if higher order characteristics of the underlying functional process beyond those of order two , should also be correctly mimicked by the functional pseudo - time series @xmath4 . in such a case , the i.i.d . resampling used to generate the @xmath171 s in step 5 can be replaced by other resampling schemes ( i.e. , block bootstrap schemes ) that are able to capture higher order dependence characteristics of the white noise series @xmath192 s as well . it should be mentioned that if @xmath24 is a stationary functional autoregressive ( far ) process , i.e. , if @xmath193 , where @xmath194 are autoregressive operators and @xmath34 an i.i.d . sequence in @xmath30 , see bosq ( 2000 ) for stability conditions of the far process , then the infinite dimensional vector process @xmath195 of scores satisfies @xmath196 , where @xmath197 , @xmath198 , are @xmath199 dimensional matrices and @xmath200 is an infinite dimensional vector of i.i.d . innovations . that is , in this case the vector process of scores is generated by a linear , ( infinite dimensional ) autoregressive process of order @xmath3 . in this case and under some obvious modifications , the sieve bootstrap procedure proposed in this paper provides an alternative way to bootstrap the far(p ) process which circumvents the problem of estimating explicitly the autoregressive operators @xmath201 , @xmath202 . asymptotic validity of such a bootstrap procedure will then require that the truncation parameter @xmath2 increases to infinity ( at some appropriate rate ) with the sample size @xmath203 . as usual , all considerations made regarding the bootstrap procedure are made conditionally on the observed functional time series @xmath204 . the generation mechanism of the pseudo - time series @xmath4 , enables us to consider the bootstrap functional process @xmath205 , where for @xmath206 , @xmath207 , with @xmath208 generated as @xmath209 , and , where the @xmath210 s are i.i.d . functional random variable taking values in the set @xmath211 with probability @xmath212 . in the above notation @xmath213 is the @xmath2-dimensional vector @xmath214 , where the unity appears in the @xmath135th position . it is easy to see that @xmath215 is a strictly stationary functional process with mean function @xmath216 for all @xmath81 , and , autocovariance operator @xmath217 given , for @xmath91 , by @xmath218 where @xmath219 is the @xmath220 autocovariance matrix at lag @xmath221 of @xmath222 . @xmath223 is a hilbert - schmidt operator since it is , for @xmath224 , a finite rank operator while for @xmath225 it is the sum of a finite rank operator and of the ( hilbert - schmidt ) empirical covariance operator of the pseudo - innovations @xmath226 if the ( estimated ) vector autoregressive process used to generate the time series of pseudo - scores @xmath227 is stable , then the dependence structure of the bootstrap process @xmath215 can be more precisely described . this is stated in the following proposition . notice that the required stability condition of the estimated autoregressive polynomial is fulfilled , if for instance , @xmath228 , @xmath229 , are the yule - walker estimators ; cf . brockwell and davis ( 1991 ) , ch . [ pr.lp_m ] if the estimator @xmath228 , @xmath167 , used in step 4 of the functional sieve bootstrap algorithm is such that @xmath230 for all @xmath231 , where @xmath232 , @xmath233 , then , conditionally on @xmath204 , the bootstrap process @xmath215 is @xmath234 approximable . the @xmath234 approximability of @xmath215 implies the norm summability @xmath235 , see hrmann et al . ( 2015 ) , which can be also easily verified since @xmath236 furthermore , and because of the @xmath234 approximability property , the bootstrap process @xmath237 possesses for every @xmath55 a spectral density operator @xmath238 defined by @xmath239 @xmath240 and @xmath241 are essentially finite rank approximations of the corresponding population operators @xmath45 and @xmath242 respectively . thus and in order for the bootstrap process @xmath237 to capture the infinite dimensional structure of the underlying functional process and the infinite order dependence structure of the vector time series generating the scores , the dimension @xmath2 as well as the autoregressive order @xmath3 , used in the functional sieve bootstrap algorithm , have to increase to infinity ( at an appropriate rate ) as the sample size @xmath5 increases to infinity . this rate should take into account the fact that the true scores and eigenfunctions appearing in the karhunen - love expansion are not observed and , therefore , sample estimates are used instead . furthermore , not only the order @xmath3 but also the dimension @xmath2 of the fitted autoregressive process has to increase to infinity with the sample size which makes the asymptotic analysis more involved . finally , the lower bound @xmath243 of the spectral density matrix @xmath244 of the scores approaches zero as the sample size @xmath203 increases to infinity . this is due to the fact that the eigenvalues @xmath245 of the spectral density operator @xmath60 converge to zero as @xmath246 . these conditions impose several restrictions regarding the behavior of @xmath2 and @xmath3 with respect to the sample size @xmath5 which are summarized in the following assumption . * assumption 4 * the sequences @xmath247 and @xmath248 satisfy @xmath249 and @xmath250 as @xmath251 such that , 1 . @xmath252 2 . @xmath253 , where @xmath254 and @xmath255 for @xmath256 . @xmath257 , where @xmath243 is the lower bound of the spectral density matrix @xmath258 given in ( [ eq.fbounds ] ) . @xmath259 , where @xmath260 , @xmath167 denote the same estimator as @xmath228 , @xmath167 , based on the true vector time series of scores @xmath181 instead of their estimates @xmath261 and @xmath262 , @xmath263 are the coefficient matrices of the best ( in the mean square sense ) linear predictor of @xmath149 based on @xmath264 , @xmath167 . assumption 4(ii ) is mainly imposed in order to control the error made by the fact that the bootstrap procedure is based on estimated scores and eigenfunctions instead on the unobserved true quantities in a context where the dimension @xmath2 and the autoregressive order @xmath3 , both , increase to infinity . part ( iii ) relates the rate of increase of the autoregressive order @xmath3 to the lower bound of the spectral density matrix @xmath258 and the decay of the norm of the autoregressive matrices to zero . part ( iv ) is a requirement on the estimator @xmath265 , @xmath167 based on the true scores . this is essentially a requirement on the estimator @xmath228 applied , after ignoring the error caused by the fact that the estimated scores are used . notice that assumptions on the estimators of the autoregressive parameters are common in the autoregressive - sieve bootstrap literature for real valued - random variablees ; see kreis et al . ( 2011 ) and meyer and kreiss ( 2015 ) . however , the situation here is different since in our context , not only the order @xmath3 but also the dimension @xmath2 of the vector autoregression has to increase to infinite with the sample size . this justifies the additional factor @xmath266 appearing in ( iv ) . under the condition that @xmath2 and @xmath161 increase to infinity at an appropriate rate with @xmath5 such that assumption 4 is satisfied , the following proposition can be established which shows that the spectral density operator @xmath267 of the bootstrap process @xmath215 converges , in hilbert - schmidt norm , to the spectral density operator @xmath268 of the underlying functional process @xmath24 . [ pr.f ] under assumptions 1 , 3 and 4 and assumption 2 with @xmath103 , we have , that , as @xmath251 , @xmath269}\|{\mathcal f}^\ast _ { \omega , m } - { \mathcal f } _ \omega\|_{hs } \ \rightarrow \ 0,\ ] ] in probability . from the above proposition and the inversion formulae of fourier transforms , we immediately get for the covariance operators @xmath270 and @xmath52 of the bootstrap process @xmath237 and of the underlying process @xmath24 , that @xmath271 in probability as @xmath272 . thus the bootstrap process @xmath215 , imitates asymptotically correct the entire infinite dimensional autocovariance structure of the functional process @xmath24 . this allows for the use of the bootstrap functional time @xmath273 to approximate the distribution of statistics based on the functional time series @xmath204 . examples of such statistics are discussed in the next section . in this section we investigate the validity of the functional sieve bootstrap applied in order to approximate the distribution of some statistic @xmath274 of interest , when the bootstrap analogue @xmath275 is used . consider the distribution of the functional fourier transform @xmath276,\ ] ] and notice that the sample mean @xmath277 is a special case of ( [ eq.fft ] ) . in order to investigate the limiting distribution of @xmath278 , we fix some notation . we say that a random element @xmath279 , follows a complex gaussian distribution with mean @xmath280 and covariance @xmath281 , we write @xmath282 , if @xmath283 under a range of different weak dependence assumptions on the functional process @xmath23 , it has been shown that @xmath284 as @xmath251 . for @xmath285 , such a limiting behavior has been established for linear functional processes by merlevde et al . ( 1997 ) and for @xmath286 approximable processes by horvth et al . panaretros and tavakoli ( 2013 ) derived the above limiting distribution of @xmath287 for @xmath61 $ ] , under a summability condition of the functional cumulants , while more general results for the same statistic and under weaker conditions , have been recently obtained by cerovecki and hrmann ( 2015 ) . the following theorem establishes asymptotic validity of the functional sieve bootstrap procedure for the class of functional fourier transforms , when the bootstrap statistic @xmath288 is used to approximate the distribution of the statistic @xmath289 . [ th.th_main ] under assumptions 1 , 3 and 4 and assumption 2 with @xmath290 , we have for @xmath291 $ ] , that , as @xmath272 , @xmath292 in probability . in order to investigate the finite sample behavior of the functional sieve bootstrap ( fsp ) we have performed numerical simulations using time series stemming from a first order functional moving average process given by @xmath293 to specify the operator @xmath294 we adopted the simulation set - up of aue et al . ( 2015 ) and set @xmath295 , where @xmath296 , @xmath297 with @xmath298 , @xmath299 and @xmath300 fourier basis functions on the interval @xmath301 $ ] . for @xmath302 where @xmath303 with @xmath304 , the operator @xmath305 considered , acts as follows : @xmath306 . furthermore , the i.i.d . innovations @xmath34 in ( [ eq.ma1sim ] ) are obtained as @xmath307 , where the random variables @xmath308 are i.i.d . gaussian with mean zero and standard deviation equal to @xmath309 . we are interested in estimating the standard deviation of the sample mean @xmath310 , calculated for time series of length @xmath311 observations and for @xmath312 , @xmath313 , @xmath314 , equidistant time points in the interval @xmath301 $ ] . the exact standard deviation of the sample mean is estimated using 20,000 replications of the moving average model structure ( [ eq.ma1sim ] ) considered . the finite sample performance of the fsb is also compared with that of three other bootstrap procedures , the moving block bootstrap ( mbb ) , the tapered block bootstrap ( tbb ) and the stationary bootstrap ( sb ) . all estimates presented are based on @xmath315 replications and @xmath316 bootstrap repetitions . although the development of a data driven procedure that automatically selects @xmath3 and @xmath2 is a problem of future research , we experimented with several values of @xmath2 and @xmath3 where we found that for this model and for the sample size considered , a vector autoregression of order @xmath317 describes appropriately the temporal dependence of the vector time series of scores . table 1 shows the fsb estimates obtained using different values of the bootstrap parameters @xmath2 and @xmath3 . in figure 1 the results obtained using the fsb procedure are compared with those of the three aforementioned block bootstrap methods . in particular , figure 1 shows the fsb estimates of the standard deviation of the sample mean for @xmath318 and @xmath317 and the corresponding estimates of the mbb , of the tbb and of the sb . the bootstrap parameters of these block bootstrap procedures have been selected as follows . for the mbb we calculated the corresponding estimates for a range of values of the block size @xmath319 in the set @xmath320 and we selected the block size @xmath319 which minimizes the overall relative bias given by @xmath321 , where @xmath322 denote the moving block bootstrap estimator of the standard deviation using the block size @xmath319 and @xmath323 the corresponding estimated exact standard deviation . the block size for the mbb selected according to this procedure was equal to @xmath324 . the same procedure has been applied in order to choose the bootstrap parameters for the other two block bootstrap methods , that is of the tbb and of the sb procedure . for the tbb the block size selected according to this criterion was @xmath325 , while the same criterion delivered a probability of @xmath326 for the geometric distribution involved in the sb procedure . thus the corresponding block bootstrap estimator presented in figure 1 are , in the described sense , the ( overall ) less biased block bootstrap estimators . * please insert table 1 and figure 1 about here * as table 1 shows the fsb estimates are quite good even for functional time series consisting of @xmath311 observations and seem not to be very sensitive with respect to the different choices of the parameter @xmath2 used to truncate the karhunen - love expansion . notice that as @xmath2 increases the variability of the fsb estimates increases too . this is expected since the number of parameters to be estimated for the vector autoregression fitted to the time series of scores equals @xmath327 . the good finite sample behavior of the fsb method is also demonstrated in figure 1 where comparisons of the fsb estimates with those obtained using the three different block bootstrap methods are given . as this figure shows , between the three bootstrap estimators considered , the mbb estimator seems to behave better that the sb estimator , while both estimators are more biased compared to the tbb estimator . the later method performs best among the three different block bootstrap methods considered . however , all block bootstrap estimates are quite biased and they are clearly outperformed by the fsb estimates for all time points @xmath328 in the interval @xmath301 $ ] considered . [ le.a1 ] let assumption 1 , 2 and 3 be satisfied . denote by @xmath329 , @xmath78 , the coefficients matrices of the power series @xmath330 , where @xmath331 , @xmath231 , and let @xmath332 . then , * proof : * consider ( i ) and ( ii ) . let @xmath336 be the class of all @xmath220 matrix - valued functions on @xmath337 $ ] with @xmath338-valued fourier coefficient matrices @xmath339 satisfying the condition @xmath340 , where @xmath341 is independent of @xmath2 . then , @xmath342 since the autocovariance matrix function of @xmath343 satisfies @xmath344 , see ( [ eq.gammasum ] ) . furthermore , @xmath345 , with @xmath346 the optimal factor of @xmath347 ; see cheng and pourahmadi ( 1993 ) , p. 116 . from the boundeness conditions it follows that @xmath348 for all @xmath349 , and , therefore , @xmath350 is invertible with inverse denoted by @xmath351 . notice that @xmath352 . according to wiener and masani ( 1958 ) , theorem 5.5 and theorem 5.7 , there exist sequences @xmath353 and @xmath354 which are independent of @xmath7 such that for all @xmath81 , @xmath355 and @xmath356 , where @xmath357 and the infinite sums are @xmath358-convergent . the coefficients in the autoregressive and the wold representation are obtained by setting @xmath359 , @xmath360 and @xmath361 , where @xmath362 , @xmath363 and @xmath364 , @xmath365 , are the fourier coefficients of @xmath346 and @xmath351 , respectively . since @xmath352 , we get that @xmath366 and @xmath367 are bounded uniformly in @xmath2 . in ( iii ) the lower bound follows from the regularity of the infinite dimensional process of scores @xmath368 which in turn follows from the regularity of @xmath23 . for the upper bound , let @xmath369 , @xmath263 , be the ( positive ) eigenvalues of @xmath370 . then , since @xmath371 we have @xmath372 . @xmath373 the following version of baxter s inequality is very useful in our setting because it relates the approximation error of the coefficient matrices of the finite predictor and of the autoregressive - representation of the @xmath2-dimensional process of scores to the lower bound of the spectral density matrix @xmath374 . it is an immediate consequence of lemma [ le.eigen ] and theorem 3.2 of meyer et al . ( 2016 ) . [ le.a2 ] let assumption 1 , 2 and 3 be satisfied . then there exists a constant @xmath375 which does not depend on @xmath2 , such that for all @xmath376 , @xmath377 where @xmath243 is given in lemma [ le.eigen ] . the following lemma provides a useful bound between the estimated matrices of the autoregressive parameters based on the vector of scores @xmath134 and on the vector of their estimates @xmath162 , @xmath148 . it deals with the case of the yule - walker estimators and can be established for other estimators , like least squares , by similar arguments . * proof : * we first show that @xmath382 where @xmath383 , @xmath384 , @xmath385 , @xmath386 and @xmath387 , @xmath388 . since @xmath389 it suffices to consider only one of the two terms on the right hand side of the last bound . by the triangular and the cauchy - schwarz inequality we have @xmath390 with the @xmath391 term uniformly in @xmath221 . the assertion follows because by assumption 1 , @xmath392 ; see hrmann and kokoszka ( 2010 ) . we next proof the assertion for @xmath393 . for this recal that @xmath394 and @xmath395 . furthermore , @xmath396 and using the analogue expression for @xmath397 we get @xmath398 where the last equality follows using ( [ eq.difgamma ] ) and the bound @xmath399 . by the above arguments and using expression ( [ eq.difgamma ] ) we obtain @xmath400 extension of the assertion for @xmath401 follows using the above bound for @xmath402 , equation ( [ eq.difgamma ] ) and durbin - levinson s recursion formulae for the yule - walker estimators ; see brockwell and davis ( 1991 ) , ch . 11.4 . @xmath373 [ le.a4 ] let assumption 1 and 2 ( with @xmath103 ) be satisfied and @xmath403 , @xmath233 . there exists @xmath404 and a positive constant @xmath341 which does not depend on @xmath2 such that for @xmath405 and all @xmath406 , @xmath407 * proof : * we first show that the assertion is true for @xmath231 . since @xmath408 for @xmath231 it follows by the minimum modulus principle for holomorphic functions that @xmath409 . now , recall that for @xmath410 $ ] , @xmath411 . let @xmath412 be the largest eigenvalue of @xmath413 , we then have @xmath414 for some constant @xmath415 independent of @xmath2 . notice that the first inequality follows by lemma [ le.a1](iii ) and the last by the fact that @xmath416 is bounded uniformly in @xmath2 ; see lemma [ le.eigen ] . thus @xmath417 } |det ( a_{p , m}(e^{-i\omega})|^2 \geq c m^{-1}$ ] which implies that @xmath418 with some constant @xmath375 independent of @xmath2 . extension of this lower bound to the slightly larger region @xmath419 and for all @xmath406 for some @xmath404 , follows then exactly along the same lines as the proof of lemma 3.2 of meyer and kreiss ( 2015 ) ; see also lemma 2.3 of kreiss et al . @xmath373 to state the next lemma we first fix the following notation . @xmath420 , @xmath421 , @xmath422 and @xmath423 @xmath78 denote the coefficient matrices in the power series expansions of @xmath330 , @xmath424 , @xmath425 and @xmath426 , respectively , @xmath231 . we set @xmath427 . furthermore , @xmath428 , @xmath429 , @xmath430 and @xmath431 , while @xmath432 and @xmath433 with @xmath434 and @xmath435 where @xmath436 denotes expectation with respect to the measure assigning probability @xmath437 to each @xmath438 , @xmath439 . * proof : * to see ( i ) let @xmath446 be the @xmath447th element of a matrix @xmath448 and notice that by cauchy s inequality for holomorphic functions we have @xmath449 with an obvious notation for @xmath450 and @xmath451 . by theorem 2.12 of ipsen and rehman ( 2008 ) we get that @xmath452 where @xmath453 and @xmath454 denotes the spectral norm ( i.e. , the largest singular value ) of the matrix @xmath448 . since @xmath455 and using the bound @xmath456 for the largest singular value of a non - singular matrix @xmath448 , see merikoski and kumar ( 2005 ) , p. 373 , we get by straightforward calculations and in view of lemma [ le.a4 ] and the constant @xmath341 appearing there , that @xmath457 since @xmath458 and @xmath459 uniformly in @xmath2 . thus @xmath460 and assumption 4(iv ) lead to the bound @xmath461 from which we derive using ( [ eq.det-bound ] ) that @xmath462 and by lemma [ le.a4 ] that @xmath463 furthermore , by lemma [ le.a4 ] and the bound ( [ eq.det-bound ] ) we get @xmath464 thus and using equation ( [ eq.cauchy ] ) , we conclude that @xmath465 by assumption 4 . consider ( ii ) so we have , @xmath466 with an obvious notation for @xmath467 , @xmath468 . we show that all three terms converge to zero in probability . by the triangular inequality and in order to show @xmath469 , it suffices to show that @xmath470 . for this we use the bound @xmath471 since by straightforward calculations it yields that @xmath472 , we get by assumption 4(iv ) and cauchy - schwarz s inequality that @xmath473 for the second term on the right hand side of ( [ eq.bound1 ] ) we get by replacing @xmath474 by @xmath475 and using lemma [ le.a2 ] , that @xmath476 by assumption 4 . finally and by the same assumption , we get for the third term of ( [ eq.bound1 ] ) using @xmath477 which converges to zero in probability . since the term @xmath478 is easier to deal with using similar arguments as for the term @xmath479 , we consider the term @xmath480 . using @xmath481 we have that @xmath482 since @xmath483 uniformly in @xmath2 and by similar arguments as above , we have @xmath484 thus we conclude using assumption 4 , that @xmath485 . consider ( iii ) . by ( i ) it suffices to show that @xmath486 . for this notice that @xmath487 by lemma [ le.a3 ] and by cauchy s inequality for holomorphic functions we get for the @xmath447th element of the matrices @xmath423 and @xmath488 , that @xmath489 and @xmath490 from the above bound and by lemma [ le.a3 ] and lemma [ le.a4 ] , we get by similar arguments as those leading to the bound ( [ eq.det-bound ] ) , that , uniformly in @xmath138 , @xmath491 that is @xmath492 from which we get @xmath493 by assumption 4 ( ii ) . to establish ( iv ) notice that using ( ii ) it suffices to show that @xmath494 . by the triangular inequality it suffices to show that @xmath495\big(\widehat{e}_{t , p}(m)-\widetilde{e}_{t , p}(m)\big)\| \stackrel{p}{\rightarrow } 0.\ ] ] since the above term can be bounded by @xmath496 we show that both terms above converge to zero in probability . we use the bound @xmath497 from lemma [ le.a3 ] we get by straightforward calculations that , @xmath498 , @xmath499 and @xmath500 similar arguments yield @xmath501 and @xmath502 . * proof of lemma [ le.eigen ] : * expression ( [ eq.gammasum ] ) imediately leads , for all @xmath61 $ ] , to an upper bound of @xmath503 . to derive a lower bound , recall that @xmath504 and observe that @xmath505 let @xmath506 , @xmath139 , be the eigenvalues of @xmath507 ( including multiplicity ) . it suffices to show that @xmath508 for all frequencies @xmath509 $ ] , i.e. , that the eigenvalues of the spectral density matrix @xmath510 are uniformly ( in @xmath57 ) bounded away from zero . for this let @xmath511 be the corresponding normalized eigenvectors . then for every @xmath512 , we have @xmath513 by the positivity of @xmath60 , where @xmath514 and @xmath515 . because of the norm summability of the autocovariance matrix function @xmath516 , the eigenvalues @xmath517 , @xmath263 , are continuous functions of @xmath518 . let @xmath519 and notice that @xmath520 is continuous in @xmath57 and @xmath521 for all @xmath61 $ ] . define @xmath522}\delta_m(\omega)$ ] which is positive by the continuity of @xmath523 . hence @xmath524 for all @xmath509 $ ] . @xmath373 * proof of proposition [ pr.lp_m ] : * recall the definition of @xmath525 and observe that @xmath526 , where @xmath527 and the power series @xmath528 converges for @xmath231 . write @xmath529 and define @xmath530 where for each @xmath81 , @xmath531 is an independent copy of @xmath532 . notice that @xmath533 by minkowski s inequality we have @xmath534 and evaluating the first expectation term we get using @xmath535 and the submultiplicative property of the frobenius matrix norm , that @xmath536 where @xmath537 . an identical expression appears for the second expectation term on the right hand side of ( [ eq.l2mapp ] ) . applying minkowski s inequality again we get by the exponential decay of @xmath538 , that @xmath539 @xmath373 * proof of proposition [ pr.f ] : * recall that the spectral density operator @xmath67 can be expressed as @xmath540 . define for @xmath405 , @xmath541 and verify that since @xmath542 , the following expression is also valid for @xmath543 , @xmath544 where @xmath545 , @xmath231 . let @xmath546 where @xmath547 , @xmath231 . finally recall that @xmath548 then , @xmath549 the first term on the right hand side above is bounded by @xmath550 furthermore , @xmath551 where @xmath552 are i.i.d . random variables taking values with probability @xmath553 in the set @xmath554 and @xmath555 . then @xmath556 in probability , since @xmath557 and the operator @xmath70 is hilbert - schmidt . furthermore , @xmath558 in probability , where the last equality follows by straightforward calculations and using @xmath559 . similarly , and by the same arguments as above and using lemma [ le.a5 ] , we get @xmath560 consider next the second term on the right hand side of ( [ eq.speczer ] ) . for this term we get @xmath563 , i.e. , @xmath564 converges to zero in probability by lemma [ le.a5](v ) and ( vi ) . for the third and last term on the right hand side of ( [ eq.speczer ] ) we obtain @xmath565 as @xmath249 , since @xmath566 is a complete orthonormal basis of @xmath567 . @xmath373 * proof of theorem [ th.th_main ] * let @xmath568 where @xmath569 , @xmath160 with @xmath570 , where @xmath571 , @xmath167 are the estimators of the autoregressive parameter matrices based on the vector time series of true scores @xmath149 , @xmath148 and @xmath572 are obtained by i.i.d . resampling from the centered residuals @xmath573 , @xmath439 . that is , the pseudo - variable @xmath574 is obtained using the true eingefunctions @xmath77 and the true scores @xmath575 instead of their estimates @xmath184 and @xmath576 respectively . decompose then @xmath577 as @xmath578 with an obvious notation for @xmath579 , @xmath580 , @xmath581 and @xmath582 . notice that the terms @xmath580 and @xmath583 are due to the fact that , in the bootstrap procedure , the unknown scores and eigenfunctions are replaced by their sample estimates , while @xmath584 is due to the @xmath2-dimensional approximation of the infinite dimensional structure of the underlying process . the assertion of the theorem follows then from lemma [ le.le1th_main ] , [ le.le2th_main ] , [ le.le3th_main ] and [ le.le4th_main ] and slutsky s lemma . @xmath373 * proof : * note that @xmath586 using @xmath587 , see hrmann and kokoszka ( 2010 ) , we get @xmath588 where the last equality follows because @xmath589 . furthermore , @xmath590 follows using similar arguments and since @xmath591 , where @xmath592 . @xmath373 * proof : * we have @xmath594 with an obvious notation for @xmath595 and @xmath596 . we consider @xmath597 only since @xmath598 can be handled similarly . for this term we have @xmath599\widetilde{\psi}_{l+s - t , p}(m)^\top { \bf 1}_j \end{aligned}\ ] ] and , using lemma [ le.a1 ] and [ le.a5 ] , we get for the first term on the right hand side of ( [ eq.dnm1 ] ) , that , this term is bounded by @xmath600 in probability . the second term of ( [ eq.dnm1 ] ) is bounded by @xmath601 which converges to zero in probability , because @xmath602 in probability . this can be seen using @xmath603 we then have @xmath604 furthermore , @xmath605 where the last equality follows using lemma [ le.a1 ] and [ le.a3 ] . finally , @xmath606 in probability , since @xmath607 and @xmath608 by similar arguments we get @xmath609 , in probability . @xmath610 * proof : * write @xmath615 where @xmath616 with @xmath617 , @xmath618 , a random element in @xmath13 . notice that @xmath619 , while using @xmath620 , @xmath621 , we get @xmath622 with an obvious notation for @xmath623 . it is easily seen that @xmath624 and therefore @xmath625 in probability , by lemma [ le.a5 ] . hence and using @xmath626 as @xmath627 , we get that @xmath628 in probability , as @xmath251 . let @xmath629 , define @xmath630 and @xmath631 . it easily follows by simple algebra and using lemma [ le.a5 ] that @xmath632 , in probability , that is @xmath633 . thus to prove the assertion of the lemma it suffices to show that @xmath634 . for this we show that the assumptions of theorem 2 of cerovecki and hrmann ( 2015 ) are satisfied , that is , using the notation @xmath635 , we show that the following two conditions are fulfilled , in probability . @xmath636 and @xmath637 where the operator @xmath638 is defined as @xmath639 and @xmath640 . toward this we first define @xmath641 , where @xmath642 if @xmath643 and @xmath644 if @xmath645 with @xmath646 a copy of @xmath572 which is independent of @xmath647 for @xmath648 . we show that @xmath649 where the @xmath650 term is independent of @xmath2 and @xmath3 . notice first that by minkowski s inequality @xmath651 thus @xmath652 now , since by lemma [ le.a1](ii ) , @xmath653 is bounded uniformly in @xmath2 , and , by lemma [ le.a5](ii ) and ( vi ) and lemma [ le.a1](iii ) , @xmath654 is bounded in probability , where the bound is independent of @xmath3 and @xmath2 , assertion ( [ eq.uo ] ) follows . consider next condition ( [ th2.cond1 ] ) . for positive integers @xmath655 we have that @xmath656 recall the definition of @xmath657.then we have , since @xmath658 , that @xmath659 hence @xmath660 as @xmath661 because of ( [ eq.uo ] ) . 50 aneiros - prez , g. , cao , r. and vilar - fernanz , j. m. ( 2011 ) . functional methods for time series prediction : a nonparametric approach . _ journal of forecasting _ , * 30 * , 377 - 392 . aue , a. , dubart , d. n. and hrmann , s. ( 2015 ) . on the prediction of stationary functional time series . _ journal of the american statistical association _ , * 110 * , 378 - 392 . bosq , d. ( 2000 ) . _ linear process in function spaces_. springer , berlin - heidelberg - new york . cerovecki , c. and hrmann , s. ( 2015 ) . on the clt for discrete fourier transforms of functional time series . _ preprint_. cheng , r. and pourahmadi , m. ( 1993 ) . baxter s inequality and convergence of finite predictors of multivariate stochastic processes . _ probability theory and related fields _ , * 95 * , 115 - 124 . dehling , h. , sharipov , o. s. and wendler , m. ( 2015 ) . bootstrap for dependent hilbert space valued random variables with application to von mises statistics . _ journal of multivariate analysis _ , * 133 * , 200 - 215 . fernndez de castro , b. , guillas , s. and gonzlez manteiga , w. ( 2005 ) . functional samples and bootstrap for predicting sulfur dioxide levels . _ technometrics _ , * 47 * , 212 - 222 . franke , j. and nyarigue , e. ( 2016 ) . residual - based bootstrap for functional autoregressions . _ preprint_. hrmann , s. and kokoszka , p. ( 2010 ) . weakly dependent functional data . _ annals of statistics _ , * 38 * , 1845 - 1884 . hrmann , s. and kokoszka , p. ( 2012 ) . functional time series . _ handbook of statistics : time series analysis - methods and applications _ , 157 - 186 . hrmann , s. , kidziski , l. and hallin , m. ( 2015 ) . dynamic functional principal components . _ journal of the royal statistical society : series b _ , * 77 * , 319 - 348 . horvth , l. and kokoszka , p. ( 2012 ) . _ inference for functional data with applications_. springer , berlin - heidelberg - new york . hyndman , r. j. and shang , h. l. ( 2009 ) . forecasting functional time series . _ journal of the korean statistical society _ , * 38 * , 199 - 211 . ipsen , i. c. f. and rehman , r. ( 2008 ) . perturbation bounds for determinants and characteristic polynomials . _ siam journal of matrix analysis and applications _ , * 30 * , 762 - 776 . kreiss , j .- asymptotic statistical inference for a class of stochastic processes . universitt hamburg . kreiss , j .- p . , paparoditis , e. and politis , d. n. ( 2011 ) . on the range of validity of the autoregressive sieve bootstrap . _ annals of statistics _ , * 39 * , 2103 - 2130 . mcmurry , t. and politis , d. n. ( 2011 ) . resampling methods for functional data . _ the oxford handbook of functional data analysis _ , ( f. ferraty and y. romain , eds . ) 189 - 209 . oxford university press . merikoski , j. k. and kumar , r. ( 2005 ) . upper bounds for singular values . _ linear algebra and its applications _ , 401 , 371 - 379 . merlevede , f. , peligrad , m. and utev , s. ( 1997 ) . sharp conditions for the clt of linear processes in a hilbert -space . _ journal of theoretical probability _ , * 10 * , 681 - 693 . meyer , m. and kreiss , j .- ( 2015 ) . on the vector autoregressive sieve bootstrap . _ journal of time series analysis _ , * 36 * , 377 - 397 . meyer , m. , jentsch , c. , and kreiss , j .- baxter s inequality and sieve bootstrap for random fields . _ bernoulli _ , to appear . mingotti , n. , lillo , r. e. and romo , j. ( 2015 ) . a random walk test for functional time series . uc3 m working papers , statistics and econometrics . panaretos , v. and tavakoli , s. ( 2013 ) . fourier analysis of stationary time series in function spaces . _ annals of statistics _ , * 41 * , 568 - 603 . politis , d. n. and romano , j. ( 1994 ) . limit theorems for weakly dependent hilbert space valued random variables with applications to the stationary bootstrap . _ statistica sinica _ , * 4 * , 461 - 476 . pourahmadi , m. ( 2001 ) . _ foundation of time series analysis and prediction theory_. john - wiley : new york . sharipov , o. , tewes , j. and wendler , m. ( 2016 ) . sequential block bootstrap in a hilbert space with application to change point analysis . _ the canadian journal of statistics _ , forthcoming . wiener , n. and masani , p. ( 1957 ) . the prediction theory of multivariate stochastic processes , i. _ acta mathematica _ , * 98 * , 111 - 150 . wiener , n. and masani , p. ( 1958 ) . the prediction theory of multivariate stochastic processes , ii . _ acta mathematica _ , * 99 * , 93 - 137 . shang , l. h. ( 2016 ) . resampling methods for dependent functional data . _ preprint_. zhou , t. and politis , d. n. ( 2016 ) . kernel estimation of first - order nonparametric functional autoregression model and its bootstrap approximation . _ preprint_. * table 1 : * estimated exact ( @xmath663 ) and functional sieve bootstrap ( fsb ) estimates of the standard deviation of the sample mean @xmath664 for different values of @xmath665 $ ] and for different parameters @xmath666 and @xmath3 . @xmath667 refers to the mean , while @xmath668 to the standard deviation of the fsb estimates obtained for @xmath669 replications . for different values of @xmath665 $ ] . bullets refer to the estimated exact standard deviation , while the mean estimates of the standard deviation of the fsb ( @xmath670 ) are presented by circles , of the tbb ( @xmath325 ) , by diamonds , of the mbb ( @xmath324 ) , by triangles and of the sb @xmath671 , by + " .,title="fig:",width=566 ] +

band structures are of fundamental interest to the solid - state physicist as they reveal important properties of crystals . @xcite the determination of band structures on _ ab initio _ level is therefore an important issue in theoretical solid state physics . nowadays predominantly density functional theory ( dft ) @xcite with the local density approximation ( lda ) or its variants are employed to treat this problem . @xcite dft focuses on the ground - state electron density and yields ground - state properties . in several cases , the interpretation of the kohn - sham orbital energies as quasiparticle energies also turns out to be successful and particular attention has been paid to the lda , due to its numerical simplicity . however , in insulating materials , the lda tends to underestimate the band gap , e.g. , ref . . as is well known , it is difficult to improve band structure calculations based on density functional theory . this is in particular the case when electron correlations are strong such that in quite some cases @xcite the one - electron picture becomes inappropriate and corrections are needed . nevertheless , a number of ingenious methods have been devised to improve band structure results based on density functional theory . for example , the calculations have been supplemented by a gw treatment , e.g. , refs , or the so - called lda+@xmath1 method . @xcite other improvements concern the use of optimized effective potentials @xcite such as the exact exchange potential @xcite instead of its local , simplified form . time - dependent dft is another route to treat excited states more rigorously . furthermore , lda calculations have been coupled with various extensions of the coherent potential approximation ( cpa ) like the dynamical cpa @xcite or dynamical mean field theory ( dmft ) . @xcite also faddeev s method @xcite of treating the three - particle @xmath2-matrix has been applied in conjunction with lda band structure calculations . @xcite the same holds true for different forms of the projection operator technique . @xcite yet one must face the fact that the various approximations remain uncontrolled and therefore have to be reconsidered from case to case . when one is aiming at controlled approximations , two different routes offer themselves . one is bound to quantum monte carlo ( qmc ) calculations , e.g. , ref . . this is a simple and straightforward way of dealing with the many - body problem . yet it suffers from one serious shortcoming , the so - called sign problem when treating fermions . therefore , we are pursuing the second route . it starts from hartree - fock band structures and improves them by including correlations by means of many - particle techniques similar to the ones employed for molecules . @xcite thereby one can take advantage of program packages like wannier which has been written by shukla _ et al . _ @xcite or crystal which originates from torino . @xcite both provide hartree - fock bands as well as wannier orbitals for crystals . these orbitals are especially suited for an _ ab initio _ treatment of electron correlations because of their local character . @xcite a number of correlation calculations have been performed . @xcite a local hamiltonian approach @xcite was shown to improve hartree - fock energy bands substantially . suhai used toyozawa s electronic polaron model @xcite to repartition mp2 pair energies to estimate quasiparticle band structures . @xcite by inserting the orbital energies into the energy - dependent self - energy , this model was shown to be a special case of the outer valence green s functions ( ovgfs ) @xcite that were derived in terms of crystal orbitals by liegener . @xcite by applying the approximation to the self - energy of igarashi _ et al . _ , @xcite quasiparticle band structures were obtained by albrecht _ et al._. @xcite in the present paper , we derive the algebraic diagrammatic construction ( adc ) scheme @xcite for crystals . the adc scheme has proven to be superior to the ovgf method in molecular studies , @xcite and numerous works have been carried out over the last two decades , e.g. , refs . , including studies of oligomers and clusters chosen to model infinite chains or crystals . @xcite the adc scheme is a method to approximate the feynman - dyson perturbation series for the self - energy and contains sums of certain proper and improper diagrams to infinite order . @xcite the basic properties of the adc scheme and their derivations , as applied to molecules , are found in refs . . among those size - extensivity is an important one since it is crucial when solids are considered . furthermore , the adc method is known to be robust and facilitates also to study strong electron correlations due to the efficient and stable evaluation of the one - particle green s function in terms of a hermitian eigenvalue problem . in molecules strong correlations for instance occur when inner valence electrons are treated . @xcite with the crystal orbital adc ( co - adc ) devised here , the spectral representation of the one - particle green s function is obtained . we suggest a formulation of the expressions in wannier representation instead of crystal momentum representation in order to be able to exploit the local character of the correlation hole around an electron especially in the case of crystals with a large unit cell . as a first application of the co - adc method to a three - dimensional crystalline solid , we have chosen lithium fluoride which occurs in nature as the mineral griceite . lif not only has a wide range of technological applications like in x - ray monochromators or in filters for ultraviolet radiation , e.g. , ref . ( and references . therein ) , but is also interesting for a number of fundamental physical reasons . it is considered to be the `` most ionic substance '' and a prototypical insulator which manifests in its very large fundamental band gap of@xmath3 . @xcite piacentini @xcite estimates the band gap of lif to be @xmath4 . this value is refined in ref . to @xmath5 . shirley _ et al . _ @xcite communicate @xmath6 from the dissertation of himpsel . ] @xmath6 that is the largest one found in nature apart from exotic systems . some authors even consider lif to be comparable to a he ne rare - gas solid . @xcite its optical spectra show strong excitonic effects , which complicates the experimental determination of the band gap @xcite by optical spectroscopy . similarly , many - particle effects have to be accounted for in the measurement of the widths of the f@xmath0 and f@xmath7 valence bands . lif has been thoroughly studied both experimentally and theoretically . @xcite poole _ et al . _ survey early experimental and theoretical data . @xcite recent studies of lif comprise density functional theory calculations at the lda level supplemented by an inclusion of correlation effects in terms of the gw approximation . @xcite _ ab initio _ investigations comprise hartree - fock studies of ground - state properties @xcite and of the band structure . @xcite an accurate treatment of electron correlations for the ground - state properties of lif ( and other alkali halides ) has been carried out by doll and stoll . @xcite quasiparticle band structures have been obtained by kunz @xcite on the basis of the electronic polaron model @xcite and by albrecht @xcite with the green s function approach of igarashi _ et al._. @xcite the article is structured as follows : section [ sec : band ] puts forward an _ ab initio _ foundation for band structures based on the one - particle green s function . in secs . [ sec : dynamic ] and [ sec : static ] , we derive the co - adc approximation to the self - energy in terms of wannier orbitals . section [ sec : poledyson ] combines the results of the previous sections to formulate the band structure problem as a hermitian matrix eigenvalue problem . criteria for the truncation of lattice sums are discussed in sec . [ sec : onion ] . in sec . [ sec : lif_crystal ] , the application of co - adc theory to lithium fluoride is described . conclusions are drawn in sec . [ sec : conclusion ] . the nonrelativistic hamiltonian of a crystal in atomic units @xcite reads @xmath8 \\ & & { } + \frac{1}{2 } \ , { \sum\limits}_{\scriptstyle m , n=1 \atop \scriptstyle m \neq n}^n \frac{1}{|\vec r_m - \vec r_n| } \\ & & { } \stackrel{\textstyle + \frac{1}{2 } \ , { \sum\limits}_{i , j=1}^{n_0 } { \sum\limits}_{a , b=1}^m \frac{z_{\vec r_i a } z_{\vec r_j b}}{|\vec r_{\vec r_i a } - \vec r_{\vec r_j b}| } } { \qquad\scriptstyle i \neq j \lor a \neq b \hfill } \ ; , \end{array}\ ] ] where fixed nuclei are assumed . here @xmath9 is the number of electrons in the crystal , @xmath10 denotes the number of unit cells , and @xmath11 is the number of nuclei per unit cell . @xmath12 stands for the charge of nucleus @xmath13 in unit cell @xmath14 , and @xmath15 is the distance between the @xmath16-th electron and the @xmath13-th nucleus in unit cell @xmath17 . finally , @xmath18 represents the distance between the @xmath19-th and @xmath16-th electron and @xmath20 denotes the distance between the nuclei @xmath13 , @xmath21 of charge @xmath12 , @xmath22 in unit cells @xmath14 , @xmath17 . the schrdinger equation with the hamiltonian ( [ eq : operator_hamilton ] ) can be solved for the ground state in a restricted , closed - shell hartree - fock approximation , @xcite yielding a set of spin bloch orbitals @xmath23 where @xmath24 is the crystal momentum and @xmath25 is a band index . @xcite we use these orbitals to represent the hamiltonian ( [ eq : operator_hamilton ] ) in second quantization , yielding the so - called bloch or crystal momentum representation . @xcite it is decomposed according to mller and plesset @xcite into the hartree - fock part @xmath26 and the residual interaction @xmath27 , a suitable form to apply perturbation theory : @xmath28 the energy bands ( or bloch orbital energies ) are denoted by @xmath29 , and the operators @xmath30 ( @xmath31 ) create ( annihilate ) electrons in @xmath23 . the negative of the hartree - fock potential is given by @xmath32 } \ > n_{\vec k \ , i}$ ] , the occupation numbers @xmath33 being unity for occupied and zero for unoccupied ( or virtual ) bloch orbitals ( and @xmath34 ) . generally , we use the band indices @xmath35 , @xmath36 , @xmath19 , @xmath16 , to denote occupied hartree - fock bands , @xmath37 , @xmath38 , @xmath39 , @xmath40 , for unoccupied bands , and @xmath25 , @xmath41 , @xmath42 , @xmath43 , for bands which are occupied or unoccupied . the two - electron integrals in eq . ( [ eq : bloch_hamilton ] ) are defined with respect to bloch orbitals by @xcite @xmath44 the antisymmetrized two - electron integrals are @xmath45 } = v_{\vec k_1 \ , p \ ; \vec k_2 \ , q \ ; \vec k_3 \ , r \ ; \vec k_4 \ , s } - v_{\vec k_1 \ , p \ ; \vec k_2 \ , q \ ; \vec k_4 \ , s \ ; \vec k_3 \ , r}\ ] ] and consist of a coulomb term @xmath46 and an exchange term @xmath47 . all crystal momentum vectors in eqs . ( [ eq : bloch_hamilton ] ) , ( [ eq : twoint ] ) , and ( [ eq : antitwoint ] ) are restricted to the first brillouin zone . furthermore , we assume that the first brillouin zone is discretized by the born von krmn boundary conditions , or equivalently , that numerical integrations in reciprocal space are carried out in terms of a monkhorst - pack net . @xcite in both cases , integrations over the brillouin zone are approximated by finite sums . the one - particle green s function ( or particle propagator ) in terms of bloch orbitals @xcite reads @xmath48 { \left|\right.\!\psi_0^n\!\left.\right > } \ ; , \ ] ] with wick s time - ordering operator @xmath49 . @xmath50 is the exact ground state of the @xmath9 particle system . due to translational symmetry of the hamiltonian ( [ eq : operator_hamilton ] ) , the one - particle green s function depends only on one crystal momentum , @xcite i.e. , @xmath51 . similarly , the time independence of the hamiltonian ( [ eq : operator_hamilton ] ) causes the particle propagator ( [ eq : k - prop ] ) to depend only on the time difference @xmath52 . @xcite fourier transforming @xmath53 with respect to @xmath52 yields the one - particle green s function in energy space @xmath54 which can be recast in terms of the spectral or lehmann representation as @xcite [ eq : p - spectrum ] g_pq(k , ) & = & _ n \{n+1 } + & & + _ n \{n-1 } + & = & g_pq^+(k , ) + g_pq^-(k , ) . here @xmath16 numerates the excited states of the @xmath55 particle system . the negative of the pole positions in eq . ( [ eq : p - spectral ] ) is given by either the electron affinities @xmath56 or the ionization potentials @xmath57 where @xmath58 are the energies of the different excited states of the @xmath59 particle system . the summands @xmath60 are required to ensure the convergence of the fourier transformation . whenever such a factor occurs , we implicitly take the limit @xmath61 . @xcite the pole strengths in eq . ( [ eq : p - spectral ] ) are given by the transition amplitudes @xcite @xmath62 where @xmath63 denote excited states of the @xmath59 particle system with crystal momenta @xmath64 . the pole strengths can be interpreted in terms of the spectral intensities observed in photoelectron spectroscopy experiments , similarly to the molecular case discussed in ref . . to obtain electron affinities and ionization potentials from the band structure of three - dimensional crystals , one has to specify the energy of the added ( removed ) electron at the fermi level of the neutral crystal . @xcite this is usually done by introducing chemical potentials @xmath65 which are added to the pole positions of the one - particle green s function , i.e. , @xmath66 and @xmath67 . the ( fundamental ) band gap is the smallest difference between the energies for removing an electron from and attaching an electron to an @xmath9 particle system : @xmath68 here @xmath13 stands for the largest electron affinity and @xmath69 designates its smallest ionization potential . correspondingly , @xmath70 denote the energies of the electron attachment ( removal ) state with the lowest ( highest ) energy and @xmath71 represents the ground - state energy . the self - energy @xmath72 with respect to the residual interaction @xmath27 is defined by the dyson equation @xcite [ eq : dyson ] g(k , ) = g^0(k , ) & + & g^0(k , ) ( k , ) g(k , ) + = g^0(k , ) & + & g^0(k , ) ( k , ) g^0(k , ) + & + & , which can be solved formally by @xmath73^{-1 } \ ; , \ ] ] where @xmath74\ ] ] is the free green s function with respect to the mller - plesset partition ( [ eq : bloch_hamilton ] ) . in fact , for crystals with a band gap , the occupation numbers in eq . ( [ eq : freeprop ] ) are independent of @xmath24 , i.e. , @xmath75 and @xmath76 . the self - energy can be decomposed into an @xmath77 independent part , the static self - energy @xmath78 , and an @xmath77 dependent part , the dynamic self - energy @xmath79 , @xcite @xmath80 where @xmath81 . the dynamic self - energy is considered first in the ensuing sec . [ sec : dynamic ] as the static self - energy is determined using the former one in sec . [ sec : static ] . the partition ( [ eq : bloch_hamilton ] ) of the hamiltonian has the same form and meaning as the mller - plesset partition for canonical molecular orbitals @xcite used frequently in molecular physics . therefore , all equations derived on the basis of the molecular mller - plesset partition are in complete analogy to the equations in the crystalline case . the only difference between molecules and crystals is the occurrence of composite indices in the hamiltonian ( [ eq : bloch_hamilton ] ) , consisting of a crystal momentum vector @xmath24 and a band index @xmath25 which vary independently . given an equation in terms of molecular orbitals , it can be written immediately in terms of bloch orbitals by replacing all molecular orbital indices by composite indices . afterwards one can exploit ( translational ) symmetry to simplify the equation . the dynamic self - energy @xmath82 is represented in terms of the @xmath83/@xmath84-propagator ( two - particle - one - hole / two - hole - one - particle - propagator ) and possesses the spectral representation @xcite @xmath85 where @xmath86 denote the pole positions of the @xmath83/@xmath84-propagator and the @xmath87 are termed dyson amplitudes . the retarded dynamic self - energy @xmath88 and the advanced dynamic self - energy @xmath89 are associated with excitations of the @xmath59 particle system , respectively . yet the pole positions @xmath86 do _ not _ directly correspond to the energies of physical states as the ( dynamic ) self - energy is only defined in conjunction with the dyson equation ( [ eq : dyson ] ) . the algebraic diagrammatic construction scheme is a stable and efficient method to determine the spectral representation of the dynamic self - energy ( [ eq:2h1p - spectral ] ) . the construction starts by making the ansatz @xmath90 which is termed general algebraic form or adc form . the vectors @xmath91 are called modified coupling amplitudes for the crystal orbital @xmath92 . both @xmath93 and @xmath94 are hermitian matrices , @xmath94 being referred to as modified interaction matrices @xcite while @xmath93 are assumed to be diagonal . explicit expressions for these matrices are given below in eq . ( [ eq : k - co - adc2 ] ) . the adc form ( [ eq : adc - form ] ) reproduces the analytic structure of the feynman - dyson perturbation series for the dynamic self - energy as can be seen from its expansion into a geometric series : @xmath95^n \nonumber \\ \label{eq : adcanalytic } & & { } \times ( \omega \ , { { \hbox{\boldmath{$\mathbbm{1}$}\unboldmath}}}- { \hbox{\boldmath{$k$}\unboldmath}}^{\pm}(\vec k))^{-1 } \ , \vec u_q^{\pm}(\vec k ) \ ; .\end{aligned}\ ] ] the @xmath16-th order approximation to the adc form ( [ eq : adc - form ] ) is constructed by inserting the perturbation expansions @xmath96 in terms of the residual interaction @xmath97 into eq . ( [ eq : adcsuminfinity ] ) , and comparing the resulting expressions with the diagrammatic expansion of the dynamic self - energy up to @xmath16-th order . this constitutes a scheme which is denoted by adc(@xmath16 ) . note that the analytic structure imposed by eq . ( [ eq : adcanalytic ] ) occasionally requires one to associate linear combinations of the analytic expression of several diagrams with particular terms in the above expansion . in adc(2 ) and adc(3 ) , each entry of the @xmath91 and the @xmath98 in eq . ( [ eq : adc - form ] ) is characterized by one or two , respectively , arrangements of two particles and one hole ( @xmath83 ) for @xmath99 and two holes and one particle ( @xmath84 ) for @xmath89 . in brief notation these arrangements are referred to as @xmath83- and @xmath84-configurations . all such configurations that can be formed with the one - particle basis set underlying a hartree - fock calculation constitute the configuration space in adc(2 ) and adc(3 ) . the configuration space is enlarged in adc(4 ) by @xmath100- and @xmath101-configurations . @xcite in general , every second order in adc , the configuration space enlarges by the next higher excitation class . restricting our discussion to @xmath83- and @xmath84-configurations , we can rewrite eq . ( [ eq : adc - form ] ) as @xmath102 the prime on the summation symbol indicates that the summation only runs over indices with @xmath103 and @xmath104 to avoid double counting of contributions . the corresponding ansatz for @xmath105 is formally identical to eq . ( [ eq : adc - form - expanded ] ) apart from the changed occupation numbers of the band indices @xmath106 , @xmath107 , @xmath108 , @xmath109 , @xmath110 , and @xmath111 . the summation variables @xmath112 , @xmath113 , @xmath114 , and @xmath115 are crystal momentum vectors like @xmath24 . the quantities @xmath116 , @xmath117 , @xmath93 , and @xmath94 are invariant under translations by an arbitrary lattice vector , a property which is taken into account for the internal summations in the adc form ( [ eq : adc - form - expanded ] ) . explicit molecular adc equations up to fourth order are given in ref . . here we show only the adc equations for the dynamic self - energy in terms of crystal ( bloch ) orbitals up to second order [ co - adc(2 ) ] : [ eq : k - co - adc2 ] u^+_k p ; l_1 + l_2 - k i l_1 a l_2 b & = & v^*_k p l_1 + l_2 - k i [ l_1 a l_2 b ] n_l_1 + l_2 - k i |n_l_1 a |n_l_2 b , + k^+_l_1 + l_2 - k i l_1 a l_2 b ; l_1^ + l_2^ - k i^ l_1^ a^ l_2^ b^ & = & _ l_1 + l_2 - k i l_1 a l_2 a ; l_1^ + l_2^ - k i^ l_1^ a^ l_2^ b^ ( _ l_1 a + _ l_2 b - _ l_1 + l_2 - k i ) + & & n_l_1 + l_2 - k i |n_l_1 a |n_l_2 b , + c^+_l_1 + l_2 - k i l_1 a l_2 b ; l_1^ + l_2^ - k i^ l_1^ a^ l_2^ b^ & = & 0 . the equations for @xmath118 , @xmath119 , and @xmath120 are formally identical to eq . ( [ eq : k - co - adc2 ] ) apart from the change of labels as for the adaptation of eq . ( [ eq : adc - form - expanded ] ) to @xmath121 and the occupation numbers in eqs . ( [ eq : k - co - adc2-u ] ) and ( [ eq : k - co - adc2-k ] ) which are in this case @xmath122 . expression ( [ eq : k - co - adc2 ] ) is derived in ref . utilizing a gaussian basis set expansion of the bloch orbitals . the spectral representation ( [ eq:2h1p - spectral ] ) of the dynamic self - energy is obtained from the adc form ( [ eq : adc - form ] ) by solving the hermitian eigenvalue problem @xmath123 the vector notation in eqs . ( [ eq : adc - form ] ) and ( [ eq : adc - inversion ] ) is a convenient means to sum over all intermediate crystal momenta and band indices . the diagonal matrix @xmath124 contains the eigenvalues of the secular matrix @xmath98 , and @xmath125 denotes the eigenvector matrix . the eigenvalues are the pole positions of the dynamic self - energy , i.e. , @xmath126 , while the dyson amplitudes in eq . ( [ eq:2h1p - spectral ] ) are obtained via @xmath127 where the @xmath16-th column vector @xmath128 of @xmath125 is used here together with the adjoint of @xmath129 . note that , unlike co - adc(2 ) , in co - adc(3 ) ( and higher orders ) the eigenvectors @xmath128 couple several @xmath83- , @xmath84- or higher excited configurations via the modified coupling amplitudes ( [ eq : k - co - adc2-c ] ) which are nonzero in this case . we derive in the following an expansion of the dynamic self - energy in terms of generalized wannier orbitals ; such orbitals are localized within unit cells and allow one to apply cutoff criteria inside a cell and between a cell and other cells . this affords one the ability to apply fine - grained configuration selection which is independent of the actual choice of the unit cell . in particular , it enables the treatment of crystals with large unit cells , not amenable to bloch - orbital - based approaches [ sec . [ sec : onion ] ] . @xcite the generalized wannier transformation @xcite and its inverse which mediate between bloch and wannier orbitals are [ eq : wanniertrafo ] w_r ( r ) & = & _ k _ p=1^k u_p(k ) ^-ik r _ k p(r ) , + _ k p(r ) & = & _ r _ = 1^k u^*_p(k ) ^ ik r w_r ( r ) , where @xmath130 denotes the number of bands involved in the transformation and @xmath131 is a suitable unitary matrix as generated , e.g. , by the localization procedures in refs . . both bloch and wannier orbitals are normalized to unity upon integration over the entire crystal which consists of @xmath10 unit cells . generally , we denote the wannier orbital indices that result from occupied bands with @xmath132 , @xmath133 , @xmath134 , @xmath135 , @xmath3 , while the wannier orbital indices of unoccupied bands are denoted by @xmath136 , @xmath137 , @xmath138 , @xmath139 , @xmath3 . indices of wannier orbitals from occupied or unoccupied bands are referred to by @xmath140 , @xmath141 , @xmath142 , @xmath143 , @xmath3 . for an evaluation of the one - particle green s function in terms of the feynman - dyson perturbation series , it is essential that there is no mixing between occupied and virtual wannier orbitals ; otherwise wick s theorem can no longer be applied . @xcite to fulfill this requirement , the wannier transformation ( [ eq : wannier ] ) is applied to the occupied and virtual bloch orbitals separately , yielding two independent unitary matrices , one for the occupied bands and one for the virtual bands , respectively . hence , @xmath131 in eq . ( [ eq : wanniertrafo ] ) is block - diagonal . on the one hand , we want to evaluate the co - adc equations entirely in terms of wannier orbitals , in the so - called wannier or local representation ; @xcite on the other hand , band structures are defined with respect to the crystal momentum quantum number . in what follows three different transformation schemes to switch between the local representation and the crystal momentum representation are applied to the co - adc formalism and their physical and methodological implications are discussed . the dynamic self - energy @xmath144 depends on two external bloch orbitals @xmath145 and @xmath146 . hence , carrying out the inverse wannier transformation ( [ eq : invwannier ] ) for the external orbitals in eq . ( [ eq : adc - form - expanded ] ) leads to @xmath147 for the left - hand side of eq . ( [ eq : adc - form - expanded ] ) and a similar expression for the right - hand side where the internal summations @xmath112 , @xmath113 , @xmath114 , and @xmath115 still run over the brillouin zone . in order to obtain a representation of the dynamic self - energy @xmath148 entirely in terms of wannier orbitals , it is important to note that , although matrix elements such as @xmath149 in eq . ( [ eq : adc - form - expanded ] ) only depend on three independent crystal momenta , they actually describe quantities which depend on four bloch orbitals , namely , @xmath23 , @xmath150 , @xmath151 , and @xmath152 and thus a fourfold inverse wannier transformation ( [ eq : invwannier ] ) has to be applied . this transformation of eq . ( [ eq : adc - form - expanded ] ) to the wannier representation yields @xmath153 exploiting the unitarity of @xmath154 and the basic relations @xcite _ r ^ik r & = & n_0 _ k , 0 , + _ k ^ik r & = & n_0 _ r , 0 . the check accent on the quantities @xmath155 , @xmath156 , and @xmath157 in eq . ( [ eq : mwannier ] ) indicates that the sums run over lattice vectors and wannier orbital indices rather than crystal momenta and band indices as in eq . ( [ eq : adc - form - expanded ] ) . the prime on the summation symbol indicates that @xmath158 and @xmath159 must hold to avoid double counting of contributions . the corresponding equation for @xmath160 is formally identical to eq . ( [ eq : mwannier ] ) apart from the changed occupation numbers @xmath161 , @xmath162 , @xmath163 , @xmath164 , @xmath165 , and @xmath166 . one obviously hermitian form for the dynamic self - energy emerging from eqs . ( [ eq : sigmawannier ] ) and ( [ eq : mwannier ] ) is given by [ eq : mcluster ] m_pq^(k , ) & = & _ , u_p(k ) u^*_q(k ) _ ^(k)^ + & & ( - ^ - ^)^-1 _ ^(k ) , + _ ^(k ) & = & _ r e^i k r _ the vector notation in eq . ( [ eq : mcluster ] ) combines the six internal summations over the intermediate lattice vectors and wannier orbital indices @xmath167 , @xmath168 , @xmath169 , @xmath170 , @xmath171 , and @xmath172 in eq . ( [ eq : mwannier ] ) . we refer to eq . ( [ eq : mcluster ] ) as the _ supercell form _ of co - adc . the matrix @xmath173 in eq . ( [ eq : mclusterass ] ) does not explicitly dependend on @xmath24 which implies that its eigenvalues comprise all pole positions , i.e. , pole positions for each @xmath24-point . consequently , the resulting eigenvalues also do not explicitly depend on @xmath24 . multiplying with @xmath174 from the left projects on the desired set of eigenvectors . hence , the translational symmetry among the @xmath10 unit cells of the crystal is not exploited to reduce the size of the configuration space which is unfavorable in conjunction with the cutoff criteria discussed in sec . [ sec : onion ] . in practice , ( [ eq : mcluster ] ) is applied to a molecular cluster formed by @xmath10 unit cells of a crystal for which born von krmn boundary conditions are enforced . as a consequence , only @xmath10 discrete crystal momenta are in the brillouin zone . the above given transformation from the wannier representation to the crystal momentum representation is simple and robust and can be used in conjunction with almost any electron correlation method , but it suffers from finite - size effects due to the , usually , small number of unit cells which are considered in practice ; see , e.g. , ref . . the translational symmetry of the dynamic self - energy ( [ eq : mwannier ] ) can be exploited by applying a translation by @xmath175 which removes the explicit dependence of @xmath176 on two external lattice vectors such that it only depends on the difference of the lattice vectors according to @xmath177 this facilitates to remove the lattice summation over @xmath178 and the prefactor @xmath179 in eq . ( [ eq : sigmawannier ] ) which results in ( dropping the prime on @xmath180 ) @xmath181 to continue , we give up the idea of an adc form for @xmath182 , where the modified coupling amplitudes on the left and on the right are related by hermitian conjugation , as is the case in eq . ( [ eq : mcluster ] ) , and arrive at [ eq : m_k ] m_pq^(k , ) & = & _ , u_p(k ) u^*_q(k ) _ 0 ^ , + & & ( - ^ - ^)^-1 _ ^(k ) + _ ^(k ) & = & _ r ^i k r _ r ^ , where a slightly changed definition of the transformed modified coupling amplitude @xmath183 is employed which is indicated by a tilde accent on the quantities @xmath184 , @xmath185 , and @xmath186 . because of its asymmetric nature , we denote eq . ( [ eq : m_k ] ) as the _ semitransformed form_. by exploiting the translational symmetry ( [ eq : transdyn ] ) , we remove the summation over all translationally equivalent octuples consisting of the two external wannier orbitals @xmath187 and @xmath188 of the dynamic self - energy @xmath189 and two @xmath83- or @xmath84-configurations . we arrive at a summation over @xmath190 , @xmath191 , and @xmath141 in conjunction with the remaining six internal lattice vectors and wannier orbital indices . this formulation thus provides a far better starting point for reducing the number of configurations by means of the cutoff criteria of sec . [ sec : onion ] than the supercell form . using the unsymmetric adc form ( [ eq : m_k ] ) implies that the pole positions still do not dependent explicitly on @xmath24 . moreover , the modified coupling amplitudes ( [ eq : m_k_dyson_amp ] ) are , in contrast to eq . ( [ eq : mclusteradd ] ) , no longer related by hermitian conjugation ; yet , for sure , the dynamic self - energy itself stays hermitian . this way of exploiting translational symmetry in conjunction with the wannier orbitals has already been applied successfully before to devise _ ab initio _ electron correlation methods for crystals like the local hamiltonian approach @xcite or the green s function based method of albrecht _ et al._. @xcite the internal structure of the dynamic self - energy is not regarded in the two formulas ( [ eq : mcluster ] ) and ( [ eq : m_k ] ) . expression ( [ eq : m_k ] ) results from eqs . ( [ eq : sigmawannier ] ) and ( [ eq : mwannier ] ) by exploiting the translational symmetry of @xmath116 in eq . ( [ eq : transdyn ] ) and utilizing its exclusive dependence on two external wannier orbitals . however , the translational symmetry of the matrix @xmath192 in eq . ( [ eq : m_k ] ) can also be exploited additionally which already was harnessed to derive eq . ( [ eq : adc - form - expanded ] ) from its original molecular orbital formulation . we obtain [ eq : m_k_hermitian ] m_pq^+(k , ) & = & _ , u_p(k ) u^*_q(k ) _ g_1 , g_2 , g_1^ , g_2^ _ , , , ^ , ^ , ^ |u _ ; g_1 g_2 ^+ * ( k ) + & & ( - ^+(k ) - ^+(k))^-1_g_1 g_2 ; g_1^ g_2^ ^ ^ ^ + & & |u _ ; g_1^ g_2^ ^ ^ ^^+(k ) , + ( ^+(k ) + ^+(k))_g_1 g_2 ; g_1^ g_2^ ^ ^ ^ & = & _ r ^ik r ( ^+ + ^+)_0 g_1 g_2 ; r ^ g_1^ + r ^ g_2^ + r ^ , where the identities @xmath193 , @xmath194 , and @xmath195 hold . the corresponding relations for @xmath196 are formally identical to eq . ( [ eq : m_k_hermitian ] ) apart from the changed occupation numbers @xmath197 , @xmath162 , @xmath163 , @xmath198 , @xmath165 , and @xmath199 . alternatively , we can derive eq . ( [ eq : m_k_hermitian ] ) by inserting the inverse wannier transformation ( [ eq : invwannier ] ) into the co - adc equations for bloch orbitals ( [ eq : adc - form - expanded ] ) which already are fully adapted to translational symmetry . we refer to eq . ( [ eq : m_k_hermitian ] ) as the _ fully translational symmetry adapted form_. the modified coupling amplitudes ( [ eq : m_k_h_u ] ) are constructed by considering only wannier orbitals relative to an origin cell ( or reference cell ) where the hole with index @xmath132 is assumed to reside . the external orbital index @xmath141 of a modified coupling amplitude can be viewed to represent a bloch orbital @xmath200 . this orbital interacts with the @xmath83-configurations that are pinned with one lattice vector to the origin cell and extend with the two remaining lattice vectors @xmath201 and @xmath202 over up to two different unit cells . alternatively , one can use the translational symmetry of the @xmath203 to arrive at @xmath204 now the external orbital is @xmath205 which is independent of @xmath178 and consequently eq . ( [ eq : m_k_h_u_trans ] ) is interpreted to describe the interaction of a wannier orbital in the origin cell with @xmath83-configurations centered around all lattice vectors @xmath178 which are combined to give an intermediate @xmath83-configuration with a total crystal momentum @xmath206 . the matrix @xmath207 in eq . ( [ eq : m_k_h_kc ] ) describes the coupling among the @xmath83-configurations . one of the lattice vectors of the first triple of wannier orbitals is pinned to the origin cell , and the two other lattice vectors @xmath201 and @xmath202 are offsets to it . the remaining three lattice vectors belong to a wannier orbital in an arbitrary cell @xmath178 and two further wannier orbitals in the cells @xmath208 and @xmath209 relative to the former . obviously the entire eq . ( [ eq : m_k_h_kc ] ) can be interpreted to describe the interaction of a @xmath83-configuration of wannier orbitals centered around the origin cell with another @xmath83-configuration with crystal momentum @xmath24 . due to the full exploitation of translational symmetry , also of the intermediate @xmath83-configurations , the fully translational symmetry adapted form yields the smallest configuration space in conjunction with the cutoff criteria of sec . [ sec : onion ] . in order to give explicit co - adc expressions for the matrices showing up in eq . ( [ eq : m_k_hermitian ] ) , we transform the hamiltonian ( [ eq : bloch_hamilton ] ) with the help of formula ( [ eq : invwannier ] ) to the wannier representation . a similar decomposition as in eq . ( [ eq : bloch_hamilton ] ) is chosen for a subsequent application of perturbation theory : @xmath210 with @xmath211 being the diagonal elements of the fock matrix @xmath212 . here @xmath213 } \ , n_{\vec r \ , \tau } + ( 1 - \delta_{\vec r_1 \ , \varrho \ ; \vec r_2 \ , \sigma } ) \ , f_{\vec r_1 \ , \varrho \ ; \vec r_2 \ , \sigma}$ ] comprises the negative of the hartree - fock potential and also the off - diagonal elements of the fock matrix . note that @xmath212 is block - diagonal and thus the wannier orbitals are still separated into occupied and virtual wannier orbitals . using the inverse generalized wannier transformation ( [ eq : invwannier])as is done throughout this article corresponds to the use of the full fock matrix as zeroth - order hamiltonian . in addition , a perturbative expansion of the off - diagonal elements in the fock matrix is applied in succession of the inverse wannier transformation . this conduct is equivalent to employing the partitioning of the fock matrix introduced in eq . ( [ eq : wannier_hamilton ] ) . the one - electron interaction is mediated by the off - diagonal fock matrix elements , and the two - electron interaction is represented by the two - electron integrals ; let the one - electron interaction be treated using feynman diagrams up to @xmath19-th order and account for diagrams which exclusively involve two - electron interactions of highest order @xmath16 . this separation facilitates to take higher - order terms into account for the slowlier convergent series ; the resulting scheme is denoted by co - adc(@xmath19,@xmath16 ) . the co - adc(2,2 ) approximation for wannier orbitals is obtained similarly to the case of the co - adc(2 ) approximation for bloch orbitals ( [ eq : k - co - adc2 ] ) and reads [ eq : r - co - adc2 ] |u^+_r ; g_1 g_2 g_3 & = & v^*_r g_1 [ g_2 g_3 ] n_g_1 g_3^ ^ ( _ g_2 + _ g_3 p_3 - _ g_1 ) n_g_1 |n_g_2 |n_g_3 , + the equations for @xmath214 , @xmath215 , and @xmath216 are formally identical to eq . ( [ eq : r - co - adc2 ] ) upon modifying the labels according to the adaptation of eq . ( [ eq : m_k_hermitian ] ) to @xmath89 and replacing the occupation numbers in eqs . ( [ eq : m_r_h_u ] ) and ( [ eq : m_r_h_k ] ) by @xmath217 . up to second order , the perturbative treatment of the off - diagonal elements of the fock matrix does not introduce any new diagrams for the dynamic self - energy and thus @xmath218 remains diagonal . @xcite however , in a formulation which is invariant under unitary transformations in the occupied and virtual space , this is no longer the case because the full fock matrix needs to be chosen as the zeroth - order hamiltonian in place of @xmath219 then . the static self - energy has not been determined so far . it is represented by all those diagrams in the diagrammatic expansion of the self - energy where the external points of the diagrams correspond to equal times . @xcite in a strict second - order treatment of electron correlations , using bloch orbitals ( [ eq : bloch_hamilton ] ) , the static self - energy is zero because the first static self - energy diagrams arise the earliest in third order . moreover , it turns out that the diagrammatic series for the static self - energy does not converge reasonably in many cases . @xcite however , a self - consistent solution is possible @xcite utilizing @xcite @xmath220 } \\ & & { } \times \biggl [ \frac{1}{2\pi{{\rm i } } } \oint g_{sr}(\vec k^{\prime } , \omega ) { \>\textrm{d}}\omega \biggr ] \ ; , \end{array}\ ] ] here given in terms of bloch orbitals . the contour integration in eq . ( [ eq : statics ] ) runs along the real axis and closes in the upper complex @xmath77-plane , hence enclosing only the poles of the advanced green s function @xmath221 in eq . ( [ eq : p - decomp ] ) . the self - consistent solution of eq . ( [ eq : statics ] ) is computationally expensive . yet a stable and efficient dyson expansion method ( dem ) for determining the static self - energy is devised in ref . . there , the first two terms of the dyson expansion ( [ eq : dysonexp ] ) are inserted into eq . ( [ eq : statics ] ) . the term @xmath222 in eq . ( [ eq : statics ] ) cancels the result of the contour integration over the free green s function @xmath223 . carrying out the contour integration over the product of the two free green s functions and the static self - energy yields [ eq : statick ] _ pq^(k ) & = & _ k^ _ r , s a_k p k q , k^ s k^ r _ sr^(k^ ) + b_pq(k ) , + a_k p k q , k^ s k^ r & = & v_k p k^ r [ k q k^ s ] ( _ k^ r - _ k^ s)^-1 , + b_pq(k ) & = & _ k^ _ r , s v_k p k^ r [ k q k^ s ] q_rs(k^ ) , + q_rs(k^ ) & = & g^0_ss(k^ , ) m_sr(k^ , ) g^0_rr(k^ , ) . the @xmath224 can now be determined by inserting the spectral representation of the dynamic self - energy ( [ eq:2h1p - spectral ] ) into eq . ( [ eq : kqqp ] ) . carrying out the contour integration yields [ eq : qkl_mlk - k ] q_rs(k^ ) & = & q^+_rs(k^ ) + q^-_rs(k^ ) , + q^+_rs(k^ ) & = & _ n \{n+1 } m_s^+,(n)(k^ ) m_r^+,(n)*(k^ ) , + q^-_rs(k^ ) & = & _ n \{n-1 } m_s^-,(n)(k^ ) m_r^-,(n)*(k^ ) . in order to evaluate eq . ( [ eq : qkl_mlk - k ] ) , from now on , we resort to the adc form of the dynamic self - energy in terms of bloch orbitals ( [ eq : adc - form - expanded ] ) . by affixing a bar accent to the quantities @xmath117 , @xmath93 , and @xmath94 in all following equations of this subsection , they become formally identical to the ones obtained with the fully translational symmetry - adapted form ( [ eq : m_k_hermitian ] ) . the direct usage of eq . ( [ eq : qkl_mlk - k ] ) is in practice very time consuming because the @xmath225 matrix has to be diagonalized fully ( [ eq : adc - inversion ] ) to obtain the pole positions @xmath226 of the spectral representation ( [ eq:2h1p - spectral ] ) . for the co - adc(2 ) approximation of the dynamic self - energy in crystal momentum representation ( [ eq : k - co - adc2 ] ) , however , @xmath227 remains diagonal . this leads to @xmath228 and @xmath229 where the @xmath16-th eigenvector of @xmath230 is unity for a single @xmath83- or @xmath84-configuration , respectively , and zero otherwise . in general , the @xmath224 are determined efficiently from eq . ( [ eq : qkl_mlk - k ] ) by the inversion method or by a single - vector lanczos diagonalization to circumvent a full diagonalization of @xmath225 . @xcite the inversion method has been found to be more efficient @xcite than the single - vector lanczos diagonalization , so we concentrate on the former . we define auxiliary vectors @xmath231 as the solution of the inhomogeneous systems of linear equations @xmath232 inserting the eigenvector matrix @xmath233 from eq . ( [ eq : adc - inversion ] ) into eq . ( [ eq : vr ] ) , one obtains , by using formula ( [ eq : moddysontraf ] ) , @xmath234 which reveals the usefulness of the new vectors and allows us to rewrite eq . ( [ eq : qkl_mlk - k ] ) as follows : [ eq : qrs_vr ] q_rs^+(k^ ) & = & -v_s^+(k^ ) v_r^+(k^ ) n_k^ r n _ k^ s + ( _ k^ r - _ k^ s)^-1 + & & , + q_rs^-(k^ ) & = & v_s^-(k^ ) v_r^-(k^ ) |n_k^ r | n_k^ s - ( _ k^ r - _ k^ s)^-1 + & & . we have reduced the problem of determining the @xmath235 to the problem of determining the @xmath231 . inspecting eqs . ( [ eq : vr ] ) and ( [ eq : qrs_vr ] ) , we note that we have to solve for the @xmath236 in the large @xmath83-configuration space but only for the usually small set of all occupied orbitals . conversely , the @xmath237 have to be calculated in the small @xmath84-configuration space but for the usually large set of all virtual orbitals . the diagonal parts @xmath238 in eq . ( [ eq : vr ] ) are , at least , of a magnitude of twice the hartree - fock band gap , implying usually a diagonal dominance of the full matrices in eq . ( [ eq : vr ] ) . therefore , a solution by jacobi iterations @xcite is suggested : @xmath239 which turns out to converge rapidly . the inhomogeneous linear system of equations ( [ eq : kstaticimplicit ] ) can now be solved for @xmath240 by a matrix inversion @xcite : @xmath241 where @xmath242 , @xmath243 , and @xmath244 are composed by numerating the compound indices @xmath245 and @xmath246 in @xmath240 , @xmath247 , and @xmath248 by integer numbers @xmath35 and @xmath36 , respectively . inspecting eq . ( [ eq : ka ] ) , we note that @xmath249 couples only to the components of @xmath250 where the @xmath251 indices denote a combination of particle - hole- or hole - particle - orbitals . as all quantities in eq . ( [ eq : getsigmak ] ) are hermitian , we can further restrict the band indices to @xmath252 . as soon as the number of @xmath24-points becomes large , the solution of eq . ( [ eq : getsigmak ] ) gets cumbersome as the dimension of the system of linear equations is given by the number of entries of the upper triangle ( including the diagonal entries ) of the static self - energy matrix times the number of @xmath24-points . thus an iterative linear equations solver has to be employed . @xcite the static self - energy can also be determined directly in wannier representation . having expressed the static self - energy in terms of wannier orbitals , we obtain the static self - energy in crystal momentum representation by @xmath253 to find an approximation for @xmath254 , we insert the inverse wannier transformation ( [ eq : invwannier ] ) and its hermitian conjugate into eq . ( [ eq : statick ] ) and arrive at [ eq : getsigma ] _ ^(r ) & = & _ r^ _ , _ ^(r^ ) + & & + b_(r ) , + b_(r ) & = & _ r^ _ , q _ ( r^ ) . note that the translational symmetry of @xmath255 could only be utilized once to remove one lattice summation such that two lattice summations ( one over @xmath180 and one over @xmath256 ) show up in eq . ( [ eq : getsigma ] ) while only one summation over @xmath257 is necessary in eq . ( [ eq : statick ] ) . there is no first - order contribution to the static self - energy in eq . ( [ eq : getsigma ] ) because the static self - energy used here evolves from an inverse wannier transformation of the self - energy obtained in mller - plesset partition @xcite and a subsequent perturbative expansion with respect to the off - diagonal elements of the fock matrix . for the remaining equations ( [ eq : ka ] ) and ( [ eq : kqqp ] ) , we assume the partition ( [ eq : wannier_hamilton ] ) of the hamiltonian in wannier representation , to obtain [ eq : staticcoeff ] a_0 r , g g + r^ & = & v_0 g + r^ [ r g ] + & & ( _ g + r^ - _ g ) ^-1 + & & ( n_g + r^ |n_g - |n_g + r^ n_g ) , + q_(r^ ) & & q_r^ 0 = g_^0(0 , ) + & & m_(r^ , ) g_^0(0 , ) , with the free green s function in wannier representation given by @xmath258 \ ; . \end{array}\ ] ] to determine the @xmath259 , we introduce the spectral representation of the dynamic self - energy in wannier representation [ the analog of eq . ( [ eq:2h1p - spectral ] ) ] @xmath260 which is obtained from the adc form of the dynamic self - energy in wannier representation ( [ eq : mwannier ] ) by letting @xmath261 and diagonalizing @xmath173 analogously to eq . ( [ eq : adc - inversion ] ) . since translational symmetry can not be exploited in eq . ( [ eq : mwannier ] ) as @xmath180in contrast to @xmath24is not a good quantum number , this leads to the redundancies that have already been mentioned in conjunction with the supercell form of the dynamic self - energy in eq . ( [ eq : mcluster ] ) ; namely , the full configuration space of @xmath83- and @xmath84-configurations from the @xmath10 unit cells of the crystal is necessary here . however , the cutoff criterion of sec . [ sec : onion ] , nevertheless , can be applied . the @xmath259 can now be determined by inserting eq . ( [ eq:2h1p - spectral - r ] ) into eq . ( [ eq : qtauupsilon ] ) . carrying out the contour integration yields [ eq : qkl_mlk ] q_(r^ ) & = & q_^+(r^ ) + q_^-(r^ ) , + q^+_(r^ ) & = & _ n \{n+1 } m_^+,(n)(0 ) m_^+,(n)*(r^ ) , + q^-_(r^ ) & = & _ n \{n-1 } m_^-,(n)(0 ) m_^-,(n)*(r^ ) . we proceed as before by defining the auxiliary vectors @xmath262 setting @xmath263 , and rewrite eq . ( [ eq : qkl_mlk ] ) : [ eq : qkl - r ] q_^+(r^ ) & = & - v_^+(0 ) v_^+(r^ ) n_r^ n_0 + ( _ r^ - _ 0 ) ^-1 + & & , + q_^-(r^ ) & = & v_^-(0 ) v_^-(r^ ) |n_r^ |n_0 - ( _ r^ - _ 0 ) ^-1 + & & . the algorithm for jacobi iterations now reads @xmath264 and the inhomogeneous linear system of equations ( [ eq : staticinhomo ] ) can be solved for @xmath254 by a matrix inversion @xcite : @xmath265 where a check accent is affixed to indicate that wannier orbitals rather than bloch orbitals are used . we employ a mapping of the compound indices @xmath266 and @xmath267 to the integer numbers @xmath35 and @xmath36 , respectively , for the quantities @xmath268 , @xmath269 , and @xmath270 . note that @xmath271 is not hermitian . hence , the full static self - energy matrix has to be included in @xmath272 explicitly for several lattice vectors @xmath178 . however , the property @xmath273 holds which can be utilized to reduce the number of equations in formula ( [ eq : getsigmar ] ) . nevertheless , in many cases , the linear system of equations needs to be solved iteratively . having determined approximations for the static and the dynamic self - energy in terms of wannier orbitals in eqs . ( [ eq : sigmakr ] ) and ( [ eq : m_k_hermitian ] ) , respectively , we can finally determine the positions and strengths of the poles of the one - particle green s function from eq . ( [ eq : solvedysonformal ] ) . [ the pole positions and strengths of the green s function , using bloch orbitals to represent the self - energy ( [ eq : statick ] ) and ( [ eq : adc - form - expanded ] ) , can be obtained following a nearly identical line of argument . ] to this end , the small hermitian eigenvalue problem @xmath274 with @xmath275 being the diagonal matrix of bloch orbital energies and @xmath276 denoting eigenvectors , has to be solved self - consistently , i.e. , such that the energy @xmath77 entering the dynamic self - energy @xmath277 is identical to the resulting eigenvalue @xmath77 . to avoid the unitary matrices @xmath131 and @xmath278 in eqs . ( [ eq : m_k_h_m ] ) and ( [ eq : sigmakr ] ) arising in the following expressions , we multiply eq . ( [ eq : selfconsistent ] ) from the left with @xmath278 and insert @xmath279 between the right parenthesis and the eigenvector @xmath280 on the left - hand side of the formula . defining the translational symmetry - adapted fock matrix @xmath281 , one arrives at @xmath282 where the new quantities @xmath283 , @xmath284 , @xmath285 , and @xmath286 are related to the old ones @xmath287 , @xmath240 , @xmath288 , and @xmath289 by @xmath290 , @xmath291 , @xmath292 , and @xmath293 . inserting the adc form of the dynamic self - energy ( [ eq : m_k_h_m ] ) into eq . ( [ eq : fixedomega ] ) and defining @xmath294 yields the following form of the eigenvalue problem ( [ eq : fixedomega ] ) : @xmath295 rewriting eq . ( [ eq : fixeddynamic ] ) as @xmath296 it becomes evident that eqs . ( [ eq : unfixomega ] ) and ( [ eq : evsub ] ) can be recast as a joint hermitian eigenvalue problem @xmath297 & { \hbox{\boldmath{$b$}\unboldmath}}(\vec k ) = \left ( \begin{array}{ccc } { \hbox{\boldmath{$\bar f$}\unboldmath}}(\vec k ) + { \hbox{\boldmath{$\bar \sigma$}\unboldmath}}^{\infty}(\vec k ) & { \hbox{\boldmath{$\bar u$}\unboldmath}}^{+\dagger}(\vec k ) & { \hbox{\boldmath{$\bar u$}\unboldmath}}^{-\dagger}(\vec k ) \\ { \hbox{\boldmath{$\bar u$}\unboldmath}}^+(\vec k ) & { \hbox{\boldmath{$\bar k$}\unboldmath}}^+(\vec k ) + { \hbox{\boldmath{$\bar c$}\unboldmath}}^+(\vec k ) & { \hbox{\boldmath{$0$}\unboldmath } } \\ { \hbox{\boldmath{$\bar u$}\unboldmath}}^-(\vec k ) & { \hbox{\boldmath{$0$}\unboldmath } } & { \hbox{\boldmath{$\bar k$}\unboldmath}}^-(\vec k ) + { \hbox{\boldmath{$\bar c$}\unboldmath}}^-(\vec k ) \end{array } \right ) \ ; . & \end{aligned}\ ] ] the new matrix @xmath298 is called the _ band structure matrix_. it has to be diagonalized for several @xmath24-points yielding eigenvalues @xmath299 and eigenvectors @xmath300 . formula ( [ eq : solvepadc ] ) is similar to the result for molecules . @xcite one can envisage that the diagonalization of the subblocks @xmath301 in @xmath298 is carried out first by solving eq . ( [ eq : adc - inversion ] ) ; afterwards , the full eigenvalue problem ( [ eq : solvepadc ] ) is treated with all subblocks modified accordingly . @xcite therefore , via eq . ( [ eq : adcsuminfinity ] ) , infinitely many proper self - energy diagrams are summed before the result is put in the dyson equation ( [ eq : dyson ] ) to sum all improper self - energy diagrams that derive from those contained in the adc form . the spectral representation of the one - particle green s function ( [ eq : p - spectrum ] ) reads , in terms of the eigenpairs of the band structure matrix ( [ eq : solvepadc ] ) , @xmath302 where @xmath303 denotes the transition amplitude of the @xmath16-th state . it is given by either the @xmath304 or the @xmath305 in eq . ( [ eq : transamp ] ) whether it belongs to an @xmath306 or an @xmath307 particle state . in some cases , the external wannier orbitals of the dynamic self - energy are restricted to a single cell of the crystal , i.e. , @xmath308 , the so - called one - lattice - site approximation for the dynamic self - energy . note , however , that the @xmath83- and @xmath84-configurations in neighboring unit cells are not entirely excluded . the @xmath24-dependence , in this approximation , is solely due to the one - particle matrix @xmath309 . the static self - energy can also be evaluated in one - lattice - site approximation . then the @xmath24-dependence is exclusively due to @xmath310 and this approximation to the co - adc equations becomes similar to the dynamical mean - field theory . all equations derived so far contain certain infinite lattice sums that run over the entire crystal forming the configuration space . to obtain a meaningful approximation of the one - particle green s function , the convergence of those lattice sums must be granted . sun and bartlett @xcite study the convergence properties of the lattice sums in several quantum chemical correlation methods for ground states and exited states of crystals , based on the mller - plesset partition of the hamiltonian ( [ eq : bloch_hamilton ] ) . following the reasoning of ref . , the lattice sums in co - adc in crystal momentum representation converge . moreover , the equations in local representation also converge as they evolve from the former by means of an inverse wannier transformation ( [ eq : invwannier ] ) . although the convergence of lattice sums is granted principally , we have to devise an algorithm for their proper truncation ; @xcite i.e. , a dynamical building of the configuration space is required which means to meet a chosen accuracy in the lattice sums for a given crystal . the appropriate truncation of lattice sums is an essential ingredient of all _ ab initio _ methods for crystals and corresponds to a configuration selection procedure which ensures that the configuration space is sufficiently large for the desired accuracy of the band structure but still sufficiently small to be tractable on present - day computers . in molecular physics , configuration selection has been introduced in the context of the configuration interaction method @xcite which suffers from an exponentially growing configuration space with respect to the number of atoms in molecules . the techniques discussed in ref . are not applicable here , as they are designed with the configuration interaction method in mind , but there are analogs for crystals which are similar in spirit . the incremental scheme for ground states of crystals @xcite is based on wannier orbitals . the total configuration space of a crystal is partitioned into certain subsets which are used in a subsequent calculation of the correlation energy to yield the so - called energy increments . frequently , explicit manual configuration selection is applied . this allows the partition of the configuration space to be chosen following chemical intuition . yet the situation is much more cumbersome for band structures because the number of distinct excited states grows with the number of correlated electrons , i.e. , the number of bloch orbitals and wannier orbitals which are considered in the sums in eqs . ( [ eq : adc - form - expanded ] ) and ( [ eq : m_k_hermitian ] ) , respectively . in the local hamiltonian approach , the number of states to be calculated is restricted to a fixed number which is treated in an incremental way . @xcite albrecht _ et al . _ @xcite ensure the convergence of an incremental series for the self - energy for @xmath311 , a criterion which is independent of the number of states actually described . we devise here a configuration selection procedure that is perfectly adapted to the structure of the co - adc scheme . for the equations in terms of bloch orbitals ( [ eq : statick ] ) and ( [ eq : adc - form - expanded ] ) , @xmath24-points are chosen harnessing born von krmn boundary conditions which lead to a net of equidistant @xmath24-points . @xcite in this case , configuration selection means choosing a sufficiently dense net in conjunction with a cutoff criterion which is similar to the one for co - adc in terms of wannier orbitals ( [ eq : sigmakr ] ) and ( [ eq : m_k_hermitian ] ) discussed in the next paragraph . yet beforehand , one should interpret the impact of the number of @xmath24-points used for brillouin zone integration . using a given net of @xmath10 @xmath24-points to carry out a wannier transformation ( [ eq : wannier ] ) , co - adc in terms of bloch orbitals can be immediately represented and analyzed in terms of the supercell form ( [ eq : mcluster ] ) . however , one should emphasize that the redundancies of this representation which is utilized here for interpretation only are not present in the bloch orbital formulation of co - adc . the supercell form affords a tight - binding interpretation : within the supercell , the number of distinct neighbors to the origin cell is restricted by the volume of the supercell . distinct interaction terms among the cells , i.e. , the fock matrix and two - electron integrals , are thus also resticted and the number of @xmath24-points employed can thus be understood to imply the number of nearest - neighbor cells treated in the interaction terms . these arguments generalize the tight - binding arguments given , e.g. , in ref . , in conjunction with a single brillouin zone integration of a crystal - momentum - dependent function to the multi - dimensional case . configuration selection in crystal momentum representation is not intuitive and does not allow a fine - grained selection of configurations within unit cells and between an origin cell and its neighbor cells . for crystals with large unit cells this is a significant restriction . to obtain a suitable cutoff criterion for co - adc in terms of wannier orbitals ( [ eq : sigmakr ] ) and ( [ eq : m_k_hermitian ] ) , we evaluate the second - order diagram of the retarded dynamic self - energy @xcite and examine the summand therein : @xmath312 } \ , v_{\vec r \ , \sigma \ ; \vec g_1 \ , \kappa \ ; [ \vec g_2 \ , \alpha \ ; \vec g_3 \ , \beta]}^ * \over \omega - \varepsilon_{\vec g_2 \ , \alpha } - \varepsilon_{\vec g_3 \ , \beta } + \varepsilon_{\vec g_1 \ , \kappa } } \ ; n_{\vec g_1 \ , \kappa } \ ; \bar n_{\vec g_2 \ , \alpha } \ , \bar n_{\vec g_3 \ , \beta } \ ; .\ ] ] the most delocalized occupied orbitals are the valence orbitals . therefore , the occupied valence orbitals couple to the most important @xmath83-configurations hence , @xmath77 can be assumed to be of the order of the band gap . the denominator in eq . ( [ eq : selfsecond ] ) is regarded to be constant to a reasonable degree . it is thus considered as being part of the cutoff threshold . therefore , if @xmath313 } \ , v_{\vec r \ , \sigma \ ; \vec g_1 \ , \kappa \ ; [ \vec g_2 \ , \alpha \ ; \vec g_3 \ , \beta]}^ * \ ; n_{\vec g_1 \ , \kappa } \ ; \bar n_{\vec g_2 \ , \alpha } \ , \bar n_{\vec g_3 \ , \beta}$ ] is above a certain cutoff threshold for some combination of @xmath314 and @xmath315 , we include the corresponding @xmath83-test - configuration @xmath316 in the configuration space . if the product is small for all external wannier orbitals of the self - energy , @xmath317 is neglected completely by setting all two - electron integrals containing this configuration exactly to zero . by this , only a finite range of the residual coulomb interaction , which is not treated in hartree - fock approximation , is explored . the same arguments hold for the selection of @xmath84-configurations for the evaluation of the advanced dynamic self - energy , if eq . ( [ eq : selfsecond ] ) is adapted in the same way as eq . ( [ eq : r - co - adc2 ] ) is adapted to @xmath214 , @xmath215 , and @xmath216 . convergence of the configuration space , i.e. , the number of unit cells contributing configurations to be considered in the cutoff criterion ( [ eq : selfsecond ] ) , is expected to be sufficiently quick . van der waals dispersion interactions which are present both in the ground state of the @xmath9 particle system and the states of the @xmath59 particle system are nearly equal and thus almost cancel in band structure calculations . however , the effect of the extra coulomb charge that arises in @xmath59 particle states in contrast to the @xmath9 particle ground state is long range . the potential of an extra point charge falls off like @xmath318 , where @xmath319 is the dielectric constant of the solid . this causes an overall polarization of the crystal which can be accounted for by adjusting the chemical potential @xmath65 of the added or removed electron accordingly by using a continuum approximation . @xcite this procedure is not specific to the co - adc theory advocated here to treat the short - range interactions , but is required generally . having introduced a convenient method to truncate the configuration space for co - adc , we have to elucidate the physical impact of configuration selection on band structures . to this end , we consider the fully translational symmetry adapted form ( [ eq : m_k_hermitian ] ) . the modified coupling amplitudes ( [ eq : m_k_h_u ] ) in the band structure matrix ( [ eq : solvepadc ] ) of a crystal carry two independent crystal lattice vectors @xmath320 and @xmath321 . if @xmath320 and @xmath321 are sufficiently far away from the origin cell , all two - electron matrix elements ( [ eq : m_r_h_u ] ) which contribute to the modified coupling amplitude @xmath322 fall for all @xmath141 below the configuration selection threshold and are therefore zero . because the modified interaction matrices ( [ eq : m_r_h_c ] ) vanish in the co - adc(2,2 ) approximation and eq . ( [ eq : m_k_h_kc ] ) thus denotes a diagonal matrix , the eigenvalue associated with the configuration @xmath323 is simply given by the diagonal element @xmath324 , with an eigenvector being unity on position @xmath325 and zero elsewhere . consequently , such an eigenstate of @xmath298 is removed by deleting the column and the row @xmath326 in @xmath298 because the transition amplitude ( [ eq : transamp ] ) vanishes and thus it does not contribute to the spectral representration of the one - particle green s function ( [ eq : spectral_co - adc ] ) . a further consequence of configuration selection involves the lattice summations over @xmath178 in eqs . ( [ eq : m_k_h_u ] ) and ( [ eq : m_k_h_kc ] ) . in the first place , we consider a crystal with a macroscopic lattice constant which consists of @xmath10 unit cells . as a result , only fock and two - electron matrix elements that involve wannier orbitals from the origin cell are selected . therefore , in eqs . ( [ eq : m_k_h_u ] ) and ( [ eq : m_k_h_kc ] ) , the lattice sums run over only the single term for @xmath327 with @xmath328 and the resulting band structure matrix ( [ eq : solvepadc ] ) is consequently independent of @xmath24 . as the brillouin zone contains @xmath10 crystal momenta , we obtain a @xmath10-fold degenerate spectrum similarly to simple tight - binding models . @xcite in typical crystals , interactions with neighboring unit cells are important . yet a sufficiently large supercell which consist of @xmath329 unit cells can be chosen such that interactions between supercells are negligible leading to a @xmath330 fold degenerate spectrum . in other words , a cutoff threshold , to be used for the selection of fock matrix elements and @xmath83- and @xmath84-configurations , implies a definition of degeneracy of the physical states in our model of the crystal . in order to selectively diagonalize the band structure matrix , one uses an iterative eigenvalue solver , e.g. , a block - lanczos algorithm , @xcite that is capable of exploiting the sparsity of @xmath298 . the most expensive step of the block - lanczos algorithm is a matrix times vector product denote the number of rows of the square matrix @xmath249 . then , specifically for the co - adc(2 ) approximation in terms of bloch orbitals ( [ eq : adc - form - expanded ] ) , ( [ eq : k - co - adc2 ] ) , and ( [ eq : solvepadc ] ) , a single matrix times vector operation requires @xmath331 = o(n_{\mathbf f(\vec k ) } \ , n_{\mathbf k^+(\vec k)})$ ] flops . by affixing bar accents to the matrices showing up in the expression for @xmath332 , we immediately obtain the corresponding formula for the number of flops required in the co - adc(2,2 ) approximation for wannier orbitals ( [ eq : m_k_hermitian ] ) , ( [ eq : r - co - adc2 ] ) , and ( [ eq : solvepadc ] ) . ] between @xmath298 and lanczos vectors which determines the overall performance of the eigenvalue solver . @xcite to investigate the scaling of a selective computation of eigenpairs , we consider a supercell consisting of two unit cells of an original lattice . let us assume a crystal with a macroscopic lattice constant to investigate the asymptotic scaling behavior of the problem . then , the configuration selection method of the previous paragraphs selects only configurations local to the individual unit cells . hence , the resulting configuration space scales linearly with the system size , i.e. , the number of atoms per unit cell . this is a necessary condition for the selective diagonalization of the band structure matrix ( [ eq : solvepadc ] ) to scale linearly as well . the total fock matrix of the supercell also is block - diagonal . similarly , the modified coupling amplitudes ( [ eq : m_k_h_u ] ) only contain two nonzero blocks which describe the coupling of @xmath83- and @xmath84-configurations to the fock matrix of each of the two constituting unit cells of the supercell such that the total band structure matrix decomposes into two subproblems that can be solved independently . a single matrix times vector product for the supercell requires twice as many floating point operations than are needed for a matrix times vector product for one of the two unit cells in the supercell . hence , the computation of matrix times vector products scales linearly with the system size . ) with a given accuracy , is assumed to be the same for a single unit cell and a supercell consisting of two unit cells . ] yet doubling the system size also usually means that we are interested in twice as many excited states . hence , the overall effort to determine all excited states of a crystal in a given energy range scales quadratically . it is a _ quadratically _ scaling problem , where linear scaling can only be achieved by an _ a priori _ restriction to a few excited states of the system . for typical crystals , the fock matrix of a supercell is not block - diagonal which implies another doubling of the number of floating point operations upon doubling the system size . yet this factor cancels as the number of @xmath24-points needed for a given accuracy of the integration over the brillouin zone of the supercell is halved because the volume of its brillouin zone is half the volume of the brillouin zone which corresponds to the original crystal lattice . lithium fluoride crystallizes in a face - centered - cubic ( fcc ) rocksalt structure described by the space group @xmath333 as shown in fig . [ fig : lifgeom ] . the crystal lattice has a two - atomic basis li(0,0,0 ) and f(@xmath334 ) , given in units of the lattice constant is reported therein which is consistent with early and recent experimental values . in order to allow comparison , we adopt the previously used value of @xmath335 . although the deviation is small , it has a quite noticeable influence on the fundamental band gap which decreased by @xmath336 when changing from @xmath337 to @xmath338 in a recent density functional theory study at the lda level . @xcite ] @xmath339 . to investigate a lif crystal , we have chosen the simplest possible _ ab initio _ description . the main purpose here is to demonstrate the feasibility of the co - adc formulas for a three - dimensional crystalline solid . thus , only a minimal gaussian sto-6 g basis set @xcite is used which describes the lithium atom by a @xmath340 $ ] contraction and the fluorine atom by @xmath341 $ ] . each shell of the sto-6 g basis set is constructed by fitting six gaussian functions to each occupied orbital of the isolated atoms . @xcite the @xmath342 shell of a lithium atom is rather diffuse due to the single outer valence electron . as the lithium atoms are ionized in lif crystals , we remove the two most diffuse gaussian functions in the li@xmath7 contraction because they do not reflect the physical situation of a compact li@xmath343 ion , and arrive at a @xmath344 $ ] contraction scheme . the hartree - fock equations are solved self - consistently by means of the wannier program @xcite which directly yields wannier orbitals . a finite cluster of unit cells is utilized as support for the wannier orbitals of the origin cell which , in our case , consist of up to third - nearest - neighbor cells ( 43 unit cells alltogether ) . as the program determines only occupied wannier orbitals , crystal projected atomic orbitals ( crystal paos ) @xcite@xmath345 have been devised on the basis of the projected atomic orbitals introduced by saeb and pulay @xcite and also used by hampel and werner . @xcite the resulting virtual functions are subsequently lwdin orthonormalized . @xcite for the co - adc(2,2 ) method ( [ eq : m_k_hermitian ] ) , ( [ eq : r - co - adc2 ] ) , and ( [ eq : solvepadc ] ) , the fock matrix and the two - electron integrals in wannier representation , entering the hamiltonian ( [ eq : wannier_hamilton ] ) , need to be obtained by a transformation of the corresponding quantities in terms of gaussian basis functions . this transformation is presently the bottleneck in practical computations since the wannier program @xcite carries it out for clusters only and neither uses translational nor point group symmetry . at present , this permits only the use of a minimal basis set for a three - dimensional lif crystal . the subsequent correlation calculations with the co - adc program @xcite are very efficient and as rapid as typical molecular adc calculations . .convergence of the top of the valence bands @xmath346 , the bottom of the conduction band @xmath347 , the fundamental band gap @xmath348 , and the bandwidth of the f@xmath0 valence band complex @xmath349 of a lif crystal . `` cells '' designates the number of unit cells taken into account in co - adc(2,2 ) calculations of the quasiparticle band structure where zero refers to the hartree - fock result . unity denotes the inclusion of the @xmath83- and @xmath84-configurations from the origin cell only in correlation calculations . 13 and 19 indicate the additional inclusion of configurations from nearest- and next - nearest - neighbor cells , respectively . for a lif crystal , @xmath346 and @xmath347 are both situated at the @xmath350 point . all data are given in electronvolts . [ cols="^,^,^,^,^",options="header " , ] the hartree - fock band structure that results from this minimal basis is shown in fig . [ fig : lif_19cell ] . all band energies are given with respect to the top of the valence bands which is located at the @xmath350 point and is set to zero . below zero there is a complex of three f@xmath0 valence bands shown as a closeup in fig . [ fig : lif_19cell_valence ] . they originate from the three f@xmath0 valence energy levels of the lif molecule around @xmath351 where the lowest ionization potential of the molecule is doubly degenerate . @xcite upon crystallization , this degeneracy is lifted due to interactions with neighboring atoms and , except for the high symmetry lines @xmath350l and @xmath350x , one observes three distinct f@xmath0 valence bands . as a minimal basis set is utilized , there is only a single conduction band of mainly li@xmath7 character . a slight admixture of f@xmath0 character to this band is observed in the w k panel by comparing with the band structure resulting from a larger basis set , e.g. , the one of ref . , where a conduction band of f@xmath0 character mixes with the li@xmath7 band . nevertheless , the single conduction band represents the energetically lower edge of the conduction - band complex well . lif has a direct band gap ; i.e. , the maximum of the f@xmath0 valence bands and the minimum of the li@xmath7 conduction band are located at the same crystal momentum , the @xmath350 point , here . - and @xmath84-configurations involving wannier orbitals in the origin cell and the nearest- and next - nearest - neighbor cells . ] in order to investigate electron correlations , we start from a single unit cell , to form the @xmath83- and @xmath84-configurations entering eq . ( [ eq : m_k_hermitian ] ) for @xmath116 , respectively . successively , configurations which contain wannier orbitals in nearest- and next - nearest - neighbor cells are considered . this leads to three distinct quasiparticle band structures whose convergence with respect to the number of unit cells where electron correlations are regarded explicitly is monitored for the following key quantities : the top of the valence bands , the bottom of the conduction band , and the band gap and the width of the f@xmath0 valence - band complex [ tab . [ tab : cablif ] ] . the results are compared with the plain hartree - fock data ( with configurations out of `` 0 '' unit cells ) . taking into account electron correlations of a single unit cell causes a slight upwards shift of the valence bands by @xmath352 while the conduction band essentially remains unchanged . upon inclusion of configurations in the nearest - neighbor cells , the band structure changes drastically , leading to a significant reduction of the band gap by @xmath353 . accounting for configurations which extend to next - nearest - neighbor cells has three times less impact on the band gap . therewith , it can clearly be seen that the major contributions of electron correlations have been covered and that the effect of configurations which extend to wannier orbitals beyond next - nearest - neighbor cells will yield much less significant changes of the quantities than those discernible in tab . [ tab : cablif ] . as pointed out in sec . [ sec : onion ] , the effect of the more distant unit cells can be determined using a continuum approximation . @xcite the observed correlation corrections in tab . [ tab : cablif ] with a growing number of neighboring unit cells result primarily from the formation of a quasiparticle and the polarization of the surrounding by the coulomb charge of the added particle or hole . @xcite note that , additionally , the beneficial basis set extension ( bse ) @xcite effect is noticeable in solids which leads to an improved description of electron correlations with an increasing number of unit cells taken into account . . ] the quasiparticle band structure of lif is displayed in fig . [ fig : lif_19cell ] for the f@xmath0 valence bands and the li@xmath7 conduction band . the former bands are additionally displayed on an enlarged scale in fig . [ fig : lif_19cell_valence ] . the data are taken from the computation with configurations from alltogether 19 unit cells . valence and conduction bands do not shift by the same amount , upwards and downwards , respectively . instead , the former bands are much more strongly influenced by electron correlations than the latter bands . this is due to the fact that the occupied wannier orbitals are spatially less extended than the virtual wannier orbitals . for electrons in compact orbitals , the hartree - fock approximation becomes progressively worse due to an insufficient description of the coulomb hole . @xcite although the quasiparticle bands are considerably shifted with respect to the corresponding hartree - fock bands , they essentially keep their width . we also observe this effect for ( hf)@xmath354 chains . @xcite this is in contrast to the significant reduction of bandwidths observed for quasiparticle band structures of covalently bonded polymers like _ trans_-polyacetylene @xcite and covalently bonded crystals like diamond , @xcite silicon @xcite , and germanium . @xcite the _ ab initio _ description of a lif crystal using a minimal basis set provides valuable insights . however , we can not expect full quantitative agreement neither at the hartree - fock nor at the correlation level . the accuracy of our approach can be determined by comparing with experimental and theoretical data from related studies . the fundamental band gap is accessible by photoelectron spectroscopy and has been measured to be @xmath355 . another experimentally accessible quantity , the width of the f@xmath0 valence - band complex , was determined to lie in the range @xcite of @xmath356@xmath357 and thus is inconclusive , unfortunately . with our sto-6g - like basis set , we find a hartree - fock band gap of @xmath358 [ tab . [ tab : cablif ] ] which deviates considerably from the value of @xmath359 communicated by kunz @xcite or the value of @xmath360 reported by albrecht @xcite who both use larger basis sets . for the width of the f@xmath0 valence - band complex , we find @xmath361 [ tab . [ tab : cablif ] ] in the hartree - fock approximation which is appreciably larger than the value of @xmath362 from kunz @xcite and the value of @xmath363 from albrecht . @xcite including electron correlations , the quasiparticle band gap reduces to @xmath364 [ tab . [ tab : cablif ] ] and thus falls short by @xmath365 compared to the experimental value . albrecht @xcite finds @xmath366 , and calculations of kunz @xcite yield a theoretical band gap of @xmath367 , in good agreement with the recent measurements of @xmath355 . for the f@xmath0 valence band complex , we obtain a width of @xmath368 [ tab . [ tab : cablif ] ] which is nearly unchanged compared to the hartree - fock value , a fact that has also been found in previous studies of bulk lif . @xcite consequently , our value of the width remains much higher than the value of @xmath369 from kunz @xcite or the value of @xmath370 from albrecht . @xcite the above results show that the co - adc theory can be applied without problems to compute the energy bands of three - dimensional crystalline solids . what remains to be improved considerably is the transformation of the fock matrix and the two - electron integrals from the gaussian basis set representation to a representation in terms of wannier orbitals . in this article , a general local orbital _ ab initio _ green s function method is devised for band structures . the band structure is given by the pole positions of the one - particle green s function in terms of bloch orbitals which in turn is evaluated efficiently utilizing dyson s equation . to approximate the self - energy up to @xmath16-th order , we devise a crystal orbital formulation of the well - established algebraic diagrammatic construction ( adc ) @xcite which is termed crystal orbital adc ( co - adc ) . @xcite the pole search of the one - particle green s function is recast into a hermitian eigenvalue problem which is a numerically stable and efficient formulation , permitting us to explore strong correlations that occur for the energetically lower lying bands of crystals . @xcite wannier orbitals are used in internal summations in the self - energy and allow us to exploit the fact that electron correlations are predominantly local . the lattice summations in the co - adc equations must be truncated to render the problem tractable . to this end , we devise a configuration selection procedure which can equally well be used in conjunction with the calculation of ionization potentials and electron affinities of large molecules to speed up computations . configuration selection is shown to lead to a definition of degeneracy among the states of a crystal . the derivation of the local orbital co - adc theory sets out from the equations in terms of bloch orbitals . afterwards internal indices are transformed to a wannier representation . we consider this line of argument to be compelling due to the close analogy of the equations in crystal momentum representation to the equations of molecular physics . alternatively , the derivation can be conducted by starting from co - adc in terms of wannier orbitals . a transformation of the equations to crystal momentum representation is then carried out by utilizing the wannier transformation ( [ eq : wannier ] ) . the co - adc computer program @xcite has been developed for computations based on the co - adc(2,2 ) approximation . it requires the fock matrix and the two - electron integrals in local representation which are obtained by an integral transformation using the wannier orbitals from a hartree - fock calculation with the _ ab initio _ program wannier . @xcite at present , this transformation poses a bottleneck because the current implementation has been designed for molecular clusters and does not exploit any symmetries . therefore , only a minimal basis set is used to study a lithium fluoride crystal in order to demonstrate that the method is working . a fundamental band gap of @xmath371 and a bandwidth of the f@xmath0 valence - band complex of @xmath368 is obtained . these values are to compare with the experimental data of @xmath355 and @xmath356@xmath357 and the theoretical results of @xmath372 , @xmath367 , and @xmath373 , @xmath374 for the band gap and the bandwidth , respectively . increasing the size of the basis set will improve the co - adc(2,2 ) results . yet it is demonstrated that the co - adc theory is well suited for tackling the problem of calculating energy bands of solids with controlled approximations . we would like to point out that the block of @xmath83-configurations in the band structure matrix ( [ eq : solvepadc ] ) is much larger than the block of @xmath84-configurations . in order to reduce the computational effort , the advanced part and the retarded part of the one - particle green s function can be independently evaluated diagrammatically . this leads to a different , so - called non - dyson co - adc scheme , which decouples the computation of occupied and virtual bands completely @xcite and results in two independent hermitian matrix eigenvalue problems to be solved . we would like to thank jochen schirmer for fruitful discussions and critically reading the manuscript . is indebted to thomas sommerfeld for providing his complex - symmetric block - lanczos algorithm and helpful advice .

the lacerta ob1 ( i lacertae ) association was discovered by @xcite as an aggregate of dispersing early - type stars , with an expansion time scale of a couple million years . in the review article on nearby o associations , @xcite listed a distance of 600 pc for lacob1 , estimated by means of h@xmath0 photometry by @xcite , which had quite a large uncertainty because of the presence of pre - main sequence ( pms ) objects , and a possible age spread among member stars . a relatively recent distance determination by @xcite , derived from the @xmath1 data , yielded an average distance of @xmath2 pc . a noticeable distance range is obviously expected for a nearby association , which by itself has a typical size extent of a few hundred parsecs . with a distance less than 400 pc , lacob1 ranks among the nearest ob associations in the solar neighborhood , and forms a part of the gould belt system . the interstellar matter associated with the gould belt is organized into a giant expanding ring , called the lindblad ring @xcite , in whose periphery lie the local stellar associations , including lacob1 @xcite . early radio observations of lacob1 in the 21 cm line of neutral hydrogen were carried out by @xcite , @xcite , and @xcite . @xcite divided lacob1 into two subgroups , lacob1a and lacob1b , on the basis of stellar proper motions and radial velocities . the entire lacob1 is centered around ra=22h35 m and decl=@xmath3 , and covers the large sky region @xmath4 and @xmath5 @xcite . the subgroup lacob1b is considered younger and more concentrated , distributed within a @xmath6 radius centered around @xmath7 , whereas the older lacob1a extends over the remaining region . @xcite listed 15 stars for lacob1a , and 11 stars for lacob1b . the lacob1b harbors the only o star in the region , 10lac ( o9v ; hip111841 ) . identified a total of 96 @xmath1 members for lacob1 , including 1 o , 35 b , 46 a , 1 f , 8 k , 3 m - type stars , 1 carbon star ( hip116681 ) and 1 star without spectral information ( hip111762 ) . table 1 lists these 96 stars together with their 2mass jhks photometry . the first column is the @xmath1 number , followed by ( 2 ) and ( 3 ) the star s coordinates , ( 4 ) the apparent @xmath8 magnitude , ( 5 ) @xmath9 color , ( 6 ) parallax and ( 7 ) the proper motions . columns ( 8) , ( 9 ) , and ( 10 ) are 2mass magnitudes . column ( 11 ) gives the spectral type , and the last column ( 12 ) provides some information gathered from simbad . gave a comprehensive reference list for lacob1 on the following data : ( 1 ) distance . for instance , @xcite estimated 368 pc and 603 pc , while @xcite obtained 417 pc and 479 pc for the subgroups lacob1a and lacob1b , respectively . ( 2 ) proper motions , @xmath10 mas yr@xmath11 , and @xmath12 mas yr@xmath11 . ( 3 ) radial velocity . @xcite obtained the peak around @xmath13 km s@xmath11 , whereas the @xmath1 input catalogue gave an average of @xmath14 km s@xmath11 . ( 4 ) expansion age ( e.g. , @xmath15 myr , @xcite @xcite ) . ( 5 ) stellar rotation ( e.g. , @xcite @xcite ) . ( 6 ) photometric ( e.g. , @xmath16 , @xcite @xcite ) and spectroscopic ( e.g. , @xcite @xcite , @xcite @xcite , @xcite @xcite ) studies . also useful is the review by @xcite on the physical properties and dynamical evolution of ob associations , including lacob1 . despite a considerable number of massive member stars , lacob1 is relatively devoid of cloud material . two regions both being remnant molecular clouds are known to have had recent star - forming activities , namely the bright - rimmed cloud lbn437 @xcite and the comet - shaped cloud gal110@xmath1713 @xcite . figure [ fig : lacob1co ] shows the molecular co emission in the region @xcite , along with the @xmath1 members @xcite , herbig ae / be stars and classical t tauri stars ( cttss ) @xcite in the lacob1 region . lbn437 is at the edge of an elongated molecular cloud complex kh149 @xcite , also known as gal96@xmath1715 @xcite , and on the border of the hii region s126 @xcite excited by 10lac ( see fig . [ fig : lbn437 ] ) . the southern end of lbn437 is forked into two condensations sharing the same mean radial velocity . condensation a contains a cold , elliptical dense core traced by nh@xmath18 emission , and is associatd with optical reflection nebula and luminous young stars , whereas the less massive condensation b appears not associated with any optical stars @xcite . stars associated with condensation a include lkh@xmath19233 ( or v375lac ) , a herbig a4e star @xcite showing h@xmath19 , [ oi]6300 , and [ sii]6717 emission lines in the spectrum @xcite . lkh@xmath19233 was noticed by @xcite to be an ae / be star associated with nebulosity that was later resolved by near - infrared speckle interferometry to be @xmath20 au in size @xcite . given @xmath21 for lkh@xmath19233 , and assuming a luminosity class v ( later known not to be valid ) , hence @xmath22 , @xcite estimated a distance 140860 pc to lkh@xmath19233 , depending on the adopted value of optical extinction . [ fig : lkha233dss2red ] shows the region around lkh@xmath19233 , including other fainter emission - line stars lkh@xmath19230 , lkh@xmath19231 , lkh@xmath19232 , and the luminous star hd213976 @xcite . lkh@xmath19233 is the exciting source of a series of bipolar herbig - haro objects @xcite , including hh398 and hh808 through hh814 , that stretch a few parsecs in length along roughly the direction of 65/245 ( * ? ? ? * see fig . [ fig : lkha233 ] ) . note that @xcite adopted a distance of 880 pc to lkh@xmath19233 , apparently taken from @xcite based on the inference that the b1.5 v star hd213976 , with @xmath23 and a distance modulus of 9.6 , has a negligible extinction @xmath24 @xcite , so should be in front of the dark cloud . in such a case , the cloud , and hence lkh@xmath19233 , should be at least 880 pc away . this inferred distance is however much farther than the recent @xmath1 value of 370 pc @xcite , thus the linear dimensions of the lkh@xmath19233 outflows derived by @xcite should be a couple of times shorter but still on parsec scales . most hh outflows are excited by low - mass pms stars , so the ones associated with lkh@xmath19233 are among the rarities to be related to intermediate - mass pms stars @xcite . optical polarimetric imaging taken by @xcite revealed a circumstellar disk roughly perpendicular to the outflows . a recent high angular resolution imaging by keck adaptive optics indicates that the bipolar jet of lkh@xmath19233 , redshifted in position angle of 69and blueshifted in 249 , is highly collimated , with an opening angle less than 10 , suggestive of an early accretion phase @xcite . this means the transition from highly collimated outflows typically seen in t tauri stars , to less collimated ones associated with massive young stars , must occur at a higher mass than the 4 m@xmath25 estimated for lkh@xmath19233 @xcite . between 10lac and lbn437 there is a group of pms stars spanning some 24@xmath26 ( about 2.6 pc ) across , most of which exhibit forbidden lines , indicative of youth @xcite . lkh@xmath19233 is located near the edge of lbn437 and , being the exciting source of herbig - haro objects , conceivably should be among the youngest . there are otherwise no cttss or herbig ae / be stars known inside the cloud @xcite . the formation of this chain of young stars lying between 10lac and the lbn437 cloud complex might be triggered by the radiation - driven implosion mechanism ( @xcite @xcite , @xcite @xcite , @xcite @xcite ) , in which the uv photons from a luminous star evaporate and compress a nearby molecular cloud . as the result , the cloud is shaped into a pillar , being illuminated to become a bright - rimmed cloud , and star formation may take place at the surface layer of the cloud . gal110@xmath1713 is an isolated and elongated cloud @xcite . the ctts bmand ( ra = 23h37m38.5 , decl = + 482412 , j2000 ) and three b - type stars associated with the cloud , hd222142 ( which illuminates the nebula vdb158 , @xcite @xcite ) , hd222046 , and hd222086 , all share common proper motions , suggesting a physical group @xcite . this cloud was not included in the study by @xcite , but given its distance ( @xmath27 pc , * ? ? ? * ) , cloud radial velocity ( @xmath28 km s@xmath11 , * ? ? ? * ) , proper motions @xcite , and association with young stars , it is likely a part of lacob1 . @xcite attributed the morphology and high star - forming efficiency ( 30% ) in gal110@xmath1713 to compression by a recent cloud collision . the cloud points to the central part of lacob1 where 10lac is located , similar to lbn437 and other cloud filaments in the region ( see fig . [ fig : lacob1co ] ) . alternative to a cloud collision is shock interaction from a supernova in lacob1b which shaped gal110@xmath1713 and prompted the formation of stars in the cloud . evidence in support of this supernova scenario comes from the b5v star hd201910 , a possible runaway star from a binary system in lacob1b when one of the component stars became a supernova ( @xcite @xcite , @xcite @xcite ) . @xcite and @xcite derived an expansion age of 1625 myr for lacob1a and 1216 myr for lacob1b , on the basis of stellar proper motions and radial velocities . the majority of the lacob1 members indeed was thought to be an evolved population ; e.g. , @xcite failed to find _ bright _ herbig ae / be stars in the region , and all the h@xmath19 emission - line stars these authors studied turned out to be classical be stars , i.e. , on the verge of turning off the main sequence . the kinematic ages of tens of myr , however , are much longer than the main sequence lifetime of @xmath29 myr for 10lac @xcite and the typical age of a few myr for the cttss in the region . star formation in lacob1 therefore appears not coeval , with the latest episode occurring no more than a few myr ago . kinematic ages of ob associations are often a factor of 2 less than those derived photometrically based on stellar evolution models @xcite . subgroups in an ob association may originate from a gravitationally unbound giant molecular cloud @xcite . likewise , members in a subgroup may be formed out of dispersing cloud fragments , or as a consequence of triggered star formation by an expanding ionization front . figure [ fig : cmd ] shows the color - magnitude diagram for lacob1a and for lacob1b . it is seen that the stars in the subgroup lacob1b form a clear main sequence , whereas those in lacob1a are much scattered . suspected that lacob1a might not be a physical group . in any case , care should be exercised when doing photometric dating . the scattering could be attributed partly to the distance spread among members , as lacob1a is nearby and occupies a large volume in space . the ageing of lacob1b is evidenced by deficiency of hi gas around s126 where 10lac and other luminous stars are located @xcite . lacob1a , if it is a real association , should contain some pms stars so represents a generation of stars younger than those in lacob1b . eventually the sequence of star formation reached gal110@xmath1713 , as we now witness . both lbn437 and gal110@xmath1713 have low dust extinction , just like the bright - rimmed clouds in oriob1 @xcite , supporting the notion that they are remnant clouds @xcite . such a low density condition is unfavorable for spontaneous , global cloud collapse . the ablation of molecular clouds also gives rise to a seemingly high star - formation efficiency , e.g. , 30% for gal110@xmath1713 @xcite , to be compared with a few percent typical in star - forming regions @xcite . the cloud morphology , age sequence , and spatial distribution of young stars in the vicinity of clouds suggest triggered star formation by stellar radiation , supernova shocks or cloud collision . in particular , if gal110@xmath1713 is indeed related to lacob1 , the triggering appears to have far - reaching influence out to hundreds of parsecs . the lacob1 association , with much of the cloud material already dissipated , is clearly ending its star - formation activity , and stages an interesting case of the starbirth sequence in an ob association . we thank the referee , carlos olano for very useful comments that much improved the quality of the article . this work has made use of the nasa s astrophysics data system , and of the simbad database , operated at cds , strasbourg , france . the grant nsc-95 - 2745-m-008 - 002 is acknowledged . abt , h. a. , & hunter , j. h. 1962 , , 136 , 381 aspin , c. , mclean , i. s. , & mccaughrean , m. j. 1985 , , 144 , 220 aveni , a. f. & hunter , j. h. 1969 , , 74 , 1021 bertoldi , f. 1989 , , 346 , 735 bertoldi , f. & mckee , c. f. 1990 , , 354 , 529 bijaoui , a. , lacoarret , m. , & granes , p. 1981 , a&as , 45 , 483 blaauw , a. 1958 , , 63 , 186 blaauw , a. 1961 , bull . netherlands , 15 , 265 blaauw , a. 1964 , , 2 , 213 blaauw , a. 1991 , in _ the physics of star formation and early stellar evolution _ , ed . c. j. lada & n. d. kylafis ( dordresht : kluwer ) , 125 blaauw , a. & morgan , w. w. 1953 , , 117 , 256 calvet , n. & cohen , m. 1978 , , 182 , 687 cappa de nicolau , c. & olano , c. a. 1990 , rev . astrof . , 21 , 269 chen , w. p. , lee , h. t. , & sanchawala , k. 2007 , in _ triggered star formation in a turbulent ism _ , eds . b. g. elmegreen & j. palous , iau symposium 237 , p. 278 clark , p. c. , bonnell , i. a. , zinnecker , h. & bate , m. r. 2005 , , 359 , 809 corcoran , m. , & ray , t. p. 1998 , , 336 , 535 coyne , g. , berley - mead , j. & kaufman , m. 1969 , , 74 , 103 ( see also nasa technical note d-5060 ) crawford , d. l. 1961 , , 133 , 860 crawford , d. l. & warren , w. h. 1976 , , 88 , 930 dame , t. m. , hartmann , dap , thaddeus , p. 2001 , , 547 , 792 de zeeuw , p. t. , hoogerwerf , r. , de bruijne , j. h. j. , brown , a. g. a. & blaauw , a. 1999 , , 117 , 354 dieter , n. h. 1960 , , 132 , 49 finkbeiner , d. p. 2003 , , 146 , 407 garmany , c. d. 1994 , , 106 , 25 gies , d. r. & bolton , c. t. 1986 , , 61 , 419 guetter , h. h. 1976 , , 81 , 1120 herbig , g. h. 1960 , , 4 , 337 hernndez , j. , calvet , n. , briceo , c. , hartmann , l. , & berlind , p. 2004 , , 127 , 1682 hernndez , j. , calvet , n. , hartmann , l. , briceno , c. , sicilia - aguilar , a. , & berlind , p. 2005 , , 129 , 856 hester , j. j. & desch , s. j. 2005 , in proceedings of _ chondrites and the protoplanetary disk _ , eds . alexander n. krot , edward r. d. scott , & bo reipurth , asp conf . 341 , 107 howard , w. e. 1958 , , 63 , 50 khavtassi , j. 1960 , _ atlas of galactic dark nebulae _ ( abastuman , ussr : astastumani astrophysical observatory ) lee , h .- t . , chen , w. p. , zhang , z. w. & hu , j. y. 2005 , , 624 , 808 lee , h .- t . & chen , w. p. 2007 , , 657 , 884 leinert , ch . , haas , m. , & weitzel , n. 1993 , , 271 , 535 lesh , j. r. 1969 , , 74 , 891 levato , h. & abt , h. a. 1976 , , 88 , 141 lindblad , p. o. , grape , k. , sandqvist , a. , & schober , j. 1973 , , 24 , 309 lynds , b. t. 1965 , , 12 , 163 mcgroarty , f. , ray , t. p. & bally , j. 2004 , , 415 , 189 odenwald , s. f. 1988 , , 325 , 320 odenwald , s. , fischer , j. , lockman , f. j. & stemwedel , s. 1992 , , 397 , 174 olano , c. a. 1982 , , 112 , 195 olano , c. a. , walmsley , c. m. & wilson , t. l. 1994 , , 290 , 235 perrin , m. d. , & graham , j. r. 2007 , astro - ph0707.2394 raimond , e. 1957 , bull . netherlands , 13 , 269 schaerer , d. & de koter , a. 1997 , , 322 , 598 shapless , s. 1959 , , 4 , 257 sugitani , k. , fukui , y. , & ogura , k. 1991 , , 77 , 59 van den bergh , s. 1957 , , 126 , 323 white , g. j. , casali , m. m. & eiroa , c. 1995 , , 298 , 594 whitney , b. 1949 , , 109 , 540 cccrrrrrrrll i d & ra & decl & & & & & & & & & + & h m s & & & & & & & & & & + hip 108508 & 21 58 56.6 & 47 59 00 & 8.82 & -0.019 & 1.32 ( 0.87 ) & -2.07 -3.38 & 8.66 & 8.70 & 8.69 & b3v & mrcyg , sp . binary + hip 109082 & 22 05 51.2 & 48 13 53 & 6.26 & -0.088 & 1.66 ( 0.53 ) & -1.62 -4.33 & 6.36 & 6.50 & 6.48 & b2v & v365lac , sp . binary + hip 110476 & 22 22 41.4 & 42 57 04 & 9.29 & 0.084 & 1.91 ( 1.15 ) & -0.27 -6.32 & 8.91 & 8.87 & 8.76 & b8 & bd+424370 + hip 110790 & 22 26 45.6 & 37 26 37 & 6.46 & -0.132 & 1.69 ( 0.95 ) & -0.82 -5.20 & 6.70 & 6.84 & 6.87 & b2v & double star + hip 110835 & 22 27 17.3 & 44 22 46 & 9.92 & 0.004 & 2.68 ( 1.45 ) & -0.38 -4.18 & 9.86 & 9.86 & 9.87 & b8 & bd+434205 + hip 110849 & 22 27 26.5 & 39 48 36 & 6.16 & -0.136 & 2.39 ( 0.71 ) & -0.53 -6.09 & 6.41 & 6.54 & 6.53 & b2v & hd212978 , double / multiple + hip 110953 & 22 28 47.7 & 46 37 51 & 9.10 & -0.054 & 2.25 ( 1.17 ) & 0.17 -3.67 & 9.15 & 9.21 & 9.21 & b9 & hd213190 + hip 111080 & 22 30 12.3 & 44 26 18 & 8.89 & 0.046 & 1.52 ( 1.11 ) & 0.99 -2.21 & 8.65 & 8.65 & 8.64 & b9 & hd213390 + hip 111104 & 22 30 29.3 & 43 07 24 & 4.52 & -0.086 & 2.38 ( 0.64 ) & -2.05 -5.76 & 4.99 & 4.70 & 4.75 & b2iv & hd213420 , sp . binary + hip 111139 & 22 30 54.4 & 43 25 40 & 8.29 & -0.002 & 1.19 ( 1.00 ) & 0.63 -3.39 & 8.22 & 8.27 & 8.27 & b9 & hd213484 + hip 111308 & 22 32 58.6 & 37 34 32 & 10.52 & 0.138 & 5.45 ( 1.81 ) & -0.79 -3.39 & 8.67 & 8.21 & 8.13 & b9 & bd+364868 + hip 111337 & 22 33 23.5 & 39 34 31 & 8.17 & -0.035 & 2.35 ( 1.13 ) & -1.53 -4.29 & 8.18 & 8.23 & 8.26 & b9v & hd213801 , double / multiple + hip 111429 & 22 34 30.2 & 40 46 30 & 7.02 & -0.119 & 3.45 ( 0.90 ) & -0.68 -3.45 & 7.19 & 7.27 & 7.29 & b1.5v & hd213976 , double + hip 111491 & 22 35 18.1 & 43 40 52 & 8.33 & -0.056 & 3.92 ( 0.93 ) & -0.10 -3.22 & 8.42 & 8.46 & 8.50 & b8 & hd214098 + hip 111546 & 22 35 52.3 & 39 38 04 & 5.73 & -0.160 & 5.10 ( 1.79 ) & 1.11 -4.39 & 5.78 & 5.85 & 5.70 & b2ve & hd214167 + hip 111576 & 22 36 16.7 & 40 05 20 & 8.30 & -0.099 & 3.07 ( 1.01 ) & 0.94 -3.22 & 8.50 & 8.58 & 8.59 & b6iv & hd214243 + hip 111589 & 22 36 22.3 & 37 50 32 & 6.84 & -0.131 & 2.92 ( 0.72 ) & -0.89 -5.34 & 7.12 & 7.23 & 7.25 & b2v & hd214263 + hip 111683 & 22 37 28.7 & 39 26 20 & 7.59 & -0.108 & 3.07 ( 0.79 ) & -0.34 -5.04 & 7.74 & 7.83 & 7.85 & b3v & hd214432 + hip 111841 & 22 39 15.7 & 39 03 01 & 4.89 & -0.207 & 3.08 ( 0.62 ) & -0.29 -5.70 & 5.30 & 5.44 & 5.50 & o9v & 10lac , hd214680 , double + hip 112031 & 22 41 28.7 & 40 13 32 & 5.25 & -0.137 & 2.34 ( 0.62 ) & -0.75 -5.90 & 5.48 & 5.58 & 5.62 & b2iii & 12lac , hd214993 , @xmath0 cep var . + hip 112144 & 22 42 55.4 & 37 48 10 & 6.43 & -0.119 & 2.71 ( 0.79 ) & -1.12 -5.30 & 6.61 & 6.67 & 6.67 & b1v & hd215191 + hip 112148 & 22 42 57.3 & 44 43 18 & 8.75 & 0.090 & 3.90 ( 1.34 ) & -2.82 -3.16 & 8.23 & 8.13 & 7.88 & b5:ne & hd215227 + hip 112167 & 22 43 03.4 & 38 46 07 & 8.68 & -0.050 & 1.80 ( 1.13 ) & 0.67 -4.97 & 8.73 & 8.78 & 8.81 & b8v & hd215211 + hip 112293 & 22 44 43.3 & 40 33 16 & 9.93 & 0.064 & 3.30 ( 1.61 ) & -1.66 -5.41 & 9.71 & 9.63 & 9.62 & b8 & bd+394920 + hip 112906 & 22 51 50.2 & 39 08 42 & 9.49 & -0.002 & 2.19 ( 1.41 ) & 0.30 -5.12 & 9.44 & 9.49 & 9.48 & b8 & bd+384883 + hip 113003 & 22 53 07.3 & 43 03 21 & 8.80 & 0.023 & 2.85 ( 1.12 ) & -0.18 -6.00 & 8.63 & 8.72 & 8.69 & b9 & hd216537 + hip 113110 & 22 54 21.2 & 43 31 43 & 7.77 & -0.043 & 3.18 ( 0.85 ) & 0.09 -4.79 & 7.76 & 7.87 & 7.85 & b8v & hd216684 + hip 113226 & 22 55 47.1 & 43 33 33 & 7.98 & 0.046 & 4.01 ( 1.61 ) & -0.33 -4.29 & 7.67 & 7.61 & 7.46 & b3v : n & v423lac , hd216851 + hip 113281 & 22 56 23.6 & 41 36 14 & 5.60 & -0.149 & 2.71 ( 0.69 ) & -0.99 -4.25 & 5.88 & 6.01 & 6.03 & b2iv & 16lac , hd216916 , @xmath0 cep var . + hip 113371 & 22 57 40.7 & 39 18 32 & 6.17 & -0.148 & 2.39 ( 0.66 ) & 0.46 -5.13 & 6.46 & 6.61 & 6.64 & b2iv / v & hd217101 + hip 113469 & 22 58 45.7 & 43 50 20 & 7.18 & -0.070 & 2.74 ( 0.72 ) & 0.67 -5.75 & 7.25 & 7.31 & 7.33 & b2:v & hd217227 + hip 113835 & 23 03 08.3 & 49 35 10 & 9.69 & -0.110 & 3.14 ( 1.37 ) & -2.33 -1.53 & 9.64 & 9.68 & 9.67 & b8 & bd+483916 + cccrrrrrrrll i d & ra & decl & & & & & & & & & + & h m s & & & & & & & & & & + hip 114097 & 23 06 32.2 & 51 04 38 & 7.42 & -0.108 & 2.41 ( 0.67 ) & -0.28 -4.13 & 7.63 & 7.73 & 7.76 & b2v & hd218344 + hip 114106 & 23 06 37.1 & 42 39 27 & 8.01 & -0.040 & 4.57 ( 0.87 ) & -1.81 -4.54 & 8.06 & 8.10 & 8.16 & b9 & v380and , hd218326 + hip 115067 & 23 18 23.6 & 47 15 42 & 8.66 & 0.130 & 3.77 ( 1.52 ) & -1.85 -2.62 & 8.54 & 8.57 & 8.64 & b9ii & hd219813 , double + hip 115334 & 23 21 38.7 & 47 21 04 & 8.41 & 0.048 & 3.28 ( 0.90 ) & -0.93 -5.50 & 8.21 & 8.17 & 8.14 & b9 & hd220210 + hip 106656 & 21 36 11.4 & 44 25 38 & 8.90 & 0.042 & 2.53 ( 0.95 ) & 1.02 -3.46 & 8.77 & 8.80 & 8.79 & a0 & hd205742 + hip 108841 & 22 02 54.6 & 39 33 46 & 8.21 & 0.022 & 1.71 ( 1.50 ) & -1.07 -2.91 & 8.08 & 8.09 & 8.10 & a0 & hd209483 , double + hip 108933 & 22 04 06.7 & 44 20 42 & 6.56 & 0.078 & 5.72 ( 0.67 ) & -0.53 -2.90 & 6.29 & 6.31 & 6.27 & a2 & hd209679 + hip 110033 & 22 17 12.0 & 40 58 05 & 9.48 & 0.071 & 1.80 ( 1.25 ) & 0.96 -3.41 & 9.36 & 9.39 & 9.38 & a0 & bd+404771 + hip 110373 & 22 21 21.1 & 41 47 48 & 8.34 & -0.023 & 3.25 ( 1.11 ) & -2.23 -5.81 & 8.36 & 8.42 & 8.42 & a0 & hd212153 , double / multiple + hip 110448 & 22 22 17.9 & 48 50 25 & 8.39 & 1.172 & 1.89 ( 0.79 ) & -2.02 -2.48 & 6.30 & 5.79 & 5.63 & k0 & bd+483697 + hip 110473 & 22 22 38.5 & 47 37 56 & 9.98 & 0.206 & 2.33 ( 1.92 ) & 0.12 -4.51 & 9.62 & 9.55 & 9.51 & a0 & bd+463676 + hip 110664 & 22 25 06.0 & 44 32 19 & 8.10 & -0.049 & 1.87 ( 0.82 ) & -1.85 -5.22 & 8.19 & 8.28 & 8.26 & a0 & hd212668 + hip 110700 & 22 25 43.7 & 38 49 26 & 9.48 & -0.022 & 5.20 ( 1.37 ) & -0.16 -6.18 & 9.45 & 9.51 & 9.51 & a0 & hd212732 + hip 110804 & 22 26 58.3 & 46 01 49 & 10.29 & 0.004 & 2.79 ( 1.32 ) & -2.01 -3.57 & 10.12 & 10.15 & 10.17 & a0 & bd+453940 + hip 110929 & 22 28 29.3 & 48 32 34 & 7.84 & 1.800 & 2.09 ( 1.02 ) & -1.82 -4.17 & 5.26 & 4.62 & 4.39 & k0iii : & hd213141 , double + hip 111022 & 22 29 31.8 & 47 42 25 & 4.34 & 1.679 & 2.80 ( 0.50 ) & -0.60 -3.37 & 1.32 & 0.41 & 0.27 & m0ii : & 5lac , hd213310/213311 + hip 111038 & 22 29 42.7 & 40 55 19 & 9.62 & 0.245 & 1.87 ( 1.57 ) & 0.52 -2.45 & 9.04 & 8.98 & 8.94 & a5 & bd+404831 + hip 111055 & 22 29 52.6 & 45 44 41 & 7.80 & 1.334 & 1.73 ( 0.87 ) & 1.41 -2.34 & 5.42 & 4.82 & 4.64 & k2 & hd213354 + hip 111207 & 22 31 45.3 & 43 16 52 & 9.95 & -0.027 & 1.71 ( 1.50 ) & -2.12 -3.00 & 9.87 & 9.88 & 9.88 & a0 & bd+424429 + hip 111292 & 22 32 43.1 & 46 16 21 & 10.10 & -0.028 & 2.19 ( 1.50 ) & 1.40 -5.29 & 10.01 & 10.05 & 10.08 & ap & hd213732 + hip 111329 & 22 33 19.9 & 42 23 42 & 9.19 & -0.025 & 3.20 ( 1.26 ) & 0.48 -3.63 & 9.14 & 9.17 & 9.14 & a0 & hd213800 + hip 111340 & 22 33 25.1 & 46 51 27 & 9.39 & 0.279 & 1.38 ( 1.21 ) & -1.02 -4.08 & 8.69 & 8.57 & 8.53 & a2 & hd213833 + hip 111375 & 22 33 48.1 & 41 40 28 & 9.94 & 0.001 & 2.07 ( 2.51 ) & 0.41 -5.11 & 9.94 & 10.01 & 9.99 & a0 & bd+404852 + hip 111552 & 22 35 54.5 & 43 41 26 & 9.66 & 0.017 & 2.36 ( 1.42 ) & -0.58 -4.08 & 9.57 & 9.59 & 9.58 & a0 & hd214179 + hip 111591 & 22 36 25.0 & 46 55 39 & 10.37 & 0.098 & 2.15 ( 1.67 ) & -0.31 -3.88 & 10.18 & 10.20 & 10.19 & a0 & hd214311 + hip 111762 & 22 38 22.2 & 52 22 06 & 9.98 & 1.363 & 8.98 ( 3.27 ) & -0.85 -5.59 & 7.26 & 6.53 & 6.34 & - & bd+513434 , double / multiple + hip 111814 & 22 38 54.9 & 36 55 42 & 9.72 & 0.092 & 4.06 ( 1.41 ) & -2.83 -4.82 & 9.42 & 9.42 & 9.41 & a2 & bd+364896 + hip 111916 & 22 40 12.5 & 38 58 25 & 9.59 & 0.140 & 2.79 ( 1.47 ) & -2.07 -6.15 & 9.27 & 9.25 & 9.22 & a2 & bd+384834 + hip 112016 & 22 41 22.9 & 50 05 33 & 7.91 & 0.024 & 2.38 ( 0.74 ) & 0.10 -2.74 & 7.78 & 7.81 & 7.80 & a0 & hd215025 + hip 112017 & 22 41 23.7 & 41 02 16 & 9.35 & 0.029 & 2.27 ( 1.33 ) & -1.04 -5.40 & 9.29 & 9.28 & 9.28 & a2 & hd214977 + hip 112182 & 22 43 15.2 & 43 46 25 & 10.50 & 0.108 & 3.77 ( 1.64 ) & -1.02 -5.11 & 10.14 & 10.15 & 10.10 & a0 & hd215271 + cccrrrrrrrll i d & ra & decl & & & & & & & & & + & h m s & & & & & & & & & & + hip 112212 & 22 43 35.3 & 32 49 19 & 7.27 & 1.628 & 1.92 ( 0.81 ) & -1.00 -3.11 & 4.14 & 3.31 & 2.90 & m0iii & qupeg , hd215290 + hip 112213 & 22 43 36.7 & 40 23 06 & 9.92 & 0.100 & 4.63 ( 1.67 ) & -2.60 -5.55 & 9.54 & 9.45 & 9.43 & a2 & bd+394917 + hip 112639 & 22 48 47.2 & 46 22 11 & 9.03 & 0.113 & 2.31 ( 1.18 ) & 0.61 -1.98 & 8.75 & 8.73 & 8.72 & a0 & hd216037 + hip 112700 & 22 49 21.6 & 45 53 50 & 8.48 & 0.135 & 2.02 ( 1.04 ) & 0.72 -2.33 & 8.13 & 8.13 & 8.10 & a0 & hd216107 + hip 112710 & 22 49 29.4 & 45 46 59 & 9.87 & 0.196 & 2.13 ( 1.55 ) & 0.40 -3.25 & 9.45 & 9.44 & 9.44 & a2 & hd216117 + hip 112805 & 22 50 40.5 & 51 06 58 & 8.30 & 0.149 & 2.00 ( 1.70 ) & -1.12 -6.07 & 8.06 & 8.09 & 8.05 & a0 & hd216255 , double / multiple + hip 113145 & 22 54 43.9 & 42 30 39 & 7.83 & 0.520 & 3.74 ( 0.84 ) & 1.50 -2.53 & 6.54 & 6.37 & 6.25 & a2 & hd216733 + hip 113187 & 22 55 13.8 & 46 22 20 & 8.30 & 0.092 & 1.91 ( 0.88 ) & 1.34 -2.92 & 8.02 & 8.02 & 8.03 & a0 & hd216797 + hip 113188 & 22 55 14.1 & 49 58 42 & 8.78 & 0.037 & 1.59 ( 1.08 ) & -1.97 -3.50 & 8.64 & 8.68 & 8.67 & a2 & hd216795 + hip 113208 & 22 55 31.5 & 43 17 36 & 8.35 & 0.044 & 2.99 ( 1.03 ) & 0.01 -3.65 & 8.20 & 8.21 & 8.26 & a2 & hd216815 + hip 113237 & 22 55 52.9 & 41 58 32 & 8.13 & 1.312 & 2.21 ( 0.96 ) & -0.68 -4.78 & 5.66 & 5.00 & 4.84 & k2 & hd216853 + hip 113288 & 22 56 26.0 & 49 44 01 & 4.99 & 1.778 & 1.74 ( 0.58 ) & 0.05 -2.87 & 1.79 & 0.96 & 0.72 & k5ib : & v424lac , hd216946 + hip 113411 & 22 58 06.7 & 41 56 04 & 9.01 & 0.141 & 2.96 ( 1.78 ) & 0.41 -3.98 & 9.06 & 9.12 & 9.07 & a2 & hd217161 , double + hip 113474 & 22 58 49.6 & 46 19 38 & 8.24 & 0.065 & 1.83 ( 0.89 ) & -0.01 -3.92 & 7.99 & 8.05 & 8.06 & a0 & hd217262 + hip 113731 & 23 01 58.4 & 47 01 00 & 8.41 & 0.022 & 1.81 ( 1.03 ) & -1.47 -2.48 & 8.21 & 8.23 & 8.24 & a2 & hd217713 + hip 113950 & 23 04 35.5 & 44 13 10 & 10.36 & 0.005 & 4.35 ( 2.67 ) & -0.06 -3.28 & 9.99 & 10.00 & 9.95 & a0 & bd+434383 + hip 114134 & 23 06 54.0 & 44 19 26 & 8.55 & 0.026 & 2.58 ( 1.00 ) & 0.92 -4.19 & 8.46 & 8.53 & 8.52 & a0 & hd218364 + hip 114153 & 23 07 05.6 & 46 07 47 & 10.17 & 0.184 & 2.13 ( 1.34 ) & -0.22 -4.95 & 9.82 & 9.78 & 9.79 & a0 & bd+454144 + hip 114441 & 23 10 37.7 & 46 22 40 & 8.20 & -0.052 & 2.22 ( 0.90 ) & 0.27 -3.76 & 8.24 & 8.27 & 8.30 & a0 & hd218844 + hip 114554 & 23 12 15.0 & 38 46 59 & 9.16 & 0.245 & 4.17 ( 1.46 ) & -2.40 -5.09 & 8.64 & 8.57 & 8.55 & a5 & hd219016 + hip 114593 & 23 12 52.4 & 48 17 01 & 9.73 & 0.103 & 2.25 ( 1.38 ) & 0.04 -5.32 & 9.57 & 9.60 & 9.56 & a0 & bd+474075 + hip 114625 & 23 13 15.1 & 45 50 25 & 10.42 & 0.166 & 2.41 ( 1.83 ) & -1.81 -3.19 & 10.17 & 10.20 & 10.18 & a2 & bd+454171 + hip 114642 & 23 13 26.7 & 35 45 43 & 8.65 & 1.100 & 4.72 ( 1.26 ) & -1.06 -6.02 & 6.79 & 6.29 & 6.19 & k0 & bd+344870 + hip 114890 & 23 16 19.0 & 50 01 41 & 9.17 & 0.284 & 2.48 ( 1.04 ) & -0.73 -2.65 & 8.69 & 8.65 & 8.58 & a0 & hd219574 + hip 114909 & 23 16 31.1 & 36 50 13 & 9.67 & 0.334 & 2.07 ( 1.42 ) & 1.46 -3.79 & 8.94 & 8.87 & 8.79 & f0 & bd+365034 + hip 115441 & 23 23 00.4 & 38 59 57 & 9.00 & 0.131 & 2.36 ( 1.14 ) & -0.89 -3.50 & 8.68 & 8.67 & 8.65 & a2 & bd+384988 + hip 116088 & 23 31 23.3 & 43 22 24 & 8.17 & 0.216 & 6.96 ( 0.94 ) & -1.97 -5.62 & 7.81 & 7.75 & 7.73 & a0 & hd221379 + hip 116135 & 23 31 53.1 & 47 32 52 & 9.79 & 0.200 & 2.24 ( 1.48 ) & 1.70 -2.58 & 9.47 & 9.48 & 9.41 & a2 & bd+464070 + hip 116411 & 23 35 24.2 & 36 44 54 & 9.19 & 1.628 & 3.19 ( 1.19 ) & 1.69 -2.50 & 4.85 & 3.88 & 3.59 & m2 & v391and , bd+355056 + hip 116457 & 23 35 50.9 & 47 25 39 & 8.73 & 1.071 & 1.52 ( 1.11 ) & 1.54 -2.04 & 6.83 & 6.33 & 6.22 & k0 & bd+464089 + hip 116522 & 23 36 51.8 & 43 18 32 & 7.81 & 1.198 & 2.02 ( 0.88 ) & 0.66 -5.07 & 5.79 & 5.18 & 5.09 & k2 & hd222018 + hip 116540 & 23 37 05.9 & 46 17 15 & 8.73 & -0.026 & 2.17 ( 1.13 ) & 1.06 -3.15 & 8.68 & 8.72 & 8.75 & a0 & hd222064 + hip 116681 & 23 38 45.1 & 35 46 21 & 9.74 & 2.100 & 5.50 ( 2.83 ) & 1.58 -3.66 & 4.83 & 3.94 & 3.10 & c & stand , hd222241 , carbon star +

over the past ten years , large ( supersymmetric ) extra dimensions have been subject to an increased attention , providing a simple framework for new cosmological ideas . motivated by scenarios such as the randall - sundrum model @xcite , physics in the presence of one extra dimension has been extensively analyzed and represents an interesting framework in which effects from the higher dimension can be understood . nevertheless , just as the dynamics of domain walls in a four - dimensional spacetime is in many ways very different to that of a cosmic string or point particle , the behavior of gravity near codimension - one branes is not representative of that around higher codimensional objects . models with two large extra dimensions , on the other hand , are capable of tracking some of the more interesting features of higher codimension objects , without introducing complications associated with higher - codimension branes . six - dimensional ( super)gravity is therefore a choice framework for the study of higher - dimensional effects , and presents remarkable features of its own . solutions of six - dimensional gravity , have been found in refs . @xcite , cosmological solutions in @xcite , and the stability of these models has been studied in @xcite . in six - dimensional supergravity , for instance , not only would the hierarchy problem be resolved if these dimensions had a submillimeter size @xcite , but if supersymmetry remained unbroken in the bulk at energies much lower that on the brane , the casimir energy could also be of the same order of magnitude as the observed four - dimensional cosmological constant , @xcite . another property of codimension - two objects relevant to the cosmological constant problem is their capacity of preserving a flat minkowski induced geometry in the presence of any tension , @xcite . codimension - two branes in the context of six - dimensional ( super)gravity can therefore provide potential resolutions of two of the most embarrassing problems of current particle physics and cosmology , namely the hierarchy and cosmological constant problems in scenarios where the weinberg s argument has different imports , @xcite . nevertheless , these models are faced with one great obstacle , the necessity of regularizing the brane before any question can be addressed , @xcite . distributional sources of matter on codimension - two and higher branes are indeed only well - defined as regularized objects , and one can therefore wonder whether any regularization - independent statement can even be made . from a field theory perspective , one expects the low - energy theory on such a brane to be independent of any high - energy regularization scheme , yet as of today , no regularization - invariant scheme has been proposed to study the effective theory on such branes . the only known work in this direction has been developed by goldberger and wise in 2001 ( see ref . @xcite ) , where it is pointed out that a field living on a six - dimensional flat spacetime will typically present a pathological behavior if coupling terms were to be introduced on a codimension - two surface . this pathology can however be removed by appropriate renormalization of the coupling constants . in this paper , we propose a direct extension to ref . @xcite , where we analyze couplings between bulk fields , free to live in the entire six dimensions , and brane fields , which are confined to codimension - two branes . the couplings between the two fields induce pathologies for both fields which can be absorbed by appropriate renormalization . similar ideas have been proposed as being useful for understanding black hole physics @xcite , post - newtonian corrections @xcite and brane localized kinetic terms @xcite . in what follows , we first review in section [ sectionii ] the problems arising when dealing with distributional sources on codimension - two branes , the regularization schemes that have been proposed in the literature as well as different sources of confusion which we clarify . we then explain the philosophy of our approach and discuss the main consequences . our strategy is applied in section [ section free theory ] , where a scalar field toy - model is considered . in particular , we analyze couplings between a bulk and a brane scalar field and discuss the renormalization procedure using two different techniques . the first one makes use of a conical cap to regularize the brane , while the other removes the divergences directly in the propagators . both methods give rise to the same renormalization group ( rg ) flows . this analysis is then extended to all possible relevant and marginal couplings in section [ sectioninteractions ] , where the three and four - point functions are computed as well as the loop diagrams . using these couplings , we also demonstrate that the theory is renormalizable . the second part of this paper is then dedicated to the physical implications . in section [ section em ] , we show how the same prescription remains valid when considering the more physical example of interactions between gravity and electromagnetism and finally explores the implications for localized kinetic terms on the brane which are relevant for models such as the dvali - gabadadze - porrati ( dgp ) in section [ section kinetic terms ] . we show how to make sense of kinetic couplings on the brane and deduce that only fixed functions of the kinetic terms are allowed by the renormalization procedure . after concluding in section [ section conclusion ] , we consider all possible local counterterms in appendix [ appendix ct ] and argue that no such counterterms can simultaneously absorb the bulk field divergences both in the bulk and on the brane , in complete agreement with our procedure . using the rg flows obtained for the brane couplings , we finish in appendix [ appendix generalcase ] by computing the arbitrary @xmath0-point function at all order in loops , and prove that it remains finite in the thin - brane limit , hence justifying that the theory is renormalizable against codimension - two divergences . in 1987 , geroch & traschen showed that strings and point particles do not belong to the class of metrics whose curvatures are well defined as distributions , @xcite . in their analysis , they considered a string in @xmath1-dimensions to be regularized as a cylinder of radius @xmath2 carrying energy density @xmath3 . the profile of the gravitational potential @xmath4 is obtained by solving poisson s equation @xmath5 , where @xmath3 vanishes outside the cylinder . although well - defined when the cylinder has a finite width @xmath2 , it turns out that the gravitational potential @xmath4 is not locally integrable in the thin - brane limit @xmath6 and so strings ( and point particles ) are not permitted as sources in @xmath1-dimensions , unless they are pure tension strings . issues arising from smoothing out codimension - two branes are also discussed in @xcite . despite the amount of attention branes have recently received , mainly motivated by string theory , the situation is unfortunately no different for those objects whose intrinsic codimension is equal or greater to two . from the string theory point of view , progress in these areas has mainly been achieved by neglecting the backreaction of such objects , treating them as test particles , the so - called probe - brane approximation . from a cosmological point view , however , such a procedure would miss some of their most important features and is hence not always satisfying . instead , much effort has been invested in specific regularizations of the theory , such as models arising from abelian - higgs theory , @xcite , thick brane regularizations , @xcite , capped branes , @xcite , intersecting branes , @xcite , codimension - two branes confined on codimension - one objects @xcite , etc ... in all of these examples , if the extra dimensions are compact or if an einstein - hilbert term is confined on the brane , one recovers four - dimensional gravity for the zero mode . however a logarithmic divergence appears in the first kaluza - klein mode as soon as the regularization mechanism is removed ( or the thin - brane limit is taken ) . understanding the significance of this divergence and the consequences for an observer on the brane represents the main objective of this paper . bulk fields away from a defect should be insensitive to the regularization procedure , but evaluating them on the defect itself requires knowledge of the internal structure of the defect . the philosophy of this paper is therefore to accept the presence of divergences in the thin - brane limit for bulk fields when evaluated on a codimension - two defect but to ensure that these divergences do not propagate into matter fields confined on the defect . we will show the validity of an effective field theory for such fields and present how observables on the brane remain finite after appropriate renormalization of the coupling constants . following the analysis of goldberger and wise @xcite , the brane - to - brane feynman propagator for a massless scalar field living in six - dimensional flat space - time with a conical singularity of deficit angle @xmath7 at @xmath8 is given by [ b - b prop ] d_k(0 ; 0)=-_0^ = - , where @xmath9 is the four - dimensional momentum along the brane direction , ( see eq . for more details . ) the brane - brane propagator is therefore divergent in the thin - brane limit for which the cutoff @xmath10 is sent to infinity ( or for large physical scale . ) in real space , on the other hand , the free propagator is finite outside the coincidence limit , and the presence of the logarithmic divergence in four - dimensional momentum space is merely a consequence to the fact that the gravitational potential of six - dimensional gravity behaves as @xmath11 in real space . more precisely , one can express the free brane - to - brane bulk propagator in real space as d(0,x;0,x)&= & _ 0^ + & = & _ 0^ = , where @xmath12 and @xmath13 represent directions tangent to the codimension - two brane and the two - point function is evaluated at @xmath8 along the normal direction . when the integral over the brane momentum @xmath9 is performed before that over the bulk momentum @xmath14 , the two - point function is finite everywhere . nevertheless , as soon as a source is considered at @xmath8 , the logarithmic behavior of the brane - brane two - point function in momentum space becomes relevant . this will be seen more concretely in what follows . we present here two general sources of confusion that usually appear when discovering these logarithmic divergences : * the first one is related to the nature of the divergence , and the order of magnitude at which it arises . any codimension - two object arises from an underlying theory ( abelian - higgs field theory , string theory or any other underlying theory ) which will naturally provide a regularization mechanism for the brane . we could hence argue that the notion of thin - brane limit is not physical and one should not be concerned about any thin - brane divergences . yet , such a argument would be going against the principles of effective field theory ( eft ) . even though the brane is expected to be regularized at some scale ( the string scale ) , we expect from eft that the low - energy physics is independent of the high - energy regularization mechanism . in other words , one should not need to understand the physics at string scale in order to understand and make predictions about the low - energy physics . * faced with this realization , one can hope that introducing brane localized counterterms should remove the logarithmic divergence present in without effecting the bulk propagator . unfortunately such an approach is too naive in this context , since brane - localized counterterms will not only affect the brane - brane propagator @xmath15 but also the brane - bulk @xmath16 and bulk - bulk propagators @xmath17 which were previously finite . any attempt to absorb the logarithmic divergence of the brane - brane propagator into brane localized counterterms will then automatically result in the introduction of further divergences . this argument is made more explicit in appendix [ appendix ct ] , where we consider the most general set of local counterterms ( that remain quadratic in the scalar field ) both in the bulk and on the brane , and show explicitly that no such counterterms will allow the propagator to be finite everywhere . in the philosophy we will follow , we will thus not attempt to make the bulk field propagator finite everywhere but will rather explore the consequences for observers on the brane . in a conical space - time , the propagator diverges at the tip of the cone . the aim of this paper is to explore the consequences for a scalar field living on the tip and coupled to a bulk field . we expect physically that 1 . bulk fields evaluated away from the brane should not depend on the regularization mechanism and thus be finite in the thin - brane limit . bulk fields evaluated on the brane itself are sensitive to the regularization procedure , since the position at which they are evaluated , the position of the brane , is dependent of the regularization . therefore we do not require bulk fields evaluated on the brane ( at @xmath8 ) to be finite in the thin - brane limit . brane fields , should have a well defined low - energy theory independent of the brane regularization . as long as the energy scales probed by an observer on the brane are much lower than that of the cutoff theory , the physics we will observe should be independent of the regularization , and hence finite in the thin - brane limit . in this paper , we will follow this philosophy carefully . the case of bulk fields with brane couplings was considered in @xcite . we here extend this analysis to matter fields confined to the brane and draw conclusions for brane observers . in this section we compute the propagator for scalar fields confined to a codimension - two brane and coupled with a bulk scalar field . we work in a six - dimensional flat space - time with a conical singularity located at @xmath8 : s^2=x^x^+r^2 + r^2 ^2 , with @xmath18 , and where @xmath19 is the deficit angle ( @xmath20 ) . a three - dimensional brane is located at the tip of the cone and @xmath21 represents the coordinates along the brane direction . we use the notation @xmath22 ; @xmath23 ; + @xmath24 and @xmath25 . this scalar field toy - model represents a good framework for the study of effective theories on codimension - two branes . the theory is composed of two coupled scalar fields , namely * the scalar field @xmath26 which symbolizes the bulk fields ( gravity , dilaton , gauge field ... ) and thus lives in six dimensions , * the brane field @xmath27 which symbolizes the matter fields living on the brane ( standard model ) and thus confined to a four - dimensional space - time . the action for this system can thus be taken to be of the form [ freeaction ] s=-^6 x @xmath28 $ ] , where for simplicity we have assumed the field @xmath26 to be massless in the bulk , but further extensions will be considered in section [ sectioninteractions ] . the coupling between the bulk and brane fields is symbolized by the term @xmath29 ( which can be set to zero ) . higher interactions will be considered in section [ sectioninteractions ] . to make contact with previous works in the literature , we first consider a specific thick - brane regularization mechanism , and show how the coupling constants can be renormalized in order for the theory to remain finite in the thin - brane limit . we then explore the renormalization mechanism in a more systematic way by analyzing the different classical two - point functions before turning to interactions and one - loop corrections in the following section . we point out that both methods will give rise to the same tree - level renomalized couplings . as a warm up , we follow a standard technique used in the literature to confine matter fields on a codimension - two brane , namely a thick - brane regularization in which the brane is no longer located at @xmath8 , but rather at @xmath30 . + the thin - brane limit is then recovered when @xmath31 . the action is regularized by [ regaction ] s=-^4 xr r & ( & 12 ( _ r ) ^2 + 12 ( _ ) ^2 + & & + ) , where for simplicity we omit for now any angular dependance . this leads to the following equations of motions : & & -_r(r (r))+k^2 = -@xmath32(r- ) + & & ( r- ) , [ bdycond ] @xmath33 being the eigenvalue of the four - dimensional dalembertian @xmath34 . integrating the first equation along the brane gives rise to the following jump condition : [ jump ] . r (r)|^^+_^-=-@xmath35 . one can solve for the bulk scalar field separately in the conical cap ( for @xmath36 ) and within the bulk ( @xmath37 ) . we choose the solution within the cap such that the scalar field remains finite at the tip @xmath8 . this leads to the following solutions ( r)=\ { lc a i_0 ( kr ) & r < + i_0 ( kr)+b k_0 ( k r ) & r > . , where @xmath38 and @xmath39 are the two modified bessel functions , or hyperbolic bessel functions , and @xmath40 remains finite as @xmath41 . for @xmath42 , we have set the coefficient of the divergent bessel function @xmath39 to zero , so that @xmath43 remains finite as @xmath44 . for @xmath37 , on the other hand , no such choice has been made and this solution is therefore independent of any other boundary conditions . these results will thus stand independently to any condition imposed on the fields away from the brane . the constants @xmath45 and @xmath46 are determined using the boundary condition and the jump condition . in the thin - brane limit , this leads to a&= & , + b&= & , where @xmath47 is the euler number , @xmath48 . for the bulk scalar field to be well - defined in the thin - brane limit , we require @xmath46 to remain finite when probing large physical scales @xmath49 . this will only be possible if the logarithmic divergence is reabsorbed into one of the coupling constants @xmath50 or @xmath51 . furthermore the scalar field @xmath27 should also be well - defined in that limit . this will thus be the case if both the following quantities remain finite b/&=&-(k^2+m^2 ) + ^-1&=&-(k^2+m^2)+ @xmath52@xmath53 . the logarithmic divergence can thus be absorbed into the coupling constants by arguing that their renormalized expression is related to their bare value by [ renormcouplingsthick ] _ 2= , = m^2=m_b^2 + , where the subscript @xmath54 represents the bare value , and @xmath3 is the physical scale . we therefore get the following renormalization group flows for the brane couplings [ rgflowthick ] = , = = . notice that we recover the same flow as ref . @xcite for the coupling @xmath55 which gives rise to a mass term for the bulk field on the brane . the renormalization of this coupling ensures that the bulk field to be finite away from the tip . however we wish to emphasize that this procedure does not get rid of the divergence of @xmath26 as @xmath56 . the point of the renormalization is to make sense of the bulk field away from the tip , however at the tip itself the bulk field diverges logarithmically as explained in section [ sectionii ] which is consistent with our philosophy . the key point here is that one can still make sense of the brane field @xmath27 ( @xmath27 is finite ) despite its coupling with @xmath26 . in other words , at low - energy , matter fields living on a codimension - two brane are independent of the regularization procedure , even though they couple to gravity and other bulk fields which are themselves ill - defined in the brane in the thin - brane limit . this is possible through adequate renormalization of the couplings . the renormalization for the coupling @xmath55 is already known from ref.@xcite . we show here how the renormalization is extended to the couplings for brane fields . in particular , we see that as soon as @xmath57 , the brane field acquires a mass . in this setup , we have used an artificial thick brane regularization . more fundamentally , we expect this defect to arise as the result of other fields ( abelian - higgs scalar and gauge field @xmath58 ) , providing a natural regularization . the fundamental theory is thus of the form = [ ] [ a _ ] [ ] [ ] e^i s_tot [ , a _ , , ] . the resulting field theory is obtained by integrating out the regularizing fields @xmath59 and @xmath60 . in this picture , we therefore expect that field loop integrations generate the same tree - level counterterms as those obtained in . in what follows , we shall recover the same results using a green s function approach , this uses the same technique as in @xcite . we will also discuss other interaction terms and show how the same renormalization procedures goes through , leading to a renormalizable theory . in this section we adopt a more field theoretic approach and require that the bulk - bulk propagator of the bulk field remains finite as well as the brane field propagator . we will proceed in three steps : 1 . we consider first of all the purely free theory for which the bulk and brane fields do not couple and all couplings vanish @xmath61 . in particular we recover the logarithmic divergence of the bulk field propagator when evaluated on the brane , but this divergence is only present in four - dimensional momentum space . we then consider the corrections to the bulk field propagators arising from the mass term @xmath55 on the brane . this situation is precisely that considered in @xcite , and we will follow the same approach . in particular , we will show how this brane coupling induces divergences in the bulk which can be removed by appropriate renormalization of the coupling @xmath55 , hence recovering the same result as in . we finally consider the corrections to both the bulk and brane field propagators induced by the coupling @xmath62 between the two fields . once again , these couplings will induce divergences which can be removed by renormalization of @xmath62 and @xmath63 , as in . \i ) _ free propagators _ + we concentrate first of all on the purely free theory given by the action [ freeaction2 ] s=-^4 xr r@xmath64 $ ] . the propagators for both fields satisfy r^(6d)_x d(x^a , x^a)&= & d(x^a , x^a)=i^(6)(x^a - x^a ) + _ x h(x^ , x^)&=&i^(4)(x^-x^ ) , where @xmath65 is the feynman propagator for @xmath26 and @xmath66 the one for @xmath27 . using a mixed - representation , momentum space along the directions @xmath21 and real space along the two extra dimensions , the propagators for both fields are simply [ bulk prop ] d_k(r , ; r, ) & = & -_n=-^+_0^ e^i n(-)j_|n|(q r)j_|n|(q r ) + h_k&=&- , where @xmath67 the bessel function of first kind , @xmath68 , and @xmath69 the four - dimensional momentum . notice that in this representation , in four - dimensional momentum space , the free propagator for @xmath26 is finite when at least one of the points is evaluated in the bulk ( @xmath17 and @xmath16 finite ) but it contains a logarithmic singularity when trying to evaluate both points on the brane . introducing a momentum cutoff scale @xmath10 in the evaluation of the propagator , one has [ b - b prop 2 ] d_k(0,0)=-_0^ = , which has the short distance singularity pointed out in @xcite . this divergence is usually not a problem since the two - point function is actually finite in real space , ( see section [ sectionii ] ) . however , as soon as a source is included at @xmath8 , the convolution of this two - point function will not be finite in real space . we therefore expect this divergence to affect the two - point function of both scalar fields when brane couplings are included . \ii ) _ corrections from the brane mass term @xmath55 _ + the previous two - point functions were that of the free theory for which the both fields were not coupled . we can now dress " these propagators with first of all the coupling @xmath55 : [ coupledaction ] s=-^4 xr r@xmath70 $ ] . ( 150,30)(-15,35)(0,51)(0,0)*@xmath71 * ( 40,60)(0,0)@xmath72 ( 40,50 ) ( 40,50)(3,0)10 ( 70,50 ) ( 70,60)(0,0)@xmath73 ( 85,49.5)(0,0)@xmath74 ( 100,60)(0,0)@xmath72 ( 100,50 ) ( 100,50)(10,0)3(1,0)7 ( 130,50 ) ( 130,60)(0,0)@xmath73 ( 140,49.5)(0,0)@xmath75 ( 155,60)(0,0)@xmath72 ( 155,50 ) ( 155,50)(10,0)3(1,0)7 ( 185,50)(10,0)3(1,0)7 ( 215,50 ) ( 215,60)(0,0)@xmath73 ( 185,60)(0,0)@xmath76 ( 185,50 ) ( 215,50 ) ( 185,40)(0,0)@xmath55 ( 230,50)(0,0)@xmath77 ( 245,60)(0,0)@xmath72 ( 245,50 ) ( 245,50)(10,0)3(1,0)7 ( 275,50)(10,0)3(1,0)7 ( 305,50)(10,0)3(1,0)7 ( 335,50 ) ( 335,60)(0,0)@xmath73 ( 305,60)(0,0)@xmath76 ( 275,60)(0,0)@xmath76 ( 305,50 ) ( 275,50 ) ( 305,40)(0,0)@xmath55 ( 275,40)(0,0)@xmath55 ( 350,51)(0,0)@xmath77 ( 370,50)(0,0)@xmath78 the propagator for the brane field @xmath27 remains unaffected while that for the bulk field @xmath26 gets modified to d_k(r , r)&=&d_k(r , r)-i_2 d_k(r,0 ) d_k(0,r)+ i^2_2 ^ 2 d_k(0,0 ) d_k(r,0 ) d_k(0,r)+ + & = & d_k(r , r)- d_k(r,0 ) d_k(0,r ) , as symbolized in fig . [ figlamdba2 ] . if @xmath15 was finite , this bulk field propagator would be finite at tree level as one should expect from usual field theory . in the presented case , the logarithmic divergence of @xmath15 needs to be absorbed in the coupling constant @xmath55 in the following way : [ lambda2 renorm ] _ 2()= , so that this coupling constant flows as = . we point out that this renormalization ensures the two - point function @xmath79 to be finite away from the brane . however , both the bulk - brane and the brane - brane two point functions remain ill - defined : both @xmath80 and @xmath81 contains a logarithmic dependence . once again , this is to be expected since evaluating the bulk two - point function on the brane requires knowledge about the exact brane position ( see section [ sectionii ] ) . \iii ) _ corrections from the coupling between the two fields _ + finally , we consider the corrections to these propagators arising from the coupling @xmath62 between the bulk and brane fields : [ coupledaction2 ] s=-^4 xr r@xmath82 $ ] . the tree level green s functions for this coupled theory are symbolically represented in fig . [ treelevel ] . ( 150,40)(20,-5 ) ( 5,60)(0,0)@xmath72 ( 5,50 ) ( 5,50)(3,0)15 ( 27.5,50 ) ( 50,50 ) ( 50,60)(0,0)@xmath73 ( 65,49.5)(0,0)@xmath74 ( 80,60)(0,0)@xmath72 ( 80,50 ) ( 80,50)(3,0)15 ( 125,50 ) ( 125,60)(0,0)@xmath73 ( 140,49.5)(0,0)@xmath77 ( 155,60)(0,0)@xmath72 ( 155,50 ) ( 155,50)(3,0)10 ( 185,49.5)(1,0)30 ( 215,50)(3,0)10 ( 245,50 ) ( 245,60)(0,0)@xmath73 ( 185,60)(0,0)@xmath76 ( 215,60)(0,0)@xmath76 ( 185,50 ) ( 215,50 ) ( 185,40)(0,0)@xmath62 ( 215,40)(0,0)@xmath62 ( 260,50)(0,0)@xmath77 ( 275,60)(0,0)@xmath72 ( 275,50 ) ( 275,50)(3,0)10 ( 335,50)(3,0)10 ( 395,50)(3,0)10 ( 305,50)(1,0)30 ( 365,50)(1,0)30 ( 425,50 ) ( 425,60)(0,0)@xmath73 ( 305,60)(0,0)@xmath76 ( 335,60)(0,0)@xmath76 ( 365,60)(0,0)@xmath76 ( 395,60)(0,0)@xmath76 ( 305,50 ) ( 335,50 ) ( 365,50 ) ( 395,50 ) ( 305,40)(0,0)@xmath62 ( 335,40)(0,0)@xmath62 ( 365,40)(0,0)@xmath62 ( 395,40)(0,0)@xmath62 ( 440,51)(0,0)@xmath77 ( 460,50)(0,0)@xmath78 ( 5,0 ) ( 5,0.1)(1,0)45 ( 50,0 ) ( 27.5,0 ) ( 65,0.5)(0,0)@xmath74 ( 80,0 ) ( 80,0.1)(1,0)45 ( 125,0 ) ( 140,0.5)(0,0)@xmath77 ( 155,0 ) ( 155,0.1)(1,0)30 ( 185,10)(0,0)@xmath76 ( 230,10)(0,0)@xmath76 ( 185,0)(3,0)15 ( 230,0.1)(1,0)30 ( 260,0 ) ( 185,-10)(0,0)@xmath62 ( 230,-10)(0,0)@xmath62 ( 185,0 ) ( 230,0 ) ( 207.5,0 ) + by summing these diagrams , we obtain the following tree - level green s functions g^_k(r , r ) & = & d_k(r , r)-^2 d_k(r,0)d_k(0,r ) h_k _ n0(-^2)^nd_k(0,0)^n h_k^n + & = & d_k(r , r)-d_k(r,0)d_k(0,r ) + & = & d_k(r , r)-d_k(r,0 ) d_k(0,r ) + g^_k&=&h_k@xmath83= + & = & . notice that there is now also a mixed two - point function for the bulk and brane fields : g^_k(r)=(r ) , = - . here again , if @xmath15 was finite , both green s functions would be finite at the tree level as one expects in usual field theory . in the presented case , the logarithmic divergence of @xmath15 needs to be absorbed in the coupling constants . these propagators will thus remain finite in the thin brane limit ( for @xmath84 ) , provided the coupling constants are renormalized as follows : [ renormcouplings ] _ 2()= , ( ) = , + [ renormcouplingsm ] m^2()=m^2()- , leading to the following rg flows = , = = . we therefore recover precisely the same relations between the bare and renormalized coupling constants as in the thick brane analysis and the same rg flows . this is a non - trivial check of our prescription . notice that @xmath85 , @xmath86 and @xmath87 are still divergent in the four - momentum representation , but the two - point function of the brane field @xmath88 has been made completely finite , and so are @xmath89 and @xmath90 , for @xmath91 . in summary , we find that by renormalizing the tree - level theory , the propagators of the field on the branes are finite , and the propagator in the bulk are only divergent when one point is evaluated on the brane ( and in the coincident limit ) . thus there is a consistent effective field theory on the brane and matter can be considered on a codimension - two brane in a completely meaningful regularization - invariant way . this will have important implications for observers on such a brane . before attacking this argument , let us consider in what follows all possible relevant and marginal interactions between a bulk and a brane field . we consider here an extension of the precedent toy - model where further couplings are taken into account . the effective field theory approach will remain completely consistent after the appropriate tree - level renormalization of the couplings . we also expect uv divergences to be present in loop corrections , but these can be dealt with through the usual uv renormalization mechanism . in order to avoid issues related to the uv divergences , ( which are independent of the fact that we consider a codimension - two brane ) , we restrict ourselves to the relevant and marginal operators . the most general brane interactions are then s=-^6x , with ^int_= _ 3 ^3 + _ 4 ^4+_3 ^2 , where the coupling @xmath92 is relevant while @xmath93 and @xmath94 are marginal . for each diagram in this theory , we expect two sorts of divergences to arise : * the ones associated with the usual uv divergences which appear in four dimensions when integrating over loops , * the short - distance divergences associated with the thin - brane limit . from standard four - dimensional eft , it is a well - known fact that the interactions of the type @xmath95 will induce uv divergences in the one - loop correction of both the two - point function and the four - point functions . these divergences can be absorbed by renormalization of the mass @xmath63 , the coupling @xmath93 as well as the wave - function . however , such divergences can be treated in a completely independent way to that arising at the tree - level in our codimension - two scenario . interactions of the form @xmath96 , for instance will typically induce divergences in the thin - brane limit which can be absorbed by appropriate renormalization of the coupling @xmath97 , this will be studied in the three - point functions in what follows . ( 150,80)(0,-40 ) ( -10,60)(0,0)@xmath98 ( 35,60)(1,0)30 ( 75,65)(2,1)20 ( 75,55)(2,-1)20 ( 50,60 ) ( 85,70 ) ( 85,50 ) ( 25,60)(0,0)@xmath74 ( 110,60)(0,0)@xmath74 ( 130,60)(0,0)@xmath99 ( 230,60)(0,0)@xmath100 ( 350,60)(0,0)@xmath101 ( 145,60)(1,0)30 ( 175,60)(1,1)15 ( 175,60)(1,-1)15 ( 160,60 ) ( 183,68 ) ( 183,52 ) ( 260,60)(1,0)15 ( 268,60 ) ( 290,60)(1,1)15 ( 290,60)(1,-1)15 ( 298,68 ) ( 298,52 ) ( 65,55)(0,1)10 ( 75,55)(0,1)10 ( 65,55)(1,0)10 ( 65,65)(1,0)10 ( 275,60)(3,0)6 ( -10,-10)(0,0)@xmath102 ( 75,-5)(2,1)20 ( 75,-15)(2,-1)20 ( 85,0 ) ( 85,-20 ) ( 25,-10)(0,0)@xmath74 ( 110,-10)(0,0)@xmath74 ( 145,-10)(3,0)7 ( 154,-10 ) ( 35,-10)(3,0)10 ( 50,-10 ) ( 330,-10)(3,0)7 ( 339,-10 ) ( 130,-10)(0,0)@xmath99 ( 225,-10)(0,0)@xmath103 ( 310,-10)(0,0)@xmath104 ( 415,-10)(0,0)@xmath77 perm.@xmath105 ( 165,-10)(1,0)15 ( 180,-10)(1,1)15 ( 180,-10)(1,-1)15 ( 188,-2.5 ) ( 188,-17.5 ) ( 265,-10)(1,1)15 ( 265,-10)(1,-1)15 ( 273,-2.5 ) ( 273,-17.5 ) ( 350,-10)(1,0)10 ( 360,-9.5)(1.5,1.5)5 ( 367,-2)(1,1)10 ( 360,-10)(1,-1)15 ( 372,2.5 ) ( 368,-17.5 ) ( 65,-15)(0,1)10 ( 75,-15)(0,1)10 ( 65,-15)(1,0)10 ( 65,-5)(1,0)10 ( 245,-9.5)(3,0)7 ( 254,-10 ) the diagrams involved in the three - point functions are summarized in fig . [ 3pt ] . summing these diagrams , we get , for @xmath106 , g^_k_1 , k_2 , k_3&=&_k_1 _ k_2 _ k_3 + & = & ( -i)@xmath107 + g^_k_1 , k_2 , k_3(r)&= & _ k_1(r ) _ k_2 _ k_3 + & = & ( -i)(6 _ 3 g^_k_1(r)g^_k_2g^_k_3 + 2_3 g^_k_1(r,0 ) g^_k_2g^_k_3 + & & + 2 _ 3g^_k_1(r)@xmath108 ) , where the factor @xmath109 arises from the first order expansion of @xmath110 . notice that after appropriate renormalization of @xmath62 , @xmath51 and @xmath55 , the propagators @xmath111 and @xmath112 have been made finite , however the bulk quantities evaluated on the brane @xmath113 and @xmath114 are a priori ill - defined . once again , this divergence would propagate into the three - point function for the brane field , had we not renormalized the couplings @xmath92 and @xmath94 . upon simplification of the previous expression , we find that the divergent part of @xmath115 is proportional to _ k_1 _ k_2 _ k_3_div & & @xmath116 , so that the coupling @xmath92 should be renormalized as [ beta3 ] _ 3()- d_k(0,0 ) + = _ 3()- , for any @xmath9 , and where we recall that the coupling @xmath62 has been renormalized in such a way that @xmath117 is finite ( see eqs . ( [ renormcouplings],[renormcouplingsm ] ) . ) the divergent part of @xmath118 is then proportional to _ k_1(r ) _ k_2 _ k_3_div & & . this divergence will thus be absorbed if the coupling @xmath94 is renormalized as [ rgflow3 ] _ 3()&= & , and so [ beta3_2 ] _ 3()=_3()+ , notice that this renormalization of @xmath94 is precisely the one that ensures the renormalized coupling @xmath92 in eq . to be independent of the four - momentum @xmath9 . after renormalization , the quantity @xmath119 is therefore finite [ 3ptfinite ] _ @xmath120=0 which corresponds to the following flows for @xmath92 and @xmath94 = = . once again , this tree - level renormalization leads to a perfectly well - defined notion of the three - point functions @xmath121 and @xmath122 , as long as @xmath72 is not evaluated on the brane and we work outside the coincidence limit . furthermore , this renormalization of these coupling constants @xmath92 and @xmath94 also ensures that the two additional three - point functions @xmath123 and @xmath124 are finite ( for @xmath125 ) . indeed , we have , for @xmath106 , & & g^_k_1,k_2,k_3(r_1,r_2)=_k_1(r_1 ) _ k_2(r_2)_k_3 + & & = -i(6 _ 3 g^_k_1(r_1 ) g^_k_2(r_2 ) g^_k_3 + 2 _ 3 g^_k_1(r_1 ) g^_k_2(r_2 ) g^_k_3(0 ) + & & + 2 _ 3 @xmath126 g^_k_3 ) + & & = -2i g^_k_3_i=1 ^ 2 g^_k_i(r_i)(3 _ 3 ( ) -_3 ( ) -_i=1 ^ 2 ) + & & = -2i g^_k_3_i=1 ^ 2 g^_k_i(r_i ) ( 3_3()-_3()_i=1 ^ 3 + & & + _ i=1 ^ 2 ) , which , in terms of the renormalized coupling constants , is clearly finite . the last three - point function can be expressed in a similar way : & & g^_k_1,k_2,k_3(r_1,r_2,r_3)=_k_1(r_1 ) _ k_2(r_2)_k_3(r_3 ) + & & = -i@xmath127(6 _ 3 + 2 _ 3 _ i=1 ^ 3 ) + & & = -2i @xmath127 _ i=1 ^ 3 , and is therefore also completely finite . this result is already non - trivial as it stands , since all four three - point functions have been made finite by simple tree - level renormalization of the two coupling constants @xmath92 and @xmath94 . we can however push this analysis even a step further by studying the implications for the four - point functions as well as the loop corrections to the two - point functions . the classical contributions to the four - point function @xmath128 are symbolically represented in fig . [ 4pt ] . summing these diagrams , we get for @xmath129 , & & g^_k_1,k_2,k_3,k_3= _ k_1 _ k_2_k_3_k_4 + & & = @xmath130,where we recall again that the factor @xmath109 and @xmath131 arise from the expansion to first order ( resp . second order ) of @xmath110 . the sum over perm@xmath132 is the one over the three permutations @xmath133 , for which @xmath134 . ( 150,90)(-10,-65)(-20,60)(0,0)[l]@xmath135 ( -20,-40)(0,0)[l]@xmath136 ( 60,60)(-1,1)10 ( 60,60)(-1,-1)10 ( 60,60)(1,1)10 ( 60,60)(1,-1)10 ( 55,55 ) ( 65,55 ) ( 55,65 ) ( 65,65 ) ( 165,60)(-1,1)10 ( 160,65 ) ( 165,60)(-1,-1)10 ( 160,55 ) ( 165,60)(1,0)15 ( 172,60 ) ( 180,60)(1,1)10 ( 185,65 ) ( 180,60)(1,-1)10 ( 185,55 ) ( 277,60)(-1,1)10 ( 277,60)(-1,-1)10 ( 277,60)(2,0)7(290,60)(1,0)10 ( 300,60)(1,1)10 ( 300,60)(1,-1)10 ( 272,55 ) ( 272,65 ) ( 295,60 ) ( 305,55 ) ( 305,65 ) ( 344,66)(-1,1)10 ( 350,60)(-1,-1)10 ( 350,60)(1,0)10 ( 360,60)(1,1)10 ( 360,60)(1,-1)10 ( 350,60)(-1,1)7(345,55 ) ( 340,70 ) ( 355,60 ) ( 365,55 ) ( 365,65 ) ( 334,76 ) ( 190,10)(-1,1)10 ( 190,10)(-1,-1)10 ( 190,10)(2,0)6(195,10)(200,10)(1,1)10 ( 200,10)(1,-1)10 ( 185,15)(185,5)(205,15)(205,5 ) ( 249,16)(-1,1)10 ( 255,10)(-1,-1)10 ( 255,10)(1,0)10 ( 265,10)(1,-1)10 ( 271,16)(1,1)10 ( 255,10)(-1,1)7(265,10)(1,1)7(250,5 ) ( 245,20 ) ( 260,10 ) ( 270,5 ) ( 275,20 ) ( 239,26 ) ( 281,26 ) ( 322,10)(-1,1)10 ( 322,10)(-1,-1)10 ( 322,10)(2,0)7(335,10)(1,0)10 ( 351,16)(1,1)10 ( 345,10)(1,-1)10 ( 317,5 ) ( 317,15 ) ( 340,10 ) ( 350,5 ) ( 355,20 ) ( 345,10)(1,1)7 ( 361,26 ) ( 30,60)(0,0 ) @xmath137 ( 87,50)(0,0)[l ] @xmath138 ( 70,60)(0,0)[l ] @xmath139 ( 133,60)(0,0 ) @xmath140 ( 230,60)(0,0)@xmath141 ( 320,60)(0,0)@xmath142 ( 385,60)(0,0)@xmath105 ( 130,10 ) ( 0,0)[l]@xmath143 ( 220,10)(0,0)[l]@xmath144 ( 290,10)(0,0)[l]@xmath144 ( 370,10)(0,0)[l]@xmath145 $ ] ( 60,-40)(-1,1)10 ( 60,-40)(-1,-1)10 ( 60,-40)(1,1)10 ( 60,-40)(1,-1)10 ( 55,-45 ) ( 65,-45 ) ( 55,-35 ) ( 65,-35 ) ( 140,-35)(-1,1)10 ( 140,-45)(-1,-1)10 ( 175,-35)(1,1)10 ( 175,-45)(1,-1)10 ( 150,-40)(1,0)15 ( 140,-35)(1,0)10(140,-45)(1,0)10(140,-35)(0,-1)10(150,-35)(0,-1)10 ( 165,-35)(1,0)10(165,-45)(1,0)10(165,-35)(0,-1)10(175,-35)(0,-1)10 ( 135,-30 ) ( 135,-50 ) ( 180,-30 ) ( 180,-50 ) ( 157,-40 ) ( 240,-40)(-1,1)10 ( 240,-40)(-1,-1)10 ( 240,-40)(2,0)6(250,-40)(1,1)10 ( 250,-40)(1,-1)10 ( 235,-35)(235,-45)(255,-35)(255,-45 ) ( 30,-40)(0,0 ) @xmath137 ( 87,-50)(0,0)[l ] @xmath138 ( 70,-40)(0,0)[l ] @xmath146 ( 190,-40 ) ( 0,0)[l]@xmath147 ( 270,-40 ) ( 0,0)[l]@xmath148 $ ] we now recall that the product of the two three - point functions @xmath149 is _ finite _ ( after appropriate renormalization of the couplings @xmath92 and @xmath94 in ) and is given by & & _ u= _ k_u_1_k_u_2_k_u_k_u_3_k_u_4_k_u + & & = -@xmath150 ^ 2@xmath151@xmath152 @xmath153 + & & = -(6 ^ 2 _ 3 ^ 2 g^_k_u+12 _ 3 _ 3 g^_k_u _ i=1 ^ 4 + 24 _ 3 _ 3 g^_k_u(0 ) + 4_3 ^ 2 g^_k_u(0)_i=1 ^ 4 + & & + 4 _ 3 ^ 2g^_k_u_i = u_1,u_2_j = u_3,u_4 + 4_3 ^ 2 ) g^_k_u @xmath130.as shown symbolically in fig . [ 4pt ] , after appropriate recombination of these different contributions , we can reexpress this four - point function as a finite product of these two renormalized three - point function plus a divergent piece which fixes the renormalization of the coupling @xmath93 : g^_k_1,k_2,k_3,k_4 & = & _ perm_u\{_u + 4(-i)^2_3 ^ 2@xmath130 } + & & + 4!(-i)_4@xmath154 + [ 4ptfunc ] & = & _ perm_u\{_u + @xmath154}. all terms in the previous expressions are finite apart from the ones in square brackets . the divergence of this term can once again be absorbed into @xmath93 using the following appropriate renormalization [ beta4 ] _ 4()-2_3 ^ 2()d_k(0,0)&= & _ 4()- + & = & _ 4()- , the renormalized coupling @xmath93 must flow as _ _ 4()=. in other words , as soon as cubic interactions between the bulk and the brane field are introduced , a quartic interaction for the brane field is spontaneously generated classically . this is familiar for standard eft . we can also carefully check using exactly the same technique as previously that this renormalization of the coupling @xmath93 also ensures that all remaining four - point functions @xmath155 , @xmath156 , @xmath157 and @xmath158 are completely finite classically ( provided the bulk field is evaluated away from the brane ) , g^_k_1,k_2,k_3,k_4(r ) & = & _ perm_u\ { g^_k_u_1,k_u_2,k_u(r)g^_k_u_3,k_u_4,k_u + & & -4g^_k_1(r)@xmath159 } + g^_k_1,k_2,k_3,k_4(r_1,r_2 ) & = & g^_k_1,k_2,k_1+k_2(r_1,r_2)g^_k_3,k_4,k_1+k_2 + & & + g^_k_1,k_3,k_1+k_3(r_1)g^_k_2,k_4,k_1+k_3(r_2 ) + ( 34 ) + & & -(4+(34 ) + & & + 4 ) @xmath160@xmath161g^_k_1,k_2,k_3,k_4(r_1,r_2,r_3 ) & = & _ perm_u\ { g^_k_u_1,k_u_2,k_u(r_u_1,r_u_2)g^_k_u_3,k_4,k_u(r_u_3 ) + & & -4@xmath162 g^_k_4 } + g^_k_1,k_2,k_3,k_4(r_1,r_2,r_3,r_4 ) & = & _ perm_u\ { g^_k_u_1,k_u_2,k_u(r_u_1,r_u_2)g^_k_u_3,k_u_4,k_u(r_u_3,r_u_4 ) + & & -4@xmath163 } . each of these four - point functions introduces a combination of the couplings which is completely finite once they have been renormalized as specified previously ( @xmath164 is finite ) . this non - trivial check ensures that our proposal makes sense at least at the classical level up to the four - point function . before discussion the general renormalizability of this theory for higher point functions , we present in what follows an insight into the situation at the quantum level , when loops are taken into account . at the loop level , we expect from standard field theory in four dimensions , that uv divergences will arise from the momentum integral over the loop . however no further divergences arise from the codimension - two nature of the theory , and the counterterms required to absorb the divergences are thus the usual one of four - dimensional field theory . to start with , we concentrate on the two - point function of the brane field @xmath27 . at first order in loops , ( second order for @xmath92 and @xmath94 ) one 1loop = _ loop(p , k ) g^_k g^_k , with the integrand @xmath165 being the sum over the different loop configurations : _ loop(p , k)&= & g^_pg^_k - p@xmath166 ^ 2 + & & g^_kg^_0@xmath167 ^ 2 + & & + 4 _ 3 ^ 2@xmath168 + ( -i)_4 g^_p , where for simplicity we have used the notation @xmath169 . this expression can be most easily interpreted as finite products of three - point functions and extra terms as symbolized in fig . [ figloops ] [ loop1 ] _ loop(p , k ) & = & -12 -12 + & & -2 g^_k - p -2g^_p + & & -2 g^_p . ( 150,100)(-50,-10)(-40,60)(0,0)[l]@xmath170 ( 125,60)(0,0)[l]@xmath171 ( 248,60)(0,0)[l]@xmath172 ( 20,10)(0,0)[l]@xmath171 ( 170,10)(0,0)[l]@xmath173 ( 365,60)(0,0)[l]@xmath148 $ ] ( 300,10)(0,0)[l]@xmath148 $ ] ( 65,60 ) ( 1,0)52 ( 91,68 ) ( 78,60 ) ( 104,60 ) ( 91,76 ) ( 180,60)(1,0)60 ( 210,60)(26,20)[t ] ( 186,60 ) ( 234,60 ) ( 210,70 ) ( 210,60 ) ( 196,60 ) ( 224,60 ) ( 192,56)(0,1)8 ( 200,56)(0,1)8 ( 192,56)(1,0)8 ( 192,64)(1,0)8 ( 220,56)(0,1)8 ( 228,56)(0,1)8 ( 220,56)(1,0)8 ( 220,64)(1,0)8 ( 315,60)(2,0)13(295,60)(1,0)20 ( 341,60)(1,0)20 ( 328,60)(26,20)[t ] ( 305,60 ) ( 328,70 ) ( 351,60 ) ( 90,0 ) ( 1,0)60 ( 120,0)(0,1)22 ( 120,30 ) ( 100,0 ) ( 140,0 ) ( 120,38 ) ( 120,11 ) ( 116,-4)(0,1)8 ( 124,-4)(0,1)8 ( 116,-4)(1,0)8 ( 116,4)(1,0)8 ( 116,18)(0,1)8 ( 124,18)(0,1)8 ( 116,18)(1,0)8 ( 116,26)(1,0)8 ( 120,0 ) ( 120,22)(220,0 ) ( 1,0)60 ( 250,30 ) ( 230,0 ) ( 270,0 ) ( 250,38 ) ( 250,0)(0,2)12 to clarify the discussion , we denote by @xmath10 , the cutoff scale associated with the codimension - two brane thickness , or equivalently with the integration over the momentum along the extra dimensions , while @xmath174 designates the standard four - dimensional momentum cut - off scale , in the loop integration is cutoff at the scale @xmath174 : @xmath175 . clearly these two scales could be associated with one another , simply representing the scale at which uv physics becomes important . however for sake of simplicity , we distinguish for now between these two quantities and assume that in general they could be different . in this scenario , the couplings are then flowing along two distinct directions @xmath10 and @xmath174 , and we focus our attention on the flow along the @xmath10 direction . on simple dimension grounds , we expect that the loop integral over @xmath176 will diverge logarithmically , and the one - loop contribution to the two - point function has thus a cutoff dependence of the form @xmath177 which should be absorbed by introduction of a mass counterterm of the form @xmath178 . this is standard procedure in four - dimensional field theory . we are however more concern here on the dependence of the other cutoff @xmath10 . @xmath94 and @xmath92 have been renormalized in , such that the three - point function @xmath121 is finite so the first line of is clearly finite . although the rest of expression includes divergent terms of the form @xmath179 , the combination involved @xmath180 $ ] is precisely the combination that appeared in the expression of four - point function ( [ 4ptfunc ] , [ beta4 ] ) , and is thus also finite . notice furthermore that the one - loop correction to the two remaining two - point functions @xmath181 and @xmath182 will be also be finite in the thin - brane limit ( once the loop divergences associated with @xmath174 have been taken care of ) , as shown explicitly for the four - point function . however a non - trivial feature emerges from the computation of the one - loop corrections . the expression involves terms of the form @xmath183 which also diverges logarithmically , but this time this divergence is instead associated with a ir behaviour . in this toy - model , the origin of this ir divergence is related to the fact that the bulk field @xmath26 is massless , but would disappear as soon as a small mass @xmath184 was introduced . however , if the bulk field is to mimic the graviton , this field should remain massless . physically , such ir divergences can be removed in the same way as in quantum electrodynamics , see ref . @xcite . to complete this section , we argue that this scalar field toy - model is renormalizable , provided that only relevant and marginal operators are considered . since no coupling of the form @xmath185 , @xmath186 is introduced , any further @xmath0-point function will necessarily be composed of only reducible diagrams and will thus be expressible in terms of lower - dimensional @xmath187-point functions ( @xmath188 . ) since we have shown that all of these @xmath187-point functions are finite at the classical level , any further @xmath0-point function will thus automatically be finite , without any further counterterms . we make this argument more concrete by exploring the five - point function , and showing explicitly that it remains finite in the thin - brane limit if @xmath93 is renormalized as in . a completely general argument for an arbitrary @xmath0-point function can be found in appendix [ appendix generalcase ] . in particular we show that the same will remains true at the level of any @xmath0-point function . the theory will thus be renormalizable , as one can expect from standard four - dimensional field theory intuition . the contributions to the five - point function are symbolized in fig.[fig5pt ] . ( 150,70)(-20,40)(-20,60)(0,0)[l]@xmath189 ( 130,60)(0,0)[l]@xmath190 ( 270,60)(0,0)[l]@xmath191 ( 360,60)(0,0)[l]@xmath148 $ ] ( 90,64 ) ( -1,1)10 ( 90,56 ) ( -1,-1)10 ( 94,60 ) ( 1,0)30 ( 114,60)(1,1)10 ( 114,60)(1,-1)10 ( 85,69 ) ( 85,51 ) ( 104,60 ) ( 120,60 ) ( 120,66 ) ( 120,54 ) ( 86,56)(0,1)8 ( 94,56)(0,1)8 ( 86,56)(1,0)8 ( 86,64)(1,0)8 ( 210,64 ) ( -1,1)10 ( 210,56 ) ( -1,-1)10 ( 234,64 ) ( 0,1)10 ( 258,64 ) ( 1,1)10 ( 258,56 ) ( 1,-1)10 ( 238,60 ) ( 1,0)16 ( 214,60 ) ( 1,0)16 ( 205,69 ) ( 205,51 ) ( 222,60 ) ( 234,69 ) ( 246,60 ) ( 263,69 ) ( 263,51 ) ( 206,56)(0,1)8 ( 214,56)(0,1)8 ( 206,56)(1,0)8 ( 206,64)(1,0)8 ( 230,56)(0,1)8 ( 238,56)(0,1)8 ( 230,56)(1,0)8 ( 230,64)(1,0)8 ( 254,56)(0,1)8 ( 262,56)(0,1)8 ( 254,56)(1,0)8 ( 254,64)(1,0)8 ( 320,64 ) ( -1,1)10 ( 320,56 ) ( -1,-1)10 ( 324,60 ) ( 1,0)15 ( 349,60)(1,1)10 ( 349,60)(1,-1)10 ( 339,60)(0,1)10 ( 315,69 ) ( 315,51 ) ( 331,60 ) ( 355,66 ) ( 355,54 ) ( 339,65 ) ( 316,56)(0,1)8 ( 324,56)(0,1)8 ( 316,56)(1,0)8 ( 316,64)(1,0)8 ( 339,60)(2,0)5 summing these diagrams , we obtain g^^5_k_1,,k_5&= & + & & + @xmath192g^_(k_4+k_5),k_4,k_5_i=1 ^ 3g^_k_i + & & = 5 ! ig^_(k_4+k_5 ) , k_4 , k_5 , where for simplicity we have used a specific momentum configuration , but the counting takes in account all possible permutations . the first line of the pervious expression is trivially finite , while the second line is finite only if the terms proportional to @xmath93 and @xmath193 contribute with appropriate coefficients . as can be seen in the third line of this expression , the contribution from these terms is also finite as @xmath93 is renormalized precisely so as to have @xmath194 finite . the five - point function is therefore finite at the classical level and no counterterms ought to be added . this result will remain valid for any other five - point functions ( including the ones with bulk external fields @xmath195 , etc . ) , as well as for any higher @xmath0-point function and their loop corrections , ( see appendix [ appendix generalcase ] ) . this represents a highly non - trivial check and leads to the conclusion that the theory is completely renormalizable against divergences associated with the codimension - two source . in this section , we consider the more physical scenario of a massless gauge field @xmath60 confined to the brane and coupled to gravity . this represents a more realistic framework to study the coupling between gravity and electromagnetism . we start with the following six - dimensional action s^(em)=^6 x , with @xmath196 . in what follows we work at linear order in perturbations around a flat conical background : s^2= g_a b x^a x^b = r^2+r^2 ^2+@xmath197x^x^ , where we work in de donder gauge , @xmath198 . since the stress - energy for radiation is transverse , we will have @xmath199 . the einstein s equations impose g_&= & ^2 t_^em + -12 ^(6 ) h_&= & -^2@xmath200 . using results from the previous sections , we know that @xmath201 will diverge logarithmically when evaluated at @xmath8 [ induced h ] h_(0)= @xmath202(f_f_^ -1 4 f^2 _ ) , where @xmath47 is the euler number and @xmath203 represents the thin - brane limit . this will affect the equation of motion for the photon : _ f^&=&@xmath204@xmath205@xmath206 + & = & [ maxwell1 ] _ f^-_@xmath207=0 , where in the second line , all index raising is performed with respect to the background flat metric @xmath208 . we remember that in the previous expression , @xmath209 represents the induced value of the metric perturbation evaluated on the brane and thus diverges logarithmically in the thin - brane limit . if this was the end of the story , then photons would be very sensitive to the brane thickness @xmath2 even at low - energy . however , we have learned from section [ section free theory ] , that as soon as a coupling @xmath62 is introduced between brane and bulk fields , this spontaneously generates a mass term @xmath63 for the brane field at the classical level . the situation is no different here , and the logarithmic divergence of @xmath210 on the brane will spontaneously generate @xmath211 terms on the brane . more precisely , let us consider the euler - heisenberg brane action s_(brane)=-x^4 , so that the maxwell s equations are modified to _ f^-_@xmath207-_1 _ @xmath212-_2 _ @xmath213=0 . substituting the expression for the perturbed metric on the brane , we obtain _ f^ & + & @xmath214-_1 _ @xmath215 + & -&@xmath216+_2 _ @xmath217=0 . the divergence of the graviton can thus be absorbed in the two couplings @xmath218 and @xmath219 : _ 1()=_1()+ and _ 2()=_2()- , leading to the following rg flows _ _ 1 ( ) = -1 4 _ _ 2 ( ) = . the generation of @xmath211 terms on the brane at the classical level ensures that a photon confined to a codimension - two brane and interacting with a gravitational wave will not evolve in a regularization - dependent way . in particular , this mechanisms ensures that at low - energy , there is a well defined thin - brane limit description of the codimension - two brane . of course once , @xmath211 terms are introduced , they will in turn introduce divergences on the brane , which should be absorbed with higher order terms . the proper finite theory will hence include a infinite series . as a last intriguing physical implication , we consider in this section the consequences for localized kinetic terms on the brane . localized kinetic terms are of importance when considering brane - induced einstein - hilbert terms , where the action is typically of the form s= ^d x r_(d ) + ^4 x @xmath220 . the induced einstein - hilbert term @xmath221 term is expected to be spontaneously generated at the quantum level , and represents a natural mechanism to localize gravity on a four - dimensional surface when the extra dimensions are flat and infinite @xcite . such models also represent a physical realization of the degravitation process ( see ref . @xcite ) , since in such scenarios gravity becomes fully higher - dimensional at long wavelengths , hence providing a potential explanation for the observed value of the cosmological constant . such models are also enriched with an additional interesting feature namely the possibility of having self - accelerating branches @xcite ( see refs . @xcite for ghost - free realizations ) . although such models are usually considered in the context of one large extra dimension , a simultaneous resolution of the hierarchy problem usually requires at least two extra dimensions @xcite , and the degravitation observed in the presence of only one extra dimension is only marginal . understanding this scenario in the presence of two extra dimensions is therefore an important next step . in what follows , we examine the consequences of such kinetic terms in a scalar field toy - model . we consider a massless scalar field @xmath26 living in a six - dimensional flat space - time with induced kinetic terms on a codimension - two brane s=-^6x , where @xmath222 represents the four - dimensional dalembertian @xmath223 . in order to recover on the brane the standard klein - gordon equation for the scalar field in the infrared , @xmath224 , we require that only positive powers of @xmath222 be present in @xmath225 . in particular we write f()= _ n0 c_n @xmath226^n , where @xmath227 is an arbitrary length scale ( we recall that in this six - dimensional formalism , @xmath26 has dimension mass squared and thus @xmath225 ought to be dimensionless ) . in particular , @xmath228 represents the dimensionless coupling @xmath55 that was considered in section [ section free theory ] . from that section , we know that the scalar field will be well - defined away from the brane only if the induced couplings on the brane ( in this case the function @xmath225 ) are renormalized and flow as _ f()=12 f^2 ( ) . in terms of the coefficients @xmath229 , this implies [ cn ] _ c_n()=12 _ u=0^n c_n - u()c_u ( ) . this has important consequences for these kind of theories , in particular , we are not free to choose a brane induced function of the form @xmath230 , as higher curvature terms will spontaneously be generated at the tree - level . as soon as a mass term @xmath228 and kinetic term @xmath231 are present , all the other couplings @xmath229 will flow in a non - trivial way . the solution for the two first terms is of the form c_0()&= & + c_1()&=&=c_0()^2 , where @xmath232 is a dimensionless parameter . as an example , one can choose a particular solution of , for which c_n()=^n c_0()^n+1 , and so f()=. notice that in this formalism , the function of the kinetic term is fixed by the renormalization conditions , and very few parameters can actually be tuned arbitrarily . in this sense this represents a much more satisfying candidate for theories of modified gravity than for instance @xmath233 gravities , ( see ref . @xcite for a review on such theories ) . as a natural extension , one should understand the cosmology for such a scenario and possibly the different signatures which could allow for the discrimination of codimension - two models . we have analyzed the coupling between bulk fields living in a six - dimensional flat space - time and brane fields confined onto a four - dimensional surface . due to these couplings , logarithmic divergences that arise when evaluating the bulk field on the brane generically propagate into the brane field . in this paper , we have presented a consistent renormalization mechanism at tree and one - loop level that removes any divergences simultaneously in the brane field and bulk field when the latter is evaluated away from the brane . we have also shown that any five - point function is finite at the classical level without the addition of any further counterterm , and demonstrated that this remains true for any @xmath0-point function at any order in the loop expansion , thus proving the renormalizability of the theory . we also point out the presence of ir divergences in the loop diagrams which can be dealt with the same way as for quantum electrodynamics . the same principle can be applied to more complex theories such as electromagnetism in curved space - time , for which the same prescription remains completely valid . in particular we show that at tree - level , the coupling of electromagnetism to gravity generates a quartic term for the form field in the action , giving rise to a euler - heisenberg lagrangian which is usually only generated via quantum corrections . as another physical application , we have also explored the consequences for localized kinetic terms which are relevant in scenarios such as the dgp model . in particular we show that as soon as a kinetic term are induced on the brane , one can not prevent for the generation of an infinite series of higher order terms . this provides a natural modification of gravity on the brane which might have potential interesting signatures . to our knowledge , this prescription is the only one to date capable of making sense of sources on codimension - two branes in a regularization - independent way and providing a way to derive the low - energy effective theory on such objects in a regime where interactions with the bulk can not be ignored . since the main objective of this paper was the establishment of a consistent framework to study sources on codimension - two branes , implications for braneworld physics have only been superficially addressed . armed with these new tools , extensions to more physical scenarios will however be of great importance . understanding the relevance of our results for electromagnetism and theories with induced gravity terms were beyond the scope of this paper but will present interesting extensions . more realistic interactions from the standard model would also be intriguing , and in particular consequences to the higgs physics in six - dimensional scenarios , such as the sled , @xcite should be understood in more detail . the author wishes to thank rouzbeh allahverdi , cliff burgess , stefan hofmann , justin khoury , federico piazza , andrew tolley , mark wise and mark wyman for their contribution throughout the development of this work . research at mcmaster is supported by the natural sciences and engineering research council of canada . research at perimeter institute for theoretical physics is supported in part by the government of canada through nserc and by the province of ontario through mri . in what follows , we consider a free scalar field living on a flat six - dimensional spacetime with a conical singularity at @xmath8 . as shown in and , the brane - brane propagator of this field diverges logarithmically in the thin - brane limit : @xmath234 as @xmath235 . in order to make this quantity finite , one can try to include counterterms both in the bulk and the brane . unlike in usual eft , these counterterms are not added to make the interacting theory finite , but the classical free theory itself . we consider the following most general set of counterterms to renormalize the free theory ( renormalization of the wave function and mass both in the bulk and brane ) s=-^6x . the propagator will now be instead d^_k(r , r)&=&_n=-^+_0 j_|n|(q r)j_|n|(q r ) + & = & , where @xmath68 and @xmath236 . because of the brane counterterms @xmath237 and @xmath55 , this is however not the complete two - point function . the two - point function is obtained by summing over all the interactions with the bulk coupling . this gives rise to following dressed " propagators g_k(r , r)&=&d^_k(r , r)-d_k^(r,0)d_k^(0,r ) + g_k(r,0)&=&d^_k(r,0 ) + g_k(0,0)&=&d^_k(0,0 ) . notice that in this approach , we no longer require the propagator @xmath238 to be finite in the thin brane limit but require instead that the two - point function @xmath239 between any two points ( taken in the bulk or the conical tip ) is finite @xmath240 and @xmath55 should flow in such a way that @xmath241 and @xmath242 are all finite . this implies that : * @xmath243 should be finite in the limit @xmath244 for any value of @xmath72 , * and similarly the quantity @xmath245 $ ] should be finite for any @xmath72 and @xmath73 . it will therefore only be possible to make sense of the two - point function on the brane , if one can renormalize the wave function and the mass in such a way that the ratio @xmath243 is finite . from , we get d_k^(r,0)= k_0@xmath246 , and d_k^(0,0)= _ d_k^(^-1,0)= . it is therefore clear from these expressions that no matter how the wave function and the mass renormalization flow , the quantity @xmath247 will never be finite in the thin brane limit : = , ( ) . no local counterterm ( quadratic in the field ) will thus ever make the two - point function finite everywhere both in the bulk and the brane . in this appendix , we demonstrate that the rg flows of the brane couplings @xmath51 , @xmath92 , @xmath93 , @xmath62 , @xmath55 , @xmath94 , ( [ renormcouplings ] , [ rgflow3 ] , [ beta3_2 ] and [ beta4 ] ) is sufficient to make all @xmath0-point green s functions finite in the thin - brane limit at any order in the loop expansion , hence justifying the renormalizability of the theory . we focus , in what follows , on divergences associated with the codimension - two nature of the theory and do not discuss loop divergences which can be renormalized the standard way . to simplify , we present the argument for the @xmath0-point green s functions having only brane field external legs @xmath27 . as seen in section [ sectioninteractions ] , the generalization to an arbitrary number of bulk field legs @xmath26 is straight forward . defining the generating functional @xmath248 $ ] g[j]=0|e^-i(h^int _ + j ) |0 , the @xmath0-point green s functions are expressed by g^(n)(x_1,,x_n)=.|_j=0 . technically , we are only interested in the connected green s functions which are generated by @xmath249=-i \log g[j]$ ] . the connected @xmath0-point green s function can thus be expressed in terms of the lower ones as g_c^(n)=-i@xmath250 , where @xmath251 is the binomial coefficient . we have previously demonstrated that all the connected tree - level green s functions @xmath252 were finite for . in what follows , we show that this results remains true for any green s function @xmath253 , @xmath254 at any order in the loop expansion , thus ensuring that all the connected green s is finite for any arbitrary number of external legs . the expression for the @xmath0-point function is g^(n)(x_1,,x_n)&= & _ n0(x_1)(x_n ) + & & _ i=1^ny_i @xmath255 . omitting the evaluation points @xmath256 and @xmath257 and remembering that every field is evaluated on the brane , we have g^(n)&=&_n0 ^n@xmath258^n + & = & _ n0_m=0^n_k=0^mc^m_n c^k_m _ 3^m - k_4^n - m _ ^n - k+4n - m^k among these diagrams , some of them can connect two bulk fields together , generating a singular two - point function @xmath259 . there can be @xmath260 such connections , ( with @xmath261 ) , so that g^(n)=_n0_m=0^n_k=0^m_=0^k/2c^m_n c^k_m c^2_k _ 3^m - k_4^n - m _ 3^k ^n - k+4n - m^k-2 ^2.we now consider the number of ways there is to connect the different fields together ( we recall that for now we are interested in all the possible configurations , and do not restrict ourselves to the connected ones ) . for @xmath12 fields @xmath262 , there is@xmath263 possible configurations if @xmath12 is even , and no possible configurations if @xmath12 is odd ( @xmath264 ) . there is therefore @xmath265 ways to connect the @xmath266 bulk fields in @xmath267 . to count the number of configurations in @xmath268 , we need to pick first of all the fields @xmath27 which will connect with the remaining @xmath269 fields @xmath26 . there is @xmath270 such configurations and then @xmath271 ways to connect the remaining @xmath27 together . putting all this together , we therefore get [ g(n)1 ] g^(n)&=&_n0_m=0^n_k=0^m_=0^k/2 ( c^m_n c^k_m _ 3^m - k_4^n - m_3^k c^2_k + & & p_n-2k+4n - m+2p_2 + & & @xmath272^@xmath273^k-2@xmath259^)with p_x=\ { expressing the bulk - bulk and bulk - brane propagator as g^(0)&=&-id g^ + g^(0,0)&=&d-^2 d ^2 g^ , we get g^(n)&=&_n0_m=0^n_k=0^m_=0^k/2_=0^c^m_n c^k_m _ 4^n - m_3^m - k_3^k^k-2d^k-2 @xmath272^(n+4n - m-2 ) + & & _ = f ( ) . notice that the coefficient @xmath260 does not affect the couplings , and so the summation over @xmath260 can be performed without affecting the order of the diagram _ = 0^k/2_=0^f()=_=0^k/2 _ = ^k/2 f ( ) = _ = 0^k/2(-i)^k-2p_n - m+4n+2 . we now change the summation variables to @xmath275 , with x = m-2,y = n - m+,l = k-2so that the green s function can be expressed as g^(n)&=&_x0_y0 ( p_n+3x+4y @xmath272^(n+3x+4y ) + & & _ l=0^x_=0^y c^l_x c^_y ^l _ 3^l+2_3^x - l_4^y - d ^l+ ) , finally , summing over @xmath276 and @xmath277 , we recover the familiar expressions g^(n)=_x0_y0 p_n+3x+4y @xmath272^(n+3x+4y ) @xmath278^x @xmath279^y . so the coupling constants @xmath62 , @xmath92 , @xmath92 , @xmath94 and the free bulk - bulk propagator @xmath280 come in precisely the right combination to be finite . the rg flows of @xmath94 , @xmath92 and @xmath93 indeed ensures that both expressions @xmath281 and @xmath282 are finite , see eqs . and . since the renormalization of @xmath62 , @xmath63 and @xmath55 is such that the brane propagator @xmath272 is finite , we can conclude that @xmath253 is completely finite in the thin brane limit ( up to loop diagram divergences which can be renormalized in the standard way ) . this argument thus demonstrates that the theory is renormalizable . l. randall and r. sundrum , `` a large mass hierarchy from a small extra dimension , '' phys . lett . * 83 * , 3370 ( 1999 ) [ arxiv : hep - ph/9905221 ] ; l. randall and r. sundrum , `` an alternative to compactification , '' phys . * 83 * , 4690 ( 1999 ) [ arxiv : hep - th/9906064 ] . a. salam and e. sezgin , `` chiral compactification on minkowski x s**2 of n=2 einstein - maxwell supergravity in six - dimensions , '' phys . b * 147 * , 47 ( 1984 ) . p. kanti , r. madden and k. a. olive , `` a 6-d brane world model , '' phys . d * 64 * , 044021 ( 2001 ) [ arxiv : hep - th/0104177 ] ; g. w. gibbons , r. guven and c. n. pope , `` 3-branes and uniqueness of the salam - sezgin vacuum , '' phys . b * 595 * , 498 ( 2004 ) [ arxiv : hep - th/0307238 ] ; y. aghababaie _ et al . _ , `` warped brane worlds in six dimensional supergravity , '' jhep * 0309 * , 037 ( 2003 ) [ arxiv : hep - th/0308064 ] ; h. m. lee and a. papazoglou , `` brane solutions of a spherical sigma model in six dimensions , '' nucl . phys . b * 705 * , 152 ( 2005 ) [ arxiv : hep - th/0407208 ] ; c. p. burgess , f. quevedo , g. tasinato and i. zavala , `` general axisymmetric solutions and self - tuning in 6d chiral gauged supergravity , '' jhep * 0411 * , 069 ( 2004 ) [ arxiv : hep - th/0408109 ] ; s. mukohyama , y. sendouda , h. yoshiguchi and s. kinoshita , `` warped flux compactification and brane gravity , '' jcap * 0507 * , 013 ( 2005 ) [ arxiv : hep - th/0506050 ] ; s. l. parameswaran , g. tasinato and i. zavala , `` the 6d superswirl , '' nucl . b * 737 * , 49 ( 2006 ) [ arxiv : hep - th/0509061 ] ; h. m. lee and c. ludeling , `` the general warped solution with conical branes in six - dimensional supergravity , '' jhep * 0601 * , 062 ( 2006 ) [ arxiv : hep - th/0510026 ] ; a. j. tolley , c. p. burgess , d. hoover and y. aghababaie , bulk singularities and the effective cosmological constant for higher jhep * 0603 * , 091 ( 2006 ) [ arxiv : hep - th/0512218 ] ; a. j. tolley , c. p. burgess , c. de rham and d. hoover , `` scaling solutions to 6d gauged chiral supergravity , '' new j. phys . * 8 * , 324 ( 2006 ) [ arxiv : hep - th/0608083 ] ; t. kobayashi and m. minamitsuji , `` brane cosmological solutions in six - dimensional warped flux compactifications , '' arxiv:0705.3500 [ hep - th ] ; e. j. copeland and o. seto , `` dynamical solutions of warped six dimensional supergravity , '' arxiv:0705.4169 [ hep - th ] . s. hayakawa , d. ida , t. shiromizu and t. tanaka , `` gravitation in the codimension two brane world , '' prog . theor . phys . suppl . * 148 * , 128 ( 2003 ) ; j. m. cline , j. descheneau , m. giovannini and j. vinet , `` cosmology of codimension - two braneworlds , '' jhep * 0306 * , 048 ( 2003 ) [ arxiv : hep - th/0304147 ] ; j. vinet , `` generalised cosmology of codimension - two braneworlds , '' int . j. mod . a * 19 * , 5295 ( 2004 ) [ arxiv : hep - th/0408082 ] ; c. charmousis and r. zegers , `` einstein gravity on an even codimension brane , '' phys . d * 72 * , 064005 ( 2005 ) [ arxiv : hep - th/0502171 ] ; e. papantonopoulos and a. papazoglou , `` cosmological evolution of a purely conical codimension-2 brane world , '' jhep * 0509 * , 012 ( 2005 ) [ arxiv : hep - th/0507278 ] . m. l. graesser , j. e. kile and p. wang , `` gravitational perturbations of a six dimensional self - tuning model , '' phys . d * 70 * , 024008 ( 2004 ) [ arxiv : hep - th/0403074 ] ; h. m. lee and a. papazoglou , `` scalar mode analysis of the warped salam - sezgin model , '' nucl . b * 747 * , 294 ( 2006 ) [ erratum - ibid . b * 765 * , 200 ( 2007 ) ] [ arxiv : hep - th/0602208 ] ; c. p. burgess , c. de rham , d. hoover , d. mason and a. j. tolley , `` kicking the rugby ball : perturbations of 6d gauged chiral supergravity , '' jcap * 0702 * , 009 ( 2007 ) [ arxiv : hep - th/0610078 ] ; s. l. parameswaran , s. randjbar - daemi and a. salvio , `` stability and negative tensions in 6d brane worlds , '' arxiv:0706.1893 [ hep - th ] . m. peloso , l. sorbo and g. tasinato , `` standard 4d gravity on a brane in six dimensional flux compactifications , '' phys . d * 73 * , 104025 ( 2006 ) [ arxiv : hep - th/0603026 ] . n. arkani - hamed , s. dimopoulos and g. r. dvali , `` the hierarchy problem and new dimensions at a millimeter , '' phys . b * 429 * , 263 ( 1998 ) [ arxiv : hep - ph/9803315 ] . a. albrecht , c. p. burgess , f. ravndal and c. skordis , `` natural quintessence and large extra dimensions , '' phys . d * 65 * , 123507 ( 2002 ) [ arxiv : astro - ph/0107573 ] ; d. m. ghilencea , d. hoover , c. p. burgess and f. quevedo , `` casimir energies for 6d supergravities compactified on t(2)/z(n ) with wilson lines , '' jhep * 0509 * , 050 ( 2005 ) [ arxiv : hep - th/0506164 ] ; e. elizalde , m. minamitsuji and w. naylor , `` casimir effect in rugby - ball type flux compactifications , '' phys . rev . d * 75 * , 064032 ( 2007 ) [ arxiv : hep - th/0702098 ] ; m. minamitsuji , `` casimir effect in a 6d warped flux compactification model , '' arxiv:0704.3623 [ gr - qc ] ; k. koyama , `` the cosmological constant and dark energy in braneworlds , '' arxiv:0706.1557 [ astro - ph ] . v. a. rubakov and m. e. shaposhnikov , `` extra space - time dimensions : towards a solution to the cosmological constant problem , '' phys . lett . b * 125 * , 139 ( 1983 ) . y. aghababaie , c. p. burgess , s. l. parameswaran and f. quevedo , `` towards a naturally small cosmological constant from branes in 6d supergravity , '' nucl . b * 680 * , 389 ( 2004 ) [ arxiv : hep - th/0304256 ] ; o. corradini , a. iglesias and z. kakushadze , `` toward solving the cosmological constant problem ? , '' int . j. mod . a * 18 * , 3221 ( 2003 ) [ arxiv : hep - th/0212101 ] ; s. m. carroll and m. m. guica , `` sidestepping the cosmological constant with football - shaped extra dimensions , '' arxiv : hep - th/0302067 ; i. navarro , `` codimension two compactifications and the cosmological constant problem , '' jcap * 0309 * , 004 ( 2003 ) [ arxiv : hep - th/0302129 ] ; i. navarro , `` spheres , deficit angles and the cosmological constant , '' class . quant . grav . * 20 * , 3603 ( 2003 ) [ arxiv : hep - th/0305014 ] ; h. p. nilles , a. papazoglou and g. tasinato , `` selftuning and its footprints , '' nucl . b * 677 * , 405 ( 2004 ) [ arxiv : hep - th/0309042 ] ; c. p. burgess , `` supersymmetric large extra dimensions and the cosmological constant : an update , '' annals phys . * 313 * , 283 ( 2004 ) [ arxiv : hep - th/0402200 ] ; j. m. schwindt and c. wetterich , `` the cosmological constant problem in codimension - two brane models , '' phys . b * 628 * , 189 ( 2005 ) [ arxiv : hep - th/0508065 ] ; c. p. burgess , `` supersymmetric large extra dimensions and the cosmological constant problem , '' arxiv : hep - th/0510123 ; e. k. park and p. s. kwon , `` a self - tuning mechanism in 6d gravity - scalar theory , '' arxiv : hep - th/0702171 . c. p. burgess , j. matias and f. quevedo , `` msled : a minimal supersymmetric large extra dimensions scenario , '' nucl . b * 706 * , 71 ( 2005 ) [ arxiv : hep - ph/0404135 ] ; j. vinet and j. m. cline , `` can codimension - two branes solve the cosmological constant problem ? , '' phys . d * 70 * , 083514 ( 2004 ) [ arxiv : hep - th/0406141 ] ; j. garriga and m. porrati , `` football shaped extra dimensions and the absence of self - tuning , '' jhep * 0408 * , 028 ( 2004 ) [ arxiv : hep - th/0406158 ] ; j. vinet and j. m. cline , `` codimension - two branes in six - dimensional supergravity and the cosmological constant problem , '' phys . d * 71 * , 064011 ( 2005 ) [ arxiv : hep - th/0501098 ] . g. dvali , g. gabadadze and m. shifman , `` diluting cosmological constant in infinite volume extra dimensions , '' phys . d * 67 * , 044020 ( 2003 ) [ arxiv : hep - th/0202174 ] ; o. corradini , a. iglesias and z. kakushadze , `` diluting solutions of the cosmological constant problem , '' mod . a * 18 * , 1343 ( 2003 ) [ arxiv : hep - th/0305164 ] . g. dvali , s. hofmann and j. khoury , arxiv : hep - th/0703027 . g. r. dvali , g. gabadadze and m. porrati , `` metastable gravitons and infinite volume extra dimensions , '' phys . b * 484 * , 112 ( 2000 ) [ arxiv : hep - th/0002190 ] ; g. r. dvali , g. gabadadze and m. porrati , `` 4d gravity on a brane in 5d minkowski space , '' phys . b * 485 * , 208 ( 2000 ) [ arxiv : hep - th/0005016 ] ; g. r. dvali and g. gabadadze , `` gravity on a brane in infinite - volume extra space , '' phys . d * 63 * , 065007 ( 2001 ) [ arxiv : hep - th/0008054 ] . w. d. goldberger and m. b. wise , `` renormalization group flows for brane couplings , '' phys . d * 65 * , 025011 ( 2002 ) [ arxiv : hep - th/0104170 ] . w. d. goldberger and i. z. rothstein , `` an effective field theory of gravity for extended objects , '' phys . d * 73 * , 104029 ( 2006 ) [ arxiv : hep - th/0409156 ] ; w. d. goldberger and i. z. rothstein , `` dissipative effects in the worldline approach to black hole dynamics , '' phys . d * 73 * , 104030 ( 2006 ) [ arxiv : hep - th/0511133 ] ; w. d. goldberger and i. z. rothstein , `` towers of gravitational theories , '' gen . grav . * 38 * , 1537 ( 2006 ) [ int . j. mod . d * 15 * , 2293 ( 2006 ) ] [ arxiv : hep - th/0605238 ] ; w. d. goldberger , `` les houches lectures on effective field theories and gravitational radiation , '' arxiv : hep - ph/0701129 . r. a. porto , `` post - newtonian corrections to the motion of spinning bodies in nrgr , '' phys . d * 73 * , 104031 ( 2006 ) [ arxiv : gr - qc/0511061 ] ; r. a. porto and i. z. rothstein , `` the hyperfine einstein - infeld - hoffmann potential , '' phys . lett . * 97 * , 021101 ( 2006 ) [ arxiv : gr - qc/0604099 ] ; r. a. porto and r. sturani , `` scalar gravity : post - newtonian corrections via an effective field theory approach , '' arxiv : gr - qc/0701105 ; r. a. porto , `` new results at 3pn via an effective field theory of gravity , '' arxiv : gr - qc/0701106 . f. del aguila , m. perez - victoria and j. santiago , `` bulk fields with general brane kinetic terms , '' jhep * 0302 * , 051 ( 2003 ) [ arxiv : hep - th/0302023 ] ; f. del aguila , m. perez - victoria and j. santiago , `` effective description of brane terms in extra dimensions , '' jhep * 0610 * , 056 ( 2006 ) [ arxiv : hep - ph/0601222 ] . t. gherghetta and m. e. shaposhnikov , `` localizing gravity on a string - like defect in six dimensions , '' phys . rev . lett . * 85 * , 240 ( 2000 ) [ arxiv : hep - th/0004014 ] . i. navarro and j. santiago , `` gravity on codimension 2 brane worlds , '' jhep * 0502 * , 007 ( 2005 ) [ arxiv : hep - th/0411250 ] ; c. de rham and a. j. tolley , jcap * 0602 * , 003 ( 2006 ) [ arxiv : hep - th/0511138 ] . t. kobayashi and m. minamitsuji , `` gravity on an extended brane in six - dimensional warped flux compactifications , '' phys . d * 75 * , 104013 ( 2007 ) [ arxiv : hep - th/0703029 ] ; c. p. burgess , d. hoover and g. tasinato , `` uv caps and modulus stabilization for 6d gauged chiral supergravity , '' arxiv:0705.3212 [ hep - th ] . a. e. nelson , `` a new angle on intersecting branes in infinite extra dimensions , '' phys . d * 63 * , 087503 ( 2001 ) [ arxiv : hep - th/9909001 ] ; e. gravanis and s. willison , `` intersecting hyper - surfaces in dimensionally continued topological density gravitation , '' j. math . * 45 * , 4223 ( 2004 ) [ arxiv : hep - th/0306220 ] ; i. navarro and j. santiago , `` higher codimension braneworlds from intersecting branes , '' jhep * 0404 * , 062 ( 2004 ) [ arxiv : hep - th/0402204 ] ; n. kaloper , `` origami world , '' jhep * 0405 * , 061 ( 2004 ) [ arxiv : hep - th/0403208 ] . c. de rham , g. dvali , s. hofmann , j. khoury , o. pujolas , m. redi and a. j. tolley , `` cascading dgp , '' arxiv:0711.2072 [ hep - th ] ; c. de rham , s. hofmann , j. khoury and a. j. tolley , `` cascading gravity and degravitation , '' arxiv:0712.2821 [ hep - th ] . c. de rham , t. shiromizu and a. j. tolley , `` weaker gravity at submillimetre scales in braneworlds models , '' arxiv : gr - qc/0604071 . c. deffayet , g. r. dvali and g. gabadadze , `` accelerated universe from gravity leaking to extra dimensions , '' phys . d * 65 * , 044023 ( 2002 ) [ arxiv : astro - ph/0105068 ] . c. de rham and a. j. tolley , `` mimicking lambda with a spin - two ghost condensate , '' jcap * 0607 * , 004 ( 2006 ) [ arxiv : hep - th/0605122 ] ; g. gabadadze , `` a model for cosmic self - acceleration , '' arxiv : hep - th/0612213 . h. beauchemin , g. azuelos and c. p. burgess , `` dimensionless coupling of bulk scalars at the lhc , '' j. phys . g * 30 * , n17 ( 2004 ) [ arxiv : hep - ph/0407196 ] ; j. j. van der bij and s. dilcher , `` a higher dimensional explanation of the excess of higgs - like events at cern lep , '' phys . b * 638 * , 234 ( 2006 ) [ arxiv : hep - ph/0605008 ] .

self - contentedness is a convenience for the reader : he / she is exempted from searching definitions , statements and sometimes proofs from previous papers , many times from the same author(s ) . now , if i comply to this constraint , anti - plagiarism software will warn the reader that there is a definite percent of overlap of this paper with other papers of the author . the struggle against plagiarism is definitely necessary . i just mention that it may have some inconvenience . now , i am convinced that the reader will understand these inconveniences , so that he / she will play the game of looking into the papers indicated in the references instead of throwing the paper because some context is not at hand . accordingly , i do not repeat most of the introduction of @xcite . as mentioned in the abstract , the improvement with respect to @xcite is obtained by relaxing the constraint of rotation invariance . again , we repeat that the new result is a true planar cellular automaton , which means that at each step of the computation , the set of non quiescent states has always infinitely many cycles , as already mentioned in the abstract . also , the reason for not repeating the introduction is that we make use of the same model of railway computation . also for this reason , we do not mention the reminder sections of @xcite about the railway model and its implementation in the hyperbolic tiling we use in this paper . we also not repeat the introduction to the approach to hyperbolic tilings used in this paper , again , the reader is referred to @xcite and to several papers as @xcite where other references are available . in section [ scenar ] , i indicate the new features for implementing the model with the constraint of using two states only . in section [ rules ] , we give the rules of the automaton , insisting in the way we defined these rules in a context where rotation invariance is no more required , which gives a definite advantage as indicated by the result itself : [ letheo ] there is a weakly universal cellular automaton on the pentagrid with moore neighbourhood which is truly planar and which has two states . presently , we turn to the proof of this result . as in previous papers , sensors and control devices are no more immediate neighbours of the cells where tracks arrive at the point which define a switch . this reinforce the careful study of the tracks themselves as their role for conveying key information is more and more important . here too , tracks are blank cells marked by appropriate milestones . we carefully study this point in sub - section [ tracks ] . later , in sub - sections [ roundabout ] and [ forcontrol ] , we look at the changes introduced with respect to @xcite . in this implementation , the tracks are represented in the same way as in @xcite . for the convenience of the reader , we repeat the illustration of this implementation given by figure [ elemtrack ] . however , some care has to be given to this structure . with two states , we have to thoroughly check that all constraints we have to define horizontal and vertical tracks can be satisfied . we refer the reader to sub - section [ trackrules ] were we give the rules and in which appropriate figures allow us to check the correctness of the implementation given in this paper . we just remember that for this implementation , the locomotive is implemented as a single black cell : it has the same colour as the milestones of the tracks . only the position of the locomotive with respect to the milestones allow us to distinguish it from the milestones . as clear from the next sub - section , we know that besides this * simple locomotive * , the locomotive also occurs as a * double one * in some portions of the circuit . moreover , the circuit also makes use of signals which are implemented in the form of a simple locomotive . so that at some point , it may happen that we have three simple locomotives travelling on the circuit : the locomotive and two auxiliary signals involved in the working of some switch . for aesthetic reasons , the black colour which is opposed to the blank is dark blue in the figures . the round - about replaces the crossing , a railway structure , by a structure inspired by road traffic . at a round - bout where two roads are crossing , if you want to keep the direction arriving at the round - about , you need to leave the round - about at the second road . figure [ roundaboutfig ] illustrates this features , a figure from @xcite which is repeated here for convenience . now , this strategy requires the cellular automaton knows how to count up to two . as in @xcite , we use the three auxiliary structures which are represented in figure [ roundaboutfig ] by a rhombus , a small circle and a bit larger ones . these structures are : the fixed switch , see sub - subsection [ fix ] , the doubler and the selector , see sub - subsection [ doublesel ] . when the locomotive arrives close to the round - about , it first meets the doubler : it transforms the simple locomotive into a double one which consists of two contiguous black cells occupying blank cells of the tracks . then , the locomotive may arrive at the fixed switch , depending on whether it arrived from * a * or from**b**. then , the double locomotive arrives at the first selector : the structure recognizes a double locomotive . it kills one of the cells and the surviving simple locomotive is sent further on the round - about . when it meets the second selector , the structure recognizes a simple locomotive . accordingly , it sends it on the track which leaves the round - about at that point . in order to use the illustrations of this section and of the next one , me make use of a numbering of the tiles based on what is indicated by figure [ pentanum ] . the central tile is numbered 0 . its side are increasingly and counter - clockwise numbered from 1 to 5 . side 1 is defined as follows . let @xmath0 be the upper vertex of tile 0 which is on the vertical diameter @xmath1 of the right - hand side figure . the side 1 of tile 0 is the tile which shares @xmath0 with tile 0 and which is on the left - hand side of @xmath1 . each side of tile 0 delimits a sector : sector 1 is headed by the tile 1(1 ) obtained from tile 0 by a reflection on its side 1 . the sector is delimited by the rays starting from the other end of the side 1 of tile 0 which belong to edges of tile 1(1 ) . the other sectors are obtained by appropriate rotations around the centre of tile 0 in such a way that the union of the sectors is the complement of tile 0 in the tiling . inside each sector , the tiles are numbered as indicated in the right - hand side part of figure [ pentanum ] . the tile @xmath2 of sector @xmath3 is numbered @xmath4 . each sector is panned by the same tree . in each tile , the sides are increasingly and counter - clockwise numbered from 1 to 5 . the side 1 of the tile @xmath4 is the side it shares with its father in the tree rooted in the tile @xmath5 . in section [ rules ] , we shall indicate another way to number the sides of the tiles based on this numbering too . when we shall consider the cellular automaton constructed for the proof of theorem [ letheo ] , we shall speak of cell @xmath4 instead of tile @xmath4 . in this notation , tile 0 will be denoted by 0(0 ) . once the side 1 is fixed in a cell @xmath6 , its edges are numbered according to what we mentioned . the neighbour of @xmath6 which shares the side @xmath7 with it is called neighbour @xmath7 . as we also consider as neighbours of @xmath6 the cells which share a vertex only with @xmath6 , we call these neighbours * vertex - neighbours * and we number them as follows : the vertices are increasingly and counter - clockwise numbered from 6 to 10 , vertex 6 being shared by sides 1 and 5 . accordingly , the vertex - neighbour of @xmath6 sharing its vertex @xmath8 will be called neighbour @xmath8 or vertex - neighbour @xmath8 . as the tracks are one - way and as an active fixed switch always sends the locomotive in the same direction , there is no need of the other direction : there is no active fixed switch . now , passive fixed switches are still needed as just seen in the previous paragraph . figure [ stabfix ] illustrates the passive fixed switch when there is no locomotive around . we can see that it consists of elements of the tracks which are simply assembled in the appropriate way in order to drive the locomotive to the bottom direction in the picture , whatever upper side the locomotive arrived at the switch . from our description of the working of the round - about , a passive fixed switch must be crossed by a double locomotive as well as a simple one . later , in section [ fixedswitch ] , we shall check that the structure illustrated by figure [ stabfix ] allows these crossings . note that the cell 0(0 ) in figure [ stabfix ] has six black neighbours instead of five of them in an ordinary cell of the tracks in a vertical track , see figures [ voievertd ] and [ voievertm ] , further . its black neighbours are neighbours 2 , 4 , 6 , 7 , 8 and 10 . note that neighbour 10 is a milestone belonging to the cell 1(5 ) which is an element of the second track which arrives to the centre of the switch . the doubler is illustrated by the left - hand side picture of figure [ stab_doublsel ] . note that the neighbourhood of the cell @xmath9 is different from an ordinary element of the tracks and it is also different from the cell 0(0 ) of a fixed switch . here , the black neighbours are neighbours 2 , 4 , 5 , 7 and 8 . it is slightly simpler than the doubler in @xcite although we have two states only . as will be seen in section [ rules ] , this is obtained by a particular numbering of the sides of a cell . the working of the configuration is the same as in @xcite , with this difference that the locomotive is black . the selector is illustrated by the right - hand side of figure [ stab_doublsel ] . it is a very symmetric picture , rather different from that for the selector in @xcite . in fact , the symmetry of the picture and an appropriate choice of the sides 1 allow us to make the cells 1(2 ) and 1(4 ) to react in a different way according to the number of black cells in the locomotive . whatever this number , when the first cell of the locomotive is in cell 0(0 ) , both cells 1(2 ) and 1(4 ) can see whether or not a second black cell follows the first one in the cell 1(3 ) . if a single black cell occurs , the cell 1(4 ) becomes black and the cell 1(2 ) remains white . if a second black cell occurs , the cell 1(4 ) remains white and the cell 1(2 ) turns to black . after this point , in both cases , the configuration is that of tracks so the locomotive is sent on the appropriate path . in this sub - section , we look at the decomposition of two active switches : the flip - flop and the active part of the memory switch . we follow the same implementation as in @xcite , separating the working of the switch into two separate stages . the fork is implemented as in @xcite , but the controller is a bit different . for the convenience of the reader , we reproduce the figures of @xcite which are again used by our automaton . in particular , we reproduce figure [ activswitches ] . figure [ stab_active ] is different : the fork has the same configuration , but the two possible stable configurations of the controller are different , distinguished by the colour of the cell 1(1 ) , especially when the controller has to stop the locomotive . when the simple locomotive arrives at the fork @xmath10 , it is duplicated into two simple locomotives , each one following its own path . in the active memory switch , each locomotive goes to a controller . in the flip - flop , one of these locomotives goes on its way to the controller , and the other goes to another fork @xmath11 which sends one locomotive to the other controller and the other , which is now a third locomotive , is sent to a third fork @xmath12 . the locomotives sent by @xmath12 go both to a controller , one to the black controller , the other to the white one . now , when a locomotive sent by @xmath12 arrives at a controller , it changes the black one into a white one and the white one into a black one . accordingly , what should be performed by a flip - flop is indeed performed . it is enough to manage things in such a way that the locomotives arriving to the controllers from @xmath12 arrive later than those sent by @xmath10 and by @xmath11 . in the active memory switch , the locomotive which arrives to @xmath12 is sent by the passive memory switch . in this sense , the passive memory switch is active while the active one is passive . in the fork , the locomotive , in its simple form , arrives through the cell 1(3 ) . from the cell 0(0 ) , two simple locomotives arise at the same time at the cells 1(1 ) and 1(4 ) . in the controller , the locomotive arrives through the path defined by the cells 21(4 ) , 8(4 ) , 3(4 ) and 1(4 ) , in this order . these cells constitute a branch of the tree spanning the sector . along this path , the simple locomotive arrives at cell 0(0 ) through the cell 1(4 ) . when the cell 1(1 ) is black , it goes on its way along the track through the cells 1(2 ) , 3(2 ) , 8(2 ) and 21(2 ) . when the cell 1(1 ) is white , the locomotive is stopped at the cell 1(4 ) and does not reach cell 0(0 ) . figure [ memopass ] illustrates the construction of the passive memory switch with the help of forks and controllers . that implementation is somehow different from the one indicated in @xcite . however , as in @xcite , the controllers of figure [ memopass ] are not the same as those of figure [ activswitches ] and they are different from those of @xcite . the main reason is that in the case of a passage of the locomotive through the non - selected track , in @xcite , the controller let the locomotive go , it changes the selection and it sends a signal to the other controller in order to order it to change its selection too . so that the controller has to perform three tasks at once . the present idea is to slightly simplify the task of the controller . as the locomotive should anyway not be stopped , we can place a fork on the tracks passively arriving to the switch : this is the reason for placing the forks @xmath13 and @xmath14 in figure [ memopass ] where they are on the picture . let us look at the locomotive which arrives at the fork @xmath14 on the figure . it corresponds to the non - selected track . the simple locomotive is duplicated : one goes to the fixed switch @xmath15 and goes further . the second simple locomotive goes to the controller @xmath16 . as this controller signalizes a non - selected track , the selection is changed and the controller let this locomotive go on along the track which leads it to the other controller , @xmath17 . there , at @xmath17 , the second locomotive changes the selection from selected to non - selected . note that the locomotive which goes on its way to @xmath17 passes through another fork , namely @xmath18 which sends a third locomotive to the active switch in order to change its selection too . consider the case when the simple locomotive arrives to @xmath13 , the fork which corresponds on the figure to the selected track . then , the locomotive is sent to @xmath17 where it is stopped as no change should be performed . figure [ stab_ctrlsgn ] illustrates the controller of the passive memory switch . it is presented when no locomotive is around and it presents the structure in its two basic configurations : when it is black and when it is blank . the difference is shown by the colour of the cell 1(1 ) . the locomotive arrives at the cell 0(0 ) along the path 33(2 ) , 12(2 ) , 4(2 ) and 1(2 ) in this order . if the cell 1(1 ) is blank , which signalizes that the track is not selected , then the cell 1(1 ) turns to black and the locomotive goes on its way through the cells 1(4 ) , 4(4 ) , 12(4 ) and 33(4 ) . accordingly , the constraints above described are satisfied . if the cell 1(1 ) is black when the locomotive arrives from 33(4 ) , the locomotive is stopped at 1(2 ) : it does not reach even the cell 0(0 ) and the cell 1(1 ) remains white . in section [ rules ] , we shall see that the working is slightly more complex : this is induced by technical details raised by the rules themselves . note that when the locomotive arrived through the non - selected track , the second locomotive which arrives to @xmath18 or @xmath19 in figure [ stab_ctrlsgn ] arrives to the controller through the path defined by 2(1 ) , 2(6 ) , 16(1 ) and 42(1 ) . the signal stops at the cell 2(1 ) where it is changed to blank . at the same time , the cell 1(1 ) is also changed to blank . it should be stressed that when the cell 1(1 ) is black , and when the locomotive arrives through the selected track , no locomotive is sent to the other controller as there is nothing to change . the figures of section [ scenar ] help us to establish the rules . the rules and the figures were established with the help of a computer program which checked the coherence of the rules . the program also wrote the postscript files of the pictures from the computation of the application of the rules to the configurations of the various type of parts of the circuit . the computer program also established the traces of execution which allow the reader to check the application of the rules . we have to revisit the format of the rules and also to explain what is allowed from the relaxation from rotation invariance . we remind the reader that a rule has the form , where is the state of the cell @xmath6 , is the state of the neighbour @xmath7 of @xmath6 and is the new state of @xmath6 once the rule has been applied . as the rules no more observe the rotation invariance , we may freely choose which is side 1 for each cell . we take this freedom from the format of the rule which only requires to know which is neighbour 1 . in order to restrict the number of rules , it is decided that as a general rule , for a cell which is an element of the track , side 1 is the side shared by the cell and its next neighbour on the track . there can be exceptions when the cell is in a switch or the neighbour of the central cell in a switch . in particular , when a cell belongs to two tracks , side 1 is arbitrarily chosen among the two possible cases . the milestones have their side 1 shared by an element of the track . this means that when a milestone @xmath20 is a vertex - neighbour of an element of the track @xmath21 , the side 1 of @xmath20 is not a side of @xmath21 . we have to keep in mind that there are two types of rules . those which keep the structure invariant when the locomotive is far from them , we call this type of rules * conservative * , and those which control the motion of the locomotive . these latter rules , which we call * motion rules * , are the rules applied to the cells of the tracks as well as their milestones and , sometimes to the cells of the structures which may be affected by the passage of the locomotive . table [ consrules ] provides us with rules which concern the white cells which are not on the track and who have a single or two black neighbours . it also concerns black cells which have a single or two non white neighbours . a few other conservative rules of the same kind also appear in further tables . # 1 # 2 # 3 # 4 -4pt # 1 # 2 # 3 # 4 # 4 as explained in section [ scenar ] , the construction of the tracks is very important . the figures we give in this sub - section and in the following ones are _ local views_. by that phrase , we mean that we can only see a very tiny part of the hyperbolic plane , that the central cell is not the centre of the hyperbolic plane , such a centre does not exist , but the cell on which we focus our attention . in these figures , we can see the central cell , an element of the tracks in figures [ voievertd ] and [ voievertm ] , as well as two other cells of the track , on both sides of the central cell . we need to define verticals and horizontals in the pentagrid in order to simulate the corresponding euclidean notions , which will allow us to directly implement the basic elements of the railway model , see @xcite for references . we look at verticals in sub - subsection [ les_verticales ] and we look at horizontals in sub - subsection [ les_horizontales ] , defining in these sub - subsections what we call verticals and horizontals in the pentagrid . we call * line of the pentagrid * a line of the hyperbolic plane which supports a side of a tile of the pentagrid . note that if a line @xmath1 of the hyperbolic plane supports a side @xmath22 of some tile @xmath23 , it also supports the sides of a sequence where @xmath24 for some and such that each @xmath25 has a side on @xmath1 , that @xmath25 and @xmath26 share an edge for all and that all @xmath25 are on the same side of @xmath1 . such a set of tiles is called a * tape*. it is easy to see that a line of the pentagrid @xmath1 defines two tapes : each one is in its own side of @xmath1 . say that a ray is a * ray of the pentagrid * if it is contained in a line of the pentagrid and if it is issued from a vertex of a tile of the pentagrid . we obtain a * half - tape * by replacing the line of the pentagrid @xmath1 by a ray of the pentagrid in the above definition of a tape . the tile of a half - tape defined by the ray of the pentagrid @xmath27 which contains the vertex from which @xmath27 is issued is called the * head * of the half - tape . we say that a * vertical * is a finite sequence of consecutive tiles belonging to a half - tape @xmath21 and which also belong to a branch of the fibonacci tree rooted at the head of @xmath21 . we remind the reader that the fibonacci tree is the tree which generates the restriction of the pentagrid to a sector as illustrated in figure [ pentanum ] . = 21.5pt # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 # 9 # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 8 # 1 # 2 # 3 # 4 # 5 # 6 # 7 # 1 # 2 # 3 # 4 # 5 # 6 # 1 # 2 # 3 # 4 # 5 # 1 # 2 # 3 # 4 # 1 # 2 # 3 a * vertical track * as illustrated by figures [ voievertd ] and [ voievertm ] consists in two verticals which belong to half - tapes defined on both sides of the same ray of the pentagrid @xmath27 . we say that @xmath27 is the * backbone * of the vertical track . one vertical along @xmath27 consists of blank cells while the other consists of black cells : each black cell @xmath28 is a milestone of the white cell @xmath29 which shares a side on @xmath27 with @xmath28 . if the side 1 of @xmath29 is shared with the centre of the next element of the track , @xmath28 is the neighbour 5 of @xmath29 while the second milestone of @xmath29 is its neighbour 2 . note that @xmath29 can see other milestones of its neighbouring elements of the track . in a going down track , neighbour 6 is a milestone in the next element of the tracks , it is also a neighbour of the neighbour 5 of @xmath29 . neighbours 9 and 10 of @xmath29 are milestones of the previous element of the track , neighbour 10 being also a neighbour of the neighbour 5 of @xmath29 . when the track is going up , the vertex - neighbours are 7 , a milestone of the next element of the track , and 8 with 9 which are milestones of the previous element . note that if we take the reflection in its backbone of the vertical track illustrated by figure [ voievertd ] , we obtain the image of the vertical track illustrated by figure [ voievertm ] by a reflection in a vertex of a pentagon lying on the backbone of that vertical track . this shows us that figures [ voievertd ] and [ voievertm ] illustrate all possible vertical track : going down as in figure [ voievertd ] or going up as in figure [ voievertm ] and again whichever the side where the track lies with respect to its backbone . this is why the rules of tables [ verttracktab ] and [ mounttracktab ] also apply to tracks which are displayed on another side compared to that of figures [ voievertd ] and [ voievertm ] . let us now turn to the rules displayed by table [ verttracktab ] for a locomotive which goes down and by table [ mounttracktab ] for a locomotive which goes up . when the locomotive is not there , the rule which is applied to the centre of an element of the track is rule 25 in table [ verttracktab ] : note that , as already mentioned , the milestones are neighbours 2 and 5 and that the milestones of neighbouring elements are seen by the centre as neighbours 6 , 9 and 10 when the track goes down . when a simple locomotive crosses an element @xmath21 in a vertical track , rule 26 makes the locomotive enter @xmath21 . indeed , the locomotive is seen by @xmath21 as its neighbour 4 . rule 27 makes the locomotive go out : the cell turns from white to black . rule 28 witnesses that the locomotive went out : it is seen as the neighbour 1 of @xmath21 . table [ exvertd ] shows which rule are applied on the cells of the track around the central cell of figure [ voievertd ] . we can see than the same sequence of rules : 25 , 27 , 28 and again 25 is applied along the cells of the tracks : 4(1 ) , 1(1 ) , 0(0 ) , 1(3 ) and 3(3 ) . these rules are motion rules as they apply to the centres of the elements of the tracks . they allow the locomotive to move on its way . but other rules are also needed . we can see there necessity in table [ exvertd ] which displays the rules applied to the cells 1(2 ) and 1(4 ) which are the milestones of the cell @xmath9 . when the locomotive is far from the element of the track , rule 12 applies to 1(2 ) and rule 2 applies to 1(4 ) . for that latter cell it is clear : all neighbours of 1(4 ) are white . for the cell 1(2 ) , it can see two other milestones on the half - line of black cells to which it belongs : its neighbour 2 , a milestone of the cell of the track before 0(0 ) and its neighbour 5 , a milestone of the cell of the track after 0(0 ) . this is exactly what rule 12 says : with such a neighbouring , the cell remains blank . when the locomotive approaches , the milestone can first see it as its neighbour 7 : rule 15 applies . then , the milestone can see the locomotive as its neighbour 1 : rule 16 applies . then , the locomotive is going further , it is seen as neighbour 6 : rule 17 applies . after that time , the locomotive is no more seen before a possible new passage on this very track , rule 12 applies again . all these rules : 12 , 15 , 16 and 17 leave the milestone unchanged . for the cell 1(4 ) , other rules are applied with the same effect : the state is not changed . rule 8 is applied when the locomotive is in the cell 0(0 ) and not before : it is seen as neighbour 1 of the milestone . then rule 11 is applied , indicating that the locomotive , seen as neighbour 7 , has left the cell 0(0 ) . it is no more seen so that rule 2 again is applied . table [ verttracktab ] also displays the rules when the locomotive is double . table [ exvertd ] also indicates which rules are applied to the same cells as those seen previously . to some of the rules already examined , new rules are applied due to the fact that now two contiguous black cells travel together . when the locomotive approaches the cell 0(0 ) , the same rule 26 is applied . but at the next time , instead of rule 27 , rule 29 is applied : this rule makes the second black cell enter the cell 0(0 ) as this second cell is seen as the neighbour 4 of 0(0 ) . then , rule 31 is applied which makes the second black cell leave 0(0 ) because the first cell of the locomotive is seen as the neighbour 1 of 0(0 ) . when the cell 0(0 ) is again white , rule 28 is again applied : it can see the second cell of the locomotive in neighbour 1 . after that time , rule 25 is again applied . for the milestones , we have a similar situation . for 1(2 ) , after rule 15 , rule 33 is applied : it can see the first cell of the locomotive in its neighbour 1 and the second one in its neighbour 7 . then rule 34 is applied which can see the first cell of the locomotive in the neighbour 6 of the milestone and the second cell of the locomotive in its neighbour 1 . after that , rule 17 is applied as the first cell of the locomotive is no more seen by the milestone . for the cell 1(4 ) , similarly , rule 32 is applied after rule 8 as two cells of the locomotive are seen : in the neighbour 7 of 1(4 ) and in its neighbour 1 . then rule 11 applies as a single cell of the locomotive can be seen . after that , rule 2 is again applied . in table [ mounttracktab ] , we have the rules for going up . table [ exvertm ] shows the rules which where applied in a motion illustrated by figure [ voievertm ] , where the locomotives are going up . this time , rules 35 , 41 , 42 and 48 are used . indeed , although side 1 is defined in the same way as in the case of figure [ voievertd ] , the neighbourhood of the cell 0(0 ) is not exactly the same . this , time , as already mentioned , the milestones of the cell are still its neighbours 2 and 5 , but the milestones of neighbouring elements are now seen in the neighbours 7 , 8 and 9 instead of the neighbours 6 , 9 and 10 in the previous case . that explains the differences we already noticed for a simple locomotive . note that for the milestones , the rules are the same but they are applied in the reverse order : this is explained by the symmetry we already noticed . we now investigate the construction of the horizontal tracks . for this , we have to define what we call * horizontals of the pentagrid*. these objects are not lines nor rays , they are sets of tiles defined with the help of the fibonacci tree . a * level * of the fibonacci tree is the set of nodes which are at the same distance from its root : this distance is the number of nodes on the branch going from the root to the node , the root being not taken into account . a * horizontal of the pentagrid * is a finite sequence of tiles whose image on in the fibonacci tree belong to the same level of the tree , with this constraint , that two consecutive tiles in the sequence share a vertex . a * horizontal track * @xmath21 is defined from a horizontal of the pentagrid @xmath30 which is called the * backbone * of @xmath21 . figure [ horizstab ] shows us two possible configurations of a horizontal track depending on whether the tile around which they turn corresponds to a black tile or to a white one . moreover , the figure is split into two parts : to left , the track goes in a clockwise way , to right , it goes on a counter - clockwise one . consider the case of a clockwise way around a black tile . the track consists of the following cells : 5(4 ) , 4(3 ) , 10(3 ) , 3(3 ) , 7(3 ) , 2(3 ) , 1(2 ) , 2(2 ) , 5(2 ) , 4(1 ) , 10(1 ) , 3(1 ) , 7(1 ) , 18(1 ) , 6(1 ) , 15(1 ) and 5(1 ) . their isolated milestones are , in the same order : 13(4 ) , 11(3 ) , 28(3 ) , 26(3 ) , 8(3 ) , 20(3 ) , 18(3 ) , 5(3 ) , 3(2 ) , 6(2 ) , 15(2 ) , 13(2 ) , 11(1 ) , 28(1 ) , 26(1 ) , 8(1 ) , 19(1 ) , 49(1 ) , 47(1 ) , 16(1 ) , 41(1 ) , 39(1 ) and 13(1 ) . the other milestones , those which share a side with another one , are : 2(1 ) , 1(1 ) , 0(0 ) , 1(3 ) and 2(4 ) . in our setting , we consider that 2(1 ) and 2(4 ) are white nodes and that 0(0 ) is a black one . as a consequence , 1(3 ) is a black node . if we assume that these three nodes belong to the same generation , we need to append the tiles 1(1 ) and 1(3 ) in order to avoid a cell of the track with two consecutive neighbours numbered in 1 .. 5 . indeed , in the same level of the tree , a black node is in between two white ones . the isolated milestones are not always completely surrounded by blank cells . for instance , 19(1 ) shares a side with 41(1 ) . similarly , the cells which constitute the clockwise horizontal track around a white cell are : 5(3 ) , 4(2 ) , 10(2 ) , 3(2 ) , 7(2 ) , 2(2 ) , 1(1 ) , 2(1 ) , 1(5 ) , 2(5 ) , 5(5 ) , 4(4 ) , 10(4 ) , 3(4 ) and 7(4 ) . the isolated milestones are , in the same order : 13(3 ) , 11(2 ) , 28(2 ) , 26(2 ) , 8(2 ) , 20(2 ) , 18(2 ) , 5(2 ) , 3(1 ) , 7(1 ) , 5(1 ) , 3(5 ) , 6(5 ) , 15(5 ) , 13(5 ) , 11(4 ) , 28(4 ) , 26(4 ) , 8(4 ) and 19(4 ) . the other milestones are , sharing at least a side with another one : 18(3 ) , 6(3 ) , 2(3 ) , 1(2 ) , 0(0 ) , 1(3 ) , 1(4 ) , 2(4 ) , 5(4 ) and 4(3 ) . note that 1(3 ) is almost surrounded by black cells . we assume that 2(4 ) is a black node , 0(0 ) and 2(3 ) are both white nodes , all of them the sons of a white node , here 1(3 ) . in our case , 1(2 ) and 1(4 ) can not be the centres of elements of the tracks so that we define these cells as milestones . this is a general feature : in between two consecutive nodes @xmath31 and @xmath32 on the same level which are milestones , the black node which shares a side with both @xmath31 and @xmath32 must also be a milestone . for the counter - clockwise tracks , the centres of the elements of the track are the same as in the clockwise case . the changes occur in the isolated milestones . for the black node track , 5(4 ) , 11(3 ) , 8(3 ) , 5(3 ) , 3(2 ) , 6(2 ) , 11(1 ) , 8(1 ) , 19(1 ) , 16(1 ) and 13(1 ) are eplaced by 6(4 ) , 12(3 ) , 9(3 ) , 6(3 ) , 4(2 ) , 7(2 ) , 12(1 ) , 9(1 ) , 20(1 ) , 17(1 ) and 14(1 ) respectively . similarly , for the the white node track , the milestones : 13(3 ) , 11(2 ) , 8(2 ) , 5(2 ) , 3(1 ) , 3(5 ) , 6(5 ) , 11(4 ) , 8(4 ) and 19(4 ) are replaced by 14(3 ) , 12(2 ) , 9(2 ) , 6(2 ) , 4(1 ) , 4(5 ) , 7(5 ) , 12(4 ) , 9(4 ) and 20(4 ) respectively . tables [ horizahstab ] and [ horizahdtab ] display the rules for the motion on a horizontal in a clockwise way around a black node . the corresponding motion is illustrated by figure [ horizah ] . also , tables [ exhorizahs ] and [ exhorizahd ] show the application of the rules of tables [ horizahstab ] and [ horizahdtab ] respectively , on cells of horizontal tracks and on some of their milestones . the motion is illustrated by figure [ horizh ] . we can see that the centres of the elements of the tracks are applied different rules from those we have seen previously , although horizontal tracks consists of the same elements of the tracks . let us look at this more carefully . for instance , in figure [ horizah ] , when it is visited by a simple locomotive , the cell 3(1 ) is applied rules 58 , 78 , 83 , 87 and 90 respectively . rule 58 , is a conservative rule : it leaves the cell blank and it is applied when the locomotive is not by the cell . note that the milestones are in neighbours 2 and 5 , but the cell can see the milestones of neighbouring cells in its vertex - neighbours 6 , 8 and 9 , a very slight difference with rule 35 . when the locomotive approaches 3(1 ) , it is seen in its neighbour 3 and rule 78 , is applied . then , rule 83 , is applied which makes the locomotive leave the cell . next , rule 87 , is applied : it witnesses that the locomotive is now in neighbour 1 . then , rule 90 , witnesses that the locomotive is in the vertex - number 6 of 3(1 ) . note that for the cell 4(1 ) , the neighbourhood is a bit different . besides the milestones of the element in neighbours 2 and 5 , other milestones , from neighbouring elements , can be seen in the vertex - neighbours 6 and 9 only . now , rule 58 again apply as the locomotive is seen in the vertex - neighbour of the cell . later , rule 88 , makes the locomotive enter the cell as it is seen in its neighbour 3 . rule 61 , makes it leave the cell , rule 71 , can see it in its neighbour 1 and rule 79 , can see it in its vertex - neighbour 7 . now , the cells 10(1 ) and 5(2 ) are applied the same rules as they are in the same situation , playing the same role . the locomotive enters such a cell in a standard way but it leaves the cell from an entry and not from the side which is in between two ones which are shared by a black neighbour . this can be performed in one direction and also in the opposite one as seen later , with table [ exhorizhs ] . rule 62 , is the conservative rule : it can see the milestones in the neighbours 3 , 5 and vertex - neighbours 7 and 8 . the locomotive enters through side 2 , rule 65 , and it leaves the cell by rule 73 , rule 63 , can see the leaving locomotive in the neighbour 1 of the cell . the milestones 12(1 ) and 13(1 ) are applied already used rules in other contexts in the vertical motions . we leave this study to the reader . we also do the same for the motions when the horizontal track goes through the sons of a white node : see tables [ horizwahtab ] and [ horizwhtab ] for the rules , tables [ exhorizwahs ] , [ exhorizwahd ] , [ exhorizwhs ] and [ exhorizwhd ] for the application of the rules to cells of the tracks and a few milestones which are taken from figures [ horizwah ] , and [ horizwh ] . in the abstract and in the introduction , we mentioned that we relaxed the condition on the rotation invariance of the cellular automaton constructed for the proof of theorem [ letheo ] . in this part of our study , we can see six rules which are concerned with this point . as an example : rule 21 , has a rotated neighbourhood which is also that of rule 65 : the state of the cell is the same , but the new state is different . rule 21 appeared in the motion of a simple locomotive going down a vertical track , while rule 65 appears in the motion of a simple locomotive counter - clockwise moving along a horizontal track . a higher number of rules are the rotated form of another one . for instance , rules 81 , 93 and 136 are in this case : rule 93 , and rule 136 , are both rotated forms of rule 81 , as can easily be seen . at the end of the paper we shall look at all rules which have the same neighbourhood , up to rotation , the same state but a different new state . figures [ fixeg ] and [ fixed ] illustrate how the locomotive passively crosses a fixed switch . both figures illustrate the case when the locomotive comes from the left - hand side and when it comes from the right - hand side . note that the locomotive may be simple as well as double : indeed , the fixed switch plays a role in the round - about , this is why it may be crossed by a double locomotive . table [ fixtab ] displays the rules used for the passive fixed switch . more precisely , the table gives the rules which were not yet used . in fact several rules from the vertical motions are also used in this case . as illustrated by the figures , the switch is passive and it works both for a single and a double locomotive whatever the side from which they arrive to the switch . note that the exceptional cell is the centre of the switch , the cell 0(0 ) in the pictures of figures [ fixeg ] and [ fixed ] . rule 191 , is the conservative rule which is applied when no locomotive crosses the switch . its milestones are neighbours 2 , and 5 and also the vertex - neighbours 6 , 8 , 9 and 10 which belong to milestones of neighbouring elements of the tracks . note that the cell 2(1 ) is a milestone shared by two elements , each one belonging to one of the tracks which arrive to the switch . table [ exfixs ] gives traces of executions for a simple locomotive crossing the switch , either from the left , left - hand side part of the table , or from the right , right - hand side part of the table . table [ exfixd ] does the same for the same cells when a double locomotive crosses the switch . these traces deal with the cells 1(1 ) , 1(5 ) , 0(0 ) and 1(3 ) for the tracks and with the cells 1(2 ) and 1(4 ) for the particular milestones which are not in the situation of milestones in a vertical track or around a set of cells . note that the side from which the locomotive arrives to the cell 0(0 ) is indicated by rule 193 , for the left - hand one and rule 200 , for the right hand one . for rule 193 , the locomotive can be seen in the neighbour 4 of 0(0 ) when it comes from the left - hand side . for rule 200 , it can be seen in neighbour 3 when it comes from the right - hand side . it can be checked that many rules from table [ verttracktab ] are used in these executions , this is especially the case for the conservative rules applied to all cells of the tables . the motion rules are also those of the vertical tracks for cells 1(1 ) and 1(3 ) . in section [ scenar ] , we have seen the idle configurations of the other pieces of the round - about : the doubler , see sub - subsection [ subsdoubler ] and the selector , see sub - subsection [ subssel ] . figure [ doubleur ] illustrates the motion of the locomotive through the doubler . the locomotive arrives in the form of a simple one . the creation of the second cell of the locomotive happens in a simple way : when the locomotive enters the cell 1(1 ) , at the same moment , a black cell is created in the cell 0(0 ) . that event constitutes the creation of the double locomotive , which is clear in figure [ doubleur ] as well as in table [ exdoubl ] . the rules are given in table [ doubltab ] . however , rules of previous tables are also involved , in particular rules from the motions on the vertical tracks . the cells of the track involved in the doubler are , in the order they are traversed : 2(1 ) , 1(1 ) , 0(0 ) , 1(3 ) , 3(3 ) and 8(3 ) . the milestones are in the cells : 4(5 ) , 15(1 ) , 1(5 ) , 7(1 ) , 3(1 ) , 2(2 ) , 1(2 ) , 1(4 ) , 2(3 ) , 4(3 ) , 7(3 ) and 9(3 ) . here , the conservative rule for the central cell is rule 202 : we can see the milestones in neighbours 2 , 3 and 5 as well as in vertex - neighbours 6 and 10 . the creation of the double locomotive is performed by rules 207 , and rule 206 , simultaneously applied to the cell 1(1 ) and to 0(0 ) , respectively . note that in rule 206 , the occurrence of a black cell in the vertex - neighbour 9 triggers the creation of the front cell of the locomotive . although the fork is not connected with the round - about , we place it here as it appears at that point in the table of the rules . however , it has some link with the doubler as , like for that structure , two locomotives are issued from the fork . now , in contrast with what happens in the doubler where both locomotives travel together , always in contact with one another , in the fork each created simple locomotive goes its own way : they travel in different directions . figure [ fourche ] illustrates this structure . table [ forktab ] gives the rules for the fork and table [ exfork ] displays the rules which are applied in the visit of the fork by a locomotive . the cells of the tracks involved in the structure are : 8(3 ) , 3(3 ) , 1(3 ) , 0(0 ) and then 1(1 ) , 4(1 ) , 12(1 ) and 33(1 ) in one direction and 1(4 ) , 2(5 ) , 7(5 ) , 20(5 ) and 54(5 ) in the other direction . the milestones are : 20(3 ) , 23(3 ) , 7(3 ) , 10(3 ) , 2(3 ) , 2(4 ) and then 1(2 ) , 1(5 ) , 2(2 ) , 3(1 ) , 5(2 ) , 11(1 ) , 13(2 ) and 32(1 ) in the first direction and 4(4 ) , 6(5 ) , 3(5 ) , 19(5 ) , 8(5 ) and 53(5 ) for the second one . the neighbourhood of the cell 0(0 ) is exactly that of a cell of going up vertical tracks . it is given by the conservative rule 35 : the milestones are in neighbours 2 and 5 and the cell can see the milestones of the neighbouring elements in the vertex - neighbours 7 , 8 and 9 . when the simple locomotive is in the cell 0(0 ) , rules 41 , and rule 222 , apply to cell 1(1 ) and 1(4 ) respectively , simultaneously creating a locomotive at the corresponding place . in the cell 1(1 ) , the locomotive can be seen in its neighbour 3 while in the cell 1(4 ) , it can be seen in its neighbour 2 . note that the cell 1(4 ) behaves like the cells 10(1 ) and 5(2 ) we have looked at in sub - section [ les_horizontales ] . the cell 1(4 ) allows the locomotive to perform a turn at right - angle . figure [ select ] illustrates the working of the selector of the round - about . the case of a single locomotive is illustrated by the first two rows of the figure while the case of a double locomotive is illustrated by the last row . tables [ selectstab ] and [ selectdtab ] give the rules for the locomotive and for the cells of the selector , table [ selectstab ] for a simple locomotive , table [ selectdtab ] for a double one . note the very symmetric structure of the exit configurations when no locomotive is around . as shown by the figures , the cells of tracks involved in this structure are 8(3 ) , 3(3 ) , 1(3 ) and 0(0 ) for the arrival and then 1(2 ) , 2(2 ) , 5(2 ) and 13(2 ) for the path leading to the next selector and 1(4 ) , 2(5 ) , 7(5 ) and 20(5 ) for the track leading to the continuation of the required path . the milestones are 20(3 ) , 23(3 ) , 7(3 ) , 10(3 ) , 2(3 ) and 2(4 ) for the entry path , 1(1 ) , 6(2 ) , 4(1 ) , 14(2 ) , 12(1 ) and 35(2 ) for the path leading to the next selector and 1(5 ) , 6(5 ) , 3(5 ) , 19(5 ) , 8(5 ) and 53(5 ) for the track to the required path . we remain with the milestones of cells 1(2 ) and 1(4 ) as these cells play the selection role : besides the already mentioned 1(1 ) and 2(3 ) , we have 5(3 ) , 3(2 ) and 10(2 ) for 1(2 ) ; besides the already mentioned 1(5 ) and 2(4 ) , we have 5(4 ) , 7(4 ) , 18(4 ) , 4(4 ) and 10(4 ) for 1(4 ) . the conservative rule for 0(0 ) is rule 230 , and we can see that the milestones of the cell are neighbours 4 , 5 , 7 and 8 . rule 232 , and rule 236 , are the conservative rules for 1(2 ) and 1(4 ) respectively . note that they are rotated forms of one another . when the locomotive is a simple one , 1(2 ) remains white while the locomotive passes by . this can checked on rule 243 , and on rule 248 , the locomotive being seen in the neighbour 10 of 1(2 ) for the first rule and in its neighbour 5 for the second one . for the same passage , 1(4 ) very differently reacts : rule 245 , can see the locomotive through its vertex - neighbour 8 while rule 251 , now seeing the locomotive in the neighbour 2 of the cell makes it become black . this can be checked on table [ exsels ] . when the locomotive is double , 1(4 ) remains white . it is witnessed by rule 245 again , the locomotive is in the neighbour 8 of the cell , and then by rule 269 , applied when the locomotive is in the neighbours 2 and 8 of the cell . but this time , 1(2 ) behaves differently : rule 243 again applies as the cell can only see the front of the locomotive through a vertex neighbour : it can not know whether it is simple or double . at the next time , rule 193 , applies as the locomotive can be seen in the neighbours 5 and 10 respectively , making the cell becoming black . in section [ scenar ] we have seen that we can implement both a flip - flop and the active memory switch by using the same basic devices provided that they are suitably arranged . figure [ control ] shows how the controller is working , depending on its cell 1(1 ) whose state we call the * colour * of the controller . figure [ controls ] shows how the colour of controller is changed by the arrival of the appropriate signal : we remind the reader that the signal has the form of a simple locomotive . table [ controltab ] gives the rules corresponding to the illustration given by figures [ control ] and [ controls ] . also , table [ excontrol ] gives the traces of executions of the rules applied to significant cells when the locomotive arrives to the controller while table [ excontrols ] gives similar traces when the signal arrives to the controller in order to change its colour . note that the tracks which crosses the controller passes through the cells 8(4 ) , 3(4 ) , 1(4 ) , 0(0 ) , 1(2 ) , 3(2 ) and 8(2 ) in this order . the signal for changing the colour of the cell 1(1 ) passes through the cells 16(1 ) , 6(1 ) and 2(1 ) . the milestones of the path which crosses the controller are 54(4 ) , 57(4 ) , 20(4 ) , 23(4 ) , 7(4 ) , 10(4 ) , 2(4 ) , 2(5 ) , 1(3 ) , 1(1 ) , 4(2 ) , 2(2 ) , 9(2 ) , 7(2 ) , 20(2 ) and 22(2 ) . the milestones for the path followed by the signal are the cells 41(1 ) , 44(1 ) , 15(1 ) , 18(1 ) and 2(5 ) . there are three additional milestones which are specific to the controller : the cells 4(1 ) , 10(1 ) and 6(2 ) . figure [ control ] shows us that when the cell 1(1 ) is black , the locomotive passes through the tracks , because 1(1 ) constitutes a milestone of the track . when 1(1 ) is white , the expected milestone being missing , the locomotive is stopped , see rule 315 , the rule shows that the cell 0(0 ) remains white despite the fact that the locomotive is seen in the neighbour 3 of the cell . indeed , the conservative rule for 0(0 ) when the cell 1(1 ) is white is rule 279 , that rule can be compared with the one which is used when the cell 1(1 ) is black , namely rule 58 , table [ excontrol ] gives the traces of execution for the rules which apply to the cells 1(4 ) , 0(0 ) and 1(2 ) for the tracks , the cells 2(2 ) and 1(1 ) for milestones and the cell 2(1 ) for a witness of transformations induced by the passage of the locomotive . # 1 # 2 # 3 # 4 to 25pt#1 to 25pt#2 to 25pt#3 to 25pt#4 when the colour of the controller is black , the conservative rule for 1(1 ) is rule 294 , when the locomotive crosses the controller , cell 1(1 ) can see it first through in its neighbour 1 , rule 306 , and then in its vertex - neighbour 6 , see rule 310 , that rule and figure [ control ] show us that the cell 2(2 ) became white after seeing the locomotive in 0(0 ) , which is the neighbour 7 of 2(2 ) . indeed , in this situation , rule 266 , applies to 2(2 ) . when 2(2 ) is white , applying rule 312 , allows us to restore the black state in the cell . this problem comes from the fact that rule 266 was needed in the selector for the cell 0(0 ) when a double locomotive crosses that cell . note that after rule 310 , rule 294 again applies to the cell 1(1 ) . figure [ controls ] illustrates the action of the signal on the cell 1(1 ) : when the cell is black , the signal turns it to white and conversely . table [ excontrols ] displays which rules are used to which cells when the signal arrives in order to change the colour of the cell 1(1 ) . the conservative rules for the cell 1(1 ) are rule 294 , we have already seen , when the cell is black , it is rule 280 , when the cell is white . we can see that the neighbourhood is the same for rules 294 and 280 : the milestones are in the neighbours 4 , 5 and 9 of the cell , and the difference lies in the state of the cell itself . the change from black to blank is performed by rule 302 , which can see the signal in its neighbour 2 . the reverse change is obtained by rule 290 , again as the signal is seen in the neighbour 2 of the cell . # 1 # 2 # 3 # 4 to 25pt#1 to 25pt#2 to 25pt#3 to 25pt#4 other neighbours of the cell 1(1 ) can see those changes . as an example which is not given in table [ excontrol ] , we take the cell 1(5 ) which is always blank . the conservative rule is rule 299 when the cell 1(1 ) is black , and it is rule 286 when 1(1 ) is white . when the locomotive crosses the controller in the black colour , rule 305 , can see the locomotive in its neighbour 7 , then rule 308 , can see it in its neighbour 1 and then , rule 299 is again applied . when 1(1 ) is white , rule 316 , is applied when the locomotive appears : it can be seen in its neighbour 7 and then the locomotive disappeared , so that rule 286 again applies . table [ excontrols ] shows the rules applied when the signal arrives , changing the colour of the cell 1(1 ) . as we said in section [ scenar ] , the controller of the passive memory switch works in a different way than that of the active switches . for this reason , we call it the * controller - sensor * as it does not stop the locomotive when it runs through the non - selected track . as in the case of the active switches , this controller is characterized by the state of a cell , again the cell 1(1 ) , see figure [ controlsign ] . here , the cell 1(1 ) is called the * sensor * as its role is to detect a passage , not to stop it . as in the case of the controller , we say that the state of 1(1 ) is the * colour * of the sensor . the tracks goes along the cells 33(2 ) , 12(2 ) , 4(2 ) , 1(2 ) , 0(0 ) and 1(4 ) , 4(4 ) , 12(4 ) and 33(4 ) in this order . this order is the opposite of what he have seen for the passage of a locomotive through the controller of an active switch . this raises no problem thanks to the possibility to use round - abouts if needed : the round - abouts do not use the controllers nor the controller - sensors . the difference of direction induces a sharp difference in the milestones . here the milestones of the track of the locomotive are : 34(3 ) , 86(2 ) , 13(3 ) , 31(2 ) , 5(3 ) , 10(2 ) , 2(3 ) , 2(2 ) , 2(5 ) , 3(4 ) , 5(5 ) , 11(4 ) , 13(5 ) , 32(4 ) , 34(5 ) and 87(4 ) . now , there is a common point between the controller - sensor and the controller of the active switches : it is the path of the signal for changing the colour of the sensor . the path and the milestones are the same . however a difference again : here , in the case of the controller - sensor , the colour is changed only if the sensor is black . when it is white , it is changed to black by the very passage of the locomotive , there are also additional milestones : they consists of the cells 7(2 ) , 6(2 ) , 4(1 ) , 10(1 ) and 2(4 ) . those actions are illustrated by figure [ controlsign ] . table [ passmemotab ] gives all the new rules induced by those constraints . consider the case when the sensor is white . the conservative rule for 0(0 ) is rule 317 , its milestones are in its vertex - neighbours 6 , 8 , 9 and 10 . table [ exctrlsn ] displays the rules used for the cells of the controller - sensor when the locomotive passes through the non - selected track . when the locomotive arrives at the cell 0(0 ) , it is seen from the neighbour 3 of the cell see rule 325 , here , the side 1 of cell 0(0 ) is shared with 1(5 ) . then , rule 328 , makes the locomotive leave the cell . rule 25 , witnesses that the locomotive left 0(0 ) as it can be seen in the neighbour 5 of 0(0 ) . the rule also witnesses that the sensor turned to black : the cell 1(1 ) is the neighbour 2 of 0(0 ) which is now black . rule 25 is a rule of the vertical track is also applied here because at that very moment , the cell 2(2 ) became white and the cell 2(2 ) is the vertex - number 8 of the cell which , usually , is black . after that , rule 289 , is applied . it is the conservative rule of 0(0 ) when the sensor is black : remember that the cell 1(1 ) is the neighbour 2 of 0(0 ) . the rule also witnesses that the cell 2(2 ) returned to black . we leave to the reader the examination of the application of the rules given by table [ passmemotab ] to the cells indicated in table [ exctrlsn ] . the first two rows of figure [ controlsign ] are of help for such a task . table [ exctrlbands ] gives traces of execution for the rules applied to the cells when the locomotive crosses a black sensor and for those concerned by the change of the sensor from black to white . in the latter case , the comparison of the right - hand side part of table [ exctrlbands ] with the left - hand side part of table [ excontrols ] show us that the same rules apply . this explains why the last row of figure [ controlsign ] is very similar to that of the first line of figure [ controls ] for the part of these figures involved by the cells concerned by the signal : the track where the signal goes , their milestones and the cell 1(1 ) with its neighbours . similarly , the behaviour of the cell 2(2 ) is the same in the first line of figure [ control ] and in the first one of figure [ controlsign ] . an important difference should also be noticed between the first line of figure [ controlsign ] and the first line of figure [ control ] . it is the fact that the sensor is white in the first case while it is black in the second one . also , in the first case the sensor changes its colour while its colour is unchanged in the second one . this explains why a few rules in the tables [ exctrlsn ] and [ exctrlbands ] are taken from table [ passmemotab ] . as examples , consider the cells 0(0 ) , 1(1 ) and 2(2 ) when the locomotive passes while the sensor is black . as already mentioned , the conservative rule for 0(0 ) when the cell 1(1 ) is black is rule 289 . when the locomotive is seen by 0(0 ) , the locomotive is in the neighbour 3 of 0(0 ) , so that rule 344 , is applied . at the next time , as the locomotive vanished , rule 289 is again applied . in the case of 1(1 ) when it is black , the conservative rule is rule 294 , a rule from table [ controltab ] . when the cell can see the locomotive which appears as the neighbour 6 of 1(1 ) , rule 345 , is applied . again , as the locomotive vanishes at the next time , rule 294 is again applied . at last , the conservative rule for 2(2 ) is rule 143 , a rule for the horizontal tracks . here too , the presence of the locomotive is noticed for one time only : the locomotive is in the neighbour 1 of 2(2 ) : rule 347 , is applied . at the next time , rule 143 is applied : the locomotive vanished . the tables of this section completes the proof of theorem [ letheo ] . we already mentioned in section [ rules ] that the automaton we constructed to prove theorem [ letheo ] is not rotation invariant . we used this property in order to suitably number the sides of each cell . among the rules displayed in the various tables of section [ rules ] , many of them are rotated forms of other rules . as an example , rules 3 , 39 , 126 and 286 , _ i.e. _ and respectively , are rotated form of each other . they apply to white cells with a single black neighbour , neighbour 1 , 2 , 5 and 3 respectively . we shall say that in the corresponding neighbourhoods there is a shift by 1 , 4 and 2 for rules 39 , 126 and 286 respectively with respect to that of rule 3 . # 1 # 2 # 3 # 4 # 5 # 6 # 7 now , there are 14 pairs of rules in which the neighbourhoods are rotated , the state of the cell is the same but its new state is different . tables [ diffs1 ] and [ diffs2 ] give the rules together with the cells to which they apply and in which part of the circuit . the part of the circuit is identified by a few letters as follows . for the tracks , v means vertical and h means horizontal ; d as second letter means down while u means up ; d , s as a last letter indicates a double , simple locomotive , respectively ; c , a means a clockwise , counter - clockwise running , respectively ; w means that the horizontal goes around a white node ; when w is missing , this means that the horizontal goes around a black node . the fixed switch is represented by fx , l , r indicates from the left- , right - hand side respectively . the fork is represented by fk , the doubler by d. the controller is denoted by c while the controller - sensor is defined by sn , b , w indicate the colour of the controller or the sensor and s indicates the run involving the signal . at last , sl denotes the selector , s and d indicating whether it is crossed by a simple or a double locomotive . the tables indicate that the relaxation of the rotation invariance plays an important role . however , the number of rules concerned is a bit less than ten percents of the rules . also the high number of rotated forms of rules is an interesting feature . at last , besides the first rules numbered from 1 to 10 , several rules are used in many different places . as an example , rule 190 , is used in many places : 5(3 ) in horizontal tracks which clockwise go around a white node , both for the simple and the double locomotive . the rule is also applied to the cell 1(5 ) in the controller and in the controller - sensor when the black colour is changed to the white one . we have already seen that these motions of the signal are very similar in both those cases . three rules have all their rotated forms present in the tables : rule 16 , rule 23 and rule 56 . as an interesting example , we have chosen rule 16 , whose other forms are rule 81 , rule 93 , rule 136 , and rule 320 , with respect to rule 16 , the shifts applied to rule 81 , 93 , 136 and 320 are 3 , 2 , 4 and 1 , respectively . rule 16 is present very often , in many situations , the horizontal motions excepted . rule 81 is applied to 1(1 ) and 1(3 ) each time once in the horizontal motions with the simple locomotive around a black node . rule 93 is applied once to 1(1 ) in the same motions . rule 136 is applied once to 1(3 ) in the clockwise motion only . rule 320 is applied to 5(3 ) and 13(3 ) , once in the sensor , in the passage of the locomotive , whatever the colour of the sensor . it should be noted that the rules of table [ diffs1 ] apply to many cells , in particular the first five pairs of the table . this is due to the fact that the rules mainly concern elements of the tracks . by contrast , in each pair of table [ diffs2 ] , each rule concerns at most two cells . several questions are raised by this result . how to reduce the number of states in this situation ? is it possible to obtain two states with a moore neighbourhood and still observing rotation invariance ? still hard work ahead . m. margenstern , _ cellular automata in hyperbolic spaces , vol . @xmath33 , theory _ , collection : _ advances in unconventional computing and cellular automata _ , editor : andrew adamatzky , old city publishing , philadelphia , ( 2007 ) , 422p . m. margenstern , _ cellular automata in hyperbolic spaces , vol . @xmath34 , implementation and computations _ , collection : _ advances in unconventional computing and cellular automata _ , editor : andrew adamatzky , old city publishing , philadelphia , ( 2008 ) , 360p .

type i supernovae ( sne i ) are characterized by the lack of prominent hydrogen lines in the spectra ( e.g. , * ? ? ? strong helium lines are present in the spectra of sne ib , while they are practically absent in those of sne ic . sne ib / c are further distinguished from sne ia by the lack of strong siii absorption line at 6355 . most of ordinary sne ib / c , if not all , occur in star - forming galaxies , indicating that sne ib / c have a massive star origin ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? their light curves are dominated by the energy release from radioactive @xmath0ni as in the case of sne ia @xcite , but the inferred amounts of @xmath0ni ejected by sne ib / c are similar , on average , to those of sne ii ( @xmath1;e.g . , @xcite ) rather than sne ia ( @xmath2 ; e.g. @xcite ) . the current consensus is that most of sne ib / c belong to a subset of core - collapse sne that are sn explosions via collapse of the iron cores in massive stars at their deaths . hydrogen can not be easily hidden in sn spectra ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) and sn ib / c progenitors must have lost their hydrogen envelopes by the time of explosion . there exist mainly three possible ways for massive stars to become a hydrogen - deficient sn progenitor : mass loss from single stars via stellar winds ( e.g. , * ? ? ? * ) , binary interactions ( e.g. * ? ? ? * ) , and chemically homogeneous evolution with rapid rotation @xcite . the last mode of evolution have been invoked for explaining massive blue stragglers and long gamma - ray bursts within the collapsar scenario @xcite , but is not likely to be much relevant for the majority of sne ib / c that are found in the local universe @xcite . in this review , we focus on ordinary sne ib / c : our objective here is to summarize theoretical results on sn ib / c progenitors via single and binary evolutionary paths . we emphasize that each case has its own unique prediction that can be in principle well tested by observations . sne ib / c can therefore provide an invaluable insight on massive star evolution . note that we restrict this review to the detailed properties of sne ib / c progenitors that are predicted by recent stellar evolution models . progenitors of sn iib ( i.e. , sne of which the spectra have hydrogen lines at early times , but resemble those of sne ib at later times ) are closely related to sn ib / c progenitors , and will also be discussed briefly . our discussions on observations , sn modeling , and stellar population studies will be highly biased by the selected topics we address here . for more general topics on the evolution of massive stars and sn progenitors , readers are referred to the recent reviews by @xcite , @xcite , @xcite , @xcite , @xcite and @xcite . it has been widely believed that wolf - rayet ( wr ) stars are observational counterparts of sne ib / c progenitors ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? although helium stars as wr stars can be produced by binary interactions ( e.g. , * ? ? ? * ; * ? ? ? * ) , a large fraction of wr stars are found in isolation @xcite , and must have been produced from massive single stars via mass loss due to stellar winds ( the so - called conti scenario ; @xcite ) . evolutionary models of massive stars with mass loss predict that there exists an initial mass limit for wr stars , above which stars can lose the entire hydrogen envelope during the post main sequence phases ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? a useful constraint on this mass limit can be provided by galactic wr stars . observations indicate that wr stars of wn type in our galaxy have the lower bolometric luminosity limit of @xmath3 @xcite . this roughly corresponds to 10 @xmath4 of a naked helium star , which requires an initial mass of about 25 @xmath4 . stellar evolution models indicate that non - rotating stars at solar metallicity can not lose their hydrogen envelope to become a wr star if @xmath5 @xmath4 , with the most commonly adopted mass loss rate from red supergiant stars given by @xcite . enhancement of mass loss due to rotation or pulsation compared to the de jager rate and some alternative empirical mass loss prescriptions have been invoked to resolve this discrepancy ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? once a star becomes a wr star , further mass loss due to wr winds determines its final mass . in the 80s and 90s , a fixed value of about @xmath6 or mass - dependent values have been widely used for the wr mass loss rate in most evolutionary models @xcite . later studies began to consider the wr mass loss rate as a function of the luminosity and the surface abundances of helium and metals in a more explicit way ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? more important , with the growing evidence for hydrodynamic clumping of wr wind material , recent estimates for the wr mass loss rate give significantly lower values than previously thought . several different prescriptions for the wr mass loss rate are compared in fig . [ fig : wrwind ] . the single star models produced later than 2000 predict systematically higher final masses of sn ib / c progenitors than those in the 80s and 90s , as summarized in fig . [ fig : finalmass ] . for example , with the langer s mass - dependent wr mass loss rate @xcite , a 60 @xmath4 star at solar metallicity can become a sn ib / c progenitor with an final mass as low as 4.25 @xmath4 @xcite . by contrast the models with the wr mass loss rate of @xcite give final masses higher than 10 @xmath4 , at solar metallicity . this high final mass ( @xmath7 @xmath4 ) has consequences on the sn explosion . first of all , such massive helium stars have large amounts of binding energy . this would make successful explosion of these progenitors difficult : they may collapse to a black hole , without making an ordinary sn ib / c ( e.g. * ? ? ? * and references therein ) . secondly , even if they exploded successfully , the resultant light curves would be too broad to be compatible with observations @xcite . this brings into question the importance of single wr stars as progenitors of sne ib / c at solar metallicity . however , the role of single stars at super - solar metallicity may be significant given their relatively low final masses ( fig . [ fig : finalmass ] ) . ) . the dot - dashed line and the dotted line give the wr mass loss rates by @xcite and @xcite , respectively . the solid line denotes the mass loss rate prescription given by eq . ( 1 ) : the wr mass loss rate by @xcite for @xmath8 and the mass loss rate of relatively low - mass helium stars for @xmath9 , which is based on the extreme helium stars analyzed by @xcite . the blue data points with the error bars are the mass loss rates of these extreme helium stars . the orange point with the error bars denotes the mass loss rate of the quasi - wr star hd 45166 @xcite . the dashed line is 10 times lower than the solid line : @xmath10 is the reduction factor compared to the mass loss rate given by eq . ( 1 ) . ] , asterisk : rotating models of @xcite at @xmath11 , star : rotating models of @xcite at @xmath12 with a metallicity dependent wr mass loss rate . triangle : non - rotating models of @xcite at @xmath11 , square : non - rotating models of @xcite at @xmath11 . ] wr stars have large convective cores , being close to the eddington limit , and rapidly lose the outer helium - rich layers by wr winds . this makes them almost chemically homogeneous . therefore , stellar evolution models predict that wr stars evolve systematically bluewards on the hr diagram ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , in contrast to the case of the evolution of hydrogen - rich stars which evolve redwards in general . the surface composition of chemical elements also evolves . first , as the residual of the hydrogen envelope is removed by winds , they evolve from wnl type to wn type . as they lose more mass , the products of helium burning including carbon and oxygen begin to appear at the surface , to become wc and wo stars ( fig . [ fig : surface ] ) . the general consensus is that wr stars from sufficiently high initial masses evolve according to the following order : wnl @xmath13 wn @xmath13 wc @xmath13 wo . in this scenario , wr stars should become more compact as they evolve from wnl to wo . indeed , wnl and wo stars in our galaxy have lowest and highest surface temperatures , respectively , in agreement with the theoretical prediction @xcite . but stellar evolution models have great difficulty in explaining many of the surface properties of wr stars . in particular , the observed wr stars are found to have much larger radii and lower surface temperatures than what the evolutionary models predict ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? the reason for this discrepancy is not well understood yet . inflation of the envelope with a density inversion is observed in wr star models near the eddington limit in hydrostatic equilibrium @xcite , but this is still not sufficient to fit observations as seen in fig . [ fig : hr ] . a recent suggestion is that the observationally implied inflation of wr stars may result from density inhomogeneities and the consequent enhancement of opacity in the sub - surface convective layer @xcite . this envelope inflation affects the bolometric correction , making wr stars fairly luminous in the optical ( @xmath14 ) . if the optical luminosities of the observed wr stars represented those of sn ib / c progenitors at the pre - sn stage , the previous search for a sn ib / c progenitor in pre - sn images would have been successful @xcite . as of today , only one tentative identification has been reported with the sn ib iptf13bvn @xcite . all of the other searches for sn ib / c progenitors have failed , even for the case with a very deep detection limit ( @xmath15 ; @xcite ) . this result has often been interpreted as evidence for binary star progenitors @xcite . it should be noted that the majority of the observed wr stars must be on the helium main sequence , which is still far from the final evolutionary stage . after core helium exhaustion , the evolution of the core in a wr star is dominated by neutrino cooling and undergoes rapid kelvin - helmoltz contraction . with a sufficient amount of helium in the envelope , this would lead to further expansion of the helium envelope due to the so - called mirror effect . however , single wr stars would rapidly lose helium in the envelope as implied by the high mass loss rate , and the overall radius would gradually decrease as the stellar evolution models predict . at the pre - sn stage , many wr stars would tend to become very hot , and optically faint like wo stars despite their very high bolometric luminosities @xcite . this means that the non - detection of most sn ib / c progenitors in the previous attempts does not necessarily exclude single wr progenitors , and other constraints like ejecta masses of sne ib / c should also be taken into account to better understand the nature of sne ib / c progenitors . helium star at @xmath11 , with the wr mass loss rate by @xcite . the calculation was terminated at the end of core neon burning . ] the production of he i lines is found to depend both on the total he mass @xcite and on the helium distribution in the envelope @xcite . non - thermal excitation and ionization of helium also play the key role for the formation of helium lines in sne ib @xcite . this does not only require presence of helium in the progenitors , but also strong chemical mixing between helium in the envelope and radioactive @xmath0ni produced in the innermost region of the sn ejecta @xcite . we still do not know exactly how much helium is needed for sne ib . this limit must depend on the degree of mixing of helium and nickel , which may in turn depend on the detailed structure of the progenitor and the energy and asymmetry of the explosion . many authors simply assume a certain amount of helium ( e.g. , 0.5 - 0.6 @xmath4 ) as the lower limit for sn ib progenitors ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? recently @xcite suggested 0.14 @xmath4 as the maximum possible amount of helium that can be hidden in the sn spectra , based on a spectroscopic study of several sn ib / c with relatively low inferred ejecta masses . in the most recent single star models @xcite , the total amounts of helium in sn ib / c progenitors range from 0.28 @xmath4 to 2.2 @xmath4 . this is significantly higher than the proposed limit of 0.14 @xmath4 by hachinger et al . in fact , helium mass as low as 0.14 @xmath4 is very difficult to achieve with stellar evolution models . @xcite found that a 60 @xmath4 star can become a 4.25 @xmath4 sn ib / c progenitor with a wr mass loss rate much higher than nowadays adopted , but even in this extreme case , the remaining helium mass was as large as 0.18 @xmath4 . the reason for this difficulty is largely related to the dynamical adjustment of the stellar structure of wr stars with mass loss . as shown in fig . [ fig : kipp1 ] as an example , the size of the helium - burning convective core in a wr star decreases as the wr star loses mass by winds ( fig . [ fig : kipp1 ] ) , and therefore some amount of helium can remain unburned until the end of core helium burning even if more than half of the initial mass is lost . the residual helium could be completely removed with efficient mass loss during the later evolutionary stages . the current models predict , however , that the effect of mass loss during the post helium burning phases is relatively minor mainly because of the relatively short evolutionary time . we discuss the problem of helium in sne ib / c progenitors in sect . [ sect : binaryhe ] in more detail . the mass fraction of helium in the outermost layers can also play an important role for the early time light curves and spectra of sne ib / c . @xcite showed that if the helium mass fraction is sufficiently large ( @xmath16 ) , he i lines can be produced without the contribution of non - thermal processes for several days after the shock breakout , while no helium lines are seen for a low helium mass fraction ( @xmath17 ) even with a total helium mass of about 1 @xmath4 . therefore , he i lines during the early epoch of a sn ib / c will provide an important constraint on the progenitor . recent single star models in the literature predict that helium mass fraction at the surface of sn ib / c progenitors is below 0.4 except for a limited initial mass range above the critical mass for wr star progenitors ( e.g. , * ? ? ? * ; * ? ? ? * ) , in contrast to the case of binary star models that predict the majority of sn ib / c progenitors have a surface helium mass fraction higher than 0.9 ( sect . [ sect : binaryhe ] ) . given that the total amount of helium is also systematically smaller in single star models than in binary star models as discussed below , this implies that single star evolution is probably prone to sne ic . helium star , for which the nugis & lamers wr mass loss rate was adopted . the helium - burning convective core is marked by the hatched lines . the black solid line marks the surface of the star . the calculation was terminated at the end of core neon burning . ] super - luminous supernovae ( slsne ) and very energetic explosions like gamma - ray bursts ( grbs ) can be driven by rapid rotation . the most commonly invoked mechanisms for these events include the collapsar scenario @xcite and the magnetar scenario ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? so far , all of the supernovae associated with grbs belong to type ic @xcite , and many slsne are also found to be sne ic . while the collapsar scenario still remains most popular to explain grbs @xcite , the magnetar - driven explosion is nowadays the most invoked mechanism for the slsne ic ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? the contending mechanism for slsne ic is the pair - production instability ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? it is still a matter of debate which evolutionary channels of sne ib / c progenitors can lead to rotation - driven explosions like grbs and slsne - ic . observations indicate that the majority of massive stars on the main sequence are rapid rotators , where the necessary condition for both collapsar and magnetar mechanism could be fulfilled if they retained the angular momentum until the pre - collapse stage ( e.g. , * ? ? ? however , massive stars may undergo angular momentum redistribution via mass loss due to stellar winds and/or binary interactions , and the transport of angular momentum @xcite . the angular momentum transfer may occur on a dynamical timescale in convective layers by convection . in radiative layers , rotationally - induced hydrodynamic instabilities like the shear instability and eddington - sweet circulations may transport angular momentum @xcite . dynamo actions may also occur in radiative layers according to the so - called tayler - spruit dynamo theory @xcite , which may cause strong magnetic torques across differentially rotating layers . theoretical studies indicate that , without magnetic fields , angular momentum transport is severely inhibited by the chemical stratification across the boundary between the stellar core and the hydrogen envelope ( @xmath18-barrier ; @xcite ) . single star progenitors of sne ib / c can thus retain a significant amount of angular momentum until the pre - sn stage , even though most of the initial angular momentum is lost by stellar winds @xcite : the predicted amounts of angular momentum in the cores are much more than what neutron stars can have at the break - up velocity , and enough to produce a long gamma - ray burst ( grb ) within the collapsar scenario @xcite . this means that almost all sn ib / c progenitors have enough angular momentum to form grb / slsn - ic progenitors . given that rapid rotation is only one necessary condition for collapsar / magnetar production , this should not necessarily lead to the conclusion that non - magnetic models predict too many grbs and slsne - ic compared to the observation . however , because such exotic explosions belong to a subset of sne ib / c , the result of non - magnetic models implies that grbs and slsne - ic should occur more frequently at higher metallicity as ordinary sne ib / c do , if mass loss by winds provided the main evolutionary path for their progenitors . contrary to this expectation , observations indicate that low - metallicity is preferred for both grbs and slsne - ic ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? magnetic torques can easily overcome the hindrance by the chemical stratification to the transport of angular momentum . magnetic models with the tayler - spruit dynamo predict that single wr stars rotate too slowly to produce a magnetar / collapsar @xcite . this is consistent with the fact that grbs and slsne - ic are very rare . magnetic models also better explain the spin rates of young millisecond pulsars . several authors have questioned the validity of the talyer - spruit dynamo theory @xcite and we still can not draw any solid conclusion on which case between magnetic and non - magnetic models better represents the reality . however , it is not only massive stars but also intermediate- and low - mass stars that provide evidence for very efficient transport of angular momentum in the radiative layers . such examples include slow rotation of isolated white dwarfs @xcite , the radial velocity profile in the sun @xcite , and recent asteroseismic results of low - mass stars @xcite . our tentative conclusion is that some strong braking mechanism like the tayler - spruit dynamo is actually working in stars , and that magnetic models may better explain recent observations in general . the role of magnetic fields on the evolution of stars still remains a very challenging subject of future study . some other mechanisms like baroclinic instability @xcite and pulsational instabilities @xcite may also play an important role for the transport of angular momentum , but have not been extensively studied for massive stars yet . the majority of massive stars form in binary systems ( e.g. , see * ? ? ? * for a recent observational analysis on the population of binary systems ) . a large fraction of them are believed to experience binary interactions during the course of their evolution , mainly due to the increase of stellar radius . once the more massive star ( the primary star ) fills the roche lobe in a binary system , mass transfer to the less massive star ( the secondary star ) begins . mass transfer can be unstable if the mass ratio of the stellar components ( i.e. , @xmath19 where @xmath20 and @xmath21 are the masses of the primary and secondary stars , respectively ) is sufficiently small . unstable mass transfer will lead the binary system to a contact phase , which may eventually make the stellar components merge to become a single star . although binary mergers are related to many important topics like stellar rotation , peculiar sne and long gamma - ray bursts @xcite , here we focus our discussion on non - merging systems that can produce ordinary sne ib / c . recent analyses also indicate that the fraction of stable binary systems is much higher than previously believed @xcite , and thus the event rate of sne ib / c can be well explained by binary progenitors @xcite . in the literature , mass transfer is often categorized into case a , case b and case c , depending on the evolutionary stage of the primary star when it fills the roche lobe @xcite , as the following : * case a mass transfer : mass transfer during the main sequence . * case b mass transfer : mass transfer during the helium core contraction phase . * case c mass transfer : mass transfer during the core helium burning and later evolutionary stages . some binary systems may undergo multiple mass transfer phases , depending on the initial orbital periods and masses of the stellar components . the mass transfer phases that follow case a / b mass transfer are often denoted as the following : * case ab mass transfer : mass transfer from the primary star that has previously undergone the case a mass transfer , during the helium core contraction phase . * case bb mass transfer : mass transfer from the primary star that has previously undergone case b mass transfer , during the late evolutionary stages ( mostly after core helium exhaustion for sn ib / c progenitors ) . * case abb mass transfer : mass transfer from the primary star that has previously undergone case ab mass transfer , during the late evolutionary stages ( mostly after core helium exhaustion for sn ib / c progenitors ) . helium stars as sn ib / c progenitors can be made via case b / ab mass transfer as illustrated in fig . [ fig : binary ] . the initial mass of such a helium star corresponds to the helium core mass ( @xmath22 ) of the primary star at the onset of case b / ab mass transfer . for case b systems , @xmath22 can be given by a well - defined function of the zams mass of the primary star as shown in fig . [ fig : mimf ] . the evolution of the helium stars produced via case b / ab mass transfer largely depends on their masses . stellar evolution models indicate that case bb / abb mass transfer occurs when the primary star becomes a helium giant if the helium star mass is initially less than about 3.5 - 4.0 @xmath4 , depending on the adopted mass loss rate and metallicity . more massive helium stars do not interact anymore with the secondary stars after case b / ab mass transfer , but can still lose mass further by winds . wr winds may be induced if the helium star mass is sufficiently high ( @xmath23 @xmath4 ) but the mass loss rate from less massive helium stars is not well constrained observationally because such relatively low - mass helium stars have been rarely observed ( see sects . [ sect : binarymass ] and [ sect : counterpart ] below for more discussions ) . it is not only the primary star , but also the secondary star that can produce a sn ib / c . in a close binary system , the primary star will leave a compact star remnant if it explodes as a sn ib / c via case b / ab / bb / abb mass transfer , or if it becomes a white dwarf via case bb / abb mass transfer ( see fig . [ fig : binary ] ) . unless the binary system is unbound by a strong neutron star kick , it will form a common envelope after the core hydrogen exhaustion in the secondary star . a short - period binary system consisting of a helium star plus a compact star ( white dwarf , neutron star , or black hole ) will be produced after the common envelope ejection . explosion of the helium star will then produce a sn ib / c . note also that fig . [ fig : binary ] still does not depict all the possible binary paths for sne ib / c , and there may exist other relatively rare channels . for example , some authors found that with case a mass transfer , the sn order can be reversed for a limited parameter space : the secondary star first explode as a sn ib / c , followed by sn ib / c explosion of the primary star as a helium star in isolation or in a compact binary system with a neutron star companion , depending on the impact of the neutron star kick @xcite . the contribution of each evolutionary path to the total production of sne ib / c may depend on several physical parameters . they include the so - called mass accretion efficiency ( i.e , the ratio of the accreted mass onto the secondary star to the transferred mass from the primary star ) , the specific angular momentum of any matter that is not accreted on the secondary star but lost from the binary system , distribution of neutron star kick velocities , and the common envelope ejection efficiency ( e.g. * ? ? ? it is beyond the scope of this review to discuss all the details regarding binary population with respect to sne ib / c , and readers are referred to @xcite , @xcite , @xcite , as well as the contribution by s.e . de mink in this issue . here , it may be sufficient to say that the dominant channel to sne ib / c in binary systems is the case b / bb , among others @xcite . in the sections below , we focus on the detailed properties of sn ib / c progenitors predicted from evolutionary models . to model the evolution of a binary system , we have to consider the change of the orbit due to stellar winds , mass transfer and/or gravitational wave radiation , and mass exchange between the stellar components via mass transfer . to investigate the effect of rotation , angular momentum exchange between stars and the orbit via tidal synchronization and the spin - up effect of the secondary during the mass transfer phases should also be followed ( see * ? ? ? * for a recent review ) . many evolutionary models of massive binary stars have been presented in the literature , but only a limited number of studies aimed at detailed investigation of the structure of sne ib / c progenitors near / at the pre - sn stage @xcite . before we summarize the main results of these studies , we give an example for the evolution of relatively low - mass sn ib / c progenitors , which is very different from that of massive wr stars . the evolution of a sn ib / c progenitor having @xmath24 @xmath4 in a case bb system is illustrated in figs . [ fig : hr2 ] and [ fig : kippen ] . after case b mass transfer , the primary star becomes a hot and compact helium star . the surface hydrogen and helium mass fraction at this stage is about 0.28 and 0.7 , respectively and a small amount of hydrogen of about 0.05 @xmath4 is still retained in the outermost layer . as a result , the convective helium burning core can grow with hydrogen shell burning even though the total mass somewhat decreases due to mass loss by winds ( fig . [ fig : kippen ] ) . this is contrasted to the case of mass - losing pure helium stars where the convective helium burning core shrinks in size in terms of the mass coordinate ( fig . [ fig : kipp1 ] ) . during the post - helium burning stages , the helium envelope rapidly expands and case bb mass transfer is initiated when carbon burning begins in the core ( figs . [ fig : hr2 ] and [ fig : kippen ] ) . the sn explosion will occur when the surface temperature becomes fairly low ( @xmath25 , fig . [ fig : hr2 ] ) while the star is still undergoing case bb mass transfer . this is in stark contrast to the case of massive wr star progenitors that evolve bluewards throughout and explode when they become very hot ( @xmath26 ; fig . [ fig : hr ] ) . the helium envelope expansion makes this binary progenitor bright in the optical bands compared to the case of wr progenitors , as discussed below ( sect . [ sect : binarysurface ] ) . of course , more massive progenitors having @xmath27 will become a wr star even in binary systems after case b / ab transfer , and evolve like a single wr star thereafter . plus 14 @xmath4 stars with the initial period of 5 days . convective layers are marked by green hatched lines , and semi - convection layers by red dots . the surface of the star is indicated by the black solid line . the blue and pink color shading marks net energy gain or loss from nuclear energy generation and neutrino emission . ] as mentioned above , one of the biggest uncertainties in the evolution of sn ib / c progenitors is the mass loss rate of naked helium stars . the mass loss rate of wr stars ( @xmath28 ) is relatively well known , but less luminous helium stars have not been well studied observationally . several authors therefore used extrapolated values of the nugis & lamers rate or the langer rate ( see fig . [ fig : wrwind ] ) for the whole possible range of helium star mass @xcite . on the other hand , @xcite and @xcite used a significantly reduced mass loss rate for @xmath29 as the following : @xmath30[eqhewind ] here the wr mass loss rate for @xmath31 is given by @xcite . the prescription for less luminous helium stars ( @xmath29 ) is based on the observations of extreme helium stars by @xcite . [ fig : wrwind ] indicates that the simple extrapolation of the wr mass loss rate down to @xmath32 may lead to a significant overestimate even with the nugis & lamers rate , which is about 10 times lower than that of @xcite for @xmath31 . given that the stellar wind mass loss rates of massive stars were usually overestimated before the late-90s ( see sect . [ sect : singlemass ] above ) , mass - loss rates reduced by a certain factor ( @xmath10 ) compared to that of eq . ( 1 ) were applied for some binary models of ( * ? ? ? * @xmath33 ) and for all the models of ( * ? ? ? * @xmath34 ) . note that the case of @xmath35 , for which the mass loss rate becomes comparable to the nugis & lamers rate with @xmath31 , is still compatible with the observation of extreme helium stars of @xmath29 ( fig . [ fig : wrwind ] ) . the caveat is that the extreme helium stars in the figure are giant stars at a pre - white dwarf stage , having masses of only about 0.8 @xmath4 @xcite , and can not represent ordinary sn ib / c progenitors . on the other hand , the quasi - wr star hd 45166 is currently one of the most promising observational counterparts of relatively low - mass helium stars on the helium main sequence ( sect . [ sect : counterpart ] ) . its mass is about 4.2 @xmath4 with a surface helium mass fraction of 0.5 and surface luminosity of @xmath36 . this star is likely on the helium main sequence . the estimated mass loss rate gives a better agreement with the extrapolated value of the nugis & lamers rate than the mass loss rates of extreme helium stars of @xcite . but we still have only one sample for such relatively low - mass helium stars in the core helium burning phase , and can not make a solid conclusion on which mass loss prescription is best suited for our purpose . as shown in fig . [ fig : finalmass2 ] , however , this uncertainty does not make a great difference in terms of the final masses . the predicted final masses of sn ib / c progenitors that undergo case b / bb mass transfer are given by fig . [ fig : finalmass2 ] . the models in the figure , for which the considered initial period of the orbit ranges from 4 to 7 days , were taken from @xcite and @xcite . in @xcite , conservative mass transfer ( i.e. , the mass accretion efficiency @xmath37 , which is the ratio of the transferred matter from the primary to the accreted matter onto the secondary , equals to 1.0 ) was assumed and rotation was not taken into account . @xcite considered the effects of rotation , with which @xmath37 is self - regulated by the interplay between the mass transfer from the primary star and the mass - loss enhancement due to rotation from the secondary star that is spun - up by mass and angular momentum accretion . these models indicate @xmath38 for case a mass transfer and @xmath39 for case b mass transfer , respectively . for case b systems , the final masses converge to about 3.15 @xmath4 with @xmath40 and gradually increases with increasing @xmath41 for @xmath42 5 and 10 . the result with the nugis & lamers rate is comparable to that with @xmath35 . for @xmath43 @xmath4 , helium stars produced by case b mass transfer undergoes case bb mass transfer during carbon burning and later phases . consequently the final masses for case bb systems decrease more rapidly from this point with decreasing @xmath41 , and the lower boundary of the zams mass for sn ib / c progenitors becomes about 12.5 @xmath4 , below which the primary star becomes a white dwarf . in a case a system , the primary star loses a significant fraction of the initial mass on the main sequence , and the helium core mass at the end of core hydrogen burning becomes lower than the corresponding case b system . this makes the lower limit of zams mass for sne ib / c shift to about 16 @xmath4 for case a systems @xcite . very short orbits ( typically @xmath44 ) are required for case a systems , and their contribution to the sne ib / c rate from binary systems may be smaller by about a factor of 3 4 than that of case b systems ( cf . * ? ? ? these minimum zams mass limits for sne ib / c ( i.e. , 12.5 @xmath4 and 16 @xmath4 for case b and case a , respectively ) are significantly higher than for sne iip from single stars , which is about 8 - 9 @xmath4 @xcite . note that binary interactions can make this limit for sn iip lowered even down to about 4 @xmath4 : for example , mergers of 4 @xmath4 plus 4 @xmath4 star would make a 8 @xmath4 sn iip progenitor . therefore , the stronger association of sne ib / c with younger stellar populations than sne iip in their host galaxies ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) is in qualitative agreement with the binary scenario . according to recent analyses of sn light curves @xcite , sne ib / c have typically ejecta masses of 1 @xmath4 6 @xmath4 . this means that , assuming the remnant mass of 1.4 @xmath4 , the final masses of ordinary sne ib / c range from 2.4 @xmath4 to 7.4 @xmath4 at least . if some amounts of helium were hidden in these analyses as argued by @xcite , the actual final mass would be somewhat higher . this observation is not compatible with the case of @xmath45 in fig . [ fig : finalmass2 ] , which predicts too low ejecta masses on average . the results with lower mass loss rates give a better agreement : stars of @xmath46 14 @xmath4 - 35 @xmath4 can explain the observed ejecta mass range with @xmath35 and the nugis & lamers rate for example . @xcite also gives a similar conclusion from the comparison of their stellar population model with the observation . the effect of different values of @xmath10 can be regarded as the metallicity effect , because both observations and theories indicate that wr mass loss becomes stronger for higher metallicity @xcite . the result in fig . [ fig : finalmass2 ] implies that sn ib / c ejecta should be systematically more massive for lower metallicity , which can be tested by observations . if there were a certain mass cut for the boundary between successful sn explosion and black hole formation in terms of the final mass , this metallicity dependence of final masses would lead to gradual decrease of the sn ib / c rate for decreasing metallicity ( cf . * ; * ? ? ? for example , if @xmath47 @xmath4 were the lower limit for bh formation , we would not expect a sn ib / c explosion from @xmath48 @xmath4 at solar metallicity ( @xmath49 ) and @xmath50 @xmath4 at smc metallicity ( @xmath51 ) , respectively . denotes the reduction factor that are applied to the mass loss rate given by eq . ( 1 ) : @xmath52 10 and 20 roughly corresponds to solar and smc metallicity , respectively . the presented results are based on the full binary models by @xcite for @xmath45 and the binary models and pure helium star models by @xcite for the others . the boundary line between case b and bb systems is marked by the dashed line . the data with @xmath53 @xmath4 in @xcite is based on the full binary evolution calculations but it is based on pure helium star models for @xmath54 @xmath4 assuming that the pure helium star was produced via case b mass transfer from the primary star with the corresponding zams mass in a binary system . see @xcite for more details . the result with the wr mass loss rate by @xcite at solar metallicity from unpublished binary star models ( yoon , s .- c . , in prep ) is marked by the green dashed line , for which the boundary for case b and bb shifts to about @xmath55 @xmath4 . ] as discussed in sect . [ sect : singlehe ] , the amount of helium retained in progenitors may be one of the key factors that make the difference between sne ib and sne ic . as shown in fig . [ fig : hemass ] , helium mass varies from 0.1 to 1.9 for the considered range of zams mass and mass loss rates in case b / bb systems . helium mass ( @xmath56 ) as a function of @xmath41 has a local maximum at @xmath57 @xmath4 for @xmath42 1 and 5 , and at @xmath58 @xmath4 for @xmath42 20 , respectively . the rapid decrease of @xmath56 as @xmath41 approaches 12 @xmath4 results from case bb mass transfer , while the gradual decrease of @xmath56 with increasing @xmath41 is the effect of the increasing mass loss rate . it is important to note that , even for the case of very strong mass loss ( i.e. , @xmath45 ) , the amounts of helium at the pre - sn stage are significantly greater than the upper limit of @xmath59 @xmath4 for sne ic that was suggested by @xcite , except for the extreme case bb case at around @xmath60 @xmath4 . therefore , both single and binary star models can not fulfill the condition of low amounts of helium for sn ic . this might mean that , in reality , the chemical mixing between helium and nickel in sn ejecta is not as efficient as considered by @xcite for most sne ic ( cf . * ? ? ? * ) such that more helium could be hidden . such inefficient mixing is not usually supported by observations ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , but we need a systematic study on how the mixing efficiency via the rayleigh - taylor instability depends on the pre - sn structure of sne ib / c progenitors to resolve this issue . mass loss from helium stars is another uncertain factor that should be better understood . in particular , some helium stars can closely approach the eddington limit during the final evolutionary stages if the mass is high enough . this might make the surface layers unstable to cause rapid mass eruption ( cf . * ; * ? ? ? * ) , thus removing most of helium in the envelope shortly before the sn explosion as evidenced by many sne ib / c @xcite . @xcite assumed a certain value of the ratio of the helium mass to the ejecta mass ( @xmath61 ) as the demarcation criterion between sn ib and sn ic . because the chemical mixing between helium and nickel plays an important role for having helium lines in sn spectra , using @xmath61 instead of @xmath56 may be appropriate because a lower value of @xmath61 means that helium can be more easily shielded from the gamma - ray photons produced in the innermost nickel - rich layer of the sn ejecta . interestingly , fig . [ fig : hemass ] indicates that @xmath61 does not depend on the adopted mass loss rate ( hence on metallicity ) as strongly as @xmath56 does . this is because , for a lower mass loss rate , the final mass and the co core mass become higher and compensate the higher helium mass . this has important consequences in the prediction of sn ib / c rate as a function of metallicity as discussed in sect . [ sect : fate ] below . denotes the reduction factor that are applied to the mass loss rate given by eq . ( 1 ) ( see the figure caption of fig . [ fig : finalmass2 ] ) . the nl rate means the wr mass loss rate by @xcite . the data were taken from @xcite for @xmath45 , @xcite for @xmath52 5 , 10 and 20 , and unpublished models by yoon ( in prep . ) for the nl rate . _ lower panel _ : the corresponding ratios of the helium to ejecta mass . here we assumed that the remnant neutron star mass is 1.4 @xmath4 . see fig . [ fig : finalmass2 ] for the corresponding final mass.,title="fig : " ] denotes the reduction factor that are applied to the mass loss rate given by eq . ( 1 ) ( see the figure caption of fig . [ fig : finalmass2 ] ) . the nl rate means the wr mass loss rate by @xcite . the data were taken from @xcite for @xmath45 , @xcite for @xmath52 5 , 10 and 20 , and unpublished models by yoon ( in prep . ) for the nl rate . _ lower panel _ : the corresponding ratios of the helium to ejecta mass . here we assumed that the remnant neutron star mass is 1.4 @xmath4 . see fig . [ fig : finalmass2 ] for the corresponding final mass.,title="fig : " ] if relatively low - mass helium star progenitors in binary systems have helium masses of about 1 - 1.5 @xmath4 in their envelopes as the most recent models predict , this may lead to an early plateau phase due to helium recombination for several days as shown by @xcite . this prediction has recently been tentatively confirmed with the early - time light curve of sn ib 2006lc @xcite . even in the absence of non - thermal effects , helium lines will also be observed during this phase , given that the mass fraction of helium in the envelope is very high ( @xmath62 ; see sect . [ sect : singlehe ] ) . on the other hand , sn ib / c progenitors with a compact star companion ( see fig . [ fig : binary ] ) may undergo very strong mass transfer after core helium exhaustion , because these systems have a small mass ratio and short orbital period . most of the helium envelope can be stripped off in this case , and very little helium will be left at the pre - sn stage @xcite . this evolutionary channel has been often invoked to explain sne ic having low ejecta masses and fast declining light curves ( e.g. , * ? ? ? * ) , including some ultra - faint sne ic @xcite . the frequency of such events would be rare compared to ordinary sne ib / c , and its quantitative prediction heavily depends on the uncertain parameters related to the common envelope efficiency and neutron star kick . case b / ab mass transfer in a binary system stops when the hydrogen envelope of the primary star is almost stripped off , but it does not completely remove hydrogen . the amount of hydrogen remaining in the outermost layers of the helium core immediately after the mass transfer phase is typically @xmath63 @xmath4 . @xcite found that the final amount of hydrogen at the pre - sn stage is a function of the progenitor mass as the following . at solar metallicity , which roughly corresponds to the case with @xmath64 in figs . [ fig : finalmass2 ] and [ fig : hemass ] ( see also fig . [ fig : wrwind ] ) , no hydrogen will be left for helium stars having final masses of @xmath65 @xmath4 . for the case of @xmath66 @xmath4 , case bb / abb mass transfer becomes efficient enough to completely remove hydrogen . for the final mass range in - between , the remaining hydrogen until sn explosion will be about @xmath67 @xmath4 . on the other hand , the models with the nugis and lamers mass loss rate at solar metallicity lose all the hydrogen in the envelope for the whole mass range ( @xcite ; yoon , s .- c . in prep ) . therefore , in principle , the signature of hydrogen in sn ib / c progenitors can indirectly constrain the mass loss rate from helium stars if we have good information about the metallicity in the vicinity of the sn site . this small amount of hydrogen can be easily detected in the early time spectra of the resulting sn @xcite , in which case it will be identified as a sn iib ( cf . these progenitors have relatively small radii ( @xmath68 ) compared to the yellow - supergiant sn iib progenitors ( e.g. * ? ? ? * ; * ? ? ? * ) produced via case c mass transfer @xcite , and may belong to the compact category of sn iib @xcite . some authors argue that several sne classified as type ib also have weak hydrogen absorption lines at high velocity @xcite . therefore , whether the explosion of a helium star having a thin hydrogen layer may be recognized as sn iib or sn ib depends on the details of the sn observation . the progenitor mass range for which a thin hydrogen layer is present at the pre - sn stage becomes widened and the total amount of remaining hydrogen increases with decreasing metallicity . the ratio of sn ib / c to sn iib rate should decrease with decreasing metallicity ( fig . [ fig : fate ] ) . from the above discussion , we can make a crude prediction on the sn types from case b / bb mass systems as summarized in fig . [ fig : fate ] . with case ab / abb systems , each boundary in the figure would simply shift to a higher @xmath41 . for example , the lower limit of @xmath41 for sn ib / c from case b / bb systems is about 12.5 @xmath4 , while it is about 16 @xmath4 for case ab / abb systems . because we still do not have any clear demarcation between sn ib and sn ic in terms of the progenitor structure ( sect . [ sect : binaryhe ] ) , ad - hoc assumptions of @xmath69 @xmath4 ( case i ) and @xmath70 ( case ii ) were made for the upper and lower panels , respectively . several interesting predictions can be made from this figure , which should be tested in future observations . these predictions are only relevant for ordinary sne ib / c and those associated with grbs or slsne - ic are not considered here . * for case i , the ratio of sn ic to sn ib rate from binary systems would decrease with decreasing metallicity in good agreement with @xcite and @xcite , and sn ic is hardly expected at sufficiently low metallicity . by contrast , this ratio would not necessarily decrease for decreasing metallicity with case ii , unless there existed a mass cut of @xmath41 for bh formation ( @xmath71 ) . if sn ib / c progenitors with a sufficiently high final mass could not produce an ordinary sn ib / c , the sn ic rate would significantly decrease with decreasing metallicity even with case ii . * for case ii , sne ic with hydrogen or sne iic are expected to occur at sufficiently low metallicity ( cf . @xcite indeed showed that sne iic can be produced if helium is effectively shielded from radioactive nickel . * the sn ib / sn iib ratio would decrease with decreasing metallicity for both cases , which is in good agreement with the recent observation by @xcite . * comparison of fig . [ fig : finalmass2 ] and fig . [ fig : fate ] implies that both the initial and final masses of sne ic progenitors would be systematically higher than those of sn ib progenitors , regardless of the adopted demarcation criterion of helium . this prediction is in qualitative agreement with sn observations @xcite and the stronger association of sne ic with younger stellar population than sne ib in the host galaxies ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . * for case ii , the average ejecta masses would increase following the order of sn ib , iib and ic at sufficiently low metallicity . * the average ejecta masses of sn ib / c would increase with decreasing metallicity . this effect would be more dramatic for sne ic than for sne ib ( see fig . [ fig : finalmass2 ] ) . * the comparison of fig . [ fig : finalmass2 ] and fig . [ fig : fate ] indicates that for case ii , the maximum final mass of sn ib ( and iib ) would be limited to a fairly small value ( about 7.0 @xmath4 with the assumed value of @xmath70 ) even for very low metallicity . as explained in sect . [ sect : binaryexample ] , a relatively low - mass helium star progenitor of sn ib / c in binary systems undergo rapid expansion of its envelope during the carbon burning phase and later stages . this envelope expansion becomes stronger for a lower mass star for which the carbon - oxygen core becomes more compact , following the mirror effect @xcite . within the framework discussed in sect . [ sect : fate ] , therefore , sne ib progenitors would have a more extended envelope than sn ic progenitors for most cases , which may in turn lead a more luminous early plateau in the consequent sn @xcite . the extended envelope at the pre - sn stage can make a sn ib / c progenitor fairly bright in the optical . the expected visual magnitude is about @xmath72 for the progenitors having final masses of 3 5 @xmath4 ( @xcite ) . with an o - type star companion , the visual brightness would be even higher ( @xmath73 ) . by contrast , wr star progenitors ( @xmath27 @xmath4 ) from both single and binary stars have much higher bolometric luminosities , but the expected high surface temperatures at the pre - sn stage result in fainter visual brightness ( i.e. , @xmath74 * ? ? ? * ; * ? ? ? * ) in most cases , which would make them more difficult to directly identify in pre - sn images as discussed in sect . [ sect : singlesurface ] . recently , @xcite have reported the tentative identification of the progenitor of the sn ib iptf13bvn . the estimated absolute magnitudes of the object in the optical range from -5.0 to -7 , depending on the filters , adopted extinction values , and photometry methods @xcite . @xcite argued for a single star progenitor with initial masses of 31 - 35 @xmath4 based on the non - rotating models of the geneva group . unlike more massive stars that become wo stars at the end , these models end their lives as wn stars having relatively thick helium envelopes with surface temperatures of about @xmath75 . the predicted optical magnitudes ( @xmath76 ) agree with the observation , but the final mass ( @xmath77 @xmath4 ) seems to be too high , compared to the estimated ejecta mass of the sn ( @xmath78 @xmath4 ; @xcite ) that supports the binary scenario . the optical brightness of the progenitor candidate can also be explained by a binary progenitor having an initial mass of 10 - 20 @xmath4 @xcite . if future observations find evidence for the companion star that survives the sn explosion , it will directly confirm a binary star origin of sne ib for the first time . massive stars in a close binary system before mass transfer are likely to be synchronized with the orbit because of the short tidal interaction timescale . after case b / ab mass transfer , the orbit becomes too wide to keep the tidal synchronization , and the final rotation velocity of the primary star is mainly determined by the amount of angular momentum that is retained after the mass transfer phase . binary evolution models including the effects of rotation indicate that sn ib / c progenitors via case b / ab mass transfers are slow rotators as they lose mass and angular momentum @xcite . both models with and without magnetic torques due to the tayler - spruit dynamo ( see sect . [ sect : singlerot ] for the discussion on angular momentum redistribution inside stars ) predict that the surface rotation velocity of naked helium stars on the helium main sequence after the mass transfer phase is only about @xmath79 . the specific angular momentum in the innermost region of 1.4 @xmath4 that would collapse to a neutron star significantly according to the prescription of angular momentum transfer @xcite . models without the tayler - spruit dynamo predict that neutron stars from sne ib / c in binary systems would almost reach the critical rotation . this means that the majority of sn ib / c progenitors in binary systems would also be good progenitor candidates for magnetars / collapsars without the talyer - spruit dynamo , which can not be easily reconciled with observations . models with the tayler - spruit dynamo predict that the specific angular momentum of the collapsing core would be comparable to those of single star models ( i.e. , @xmath80 ) and the resultant neutron stars would be rotating at a period of several milliseconds . as in the case of single stars , therefore , binary models with the tayler - spruit dynamo do not predict magentar / collapsar progenitors for energetic sne and/or long grbs via the standard case b / bb / ab / abb systems @xcite under normal circumstances . at sufficiently low metallicity , however , mass accreting stars in case b mass transfer systems may be spun up to undergo the chemically homogeneous evolution , which may end up as a grb @xcite . sn ib / c progenitors in very close binary systems consisting of a helium star and a low - mass main sequence star / compact object ( cf . fig . [ fig : binary ] ; sect . [ sect : binaryexample ] ) may experience strong tidal interaction , given that the orbital period can be as short as 0.1 day in this case ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . some of these systems may produce unusually rapid rotators even within the framework of the tayler - spruit dynamo , but more detailed evolutionary studies are needed to make a meaningful conclusion on this . the fact that host galaxies of broad - lined sne ic and grbs are found to be systematically overdense compared to other sdss galaxies might be an indication for the importance of the binary channel for producing rapidly rotating progenitors to produce these events @xcite . wr stars in binary systems are excellent observational counterparts for sn ib / c progenitors during the post case b / ab mass transfer phase , for @xmath81 @xcite . relatively low - mass helium stars ( @xmath82 ) in binary systems are only rarely observed . this is probably because such helium stars on the helium main sequence are very hot and faint in optical bands ( see fig . [ fig : hr2 ] ) and because many of them have bright ob - type companion stars . as mentioned above ( sect . [ sect : binarymass ] ) , the best observational counterpart of binary sn ib / c progenitors with @xmath29 is the quasi - wr ( qwr ) star hd 45166 @xcite : the primary is a helium rich 4.2 @xmath4 star with @xmath83 and the secondary is a 4.8 @xmath4 main sequence star , in a 1.596 day orbit @xcite . this system was probably produced via unstable case b / ab mass transfer and the consequent common envelope phase . this star gives important information about the mass loss rate from such a relatively low - mass helium star as discussed in sect . [ sect : binarymass ] . wr 7a is another qwr star @xcite . no companion of this star has been found so far . radial periodicity of 0.204 day has been reported , and if this is related to binarity , its companion should be a low - mass main sequence star ( @xmath84 @xmath4 ) or a compact object . to our knowledge , no estimate of the wind mass loss rate from this star has been reported yet . there also exist candidates for evolved helium giant stars beyond core helium exhaustion . they include @xmath85 sgr , ks per and lss 4300 @xcite . among these , the @xmath85 sgr system has been best studied ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? the mass , surface temperature and bolometric luminosity of the hydrogen - deficient primary of this system are @xmath86 @xmath4 , @xmath87 k and @xmath88 . it also shows radial pulsation of a 20 day period @xcite , and evidence for mass transfer and a circumbinary disk @xcite , which agrees well with the case bb mass transfer scenario @xcite . therefore , these systems can provide useful information about the progenitor evolution as well as their circumstellar environments immediately before sn explosion . the visual magnitude of this star is @xmath89 @xcite , which is consistent with the model predictions discussed in sect . [ sect : binarysurface ] . the companion stars of binary progenitors of sne ib / c will survive the sn explosion and may be found in some young sn remnants ( e.g. * ? ? ? * ; * ? ? ? * ) . there may be several types of companion stars : main sequence stars of early to late types , compact objects like white dwarfs , neutron stars and black holes , and helium stars ( sect . [ sect : binaryevol ] ; fig . [ fig : binary ] ) . in terms of stellar population , the most common type ( i.e. , more than 30% of all binary sn ib / c progenitors ) may be relatively high mass stars ( o / b type ) on the main sequence that underwent stable case b / bb / ab / abb mass transfer ( e.g. , * ? ? ? * ; * ? ? ? these massive companion stars accrete mass and angular momentum via stable mass transfer to be spun up to the critical value @xcite . at solar metallicity , however , they will lose angular momentum again via stellar winds after the mass transfer phase , and the rotation velocity shortly after the sn explosion depends on how much mass is lost by winds until that time . models by @xcite indicate that the surface rotation velocity will be about 300 - 450 @xmath90 for surviving companion of @xmath91 @xmath4 . by contrast , more massive stars loses mass and angular momentum very quickly : a 48 @xmath4 companion star would be slowed down to about 60 @xmath90 at the pre - sn stage . the transferred mass to the secondary star is enriched with the ashes of hydrogen burning compared to the initial composition @xcite . the surface composition of the surviving secondary star after the sn explosion should be therefore marked by the enhancement of helium and nitrogen . the models of @xcite give typically the mass fractions of helium and nitrogen of about 0.35 and @xmath92 ( i.e. , 4.3 times the solar value ) at the surface of surviving companion stars , respectively . > x| > x & single & binary + initial mass & @xmath50 @xmath4 at @xmath11 & @xmath93 @xmath4 + final mass & @xmath94 at @xmath11 & @xmath95 at @xmath11 + final radius & @xmath96 & @xmath97 + wind mass loss rate & @xmath98 & @xmath99 + escape velocity ( @xmath100 ) & @xmath101 & @xmath102 + circumstellar structure & @xmath103 & complex with wind - wind collision and orbital motion + optical magnitudes & @xmath104 for wo type progenitor , and @xmath105 for wn type progenitor . & @xmath106 for most helium star progenitors . it will be more luminous in optical bands with a bright companion . + light curves & relatively faint at early times . broad light curves & fairly luminous early time plateau for @xmath107 @xmath4 + spectra & no or weak helium lines for most cases , and hence biased towards the production of sne ic . & helium lines even without non - thermal processes during the early time plateau phase for @xmath108 @xmath4 + many of wr progenitors would end their lives as a wo type star that is relatively faint with optical filters ( @xmath104 ; @xcite ) . binary systems with a sufficiently high initial mass of the primary star can produce wr progenitors of which the properties would be very similar to those of single wr stars , but the presence of the companion star would result in rather complex structures of the circumstellar medium and the sn remnant @xcite . compared to the single star case , the detectability of wr progenitors can be significantly enhanced with a luminous early type companion . however , given the preference for lower masses of the initial mass function , the majority of binary progenitors should have several unique properties that are very different from wr progenitors . their final masses at the pre - sn stage are systematically lower ( @xmath109 for @xmath110 ) than those of wr progenitors ( @xmath7 @xmath4 ) . this agrees well with the recent ejecta mass estimates of ordinary sne ib / c ( @xmath111 @xmath4 ; @xcite ) . the current binary model predictions are also consistent with the observational facts that the association of the sn locations in their host galaxies with young stellar populations becomes stronger following the order of sn ii , sn ib and sn ic ( sects . [ sect : binarymass ] and [ sect : fate ] ) , which is usually interpreted within the single star scenario in the literature ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . despite their relatively low masses , the detectability of binary progenitors in optical bands is not necessarily lower than single wr star progenitors . to the contrary , a significant fraction of binary progenitors should have very high visual luminosities because relatively low - mass helium stars can rapidly expand during the late evolutionary stages , and/or because many of them would have luminous companion stars @xcite . the sn models by @xcite indicate that early time light curves and spectra would have the critical information about the nature of sn ib / c progenitors . in particular , the plateau phase due to helium recombination at early times and he i lines formed with thermal processes during this phase would be strong evidence for a relatively low - mass helium star progenitor having an extended helium envelope . as demonstrated by several recent observations(e.g . ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , the progenitor size can also be best constrained by early - time sn light curves including shock breakouts . we therefore conclude that observations of early - time light curves and spectra will be an excellent probe into the nature of sn ib / c progenitors . * a caveat in the above conclusions is that the predictions summarized in table [ tab1 ] are mostly based on the models at solar metallicity . the final masses of single star progenitors at super - solar metallicity can become as low as 5 @xmath4 , depending on the adopted mass loss rate ( fig . [ fig : finalmass ] ; * ? ? ? * ) . this value is within the ejecta mass range of ordinary sne ib / c by @xcite and @xcite , although it still can not explain sne ib / c having ejecta masses lower than about 3.5 @xmath4 . given that the final amounts of helium must also be smaller than in the case of solar metallicity ( fig . [ fig : hemass ] ) , the contribution from super - solar metallicity single stars might be particularly significant for sne ic ( cf . * ; * ? ? ? * ; * ? ? ? although the sn ib / c event rate implies the dominant role of binary systems for the production of sne ib / c @xcite , more direct evidence for binary progenitors comes from sne ib / c ejecta masses ( @xmath112 ; sects . [ sect : binarymass ] and [ sect : fate ] ) . precise estimates of sn ejecta masses can give one of the best constraints for progenitor models . the uncertainties related to the effects of asymmetry of the explosion , mixing of chemical compositions in the sn ejecta and the presence of helium on the estimates of @xmath112 using sn light curves should be clarified in the near future ( cf . * ; * ? ? ? * ; * ? ? ? * the question on how much helium can be hidden in sn ic spectra is another critical test case for sn ib / c progenitor models ( sects . [ sect : singlehe ] , [ sect : binaryhe ] and [ sect : fate ] ; fig . [ fig : fate ] ) . recent observations indicate very low helium mass ( @xmath113 ) in the ejecta @xcite . this can not be easily accommodated to the current model predictions that most sn ib / c progenitors have @xmath114 . this conflict may be due to our lack of proper knowledge on the mass loss rate from helium stars ( in particular during the post - helium burning phase ; sect . [ sect : binaryhe ] ) but a rigorous estimate of helium masses for a large sample of sne ib / c is highly required to resolve this issue . the ratio of sn ic to sn ib is another important constraint for progenitor models , which should be better estimated in the future . several authors reported that the ic to ib ratio is about 2 ( e.g. , * ? ? ? * ; * ? ? ? * ) , but @xcite point out that many sne ib had been misclassified as sne ic , suggesting a lower value . * as discussed in sect . [ sect : fate ] , future observations on sn ib / c and sn iib properties and their relative frequencies as a function of metallicity ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) would greatly help to clarify the role of mass loss from helium stars in the evolution of sn ib / c progenitors . it would be particularly important to investigate how the ejecta masses of sne ib / c and the ratio of sn ib / sn ic rate systematically depend on metallicity . * in the discussions above , we did not address progenitors of broad - lined sne ic ( sne ic - bl ) , simply because we do not have a good clue on what makes them . the association of long grbs and some sne ic - bl ( e.g. * ? ? ? * ) implies the importance of rapid rotation , and attempts have been made to explain long grbs , sne ic - bl and super - luminous sne ic within the single framework of the magnetar scenario ( e.g. , * ? ? ? binary star models including rotation do not predict any particular parameter space where unusually rapid rotation in the core can be realized from the standard case b / bb / ab / abb channels for sne ib / c @xcite , except for the so - called chemically homogeneous evolution induced by mass accretion at low metallicity @xcite . future studies should investigate more carefully the evolution of some specific binary systems where the condition of rapid rotation for magnetars can be rather easily fulfilled compared to the standard channels , such as binary systems consisting of a helium star plus a compact object in a very tight orbit ( fig . [ fig : binary ] ; sect . [ sect : binaryrot ] ) . the author is grateful to norbert langer , stan woosley , luc dessart , hyun - jeong kim , bon - chul koo , gtz grfener , jorick vink , wonseok chun , thomas tauris , rob izzard , alexandra kozyreva , and john hillier for their contribution to this work and fruitful collaboration , and john eldridge , maryam modjaz , philipp podsiadlowski , kenichi nomoto , keiichi maeda , selma de mink and cyril georgy for helpful discussions . this work was supported by basic science research ( 2013r1a1a2061842 ) program through the national research foundation of korea ( nrf ) and by research resettlement fund for the new faculty of snu . anderson , j.p . , habergham , s.m . , james , p.a . , & hamuy , m. 2012 , mnras , 424 , 1372 arcavi , i. et al . 2010 , apj , 721 , 777 barkat , z. , rakavy , g. , & sack , n. 1967 , phrvl , 18 , 379 bersten , m. , benvenuto , o.m . , & folatelli , g. et al . 2014 , apj , 148 , 68 boissier , s. , & prantozs , n. 2009 , a&a , 503 , 137 branch , d. , et al . 2002 , apj , 566 , 1005 braun , h. 1997 , phd thesis , ludwig - maximilians university of munich burrows , a. , dessart , l. , livne , e. , ott , c.d . , & murphy , j. 2007 , apj , 664 , 416 cao , y. , kasliwal , m. m. , & arcavi , i. et al . 2013 , apj , 775 , l7 cano , z. 2013 , mnras , 434 , 1098 cano , z. , maeda , k. , & schulze , s. 2014 , mnras , 785 , 95 cantiello , m. , yoon , s .- c . , langer , n. , & livio , m. 2007 , a&a , 465 , 29 cantiello , m. , mankovich , c. , bildsten , l. , christensen - dalsgaard , j. , & paxton , b. 2014 , apj , 788 , 93 chevalier , r.a . , & soderberg , a. , 2010 , apj , 711 , l40 chiosi , c. & maeder , a. 1986 , ara&a , 24 , 328 chomiuk , l. et al . 2011 , apj , 743 , 114 chornock , r. et al . 2011 , apj , 739 , 41 claeys , j.s.w . , de mink , s > e . , pols , o.r . , eldridge , j.j . , & baes , m. 2011 , a&a , 528 , a131 conti , p.s . 1976 , in proc . 20th colloq . lige , p. 193 , univ . lige corsi , a. , ofek , e.o . , & gal - yam , a. , et al . 2012 , apj , 747 , 5 crockett , r.m . , smartt , s.j . , & eldridge , j.j . 2007 , mnras , 381 , 835 crowther , p.a . 2007 , ara&a , 45 , 177 de jager , c. , nieuwenhuijzen , h. , & van der hucht , k.a . 1988 , a&as , 72 , 259 de mink , s.e . , pols , o.r . , langer , n. , izzard , r.g . 2009 , a&a , 507 , 169 de mink , s.e . , langer , n. , izzard , r.g . , sana , h. , & de koter , a. 2013 , apj , 764 , 166 de mink , s.e . , sana , h. , langer , n. , izzard , r.g . , & schneider , f.r.n . 2014 , apj , 782 , 7 deng , j.s . , qiu , y.l . , hu , j.y . , hatano , k. , & branch , d. 2000 , apj , 540 , 452 dessart , l. , hillier , d.j . , livne , e. , yoon , s .- c . , woosley , s.e . , waldman , r. , & langer , n. 2011 , mnras , 414 , 2985 dessart , l. , hillier , d. , li , c. , & woosley , s.e . 2012a , mnras , 424 , 2139 dessart , l. , hillier , d. , waldman , r. , livne , e. , & blondin , s. 2012b , mnras , 426 , l76 detmers , r.g . , langer , n. , podsiadlowski , ph . , & izzard , r.g . 2008 , a&a , 484 , 831 dewi , j.d.m . , pols , o.r . , savonije , g.j . , & van den heuvel , e.p.j . 2002 , mnras , 331 , 1027 drout , m.r . 2011 , apj , 741 , 97 dudley , r.e . , & jeffery , c.s . , 1993 , mnras , 262 , 945 eggenberger , p. , maeder , a. , & meynet , g. 2005 , a&a , 440 , l9 eggenberger , p. , montalbn , j. , & miglio , a. 2012 , a&a , 544 , l4 ekstrm , s. , georgy , c. , eggenberger , p. et al . 2012 , a&a , 537 , a146 eldridge , j.j . , fraser , m. , mound , j.r . , & smartt , s.j . 2015 , mnras , 446 , 2689 eldridge , j.j . , fraser , m. , smartt , s.j . , mound , j.r . , & crockett , r. m. 2013 , mnras , 436 , 774 eldridge , j.j . , izzard , r.g . , tout , c.a . 2008 , mnras , 384 , 1109 eldridge , j.j . , langer , n. , & tout , c.a . 2011 , mnras , 414 , 3501 eldridge , j.j . , & vink , j.s . 2006 , a&a , 452 , 295 elmhamdi , a. , danziger , i.j . , branch , d. , leibundgut , b. , baron , e. , & kirchner , r.p . 2006 , a&a , 450 , 305 filippenko , a.v . 1997 , ara&a , 35 , 309 foley , r.j . , smith , n. , ganeshalingam , m. , li , w. , chornock , r. , & filippenko , a.v . 2007 , apj , 657 , 105 frame , d.j . , cottrell , p.l . , gilmore , a.c . , kilmartin , p.m. , & lawson , w.a . 1995 , mnras , 276 , 383 fremling , c. , sollerman , j. , & taddia , f. , et al . 2014 , a&a , 565 , 114 fryer , c.l . , & heger , a. 2005 , apj , 623 , 302 fujimoto , m.y . 1993 , apj , 419 , 768 gal - yam , a. 2009 , nature , 462 , 624 gal - yam , a. , arcavi , i. , & ofek , e.o . , et al . 2014 , nature , 509 , 471 gellert , m. , rdiger , g. , & elstner , d. 2008 , a&a , 479 , l33 georgy , c. , ekstrm , s. , & meynet , g. et al . 2012 , a&a , 542 , a29 gorbikov . 2014 , mnras , 443 , 671 grfener , g. , & hamann , w .- r . 2008 , a&a , 482 , 945 grfener , g. , owocki , s.p . , & vink , j.s . 2012 , a&a , 538 , a40 grfener , g. , & vink , j.s . 2013 , a&a , 560 , a6 graham , j.f . , & fruchter , a.s . 2013 , apj , 774 , 119 groh , j.h . , oliveira , a.s . , & steiner , j.e . 2008 , a&a , 485 , 245 groh , j . h. , meynet , g. , georgy , c. , & ekstroem , s. 2013a , a&a , 558 , a131 groh , j . h. , georgy , c. , & ekstroem , s. 2013b , a&a , 558 , l1 hakobyan , a.a . , mamon , g.a . , petrosian , a.r . , kunth , d. , & turatto , m. 2009 , a&a , 508 , 1259 hamann , w .- r . , schoenberner , d. , & heber , u. 1982 , a&a , 116 , 273 hamann , w .- r . , koesterke , l. , wessolowski , u. 1995 , a&a , 299 , 151 hamann , w .- r . , grfener , g. , & liermann , a. 2006 , a&a , 457 , 1015 hachinger , s. , mazzaili , p.a . , taubenberger , s. , hillebrandt , w. , nomoto , k. , & sauer , d.n . 2012 , mnras , 422 , 70 heger , a. , langer , n. , & woosley , s.e . 2000 , apj , 528 , 368 heger , a. , fryer , c.l . , woosley , s.e . , langer , n. , & hartmann , d.h . 2004 , apj , 591 , 288 heger , a. , woosley , s.e . , & spruit , h.c . 2005 , apj , 626 , 350 hirschi , r. , meynet , g. , & maeder , a. 2004 , a&a , 425 , 649 hjorth , j. 2013 , rspta , 371 , 20120275 ibeling , d. , & heger , a. 2013 , apj , 765 , 43 inserra , c. et al . 2013 , apj , 770 , 128 ishii , m. , ueno , m. , & kato , m. 1999 , pasj , 51 , 417 izzard , r.g . , ramirez - ruiz , e. , & tout , c.a . 2004 , mnras , 348 , 1215 ivanova , n. , belczynski , k. , kalogera , v. , rasio , f.a . , & taam , r.e . 2003 , apj , 592 , 475 jeffery , c.s . , & hamann , w .- r . 2010 , mnras , 404 , 1698 justham , s. , podsiadlowski , ph . , vink , j.s . 2014 , apj , 796 , 121 kasen , d. , & bildsten , l. 2010 , apj , 717 , 245 kelly , p.l . , & kirshnir , r.p . 2012 , apj , 759 , 107 kelly , p.l . , filippenko , a.v . , modjaz , m. , & kocevski , d. 2014 , apj , 789 , 23 kim , h .- j . , yoon , s .- c . , & koo , b .- 2015 , apj , submitted kippenhahn , r. , & weigert , a. 1967 , z. astrophys . 65 , 251 kipper , t. , & klochkova , v.g . , 2012 , baltic astronomy , 21 , 219 kobulnicky , h.a . , & fryer , c.l . 2007 , apj , 670 , 747 kochanek , c. 2009 , apj , 707 , 1578 koo , b.c . , mckee , c.f . , & suh , k.w . 2011 , apj , 732 , 6 koubsky , p. , harmanec , p. , yang , s. , netolick , m. , et al . 2006 , a&a , 459 , 849 kozyreva , a. , blinnikov , s. , langer , n. & yoon , s .- c . 2014 , a&a , 565 , 70 langer , n. 1989a , a&a , 210 , 93 langer , n. 1989b , a&a , 220 , 135 langer , n. 1992 , a&a , 265 , 17 langer , n. 2012 , ara&a , 50 , 107 lauterborn , d. 1970 , a&a , 7 , 150 li , w. , leaman , j. , & chornock , r. et al . 2011 , mnras , 412 , 1441 lucy , l.b . 1991 , apj , 383 , 308 lunnan , r. et al . 2014 , apj , 787 , 138 lyman , j. , bersier , d. , james , p. et al . 2014 , arxive:1406.3667 macfadyen , a. , & woosley , s. 1999 , apj , 524 , 262 maeda , k. , katsuda , s. , bamba , a. , terada , y. , & fukazawa , y. 2014 , apj , 785 , 95 maeder , a. 1987 , a&a , 178 , 159 maeder , a. , & meynet , g. 1987 , a&a , 182 , 243 maeder , a. , & meynet , g. 2000 , ara&a , 38 , 143 maeder , a. , georgy , c. , meynet , g. , & ekstrm , s. 2012 , a&a , 539 , a110 massey , ph . 2003 , ara&a , 41 , 15 maund , j.r . , & smartt , s.j . 2005 , mnras , 360 , 288 maund , j.r . , smartt , s.j . , kudritzki , r.p . , podsiadlowski , ph . , & gilmore , g.f . , 2004 , nature , 427 , 129 maund , j.r . , smartt , s.j . , & schweizer , f. 2005 , apj , 630 , 33 mazzali , p.a . , rpke , f.k . , benetti , s. , & hillebrandt , w. 2007 , science , 315 , 825 mazzali , p.a . , mcfadyen , a.i . , woosley , s.e . , pian , e. , & tanaka , m. 2014 , mnras , 443 , 67 meynet , g. , & maeder , a. 1997 , a&a , 321 , 465 meynet , g. , & maeder , a. 2003 , a&a , 404 , 975 meynet , g. , & maeder , a. 2005 , a&a , 429 , 581 meynet , g. , maeder , a. , schaller , g. , schaerer , d. , & charbonnel , c. 1994 , a&as , 103 , 97 modjaz , m. , blondin , s. , & kirshner , r.p . 2014 , apj , 147 , 99 modjaz , m , kewley , l. , bloom , j.s . , filippenko , a.v . , perley , d. , & silverman , j.m . 2011 , apj , 731 , 4 modjaz , m , kewley , l. , & kirshner et al . 2008 , aj , 135 , 1136 modjaz , m , li , w. , & butler , n. et al . 2009 , apj , 702 , 226 netolick , m. , bonneau , d. , cehsneau , o. et al . 2009 , a&a , 499 , 827 nicholl , m. et al . 2013 , nature , 502 , 346 nomoto , k. , yamaoka , h. , pols , o.r . , van den heuvel , e.p.j . , iwamoto , k. , kumagai , s. , & shigeyama , t. 1994 , nature , 371 , 227 nugis , t. , & lamers , h.j.g.l.m . 2000 , a&a , 360 , 227 oliveira , a.s . , steiner , j.e . , & cieslinski , d. , 2003 , mnras , 346 , 963 pereira , c.b . 1998 , a&a , 338 , l91 petrovic , j. , langer , van der hucht , k.a . 2005a , 435 , 1013 petrovic , j. , langer , yoon , s .- c . , hger , a. 2005b , 435 , 247 petrovic , j. , pols , o. , & langer , n. 2006 , a&a , 450 , 219 piro , a. , & morozova , v.s . 2014 , apj , 792 , l11 podsiadlowski , ph . , joss , p.c . , & hsu , j.j.l . 1992 , apj , 391 , 246 podsiadlowski , ph . , hsu , j.j.l . , joss , p.c . , & ross , r.r . 1993 , nature , 364 , 509 podsiadlowski , ph . , ivanova , n. , justham , s. , rappaport , s. 2010 , mnras , 406 , 840 pols , o.r . 1994 , a&a . 290 , 119 pols , o.r . , & dewi , j.d.m . 2002 , pasa , 19 , 233 prieto , j.l . , stanek , k.z . , & beacom , j.f . 2008 , apj , 673 , 999 saio , h. 1995 , mnras , 277 , 1393 salasnich , b. , bressan , a. , & chiosi , c. 1999 , a&a , 342 , 131 sana , h. et al . 2012 , science , 337 , 444 sander , a. , hamann , w .- r . , & todt , h. 2012 , a&a , 540 , a144 sanders , n.e . , soderberg , a.m. , & levesque , e.m . 2012 , apj , 758 , 132 scalzo , r.a . , ruiter , a.j . , & sim , s.a . 2014 , mnras , 445 , 2535 schaeffer , r. , casse , m. , & cahen , s. 1987 , apjl , 316 , 31 schaerer , d. , charbonnel , c. , meynet , g. , maeder , a. , & schaller , g. 1993a , a&as,102 , 339 schaerer , d. , meynet , g. , maeder , a. , & schaller , g. 1993b , a&as , 98 , 523 schaller , g. , schaerer , d. , meynet , g. , & maeder , a. 1992 , a&as , 96 , 269 schoenberner , d. , & drilling , j.s . 1983 , apj , 268 , 225 schwartz , r.d . , persson , s.e . , & hamann , f.w . , 1990 , aj , 100 , 793 smartt , s.j . 2009 , ara&a , 47 , 63 smith , n. , li , w. , filippenko , a.v . , & chornock , r. 2011 , mnras , 412 , 1522 smith , n. 2014 , ara&a , 53 , 487 soderberg , a.m. , et al . 2008 , nature , 453 , 469 spencer , j. , baron , e. 2010 , apj , 718 , 957 spruit , h.c . 2002 , a&a , 381 , 923 steiner , j.e . , & oliveira , a.s . 2005 , a&a , 444 , 895 stritzinger , m. , leibundgut , b. , walch , s. , & contardo , g. 2006 , a&a , 450 , 241 suijs , m.p.l . , langer , n. , poelarends , a.j . 2008 , a&a , 481 , 87 taddia , f. , sollerman , j. , & leloudas , g. et al . 2014 , arxiv:1408.4084 tauris , t.m . , langer , n. , moriya , t.j . , podsiadlowski , ph . , yoon , s .- c . , blinnikov , s.i . 2013 , apjl , 778 , l23 townsend , r. , & macdonald , j. 2008 , in massive stars as cosmic engines , iau symp . 250 , p. 161 vanbeveren , d. , de loore , c. , & van rensbergern , w. 1998 , a&ar , 9 , 63 vanbeveren , d. , van bever , j. , & belkus , h. 2007 , apj , 662 , 107 van blerkom , d. , 1978 , apj , 225 , 175 van den berg , s. , li , w. , filippenko , a.v . 2005 , pasp , 117 , 773 van den heuvel , e.p.j . , & yoon , s .- c . 2007 , ap&ss , 311 , 177 van der hucht , k.a . 2001 , new . 45 , 135 van dyk , s.d . , zheng , w. , & fox , o.d . , 2014 , apj , 147 , 37 van loon , j.th . , cioni , m .- r.l . , zijlstra , a.a . & loup , c. 2005 , a&a , 438 , 273 van loon , j.th . 2008 , a&a , 487 , 1055 vink , j.s . , & de koter , a. 2005 , a&a , 442 , 587 wellons , s. , soderberg , a.m. , & chevalier , r.a . 2012 , apj , 752 , 17 wellstein , s. , & langer , n. 1999 , a&a , 350 , 148 wellstein , s. 2001 , phd thesis , university of potsdam wellstein , s. , langer , n. , & braun , h. 2001 , a&a , 369 , 939 wheeler , j.c . , yi , i. , hflich , p. , & wang , l. 2000 , apj , 537 , 810 wheeler , j.c . , johnson , v. , & clocchiatti , a. 2014 , arxiv:1411.5975 willis , a.j . , & stickland , d.j . , 1983 , mnras , 203 , 619 woosley , s.e . 1993 , apj , 405 , 273 woosley , s.e . 2010 , apj , 719 , 204 woosley , s.e . , & bloom , j. s. 2006 , ara&a , 44 , 507 woosley , s.e . , & eastman , r.g . 1997 , in thermonuclear supernovae , nato asi proc . sereis c. , vol . 486 , p. 821 , ruiz - lapuente , p. , canal , r. , isern , j. , eds . , kluwer , dortrecht woosley , s.e . , langer , n. , & weaver , t.a . 1993 , apj , 411 , 823 woosley , s.e . , langer , n. , & weaver , t.a . 1995 , apj , 448 , 315 woosley , s.e . , & heger , a. 2006 , apj , 637 , 914 yoon , s .- c . , & langer , n. 2005 , a&a , 443 , 643 yoon , s .- c . , langer , n. , norman , c. 2006 , a&a , 460 , 199 yoon , s .- c . , & cantiello , m. 2010 , apjl , 717 , l62 yoon , s .- c . , woosley , s.e . , & langer , n. 2010 , apj , 725 , 940 yoon , s .- c . , dierks , a. , & langer , n. 2012a , a&a , 542 , a113 yoon , s .- c . , grfener , g. , vink , j.s . , kozyreva , a. , & izzard , r.g . 2012b , a&a , 544 , l11 zahn , j .- p . , brun , a.s . , & mathis , s. 2007 , a&a , 474 , 145

it is well known that the fiber optical parametric amplifiers ( fopas ) , employing @xmath0 nonlinearity based four wave mixing ( fwm ) to transfer energy from one or two strong pump fields to a weak signal field , can be used in the conventional optical communication systems for signal - processing applications , such as optical amplification , phase conjugation , optical limiters , 3r generators , and time - domain de - multiplexer etc . in fact , similar to the @xmath1 crystals based optical parametric amplifiers ( opas ) @xcite , fopas are also candidates for generating non - classical light for quantum communication @xcite . in quantum communication , information can be encoded discretely on each photon or continuously in quadrature - phase amplitudes of an optical field . the former requires single - photon sources that can be generated with weak nonlinear interaction @xcite , but the latter relies on strong interaction to generate quantum correlation between the amplitudes of optical fields for entanglement @xcite . moreover , continuous variable ( cv ) entanglement offers several advantages over its discrete variable counterpart . the main advantage is the experimental feasibility in unconditional production of entangled states which in turn enables quantum information protocols to be accomplished unconditionally @xcite . the past decade has seen the growing interest in developing nonclassical light sources via the fwm in fibers , because they are compatible with the fiber network @xcite . in order to enhance the fwm and to suppress the raman scattering @xcite , a pulse pumped fopa is preferred because the peak power of pulsed laser is very high even at a modest average power . however , for the majority of the experiments performed to date , the fwm in fiber is in the low gain regime , and the fiber based nonclassical light sources with vacuum injection are in the domain of a few photons and thus can be only employed for discrete variable information encoding . in order to develop the fiber based source of the cv nonclassical light , the pulse pumped fopa needs to be operated in the high gain regime . up to now , only a couple of experiments reported the generation of cv nonclassical light via the pulse pumped fwm in fibers @xcite . in 2001 , sharping and co - works realized the pulsed twin beams by using the fopa for the first time @xcite . the amplified signal and generated idler beams bear a strong quantum correlation in intensity ( photon - number ) fluctuations in the sense that the quantum noise level of their intensity difference is lower than the shot - noise limit ( snl ) . in that experiment , the gain of fwm is less than 3 and the observed intensity difference noise of twin beams is only about 1.1 db ( 2.6 db after correction for losses ) lower than the snl . recently , using the fopa with photon number gain of about 16 db , our group demonstrated the twin beams with the intensity - difference noise below the snl by 3.1 db ( 10.4 db after correction for losses ) @xcite . the results indicate that compared with its @xmath1 crystal counterparts @xcite , the pulse pumped high gain fopa is a simple and promising system for generating the cv nonclassical light , which has the potential applications in realizing the quantum communication protocols and in performing the high precision measurement . there have been a few theoretical papers analyzing the generation of pulsed cv nonclassical light via the high gain @xmath1 crystals based opas @xcite . for the pulse pumped high gain fopas , however , the investigation is mostly within the domain of classical characteristics @xcite and the quantum theory has not been worked out yet . this is different from the case of continuous wave ( cw ) or quasi - cw pumped high gain fopas @xcite . in order to understand the high gain fopa for further reducing the intensity difference noise of pulsed twin beams , we focus on developing a multi - mode quantum theory in this paper . the rest of the paper is organized as follows . after briefly introducing the quantum description of a single - mode fopa in sec . ii , we deduce a multi - mode quantum theory of a pulse pumped non - degenerate fopa in sec . the evolution of non - degenerate signal and idler beams is described by the bogoliubov transformation in a multi - frequency mode form . according to the hamiltonian of fwm in fibers , we derive the four transformation functions of the bogoliubov transformation in the analytical form of infinite series . the calculation shows that the noise performance of the high gain fopa highly depends on the joint spectral function ( jsf ) . in sec . iv , we show that the fopa with factorized jsf can be described by a quantum model of single - mode fopa , which is similar to the model given in sec . ii , because both signal and idler beams can be in single temporal mode , respectively . the results illustrate the validity of the multi - mode theory . in sec . v , we focus on analyzing the fopa with a very broad gain bandwidth . in this situation , the jsf is non - factorable . the calculated results show that although the pulse pumped high gain fopa generally has a noise figure greater than the 3 db quantum limit , the intensity difference noise of the twin beams , weakly depends on the jsf , can be significantly reduced to lower than the snl . to closely resemble the real experimental situation , the dependence of the intensity difference noise of twin beams is numerically studied by considering the influences of experimental imperfections , including the quantum efficiencies of detectors , the collection efficiency of twin beams , and the excess noise of input signal . finally , we conclude in sec . [ conclusion ] . we start with the quantum description of single - mode linear fopas , whose pumps are single frequency lasers with negligible depletions . in this section , we will introduce the bogoliubov transformation in the single frequency mode form and the analytical expressions of key parameters of the fopa , including the photon number gain , the noise figure , and the intensity difference noise of the twin beams . figure [ fopa](a ) is a typical diagram of the non - degenerate phase - insensitive fopa , which is referred to as the fopa hereinafter for brevity . when the wavelengths of the pump are properly set , the phase matching of fwm in fiber is satisfied . in the fwm process , two pump photons at frequencies @xmath2 and @xmath3 are simultaneously scattered into a pair of signal and idler photons at frequencies @xmath4 and @xmath5 via the @xmath0 nonlinearity , and the energy conservation relation @xmath6 is satisfied . since each generated signal photon is always accompanied by the birth of an idler photon , the two form correlated twin beams . a weak input is injected into the fopa from either the signal or the idler field . for convenience , we refer the field with weak injection as the signal , while the other with vacuum input as idler . when the weak signal is amplified by the strong pump field via the fwm process , an idler beam is generated , and the frequency overlap between the amplified signal and generated idler fields is negligible for the non - degenerate fopa . ( a ) a typical diagram of the non - degenerate phase - insensitive fopa . @xmath7 and @xmath8 , filters ; d1 and d2 , detectors ; @xmath9 and @xmath10 , radio frequency amplifiers ; dds , differential detection system . ( b ) the sketch map of the frequency modes for the single frequency laser pumped fopa . two pump photons of a single frequency laser uniquely defines the single frequency modes of a pair of signal and idler photons . ( c ) the sketch map of the frequency modes for the pulse pumped fopa . a single frequency signal mode ( dotted arrow ) will be correlated to many idler frequency modes ( enveloping curve in gray ) , and vice versa . since the pump is in multi - frequency modes , the pulsed signal and idler beams are also in multi - frequency modes ( enveloping curves in black).,scaledwidth=65.0% ] when the fopa is pumped with a single frequency cw laser , as illustrated in fig . 1(b ) , the two pump photons have the same frequency mode ( @xmath11 ) , a pair of signal and idler photons in the frequency modes @xmath4 and @xmath5 are uniquely defined . in this case , the input - output relation of the single mode fopa is given by the well - known bogoliubov transformation : @xmath12 where @xmath13 and @xmath14 refer the input and output modes , the subscript @xmath15 respectively denote the signal and idler fields , the coefficients @xmath16 and @xmath17 are determined by the parametric gain of fwm and satisfy the relation @xmath18 . for input signal in a coherent state @xmath19 with photon number much greater than 1 , i.e. , @xmath20 it is straightforward to deduce the expressions of all the parameters of the fopa by calculating the quantum average of the corresponding operators over the state @xmath21 , where @xmath22 denotes the vacuum input of the idler field and the superscript @xmath23 denotes the single frequency mode case . the amplified signal ( generated idler ) beam is then measured by the detector d1(2 ) after passing through the filter @xmath24 to reject the strong pump ( fig . 1(a ) ) . the average photon number @xmath25 and the photon number fluctuation @xmath26 of the signal ( idler ) fields are converted to the dc and ac currents of d1 and d2 , respectively . for identical d1 and d2 , when the efficiencies of the detector d1(2 ) and filter @xmath24 are ideal , the average photon numbers of the amplified signal and generated idler beams can be written as [ sig_i ] @xmath27 @xmath28 where @xmath29 is the photon number operator of signal ( idler ) beam . in our theoretical model , we assume the condition @xmath30 is always satisfied , so the contribution of spontaneous terms @xmath31 in the right hand side of eqs . ( [ sig_i]a ) and ( [ sig_i]b ) are negligible . according to the definition of the photon number gain the ratio between the average photon numbers of amplified signal and input signal @xmath32 we have @xmath33 . moreover , according to the definition of the intensity noise of the amplified signal and generated idler fields , we have [ sig_delta_i ] @xmath34 @xmath35 after using the snl of signal ( idler ) beam @xmath36 to normalize the intensity noise @xmath37 , we obtain @xmath38 indicating the normalized intensity noise of the amplified signal beam @xmath39 is equal to that of the generated idler beam @xmath40 . notice that they are both above the snl because of amplification . as for the noise figure ( nf ) of the fopa , according to the definition @xcite @xmath41 with @xmath42 and @xmath43 respectively denoting the signal to noise ratio of the input and amplified signal beams , the nf of the single mode fopa is given by @xmath44 which shows that the parametric amplification process adds the excess noise to the amplified signal beam and the 3 db quantum limit of nf can be achieved in the high gain limit . in the ideal case , the photocurrents of d1 and d2 are highly correlated and can be subtracted from each other to yield quantum - noise reduction , i.e. , @xmath45 . in practice , the measured noise reduction is very sensitive to the collection efficiency of twin beams and the response function of differential detection system ( dds ) , consisting of d1 and d2 followed by the radio frequency ( rf ) amplifiers @xmath9 and @xmath10 , and the subtractor ( see fig . 1(a ) ) . for brevity , we assume throughout this paper that the transmission efficiencies of the filters @xmath7 and @xmath8 at their central wavelengths are perfect . in this situation , the collection efficiencies of the signal and idler beams at the single frequency @xmath4 and @xmath5 are perfect , so we only need to analyze the influence of dds with non - ideal response . the response function of the dds is described by @xmath46 and @xmath47 , which are determined by the quantum efficiencies and electrical circuit of d1 and d2 , respectively . here @xmath48 denotes the radio frequency . the dc response of the dds , @xmath49 , is associated with the quantum efficiency through the relation @xmath50 , where @xmath51 is the electron charge , @xmath52 is the quantum efficiency of d1 ( d2 ) , and @xmath53 is the reduced planck constant ; while the ac response , @xmath54 ( @xmath55 ) , is not only determined by the response of d1 ( d2 ) @xmath56 , but also determined by that of the rf amplifier @xmath57 . for convenience , the ratio @xmath58 between the ac response @xmath46 and @xmath47 , defined as @xmath59 which can be adjusted by changing the electronic gain of the amplifier @xmath60 or @xmath61 . in practice , the quantum efficiencies of d1 and d2 are not perfect , i.e. , @xmath62 and @xmath63 . modeling the non - ideal detectors as the beam splitters , which couple the vacuum mode @xmath64 and @xmath65 to the signal and idler fields , the field operator incident on d1 and d2 are given by [ eq : sm_operator ] @xmath66 @xmath67 in this situation , the measured photon numbers of the amplified signal and idler beams in eq . ( 3 ) are rewritten as [ sm_isi ] @xmath68 @xmath69 and the intensity noise of individual signal and idler beams in eq . ( 5 ) is rewritten as [ deltai_sm ] @xmath70 @xmath71 for the operator of the measured intensity difference of the twin beams @xmath72 the normalized intensity difference noise of the twin beams is expressed as @xmath73 where @xmath74 is the corresponding snl of the detected signal ( idler ) beam and @xmath75 is the quantum correlation term . substituting eqs . ( [ sm_isi])-([deltai_sm ] ) into eq . ( [ eq : electrical_gain ] ) , we obtain @xmath76 to illustrate the factors influencing the value of @xmath77 , we first plot @xmath78 as a function of the gain @xmath79 by assuming the ratio @xmath58 in eq . ( 11 ) is equal to 1 and by varying the efficiencies @xmath80 and @xmath81 , as shown in fig . when the efficiencies of both d1 and d2 are @xmath82 or @xmath83 , @xmath84 decreases with the increase of @xmath79 ; for a given @xmath79 , @xmath84 obtained for @xmath85 is lower than that for @xmath86 . when efficiencies of d1 and d2 are @xmath87 and @xmath88 , respectively , @xmath84 is higher than that for @xmath89 in the high gain regime , however , @xmath84 can be slightly lower than that for @xmath90 when the gain is within the regime of @xmath91 . therefore , for the case of @xmath92 , @xmath84 decreases with @xmath79 and increases with the decrease of @xmath52 ; while for the case of @xmath93 , @xmath84 does not always decrease with @xmath79 : after obtaining the minimum @xmath84 at a certain gain , @xmath84 will start to increase with @xmath79 . ( a ) the noise reduction @xmath84 as a function of the gain @xmath79 under different quantum efficiencies for the cases of ( a ) @xmath94 and ( b ) @xmath95 . , scaledwidth=55.0% ] eq . ( [ eq : r_sm_eta ] ) shows that for the given values of @xmath80 , @xmath81 and @xmath96 , @xmath84 can be minimized when the ratio @xmath58 takes the optimized value @xmath97 we then plot @xmath84 as a function of @xmath79 by varying the efficiencies @xmath80 and @xmath81 for the case of @xmath95 . as shown in fig . 2(b ) , no matter @xmath80 and @xmath81 are equal or not , @xmath84 always decreases with the increase of @xmath79 . moreover , for a fixed value of @xmath79 , our calculation indicates that @xmath84 decreases with the increase of @xmath98 and @xmath99 . we note that it is impossible to obtain the ideal noise reduction @xmath100 for the input signal with photon number @xmath101 , even if the detectors d1 and d2 are perfect . to clearly illustrate this point , let s look at the expressions of @xmath78 when the condition @xmath102 is fulfilled . for the case of @xmath103 , eq . ( [ eq : r_sm_eta ] ) is rewritten as @xmath104 while for the case of @xmath94 , eq . ( [ eq : r_sm_eta ] ) is rewritten as @xmath105 both eqs . ( [ r_sm_r ] ) and ( [ r_t_sm ] ) show the intensity difference noise of twin beams is always less than the snl ( 1 or @xmath106 db ) , however , the condition @xmath107 is always satisfied . moreover , a comparison between eqs . ( [ r_sm_r ] ) and ( [ r_t_sm ] ) indicates that in the high gain limit , the noise reduction can be improved up to 3 db by optimizing the ratio @xmath58 . when the fopa is pumped by a strong pulsed field having multiple frequency components , such as a mode - locked laser with a pulse width of a few pico - seconds , a single frequency signal mode is correlated to many idler frequency modes , and vice versa ( see the dotted arrow and the corresponding enveloping curve in fig.1 ( c ) ) . in this situation , both amplified signal and generated idler beams are pulsed fields with broad bandwidth , as illustrated by the spectrally broadened curves in fig . therefore , it is necessary to deal with the multi - frequency modes . in this section , we first extend the single mode bogoliubov transformation to the multi - frequency domain . then , according to the hamiltonian of fwm in fiber , we deduce the four transformation functions of the multi - mode bogoliubov transformation , from which the general expressions of the key parameters of a pulse pumped high gain fopa can be derived . an optical field propagating along the fiber ( denoted as the z direction ) can be quantized by using one dimensional approximation if the polarization mode is well defined . therefore , the pulsed field operator in optical fiber is written as @xcite @xmath108 where the positive frequency field operator @xmath109 and negative frequency operator @xmath110 satisfy the relation @xmath111^\dagger$ ] , and @xmath109 is given by @xmath112 here the annihilation operator at frequency @xmath113 , @xmath114 , satisfies the commutation relation @xmath115=\delta(\omega-\omega')$ ] . note that @xmath116 , so the distance in propagation is equivalent to a delay in time . for the pulse pumped fopa , the strong pump pulses remain classical , but the signal and idler fields are quantized . extending bogoliubov transformation in eq . ( [ single_trans ] ) to the multi - frequency domain and using eq . ( 22 ) , we write the input - output relation as @xcite : [ eq : multimode_trans ] @xmath117 @xmath118 where the unitary evolution operator @xmath119 is determined by the hamiltonian of the fopa(see later for the explicit form ) , the footnotes @xmath120 denote the frequency ranges of the signal and idler fields , and the four green functions @xmath121 are referred to as the transformation functions . to ensure the satisfaction of the commutation relations of the operators @xmath122 and @xmath123 , the transformation functions should be constrained by the relations : [ eq : fwm_opa_unitarity_conditions ] @xmath124 the weak input signal pulses , synchronized with the pump pulses , is ideally in a multi - mode coherent state @xmath125d\omega_s \big\ } | 0 \rangle,\ ] ] where @xmath126 is the complex amplitude , and the frequency distribution function @xmath127 satisfies the normalization condition @xmath128 so that the photon number of the input signal is @xmath129 , which is the same as eq . ( [ input - sig ] ) . in general , the bandwidth of the input signal is much smaller than that of the central frequency . so the quasi - monochromatic approximation applies , and the frequency integral range from @xmath106 to @xmath130 can be treated as from @xmath131 to @xmath130 . for the sake of brevity , all the integral ranges from @xmath131 to @xmath130 will be omitted hereinafter . in contrast to the single mode fopa , wherein the filters @xmath132 and @xmath133 with ideal transmission efficiency at the central frequencies result in perfect collection efficiency of twin beams , the collection efficiency of the broad band multi - mode twin beams is associated with the spectra of @xmath7 and @xmath8 . in this case , the field operators incident on the detectors d1 and d2 are given by [ eq : vacuum_mixing ] @xmath134 @xmath135 where the complex function @xmath136 ( @xmath137 ) describes the spectrum of the filter @xmath138 . eq . ( [ eq : vacuum_mixing ] ) implies that the measured quantum noise property of the pulse pumped fopa is not only influenced by the quantum efficiencies of d1 and d2 , but also influenced by the spectral functions @xmath139 and @xmath140 . for the ideal detectors , we should have @xmath102 , while for the perfect collection of the twin beams , we should have @xmath141 . similar to the single mode fopa , the key parameters of the pulse pumped fopa can be obtained by calculating the quantum average of the corresponding operators over the state @xmath142 . in the next subsection , we will derive the analytical forms of @xmath121 , which are the key for deriving the formulas of the operators of twin beams ( see eq . ( [ eq : multimode_trans ] ) ) . the unitary operator in eq . ( [ eq : multimode_trans ] ) is determined by the hamiltonian of parametric process . so let s begin with the hamiltonian of the co - polarized pulse pumped fwm in single - mode optical fibers @xcite @xmath143,\ ] ] where @xmath144 is a constant determined by experimental details and the units of quantized optical fields , @xmath0 is the 3rd - order real nonlinearity , @xmath145 and @xmath146 are the classical pump fields , [ eq : ea ] @xmath147 and @xmath148 are the quantized negative - frequency field operator of the signal and idler beams , respectively . the gaussian shaped strong pump pulses propagating along the fiber can be written as @xcite @xmath149 where @xmath150 , @xmath151 and @xmath152 are the bandwidth , central frequency , and wave vector of the pump , respectively , and @xmath153 is related to the peak power through the relation @xmath154 . the nonlinear coefficient @xmath155 is expressed as @xmath156 , where @xmath157 denotes the effective mode area and @xmath158 is the speed of light in vacuum . after substituting eqs . ( [ eq : ea ] ) and ( [ eq : pump ] ) into eq . ( [ eq : hamiltonian_0 ] ) and changing @xmath159 in eq . ( [ eq : hamiltonian_0 ] ) to @xmath160 , we carry out the integral over the whole fiber length @xmath161 ( from -l/2 to l/2 ) and arrive at the hamiltonian in the time dependent form @xmath162 where @xmath163 is the phase mismatching term , and the term @xmath164 is originated from the self - phase modulation of pump . to obtain the formula of the unitary evolution operator @xmath165 , we need to carry out the integration over the time and over all the possible combinations of @xmath2 and @xmath3 within the pump bandwidth . since the time integral gives rise to the @xmath166 function in @xmath167 to guarantee the energy conservation at single - photon level , we arrive at @xmath168\big\},\ ] ] where the so called two - photon joint spectral function ( jsf ) of fwm has the form of @xcite @xmath169 which is determined by the pump envelop function @xmath170 and phase matching function @xmath171 , and is referred to as the probability amplitude of simultaneously finding a pair of signal and idler photons within the frequency range of @xmath172 and @xmath173 , respectively . the coefficient @xmath174 in eq . ( [ eq : nondegererate_evolution_operator ] ) determines the gain of fwm , and the coefficient @xmath175 in eq . ( [ jsf_def ] ) is a constant used to ensure the satisfaction of the normalization condition @xmath176 . according to eq . ( [ jsf_def ] ) , the jsf is generally asymmetry since @xmath177 is frequency dependent . therefore , @xmath178 is generally not satisfied . to obtain the formula of the transformation functions , we rewrite the input - output relation by substituting eqs . ( [ eq : nondegererate_evolution_operator ] ) into eqs.([eq : multimode_trans ] ) and using baker - hausdorff lemma : @xmath179\nonumber\\ & & \ \ + \frac{1}{2!}[\hat{a}_j(\omega_j'),[\hat{a}_j(\omega_j'),\hat{b}]]+\frac{1}{3!}[\hat{a}_j(\omega_j'),[\hat{a}_j(\omega_j'),[\hat{a}_j(\omega_j'),\hat{b}]]]+ ... \ , ~~ \label{lieseries } \end{aligned}\ ] ] where the operator @xmath180 is expressed as @xmath181.\ ] ] then the transformation functions can be written in the form of an infinite series [ eq : trans_intergral ] @xmath182\nonumber\\ & & [ \psi^*(\omega_2,\omega_1)\psi^*(\omega_4,\omega_3)\psi^*(\omega_6,\omega_5 ) \cdot\cdot\cdot \psi^*(\omega_s',\omega_{2n-1 } ) ] \}\end{aligned}\ ] ] @xmath183\nonumber\\&&[\psi(\omega_s,\omega_1)\psi(\omega_2,\omega_3)\psi(\omega_4,\omega_5 ) \cdot\cdot\cdot \psi(\omega_{2n},\omega_i ' ) ] \}\end{aligned}\ ] ] @xmath184\nonumber\\ & & [ \psi^*(\omega_1,\omega_2)\psi^*(\omega_3,\omega_4)\psi^*(\omega_5,\omega_6 ) \cdot\cdot\cdot \psi^*(\omega_{2n-1},\omega_i ' ) ] \}\end{aligned}\ ] ] @xmath185\nonumber \\ & & [ \psi(\omega_1,\omega_i)\psi(\omega_3,\omega_2)\psi(\omega_5,\omega_4 ) \cdot\cdot\cdot \psi(\omega_s',\omega_{2n } ) ] \}.\end{aligned}\ ] ] eq . ( [ eq : trans_intergral ] ) clearly shows that the transformation functions satisfy the relations [ eq : signal_idler_correlation ] @xmath186 @xmath187 @xmath188 indicating that there is a correlation between the signal and idler beams . it is worth noting that bogoliubov transformation in multi - frequency modes ( eq ( [ eq : multimode_trans ] ) ) with the transformation function shown in eq . ( [ eq : trans_intergral ] ) is also suitable for describing the single mode fopa . for example , if the jsf in eq . ( [ jsf_def ] ) takes the limit of a single frequency laser pumped case , i.e. , @xmath189 , with @xmath190 and @xmath191 denoting the single frequencies of the signal and idler beams , the input - output relation of eq ( [ eq : multimode_trans ] ) would transform into @xmath192 with @xmath193 and @xmath194 , which is exactly the same as eq . ( [ single_trans ] ) . hence , eqs . ( [ eq : multimode_trans ] ) and ( [ eq : trans_intergral ] ) can be viewed as a general description of the input - output relation of an fopa . for the measurements of the signal and idler fields , in addition to taking the quantum efficiencies , @xmath80 and @xmath81 , and the spectrum of filter @xmath24 into account , we need to consider the response times of the detectors d1 and d2 as well . when the response time of each detector is much longer than the pump pulse duration , but is much shorter than the period of two adjacent pump pulses , the average photon number of the amplified signal ( generated idler ) beam per pulse can be expressed as @xmath195 where @xmath196 with @xmath197 is the photon number operator of the pulsed field . the time integral ranges in eqs . ( [ i_j ] ) and ( [ field - operator ] ) are omitted , because they can be treated as from @xmath131 to @xmath130 . with the filter operator incident on d1 and d2 ( eq . ( [ eq : vacuum_mixing ] ) ) , we substitute eqs . ( [ eq : multimode_trans ] ) , ( [ eq : coh_state ] ) , ( [ field - operator ] ) into eq . ( [ i_j ] ) , and obtain the expressions of the detected photon number [ eq : signal_idler_average photon number ] @xmath198 @xmath199 where [ spectra ] @xmath200 and @xmath201 are the power spectra of the signal and idler beams . the first terms in the right hand side of eqs . ( [ eq : signal_idler_average photon number]a ) and ( [ eq : signal_idler_average photon number]b ) are originated from the stimulated emission of input signal , while the second terms are from the spontaneous emission . since we have assumed that the photon number of the weak input signal per pulse is much greater than 1 , the spontaneous emission terms are negligible . using eq . ( [ eq : is ] ) and the definition of photon number gain in eq . ( [ def_g ] ) , we find the expression of the photon number gain of the pulse pumped fopa is @xmath202 for the case of @xmath203 and @xmath204 . to compute the intensity noise of the amplified signal ( generated idler ) beam @xmath205 we substitute eqs . ( [ eq : multimode_trans ] ) , ( [ eq : coh_state ] ) , ( [ eq : vacuum_mixing ] ) and ( [ i_j])-([field - operator ] ) into eq . ( [ eq : fluctuation_general ] ) , and arrive at the detailed expression [ eq : photon_number_fluctuation_fwm_opa ] @xmath206 @xmath207 with [ eq45 ] @xmath208 @xmath209 @xmath210 @xmath211 and [ eq46 ] @xmath212\end{aligned}\ ] ] @xmath213.\end{aligned}\ ] ] accordingly , the normalized intensity noise of signal ( idler ) beam is @xmath214 moreover , we have @xmath215 by substituting eqs . ( [ eq : signal_idler_average photon number ] ) and ( [ rsandri ] ) into eq . ( [ snr - out ] ) . according to the definition in eq . ( [ eq : nf1 ] ) , we obtain the formula of noise figure @xmath216 eqs . ( [ eq : photon_number_fluctuation_fwm_opa])-([rsandri ] ) indicate that the intensity noises of the individual beams @xmath217 are not only related to detection efficiencies , @xmath80 and @xmath81 , but also to the spectral functions of filters , @xmath139 and @xmath140 . when @xmath80 and @xmath81 are very low , i.e. , @xmath218 ( @xmath137 ) , the intensity fluctuation @xmath26 in eq . ( [ eq : photon_number_fluctuation_fwm_opa ] ) is dominated by the term @xmath219 ( @xmath137 ) , while the contribution of the term @xmath220 ( @xmath221 ) is negligible . on the other hand , when the bandwidth of the filters @xmath7 and @xmath8 are much narrower than that of input signal , @xmath26 is also dominated by the term @xmath219 ( @xmath137 ) . since the non - ideal detector induces the vacuum noise to the individual signal or idler field , the former is easy to understand . however , the understanding of the latter is not so straightforward . we think this is because , when the signal and idler beams are in multi - frequency mode , the filter @xmath7 ( @xmath8 ) with bandwidth narrower than that of the amplified signal ( generated idler ) beam can also be viewed as a loss , which introduce vacuum noise to the signal ( idler ) beam as well . hence , the term @xmath219 ( @xmath137 ) in eqs . ( [ eq : h_vac_s ] ) and ( [ eq : h_vac_i ] ) is originated from the loss induced vacuum . if the detection and collection efficiencies of the signal and idler beam are ideal , we would have @xmath222 ( @xmath137 ) . finally , we derive the expression of the intensity difference noise of the twin beams by using the definition @xmath223 where the operator @xmath224 is the photon number difference with a weight factor of ac response ratio @xmath58 . using eqs . ( [ eq : multimode_trans ] ) , ( [ eq : vacuum_mixing ] ) , ( [ i_j])-([field - operator ] ) and ( [ eq : fluctuation_general ] ) , eq . ( [ difference - fluctuation_general ] ) is transformed into @xmath225 - ( i_s - ri_i)^2\nonumber\\ = & & \delta i_s^2+r^2\delta i_i^2 - 2r(h_{si1}+h_{si2})|\alpha|^2 , \end{aligned}\ ] ] where @xmath226 and @xmath227 are positive terms originated from the quantum correlation between signal and idler twin beams . it is obvious that @xmath228 is less than the sum of intensity noise of individual beams @xmath229 due to the correlation of twin beams . consequently , similar to eq . ( [ eq : electrical_gain ] ) , the general expression of the normalized intensity difference noise for pulsed twin beams @xmath230 |\alpha|^2 } { i_s + r^2 i_i } \ ] ] is obtained by substituting eqs . ( [ eq : photon_number_fluctuation_fwm_opa ] ) and ( [ eq : h_si1])-([eq : h_si2 ] ) into eq . ( [ eq : it_h ] ) . in addition , @xmath231 can be further minimized if the ac response ratio @xmath58 takes the optimized value @xmath232 ^ 2 } } { 2i_i(h_{si1}+h_{si2})}\nonumber\\ & & + \frac{i_i(h_{s1}+h_{s2}+h_{vs})-i_s(h_{i1}+h_{i2}+h{vi})}{2i_i(h_{si1}+h_{si2})}.\end{aligned}\ ] ] from the formulas of the photon number gain , noise figure and intensity difference noise in eqs . ( [ gdef ] ) , ( [ nf2 ] ) and ( [ eq : r_t2 ] ) , one sees that the key parameters of the fopa are determined by the transformation functions in eq . ( [ eq : trans_intergral ] ) , which highly depend on the jsf in eq . ( [ jsf_def ] ) . in the following sections , we will focus on studying the noise characteristics of the pulse pumped fopas with two typical kinds of jsf . we first study the fopa with a spectrally factorable jsf . under this condition , both the signal and idler modes are in single temporal mode , so we expect to recover the results for single mode fopa in sec . [ single_mode ] . for the sake of brevity , in this section , we assume the quantum efficiencies of d1 and d2 are perfect , and the collection efficiency of the twin beams is ideal . a factorable jsf , obtained by properly regulating the pump pulses and tailoring the dispersion of optical fibers , is written as @xcite @xmath233 where @xmath234 and @xmath235 with the normalization condition of @xmath236 and @xmath237 are the probability amplitude of finding a pairs of signal and idler photons within the frequency range of @xmath238 and @xmath239 , respectively . in fact , looking at eq . ( [ eq : factotized jsf ] ) from another point of view , we find @xmath234 and @xmath235 can also be viewed as the gain spectra in signal and idler fields , respectively . with the jsf in eq . ( [ eq : factotized jsf ] ) , the transformation functions in eq . ( [ eq : trans_intergral ] ) are simplified to [ eq : fac_trans_func ] @xmath240 note that in this section , we will use the superscript @xmath241 to represent the result of an fopa with the factorable jsf . with the jsf in eq . ( [ eq : factotized jsf ] ) , the average photon numbers of the amplified signal and generated idler beams per pulse in eq.([eq : signal_idler_average photon number ] ) are simplified to [ eq : factorized_signal_avg_ph_number ] @xmath242 where the coefficient @xmath243 is determined by the matching between the spectra of signal injection @xmath244 ( see eq . ( [ eq : coh_state ] ) ) and that of the gain in signal band @xmath245 . therefore , we obtain the photon number gain of the fopa @xmath246 by substituting eqs . ( [ eq : factorized_signal_avg_ph_number ] ) and ( [ input - sig ] ) into eq . ( [ def_g ] ) . to show the factors influencing the noise performance of the fopa with factorable jsf , we first substitute the simplified transformation function in eq . ( [ eq : fac_trans_func ] ) into eqs.([eq45])-([eq46 ] ) and ( [ eq : h_si1])-([eq : h_si2 ] ) . for the case of @xmath247 , the expressions of the terms in eqs . ( [ eq45])-([eq46 ] ) and ( [ eq : h_si1])-([eq : h_si2 ] ) , which determine @xmath248 ( @xmath137 ) in eq . ( [ eq : photon_number_fluctuation_fwm_opa ] ) and @xmath249 in eq . ( [ eq : it_h ] ) , are simplified to @xmath250 @xmath251 @xmath252 @xmath253 consequently , the normalized intensity noise of individual signal and idler beams are @xmath254 and @xmath255 respectively . moreover , according to eq . ( [ eq : factorized_signal_avg_ph_number ] ) , the noise figure in eq . ( [ nf2 ] ) and the normalized intensity difference noise of twin beams in eq . ( [ eq : r_t2 ] ) can be respectively simplified to @xmath256 and @xmath257 equations ( [ rs_fac])-([eq : rt_fac ] ) show that except for the normalized intensity noise of idler beam @xmath258 , the remaining three parameters , @xmath259 , @xmath260 and @xmath261 , are associated with the matching coefficient @xmath262 . for the case of @xmath263 , if we respectively replace the terms @xmath264 and @xmath265 with @xmath266 and @xmath31 , the analytical expressions of the key parameters ( eqs . ( [ rs_fac])-([eq : rt_fac ] ) ) will be the same as those of the single frequency mode fopa in sec . [ single_mode ] . the results indicate that our multi - mode theory of the pulse pumped fopa are valid . an marked difference between the fopas described by single temporal mode and single frequency mode is the spectra of signal and idler beams . the spectra of the amplified signal and generated idler beams obtained by substituting eq . ( [ eq : fac_trans_func ] ) into eq.([spectra ] ) are written as [ spectrum_fa ] @xmath267 and @xmath268 respectively . from eq . ( [ spectrum_fa ] ) , one sees that the spectrum of generated idler is always the same as the spectrum of the gain in idler field , however , different from the fopa pumped with a single frequency laser , the spectra of the injected signal and amplified signal beams are different unless the condition @xmath263 is fulfilled . to further understand the difference between the fopas described by single temporal mode and single frequency mode , we then compare the noise performance of the two cases . we find that for a given value of @xmath269 ( or @xmath270 ) , the noise figure and the intensity difference noise of the twin beams in eqs . ( [ eq : nf_fac ] ) and ( [ eq : rt_fac ] ) are worse than that predicted by eqs ( [ nf_s_sm ] ) and ( [ r_t_sm ] ) unless @xmath263 . in practice , it is very difficult to maintain the spectral matching condition @xmath271 in the high gain regime due to the self - phase modulation and cross - phase modulation induced spectral broadening @xcite . hence , it is very challenging to experimentally realize a pulse pumped fopa capable of being described by a single mode theory . for the conventional optical communication systems , an fopa with an extremely broad gain bandwidth is desirable . when the gain bandwidth of the fopa is much broader than the bandwidths of the pump and input signal , the jsf can be simplified by assuming the perfect phase matching condition @xmath272 . in this case , the jsf in eq . ( [ jsf_def ] ) is rewritten as @xcite @xmath273 which is obviously non - factorable . in this section , we focus on numerical investigation of the quantum noise performance of this kind of fopa . with the jsf in eq . ( [ eq : jsf_pm ] ) , the transformation functions in eq.([eq : trans_intergral ] ) are reformulated as [ trans_unfac ] @xmath274 where @xmath275 , and the superscript @xmath276 represents the result for the fopa with broad gain bandwidth . for the sake of convenience , the frequency variables @xmath277 and @xmath278 used in previous sections will be replaced with @xmath113 and @xmath279 hereinafter . we first study the factors influencing the key parameters of the fopa by assuming the detectors are perfect , and collection efficiency of the twin beams is ideal , i.e. , @xmath102 and @xmath280 . without loss of the generality , we assume the spectral function of the pulsed input signal can be described by the gaussian function @xmath281 where @xmath282 denotes the bandwidth . from eqs . ( [ trans_unfac ] ) and ( [ eq : guassian_input_signal_spectrum ] ) , we can rewrite the power spectra of the amplified signal and generated idler beams in eq . ( [ spectra ] ) as [ spectra_unfac ] @xmath283 for all the terms of the infinite series in eqs . ( [ spectra_unfac]a ) and ( [ spectra_unfac]b ) , whose order corresponds to @xmath284 or @xmath285 , their bandwidths are larger than that of the input power spectrum @xmath286 . moreover , since the bandwidth of the term in infinite series increases with the order determined by the integers @xmath287 and @xmath288 , the bandwidths of both signal and idler twin beams increase with @xmath289 . using eq . ( [ spectra_unfac ] ) , the average photon numbers of amplified signal field and generated idler beams in eq . ( [ eq : signal_idler_average photon number ] ) can be rewritten as [ i_unfac ] @xmath290 where @xmath291 describes the ratio of the pump bandwidth to the input signal bandwidth . for all the numerically simulated results presented hereinafter , the coefficient @xmath289 is within the range of @xmath292 , and the series in eqs . ( [ trans_unfac ] ) , ( [ spectra_unfac ] ) and ( [ i_unfac ] ) are calculated up to the order of @xmath293 and @xmath294 . to ensure the correctness of this truncation approximation , we numerically verify the general relation @xmath295 ( @xmath129 ) , which is the inherent nature of signal and idler twin beams originated from the energy conservation of fwm . after calculating the series in eq . ( [ i_unfac ] ) to the order of @xmath293 and @xmath294 , we find that the difference between the calculated result of @xmath296 and 1 " is less than @xmath297 , showing the validity of the approximation . figures [ spectrum](a ) and [ spectrum](b ) show the normalized power spectra of the amplified signal and generated idler beams . in the plots , eqs . ( [ spectra_unfac]a ) and ( [ spectra_unfac]b ) are calculated by varying the value of @xmath289 under the condition @xmath298 . as a comparison , we also show the power spectrum of input signal @xmath299 in fig . [ spectrum](a ) . it is obvious that the bandwidths of both signal and idler twin beams are broader than that of the input signal and they increase with @xmath289 , showing the spectrum broadening effect . moreover , in the low gain regime , the spectrum of amplified signal is mainly determined by the input signal , so the bandwidth of generated idler beam , which is the convolution of the spectra of the input signal and pump , is greater than that of the signal beam . however , in the high gain regime , the bandwidths of signal and idler twin beams are about equal because both the spectra of signal and idler beams are the convolution of the spectra of pump and its counterparts . the normalized power spectra of ( a ) amplifed signal and ( b ) generated idler beams at different @xmath289 under the condition @xmath298 . as a comparison , the power spetrum of input signal @xmath286 is plotted in ( a ) as well.,scaledwidth=48.0% ] figure [ fig : photon_number_gain_no_filtered ] plots the dependence of the photon number gain @xmath300 upon the pump power for the different ratio @xmath291 . in the calculation , @xmath301 is changed by varying the peak pump power @xmath302 . one sees that at a fixed peak power @xmath302 , @xmath270 increases with the decrease of @xmath303 because the overlap of temporal mode , which is required for maximizing the gain of fwm , is improved by decreasing the bandwidth of pump . when @xmath303 is less than @xmath304 , further decreasing the ratio to @xmath305 does not result in an obvious increase in @xmath270 , because no more space is left for improving the temporal mode overlapping . the result in fig . [ fig : photon_number_gain_no_filtered ] implies the pump with a longer pulse duration ( narrower bandwidth ) gives a better mode overlapping . however , it is worth noting that the pump with a short pulse duration helps to obtain a high gain under the condition of low average pump power , which is of practical importance . photon number gain @xmath270 versus the peak pump power for the different ratio @xmath291 . in the calculation , @xmath272 , @xmath306 , the parameter @xmath307 , and @xmath308.,scaledwidth=33.0% ] to investigate the quantum noise performance of the fopa with broad gain bandwidth , we first deduce the formulas of terms @xmath220 ( @xmath137 , @xmath309 ) and @xmath219 ( @xmath137 ) , which determine @xmath248 ( @xmath137 ) in eq . ( [ eq : photon_number_fluctuation_fwm_opa ] ) . by substituting eqs . ( [ trans_unfac ] ) and ( [ eq : guassian_input_signal_spectrum ] ) into eqs . ( [ eq45])-([eq46 ] ) , we obtain @xmath310 @xmath311 @xmath312 @xmath313 with the above , we then numerically study the dependence of the normalized intensity noise of individual beams @xmath314 ( @xmath137 ) upon the ratio @xmath303 . figure [ fig : relative_intensity_noise_no_filtered](a ) plots @xmath315 and @xmath316 versus @xmath270 for the cases of @xmath317 , @xmath318 , and @xmath319 , respectively . from fig . [ fig : relative_intensity_noise_no_filtered](a ) , one sees that for a fixed value of @xmath270 ( @xmath303 ) , both @xmath315 and @xmath316 increase with @xmath303 ( @xmath270 ) . moreover , for the given values of @xmath303 and @xmath270 , @xmath316 is greater than @xmath315 . this is different from the results of @xmath320 ( @xmath137 ) for a single mode fopa [ see eq . ( [ r_cw ] ) ] , but has been experimentally confirmed in references @xcite and @xcite . ( a ) the normalized intensity noise of the individual signal/ idler beams @xmath321 ( @xmath137 ) and ( b ) noise figure nf versus the photon number gain @xmath270 for the different ratio @xmath291 . in this calculation , @xmath272 , and @xmath247.,scaledwidth=33.0% ] we also compute the noise figure @xmath322 in eq . ( [ nf2 ] ) . figure [ fig : relative_intensity_noise_no_filtered](b ) plots @xmath322 versus @xmath270 for @xmath317 , @xmath318 , and @xmath319 , respectively . one sees that for a fixed ratio @xmath303 , @xmath322 increases with @xmath270 in the low gain regime , and approaches a constant in the high gain regime . while for a fixed @xmath270 , the value of nf increases with the ratio @xmath303 . it is obvious that for the pulsed pump case with @xmath323 , @xmath322 in the high gain regime is always greater than the well known 3db - limit of a single mode fopa . the normalized intensity difference noise of the twin beams @xmath231 versus photon number gain @xmath270 for @xmath324 , @xmath325 and @xmath326 when @xmath58 is set to ( a ) @xmath94 and ( b ) @xmath327 , respectively.,scaledwidth=55.0% ] the optimized value @xmath328 and the photon number ratio @xmath329 as a function of @xmath270 . in this calculation , the phase mismatching term of the fopa is @xmath272 , and the detection efficiencies and the filters satisfy @xmath247.,scaledwidth=40.0% ] to study the factors influencing the normalized intensity difference noise of the twin beams @xmath231 in eq . ( [ eq : r_t2 ] ) , we need to calculate @xmath330 and @xmath331 as well . by substituting eqs . ( [ trans_unfac ] ) and ( [ eq : guassian_input_signal_spectrum ] ) into eqs . ( [ eq : h_si1 ] ) and ( [ eq : h_si2 ] ) , we find the following relation @xmath332 in figs . [ fig : twin_beams_no_filtered](a ) and [ fig : twin_beams_no_filtered](b ) , we plot @xmath231 as a function of @xmath270 for different values of @xmath303 when the ac response ratio @xmath58 takes the values of @xmath94 and @xmath327 , respectively . [ fig : twin_beams_no_filtered](a ) shows @xmath231 is independent on @xmath303 and always decreases with the increase of @xmath270 ; while fig . [ fig : twin_beams_no_filtered](b ) demonstrates that @xmath231 still decreases with the increase of @xmath270 , but @xmath231 can be further decreased and depends upon the ratio @xmath303 . for a fixed gain @xmath270 , @xmath231 decreases with @xmath303 , which is different from the case of @xmath94 . moreover , when @xmath303 is very small , for example , @xmath324 , the difference between the @xmath231-values with @xmath94 and @xmath327 is about 3 db in the high gain limit ; when @xmath303 is very big , for example , @xmath326 , @xmath231-values with @xmath94 and @xmath327 are almost the same . the results in figs . [ fig : twin_beams_no_filtered](a ) and [ fig : twin_beams_no_filtered](b ) not only indicate the noise reduction of @xmath231 can be improved by properly reducing the ratio @xmath303 , but also imply that @xmath328 might depend upon @xmath303 and @xmath328 might be very close to 1 when the value of @xmath303 is greater than 10 . this is shown in fig . [ fig6 ] , in which we plot @xmath328 as a function of @xmath270 for @xmath324 , @xmath325 and @xmath326 . in addition , we also plot the photon number ratio @xmath329 versus @xmath270 in fig . one sees that @xmath328 is always within the range of @xmath329 and @xmath333 . for a fixed value of @xmath270 , @xmath328 decreases with the increase of @xmath303 . as we expected , @xmath328 obtained for the case of @xmath326 is about 1 if @xmath270 is greater than @xmath334 . moreover , in the high gain limit , both @xmath328 and @xmath329 approach @xmath333 , which is irrelevant to @xmath303 . in practice , the ideal condition @xmath247 can never be fulfilled , the measured intensity difference noise @xmath231 is closely related to the loss of the fopa system . so we need to study the influence of the quantum efficiencies ( @xmath80 and @xmath81 ) and collection efficiency of twin beams . in this subsection , we assume the filters placed at the output port of the fopa ( see fig . 1 ) have a gaussian shaped spectrum @xmath335 where @xmath336 and @xmath337 denote the central frequency and bandwidth of the filter @xmath338 @xmath339 , respectively . using the transformation functions , input signal spectrum and the spectrum of @xmath338 @xmath339 in eqs . ( [ trans_unfac ] ) , ( [ eq : guassian_input_signal_spectrum ] ) and ( [ eq : filter_shape_guassian ] ) , respectively , the measured photon numbers of the amplified signal and generated idler beams in eq . ( [ eq : signal_idler_average photon number ] ) are rewritten as [ eq : perfect_phase_matching_filtered_signal_ave_pho_num ] @xmath340\ \}^{-1/2 } \\ & & i'^{(b)}_i=\eta_i \sum_{n_1=0}^{\infty } \sum_{n_2=0}^{\infty } \frac{|g'|^{2n_1 + 2n_2 + 2}}{(2n_1 + 1)!(2n_2 + 1 ) ! } \nonumber\\&&\ \ \ \ \ \{\ s^2[(4n_1 + 2)p^2 + 1][(4n_2 + 2)p^2 + 1 ] \nonumber\\&&\ \ \ \ \ \ \ + [ 2(n_1+n_2 + 1)p^2 + 1]\ \}^{-1/2},\end{aligned}\ ] ] for a fixed value of @xmath351 , if the bandwidth of filters @xmath24 is much narrower than that of the input signal , namely , @xmath352 , we have @xmath353 . in this case , the collection efficiency of the twin beams is very low due to the multi - mode nature of the signal and idler fields @xcite . since the narrow band filters @xmath24 introduce the vacuum noise into the individual signal and idler beams , for a given @xmath289 , the value of @xmath354 @xmath339 in eqs ( [ hvacs_f])-([hvaci_f ] ) approaches to @xmath355 and is dominant over the terms @xmath356 ( @xmath357 ) and @xmath358 ( @xmath357 ) ( eqs . ( [ hs1_f])-([hi2_f ] ) and ( [ hsi1_f])-([hsi2_f ] ) ) . hence , the normalized intensity noise and normalized intensity difference noise approach to the snl , i.e. @xmath359 and @xmath360 . in general , the collection efficiency of the twin beams is not only associated with the ratio @xmath361 , but also depends on the ratio @xmath303 . to reduce the detrimental effect of the collection efficiency on @xmath231 , both @xmath361 and @xmath303 should be small enough unless the condition @xmath362 or @xmath305 is satisfied @xcite . to illustrate this point , we plot the contour of the normalized intensity difference noise @xmath231 as a function of @xmath303 and @xmath361 in fig . [ contour](a ) . in the calculation , we assume @xmath363 , @xmath94 , and the quantum efficiencies of d1 and d2 are perfect . for each combination of @xmath361 and @xmath303 , we compute the terms in eqs . ( [ eq : perfect_phase_matching_filtered_signal_ave_pho_num])-([hsi2_f ] ) and substitute the results into eq . ( [ eq : r_t2 ] ) . [ contour](a ) shows that @xmath231 depends on both @xmath361 and @xmath303 . we notice that with the increase of @xmath303 , the dependence of @xmath231 upon @xmath361 becomes stronger . when @xmath303 is less than @xmath364 , at a fixed @xmath303 , @xmath231 increases with @xmath361 , but the noise reduction condition of @xmath365 is always achievable . however , when @xmath303 is greater than @xmath364 , to obtain the noise reduction of @xmath365 , the value of @xmath361 should be smaller than a certain value , which decreases with the increase of @xmath303 . in real experiment , it is necessary to use the filter @xmath24 with a certain bandwidth to prevent the strong pump background from reaching the detectors d1 and d2 . so , to improve the noise reduction of twin beams , the pump with a pulse duration longer than that of the input signal , i.e. , @xmath366 , is desirable . contour plot of ( a ) the normalized intensity difference noise @xmath231 , ( b ) @xmath367 , and ( c ) @xmath368 as a function of the ratios @xmath303 and @xmath361 . in this calculation , @xmath369 , @xmath363 , and @xmath94 , and @xmath370 , scaledwidth=85.0% ] figure [ contour](a ) also shows that the contour @xmath371 divides the plot into two zones according to @xmath372 or @xmath373 . for the former case , there is a one - to - one correspondence between the combination of @xmath361 , @xmath303 , and @xmath231 ; while for the latter case , each @xmath231 maps two values of @xmath361 at a fixed value of @xmath303 : one is very close to @xmath106 , the other is larger . to better understand fig . [ contour](a ) , we also plot the contours of the sums of the positive terms and negative terms in eq . ( [ eq : r_t2 ] ) in figs . [ contour](b ) and [ contour](c ) , which respectively illustrate the influence of collection efficiency upon the intensity noise of individual beams and upon the correlation of twin beams . for both figs . [ contour](b ) and [ contour](c ) , the contour decreases with the increase of @xmath361 ( @xmath303 ) at a fixed value of @xmath303 ( @xmath361 ) . however , since the correlation of twin beams highly relies on the collection efficiency , the dependence of the contour on @xmath361 and @xmath303 in fig . [ contour](c ) is stronger than that in fig . [ contour](b ) . in particular , when @xmath303 increases , the value of @xmath374 in fig . [ contour](c ) rapidly decreases with the increase of @xmath361 , which is responsible for the multivalued mappings in the zone of @xmath373 ( see fig . [ contour](a ) ) . the normalized intensity difference noise of the twin beams @xmath231 versus @xmath270 for ( a ) @xmath375 , ( b ) @xmath325 , respectively . for each setting of @xmath303 , @xmath361 in eq.([eq : perfect_phase_matching_filtered_signal_ave_pho_num ] ) and ( [ hs1_f])-([hvaci_f ] ) is @xmath376 , @xmath377 , and @xmath378 , respectively . in this calculation , @xmath369 . , scaledwidth=66.0% ] to further demonstrate the dependence of the noise reduction of @xmath231 upon the photon number gain @xmath270 , we plot @xmath231 as a function of @xmath270 for @xmath376 , @xmath377 and @xmath378 . in figs . [ fig : twin_beams_filter_bandwidth](a ) and [ fig : twin_beams_filter_bandwidth](b ) , we assume @xmath102 and @xmath94 , and @xmath303 is fixed at @xmath324 and @xmath325 , respectively . for each combination of @xmath361 and @xmath303 , we compute the terms in eqs . ( [ eq : perfect_phase_matching_filtered_signal_ave_pho_num])-([hvaci_f ] ) by changing @xmath289 ( @xmath270 ) and substitute the results into eq.([eq : r_t2 ] ) . it is obvious that for @xmath303 and @xmath361 with fixed values , @xmath231 decreases with the increase of @xmath270 . moreover , at a fixed value of @xmath270 and @xmath361 ( @xmath303 ) , @xmath231 increases with the increase of @xmath303 ( @xmath361 ) due to the decreased collection efficiency of twin beams . the normalized intensity difference noise of the twin beams @xmath231 versus g for different setting of @xmath80 and @xmath81 . for the plots in ( a ) , @xmath379 ; while for the plots in ( b ) , r is optimized for minimizing the measured @xmath231 . in this calculation , @xmath318 and @xmath380.,scaledwidth=66.0% ] having understood the influence of the collection efficiency of twin beams on @xmath231 , we then study the influence of the detection efficiency on @xmath231 when the values of @xmath303 and @xmath361 are fixed . without loss of generality , we assume @xmath325 and @xmath377 . when the ac response ratio is set to @xmath94 , we plot @xmath231 as a function of @xmath270 by varying @xmath80 and @xmath81 , as shown in fig . [ fig : twin_beams_filtered_effeciency](a ) . it is obvious that for the case of @xmath92 , @xmath231 decreases with the increase of @xmath52 at a fixed value of @xmath270 . however , for the case of @xmath93 , @xmath231 is not a monotonically decreasing function of @xmath270 and @xmath52 . depending on @xmath270 , the correlation of photon currents originated from the quantum correlation of the signal and idler beams might become weak or strong due to the unbalanced detection efficiencies . as shown in fig . [ fig : twin_beams_filtered_effeciency](a ) , @xmath231 for the case of @xmath87 and @xmath88 is lower than that for @xmath90 when @xmath270 is less than 18 ; however , @xmath231 becomes even higher than that for @xmath89 when @xmath270 is greater than 30 . making comparison between fig . [ fig : twin_beams_filtered_effeciency](a ) and fig . 2(a ) , which are obtained for fopa with non - factorable jsf and with single frequency pump , respectively , we find there are two differences : ( i ) for @xmath270 , @xmath80 and @xmath81 with certain values , @xmath231 in fig . [ fig : twin_beams_filtered_effeciency](a ) is larger than that in fig . 2(a ) because the collection efficiency of the pulsed twin beams is smaller ; ( ii ) the difference between the curve for @xmath87 and @xmath88 and that for @xmath90 ( @xmath89 ) in fig . [ fig : twin_beams_filtered_effeciency](a ) is more apparent than that in fig . therefore , for the pulse pumped fopa , the measured @xmath231 can be minimized by properly adjusting the efficiency @xmath80 or @xmath81 @xcite . however , this method will introduce extra loss , and the minimized @xmath231 can only be obtained for @xmath270 with a specified value . if the ratio @xmath58 is adjustable , we can minimize @xmath231 by optimizing the value of @xmath58 without introducing the vacuum noise . in this case , as shown in fig . [ fig : twin_beams_filtered_effeciency](b ) , for fixed values of @xmath80 and @xmath81 , @xmath231 always decreases with the increase of @xmath270 ; while for a fixed value of @xmath270 , @xmath231 always decreases with the increase of @xmath98 and @xmath99 . hence , for both the signal and idler channels , the higher the efficiencies are , the better the noise reduction is . finally , we analyze the influence of the excess noise of input signal on @xmath231 , since the noise of input signal pulses originated from the mode - locked fiber lasers is often higher than the snl @xcite . when the excess noise is included , the input signal can be viewed as a mixture of coherent states , whose density operator and average photon number are written as @xmath381 and @xmath382 respectively , where @xmath383 is a complex random variable and @xmath384 is the classical probability density . consequently , the formulas of the average photon number of amplified signal and generated idler beams , @xmath385 and @xmath386 , can be obtained by replacing the term @xmath387 in eq . ( [ eq : signal_idler_average photon number ] ) with @xmath388 . we then calculate the quantum noise @xmath26 in eq . ( [ eq : fluctuation_general ] ) and @xmath389 in eq . ( [ difference - fluctuation_general ] ) by taking the quantum average over the density operator @xmath390 . after some algebra , we get the normalized intensity noise of the amplified signal ( generated idler ) beam @xmath391 and the normalized intensity difference noise of twin beams @xmath392 where @xmath393 , and the inequality @xmath394 holds for the input signal with excess noise . in eqs . ( [ rj - excess ] ) and ( [ rt - excess ] ) , the first terms @xmath395 and @xmath396 in the right hand sides are the corresponding noise for input signal in a pure coherent state with @xmath397 ( see eqs . ( [ rsandri ] ) and ( [ eq : r_t2 ] ) ) ; while the second terms in the right hand sides are originated from the excess noise of input signal . equations ( [ rj - excess ] ) and ( [ rt - excess ] ) indicate that the side effect of the excess noise can be eliminated under the condition @xmath398 , and the experiment in ref . @xcite has verified this for the case of @xmath94 . in the low gain regime , the value of @xmath398 is different from @xmath328 for minimizing @xmath231 ( see fig . [ fig : twin_beams_no_filtered](b ) ) , so the excess noise the input signal is deleterious for the noise reduction of @xmath231 . however , in the high gain limit , we have @xmath399 . in this case , eliminating the side effect of excess signal noise and minimizing @xmath231 can be achieved simultaneously . therefore , a high gain fopa is desirable for improving the noise reduction of twin beams . in conclusion , we have developed the multi - mode quantum theory for analyzing the noise characteristics of a non - degenerate phase insensitive fopa pumped by a pulsed laser . the calculation shows that both the noise figure and the intensity difference noise of twin beams depend on the jsf . in the high gain regime , the noise figure is generally greater than the 3db quantum limit unless the jsf is factorable and the spectrum of the injected signal well matches the gain spectrum in signal band ; while the intensity difference noise can be significantly less than the snl . to closely resemble the experiments , the quantum noise of twin beams generated by a broadband fopa is numerically studied by taking the real experimental conditions into account . in addition to the influences of the quantum efficiency of detectors and the excesses noise of input signal , the influence of the collection efficiency of the twin beams , determined by the ratio of the pump bandwidth to signal bandwidth @xmath303 and the ratio of the signal bandwidth to filter bandwidth @xmath361 , are carefully analyzed as well . our theoretical investigation is of practical importance , it not only serves as a guideline for optimizing the measured noise reduction of the twin beams , but also helps for understanding the quantum noise preformance of pulse pumped fopa . instead of using the bloch - messiah reduction or schmidt decomposition of the jsf , which has been exploited to theoretically analyze the pulsed cv nonclassical light via the high gain spontaneous parametric down conversions in refs . @xcite , our transformation functions of the bogoliubov transformation are expressed in the infinite series of the gain coefficient @xmath269 . the accuracy of our numerical simulation can be controlled by the two parameters : one is the precision of the jsf determined by the step size of the frequency range , the other is the order of the series determined by the gain of fwm . so our method has a greater flexibility . we believe our theory can be extended to a more generalized opa , including the degenerate opa and the phase - sensitive opa . moreover , our theory can also be used to study the quantum noise of other quantities , such as the noise correlation between the quadrature amplitudes of the signal and idler beams . however , it is worth pointing out that the fopa in our model is simplified as a linear amplifier , and we only considered the fwm process . therefore , our calculation only quantitatively explains the results obtained under the condition of negligible pump depletion @xcite . to model the fopa system more accurately , other effects like raman effect , higher - order fwm and gain saturation effect should be included as well @xcite . this work was supported in part by the state key development program for basic research of china ( no . 2010cb923101 ) and by the specialized research fund for the doctoral program of higher education of china ( no . 20120032110055 ) .

nonlinear globally coupled systems under the influence of noise have been an active field of research over the last few decades @xcite . for additive noise there is a natural and far - reaching analogy to equilibrium thermodynamics @xcite . we consider an array of @xmath1 harmonically coupled overdamped anharmonic oscillators subject to additive noise which is governed by the system of langevin equations @xmath2 for @xmath3 . each of the oscillators is harmonically coupled to all the others ( global coupling ) , the total strength of the coupling is @xmath4 . the additive noise @xmath5 is a zero mean gaussian white process with autocorrelation @xmath6 where @xmath7 denotes the noise strength . the steady state of an isolated system without noise ( @xmath8 ) undergoes a pitchfork bifurcation if the control parameter @xmath9 changes the sign . the isolated system ( @xmath10 ) describes diffusion in a potential with one or two minima depending on the sign of @xmath9 which is thoroughly studied since @xcite . in the limit @xmath11 the array exhibits a continuous phase transition with a critical point @xmath12 and the mean field critical exponent @xmath13 such that for @xmath14 the stationary probability density @xmath15 is symmetric with respect to @xmath16 , and for @xmath17 it is non - symmetric such that the order parameter @xmath18 . the dependence of @xmath0 on the other parameters has been investigated already by a number of authors @xcite . kometani and shimizu @xcite closed the equation of motion for the moments using a decoupling which is correct for gaussian distributed variables . the critical point is determined by the occurrence of a non - trivial solution . this yields , in our notation , @xmath19 which is asymptotically correct for weak noise , see below . desai and zwanzig @xcite described the system by a self - consistent dynamic mean field theory . they evaluated numerically the correct phase transition condition and observed that the critical point deviates from the result of a gaussian approximation . dawson @xcite correctly claimed existence and uniqueness of the critical parameter . we show that he used a wrong argument and give a different proof of uniqueness . he proved in the limit of infinitely many oscillators that the fluctuations at the critical point are non - gaussian and occur at a slower time scale than the noncritical fluctuations . furthermore , he observed that for @xmath20 the critical point can be computed up to a quadrature . van den broeck and collaborators @xcite also gave numerical results for the parameter dependence of @xmath0 in mean field theory and compared this with simulations for a system with nearest neighbor coupling in @xmath21 . implicitly they state that for strong coupling ( @xmath22 ) the critical point is @xmath23 , cf . also @xcite . shiino @xcite proved a @xmath24-theorem showing thus that the stationary state is asymptotically reached for long times and analyzed the stability of the trivial and the symmetry breaking solutions . the harmonic coupling between the constituents of an array was introduced by kometani and shimizu @xcite to describe the interaction between myosin and actin filaments in muscle contraction . in this context the variables @xmath25 represent velocities rather than coordinates and the system can be regarded as an early example of a canonical - dissipative system @xcite , which have been proposed to describe , e.g. , swarm dynamics @xcite . the harmonic coupling between nearest neighbors of a regular lattice can be conceived as discretization of the laplace operator . thus , in the continuum limit there is a relation to mean field solutions of a class of models described by stochastic partial differential equations , see e.g. @xcite . there is further a relation to the discretized version of the @xmath26-ginzburg - landau model , in different context also known as soft - spin ising model . for example , in @xcite the authors studied spin glasses , where the coupling strength for each pair of coordinates @xmath25 , @xmath27 is an independent gaussian distributed random variable . noise induced phenomena in a double - well potential are still a topic of recent research . for example , in @xcite the authors investigated a fokker - planck equation driven by dynamical constraints , modeling many - particle storage systems . for a similar system driven by multiplicative noise instead of additive noise also continuous phase transitions the critical exponent of the order parameter @xmath28 undergoes a transition from a constant ( though non - mean field ) value towards a parameter - dependent value when @xmath29 exceeds a threshold @xcite . also higher moments @xmath30 show such transitions @xcite . it is a natural question whether this behavior is robust against additive noise , since in natural systems additive noise is apparently unavoidable . although many papers study systems with additive and multiplicative noise ( for early refs . see @xcite ) this question has not been explicitly addressed . to determine critical exponents it is advantageous to know analytically the parameter dependence of the critical manifold . besides our general interest this is an additional motivation to study the present system . the paper is organized as follows . in sec . [ sec : model ] we reformulate the model in the fokker - planck picture and explain the self - consistent mean field approach which becomes exact in the limit of infinite system size . we show that the self - consistency condition is equivalent with the stationarity condition for the center of mass variable . in sec . [ sec : critmnf ] we prove the existence of a well - behaved critical manifold in the parameter space which separates the regime with @xmath31 from the regime with broken symmetry . the critical manifold is determined by an implicit integral equation , the phase transition condition ( ptc ) . to show its wellbehavedness it is necessary to know that the critical parameter @xmath12 is bounded by values which are asymptotically reached in the limits of strong and weak noise , respectively . the wellbehavedness of the critical manifold allows to reduce the number of parameters by rescaling and implies that @xmath32 is only a function of the ratio @xmath33 . thus it exhibits the same behavior for strong coupling as for weak noise , and for weak coupling as for strong noise . exploiting that the critical manifold is well - behaved , together with exact relations of moments on this manifold we derive an ordinary differential equation ( ode ) for @xmath32 as a function of @xmath34 . to determine a solution for small and large values of @xmath34 we use a series expansion . the results from the ode are the same as from an evaluation of the ptc using the laplace method but easier to obtain . the asymptotic behavior of the series agrees with that of the numerical solution of the ptc . we were not able to prove convergence of these series , but for small @xmath34 a pad approximant agrees well with the numerical solution . we also give results for a series expansion around @xmath35 which is analytically treatable . in sec . [ sec : critbeh ] we determine the behavior of the order parameter and the susceptibility near the critical manifold . thanks to the boundedness of @xmath0 the existence of a tricritical point is excluded and the critical exponents are the mean field exponents , as expected . we derive the amplitudes of the power laws of order parameter and susceptibility in closed form in terms of @xmath0 . the amplitude ratio of the susceptibilities just above and below the critical point is universal as expected by analogy with equilibrium second order phase transitions . several detailed calculations and technical discussions are deferred to the appendices . in appendix [ sec : moments ] recursion relations for the moments are given . in appendix [ sec : proofac ] we derive in some detail the bounds of @xmath0 . in appendix [ sec : laplace ] the phase transition condition is evaluated by the laplace method up to the third leading order in the limits of weak and strong noise , respectively . the fokker - planck equation corresponding to eq . is @xmath36 p(\mathbf{x } , t)\big\}. \label{eq : fpldim}\end{aligned}\ ] ] integrating over all coordinates but @xmath37 we find @xmath38 p_1^l(x_1 , t)\big\ } , \label{eq : fpstatldim}\end{aligned}\ ] ] where @xmath39 is the conditional expectation value of @xmath40 given @xmath37 and @xmath41 is the probability distribution of @xmath37 for the system with @xmath1 constituents . for @xmath42 we assume independence of coordinates , i.e. @xmath43 and find the one - particle distribution as the solution of @xmath44 p(x , t ) \big\ } , \label{eq : statsol}\end{aligned}\ ] ] where we wrote @xmath45 instead of @xmath46 . besides eq . there is a more rigorous approach . following @xcite , we introduce the empirical distribution defined by @xmath47}(x_i(t)),\end{aligned}\ ] ] which is an probability - distribution - valued random variable . provided that for @xmath48 all coordinates are independent and identically distributed with distribution @xmath49 , it is a result of @xcite that in the limit @xmath42 both distributions converge weakly to @xmath45 , which is the unique solution of eq . with initial condition @xmath50 . hence , for large @xmath1 we can interpret @xmath45 either as the empirical distribution of the many coordinates or as the probability distribution of a single coordinate . for an infinitely large system the harmonic coupling in eq . of the site @xmath51 to all other sites becomes @xmath52 where @xmath53 is the mean field exerted by all other sites to the system at site @xmath51 . it was shown in @xcite that the time dependent solution of eq . approaches the stationary solution for long times . in the stationary case @xmath54 is constant and can be considered as a parameter . we find the stationary solution of eq . @xmath55\label{eq : statdistr},\end{aligned}\ ] ] with normalization @xmath56 , \label{eq : norm}\end{aligned}\ ] ] satisfying @xmath57 the mean field should solve the self - consistency equation @xmath58 obviously the mean field is related to the center of mass coordinate @xmath59 obeying the langevin equation @xmath60 for the infinite system the noise term vanishes due to the law of large numbers and eq . becomes @xmath61 in the stationary case @xmath62 . this is equivalent to the self - consistency condition as can be seen writing @xmath63 \right\ } \notag \\ & \ ; \ ; \ ; \ ; \ ; \times \exp \left [ \frac{2d}{\sigma^2}\big(-\frac{1}{2}x^2+mx \big ) \right ] \notag \\ & = d\left(\langle x\rangle - m\right ) = 0 . \label{eq : limr}\end{aligned}\ ] ] the second equality follows by partial integration and observing that the boundary term vanishes since @xmath64 decays exponentially fast for @xmath65 . we introduce the function @xmath66 satisfying the symmetry @xmath67 the self - consistency equation now reads @xmath68 which obviously has a solution @xmath69 . for @xmath70 the curvature of @xmath71 is negative as shown in @xcite using a simple version of the griffiths - hurst - sherman inequality @xcite . if the derivative of @xmath72 at @xmath69 is larger than one , there exists exactly one positive solution to eq . then , by symmetry we also have a negative solution . otherwise @xmath69 is the only solution . shiino @xcite showed that the solutions with @xmath73 are stable if they exist whereas the @xmath69 solution is unstable in that case . the phase transition condition @xmath74 can be written as @xmath75 where @xmath76 } { \int_{-\infty}^{\infty}{\mathop{\!\mathrm{d}\!}}x \ , \exp \left [ \frac{2}{\sigma^2 } ( \frac{a - d}{2}x^2 - \frac{1}{4 } x^4 ) \right ] } -1 . \label{eq : varphi}\end{aligned}\ ] ] this defines the critical manifold in the space spanned by @xmath77 . we observe immediately that on the critical manifold we have @xmath78 in the following we show that the critical manifold is well - behaved : given any two of the parameters @xmath79 or @xmath80 , there exists a unique value , the critical value , of the third parameter which solves eq . . for @xmath81 there is no solution to . therefore we consider @xmath82 and furthermore @xmath83 since there are only contributions in @xmath84 , negative @xmath80 is equivalent . @xmath85 is continuous and continuously differentiable in @xmath79 and @xmath80 . @xmath85 is even @xmath86 on @xmath87 . by asymptotic evaluation of the integrals in we find @xmath88 because of continuity , for every @xmath89 there exists an @xmath9 satisfying eq . . since @xmath90 this solution is unique , @xmath91 . the critical parameter @xmath0 is bounded by @xmath92 as proven in appendix [ sec : proofac ] . these are the best possible bounds since the upper and the lower bound are asymptotically reached for weak or strong noise , respectively , as shown in sects . [ subsec : wn ] and [ subsec : sn ] below . since @xmath93 we consider @xmath85 on @xmath94 . because of eq . we can apply the implicit function theorem : there is locally around a solution of eq . a unique @xmath86 function @xmath95 satisfying @xmath96 since there exists a unique solution @xmath97 to eq . for every @xmath89 the function @xmath98 is globally uniquely defined on @xmath99 and is in @xmath86 . looking at relation we see furthermore that @xmath100 assumes all values in @xmath101 even if one of the variables @xmath4 or @xmath80 is fixed to some value . we consider a fixed @xmath102 and exploit that @xmath103 for all @xmath4 . therefore @xmath104 solving for @xmath105 and inserting the relations and we find @xmath106 where the negativity is guaranteed by inequality . because of the monotonicity and surjectivity of @xmath107 there exists an inverse function @xmath108 for all @xmath109 , i.e. , we have the function @xmath110 satisfying @xmath111 analogously , for fixed @xmath112 we calculate @xmath113 \notag \\ & - \frac{2d_0}{\sigma^2 } \left [ ( f_a - d_0)\langle x^2\rangle^2 - \frac{1}{2}\langle x^2 \rangle \langle x^4 \rangle \right ] \notag \\ & + \frac{2d_0}{\sigma^4}\left ( \langle x^4 \rangle - \langle x^2 \rangle^2 \right ) \frac{\partial f_a}{\partial \left ( { 1}/{\sigma^2}\right)}. \label{eq : calcdsigmafa}\end{aligned}\ ] ] inserting the expressions , and we find @xmath114 where inequality was used . hence we find @xmath115 from the monotonicity there follows the existence of a function @xmath116 satisfying @xmath117 in analogy to the previous case . + hence , given any two parameters , positive , there exists a unique critical value of the third parameter denoted by @xmath118 , @xmath119 , or @xmath120 . because of the monotonicity of @xmath85 with respect to @xmath9 and of @xmath118 with respect to @xmath4 and @xmath80 we conclude that there exists a pair of nonzero solutions @xmath121 of eq . if and only if one of the following equivalent conditions is satisfied @xmath122 with an arbitrary @xmath123 we can rescale variables and parameters as @xmath124 which leads to a system of langevin equations equivalent to eqs . @xmath125 for @xmath126 , with @xmath127 hence we have not three but only two independent parameters . in a similar way , also the general case with a coefficient of the cubic term in eqs . can be treated , ending with only two independent parameters . we observe that @xmath128 given by eq . is invariant under this rescaling . this allows the following argument . we set @xmath129 such that @xmath130 and find @xmath131 . therefore @xmath132 fig . [ fig : log - log - plot ] shows @xmath32 as a function of @xmath133 in a log - log plot for different values of @xmath4 as obtained by numerical solution of eq . . it confirms that @xmath32 depends only on the ratio of @xmath80 and @xmath4 as predicted by eq . . therefore @xmath134 exhibits the same asymptotic behavior for strong coupling as for weak noise , and for weak coupling as for strong noise , respectively . in dependence of noise strength @xmath34 as a log - log plot . here the equation was numerical solved for @xmath135 @xmath136 , @xmath137 @xmath138 , @xmath139 @xmath140 , and @xmath141 @xmath142 , each for several values of @xmath80 . the solid line describes the function @xmath143 , the limit of weak noise , the dashed line the function @xmath144 , the asymptote for strong noise.[fig : log - log - plot ] ] dawson @xcite uses a scaling where @xmath145 and claims that , given @xmath146 , the critical noise strength @xmath147 is bounded by @xmath148 , cf . @xmath149 there . we show that this assertion is not true . from eqs . and we have @xmath150 , where @xmath116 is defined by eq . . we now choose @xmath141 and @xmath151 , where @xmath152 can have any positive value and @xmath118 is defined by eq . . since @xmath153 it holds that @xmath154 which can be chosen beyond the bonds asserted in @xcite . however the property that for any @xmath155 there exists a unique critical @xmath147 remains true , as we have shown in sec . [ subsec : cmgeneral ] . at @xmath156 the stationary probability density @xmath157 for the coordinate of an arbitrary constituent is qualitatively different for weak and for strong noise , cf . [ fig : symbreak ] . for weak noise @xmath157 is approximately a gaussian centered at @xmath16 , whereas for strong noise it is the sum of two equally weighted narrow peaks located at @xmath158 . for @xmath17 the symmetry is broken in different ways . in the weak noise limit @xmath157 is still a gaussian but centered at @xmath18 , whereas in the strong noise limit the two narrow peaks stay at @xmath158 but their weights become unequal such that @xmath18 also in this case . the critical parameter @xmath12 is bounded between its limit values for strong and weak noise respectively , as will be shown rigorously in appendix [ sec : proofac ] . exploiting that at criticality the even moments are explicitly known we give a simple handwaving argument which leads to the correct leading behavior for weak and strong noise . especially we use eq . , that is @xmath159 and , cf . appendix [ sec : moments ] , @xmath160 . the fourth cumulant , the kurtosis , of a gaussian is zero and therefore @xmath161 . comparing this with the above expressions we obtain for weak noise in leading order @xmath162 , which is the upper bound in the inequality furthermore , for a symmetric probability density of two narrow peaks , the variance of @xmath163 is approximately zero and therefore @xmath164 . comparing with the above expressions we obtain for strong noise in leading order @xmath165 , which is the lower bound for @xmath0 in . in eqs . and we have ordinary differential equations for @xmath0 as a function of @xmath4 and @xmath166 . by substituting @xmath167 and @xmath168 we find from either of these equations @xmath169 in the next two sections we will systematically study the behavior of the critical parameter for weak and strong noise using eq . . the same results can be obtained by asymptotic evaluation of the integrals in the phase transition condition as shown in appendix [ sec : laplace ] . because of inequality it holds that @xmath170 that means we can continuously extend @xmath171 to @xmath172 . to obtain the asymptotic behavior of @xmath171 for weak noise we make the ansatz @xmath173 inserting this series in the differential equation and comparing coefficients in powers of @xmath174 we find the recursion relation @xmath175a_{n - k+1}a_{k } \label{eq : coeffweaknoise}\end{aligned}\ ] ] for @xmath176 with initial condition @xmath177 . for @xmath178 the three leading terms as @xmath179 are @xmath180 which coincides with the result in appendix [ sec : laplacewn ] . we observe that the upper bound in is reached asymptotically . in fig . [ fig : weaknoiselin ] we see @xmath178 as a function of @xmath174 . the series up to @xmath181 coincides with the numerical solution of eq . only for very small values of @xmath174 and taking into account more terms does not seem to improve the result for larger values of @xmath174 . the figure shows also the pad approximant @xmath182 which coincides much better with the numerical results . the pad approximant @xmath183 is a rational function @xmath184 , where @xmath185 and @xmath186 are polynomials of degree @xmath187 and the taylor series of @xmath183 agrees with the series up to @xmath188 @xcite . as a function of @xmath174 . the point with @xmath189 is known exactly @xmath190 . the figure shows the numerical solution @xmath191 of eq . , the series for weak noise up to @xmath181 ( thin solid line , see also insert ) , the corresponding pad approximant @xmath192 ( thick solid line ) , and the series around the exactly known point @xmath190 up to 10th order ( dashed line).[fig : weaknoiselin ] ] we have not been able to prove convergence of the series near @xmath193 . it is clear by its definition via the implicit function that @xmath194 is an analytic function for any positive @xmath80 , but at @xmath195 we do nt know . nevertheless , eq . has a meaning as it correctly describes the asymptotic behavior of @xmath178 obtained by numerically solving the ptc for @xmath179 as @xmath196 where the symbol @xmath197 means @xmath198 , cf . [ fig : log - log - plot ] . the coefficients of the higher order terms in eq . give systematic corrections in the sense that @xmath199 and @xmath200 , cf . [ fig : numasympwn ] . for small @xmath174 . the symbols show the results for @xmath201 and @xmath202 obtained from the numerical solution of eq . . the predicted behavior @xmath203 ( solid line ) and @xmath204 ( dashed line ) is confirmed.[fig : numasympwn ] ] in the limit of strong noise @xmath205 , again motivated by , we use a series ansatz for @xmath178 as @xmath206 , @xmath207 inserting in eq . and comparing coefficients in powers of @xmath174 leads to @xmath208 and to the backward recursion @xmath209 for @xmath210 with initial condition @xmath211 . the three leading terms for @xmath178 as @xmath212 are @xmath213 the same coefficients are obtained via laplace s method , cf . appendix [ sec : laplacesn ] . in the case of strong noise we reach the lower bound in asymptotically . in fig . [ fig : strongnoiseinv ] we see @xmath214 as a function of @xmath215 , where the numerical solution of eq . is compared with the series . the series agrees with the numerical results only for very small values of @xmath215 . here , also the pad approximants do not work as well as for weak noise , since they have many poles within the region of interest . as a function of @xmath215 . the figure shows the numerical solution @xmath191 of eq . compared with the series for strong noise ( solid line ) up to @xmath216.[fig : strongnoiseinv ] ] as for weak noise we have not been able to prove convergence of the series . nevertheless , eq . describes the asymptotics of @xmath178 for @xmath212 as @xmath217 , cf . [ fig : log - log - plot ] , and the higher order coefficients in eq . give systematic corrections in the sense @xmath218 , and @xmath219 , cf . [ fig : numasympsn ] . for large @xmath174 . the symbols show the results for @xmath220 and @xmath221 obtained from the numerical solution of eq . . the predicted behavior @xmath222 ( solid line ) and @xmath223 ( dashed line ) is confirmed.[fig : numasympsn ] ] dawson @xcite observed that , in our notation , considering only the sub - manifold of the critical manifold defined by the condition @xmath35 , it is possible to obtain an explicit expression for the critical value of @xmath9 as a function of @xmath80 . substituting @xmath224 in eq . and solving eq . for @xmath9 under the restriction that @xmath35 yields the critical parameter @xmath225 in terms of the rescaled parameters @xmath226 and @xmath227 this reads @xmath228 using the ansatz @xmath229 we find with the ode the recursion formula @xmath230\big\ } , \end{aligned}\ ] ] for @xmath231 with initial conditions @xmath232 in fig . [ fig : weaknoiselin ] we see good agreement between the numerical solution of the phase transition condition and the series up to the tenth order term . to calculate the behavior of the order parameter @xmath54 for @xmath9 close to the critical value @xmath0 it is convenient to introduce the notation @xmath233 . \nonumber\end{aligned}\ ] ] @xmath71 defined in eq . can be expressed as @xmath234 and the @xmath235th moment of the probability density @xmath236 as @xmath237 expanding the right hand side of eq . for small @xmath54 , according to we obtain the self - consistency equation @xmath238 where @xmath239 . equation has always the trivial solution @xmath69 . for @xmath240 there is a pair of nontrivial real solutions @xmath241 since the denominator of the radicand is always positive as proven in appendix [ sec : proofac ] and the numerator of the radicand is positive if and only if @xmath242 . this follows from the monotonicity of the second moment as a function of @xmath9 , cf . hence the expansion is sufficient to determine the leading behavior of @xmath54 close to the critical point and we can exclude the existence of a tricritical point . we now expand the right hand side of eq . for small @xmath243 , exploiting that at @xmath91 all even moments can be determined recursively from eq . , see appendix [ sec : moments ] . inserting eqs . , and in eq . yields in leading order @xmath244 hence we have found the typical mean field exponent @xmath13 and an analytic expression for the amplitude in terms of the critical parameter . inserting @xmath0 in the limit of weak noise from eq . leads to @xmath245 in the limit of strong noise we obtain with eq . @xmath246 in fig . [ fig:2 ] these analytical results are compared for a typical parameter setting with the numerical evaluation of the self - consistency equation . for and @xmath247 determined by numerical evaluation of eq . compared to the analytical results from eqs . and for weak noise ( dashed line ) and for strong noise ( solid line ) . coupling strength @xmath141.[fig:2 ] ] in this section we observe that the susceptibility is diverging at the critical point as @xmath248 with the amplitudes @xmath249 and @xmath250 for @xmath17 and @xmath251 , respectively . we explicitly calculate the amplitudes @xmath252 in terms of @xmath0 and find a universal ratio between them . the whole procedure , as well as the results , are in complete analogy to equilibrium thermodynamics . however the calculation is explicitly possible here , and up to our knowledge it has not been done in this context before . we introduce an external field @xmath253 in eq . @xmath254 the susceptibility is defined as the response of the system to a small external field @xmath255 we investigate the susceptibility close to the critical point . in analogy to eq . we find the self - consistency equation for small @xmath54 and @xmath253 @xmath256 taking the derivative with respect to @xmath253 at @xmath257 we obtain @xmath258 at the critical point we find @xmath259 which can be satisfied only asymptotically by @xmath260 . below the critical point we have @xmath69 and . therefore becomes @xmath261 for small @xmath262 we find the leading behavior of the susceptibility close to the critical point using eq . @xmath263 for @xmath264 we obtain in leading order with the help of eqs . , and @xmath265 since @xmath266 we have @xmath267 which is universal , i.e. not depending on parameters . in this paper we have proved an upper and a lower bound for the critical parameter @xmath0 . these bounds are optimal since they are asymptotically reached for weak and for strong noise , respectively . we found an ordinary differential equation describing the critical point @xmath32 as a function of @xmath34 , which allows to explicitly give a recursion formula for all coefficients of the asymptotic expansion of @xmath0 for weak and strong noise as well as for an expansion around a special point , where @xmath0 is known exactly . in the limits of weak and strong noise the mechanism of symmetry breaking is qualitatively different . for weak noise and @xmath9 close to @xmath0 the stationary distribution of the coordinates @xmath157 is a gaussian . below the critical point the gaussian is centered around zero . above , for @xmath17 , the gaussian is shifted in positive or negative direction , the symmetry is broken . for strong noise and @xmath9 close to @xmath0 @xmath157 consists of two narrow peaks located symmetrically with respect to zero . the symmetry is broken such that for @xmath17 one of the peaks gains a larger weight than the other . we have proved that the critical manifold is well behaved , that is if two of the three positive parameters @xmath79 , and @xmath80 are given there exists a unique critical value of the third . the proof hinges on the knowledge of the above mentioned boundaries of @xmath0 . the wellbehavedness of the critical manifold allows to reduce the number of parameters and implies certain scaling properties . for example , the ratio @xmath32 depends on noise strength and coupling constant only as a function of @xmath268 , and the limits of weak noise or strong coupling and strong noise or weak coupling are equivalent . we further have determined the critical behavior of order parameter and susceptibility . as well known , they follow as a function of @xmath269 power laws with the mean field exponents . we have calculated the amplitude of the order parameter in terms of the critical parameter @xmath0 and explicitly in the limits of weak and strong noise and found for the amplitude ratio of the susceptibilities the universal law @xmath270 . it is a natural question whether for systems with higher order nonlinearity similar results can be obtained . it is further desirable to study the critical manifold of a system with both additive and multiplicative noise . thanks is due to klaus kroy for valuable discussions . r. k. thanks the imprs mathematics in the sciences for funding . we need the exact recursive relations between moments which are derived by dawson @xcite exploiting the ito - formula . to keep the paper self - contained we rederive these relations in our notation using a different argument . by partial integration of the right hand side of eq . we find @xmath272 \nonumber \\ = & \frac{1}{k+1 } \frac{2}{\sigma^2 } \times \notag \\ & \int_{-\infty}^{\infty } { \mathop{\!\mathrm{d}\!}}x \ ; \left ( -(a - d)x^{k+2 } + x^{k+4 } -dmx^{k+1 } \right ) \notag \\ & \times \exp\left [ \frac{2}{\sigma^2 } \left ( dmx + \frac{a - d}{2}x^2 ) - \frac{1}{4}x^4 \right ) \right]\end{aligned}\ ] ] for @xmath273 , @xmath274 . dividing by the normalization @xmath275 we obtain @xmath276 in fact eq . is also true for @xmath277 , then @xmath278 , cf . eq . . for @xmath14 we have @xmath279 , such that all odd moments are zero by symmetry . in that case the recursion formula simplifies for all even moments to @xmath280 with @xmath281 . at @xmath91 we have already calculated @xmath282 in eq . and know @xmath283 since @xmath284 is normalized . therefore it is possible to calculate all even moments using . especially one finds @xmath285 for small @xmath286 , i.e. above but close the critical point , we obtain @xmath287 using eqs . and . since the variance of @xmath163 is larger than zero for any extended distribution we have @xmath288 at @xmath289 we obtain with eqs . and @xmath290 which gives the lower bound @xmath291 to obtain the upper bound we use the inequality @xmath292 which states that the kurtosis of @xmath271 is negative at @xmath91 , see below . again with eqs . and we find @xmath293 which gives @xmath294 to show we substitute @xmath295 such that the new coordinate @xmath296 has variance one . we denote the stationary distribution of the new coordinate by @xmath297 . the inequality in the @xmath271 coordinate is equivalent to the same expression in the new coordinate @xmath296 @xmath298 now we compare the distribution @xmath297 with the gaussian distribution with variance one denoted by @xmath299 . in the following we only consider the critical point @xmath91 where both distributions have zero mean and are symmetric under the transformation @xmath300 . we look at the intersection of both curves and distinguish two cases . there can be either two or four intersecting points . for two intersection points we use the theorem @xcite : if a symmetric zero - mean probability distribution @xmath297 intersects with the standard normal distribution @xmath299 in exactly two points @xmath301 , then @xmath302 for all @xmath303 if and only if the kurtosis of @xmath297 is negative . in the present situation @xmath297 decays as @xmath304 , as @xmath305 and therefore @xmath302 for large enough @xmath306 . hence we can apply the theorem and is satisfied . in the case of four intersection points of @xmath297 and @xmath299 we use the following theorem from @xcite : suppose two probability densities @xmath299 and @xmath297 with zero mean and the same variance are given . let @xmath307 be their respective third and fourth moments . then we have a sufficient condition for : there should exist four abscissae @xmath308 such that + @xmath309 @xmath310 and ( iii ) @xmath311 and @xmath312 are not both strictly positive or both strictly negative . in the present case we have @xmath313 . furthermore @xmath314 and @xmath315 since both @xmath299 and @xmath297 are even functions . hence @xmath316 and we can apply the theorem . therefore @xmath317 . there . by numerical evidence he conjectured that there is strict inequality . we prove this conjecture . ] we can follow the lines in @xcite to prove even strict inequality . consider the function @xmath318 . for any @xmath319 the functions @xmath320 and @xmath321 have either the same sign or at least one of them is zero . thus @xmath322 . in the present situation , since both functions are continuous and zero only if @xmath323 , there exists @xmath324 such that @xmath325 for @xmath326 , @xmath327 $ ] for @xmath328 . hence we have @xmath329 expanding the polynomial @xmath321 and performing the integral in we find @xmath330 where we used that odd moments of @xmath297 and @xmath299 are zero and that both distributions have variance one and are normalized . since the kurtosis of a gaussian is zero , by we follow that is true , which completes the proof of the inequality . in the following we solve the ptc asymptotically for strong and for weak noise using laplace s method to evaluate the integrals in eq . . independently of the simpler method using the ode we obtain the same results . when using laplace s method we explicitly use the different shapes of the distribution @xmath157 for weak and strong noise , discussed in section [ subsec : scaling ] . in the next section we explicitly calculate the integral @xmath331 for large @xmath332 , where @xmath333 is a polynomial of degree four . in the succeeding two sections we apply the results of this calculation to asymptotically solve the ptc for weak and for strong noise . we evaluate an integral of the form @xmath334 for large @xmath332 applying the laplace method @xcite . we denote @xmath335 suppose that @xmath333 has its global maximum at @xmath16 . expanding around zero yields @xmath336 now we change the integration boundaries in to @xmath337 and @xmath338 for some @xmath339 . doing so , we make only exponentially small errors for @xmath340 , @xmath341 inserting the expansion yields @xmath342 \notag \\ & \times \exp\left(\lambda \frac{1}{2 } \phi^{(2)}_0 x^2\right ) . \label{eq : intlp3}\end{aligned}\ ] ] we are interested in the case where @xmath343 is a polynomial of fourth order . we expand @xmath344 for small @xmath271 and neglect terms which give either a vanishing contribution or contributions of the order @xmath345 when the gaussian integrals are performed . in view of @xmath346 = \frac{(2n-1)!!}{2^n}\frac{1}{\alpha^n}\sqrt{\frac{\pi}{\alpha } } \label{eq : gauss4},\\ & \int_{-\infty}^{\infty } { \mathop{\!\mathrm{d}\!}}x \ , x^{2n+1}\exp\left [ -\alpha x^2 \right ] = 0 \end{aligned}\ ] ] we should take into account @xcite @xmath347 now we plug the expansion into eq . and extend the integration range to @xmath348 . again the change of the integration range produces only an exponentially small error for @xmath349 . evaluating the gaussian integrals leads to @xmath350 . \notag\end{aligned}\ ] ] in the limit of weak noise we set @xmath351 and use the ansatz @xmath352 which is equivalent to eq . . we define @xmath353 . \label{eq : integr1}\end{aligned}\ ] ] and rewrite the ptc as @xmath354 to express the right hand side of eq . in a controllable way as a series in @xmath84 . + in view of the limit it is clear that @xmath355 is locally uniformly converging to @xmath356 as @xmath357 such that we can use laplace s method @xcite to evaluate the integral . for @xmath333 defined by we find @xmath358 and @xmath359 inserting eq . into the ptc leads to @xmath360 . \notag\end{aligned}\ ] ] inserting the ansatz into eq . and expanding in powers of @xmath7 leads to @xcite @xmath361 comparing coefficients in powers of @xmath84 , we find @xmath362 which is equivalent to the result . in the limit of strong noise we set @xmath363 and use the ansatz @xmath364 which is equivalent to eq . we substitute and define @xmath365 \label{eq : intstrongnoise}\end{aligned}\ ] ] with the function @xmath366 that has its local maxima at @xmath367 by the inequality we find that @xmath368 . analogously to the weak noise case we rewrite the ptc as @xmath369 for strong noise , @xmath370 is locally uniformly converging to @xmath371 we will perform laplace approximation around the positive maximum , an equal contribution comes from the other maximum . instead of @xmath372 in we use @xmath373 with @xmath374 by with @xmath375 instead of @xmath376 we obtain up to @xmath377 @xmath378 \notag \\ = & \ , \sigma^2 \frac{\tilde{a}^2}{2 } + \frac{1}{2 } \ln ( 2\pi ) - \frac{1}{2}\ln ( \sigma^{2}\tilde{a } ) \notag \\ & + \ln \bigg [ 1 + \sigma^{-2 } \frac{3}{8 } \frac{1}{\tilde{a}^2 } + \sigma^{-4 } \frac{105}{128 } \frac{1}{\tilde{a}^4 } \bigg ] , \label{eq : lnistrnoise}\end{aligned}\ ] ] with @xmath379 inserting into the ptc we obtain @xmath380 . \notag\end{aligned}\ ] ] inserting and furthermore the ansatz for @xmath0 eq . we expand in powers of @xmath381 and obtain @xcite @xmath382 comparing coefficients in powers of @xmath7 leads to @xmath383 which is equivalent to eq . . 25ifxundefined [ 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 10.1103/revmodphys.79.829 [ * * , ( ) ] @noop _ _ ( , , ) @noop _ _ ( , , ) link:\doibase 10.1016/s0031 - 8914(40)90098 - 2 [ * * , ( ) ] link:\doibase 10.1007/bf01013146 [ * * , ( ) ] link:\doibase 10.1007/bf01020331 [ * * , ( ) ] link:\doibase 10.1007/bf01010922 [ * * , ( ) ] link:\doibase 10.1103/physreve.49.2639 [ * * , ( ) ] link:\doibase 10.1103/physreve.54.6918 [ * * , ( ) ] link:\doibase 10.1016/0375 - 9601(85)90345 - 7 [ * * , ( ) ] link:\doibase 10.1103/physreva.36.2393 [ * * , ( ) ] @noop _ _ ( , , ) link:\doibase 10.1103/physreve.64.021110 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.98.230601 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.47.359 [ * * , ( ) ] link:\doibase 10.1137/110851882 [ * * , ( ) ] link:\doibase 10.1103/physreve.65.046110 [ * * , ( ) ] link:\doibase 10.1103/physreve.72.056102 [ * * , ( ) ] link:\doibase 10.1103/physreve.75.031105 [ * * , ( ) ] link:\doibase 10.1007/bf01608495 [ * * , ( ) ] @noop _ _ ( , , ) link:\doibase 10.1109/97.803435 [ * * , ( ) ] http://www.jstor.org/stable/2980484 [ * * , ( ) ] @noop _ _ ( , , )

the theory of support varieties for finitely generated modules over the group algebra of a finite group began back in the nineteen eighties with the work of alperin and evens @xcite , avrunin and scott @xcite , carlson @xcite , among others . an essential ingredient in its development was carlson s anticipation that for elementary abelian @xmath1-groups the cohomological definition of support , which takes its roots in quillen s fundamental work on mod @xmath1 group cohomology @xcite , gave the same answer as the `` rank '' definition through restriction to cyclic shifted subgroups . to deal with infinitely generated modules , benson , carlson and rickard @xcite found that they had to introduce cyclic shifted subgroups for generic points of subvarieties , defined over transcendental extensions of the field of coefficients . correspondingly , all the homogeneous prime ideals in the cohomology ring are involved , not just the maximal ones . this enabled them to classify the tensor ideal thick subcategories of the stable category @xmath2 of finitely generated @xmath3-modules @xcite . it seemed plausible that one should be able to use modifications of the same techniques to classify the tensor ideal localising subcategories of the stable category @xmath4 of all @xmath3-modules , but there were formidable technical obstructions to realising this , and it was not until more than a decade later that this was achieved by the first three authors of this paper @xcite , using a rather different set of ideas than the ones in @xcite . the basic strategy was a series of reductions , via changes of categories , that reduced the problem to that of classifying the localising subcategories of the derived category of differential graded modules over a polynomial ring , where it was solved using methods from commutative algebra . a series of papers @xcite , established machinery required to execute this strategy . in this paper we give an entirely new , and conceptually different , proof of the classification of the tensor ideal localising subcategories of @xmath4 from @xcite . it is closer in spirit to the proof of the classification of the tensor ideal thick subcategories of @xmath2 from @xcite , and rooted essentially in linear algebra . the crucial new idea is to introduce and study @xmath0-cosupports for representations . the inspiration for this comes from two sources . the first is the theory of @xmath0-points developed by friedlander and the fourth author @xcite as a suitable generalisation of cyclic shifted subgroups . whereas carlson s original construction only applied to elementary abelian @xmath1-groups , and required an explicit choice of generators of the group algebra , @xmath0-points allow for a `` rank variety '' description of the cohomological support for any finite group scheme ; see ( * ? ? ? * theorem 3.6 ) and section [ se : pi - points ] of this paper . based on this , in @xcite the @xmath0-support of a module over a finite group scheme @xmath5 defined over a field @xmath6 is introduced and used to classify the tensor ideal thick subcategories of @xmath2 ( with an error corrected in a sequel @xcite to this paper ) . but the tensor ideal localising subcategories of @xmath4 remained inaccessible by these techniques alone , even for finite groups . what is required is a @xmath0-version of the notion of cosupports introduced in @xcite . the relevance of @xmath0-cosupport is through the following formula for the module of homomorphisms , proved in section [ se : support - and - cosupport ] . for any finite group scheme @xmath5 over @xmath6 , and @xmath3-modules @xmath7 and @xmath8 , there is an equality @xmath9 to be able to apply this formula to cohomological cosupport developed in [ 9 ] one needs to identify the two notions of cosupport . our strategy for making this identification is to prove that @xmath0-cosupport detects projectivity : a @xmath3-module is projective if and only if its @xmath0-cosupport is empty . the desired classification result would then follow from general techniques developed in @xcite . in section [ se : finite - groups ] we prove such a detection theorem for projectivity for _ finite groups_. besides yielding the desired classification theorem for @xmath4 from @xcite , it implies that cohomological support and cosupport coincide with @xmath0-support and @xmath0-cosupport , respectively . this remarkable fact is a vast generalisation of carlson s original anticipation . the different origins of the two notions are reflected in the fact that phenomena that are transparent , or at least easy to detect , for one may be rather opaque and difficult to verify for the other . see section [ se : applications ] for illustrations . the corresponding detection theorem for arbitrary finite group schemes has turned out to be more challenging , and is dealt with in the sequel to this paper @xcite , using a different approach , where again @xmath0-support and @xmath0-cosupport play a crucial role . this brings us to the second purpose of this paper : to lay the groundwork for the proof in @xcite . for this reason parts of this paper are written in the language of finite group schemes . however , we have attempted to present it in such a way that the reader only interested in finite groups can easily ignore the extra generality . this section summarises basic property of modules over affine group schemes ; for details we refer the reader to jantzen @xcite and waterhouse @xcite . let @xmath6 be a field . affine group scheme _ @xmath5 over @xmath6 is a functor from commutative @xmath6-algebras to groups , with the property that , considered as a functor to sets , it is representable as @xmath10 . the commutative @xmath6-algebra @xmath11 has a comultiplication coming from the multiplicative structure of @xmath5 , and an antipode coming from the inverse . this makes @xmath11 into a commutative hopf algebra called the _ coordinate algebra _ @xmath12 $ ] of @xmath5 . conversely , if @xmath12 $ ] is a commutative hopf algebra over @xmath6 then @xmath13,-)$ ] is an affine group scheme . this work concerns only affine group schemes so henceforth we drop the qualifier `` affine '' . a group scheme @xmath5 over @xmath6 is _ finite _ if @xmath12 $ ] is finite dimensional as a @xmath6-vector space . the @xmath6-linear dual of @xmath12 $ ] is then a cocommutative hopf algebra , called the _ group algebra _ of @xmath5 , and denoted @xmath3 . we identify modules over a finite group scheme @xmath5 with modules over its group algebra @xmath3 ; this is justified by ( * ? ? ? * i.8.6 ) . thus , we will speak of @xmath3-modules ( rather than @xmath5-modules ) , and write @xmath14 for the category of @xmath3-modules . let @xmath5 be a finite group scheme over a field @xmath6 . if @xmath15 is a field extension of @xmath6 , we write @xmath16 $ ] for @xmath17 $ ] , which is a commutative hopf algebra over @xmath15 . this defines a group scheme over @xmath15 denoted @xmath18 , and we have a natural isomorphism @xmath19 . for each @xmath3-module @xmath7 , we set @xmath20 viewed as @xmath21-modules . when @xmath15 or @xmath7 is finite dimensional over @xmath6 , these are related as follows . [ re : finite - dimensional ] for any @xmath3-module @xmath7 , there is a natural map of @xmath21-modules @xmath22 this is a bijection when @xmath15 or @xmath7 is finite dimensional over @xmath6 . then @xmath23 is a direct sum of copies of @xmath24 as a @xmath21-module , for @xmath25 is a direct sum of copies of @xmath15 , as a @xmath15-vector space . the assignments @xmath26 and @xmath27 define functors from @xmath14 to @xmath28 that are left and right adjoint , respectively , to restriction of scalars along the homomorphism of rings @xmath29 . the result below collects some basic facts concerning how these functors interact with tensor products and homomorphisms . in what follows , the submodule of @xmath5-invariants of a @xmath3-module @xmath7 is denoted @xmath30 ; see ( * ? ? ? * i.2.10 ) for the construction . [ le : invariants ] let @xmath5 be a finite group scheme over @xmath6 and @xmath15 an extension of the field @xmath6 . let @xmath7 and @xmath8 be @xmath3-modules . there are natural isomorphisms of @xmath21-modules : 1 . 2 . @xmath32 when @xmath7 is finite dimensional over @xmath6 . 3 . @xmath33 . there are also natural isomorphisms of @xmath15-vector spaces : 1 . 2 . @xmath35 . the isomorphisms in ( i ) , ( iii ) and ( iv ) are standard whilst ( v ) is the special case @xmath36 and @xmath37 of ( iv ) . the isomorphism in ( ii ) can be realised as the composition of natural maps @xmath38 where the last map is an isomorphism as @xmath7 is finite dimensional over @xmath6 . we recall some important classes of finite group schemes relevant to this work . [ ex : finite - groups ] a finite group @xmath5 defines a finite group scheme over any field @xmath6 . more precisely , the group algebra @xmath3 is a finite dimensional cocommutative hopf algebra , and hence its dual is a commutative hopf algebra which defines a group scheme over @xmath6 ; it is also denoted @xmath5 . a finite group @xmath39 is an _ elementary abelian @xmath1-group _ if it is isomorphic to @xmath40 , for some prime number @xmath1 . the integer @xmath41 is then the _ rank _ of @xmath39 . over a field @xmath6 of characteristic @xmath1 , there are isomorphisms of @xmath6-algebras @xmath42\cong k^{\times r}\quad\text{and}\quad ke\cong k[z_1,\dots , z_{r}]/(z_1^p,\dots , z_{r}^p).\ ] ] the comultiplication on @xmath43 is determined by the map @xmath44 and the antipode is determined by the map @xmath45 . [ ex : ga ] fix a positive integer @xmath41 and let @xmath46 denote the finite group scheme whose coordinate algebra is @xmath47= k[t]/(t^{p^r})\ ] ] with comultiplication defined by @xmath48 and antipode @xmath49 . there is an isomorphism of @xmath6-algebras @xmath50/(u_0^p,\dots , u_{r-1}^p).\ ] ] we note that @xmath46 is the @xmath41th frobenius kernel of the additive group scheme @xmath51 over @xmath6 ; see , for instance , ( * ? ? ? * i.9.4 ) [ ex : quasi - elementary ] following bendel @xcite , a group scheme over a field @xmath6 of positive characteristic @xmath1 is said to be _ quasi - elementary _ if it is isomorphic to @xmath52 . its group algebra structure is the same as that of an elementary abelian @xmath1-group . a finite group scheme @xmath5 over a field @xmath6 is _ unipotent _ if its group algebra @xmath3 is local . quasi - elementary group schemes are unipotent . also , the group scheme over a field of positive characteristic @xmath1 defined by a finite @xmath1-group is unipotent . in the rest of this paper @xmath5 denotes a finite group scheme defined over a field @xmath6 of positive characteristic @xmath1 . we recall the notion of @xmath0-points and basic results about them . the primary references are the papers of friedlander and pevtsova @xcite . a @xmath0-_point _ of @xmath5 , defined over a field extension @xmath15 of @xmath6 , is a morphism of @xmath15-algebras @xmath53/(t^p ) { \longrightarrow}kg_k\ ] ] that factors through the group algebra of a unipotent abelian subgroup scheme @xmath54 of @xmath55 , and such that @xmath21 is flat when viewed as a left ( equivalently , as a right ) module over @xmath56/(t^{p})$ ] via @xmath57 . it should be emphasised that @xmath54 need not be defined over @xmath6 ; see examples [ ex : klein ] . restriction along @xmath57 defines a functor @xmath58/(t^{p})\,.\ ] ] the result below extends ( * ? ? ? * theorem 4.6 ) , that dealt with @xmath59 . [ th : pi - equivalence ] let @xmath60/(t^p)\to kg_k$ ] and @xmath61/(t^p)\to lg_l$ ] be @xmath0-points of @xmath5 . then the following conditions are equivalent . 1 . for any finite dimensional @xmath3-module @xmath7 , the module @xmath62 is projective if and only if @xmath63 is projective . 2 . for any @xmath3-module @xmath7 , the module @xmath62 is projective if and only if @xmath63 is projective . 3 . for any finite dimensional @xmath3-module @xmath7 , the module @xmath64 is projective if and only if @xmath65 is projective . 4 . for any @xmath3-module @xmath7 , the module @xmath64 is projective if and only if @xmath65 is projective . the equivalence of ( i ) and ( ii ) is proved in ( * ? ? ? * theorem 4.6 ) . the equivalence of ( iii ) and ( iv ) can be proved in exactly the same way . ( i)@xmath66(iii ) since @xmath7 is finite dimensional , @xmath23 is a direct sum of copies of @xmath59 , by remark [ re : finite - dimensional ] . hence @xmath64 is projective if and only if @xmath62 is projective . the same is true of @xmath65 and @xmath63 . [ de : pi ] when @xmath0-points @xmath57 and @xmath67 satisfy the conditions of theorem [ th : pi - equivalence ] , they are said to be _ equivalent _ , and denoted @xmath68 . for ease of reference , we list some basic properties of @xmath0-points . [ re : pi - basics ] ( 1 ) let @xmath60/(t^p)\to kg_k$ ] be a @xmath0-point and @xmath69 a field extension of @xmath15 . then @xmath70/(t^p)\to lg_l$ ] is a @xmath0-point and it is easy to verify , say from condition ( i ) of theorem [ th : pi - equivalence ] , that @xmath71 . \(2 ) every @xmath0-point of a subgroup scheme @xmath72 of @xmath5 is naturally a @xmath0-point of @xmath5 . this follows from the fact that an embedding of group schemes always induces a flat map of group algebras . \(3 ) every @xmath0-point is equivalent to one that factors through a quasi - elementary subgroup scheme over the same field extension ; see ( * ? ? ? * proposition 4.2 ) . the cohomology of @xmath5 with coefficients in a @xmath3-module @xmath7 is denoted @xmath73 . it can be identified with @xmath74 . recall that @xmath75 is a @xmath6-algebra that is graded - commutative ( because @xmath3 is a hopf algebra ) and finitely generated , as was proved by friedlander and suslin ( * ? ? ? * theorem 1.1 ) . let @xmath76 denote the set of homogeneous prime ideals @xmath77 that are properly contained in the maximal ideal of positive degree elements . given a @xmath0-point @xmath60/(t^{p})\to kg_k$ ] we write @xmath78 for the composition of homomorphisms of @xmath6-algebras . @xmath79/(t^{p})}(k , k),\ ] ] where the second map is induced by restriction . by frobenius reciprocity and the theorem of friedlander and suslin recalled above , @xmath80/(t^{p})}(k , k)$ ] is finitely generated as a module over @xmath81 . since the former is nonzero , it follows that the radical of the kernel of the map @xmath82/(t^{p})}(k , k)$ ] is a prime ideal different from @xmath83 and hence that the radical of @xmath84 is a prime ideal in @xmath85 , different from @xmath86 . [ rem : generic - points ] fix a point @xmath87 in @xmath88 . there exists a field @xmath15 and a @xmath0-point @xmath89/(t^p){\longrightarrow}kg_k\ ] ] such that @xmath90 . in fact , there is such a @xmath15 that is a finite extension of the degree zero part of the homogenous residue field at @xmath87 ; see ( * ? ? ? * theorem 4.2 ) . it is shown in ( * ? ? ? * corollary 2.11 ) that @xmath68 if and only if there is an equality @xmath91 in this way , the equivalence classes of @xmath0-points are in bijection with @xmath92 . let @xmath5 be a finite group scheme over a field @xmath6 . taking a @xmath0-point @xmath57 to the radical of @xmath84 induces a bijection between the set of equivalence classes of @xmath0-points of @xmath5 and @xmath92 . we illustrate these ideas on the klein four group that will be the running example in this work . [ ex : klein ] let @xmath93 and @xmath6 a field of characteristic two . the group algebra @xmath94 is isomorphic to @xmath95/(x^{2},y^{2})$ ] , where @xmath96 and @xmath97 correspond to the generators of @xmath98 . let @xmath99 denote the jacobson radical of @xmath94 . it is well - known that @xmath100 is the symmetric algebra on the @xmath6-vector space @xmath101 ; see , for example , ( * ? ? ? * corollary 3.5.7 ) . thus @xmath100 is a polynomial ring over @xmath6 in two variables in degree one and @xmath102 . the @xmath0-point corresponding to a rational point @xmath103\in\bbp^{1}_{k}$ ] ( using homogeneous coordinates ) is represented by the map of @xmath6-algebras @xmath104/(t^{p}){\longrightarrow}k[x , y]/(x^{2},y^{2})\quad\text{where $ t\mapsto ax + by$.}\ ] ] more generally , for each closed point @xmath105 there is some finite field extension @xmath15 of @xmath6 such that @xmath106 contains a rational point @xmath107 $ ] over @xmath87 ( with @xmath108 acting transitively on the finite set of points over @xmath87 ) . then the @xmath0-point corresponding to @xmath87 is represented by the map of @xmath15-algebras @xmath109/(t^{p}){\longrightarrow}k[x , y]/(x^{2},y^{2})\quad\text{where $ t\mapsto a'x + b'y$.}\ ] ] now let @xmath15 denote the field of rational functions in a variable @xmath110 . the generic point of @xmath111 then corresponds to the map of @xmath15-algebras @xmath109/(t^{p}){\longrightarrow}k[x , y]/(x^{2},y^{2 } ) \quad\text{where $ t\mapsto x+sy$.}\ ] ] as before , @xmath5 is a finite group scheme over a field @xmath6 of positive characteristic @xmath1 . in this section , we introduce a notion of @xmath0-cosupport of a @xmath3-module , by analogy with the notion of @xmath0-support introduced in @xcite . the main result , theorem [ th : tensor - and - hom - pi ] , is a formula that computes the @xmath0-cosupport of a function object , in terms of the @xmath0-support and @xmath0-cosupport of its component modules . [ de : cosupport ] the _ @xmath0-cosupport _ of a @xmath3-module @xmath7 is the subset of @xmath88 defined by @xmath112 here @xmath113/(t^p)\to kg_k$ ] denotes a representative of the equivalence class of @xmath0-points corresponding to @xmath87 ; see remark [ rem : generic - points ] . the definition is modelled on that of the _ @xmath0-support _ of @xmath7 , introduced in @xcite as the subset @xmath114 this is denoted @xmath115 in @xcite ; our notation is closer to the one used in for cohomological support . for later use , we record the well - known property that the module of homomorphisms preserves and detects projectivity . [ le : end - projectivity ] let @xmath7 and @xmath8 be @xmath3-modules . 1 . if @xmath7 or @xmath8 is projective , then so is @xmath116 . @xmath7 is projective if and only if @xmath117 is projective . we repeatedly use the fact that a @xmath3-module is projective if and only if it is injective ; see , for example , ( * ? ? ? * lemma i.3.18 ) . \(i ) the functor @xmath118 takes injective @xmath3-modules to injective @xmath3-modules because it is right adjoint to an exact functor . thus , when @xmath8 injective , so is @xmath116 . the same conclusion follows also from the projectivity of @xmath7 because @xmath119 takes projective modules to injective modules , as follows from the natural isomorphisms : @xmath120 the first and the last isomorphisms are by adjunction , and the one in the middle holds because @xmath3 is cocommutative . \(ii ) when @xmath7 is projective , so is @xmath121 , by ( i ) . for the converse , observe that when @xmath117 is projective , so is @xmath122 , since @xmath123 preserves projectivity being the left adjoint of an exact functor . it remains to note that @xmath7 is a direct summand of @xmath122 , because the composition of the homomorphisms @xmath124 of @xmath3-modules equals the identity on @xmath7 . we now work towards theorem [ th : tensor - and - hom - pi ] that gives a formula for the cosupport of a function object , and the support of a tensor product . these are useful for studying modules over finite group schemes , as will become clear in section [ se : applications ] ; see also @xcite . the proof of theorem [ th : tensor - and - hom - pi ] is complicated by the fact that , in general , a @xmath0-point @xmath60/(t^p)\to kg_k$ ] does not preserve hopf structures , so restriction along @xmath57 does not commute with taking tensor products , or the module of homomorphisms . to deal with this situation , we adapt an idea from the proof of ( * ? ? ? * lemma 3.9)see also ( * ? ? ? * lemma 6.4 ) and ( * ? ? ? * lemma 6.4)where the equivalence of ( i ) and ( iii ) in the following result is proved . the hypothesis on the algebra @xmath125 is motivated by remark [ re : pi - basics](3 ) and example [ ex : quasi - elementary ] . [ le : pi - proj ] let @xmath15 be a field of positive characteristic @xmath1 and @xmath125 a cocommutative hopf @xmath15-algebra that is isomorphic as an algebra to @xmath126/(t_{1}^{p},\dots t_{r}^{p})$ ] . let @xmath53/(t^p ) { \longrightarrow}a\ ] ] be a flat homomorphism of @xmath15-algebras . for any @xmath125-modules @xmath7 and @xmath8 , the following conditions are equivalent : 1 . @xmath127 is projective . @xmath128 is projective . @xmath129 or @xmath130 is projective . as noted before , ( i ) @xmath66 ( iii ) is ( * ? ? ? * lemma 3.9 ) ; the hypotheses of _ op . _ includes that @xmath7 and @xmath8 are finite dimensional , but that is not used in the proof . we employ a similar argument to verify that ( ii ) and ( iii ) are equivalent . let @xmath131 be the antipode of @xmath125 , @xmath132 its comultiplication , and set @xmath133 , the augmentation ideal of @xmath125 . identifying @xmath134 with its image in @xmath125 , one has @xmath135 see ( * ? ? ? * i.2.4 ) . recall that the action of @xmath136 on @xmath137 is given by multiplication with @xmath138 , so that for @xmath139 and @xmath140 one has @xmath141 given a module over @xmath142 , we consider two @xmath56/(t^{p})$]-structures on it : one where @xmath134 acts via multiplication with @xmath143 and another where it acts via multiplication with @xmath144 . we claim that these two @xmath56/(t^{p})$]-modules are both projective or both not projective . this follows from a repeated use of ( * ? ? ? * proposition 2.2 ) because @xmath145 can be represented as a sum of products of nilpotent elements of @xmath142 , and each nilpotent element @xmath146 of @xmath142 satisfies @xmath147 . we may thus assume that @xmath134 acts on @xmath137 via @xmath144 . there is then an isomorphism of @xmath56/(t^{p})$]-modules @xmath148 where the action of @xmath56/(t^{p})$ ] on the right hand side is the one obtained by viewing it as a hopf @xmath15-algebra with comultiplication defined by @xmath48 and antipode @xmath49 . it remains to observe that for any @xmath56/(t^{p})$]-modules @xmath149 , the module @xmath150 is projective if and only if one of @xmath151 or @xmath98 is projective . indeed , if @xmath151 or @xmath98 is projective , so is @xmath152 , by lemma [ le : end - projectivity ] . as to the converse , every @xmath56/(t^{p})$]-module is a direct sum of cyclic modules , so when @xmath151 and @xmath98 are not projective , they must have direct summands isomorphic to @xmath56/(t^{u})$ ] and @xmath56/(t^{v})$ ] , respectively , for some @xmath153 . then @xmath154/(t^{u}),k[t]/(t^{v}))$ ] is a direct summand of @xmath152 . the former can not be projective as a @xmath56/(t^{p})$]-module because its dimension as a @xmath15-vector space is @xmath155 , while the dimension of any projective module must be divisible by @xmath1 : projective @xmath56/(t^{p})$]-modules are free . the first part of the result below is ( * ? ? ? * proposition 5.2 ) . [ th : tensor - and - hom - pi ] let @xmath7 and @xmath8 be @xmath3-modules . then there are equalities 1 . @xmath156 , 2 . @xmath157 . we prove part ( ii ) . part ( i ) can be proved in the same fashion ; see ( * ? ? ? * proposition 5.2 ) . fix a @xmath0-point @xmath158/(t^{p})\to kg_k$ ] . by remark [ re : pi - basics](3 ) , we can assume @xmath57 factors as @xmath56/(t^{p})\to kc\to kg_k$ ] , where @xmath54 is a quasi - elementary subgroup scheme of @xmath55 . as noted in lemma [ le : invariants](iii ) , there is an isomorphism of @xmath21-modules @xmath159 we may restrict a @xmath21 module to @xmath56/(t^{p})$ ] by first restricting to @xmath160 , and @xmath161 commutes with this operation . thus the desired result follows from the equivalence of ( ii ) and ( iii ) in lemma [ le : pi - proj ] , applied to the map @xmath56/(t^{p})\to kc$ ] , keeping in mind the structure of @xmath160 ; see example [ ex : quasi - elementary ] . next we record , for later use , some elementary observations concerning support and cosupport . the converse of ( i ) in the result below also holds . for finite groups this is proved in theorem [ th : finite - groups ] below , and for finite group schemes this is one of the main results of @xcite . [ le : cosupp - k ] let @xmath7 be a @xmath3-module . 1 . @xmath162 when @xmath7 is projective . 2 . @xmath163 when @xmath7 is finite dimensional . 3 . @xmath164 . part ( ii ) is immediate from definitions , given remark [ re : finite - dimensional ] . for the rest , fix a @xmath0-point @xmath60/(t^{p})\to kg_k$ ] . \(i ) when @xmath7 is projective , so are the @xmath21-modules @xmath165 and @xmath166 , and restriction along @xmath57 preserves projectivity , as @xmath57 is a flat map . this justifies ( i ) . \(iii ) evidently , @xmath167 equals @xmath15 and @xmath168 is non - projective . since @xmath57 was arbitrary , one deduces that @xmath169 is all of @xmath88 . that this also equals the @xmath0-cosupport of @xmath6 now follows from ( ii ) [ co : cosuppm * ] if @xmath7 is a @xmath3-module , then there is an equality @xmath170 this follows from theorem [ th : tensor - and - hom - pi](ii ) and lemma [ le : cosupp - k](iii ) . the equality of @xmath0-support and @xmath0-cosupport for finite dimensional modules , which holds by lemma [ le : cosupp - k](ii ) , may fail for infinite dimensional modules . we describe an example over the klein four group ; see example [ ex : t(i ) ] for a general construction . [ ex : klein2 ] let @xmath98 be the klein four group @xmath171 and @xmath6 a field of characteristic two . the @xmath0-points of @xmath94 were described in example [ ex : klein ] . we keep the notation from there . let @xmath7 be the infinite dimensional @xmath3-module with basis elements @xmath172 and @xmath173 for @xmath174 and @xmath3-action defined by @xmath175 where @xmath176 is interpreted as the zero vector . a diagram for this module is as follows : @xmath177 \overset{u_0}{\circ } \arrow[dash]{dr}[swap]{x } & & \overset{u_1}{\circ}\arrow[dash]{dl}{y } \arrow[dash]{dr } & & \overset{u_2}{\circ } \arrow[dash]{dl } \arrow[dash]{dr}&&\dots \\ & \underset{v_0}{\circ}&&\underset{v_1}{\circ } & & \underset{v_2}{\circ}&&\dots \end{tikzcd}\ ] ] the @xmath0-support of @xmath7 is the closed point @xmath178\}$ ] of @xmath111 whilst its @xmath0-cosupport contains also the generic point . given a finite field extension @xmath15 of @xmath6 , it is not hard to verify that for any rational point @xmath103\in\bbp^{1}_{k}$ ] the image of multiplication by @xmath179 on @xmath24 is its socle , that is spanned by the elements @xmath180 . for @xmath103 \ne [ 0,1]$ ] , this is also the kernel of @xmath179 , whilst for @xmath103=[0,1]$ ] it contains , in addition , the element @xmath181 . in view of remark [ re : finite - dimensional ] , this justifies the assertions about the closed points of @xmath182 . for the generic point let @xmath15 denote the field of rational functions in a variable @xmath110 . it is again easy to check that the kernel of @xmath183 and the image of multiplication by @xmath183 on @xmath59 are equal to its socle . so the generic point is not in @xmath184 . for cosupport , consider the @xmath6-linear map @xmath185 defined as follows . given a rational function @xmath186 , its partial fraction expansion consists of a polynomial @xmath187 plus the negative parts of the laurent expansions at the poles . if @xmath188 we define @xmath189 to be @xmath190 . by definition , @xmath191 ; using this it is easy to calculate that @xmath192 is in the kernel of @xmath183 . on the other hand , any function in the image of @xmath183 lands in the socle of @xmath7 . it follows that the kernel of @xmath183 is strictly larger than the image , and so the generic point is in the @xmath0-cosupport of @xmath7 . the focus of the rest of the paper is on finite groups . in this section we prove that a module over a finite group is projective if and only if it has empty @xmath0-cosupport . the key ingredient is a version of dade s lemma for elementary abelian groups . this in turn is based on an analogue of the kronecker quiver lemma ( * ? ? ? * lemma 4.1 ) . [ le : kronecker ] let @xmath6 be an algebraically closed field , @xmath193 a non - trivial extension field of @xmath6 , and let @xmath194 be @xmath6-vector spaces . if there exist @xmath6-linear maps @xmath195 with the property that for every pair of scalars @xmath196 and @xmath197 in @xmath193 , not both zero , the linear map @xmath198 is an isomorphism , then @xmath199 . note that @xmath200 here really means @xmath201 since this is the way @xmath193 acts on homomorphisms . since @xmath6 is algebraically closed , we may as well assume that @xmath202 , a simple transcendental extension of @xmath6 , since further extending the field only strengthens the hypothesis without changing the conclusion . use @xmath203 to identify @xmath98 and @xmath204 . then @xmath192 is a @xmath6-endomorphism of @xmath98 with the property that for all @xmath205 the endomorphism @xmath206 of @xmath207 is invertible . the action of @xmath192 makes @xmath98 into a @xmath208$]-module , with @xmath134 acting as @xmath192 does . since @xmath209 is invertible for all @xmath210 , and @xmath6 is algebraically closed , @xmath98 is a @xmath211-module . consider the homomorphism @xmath212 of rings that is induced by the assignment @xmath213 . it is not hard to verify that its kernel is the ideal generated by @xmath214 , so there is an exact sequence of @xmath215-modules @xmath216 applying @xmath217 and using adjunction yields an exact sequence @xmath218 thus @xmath214 is not invertible on @xmath219 . if @xmath98 is nonzero then it has @xmath211 as a summand as a @xmath211-module , so that @xmath214 , that is to say , @xmath220 , is not invertible . this contradicts the hypothesis . the result below , and its proof , are modifications of ( * ? ? ? * theorem 5.2 ) . given an @xmath6-algebra @xmath11 and an extension field @xmath15 of @xmath6 , we write @xmath221 for the @xmath15-algebra @xmath222 and @xmath23 for the @xmath221-module @xmath223 . [ th : dadehom ] let @xmath6 be a field of positive characteristic @xmath1 and set @xmath224/(t_1^p,\dots , t_r^p).\ ] ] let @xmath15 be an algebraically closed field of transcendence degree at least @xmath225 over @xmath6 . then an @xmath11-module @xmath7 is projective if and only if for all flat maps @xmath60/(t^r)\to r_k$ ] the module @xmath64 is projective . if @xmath7 is projective as an @xmath11-module , then @xmath226 is projective as an @xmath221-module , and because @xmath57 is flat , it follows that @xmath64 is projective . assume that @xmath64 is projective for all @xmath57 as in the statement of the theorem . we verify that @xmath7 is projective as an @xmath11-module by induction on @xmath41 , the case @xmath227 being trivial . assume @xmath228 and that the theorem is true with @xmath41 replaced by @xmath225 . it is easy to verify that the hypothesis and the conclusion of the result are unchanged if we pass from @xmath6 to any extension field in @xmath15 . in particular , replacing @xmath6 by its algebraic closure in @xmath15 we may assume that @xmath6 is itself algebraically closed . the plan is to use lemma [ le : kronecker ] to prove that @xmath229 . since @xmath11 is artinian it would then follow that @xmath7 is injective , and hence also projective , because @xmath11 is a self - injective algebra . we first note that for any extension field @xmath193 of @xmath6 , tensoring with @xmath193 gives a one - to - one map @xmath230 , which we view as an inclusion , and a natural isomorphism of @xmath193-vector spaces @xmath231 these remarks will be used repeatedly in the argument below . they imply , for example , that @xmath232 if and only if @xmath233 . let @xmath234 be the bockstein map ; see @xcite or , for a slightly different approach , ( * ? ? ? * i.4.22 ) . recall that this map is semilinear through the frobenius map , in the sense that @xmath235 fix an extension field @xmath193 of @xmath6 in @xmath15 that is algebraically closed and of transcendence degree @xmath236 . choose linearly independent elements @xmath237 and @xmath238 of @xmath239 . the elements @xmath240 and @xmath241 of @xmath242 induce @xmath6-linear maps @xmath243 let @xmath196 and @xmath197 be elements in @xmath193 , not both zero , and consider the element @xmath244 where @xmath245 is the radical of the ring @xmath246 . it defines a linear subspace of codimension one in the @xmath193-linear span of @xmath247 in @xmath246 . let @xmath248 be the subalgebra of @xmath246 generated by this subspace and view @xmath249 as an @xmath248-module , by restriction of scalars . as an @xmath248-module , @xmath249 is projective . indeed , @xmath248 is isomorphic to @xmath250/(z_{1}^{p},\dots , z_{r-1}^{p})$ ] , as an @xmath193-algebra . it is not hard to verify that the hypotheses of the theorem apply to the @xmath248-module @xmath249 and the extension field @xmath251 . since the transcendence degree of @xmath15 over @xmath193 is @xmath252 , the induction hypothesis yields that the @xmath248-module @xmath249 is projective , as claimed . we give @xmath253 the structure of a hopf algebra by making the generators @xmath254 primitive : that is , comultiplication is determined by the map @xmath255 and the antipode is determined by the map @xmath256 . note that @xmath257 has a natural structure of an @xmath246-module , with action induced from the left hand factor . the @xmath246-module @xmath258 , with the diagonal action , is projective . indeed , it is not hard to see the comultiplication on @xmath253 induces one on @xmath248 , so the latter is a sub hopf - algebra of the former . as @xmath249 is projective as an @xmath248-module , by the previous claim , so is @xmath259 for any @xmath248-module @xmath8 ; see lemma [ le : end - projectivity ] . since projective @xmath248-modules are injective , the desired claim is then a consequence of the following standard isomorphisms of functors @xmath260 on the category of @xmath246-modules . the bockstein of the element is @xmath261 and is represented by an exact sequence of the form @xmath262 for any @xmath246-module @xmath8 the hopf algebra structure on @xmath246 gives a map of @xmath193-algebras @xmath263 such that the two actions of @xmath264 on @xmath265 coincide , up to the usual sign ( * ? ? ? * corollary 3.2.2 ) . what this entails is that the map @xmath266 may be described as splicing with the extension @xmath267 which is obtained from the exact sequence by applying @xmath268 . by the preceding claim , the modules in the middle are projective , and so the element @xmath269 induces a stable isomorphism @xmath270 it follows that @xmath200 is an isomorphism for all @xmath271 in @xmath272 not both zero . thus lemma [ le : kronecker ] applies and yields @xmath273 as desired . the theorem below is the main result of this work . several consequences are discussed in the subsequent sections . [ th : finite - groups ] let @xmath6 be a field and @xmath5 a finite group . for any @xmath3-module @xmath7 , the following conditions are equivalent . 1 . @xmath7 is projective . 2 . @xmath274 3 . @xmath275 . we may assume that the characteristic of @xmath6 , say @xmath1 , divides the order of @xmath5 . the implications ( i)@xmath276(ii ) and ( i)@xmath276(iii ) are by lemma [ le : cosupp - k ] . ( iii)@xmath276(i ) let @xmath39 be an elementary abelian @xmath1-subgroup of @xmath5 and let @xmath277 denote @xmath7 viewed as a @xmath43-module . the hypothesis implies @xmath278 , by remark [ re : pi - basics](2 ) , and then it follows from theorem [ th : dadehom ] that @xmath277 is projective . theorem ( * ? ? ? * theorem 1 ) thus implies that @xmath7 is projective . ( ii)@xmath276(i ) when @xmath274 , it follows from theorem [ th : tensor - and - hom - pi ] that @xmath279 the already settled implication ( iii)@xmath276(i ) now yields that @xmath117 is projective . hence @xmath7 is projective , by lemma [ le : end - projectivity ] . the final part of this paper is devoted to applications of theorem [ th : finite - groups ] . we proceed in several steps and derive global results about the module category of a finite group from local properties , including a comparison of @xmath0-support and @xmath0-cosupport with cohomological support and cosupport . in the next section we consider the classification of thick and localising subcategories of the stable module category . from now on @xmath5 denotes a finite group , @xmath6 a field of positive characteristic dividing the order of @xmath5 . let @xmath4 be the stable module category of all ( meaning , also infinite dimensional ) @xmath3-modules modulo the projectives ; see , for example , @xcite . this is not an abelian category ; rather , it has the structure of a compactly generated tensor triangulated category and comes equipped with a natural action of the cohomology ring @xmath85 ; see ( * ? ? ? * section 10 ) . this yields the notion of cohomological support and cosupport developed in . more precisely , for each homogeneous prime ideal @xmath87 of @xmath85 that is different from the maximal ideal of positive degree elements , there is a distinguished object @xmath280 . using this one defines for each @xmath3-module @xmath7 its _ cohomological support _ @xmath281 the result below reconciles these notions with the corresponding notions defined in terms of @xmath0-points . [ th : supp - pisupp ] let @xmath5 be a finite group and @xmath7 a @xmath3-module . then @xmath282 regarded as subsets of @xmath92 . we use the fact that @xmath283 ; see ( * ? ? ? * proposition 6.6 ) . then using theorems [ th : finite - groups ] and [ th : tensor - and - hom - pi](ii ) one gets @xmath284 this gives the equality involving cosupports . in the same vein , using theorems [ th : finite - groups ] and [ th : tensor - and - hom - pi](i ) one gets @xmath285 this gives the equality involving supports . here is a first consequence of this result ; we are unable to verify it directly , except for closed points in the @xmath0-support and @xmath0-cosupport . [ co : maximal - elements ] for any @xmath3-module @xmath7 the maximal elements , with respect to inclusion , in @xmath286 and @xmath287 coincide . given theorem [ th : supp - pisupp ] , this follows from . we continue with two useful formulas for computing cohomological supports and cosupports ; they are known from previous work and are now accessible from the perspective of @xmath0-points . [ co : tensor - and - hom ] for all @xmath3-modules @xmath7 and @xmath8 there are equalities 1 . 2 . @xmath289 . this follows from theorems [ th : tensor - and - hom - pi ] and [ th : supp - pisupp ] . we wrap up this section with a couple of examples . the first one shows that the @xmath0-support of a module @xmath7 may be properly contained in that of @xmath121 . [ ex : klein3 ] let @xmath98 be the klein four group @xmath290 and @xmath6 a field of characteristic two . thus , @xmath291 , and a realisation of points of @xmath111 as @xmath0-points of @xmath94 was given in example [ ex : klein ] . let @xmath7 be the infinite dimensional @xmath94-module described in example [ ex : klein2 ] . as noted there , the @xmath0-support of @xmath7 consists of a single point , namely , the closed point @xmath292 $ ] . we claim @xmath293\}\cup\{\text{generic point of $ \bbp^{1}_{k}$}\}\,.\ ] ] indeed , since the @xmath0-cosupport of @xmath7 contains @xmath292 $ ] , by example [ ex : klein2 ] , it follows from theorem [ th : tensor - and - hom - pi ] that the @xmath0-cosupport of @xmath117 is exactly @xmath178\}$ ] . corollary [ co : maximal - elements ] then implies that @xmath292 $ ] is the only closed point in the @xmath0-support of @xmath117 . it remains to verify that the latter contains also the generic point . let @xmath15 be the field of rational functions in a variable @xmath110 . the @xmath0-point defined by @xmath56/(t^{p})\to k[x , y]/(x^{2},y^{2})$ ] with @xmath294 corresponds to the generic point of @xmath111 ; see example [ ex : klein ] . the desired result follows once we verify that the element @xmath295 of @xmath296 is in the kernel of @xmath183 but not in its image . it is in the kernel because @xmath297 is a @xmath94-module homomorphism . suppose there exists an @xmath192 in @xmath296 with @xmath298 . then for each @xmath299 , the identity @xmath300 yields @xmath301 noting that @xmath302 , by convention , it follows that @xmath303 modulo the submodule @xmath304 of @xmath59 . this can not be , as @xmath192 is in @xmath296 . in example [ ex : klein2 ] it is proved that , for @xmath7 as above , @xmath305 . the remark below is a conceptual explanation of this phenomenon , since @xmath7 is of the form @xmath306 for @xmath307 $ ] in @xmath111 . [ ex : t(i ) ] given @xmath308 , there is a @xmath3-module @xmath306 which is defined in terms of the following natural isomorphism @xmath309 where @xmath310 denotes the injective envelope of @xmath311 , @xmath312 is tate cohomology , and @xmath313 is the set of homomorphisms in @xmath314 ; see @xcite . the cohomological support and cosupport of this module have been computed in and , respectively . combining this with theorem [ th : supp - pisupp ] gives @xmath315 the results of this section concern the triangulated category structure of the stable module category , @xmath4 . recall that a full subcategory @xmath316 of @xmath4 is _ localising _ if it is a triangulated subcategory and is closed under arbitrary direct sums . in a different vein , @xmath316 is _ tensor ideal _ if for all @xmath54 in @xmath316 and arbitrary @xmath7 , the @xmath3-module @xmath317 is in @xmath316 . following , we say that the stable module category @xmath4 is _ stratified _ by @xmath85 if for each homogeneous prime ideal @xmath87 of @xmath85 that is different from the maximal ideal of positive degree elements the localising subcategory @xmath318 admits no proper non - zero tensor ideal localising subcategory . [ th : stratification ] let @xmath6 be a field and @xmath5 a finite group . then the stable module category @xmath4 is stratified as a tensor triangulated category by the natural action of the cohomology ring @xmath85 . therefore the assignment @xmath319 induces a one to one correspondence between the tensor ideal localising subcategories of @xmath4 and the subsets of @xmath76 . it suffices to show that @xmath320 whenever @xmath321 are @xmath3-modules with @xmath322 ; see . by adjointness , this is equivalent to @xmath323 being non - projective . thus the assertion follows from corollary [ co : tensor - and - hom ] , once we observe that @xmath324 implies @xmath325 . but this is again a consequence of corollary [ co : tensor - and - hom ] , since @xmath326 is non - projective by lemma [ le : end - projectivity ] . the second part of the assertion is a formal consequence of the first ; see ( * ? ? ? * theorem 3.8 ) . the inverse map sends a subset @xmath327 of @xmath76 to the subcategory consisting of all @xmath3-modules @xmath7 such that @xmath328 . the next results concern @xmath2 , the full subcategory of @xmath329 consisting of finite dimensional modules . a _ tensor ideal thick subcategory _ @xmath316 of @xmath2 is a triangulated subcategory that is closed under direct summands and has the property that for any @xmath54 in @xmath316 and finite dimensional @xmath3-module @xmath7 , the @xmath3-module @xmath317 is in @xmath316 . the classification of the tensor ideal thick subcategories of @xmath2 is the main result of @xcite and can be deduced from the classification of the tensor ideal localising subcategories of @xmath4 . this is based on the following lemma . [ th : thick ] let @xmath5 be a finite group and @xmath6 a field . then the assignment induces a one to one correspondence between the tensor ideal thick subcategories of @xmath2 and the specialisation closed subsets of @xmath92 . let @xmath331 be the assignment . by lemma [ le : thick ] , when @xmath316 is a tensor ideal thick subcategory of @xmath2 , the subset @xmath332 of @xmath88 is a union of zarksi - closed subsets , and hence specialisation closed . thus @xmath331 restricted to @xmath333 has the desired image . let @xmath334 be the map from specialisation closed subsets of @xmath88 to @xmath333 that assigns @xmath327 to the subcategory with objects @xmath335 this is readily seen to be a tensor ideal thick subcategory . the claim is that @xmath331 and @xmath334 are inverses of each other . fix a tensor ideal thick subcategory @xmath316 of @xmath333 . evidently , there is an inclusion @xmath338 . to prove that equality holds , it suffices to prove that if @xmath7 is a finite dimensional @xmath5-module with @xmath339 , then @xmath7 is in @xmath316 . let @xmath340 be the tensor ideal localising subcategory of @xmath4 generated by @xmath316 . from the properties of support , it is easy to verify that @xmath341 , and then theorem [ th : stratification ] implies @xmath7 is in @xmath340 . since @xmath7 is compact when viewed as an object in @xmath329 , it follows by an argument analogous to the proof of ( * ? ? ? * lemma 2.2 ) that @xmath7 is in @xmath316 , as desired . part of this article is based on work supported by the national science foundation under grant no.0932078000 , while db , sbi , and hk were in residence at the mathematical sciences research institute in berkeley , california , during the 20122013 special year in commutative algebra . the authors thank the centre de recerca matemtica , barcelona , for hospitality during a visit in april 2015 that turned out to be productive and pleasant . sbi and jp were partly supported by nsf grants dms-1503044 and dms-0953011 , respectively . we are grateful to eric friedlander for comments on an earlier version of this paper . d. j. benson , _ representations and cohomology i : basic representation theory of finite groups and associative algebras _ , cambridge studies in advanced mathematics , vol . 30 , second edition , cambridge university press , 1998 . j. f. carlson , _ the complexity and varieties of modules _ , integral representations and their applications , oberwolfach , 1980 , lecture notes in mathematics , vol . 882 , springer - verlag , berlin / new york , 1981 , pp . 415422 . a. neeman , _ the connection between the k - theory localization theorem of thomason , trobaugh and yao and the smashing subcategories of bousfield and ravenel _ , ann . . norm . ( 4 ) * 25 * ( 1992 ) , 547566 .

since the discovery of the fractional quantum hall effect@xcite , the properties of two - dimensional ( 2d ) interacting electron systems in a strong magnetic field have attracted persistent attention . historically , almost all the efforts were directed toward the study of the strong magnetic field case when only the lowest landau level is occupied . some attempts were undertaken @xcite to extend the analysis to a larger filling factor @xmath5 . however , they were all limited to the case when the energy of the electron - electron interaction is much smaller than the inter - landau level spacing , @xmath6 , with @xmath7 and @xmath8 being the magnetic length and the background dielectric constant respectively . under this condition , one can describe the system in terms of one landau level only ; the effect of mixing of the other landau levels was either neglected or taken into account by perturbation theory@xcite . in a weak magnetic field , the typical coulomb energy exceeds the cyclotron energy @xmath2 . in this case , one can start from the fermi liquid theory in zero magnetic field . the concept of fermi liquid usually enables one to neglect the interactions between quasiparticles when determining their energy spectrum in the vicinity of the fermi level . this concept is based , first , on the screening of coulomb interaction , and , second , on the constraint on the phase space allowed for a scattering event@xcite . landau quantization of the non - interacting quasiparticles in two dimensions creates a system of discrete levels . the existence of energy gaps between the levels affects adversely the screening , which involves the low - energy excitations of the electron system . in addition , if the fermi level coincides with a landau level , the system of non - interacting quasiparticles becomes infinitely degenerate , and even a weak interaction lifts this degeneracy . this may result in a non - trivial fine structure of the partially filled landau level ( pfll ) . in this paper , we study the low - energy properties of the system with a partially filled high landau level@xcite ( the level index @xmath9 equals to the integer part of @xmath10 $ ] ) . by integrating out all the other degrees of freedom , we derive the effective interaction between the electrons occupying this level . this procedure is valid for a weakly interacting electron gas @xmath11 , and at sufficiently large filling factors , @xmath12 . ( here @xmath13 is the conventional parameter characterizing the interaction strength . ) the effective hamiltonian enables us to develop a microscopic theory of the thermodynamic and tunnelling densities of states , and a description of spin excitations . in the tunnelling density of states , we find a pseudo - gap with a characteristic width @xmath14 , which confirms the result of the hydrodynamic approach@xcite . the thermodynamic density of states and spin excitations are characterized by the energy scale @xmath15 . both energies are smaller than @xmath2 . it makes our theory complementary to the conventional fermi liquid picture@xcite which is valid in the energy range @xmath16 . the paper is arranged as follows . [ peds ] is a qualitative discussion of the energy scales relevant for the problem , and the summary of our main results . [ sec : hamiltonian ] is devoted to the rigorous derivation of the effective hamiltonian describing the low - energy physics . in secs . [ sec : td ] and [ sd ] we apply the effective hamiltonian to study the thermodynamic and the tunnelling density of states . the spin excitations are discussed in sec . let us first consider an incompressible electron liquid with an integer filling factor @xmath17 . because of the gap @xmath2 in the spectrum , an external in - plane electric field @xmath18 can not be screened by the 2d electron system . instead , it causes only a finite polarization per unit area , @xmath19 . the polarizability of the incompressible 2d electron liquid reduces the interaction @xmath20 between two point charges embedded into it , @xmath21 where the two - dimensional dielectric function @xmath22 is related to the polarizability @xmath23 by@xcite @xmath24 at small wave vectors , the matrix elements of the dipolar moment between adjacent landau levels , @xmath25 , give the main contribution to the polarizability , @xmath26 ; here @xmath27 is the electron density on a landau level . the characteristic spreading of the electron wave function on a high landau level is equal to the cyclotron radius @xmath28 , and the estimate of the dipolar moment is @xmath29 . substituting @xmath23 into eq . ( [ chi ] ) , we find : @xmath30 where @xmath31 is the bohr radius . ( [ chilong ] ) is valid only at small wave vectors , @xmath32 . in the opposite limit , @xmath33 , a large number of landau levels participate in polarization of the 2d electron liquid . therefore , the standard thomas - fermi screening holds : @xmath34 formulas ( [ chilong ] ) and ( [ chishort ] ) match at @xmath35 ; the corresponding value of the dielectric constant , @xmath36 , is large in the weak magnetic field limit , where @xmath37 . the dielectric function for arbitrary @xmath38 is given by eq . ( [ simpleepsilon ] ) , see sec . [ sec : hamiltonian ] . as it follows from eq . ( [ chilong ] ) , polarization is irrelevant only for interactions on a very large length scale , @xmath39 , where @xmath20 is given by the unscreened coulomb interaction , see fig . [ figpot ] . at a smaller scale , @xmath40 , polarization is important , and eqs . ( [ potentialq ] ) and ( [ chilong ] ) yield : @xmath41 a t @xmath42 , according to eq . ( [ chishort ] ) , thomas - fermi screening takes placer , and the potential has the form @xmath43 the renormalized potential eq . ( [ potentialq ] ) is significantly smaller than @xmath2 . therefore , unlike the bare coulomb potential , the renormalized interaction does not mix landau levels . this observation enables us to construct an effective theory of the low energy properties of the electron system . similar to the strong magnetic field case , @xmath44 , the corresponding hamiltonian is just the energy of interaction between the electrons restricted to a single , partially filled landau level . the main difference is that in our case the interaction potential ( [ potentialq ] ) is much smaller than bare coulomb potential , and the pfll has a high index @xmath9 . although still unsolvable , this low - energy theory is much simpler than the original one , where all the landau levels were relevant . armed with the effective theory , we are able to estimate the characteristic energy scales which control the low - energy dynamics . in order to make these estimates , we employ the hartree - fock trial function , based on the one - electron wave functions of the @xmath45-th landau level . ( see secs . [ sec : td ] and [ sd ] for more details . ) at small electron concentration on the pfll , @xmath46 , the separation between electrons on this level exceeds the size of the wave function , @xmath47 . in this case the system under consideration is equivalent to a classical crystal of point charges interacting via potential ( [ potentialq ] ) . in the extremely dilute limit , @xmath48 , the polarization is not important , and the usual coulomb repulsion between the electrons of the pfll is not modified . in this limit the electron system of the pfll does not differ from a wigner crystal of electrons of the lowest landau level in the strong magnetic field regime . however , due to the polarization effect , there is a wide region of electron densities in the weak field regime , @xmath49 , where the coulomb repulsion in the crystal is replaced by the logarithmic potential ( [ potlong ] ) . in this case , the static properties of the electron crystal are equivalent to those of the vortex lattice in a thin superconducting film@xcite . we apply the developed picture to study of the energy dependence of the tunnelling density of states . we find a suppression of the tunnelling density of states ( one - electron spectral density ) at energies close to the fermi level @xmath50 , which is similar to the known results@xcite for a small filling factor , @xmath51 . in the latter case , despite rather different approaches , all the authors@xcite find a `` gap '' region of width @xmath52 ( here @xmath53 is the total electron density at @xmath54 ) . such a gap can be easily understood : @xmath55 equals the coulomb interaction energy of an extra electron with a `` frozen '' 2d electron system . if there were no relaxation of the system at all , the density of states would vanish at energies @xmath56 . the relaxation processes smear out the threshold in the density of states ; however , the suppression is still strong at these energies , because the characteristic relaxation time exceeds @xmath57 in a strong magnetic field . for a weak magnetic field , there are two stages in the evolution of the system after an electron tunnelled in . on the short time scale , @xmath58 , the polarization of the medium is formed , and the potential induced by the additional electron acquires the form ( [ potentialq ] ) . at larger time scales , only the electrons of the pfll can be re - distributed . the latter processes are slow , which causes the suppression of the tunnelling density of states . the corresponding gap is determined by the interaction of the tunneled electron with its nearest neighbors , @xmath59 . for a broad range of the filling factors of the pfll , @xmath60 , the width of the gap is : @xmath61 it is worth noticing that @xmath55 depends on the interaction strength and on the electron density on the pfll only logarithmically . at larger concentrations , @xmath62 , the single electron wave functions are still orthogonal to each other , but the densities corresponding to these wave functions have a significant spatial overlap . because of the strong repulsion between electrons at short distances , one could naively expect that the interaction on these distances gives the main contribution to the energy of the many - electron state . however , for the fully spin polarized electron system of the pfll , the orbital part of the many - electron wave function is antisymmetric , and vanishes whenever two electrons have the same coordinates . this suppression of the wave function amplitude compensates the large value of the interaction potential at @xmath63 . at larger distances , @xmath64 , the interaction ( [ potshort ] ) decays rapidly , and the corresponding contribution to the hartree - fock energy turns out to be smaller than the contribution of the long - range interaction ( [ potlong ] ) , see secs . [ sec : td ] and [ sd ] . this results in the saturation of the logarithmic growth of the gap @xmath55 with the electron density : at @xmath62 the gap in the tunnelling density of states is @xmath65 this estimate was obtained previously in the framework of the hydrodynamic approach@xcite . we turn now to the discussion of the spin - flip excitations . the orbital part of the many - electron wave function of such an excited state is no longer antisymmetric . therefore , the energy of such an excitation , @xmath66 , is determined not only by the bare zeeman energy but also by the extra interaction energy associated with the change in the structure of the orbital wave function . because the orbital part of the wave function is not antisymmetric , the interaction on a short range ( @xmath67 ) now does contribute to the energy of the state . this corresponds to the well known exchange enhancement of the @xmath68-factor@xcite . following ref . we analyze the exactly solvable case of a pfll which has a completely filled spin - polarized sublevel . flipping the spin of a single electron affects the exchange interaction of this electron with all the other electrons of the _ same _ landau level . all the lower landau levels carry equal number of electrons of both spin polarizations , and do not contribute to the energy associated with the spin flip . therefore , only @xmath69-th fraction of the total electron density participates in the exchange enhancement , and the corresponding contribution to @xmath66 is approximately @xmath45 times smaller than the exchange energy per electron at zero magnetic field . for a 2d electron gas with fermi wave vector @xmath70 , the latter energy is of the order @xmath71 , and the contribution to @xmath66 is of the order @xmath72 . the rigorous result obtained in sec . [ sw ] differs from this estimate only by a logarithmic factor . for the effective @xmath68-factor defined by the relation @xmath73 we obtain @xmath74 if the filling factor @xmath75 is odd . here @xmath76 is determined by the zeeman splitting in the absence of the interaction , and it is usually small ( in @xmath77 the value of @xmath76 is only @xmath78 ) . for the even filling factors , the value of @xmath66 is determined solely by @xmath76@xcite . we see that in the weak magnetic field , the energy scale of charge excitations @xmath79 is parametrically smaller than the energy scale for spin excitations @xmath15 . this is qualitatively different from the situation at low filling factor , @xmath51 , where both these excitations are characterized by the same energy scale @xmath80 . we start from the full hamiltonian of the system @xmath81 where the hamiltonian for non - interacting electrons , @xmath82 , is given by @xmath83 \psi_{n , k}^{\dagger}\psi_{n , k}. \label{h0}\ ] ] here @xmath84 and @xmath85 are the landau level index and the guiding center coordinate respectively ( we adopt the landau gauge ) . the fermionic field operators in eq . ( [ h0 ] ) satisfy the standard anticommutation relations @xmath86 , and @xmath87 is the chemical potential ( we include the energy @xmath88 into @xmath87 ) . to simplify the notation , we omit the spin indices . we will also neglect the zeeman term @xmath89 in all the intermediate calculations . this is legitimate because @xmath90 and therefore all the effects associated with the small zeeman energy can be included in the effective hamiltonian after the fast degrees of freedom are integrated out . electron - electron interaction is described by @xmath91 where @xmath92 stands for the normal ordering , @xmath93 is the electron density operator , @xmath94 , and @xmath95 is the average electron density . the electron annihilation operator @xmath96 is related to the operators @xmath97 by @xmath98 , where @xmath99 are the single electron wave functions in a magnetic field . in order to develop the effective low - energy theory , we notice that the expression for the partition function , @xmath100 , can be written as @xmath101 where @xmath102 is the trace over the fock space of the pfll ( @xmath45 is the index of the pfll , @xmath103 is the inverse temperature ) and @xmath104 means trace over all the other landau levels . at low temperatures , @xmath105 , eq . ( [ z1 ] ) acquires the form @xmath106 where the effective hamiltonian @xmath107 is defined as @xmath108 let us emphasize that , in general , the low frequency properties can be described only by means of the effective action which includes a retarded interaction . the description by an effective hamiltonian with _ instantaneous _ interaction is possible only if the characteristic time of the retardation is much smaller than @xmath109 , where @xmath110 is the maximal energy scale which is considered within the effective theory . this condition is met for the problem under consideration . indeed , the summation in eq . ( [ heff ] ) involves inter - landau level transitions only . these transitions are associated with the energy gaps not smaller than @xmath2 . correspondingly , the characteristic time of the retardation related to such a transition is @xmath111 . on the other hand , we saw at the sec . [ peds ] that the largest characteristic energy @xmath112 is still parametrically smaller than @xmath2 , which enables us to neglect the retardation . we now turn to the calculation of the effective hamiltonian ( [ heff ] ) . the simplest way to proceed is to use the hubbard - stratonovich transformation : first , we introduce auxiliary scalar field @xmath113 to decouple quartic in @xmath96 interaction part of the hamiltonian ; second , we perform the summation over all the fermionic degrees of freedom not belonging to pfll . after that , we integrate out auxiliary fields by means of the saddle point approximation and thus obtain the effective hamiltonian @xmath114 . below we implement this procedure . performing the first step of the hubbard - stratonovich transformation , we find for @xmath115 from eq . ( [ heff ] ) @xmath116 where @xmath117 is the imaginary time ordering , and @xmath118 vector @xmath119 represents time and space coordinates , @xmath120 . the domain of integration over @xmath121 is : @xmath122,\ x \in [ 0 , l_x],\ y \in [ 0 , l_y]$ ] , where @xmath123 is the geometrical size of the 2d system . we will omit the planck constant in all the intermediate calculations . the measure of the functional integral in eq . ( [ decoupling ] ) is determined by the condition @xmath124 and the function @xmath125 is defined by the equation @xmath126 the density operator is @xmath127 where the fermionic operators in matsubara representation are @xmath128 , and @xmath129 . by introducing the infinitesimal positive time shift @xmath130 into definition ( [ rho ] ) we preserve the normal ordering of the fermionic operators in the interaction hamiltonian , see eq . ( [ hi ] ) . after the decoupling is performed , the fermionic part of the hamiltonian in eq . ( [ decoupling ] ) becomes quadratic , which enables us to make the second step of the hubbard - stratonovich transformation , i.e. to carry out the trace @xmath104 over the `` fast '' fermionic degrees of freedom . the result depends solely on the variables belonging to the pfll : @xmath131 \phi(\mbox{\boldmath $ \xi_1 $ } ) \tilde{g}_{\phi}(\mbox{\boldmath $ \xi_1$},\mbox{\boldmath $ \xi_2 $ } ) \phi(\mbox{\boldmath $ \xi_2 $ } ) , } \label{f2}\\ & & \displaystyle{g_n(\mbox{\boldmath $ \xi_1$},\mbox{\boldmath $ \xi_2$})=-p_n(\mbox{\boldmath $ r_1$},\mbox{\boldmath $ r_2 $ } ) e^{(\tau_2-\tau_1)(n\omega_c-\mu)}\theta ( \tau_1-\tau_2-\epsilon ) } , \label{gn}\end{aligned}\ ] ] [ ugly1 ] where the @xmath118 vector @xmath132 is @xmath133 . in eqs . ( [ ugly1 ] ) @xmath134 is the projection of the fermionic operator ( in matsubara representation ) on the pfll , @xmath135 , operator @xmath136 is defined as @xmath137 . projections of the density operator @xmath93 and of the hamiltonian @xmath82 on the pfll are given by @xmath138 and by @xmath139 respectively . ( [ gn ] ) , @xmath140 is the projection operator onto the pfll@xcite ; @xmath141 is the average electron concentration on the pfll . operator @xmath142 in eq . ( [ f2 ] ) is defined in such a way @xmath143 . this follows from the anticommutation relation for the fermionic operators @xmath144 . the green s function @xmath145 is the kernel of the integral operator : @xmath146^{-1}\hat{p}_{n}^{\perp } , \label{g}\ ] ] here @xmath147 is the projection operator onto the space of functions _ orthogonal _ to the states of the pfll @xcite , and @xmath148 is the hamiltonian of an electron in the magnetic field and in the external potential @xmath149 : @xmath150 the function @xmath151 describes the evolution in the external field @xmath149 of electron states constrained to the landau levels @xmath152 . this constraint is implemented by the introduction of the projection operators @xmath147 into eq . ( [ g ] ) . at @xmath153 the function @xmath151 is defined as @xmath154 in accordance with eq . ( [ rho ] ) . the first term in eq . ( [ f ] ) results from the summation over the fermionic states with @xmath152 , and represents the `` thermodynamic potential '' of electrons in these states in the field @xmath149 . the function @xmath155 is defined as the kernel of the operator @xmath156 with @xmath157 from eq . ( [ g ] ) . here , the factor @xmath158 comes from the spin degeneracy . in order to perform the integration over the auxiliary field in eq . ( [ ugly ] ) , we employ the saddle point approximation . it means that one has to expand the functional @xmath159 , see eq . ( [ f ] ) , and the operator @xmath160 , see eq . ( [ f2 ] ) , up to the second order in @xmath149 . we start from the expansion of the operator @xmath161 . it contains explicitly a factor , which is bilinear in @xmath149 , and therefore we can replace @xmath151 in eq . ( [ f2 ] ) by the green s function @xmath162 of an electron in the absence of an external field , @xmath163 . then , eq . ( [ g ] ) enables us to find the explicit form of @xmath162 : @xmath164}. \label{g0}\end{aligned}\ ] ] when expanding the first term in @xmath159 , we use definition ( [ f ] ) and the solution of eqs . ( [ g ] ) , ( [ hphi ] ) up to the second order in @xmath149 , and we obtain : @xmath165 where @xmath166nl_xl_y/\pi\ell^2 $ ] is the thermodynamic potential of the system of noninteracting electrons on the completely filled landau levels . the third term in the functional @xmath159 , see eq . ( [ f ] ) , is already bilinear in @xmath149 . substitution of eq . ( [ fe ] ) into eq . ( [ f ] ) yields @xmath167 where the second term describes fluctuations of the field @xmath149 which are renormalized due to the integrated out degrees of freedom . the kernel @xmath168 equals to @xmath169 where @xmath170 coincides with the dielectric function calculated in the random phase approximation@xcite : @xmath171 here , the polarization operator @xmath172 in eq . ( [ fullepsilon ] ) is given by @xmath173 \left[\tilde{g}_{0}(0 , 0 ; \mbox{\boldmath $ r$ } , \tau)+g_n(0 , 0 ; \mbox{\boldmath $ r$ } , \tau ] \right],}\label{polarization}\end{aligned}\ ] ] where the functions @xmath174 and @xmath175 are defined by eqs . ( [ g0 ] ) and ( [ gn ] ) respectively . the explicit form and the asymptotic behavior of @xmath172 are presented in appendix [ ap1 ] . within the approximations made , we can transform eq . ( [ ugly ] ) to the form of a gaussian integral over the auxiliary field @xmath149 . the result is @xmath176 } \nonumber \\ & & \displaystyle{\times\exp\left(\frac{1}{2}\!\int\ ! d^3\mbox{\boldmath $ \xi_1$}\!\int\ ! d^3\mbox{\boldmath $ \xi_2 $ } \phi(\mbox{\boldmath $ \xi_1 $ } ) \phi(\mbox{\boldmath $ \xi_2 $ } ) { \cal s}(\mbox{\boldmath $ \xi_1\ ! -\!\xi_2$})\right ) } , \label{almostthere}\end{aligned}\ ] ] where @xmath177 is obtained from operator @xmath142 , see eq . ( [ f2 ] ) , by replacing @xmath178 with @xmath174 . before we proceed further , let us emphasize that the saddle point approximation can be justified only if the characteristic value of the fluctuations of the auxiliary field @xmath149 is small enough . the estimate which will be presented later in this section shows that the expansion quadratic in @xmath149 is parametrically valid for the weakly interacting electron system at large filling factors . now , we are in the position to perform the actual integration in eq . ( [ almostthere ] ) . operator @xmath179 , see eq . ( [ f1 ] ) , is a linear functional of @xmath149 , whereas operator @xmath177 , see eq . ( [ f2 ] ) , is quadratic in @xmath149 ; because the fluctuations of field @xmath149 are small , the typical value of @xmath142 is much smaller than that of @xmath179 . if we neglect the term @xmath177 at all , the functional integration in eq . ( [ almostthere ] ) can be easily performed . the calculation is further simplified if we approximate @xmath180 . it is valid approximation for the description of the low - energy dynamics of electrons belonging to pfll : @xmath181 is rapidly decaying function of time : @xmath182 at @xmath183 , as it follows from eqs . ( [ renorm ] ) - ( [ polarization ] ) . the integration over field @xmath149 then results in an exponential @xmath184 , with the hamiltonian of density - density interaction , @xmath185 the operator ( [ heint ] ) acts within the fock space of pfll , and the renormalized pair interaction @xmath20 is related to @xmath186 by @xmath187 substitution of eq . ( [ renorm ] ) into eq . ( [ us ] ) yields : @xmath188 here the static dielectric function @xmath189 describes the renormalization of the bare coulomb potential due to the integrated out degrees of freedom . this function can be easily found with the help of eqs . ( [ fullepsilon ] ) and ( [ stat ] ) , and it has the form @xmath190 \label{simpleepsilon}\ ] ] for the large filling factor @xmath191 and for values of wave vector @xmath38 much smaller than fermi wave vector , @xmath70 . in eq . ( [ simpleepsilon ] ) , @xmath31 is the bohr radius and @xmath192 is the zeroth order bessel function . asymptotic behavior of the renormalized potential @xmath20 was discussed in sec . [ peds ] , see also fig . [ figpot ] . the term @xmath177 from eq . ( [ almostthere ] ) can be taken into consideration by means of the perturbation theory using the small parameters @xmath193 and @xmath194 . we show in appendix [ ap0 ] that this operator generates term of the effective hamiltonian which is linear in density . this linear term can be trivially included into the free electron hamiltonian @xmath195 by a shift of the chemical potential . finally , substituting the resulting expression for @xmath115 in eq . ( [ heff ] ) and restoring the zeeman energy term , we obtain the expression for the effective hamiltonian : @xmath196 where @xmath197 is defined by eq . ( [ heint ] ) , the explicit expression for the shifted chemical potential @xmath198 is presented in appendix [ ap0 ] , see eq . ( [ mufinal ] ) , and the thermodynamic potential @xmath199 of the electrons on the filled landau levels is given by eq . ( [ omegafinal ] ) . we will not need concrete values of @xmath198 and @xmath199 in the further calculations . hamiltonian ( [ he ] ) is the main result of this section . the physical meaning of this hamiltonian is that the low - frequency dynamics of the system is described by the electrons belonging to the upper landau level . interaction between these electrons is renormalized due to the large polarizability of all the other landau levels . let us now discuss the condition of the validity of the approximations we made . as we mentioned earlier , the saddle point approximation can be justified only if the characteristic value of the fluctuations of the auxiliary field @xmath149 is small enough . below we show that the fluctuations at all the spatial and time scales are parametrically small for a weakly interacting electron system at large filling factors . the magnitude of the fluctuations localized within the spatial range @xmath200 and within the time interval @xmath201 can be estimated from eqs . ( [ almostthere ] ) and ( [ renorm ] ) as @xmath202 further estimates depend on the relation between the scale @xmath200 and the cyclotron radius @xmath47 . if this spatial scale is large , @xmath203 , only the dipole transitions between the nearest landau levels are induced . the corresponding off - diagonal matrix elements of the hamiltonian are of the order of @xmath204 . the mixing of the landau levels by the field @xmath149 is small if @xmath205 combining eq . ( [ estimate ] ) with eq . ( [ shfl ] ) and using the results ( [ fullepsilon ] ) and ( [ hydrodynamic ] ) for the dielectric function , we find the condition of small fluctuations [ cond ] @xmath206 therefore , the long - range fluctuations are not `` dangerous '' in the weak magnetic field regime . let us now analyze the short range fluctuations , @xmath207 . in this case transitions between distant landau levels are possible , and the requirement of the smallness of fluctuations coincides with the standard one for the validity of the random phase approximation at zero magnetic field , @xmath208 thus , the saddle point approximation is valid for the weakly interacting electrons in a weak magnetic field . in the following sections , we will apply the effective hamiltonian ( [ he ] ) to describe various physical effects associated with the pfll . in this section , we evaluate the ground state energy and thermodynamic density of states @xmath209 as a function of the filling factor of the partially filled landau level @xmath210 . we consider explicitly the case @xmath211 . system at the filling factors @xmath212 can be analyzed with the help of the electron - hole symmetry , and the the thermodynamic density of states for this case can be obtained from the results for @xmath213 by the replacement @xmath214 . we assume that the ground state is spin polarized at @xmath215 , and thus omit the irrelevant zeeman term@xcite . the chemical potential @xmath198 for the given electron concentration on the pfll @xmath216 is found by differentiating the ground state energy of hamiltonian ( [ he ] ) with respect to the number of electrons @xmath217 : @xmath218 where @xmath219 is the wave function of the ground state of the system with @xmath217 electrons on the pfll . evidently , it is sufficient to consider only the filling factors @xmath220 . at larger filling factors , @xmath221 , one can use the electron - hole transformation within the spin sublevel , @xmath222 , and study the system of holes with the filling factor @xmath223 described by the hamiltonian @xmath224\ ! \left[\hat{\rho}_n^h(\mbox{\boldmath $ r_2 $ } ) - n_h^n\right]\ ! : . \nonumber\end{aligned}\ ] ] here @xmath225 is the hole density operator @xmath226 and @xmath227 is the average density of holes at the pfll . the first term in @xmath197 corresponds to the shift of the chemical potential of the completely filled landau level due to the exchange interaction@xcite : @xmath228 here @xmath229 is the projector operator on the pfll@xcite . the integration in eq . ( [ exch ] ) can be performed with the help of the explicit form ( [ property ] ) for @xmath229 . under the conditions ( [ cond ] ) , the calculation yields @xmath230 it follows from eq . ( [ hh ] ) that filling of one spin sublevel results in the shift @xmath231 of the chemical potential . this enables us to find the average thermodynamic density of states which is defined by the relation @xmath232 with the help of eq . ( [ exch1 ] ) , we obtain @xmath233^{-1}. \label{thermdensity}\ ] ] it is worth noticing , that the thermodynamic density of states ( [ thermdensity ] ) does not depend on magnetic field at all . we will see below that for a broad range of filling factors of the pfll @xmath234 , the actual value of the thermodynamic density of states is close to the average one given by eq . ( [ thermdensity ] ) . the true ground state of the hamiltonian ( [ he ] ) for non - integer filling factors is not known . in order to estimate the ground state energy , we use the hartree - fock trial function analogous to that used by maki and zotos@xcite for the lowest landau level @xmath235 where @xmath236 is the normalized one - electron wave function of a coherent state@xcite on the pfll with the guiding center localized about point @xmath237 . in the landau gauge , @xmath238 , @xmath239 , this function has the form @xmath240{\bf z}/2\ell^2+ixy/2\ell^2 } g\left({\mbox{\boldmath $ r - r$}}\right ) , & \nonumber\\ & g\left(\mbox{\boldmath $ r$}\right)=\frac{1}{\sqrt{n!}}\left(\frac{x - iy}{\sqrt{2}\ell}\right)^n e^{-{\bf r}^2/4\ell^2}. & \label{coherentstate}\end{aligned}\ ] ] the overlap between two functions ( [ coherentstate ] ) rapidly decreases with the distance between their guiding centers : @xmath241{\bf z}/2\ell^2}e^{-({\bf r}_1-{\bf r}_2)^2/4\ell^2}. \label{overlap}\ ] ] equation ( [ overlap ] ) shows that the overlap is exponentially small even when the distance between the guiding centers is smaller than @xmath47 and the electrons are not separated in space . for small filling factors @xmath242 , we can choose the guiding centers @xmath243 separated by the distance much larger than the magnetic length @xmath7 which enables us to neglect the non - orthogonality of the coherent states . then , the expression for the energy @xmath244 given by eq . ( [ mu ] ) acquires a simple form @xmath245 where hartree @xmath246 and fock @xmath247 potentials are given by @xmath248 ^ 2 e^{-q^2\ell^2+i{\bf qr } } } \label{hartree}\end{aligned}\ ] ] and @xmath249 ^ 2 e^{-q^2\ell^2-{\bf q}{\bf r}-{\bf r}^2/2\ell^2}},\nonumber\\ \label{fock}\end{aligned}\ ] ] respectively . here @xmath250 is the laguerre polynomial@xcite . we used eqs . ( [ renpotential ] ) and ( [ coherentstate ] ) , when deriving eqs . ( [ hartree ] ) and ( [ fock ] ) . small corrections appearing due to the non - orthogonality of the functions ( [ coherentstate ] ) were discussed in ref . . now , the energy given by eq . ( [ hf_energy ] ) should be minimized with respect to the positions of the guiding centers @xmath251 . the best configuration corresponds to the guiding centers arranged in a triangular lattice@xcite . below we present analytical results for the energy of the variational ground state ( [ trialfunction ] ) in different domains of filling factor @xmath252 . _ dilute system , @xmath253 _ in this case , the lattice constant @xmath254 is large , @xmath255 , so that the fock potential ( [ fock ] ) is exponentially small and the hartree term @xmath256 coincides with the interaction potential @xmath257 . in this limit , two situations may be distinguished . for an extremely dilute system , @xmath258 , the lattice constant is larger than @xmath259 , and therefore the coulomb potential is not renormalized . in this case , the ground state energy coincides with that of the wigner crystal on a neutralizing background , @xmath260 where @xmath261 is a numerical constant @xcite , and @xmath262 is the number of states on a landau level . the use of eq . ( [ mu ] ) immediately yields for the thermodynamic density of states in this regime @xmath263 the negative thermodynamic density of states for an analogous system was considered earlier in ref .. for a moderately dilute system @xmath264 , the typical distance between electrons on the pfll is much smaller than @xmath259 . at such distances the potential of interaction is strongly renormalized by the screening and it is given by eq . ( [ potlong ] ) . with the logarithmic accuracy the energy of the system is given by @xmath265 \label{dilute_energy}\ ] ] and , correspondingly , the thermodynamic density of states for moderately dilute system takes the form : @xmath266 _ `` dense '' limit , @xmath267 _ in this case the distance between the nearest guiding centers is smaller than @xmath47 . the asymptotic behavior of the hartree and fock potentials depends on the relation between the magnetic length @xmath7 and the bohr radius @xmath268 . we restrict ourselves to the case @xmath269 , or @xmath270 which enables us to simplify eqs . ( [ hartree ] ) and ( [ fock ] ) , @xmath271 for the range of distances @xmath272 . the resulting potential @xmath273 is a smooth function of @xmath274 at @xmath275 . this enables us to find the analytical expression for the ground state energy ( [ hf_energy ] ) in the region @xmath276 . first , we can approximate @xmath277 , as the potential varies only logarithmically over the lattice cell . second , we notice that the spatial average of the hartree potential is exactly equal to the average of the bare potential @xmath20 , and these two terms cancel each other in eq . ( [ hf_energy ] ) for @xmath244 . finally , the integral of the fock potential is proportional to the exchange shift , @xmath278 , see eq . ( [ exch1 ] ) . the resulting energy of the system is @xmath279 , } & \nonumber\end{aligned}\ ] ] which yields the thermodynamic density of states of the form @xmath280^{-1}\times } & \nonumber \\ & \displaystyle{\left[1-\frac{3\sqrt{2 } } { 32\pi r_s\ln \left(r_s^{-1}\right ) } \left(\delta\nu n\right)^{-1}\right ] . } & \label{dtd}\end{aligned}\ ] ] the leading term in eq . ( [ dtd ] ) coincides with eq . ( [ thermdensity ] ) . the filling factors @xmath281 can be considered similarly by using the electron - hole symmetry , see eq . ( [ hh ] ) , and the trial function of type ( [ trialfunction ] ) . the overall dependence of the thermodynamic density of states on the filling factor is shown in fig . the value of the one electron spectral density @xmath282 can be measured in the tunnelling experiments@xcite . for the electron states on the pfll , @xmath283 is defined by relations : @xmath284 here , @xmath285 is an eigenstate of the _ full _ hamiltonian ( [ initial_h ] ) , @xmath286 is the corresponding eigenvalue and @xmath287 stands for the ground state of the system . functions @xmath288 and @xmath289 describe the introduction of an extra electron or an extra hole onto the pfll respectively . we are interested in the spectral density at energies @xmath290 . all the behavior of the system at such low energies can be described by the effective hamiltonian ( [ he ] ) . following the method of sec . [ sec : hamiltonian ] , we can express the spectral density ( [ sddef ] ) in terms of the eigenstates of the _ effective _ hamiltonian as @xmath291 where @xmath292 denotes the eigenstates of the effective hamiltonian ( [ he ] ) and @xmath293 is the corresponding eigenvalue , @xmath294 is the ground state . the @xmath295-factor in eq . ( [ sdeff ] ) describes the overlap between the low - energy eigenstates of the full hamiltonian and the eigenstates of the effective hamiltonian . the value of @xmath295-factor coincides with the quasiparticle weight for a @xmath296 degenerate plasma in the absence of the magnetic field , @xmath297 now we use eq . ( [ sdeff ] ) to evaluate the spectral density . our analysis is based on the trial function ( [ trialfunction ] ) . even though we are not able to find the details of the energy dependence of the spectral density , this approach provides a reliable estimate for the energy scales involved . because the trial function ( [ trialfunction ] ) is constructed from the coherent states ( [ coherentstate ] ) , it is convenient to rewrite eq . ( [ sdeff ] ) in terms of the operators @xmath298 , see eq . ( [ c_s_creation ] ) , creating or annihilating an electron in the coherent state : [ sdc ] @xmath299 here we used the representation of the projection operator in terms of the coherent states ( [ coherentstate ] ) : @xmath300 let us concentrate below on the calculation of the spectral density for the hole excitations @xmath289 . ( the calculation procedure for @xmath288 is similar and will be briefly outlined later . ) at small time @xmath301 , expression ( [ sdcs2 ] ) can be approximated by the formula @xmath302 ^ 2\right ) \label{sdcs_exp}\ ] ] with parameters @xmath303 \right|0\right>}{\left<0\left| \psi_{\bf r}^\dagger\psi_{\bf r } \right|0\right>}};\label{e } \\ & & \displaystyle { \left[d({\bf r})\right]^2 = \frac{\left<0\left|\psi_{\bf r}^\dagger \left[{\hat{h}_{eff}},\left[{\hat{h}_{eff}},\psi_{\bf r } \right]\right ] \right|0\right>}{\left<0\left|\psi_{\bf r}^\dagger \psi_{\bf r } \right|0\right > } - \left[e^- ( { \bf r})\right]^2 . } \nonumber\end{aligned}\ ] ] substitution of eq . ( [ sdcs_exp ] ) into eq . ( [ sdcs ] ) gives the result @xmath304 ^ 2 } \!\right\ } \label{sdfinal}\ ] ] for the spectral density . now we use maki - zotos trial state ( [ trialfunction ] ) to find the functions in the integrand in eq . ( [ sdfinal ] ) . at a small filling of the pfll , @xmath305 , we obtain with the help of eq . ( [ overlap ] ) @xmath306 which describes the `` density of the centers of orbits '' in the wigner crystal . energy @xmath307 defined in eq . ( [ e ] ) turns out to be independent on the position @xmath237 because all the electrons in the wigner crystal have the same energy : @xmath308 here the hartree - fock energy @xmath309 is defined by eq . ( [ hf_energy ] ) , and the second term describes the neutralizing background . the parameter @xmath310 is also independent of @xmath274 and it is given by @xmath311 ^ 2 . \label{width}\ ] ] because none of the parameters in the exponent in eq . ( [ sdfinal ] ) depend on @xmath312 , the integration can be easily performed , and it yields : @xmath313 let us now discuss the physical meaning of the energies @xmath314 , @xmath315 . after an electron is removed from the pfll , the system acquires extra energy equal to @xmath316 due to the interaction between the hole and all the electrons of the pfll . if all the other electrons of the pfll were `` frozen '' , this state would give a contribution @xmath317 to the spectral density . ( classical model of the frozen electrons was used in ref . for the electrons on the lowest landau level . ) however , the state formed right after the tunnelling is not an eigenfunction . the decay of this initial state leads to the finite quantum width @xmath318 of the spectral density , see fig . [ figds ] a. to find the region of applicability of expression ( [ sdfinal1 ] ) we notice that the expansion in @xmath301 we used in eq . ( [ sdcs_exp ] ) is valid for @xmath319 . the characteristic time determining the value of the integral in eq . ( [ sdcs ] ) can be estimated as @xmath320 ) is applicable for @xmath321 . at the boundary of applicability , the spectral density ( [ sdfinal1 ] ) is proportional to @xmath322 $ ] , and therefore eq . ( [ sdfinal1 ] ) describes the main contribution to the spectral density only if @xmath323 . substituting the explicit expression for hartree - fock potential ( [ hf_energy ] ) into eqs . ( [ energies ] ) and ( [ width ] ) , we see that the ratio @xmath324 is maximal at @xmath325 , where its value is @xmath326 . furthermore , we believe that the use of the hartree - fock function sets the upper limit for @xmath315 . this is because trial function ( [ trialfunction ] ) does not take into account the correlations in the motion of centers of orbits , and thus overestimates zero - point fluctuations of the electron density . the use of a more sophisticated@xcite trial function , or of a phenomenological model that has the correct spectrum of excitations of the wigner crystal@xcite would restrict the summation in eq . ( [ width ] ) to the nearest neighbors only , and further reduce the value of @xmath315 . the spectral density of holes has a sharp peak , because a hole corresponds to a vacancy in the wigner crystal . in contrast , the tunnelling electron may create an interstitial at an arbitrary position in the elementary cell . this gives rise to a finite width of the spectral density @xmath288 , even for the `` frozen '' , @xmath327 , crystal . the energy band for an interstitial is given by @xmath328\!- \!n_e^n \!\int \!d^2\mbox{\boldmath $ r$}u(r ) - \!\mu^\ast . \label{e+}\ ] ] the leading term in the quantum width @xmath315 for electrons does not depend on @xmath237 , and it is still given by eq . ( [ width ] ) . to calculate the spectral density @xmath288 , one can use eq . ( [ sdfinal ] ) with @xmath307 and @xmath329 replaced by @xmath330 and @xmath331 respectively . the width of the electron spectral density is determined by the energy band ( [ e+ ] ) . the abrupt edge of the spectral density at @xmath332 is smeared only by a small quantum width @xmath315 , see fig . [ figds ] a. the minimal energy necessary to create an interstitial , @xmath333 , is of the order of energy @xmath334 needed to create a vacancy . therefore the difference @xmath335 exceeds significantly @xmath315 and thus determines the pseudogap in the spectral density ( [ sddef ] ) . using eqs . ( [ energies ] ) , ( [ e+ ] ) , and the asymptotic behavior of the hartree - fock potential @xmath336 , see sec . [ sec : td ] , we obtain the width of the pseudogap for different filling factors : @xmath337 } , \quad & n^{-3}r_s^{-2}\ll \delta\nu \ll n^{-1}\\ \displaystyle{\frac{\hbar\omega_c}{2n } \ln\left(nr_s\right ) } , \quad & n^{-1 } \ll \delta\nu \ll 1/2 , \end{array } \right .. \label{gapas}\ ] ] the leading term in the width of the gap agrees with the result of ref . for the broad range of filling factors @xmath338 . the overall dependence of the characteristic energy scales on the filling factor @xmath252 is shown in fig . [ figds ] b. it is important to notice that in the dominant region of filling factors , @xmath339 , the width of the gap @xmath55 is determined by the renormalized interaction ( [ potlong ] ) at distances @xmath340 , which are large compared to the characteristic distance between the electrons of pfll . the gap is insensitive to the short range correlations in the wave function of the ground state . therefore , we anticipate that the value of @xmath55 we found is robust : it is not an artifact of the maki - zotos trial function , but it should be the same ( with logarithmic accuracy ) for any state homogeneous on the macroscopic scale @xmath341 . in this section we analyze the simplest excitations of the pfll at @xmath342 . this problem is similar to the one studied in ref . . the only difference is that the interaction potential is renormalized due to the polarizability of the other landau levels . our goal is to study how this renormalization affects the spin excitations . the energy of the fully spin polarized ground state@xcite @xmath343 at @xmath344 is @xmath345 where @xmath278 is defined by eq . ( [ exch1 ] ) . the eigenstates of the hamiltonian with an extra electron or hole @xmath346 have the energies @xmath347 respectively . equations ( [ eh ] ) enable us to relate the width of the spin gap @xmath348 to the value of @xmath278 , i.e. @xmath349 . substituting @xmath66 in eq . ( [ gdef ] ) and using eq . ( [ exch1 ] ) , we obtain the result ( [ geff ] ) for the effective @xmath68-factor . it is worth noticing that the effective @xmath68-factor is independent on @xmath45 in the limit of a weak magnetic field due to the thomas - fermi screening@xcite . ( without screening , this factor would logarithmically diverge with @xmath350 . ) the exchange enhancement is maximal if the filling factor is odd . when the both spin sublevels are either empty or completely filled ( @xmath75 is even ) , the spin splitting is determined by the bare @xmath68 - factor@xcite . let us turn to the consideration of neutral excitations spin waves@xcite which at @xmath342 are described by the wave functions : @xmath351 wave functions ( [ swf ] ) are eigenstates of the hamiltonian ( [ he ] ) with the energies @xmath352 where the energy of the spin wave @xmath353 is given by @xmath354 ^ 2 e^{-q^2\ell^2/2+i{\bf kq}\ell^2}. } & \nonumber\end{aligned}\ ] ] for small wave vectors , @xmath355 , the spin wave has a quadratic dispersion relation which is typical for ferromagnets : @xmath356 at @xmath357 the energy of the spin wave is determined by the bare @xmath68-factor in agreement with larmor s theorem . it is worth mentioning that the long - wavelength domain of the spin wave spectrum is controlled by the bare coulomb interaction . this is because the spatial scale @xmath358 important for the spectrum , see eq . ( [ wave ] ) , is much smaller than the screening radius @xmath268 . in the region @xmath359 the energy of the spin wave reveals oscillatory behavior @xmath360 . \label{waveint}\ ] ] here @xmath361 is a smooth function with asymptotes @xmath362 spin waves with extremely large wave vectors , @xmath363 , correspond to almost independent electron and hole , see eq . ( [ eh ] ) . in this limit , the energy @xmath364 approaches @xmath66 : @xmath365 in this paper we constructed a theory describing low - energy excitations in a 2d electron liquid in a weak magnetic field . we have shown that all the excitations with energy smaller than cyclotron energy @xmath2 can be described by the effective hamiltonian acting in the fock space of the partially filled landau level only . starting from first principles , we obtained the explicit form of this hamiltonian by integrating out all the other degrees of freedom . armed with this effective theory , we have been able to make important predictions : + 1 ) we found that the _ thermodynamic density of states _ is negative for non - integer filling factors . for a broad range of filling factors , the value of the thermodynamic density of states was found to be independent on the value of magnetic field , see eq . ( [ thermdensity ] ) . this effect may be revealed in magnetocapacitance measurements@xcite in the weak magnetic field regime . + 2 ) the _ tunnelling density of states _ was shown to have a gap at the fermi level . for a broad range of filling factors , the width of the gap , see eq . ( [ gap2 ] ) , was shown to be consistent with that predicted by hydrodynamic approach of ref . . the gap can be observed in tunnelling experiments@xcite in a weak magnetic field . evidence of suppression of the tunnelling density of states for the filling factors @xmath366 was reported recently by turner _ this suppression may be associated with the gap predicted by our theory . however , the observed width of the gap is twice larger than that given by eq . ( [ gap2 ] ) . + 3 ) the _ exchange enhancement _ of the effective @xmath68-factor remains strong in the weak field regime , and at all odd filling factors it takes a universal value , see eq . ( [ geff ] ) . we found that the energy scale of charge excitations is parametrically smaller than the energy scale for spin excitations . this is qualitatively different from the situation at low filling factors , @xmath51 , where both these excitations are characterized by the same energy scale @xmath80 . none of the aforementioned effects could be obtained in the framework of the landau quantization of the spectra of quasiparticles in the conventional fermi liquid theory . this theory describes adequately the excitations with energies larger than @xmath2 . thus , our description of the low - energy properties of interacting 2d electron gas is complementary to the fermi liquid picture . helpful discussions with b. l. altshuler , b. i. halperin , a. i. larkin , l. s. levitov , yu . b. lyanda - geller , a. h. macdonald , b. i. shklovskii , and x. zhu are acknowledged with gratitude . we are thankful to f. w. j. hekking for critical reading the manuscript and valuable comments . this work was supported by nsf grant dmr-9423244 . integration in eq . ( [ polarization ] ) immediately yields @xmath368 , } \label{pol_arb_k}\end{aligned}\ ] ] where @xmath369 is the laguerre polynomial@xcite . further simplification is possible for @xmath370 . under this condition , we can use the asymptotic expression for the laguerre polynomial@xcite @xmath371^{m/2 } } , \label{laguerre}\ ] ] where @xmath372 is the bessel function . this expression is applicable if @xmath373 . using eq . ( [ laguerre ] ) and introducing a new index @xmath374 we approximate expression ( [ pol_arb_k ] ) as : @xmath375 ^ 2\right ] . } & \label{pol1}\end{aligned}\ ] ] the terms giving the main contribution to the sum in eq . ( [ pol1 ] ) are characterized by value @xmath376 . this enables us to approximate @xmath377 and perform the summation over @xmath378 . @xmath379 ^ 2 . \label{pol2}\ ] ] equation ( [ pol1 ] ) can be transformed into an integral form : substituting the identity @xmath380 ^ 2=\int_0^\pi \frac{dy}{\pi}\cos n y \ { \cal j}_{0}\left(2x\sin\frac{y}{2}\right ) \label{jsquared}\ ] ] into eq . ( [ pol1 ] ) and performing the summation over @xmath381 , one easily obtains : @xmath382 . \label{pol3}\ ] ] equation ( [ pol3 ] ) enables one to obtain the asymptotic form of the polarization operator in different regimes . in the static limit , @xmath383 , one finds from eqs . ( [ pol3 ] ) and ( [ jsquared ] ) : @xmath384 . \label{stat}\ ] ] for the high - frequency domain , @xmath385 , only a small vicinity of the point @xmath386 contributes to the integral in eq . ( [ pol3 ] ) , and @xmath387 coincides with the result for 2d electron gas in the absence of the magnetic field:@xcite @xmath388 . \label{dynamical}\ ] ] in the hydrodynamic limit , @xmath389 , one can expand the bessel function in eq . ( [ pol3 ] ) in a taylor series , which yields @xmath390 . \label{hydrodynamic}\ ] ] in the opposite limit , @xmath391 , the main contribution to the integral in eq . ( [ pol3 ] ) comes from the vicinities of points @xmath392 and @xmath393 . the calculation gives @xmath394 , } & \nonumber\\ & \displaystyle{b(q,\omega)=1 + \frac{{q^2v_f^2 } \omega_c^2\left(4\omega^2 - q^2 v_f^2\right)}{8\left(\omega^2+q^2v_f^2\right)^3}. } & \label{shortrange}\end{aligned}\ ] ] in this appendix we present the details of the derivation of the effective hamiltonian ( [ he ] ) . we start from eq . ( [ almostthere ] ) . we notice that operators @xmath179 and @xmath177 , see eqs . ( [ f1 ] ) and ( [ f2 ] ) , are integrals of the matsubara operators over the time interval @xmath395 $ ] . let us divide this time interval into @xmath217 small intervals @xmath396 , m=1,2\dots m-1 $ ] , where @xmath397 and @xmath398 . we imply that @xmath399 must be smaller than the characteristic time of the low - energy dynamics , but much larger than @xmath400 . then , the chronologically ordered exponent in eq . ( [ almostthere ] ) can be factorized as @xmath401 where operators @xmath402 and @xmath403 have the form @xmath404[\tau_{m}-\xi^0]\right ) , } \nonumber\\ & & \displaystyle{\hat{f}_2^{(m ) } = \!\int\ ! d^3\mbox{\boldmath $ \xi_1 $ } \int\ ! d^3\mbox{\boldmath $ \xi_2 $ } \left [ \bar{\psi}_n(\mbox{\boldmath $ \xi_1^+$})\psi_n(\mbox{\boldmath $ \xi_2 $ } ) - 2g_n(\mbox{\boldmath $ \xi_2 $ } , \mbox{\boldmath $ \xi_1 $ } ) \right ] \times } \nonumber \\ & & \displaystyle { \quad \phi(\mbox{\boldmath $ \xi_1 $ } ) \tilde{g}_{0}(\mbox{\boldmath $ \xi_1$},\mbox{\boldmath $ \xi_2 $ } ) \phi(\mbox{\boldmath $ \xi_2 $ } ) \theta\left([\xi^0_1-\tau_{m-1}][\tau_{m}-\xi^0_1]\right ) . } \label{f2a}\end{aligned}\ ] ] definitions ( [ f2a ] ) differ from the corresponding definitions ( [ f1 ] ) and ( [ f2 ] ) by the constraint on the time domain of integration . we do not need to impose an additional constraint on the time component of in the second of eqs . ( [ f2a ] ) because @xmath174 decays rapidly at @xmath405 , see eq . ( [ g0 ] ) . now , we substitute eq . ( [ split ] ) into eq . ( [ almostthere ] ) and perform the integration over the field @xmath149 . it yields @xmath406 where average @xmath407 of an arbitrary functional @xmath408 over the field @xmath149 is defined as @xmath409 the normalization factor @xmath410 describes the fluctuations of field @xmath149 around the saddle point : @xmath411 an explicit expression for @xmath412 is presented in subsection [ calomega ] . because @xmath399 is small , we can expand every factor in eq . ( [ lambdaap ] ) up to the first order in @xmath399 : @xmath413 this relation defines the operator @xmath414 . let us notice that this operator actually does not depend on @xmath415 because it contains the same number of creation and annihilation operators taken at the same moment of time . having this in mind , we substitute eq . ( [ expand ] ) into eq . ( [ lambdaap ] ) and obtain @xmath416 finally , we take the limit @xmath417 in eq . ( [ lambdaap1 ] ) and substitute the result in eq . ( [ heff ] ) . it yields @xmath418 now , one has to perform the actual calculation of the average in the left hand side of eq . ( [ expand ] ) in order to find the operator @xmath414 . it can be done by using perturbation theory . the leading terms in the small parameters @xmath193 and @xmath194 are @xmath419 to find the operator @xmath420 , only terms linear in @xmath421 should be retained . calculation of the right - hand side of eq . ( [ perturbation ] ) is carried out with the help of eqs . ( [ f2a ] ) and of the correlation functions of the field @xmath149 : @xmath422 the latter expression follows from eq . ( [ average ] ) and ( [ renorm ] ) . in the second of eqs . ( [ corphi ] ) , we explicitly separated a term which is singular at @xmath423 . in eqs . ( [ f2av ] ) , @xmath425 coincides with the well - known exchange correction to the chemical potential : @xmath426 where @xmath174 is defined by eq . ( [ g0 ] ) , and @xmath427 are the correlation corrections to the chemical potential found in the random phase approximation : @xmath428 when deriving eqs . ( [ muex ] ) , ( [ muc ] ) , we use eqs . ( [ equal_times ] ) and ( [ g0 ] ) . finally , substitution of eq . ( [ f2av ] ) into eq . ( [ perturbation ] ) enables us to find the operator @xmath414 . comparing eq . ( [ apanswer ] ) and eq . ( [ he ] ) , we obtain the expressions for the parameters of the hamiltonian @xmath114 : @xmath429 the following subsections are devoted to the explicit calculation of the parameters of the effective hamiltonian in terms of the filling factor and interaction strength . we will assume that condition ( [ cond3 ] ) is met . equations ( [ deltaomega ] ) , ( [ normalization ] ) and ( [ renorm ] ) yield @xmath430 with the help of eqs . ( [ fullepsilon ] ) , ( [ polarization ] ) and ( [ equal_times ] ) , one can transform eq . ( [ do ] ) into a more convenient form : @xmath431 expression ( [ omex ] ) corresponds to the first order exchange correction to the ground state energy , and @xmath432 . equation ( [ omc ] ) is the correlation correction to the ground state energy equivalent to the sum of ring diagrams@xcite . first , we evaluate the exchange energy ( [ omex ] ) . using the explicit expression for the green s function ( [ g0 ] ) , the property of the landau level wave functions , @xmath433 and the identity for the laguerre polynomials , @xmath434 we obtain @xmath435 ^ 2 . \label{omex1}\ ] ] integration in eq . ( [ omex1 ] ) can be easily performed , which yields for @xmath191 : @xmath436 . \label{omex2}\ ] ] the first term in brackets coincides with the exchange energy of 2d electron gas at zero magnetic field , and the second term appears due to the confinement of the electron wave functions by the magnetic field . the logarithmical factor @xmath437 arises due to the integration over relatively large spatial scales @xmath438 in eq . ( [ omex ] ) . on these scales , the screened interaction potential is significantly smaller than the bare potential appearing in eq . ( [ omex ] ) . therefore , it is plausible to anticipate that the correlation term of the thermodynamic potential , which accounts for the screening effect , should partially compensate the large logarithmic factor in eq . ( [ omex2 ] ) , and lead to the replacement @xmath439 in the argument of the logarithm . the calculation of the correlation energy ( [ omc ] ) explicitly demonstrates this . in zero magnetic field the contribution of the correlation energy@xcite is only of the order of @xmath440 . the situation changes in the magnetic field , where a contribution proportional to @xmath441 appears . this term arises mainly from integration over the domain of wave vectors @xmath442 in eq . ( [ omc ] ) , where we can use the asymptotic expression ( [ shortrange ] ) for @xmath172 . it yields with the logarithmic accuracy @xmath443 finally , with the help of eqs . ( [ omex2 ] ) and ( [ omec2 ] ) , we obtain for the thermodynamic potential ( [ omegai ] ) , @xmath444 \right\}. } & \label{omegafinal}\end{aligned}\ ] ] this result contradicts ref . where the contributions ( [ omex ] ) and ( [ omc ] ) were evaluated and a correction @xmath445 was obtained instead of @xmath446 . we start from the calculation of the exchange correction to the chemical potential . with the help of eqs . ( [ g0 ] ) , ( [ property ] ) and ( [ lid ] ) we transform eq . ( [ muex ] ) to the form @xmath447 after integration in eq . ( [ muex1 ] ) we obtain for @xmath448 with logarithmic accuracy : @xmath449 the correlation shift of the chemical potential , eq . ( [ mup ] ) , is calculated with the help of eqs . ( [ property ] ) and ( [ renpotential ] ) . with the logarithmic accuracy , the calculation gives @xmath450 it can be checked by an explicit calculation that the other correlation correction to the chemical potential , @xmath451 , see eq . ( [ muc ] ) , contains an additional small factor in comparison with eq . ( [ mup2 ] ) : @xmath452 , and , therefore , it can be neglected . finally , with the logarithmic accuracy , we have for @xmath198 appearing in eq . ( [ mustar ] ) the result : @xmath453 , edited by r. e. prange and s. m. girvin ( springer - verlag , new york , 1990 ) , 2nd ed . a. h. macdonald and s. m. girvin , phys . b * 33 * , 4009 ( 1986 ) ; r. morf and n. dambrumenil , preprint ( 1994 ) ; l. belkhir and j. jain , preprint ( 1994 ) . a. h. macdonald , phys . b * 30 * , 4392 ( 1984 ) ; d. j. yoshioka , j. phys . soc . japan , * 53 * , 3740 ( 1984 ) . e. g. _ g. baym and c. pethick , _ landau fermi - liquid theory : concepts and applications _ ( new york , 1991 ) . earlier , the electron system for the large integer filling factors was considered in : a. p. smith , a. h. macdonald and g. gumbs , phys . b * 45 * , 8829 ( 1992 ) . i. l. aleiner , h. u. baranger and l. i. glazman , phys . * 74 * , 3435 ( 1995 ) . t. k. lee and j. j. quinn , phys . b * 11 * , 2144 ( 1975 ) . l. v. keldysh , pisma zh . . fiz . * 30 * , 245 ( 1979 ) ( jetp lett . * 30 * , 224 ( 1979 ) ) . d. s. fisher , phys . b * 22 * , 1190 ( 1980 ) . s .- e. yang and a. h. macdonald , phys lett . * 70 * , 4110 ( 1993 ) ; y. hatsugai , p .- a . bares , and x. g. wen , phys . 71 * , 424 ( 1993 ) ; song he , p. m. platzman , and b. i. halperin , phys . lett . * 71 * , 777 ( 1993 ) ; y. b. kim and x. g. wen , phys . b * 50 * , 8078 ( 1994 ) . a. l. efros and f. g. pikus , phys . rev . b * 48 * , 14694 ( 1993 ) . p. johansson and j. m. kinaret , phys . lett . * 71 * , 1435 ( 1993 ) ; phys . b * 50 * , 4671 ( 1994 ) . t. ando and y. uemura , j. phys . japan * 37 * , 1044 ( 1974 ) . c. kallin and b. i. halperin , phys . b * 30 * , 5655 ( 1984 ) . in the coordinate representation the operators @xmath147 and @xmath454 have the form @xmath455 and @xmath456 respectively . g. d. mahan , _ many particle physics _ ( new york,1990 ) pp . 430 - 444 we believe that the electron system of the pfll has a maximal spin in the ground state . it was proven not to be the case for the lowest landau level [ see s. l. sondhi , a. karlhede , s. a. kivelson and e. h. rezayi , phys . b * 47 * , 16419 ( 1993 ) ] where the states with the filling factors close to @xmath457 are associated with creation of topological defects skyrmions at @xmath458 , or antyskyrmions at @xmath44 . however , it was shown recently [ x. g. wu and s. l. sondhi , preprint ( 1995 ) ] that the ground state does not contain skyrmions at any @xmath252 for the landau levels with higher indices . k. maki and x. zotos , phys . b * 28 * , 4349 ( 1983 ) . s. kivelson , c. kallin , d. p. arovas , and j. r. schrieffer , phys . b * 36 * , 1620 ( 1987 ) . h. bateman and a. erdelyi , _ higher transcendental functions _ , mcgraw - hill book co. , 1953 . l. bonsall and a. a. maradudin , phys . b * 15 * , 1959 ( 1977 ) . m. s. bello , e. i. levin , b. i. shklovskii , and a. l. efros , zh . 80 * , 1596 ( 1981 ) ( sov . jetp * 53 * , 822 ( 1981 ) ) . r. c. ashoori , j. a. lebens , n. p. bigelow and r. h. silsbee , phys . rev . lett . * 64 * , 681 ( 1990 ) ; phys . b * 48 * , 4616 ( 1993 ) ; j. p. eisenstein , l. n. pfeiffer , and k. w. west , phys . lett . * 69 * , 3804 ( 1992 ) ; surf . sci . * 305 * , 393 ( 1994 ) ; k. m. brown , n. turner , j. t. nicholls , e. h. linfield , m. pepper , d. a. ritchie , and g. a. c. jones , phys . b * 50 * , 15465 ( 1994 ) . n. turner , j. t. nicholls , k. m. brown , e. h. linfield , m. pepper , d. a. ritchie , and g. a. c. jones , preprint , ( 1995 ) . p. k. lam and s. m. girvin , phys . b * 30 * , 473 ( 1984 ) . previously , the effect of screening on the effective @xmath68 - factor was considered in refs . and , however , no explicit analytical expression was given . a. bychkov , s. v. iordanskii , and g. m. eliashberg , pisma zh . . fiz . * 33 * , 152 ( 1981 ) [ jetp lett . * 33 * , 143 ( 1981 ) ] . j. p. eisenstein , l. n. pfeiffer , and k. w. west , phys . b * 50 * , 1760 ( 1994 ) . the evaluation of the analogous sum was performed by s. simon and b. halperin , phys . b * 48 * , 17368 ( 1993 ) . f. stern , phys . lett . * 18 * , 546 ( 1967 ) . t. ando , a. b. fowler , and f. stern , rev . mod . phys . * 54 * , 437 ( 1982 ) . a. isihara and d. y. kojima , phys . b. * 19 * , 846 ( 1979 ) .

the development of the phase - field method has led to tremendous progress in the modelling of pattern formation during solidification , due to its capability to simulate complex time - dependent and three - dimensional morphologies with relatively simple numerical codes @xcite . the general principle of this method is to describe a multi - phase system by a set of phase fields which take constant values in each of the bulk phases and vary smoothly through interfaces of a characteristic thickness @xmath0 . the equations of motion for the phase fields and their coupling to the local thermodynamic state variables ( temperature , density , composition etc . ) can be obtained , following the basic principles of out - of - equilibrium thermodynamics , by taking a variational derivative of a free energy functional , which is generally of the ginzburg - landau type . mean - field approximations can be used to relate the parameters that appear in this functional to microscopic quantities , and the phase fields can often be interpreted as order parameters . generally , the equations that result from the straightforward application of these principles are not suitable for obtaining quantitatively accurate simulation results on solidification microstructures . the reason is that the characteristic natural thickness of the diffuse solid - liquid interfaces is a few times the interatomic distance , whereas solidification patterns typically exhibit length scales ranging from @xmath1 to @xmath2 @xmath3 m . even with the help of modern computers and multi - scale algorithms , both of these scales can not be resolved at the same time . therefore , in order to simulate solidification microstructures , the thickness of the diffuse interfaces in the phase - field model has to be artificially enlarged , sometimes by two or three orders of magnitude . quantitative results can only be expected if both the equilibrium and kinetic properties of the interfaces remain unchanged under this procedure . to achieve this goal , it is helpful to adopt a phenomenological point of view : the phase field is seen as a smoothed indicator function ( as opposed to a physical order parameter or density ) , and all equilibrium quantities and transport coefficients are interpolated between the phases with smooth functions of the phase fields that can be freely chosen . this freedom can be exploited to construct phase - field models with special properties . in particular , rescaling of the interface thickness is greatly simplified in models where bulk thermodynamics and interfacial properties can be controlled separately . the phase - field models for the solidification of a pure substance published in the literature are all quite similar @xcite , in the sense that they use the same set of fundamental fields ( a phase field and the temperature field ) , and that the structure of the equations is the same . one reason for this universality is that a simple and intuitive formulation of the model in terms of these fields yields indeed , as will be detailed below , a model in which this separation is achieved . therefore , the development of the `` thin interface limit '' @xcite which has paved the way to quantitative simulations of dendrites @xcite did not require a change in the model formulation . the situation is more complicated for alloy phase - field models . two different approaches with quite distinct philosophies have been pursued in parallel . the first , which will be called the `` coarse - graining '' approach in the following , generalizes the pure substance model by introducing a concentration field in addition to the temperature field and by writing down a free energy functional that depends on the phase field , the temperature , and the concentration @xcite . the local values of these three fields are in principle the coarse - grained counterparts of the microscopic structural order parameter , temperature , and concentration fields , and the interface is seen as a narrow region in space where all of these quantities can exhibit rapid spatial variations . in contrast , the second approach , called `` two - phase '' approach in the following , treats the interfaces as a mixture of two phases , each of which retains its bulk properties even inside the interface @xcite . in this approach , the phase field represents the volume fraction of one of the phases . moreover , a separate concentration field is defined for each phase , and the physical concentration field is obtained by a weighted average depending on the value of the phase field . the introduction of two separate concentration fields adds a supplementary degree of freedom which has to be removed from the problem ; this is done either by a specific partition relation @xcite or by the condition of local equilibrium between the coexisting phases @xcite . in the models of the `` coarse - graining '' type , the model structure generally leads to an intrinsic coupling between bulk and interface properties , which makes simulation results dependent on the chosen interface thickness . only for the case of dilute binary alloys , a specific interpolation of the thermodynamic properties through the interface has been developed @xcite which overcomes this constraint and makes quantitative simulations possible . in the two - phase approach , which is more phenomonological from the outset , bulk and interfacial properties are decoupled by construction . however , the removal of the extra degree of freedom introduced by the model formulation generally requires the solution of a nonlinear equation in each point of the interface , and thus adds significant computational complexity . the purpose of the present paper is to show that the coarse - graining approach can be easily extended to more complex alloy systems if instead of a free energy functional a grand - potential functional is used to generate the equations of motion . furthermore , an analysis of the resulting model shows that it is in fact perfectly equivalent to the two - phase model , which offers the possibility to reinterpret and simplify the latter . the fundamentals underlying these findings can be stated quite simply . the motion of interfaces is controlled by the transport of a conserved extensive quantity : energy for a pure substance , and chemical species for isothermal alloy solidification . in sharp - interface models , this fact is expressed by two separate laws : a transport equation in the volume and a conservation law of stefan type at the moving boundary . in contrast , the two - phase equilibrium at interfaces is controlled by the intensive quantity that is conjugate to the conserved one : temperature for pure substances and chemical potential for alloys . it turns out that the pure substance model has been formulated from the start in terms of a phase field and the intensive variable ( temperature ) , whereas alloy phase field models are traditionally formulated in terms of a phase field and the composition , which is a density of the extensive variable ( number of solute atoms ) . to obtain a model for alloys that has properties analogous to those of the pure substance model , it is sufficient to choose a formulation in terms of a phase field and the chemical potential , and to switch to the appropriate thermodynamic potential , which is the grand potential of course , the fact that the quantity analogous to the temperature is the chemical potential is well known @xcite and has been extensively used in sharp - interface models as well as in a few specific phase - field models @xcite , but its general consequences have so far not been fully appreciated in the framework of phase - field models . it should be mentioned that a second obstacle for obtaining quantitative results on alloys is the strong contrast of the solute diffusivities between solid and liquid , which generates spurious solute trapping when the interface thickness is scaled up . this problem was solved by the so - called antitrapping current , which was introduced first in the coarse - graining approach @xcite , but has also been incorporated in the two - phase model @xcite . since the results of the present paper do not introduce major changes on this point , the results of ref . @xcite will be taken over without a detailed discussion . after a change of variables , the model is almost identical to the model of ref . @xcite , such that the asymptotic analysis developed there remains valid . this will be illustrated by numerical simulations that explicitly test the independence of simulation results of the interface thickness for the case of a lens - shaped phase diagram . the remainder of the paper is structured as follows . in sec . [ sec_pure ] , the standard phase - field model for the solidification of pure substances is reviewd for reference . next , the grand potential formulation is introduced and motivated in sec . [ sec_grand ] , and illustrated by several examples in sec . [ sec_examples ] : a model with two parabolic free energy functions , a dilute alloy , and an alloy with a lens - shaped phase diagram . the relation of this model to other phase - field models is clarified in sec . [ sec_relations ] , and an example for numerical simulations is presented in sec . [ sec_numerics ] . finally , the implications of the present findings for the further developments of alloy phase field models are discussed in sec . [ sec_final ] . the minimal model of solidification is considered , which implies some standard simplifications : the densities of the solid and the liquid are taken to be equal , and heat transport is assumed to take place by diffusion only . as a consequence , no motion of matter needs to be considered , and the only transported extensive quantity is heat . under these assumptions , the state of an inhomogeneous two - phase system can be completely specified on a coarse - grained ( mesoscopic ) scale by two fields : a phase field @xmath4 which indicates the local state ( liquid or solid ) of matter , and the internal energy density @xmath5 . the conservation law corresponding to heat transport is the conservation of energy , which writes _ t e = - j_e . [ eq : heatconservation ] the heat current @xmath6 is given by fourier s law , j_e = - ( , e)t [ eq : heatcurrent ] where @xmath7 is the heat conductivity , which in general may depend on both variables , and @xmath8 is the local temperature . if @xmath5 is chosen as dynamic variable , the corresponding thermodynamic potential is the entropy . consequently , the system is naturally described by an entropy functional [ , e ] = _ v s(,,e , e ) where the local entropy density @xmath9 depends on the fields as well as their gradients . using the thermodynamic definition of the temperature , = , the evolution equations for the fields @xmath5 and @xmath4 can be written in a variational form , _ t = m _ _ t e = ( ( ) t ) = - ( ( ) t^2 ) . [ eq : eeq ] the first equation expresses the local maximization of the entropy , which occurs at a rate given by the constant @xmath10 ; the second is identical to the conservation law , eq . ( [ eq : heatconservation ] ) . while this is a perfectly viable starting point which has been explored by several authors @xcite , this formulation is rarely used in practice . simulations are almost always carried out with models formulated in the variables @xmath4 and @xmath8 that can be obtained from free energy functionals . besides historical reasons ( the first phase - field models for solidification were formulated in this language @xcite ) , there are also formal considerations which make this approach preferrable . the main reason to choose the intensive variable @xmath8 instead of the extensive variable @xmath5 is that it directly controls the two - phase equilibrium , which makes it easier and more intuitive to identify the driving forces in the model . this point will be illustrated by obtaining the equations of motion from a free energy functional that is constructed using a purely phenomenological point of view . it will be shown below that the standard phase - field model of solidification is easily obtained as a special case . let @xmath11 and @xmath12 be the free energy densities of pure solid and liquid , respectively , and let the corresponding equilibrium values of the phase field be @xmath13 . the free energy is given by [ , t]=_v f(,,t ) = _ v f_int ( , ) + g_s ( ) f_s(t ) + ( 1-g_s ( ) ) f_l(t ) , [ eq : ffunctional ] where the weighting function @xmath14 is given by g_s ( ) = , with @xmath15 a function that satisfies @xmath16 and @xmath17 ; hence , @xmath18 in the solid and @xmath19 in the liquid . the term @xmath20 is given by f_int = 12 ( ) ^2 + h f_dw ( ) , where @xmath21 and @xmath22 are constants of dimension energy per unit length and energy per unit volume , respectively , and @xmath23 is a double - well function with minima at @xmath24 . the motivations for this formulation are easily understood and are common to many phase - field models . the term @xmath20 creates domains where the phase field is close to its equilibrium values @xmath13 ( the minima of the double - well function ) , separated by diffuse interfaces . therefore , far from the interfaces , the free energy density reduces to the one of the corresponding bulk phase . the term @xmath20 contributes to the free energy only inside the interfaces ; this excess free energy represents the surface tension . the function @xmath14 interpolates between the two free energy densities through the diffuse interface . a variation of the free energy functional with respect to the two fields @xmath4 and @xmath8 yields & = & _ v \{-^2 + hf_dw ( ) + } ( x ) + & & + \{g_s ( ) + ( 1-g_s ( ) ) } t(x ) , [ eq : fvariation ] where the prime stands for derivation with respect to @xmath4 . the variation of @xmath25 with respect to @xmath4 is the driving force for the phase transition . two features are noteworthy : ( i ) the requirement that @xmath17 ensures that the driving force vanishes outside of the interfacial regions , and ( ii ) since , at the melting temperature @xmath26 , @xmath27 , the `` thermodynamical '' part of the driving force is identically zero , independently of the value of @xmath4 . the latter property implies that the equilibrium interface profile in @xmath4 can be calculated from the term @xmath20 alone . this can be shown by seeking the equilibrium solution for a planar interface along the @xmath28 direction , which can be obtained from the condition that the variation of @xmath25 with repect to @xmath4 must vanish . at @xmath26 , this condition yields -_xx + hf_dw ( ) = 0 [ eq : intgood ] which implies that the solution of this equation is independent of the free energies @xmath11 and @xmath12 . as a consequence , the surface free energy @xmath29 ( defined as the excess free energy due to the presence of the interface ) is also independent of the bulk free energies and is given by = i = i h w , [ eq : gammarelation ] where @xmath30 is a numerical constant that depends on the shape of the double well function @xmath31 , and the interface thickness @xmath0 is defined by w = . [ eq : wdef ] therefore , the surface tension @xmath29 and the interface thickness @xmath0 can be freely chosen by appropriately fixing the two constants @xmath21 and @xmath22 , independently of the bulk properties . as stated in the introduction , the interface properties can thus be controlled independently of the bulk thermodynamics . the equation of motion for the phase field is obtained from the free energy functional by the standard variational procedure , _ t = - m _ , [ eq : phiequation ] which expresses the fact that the system seeks to minimize its local free energy at a rate which is controlled by the constant @xmath10 . to obtain an evolution equation for the temperature field , the starting point is the observation that , by definition , the variation of @xmath25 with respect to @xmath8 is equal to the negative of the local entropy density , s(t , ) = - = - g_s ( ) - ( 1-g_s ( ) ) . [ eq : sdensity ] since @xmath20 was chosen independent of temperature , @xmath9 is a _ local _ function of @xmath4 and @xmath8 ( no gradients are involved ) and can be simply seen as the interpolation of the bulk entropy densities , s(,t ) = g_s ( ) s_s(t ) + ( 1-g_s ( ) ) s_l(t ) . [ eq : softandphi ] the use of the thermodynamic identity @xmath32 ( valid at constant density ) yields _ t e = t _ t s = t(_t + _ t t ) . [ eq : eevolution ] note that in writing down the second equality , it is assumed that there is no entropy production due to local dissipation , which is equivalent to the hypothesis that the transformations are reversible on the mesoscopic scale of a coarse - graining cell . furthermore , the definition of the specific heat per unit volume is ( , t ) = t = -t g_s ( ) - t ( 1-g_s ( ) ) . [ eq : cdef ] the equations ( [ eq : eevolution ] ) and ( [ eq : cdef ] ) can be combined with the energy conservation law , eq . ( [ eq : heatconservation ] ) , where the heat conductivity @xmath33 now depends on the variables @xmath4 and @xmath8 , to yield ( , t)_t t = - t _ t + , which is the desired evolution equation for the temperature field . this equation can be further simplified by writing the heat conductivity as the product of the specific heat and the thermal diffusion coefficient @xmath34 , and by using eq . ( [ eq : softandphi ] ) for the entropy density . the result is _ t t = \ { + t[s_l(t)-s_s(t)]_t}. [ eq : tequation ] note that all quantities that appear in this equation except for the thermal diffusivity can be obtained from the bulk free energy densities @xmath11 and @xmath12 . the free energy functional used in the standard formulation of the phase - field model can be obtained from eq . ( [ eq : ffunctional ] ) by linearizing the free energy density @xmath35 around the melting temperature @xmath36 , f(,,t ) = f(,,t_m ) + .|_t_m(t - t_m ) . using the definition of the latent heat per unit volume , @xmath37 $ ] , as well as the fact that the free energies of solid and liquid are equal at @xmath36 , @xmath27 , this expansion yields = _ v 12 ( ) ^2 + h f_dw ( ) + g()(t - t_m ) , [ eq : flinear ] where the constant @xmath38/2 $ ] has been disregarded for simplicity . the equation of motion for the phase field is then obtained from this linearized functional by a variational derivative . to obtain an equation of motion for the temperature , it is usually assumed that the specific heat is independent of both temperature and the phase field , @xmath39 . then , it is easy to `` guess '' the correct equation by realizing that the latent heat released or consumed during the phase transformation appears as a source term in the diffusion equation for the temperature , _ t t = ( d_t(,t)t ) + _ t . [ eq : tapprox ] this equation is indeed obtained from eq . ( [ eq : tequation ] ) when a constant specific heat is inserted and the approximation @xmath26 is made in the second term on the right hand side . note that , in contrast , the correct general form could not have been easily guessed from eq . ( [ eq : tapprox ] ) . the underlying reason is that the linearized free energy functional formally yields a specific heat which is zero since all second derivatives with respect to the temperature vanish ; therefore , thermodynamic consistency between the linearized functional , eqs . ( [ eq : flinear ] ) , and the evolution equation for the temperature , eq . ( [ eq : tapprox ] ) , has been lost . the equations are nevertheless correct , since for the case of constant specific heat and constant latent heat , that is @xmath40=l$ ] , the internal energy density and the temperature are linearly related , e(,t ) = + ( t - t_m)- l. [ eq : etlinear ] then , eq . ( [ eq : tapprox ] ) can be directly obtained by combining the time derivative of eq . ( [ eq : etlinear ] ) with the energy conservation law , eq . ( [ eq : heatconservation ] ) . a binary alloy is a mixture of two pure substances a and b. for simplicity , it is assumed here that the atomic volume @xmath41 of both pure substances and of the mixture are all the same , and that hence the total number density of the alloy is a constant and equal in solid and liquid . then , the only new field needed is the local composition ( atomic fraction ) @xmath42 of `` solute '' ( b ) atoms . furthermore , for constant atomic volume the chemical potentials of a and b atoms are not independent since removal of an a atom implies the addition of a b atom . this means that the only new intensive variable that needs to be considered is the chemical potential @xmath3 of the solute atoms . the starting point of the coarse - graining approach , as pioneered in refs . @xcite , is a free energy functional that depends on the variables @xmath4 , @xmath8 , and @xmath42 . the chemical potential is then defined as the functional derivative of the free energy functional with respect to @xmath42 . since @xmath42 is dimensionless , the chemical potential obtained by this procedure has the dimensions of energy per unit volume . this convention obscures the thermodynamic roles of the two variables : the relevant extensive variable from which a density should be defined is the number of b atoms . therefore , the variable that is analogous to the internal energy density in the pure substance model is the number density of b atoms , = , [ eq : rhodef ] where @xmath41 is the atomic volume ( the constant volume occupied by one a or b atom ) . then , the chemical potential defined by = = . [ eq : mudef ] has the dimension of an energy , as is standard in basic thermodynamics . since the nature of the variables is important for establishing the analogy between pure substance and alloy models , this convention for the chemical potential is adopted for the remainder of the paper . however , since it is customary to express free energy densities in terms of the composition rather than the number density , both @xmath43 and @xmath42 will be used in the following for ease of presentation , keeping in mind that the two variables are simply related by eq . ( [ eq : rhodef ] ) . the number density is a conserved quantity , which implies _ t = -j _ , [ eq : massconservation ] where @xmath44 is the mass current . for isothermal solidification , where the only thermodynamic driving force for mass diffusion is the gradient of the chemical potential , the mass current is given by j_= - m(,t , c ) , [ eq : masscurrent ] where @xmath45 is the atomic mobility . combining these equations and the definition of @xmath3 yields the equation of motion for @xmath43 , _ t = ( m(,t , c ) ) . [ eq : ceq ] it will now be shown that this approach generally leads to a model in which bulk and interface properties do not decouple . to this end , it is useful to start again from the variation of the free energy functional , now in the variables @xmath4 and @xmath43 . since isothermal solidification is considered , there is no variation with respect to temperature . in order to simplify the notations , the variable @xmath8 ( which becomes a simple parameter for isothermal solidification ) will be dropped from the free energy densities and the mobility from now on . the variation of @xmath25 is & = & _ v \{-^2 + hf_dw ( ) + } ( x ) + & & + \{g_s ( ) + ( 1-g_s ( ) ) } ( x ) . [ eq : fvariation2 ] a crucial difference which the pure substance case is obvious : there is no simple argument which ensures that the `` thermodynamic driving force '' term proportional to @xmath46 vanishes . indeed , for two - phase equilibrium in an alloy , both the free energy density and the concentration vary across the interface . the value of all these quantities in the bulk phases at two - phase coexistence are obtained from two conditions : ( i ) the chemical potential must be the same in both phases , and ( ii ) the grand - potential density @xmath47 must also be the same . given the curves of free energy versus composition , the graphical interpretation of these two conditions is the well - known common tangent construction . let us examine the consequence of these conditions for the phase - field model outlined above . since the two variables phase field and concentration vary through the interface , the equilibrium interface profile is given by two coupled nonlinear differential equations . one is obtained from the condition of constant chemical potential , which remains valid in the diffuse interface picture , and reads = g_s ( ) + ( 1-g_s ( ) ) = _ eq(t ) [ eq : mucondition ] where @xmath48 is the equilibrium value obtained from the common tangent construction . this equation defines an implicit relation between the composition @xmath42 and the phase field @xmath4 . the equation for the phase field , obtained as before from the condition that the variation of @xmath25 vanishes , is -_xx + hf_dw ( ) + = 0 , where @xmath49 denotes the `` thermodynamic part '' of the free energy density . obviously , this equation becomes identical to eq . ( [ eq : intgood ] ) only if the third term is identically zero . the physical meaning of this condition can be made transparent by remarking that , since the concentration and the phase field are not independent variables any more under the constraint of eq . ( [ eq : mucondition ] ) , the variation of @xmath50 with respect to @xmath4 , taking into account the constraint of eq . ( [ eq : mucondition ] ) , is = + = + . therefore , if = ( f_th-_eq)0 , [ eq : grandcondition ] the equation for the equilibrium phase field profile reduces to eq . ( [ eq : intgood ] ) ; in other words , the quantity @xmath51 must be constant through the interface . far from the interfaces , where @xmath20 does not contribute , this quantity is equal to the grand potential density . note that the common tangent construction implies that the two _ bulk _ values of the grand potential must be equal . however , for a general choice of free energy functions , there is no reason for this condition to be valid _ throughout the whole interface_. as a consequence , the interface equation and all quantities that are obtained from its solution ( surface tension , kinetic coefficients etc . ) depend on the bulk free energy densities . for realistic values of the interface thickness , this dependence is small , but when the interface thickness is upscaled , large errors can occur . this fact has been recognized by several authors , and so far two different strategies have been followed to cure this problem . the first is to develop specifically designed free energy functionals that satisfy the condition of eq . ( [ eq : grandcondition ] ) , but are valid only for certain choices of bulk free energies ( see below ) . the second strategy is the one of the two - phase model @xcite , in which two separate concentration fields , one for each phase , are used ; the supplementary degree of freedom is then eliminated in such a way that eq . ( [ eq : grandcondition ] ) is satisfied . the new idea put forward here is that a general solution to this problem can also be obtained in the coarse - graining spirit ( using a single concentration field ) when the model is derived from a grand - potential functional instead of a free energy functional . indeed , the model formulated in the variables @xmath4 and @xmath43 is equivalent to the pure substance model formulated in terms of @xmath4 and @xmath5 : it has the same variational structure [ compare eqs . ( [ eq : eeq ] ) and ( [ eq : ceq ] ) ] , and both @xmath43 and @xmath5 are densities of extensive variables . to obtain the equivalent of the more successful pure substance models formulated in the variables @xmath4 and @xmath8 , alloy models should be formulated in the variables @xmath4 and @xmath3 ; the corresponding thermodynamic potential is the grand potential . grand - potential functionals have been used extensively in classical density functional theory ( see @xcite for a review ) . in the context of phase - field models , a grand potential functional has been introduced to study solidification with density change @xcite . however , in all the cited works , the density is retained as the fundamental field that is used to evaluate the functional . in contrast , the grand potential in its role as a thermodynamic potential depends on the chemical potential . if the goal is to have a complete formulation of the problem in terms of the dynamical variable @xmath3 , the grand - potential functional @xmath52 should be a functional of the field @xmath3 . in thermodynamic equilibrium , this field is just a constant which is equal to the thermodynamic equilibrium chemical potential , but out of equilibrium @xmath3 can depend on space and time . therefore , the field @xmath43 that appears in the free energy density needs to be eliminated in favor of @xmath3 . this is simple if the values of @xmath43 and @xmath3 are related by a local and invertible function . a free energy functional of the form of eq . ( [ eq : ffunctional ] ) , taken with free energy densities that depend on @xmath8 and @xmath42 , is a good starting point since it contains no nonlocal terms in @xmath43 such as @xmath53 . moreover , for functions @xmath54 and @xmath55 that are convex in @xmath42 , the relation between @xmath3 and @xmath42 is monotonous and hence invertible . thus , it is possible to switch from @xmath42 to @xmath3 as the dynamic field . after this operation , the number density is not a fundamental free field any more , but is obtained as a local functional derivative of the grand - potential functional with respect to the local chemical potential , @xmath56}{\delta \mu}. \label{eq : rhodepdef}\ ] ] note that the above requirements ( no square gradient terms in @xmath42 and convex free energy functions ) implies that the present method can not be applied to systems that exhibit phase separation . in analogy with eq . ( [ eq : ffunctional ] ) , the grand - potential functional is & = & _ v ( , , ) + & = & _ v _ int ( , ) + g_s ( ) _ s ( ) + ( 1-g_s ( ) ) _ l ( ) , [ eq : gfunctional ] where @xmath57 is identical to @xmath20 , and the grand potential densities of the bulk phases are obtained by a legendre transform of the free energies , _ ( ) = f_(c)-(=s , l ) . [ eq : legendre ] this procedure can be easily performed for any convex free energy function , either analytically or numerically . note that this transformation implicitly uses the equivalence between statistical ensembles ( canonical and grand canonical ) on the mesoscopic scale . this is consistent with the general philosophy of the coarse - graining approach , which assumes that thermodynamic quantities can be defined on the scale of a coarse - graining cell . the variation of the grand - potential functional ( at constant temperature ) is & = & _ v \{-^2 + hf_dw ( ) + } ( x ) + & & + \{g_s ( ) + ( 1-g_s ( ) ) } ( x ) . [ eq : gvariation ] since , at solid - liquid coexistence , @xmath58 , the equilibrium interface equation obtained from the condition of vanishing variation with respect to @xmath4 is identical to eq . ( [ eq : intgood ] ) , as desired . furthermore , using eq . ( [ eq : rhodepdef ] ) , the variation of @xmath52 with respect to @xmath3 yields ( , ) = - = - g_s ( ) - ( 1-g_s ( ) ) , [ eq : rhoofmu ] which is the equivalent of eq . ( [ eq : sdensity ] ) for the entropy density . note that this can also be rewritten as ( , t , ) = g_s ( ) _ s(t , ) + ( 1-g_s ( ) ) _ l(t , ) [ eq : rhoofmuandphi ] with @xmath59 ( @xmath60 ) . it is useful to restate this equation in terms of @xmath42 for future use , c ( , ) = ( , ) = g_s ( ) c_s ( ) + ( 1-g_s ( ) ) c_l ( ) , [ eq : cofmuandphi ] where obviously @xmath61 . two - phase coexistence is characterized by a constant chemical potential , @xmath62 ; the corresponding composition profile through a solid - liquid interface is c_eq ( ) = c(,_eq(t ) ) = g_s ( ) c_s^eq(t ) + ( 1-g_s())c_l^eq(t ) . [ eq : cprofile ] the equations of motion for @xmath4 and @xmath3 are now formulated following the same steps as for the pure substance model . the phase field evolves toward a minimum of the grand potential , _ t= -m _ = m _ ; [ eq : phievolution ] this equation shows that the thermodynamic driving force for the phase transition is the difference in grand potential densities . the evolution equation for the chemical potential is obtained by taking the time derivative of eq . ( [ eq : rhoofmuandphi ] ) , which yields _ t = ( _ t + _ t ) . [ eq : rhoevolution ] it is useful to define the quantity ( , ) = = g_s ( ) + ( 1-g_s ( ) ) , [ eq : chidef ] which will play a role similar to the specific heat in the pure substance model . the symbol @xmath63 is chosen here because this quantity can be seen as a generalized susceptibility @xcite . furthermore , the mobility can be written as the product of @xmath64 and a solute diffusion coefficient @xmath65 . indeed , in the bulk phases , for monotonous ( and hence invertible ) functions @xmath66 , _ ( ) = = = , [ eq : darken ] which is the well - known thermodynamic factor ( darken factor @xcite ) . the combination of these definitions and eq . ( [ eq : rhoevolution ] ) with the mass conservation law , eq . ( [ eq : massconservation ] ) yields _ t= \{- _ t } , [ eq : muevolution ] an equation completely equivalent to eq . ( [ eq : tequation ] ) for the temperature in the pure substance model . the model is completely specified for _ any _ set of free energy functions for solid and liquid by the definitions of the grand - potential functional , eqs . ( [ eq : gfunctional ] ) and ( [ eq : legendre ] ) , and the evolution equations ( [ eq : phievolution ] ) and ( [ eq : muevolution ] ) . in order to illustrate some of its properties , it is useful to work out several explicit examples . first , it is shown that the equivalent of the linearized pure substance model is obtained from parabolic free energies . next , it will be shown that the dilute alloy model of ref . @xcite can be recovered using this formalism . finally , the more general case of an ideal solution model will be treated . the simplest phenomenological approximation for free energy functions for fixed equilibrium compositions @xmath67 and @xmath68 at some temperature @xmath8 are two parabolas , f_(c ) = 12 _ ( c - c_^eq)^2 , ( = s , l ) where @xmath69 and @xmath70 are constants with dimension energy per unit volume . the chemical potential in each phase is = = _ ( c - c _ ) , [ eq : muofc ] which can of course be inverted to yield @xmath42 as a function of @xmath3 in each phase , c= + c_^eq . [ eq : cofmu ] the grand potential densities are then obtained from the legendre transform , @xmath71 , where eq . ( [ eq : cofmu ] ) is used to switch variables from @xmath42 to @xmath3 , _ ( ) = -12 - c_^eq . of course , the use of the definition @xmath72 together with @xmath73 yields again eq . ( [ eq : cofmu ] ) . the equilibrium chemical potential for two - phase coexistence is obtained by the condition @xmath74 . the solution @xmath75 corresponds to the common tangent between the bottoms of the parabolic wells , and the equilibrium compositions are equal to @xmath67 and @xmath68 . for @xmath76 , a second solution exists which corresponds to a common tangent that is tilted and yields different values for the equilibrium compositions ; this solution is not of interest here . since @xmath75 , we have @xmath77 , that is , in this model @xmath3 is directly the deviation from the equilibrium value of the chemical potential . inserting these grand potential densities in eqs . ( [ eq : cofmuandphi ] ) and ( [ eq : chidef ] ) yields the expressions for @xmath42 and @xmath63 as a function of @xmath4 and @xmath3 , c ( , ) = + c_s^eqg_s()+c_l^eq(1-g_s ( ) ) [ eq : cofmuandphiparabolas ] ( , ) = g_s()+(1-g_s ( ) ) . it is easy to see that when @xmath78 , the resulting model is equivalent to the standard pure substance model . indeed , in this case the difference between the compositions in the two phases is independent of @xmath3 and identical to @xmath79 , and the susceptibility is just a constant , @xmath80 . this corresponds to the approximations of constant latent heat and constant specific heat , respectively . it is easily verified that eq . ( [ eq : cofmuandphiparabolas ] ) becomes identical to eq . ( [ eq : etlinear ] ) , with @xmath81 and @xmath63 replacing @xmath82 and @xmath83 , respectively . this analogy has been used in several phase - field models @xcite . the present formalism makes it possible to generalize this model and to use @xmath76 , that is , parabolas with different curvatures . the phase - field model for a dilute alloy of ref . @xcite starts from the free energy densities f_(t , c ) = f_^a(t ) + _ c + ( cc - c ) , (= s , l ) [ eq : fdensitydilute ] where @xmath84 and @xmath85 are the free energy densities of pure a , @xmath69 and @xmath70 are again constants with dimension energy per unit volume , @xmath86 is boltzmann s constant , and the last term in eq . ( [ eq : fdensitydilute ] ) is the dilute limit of the entropy of mixing . the calculation of the grand potential densities is straightforward and yields _ = f_^a - ( ) . the compositions as a function of the chemical potential are given by c _ [ eq : cofmudilute ] it is obvious that these compositions satisfy , for any value of the chemical potential , the partition relation c_s = k c_l with the partition coefficient @xmath87 given by k= ( ) . [ eq : kdef ] the equilibrium chemical potential for a given temperature is again obtained from the condition @xmath88 which yields f_s^a(t)-f_l^a(t)=. [ eq : mueqdilute ] the left hand side can be expanded for temperatures close to the melting temperature @xmath36 ; the right hand side can be rewritten in terms of @xmath89 and the partition coefficient . the result is ( t - t_m ) = c_l^eq(k-1 ) . if , on the right hand side of this equation , @xmath8 is approximated by @xmath36 , which is justified in the limit @xmath90 , the standard dilute alloy phase diagram is obtained , c_l^eq = , where @xmath91 is the liquidus slope . when these expressions are now used in the complete grand - potential functional , the composition as a function of @xmath4 and @xmath3 becomes c ( , ) = g_s ( ) ( ) + ( 1-g_s ( ) ) ( ) . [ eq : cofphiandmudilute ] a specific feature of the dilute alloy model is that this expression can be factorized into two functions that depend only on @xmath4 and @xmath3 , respectively . moreover , the latter can again be rewritten in terms of the liquid composition , which yields c(,)=c_l ( ) . [ eq : csimplified ] as a result , the expression for the susceptibility is also quite simple and reads ( , ) = ( ) = . [ eq : chidilute ] inserting these expressions in eqs . ( [ eq : phievolution ] ) and ( [ eq : muevolution ] ) leads to model equations identical to those of ref . the difference of the grand potential densities is @xmath92.\end{aligned}\ ] ] the free energy difference can be expressed in terms of the equilibrium chemical potential using eq . ( [ eq : mueqdilute ] ) , and after some algebra one obtains _ s()-_l ( ) = c_l^eq(1-k ) . as in ref . @xcite , two dimensionless variables are now introduced . the first , u = [ eq : udef ] is the dimensionless deviation of the chemical potential from its equilibrium value . this variable can also be expressed in terms of the composition and the equilibrium composition at two - phase coexistence using the definition of eq . ( [ eq : cprofile ] ) as u=. the second dimensionless variable , u = [ eq : udef ] is a dimensionless supersaturation . this can be seen by inserting the identity @xmath93 obtained from the preceding equation , which yields u = . [ eq : uofceq ] when the driving force ( the grand potential difference ) is expressed in either of these variables , the same evolution equation for the phase field as in ref . @xcite is obtained . matters are slightly more complicated for the second evolution equation . @xcite uses an evolution equation for the composition ( or , equivalently , for the dimensionless supersaturation @xmath94 ) rather than for the chemical potential . for the dilute alloy model , such an equation can be obtained starting from the general evolution equation of the chemical potential , eq . ( [ eq : muevolution ] ) , or ( more simply ) from the original formulation of the mass conservation law , eq . ( [ eq : massconservation ] ) . for this purpose , the chemical potential has to be eliminated in favor of @xmath42 or @xmath94 . this is possible because eqs . ( [ eq : udef ] ) and eq . ( [ eq : udef ] ) can be inverted and combined with eq . ( [ eq : csimplified ] ) to yield = _ eq + and c = c_l^eq . a straightforward calculation then yields the variational form ( without the antitrapping current ) of the evolution equation for the composition of ref . these steps will be discussed in more details later on ; here , it is important to stress that the possibility to switch from an evolution equation for @xmath3 to one for @xmath42 by an exact transformation is specific to the dilute alloy model : this procedure only works because the function @xmath95 is easily inverted , which is not the case in more general models , as will be seen below . in an ideal solution model , the free energy is a weighted average of the pure substance free energies of a and b plus an entropy of mixing term , f_(t , c)= ( 1-c)f_^a(t)+cf_^b(t)+ . [ eq : fideal ] contact with the notations of the previous examples can be made by setting @xmath96 ; the free energy becomes f_(t , c)= f_^a(t)+c_(t)+ . the chemical potential is = = _ + . this can be inverted to yield the conentration in each phase as a function of @xmath3 , c _ ( ) = . the grand potential densities are _ s , l = f_^a + ( 1-c ) = f_^a - . [ eq : omegadensitiesideal ] as before , these expressions can now be used to define the interpolated composition and the susceptibility , c(,)=g_s ( ) + ( 1-g_s ( ) ) , [ eq : cofphiandmuideal ] ( , ) = , where the latter has been expressed in terms of the functions @xmath97 because this leads to a simpler expression . the equations of motion for the ideal solution model are obtained by inserting these expressions in the general evolution equations , eqs . ( [ eq : phievolution ] ) and ( [ eq : muevolution ] ) . the equilibrium chemical potential and the phase diagram can be calculated analytically , but the resulting expressions are quite complicated and not of interest here . the important point here is that for this model it is not possible to transform the evolution equation for @xmath3 into one for the composition @xmath42 : whereas the functions @xmath98 for the composition in each phase as a function of @xmath3 can be easily inverted , the same is not true for the interpolated composition given by eq . ( [ eq : cofphiandmuideal ] ) . as a consequence , it is easy to obtain @xmath42 from @xmath3 for given @xmath4 and @xmath8 , but hard to obtain @xmath3 from @xmath42 . in the two - phase model @xcite , the interface region is seen as a phenomenological superposition of the two bulk phases , with a weighting function @xmath99 that interpolates between liquid ( @xmath100 ) and solid ( @xmath101 ) . the main difference to the coarse - graining approach is that the two - phase model starts with two separate composition fields for the solid and the liquid , @xmath102 and @xmath103 . the free energy density and the `` true '' composition ( in the coarse - graining sense ) are then written as f(,c ) = h_s()f_s(c_s ) + ( 1-h_s())f_l(c_l ) , c = h_s ( ) c_s + ( 1-h_s ( ) ) c_l . [ eq : cdefmixture ] since there are two equations but three variables ( @xmath4 , @xmath102 , and @xmath103 ) , an additional condition is needed to close the system : the chemical potentials of the two coexisting phases are required to be the same , = . [ eq : muequal ] taking this implicit relationship between @xmath102 and @xmath103 into account , the number of independent fields is reduced to two , and two evolution equations for , say , @xmath4 and @xmath42 can be written down . this approach is completely equivalent to the grand - potential formalism outlined above , with the difference that the dynamical variable is @xmath42 instead of @xmath3 . to see this , set @xmath104 . equation ( [ eq : cdefmixture ] ) becomes identical to eq . ( [ eq : cofmuandphi ] ) . furthermore , since the compositions @xmath102 and @xmath103 in eq . ( [ eq : cofmuandphi ] ) are defined as functions of the variable @xmath3 , eq . ( [ eq : muequal ] ) is automatically satisfied . finally , since all the equations of ref . @xcite are developed by analogy with the pure substance model , it is not surprising that the evolution equations , eqs . ( 31 ) and ( 32 ) of ref . @xcite are identical to the evolution equations ( [ eq : phievolution ] ) and ( [ eq : muevolution ] ) of the present paper once all the notations have been properly translated : the driving force for the phase transformation is the difference in grand potential density , and the quantity @xmath105 that appears in the evolution equation for the concentration in ref . @xcite is identical to @xmath106 here . in summary , the two - phase model can be obtained in a fully variational manner from a grand - potential functional . note that , in this point of view , the introduction of two separate composition fields is not necessary any more : the fundamental dynamical field is the chemical potential , and the two fields @xmath102 and @xmath103 are simply obtained as the derivatives of the bulk grand potential densities with respect to @xmath3 , whereas the `` real '' composition at any space point is given by the functional derivative of the grand - potential functional . at this point , it is useful to discuss the respective merits of the grand - potential and the two - phase approaches . at first glance , the former seems to be more advantageous : the number of fundamental fields is equal to two ( as compared to three for the two - phase model ) , and whereas @xmath3 has to be obtained from @xmath102 and @xmath103 in the two - phase model by solving eq . ( [ eq : muequal ] ) , @xmath102 and @xmath103 are obtained from @xmath3 by simple derivatives in the grand - potential formalism . however , an analysis in terms of computation time reveals that matters are not so simple . for the sake of concreteness , consider the simplest case of constant but distinct solute diffusivities in the solid and the liquid . in this case , the evolution equation for the composition reduces , in the bulk phases , to the simple linear diffusion equation . in contrast , the evolution equation for @xmath3 has several nonlinearities ( even in the bulk ) due to the presence of the factors @xmath64 , which are in general nonlinear functions of @xmath3 . this means that the numerical effort to integrate the equation for @xmath3 in the bulk is higher that the one for integrating an equation for @xmath42 . in the interfaces , the grand potential formulation does not require much additional effort , whereas the nontrivial eq . ( [ eq : muequal ] ) has to be solved in the two - phase model . however , since the interface regions usually represent only a small fraction of all the grid points in a numerical simulation , the equation of motion for @xmath3 may not always be advantageous from a computational point of view . the computational disadvantage of the grand - potential formalism in the bulk can be alleviated by an additional step , which also brings to light the direct relation of this approach to the phenomenological two - phase model developed by the access group @xcite . the idea is to make a change of variables and to replace the chemical potential field by another continous field that plays the same role . in order to obtain the standard diffusion equation for this field in the liquid phase , the appropriate field is the density ( or the composition ) in the liquid , @xmath107 or @xmath103 . indeed , under the hypothesis that @xmath108 is an invertible function of @xmath43 , the function @xmath109 is just its inverse function , according to the properties of the legendre transform . furthermore , @xmath110 is also a function of @xmath3 and can therefore expressed as a function of @xmath107 , @xmath111 eliminating @xmath3 in favor of @xmath107 in eq . ( [ eq : muevolution ] ) yields @xmath112 + \frac{g'(\phi)}{2}\left[\rho_s(\rho_l)-\rho_l\right]\partial_t\phi\right\},\ ] ] where @xmath113 is the susceptibility of the liquid phase . unsing the fact that @xmath114 by definition , this equation can be rewritten as @xmath115 } \times \nonumber \\ & & \biggl\{\vec\nabla\cdot\left[d(\phi,\rho_l)\chi_l(\rho_l)[1-(1-g_s(\phi)\chi_s/\chi_l)]\vec\nabla\rho_l\right ] + \nonumber \\ & & \quad \frac{g'(\phi)}{2}\left[\rho_s(\rho_l)-\rho_l\right]\partial_t\phi\biggr\}.\end{aligned}\ ] ] for a dilute alloy , @xmath116 and @xmath117 according to eqs . ( [ eq : kdef ] ) and ( [ eq : chidilute ] ) , and the above model becomes identical to the one of ref . the grand - potential formalism thus allows for a generalization of this model to arbitrary phase diagrams . further progress can be made by a simple approximation . the crucial point is the relation between composition and chemical potential : its nonlinearity penalizes the grand - potential formalism in the bulk and makes the resolution of eq . ( [ eq : muequal ] ) in the two - phase model non - trivial . the idea is thus to replace the exact relation between chemical potential and composition _ in the interfaces _ by an approximate one that will make it possible to write a simple equation for the composition . this approximation will be called _ local supersaturation approximation _ : it exploits the fact that , for slow solidification , the chemical potential in the interfaces is close to the value for two - phase equilibrium . this suggests to use a taylor expansion of the composition around the equilibrium composition profile in the interface , c - c_eq ( ) = .|__eq(-_eq ) = ( , _ eq)(-_eq ) [ eq : lsa ] which can be easily inverted to yield = _ eq + . furthermore , an expansion of the driving force around the equilibrium chemical potential yields _ s ( ) - _ l ( ) = ( -_eq ) . inserting these expression in the evolution equation for the phase field and the mass conservation law yields a simple set of equations for @xmath4 and @xmath42 , _ t= ^2-hf_dw -c [ eq : philsa ] _ t c = . [ eq : clsa ] note that the second equation can also be rewritten after applying the chain rule as _ t c = d ( ) c + , where @xmath118 . this latter form displays explicitly the two driving forces for solute diffusion that establish the equilibrium solute profile and that lead to solute redistribution out of equilibrium : composition gradients and differences between the thermodynamic potentials in the two phases . before proceeding further , it is useful to relate the quantities that appear in the above equations to the phase diagram of the binary alloy , characterized by the curves @xmath119 and @xmath120 , or equivalently by the coexistence line @xmath48 . the quantities @xmath121 are related to the liquidus and solidus slopes , @xmath122 , = = .|__eq = _ ^eq . moreover , the quantity @xmath123 can be evaluated using a clausius - clapeyron relation for the @xmath3-@xmath8 coexistence line , = - = - . combining these results yields _ s , l^eq = - . thus , the susceptibilities are inversely proportional to the liquidus and solidus slopes , and therefore @xmath124 . defining an effective partition coefficient by the ratio of the liquidus solpes @xmath125 , the evolution equations of the phase - field model can be further simplified . indeed , the susceptibility along the equilibrium profile is ( , _ eq ) = _ s^eqg_s ( ) + _ l^eq(1-g_s ( ) ) = _ l^eq[1-(1-k_m)g_s ( ) ] , and the evolution equations for @xmath4 and @xmath42 become _ ^2-hf_dw - [ eq : philsakm ] _ t c = . [ eq : clsakm ] these equations are very similar to the ones of the dilute alloy model , except that the partition coefficient @xmath87 has been replaced by the effective partition coefficient @xmath126 which depends on the temperature . in that sense , this approach bears some similarity with the method used in ref . @xcite to construct a quantitative phase - field model for arbitrary phase diagrams : the free energy curves are first approximated by a dilute alloy phase diagram with `` effective '' ( temperature - dependent ) partition coefficient , melting temperature , and liquidus slope ; the equations of motion for the dilute alloy model are then applied with these effective parameters . here , the approximation is directly in the evolution equations , and can be applied in a straightforward manner for arbitrary free energy functions . also note that the two appromixations are not completely equivalent . for instance , the effective partition coefficient in ref . @xcite is defined by the ratio of the compositions , @xmath127 , which is equal to the ratio @xmath128 only for a dilute alloy . it should also be noted that this approximation is _ not _ equivalent to the approximation of constant concentration gap @xmath81 ( equivalent to parallel liquidus and solidus slopes ) used in refs . this can be easily seen from eq . ( [ eq : lsa ] ) by considering a constant chemical potential deviation @xmath129 ( generated , for example , by a local curvature of the interface ) : the shifts in concentration on the two sides of the interface are proportional to @xmath130 , respectively , and hence inversely proportional to the liquidus slopes , as they should be . the above approximation has been called `` local '' for two reasons . first , this emphasizes the fact that the approximation of the relationship between composition and chemical potential is needed only in the interfaces , while it leaves the bulk evolution equations unchanged . second , it is anticipated that this method should be applicable to situations in which the temperature field varies slowly with time and over large length scales , such as in directional solidification or in thermosolutal models with realistic values of the lewis number . for this purpose , it should be sufficient to apply the above equations with the _ local _ value of the temperature for each point of the interface , which implies that @xmath81 , @xmath130 and @xmath126 are not constants but vary between different interface points . the next step is to incorporate into the model two features that have been widely used to increase the precision and performance of phase - field models . the first is motivated by the fact that a non - variational model can be more efficient for computational purposes than a strictly variational formulation . this was first highlighted by karma and rappel @xcite , and their method has since been used in many other models . in the terms of the present formulation , it amounts to keeping the interpolation function @xmath15 in the evolution equation of the phase field , but using a different interpolation function for the number density and the susceptibility . let @xmath131 be a function that has the property @xmath132 , and let h_s ( ) = . then , the concentration is defined as ( , ) = = _ s()h_s ( ) + _ l()(1-h_s ( ) ) [ eq : credef ] instead of eq . ( [ eq : rhoofmuandphi ] ) ; as a consequence , the equilibrium composition profile given by eq . ( [ eq : cprofile ] ) is also modified and becomes c_eq ( ) = c_s^eqh_s ( ) + c_l^eq(1-h_s ( ) ) . the susceptibility is still defined as the derivative of the number density with respect to the chemical potential and becomes ( , t,)== h_s ( ) _ s(t , ) + ( 1-h_s ( ) ) _ [ eq : chiredef ] the second feature is the so - called antitrapping current , which was developed in order to counterbalance spurious solute trapping @xcite . indeed , if the solute diffusivity in the solid is substantially lower than in the liquid , as is usually the case in alloy solidification , the upscaling of the interface thickness magnifies the solute trapping effect , whose magnitude is proportional to the interface thickness @xcite . to restore local equilibrium at the interface , as appropriate for low - speed solidification , an additional solute current is introduced which `` pushes '' the solute out of the freezing material , and which is given by j_at = aw[_l()-_s()]n _ t , where @xmath133 is the unit normal vector pointing from the solid to the liquid , @xmath0 is the interface thickness , and @xmath134 is a constant to be determined by a matched asymptotic analysis @xcite . the current thus defined is proportional to the interface velocity ( _ via _ the factor @xmath135 ) and to the composition difference between solid and liquid , and is directed from the solid to the liquid for a solidifying interface , for which @xmath136 . the mass conservation law , eq . ( [ eq : massconservation ] ) , reads now _ t = - ( j_+ j_at ) = ( m(,)- aw[_l()-_s()]n _ t ) . taking these modifications into account , the evolution equation for the chemical potential , eq . ( [ eq : muevolution ] ) , is replaced by @xmath137\hat n \partial_t\phi\right ] \nonumber \\ & & \qquad\qquad\quad \mbox { } - \frac{h'(\phi)}{2}\left[\rho_s(\mu)-\rho_l(\mu)\right]\partial_t\phi\biggr\};\end{aligned}\ ] ] the evolution equation for @xmath4 remains unchanged . in the local supersaturation approximation , the composition difference in the expression for the antitrapping current is approximated by its equilibrium value , @xmath138 . the evolution equation for the concentration , eq . ( [ eq : clsakm ] ) , is then replaced by _ [ eq : clsakmredef ] as usual , the parameters of the phase - field model have to be related to the quantities that appear in the sharp - interface theories by matched asymptotic analysis . the complete asymptotic analysis for the general eqs . ( [ eq : phievolution ] ) and ( [ eq : muevolution ] ) will be presented elsewhere . here , only the behavior of the model in the local supersaturation approximation will be discussed , because for its analysis the analogy to the dilute alloy model @xcite can be exploited . in order to apply directly the results of ref . @xcite , it is useful to cast the evolution equations in dimensionless form . from eqs . ( [ eq : philsa ] ) and ( [ eq : clsa ] ) , it is clear that the dimensionless variable that generalizes the quantity @xmath94 of the dilute alloy model defined by eq . ( [ eq : udef ] ) is u = = . indeed , this expression can be rewritten using the fact that @xmath139 $ ] , where @xmath140 is the standard partition coefficient , u = , which reduces to eq . ( [ eq : uofceq ] ) for a dilute alloy , since @xmath141 in this case . in terms of this variable , eq . ( [ eq : clsa ] ) becomes @xmath142\partial_t u & = & \vec\nabla\bigl\ { d(\phi)[1-(1-k_m)h_s(\phi ) ] \vec\nabla u \nonumber \\ & & \quad \mbox{}+ aw\hat n[1+(1-k_m)u]\partial_t\phi\bigr\ } \nonumber \\ & & \mbox{}+ [ 1+(1-k_m)u]\partial_t h_s(\phi).\end{aligned}\ ] ] a form formally identical to eq . ( 69 ) in ref . @xcite is obtained by choosing a particular interpolation for the diffusion coefficient @xmath143 , namely by setting d()[1-(1-k_m)h_s ( ) ] d_l q ( ) , where @xmath144 is the solute diffusivity in the liquid , supposed to be constant . note that , since the left hand side is actually the product of the diffusivity and the susceptibility , strictly speaking the function @xmath145 is an interpolation of the _ mobility _ rather than of the diffusion coefficient . the final result for the evolution equation is @xmath142\partial_t u & = & \vec\nabla\left ( d_l q(\phi ) \vec\nabla u + aw\hat n[1+(1-k_m)u]\partial_t\phi\right ) \nonumber \\ & & \mbox{}+ [ 1+(1-k_m)u]\partial_t h_s(\phi ) . \label{eq : ulikedilute}\end{aligned}\ ] ] this equation is indeed identical to the one used in the asymptotics of the dilute alloy model , except that the dilute alloy partition coefficient @xmath87 is replaced by the ratio of the solidus and liquidus slopes @xmath126 . this is very natural , since this quantity controls how the composition difference between solid and liquid depends on the chemical potential at the interface . the replacement of @xmath42 by @xmath94 in the evolution equation for the phase field leads to _ ^2-hf_dw - u. this equation is now divided by the constant @xmath22 , which amounts to non - dimensionalizing the free energy and grand potential densities ( since @xmath22 has dimension of energy per unit volume ) . furthermore , the function @xmath146 is now chosen to be the standard fifth - order polynomial @xmath147 , the function @xmath148 , and the double - well function to be @xmath149 . the resulting equation reads _ t= w^2 ^ 2 + -^3 -(1-^2)^2u , [ eq : phinodim ] where @xmath150 is the relaxation time for the phase field , @xmath151 is the interface thickness defined by eq . ( [ eq : wdef ] ) , and the constant @xmath152 is given by = . equation ( [ eq : phinodim ] ) is identical to the standard evolution equation for the phase field @xcite . as announced previously , the results of the asymptotic analysis of ref . @xcite can now be exploited since eqs . ( [ eq : ulikedilute ] ) and ( [ eq : phinodim ] ) are identically to the model analyzed in this reference . therefore , the variable @xmath94 obeys , in the liquid , the free boundary problem _ t u = d_l^2 u , u_int = - d_0k - v_n , [ eq : ugitho ] v_n = - d_l_n u|_int , [ eq : ustefan ] where @xmath153 and @xmath154 are the local curvature and interface velocity , respectively , @xmath155 is the capillary length , and @xmath156 the kinetic coefficient . equation ( [ eq : ugitho ] ) is the generalized gibbs - thomson equation , and eq . ( [ eq : ustefan ] ) is the stefan boundary condition that describes mass conservation at the phase boudary . in terms of the phase - field parameters , the capillary length and the kinetic coefficient are given by d_0 = a_1 [ eq : d0defpf ] = a_1(1-a_2 ) with @xmath157 and @xmath158 ; these values are identical to those obtained by karma and rappel @xcite . in terms of physical quantites , this expression for the capillary length is in fact identical to the standard thermodynamic definition @xcite . indeed , the number @xmath157 quoted above is equal to @xmath159 , where @xmath30 , the numerical constant defined in eq . ( [ eq : gammarelation ] ) , is equal to @xmath160 for the standard fourth - order double well potential used here . with the help of these relations , eq . ( [ eq : d0defpf ] ) can be rewritten as d_0 = . then , the use of eqs . ( [ eq : gammarelation ] ) and ( [ eq : darken ] ) yields d_0 = = . [ eq : d0def ] the relations found in the preceding section are now used to perform some illustrative simulations on a concrete model system . the ideal solution model is chosen , with the same parameters as those used in ref . @xcite to model the nickel - copper alloy . this alloy exhibits a typical lens - shape phase diagram with a rather narrow coexistence zone . concretely , the free energy densities given by eq . ( [ eq : fideal ] ) are used ; the free energy differences ( between solid and liquid ) of the pure substances are given by f_s^ni(t)-f_l^ni(t)=(t - t_m^ni ) , f_s^cu(t)-f_l^cu(t)=(t - t_m^ni ) , with the melting temperatures @xmath161 k and @xmath162 k and the latent heats @xmath163 j/@xmath164 and @xmath165 j/@xmath164 . the molar volume is taken to be @xmath166 @xmath164 , and the surface tension @xmath167 j/@xmath168 . for a temperature of @xmath169 k , the equilibrium concentrations are @xmath170 and @xmath171 , the partition coefficient is @xmath172 , the ratio of the liquidus slopes is @xmath173 , and the capillary length calculated by eq . ( [ eq : d0def ] ) is @xmath174 cm . isothermal dendritic solidification is simulated in two dimensions . the anisotropy needed to obtain stable dendritic growth is introduced in the standard way @xcite by making the surface tension dependent on the angle @xmath175 between the interface normal and a crystallographic axis , here chosen to coincide with the @xmath28 direction . a standard fourfold anisotropy , @xmath176 is used , with @xmath177 . the initial composition of the liquid is chosen as @xmath178 , which corresponds to a supersaturation of @xmath179 . four simulations are carried out with the equations of motion in the local supersaturation approximation , with @xmath180 , and @xmath181 , which corresponds to interface widths of @xmath182 and @xmath183 nm . for each simulation , the relaxation time of the phase - field equation is chosen such as to eliminate interface kinetics ( @xmath184 for all orientations ) . the steady - state growth velocity of the dendrites is measured as described in ref . @xcite , and the result is displayed in fig . [ figvel ] . it exhibits the behavior that is typical for quantitative phase - field models : the simulation results converge to a constant value with decreasing interface thickness @xmath0 , and the convergence is roughly quadratic in @xmath0 , which can be expected since all terms linear in @xmath0 have been eliminated by the model formulation . it is not surprising to find such behavior , since the model used here is essentially identical to the dilute alloy model for which rapid convergence with @xmath0 has been demonstrated @xcite . moreover , the parameters chosen here are in a region of the phase diagram in which the dilute approximation should still work quite well . however , this is not a limitation of the approach : simulations at @xmath185 k , where the ratio of the liquidus slopes @xmath186 is very different from the partition coefficient @xmath187 , yield a similar convergence plot . this shows that the model can be applied to alloys with arbitrary phase diagram . k , for different choices of the interface width @xmath0 . ] the most important conclusions of the present work can be summarized as follows . 1 . a phase - field model for alloy solidification has been obtained in a completely variational framework , starting from a phenomenological grand - potential functional that is a simple sum of bulk and interface contributions . in this model , the two dynamic variables are the phase field and the chemical potential field . a complete analogy can be established with the standard phase - field model for the solidification of a pure substance , in which the variables are the phase field and the temperature field . the main difference between the two sets of equations is that the relation between the composition and the chemical potential ( the extensive and the intensive variable ) is nonlinear , whereas the relation between temperature and internal energy is usually assumed to be linear . the resulting model is shown to be completely equivalent to the model of kim , kim and suzuki @xcite , with the chemical potential replacing the composition as the dynamic field . as a result , in the present model no `` partitioning '' of the solute into coexisting phases is necessary ; the ( auxiliary ) compositions in each of the phases can be simply obtained from the chemical potential . with an additional change of variables , the phenomenological access model @xcite can also be recovered and extended to general alloy phase diagrams . these developments show that these two - phase models , despite a seemingly quite different starting point , can in fact also be obtained by a coarse - graining procedure if the appropriate thermodynamic potential is used . 3 . the equations of the new model can also be written in terms of a phase field and the composition field . they thus have essentially the same computational complexity as the original phase - field models for alloy solidification @xcite derived in the coarse - graining framework . however , in contrast to the latter models , they have a decisive property which is required for quantitative simulations : bulk and interface thermodynamic properties can be adjusted independently . this difference in behavior is due to the different interpolations of the relevant thermodynamic potentials ( free energies in refs . @xcite , grand potentials here ) . 4 . with an additional approximation a linearization of the relation between chemical potential and composition inside the interfaces the model becomes equivalent to the quantitative dilute alloy model studied in refs . this feature makes it possible to include the additional antitrapping current as in that model , and to apply the detailed asymptotic analysis of ref . @xcite . as demonstrated for one particular case here , efficient and accurate simulations are thus possible for arbitrary alloy phase diagrams . numerous interesting perpectives for future work arise from the present results . first , the model has been worked out here for isothermal solidification only , but it can be easily extended to non - isothermal situations : coupled equations for the phase field , the temperature field and the chemical potential field can be developed by following the same steps as done here for each transport field separately , taking also into account off - diagonal elements in the onsager matrix of transport coefficients as well as cross - derivatives of the thermodynamic potentials . this is anticipated to yield a generalization of the thermosolutal model of ref . . moreover , the generalization of the formalism to multi - component systems should be straightforward . this is particularly interesting because general models for multi - component multi - phase solidification have been developed in the two - phase framework @xcite . in these models , the determination of the compositions in the individual phases for given global composition requires to solve a generalization of eq . ( [ eq : muequal ] ) , which represents a set of coupled nonlinear equations ( one for each component ) . a formulation in terms of the chemical potential would completely avoid this problem and thus potentially offer important gains in computational performance . it should also be mentioned that the two - phase approach has been used in other contexts . one example is the treatment of fluid flow in solidification @xcite and two - phase flows @xcite , where two separate velocity fields , one for each phase , are introduced , in contrast to `` coarse - graining '' models in which a single velocity field is used @xcite . it would be interesting to reassess the relations between these different models in the light of the present findings .

a common problem in statistical physics is that the probability distribution functions ( pdfs ) are always too complicated for practical calculations and we need to replace them by more tractable approximations . a possible solution is to identify a family of trial distributions @xmath0 , where x are parameters characterized systems and select the member of the family that is closest to the exact distribution @xmath1 . the problem , of course , is that it is not clear what one means by ` closest ' . one could minimize @xmath2 ^{2},\ ] ] but why this particular functional and not another ? and also , why limit oneself to an approximation by a single member of the trial family ? why not consider a linear combination of the trial distributions , some kind of average over the trial family ? but then , how should we choose the optimal weight assigned to each @xmath3 ? we propose to tackle these questions using the method of maximum relative entropy ( which we abbreviate as me ) and information geometry @xcite . the me method , which is developed in @xcite-@xcite , has historical roots in the earlier method of maximum entropy that was pioneered by e. t. jaynes and is commonly known as maxent @xcite . the me method is designed for updating probabilities from arbitrary priors for information in the form of arbitrary constraints and it includes bayes rule and the older maxent as special cases @xcite , @xcite . the purpose of this paper is to develop a me based method to generate optimal approximations ( brief accounts of some of the results discussed below have previously been presented in @xcite and @xcite ) . the general formalism , which is the main result of this paper , is developed in section [ sec2 ] . in section [ sec2 - 1 ] we justify the use of relative entropy as the unique and natural criterion to select the preferred approximation , which is labeled by some parameters . the optimal approximation is obtained in section [ sec2 - 3 ] by marginalizing over the variational parameters . a suitably weighed average over the whole family of trial distributions with the optimal weight provides an optimal approximation . in the second part of the paper we demonstrate the proposed formalism by applying it to simple classical fluids , a well studied field in the past barkerhenderson76-@xcite . to approximate the behavior of simple fluids we chose trial distributions that describe hard spheres @xcite-@xcite ( section [ sec3 ] ) . the me formalism is first used ( section [ sec4 ] ) to select the preferred value of the hard - sphere diameter . this is equivalent to applying the bogoliubov variational principle and reproduces the results obtained by mansoori et al . @xcite whose variational principle was justified by a very different argument . an advantage of the variational or the me methods over the perturbative approaches such as barker and henderson ( bh ) @xcite and of weeks , chandler and anderson ( wca ) @xcite is that there is no need for ad hoc criteria dictating how to separate the intermolecular potential into a strong short range repulsion and a weak long range attraction . on the other hand , a disadvantage of the standard variational approach is that it fails to take the softness of the repulsive core into account . at high temperatures this leads to results that are inferior to the perturbative approaches . in the standard variational a single preferred value of the hard - sphere diameter is selected . but , as discussed in @xcite and caticha00 , in the me method non - preferred values are not completely ruled out . this allows us ( in section [ sec5 ] ) to marginalize over hard sphere diameter to obtain an optimal hard - sphere approximation with a suitable weighting . that leads to significant improvements over the standard variational method . in section [ sec6 ] we test our method by comparing its predictions for a lennard - jones model for argon with molecular dynamics simulation data ( verlet68 , @xcite ) . we find that the me predictions for thermodynamic variables and for the radial distribution function are considerable improvements over the standard bogoliubov variational result , and are comparable to the perturbative results barkerhenderson76 @xcite . ( for a recent discussion of some of the strengths and limitations of the perturbative approach see @xcite . ) . despite the shortcomings of perturbation theory it remains very popular because it provides quantitative insights at a much lower computational cost than dynamical simulations . for recent applications to the glass transition and other more complex systems see @xcite - @xcite . although in this work the me is not applied to such complex problems one may fully expect that the information theory based me method will yield insights not only the thermodynamic behavior of complex systems but also about the approximation methods needed to analyze them . finally , our conclusions and some remarks on further improvements are presented in section [ sec7 ] . consider a system with microstates labeled by @xmath4 ( for example , the location in phase space or perhaps the values of spin variables ) . let the probability that the microstate lies within a particular range @xmath5 be given by the intractable canonical distribution @xmath6where @xmath7 is partition function , in which @xmath8 is the hamiltonian and free energy @xmath9 is defined through this partition function . the goal is to generate an approximation @xmath10 that is optimal in the sense that is `` closest '' to the `` exact '' distribution @xmath11 . it includes two steps . the first is to select a preferred trial distribution from a family of trials @xmath12 , each member of the trial family being labeled by one or more parameters @xmath13 . more generally , one could define the trial family in a non - parametric way by specifying various constraints . the second is to marginalize over parameters @xmath13 to obtain an optimal distribution . * relative entropy as the selection criterion . * the selection of the preferred approximation is achieved by ranking the distributions @xmath10 according to increasing _ preference _ , a real number @xmath14 $ ] which we call the `` entropy '' of @xmath15 . the numbers @xmath16 $ ] are such that if @xmath17 is preferred over @xmath18 , then @xmath19>\mathcal{d}[p_{2}]$ ] . thus , by design , the `` preferred '' approximation @xmath15 is that which maximizes the entropy @xmath16 $ ] . next we determine the functional form of @xmath16 $ ] . this is the general rule that provides the criterion for preference ; in our case it defines what we mean by the `` closest '' or `` preferred '' approximation . the basic strategy @xcite is one of induction : ( 1 ) if a general rule exists , then it must apply to special cases ; ( 2 ) if in a certain special case we know which is the best approximation , then this knowledge can be used to constrain the form of @xmath16 $ ] ; and finally , ( 3 ) if enough special cases are known , then @xmath20 $ ] will be completely determined . the known special cases are called the axioms of me and they reflect the conviction that whatever information was originally codified into the exact @xmath11 is important and should be preserved . the selected trial distribution should coincide with the exact one as closely as possible and one should only tolerate those minimal changes that are demanded by the information that defines the family of trials . three axioms and their consequences are listed below . detailed proofs and more extensive comments are given in caticha03 and @xcite . * axiom 1 : locality*. _ local information has local effects . _ if the constraints that define the trial family do not refer to a certain domain @xmath21 of the variable @xmath4 , then the conditional probabilities @xmath22 need not be revised , @xmath23 . the consequence of the axiom is that non - overlapping domains of @xmath4 contribute additively to the entropy : @xmath20=\int dq\,f(p(q))$ ] where @xmath9 is some unknown function . * axiom 2 : coordinate invariance . * _ the ranking should not depend on the system of coordinates . _ the coordinates that label the points @xmath4 are arbitrary ; they carry no information . the consequence of this axiom is that @xmath20=\int dq\,p(q)f(p(q)/m(q))$ ] involves coordinate invariants such as @xmath24 and @xmath25 , where the function @xmath26 is a density , and both functions @xmath27 and @xmath28 are , at this point , unknown . next we make a second use of axiom 1 ( locality ) . when there are no constraints at all and the family of trials includes the exact @xmath11 the selected trial should coincide with @xmath11 ; that is , the best approximation to @xmath11 is @xmath11 itself . the consequence is that up to normalization the previously unknown density @xmath26 is the exact distribution @xmath11 . * axiom 3 : consistency for independent subsystems*. _ when a system is composed of subsystems that are independent it should not matter whether the approximation procedure treats them separately or jointly . _ specifically , if @xmath29 , and the exact distributions for the subsystems , @xmath30 and @xmath31 , are respectively approximated by @xmath32 and @xmath33 , then the exact distribution for the whole system @xmath34 should be approximated by @xmath35 . this axiom restricts the function @xmath28 to be a logarithm . the overall consequence of these axioms is that the trial approximations @xmath36 should be ranked relative to the exact @xmath11 according to their ( relative ) entropy , @xmath37=-\int dq\,p(q)\log \frac{p(q)}{p(q)}. \label{s[p|p]}\]]the derivation has singled out the relative entropy @xmath38 $ ] as _ the unique functional to be used for the purpose of selecting a preferred approximation_. other functionals , may be useful for other purposes , but they are not a generalization from the simple cases described in the axioms above . * remark . * suppose a member of a family of trial canonical distributions @xmath12 with hamiltonian @xmath39 that are conditional probability distributions and depend on parameters @xmath40 , are given by @xmath41where @xmath42 , in which free energy @xmath43 is also defined . the preferred trial is then selected by maximizing @xmath44 $ ] . substituting eq . ( [ p(q ) ] ) and eq . ( [ ptheta(q ) ] ) into eq . ( [ s[p|p ] ] ) gives , @xmath45=\beta \left ( \langle h_{\theta } -h\rangle _ { \theta } -f_{\theta } + f\right ) \ , , \label{s[ptheta|p]}\]]where @xmath46 refers to averages computed with the distribution @xmath47 . the inequality @xmath44\leq 0 $ ] , can then be written as @xmath48thus , maximizing @xmath44 $ ] is equivalent to minimizing the quantity @xmath49 . this form of the variational principle and its use to generate approximations is well known . it is usually associated with the name of bogoliubov @xcite and it is the main technique to generate mean field approximations for discrete systems of spins on a lattice . the extent to which the preferred @xmath13 is preferred over other values ( @xcite , @xcite ) is expressed by the probability of @xmath13 , @xmath50 . the original me problem of assigning a probability to @xmath4 is now broadened into assigning probabilities to both @xmath4 and @xmath13 . in this section we use me again to find the preferred joint distribution @xmath51 . note that this is the kind of problem where the bayesian interpretation of probabilities is essential . within a frequentist interpretation it makes no sense to talk about @xmath50 or about @xmath52 because @xmath13 is not a random variable ; the value of @xmath13 is unknown but it is not random . to proceed we must ask a question . what is the prior distribution , that is , what do we know about @xmath4 and @xmath13 before the trial family is specified ? the joint prior @xmath53 can be expressed according to the product rule as , @xmath54 , where @xmath55 is conditional probability of observing system in state @xmath4 given parameter @xmath56 . our goal is to determine the preferred @xmath57 that is closest to the prior @xmath53 that reflects our initial knowledge about @xmath4 and ignorance about the @xmath13s . initially we know nothing about @xmath13 , not even how it is related to @xmath4 . the prior that represents this state of knowledge is a product@xmath58indeed , when @xmath53 is a product no correlations between @xmath13 and @xmath4 are introduced which means that information about @xmath4 tells us nothing about @xmath13 and vice versa . the first factor in @xmath53 reflects our prior knowledge about @xmath4 : the distribution for @xmath4 is known to be the exact @xmath11 . the second factor reflects our complete ignorance about @xmath13 : we choose @xmath59 to be as uniform as possible . our method applies whether @xmath13 is a discrete or a continuous variable . when @xmath13 is a continuous variable . then the uniform distribution @xmath60 is such that makes equal volumes in @xmath13 space equally likely . to define these volumes we apply method of information geometry @xcite and note that distances in @xmath61-space are uniquely defined because the @xmath13s are labels on probability distributions . cases where @xmath13 is a discrete variable are simpler . the relevant entropies involve sums over @xmath62 rather than integrals and the natural uniform distribution is @xmath63 . in what follows we concentrate on the more challenging continuous case . the unique distance between @xmath13 and @xmath64 is given by the fisher - rao metric @xcite , @xmath65 , where @xmath66accordingly , the volume of a small region @xmath67 is @xmath68 , where @xmath69 is the determinant of @xmath70 . up to an irrelevant normalization , the distribution @xmath59 that is uniform in @xmath13 is given by @xmath71 . the preferred approximation @xmath57 to the joint distribution @xmath72 is then obtained maximizing the entropy@xmath73=-\int dq\,d\theta \,p(\theta ) p_{\theta } ( q)\,\log \frac{p(\theta ) p_{\theta } ( q)}{\gamma ^{1/2}(\theta ) p(q ) } , \label{sigma[p]}\]]by varying @xmath50 subject to @xmath74 . the final result for the probability that @xmath13 lies within the small volume @xmath75 is @xmath76}\gamma ^{1/2}(\theta ) d\theta , \label{p(theta)}\]]where @xmath77 $ ] is given in eq . ( [ s[ptheta|p ] ] ) and @xmath78 is a normalization constant . note also that the density @xmath79 $ ] is a scalar function and the presence of the jacobian factor @xmath80 makes eq . ( [ p(theta ) ] ) manifestly invariant under changes of the coordinates @xmath13 . ( [ p(theta ) ] ) expresses the degree to which values of @xmath13 away from the preferred value are ruled out ; it tells us that the preferred value of @xmath13 is that which maximizes the probability density @xmath79 $ ] . finally , now that we have determined the preferred joint distribution @xmath81 we can marginalize @xmath13 and use the average @xmath82as the best approximation we can construct out of the given trial family . this approximation is expected to be better than any individual @xmath47 for the same reason that the mean is expected to be a better estimator than the mode it minimizes the variance . this concludes the first part of our paper . to summarize : our main results consist in the justification of the relative entropy eq . ( [ s[p|p ] ] ) as the uniquely natural functional to select the preferred approximations and the derivation of a quantitative measure of the degree to which the various trials are preferred , eq . ( [ p(theta ) ] ) . the final result for the best approximation is eq . ( [ pbar ] ) . next we illustrate how this me formalism is used in a specific example : simple fluids . a simple fluid composed of @xmath83 single atom molecules is described by the hamiltonian @xmath84where @xmath85 and the many - body interactions are approximated by a pair interaction , @xmath86 where @xmath87 . the probability that the positions and momenta of the molecules lie within the phase space volume @xmath88 is given by canonical distribution @xmath89where @xmath90 . for fluids dominated by pair interactions most thermodynamic quantities of interest can be written in terms of the one- and two - particle density distributions @xmath91where @xmath92and @xmath93the two - particle correlation function , @xmath94measures the extent to which the structure of liquids deviates from complete randomness . if the fluid is homogeneous and isotropic @xmath95 and @xmath96 where @xmath97 is the bulk density and @xmath98 is the radial distribution function ( rdf ) . then , the pressure is given by @xmath99where @xmath100 @xcite-@xcite . to account for the short - distance repulsion we consider a family of trials composed by distributions that describe a gas of hard spheres of diameter @xmath101 . for each @xmath102 the hamiltonian is @xmath103with @xmath104where @xmath105and the corresponding probability distribution is @xmath106the partition function and the free energy @xmath107 are @xmath108 @xmath109 @xmath110 two objections that can be raised for choosing @xmath111 as trials are , first , that they do not take the long - range interactions into account ; and second , that the actual short range potential is not that of hard spheres . these are points to which we will return later . a third objection , and this is considerably more serious , is that the exact hard - sphere rdf is not known . however , it can be calculated within the approximation of percus and yevick ( py ) for which there exists an exact analytical solution ( @xcite , @xcite , @xcite ) which is reasonably simple and in good agreement with numerical simulations over an extended range of temperatures and densities , except perhaps at high densities . there are several successful proposals @xcite to improve upon the py rdf but they also represent an additional level of complication . the simpler py rdf is sufficiently accurate for our current objective to illustrate the application and study the broad features of the me approach . the py rdf can be written in terms of the laplace transform of @xmath112 @xcite , @xmath113 } , \label{g(s)}\]]where @xmath114 is a dimensionless variable @xmath115 , @xmath116 , \]]@xmath117and @xmath118 is the packing fraction , @xmath119the rdf @xmath120 is obtained from the inverse transform using residues @xcite . the equation of state can then be computed in two alternative ways , either from the `` pressure '' equation or from the `` compressibility '' equation but , since the result above for @xmath120 is not exact , the two results do not agree . it has been found that better agreement with simulations and with virial coefficients is obtained taking an average of the two results with weights 1/3 and 2/3 respectively . the result is the carnahan - starling equation of state , @xcite - kalikmanov02 @xmath121the free energy , derived by integrating the equation of state , is @xmath122 , \mathrm{\ } \label{hs - free energy}\]]where @xmath123 , and the entropy is @xmath124it must be remembered that these expressions are not exact . they are reasonable approximations for all densities up to almost crystalline densities ( about @xmath125 ) . however , they fail to predict the face - centered - cubic phase when @xmath118 is in the range from @xmath126 up the close - packing value of @xmath127 . as discussed in section [ sec2 ] , the trial @xmath128 that is closest to the exact @xmath129 is found by maximizing the relative entropy @xmath130 $ ] , eq . ( [ s[p|p ] ] ) , with @xmath131 given by eq . ( [ phs ] ) and @xmath132 given by eq . ( [ exact dist ] ) . according to eq . ( [ f_u - general ] ) , it is equivalent to minimize @xmath133over all diameters @xmath102 , where @xmath134 is computed with @xmath111 . thus , the variational approximation to the free energy is @xmath135where @xmath136 is the preferred diameter . to calculate @xmath137 use @xmath138 \mathrm{~,}\]]where @xmath139 . but @xmath140 for @xmath141 while @xmath142 for @xmath143 , therefore @xmath144with @xmath145where we have assumed that the fluid is isotropic and homogeneous , @xmath146 , and introduced the hard - sphere rdf @xmath147note that the approximation does not consist of merely replacing the exact free energy @xmath9 by a hard - sphere free energy @xmath148 which neglects the effects of long range attraction ; @xmath9 is approximated by @xmath149 which includes attraction effects through the @xmath150 term in eq . ( [ f_u ] ) . this addresses the first of the two objections mentioned earlier : the real fluid with interactions given by @xmath151 is not being replaced by a hard - sphere fluid . the internal energy is approximated by @xmath152 and not by @xmath153 . to calculate @xmath150 it is convenient to write it in terms of @xmath154 , the inverse laplace transform of @xmath155 , @xmath156for example , for a lennard - jones potential , @xmath157 , \label{lj pot}\]]we have @xmath158 .\]]then , using equations eq . ( [ f_u ] ) and ( [ g(s ) ] ) gives @xmath159finally , it remains to minimize @xmath137 in eq . ( [ f_u ] ) to determine the preferred diameter @xmath136 . this is done numerically in an explicit example for argon in section [ sec6 ] . the me method as pursued in the last section has led us to determine a preferred hard - sphere diameter . it fails to take the softness of the repulsive core into account . as discussed in section sec2 - 3 , our best assessment of the distribution of @xmath160 is given by the marginal over @xmath102 , @xmath161 the corresponding best approximation to the rdf is obtained using eq . ( [ numberdensity ] ) , ( [ rdf - general ] ) , and ( [ rdf - hs ] ) @xmath162 by averaging over all hard - sphere diameters we are effectively describing a soft - core potential . since @xmath163 takes into account soft - core effects while @xmath164 does not , we expect that it will lead to improved estimates for all other thermodynamic quantities . however , we should emphasize that the distribution over the hard - sphere diameters @xmath165 is not being introduced in an ad hoc way in order to fix the variational method . the introduction of @xmath165 is mandated by the me method ( section [ sec2 - 3 ] ) . the distribution of diameters is given by eq . ( [ p(theta ) ] ) @xmath166 } } { \zeta } \gamma ^{1/2}\left ( r_{d}\right ) dr_{d}=\frac{e^{-\beta f_{u}}}{\zeta _ { u}}\gamma ^{1/2}\left ( r_{d}\right ) dr_{d } , \label{pd(rd)}\]]where @xmath167 = \beta \left ( f - f_{u}\right ) $ ] , the partition functions @xmath78 and @xmath168 are given by @xmath169and the natural distance @xmath170 in the space of @xmath102s is given by the fisher - rao metric , @xmath171 a convenient way to calculate the fisher - rao metric is to express it as a second derivative of the entropy eq . ( [ s[p|p ] ] ) of @xmath172 relative to @xmath173 , @xmath174 \right\vert _ { r_{d}^{\prime } = r_{d } } , \label{det g}\]]where @xmath175 = \beta \left [ \left . { r_{d}^{\prime } } ^{r_{d}}-\langle \left . u_{hs}\right\vert _ { r_{d}^{\prime } } ^{r_{d}}\rangle _ { r_{d}^{\prime } } \right ] , \]]and @xmath176 is the average over @xmath177 . as we argued above eq . ( [ f_u ] ) the expectation of the potential energy @xmath178 vanishes because the product @xmath179 vanishes for both @xmath180 and @xmath181 . similarly , @xmath182 when @xmath183 . however , when @xmath184 the expectation @xmath185 diverges . the divergence is a consequence of the unphysical nature of the hard - sphere model . for more realistic continuous potentials the distance between @xmath186 and @xmath102 is the same as the distance between @xmath187 and @xmath102 . we can then always choose @xmath188 and define @xmath189 in eq . ( [ det g ] ) as the limit @xmath190 . then , using eq . ( [ hs - free energy ] ) for @xmath148 , we have @xmath191 to summarize , the distribution of diameters @xmath165 is given by eq . ( [ pd(rd ) ] ) with @xmath137 given by equations ( [ f_u ] ) , ( [ hs - free energy ] ) , ( [ f_u - lt ] ) and @xmath192 given by eq . ( [ det g - exact ] ) . our best approximation to the `` exact '' @xmath193 is the @xmath194 given in eq . ( [ pbar(q ) ] ) . however , there is a problem . since the free energy @xmath137 is an extensive quantity , @xmath195 , for large @xmath83 the distribution @xmath196 is very sharply peaked at the preferred diameter @xmath136 . when choosing a _ single _ preferred diameter for a macroscopic fluid sample we find that me confers overwhelming probability to the preferred value . this is not surprising . the same thing happens when we calculate the global temperature or density of a macroscopic sample and yet local fluctuations can be important . the question then , is whether local fluctuations are relevant to the particular quantities we want to calculate . we argue that they are . from the very definition of @xmath197 as the probability that given an atom at a certain place another atom will be found at a distance @xmath198 , it is clear that @xmath197 refers to purely local behavior and should be influenced by local fluctuations . to the extent that the preferred diameter @xmath136 depends on temperature and density we expect that local temperature and/or density fluctuations would also induce local diameter fluctuations . for the purpose of calculating @xmath197 the system is effectively reduced to the small number of atoms @xmath199 in the local vicinity of the reference atom at the origin . in order to develop a systematic , fully me method for the determination of the effective number of particles @xmath199 that are locally relevant , one needs to use trial probability distributions that allow inhomogeneities in the hard sphere diameters . yet determination of such trial probability distributions requires further investigations . since we are interested in demonstrating me formalism for optimizing approximations in this work only , we consider a rather simple approach to estimate @xmath199 . based on our me approach to a mean field approximation for fluids @xcite , we have shown the rdf is given by @xmath200for a sufficiently dilute gas , the results of the mean field approximation are comparable with experimental results . the @xmath201 is approximately close to the true , exact rdf @xmath202 for @xmath203 , the lennard - jones parameter . a fluid of hard spheres gives @xmath204 for @xmath205 and can not reproduce the behavior of eq . ( [ estimaterule ] ) . however , once we recognize that we can use a statistical mixture , eq . ( [ pbar(q ) ] ) , we can tune the size @xmath206 of the cell and thereby change the width of @xmath207 so that the rdf @xmath208 of eq . ( [ ghs bar ] ) reproduces the known short - distance behavior of eq . ( [ estimaterule ] ) . one of the difficulties in testing theories about fluids against experimental data is that it is not easy to see whether discrepancies are to be blamed to a faulty approximation or to a wrong intermolecular potential . this is why theories are normally tested against molecular dynamics numerical simulations where there is control over the intermolecular potential . in this section we compare me results against simulation results @xcite for a fluid of monoatomic molecules interacting through a lennard - jones potential , eq . ( [ lj pot ] ) . the parameters @xmath209 and @xmath210 ( the depth of the well , @xmath211 , and the radius of the repulsive core , @xmath212 , respectively ) are chosen to model argon : @xmath213 @xmath214 and @xmath215 . * the free energy @xmath137 . * figure [ fig1].(a ) shows the free energy @xmath216 as a function of hard - sphere diameter @xmath102 for argon at a fixed density of @xmath217 for different temperatures . figure [ fig1].(b ) shows @xmath218 as a function of @xmath102 for several densities at fixed @xmath219 @xmath220 . since the critical point for argon is at @xmath221 @xmath220 and @xmath222 all these curves , except that at @xmath223 @xmath220 , lie well within the liquid phase . the increase of @xmath216 for high values of @xmath102 is due to short range repulsion between the hard spheres described by @xmath224 . the increase for low @xmath102 is due to the lennard - jones short - range repulsion as described by @xmath225 . the preferred @xmath102 is that which minimizes @xmath137 and depends both on temperature and density . the preferred diameter decreases as the temperature increases because atoms with higher energy can penetrate deeper into the repulsive core . the dependence with density is less pronounced . * the distribution of diameters @xmath165 . * in section [ sec5 ] we argued that the effective number of molecules that is relevant to the local structure of the fluid is not the total number of molecules in the system @xmath83 , but a smaller number , @xmath206 . in fig [ fig3].(a ) we plot the distribution of diameters @xmath226 for different temperatures , for a fixed fluid density of @xmath227 , and for an arbitrarily chosen @xmath228 . as expected the distribution shifts to higher diameters as the temperature decreases . notice also that the distribution becomes narrower at lower temperatures in agreement with the fact that a hard - sphere approximation is better at low @xmath229 @xcite . figure [ fig3].(b ) shows that increasing @xmath199 ( with fixed density @xmath97 ) decreases the width of @xmath165 ( solid lines ) and induces a slight shift of the whole distribution . this is due to the dependence @xmath230 of the fisher - rao measure @xmath231 in [ det g - exact ] . figure [ fig3].(b ) also explores the influence of @xmath232 by comparing the actual distributions @xmath165 ( solid lines ) with the distributions @xmath233 ( dotted lines ) which are obtained by setting @xmath234 in eq . ( [ pd(rd ) ] ) . the effect of @xmath235 is to shift the distribution slightly to higher @xmath102 . * the radial distribution function . * we are finally ready to calculate the radial distribution @xmath98 for argon . we start by estimating the number of molecules @xmath236 that are locally relevant ; as explained earlier we choose @xmath236 so that our best approximation @xmath208 , eq . ( [ ghs bar ] ) , reproduces the known short - distance behavior @xmath201 , eq . ( [ estimaterule ] ) , for @xmath237 . we have found that the estimates for @xmath238 need not be very accurate but that they must be obtained for each value of the temperature and density . in fig [ fig5 ] we show an example of the short - distance behavior of @xmath239 for three values of @xmath199 at @xmath219 @xmath220 and @xmath217 ; using a chi - square fit in the range from @xmath240 to @xmath241 @xmath242 the selected best value of @xmath199 is around 38000 . in figures [ fig6].(a)-(d ) we compare three different ways to calculate the rdf . the solid line is verlet s molecular dynamics simulation verlet68 ; it plays the role of experimental data against which we compare our theory . the dotted line is @xmath243 for the hard - sphere fluid with preferred diameter @xmath136 . this curve , calculated from inverse of eq . ( [ g(s ) ] ) , is also the result of the variational method and coincides with the me result for a macroscopically large @xmath244 . the dashed line is the averaged @xmath163 of the extended me analysis . figures . [ fig6].(a)-(c ) were plotted at three different temperatures @xmath219 , @xmath245and @xmath246 k at the density @xmath217 . figure [ fig6].(d ) we changed the density and the temperature to @xmath247 and @xmath248 @xmath249 . the agreement between the me curve and verlet s data is good . the vast improvement over the simpler variational method calculation is clear . one might be tempted to dismiss this achievement as due to the adjustment of the parameter @xmath199 but this is not quite correct : @xmath236 has not been adjusted , it has been calculated by fitting @xmath163 to optimal mean field rdf @xmath250 for @xmath203 . indeed , despite the fact that the hard - sphere trial solutions that we employ are mere approximations , the functional form of the whole curve @xmath163 in eq . ( [ ghs bar ] ) is reproduced quite well . however , one may note that the agreement between the me prediction and verlet s data becomes worse when the fluid density is decreased or the temperature is increased . the reason has been spelled out by in the studies of wca @xcite they demonstrate that both the repulsive and attractive forces contribute to the fluid structure when fluid is at low and moderate densities ( @xmath251 ) . however , when the fluid density is high enough ( @xmath252 ) , the repulsive force becomes dominant . because the hard - sphere approximation does not include the attractive force , the hard - sphere rdf does not take the attractive force into account , and this error propagates into our me prediction . the same discrepancy is also revealed in the wca theory for low density @xcite . * the equation of state . * finally we use the rdf to calculate the equation of state from the pressure equation , eq . ( [ eq of state ] ) . in fig [ fig10 ] we compare the equation of state derived from the @xmath197 obtained from verlet s simulation with calculations using the eme and variational methods and the perturbative theories of barker and henderson @xcite and of weeks , chandler and anderson @xcite , at @xmath253 @xmath249 . the eme results constitute a clear improvement over the plain variational calculation . for low densities all four methods agree with each other but differ from the simulation . a better agreement in this region would probably require a better treatment of two - particle correlations at long distances . at intermediate densities the best agreement is provided by the eme and bh results , while the wca theory seems to be the best at high densities . also shown in fig [ fig10].(a ) are experimental data on argon @xcite . the discrepancy between the experimental curve and the verlet simulation is very likely due to the actual potential not being precisely of the lennard - jones type . in fig [ fig10].(b ) we plot the eme equation of state for three different isotherms ( @xmath254 , @xmath255 and @xmath256 ) . to compare to the simulation of hansen and verlet @xcite we plot @xmath257 ( rather than @xmath258 ) as a function of density @xmath259 because this kind of plot exhibits the characteristic van der waals loop that signals the liquid - gas transition as the temperature drops . a more exhaustive exploration lies , however , outside the scope of this paper . the goal of this paper has been to use the eme method to generate approximations and show that this provides a generalization of the bogoliubov variational principle . this addresses a range of applications that lie beyond the scope of the traditional maxent . to test the method we considered simple classical fluids . when faced with the difficulty of dealing a system described by an intractable hamiltonian , the traditional approach has been to consider a similar albeit idealized system described by a simpler more tractable hamiltonian . the approach we have followed here departs from this tradition : our goal is not to identify an approximately similar hamiltonian but rather to identify an approximately similar probability distribution . the end result of the eme approach is a probability distribution which is a sum or an integral over distributions corresponding to different hard - sphere diameters . while each term in the sum is of a form that can be associated to a real hard - sphere gas , the sum itself is not of the form @xmath260 , and can not be interpreted as describing any physical system . as far as the application to simple fluids is concerned the results achieved in this paper represent progress but further improvements are possible by using better approximations to the hard - sphere fluid and by choosing a broader family of trial distributions . an important improvement would be to use trial probability distributions that allow inhomogeneities in the hard sphere diameters . this would lead to a systematic , fully eme method for the determination of the effective number of particles @xmath199 that are locally relevant . many perturbative approaches to fluids had been proposed , and a gradual process of selection over many years of research led to the optimized theories of bh and wca . the variational approach was definitely less satisfactory than these bestperturbation theories . with our work , however , the situation has changed : the eme - improved variational approach offers predictions that already are competitive with the best perturbative theories . and , of course , the potential for further improvements of the eme approach remains , at this early date , far from being exhausted . the authors acknowledge r. scheicher and c .- w . hong for their valuable assistance and advice with the numerical calculations . n. n. encov : _ statistical decision rules and optimal inference , transl . math . monographs , vol . 53 _ , am . ( providence , 1981);l . l. campbell , proc . am . ( 1986 ) 135 ; for applications to statistics see s. amari , _ differential - geometrical methods in statistics _ , springer - verlag ( new york , 1985 ) ; s. amari and h. nagaoka , _ methods of information geometry _ , oxford ( providence , 2000 ) ; for a brief derivation see a. caticha , change , time and information geometry , in : a mohammad - djafari ( ed ) , bayesian methods and maximum entropy in science and engineering , aip conf . proc . * 568 * , 2001 ( arxiv : math - ph/0008018 ) . a. caticha , relative entropy and inductive inference , in : g. erickson , y. zhai ( eds ) , bayesian inference and maximum entropy methods in science and engineering , aip conf . proc . * 707 * , 2004 ( arxiv : physics/0311093 ) ; a. caticha and a. giffin , updating probabilities , in : k.h . knuth , a. caticha , a. giffin , c. c. rodriguez , j. l. j. center ( eds . ) , bayesian inference and maximum entropy methods in science and engineering , aip conf . proc . * 954 * , 2007 ( arxiv : physics/0608185v1 ) . a. caticha , lectures on probability , entropy , and statistical physics , in maxen08 , so paulo , 2008 ( arxiv.org/abs/0808.0012 ) ; an application to data analysis is given in a. caticha and r. preuss , phys . e * 70 * ( 2004 ) 0461271 . e. t. jaynes , phys . rev . * 106 * ( 1957 ) 620 ; e. t. jaynes phys * 108 * ( 1957 ) 171 ; e. t. jaynes in : r d rosenkrantz ( ed ) , _ e. t. jaynes : papers on probability , statistics and statistical physics _ , dordrecht ( reidel , 1983 ) ; e. t. jaynes : _ probability theory : the logic of science _ , cambridge university press ( cambridge , 2003 ) . tseng , a. caticha , maximum entropy approach to a mean field theory for fluids . in : c. j. williams ( ed ) , bayesian inference and maximum entropy methods in science and engineering , aip conf . * 659 * , 2003 ( arxiv : cond - mat/0212198 ) . tseng , a. caticha , maximum entropy approach to the theory of simple fluids , in : g. erickson and y. zhai ( eds ) , bayesian inference and maximum entropy methods in science and engineering , aip conf . 707 * , 2004 ( arxiv : cond - mat/0310746 ) . caticha a , maximum entropy , fluctuations and priors , in : a mohammad - djafari ( ed ) , bayesian inference and maximum entropy methods in science and engineering , aip conf . proc . * 568 * , 2001 ( arxiv : math - ph/0008017 ) . as a function of hard - sphere diameter @xmath101 for argon at a density of @xmath261 for different temperatures . the best @xmath102 is that which minimizes @xmath137 . ( b ) : @xmath137 as a function of @xmath102 for argon at @xmath219 @xmath220 for different densities.,width=480,height=336 ] for argon for several temperatures at density @xmath261 for @xmath228 . ( b ) : @xmath262 for various @xmath238 at @xmath219 @xmath220 and @xmath263 . by setting @xmath264 ( dotted lines ) we see that the effect of the @xmath265 factor is to cause a slight shift of the distribution.,width=480,height=336 ] ; ( b ) verlet s molecular dynamics simulation ; and ( c ) the improved eme analysis , for argon at ( a ) : density @xmath263 , temperature @xmath219 @xmath267 , and effective particle number @xmath268 . ( b ) : @xmath263 , @xmath269 @xmath267 , and @xmath270 . ( c ) : @xmath261 , @xmath271 @xmath267 , and @xmath272 . ( d ) : @xmath273 , @xmath248 @xmath267 , and @xmath274.,width=192,height=192 ] ; ( b ) verlet s molecular dynamics simulation ; and ( c ) the improved eme analysis , for argon at ( a ) : density @xmath263 , temperature @xmath219 @xmath267 , and effective particle number @xmath268 . ( b ) : @xmath263 , @xmath269 @xmath267 , and @xmath270 . ( c ) : @xmath261 , @xmath271 @xmath267 , and @xmath272 . ( d ) : @xmath273 , @xmath248 @xmath267 , and @xmath274.,title="fig:",width=192,height=192 ] + ; ( b ) verlet s molecular dynamics simulation ; and ( c ) the improved eme analysis , for argon at ( a ) : density @xmath263 , temperature @xmath219 @xmath267 , and effective particle number @xmath268 . ( b ) : @xmath263 , @xmath269 @xmath267 , and @xmath270 . ( c ) : @xmath261 , @xmath271 @xmath267 , and @xmath272 . ( d ) : @xmath273 , @xmath248 @xmath267 , and @xmath274.,width=192,height=192 ] ; ( b ) verlet s molecular dynamics simulation ; and ( c ) the improved eme analysis , for argon at ( a ) : density @xmath263 , temperature @xmath219 @xmath267 , and effective particle number @xmath268 . ( b ) : @xmath263 , @xmath269 @xmath267 , and @xmath270 . ( c ) : @xmath261 , @xmath271 @xmath267 , and @xmath272 . ( d ) : @xmath273 , @xmath248 @xmath267 , and @xmath274.,width=192,height=192 ] @xmath267 . also shown are levelt s experimental results . ( b ) : @xmath275 versus the reduced density @xmath276 calculated using the eme method ( solid line ) and compared to the hansen - verlet simulation for three different isotherms . the graph shows the appearance of the liquid - gas van der waals loop as the temperature drops . , width=480,height=336 ]

more than 20 years ago , hong , ou , and mandel showed experimentally that mutually strongly quantum correlated ( entangled ) photon pairs can be emitted in the nonlinear process of parametric down - conversion @xcite at the single - photon level . for the occurrence of entangled photon pairs , the spontaneous character of the process is important . entanglement of two photons comprising a photon pair might occur for various physical quantities like frequencies , emission angles ( wave - vectors ) , or polarizations . perhaps most interestingly , entangled photon pairs manifest themselves in time domain , where both photons are detected within a relatively narrow time window . this is a direct consequence of the ` point ' character of the emission of a photon pair in the time domain . the width of the time window characterized by an entanglement time is typically on the order of several hundreds of fs , and is experimentally observable using hong - ou - mandel interferometer . in the time that has intervened since the original predictions , entangled photon pairs have been used for numerous experiments demonstrating both fundamental aspects of their physical properties @xcite and their potential for applications . these also include tests of bell inequalities @xcite , quantum teleportation @xcite , the generation of greenberger - horne - zeilinger states @xcite or quantum computation @xcite that can exploit photon pairs . quantum cryptography with photon pairs @xcite may be considered the most important application . we mention that applications to metrology have also been suggested @xcite . most researchers in the field have focused their attention for the most part on bulk nonlinear crystals , pumped by intense laser beams in type i and type ii configurations . bright sources of polarization - entangled photon pairs have been fabricated using two nonlinear type i crystals , mutually rotated by 90 degrees @xcite . the nonlinear processes in an optical cavity have also been used to enhance the photon - pair generation rate @xcite . in addition , periodically - poled materials may be used to increase the photon - pair generation rate in materials where phase matching is not naturally available @xcite . at present , even sources of photon pairs pumped by laser diodes have been developed @xcite . on the fabrication front , techniques of structures composed of nonlinear thin layers ( of width of several tens or hundreds of nm ) have developed to the point that useful structures with pre - defined properties may be achieved rather easily @xcite . these nonlinear photonic - band - gap structures @xcite are very promising as sources of photon pairs , as has recently been shown in @xcite . despite the small amount of nonlinear material embedded inside them , they can generate photon pairs with relatively high efficiency thanks to the constructive interference that involves both the pump and the down - converted fields , their spatial inhomogeneity nothwithstanding . the enhancement of the photon - pair generation rate has been predicted to be several hundreds and even thousands of times larger than photon - pair generation rates in nonlinear , bulk materials @xcite . moreover , the spectral and spatial characteristics of the down - converted fields depend on details of the structure , a fact that might be used to control the process , at least to some extent . for example , photon pairs with very narrow spectra may be obtained from suitable structures . we note that along the same vein , four - wave mixing in photonic - band - gap nonlinear fibers is also promising as a modern source of photon pairs @xcite . in this paper , we present a quantum model of photon - pair emission in a nonlinear , one - dimensional photonic - band - gap structure , based upon a perturbative solution of the schrdinger equation . this model extends those developed for bulk nonlinear materials in @xcite . it has been shown in @xcite that this approach is compatible with models based on methods of classical nonlinear optics ( multiple - scale spatial and temporal expansion methods ) with a specific kind of stochastic averaging concerning emitted spectra . the quantum model , however , provides a complete description of the nonlinear process . the paper is organized as follows . in sec ii , the model is presented in three steps . description of a quantum field in a layered medium in subsec . iia is followed by the description of nonlinear quantum interactions in subsec . iib . in subsec . iic measurable characteristics of the generated down - converted fields are determined . the behavior of physical quantities characterizing a photon pair are discussed in sec . iii , both for cw and pulsed pumping regimes . iv contains our conclusions . we consider a stack of @xmath0 nonlinear layers in which spontaneous parametric down - conversion may occur . a sketch of the system is shown in fig . the @xmath1th layer begins at @xmath2 and ends at @xmath3 , its length is denoted as @xmath4 , @xmath5 . linear indices of refraction of pump , signal , and idler fields in the @xmath1th layer are denoted as @xmath6 , @xmath7 , and @xmath8 , respectively . symbols @xmath9 , @xmath10 , and @xmath11 ( @xmath12 , @xmath13 , and @xmath14 ) mean indices of refraction in front of ( beyond ) the sample . the symbol @xmath15 is used for the nonlinear tensor of @xmath1th layer ; symbols @xmath16 , @xmath17 , and @xmath18 stand for pump- , signal- , and idler- field wave - vectors in the @xmath1th layer . the structure is pumped by a classical strong pump field that propagates under the angle @xmath19 with respect to @xmath20 axis ( see fig . [ fig2 ] ) . its wave - vector @xmath21 lyes in the @xmath22 plane , whereas its electric - field amplitude @xmath23 is perpendicular to the wave - vector @xmath21 . the pump - field positive - frequency electric - field amplitude at the input of the structure at @xmath24 is denoted as @xmath25 and can be conveniently described using a positive - frequency electric - field amplitude spectrum @xmath26 determined as follows : @xmath27 if we assume a gaussian time profile and a linear polarization of the incident electric - field amplitude @xmath28 in the direction rotated by an angle @xmath29 with respect to an incident te - wave polarization direction @xmath30 ( @xmath31 denotes an incident tm - wave polarization direction , see fig . [ fig2 ] ) we have : @xmath32 \nonumber \\ & & \mbox { } \times \xi_p \exp \left ( - \frac{1+ia_p}{\tau_p^2}t^2 \right ) \exp(-i\omega_p^0 t ) , \nonumber \\ & = & { \bf e}_{p_f , te}^{(+)}(z_0,t ) + { \bf e}_{p_f , tm}^{(+)}(z_0,t ) , \label{2}\end{aligned}\ ] ] where @xmath33 is the pump - pulse amplitude , @xmath34 pulse duration , @xmath35 central frequency , and @xmath36 denotes a chirp parameter of the pulse . the spectrum @xmath26 determined by eq . ( [ 1 ] ) is given as : @xmath37 \nonumber \\ & & \mbox { } \hspace{-2.5 cm } \times \xi_p \frac{\tau_p}{\sqrt{2(1+ia_p ) } } \exp \left ( - \frac{\tau_p^2}{4(1+ia_p ) } ( \omega_p-\omega_p^0)^2 \right ) \nonumber \\ & = & { \bf e}_{p_f , te}^{(+)}(z_0,\omega_p ) + { \bf e}_{p_f , tm}^{(+)}(z_0,\omega_p).\end{aligned}\ ] ] on the other hand , the following spectrum corresponds to cw pumping : @xmath38 \nonumber \\ & & \mbox { } \times \xi_p \delta(\omega_p-\omega_p^0 ) ; \label{4}\end{aligned}\ ] ] @xmath39 means a dirac delta function . the pump field incident on the structure is scattered at each boundary inside the structure , so as to achieve a certain profile along the @xmath20 axis , provided that the nonlinear interaction does not lead to pump - field depletion . scattering of the pump field is conveniently described using its decomposition into monochromatic waves . the positive - frequency electric - field amplitude @xmath40 of a monochromatic component at frequency @xmath41 with polarization @xmath42 ( @xmath43 te , tm ) can be written as follows @xcite : @xmath44 \nonumber \\ & & \mbox { } \hspace{-2cm}+ \sum_{l=1}^{n } { \rm rect}_{z_{l-1},z_l}(z ) \nonumber \\ & & \mbox { } \hspace{-2 cm } \times \left [ a_{p_f,\alpha}^{(l)}(\omega_p){\bf e}_{p_f,\alpha}^{(l)}(\omega_p ) \exp(i{\bf k}_{p , z}^{(l)}(z - z_{l-1 } ) ) \right . \nonumber \\ & & \mbox { } \hspace{-2cm}\left . + a_{p_b,\alpha}^{(l)}(\omega_p ) { \bf e}_{p_b,\alpha}^{(l)}(\omega_p ) \exp(-i{\bf k}_{p , z}^{(l)}(z - z_{l-1 } ) ) \right ] \nonumber \\ & & \mbox { } \hspace{-2cm}+ { \rm rect}_{z_n,\infty}(z ) \nonumber \\ & & \mbox { } \hspace{-2cm}\times \left [ a_{p_f,\alpha}^{(n+1)}(\omega_p ) { \bf e}_{p_f,\alpha}^{(n+1)}(\omega_p ) \exp(i{\bf k}_{p , z}^{(n+1)}(z - z_n ) ) \right . \nonumber \\ & & \mbox { } \hspace{-2cm}\left . + a_{p_b,\alpha}^{(n+1)}(\omega_p ) { \bf e}_{p_b,\alpha}^{(n+1)}(\omega_p ) \exp(-i{\bf k}_{p , z}^{(n+1)}(z - z_n ) ) \right ] ; \nonumber \\ & & \alpha = te , tm ; \label{5}\end{aligned}\ ] ] the function @xmath45 equals one for @xmath46 and is zero otherwise . polarization vectors of @xmath47-waves in @xmath1th layer are denoted as @xmath48 and @xmath49 for forward- and backward - propagating fields with respect to @xmath20 axis , respectively , and they are frequency dependent . the symbol @xmath50 denotes a @xmath20 component of the pump - field wave - vector @xmath51 in @xmath1th layer and is determined by the expression @xmath52 where the wave - vector @xmath51 propagates in the @xmath53th layer under the angle @xmath54 with respect to @xmath55 axis . the angles @xmath54 fulfill the snell law at the boundaries , i.e. @xmath56 @xmath57 . the symbol @xmath58 is identified as an electric - field amplitude of the pump field at frequency @xmath59 with polarization @xmath42 incident on the structure from the left - hand side , whereas the symbol @xmath60 describes an electric - field amplitude of the pump field at frequency @xmath41 with polarization @xmath42 incident from the right - hand side . the remaining amplitudes @xmath61 and @xmath62 are determined using relations at the boundaries and free - field propagation inside the layers : @xmath63 we note that the coefficients @xmath64 and @xmath65 for @xmath66 describe the corresponding electric - field amplitudes at the beginning of @xmath53th layer . assuming te and tm waves , the boundary transfer matrices @xmath67 and @xmath68 have the form : @xmath69 @xmath70 and @xmath71 . the free - field propagation matrices @xmath72 can be written as : @xmath73 the positive - frequency electric - field operators @xmath74 and @xmath75 for the signal and idler fields can be decomposed into te- and tm - wave contributions @xmath76 , @xmath77 , @xmath78 , and @xmath79 and expressed as follows @xcite : @xmath80 \exp(-i\omega_m t ) \nonumber \\ & & \hspace{-17 mm } = \hat{\bf e}_{m , te}^{(+)}(z , t ) + \hat{\bf e}_{m , tm}^{(+)}(z , t ) \nonumber \\ & & \hspace{-17 mm } = \frac{1}{\sqrt{2\pi } } \int_{0}^{\infty } d\omega_m \hat{\bf e}_m^{(+)}(z,\omega_m ) \nonumber \\ & & \hspace{-17 mm } = \frac{1}{\sqrt{2\pi } } \int_{0}^{\infty } d\omega_m \left [ \hat{\bf e}_{m , te}^{(+)}(z,\omega_m ) + \hat{\bf e}_{m , tm}^{(+)}(z,\omega_m ) \right ] , \nonumber \\ & & m = s , i . \label{11}\end{aligned}\ ] ] the permittivity of vacuum is denoted as @xmath81 ; @xmath82 means speed of light in vacuum , @xmath83 is the reduced planck constant , and @xmath84 denotes the area of the transverse profile of a beam . the symbols @xmath85 and @xmath86 mean polarization vectors of mode @xmath87 of @xmath42 wave propagating forward and backward with respect to the @xmath20 axis . annihilation operators of signal [ @xmath88 , @xmath89 and idler [ @xmath90 , @xmath91 photons in @xmath42 wave introduced in eq . ( [ 11 ] ) can be expressed in the same way as the pump - field amplitude @xmath92 : @xmath93 \nonumber \\ & & \hspace{-8 cm } \mbox { } + \sum_{l=1}^{n } { \rm rect}_{z_{l-1},z_l}(z ) \nonumber \\ & & \hspace{-8 cm } \mbox { } \times \left [ \hat{a}_{m_f,\alpha}^{(l)}(\omega_m ) { \bf e}_{m_f,\alpha}^{(l)}(\omega_m ) \exp(i{\bf k}_{m , z}^{(l)}(z - z_{l-1 } ) ) \right . \nonumber \\ & & \hspace{-8 cm } \mbox { } \left . + \hat{a}_{m_b,\alpha}^{(l)}(\omega_m ) { \bf e}_{m_b,\alpha}^{(l)}(\omega_m ) \exp(-i{\bf k}_{m , z}^{(l)}(z - z_{l-1 } ) ) \right ] \nonumber \\ & & \hspace{-8cm}\mbox { } + { \rm rect}_{z_n,\infty}(z ) \nonumber \\ & & \hspace{-8 cm } \mbox { } \times \left [ \hat{a}_{m_f,\alpha}^{(n+1)}(\omega_m ) { \bf e}_{m_f,\alpha}^{(n+1)}(\omega_m ) \exp(i{\bf k}_{m , z}^{(n+1)}(z - z_n ) ) \right . \nonumber \\ & & \hspace{-8cm}\mbox { } \left . + \hat{a}_{m_b,\alpha}^{(n+1)}(\omega_m ) { \bf e}_{m_b,\alpha}^{(n+1)}(\omega_m ) \exp(-i{\bf k}_{m , z}^{(n+1)}(z - z_n ) ) \right],\nonumber \\ & & \hspace{-6 cm } m = s , i ; \;\ ; \alpha = te , tm . \label{12}\end{aligned}\ ] ] the polarization vectors @xmath94 and @xmath95 give the polarization directions of mode @xmath96 with @xmath42 wave in @xmath1th layer propagating forward and backward , respectively , whereas @xmath97 is a @xmath20 component of the wave vector @xmath98 of this mode : @xmath99 the angle @xmath100 characterizes the direction of propagation of mode @xmath87 with respect to the @xmath20 axis . the angles @xmath101 are given by the snell law at the boundaries , i.e. @xmath102 @xmath103 , where @xmath104 stands for the angle of incidence of mode @xmath87 . the operators @xmath105 and @xmath106 obey the following commutation relations : @xmath107 & = & \delta_{\alpha,\alpha ' } \delta_{m , m ' } \delta_{l , l ' } \delta(\omega_m-\omega'_m ) , \nonumber \\ { } [ \hat{a}_{m_f,\alpha}^{(l)}(\omega_m ) , \hat{a}_{m'_f,\alpha'}^{(l')}(\omega'_m ) ] & = & 0 , \nonumber \\ { } [ \hat{a}_{m_b,\alpha}^{(l)}(\omega_m ) , \hat{a}_{m'_b,\alpha'}^{(l')\dagger}(\omega'_m ) ] & = & \delta_{\alpha,\alpha'}\delta_{m , m ' } \delta_{l , l ' } \delta(\omega_m-\omega'_m ) , \nonumber \\ { } [ \hat{a}_{m_b,\alpha}^{(l)}(\omega_m ) , \hat{a}_{m'_b,\alpha'}^{(l')}(\omega'_m ) ] & = & 0 , \nonumber \\ { } [ \hat{a}_{m_f,\alpha}^{(l)}(\omega_m ) , \hat{a}_{m'_b,\alpha'}^{(l')}(\omega'_m ) ] & = & 0 , \nonumber \\ { } [ \hat{a}_{m_f,\alpha}^{(l)}(\omega_m ) , \hat{a}_{m'_b,\alpha'}^{(l')\dagger}(\omega'_m ) ] & = & 0 ; \nonumber \\ & & \hspace{-4.5 cm } m , m'=s , i ; \;\ ; \alpha = te , tm ; \ ; \ ; l , l'=0,\ldots , n+1 .\end{aligned}\ ] ] the photonic - band - gap structure imposes the following relations among the operators @xmath105 and @xmath106 acting in @xmath1th layer : @xmath108 transfer matrices @xmath109 at the boundaries and free - field propagation matrices @xmath110 for @xmath111 and @xmath112 are defined in the same way as those given in eqs . ( [ 9 ] ) and ( [ 10 ] ) for the pump - field amplitudes . the hamiltonian @xmath113 describing spontaneous parametric down - conversion in a nonlinear medium of volume @xmath114 at time @xmath115 can be written as : @xmath116 , \label{17}\end{aligned}\ ] ] where @xmath117 denotes a third - order tensor of nonlinear coefficients , the symbol @xmath118 is shorthand of the tensor @xmath119 with respect to its three indices , and @xmath120 stands for a hermitian conjugated term . the negative - frequency electric - field operators @xmath121 for @xmath122 have been introduced in eq . ( [ 17 ] ) ( @xmath123 ) . decomposing the interacting fields into te and tm waves , and using the inverse fourier transformation in eq . ( [ 17 ] ) , we arrive at : @xmath124 . \label{18}\end{aligned}\ ] ] integrations over the variables @xmath125 and @xmath126 in eq . ( [ 17 ] ) impose conditions for @xmath125 and @xmath126 component of wave - vectors ; @xmath127 [ @xmath128 is assumed ] and @xmath129 . the latter @xmath39-function provides the following relation between the angles @xmath130 and @xmath131 of modes @xmath132 and @xmath133 in the @xmath1th layer : @xmath134 , \nonumber \\ & & l=0,\ldots , n+1 . \label{19}\end{aligned}\ ] ] solution of the schrdinger equation to first order in nonlinear perturbation together with the assumption of incident vacuum state @xmath135 in signal and idler fields provides the output state @xmath136 of signal field with @xmath137 polarization and idler field with @xmath138 polarization in the form : @xmath139 \nonumber \\ & & \hspace{-1cm}\mbox { } \times { \rm sinc}\left[\frac{1}{2}(k_m^{(l ) } - k_n^{(l ) } - k_o^{(l ) } ) l_l \right ] \delta(\omega_p-\omega_s-\omega_i ) \nonumber \\ & & \hspace{-1 cm } \mbox { } \times \hat{a}_{n,\beta}^{(l)\dagger}(\omega_s ) \hat{a}_{o,\gamma}^{(l)\dagger}(\omega_i ) |{\rm vac } \rangle . \label{20}\end{aligned}\ ] ] the wave - vectors @xmath140 introduced in eq . ( [ 20 ] ) are defined as @xmath141 and @xmath142 for @xmath143 . the operators @xmath144 and @xmath145 for @xmath42 waves in mode @xmath87 in @xmath1th layer can be expressed in terms of the operators @xmath146 and @xmath147 . these relations can be , e.g. , written in the form : @xmath148 \nonumber \\ & & \hspace{-3cm}\mbox { } \times \pmatrix { 1/({\cal s}_{m,\alpha})_{11}(\omega_m ) & - ( { \cal s}_{m,\alpha})_{12}(\omega_m ) / ( { \cal s}_{m,\alpha})_{11}(\omega_m ) \cr 0 & 1 } \nonumber \\ & & \hspace{-3 cm } \mbox { } \times \pmatrix{\hat{a}_{m_f,\alpha}^{(n+1)}(\omega_m ) \cr \hat{a}_{m_b,\alpha}^{(0)}(\omega_m ) } , \nonumber \\ & & \hspace{-2 cm } m = s , i ; \;\ ; \alpha = te , tm ; \;\ ; l=1,\ldots , n . \label{21}\end{aligned}\ ] ] the matrix @xmath149 used in eq . ( [ 21 ] ) describes the propagation of @xmath42 wave in field @xmath87 through the whole structure : @xmath150 , \nonumber \\ & & m = s , i ; \;\ ; \alpha = te , tm . \label{22}\end{aligned}\ ] ] similarly , the pump - field amplitudes @xmath64 and @xmath151 for @xmath42 waves in @xmath1th layer can be determined from the amplitudes of the incident fields @xmath152 and @xmath153 as follows : @xmath154 \nonumber \\ & & \hspace{-3cm}\mbox { } \times \pmatrix { 1 & 0 \cr - ( { \cal s}_{p,\alpha})_{21}(\omega_p ) /({\cal s}_{p,\alpha})_{22}(\omega_p ) & 1/({\cal s}_{p,\alpha})_{22}(\omega_p ) } \nonumber \\ & & \hspace{-3 cm } \mbox { } \times \pmatrix{a_{p_f,\alpha}^{(0)}(\omega_p ) \cr a_{p_b,\alpha}^{(n+1)}(\omega_p ) } , \nonumber \\ & & \hspace{-.5 cm } \alpha = te , tm ; \;\ ; l=1,\ldots , n . \label{23}\end{aligned}\ ] ] the matrix @xmath155 describes the propagation of a classical pump field with polarization @xmath42 through the whole structure , i.e. @xmath156 . \label{24}\ ] ] we note that the expression in eq . ( [ 20 ] ) for the output state @xmath157 including relations written in eqs . ( [ 21]-[24 ] ) can be formally recast into a compact form using the so - called left - to - right ( @xmath158 ) and right - to - left ( @xmath159 ) modes introduced in the classical electro - magnetic theory of layered structures ( for details , see @xcite ) . fields exiting the structure at @xmath160 are described by @xmath161 functions , whereas @xmath162 functions are appropriate for fields exiting at @xmath163 . in classical theory , this corresponds to the picture in which every nonlinear layer is a source ( emitting dipole ) of photon pairs @xcite and the fields leaving the structure are given as a sum of contributions from all layers . we assume that the outgoing signal ( idler ) field is detected using an analyzer with the polarization that forms an angle @xmath164 ( @xmath165 ) with respect to the te - wave polarization direction ( see fig . [ fig2 ] ) . in order to get the right wave - function , we transform the operators of the outgoing signal ( idler ) fields into the basis with the polarization vectors given by angles @xmath166 ( @xmath165 ) and @xmath167 ( @xmath168 ) using the following formulas : @xmath169 substituting relations in eqs . ( [ 21 ] ) , ( [ 23 ] ) , and ( [ 25 ] ) into the expression @xmath170 determined using eq . ( [ 20 ] ) , we arrive at the output state @xmath171 describing a signal photon polarized along the angle @xmath166 and an idler photon polarized along the angle @xmath165 : @xmath172 . \label{26}\end{aligned}\ ] ] the function @xmath173 introduced in eq . ( [ 26 ] ) has the meaning of probability amplitude of having a signal photon in field @xmath87 polarized along the angle @xmath174 and its entangled idler twin in field @xmath175 polarized along the angle @xmath165 at the output from the photonic - band - gap structure . we are interested only in the part @xmath176 of the output state @xmath177 given in eq . ( [ 26 ] ) that describes the generated photon - pairs . including the time evolution of the free - fields outside the photonic - band - gap structure , we can write : @xmath178 where @xmath179 the creation operators @xmath180 , @xmath181 , introduced in eq . ( [ 28 ] ) describe the linearly polarized photons outside the photonic - band - gap structure with polarization angles @xmath166 and @xmath165 , and are given as : @xmath182 we note that we do not explicitly express the dependence of quantities on the polarization angles @xmath166 and @xmath165 below . the mean number @xmath183 of photon pairs that have a signal photon in the frequency interval @xmath184 around frequency @xmath185 and its twin idler photon in frequency interval @xmath186 around frequency @xmath187 in field @xmath188 is given by @xmath189 where the operator @xmath190 , the density of photons , is defined as @xmath191 using eq . ( [ 28 ] ) , the expression for @xmath183 in eq . ( [ 30 ] ) can be written as : @xmath192 we get the following expressions for , e.g. , the mean number @xmath193 of signal photons in frequency interval @xmath194 around frequency @xmath185 : @xmath195 the overall number @xmath196 of photon pairs emitted into field @xmath197 is determined by the expression @xmath198 the signal - field energy spectrum @xmath199 of field @xmath197 can be easily determined using the expression @xmath200 the energy spectra @xmath201 ( @xmath202 ; @xmath203 ) characterizing the outgoing down - converted fields without an inclusion of pair entanglement can be evaluated using the signal - field spectra @xmath204 determined in eq . ( [ 35 ] ) ; the idler - field spectra @xmath205 are determined analogously ; @xmath206 the properties of the down - converted fields in the time domain can be conveniently described using a two - photon amplitude @xmath207 giving the probability amplitude of detecting a signal photon at time @xmath208 and an idler photon at time @xmath209 : @xmath210 assuming the state @xmath211 given in eq . ( [ 28 ] ) , the expression in eq . ( [ 37 ] ) for the two - photon amplitude @xmath212 can be rearranged into the form : @xmath213 where the ` fourier transform ' @xmath214 of the function @xmath173 has been introduced : @xmath215 the photon flux of , e.g. , the signal photons , @xmath216 at time @xmath208 over the transverse profile of area @xmath84 in field @xmath188 is defined as @xmath217 the photon flux @xmath218 can be determined using the function @xmath173 ; @xmath219 if the spectrum of the idler field is narrow , we may use the alternative expression @xmath220 entanglement of the signal and idler photons in a pair can be detected in a hong - ou - mandel interferometer ( see fig . [ fig3 ] ) . in order to achieve this type of interference between the signal and idler photons , we rotate polarizations of both photons such that they are the same , and subsequently introduce a relative time delay @xmath221 between the two photons . then two photons arrive at a 50/50 % beam - splitter whose output ports are monitored by detectors . the measured coincidence - count rate @xmath222 is given by the number of simultaneously detected photons at detectors @xmath223 and @xmath224 , placed at the output ports of the beam - splitter in a given time interval . there occurs quantum interference between two paths leading to a coincidence count ; either a signal photon is detected by detector @xmath225 and its twin idler photon by detector @xmath226 or vice versa . the normalized coincidence - count rate @xmath227 in this interferometer can be expressed as follows : @xmath228 where @xmath229 \label{44}\end{aligned}\ ] ] and @xmath230 the symbol @xmath231 stands for real part , and the two - photon amplitude @xmath212 has been introduced in eq . ( [ 37 ] ) . using the expressions in eqs . ( [ 44 ] ) and ( [ 45 ] ) , we arrive at : @xmath232 , \nonumber \\ & & \\ r_0^{mn } & = & \frac{\hbar^2}{8\epsilon_0 ^ 2 c^2 { \cal b}^2 } \int_{0}^{\infty } d\omega_s \ , \int_{0}^{\infty } d\omega_i \ , \ ; \omega_s \omega_i |\phi^{mn}(\omega_s,\omega_i)|^2 . \nonumber \\ & & \end{aligned}\ ] ] if cw pumping is considered , the spectrum is described using eq . ( [ 4 ] ) . as a consequence the following expressions are obtained : @xmath233 in this case the second power of the dirac delta function @xmath234 is formal , so that the above expression should be replaced by : @xmath235 where the period of nonlinear interaction goes from @xmath236 to @xmath237 . expressions for the physical quantities determined above have to be normalized by @xmath238 , which indicates that these quantities are related to the period of @xmath239 . as an example , we study the properties of a nonlinear photonic - band - gap structure composed of 25 nonlinear layers of gan of thickness of 117 nm among which there are 24 linear layers of aln with a thickness of 180 nm . material characteristics of gan and aln can be found , e.g. , in @xcite . this structure is resonant for the pump field at wavelength of 664.5 nm , designed to correspond to the first resonance peak near the band edge . this situation favors the efficient generation of photon pairs at roughly double the pump wavelength , at angles that correspond to the first , second , and third resonance peaks , respectively , for the down - converted fields . we assume that the gan crystallographic @xmath125 axis coincides with the @xmath20 axis of propagation ( see fig . [ fig1 ] ) . in this configuration there is strong nonlinear interaction between te components of the pump , signal , and idler beams . the nonlinear coefficient @xmath240 for this geometry is chosen to be 10 pmv@xmath241 . we first consider pumping by a cw laser tuned at first resonance peak , i.e. @xmath242 nm . a photon pair whose signal photon is emitted along the angle @xmath243 may be described by a two - photon amplitude @xmath244 with typical shapes both in time and spectral domains . in the spectral domain , the two - photon amplitude @xmath245 can be written in the form @xmath246 , where the function @xmath247 is linearly proportional to the spectral amplitude of the signal field . this is a consequence of stationarity of the emitted down - converted fields . the probability @xmath248 of detecting a signal photon at time @xmath249 and an idler photon at time @xmath250 is shown in fig . [ fig4 ] for mode @xmath251 ( both photons exit the structure at @xmath160 ) . the shape of the probability @xmath252 can be easily understood when we consider the fact that both photons in a pair are born together at the same instant of time , and then propagate independently inside the structure . this independent zig - zag propagation increases the average time delay between the two photons , which takes into account all possible realizations of the random process of propagation of a photon that undergoes multiple bounces inside the structure . the random zig - zag propagation causes interference between different paths . this leads to a typical structure with local minima and maxima along the direction @xmath253 ; as shown in fig . [ fig4]b . also , the greater the difference @xmath254 the smaller the value of the two - photon amplitude @xmath255 because the propagation with more zig - zags is less probable . we note that typical stripes along the direction @xmath256 reflect stationarity of the process . typical spectral properties of this structure are documented in fig . [ fig5 ] , where the signal - field energy spectrum @xmath257 is depicted as a function of the angle @xmath258 for field @xmath251 . the spectrum @xmath259 giving the probability of emission of a signal photon at frequency @xmath185 along the angle @xmath260 is a complex function in its variables because it is built up by interference of fields coming from the multilayer structure . a maximum photon - pair generation rate is obtained for @xmath261 deg for degenerate frequencies ( @xmath262 ) . photons from a generated photon pair can exit the structure also at @xmath24 ; their energy spectra @xmath263 , @xmath264 and @xmath265 are similar to that shown in fig . [ fig5 ] . this is because efficient generation of photon pairs occurs provided both the signal and idler fields are resonant in the structure so that the generated photons exit the structure at @xmath24 or @xmath160 with comparable probabilities . the signal - field intensity transmission @xmath266 as a function of normalized frequency @xmath267 and angle @xmath243 is shown in fig . [ fig6 ] for comparison . the angle - dependence of the signal - field fwhm ( full - width at half maximum ) @xmath268 of the energy spectrum in wavelengths and value of energy spectrum @xmath269 at the central frequency @xmath270 are shown in fig . high photon - pair generation rates are observed for the frequency - degenerate case ( @xmath271 ) for values of the angle @xmath258 approximately equal to 14 , 35 , and 55 degrees . for these values of the angle @xmath258 the peaks of the signal - field intensity transmission @xmath266 cross the value @xmath185 equal to @xmath272 ( see fig . [ fig6 ] ) . the first and highest peaks in the signal - field energy spectra @xmath273 ( see fig . [ fig7]b ) correspond to the case when both the signal and idler photons are tuned at the first resonance near the band edge . at this frequency , localization of the down - converted fields is maximum , and the nonlinear process is strongest . the other two sets of peaks in the signal - field energy spectra @xmath273 correspond to the second and third peaks in the intensity transmission @xmath266 . at these frequencies , the nonlinear interaction is somewhat weaker due to smaller field intensities , i.e. , slightly worse localization of the down - converted fields . the spectral fwhms for these frequency - degenerate emissions lye in the interval from 10 nm to 15 nm ( see fig . [ fig7]a ) . for these values of the angle @xmath243 the best available constructive interference in the structure occurs . in a relatively narrow frequency interval the signal and idler fields are enhanced by constructive interference . outside this frequency interval , destructive interference occurs . for the other values of the angle @xmath258 constructive interference is less pronounced , the range of frequencies having constructive interference is broader , and spectral shapes with several peaks may occur ( see fig . [ fig5]b ) . we thus obtain down - converted fields with spectral fwhms reaching even 80 nm ( see fig . [ fig7]a ) . in this case the fwhm indicates the entire active width , which might include several peaks . however , photon - pair generation rates are low in this case . quantum correlations ( entanglement ) between the signal and idler photons in a photon pair are visible in the profile of the normalized coincidence - count rate @xmath274 in a hong - ou - mandel interferometer ( see fig . [ fig8 ] for mode @xmath275 ) . for values of the angle @xmath258 with strong constructive interference ( @xmath276 deg ) this profile is given by a relatively broad dip ( see fig . [ fig8]b ) . on the other hand , oscillating curves with several local minima and maxima characterize profiles of the normalized coincidence - count rates @xmath274 for other values of the angle @xmath258 . a typical profile for this case is shown in fig . [ fig8]b . these profiles reflect broader spectra of the down - converted fields along these angles as well as a nonzero difference of the central frequencies @xmath270 and @xmath277 of the signal and idler fields . period of these oscillations is proportional to @xmath278 . position @xmath279 of the central dip , its fwhm @xmath280 , and visibility @xmath281 of the normalized coincidence - count rate @xmath227 as functions of the angle @xmath258 are shown in fig . [ fig9 ] . we can see in fig . [ fig9]a that the maximum overlap of the signal and idler photon fields is reached for a nonzero mutual time delay @xmath221 for nonsymmetric modes @xmath282 and @xmath283 . this means that the signal and idler photons exit the structure with a small mutual time delay . this nonzero mutual time delay means that interference effects inside the structure prefer the generation of photon pairs from nonlinear layers positioned closer to one edge of the structure , or in any case approximately co - located . the visibility @xmath281 of the coincidence - count rate @xmath227 equal to one is naturally observed for symmetric modes @xmath275 and @xmath284 . values of the visibility @xmath281 better than 0.9 can also be reached in nonsymmetric modes @xmath282 and @xmath283 provided that the fields are generated along the angles @xmath258 in the vicinity of those with a strong constructive interference . the structure of quantum correlations between the signal and idler fields in time domain is complex and prevents the visibility @xmath281 from reaching higher values along the angles @xmath258 , where there are broad spectra of the down - converted fields . the width @xmath285 of the dip in the normalized coincidence - count rate @xmath274 determines the entanglement time , i.e. the time interval in which both photons can be detected . the fwhm @xmath285 of the central dip for the structure ranges from 50 fs for values of the angle @xmath258 with broad spectra of the down - converted fields , to 150 - 250 fs for values of the angle @xmath243 at where strong constructive interference occurs ( see fig . [ fig9]c ) . we now assume that the structure is pumped by an ultrashort pulse with an unchirped gaussian profile [ see eq.([2 ] ) ] having pulse duration @xmath286 equal to 200 fs , and central ( carrier ) frequency @xmath35 at the wavelength of 664.5 nm . the squared modulus @xmath287 fits well into the peak of resonance of the pump - field intensity transmission @xmath288 , as depicted in fig . [ fig10 ] . only signal and idler photons with frequencies @xmath185 and @xmath289 for which the sum @xmath290 lyes within the pump - pulse spectrum may be generated . the allowed values of the frequencies @xmath185 and @xmath187 are indicated in the graph of the probability @xmath291 of emitting a signal photon at frequency @xmath185 and its twin at frequency @xmath187 . a typical shape of the probability latexmath:[$ @xmath251 . we note that the shape of the probability @xmath293 in fig . [ fig11 ] may be considered as being composed of many curves defined above the lines @xmath294 , and it may be assumed to be linearly proportional to those characterizing cw pumping at the frequency @xmath295 . the probability @xmath252 of detecting a signal photon at time @xmath208 and its twin at time @xmath250 depicted in fig . [ fig12 ] shows that both signal and idler photons occur in sharp time windows as a consequence of pumping by an ultrashort pulse . the probability @xmath252 has a typical droplet shape that reflects the fact that the longer the photons from a pair propagate inside the structure , the greater the average difference of the occurrence times @xmath208 and @xmath250 of the photons . the structure with local minima and maxima farther from the diagonal in fig . [ fig12 ] reflects interference in the layered structure . because the pump - pulse amplitude spectrum @xmath296 fits well into the peak of resonance of the structure , and the emitted frequencies of the down - converted fields have also to fit into the structure , the energy spectra of the signal and idler fields with pulsed pumping are very similar to those obtained for cw pumping also , the behavior of photon pairs generated by pulsed pumping in the hong - ou - mandel interferometer is similar to that observed with cw pumping ; only the visibilities @xmath281 of the normalized coincidence - count rates @xmath274 are slightly worse . the signal and idler fields are now emitted in the form of ultrashort , pulsed fields in multimode fock states . a typical dependence of the signal - field photon flux @xmath297 at time @xmath208 and angle @xmath258 is shown in fig . [ fig13 ] . [ fig13 ] suggests that the photon flux @xmath298 typically spreads to longer times as a consequence of zig - zag movement , i.e. multiple scattering , of photons inside the structure . time delay @xmath299 and fwhm @xmath300 of the pulsed signal field as well as photon flux @xmath301 at the center of the pulsed field as they depend on the angle @xmath260 are depicted in fig . [ fig14 ] . the time delay @xmath299 of the signal pulse with respect to the pump pulse ranges from 70 fs to 200 fs . it reaches its maximum values for values of the angle @xmath258 with strong constructive interference ( @xmath276 deg ) . fwhm @xmath300 of the signal pulse goes from 270 fs to 320 fs . the narrower the spectrum of the signal field ( given by @xmath302 in wavelengths ) the greater the signal - field pulse duration @xmath300 ( compare figs . [ fig7]a and [ fig14]b ) . the largest value of the photon flux @xmath303 can be reached for the angle @xmath258 equal to 13.8 deg ( see fig . [ fig14]c ) . these results clearly show that the considered photonic - band - gap structure maintains a pulsed character of the nonlinear process and allows generation of pulsed down - converted fields with an ultrashort duration . such pulsed fields are required in many experiments with photon pairs in which time synchronization of several photon pairs is necessary . the model that we have developed is one dimensional with respect to the spatial coordinates , and does not include all aspects ( especially those related to transverse profiles of the interacting optical fields ) of the nonlinear process in a real structure . for this reason we do not determine the absolute values of photon - pair generation rates . instead , we judge the efficiency of the suggested structure ( stemming from constructive interference of the interacting optical fields ) with respect to an ideal reference structure which fully exploits the nonlinearity , but does not rely on interference . this reference structure has formally all linear indices of refraction equal to one , so effects on the boundaries between layers are suppressed . it is also assumed that the nonlinear process is fully phase matched . the orientations of nonlinear layers , as well as polarizations of the interacting optical fields , are such that the greatest nonlinear effect occurs . a photon pair emitted from this structure is described by the following output state @xmath304 [ compare eq . ( [ 20 ] ) ] : @xmath305 where @xmath306 [ @xmath307 denotes a creation operator of a photon in the signal [ idler ] field . the function @xmath308 used in eq . ( [ 50 ] ) gives the maximum value from the elements of a tensor in its argument . the relative photon - pair generation rate @xmath309 of a pair with a signal photon at frequency @xmath185 in mode @xmath197 is then given as : @xmath310 where the signal - field energy spectrum @xmath311 is given in eq . ( [ 35 ] ) , and the signal - field energy spectrum @xmath312 of the reference structure is determined in the same way using the output state @xmath304 written in eq . ( [ 50 ] ) . the relative photon - pair generation rates @xmath313 , @xmath314 , @xmath315 , and @xmath316 for the signal - field frequency @xmath317 as a function of the angle @xmath258 are shown in fig . [ fig15 ] assuming cw pumping . the relative photon - pair generation rates for pulsed pumping are lower in comparison with cw pumping , because frequencies at the edges of the pump - field spectrum have lower efficiencies than those in the middle . despite this , the considered pulsed pumping has the overall relative photon - pair generation rate @xmath318 equal to 9 , whereas @xmath319 for cw pumping for @xmath261 deg . the structure that we consider is approximately 12 times more efficient than the reference structure , provided we consider all the emitted photon pairs ( given by the overall relative photon - pair generation rate @xmath318 in fig . [ fig15 ] ) . even when we restrict ourselves only to the mode @xmath251 , we still have an enhancement of the nonlinear process by a factor of three . for comparison purposes , one layer of gan of thickness @xmath320 nm ( i.e. , containing the same amount of nonlinear material as our structure ) has the overall relative photon - pair generation rate @xmath318 equal to 0.09 for values of the angle @xmath243 less than 20 deg . the inclusion of linear layers of aln inside the structure and the accompanying interference effects thus increase the photon - pair generation rates by two orders of magnitude . we have developed a quantum model of spontaneous parametric down - conversion ( generating entangled photon pairs ) in one - dimensional , nonlinear , layered media . using the model we have determined a two - photon amplitude , and we have provided measurable characteristics of the down - converted fields : marginal signal and idler energy spectra , time - dependent photon fluxes of the signal and idler fields , coincidence - count interference patterns in the hong - ou - mandel interferometer , and photon - pair generation rates . a specific structure was suggested as an efficient source of photon pairs , and interference effects of the interacting fields can enhance the photon - pair generation rates by as much as two hundred times . the widths of the spectra of the down - converted fields and entanglement time of photons in a pair depend strongly on the angle of emission , and vary from 10 nm to 80 nm and from 50 fs to 250 fs , respectively . pumping the structure with an ultrashort pulse of time duration of several hundreds of fs can lead to the generation of pulsed signal and idler fields extending over several hundreds of fs . hong , z.y . ou , and l. mandel , phys . rev . lett . * 59 * , 2044 ( 1987 ) . l. mandel , e. wolf , _ optical coherence and quantum optics _ ( cambridge univ . press . cambridge , 1995 ) . j. peina , z. hradil , and b. juro , _ quantum optics and fundamentals of physics _ ( kluwer , dordrecht , 1994 ) . d. bouwmeester , j .- w . pan , k. mattle , m. eibl , h. weinfurter , and a. zeilinger , nature * 390 * , 575 ( 1997 ) . d. bouwmeester , j .- w . pan , m. daniell , h. weinfurter , and a. zeilinger , phys . . lett . * 82 * , 1345 ( 1999 ) . d. bouwmeester , a. ekert , and a. zeilinger ( eds . _ the physics of quantum information _ ( springer , berlin , 2000 ) . d. bru , n. ltkenhaus , in _ applicable algebra in engineering , communication and computing _ vol . 10 ( springer , berlin , 2000 ) ; p. 383 . a. migdall , physics today * 1 * , 41 ( 1999 ) . kwiat , e. waks , a.g . white , i. appelbaum , and p.h . eberhard , phys . a * 60 * , r773 ( 1999 ) . kwiat , nature * 412 * , 866 ( 2001 ) . y. nambu , k. usami , y. tsuda , k. matsumoto , k. nakamura , phys . rev . a * 66 * , 033816 ( 2002 ) . shapiro and n.c . wong , j. opt . b : quantum semiclass . opt . * 2 * , l1 ( 2000 ) . kuklewicz , m. fiorentino , g. messin , f.n.c . wong , and j. shapiro , opt . express * 13 * , 127 ( 2005 ) . p. trojek , c. schmid , m. bourennane , h. weinfurter , and c. kurtsiefer , opt . express * 12 * , 276 ( 2004 ) . m. bertolotti , c.m . bowden , and c. sibilia , _ nanoscale linear and nonlinear optics _ , aip vol . 560 ( aip , melville , 2001 ) . e. yablonovitch , phys . 58 , 2059 ( 1987 ) . s. john , phys . 58 , 2486 ( 1987 ) . joannopoulos , r.d . meade , and j.n . winn , _ photonic crystals _ ( princeton university press , princeton , 1995 ) . k. sakoda , _ optical properties of photonic crystals _ ( springer , newyork , 2005 ) , 2nd ed . vamivakas , b.e.a . saleh , a.v . sergienko , and m.c . teich , phys . a 70 , 043810 ( 2004 ) . m. centini , j. peina jr . , l. sciscione , c. sibilia , m. scalora , m.j . bloemer , and m. bertolotti , phys . rev . a * 72 * , 033806 ( 2005 ) . x. li , p.l . voss , j.e . sharping , and p. kumar , phys . lett . * 94 * , 053601 ( 2005 ) . j. fulconis , o. alibart , w.j . wadsworth , p. s. russell , and j. g. rarity , opt . express * 13 * , 7572 ( 2005 ) . t.e . keller and m.h . rubin , phys . a 56 , 1534 ( 1997 ) . j. peina jr , a.v . sergienko , b.m . jost , b.e.a . saleh , and m.c . teich , phys . a * 59 * , 2359 ( 1999 ) . g. di giuseppe , l. haiberger , f. de martini and a.v . sergienko , phys . a * 56 * , r21 ( 1997 ) . grice , r. erdmann , i.a . walmsley , and d. branning , phys . a * 57 * , r2289 ( 1998 ) . p. yeh , _ optical waves in layered media _ ( wiley , new york , 1988 ) . w. vogel , d.g.welsch , s. walentowicz , _ quantum optics _ ( wiley - vch , weinheim , 2001 ) . g. daguanno , n. mattiucci , m. scalora , m.j . bloemer , and a.m. zheltikov , phys . e * 70 * , 016612 ( 2004 ) . sanford , a.v . davidov , d.v . tsvetkov , a.v . dmitriev , s. keller , u.k . mishra , s.p . denbaars , s.s . park , j.y . han and r.j . molnar , j. apl . phys . * 97 * , 053512 ( 2005 ) . j. miragliotta , d.k . winckenden , t.j . kistenmacher , w. a. bryden , j. opt . b * 10 * , 1447 ( 1993 ) .

the experimental success in the coherent coupling of quantum dots @xcite has enabled measuring the transport properties of systems with a molecule - like level structure . recently , further progress in this direction has been attained by the reproducible measurement of currents through molecules which are coupled to metallic leads @xcite . together with these experimental achievements , new theoretical interest in the transport properties of such nanoscale systems emerged @xcite . one particular field of interest is the interplay of the electron transport and excitations by an oscillating gate voltage , a microwave field , or an infrared laser , respectively . such excitations bear intriguing phenomena like photon - assisted tunneling @xcite and the adiabatic @xcite and non - adiabatic @xcite pumping of electrons . a prominent example for the control of quantum dynamics is the so - called coherent destruction of tunneling , i.e. , the suppression of the tunneling dynamics of a particle in a double - well potential @xcite , in a two - level system @xcite , or in a superlattice @xcite . recently , coherent destruction of tunneling has also been found for the dynamics of two interacting electrons in a double quantum dot @xcite . moreover , it has been demonstrated that a corresponding transport effect exists : if two leads are attached to the ends of a tunneling system , then a proper driving field can be used to suppress the current even in the presence of a large transport voltage @xcite . moreover , in such a system the corresponding shot noise level _ a priori _ can be controlled by proper ac fields @xcite . within this work , we provide more details on this noise control scheme and also explore its limitations . an intuitive description of the electron transport through time - independent mesoscopic systems is provided by the landauer scattering formula @xcite and its various generalizations . both the average current @xcite and the transport noise characteristics @xcite can be expressed in terms of the quantum transmission coefficients for the respective scattering channels . by contrast , the theory for driven quantum transport is less developed . scattering of a single particle by an arbitrary time - dependent potentials has been considered @xcite without relating the resulting transmissions to a current between electron reservoirs . such a relation is indeed non - trivial since the driving opens inelastic transport channels and , therefore , in contrast to the static case , an _ ad hoc _ inclusion of the pauli principle is no longer unique . this gave rise to a discussion about `` pauli blocking factors '' @xcite . in order to avoid such conflicts , one should start out from a many - particle description . in this spirit , within a green function approach , a _ solution for the current through a time - dependent conductor has been presented , e.g. , in refs . and without taking advantage of the full floquet theory for the wire . nevertheless in some special cases like , e.g. , for conductors consisting of a single level @xcite or for the scattering by a piecewise constant potential @xcite , an explicit solution becomes feasible . moreover , for large driving frequencies , the driving can be treated within a self - consistent perturbation theory @xcite . the spectral density of the current fluctuations has been derived for the low - frequency ac conductance @xcite and the scattering by a slowly time - dependent potential @xcite . for arbitrary driving frequencies , the noise has been characterized by its zero - frequency component @xcite . a remarkable feature of the current noise in the presence of time - dependent fields is its dependence on the phase of the transmission _ amplitudes _ @xcite . by clear contrast , both the noise in the static case @xcite and the current in the driven case @xcite depend solely on transmission _ probabilities_. within this work , we _ derive _ within a floquet approach explicit expressions for both the current and the noise properties of the electron transport through a driven nanoscale conductor under the influence of time - dependent forces . this generalizes recent approaches since the presented floquet formalism is applicable to arbitrary periodically driven tight - binding systems and , in particular , is valid for arbitrary driving strength and , as well , extends beyond the adiabatic regime . the dynamics of the electrons is solved by integrating the heisenberg equations of motion for the electron creation / annihilation operators in terms of the single - particle propagator . for this propagator , in turn , we provide a solution within a generalized floquet approach . such a treatment is valid for effectively non - interacting electrons , i.e. , when no strong correlations occur . disregarding these interactions also implies that the displacement currents are not taken into account entirely . as a consequence , the ac component of the electrical current _ inside _ the nanoconductor may deviate from the particle current @xcite . this paper is organized as follows . after introducing in sec . [ sec : model ] a model for the leads and the conductor under the influence of external fields , we derive in sec . [ sec : scattering ] for a situation with time - periodic but otherwise arbitrary driving general expressions for the current and its noise and establish a connection to a floquet eigenvalue equation . in sec . [ sec : approximations ] , we consider some special cases and approximations . section [ sec : application ] is devoted to the influence of an electromagnetic dipole field on a conductor consisting of a few tight - binding levels . situations with an ac transport voltage are addressed in appendix [ app : ac.voltage ] , while in appendix [ appendix : xi ] , we detail an alternative derivation which has been introduced in ref . . we start out by introducing a model for the central conductor ( `` wire '' ) under the influence of an external driving field like , e.g. , a molecular wire subject to laser radiation or coupled quantum dots @xcite driven by microwaves or an oscillating gate voltage . the conductor is attached by tunneling couplings to external leads . the entire setup of our nanoscale system is described by the time - dependent hamiltonian @xmath0 where the different terms correspond to the wire , the leads , and the wire - lead couplings , respectively . we focus on the regime of coherent quantum transport where the main physics at work occurs on the wire itself . in doing so , we neglect other possible influences originating from driving induced hot electrons in the leads , dissipation on the wire and , as well , electron - electron interaction effects . then , the wire hamiltonian reads in a tight - binding approximation with @xmath1 orbitals @xmath2 @xmath3 for a molecular wire , this constitutes the so - called hckel description where each site corresponds to one atom . the fermion operators @xmath4 , @xmath5 annihilate and create , respectively , an electron in the orbital @xmath2 . the influence of an applied ac field with frequency @xmath6 results in a periodic time - dependence of the wire hamiltonian : @xmath7 . the leads are modeled by ideal electron gases , @xmath8 where @xmath9 ( @xmath10 ) creates an electron in the state @xmath11 ( @xmath12 ) in the left ( right ) lead . the tunneling hamiltonian @xmath13 establishes the contact between the sites @xmath14 , @xmath15 and the respective lead , as sketched in fig . [ fig : levels ] . this tunneling coupling is described by the spectral density @xmath16 of lead @xmath17 , @xmath18 . if the lead modes are dense , @xmath19 becomes a smooth function . level structure of a nano - conductor with @xmath20 orbitals . the end sites are coupled to two leads with chemical potentials @xmath21 and @xmath22.,scaledwidth=45.0% ] to fully specify the dynamics , we choose as an initial condition for the left / right lead a grand - canonical electron ensemble at temperature @xmath23 and electro - chemical potential @xmath24 , respectively . thus , the initial density matrix reads @xmath25 where @xmath26 is the number of electrons in lead @xmath17 and @xmath27 denotes the boltzmann constant times temperature . an applied voltage @xmath28 maps to a chemical potential difference @xmath29 with @xmath30 being the electron charge . then , at initial time @xmath31 , the only nontrivial expectation values of the wire operators read @xmath32 where @xmath33)^{-1}$ ] denotes the fermi function . in our model hamiltonian , the leads are time - independent . thus , it seemingly can not describe ac transport voltages . such a situation , however , can be mapped by a gauge transformation to one with time - independent chemical potentials as demonstrated in appendix [ app : ac.voltage ] . due to their experimental accessibility , the central quantities in a quantum transport problem are the stationary current and the low - frequency part of its noise spectrum . within a scattering picture of _ non - driven _ mesoscopic transport , both quantities can be expressed in terms of a transmission function @xmath34 which reflects the probability that an electron is transmitted from one lead to the other @xcite . due to energy conservation , the reversed process , occurs with equal probability . this is no longer true for driven systems and , consequently , the scattering approach needs to be generalized . thus , in this section , we _ derive _ expressions for the currents and its noise properties for the transport through the time - dependent system modeled above . in the so - called wide - band limit , the more compact derivation presented in ref . becomes possible , cf . appendix [ appendix : xi ] . we will show that the average electrical current contains only transition probabilities and , thus , resembles a scattering formula . in clear contrast to the static two - terminal case , however , we will find that the noise depends in addition also on the phases of the scattering matrix . to avoid the explicit appearance of commutators in the definition of correlation functions , we perform the derivation of the central transport quantities in the heisenberg picture . as a starting point we choose the operator @xmath35 that describes the charge accumulated in lead @xmath17 with respect to the initial state . due to total charge conservation , @xmath36 equals the net charge transmitted across the contact @xmath17 ; its time derivative defines the corresponding current @xmath37 the current noise is described by the symmetrized correlation function @xmath38_+ \big\rangle \label{eq : s}\ ] ] of the current fluctuation operator @xmath39 , where the anticommutator @xmath40_+=ab+ba$ ] ensures hermiticity . it can be shown that at long times , @xmath41 shares the time - periodicity of the driving @xcite . therefore , it is possible to characterize the noise level by the zero - frequency component of @xmath42 averaged over the driving period , @xmath43 we find below that for two - terminal devices @xmath44 is independent of the contact @xmath17 , i.e. , @xmath45 . the evaluation of the zero - frequency noise @xmath46 directly from its definition can be tedious due to the explicit appearance of both times , @xmath47 and @xmath48 . this inconvenience can be circumvented by employing the relation @xmath49 which follows from the integral representation of eqs . and , @xmath50 , in the limit @xmath51 . by averaging eq . over the driving period and using @xmath52 , we obtain @xmath53 where @xmath54 denotes the charge fluctuation operator and @xmath55 the time average . the fact that the time average can be evaluated from the limit @xmath56 allows to interpret the zero - frequency noise as the `` charge diffusion coefficient '' . as a dimensionless measure for the _ relative _ noise strength , we employ the so - called fano factor @xcite @xmath57 where @xmath58 denotes the time - average of the current expectation value @xmath59 . note that in a two - terminal device , the absolute value of the average current is independent of the contact @xmath17 . in order to take the exclusion principle properly into account , we have formulated the transport problem under consideration in terms of second quantization . nevertheless , in the absence of interactions , both the current and its noise can be traced back to the solution of the corresponding single - particle problem . thus , our next step is to relate the expectation value and the variance of the charge operator to the transmission of electrons from one lead to the other . for that purpose , we start from the heisenberg equations of motion @xmath60 for these coupled linear equations , the formal solution @xmath61 involves the propagator @xmath62 of the corresponding single - particle problem . we insert into and use the initial condition to obtain for the transfered charge at long times [ i.e. , in the limit @xmath63 , where all transients die out and , in particular , the second line in eq . becomes irrelevant ] the expectation value @xmath64 to symmetrize this expression , we eliminate the backscattering terms , i.e. , the contributions with @xmath65 , by employing the completeness relation @xmath66 where @xmath67 , @xmath68 , and @xmath69 denote the projectors onto the states of the left lead , the right lead , and the wire , respectively . then , from the time derivative of eq . , we find for the current through the left contact the result @xmath70 and _ mutatis mutandis _ for the current through the right contact . this expression already obeys the `` scattering form '' with the _ time - dependent _ transmission @xmath71 of electrons with energy @xmath72 from lead @xmath17 to lead @xmath73 . at asymptotic times , the transitions from the lead state @xmath74 to the lead state @xmath75 and the wire state @xmath2 happen with the rates @xmath76 the last term in the current describes a periodic charging of the wire stemming from the external driving . with an average over one driving period , this contribution vanishes and , thus , the dc current reads @xmath77 with @xmath78 denoting the time average of the rate . interchanging in eq . @xmath79 and @xmath80 yields the negative current @xmath81 . thus , as expected from total charge conservation , the average current is , besides its sign , independent of the contact at which it is evaluated . we emphasize that obeys the form of the current formula obtained for a _ static _ conductor within a scattering formalism . in particular , consistent with refs . and , no `` pauli blocking factors '' @xmath82 appear in our derivation . in contrast to a static situation , this is in the present context relevant since for a driven system generally @xmath83 , such that a contribution proportional to @xmath84 would not cancel @xcite . the zero - frequency noise @xmath46 is conveniently derived from the charge fluctuation with the help of relation . expressing the charge fluctuation by the heisenberg operators yields for the initial condition after some algebra @xmath85 by using the completeness relation , we have achieved a form which is , besides the appearance of @xmath69 , symmetric under exchanging @xmath86 . here , @xmath87 is a shorthand notation for @xmath62 and @xmath88 . taking the time derivative and averaging over the driving period yields @xmath89 where we have defined @xmath90 the contributions in eq . which contain the projector @xmath69 on the wire states do not contribute to the zero - frequency noise . this can be demonstrated by inserting for the propagator the explicit expressions and which we derive in the next subsection . interestingly enough , the noise @xmath46 depends on both the diagonal and the off - diagonal elements of the projector @xmath91 . by contrast , the current depends only on the diagonal elements of this operator . as a consequence , in the presence of driving it is not possible to express the noise solely by transmission _ probabilities _ ; cf . eq . , below . the evaluation of the rates @xmath92 and @xmath93 involves the matrix elements of the time - evolution operator @xmath62 with the wire and the lead states . in the following , we eliminate the lead states and will find expressions for the rates that depend explicitly only on the propagator for the wire electrons and the spectral density of the couplings to the leads . we start from the schrdinger equation for the propagator , @xmath94 , where @xmath95 is the single - particle hamiltonian underlying . formal integration with the initial condition @xmath96 results in the dyson equation @xmath97 where @xmath98 denotes the propagator in the absence of the wire - lead coupling . we emphasize that due to the explicit time - dependence of the wire hamiltonian , the integral in is not a mere convolution . using @xmath99 $ ] , we find for the transition matrix elements the relations @xmath100 and @xmath101 where @xmath102 denotes the wire site attached to lead @xmath17 , i.e. , @xmath103 and @xmath104 . at this stage , it is convenient to make use of the time - periodicity of the hamiltonian , @xmath105 . this has the consequence @xcite that @xmath106 and , thus , the retarded green function @xmath107 can be decomposed into a fourier series , @xmath108 , with the coefficients @xmath109 physically , @xmath110 describes the propagation of an electron with initial energy @xmath72 under the absorption ( emission ) of @xmath111 photons for @xmath112 ( @xmath113 ) . we emphasize that generally all sidebands @xmath114 contribute to the green function and that , consequently , the @xmath115-summations are unrestricted . after making use of eqs . and , the transition amplitudes and become @xmath116 and @xmath117 respectively . since below we restrict ourselves to asymptotic times , @xmath63 , we have shifted the lower limit of the integrals accordingly . moreover , in order to perform the @xmath118-integration in eq . , we have introduced a converging factor @xmath119 and will finally consider the limit @xmath120 . for the further evaluation of the average current , we insert the transition amplitude into . after taking the time derivative , averaging over time @xmath47 , and considering the limit @xmath120 , we find @xmath121 and the corresponding expression for @xmath122 . we have introduced the notation @xmath123 . by use of the spectral density , we replace the remaining sums over the lead states by energy integrals and obtain as our first main result the dc current @xmath124 where @xmath125 denote the transmission probabilities for electrons from the right lead , respectively from the left lead , with initial energy @xmath72 and final energy @xmath126 , i.e. , the probability for an scattering event under the absorption ( emission ) of @xmath111 photons if @xmath112 ( @xmath113 ) . for a static situation , the transmissions @xmath127 and @xmath128 are identical and contributions with @xmath129 vanish . thus , it is possible to write the current as a product of a _ single _ transmission @xmath130 and the difference of the fermi functions , @xmath131 . we emphasize that in the driven case this is no longer true . although below we focus on the computation of dc currents , we here continue the derivation of the transport quantities by presenting explicit expressions for the ac currents . we restrict ourselves to @xmath132 since @xmath133 simply follows by proper index replacements . evaluating @xmath132 , we consider also the last term in eq . which describes a periodic charging / discharging of the wire . apart from the time average we perform the same steps as in the derivation of the dc current and obtain @xmath134 where @xmath135 denotes the charge oscillating between the left lead and the wire . obviously , since @xmath136 is time - periodic and bounded , its time derivative can not contribute to the average current . the corresponding charge arising from the right lead , @xmath137 , is _ a priori _ unrelated to @xmath136 ; the actual charge on the wire reads @xmath138 . the time - dependent current is determined by the time - dependent transmission @xmath139^ * \big [ \gamma_l ( \epsilon+k'\hbar \omega ) + \frac{i}{\pi } \mathcal{p } \!\!\ ! \int d\epsilon ' \frac{\gamma_l ( \epsilon')}{\epsilon'-\epsilon - k'\hbar\omega } \big ] .\end{aligned}\ ] ] the corresponding expression for @xmath140 follows from the replacement @xmath141 . note that in the wide - band limit @xmath142 , @xmath18 , the contribution from the principal value integral vanishes . in order to obtain the zero - frequency noise @xmath46 , we evaluate the rates @xmath93 . this step is performed along the lines of reasoning for the evaluation of @xmath143 ( although the actual calculation is far more tedious ) : we insert the transition amplitude into , take the derivative with respect to @xmath47 , and average over one driving period . finally , we employ the relation @xmath144^{-1 } = ( 2\pi)^3 \delta(\epsilon - a)\ , \delta(\epsilon'-b)\ , \delta(a - b)$ ] to perform the limit @xmath120 and find @xmath145^ * g^{(k')}_{1n}({\epsilon}_{q } ) \big|^2 \delta(\epsilon_{q ' } - \epsilon_q - k\hbar\omega ) , \\ \label{eq : w : llr } w_{lq',rq}^{l } & = \frac{2 \pi}{\hbar } |v_{lq'}|^2 |v_{rq}|^2 \sum_k \big| \sum_{k ' } \gamma_l \big ( \epsilon_{q}+k'\hbar\omega \big ) \big [ g^{(k'-k)}_{11}({\epsilon}_{q ' } ) \big]^ * g^{(k')}_{1n } ( { \epsilon}_{q } ) -i g^{(k)}_{1n } ( { \epsilon}_{q } ) \big|^2 \delta(\epsilon_{q'}-{\epsilon}_{q}-k\hbar\omega ) .\end{aligned}\ ] ] the corresponding expressions for @xmath146 and @xmath147 follow from the replacement @xmath141 . inserting these into the noise expression we arrive at our central result @xmath148^ * \big|^2 f_r ( \epsilon ) \bar f_r ( \epsilon_k ) \\ + & \gamma_r ( \epsilon_k ) \gamma_l ( \epsilon ) \big| \sum_{k ' } \gamma_l(\epsilon_{k ' } ) g^{(k'-k)}_{1n}(\epsilon_k ) \big [ g_{11}^{(k')}(\epsilon ) \big]^ * -i g^{(-k)}_{1n}(\epsilon_k ) \big|^2 f_l(\epsilon ) \bar f_r(\epsilon_k ) \big\ } \\ + & \text{same terms with the replacement $ ( l,1 ) \leftrightarrow ( r , n)$ } . \end{split}\ ] ] we have defined @xmath149 and replaced the sums over the lead states by energy integrations using the spectral density . in order to evaluate the expressions for @xmath58 and @xmath46 further , we derive an eigenfunction representation for the green function . it is well - known that beyond the adiabatic limit , the eigenfunctions of the hamiltonian are not of particular use rather a proper basis is provided by a floquet ansatz @xcite . let us start from the schrdinger equation for the propagator , @xmath150 for @xmath151 , and @xmath152 where @xmath102 is defined by @xmath103 and @xmath104 . to eliminate the lead states in the second line of eq . , we insert and replace by use of the spectral density the sum over the lead states by an energy integral . then the last term in eq . becomes @xmath153 within the present context , we are mainly interested in the influence of the driving field on the conductor and not in the details of the coupling to the leads . therefore , we choose for @xmath19 a rather generic form by assuming that in the relevant regime , it is practically energy - independent , @xmath154 this so - called wide - band limit allows further progress since we now can perform in eq . the remaining energy integration to obtain @xmath155 and , consequently , eq . becomes @xmath156 equations and , together with the initial conditions @xmath157 fully determine the propagator . solving this linear set of equations is equivalent to computing a complete set of solutions for the equation @xmath158 where the self - energy @xmath159 results from the coupling to the leads . equation is linear and possesses time - dependent , @xmath160-periodic coefficients . thus , it is possible to construct a complete solution with the floquet ansatz @xmath161 |u_\alpha(t)\rangle , \\ = & \sum_{k=-\infty}^\infty |u_{\alpha , k}\rangle\exp(-ik\omega t ) .\end{aligned}\ ] ] the so - called floquet states @xmath162 obey the time - periodicity of @xmath163 and have been decomposed into a fourier series . in a hilbert space that is extended by a periodic time coordinate , the so - called sambe space @xcite , they obey the floquet eigenvalue equation @xcite @xmath164 due to the brillouin zone structure of the floquet spectrum @xcite , it is sufficient to compute all eigenvalues of the first brillouin zone , @xmath165 . since the operator on the l.h.s . of eq . is non - hermitian , the eigenvalues @xmath166 are generally complex valued and the ( right ) eigenvectors are not mutually orthogonal . thus , to determine the propagator , we need to solve also the adjoint floquet equation yielding again the same eigenvalues but providing the adjoint eigenvectors @xmath167 . it can be shown that the floquet states @xmath168 together with the adjoint states @xmath167 form at equal times a complete bi - orthogonal basis : @xmath169 and @xmath170 . a proof requires to account for the time - periodicity of the floquet states since the eigenvalue equation holds in a hilbert space extended by a periodic time coordinate @xcite . using the floquet equation , it is straightforward to show that with the help of the floquet states @xmath168 the propagator can be written as @xmath171 where the sum runs over all floquet states within one brillouin zone . consequently , the fourier coefficients of the green function [ cf . . ] read @xmath172 inserting them into eqs . and yields explicit expressions for the current and the noise , respectively . in the previous section , the dc current and the zero - frequency noise have been derived for a periodic but otherwise arbitrary driving . within the wide - band limit , both quantities can be expressed in terms of the solutions of the floquet equation , i.e. , the solution of a non - hermitian eigenvalue problem in an extended hilbert space . thus , for large systems , the numerical computation of the floquet states can be rather costly . moreover , for finite temperatures , the energy integration in the expressions and has to be performed numerically . therefore , approximation schemes which allow a more efficient computation are of much practical use . before introducing various approximation schemes for the wire propagator , we discuss two particular cases for which current and noise assume more intuitive expressions . in doing so , we define quantities to which we will refer later in this section . for consistency , the expressions and for the dc current and the zero - frequency noise , respectively , must coincide in the undriven limit with the corresponding expressions of the time - independent scattering theory @xcite . this is indeed the case since the static situation is characterized by two relations : first , in the absence of spin - dependent interactions , we have time - reversal symmetry , @xmath173 and , second , all sidebands with @xmath129 vanish , i.e. , @xmath174 , where @xmath175 and @xmath176 is the green function in the undriven limit . then the current assumes the known form @xmath177 .\ ] ] moreover in a static situation , the relation @xcite @xmath178 \\ + & t(\epsilon ) \left [ 1-t(\epsilon ) \right ] \left [ f_r ( \epsilon ) - f_l ( \epsilon ) \right]^2 \big\ } . \end{split}\ ] ] for zero temperature , the terms in the first line vanish and pure shot noise remains . in contrast , for zero voltage , @xmath179 and the terms in the first line constitute equilibrium quantum noise . obviously if both voltage and temperature are zero , not only the current but also the noise vanishes . in the presence of driving , this is no longer the case . this becomes particularly evident in the high - frequency limit studied in sec . [ sec : approx : hf ] . it is known that in the adiabatic limit , i.e. , for small driving frequencies , the numerical solution of the floquet equation becomes infeasible because a diverging number of sidebands has to be taken into account . in more mathematical terms , floquet theory has no proper limit as @xmath180 @xcite . the practical consequence of this is that for low driving frequencies , it is favorable to tackle the transport problem with a different strategy : if @xmath181 is the smallest energy - scale of the hamiltonian , one computes for the `` frozen '' hamiltonian at each instance of time the current and the noise from the static expressions and being followed up by time - averaging . many phenomena can be discussed in the limit of very large ( practically infinite ; subscript @xmath182 ) voltages such that @xmath183 and @xmath184 in the relevant energy range . then , the dc current becomes @xmath185 remains , thus , @xmath186 . to derive this expression , we again have used the completeness relation and the fact that terms containing the projector on the wire states do not contribute to time averages . expressing @xmath187 and @xmath188 by the green functions yields @xmath189^ * \big|^2 , \end{split}\ ] ] where @xmath190 . these expressions make explicit that @xmath191 and @xmath192 . consequently , for infinite voltage the fano factor can not exceed unity . in the limit of a weak wire - lead coupling , i.e. , for coupling constants @xmath193 which are far lower than all other energy scales of the wire hamiltonian , it is possible to derive within a master equation approach a closed expression for the dc current @xcite . the corresponding approximation within the present floquet approach is based on treating the self - energy contribution @xmath194 in the non - hermitian floquet equation as a perturbation . then , the zeroth order of the floquet equation @xmath195 describes the driven wire in the absence of the leads , where @xmath196 are the `` usual '' floquet states with quasienergies @xmath197 . in the absence of degeneracies the first order correction to the quasienergies is @xmath198 where @xmath199 since the first - order correction to the floquet states will contribute to neither the current nor the noise , the zeroth - order contribution @xmath200 and @xmath128 accordingly . the transmission exhibits for small values of @xmath193 sharp peaks at energies @xmath201 and @xmath202 with widths @xmath203 and @xmath204 . therefore , the relevant contributions to the sum come from terms for which the peaks of both factors coincide and , in the absence of degeneracies in the quasienergy spectrum , we keep only terms with @xmath205 then , the fraction in is a lorentzian and can be approximated by @xmath206 provided that @xmath207 is small . consequently , the energy integration in can be performed even for finite temperature and we obtain for the dc current the expression @xmath208 .\ ] ] the coefficients @xmath209 denote the overlap of the @xmath115th sideband @xmath210 of the floquet state @xmath211 with the first site and the last site of the wire , respectively . we have used @xmath212 which follows from . expression has been derived in a prior work @xcite within a rotating - wave approximation of a floquet master equation approach . within the same approximation , we expand the zero - frequency noise to lowest order in @xmath213 : after inserting the spectral representation of the green function , we again keep only terms with identical floquet index @xmath214 and identical sideband index @xmath115 to obtain @xmath215 of particular interest for the comparison to the static situation is the limit of a large applied voltage such that practically @xmath216 and @xmath217 . then , in eqs . and , the sums over the sideband indices @xmath115 can be carried out such that @xmath218 these expressions resemble the corresponding expressions for the transport across a _ static _ double barrier @xcite . if now @xmath219 for all floquet states @xmath211 , we find @xmath220 . this is in particular the case for systems obeying reflection symmetry @xcite . in the presence of such symmetries , however , the existence of exact crossings , i.e. degeneracies , limits the applicability of the weak - coupling approximation . in many experimental situations , the driving field acts as a time - dependent gate voltage , i.e. , it merely shifts all on - site energies of the wire uniformly . thus , the wire hamiltonian is of the form @xmath221 where , without loss of generality , we restrict @xmath222 to possess zero time - average . a particular case of such a homogeneous driving is realized with a system that consists of only one level @xcite . then trivially , the time and the position dependence of the floquet states factorize and , therefore , the dc current can be obtained within the formalism introduced by tien and gordon @xcite . here , we establish the relation between such a treatment and the present floquet approach . since the time - dependent part of the hamiltonian is proportional to the unity operator , the solution of the floquet equation is , besides a phase factor , given by the eigenfunctions @xmath223 of the static operator @xmath224 , @xmath225 and @xmath226 . the quasienergies @xmath227 coincide with the eigenvalues of the static eigenvalue problem . note that @xmath228 obeys the @xmath229-periodicity of the driving field since the time - average of @xmath222 vanishes by definition . thus , the phase factor in the floquet states can be written as a fourier series , @xmath230 and , consequently we find @xmath231 and the adjoint states accordingly . then , the green function becomes @xmath232 where @xmath176 denotes the green function in the absence of the driving field . inserting into and employing the sum rule @xmath233 , yields @xmath234 , \label{eq : itucker}\ ] ] where @xmath130 is the transmission in the absence of the driving . this expression allows the interpretation , that for homogeneous driving , the floquet channels contribute _ independently _ to the current @xmath235 . for the special case of a one - site conductor and a sinusoidal driving , this relation to the static situation has been discussed in refs . and . addressing the noise properties , we obtain by inserting the green function into the expression @xmath236 \big|^2 f_l(\epsilon ) \bar f_r(\epsilon+k\hbar\omega ) \\ + & \,\text{same terms with the replacement $ ( l,1 ) \leftrightarrow ( r , n)$ } \big\ } . \end{split}\ ] ] while the term in the first line contains only the static transmission at energies shifted by multiples of the photon energies , the contribution in the second line can not be brought into such a convenient form . the reason for this is that the sum over @xmath237 inhibits the application of the relation . as a consequence , in clear contrast to the dc current , the zero - frequency noise can not be interpreted in terms of independent floquet channels . only in the limit of large driving frequencies , we find below that the channels become effectively independent and reduces to an expression that depends only on the transmission in the absence of the driving and the fourier coefficients @xmath238 , cf . next subsection . expressions for the dc current and the noise that depend only on the static transmission have been derived by tucker and feldman @xcite within a tien - gordon approach @xcite . the central approximation of this approach is the description of a time - dependent chemical potential by an effective electron distribution . while this yields the correct expression for the dc current , it does not capture the interference terms in the noise formula . this reveals that a tien - gordon - like approach yields the correct dc current while for the noise ( and other higher - order correlation functions ) it is only valid in a high - frequency limit . for large voltages where @xmath217 and @xmath216 , the sums over the fourier coefficients in eqs . and can be evaluated with the help of the sum rule @xmath239 . then both the dc current and the zero - frequency noise become identical to their value in the absence of the driving . this means that for a sufficiently large transport voltage , a time - dependent gate voltage has no influence on the average current and the zero - frequency noise . many effects occurring in driven quantum systems , such as coherent destruction of tunneling @xcite or current and noise control @xcite , are most pronounced for large excitation frequencies . thus , it is particularly interesting to derive for the present floquet approach an expansion in terms of @xmath240 . thereby , the driven system will be approximated by a static system with renormalized parameters . such a perturbation scheme has been developed for two - level systems in ref . and applied to driven tunneling in bistable systems @xcite and superlattices @xcite . for open quantum system , the coupling to the external degrees of freedom ( e.g. , the leads or a heat bath ) bears additional complications that have been solved heuristically in ref . by replacing the fermi functions by effective electron distributions . in the following , we present a rigorous derivation of this approach based on a perturbation theory for the floquet equation . we assume a driving that leaves all off - diagonal matrix elements of the wire hamiltonian time - independent while the tight - binding levels undergo a position - dependent , time - periodic driving @xmath241 with zero time - average . then , the wire hamiltonian is of the form @xmath242 if @xmath181 represents the largest energy scale of the problem , we can in the floquet equation treat the _ static _ part of the hamiltonian as a perturbation . correspondingly , the eigenfunctions of the operator @xmath243 determine the zeroth order floquet states @xmath244 we have defined the function @xmath245 which is @xmath229-periodic due to the zero time - average of @xmath246 . as a consequence of this periodicity , to zeroth order the quasienergies are zero ( @xmath247 ) and the floquet spectrum is given by multiples of the photon energy , @xmath248 . each @xmath249 defines a degenerate subspace of the extended hilbert space . if now @xmath181 is larger than all other energy scales , the first - order correction to the floquet states and the quasienergies can be calculated by diagonalizing the perturbation in the subspace defined by @xmath250 . thus , we have to solve the time - independent eigenvalue equation @xmath251 the time - independent effective hamiltonian @xmath252 is defined by the matrix elements of the original static hamiltonian @xmath253 with the zeroth order floquet states , @xmath254 the @xmath47-integration constitutes the inner product in the hilbert space extended by a periodic time coordinate @xcite . to first order in @xmath240 , the quasienergies @xmath255 are given by the eigenvalues of the static equation and , consequently , the corresponding floquet states read @xmath256 are @xmath229-periodic , allows to write in the time - dependent phase factor as a fourier series , @xmath257 thus , @xmath258 and the green function for the high - frequency driving reads @xmath259 where @xmath260 denotes the green function corresponding to the static hamiltonian @xmath252 with the self - energy @xmath261 . finally , substituting @xmath262 and using the sum rule @xmath263 , we obtain @xmath264 the effective transmission @xmath265 and @xmath266 follows from the replacement @xmath267 . in order to derive a high - frequency approximation for the zero - frequency noise @xmath46 , we insert the green function into and neglect products of the type @xmath268 for @xmath269 . employing the above sum rule for the fourier coefficients @xmath270 , we obtain for the noise the static expression but with the transmission @xmath130 and the fermi functions @xmath271 replaced by the effective transmission @xmath272 and the effective distribution function , respectively . the fact that @xmath273 is generally not a mere heaviside step function has an intriguing consequence : in the presence of driving , the noise remains finite even if both voltage and temperature are zero . two differences between the high - frequency approximation and the homogeneous driving , cf . sec . [ sec : approx : homo ] , are worth mentioning : first , the static transmission is now replaced by an effective transmission which can be considerably influenced by the driving , see below . second , in general @xmath274 such that @xmath275 . this means that the driving can create an effective bias and thereby create a non - adiabatic pump current . by contrast eq . reveals that a homogeneous driving can not create such a pump current . moreover , if all @xmath276 are identical as in the case of a homogeneous driving , the effective hamiltonian @xmath252 equals the original static hamiltonian . then , also the second line of eq . can be written in terms of the static transmission @xmath130 . in this section , we apply the formalism derived in secs . [ sec : scattering ] and [ sec : approximations ] to study the conduction and noise properties of a nanoscale conductor under the influence of an electromagnetic field . as an elementary model that captures the essential features of a molecular wire @xcite , we employ a tight - binding model composed of @xmath1 sites as sketched in fig . [ fig : levels ] . each orbital is coupled to its nearest neighbor by a hopping matrix element @xmath277 , thus , the single - particle wire hamiltonian reads @xmath278\,|n\rangle\langle n| , \end{split}\ ] ] where @xmath279 denote the on - site energies of the tight - binding levels . within a dipole approximation , the oscillating electromagnetic field causes the time - dependent level shifts @xmath280 with @xmath281 the scaled position of site @xmath2 . since typical laser frequencies are below the work function of a usual metal , we assume that the radiation does not penetrate the leads and that , consequently , the leads stay in thermal equilibrium . the energy @xmath282 denotes the electrical field amplitude multiplied by the electron charge and the distance between two neighboring sites and , thus , depends implicitly on the length of the sample . this model describes , as well , an array of coherently coupled quantum dots @xcite under the influence of microwave radiation . the dipole approximation inherent to the driving neglects the propagation of the electromagnetic field and , thus , is valid only for wavelengths that are much larger than the size of the sample @xcite . this condition is indeed fulfilled for both applications we have in mind : for molecular wires , we consider frequencies up the optical spectral range , i.e. , wavelengths of the order @xmath283 and samples that extend over a few nanometers . coupled quantum dots typically @xcite have a distance of less than @xmath283 while the coupling matrix element @xmath277 is of the order of @xmath284 which corresponds to a wavelength of roughly @xmath285 . we assume that the wire couples equally strong to both leads , thus , @xmath286 . an applied transport voltage @xmath28 is mapped to a symmetric shift of the leads chemical potentials , @xmath287 . moreover , for the evaluation of the dc current and the zero - frequency noise , we restrict ourselves to zero temperature . the zero - temperature limit is physically well justified for molecular wires at room temperature and for quantum dots at helium temperature since in both cases thermal electron excitations do not play a significant role . for a wire described by the hamiltonian , it has been found @xcite that a dipole force of the form suppresses the transport if the ratio @xmath288 is close to a zero of the bessel function @xmath289 ( i.e. , values 2.405 .. , 5.520 .. , 8.654 .. , ) . moreover , in the vicinity of such suppressions , the shot noise characterized by the fano factor assumes two characteristic minima . these suppression effects are most pronounced in the high - frequency regime , i.e. , if the energy quanta @xmath181 of the driving exceed the energy scales of the wire . thus , before going into a detailed discussion , we start with a qualitative description of the effect based on the static approximation for a high - frequency driving that has been derived in sec . [ sec : approx : hf ] . let us consider first the limit of a voltage which is so large that in eq . , @xmath290 can be replaced by unity . then , the average current is determined by the effective hamiltonian @xmath291 which has been derived by inserting the driving into eqs . and . then , obviously @xmath252 is identical to the hamiltonian in the absence of the driving field but with the tunnel matrix element renormalized according to @xmath292 since the bessel function @xmath289 assumes values between zero and one , the amplitude of the driving field allows to switch the absolute value of the effective hopping on the wire , @xmath293 , between 0 and @xmath277 . since the transmission of an undriven wire is proportional to @xmath294 , the effective transmission @xmath272 acquires a factor @xmath295 . this renormalization of the hopping results finally in a current suppression@xcite . for the discussion of the shot noise , we employ the fano factor as a measure . in the limit of large applied voltages , we have to distinguish two limits : ( i ) weak wire - lead coupling @xmath296 ( i.e. , weak with respect to the effective hopping ) and ( ii ) strong wire - lead coupling @xmath297 . in the first case , the tunnel contacts between the lead and the wire act as `` bottlenecks '' for the transport . in that sense they form barriers . thus qualitatively , we face a double barrier situation and , consequently , expect the shot noise to exhibit a fano factor @xmath298 @xcite . in the second case , the links between the wire sites act as @xmath299 barriers . correspondingly , the fano factor assumes values @xmath300 for @xmath301 ( single barrier ) and @xmath302 for @xmath20 ( double barrier ) @xcite . at the crossover between the two limits , the conductor is optimally `` barrier free '' such that the fano factor assumes its minimum . in order to be more quantitative , we evaluate the current and the zero - frequency noise in more detail thereby considering a finite voltage . this requires a closer look at the effective electron distribution ; in particular , we have to quantify the concept of a `` practically infinite '' voltage . in a static situation , the voltage can be replaced by infinity , @xmath303 , if all eigenenergies of the wire lie well inside the range @xmath304 $ ] . in contrast to the fermi functions , the effective electron distribution which is decisive here , decays over a broad range in multiple steps of size @xmath181 . since for our model , @xmath272 is peaked around @xmath305 , we replace here the effective electron distributions by their values for @xmath305 , @xmath306 for zero temperature . we have inserted the coefficients @xmath307 and @xmath308 which have been computed directly from their definition ; @xmath309 denotes the @xmath115th order bessel functions of the first kind . the current , the noise , and the fano factor are given by the static expressions and with the transmission and the electron distribution replaced by the corresponding effective quantities , @xmath310 and @xmath311 , respectively . thus , we obtain @xmath312 respectively , where the subscript @xmath182 denotes the corresponding quantities in the infinite voltage limit , @xmath313 , \end{aligned}\ ] ] and @xmath314 . the factor @xmath315 reflects the influence of a finite voltage ; @xmath316 denotes the largest integer not exceeding @xmath317 . since @xmath318 for @xmath319 and @xmath320 , we find @xmath321 if @xmath322 . this means that for small driving amplitudes @xmath323 , we can consider the voltage as practically infinite . with an increasing driving strength , @xmath324 decreases and , thus , the current becomes smaller by a factor @xmath324 but still exhibits suppressions . by contrast , since @xmath325 for all situations considered here [ cf . the remark after eq . ] , we find from eq . that the fano factor will increase with smaller @xmath324 . the qualitative discussion of the current and noise suppressions can be corroborated by exact numerical results . for this purpose , we have solved numerically the floquet equation . with the resulting floquet states and quasienergies , we obtained the green function . in the zero temperature limit considered here , the fermi functions in the expressions for the average current and the zero - frequency noise become step functions . the remaining energy integrals can be performed analytically since the integrands are rational functions . figure [ fig : hfr ] depicts the average current , the zero - frequency noise , and the corresponding fano factor for a wire that consists of @xmath20 sites with on - site energies @xmath326 as sketched in fig . [ fig : levels ] . the driving frequency @xmath327 lies above all transition energies of the wire states and the applied voltage @xmath328 is relatively large . this particular value of the voltage has been selected to avoid chemical potentials to lie close to multiples of @xmath181 , i.e. , close to the steps of the effective electron distribution . the wire - lead coupling @xmath329 is sufficiently weak , such that in the absence of the driving , the transport is dominated by resonant tunneling . correspondingly , the current is essentially determined by the hopping rate @xmath330 of the electrons from the lead to the wire . the noise exhibits a fano factor @xmath302 which is the characteristic value for the transport across a double barrier @xcite . with an increasing driving amplitude , the current becomes smaller until it reaches its minimum when the ratio @xmath288 assumes a zero of the bessel function @xmath289 . note that while the analytical treatment within a high - frequency approximation predicts a vanishing current , the exact result is still roughly 1% of the value in the absence of the driving . close to the current suppression , the effective tunnel matrix element is much smaller than the wire - lead coupling @xmath331 and the connections to the central site of the wire form a double barrier . consequently , we again find a fano factor @xmath302 . at the crossover @xmath332 , the effective barriers vanish and , therefore , the fano factor assumes its minimum . these exact numerical results are well reproduced by the expressions obtained within a high - frequency approximations for finite voltage . figure [ fig : hfr ] also reveals that for small driving amplitudes , @xmath323 , the assumption of a practically infinite voltage yields the correct results . by contrast , for larger driving amplitudes , @xmath333 , the fano factor can assume values @xmath334 , i.e. , the shot noise becomes even larger than in the static situation . time - averaged current ( a ) , zero - frequency noise ( b ) , and fano factor ( c ) for a conductor consisting of @xmath20 sites with equal on - site energies , @xmath326 , as a functions of the driving amplitude @xmath282 . the driving frequency is @xmath327 , the wire - lead coupling is @xmath329 , and the chemical potentials are @xmath335 . the exact numerical results ( solid lines ) are compared to the high - frequency approximation for finite ( dashed ) and infinite voltage ( dash - dotted ) . ] as the coupling strength @xmath331 is lowered , the distance between a pair of minima of the fano factor becomes smaller until the minima finally vanish @xcite ( not shown ) . in this limit , the current and the noise are given by the weak coupling results and , respectively . the corresponding fano factor @xmath220 [ cf . the discussion after eq . ] is independent of the driving amplitude . figure [ fig : lf]a depicts the behavior for a driving frequency which is of the order of the wire excitations , @xmath336 . then , the high - frequency approximation is no longer applicable . nevertheless , the average current exhibits clear minima with a reduction of the order 50% . compared to the high - frequency case , these minima are shifted towards smaller driving amplitudes , i.e. , they occur for ratios @xmath288 slightly below the zeros of the bessel function @xmath289 . at the minima of the current , the fano factor ( solid line in fig . [ fig : lf]b ) still assumes a maximum with a value close to @xmath302 . although , the sharp minima close to the current suppressions have vanished , in - between the maxima the fano factor assumes remarkably low values ( @xmath337 ) . figure [ fig : lf]b also reveals that already for @xmath338 , the high - frequency regime is reached . ( a ) time - averaged current ( solid line ) and zero - frequency noise ( dashed line ) as a function of the driving amplitude for the driving frequency @xmath336 . ( b ) corresponding fano factor for the same data ( solid line ) and for the driving frequencies @xmath339 ( broken ) and @xmath340 ( dash - dotted ) . all other parameters are as in fig . [ fig : hfr ] . ] for strong wire - lead coupling , it is possible to choose a driving frequency that is large with respect to the wire excitations , but small as compared to the coupling @xmath331 , thus @xmath341 . figure [ fig : lg ] depicts the current and the fano factor in this limit for wires with a different number of sites . the qualitative difference between these cases can be explained by the fact that due to the strong coupling , the first and the last wire site hybridize with the leads . then the setup behaves similar to a wire with @xmath342 sites and a weak wire - lead coupling @xmath343 . this means that for @xmath301 the wire acts as point contact while for @xmath20 , we qualitatively have resonant transport through a single level . in both cases remains no tunneling matrix element of the wire that could be renormalized and , consequently , for @xmath344 the current suppressions vanish in the strong - coupling limit ( cf . [ fig : lg]a ) . this scenario is also reflected in the behavior of the fano factor ( fig . [ fig : lg]b ) which exhibits the characteristic values @xmath300 ( point contact ) for @xmath301 and @xmath302 ( single resonant level ) for @xmath20 . finally , for @xmath345 we observe the behavior of a driven wire with two sites and weak coupling @xcite . then , a vanishing effective hopping @xmath346 corresponds to a point contact , thus , @xmath300 . although the behavior of the fano factor can be explained by drawing analogies to a weakly coupled wire with @xmath342 sites , the global decay of the current with the driving amplitude , cf . [ fig : lg]a , is not within the scope of this intuitive picture . time - averaged current ( a ) and fano factor ( b ) as a function of the driving amplitude @xmath282 for a wire with @xmath347 sites and the wire - lead coupling strength @xmath348 . the other parameters are @xmath326 , @xmath327 and @xmath349 . ] so far , we have assumed that all on - site energies of the wire are identical . in an experimental setup , however , the applied transport voltage acts also a static dipole force which rearranges the charge distribution in the conductor and thereby causes an internal potential profile @xcite . the self - consistent treatment of such effects is , in particular in the time - dependent case , rather ambitious and beyond the scope of this work . thus , here we only derive the consequences of a static bias without determining its shape from microscopic considerations . we assume a position - dependent static shift of the on - site energies by an energy @xmath350 , i.e. , for a wire with @xmath20 sites , @xmath351 figure [ fig : bias]a demonstrates that the behavior of the average current is fairly stable against the bias . in particular , we still find pronounced current suppressions . note that since @xmath352 a high - frequency approximation is still possible . as a main effect of the bias , we find reduced current maxima while the minima remain . by contrast , the minima of the fano factor ( fig . [ fig : bias]b ) become washed out : once the bias becomes of the order of the wire - lead coupling , @xmath353 , the structure in the fano factor vanishes and we find @xmath302 for all driving amplitudes @xmath323 [ cf . the discussion discussion after eq . ] . interestingly , the value of the fano factor at current suppressions is bias independent . time - averaged current ( a ) and fano factor ( b ) as a function of the driving amplitude @xmath282 for a wire with @xmath20 sites in the presence of an internal bias . the on - site energies are @xmath354 , @xmath355 , @xmath356 . all other parameters are as in fig . [ fig : hfr ] . ] we have derived with eqs . , , and expressions for the dc current , the zero - frequency noise , and the time - dependent current for the electron transport through ac - driven nanoscale systems . a cornerstone of our approach is the relation of the propagator to a non - hermitian floquet equation . this yields explicit formulae for the current and the noise . moreover , the connection to floquet theory allows to elucidate various approximation schemes that enable an efficient computation and , in addition , provide physical insight . above all , a high - frequency approximation has emerged to be very useful : within an expansion in @xmath240 , the driven transport problem can be approximated by a time - independent transport problem with a renormalized tunneling and effective distribution functions for the lead electrons . the conductance properties of the latter can be derived with standard methods . moreover , for the case of a time - dependent gating voltage , we have revealed the limitations of the tien - gordon approach : while such a treatment provides the correct expression for the current , it neglects interferences of different floquet channels . a detailed investigation of the recently found shot noise suppression provided a deeper understanding of this effect . in particular , the analytical treatment within a high - frequency approximation can explain the characteristic emergence of the current suppressions which are accompanied by a noise maximum and two remarkably low minima . a numerical study fully confirmed the analytical results . for lower driving frequencies , i.e. , beyond the high - frequency limit , the current suppressions become considerably less pronounced . by contrast , the shot noise suppression turned out to be more stable . thus , since the current stays remarkably large while the noise is controllable , this regime is particularly promising for applications . at first sight , in the limit of strong wire - lead coupling these phenomena appear quite different . a closer look , however , revealed that the strong coupling entails a hybridization of the first and the last site with the respective lead . therefore , the wire behaves qualitatively like a weakly coupled wire with two sites less . moreover , we have found that the noise suppressions are quite sensitive to an internal bias . once the on - site energies of neighboring sites have differences of the order of the wire - lead coupling energy , the minima of the fano factor vanish . a most interesting application of these results is the development of current sources with a controllable noise level . we thank gert - ludwig ingold , jrg lehmann , and michael strass for helpful discussions . this work has been supported by the volkswagen - stiftung under grant no . i/77 217 , a marie curie fellowship of the european community program ihp under contract no . hpmf - ct-2001 - 01416 ( s.c . ) , and the dfg sonderforschungsbereich 486 . within this work , we focus on models where the driving enters solely by means of time - dependent matrix elements of the wire hamiltonian while the leads and the wire - lead couplings remain time - independent . an _ a priori _ different type of driving is the application of a time - dependent transport voltage . in this appendix , we demonstrate that a setup with an oscillating transport voltage can be mapped by a gauge transformation to a hamiltonian of the form . consequently , it is possible to apply the formalism derived derived in sec . [ sec : scattering ] also to situations with an oscillating transport voltage . we restrict the discussion to the case where the electron energies of only the left lead are modified by an external @xmath160-periodic voltage @xmath357 with zero time - average , thus in the left lead @xmath358 the generalization to a situation where also the levels in the right lead are @xmath160-periodically time - dependent , is straightforward . since an externally applied voltage causes a potential drop along the wire @xcite , we have to assume for consistency that for an ac voltage , the wire hamiltonian also obeys a time - dependence . ignoring such a time - dependent potential profile enables a treatment of the transport problem within the approach of refs . and . in the general case , however , we have to resort to the approach put forward with this work . we start out by a gauge transformation of the hamiltonian with the unitary operator @xmath359 where @xmath360 describes the phase accumulated from the oscillating voltage . the transformation has been constructed such that the new hamiltonian @xmath361 possesses a time - independent tunnel coupling . since , the operator @xmath362 transforms as @xmath363 , the matrix elements @xmath364 of the wire hamiltonian acquire an additional time - dependence , @xmath365 the second term in the hamiltonian stems from @xmath366 . owing to the zero time - average of the voltage @xmath357 , the phase @xmath367 is @xmath160-periodic . therefore , the transformed wire hamiltonian is also @xmath160-periodic while the contact and the lead contributions are time - independent , thus , @xmath368 is of form . in ref . , the expressions and for the current and the noise in the wide - band limit have been derived by eliminating the leads in favor of a stochastic operator . in this appendix , we detail this approach . like in section [ sec : scattering ] , we start here also from the heisenberg equations for the annihilation operators . the ones for the lead operators , eq . , are easily integrated to read @xmath369 and @xmath370 accordingly . inserting into the heisenberg equations for the wire operators yields @xmath371 owing to the wide - band limit , the dissipative terms are memory free . within the chosen grand canonical ensembles the operator - valued gaussian noise @xmath372 obeys @xmath373 the current operator then assumes the form @xmath374 the homogeneous set of equations that corresponds to coincides with the equations of motion and which are solved by the floquet states @xmath168 . thus , the floquet states @xmath168 together with the adjoint states @xmath167 , allow to write the solution of in closed form . in the asymptotic limit @xmath63 , it reads @xmath375 where @xmath376 is the propagator for the wire electrons . to obtain the current @xmath132 , we insert the operator into the expression and use the expectation values . with the green function , we find the still unsymmetric expression @xmath377 f_l(\epsilon ) \big\ } . \end{split}\ ] ] for a symmetrization , we eliminate the backscattering terms , i.e. , terms containing @xmath378 , by use of the relation @xcite @xmath379 which follows readily from the floquet representation of the propagator and the floquet eigenvalue equation together with its adjoint . a subsequent fourier transformation with respect to @xmath380 yields by inserting the matrix element @xmath381 , we obtain from for the time - dependent current the symmetric expression . to derive an expression for the zero frequency noise , we insert the operator into the definition of the current - current correlation function and integrate over the times @xmath382 and @xmath47 . again , we employ the relation to bring @xmath383 into the symmetric form .

granular materials consisting of macroscopic particles or grains can exhibit behavior reminiscent of conventional phases of matter . sand , for instance , can flow like a liquid under some conditions . dilute granular systems , or _ gases _ , have been extensively studied both experimentally and theoretically , in large part as simple model systems exhibiting nonequilibrium and dissipative behavior . these systems are intrinsically dissipative and out of equilibrium , even though it is tempting to apply such equilibrium notions as temperature . since the collisions in such a gas are inelastic , it is necessary to supply energy , or to drive them in order to maintain a gas - like steady state . otherwise , _ inelastic collapse _ can occur , in which all motion ceases after only a finite time@xcite . in principle , it is possible to heat or drive the system uniformly throughout the container ( _ uniform heating _ ) , as has been done in simulations@xcite , and as is assumed in analytic theories@xcite . in experiments , however , one usually drives a granular gas by shaking or vibrating the walls of the container . such _ boundary heating _ means that the energy is inserted in a spatially inhomogeneous way@xcite . as a consequence , the gas will develop a gradient in density and mean kinetic energy . even for uniform heating , however , significant deviations from equilibrium gases , e.g. , in density correlations , are observed@xcite . it is often assumed that in the bulk , where the boundaries are far away and spatial gradients are small on the scale of the mean free path of the particles , both heating methods should give the same behavior . here , we show that there are striking differences in the behavior of granular gases heated uniformly or through the boundary . there has been much interest recently in another aspect of these non - equilibrium gases : namely , the velocity distributions . this is in part because they deviate from the gaussian distributions that one would expect if the collisions were elastic . it has been suggested that the velocity distributions can be described by a sort - of _ stretched _ gaussian , of the form @xmath6 $ ] , where @xmath7 is sometimes called the granular temperature . rouyer and menon @xcite have reported such a distribution with exponent @xmath8 over the whole observed range of velocities , which was unaffected by changes in amplitude and frequency of driving . this observation was particularly intriguing , given the theoretical predictions obtained before by van noije and ernst @xcite . they developed a kinetic theory that predicts a high - velocity tail described by a distribution with an exponent @xmath9 . unfortunately , results for velocity distributions are never unambiguous . both in simulation and experiment , different setups and driving mechanisms usually give different behavior of the velocity distribution . for a setup where particles on a horizontal plate were driven in the vertical direction , olafsen and urbach @xcite found a crossover from exponential to gaussian distributions as the amplitude of the driving was increased . the result of rouyer and menon @xcite was obtained for a different configuration where particles were confined between two vertical plates and driven in the vertical direction . although the exponent @xmath8 they find is reminiscent of theoretical results obtained by van noije and ernst @xcite , both results are actually inconsistent in that the exponent @xmath9 can only describe the high - velocity tail of the distribution . such a high - velocity tail was actually obtained by moon et al . @xcite . using a simulation with a variation on uniform heating they find a velocity distribution that has an exponent of @xmath10 for low velocities , but crosses over to an exponent of @xmath8 for the high - velocity tail . the results of rouyer and menon are also limited in the sense that they were obtained for almost elastic particles with @xmath11 . blair and kudrolli @xcite use a different setup where particles move along an inclined plane . friction with the plane during collisions reduces the coefficient of restitution to @xmath12 . they find the distribution with exponent @xmath8 only in the very dilute case . otherwise , the distributions deviate strongly from both gaussian and the distribution obtained by rouyer and menon . it is interesting to notice that for the denser case blair and kudrolli find a distribution with a crossover much like the type observed by moon et al . at present there is no agreement on what the velocity distribution of a granular gas looks like exactly nor is it clear what causes the velocity distributions to deviate from a gaussian . puglisi et al . @xcite have suggested that the deviations are caused by the spatial correlations in the gas . they propose that for a region of uniform density the velocity distribution of the gas actually is gaussian , with a density dependent width . the spatial correlations cause density fluctuations and they claim that the non - gaussian distributions arise as an average over the velocity distributions over these regions of different density . it is true that an average over gaussian distributions with different widths in general yields a stretched gaussian , but experiments performed by olafsen and urbach @xcite showed that this does not happen in granular gases . they find that even within small windows of uniform density the velocity distributions remained non - gaussian . here , we show that non - gaussian behavior arises even in a simple model with no spatial degrees of freedom . we study behavior of the velocity distributions of the granular gas as a function of @xmath1 , the area fraction , and @xmath0 , the coefficient of restitution . specifically , we consider the effect on the velocity distribution of driving the gas by heating uniformly , as is assumed in theory and many prior simulations , and by heating through a boundary , as is done in most experiments . we will show that there exists clear qualitative difference between the velocity distributions for uniform and boundary heating . furthermore , we will show that there is no evidence for a universal velocity distribution with a constant exponent @xmath8 . we demonstrate instead that a family of distributions showing apparent exponents covering a wide range of values @xmath13 is expected , depending on both material and experimental conditions . furthermore , we show that the velocity distribution is governed primarily by the relative importance of collisions to heating , _ i.e. _ , the way in which energy flows through the system of particles . specifically , we introduce a new parameter @xmath2 , which measures the ratio between mean numbers of heating events @xmath14 and mean number of collisions @xmath15 experienced by a typical particle . these theoretical observations can explain both the observed non - gaussian behavior , as well as the ambiguities in the experimental and theoretical literature on dissipative gases to date . we also demonstrate that the behavior of the velocity distributions can be captured quantitatively by a simple model that takes only @xmath0 and @xmath3 into account , with no spatial degrees of freedom . we use an event - driven algorithm to simulate @xmath16 particles of radius @xmath17 moving in a two - dimensional box . particles gain energy by heating and lose energy through inelastic collisions . when two particles @xmath18 and @xmath19 collide their final velocities depend on their initial velocities in the following way : @xmath20 where @xmath21 is the coefficient of restitution and @xmath22 is the unit vector connecting the centers of particles @xmath18 and @xmath19 . for uniform heating we adapted an one - dimensional algorithm described in @xcite . when heating uniformly , each individual particle is heated by adding a random amount to the velocity of each particle during a time step @xmath23 : @xmath24 where @xmath25 is a vector whose components are uniformly distributed between @xmath26 and @xmath27 and @xmath28 is proportional to the heating rate . after heating the system is transferred to the center - of - mass frame . particles move in a box with sides @xmath29 . we use periodic boundary conditions to simulate bulk behavior . the time step @xmath23 is chosen in such a way that on average the number of collisions per time step is less than one . it should be noted that this heating mechanism is significantly different from the spatially homogeneous heating used in some experiments @xcite . in the experiment all particles feel the same forcing , so the motion of the neighboring particles is strongly correlated in space and time . in uniform heating however , all particles are independently driven by a stochastic source and as a consequence correlations are very weak . when heating through the boundaries , particles gain velocity upon collision with the boundary . for simplicity , we assume that the collision with the boundary is elastic . in that case , a collision occurs by reflecting @xmath30 , the component of the velocity perpendicular to the boundary . heating occurs by adding a random amount of velocity to @xmath30 . then after collision with the boundary one has : @xmath31 particles move in a circular box of radius @xmath32 . a symmetrical container has the advantage that it allows us to examine density and granular temperature gradients along a single coordinate @xmath17 , the distance from the center of the box , as in the one - dimensional case @xcite . we start the simulation by distributing the particles uniformly over the box . when using boundary heating , we give each particle a small , uniformly distributed velocity to enable particles to reach the boundary . then particles are heated and we allow the system to reach a steady state before taking data . for both uniform heating and boundary heating , data is taken periodically every time step @xmath23 . for uniform heating , data is taken when the particles are heated , so @xmath23 equals the time between heating events . one of the first striking differences between uniform heating and boundary heating is clustering . when heating through the boundary , a stable liquid - like cluster surrounded by a hot gaseous state will form for low coefficients of restitution @xmath0 or high area fraction @xmath1 . this occurs as particles are compressed in the center of the box by the pressure of particles moving in from the boundary . as the cluster grows in size , it can no longer be destroyed by the impact of high velocity particles and the cluster remains stable . a typical example is shown in fig . [ fig : cluster ] . we do not find these clusters when heating uniformly . this is because all particles are heated all the time , which prevents the collapse to a cluster . in experiments that find clusters in homogeneously driven systems @xcite this is different , because here the driven motion of neighboring particles is highly correlated . the formation of a cluster spoils measurement of the velocity distribution because the system is no longer in a pure gas - like state . to avoid values of @xmath1 and @xmath0 corresponding to the formation of clusters in our simulation , we constructed a phase diagram . we did this by counting for every particle the number @xmath33 of neighbors with their center within a distance @xmath34 from that particle . when the gas is in a hexagonal close packed state @xmath35 . we obtained the distribution @xmath36 for different values of @xmath1 and @xmath0 . an example for @xmath37 and @xmath38 is shown in figure [ fig : densityhist ] . , @xmath39 and @xmath40 . the circles indicate the current positions of the particles , while the lines indicate the direction and magnitude of their velocity . ] of a given particle for @xmath37 , @xmath38 and @xmath41(@xmath42 ) , @xmath43(@xmath44 ) , @xmath45(@xmath46 ) , @xmath47(@xmath48 ) and @xmath49(@xmath50 ) . on average @xmath51 for @xmath38 . ] for @xmath52 the distribution corresponds to a state with the particles uniformly distributed over the box and the peak of the distribution at the mean value @xmath53 . for @xmath41 the distribution becomes bimodal , with a broad peak at high @xmath33 corresponding to the densely - packed cluster and a peak at @xmath54 corresponding to the surrounding dilute gas . the distribution shows a continuous transition for @xmath0 in between , which makes it hard to pinpoint an exact value of @xmath0 for which the gas enters the clustered state . still , by looking at the shape of the distributions , it can be argued that the transition occurs somewhere between @xmath55 and @xmath56 . this was repeated for different values of @xmath1 , which allowed us to determine a sort of phase diagram . specifically , we determined the limit of a pure gas - like phase , and all results presented below were obtained in this state . . ( b ) @xmath57\}$ ] versus @xmath58 . data for both figures is taken for @xmath37 and for @xmath59 and @xmath60(@xmath42 ) , @xmath61(@xmath44 ) , @xmath62(@xmath46 ) , @xmath63(@xmath48 ) , @xmath64(@xmath50 ) . a gaussian is shown as a solid line with slope @xmath65 and the distribution obtained by rouyer and menon is shown as the dashed line with slope @xmath66 . ] . ( b ) @xmath57\}$ ] versus @xmath58 . the dashed lines have slope @xmath65 and @xmath66 . data for both figures is taken for @xmath37 and for @xmath67 and @xmath38(@xmath42 ) , @xmath68(@xmath44 ) , @xmath69(@xmath46 ) , @xmath70(@xmath48 ) . ] as a function of distance @xmath17 to the center of the box . data taken for @xmath37 , @xmath59 and @xmath52(@xmath42 ) , @xmath45(@xmath44 ) , @xmath71(@xmath46 ) , @xmath61(@xmath48 ) , @xmath72(@xmath50 ) and @xmath62(@xmath73 ) . ( b ) the mean kinetic energy @xmath74 per particle as a function of @xmath17 , for the same values of @xmath1 and @xmath0 . the dashed lines indicate the concentric rings , within which the velocity distributions were separately calculated . ] the velocity distributions @xmath75 for uniform heating are shown in figs . [ fig : uniform_phi ] and [ fig : uniform_eta ] . the velocity component @xmath76 is scaled by @xmath77 and the maximum of the distribution @xmath78 is scaled to be unity . for a broad range of the parameters @xmath1 and @xmath0 the velocity distributions are very close to gaussian . for @xmath60 the velocity distributions can be fitted by a distribution with @xmath10 . this decreases only slightly for @xmath79 , which can be fitted by a distribution with @xmath80 . values of @xmath81 are found to be independent of @xmath1 . these exponents are constant over the entire observed range of velocities and we see no high - velocity tails with @xmath8 for the range of @xmath1 and @xmath0 we examined . as a function of distance @xmath17 to the center of the box . data taken for @xmath37 , @xmath52 and @xmath38(@xmath42 ) , @xmath68(@xmath44 ) , @xmath69(@xmath46 ) and @xmath70(@xmath48 ) . ( b ) the mean kinetic energy @xmath82 per particle as a function of @xmath17 , for the same values of @xmath1 and @xmath0 . note that even for the dilute case @xmath83 the mean kinetic energy profile is not constant , but drops at the boundary of the box . the profile only becomes constant after a certain distance into the container , that corresponds to the mean free path of particles leaving the boundary . ] and [ fig : profile_eta ] , _ i.e. _ , for @xmath84(@xmath42 ) , @xmath85(@xmath44 ) , @xmath86(@xmath46 ) , @xmath87(@xmath48 ) and @xmath88(@xmath50 ) , where @xmath17 is distance to the center . data were taken for @xmath37 , @xmath39 and @xmath60 . ( a ) @xmath89 . ( b ) @xmath57\}$ ] versus @xmath58 . ] for boundary heating the gas develops a gradient in both density and mean kinetic energy as shown in figures [ fig : profile_phi ] and [ fig : profile_eta ] . ideally , we want to measure velocity distributions in a region where the gradient is small . to this end we divided the box in five rings of width @xmath63 . these rings are indicated in figures [ fig : profile_phi ] and [ fig : profile_eta ] . only for values of @xmath1 and @xmath0 close to the clustering state , does the density within a ring vary by more than @xmath90 . the velocity distributions @xmath75 for particles within the different rings are shown in figure [ fig : boundaryexample ] . figure [ fig : boundaryexample](a ) shows @xmath75 with the velocity component @xmath76 scaled by @xmath77 and the maximum of the distribution scaled to be unity . all distributions seem to have the same shape . this is remarkable , since naively one would expect the dynamics of the particles to be different for the different rings . in the center the density is high and particles have many collisions , causing strong spatial correlations . along the boundary the gas is dilute . here , many particles have experienced recent collisions with the boundary , and spatial correlations should be much weaker . it is has been suggested that these correlations cause the velocity distributions to deviate from gaussian . but contrary to the expectations , within the container the shape of the distributions seems to be determined by the granular temperature @xmath91 only . this is a feature that is observed for all values of @xmath1 and @xmath0 , even close to the cluster state . figure [ fig : boundaryexample](b ) shows the behavior of the exponent @xmath81 . this behavior is very different from the case of uniform heating . for uniform heating @xmath81 has the same value over the entire observed range of velocities . for boundary heating , on the other hand , @xmath81 has a constant value @xmath92 over the low - velocity range but crosses over to different value @xmath93 when above a critical velocity @xmath94 . for the two innermost rings @xmath95 and @xmath96 . for the third ring @xmath97 and the outer ring has @xmath8 . for the outer two rings the velocity distributions actually undergo two crossovers . first it crosses over to @xmath98 and then the exponent increases again to @xmath99 . in figure [ fig : boundary2 ] we show the effect of a change in @xmath1 and @xmath0 on the shape of the velocity distributions . here we focus on the velocity distribution as measured in the ring with @xmath100 . this has the advantage of good statistics , but for values of @xmath1 and @xmath0 close to a cluster , we might see effects due to the density gradient in the gas . as shown in figure [ fig : boundary2](a ) the exponent @xmath97 except for @xmath101 , where @xmath102 . for @xmath52 there is no crossover in the observed range of velocities . in the other distributions one does observe a crossover and the point where it occurs shifts down to lower velocities as @xmath0 is decreased . it is clear that the distribution for velocities above the crossover can not be described by a single exponent . for low enough @xmath0 , the distribution seems to approach a constant exponent for high velocities . this exponent decreases from @xmath103 for @xmath41 to @xmath96 for @xmath101 . in figure [ fig : boundary2](b ) @xmath97 for @xmath83 and @xmath59 , @xmath104 for @xmath105 and @xmath95 for @xmath39 . for every @xmath1 does one observe a crossover and the velocity at which the crossover occurs shifts only a bit as @xmath1 is varied . the distributions approach a constant exponent for high velocities . this exponent goes down from @xmath106 for @xmath83 to @xmath96 for @xmath39 . in general , the deviations from gaussian become more pronounced as dissipation increases , i.e. as @xmath1 increases or as @xmath0 decreases . \}$ ] versus @xmath58 for @xmath37 , @xmath59 and @xmath52(@xmath42 ) , @xmath45(@xmath44 ) , @xmath71(@xmath46 ) , @xmath61(@xmath48 ) , @xmath72(@xmath50 ) and @xmath62(@xmath107 ) . ( b ) @xmath57\}$ ] versus @xmath58 for @xmath37 , @xmath41 and @xmath83(@xmath42 ) , @xmath69(@xmath44 ) , @xmath108(@xmath46 ) and @xmath68(@xmath48 ) . ] to test whether the velocity distributions we find here are only observed for this specific driving mechanism of heating through a circular boundary , we constructed different systems that drive through boundaries in a different way . for instance , we constructed a box with periodic boundary conditions that includes a small circular region around the center . within this circular region particles are uniformly driven but outside of the region they are not heated at all . for particles within the circular region we observe velocity distributions that are gaussian . on the other hand , for particles outside of the circular region we observe the same non - gaussian velocity distributions as seen in the case of a circular boundary . for uniform heating the distribution of velocities of particles that are heated is gaussian , because the velocities undergo a random walk . this is not the case for the particles heated at the boundary . to see if velocity distributions are sensitive to the precise way of heating at the boundary , we changed our heating algorithm so that , when a particle hits a boundary , it s new velocity is drawn from a gaussian distribution this has a modest effect on the far end of the high - velocity tail , but leaves all major differences between uniform and boundary heating intact . for different values of @xmath16 . ( b ) @xmath57\}$ ] versus @xmath58 . data is taken for @xmath39 , @xmath60 and @xmath109(@xmath42 ) , @xmath110(@xmath44 ) , @xmath111(@xmath46 ) , @xmath112(@xmath48 ) and @xmath113(@xmath50 ) . ] finally , we studied the behavior of the velocity distribution for different particle numbers @xmath16 . figure [ fig : diffn](a ) shows that , as @xmath16 increases , the velocity distributions fall off more rapidly , but the high - velocity tails grow longer . this is shown more clearly in figure [ fig : diffn](b ) . for the high - velocity tail we find @xmath114 for @xmath109 decreasing to @xmath115 for @xmath116 . the crossover shifts to lower velocity and becomes sharper as @xmath16 is increased . as we will demonstrate in more detail below , for boundary heating there is no thermodynamic limit . instead of approaching a limiting velocity distribution as @xmath16 is increased , we find that the shape of the velocity distribution depends not only on @xmath0 and @xmath1 , but also on @xmath16 . \}$ ] versus @xmath117 for @xmath37 , @xmath39 and @xmath52 ( @xmath42 ) , @xmath118 , @xmath39 and @xmath52 ( @xmath44 ) , @xmath37 , @xmath39 and @xmath60 ( @xmath46 ) , @xmath37 , @xmath119 and @xmath52 ( @xmath48 ) . the solid lines correspond to the fit as made by rouyer and menon and has an exponent @xmath120 . the range of the solid lines corresponds to half the range used by rouyer and menon in their fit , but contains about 80@xmath121 of their data points ] . [ fig : rouyer ] for their experiments rouyer and menon used @xmath16 particles with @xmath122 , where @xmath123 and @xmath124 @xcite . in figure [ fig : rouyer ] we plotted the velocity distribution for @xmath52 , @xmath39 and several values of @xmath16 . the solid black lines indicated the fit with @xmath120 as made in @xcite . this line clearly coincides with the high - velocity tail of the velocity distribution found by simulation . this suggests that it is possible that instead of a universal distribution with @xmath8 , they only observed a part of a more complex velocity distribution , with two exponents and a crossover . for higher @xmath1 the high - velocity tails have exponents that are smaller than the @xmath8 observed for @xmath39 . for different values of @xmath23 . ( b ) @xmath57\}$ ] versus @xmath117 . data is taken for @xmath37 , @xmath59 , @xmath101 and @xmath125(@xmath42 ) , @xmath108(@xmath44 ) , @xmath68(@xmath46 ) , @xmath126(@xmath48 ) , @xmath127(@xmath50 ) , @xmath128(@xmath129 ) and @xmath130(@xmath73 ) . ] the main difference between uniform and boundary heating is that in the first case heating takes place homogeneously throughout the box , whereas in the latter case energy is injected inhomogeneously at the boundaries . this is not the direct cause for the difference in velocity distributions . this is demonstrated by the occurrence of non - gaussian velocity distributions in a system studied by moon et . they study a uniformly driven gas , but nevertheless find velocity distributions with a clear crossover . the main difference is the way they implement uniform heating . every time @xmath23 , they select at random two particles and add to these particles a random but opposite velocity to conserve the total momentum . on average heating is spatially homogeneous and in the limit of very small @xmath23 the behavior of both their and our uniform heating algorithm is the same . for small enough @xmath23 the number of collisions between heating events is smaller than one and the heating dominates over the dissipative behavior of the gas . when @xmath23 becomes larger , many collisions might happen between two heating events . apperently this changes the velocity distributions . in figure [ fig:2point ] we show the velocity distribution for @xmath37 , @xmath59 and @xmath101 . the gas is heated using the two - point heating algorithm described above , varying the time between heatings , @xmath23 . for @xmath125 the distribution has a exponent @xmath114 that is constant approximately over the observed range . when @xmath23 is reduced a clear crossover develops . the behavior of the velocity distribution for velocities higher than the crossover velocity is more complicated than in boundary heating there is also a sharp kink at the high - velocity end , that we have been unable to explain so far . this can explain the differences between velocity distributions for uniform and boundary heating . when heating uniformly all the particles are heated every timestep @xmath23 , while only a few collisions occur in that time . the system is dominated by heating and we see a velocity distribution with a constant exponent over the entire observed velocity range . if we are heating through a boundary , there are many collisions in the dense region in the center of the box , but only a few particles escape to the boundary to get heated . in this case , the system is dominated by dissipation and we see in general a velocity distribution that strongly deviates from a gaussian . this difference between uniform and boundary heating can be quantified by the ratio @xmath2 , where @xmath14 and @xmath15 are the mean number of heatings and the number of collisions a particle has experienced , respectively . for a uniformly heated gas of @xmath37 , @xmath59 and @xmath101 we find that @xmath131 . for the same parameters but heating through the boundary we find @xmath132 . for again the same parameters but using two - point heating with @xmath133 we find @xmath134 . we tested the effect of changing @xmath3 on velocity distributions obtained with boundary heating . when increasing the number of particles @xmath16 or the area fraction @xmath1 , the average number of collisions increases . one can show in a mean field approximation that @xmath135 . the average distance a particle travels between collisions is given by @xmath136 . the average distance a particles travels between heatings @xmath137 is given by the dimensions of the container . for a box of area @xmath138 the average distance between boundaries is given by @xmath139 . finally , we know @xmath140 . our simulation obeys this approximation very well . in fig . [ fig : collapse ] we show velocity distributions for @xmath60 and different combinations of @xmath16 and @xmath1 . we measure the heating - dissipation ratio @xmath3 in the simulation and show velocity distributions with the same @xmath3 on top of each other . for @xmath141 and @xmath142 we find excellent collapse for different @xmath16 and @xmath1 , even when we scale the system by a factor @xmath143 . for @xmath144 , where spatial correlations become very strong , we still find reasonable collapse . as we increase @xmath3 we observe the usual pattern , where a crossover appears in a distribution that was initially close to a gaussian . this means that the velocity distributions are almost exclusively controlled by two parameters ; the coefficient of restitution @xmath0 , which is a material parameters , and the heating - dissipation ratio @xmath3 , which depends on experimental conditions . . distributions with the same @xmath3 are shown on top of each other . ( a ) @xmath141 and we show @xmath109 and @xmath145(@xmath42 ) , @xmath146 and @xmath147(@xmath44 ) , @xmath148 and @xmath149(@xmath150 ) . ( b ) @xmath142 and we show @xmath109 and @xmath151(@xmath42 ) , @xmath146 and @xmath152(@xmath44 ) , @xmath153 and @xmath59(@xmath150 ) . ( c ) @xmath144 and we show @xmath109 and @xmath154(@xmath42 ) , @xmath146 and @xmath155(@xmath44 ) , @xmath153 and @xmath38(@xmath150 ) . inset : heating - dissipation ratio @xmath3 for @xmath148(@xmath42 ) , @xmath153(@xmath44 ) , @xmath146(@xmath150 ) and @xmath109(@xmath156 ) for several values of @xmath1 . the line is a fit of the form @xmath157 . ] to study the dependence of the velocity distributions on the ratio @xmath3 we constructed a model of the inelastic gas inspired by work done by ulam @xcite on elastic gases . ulam rederived the maxwell - bolzmann distribution for the velocities of particles in a perfect gas using the following very simple model : he considered @xmath16 particles with random initial velocities . at every time step a pair of particles was selected and the velocities of the particles were changed as if they had collided elastically . ulam found that , regardless of the initial velocities , the system quickly evolved to a state where on average all energy was distributed equally over all particles and the deviation of the velocities of the particles around the average velocity of the system was given by the maxwell - bolzmann distribution . this approximation of random collisions is only justified when the gas is sufficiently dilute . the same approximation can also be made for the granular gas by replacing the elastic collisions between particles with inelastic collisions . since the gas cools down without any further energy input , heating has to be included as well . this allows us to study the dissipative behavior of the granular gas while neglecting all spatial correlations that develop in the more detailed simulations described before . in particularly we can study the effect on the velocity distributions of changing the fraction @xmath2 described in the previous section by varying the ratio of heatings and collisions . we adapted ulam s procedure in order to model a two - dimensional granular gas consisting of @xmath16 particles . equation [ eqn : inelas ] can be cast into the following form : @xmath158 where @xmath159 and @xmath159 are the velocities of particles @xmath18 and @xmath19 , @xmath0 is the coefficient of restitution and @xmath160 is the angle between the separation vector @xmath161 and a reference axis . collisions in our model occur by selecting at random particles @xmath18 and @xmath19 and an uniformly distributed impact parameter @xmath162 . we then use the above collision rule with @xmath163 , where @xmath164 is the velocity in the center - of - mass frame and @xmath165 is a unit vector along the reference axis . we discard values of @xmath160 corresponding to @xmath166 as these represent unphysical collisions . we heat the gas by selecting a random particle @xmath18 and adding a random amount of velocity according to equation [ eqn : uniheat ] . to prevent the velocities from running away , we subtract the center - of - mass velocity after heating . in a single time step , we have let @xmath167 particles collide and we heat @xmath168 particles . this gives us @xmath169 . when we use inelastic instead of elastic collisions and drive the system , we find that the gas heats up until it reaches a steady state , where on average the energy lost in collisions is compensated by the heat inserted into the system through our driving mechanism . as @xmath168 is chosen increasingly bigger in comparison to @xmath170 , the granular temperature @xmath171 of the gas increases accordingly . and @xmath101 . ( a ) @xmath78 . ( b ) @xmath57\}$ ] versus @xmath58 . data is for different values of @xmath172 : @xmath173 ( @xmath42 ) , @xmath174 ( @xmath44 ) , @xmath175 ( @xmath46 ) , @xmath176 ( @xmath48 ) , @xmath177 ( @xmath50 ) , @xmath178 ( @xmath129 ) and @xmath179 ( @xmath73 ) . ] in figure [ fig : diff_f ] we plotted the result for an inelastic gas with @xmath101 . we varied the number of heatings and the number of collisions in a single time step from @xmath180 and @xmath181 to @xmath182 and @xmath183 . as @xmath3 is lowered , the velocity distributions develop a crossover and for @xmath184 the distributions are strongly non - gaussian , similar to the velocity distributions obtained for @xmath185 in two - point heating . in figure [ fig : diffeta ] we keep @xmath186 fixed and vary @xmath0 . we see that the crossover point shifts down as @xmath0 is lowered and that the kink in the velocity distribution shifts down . and @xmath186 . ( a ) @xmath78 . ( b ) @xmath57\}$ ] versus @xmath58 . data is for @xmath52 ( @xmath42 ) , @xmath71 ( @xmath44 ) , @xmath72 ( @xmath46 ) , @xmath187 ( @xmath48 ) and @xmath64 ( @xmath50 ) . ] we also compared velocity distributions found in the model with those acquired by simulation . to this end we measured @xmath3 in simulations using different ways of heating . for a uniformly heated gas of @xmath37 , @xmath59 and @xmath101 we found @xmath131 . for the same parameters but heating through the boundary we found @xmath132 . for two - point heating with @xmath133 we found @xmath134 . we ran the model with the corresponding set of parameters and measured velocity distributions . the results are shown in figure [ fig : fcompare ] . the velocity distributions agree qualitatively , but especially for the case of q=0.08 there is considerable quantitative difference . still , the model is able to generate non - gaussian velocity distributions , while neglecting all spatial correlations . this suggests that the non - gaussian behavior of the velocity distributions is not caused by spatial fluctuations and correlations in the gas . instead it is the flow of energy through the system , mediated by the inelastic collisions , that determines the shape of the velocity distribution . this also means that there is no a priori difference between inhomogeneous and homogeneous heating : the difference in the shape of the velocity distribution , particularly the occurrence of the crossover , is only a consequence of the fact that in inhomogeneous heating in general the number of collisions between particles exceeds the number of particles being heated . , @xmath59 and @xmath101 . ( a ) @xmath78 for @xmath131 ( solid ) , @xmath132 ( dashed ) and @xmath134 ( dotted ) . results on the bottom are obtained for simulation , results on the top for the model . ( b ) @xmath57\}$ ] versus @xmath58 . the symbols shown are velocity distributions acquired by simulation for @xmath131 ( @xmath42 ) , @xmath188 ( @xmath44 ) , @xmath189 ( @xmath46 ) . the lines show the velocity distributions found in the model for the same values of @xmath3 ( solid , dotted , dashed ) . ] due to the idealized nature of our system , it is not possible to do a direct comparison between our simulation and experiments . in the experiment of rouyer and menon the driving is through the boundary , but there are some significant differences between their heating mechanism and the one we use in simulations with boundary heating . due to gravity and the geometry of the setup the injection of energy in the experiment of rouyer and menon is mainly in the vertical direction . this energy is transferred into the horizontal direction by collisions between particles . another difference is that in the experiment the frequency of driving is relatively low . because of this the dynamics of the gas close to the driving boundary is strongly dependend on the phase of the driving cycle . in fact , it has been shown in simulation that for a system similar to the experiment by rouyer and menon , a shockwave propagates up through the gas @xcite . at a certain distance from the boundary the time dependence has decayed away and the gas enters a steady state . it is in this steady state that the velocity distributions are measured . it is not yet established how this time dependence and the occurence of a shock wave influences the velocity distributions in the steady state . a priori it is not clear if it is possible to compare velocity distributions in systems with a strong time - dependence , like the experiments , with those that have no time dependence , as is the case in our simulations . there are , however , reasons to assume this is possible . the velocity distributions in the experiment of rouyer and menon are measured only in the direction orthogonal to the driving direction . simulations @xcite show that the effect of the shock in the direction orthogonal to the shock are usually relatively weak and decay rapidly in height . in the steady state , influence of the shock is absent in the orthogonal direction , even while it still may be apparent in the direction perpendicular to the driving direction . so , if we only look at velocity distributions in the orthogonal direction and in the steady state , a comparison between the experiment and our simulations is be valid . it is also not clear how in the experiments the dynamics of the gas shape the velocity distribution and whether it is controlled by the parameter @xmath3 . we speculate that in the steady state the system behaves in fact like a one dimensional inelastic gas . fast upward moving particles inject energy in the orthogonal direction when colliding with particles in the steady state , effectively functioning as a heat source . in this picture the mean number of collisions between fast upwards moving particles and particles in the steady state would be @xmath14 , the average number of heatings , and collisions between the particles in the steady state mutually would be @xmath15 , the average number of collisions . one way of changing the shape of the velocity distribution would be changing the fraction of particles in the steady state . more particles in the steady state would lower @xmath14 and increase @xmath15 leading to more non - gaussian velocity distributions . the above considerations not only apply to the experiment of rouyer and menon but to most of the other experiments as well . in the experiments of blair and kudrolli and those of olafsen and urbach velocity distributions are measured orthogonal to the driving direction . in both cases it is not so much the collisions with the bottom plate that drive the gas in the ortogonal directions , but mainly collisions between fast upward moving particles with particles that have low velocities in the orthogonal directions . it is in these experiments rather than those of rouyer and menon that we find a similar dependence on @xmath0 and @xmath3 as we describe in this article . in the setup of olafsen and urbach @xcite velocity distributions go from non - gaussian to gaussian when a rough plate is used instead of a flat plate . on a flat plate , energy is only injected in the in - plane directions by off - angle collisions between neighboring particles . with a rough plate , energy is injected directly into the directions parallel to the plate every time a particle collides with the plate , effectively increasing the number of heatings over collisions . baxter and olafsen @xcite observe the same behavior in a system where a layer of heavy particles is inserted between the other layer of particles and a flat bottom plate . particles from the upper layer have off - angle collisions with the layer of heavy particles , injecting energy in the in - plane directions every cycle . particles in the upper layer show gaussian velocity distributions , whereas particles in the lower layer have non - gaussian velocity distributions . most convincing is the experiment by blair and kudrollli @xcite . here the number of collisions is increased by adding more particles . as a result , their velocity distributions develop the same crossover we see both in our simulations and model . the reason why these transitions are not visible in the experiment by rouyer and menon as they increase number of particles , is that in the first case the effective coefficient of restitution is much lower , @xmath190 , due to friction with the inclined plane . again , one of the main problems considering the velocity distributions in granular gases is that different setups and experiments usually find different velocity distributions . as we have shown in this section , our finding of the controlling parameter @xmath3 could ultimately explain these seemingly inconsistent results . we compared the velocity distributions of a granular gas that was driven by uniform heating and by heating through the boundary . although it is usually assumed that uniform heating yields the same behavior as boundary heating when the boundary is far away and the spatial gradients are small , we find that there are clear qualitative differences . when driven through the boundary , for instance , the gas can form coexisting cool liquid - like clusters surrounded by a hot gaseous state for certain values of @xmath1 and @xmath0 . such clusters do not occur in our simulations with uniform heating . the difference between uniform heating and boundary heating also extends to the velocity distributions . for uniform heating , we find velocity distributions that are close to gaussian over the observed range of velocities and for a wide range of @xmath1 and @xmath0 . when heating through the boundaries , on the other hand , we find that velocity distributions are stretched gaussians that usually cross over from one exponent @xmath81 to another for the high - velocity tail . for the high - velocity tail we find that the exponent @xmath81 varies over a wide range . for strong dissipation this tail can not be described by a single exponent . this means that there is no universal velocity distribution with @xmath8 , as was proposed by rouyer and menon . instead , the velocity distribution found by rouyer and menon may be just one out of the many distributions described here . the apparent universality may be due to the use of a narrow range of parameters ( nearly elastic particles were used with @xmath11 and vary @xmath191 ) . we find that varying @xmath1 only has a minor effect on the shape of the distribution when @xmath0 is so close to the elastic limit . for low @xmath0 , we instead find velocity distributions that look much like the distributions found by blair and kudrolli for gases with @xmath190 . furthermore , we demonstrate that the distribution of velocities for dissipative gases , while not universal in form , depends only on two parameters : the coefficient of restitution @xmath0 ( a material parameter ) and @xmath2 , the average ratio of heatings and collisions in the gas ( a function of experimental conditions ) . we find that velocity distributions range from gaussian for @xmath4 , where heating dominates dissipation , to strongly non - gaussian for @xmath184 , where the dynamics of the gas is dominated by the dissipative collisions between particles . the differences in distributions we find between uniform and boundary heating can then be understood as different limits of @xmath3 , for @xmath4 and @xmath5 respectively . we can control the parameter @xmath3 more easily using two - point heating and here we observe the transistion form gaussian to strongly non - gaussian behavior directly as we change @xmath3 . furthermore , a simple model of a driven , inelastic gas without spatial degrees of freedom reproduces the entire family of velocity distributions we find in simulation , as we vary @xmath0 and @xmath3 . this means that the velocity distributions are non - gaussian not because of spatial correlations . rather , it is the cascade of energy from a few high - energy particles to the slow - moving bulk of the gas that is the key determinant of the non - gaussian velocity distributions . these observations should aid in the construction of a kinetic theory of dissipative gases . 99 s. mcnamara , w.r . young , phys . e 50 , r28 ( 1994 ) t. zhou , l.p . kadanoff , phys . e 54 , 623 ( 1996 ) d.r.m . williams , f.c . mackintosh , phys . e , 54 , r9 s.j . moon , m.d . shattuck , j.b . swift , phys . e 64 , 031303 ( 2001 ) t.p.c . van noije , m.h . ernst , granular matter 1 , 57 ( 1998 ) y.s . du , h. li , l.p . kadanoff , phys . 74 , 1268 ( 1995 ) f. rouyer , n. menon , phys . 85 , 3676 ( 2000 ) a. kudrolli and j. henry , phys . e 62 , r1489 ( 2000 ) ; d.l . blair and a. kudrolli , phys . e 64 , 051301 ( 2001 ) w. losert , d.g.w . cooper , j. delour , a. kudrolli and j.p . gollub , chaos 9 , 682 ( 1999 ) s. luding , e. clment , a. blumen , j. raichenbach , j. duran , phys . e 49 , 1634 ( 1994 ) j.s . olafsen and j.s . urbach , phys . e 60 , r2468 ( 1999 ) ; j.s . olafsen and j.s . urbach , phys . 81 , 4369 ( 1998 ) ; a. prevost , d. a. egolf , j. s. urbach , phys . 89 , 084301 ( 2002 ) ; g. w. baxter , j. s. olafsen , nature 425 , 680 ( 2003 ) a. puglisi , v. loreto , u. m. b. marconi , a. petri and a. vulpiani , phys . lett 81 , 3848 ( 1998 ) s. ulam , adv . math . 1 , 7 ( 1980 ) j. bougie , s. j. moon , j. b. swift , h. l. swinney , phys . e 66 , 051301 ( 2002 )

multi - user mimo ( mu - mimo ) is a key building block of the next generation wireless communication system . in a mu - mimo downlink system , a base station ( bs ) equipped with multiple antennas simultaneously transmits data to a group of multiple antenna users . the multi - user interference , which is the major performance limiting factor of mu - mimo systems , must be managed intelligently using the physical layer techniques such as beamforming . in general , there are two major objectives in the beamformer design problem . one is to maximize a system utility ( e.g. , throughput ) under some power constraint , while the other one is to minimize the system transmission power subject to qos requirements @xcite . although both formulations are justifiable and well - studied , the latter is more appropriate for scenarios where the users need a guaranteed qos level . although the qos - constrained power minimization ( qcpm ) problem is globally solvable in polynomial time for miso system @xcite , this problem is highly non - convex and difficult to solve in mimo systems . in fact , it has been shown that when the bss and the users are equipped with more than two antennas , the problem becomes np - hard @xcite . therefore , many algorithms have been proposed to solve this problem approximately . for example , the references @xcite and @xcite propose algorithms based on interference nulling , which can completely eliminate the inter - user interference . in these algorithms , the search space of transmit and receive beamformers is limited to zero forcing transceivers , leading to simple but suboptimal solutions . in addition , such interference nulling based methods require that the number of transmit antennas is no less than the total number of active users , which is impractical in many scenarios . another approach for solving the mu - mimo downlink qcpm problem is based on the iterative optimization methods @xcite . references @xcite provide iterative algorithms that update the transmit beamformers , receive beamformers , and transmit powers by switching between the downlink and the dual uplink channels . central to these methods is the notion of uplink - downlink duality ( udd ) theory @xcite , which guarantees that a set of target sinr levels is achievable in the downlink channel if and only if the same set of sinr levels is achievable in the corresponding dual uplink channel . we refer to such algorithms as udd algorithms . the udd algorithm was first proposed in @xcite for the multi - antenna case , where the _ receive _ ( resp . transmit ) beamformer update is followed immediately by the _ transmit _ receive ) power update . the udd algorithm of @xcite differs from that of @xcite in the order of updating transmit / receive beamformers and powers . in the udd algorithm of @xcite , the _ transmit _ ( resp . receive ) power is updated exactly after the _ transmit _ ( resp . receive ) beamformer . importantly , it is shown in @xcite that the udd algorithm monotonically decreases the total power consumption while satisfying the qos constraints . however , the algorithms in @xcite can only apply to the _ single stream case_. in @xcite the udd algorithm has been generalized to the multiple stream case under the assumption of no joint detection at the receivers ( i.e. , _ inter - stream _ interference is considered ) . notice that , to the best of our knowledge , it is still not known whether the udd algorithm converges to a kkt solution . different from the previous works , the references @xcite and @xcite have proposed an iterative algorithm named mmse - socp , which aims to solve the qcpm problem directly . the algorithm consists of the following two key steps : 1 ) fixing the transmit power , update the receive beamformers using mmse receiver , and 2 ) fixing the receive beamformers , update the transmit beamformers by solving the power minimization problem with respect to the transmit beamformers . the authors show that the total transmit power monotonically decreases and thus converges . however , the convergence of the mmse - socp iterates to a kkt solution is not known . nevertheless , the authors of @xcite and @xcite observe that the mmse - socp algorithm always generates a sequence that converges to a unique solution , irrespective of the initial point ( or at least with high probability , see @xcite ) . hence , a conjecture has been made in @xcite , stating that the mmse - socp algorithm probably reaches a local optimum solution of the qcpm problem in this paper , we settle the convergence issue related to the mmse - socp and the udd algorithm . we show that both algorithms converge globally to the set of kkt solutions , regardless of the number of streams intended for each user . we start by analyzing the kkt conditions of the qcpm problem of the single stream case . based on the analysis , we propose a novel iterative algorithm called mmse - dual algorithm , which is essentially equivalent to the mmse - socp algorithm , but with the added benefits of having almost closed form updates and a lower complexity . through the mmse - dual algorithm , we reveal some connections between the udd algorithm@xcite and the mmse - socp algorithm @xcite . more importantly we prove that both the mmse - dual algorithm and the udd algorithm monotonically converge to the set of kkt solutions of the qcpm problem . in addition , we extend the algorithms to the multiple stream case and prove that they can also reach a kkt point of the qcpm problem under some mild conditions . as will be seen later , the mmse - dual algorithm has a lower complexity than the udd algorithm if the number of streams is greater than the number of transmit antennas . moreover , although both algorithms require feasible initialization , it is easier for the mmse - dual algorithm to obtain a feasible initialization ( see remark [ rem : feasibility ] in section iv ) . the remainder of this paper is organized as follows . in section ii we give the formulation of the qcpm problem and provide a brief review of the existing algorithms . in section iii we propose the mmse - dual algorithm which reveals the relation between the mmse - socp and the udd algorithm . then , we state the convergence results of the three algorithms in section iv and extend the algorithms to the multiple stream case in section v. finally , section vi presents some simulation results and section vii concludes the paper . _ notations _ : throughout this paper , we use uppercase bold letters for matrices , lowercase bold letters for column vectors , and regular letters for scalars . the superscript @xmath0 is used to denote the hermitian transpose of a matrix . for a complex number @xmath1 , @xmath2 and @xmath3 denote its phase angle and imaginary part , respectively . for a function @xmath4 , @xmath5 denotes its gradient with respect to the variable @xmath6 . for a matrix @xmath7 , @xmath8 indicates that @xmath7 is a positive semidefinite matrix . @xmath9 denotes the identity matrix of an appropriate size . the circularly symmetric complex gaussian distribution is represented by @xmath10 , where @xmath11 is the mean and @xmath12 is the variance of the distribution . the notations @xmath13 and @xmath14 represent the trace and the determinant operator , respectively . consider a multi - user mimo downlink system with @xmath15 users , where the bs is equipped with @xmath16 antennas and each user @xmath17 is equipped with @xmath18 antennas . let us use @xmath19 to denote the set of all users . assume for now that the bs transmits the single stream @xmath20 to the intended receiver @xmath17 with no multiplexing ( the multiple stream case will be considered in section vi ) . let us also assume that the bs utilizes the transmit beamformer @xmath21 to send the data stream @xmath20 to user @xmath17 . on the other side , user @xmath17 utilizes the receive beamformer @xmath22 to estimate its transmitted data stream . the estimated data stream @xmath23 can be mathematically expressed as@xmath24 where @xmath25 denotes the channel matrix from the bs to the receiver @xmath17 ; @xmath26 is the additive white gaussian noise ( awgn ) with distribution @xmath27 . the data streams @xmath20 s are i.i.d . and independent of the noise level ; and have distribution @xmath28 . we are interested in designing the transmit and receive beamformers to minimize the transmit power while the users qos requirements are satisfied . let us consider the signal - to - interference - plus - noise ratio ( sinr ) as the qos measure . the sinr of user @xmath17 is given by : @xmath29 mathematically , the qcpm problem can be written as @xmath30 where @xmath31 is the intended sinr level of user @xmath17 ; the set of all transmit beamformers ( resp . receive beamformers ) is denoted by @xmath32 ( resp . @xmath33 ) . throughout the paper , we assume that problem ( p1 ) is feasible and @xmath34 for all @xmath17 . in this subsection , we briefly review the existing mmse - socp algorithm @xcite and the udd algorithm @xcite . the mmse - socp algorithm alternates between the following two steps : * fixing all the transmit beamformers , update the receive beamformers using the mmse receiver , i.e. , @xmath35 * fixing all the receive beamformers , update the transmit beamformers by solving @xmath36 which can be transformed to a second - order cone program ( socp ) @xcite . the above socp has @xmath37 unknowns and can be efficiently solved via interior - point algorithm ; each iteration of the interior - point algorithm has computational complexity of @xmath38@xcite . let @xmath39 and @xmath40 denote the normalized beamformer , i.e. , @xmath41 and @xmath42 with @xmath43 . we refer @xmath44 and @xmath45 as the normalized beamformers , @xmath46 and @xmath47 as the power consumption . using these notations , the downlink channel model can be rewritten as @xmath48 the udd algorithm is established by introducing a _ virtual _ uplink channel , which can be constructed through the following three steps : 1 ) reverse the directions of all links ; 2 ) replace the channel matrices by their conjugated transposed version ( i.e. , @xmath49 @xmath50 ) ; 3 ) exchange the role of transmit beamformer and receive beamformer . mathematically , the virtual uplink channel model can be represented as @xmath51 where @xmath52 is the virtual uplink channel awgn . in terms of the channel models and , the sinrs for the downlink and uplink are respectively expressed as @xmath53 and @xmath54 then , fixing the beamformers , the _ downlink _ power minimization problem can be written as @xmath55 its dual problem can be obtained by using the lagrange duality theory with @xmath56 corresponding to the lagrangian multiplier of the @xmath17th qos constraint@xcite@xmath57 it can be shown that problem is equivalent to the following _ uplink _ weighted power minimization problem@xcite @xmath58 by noting that all the inequality constraints of both problem and must hold with equality at the optimality and furthermore the corresponding system of linear equations with respect to @xmath59 has a unique solution ( * ? ? ? * lemmas 1 & 2 ) . to summarize , the classical lagrange duality theory leads to the well - known uplink - downlink duality theorem : ( uplink - downlink duality @xcite ) : _ for any given set of normalized beamformers @xmath60 and @xmath61 , a set of given sinr values @xmath62 is achievable in the downlink using the total power consumption @xmath63 if and only if the same set of sinr values is achievable in the dual uplink channel with the weighted total power of @xmath64 . _ the udd algorithm is based on the uplink - downlink duality theory . we summarize the udd algorithm requires uplink / downlink power update only once at each iteration and thus is more efficient . however , the convergence result to be shown later in proposition 2 also applies to the udd algorithm in @xcite . ] @xcite in table [ tab : pseudo_code_mmse_socp ] . it should be pointed out that , the udd algorithm requires a feasible initialization @xcite . otherwise , the steps 7 and 11 in the algorithm are not well - defined . + it is also worth noting that the work @xcite ( resp . @xcite ) only shows that the mmse - socp ( resp . udd ) algorithm keeps the qcpm objective function nonincreasing as the iteration proceeds . in this paper , we reveal the connection between the mmse - socp and the udd algorithm . moreover prove that the two algorithms can monotonically converge to a kkt solution of the qcpm problem . * * input * : @xmath65 , @xmath66 , @xmath67 , @xmath68 * * output * : the beamformers @xmath69 and @xmath70 . * set @xmath71 and @xmath72 * initialize @xmath73 , @xmath68 . * * repeat * * @xmath74 ; * @xmath75 , @xmath50 * @xmath76 , @xmath50 * update @xmath47 s by solving //uplink power allocation * @xmath77 , @xmath50 * @xmath78 , @xmath50 * @xmath79 , @xmath50 * update @xmath46 s by solving //downlink power allocation * @xmath80 , @xmath50 * * until * some convergence criterion is met + in this section , we first analyze the kkt conditions of problem ( p1 ) . based on the results of the kkt analysis , we then present a new iterative power minimization algorithm , dubbed mmse - dual , to solve the system of kkt equations . moreover , the proposed algorithm reveals some connections between the mmse - socp and the udd algorithm . first , let us define the lagrange function associated with problem ( p1 ) as @xmath81 where @xmath82^t$ ] is the lagrange multiplier vector . for a given optimal primal dual tuple @xmath83 , the kkt conditions of problem ( p1 ) are given by [ eqkkt ] @xmath84 where @xmath85 and @xmath86 are the first - order optimality conditions with respect to the receive and transmit beamformers , respectively . the equation @xmath87 is the complementary condition ; and the equations @xmath88 and @xmath89 are the primal and dual feasibility conditions . in the sequel , we analyze the above kkt system . [ lemma : lambdapositive ] _ for any primal - dual tuple @xmath83 that satisfies , we have @xmath90 , and all the qos constraints hold with equality . _ we prove this using contradiction . assume the contrary that one of the optimal lagrange multipliers , say @xmath91 , equals to zero . multiplying by @xmath92 yields @xmath93 this implies @xmath94 , which contradicts the assumption that @xmath95 . let us define @xmath96 and @xmath97 [ lemma : akuk ] _ for any primal - dual tuple @xmath83 that satisfies , the minimum eigenvalue of @xmath98 is zero . furthermore , the optimal normalized receive beamformers are given by @xmath99 _ clearly , implies that @xmath98 must have at least one zero eigenvalue . on the other hand , since @xmath100 @xmath98 has at most one nonpositive eigenvalue . hence , the minimum eigenvalue of @xmath98 is zero . furthermore , the equation implies that @xmath101 and therefore @xmath102 combining and lemma [ lemma : lambdapositive ] , it can be verified that @xmath103 is the unique solution of up to scaling . defining @xmath104 and @xmath105 the next lemma follows directly from the kkt equations and . the proof of this lemma is similar to the proof of lemma [ lemma : akuk ] and thus omitted from the manuscript . [ lemma : bkvk]_for any primal - dual tuple @xmath83 that satisfies , the minimum eigenvalue of @xmath106 is zero . moreover , the optimal transmit beamformers are given by @xmath107 where the coefficients @xmath108 is chosen such that is satisfied . _ _ for any primal - dual tuple @xmath83 that satisfies , the lagrange multipliers @xmath109 satisfies the system of equations @xmath110 where @xmath111 . moreover , the unique solution of can be obtained by a fixed point iteration . _ from lemma [ lemma : bkvk ] , we have @xmath112 it follows from schur complement that @xmath113\succeq 0.\ ] ] since @xmath114 is positive definite , using schur complement again implies @xmath115 for any primal - dual tuple @xmath83 that satisfies the kkt condition , the above inequality must hold with equality ; otherwise @xmath116 by the schur complement , which contradicts lemma [ lemma : bkvk ] . hence , it holds that @xmath117 on the other hand , @xmath118 where the second equality is due to the identity @xmath119 ( * ? ? ? hence can be rewritten as @xmath120 which implies . moreover , the right hand side of is a standard function is a standard function if it satisfies 1 ) @xmath121 ; 2 ) @xmath122 for @xmath123 ; 3 ) @xmath124 for @xmath125 . if @xmath126 is a standard function , the system of equations @xmath127 has a unique solution which can be obtained by a fixed point algorithm . see more details in @xcite . ] of @xmath128 . hence , the solution of is unique and can be obtained by a fixed point algorithm @xcite . based on the results of the kkt analysis , we here present a new iterative power minimization algorithm , dubbed mmse - dual algorithm , to solve the system of kkt equations ( [ neq : kkt1]-[neq : kkt5 ] ) . our proposed mmse - dual algorithm first updates the receive beamformers using the mmse receiver , followed by the update of the transmit beamformers using equations and . the algorithm is outlined in table [ tbl : pseudo_code_mimo ] . in this table , @xmath129 denotes the total number of fixed point iterations for calculating the optimal lagrange multipliers . before stating the properties of the proposed algorithm , let us first see how the mmse - dual algorithm plays a key role in establishing the connection between the mmse - socp algorithm and the udd algorithm . first notice that in the mmse - dual algorithm , the procedure of updating the transmit beamformers ( i.e. , steps 7 - 13 in table ii ) is equivalent to solving @xmath130 this follows from the fact that the kkt condition of ( p2 ) are identical to ( [ neq : kkt2]-[neq : kkt5 ] ) and strong duality holds for problem ( p2)@xcite . since the updates of the receive beamformer in mmse - dual is the same as the receiver update in the mmse - socp algorithm , the mmse - dual algorithm is in essence the mmse - socp algorithm . the only difference of the mmse - dual algorithm with the mmse socp @xcite is that , instead of updating the transmit beamformers by directly solving the socp ( p2 ) , we use semi - closed form computation and . it is not hard to see that the complexity of each iteration of the fixed point algorithm is dominated by the computation of @xmath131 , which is @xmath132 . hence , the mmse - dual algorithm has lower complexity than the mmse - socp algorithm . next we explore the relation between the mmse - dual and the udd algorithm . comparing the algorithms in table i and ii , it is not difficult to see that , the dual variables @xmath128 and the auxiliary variables @xmath108 in the mmse - dual algorithm respectively play the role of the dual uplink transmit power and the downlink transmit power in the udd algorithm . this implies that , although the udd algorithm is developed from a different point of view , it works in a similar way as the mmse - dual algorithm towards solving the kkt system . furthermore , the constraints of problem must be satisfied with equality at the optimality . hence it is readily seen that the mmse - dual algorithm updates the auxiliary variables @xmath108 in the same way as the downlink transmission power @xmath133 in the udd algorithm ; see step 11 in table i and step 13 in table ii . the only difference between the two algorithms lies in the update of the dual variables @xmath128 ( or equivalently the dual uplink transmit power @xmath47 in udd ) . in the mmse - dual algorithm , the lagrange multipliers @xmath128 are updated by the fixed point algorithm independent of the current transmit beamformers , while the update of uplink transmit power in the udd algorithm depends on the current transmission beamformers ; see steps 7 - 10 in table ii and step 7 in table i. note that , if problem is feasible , the update of the uplink transmit power in the udd algorithm is equivalent to solving a linear system of @xmath59 which has complexity of @xmath134 . hence , when @xmath135 , the udd algorithm and the mmse - dual algorithm have comparable complexity . but in the general single stream case where @xmath136 , the udd algorithm has lower complexity than the mmse - dual algorithm . * * input * : @xmath65 , @xmath66 , @xmath67 , @xmath68 * * output * : the beamformers @xmath69 and @xmath70 . * set @xmath71 and @xmath72 * initialize @xmath73 and @xmath91 , @xmath68 . * * repeat * * @xmath74 ; * @xmath137 , @xmath50 * @xmath138 , @xmath50 * * for * @xmath139 to @xmath129//fixed point algorithm * @xmath140 * @xmath141 , @xmath50 * * end for * * @xmath142 , @xmath50 * @xmath79 , @xmath50 * solve for the linear system of @xmath108 : * @xmath143 , @xmath50 * @xmath144 , @xmath50 * * until * some convergence criterion is met + in a nutshell , similar to the mmse - dual algorithm , the mmse - socp algorithm and the udd algorithm both work towards solving the kkt system . in the next section , we rigorously show that the three algorithms indeed monotonically converge to a kkt solution . in this section , we establish the convergence of the three algorithms . compared to the existing convergence results of the mmse - socp algorithm in @xcite and the udd algorithm in @xcite , we show below a stronger convergence result that every limit point of the three algorithms is a kkt solution of . before stating the convergence results , we first present three lemmas which will be used later in the convergence proof . [ lem : phase - rotation ] for any feasible @xmath70 , the optimal solution to problem is unique up to phase rotation . let @xmath145 and @xmath146 be any two optimal solutions to problem . moreover , we define @xmath147 and @xmath148 , @xmath149 . in the following , we prove @xmath150 , @xmath149 . it is noted that any feasible solution to the following socp @xmath151 is feasible to problem . it follows that the optimal value of problem is not greater than that of problem . on the other hand , it is easily seen that the optimal solution @xmath145 to problem is a feasible solution to problem . hence , @xmath145 is an optimal solution to problem . similarly , we infer that @xmath146 is an optimal solution to the following socp @xmath152 by comparing the above two socp formulations and noting @xmath153 , we conclude that @xmath154 is an optimal solution to problem . since the socp problem has a strictly convex objective function , it has a unique solution . thus we have @xmath150 , @xmath149 , implying that the optimal solution to problem is unique up to phase rotation . this completes the proof . [ lem : lemma4-prop1 ] if @xmath70 and @xmath69 satisfy the sinr constraints of problem ( p1 ) with equality and for each @xmath17 @xmath155 is an mmse - receiver , i.e. , @xmath156 for all @xmath17 with @xmath157 being an arbitrary nonzero constant , then we have @xmath158 since @xmath70 and @xmath69 satisfy the constraints of problem ( p1 ) with equality , we have @xmath159 which can be rearranged as @xmath160 by substituting the mmse receiver @xmath156 into and noting that @xmath161 , @xmath50 , we get @xmath162 since we have @xmath163 ( due to @xmath164 ) , it follows that @xmath165 this implies that for each @xmath17 the matrix @xmath166 is positive semidefinite . hence , from , we obtain @xmath167 which completes the proof . due to the uplink downlink duality , we also have the following lemma . the proof of the lemma is similar to that of lemma [ lem : lemma4-prop1 ] and thus omitted for brevity . [ lem : lemma5-prop2 ] if @xmath70 and @xmath69 satisfy the sinr constraints of the virtual uplink weighted power minimization problem s . ] with equality and moreover for each @xmath17 , @xmath73 is a ( virtual uplink ) mmse - receiver , i.e. , @xmath168 with @xmath169 being an arbitrary nonzero constant , then we have @xmath170 where @xmath114 s are defined as in section iii . [ rem : remak - seciv ] lemmas [ lem : lemma4-prop1 ] and [ lem : lemma5-prop2 ] indicate that , for the downlink / virtual uplink power minimization problem , if a set of transmit and receive beamformers satisfies the sinr constraints with equality and moreover complies with the mmse receiver structure , the first order optimality condition with respect to the downlink / virtual uplink receive beamformers follows . in other words , if a set of transmit and receive beamformers fulfills both the assumptions of lemma [ lem : lemma4-prop1 ] and lemma [ lem : lemma5-prop2 ] , then it is a kkt solution to the kkt system . in this subsection , we study the convergence behavior of the mmse - socp ( or equivalently mmse - dual ) algorithm . [ prop1-seciv ] _ let @xmath171 denote a sequence generated by the mmse - dual algorithm ( or equivalently the mmse - socp algorithm ) . suppose @xmath172 is feasible for problem ( p1 ) , then every limit point of @xmath171 is a kkt point of ( p1 ) . _ here we prove the result for the mmse - socp algorithm . the proof for the mmse - dual algorithm follows immediately due to its equivalence to mmse - socp algorithm . the iterations of the mmse - socp algorithm are illustrated as @xmath173 , where the two arrows correspond to the two update rules shown in sec . ii.b in order . since the objective function is coercive , @xmath174 is bounded and consequently @xmath175 is bounded as well . hence , the sequence @xmath176 has at least one limit point . consider a subsequence @xmath177 converging to the limit point @xmath178 . moreover , by further restricting to another subsequence , we can assume that @xmath179 converges to a limit point @xmath180 . in the sequel , we first prove that @xmath181 ( up to a phase rotation ) . clearly , @xmath182 since the objective value is decreasing and it is bounded from below . now consider a fixed transmit beamformer @xmath183 so that @xmath184 where @xmath185 is defined in . due to the continuity of the sinr function , there exists an index @xmath186 so that for all @xmath187 , @xmath188 . since at each iteration of the algorithm , the transmit beamformers are updated after the receive beamformers , for all @xmath187 we have @xmath189 letting @xmath190 implies @xmath191 furthermore , according to the update rule , we have @xmath192 and thus @xmath193 combining with the fact that holds for any @xmath183 satisfying , we obtain such that @xmath194 , we can scale up @xmath183 with any constant @xmath195 so that the scaled @xmath183 ( i.e. , @xmath196 ) satisfies . furthermore , analogous to , we have @xmath197 , @xmath198 , implying @xmath199 . therefore , combining with the fact that holds for any @xmath183 satisfying implies . similar arguments are also used in the proof of proposition 2 . ] @xmath200 on the other hand , since the update of receive beaformers using mmse receivers keeps the sinr feasibility , we have @xmath201 . letting @xmath202 , we obtain @xmath203 combining , and , we infer that @xmath180 is also an optimal solution to problem . according to lemma [ lem : phase - rotation ] , we have @xmath181 up to an appropriate phase rotation . next , with @xmath204 , we prove that the limit point @xmath205 is a kkt point of ( p1 ) . based on the receive beamformer update rule in the algorithm , @xmath206 , where @xmath207 . letting @xmath190 implies @xmath208 with @xmath209 . on the other hand , implies that there exists a set of multipliers @xmath210 so that @xmath211 by lemma [ lem : lemma4-prop1 ] , we infer from and that @xmath212 clearly , the equations , and imply that the limit point @xmath205 is a kkt point of ( p1 ) . the following theorem states our convergence result of the udd algorithm : _ let @xmath213 denote a sequence generated by the udd algorithm . suppose @xmath214 is feasible for problem ( p1 ) , then every limit point of @xmath215 is a kkt point of ( p1 ) . _ for the ease of understanding , the iterations of the udd algorithm are visually presented in fig . [ fig : udd - flow - chart ] where each arrow indicates an update rule as labeled . let @xmath216 be the sequence generated by the udd algorithm . clearly , the sequences @xmath217 and @xmath218 are bounded . on the other hand , since the objective functions of problems and are both coercive , @xmath219 and @xmath220 are also bounded . hence , the sequence @xmath221 is bounded . it follows that there exists a subsequence @xmath222 converging to a limit point @xmath223 . furthermore , by further restricting to a subsequence , we can assume that the subsequence @xmath224 converges to some limit point @xmath225 . according to @xcite , we have the monotonic convergence , i.e. , @xmath226 , implying that @xmath227 . first , we prove that @xmath228 is the optimal solution of problem with @xmath44 s and @xmath45 s being fixed to @xmath229 s and @xmath230 s respectively , and consequently @xmath231 , @xmath232 . let @xmath233 . due to the continuity of the sinr functions , we can always find an integer @xmath234 , such that for all @xmath235 , @xmath236 step 11 of the udd algorithm [ see ( a ) in fig . [ fig : udd - flow - chart ] ] implies @xmath237 and @xmath238 taking limit as @xmath239 yields @xmath240 and @xmath241 since at the optimality of problem the sinr constraint must hold with equality , and implies @xmath242 similarly , we next show that @xmath243 is the optimal solution of problem with @xmath44 s and @xmath45 s being fixed to @xmath244 s and @xmath230 s respectively , and moreover @xmath245 , @xmath232 . let @xmath246 due to the continuity of the sinr functions , we can always find @xmath234 , for all @xmath235 , such that @xmath247 due to step 7 of the udd algorithm [ see ( c ) in fig . [ fig : udd - flow - chart ] ] , we have @xmath248 and @xmath249 hence , we infer that @xmath250 and @xmath251 now we show @xmath252 after doing an appropriate phase rotation . due to the feasible initialization , we have for any @xmath253 that [ see ( b ) in fig . [ fig : udd - flow - chart ] ] @xmath254 taking limit as @xmath239 yields @xmath255 where the equality is due to . in the following , we show that the inequality in actually achieves equality . define @xmath256 and assume for contrary that there exists one @xmath257 that is strictly greater than @xmath67 . by the uplink - downlink duality theory , there exists @xmath258 such that @xmath259 for all @xmath17 and @xmath260 . since there exists one @xmath17 for which @xmath261 , the total power @xmath262 can be further decreased by reducing @xmath263 . hence , and imply @xmath264 . this yields a contradiction due to the fact that @xmath265 . hence , we have @xmath266 note that steps 5 - 6 of the udd algorithm [ see ( b ) in fig . [ fig : udd - flow - chart ] ] imply @xmath267 where @xmath268 is normalization factor . from the above equation , we infer that @xmath269 maximizes @xmath270 with respect to @xmath44 . it follows from that @xmath271 also maximizes @xmath270 . thus we have @xmath272 for some complex valued scalar @xmath273 with @xmath274 . next , we show @xmath275 and further @xmath276 , @xmath232 . step 9 of the udd algorithm [ see ( d ) in fig . [ fig : udd - flow - chart ] ] implies @xmath277 taking limit as @xmath239 , we have @xmath278 and thus @xmath279 by noting @xmath280 . combining , , and @xmath281 , we infer that both @xmath282 s and @xmath283 s are the optimal solutions to problem with there @xmath44 s and @xmath45 s replaced by @xmath269 s and @xmath284 s respectively . since problem has a unique solution@xcite , we conclude @xmath285 , @xmath232 , and @xmath286 now we are ready to end up the proof . by noting @xmath287 and @xmath288 , step 5 - 8 of the udd algorithm implies @xmath289 where @xmath290 is a normalization factor . on the other hand , step 9 - 12 of the udd algorithm implies @xmath291 where @xmath292 is a normalization factor . , , , and can be equivalently written as @xmath293 where @xmath294 , @xmath295 . combining lemma [ lem : lemma4-prop1 ] , , and , we obtain @xmath296 similarly , lemma [ lem : lemma5-prop2 ] , , and imply @xmath297 it can be readily seen that , , , and implies the kkt condition with @xmath298 for all @xmath17 . thus the proof is completed . [ rem : feasibility ] although both algorithms require feasible initialization , it is easier for the mmse - dual algorithm than the udd algorithm to obtain a feasible initialization . for example , when @xmath299 , it is guaranteed that problem ( p1 ) with any given nonzero @xmath155 s is feasible ( e.g. , zero - forcing solution for @xmath73 s ) and thus the mmse - dual algorithm can be randomly initialized in this case . however , random initialization for the udd algorithm in this case may fail . this is also verified with a specific example in section vi . now we consider the extension of the two algorithms to the multiple stream case . differently from @xcite , we assume that joint detection is employed at receivers . let @xmath300 be the transmit beamformer for user @xmath17 . in this case , each user s achieved rate is given by @xmath301 where @xmath302 is the interference plus noise covariance matrix . we are interested in solving the following rate constrained power minimization problem @xmath303 where @xmath304 represents the rate requirement for user @xmath17 . it is known that the streams of user @xmath17 can be decoded sequentially without loss of information using mmse receiver coupled with sequential interference cancelation ( sic ) technique@xcite . indeed , it is easily verified that @xmath305 where @xmath306 @xmath307 is the @xmath308-th column of @xmath309 , i.e. , the transmit beamformer for stream @xmath308 , and @xmath310 is referred to as mmse - sic receiver . define @xmath311 , @xmath312 . with equal rate allocation @xmath313 across multiple streams , liu . al ( * ? ? ? * theorem 4 ) proved that the following sinr - constrained power minimization problem @xmath314 can achieve the same optimal solution as that of problem . however , we still can not solve problem to global optimality . fortunately , the special structure of problem allows us to apply the udd algorithm or mmse - dual algorithm . since the udd algorithm or mmse - dual algorithm reaches a kkt point of problem , a question arises : whether a kkt point of problem is also a kkt point of problem . in the following proposition , we show the kkt equivalence of the two problems under a mild set of conditions . the proof is relegated to appendix . [ prop : prop3 ] let @xmath315 be a kkt point of problem , and let @xmath316 suppose @xmath317 satisfies @xmath318 , @xmath319 , @xmath320 , then 1 . @xmath317 is a kkt point of problem . 2 . a kkt point of can be obtained by solving problem using either the mmse - dual or the udd algorithm . it can be shown that the condition @xmath318 , @xmath319 , @xmath320 is equivalent to @xmath321 , @xmath319 , @xmath320 ; see in appendix . this means that , for each user @xmath17 , detection of all the second , third , ... , and the @xmath322-th symbols are interfered by the first symbol , which is generally true in the multi - stream scenario . in the proof of proposition [ prop : prop3 ] , the assumption @xmath318 , @xmath319 , @xmath320 is used to obtain @xmath323 , @xmath320 , which finally leads to part 1 ) . it is worthy mentioning that , if this assumption is relaxed , there indeed exists some example ( shown in the end of this paper ) where @xmath323 , @xmath320 may not hold and in this case the kkt point @xmath317 of problem is not a kkt point of problem . let @xmath324 . similar to the single stream case , it can be shown that , for the multi - stream case , the complexity of each fixed point iteration in the mmse - dual algorithm is as in the single stream case . ] @xmath325 , while solving the system of equations for the @xmath326 dual variables in the udd algorithm requires the complexity of @xmath327 . hence , in the general multi - stream case where @xmath328 , the mmse - dual algorithm has a lower complexity than the udd algorithm . in this section , we numerically evaluate the performance of the mmse - dual algorithm and the udd algorithm . note that , we here only provide the convergence performance of the two algorithms in the single stream case but similar convergence behavior is also observed in the multiple stream case . the qos level and noise variance are set equally to be @xmath329 and @xmath12 across all the users . uncorrelated fading channel model for all channel matrices between users and the bs are assumed . each channel coefficient is generated from the zero mean complex gaussian distribution with unit variance . we also set @xmath330 for the number of fixed point iterations . in our first numerical experiment , we study the effect of the infeasible initialization on the algorithms . in the experiment , the bs serves two users , all equipped with two antennas , i.e. , @xmath331 . we set @xmath332 , @xmath333 , and the channel matrices between the bs and the two users are respectively @xmath334,\ ] ] @xmath335.\ ] ] with initial transmit beamformers @xmath336^t,\ ] ] @xmath337^t,\ ] ] the corresponding normalized mmse receivers are calculated as @xmath338^t,\ ] ] @xmath339^t,\ ] ] leading to sinr values , @xmath340 and @xmath341 , which are both smaller than the required sinr value @xmath332 . furthermore , it can be easily checked that the linear system on @xmath342 and @xmath343 with the fixed initial beamformers @xmath344 has negative solutions @xmath345 and @xmath346 . therefore , the uplink power update can not be done in the udd algorithm . while the udd algorithm fails with this initialization , the mmse - dual algorithm can quickly reach a feasible point in the second iteration and then exhibits a monotonic convergence in subsequent iterations , as shown in fig . [ fig : fig2 ] . in this set of numerical experiments , we randomly initialize the mmse - dual algorithm . the udd algorithm is initialized by a feasible point obtained by few iterations of the mmse - dual algorithm . figure [ fig : fig3 ] shows that the mmse - dual algorithm and the udd algorithm have a very similar convergence behavior . , @xmath347 , @xmath348 , @xmath332 , @xmath333.,width=259 ] figures [ fig : fig4 ] and [ fig : fig5 ] show that the two algorithms can almost always converge to a same objective function value which may be global optimum regardless of initial points ( different initial points are denoted by circles ) . however , in an extremely rare case , local convergence for the mmse - dual algorithm ( also for the udd algorithm ) was observed in fig . [ fig : fig6 ] where two different initial points resulted in two different objective values upon convergence . , @xmath349 , @xmath348 , @xmath332 , @xmath333 . ] , @xmath349 , @xmath348 , @xmath332 , @xmath333 . ] , @xmath350 , @xmath351 , @xmath352 , @xmath353 . ] in this paper , we have considered the sinr constrained power minimization problem for mu - mimo system . based on the kkt analysis of the power minimization problem , we propose the mmse - dual algorithm . although the latter algorithm is in essence the mmse - socp algorithm in @xcite , it connects the mmse - socp algorithm and the udd algorithm . it is shown that the udd algorithm@xcite also works towards the kkt condition of the qcpm problem as the mmse - dual / the mmse - socp algorithm . furthermore , we theoretically prove that all three algorithms can monotonically converge to a kkt point of the qcpm problem . our numerical experiments show that the three algorithms almost always converge to a same value which may be the optimal value , but local convergence of these algorithms is also observed . the proof of the second part of proposition 3 is trivial once the first part is proven . hence , we here focus on proving the first part of proposition 3 . let @xmath354 be the lagrange multiplier associated with the constraint of problem indexed by @xmath355 . the kkt condition of problem is given by @xmath356 in the following , we are going to show that given any tuple @xmath357 satisfying the above system , @xmath358 must also satisfy the kkt condition for problem . the proof is divided into two steps . in the first step , we prove that can be equivalently rewritten as the first order optimality condition of problem by using and , and showing @xmath323 , @xmath50 . from , we infer that @xmath359 must be in the form of mmse receiver @xmath360 where @xmath361 is an arbitrary nonzero scalar . by taking @xmath362 and using , we have @xmath363 moreover , from and , we have @xmath364 it follows that @xmath365 where in the second equality we have used woodbury identity@xcite . similarly , we have @xmath366 where we have used the relation @xmath367 in the first and third equality , and the woodbury identity in the last equality . by using , we have @xmath368 using , we get @xmath369 and using , we obtain @xmath370 by defining @xmath371 and substituting , , and into , we have @xmath372 by noting that @xmath373 we rewrite as @xmath374 considering with @xmath375 and @xmath376 , we have @xmath377 left - multiplying @xmath378 on both sides of and using yeild @xmath379 implying @xmath380 due to @xmath381 by assumption . recursively , we infer from that @xmath382 , by using the condition @xmath383 , @xmath384 . hence , by letting @xmath385 for all @xmath17 , and using @xmath386 $ ] , we can compactly write as @xmath387 note that we have @xmath388 and @xmath389 substituting and into yields @xmath390 which is the first order optimality condition of problem . in the second step , we prove by using that @xmath317 satisfies the constraints of problem with equality . since @xmath391 , @xmath312 , @xmath50 , we have @xmath392 where we use the identity @xmath393@xcite in the first equality and in the second equality . by summing over @xmath312 , we obtain @xmath394 combining , , and together with @xmath395 @xmath50 , we concludes that @xmath317 and @xmath128 satisfy the kkt condition of problem . 1 m. hong and z .- q . luo , _ signal processing and optimal resource allocation for the interference channel _ , academic press library in signal processing , elsevier , 2013 . a. wiesel , y. c. eldar , and s. shamai(shitz ) , `` linear precoding via conic optimization for fixed mimo receivers , '' _ ieee transactions on signal processing _ 1 , pp . 161176 , 2006 . m. razaviyayn , m. hong , and z .- q . luo , `` linear transceiver design for a mimo interfering broadcast channel achieving max - min fairness , '' _ signal processing _ , vol . 3327 - 3340 , dec . 2013 . q. h. spencer , a. l. swindlehurst and m. haardt , `` zero - forcing methods for downlink spatial multiplexing in multiuser mimo channels , '' _ ieee trans . signal process . 2 , feb . 2004 . z. g. pan , k. k. wong , and t. s. ng , `` generalized multiuser orthogonal space division multiplexing , '' _ ieee trans . wireless commun . 3 , no . 6 , pp . 1 - 5 , nov . a. m. khachan , a. j. tenebaum , r. s. adve , `` linear processing for the downlink in multiuser mimo systems with multiple data streams , '' presented at ieee conference on communications , vol . 9 , pp . 4113 - 4118 , june 2006 . h. dahrouj and w. yu , `` coordinated beamforming for the multicell multi - antenna wireless system , '' _ ieee trans . wireless commun . _ , vol . 9 , no . 5 , may 2010 . chang , l. tassiulas , f. farrokhi , `` joint transmitter receiver diversity for efficient space division multiaccess , '' _ ieee trans . wireless commun . _ , vol . 1 , no . 1 , pp . 16 - 27 , jan . 2002 . m. codreanu , a. tolli , m. juntti , m. latva - aho , `` joint design of tx - rx beamformers in mimo downlink channel , '' _ ieee trans . signal process . 9 , pp . 4639 - 4655 , sept . 2007 . b. song , r. cruz , and b. d. rao . `` network duality for multiuser mimo beamforming networks and applications , '' _ ieee trans . _ , vol 50 , no . 618 - 630 , mar . e. visotsky and u. madhow , `` optimum beamforming using transmit antenna arrays , '' presented at ieee vehicular technology conference , vol . 851 - 856 , may 1999 . wong , g. zheng , and t .- s . ng , `` convergence analysis of downlink mimo antenna system using second - order cone programming , '' presented at the the 62nd ieee vehicular technology conference , pp . 492 - 496 , sept . 2005 . h. boche and m. schubert , `` a general duality theory for uplink and downlink beamforming , '' presented at ieee vehicular technology conference ( vtc ) , vancouver , canada , vol . 87 - 91 , sept . 2002 . p. viswanath and d. n. c. tse , `` sum capacity of the vector gaussian broadcast channel and uplink - downlink duality , '' _ ieee trans . info . theory _ , 49 , no . 8 , pp . 1912 - 1921 , aug . 2003 . r. hunger and m. joham , `` a general rate duality of the mimo multiple access channel and the mimo broadcast channel , '' presented at ieee global telecommunications conference , pp . 1 - 5 , nov . m. bengtasson and b. ottersten , `` optimal and suboptimal transmit beamforming , '' chapter 18 in _ handbook of antennas in wireless communications _ , l. c. godara , ed . , crc press , aug . r. d. yates , `` a framework for uplink power control in cellular radio systems , '' _ ieee j. sel . areas commun . 13 , no . 7 , pp . 13411347 , sept . m. varanasi and t. guess , `` optimum decision feedback multiuser equalization with successive decoding achieves the total capacity of the gaussian multiple - access channel , '' presented at asilomar conference on signals , systems , and computers , vol . 2 , pp . 1405 - 1409 , 1997 . d. bertsekas , _ nonlinear programming _ , 2nd ed . belmont , ma : athena scientific , 1999 . a. liu , y. liu , h. xiang and w. luo . `` mimo b - mac interference network optimization under rate constraints by polite water - filling and duality , '' _ ieee trans . signal process . k. b. petersern and m. s. pedersern , _ the matrix cookbook_. http://matrixcookbook.com , nov . s. boyd and l. vandenberghe , _ convex optimization_. cambridge u.k . : cambridge univ . press , 2004 . in the proof of proposition [ prop : prop3 ] , the assumption @xmath318 , @xmath319 , @xmath320 is used to obtain @xmath323 , @xmath320 , which finally leads to part 1 ) of proposition [ prop : prop3 ] . intuitively , it may hold that @xmath323 for the case when @xmath396 , even if the assumption does not hold . however , it is worthy mentioning that , if the assumption is relaxed , there indeed exists some counter example as shown below , where , @xmath323 , @xmath320 may not hold when @xmath396 , @xmath320 and as a result the kkt point @xmath317 of problem is not a kkt point of problem . let us consider the special case the point to point mimo system with two streams . in this case , the rate - constrained power minimization problem boils down to @xmath397 with equal rate allocation for the two streams , the optimal solution to problem can be found by solving the following problem @xmath398 where @xmath399 . let @xmath400 and @xmath401 be the lagrange multiplier associated with the first and second constraint of problem , respectively . in the following , we show that there may exist the case @xmath402 when @xmath403 . the kkt system of problem is given by @xmath404 for the special case , we relax the assumption , i.e. , let @xmath405 , which holds if @xmath406 and @xmath407 are two orthogonal eigenvectors of @xmath408 . in the following , we let @xmath406 and @xmath407 be two orthogonal eigenvectors of @xmath408 , such that @xmath409 and @xmath410 . hence , we have @xmath411 and @xmath412 where @xmath413 and @xmath414 are the two corresponding eigenvalues of @xmath408 . furthermore , we let @xmath415 @xmath416 and @xmath417 now we are ready to show that @xmath418 defined above satisfies the kkt system ( [ exkkt1]-[exkkt7 ] ) . let us start with examining and . first , since @xmath405 , we have @xmath419 . thus , @xmath420 which is clearly orthogonal to @xmath421 and @xmath422 due to @xmath405 . it follows that @xmath423 where the first equality is due to @xmath424 and , the second equality is due to + @xmath410 , and the last equality is due to . thus , follows . similarly , we can verify that @xmath421 and @xmath407 satisfy . next let us check and . first , due to and @xmath410 , we have @xmath425 . it follows that @xmath426 where the second equality is due to and the last equality follows from @xmath427 and @xmath411 . thus , follows . similarly we can verify .

markov decision processes ( mdps ) are extensively used to solve sequential stochastic decision making problems in robotics @xcite and other disciplines @xcite . a solution to an mdp problem instance provides a policy mapping states into actions with the property of optimizing ( e.g. , minimizing ) in expectation a given objective function . in many practical situations a formulation based on expectation only is , however , not sufficient . this is the case , in particular , when variability in the system s behavior can cause a highly undesirable outcome . for example , in an autonomous navigation system , a robot attempting to minimize the expected length of the traveled path will likely travel close to obstacles , and a large deviation from the planned path may result in a collision causing a huge loss ( e.g. , damage to an expensive robot or failure of the whole mission altogether ) . metrics that study deviations from the expected value are often referred to as _ risk metrics_. typical solution algorithms for mdp problems , like value iteration or policy iteration , aim exclusively at optimizing the expected cost . the computed policies are therefore labeled as _ risk neutral_. the problem of quantifying risk associated with random variables has a rich history and is often related to the problem of managing financial assets @xcite . in fact , many risk - related studies motivated by financial problems have recently found applications in domains such as robotics @xcite . the term _ risk aversion _ refers to a preference for stochastic realizations with limited deviation from the expected value . in risk averse optimal control one may prefer a policy with higher cost in expectation but lower deviations to one with lower cost but possibly larger deviations . particularly in the context of robotic planning , introducing risk aversion in mdps is crucial to guarantee mission safety . however introducing risk aversion in mdps creates a number of additional theoretical and computational hurdles . for example , in risk averse mdps , optimal policies are not guaranteed to be markov stationary but are instead history dependent . average value at risk ( avar also known as conditional value at risk or cvar ) is a risk metric that has gained notable popularity in the area of risk averse control @xcite , @xcite . for a given random value and a predetermined confidence level , the avar is the _ tail average _ of the distribution exceeding a given confidence level ( see section [ sec : preliminaries ] for a formal definition ) . risk averse policies considering the avar metric have been studied for the case of mdps with finite horizon and discounted infinite horizon cost criteria . in this paper we instead consider how the avar metric can be applied when an undiscounted , total cost criterion is considered . in fact , such cost criterion appears particularly useful and natural for robotic applications , whereby one is usually interested in optimizing the _ undiscounted , total cost _ accrued during a mission until a _ random _ , mission - dependent stopping time . ( as an aside , the total cost criterion is the typical cost model for stochastic shortest path problems , see , e.g. , @xcite . ) the contribution of this paper is three - fold : * we identify conditions for the underlying mdp ensuring that the avar mdp problem is well defined when the total cost criterion is used . * we define a surrogate mdp problem that can be efficiently solved and whose solution approximates the optimal policy for the original problem with arbitrary precision . * we validate our findings on a rapid robotic deployment task where the objective is to maximize the mission success rate under a given temporal deadline @xcite . the rest of the paper is organized as follows . we discuss related work in section [ sec : related ] and provide some background about risk metrics and mdps in section [ sec : preliminaries ] . in section [ sec : formulation ] we formulate the risk - averse , total cost mdp problem we wish to solve . in section [ sec : risk_aversion ] we propose and analyze an approximation strategy for the problem , and in section [ sec : avar ] we provide an algorithmic solution . simulation results for a rapid deployment problem are given in section [ sec : simulations ] , and conclusions and future work are discussed in section [ sec : conclusions ] . for a general introduction to mdps the reader is referred to textbooks such as @xcite or more recent collections such as @xcite . as pointed out in the introduction , risk aversion in mdps has been studied for over four decades , with earlier efforts focusing on exponential utility @xcite , mean - variance @xcite , and percentile risk criteria @xcite . with regard to mean - variance optimization in mdps , it was recently shown that computing an optimal policy under a variance constraint is np - hard @xcite . recently , average value at risk was introduced in @xcite in order to model the tail risk of a random outcome and to address some key limitations of the prevailing value - at - risk metric . efficient methods to compute avar are discussed in @xcite . leveraging the recent strides in avar risk modeling , there have been a number of efforts aimed at embedding the avar risk metric into risk - sensitive mdps . in @xcite the authors address the problem of minimizing the avar of the discounted cost over a finite and an infinite horizon , and propose a dynamic programming approach based on state augmentation . similar techniques can be found in @xcite , where the authors propose a dynamic programming algorithm for finite - horizon , avar - constrained mdps . the algorithm is proven to asymptotically converge to an optimal risk - constrained policy . however , the algorithm involves computing integrals over continuous variables ( algorithm 1 in @xcite ) and , in general , its implementation appears quite challenging . a different approach is taken by @xcite where the authors consider a finite dimensional parameterization of the control policies , and show that an avar mdp can be optimized to a _ local _ optimum using stochastic gradient descent ( policy gradient ) . however this approach imposes additional restrictions to the policy space and in general policy gradient algorithms only converge to a local optimum . haskell and jain recently considered the problem of risk aversion in mdps using a framework based on occupancy measures @xcite ( closely connected to our recent works where constrained mdps are used to solve the multirobot rapid deployment problem @xcite ) . while their findings are only valid for the case where an infinite horizon discounted cost criterion is considered , the solution we propose uses some of the ideas introduced in @xcite . in this section we summarize some known concepts about risk metrics and mdps . the reader is referred to the aforementioned references for more details . consider a probability space @xmath0 , and let @xmath1 be the space of all essentially bounded random variables on @xmath2 . a _ risk function _ ( or risk metric ) @xmath3 is a function that maps an uncertain outcome @xmath4 onto the real line @xmath5 . a risk function that is particularly popular in many financial applications is the _ value at risk_. for @xmath6 the _ value at risk _ of @xmath7 at level @xmath8 is defined as @xmath9 here var@xmath10 represents the percentile value of outcome @xmath11 at confidence level @xmath8 . despite its popularity , var@xmath12 has a number of limitations . in particular , var is not a _ coherent _ risk measure @xcite and thus suffers from being unstable ( high fluctuations under perturbations ) when @xmath11 is not normally distributed . more importantly it does not quantify the losses that might be incurred beyond its value in the @xmath8-tail of the distribution @xcite . an alternative measure that overcomes most shortcomings of var is the _ average value at risk _ , defined as @xmath13 where @xmath6 is the confidence level as before . intuitively , @xmath14 is the expectation of @xmath11 in the conditional distribution of its upper @xmath8-tail . for this reason , it can be interpreted as a metric of how bad is bad . " @xmath14 can be equivalently written as @xcite @xmath15\right\},\ ] ] where @xmath16 . this paper relies extensively on eq . and aims at devising efficient methods to approximate the expectation in eq . when the random variable @xmath11 is the total cost of an mdp . furthermore , it has been recently shown in @xcite that optimizing the cvar of total reward is equivalent to optimizing the worst - case ( robust ) expected total reward of a system whose model uncertainty is subjected to a trajectory budget . this finding corroborates the fact that a cvar risk metric models both the variability of random costs , as well as the robustness to system transition errors . for a finite set @xmath17 , let @xmath18 indicate the set of mass distributions with support on @xmath17 . a finite , discrete - time markov decision process ( mdp ) is a tuple @xmath19 where * @xmath20 , the state space , is a finite set comprising @xmath21 elements . * @xmath22 , the control space , is a collection of @xmath21 finite sets @xmath23 . set @xmath24 , @xmath25 , represents the actions that can be applied when in state @xmath26 . the set of allowable state / action pairs is defined as @xmath27 * @xmath28 is the transition probability from state @xmath29 to state @xmath30 when action @xmath31 is applied . according to our definitions , @xmath32 . * @xmath33 is a non - negative cost function . specifically , @xmath34 is the cost incurred when executing action @xmath35 at state @xmath29 . let @xmath36 and note that the maximum is attained as @xmath37 is a finite set . define the set @xmath38 of admissible histories up to time @xmath39 by @xmath40 , for @xmath41 , and @xmath42 . an element of @xmath43 has the form @xmath44 , and records all states traversed and actions taken up to time @xmath39 . in the most general case a policy is a function @xmath45 , i.e. , it decides which action to take in state @xmath46 considering the entire state - action history . note that according to this definition a policy is in general randomized . let @xmath47 be the set of all policies , i.e. , including history - dependent , randomized policies . it is well known that in the standard mdp setting where an expected cost is minimized there is no loss of optimality in restricting the optimization over deterministic , stationary markovian policies , i.e. , policies of the type @xmath48 . however , in the risk - averse setting one needs to consider the more general class of history - dependent policies @xcite . this is achieved through a state augmentation process described later . following @xcite , we define the countable space @xmath49 , where @xmath50 , is the sample space and @xmath51 is the borel field on @xmath52 . specific trajectories in the mdp are written as @xmath53 , and we denote by @xmath54 and @xmath55 the state and actions at time @xmath39 along trajectory @xmath56 . in general the exact initial state @xmath57 is unknown . rather it is described by an initial mass distribution @xmath58 over @xmath20 , i.e. , @xmath59 . a policy @xmath60 and initial distribution @xmath58 induce a probability distribution over @xmath61 , that we will indicate as @xmath62 . in this paper we focus on transient total cost mdps , defined as follows . consider a partition of @xmath20 into sets @xmath63 and @xmath64 , i.e. , @xmath65 and @xmath66 . transient _ mdp is an mdp where _ each _ policy @xmath60 satisfies the following two properties : * @xmath67 < \infty$ ] for each @xmath68 , i.e. , the state will eventually enter set @xmath64 , and * @xmath69 for each @xmath70 , @xmath71 , @xmath35 , i.e. , once the state enters @xmath64 it can not leave it . a transient , _ total cost _ mdp is a transient mdp where * @xmath72 for each @xmath73 , i.e. , once the state enters @xmath64 no additional cost is incurred , * the cost associated with each trajectory @xmath56 is given by @xmath74 note that the cost @xmath75 is a random variable depending on both the policy @xmath60 and the initial distribution @xmath58 . the name total cost stems from the fact that an ( undiscounted ) cost is incurred throughout the lifetime " of the system ( i.e. , until the state hits the absorbing set ) . transient , total cost mdps ( closely related to stochastic shortest path problems , e.g. , @xcite ) represent an alternative to the more commonly used discounted , infinite - horizon mdps or finite horizon mdps . as outlined in the introduction , for many robotic applications the total cost , i.e. , @xmath75 , is the most appropriate cost function . we justify this statement by noting that most robotic tasks have finite duration but such duration is usually not known in advance . in these circumstances the finite horizon cost is inappropriate because one can not define the length of the finite horizon up front . similarly , the discounted infinite horizon cost is also ill suited because the task does not continue forever and the cost will not be exponentially discounted over time . without loss of generality , we assume that set @xmath64 consists of a single absorbing state @xmath76 equipped with a single action @xmath77 , i.e. , @xmath78 and @xmath79 with @xmath80 . in the following , with a slight abuse of notation , we denote by @xmath81 the set @xmath82 , i.e. , we exclude the absorbing state from the definition of @xmath81 . moreover , we assume that for a transient , total cost mdp @xmath83 , i.e. , the probability of starting at the absorbing state is zero . in fact , whenever @xmath84 the resulting state trajectory will deterministically remain in @xmath76 , and the corresponding cost is zero . our problem formulation relies on the following two technical assumptions necessary to establish an a - priori upper bound on the total cost incurred by any trajectory obtained under any policy , and to define an approximate problem that can be efficiently solved . the first assumption simply requires that all costs in the transient states are positive ( recall that we excluded @xmath76 when re - defining @xmath81 . ) as it will be shown later , this assumption ensures a non - zero discretization step when approximating the cumulative cost accrued by a system throughout the trajectory @xmath56 until it is absorbed in @xmath76 . [ ass : nn ] all costs in @xmath85 except for state @xmath76 are positive and bounded , i.e. , @xmath86 . when considering cost criteria like finite horizon or discounted infinite horizon with a finite state space , an a - priori upper bound on the accrued cost can be immediately established assuming that all costs are finite ( a fact crucially exploited in @xcite ) . however , the situation is more complex when considering the total cost case , because without introducing further hypotheses on the structure of the mdp a malicious adversary could establish an history - dependent policy capable of invalidating any a - priori established bound on the cost , in the general case one could devise a history dependent policy ensuring that every trajectory generated by the policy is not absorbed in @xmath76 in less than @xmath87 steps , thus invalidating the bound . ] . the second assumption then adds a global reachability structure " to the mdp problem . to this end , in the following , it will be useful to consider the markov chain generated by the mdp when an input is selected for each state . for an mdp @xmath85 , select @xmath88 . the selected inputs and the transition probabilities in @xmath85 define a finite markov chain that we indicate as @xmath89 . the state space of @xmath90 is equal to @xmath20 and for two states @xmath91 the transition probability @xmath92 is defined as @xmath93 where @xmath94 is the input selected in the definition of @xmath89 and @xmath95 is the transition probability of the associated mdp . [ ass : irr ] let @xmath89 be the markov chain induced by the @xmath21 inputs @xmath94 . then the absorbing state @xmath76 , under markov chain @xmath89 , is reachable from any state @xmath68 , for all @xmath96 . we recall that a state @xmath97 in a markov chain is said reachable from another state @xmath98 if there exists an integer @xmath99 such that the probability that the chain will be in state @xmath97 after @xmath100 transitions is positive @xcite . note that when assumption [ ass : irr ] holds , under every policy there is a path of non - zero probability connecting every state to @xmath76 . therefore , it is impossible to devise a policy that prevents sure absorption for an arbitrary number of steps . this holds for all policies , including history dependent policies ( see figure [ fig : assumption2 ] ) . building upon the previous material , we can now define the problem we aim to solve in this paper : _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ * risk - averse , total cost mdp * given a transient total cost mdp satisfying assumptions [ ass : nn]-[ass : irr ] , and an initial distribution @xmath58 , determine a policy @xmath60 that minimizes @xmath101 , i.e. , find @xmath102 _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ note that , under the assumption of transient total cost mdp , one can easily verify that @xmath103<\infty$ ] . since , by equation , @xmath104 $ ] , one obtains @xmath105 for all @xmath56 , as well . however , to derive an optimization algorithm for the computation of @xmath106 it is necessary to formulate an a - priori upper bound for the optimal cost in . assumptions [ ass : nn ] and [ ass : irr ] are introduced to ensure that such bound exists and can be computed . in this section we study an approximation strategy for the risk averse total cost mdp in equation . similar to the method presented in @xcite , we aim at solving the problem by using the concept of _ occupation measures_. however , unlike for the cases studied in @xcite , in total cost mdps an explicit upper bound for the accrued cost is not available , which makes the solution strategy in @xcite not applicable . our strategy is to find a surrogate to problem . by imposing an effective horizon , we construct a total cost mdp with time - out and recast this problem into a bilinear programming problem . furthermore we characterize the sub - optimality gap for such surrogate approximation . we start with a technical result characterizing the convergence rate to the absorbing state . consider a selection of inputs @xmath107 and the corresponding markov chain @xmath89 . for each state @xmath68 , let @xmath108 denote the simple ( i.e. , without cycles ) path from @xmath29 to @xmath76 of lowest , strictly positive probability . note that @xmath108 exists due to assumption [ ass : irr ] . let @xmath109 be the probability of the path , i.e. , the product of the probabilities of all the transitions along the path . since there are @xmath21 nodes and by definition the path is simple , @xmath108 includes at most @xmath110 transitions between @xmath21 nodes . let @xmath111 note that the minimum is achieved as the minimization is over a finite set , and that @xmath112 is strictly positive due to assumption [ ass : irr ] . the constant @xmath112 lower bounds the probability that , under _ any _ policy @xmath113 , the absorbing state is reached in no more than @xmath21 steps , from _ any _ state @xmath68 . we are now in a position to characterize the convergence rate to the absorbing state . [ lem : tech1 ] for any policy @xmath113 and initial distribution @xmath58 , @xmath114\leq(1 - \gamma)^k,\,\ , \forall k\in \mathbb{n}.\ ] ] the claim is proven by induction on @xmath100 . base case : we prove that @xmath115 \leq 1- \gamma.\ ] ] indeed , @xmath116 = \sum_{x\in x_t } \pr_{\beta}^{\pi}[x_n\neq x_m|x_0 = x]\pr_{\beta}^{\pi}[x_0 = x]$ ] . because of assumption [ ass : nn ] , for any policy @xmath60 , @xmath117 \geq \gamma$ ] , and the base case follows . for the inductive step , assume that @xmath118<(1 - \gamma)^k$ ] , for some @xmath119 . then , @xmath120 = \pr_{\beta}^{\pi}[x_{(k+1)n } \neq x_m |x_{kn}\neq x_m]\pr_{\beta}^{\pi}[x_{kn}\neq x_m]$ ] . by definition of @xmath112 , @xmath121 \leq ( 1-\gamma)^{k+1},\ ] ] and the claim follows . our solution strategy is to solve a _ surrogate _ problem , whereby after a deterministic number @xmath122 of steps , the state moves to the absorbing state @xmath76 surely . in other words , @xmath123 acts as a timeout " for the mdp problem . the surrogate problem is simpler to solve , and we will show in the following that its solution can approximate the solution of the original problem with arbitrary precision . denote by @xmath124}(\omega)$ ] the total cost for such surrogate problem . additionally , for the original problem , let @xmath125 denote the _ absorbing time _ , i.e. , the time at which the state reaches @xmath76 . if @xmath126 , then the two processes coincide and then @xmath124}(\omega)=c(\omega)$ ] . otherwise , for each trajectory @xmath56 such that @xmath127 , the random process is stopped after @xmath123 steps , and the state goes , deterministically , to @xmath76 at stage @xmath128 . in such a case one has @xmath124}(\omega)\leq c(\omega)$ ] . we want to characterize the relation between @xmath129}(\omega))$ ] ( i.e. , the risk for the surrogate problem ) and @xmath101 ( i.e. , the risk for the original problem ) . to this end , let @xmath130 be the total cost for the original problem up to time @xmath123 , i.e. , @xmath131 the following lemma shows the equivalence between @xmath124}(\omega)$ ] and @xmath130 . [ lem : tech2 ] for any policy @xmath132 and any trajectory @xmath56 , @xmath124}(\omega ) = c_{d}(\omega)$ ] . given a policy @xmath60 , for any trajectory @xmath56 , the cost cumulated up to time @xmath123 is the same for both the original and the surrogate problem . after time @xmath123 , both @xmath133 and @xmath124}(\omega)$ ] do not cumulate any additional cost , then the claim follows . the following theorem represents the main result of this section [ them : sub ] the surrogate problem approximates the original problem according to @xmath134}(\omega ) ) \leq \min_{\pi } \textrm{avar}_{\tau}(c(\omega))\\ & \qquad \qquad \leq \min_{\pi } \textrm{avar}_{\tau}(c^{[d]}(\omega ) ) + \frac{n\overline{k}}{1 - \tau}\ , \frac{(1- \gamma)^{\lfloor ( d+1)/n\rfloor}}{\gamma}. \end{split}\ ] ] the left inequality is proven by noticing that @xmath135}(\omega))$ ] , where the equality follows from lemma [ lem : tech2 ] . we now prove the right inequality . for any @xmath136 and policy @xmath60 , one has @xmath137=\mathbb{e}[(c(\omega)-s)^+\mid t^*(\omega)\leq d ] \mathbb p(t^*(\omega)\leq d ) \nonumber\\ & + \mathbb{e}[(c(\omega)-s)^+\mid t^*(\omega ) > d ] \mathbb p(t^*(\omega ) > d ) .\label{eq : decompose } \end{aligned}\ ] ] let @xmath138 be the tail cumulated cost , and , as before , @xmath139 . since the function @xmath140 is sub - additive , i.e. , @xmath141 and the expectation operator preserves monotonicity , one obtains the inequality @xmath142 \\ & = \mathbb{e}[(c_d(\omega)+ c_l(\omega)-s)^+\mid t^*(\omega ) > d ] \\ & \leq\mathbb{e}[(c_d(\omega)-s)^+ \mid t^*(\omega ) > d ] + \mathbb{e}[c_l(\omega ) \mid t^*(\omega ) > d ] . \end{split}\ ] ] furthermore , for each trajectory in the event set @xmath143 , one has @xmath144= \mathbb{e}\left[\left(c_d(\omega)-s\right)^+\mid t^*(\omega)\leq d \right].\ ] ] collecting the results so far , one has the following inequalities : @xmath145\\ & \leq \mathbb{e}[(c_d(\omega)-s)^+\mid t^*(\omega)\leq d ] \mathbb p(t^*(\omega)\leq d ) \\ & \qquad + \mathbb{e}[(c_d(\omega)-s)^+ \mid t^*(\omega ) > d ] \mathbb p(t^*(\omega ) > d ) + \\ & \qquad + \mathbb{e}\left[c_l(\omega)\mid t^*(\omega ) > d \right]\mathbb p(t^*(\omega ) > d)\\ & = \mathbb{e}[(c_d(\omega)-s)^+ ] + \mathbb{e}\left[c_l(\omega)\mid t^*(\omega ) > d \right]\mathbb p(t^*(\omega ) > d)\\ & \leq \mathbb{e}[(c_d(\omega)-s)^+ ] + \mathbb{e}\left[c_l(\omega)\right ] . \end{split}\ ] ] equation implies @xmath146\right\}\\ & \leq \min_{\pi } \ , \min_{s\in \mathbb{r}}\left\{s+\frac{1}{1-\tau } \left(\mathbb{e}[(c_d(\omega)-s)^+ ] + \mathbb{e}\left[c_l(\omega)\right]\right)\right\}\\ & \leq \min_{\pi } \ , \min_{s\in \mathbb{r}}\left\{s+\frac{1}{1-\tau } \left(\mathbb{e}[(c_d(\omega)-s)^+ ] \right)\right\ } \\ & \qquad + \max_{\pi } \ , \frac{1}{1-\tau}\mathbb{e}\left[c_l(\omega)\right]\\ & = \min_{\pi } \ , \min_{s\in \mathbb{r}}\left\{s+\frac{1}{1-\tau } \left(\mathbb{e}[(c^{[d]}(\omega)-s)^+ ] \right)\right\ } \\ & \qquad + \max_{\pi } \ , \frac{1}{1-\tau}\mathbb{e}\left[c_l(\omega)\right]\\ & = \min_{\pi } \textrm { avar}_{\tau}(c^{[d]}(\omega ) ) + \max_{\pi } \ , \frac{1}{1-\tau}\mathbb{e}\left[c_l(\omega)\right ] , \end{split}\ ] ] where the second to last equality follows from lemma [ lem : tech2 ] . we are left with the task of upper bounding @xmath147 $ ] . to this purpose , one can write @xmath148\leq \overline{k } \sum_{t = d+1}^{\infty } \ , \pr_{\beta}^{\pi}(x_t \neq x_m).\ ] ] note that the result in lemma [ lem : tech1 ] implies @xmath149 \leq\!\ ! \!\sum_{k=\lfloor ( d+1)/n\rfloor}^{\infty}\!\!\!\!\sum_{t = kn}^{(k+1)n-1}\pr_{\beta}^{\pi}[x_{t}\neq x_m]\\ & \leq \sum_{k=\lfloor ( d+1)/n\rfloor}^{\infty } n(1-\gamma)^k = n \frac{(1- \gamma)^{\lfloor ( d+1)/n\rfloor}}{\gamma}. \end{split}\ ] ] the claim then follows immediately , as the above upper bound is policy - independent . note that according to theorem [ them : sub ] , as @xmath150 , the optimal cost of the surrogate problem recovers the optimal cost of the original problem , i.e. , the surrogate problem provides a consistent approximation to the original problem , with a sub optimality factor that is computable from problem data . lemma [ lem : tech2 ] and theorem [ them : sub ] ensure that @xmath124}(\omega)$ ] can approximate @xmath75 with arbitrary precision for a sufficiently large value of @xmath123 . in the next section we then show how to solve the minimization problem : @xmath151}(\omega)).\ ] ] leveraging the surrogate problem from the previous section , we can now adapt the results proposed in @xcite to solve problem . an essential step to solve this optimization problem is to compute @xmath152}(\omega)-s)^+]$ ] , which entails deriving the probability distribution for the possible costs generated by the random variable @xmath124}(\omega)$ ] . this problem can be solved by suitably augmenting the state space as described in the following , and then using occupancy measures . in the space of occupancy measures , an optimal policy is determined through the solution of a bilinear program , as explained below . for a given policy @xmath60 and initial distribution @xmath58 , we define the occupancy measure for @xmath153 as @xmath154.\ ] ] note that @xmath155 is non negative but is in general not a probability itself . in the following we will use occupancy measures to determine the probability distribution of the total costs @xmath124}(\omega)$ ] and then to compute the needed expectation . according to the definition , occupancy measures depend on the policy @xmath60 and the initial distribution @xmath58 . given an absorbing mdp @xmath19 , we define a new _ state - augmented _ absorbing mdp with additional state components that track the cumulated total cost and current stage . although the original mdp @xmath85 is finite and absorbing , the set of costs @xmath124}(\omega)$ ] generated by all possible policies can be very large , and this can subsequently lead to a linear program with an unmanageable number of decision variables . to counter this problem , we introduce a discretized approximation for @xmath124}(\pi,\beta)$ ] whose error can be arbitrarily bounded . to this end , we set @xmath156 , where @xmath157 is a parameter describing the desired number of discretized values for the cumulated cost . due to assumption [ ass : nn ] , @xmath158 is strictly positive . the effective number of different values is @xmath159 this value may be higher than @xmath160 due to our definition of @xmath158 . we then define a new mdp @xmath161 as follows . its state space is @xmath162 , where @xmath163 and @xmath164 . elements in the augmented states will be indicated as @xmath165 . as clarified in the following , the two additional components store the cumulated running cost ( @xmath30 ) and current stage ( @xmath166 ) . recall that in the surrogate problem , after @xmath123 steps , the state is guaranteed to have entered the absorbing set , i.e. , it is guaranteed that @xmath167 . thus the value of the @xmath166 component is in @xmath164 . on the other hand the input sets are defined as @xmath168 . @xmath169 and @xmath170 induce a new set @xmath171 . the new cost function @xmath172 is @xmath173 . the transition probability function is modified as follows : @xmath174 as evident from the definition of the new transition function , the new variables included in the state stores the discretized . ] running cost and the stage . consistently with our definition of the surrogate problem , the revised transition function includes a timeout that imposes a transition to the absorbing state @xmath76 after @xmath123 steps , and from that point onwards the accrued cost does not change . note also that the additional state components @xmath30 and @xmath166 are deterministic functions of the previous state and control input @xmath175 . extending the formerly introduced notation , for a given trajectory @xmath56 of @xmath176 , we write @xmath177 for the second component of the state at time @xmath39 and @xmath178 for the third component . finally , for a given initial distribution @xmath58 on @xmath20 , we define the following new initial distribution @xmath179 on @xmath169 , @xmath180 note that the properties of @xmath85 carry over to @xmath176 . in particular , if assumptions [ ass : nn]-[ass : irr ] hold for @xmath85 then they hold for @xmath176 too , and , if @xmath85 is absorbing , then @xmath176 is also absorbing . thus we indicate with @xmath181 its transient set of states . for a given realization @xmath56 , consider now @xmath182}(\omega ) = \sum_{i=0}^tc(x_i , u_i)$ ] , i.e. , the true cumulative cost of the surrogate mdp problem without discretization . the following theorem establishes that even though the approximation error introduced by discretizing the running cost grows linearly with @xmath39 , it is possible to bound it with arbitrary precision . [ th : approx ] for each @xmath183 and each @xmath184 , there exists a discretization step @xmath158 such that @xmath185}(\omega)| < \varepsilon$ ] . _ proof_. pick @xmath186 . let @xmath187}(\omega)-\zeta y_t(\omega)$ ] be the approximation error at time @xmath39 . note that by definition @xmath188 and @xmath189}(\omega)-\zeta y_0(\omega)=0 $ ] . from the definition of the transition probability function , @xmath190 , it follows that @xmath191 , which implies @xmath192 . since for @xmath193 we have @xmath194 , the claim follows . @xmath195 + a key step towards the solution of problem in is therefore to derive the statistical description of the discretized total cost @xmath196 that is used to approximate @xmath124}(\omega)$ ] . this objective can be achieved by exploiting the occupancy measures for the state - augmented mdp @xmath197 . for @xmath198 , the occupancy measure on @xmath197 induced by a policy @xmath60 and an initial distribution @xmath58 is given as : @xmath199.\nonumber \ ] ] the occupancy measure , @xmath200 , is a vector in @xmath201 , i.e. , it is a vector with @xmath202 non negative components . the set of legitimate occupancy vectors is constrained by the initial distribution @xmath58 and defined by the policy @xmath60 . it is well known @xcite that these constraints can be expressed as follows : @xmath203=\beta(x , y , z)~\forall ( x , y , z)\in { x}'^t \nonumber\end{aligned}\ ] ] where @xmath204 if and only if @xmath205 . for @xmath206 we introduce random variables @xmath207 with the property that @xmath208 $ ] . this is easily achieved using occupancy measures , i.e. , @xmath209 where @xmath210 is the indicator function equal to 1 when its argument is true , and 0 otherwise . note that by definition @xmath207 is equal to @xmath211 $ ] , and by theorem [ th : approx ] @xmath196 approximates @xmath124}(\omega)$ ] with arbitrary precision . combining the above definitions we then get to the following problem whose solution approximates the solution to : @xmath212 } s+\frac{1}{1-\tau } \sum_{y \in \mathbb{n}_n}(y - s)^+\theta(y ) \\ & \textrm{s.t . } \nonumber \\ & \sum_{(x',y',z')\in { x}'^t } \sum_{u\in a(x',y',z')}\rho(x',y',z',u)[\delta_{(x , y , z)}(x',y',z')- \nonumber \\ & p'((x',y',z')|(x , y , z),u)]=\beta(x , y , z)~\forall ( x , y , z)\in { x}'^t \nonumber\\ & ~ ~ ~ \theta(k ) = \!\!\!\!\ ! \sum_{(x , y , z , u)\in \mathcal{k}'}\!\!\!\!\!i(y = k~\wedge z = d ) \rho(x , y , z , u),\ , 0 \leq k \leq n.\nonumber\end{aligned}\ ] ] when comparing this last optimization problem with , the reader will note that the variable @xmath213 is constrained in the interval @xmath214 $ ] . indeed , the objective function is continuous with respect to @xmath213 , and it is straightforward to verify that the partial derivative of the objective function with respect to @xmath213 is negative for @xmath215 and positive for @xmath216 . the objective function given in eq . is concave with respect to @xmath217 and is defined over a convex feasibility set @xcite . to the best of our knowledge , there exist no efficient methods to determine the global minimum for this class of problems . hence , the problem is approximately solved fixing different values of @xmath213 within the range @xmath214 $ ] , and then solving the corresponding linear problem over the optimization variables @xmath200 and @xmath218 . comparing the problem in eq . with the one in eq . one might initially think that the objective function in eq . does not depend on the policy @xmath60 . however , the dependency on @xmath60 is carried over by the occupancy measure @xmath200 , as evident from eq . . moreover , it is well known from the theory of constrained mdps @xcite that there is a one to one correspondence between policies and occupancy measures , i.e. , every policy defines a unique occupancy measure and every occupancy measure induces a policy . to illustrate the performance of the proposed algorithm , we adopt the rapid deployment scenario considered in @xcite . a graph is used to abstract and model the connectivity of a given map of an environment ( see , e.g. , @xcite ) . one robot is positioned at a start vertex and is tasked to reach the goal vertex within a given temporal deadline while providing some guarantee about its probability of successfully completing the task . when moving from vertex to vertex , the robot can choose from a set of actions , each trading off speed with probability of success . in particular , actions with rapid transitions between two vertices have higher probability of failure ; and conversely when the robot moves slowly between two vertices it has a higher probability of success . in this scenario , _ failure _ means that the robot does not move ( e.g. , fails to pass through an opening ) , so elapsed time increases without making progress towards the goal . with a given temporal deadline @xmath219 and success probability @xmath220 , the robot is tasked to reach the target vertex safely " ( such that the true mission success probability is at least @xmath220 ) , while satisfying the temporal constraint . from a design perspective it is of interest to know if there exists a policy @xmath60 achieving this objective , and to compute it . if the policy does not exist , it is of interest to know how to modify the parameters in order to make the task feasible . in our previous work we solved this problem by modeling it using constrained markov decision processes ( cmdp ) . in the cmdp approach , one maximizes the probability of success while imposing a constraint on the temporal deadline . however , this method only returns risk - neutral policies , i.e. , the resultant policies only guarantee that the temporal deadline is met in expectation , and there is no explicit control on the tail probability of the constraint . as a radical departure from the original problem formulation , the avar minimization method proposed in eq . searches for a policy that is feasible with respect to the temporal deadline constraint with finite expectation , @xmath221 $ ] for @xmath222 $ ] . therefore , if the solution to the avar minimization problem is bounded above by the temporal deadline , then the corresponding minimizer is also a feasible policy to the original problem . ] and systematically controls the worst - case variability of total travel time . note that a policy with low success probability will have large tail probability in total travel time even if the expected temporal deadline is met . therefore the optimal policies from avar minimization will have _ high success probability_. this motivates the application of avar minimization to rapid robotic deployment . first , note that , in the devised setting , the robot will eventually reach the final goal with positive probability . however , due to possible failures one can not put an a - priori bound on the random total travel time . therefore , the total cost criterion is indeed a natural choice for this task . moreover , assumptions [ ass : nn ] and [ ass : irr ] are easily justified because the immediate cost function ( i.e. , time to move ) is always positive and the global reachability property follows from the graph structure . to illustrate the performance of risk - averse deployment , two different policies are compared . here both policies are computed using unconstrained stochastic control methodologies for which the immediate cost is the travel time between two vertices , and the actions correspond to all possible node transitions on the graph . the first is the classic risk - neutral policy obtained with value iteration . the second is a risk averse policy obtained with the algorithm presented in this paper using @xmath223 . for each policy , 1000 executions are run , and the distribution of total travel time is reported . figures [ fig : riskneutral ] and [ fig : riskaverse ] show the distributions for the two cases . the risk - neutral policy obtains a lower expected cost , but has a longer tail , as evidenced by the 61 instances with a cost larger than or equal to 15 ( notice that @xmath224 is the desired time of completion in this example ) . moreover , as evidenced by the shape of the histogram , costs are more spread out . the risk averse policy , on the other hand , results in less variability as desired . less than 30 instances have a cost larger or equal than 15 , a reduction of the weight of the tail by more than one half . . ] importantly , when computing a risk - neutral policy using classic methods like policy iteration or value iteration , one is merely provided with a policy that minimizes the expected cost ( in our case time to completion ) , and no additional information is readily available . with our approach instead , one not only obtains a policy minimizing the avar criterion , but a statistical description of the costs is also obtained as a byproduct . that is to say , for each discretized completion time @xmath100 , the probability @xmath225}=k]$ ] is computed as well , thus unveiling the relationship between the time to complete the deployment task and its probability . this is shown in figure [ fig : probcomparison ] for different values of @xmath8 . hence , if the computed policy does not meet the desired performance , the designer has information on how to tune @xmath219 and @xmath220 . . in this paper we have considered how risk aversion in mdps can be introduced jointly with the avar risk metric under the total cost criterion . our results advance the state of the art because avar has only been previously considered in mdps with finite horizon or discounted infinite horizon cost criteria . such extension is _ important _ as the total cost criterion appears as a natural model for robotic applications , and is _ non - straightforward _ as current algorithms , e.g. , from @xcite and @xcite , only work with bounded cumulated costs ( which is not the case for total cost formulations ) . under two mild assumptions , an approximation algorithm with provable sub - optimality gap was provided . furthermore , a rapid deployment scenario was used to demonstrate that risk - aversion gives more informative policies when compared to traditional risk - neutral formulations . while our findings focus on risk averse mdps with an avar risk metric , our approach can be easily extended along multiple dimensions . in particular , by exploiting the results presented in @xcite , it is possible to use our approximation for a broader range of risk metrics , i.e. , metrics that are uniformly continuous and law invariant . moreover , since the algorithm we considered is based on occupancy measures , it can be easily extended to the cmdp case . this will be the focus of future work .

the weak equivalence principle states that the trajectories of small test bodies , neither affected by gravitational tidal forces nor by forces other than gravity , are independent of their internal structure and composition @xcite . experimentally , this principle is confirmed with extremely high precision @xcite ; in gravity theory it has been implemented already by newton who postulated that gravitational mass should equal inertial mass , and then by einstein who formulated the motion of test bodies in terms of geodesics on lorentzian spacetime . these trajectories extremize the lorentzian length integral which is interpreted physically as proper time . the fundamental geometric object entering this construction is the lorentzian spacetime metric . this observation led einstein to the development of a gravity theory that determines the metric and so provides a dynamical background geometry for point particles , observers and physical fields . the essential point in einstein s implementation of the weak equivalence principle by a clock postulate is the use of geometric concepts , not the particular choice of metric geometry . indeed , more general geometries can be used which automatically realize this principle . here we consider finsler geometry @xcite which generalizes metric geometry by providing a very general length functional for curves @xmath0 on a manifold @xmath1 , @xmath2=\int d\tau\ , f(\gamma(\tau ) , \dot \gamma(\tau))\,.\ ] ] the finsler function @xmath3 maps the points of the curve and the attached four - velocities into real numbers and is homogeneous of degree one in its second argument to ensure the reparametrization invariance of @xmath4 $ ] . the usual length measure associated to a lorentzian metric @xmath5 is obtained for the special case @xmath6 . physically , we interpret the integral above as a generalized clock postulate and point particle action . on this basis we will develop a consistent gravity theory which determines the finsler function dynamically . our construction builds on the precise definition of physical finsler spacetimes in @xcite , where we established a minimal set of requirements on the function @xmath3 so that it can describe a geometric spacetime background suitable for physics . in particular , finsler spacetimes provide a well - defined notion of causality and the possibility to formulate field theory actions . a generalization of gravity based on finsler geometry has the potential to explain various issues that are not naturally explained by einstein gravity . indeed , it has been argued that finsler geometry in principle can address the rotational curves of galaxies and the acceleration of the universe without introducing dark matter @xcite or dark energy @xcite , and that it admits sufficiently complex causal structures that allow a consistent geometric explanation of superluminal neutrino propagation @xcite . here we will find that finsler gravity may also explain the fly - by anomaly @xcite in the solar system . our presentation is structured as follows . we will begin in section [ sec : fstreview ] with a brief review of our definition of finsler spacetimes and of the mathematical tools and geometric objects needed . moreover , we will define field theory action integrals on finsler spacetimes . equipped with these concepts we will show in section [ sec : obs ] how to model observers on finsler spacetimes and discuss how they perform measurements . we will prove that two different observers are related by a transformation composed out of a certain parallel transport and a lorentz transformation . the set of these transformations has the algebraic structure of a groupoid that reduces to the usual lorentz group in the metric geometry limit . in section [ sec : gravdyn ] we will present our new theory of finsler gravity , including a matter coupling principle , which geometrically extends einstein gravity without introducing new fundamental scales . we will derive the finsler gravity field equation by variation and prove that it reduces consistently to the einstein field equations in the metric limit . symmetries of finsler spacetimes will be introduced in section [ sec : fsym ] . the maximally symmetric solution of vacuum finsler gravity turns out to be standard minkowski spacetime . section [ sec : corr ] considers a spherically symmetric perturbation around this vacuum which is found to be a refinement of the schwarzschild solution that shows how finsler gravity could resolve the fly - by anomaly . we conclude in section [ sec : disc ] . appendix [ app : details ] presents technical details of our derivations . the central idea behind the generalization of einstein gravity presented in this article is the description of spacetime and its dynamics by finsler geometry instead of metric geometry . in this section we will review the basic geometric concepts available on finsler spacetimes @xcite . these were introduced as a generalization of lorentzian metric spacetimes ; they allow full control of the null geometry and a clean definition of causality , which is essential for the description of light and observers . in particular , we will describe non - linear connections and curvature ; moreover , we will explain how to obtain well - defined field theory actions . the definition of finsler spacetimes involves the eight - dimensional tangent bundle @xmath7 which is the union of all tangent spaces to the underlying four - dimensional event manifold @xmath1 . thus , any point @xmath8 is a tangent vector to @xmath1 at some point @xmath9 ; there is a natural projection @xmath10 . it is convenient to use the so - called induced coordinates on the tangent bundle which are constructed as follows . let @xmath11 be coordinates on some open neighbourhood @xmath12 of @xmath13 . with respect to these we can express @xmath14 ; the induced coordinates of @xmath15 are then @xmath16 . the corresponding induced coordinate basis of @xmath17 will be denoted by @xmath18 and that of its dual @xmath19 by @xmath20 . * definition 1 . * _ geometric structures implemented in every tangent space @xmath21 by definition 1 of finsler spacetimes . the solid lines show the guaranteed cone of timelike vectors with the shells of unit timelike vectors and null boundary . the dotted lines indicate a potentially more complex null structure._,scaledwidth=30.0% ] lorentzian metric spacetimes @xmath22 arise from definition 1 in case @xmath23 ; then @xmath24 and @xmath25 is the set of unit @xmath5-timelike vectors . for general function @xmath26 , the relation between the inverse finsler metric @xmath27 and the inverse lagrange metric @xmath28 is given by @xmath29 the definition of finsler spacetimes guarantees a well - defined causal structure by the existence of timelike vectors that form an open convex cone with null boundary in every local tangent space , see figure [ fig : nullstructure ] . it provides full control of the geometry along the null directions where @xmath30 : we could show that the geometric concepts of connections and curvature that will be discussed below can be extended to this set , which is not possible in standard textbook formulations of finsler spaces . furthermore , by constructing a theory of electrodynamics on finsler spacetimes we proved that light propagates along null directions . for further details see @xcite . the geometry of finsler spacetimes is formulated by extending the standard language of finsler geometry , see e.g. @xcite ; this is a special geometry on the tangent bundle based on the finsler function @xmath3 that appears in the clock postulate . the basic geometric object deduced from the finsler function is the cartan non - linear connection . any non - linear connection on @xmath7 is equivalent to a unique decomposition of all tangent spaces to @xmath7 as @xmath31 , see figure [ fig : verthori ] . while the vertical bundle @xmath32 is canonically spanned by @xmath33 , the horizontal bundle @xmath34 is spanned by @xmath35 where the @xmath36 are the coefficients of the non - linear connection . then the dual bundle @xmath37 is spanned by @xmath38 , and @xmath39 by @xmath40 . for the cartan non - linear connection , the connection coefficients are determined by the fundamental functions @xmath3 or @xmath26 as @xmath41=\frac{1}{4}\bar\partial_b\big[g^{laq}\big(y^p\partial_p\bar\partial_ql-\partial_ql\big)\big ] .\ ] ] the equality of the two expressions for @xmath42 follows from the proof of theorem 2 in @xcite which makes particular use of the euler theorem @xmath43 for @xmath44-homogeneous functions . note that the right hand side is also valid on the null structure where @xmath3 is not even differentiable . _ tangent bundle geometry : decomposition of @xmath45 into horizontal and vertical parts._,scaledwidth=60.0% ] under tangent bundle changes of coordinates induced from coordinate transformations @xmath46 of the manifold , the non - linear connection coefficients @xmath42 transform in such a way that the horizontal / vertical basis @xmath47 of @xmath17 and the dual basis @xmath48 transform as vectors or one - forms over the manifold , @xmath49 tensors on @xmath7 that obey the standard tensor transformation law under manifold induced coordinate changes are called distinguished , or , in short , d - tensors . an example for a d - vector is the horizontal lift @xmath50 of a vector @xmath51 in @xmath21 to @xmath52 , @xmath53 the inverse map is given by the pushforward along the projection , @xmath54 . the horizontal tangent spaces to @xmath7 can thus be identified with the tangent spaces to the manifold @xmath1 ; this will be important for the identification of field components of tensors over the tangent bundle with field components measured over the spacetime manifold . a d - tensor of particular importance for our construction of gravitational dynamics for finsler spacetimes is the curvature of the cartan non - linear connection which measures the integrability of the horizontal bundle @xmath34 . it is defined as @xmath55^a =- y^q\big(\delta_b\gamma^{\delta a}{}_{qc}-\delta_c\gamma^{\delta a}{}_{qb}+\gamma^{\delta a}{}_{pb}\gamma^{\delta p}{}_{qc}-\gamma^{\delta a}{}_{pc}\gamma^{\delta p}{}_{qb}\big),\ ] ] using as a short - hand notation the generalized christoffel symbols @xmath56 which are related to the non - linear connection coefficients via @xmath57 . the proof of the above equality is given in theorem 2 of @xcite . in order to formulate covariant differential equations in the language of finsler geometry , a linear covariant derivative @xmath58 that acts on tensor fields over @xmath7 is needed . this must be compatible with the horizontal / vertical structure so that the identification of horizontal tangent spaces with tangent spaces to the manifold stays intact under transports . in the literature exist four such covariant derivatives with slightly different properties . we will employ the cartan linear covariant derivative defined by @xmath59 in later sections we will often need the cartan linear covariant derivative with respect to horizontal directions ; we will then use the abbreviation @xmath60 . the action of @xmath58 on the following sasaki type metric on @xmath7 , @xmath61 gives zero . hence this metric is covariantly constant with respect to the cartan linear connection ; it will play a crucial role as an integration measure on @xmath7 . this completes our quick review of geometric objects needed in this article . for later use observe that all the objects definable from @xmath3 , i.e. , the finsler metric @xmath27,the cartan non - linear connection @xmath42 , its curvature , and the cartan linear connection , are invariant under the transformation @xmath62 . in the metric limit @xmath23 they reduce to known constructions from metric geometry : the curvature essentially becomes the riemann curvature tensor , @xmath63 ; the generalized christoffel symbols become the usual christoffel symbols , @xmath64 ; the non - linear connection now is a linear connection , @xmath65 ; the cartan linear covariant derivative in horizontal directions becomes the levi - civita covariant derivative while it becomes trivial in vertical directions . we saw that all geometric objects on finsler spacetimes are homogeneous tensor fields on the tangent bundle ; the same will be true for physical fields in our construction . the formulation of field theory actions needs the well - defined integration procedure developed in @xcite . since integrals over the tangent bundle over homogeneous functions diverge , we must consider integrals over the unit tangent bundle @xmath66 on this domain we have a natural volume element constructed from the pull back of the sasaki - type metric ( [ eq : sasaki ] ) . thus , actions arise as integrals of scalar functions @xmath67 on @xmath7 restricted to @xmath68 as @xmath69 . to perform such integrals explicitly , we change coordinates on @xmath7 from the induced @xmath70 to the more convenient coordinates @xmath71 , where @xmath72 , the @xmath73 are zero - homogeneous , and @xmath74 . now @xmath68 is defined by @xmath75 and described by coordinates @xmath76 . the coordinate transformation matrices on @xmath7 are @xmath77,\qquad \frac{\partial z^a}{\partial \hat z^b } = \left[\begin{array}{c|c } \delta^a_b&\begin{array}{cc}0\quad & \;0 \end{array}\\ \hline \hat\partial_by^a & \begin{array}{cc } \partial_{u^\beta } y^a & \frac{y^a}{r } \end{array } \end{array}\right].\ ] ] and satisfy the invertibility properties [ eq : invvv ] @xmath78 = \left[\begin{array}{c|c}\delta^a_b & 0 \\ \hline \begin{array}{c}0 \\ 0 \end{array } & \begin{array}{cc } \delta^\alpha_\beta & 0 \\ 0 & 1 \end{array}\end{array}\right],\ ] ] @xmath79 = \left[\begin{array}{c|c}\delta^a_b & 0 \\ \hline 0 & \delta^a_b \end{array}\right].\ ] ] we now calculate the pullback of the sasaki type metric to @xmath68 in order to determine the relevant volume form . first we transform ( [ eq : sasaki ] ) to the new coordinates @xmath80 , which yields @xmath81 in terms of @xmath82 and @xmath83 . then the pull - back to @xmath68 with @xmath75 becomes @xmath84 using the shorthand notation @xmath85 and @xmath86 , a well - defined integral over @xmath68 of a homogeneous tangent bundle function @xmath67 now reads @xmath87 for tangent bundle functions @xmath88 that are homogeneous of degree @xmath44 the following formulae for integration by parts hold [ eq : intbp ] @xmath89_{|\sigma}\,,\\ \int_\sigma d^4\hat x d^3u\ \sqrt{g^fh^f}{}_{|\sigma}\big(\bar\partial_a a^a\big)_{|\sigma } & = & - \int_\sigma d^4\hat x d^3u\ \sqrt{g^fh^f}{}_{|\sigma}\big[\big(g^{f\,pq}\bar\partial_ag^f_{pq } - ( m+3 ) y^p g^f_{pa}\big)a^a\big]_{|\sigma}\,,\end{aligned}\ ] ] where @xmath90 and @xmath91 . these formulae can be proven with the help of the coordinate transformation relations ( [ eq : invvv ] ) . the definitions and mathematical techniques presented in this section are the foundation of the following new developments . in the next section we present how observers are modelled and perform measurements before we turn to the construction of a gravity theory for finsler spacetimes . in order to study physics on general finsler spacetimes it is necessary to define a mathematical model of physical observers and to determine how they measure time , spatial distances , and physical fields . guided by general relativity , freely falling observers move on trajectories that extremize the proper time integral ; variation of ( [ eq : fact ] ) here leads to finsler geodesics . moreover a model of observers requires four tangent vectors that build an orthonormal frame ; then measurable quantities are the components of physical fields with respect to this frame , evaluated at the observers position . we explicitly calculate the illustrative example of an observer s measurement of the speed of light on which the results of @xcite are based . to compare measurements of different observers it is necessary to communicate the results obtained by one observer to another . this communication is realized by a certain class of transformations between different observers ; we will show that these transformations have the algebraic structure of a groupoid that generalizes the usual lorentz group in metric geometry . an observer moves along a spacetime curve @xmath92 in @xmath1 with timelike tangents . the parametrization can be chosen so that @xmath93 is unit timelike ; according to definition 1 we now have @xmath94 and the signature of @xmath95 is lorentzian . for points of the tangent bundle @xmath96 , which is analogous to the coordinate representation @xmath97 . ] then the clock postulate ( [ eq : fact ] ) tells us that @xmath98 must be interpreted as the local unit time direction of the observer . we may write the normalization condition in the form @xmath99 using the horizontal lift @xmath100 of @xmath101 , see ( [ eq : horlift ] ) . to identify the three - space seen by an observer , we will complete @xmath102 to a four - dimensional basis @xmath103 of @xmath104 ; as explained before the projections of the @xmath105 for @xmath106 by @xmath107 into @xmath108 then are identified as the spatial tangent directions to the manifold . we determine the three horizontal vectors @xmath105 by the condition @xmath109 . this construction is justified by the observation that a horizontal three - space is defined by a conormal horizontal one - form . the only linearly independent one - form available in terms of geometric data is the vertical form @xmath110 . this can be mapped globally to the horizontal one - form @xmath111 , which is proportional to the cartan one - form known from finsler geometry and is a lagrangian analogue of the poincar one - form in hamiltonian mechanics . the condition @xmath112 on @xmath105 is equivalent to that stated above in terms of the finsler metric . we remark that the @xmath105 may depend less trivially on @xmath113 than in lorentzian geometry because of their defining equation @xmath114 . the definition of unique unit directions in the three - dimensional span @xmath115 requires orthonormalization . for this purpose we use the finsler metric to set @xmath116 , assuming definite signature . the choice of the metric @xmath27 for orthonormalization is preferred over that of @xmath28 , since only @xmath27 is invariant under @xmath62 as are the geometrical objects on finsler spacetimes , see section [ sec : basic ] . we will now prove a useful theorem on the relation between the signatures of the metrics @xmath28 and @xmath27 where both are defined ; a corollary will then confirm our assumption of a definite signature of the finsler metric on @xmath115 . * theorem 1 . * _ on the set @xmath117 the metric @xmath28 is nondegenerate of signature @xmath118 for natural numbers @xmath119 with @xmath120 . then the finsler metric has the same signature where @xmath121 , and reversed signature @xmath122 where @xmath123 . _ the observer s time direction @xmath102 is in @xmath124 which is contained in @xmath117 . this tells us that the metric @xmath125 has signature @xmath126 for @xmath127 and @xmath128 for @xmath129 . we also know that @xmath99 . hence we conclude from theorem 1 : * corollary . * _ the finsler metric @xmath130 evaluated at the tangent bundle position of an observer has lorentzian signature @xmath126 , and the unit spatial directions satisfy @xmath131 . _ _ proof of theorem 1 . _ by the definition of finsler spacetimes the metric @xmath28 is non - degenerate on @xmath132 , hence also on the smaller set excluding the null structure on which @xmath27 is defined . now observe that if an inner product is given by a matrix @xmath133 and @xmath134 has indefinite signature @xmath118 , then the signature of @xmath135 is found to be @xmath136 for @xmath137 and @xmath118 for @xmath138 ; for @xmath139 the result is the once degenerate signature @xmath140 . this can be seen in a sylvester normal form basis for @xmath141 by using the remaining @xmath142 freedom . we can apply this result to our situation by identifying @xmath143 from equation ( [ eq : gftogl ] ) . it follows that @xmath144 ; then the signature of @xmath27 is the signature of @xmath141 , i.e. , that of @xmath28 up to the sign of @xmath26 . @xmath145 we summarize our construction of observers frames and measurements into a precise definition . * definition 2 . * the minkowski metric @xmath146 in this definition has signature @xmath128 . we emphasize again that the frame @xmath147 in @xmath104 can be identified one to one with a frame @xmath148 in @xmath108 , or reversely by the horizontal lift . from ( [ eq : gftogl ] ) , one can show that the observer frame has the nice property to diagonalize @xmath27 and @xmath28 simultaneously . we will now discuss the measurement procedure in more detail and present the example of how an observer measures the speed of light . definition 2 of the observer frame includes the statement that a physical observable is given by the components of a horizontal tensor field with respect to the observer s frame , evaluated at her position on the tangent bundle , i.e. , at her position on the manifold and her four - velocity . the motivation for this is as follows . the geometry of finsler spacetimes is formulated on the tangent bundle @xmath7 , and hence matter tensor fields coupling to this gravitational background must also be defined over @xmath7 . not all such tensor fields can be interpreted as tensor fields from the perspective of the spacetime manifold @xmath1 . this interpretation requires that the tensor fields be horizontal ; then they are multilinear maps built on the horizontal space @xmath149 and its dual which are identified with the tangent space @xmath150 and its dual . consider the example of a 2-form field @xmath151 over @xmath7 ; in the horizontal / vertical basis this expands as @xmath152 only the purely horizontal part @xmath153 has a clear interpretation . note that such horizontal tensor fields are automatically d - tensor fields , and have the same number of components as a tensor field of same rank on @xmath1 . the difference is that the components depend on the tangent bundle position . the measurement of a horizontal tensor field by an observer at the tangent bundle position @xmath154 clearly requires an observer frame of @xmath104 in order to read out the components . we emphasize that the dependence of observables on the four - velocity of the observer is not surprising . neither is it problematic as long as observers can communicate their results . in general relativity , observables are the components of tensor fields over @xmath1 with respect to the observer s frame in @xmath108 ; they clearly depend on @xmath101 which induces the splitting of @xmath108 into time and space directions . on finsler spacetimes the dependence of observables on the observer s four - velocity is not only present in the time / space split of @xmath104 , but also in the argument of the tensor field components . as a simple example we discuss the measurement of the spatial velocity of a point particle that moves on a worldline @xmath155 with horizontal tangent @xmath156 this can be expanded in the orthonormal frame of an observer as @xmath157 , where we recall that @xmath158 is the observer s four velocity . the time @xmath159 passes while the particle moves in spatial direction @xmath160 , so the spatial velocity @xmath161 and its square @xmath162 are @xmath163 as a consequence of this formula we may derive the speed of light seen by a given observer . as discussed in @xcite light propagates on null worldlines @xmath155 with @xmath164 which is equivalent to @xmath165 . we can use this fact to replace the finsler metric in the formula for the velocity above . taylor expanding @xmath166 around @xmath167 yields @xmath168 evaluating this formula at the position of the observer @xmath169 and dividing by @xmath170 , we immediately obtain an expression for the speed of light @xmath171 , i.e. , the speed of light traveling in spatial direction @xmath172 and measured by the observer @xmath154 : @xmath173 the @xmath159 are determined by solving the null condition @xmath174 ; on a generic finsler spacetime there can be more than one solution since the null structure can be very complicated . from equation ( [ eq : speedlight ] ) we see that the measured speed of light depends on the higher than second order derivatives of the squared finsler function ; these vanish in the metric limit where we thus reobtain @xmath175 independent of the observer and the spatial direction of the light ray . the formulae ( [ eq : v2 ] ) and ( [ eq : light ] ) enable us to compare experimental results on particle and light velocities with predictions on specific finsler spacetime models . in @xcite we used this to study the possibility of superluminal particle propagation . we already stressed the importance that observers should be able to communicate their measurements . consider two observers whose worldlines meet at a point @xmath176 . since observers by definition 2 measure the components of horizontal tensor fields in their frame and at their tangent bundle position , we need to determine which transformation uniquely maps an observer frame @xmath147 in @xmath52 to a second observer frame @xmath177 in @xmath178 . their respective four - velocities , or time directions , @xmath179 and @xmath180 generically are different , so that the two observer frames are objects in tangent spaces to @xmath7 at different points . as a consequence , we will now demonstrate that the transformations between observers consist of two parts : the first is a transport of the frame @xmath147 from @xmath181 to @xmath182 , the second will turn out to be a lorentz transformation . * theorem 2 . * _ proof . we first show that the curve @xmath183 required in _ ( i ) _ exists . the verticality condition @xmath184 implies @xmath185 ; the definition of the cartan linear connection ( [ eq : cartanlin ] ) then tells us that @xmath186 is equivalent to solving @xmath187 . this has a unique solution connecting @xmath181 to any point @xmath182 in a sufficiently small neighbourhood in @xmath21 . now let @xmath188 be the parallelly transported vector fields @xmath189 with @xmath190 . the properties of the cartan linear connection ensure that the @xmath191 are horizontal fields . observe also that @xmath192 along the curve @xmath183 since @xmath27 is covariantly constant under @xmath58 . it follows that @xmath193 is independent of @xmath194 , and holds in particular at the final point of the transport @xmath195 . now @xmath196 and @xmath177 are orthonormal frames with respect to @xmath27 in @xmath178 ; hence they are related by a unique lorentz transformation as stated in point _ ( ii ) _ of the theorem . @xmath145 the procedure described in theorem 2 provides a map between the frames of two observers at the same point of the manifold @xmath176 , but with different four - velocities @xmath197 ; we display the two parts of this procedure as @xmath198 , i.e. , as parallel transport followed by lorentz transformation , which is illustrated in figure [ fig : trafo ] . the combined maps transform observers uniquely into one another as long as the autoparallel @xmath183 connecting the vertically different points in @xmath7 exists and is unique . this is certainly the case if @xmath181 and @xmath182 are sufficiently close to each other . whether the geometric structure of a specific , or maybe all , finsler spacetimes is such that unique transformations between all observers exist requires requires further investigation . in the observer transformations on generic finsler spacetimes there appears an additional ingredient that is not present on metric spacetimes . before applying the lorentz transformation to the frame , one has to perform a parallel transport in the vertical tangent space . in the metric limit the vertical covariant derivative becomes trivial so that the parallely transported frame does not change at all along the curve @xmath183 . in this special case the transformation of an observer thus reduces to @xmath199 which is fully determined by a lorentz transformation . _ transformation between two observer frames : the frame @xmath147 in @xmath52 is first parallely transported to @xmath200 in @xmath178 , second lorentz transformed into the final frame @xmath177._,scaledwidth=55.0% ] the observer transformations on finsler spacetimes essentially have the algebraic structure of a groupoid that reduces to the lorentz group in the metric limit . we first review the general definition of a groupoid and then show how this applies to our case . * definition 3 . * _ a groupoid @xmath201 consists of a set of objects @xmath202 and a set of arrows @xmath203 . every arrow @xmath141 is assigned a source @xmath204 and a target @xmath205 by the maps @xmath206 and @xmath207 ; one writes this as @xmath208 . for arrows @xmath141 and @xmath209 whose source and target match as @xmath210 there exists an associative multiplication @xmath211 with @xmath212 a unit map @xmath213 where @xmath214 exists so that @xmath215 for every arrow @xmath141 exists an inverse arrow @xmath216 that satisfies @xmath217 _ groupoids are generalizations of groups . these can be expressed as groupoids with a single object in @xmath202 ; then the arrows correspond to group elements all of which can be multiplied since sources and targets always match . the multiplication is associative , the identity element and inverse elements exist . consider @xmath218 as the set of unit timelike vectors which contains the different four - velocities of observers at the point @xmath176 . let the arrows in @xmath203 be the set of all maps between two observer frames at @xmath219 which are defined by the procedure stated in theorem 2 . in case the involved vertical autoparallels connect the four - velocities uniquely , the sets @xmath202 and @xmath203 define a groupoid : source and target of a map @xmath220 between two frames @xmath221 and @xmath222 are simply given by @xmath223 and @xmath224 ; the multiplication @xmath225 is defined by applying the procedure of theorem 2 to construct the map between @xmath226 and @xmath227 , which gives the properties ( [ eq : mult ] ) ; we choose the unit map @xmath228 that provides ( [ eq : one ] ) as @xmath229 , i.e. , as trivial parallel transport of the frame @xmath221 with respect to the cartan linear connection along the vertical autoparallel that stays at @xmath181 followed by the identity lorentz transformation . finally , we define the inverse @xmath230 , where @xmath231 denotes parallel transport backwards along the unique vertical autoparallel connecting @xmath181 and @xmath182 which is also used for @xmath232 ; to check the properties ( [ eq : inv ] ) , one simply shows that parallel transport of the frames and lorentz transformation commute . thus we have shown the following result : * theorem 3 . * _ on finsler spacetimes @xmath233 the transformations between observer frames @xmath234 at @xmath176 that are attached to points @xmath235 define a groupoid @xmath201 under the condition that any pair of points in @xmath236 can be connected by a unique vertical autoparallel of the cartan linear connection . _ we already discussed that the transformations of observer frames reduce to the form @xmath237 in the limit of metric geometry . hence the only information contained in the reduced groupoid @xmath238 with @xmath239 and @xmath240 is given by the lorentz transformations . in mathematically precise language this can be expressed as the equivalence of @xmath238 to the lorentz group seen as a groupoid @xmath241 with a single object @xmath242 and arrows @xmath243 . the functor @xmath244 establishing the equivalence can be defined by the projection @xmath245 and by @xmath246 . indeed , @xmath247 can be checked to be injective , full and essentially surjective , and so it makes @xmath248 and @xmath241 equivalent . see @xcite for details on the required mathematical definitions . we have now reviewed the basic concept of finsler spacetime geometry and laid the foundations for the interpretation of physics on these backgrounds . we have seen that well - defined observers exist which communicate with each other by means of groupoid transformations that generalize the lorentz group in metric geometry . as emphasized in previous work @xcite , physical predictions in our generalized geometric framework require gravitational dynamics for the fundamental geometry function @xmath26 to determine specific spacetime solutions . in this section we for the first time present a finsler gravity action along with a consistent minimal coupling principle between gravity and matter . we begin our presentation with the variation of the pure gravity action , before the coupling of finsler gravity to matter is discussed in section [ sec : gravmat ] ; the full field equation is derived in section [ sec : gravmateq ] . moreover , we prove that the finsler gravity field equation becomes equivalent to the einstein equations in the metric limit . on finsler spacetime the simplest curvature scalar built from the non - linear curvature tensor @xmath249 , which is relevant for the tidal acceleration of finsler geodesics , is @xmath250 . this contains the lowest number of derivatives on the fundamental function @xmath26 without involving additional d - tensors besides the curvature , like @xmath251 or @xmath252 . recall that integrals are well - defined over the seven - dimensional unit tangent bundle @xmath68 with coordinates @xmath76 , as discussed in section [ sec : int ] . these two facts directly lead to our finsler gravity action @xmath253=\int d^4\hat xd^3u \left[\sqrt{g^fh^f}\ \mathcal{r}\right]_{|\sigma}\,.\ ] ] the gravitational field equation in vacuum now is obtained by variation with respect to @xmath26 . to perform this variation for an @xmath44-homogeneous function @xmath254 on @xmath7 restricted to @xmath68 it is useful to realise that @xmath255 where @xmath256 is the homogeneity of @xmath26 . with the help of this formula and the results for integration by parts in ( [ eq : intbp ] ) we can derive the vacuum field equations in three steps . the first uses the variation formula above with @xmath257 and @xmath258 , which yields @xmath259=\int d^4\hat xd^3u \left[\delta\big(\sqrt{g^fh^f}\ \mathcal{r}\big)-\frac{5}{n}\sqrt{g^fh^f}\ \mathcal{r } \frac{\delta l}{l}\right]_{|\sigma}\,.\end{aligned}\ ] ] the second step is the variation of the volume element which leads to @xmath260=\int d^4\hat xd^3u \sqrt{g^fh^f}_{|\sigma } \left [ \big(g^{f\,pq}\delta g^f_{pq}-\frac{6}{n}\frac{\delta l}{l}\big)\mathcal{r}+y^b\delta r^a{}_{ab } \right]_{|\sigma}\,,\ ] ] while in the third step we use the following identities [ eq : calcgrav ] @xmath261_{|\sigma } = \int d^4\hat xd^3u \left[\sqrt{g^fh^f}\ g^{f\,ab}\bar\partial_a\bar\partial_b\mathcal{r}\frac{\delta l}{nl}\right]_{|\sigma}\,,\ ] ] @xmath262_{|\sigma } = \int d^4\hat xd^3u \left[\sqrt{g^fh^f}\ 2g^{f\,ab}\big(\nabla_as_b+s_as_b+\bar\partial_a\nabla s_b\big)\frac{\delta l}{nl}\right]_{|\sigma}\,,\ ] ] to arrive at the final form of the variation of the finsler gravity action ( [ eq : gravact ] ) : @xmath263=\int d^4\hat xd^3u \sqrt{g^fh^f}_{|\sigma}\left [ g^{f\,ab}\bar\partial_a\bar\partial_b\mathcal{r}-6\mathcal{r}+2g^{f\,ab}\big(\nabla_as_b+s_as_b+\bar\partial_a\nabla s_b\big)\right]_{|\sigma}\frac{\delta l}{nl}\,.\ ] ] for further details of this variation we refer the reader to appendix [ app : vargrav ] . now we can read off the vacuum field equation on @xmath68 as @xmath264_{|\sigma } = 0 \,.\ ] ] observe that all terms in the bracket are zero - homogeneous on @xmath7 , except the second term @xmath265 that has homogeneity two . since @xmath266 we can replace the second term by @xmath267 which is now also zero - homogeneous . hence the equation can be lifted to @xmath7 in the form @xmath268 it seems as if this equation could be invalid on @xmath269 where @xmath3 is not differentiable so that the finsler metric @xmath27 does not exist . however , this is not the case : the equation is valid also on the null structure . to see this , one expresses @xmath27 through @xmath28 with the help of formula ( [ eq : gftogl ] ) and multiplies by @xmath270 . the resulting equation is well - defined whereever @xmath28 is nondegenerate , and in particular on the null structure . note that equation ( [ eq : gravvac ] ) is invariant under the transformation @xmath271 , which will be a guiding principle for matter coupling below . in the metric limit @xmath272 , the tensors in the the finsler gravity equation reduce as @xmath273 and @xmath274 , where @xmath275 is the ricci tensor of the metric @xmath5 . accordingly , the field equation becomes @xmath276 which is equivalent to the einstein vacuum equations @xmath277 by differentiating twice with respect to @xmath278 . we conclude that a family of solutions of the finsler gravity vacuum equation ( [ eq : gravvac ] ) is induced by solutions @xmath279 of the vacuum einstein equations via the fundamental functions @xmath280 . in section [ sec : corr ] we will present a solution of the finsler gravity vacuum equation beyond metric geometry . above we have achieved a consistent generalization of vacuum einstein gravity from metric spacetimes to finsler spacetimes . next we will show that this generalization can be completed by the coupling of matter fields . for this purpose we will discuss a minimal coupling principle that generates consistent matter field actions on finsler spacetimes from their well - known counterparts on metric spacetimes . in the discussion we restrict our attention to @xmath281-form fields ; spinor fields have to be investigated further . in section [ sec : gravmateq ] we will deduce the complete gravity equations with energy - momentum source term . consider an action @xmath282 $ ] for a physical @xmath281-form field @xmath283 on a lorentzian spacetime @xmath22 , @xmath284 = \int_m d^4x\ \sqrt{g}\ \mathcal{l}(g,\phi , \mathrm{d}\phi)\,.\ ] ] the corresponding matter action on finsler spacetime is obtained by lifting @xmath285 to the tangent bundle @xmath7 equipped with the sasaki - type metric @xmath286 defined in ( [ eq : sasaki ] ) in the following way : a. consider the lagrangian density @xmath287 of the standard theory on @xmath1 as a contraction prescription that forms a scalar function from various tensorial objects ; b. replace the lorentzian metric @xmath288 in @xmath287 by the sasaki - type metric @xmath286 ; c. replace the @xmath281-form field @xmath289 on @xmath1 by a zero - homogeneous @xmath281-form field as in ( [ eq:2form ] ) , for example , is zero - homogeneous if and only if its components @xmath290 , @xmath291 and @xmath292 have the homogeneities @xmath293 , @xmath294 and @xmath295 , respectively ; these are cancelled by the homogeneity of @xmath296 . ] @xmath297 on @xmath7 ; d. introduce lagrange multipliers @xmath298 for all not purely horizontal components of @xmath151 ; e. finally integrate over the unit tangent bundle @xmath68 with the volume form given by the pull - back @xmath299 of the sasaki - type metric . the result of this procedure is the finsler spacetime field theory action @xmath300=\int_\sigma d^4\hat xd^3u \big[\sqrt{g^fh^f } \big(\mathcal{l}(g,\phi , \mathrm{d}\phi)+\lambda ( 1-p^h)\phi \big)\big]_{|\sigma}\,.\ ] ] the projection @xmath301 projects to the purely horizontal part of the @xmath281-form @xmath151 ; in the example of the general two - form on @xmath7 displayed in equation ( [ eq:2form ] ) we have @xmath302 the lagrange multiplier guarantees that the on - shell degrees of freedom of @xmath151 are precisely those with a clear physical interpretation as fields along the manifold @xmath1 , as discussed in section [ sec : measure ] . the minimal coupling principle for matter to finsler spacetime presented above is a slightly refined version of that in @xcite where we discussed electrodynamics on finsler spacetime . the only modification here is in the definition of the sasaki - type metric @xmath286 that is now defined in terms of the finsler metric @xmath27 instead of @xmath28 . this change ensures that the resulting matter action @xmath303 $ ] is invariant under @xmath62 in the same way as the pure gravity action . nevertheless , the results obtained for electrodynamics on finsler spacetime in @xcite are unchanged by the refined coupling principle presented here . the matter field equations obtained by extremizing the action with respect to the @xmath281-form field @xmath151 and the lagrange multiplier @xmath298 can be studied most easily if expressed in components with respect to the horizontal / vertical basis . the calculation is performed in detail in appendix [ app : varmat ] . we display the results with the convention that barred indices denote vertical components , unbarred indices now denote horizontal components , and capital indices both horizontal and vertical components . variation with respect to the lagrange multiplier yields the constraints @xmath304 variation for the purely horizontal components of @xmath151 gives @xmath305 which determines the evolution of the physical field components , while variation with respect to the remaining components produces @xmath306 which fixes the components of the lagrange multiplier . the @xmath307 are the commutator coefficients of the horizontal / vertical basis . our coupling principle is consistent with the metric limit , i.e. , the equations of motion obtained from the finsler spacetime action reduce to the equations of motion on lorentzian spacetime in the case @xmath272 and @xmath308 . then we have the geometric identity @xmath274 ; moreover @xmath309}\ ] ] using the constraints ( [ eq : cons ] ) and the fact that the horizontal derivative acts as a partial derivative on the @xmath278-independent @xmath281-form components . finally , @xmath310 because , as a consequence of our coupling principle where the sasaki - type metric is block - diagonal in the horizontal / vertical basis , the vertical index of @xmath311 must appear in @xmath312 contracted via @xmath27 into either a vertical derivative or into components of @xmath151 with at least one vertical index . in the metric limit , vertical derivatives give zero , while the constraints ( [ eq : cons ] ) guarantee that all components of @xmath151 with at least one vertical index vanish . combining these observations shows that equation ( [ eq : eom1 ] ) reduces to @xmath313 where @xmath58 now operates in the same way as the levi civita connection of the metric @xmath5 . again , as a consequence of our minimal coupling principle with the block - diagonal form of the sasaki - type metric in the horizontal / vertical basis , we can conclude in the metric limit that @xmath314 so that ( [ eq : red ] ) becomes equivalent to the standard @xmath281-form field equation of motion on metric spacetime . our minimal coupling procedure for matter fields to finsler spacetime can be applied immediately for instance to the scalar field , as done in @xcite . note that it can be easily extended to the case of interacting form fields of any degree with metric spacetime action @xmath315 = \int_m d^4x\ \sqrt{g}\ \mathcal{l}(g,\phi_1 , \mathrm{d}\phi_1,\phi_2,\mathrm{d}\phi_2, ... )\,.\ ] ] the minimal coupling procedure then leads to the action @xmath316\nonumber\\ & = & \int_\sigma d^4\hat xd^3u \big[\sqrt{g^fh^f } \big(\mathcal{l}(g,\phi_1 , \mathrm{d}\phi_1,\phi_2 , \mathrm{d}\phi_2, ... )+\sum_i \lambda_i ( 1-p^h)\phi_i \big)\big]_{|\sigma}\,.\end{aligned}\ ] ] the equations of motion for each field @xmath317 have the same form as in the single field case , and the metric limit leads to the standard field equations by arguments that proceed in a completely analogous way as before . in the standard formulation of electrodynamics , the action is a functional @xmath318 $ ] of a one - form potential @xmath141 , but the classical physical field is @xmath319 . our minimal coupling principle to obtain an action on finsler spacetimes can not be applied immediately to this situation : the problem is that the lagrange multiplier then only kills the vertical components of the lift of @xmath141 , but does not guarantee that the lift of @xmath3 is purely horizontal . this problem is solved in @xcite by starting from an equivalent interacting action of the form @xmath320 $ ] which provides the complete set of maxwell equations @xmath319 and @xmath321 by variation . now the minimal coupling principle entails that both the lifted fields @xmath141 and @xmath3 are purely horizontal and can be interpreted physically . we are now in the position to study the interplay between the matter actions @xmath322 introduced in ( [ eq : matact ] ) and the pure finsler gravity action @xmath323 in ( [ eq : gravact ] ) . their sum provides a complete description of gravity and classical matter fields on finsler spacetimes : @xmath324 & = & \kappa ^{-1}s_g[l]+s_m[l,\phi,\lambda ] \nonumber\\ & = & \kappa^{-1}\int d^4\hat xd^3u \left[\sqrt{g^fh^f}\ \mathcal{r}\right]_{|\sigma}+ \int d^4\hat xd^3u\ \big[\sqrt{g^fh^f}\ ( \mathcal{l}(g,\phi , \mathrm{d}\phi)+\lambda(1-p^h)\phi)\big]_{|\sigma}\,.\end{aligned}\ ] ] as usual , the matter field equations following from this are the same as for the pure matter action . the gravitational field equations are obtained by variation with respect to the fundamental geometry function @xmath26 . the variation of @xmath322 with respect to @xmath26 is @xmath325 and leads us to the definition of the energy momentum scalar @xmath326 on the unit tangent bundle as @xmath327 with this definition the complete gravitational field equations on finsler spacetime including energy - momentum sources formally become @xmath328_{|\sigma}=-\kappa t_{|\sigma}\,.\ ] ] as in the vacuum case with @xmath329 , these equations can be lifted to @xmath7 . the terms in the bracket on the left hand side are all zero - homogeneous and can be lifted trivially . the terms in @xmath330 without the restriction on the right hand side in principle can result from variation with different homogeneities ; to lift these one simply multiplies each term by the appropriate power of @xmath3 in order to make it zero homogeneous . this is the same procedure applied in section [ sec : gravvac ] to the gravity side . the gravitational constant @xmath331 will now be determined so that the gravitational field equation on finsler spacetimes becomes equivalent to the einstein equations in the metric limit . variation with respect to @xmath26 of the concrete form of the matter action in ( [ eq : compact ] ) and performing the metric limit , i.e. , @xmath332 for observers and @xmath333 , the gravity equation ( [ eq : fgrav ] ) becomes @xmath334 the detailed calculation of this result is involved and can be found in appendix [ app : varmat ] . introducing the standard energy momentum tensor of @xmath281-form fields on lorentzian metric spacetimes @xmath335 and its trace @xmath336 we can rewrite the equation above as @xmath337 if evaluated at @xmath5-timelike observer four - velocities @xmath278 . now we take a second derivative with respect to @xmath278 , contract with @xmath338 , reinsert the result , and conclude @xmath339 since there is no @xmath278-dependence beyond the explicit one , a second derivative with respect to @xmath278 yields the einstein equations , if we choose the gravitational constant @xmath340 . the gravity equation on finsler spacetime including the coupling to matter therefore is @xmath341 with zero homogeneous source function @xmath330 on @xmath7 . observe that this field equation including the matter part is invariant under @xmath271 by construction of the coupling principle , as the vacuum equation is . this leads to the interesting conclusion that every solution @xmath279 of the einstein equations induces a family @xmath342 of solutions of the finsler gravity solution with @xmath343 . in order to find further solutions of this highly complicated partial differential equation we will study symmetries of finsler spacetimes in the next section . then we present a solution of the linearised finsler gravity equation in section [ sec : corr ] which turns out to be a geometric refinement of the linearised schwarzschild solution of general relativity . in the previous section we have deduced the finsler gravity field equation including matter sources and shown that it is consistent with the einstein equations in the metric geometry limit . our new field equation is a highly complex differential equation ; in order to simplify the task of finding analytic solutions we wish to consider symmetric spacetimes . we begin this section by defining symmetries of finsler spacetimes and show how this concept is a generalisation of the symmetries of lorentzian spacetimes . we explicitly present the general structure of the fundamental geometry function @xmath26 for the spherically , cosmologically and maximally symmetric case . in the next section we will then use our results about symmetric finsler spacetimes to solve the linearised finsler gravity equation . on a symmetric lorentzian manifold @xmath22 , the metric is invariant under certain diffeomorphisms ; similarly we wish to define symmetries of a finsler spacetime @xmath233 as an invariance of the fundamental geometry function @xmath26 . consider a diffeomorphism generated by the vector field @xmath344 ; this acts as a coordinate change on local coordinates on @xmath1 as @xmath345 , and on the induced coordinates on the tangent bundle @xmath7 as @xmath346 . hence the diffeomophism on @xmath1 induces a diffeomorphism on @xmath7 that is generated by the vector field @xmath347 , called the complete lift of @xmath51 . the idea of implementing symmetries via complete lifts in a finsler geometry setting appears already in @xcite ; here we want to make this concept precise for finsler spacetimes : * definition 4 . * _ a symmetry of a finsler spacetime @xmath233 is a diffeomorphism generated by a vector field @xmath348 over the tangent bundle @xmath7 so that @xmath349 and @xmath348 is the complete lift @xmath350 of a vector field @xmath51 over @xmath1 . a finsler spacetime is called symmetric if it possesses at least one symmetry . _ the following theorem summarizes important properties of finsler spacetime symmetries . the symmetry generators form a lie algebra with the commutator of vector fields on @xmath7 , and they are isomorphic to a lie algebra of vector fields on @xmath1 which becomes the usual symmetry algebra of lorentzian manifolds in the metric geometry limit . this not only shows that definition 4 of symmetry is consistent with that of lorentzian spacetimes , but also that the usual killing vectors , e.g. , those for spherical symmetry , can be used to study symmetries of finsler spacetimes . * theorem 4 . * _ proof . _ _ ( i ) _ let @xmath351 ; then @xmath349 and @xmath352 . both properties are linear , so that @xmath353 is a vector subspace of the lie algebra of all vector fields on @xmath7 . it remains to be proven that the commutator of two elements @xmath354 closes in @xmath353 . it is clear that @xmath355(l)=0 $ ] ; to show that @xmath356)^c=[y_1,y_2]$ ] , one uses that @xmath357 for some vector fields @xmath358 on @xmath1 and that @xmath359=[x_1,x_2]^c$ ] , see @xcite . _ ( ii ) _ the inverse for @xmath107 on @xmath360 is given by the complete lift , hence @xmath353 and @xmath360 are isomorphic as vector spaces . the lie algebra structure is preserved in both directions because of @xmath359=[x_1,x_2]^c$ ] , and hence also @xmath361=[\pi_*y_1,\pi_*y_2]$ ] . _ ( iii ) _ for @xmath362 , we have @xmath363 . in the metric geometry limit @xmath23 , and hence @xmath364 . since the lie - derivative of the metric @xmath5 does not depend on the fibre coordinates of the tangent bundle we conclude @xmath365 . this is the condition that defines @xmath51 as the symmetry generator of a metric spacetime . we now wish to study the implications of spherical , cosmological and maximal symmetry for the fundamental function @xmath26 of a finsler spacetime . consider a finsler spacetime @xmath233 and coordinates @xmath366 on its tangent bundle . spherical symmetry is defined by the following three vector fields , that generate spatial rotations and form the algebra @xmath367 , @xmath368 their complete lifts are obtained via the procedure described in the previous section [ eq : ssymc ] @xmath369 applying the symmetry condition @xmath370 implies @xmath371 , while using @xmath372 and @xmath373 to deduce @xmath374 and @xmath375 yields @xmath376 in order to analyze the implications of these equations on @xmath26 we introduce new coordinates @xmath377 while keeping @xmath378 . the associated transformation of the derivatives @xmath379 makes the equations ( [ eq : spherical ] ) equivalent to the simple constraints @xmath380 and @xmath381 . hence we conclude from the analysis of the symmetry conditions @xmath382 that the most general spherically symmetric finsler spacetime is described by a fundamental function which is @xmath256-homogeneous in @xmath383 and of the form @xmath384 where @xmath385 is defined in ( [ eq : c1 ] ) . after our discussion of the spherically symmetric case in full detail above , we will now present the results of a similar analysis first for cosmologically and second for maximally symmetric finsler spacetimes . cosmological symmetry describes an isotropic and homogeneous spacetime . this is a much more symmetric situation than in the spherically symmetric scenario , and is implemented by requiring the following six vector fields to be symmetry generators , see @xcite , [ eq : csym ] @xmath386 where we write @xmath387 and @xmath388 is constant . the complete lifts of these vector fields are listed in appendix [ app : compl ] . applying the symmetry conditions @xmath389 to the fundamental function @xmath26 and introducing the new coordinates @xmath390 while keeping @xmath391 yields the following result : the cosmological fundamental function @xmath26 is @xmath256-homogeneous in @xmath392 and has the form @xmath393 the constant @xmath388 only appears in the expression for the coordinate @xmath394 . the value of @xmath394 can be understood as the metric length measure on a three - dimensional manifold of constant curvature @xmath388 . the same metric appears in the spatial part of the standard robertson walker metric . for the study of maximally symmetric finsler spacetimes we use some notation from @xcite , where such spacetimes are constructed from embeddings into a five - dimensional manifold . symmetry vectors generating maximal symmetry are given by @xmath395 with @xmath396 , constant @xmath397 , and constant @xmath398 matrices @xmath135 and @xmath399 . there are four linearly independent vector fields @xmath400 and six @xmath401 by requiring the condition @xmath402 ; their complete lifts are @xmath403 evaluating the symmetry conditions @xmath404 and @xmath405 on the fundamental function , and introducing new coordinates @xmath406 where @xmath407 runs over any three indices in @xmath408 , yields the following result : the maximally symmetric fundamental function @xmath26 is @xmath256-homogeneous in @xmath409 , and of the form @xmath410 the final equality is obtained from euler s theorem for homogeneous functions . observe that the maximally symmetric fundamental function always describes a metric geometry , see ( [ eq : c3 ] ) . hence all maximally symmetric finsler spacetimes are lorentzian spacetimes , and the gravity equation ( [ eq : complgrav ] ) is equivalent to einstein s equations . thus we can immediately conclude that the only maximally symmetric , source free vacuum solution of our finsler gravity equation is the minkowski metric induced fundamental function @xmath411 and its powers . in the expression for @xmath409 above this corresponds to @xmath412 and @xmath413 . maximally symmetric spacetimes with @xmath414 can only be obtained as solutions of the finsler gravity equation by adding a cosmological constant term , similarly as in general relativity . in this section we will study finsler spacetimes that describe mild deviations from lorentzian geometry . in this situation , the complicated finsler gravity field equation allows a simplified treatment . after a general discussion of the linearised field equation , we will employ what we learned about spacetime symmetries to present a spherically symmetric solution . this particular model turns out to be a refinement of the linearised schwarzschild solution of general relativity , and we will argue that it should be capable of modelling unexplained effects in the solar system like the fly - by anomaly . recall that the fundamental functions @xmath415 and @xmath416 define the same geometry , and that this is respected by the finsler gravity field equation . hence the following class of fundamental functions gives us good control over deviations from lorentzian metric geometry , @xmath417 here , @xmath418 is a @xmath419-homogeneous function that causes the finsler modifications of the lorentzian metric spacetime @xmath22 . the abbreviation @xmath286 should not be confused with the sasaki - type metric on @xmath7 . recall the finsler gravity vacuum field equation from ( [ eq : gravvac ] ) : @xmath420 we will now expand this equation to linear order in the modification @xmath418 , where @xmath421 . in the following calculations we suppress all higher order terms . we introduce @xmath422 and @xmath423 to expand the finsler function and finsler metric as @xmath424 the coefficients of the nonlinear connection are calculated from equation ( [ eq : nonlin ] ) : @xmath425 here , the @xmath426 are the christoffel symbols of the metric @xmath5 , and in the following @xmath58 acts as the levi civita connection . note that the zeroth order term , for @xmath427 , is the metric linear connection . the curvature and the tensor @xmath428 can be expressed with help of the shorthand notation @xmath429 as @xmath430- \nabla_a ( y^by^ct^a{}_{bc})+ \nabla(y^ct^a{}_{ac})\,,\\ s_a&=&\gamma^{\delta p}{}_{pa}-\bar\partial_p n^p{}_a \simeq -y^q\bar\partial_at^p{}_{pq}\,.\end{aligned}\ ] ] the zeroth order term in @xmath265 is determined by the ricci tensor of @xmath5 , while @xmath431 . collecting all terms in the gravitational field equation ( [ eq : gravvac2 ] ) finally yields @xmath432+\frac{6}{g}y^ay^br_{ab}[g]\\ & + & \big[2 \frac{g}{|g| } l^{ab}r_{ab}[g]+\frac{6g-2l}{g^2}y^ay^br_{ab}[g]+g^{ab}\bar\partial_a\bar\partial_b(-\nabla_a ( y^by^ct^a{}_{bc})+\nabla(y^ct^a{}_{ac}))\nonumber \\ & & \quad- \frac{6}{g } ( -\nabla_a ( y^by^ct^a{}_{bc})+\nabla(y^ct^a{}_{ac}))-2 g^{ab}(\nabla_ay^q\bar\partial_bt^p{}_{pq}+\bar\partial_a\nabla y^q\bar\partial_bt^p{}_{pq})\big]\,.\nonumber\end{aligned}\ ] ] the zeroth order contribution in the first line is equivalent to the einstein vacuum equations , as discussed in section [ sec : gravvac ] . the first order terms in square brackets determine the finsler modification of the unperturbed metric background solution . the details of how to rewrite the different terms of this equation in terms of the perturbation @xmath433 in the fundamental function @xmath26 , see ( [ eq : lpert ] ) , instead of @xmath434 can be found in appendix [ app : lin ] . we will now use our results on symmetries and on the linearisation of vacuum finsler gravity around metric spacetimes to derive a particular model that refines the linearised schwarzschild solution and can be used to study solar system physics . recall from section [ sec : sphsym ] that the dependence of the general spherically symmetric fundamental function in tangent bundle coordinates induced by @xmath435 is restricted to @xmath436 where @xmath437 . we wish to study such a spherically symmetric fundamental function that describes a finsler modification of lorentzian geometry . for simplicity , we consider a bimetric four - homogeneous finsler spacetime that perturbs the maximally symmetric vacuum solution of finsler gravity which is given by minkowski spacetime . we assume @xmath438 with @xmath439 . this ansatz has the explicit form @xmath440y^{t^2}+[1+b(r)]y^{r^2}+[1+c(r)]r^2w^2\big).\ ] ] observe that the function @xmath441 can not be transformed away by defining a new radial coordinate . although this could remove @xmath441 from the metric in the right hand bracket , such a coordinate change would generate extra terms in the metric appearing in the left hand bracket . therefore , the existence of the function @xmath441 as a physical degree of freedom is a finsler geometric effect that appears as a consequence of the bimetric spacetime structure assumed here . we will now solve the linearised finsler gravity equation ( [ eq : line ] ) for @xmath442 , @xmath443 and @xmath441 with the ansatz ( [ eq : bisphl ] ) . sorting the equation with respect to powers in @xmath444 and @xmath445 gives rise to three equations that have to be satisfied : @xmath446 the solution of these equations is @xmath447 we will now study the properties of this specific first order finsler spacetime solution and compare it to the linearized schwarzschild spacetime . we use the linearised expression for the non - linear connection coefficients in ( [ eq : nlinfirst ] ) to analyze the finsler geodesic equation that is derived by extremizing the proper time integral ( [ eq : fact ] ) . for a curve with coordinates @xmath448 this has the form @xmath449 . as usual in spherical symmetry , setting @xmath450 solves one of the four component equations immediately ; the remaining equations are @xmath451 from these equations we find two constants of motion @xmath452 these can be used to deduce the orbit equation from the affine normalization condition that @xmath453 along the finsler geodesic ; we employ ( [ eq : f2 ] ) and write @xmath454 for the sign of @xmath455 to obtain @xmath456 the geodesic equations , the constants of motion and the orbit equation are well suited to compare the bimetric linearised finsler solution with the linearised schwarzschild solution . to see the differences to this solution of einstein gravity we first note that @xmath457 can be absorbed into a redefinition of @xmath458 , hence can be assumed to be zero . second we introduce the schwarzschild radius @xmath459 to redefine @xmath460 , @xmath461 and @xmath462 in terms of dimensionless small constants @xmath463 , @xmath464 and @xmath465 . then the orbit equation becomes @xmath466 in the special case @xmath467 this is precisely the orbit equation in the linearized schwarzschild geometry , see @xcite ; the same limit also applies to the geodesic equations and the constants of motion . the finsler geometric refinements of the metric schwarzschild geometry are encoded in the constants @xmath468 and @xmath465 . these can in principle be fitted to data from solar system experiments . indeed , there are certain observations that can not be fully explained by the schwarzschild solution @xcite , for instance , the fly - by anomaly : for several spacecrafts it has been reported that swing - by manoeuvres lead to a small unexplained velocity increase . this corresponds to a change in the shape of the orbit of the spacecraft . such a change can in principle be modelled by finsler refinements ; the perturbations @xmath463 , @xmath464 and @xmath465 certainly provide possibilities to alter the wideness of the swing - by orbit as compared to that expected from einstein gravity . this can be confirmed by simple numerical calculations , see figure [ fig : flyby ] . _ numerical fly - by solutions of the geodesic equations for linearized schwarzschild geometry ( dashed line ) and the bimetric finsler refinement ( solid line ) with @xmath469 , @xmath470 and @xmath471 . the mass is centred at the origin an has schwarzschild radius @xmath472 . the initial conditions are @xmath473 , @xmath474 , @xmath475 , @xmath476 and @xmath477 for both curves , and @xmath478 is calculated from the respective unit normalization condition @xmath453._,scaledwidth=50.0% ] we have seen that finsler geometries exist that are extremely close to metric geometries . our specific example of a spherically symmetric bimetric perturbation around minkowski spacetime could be reinterpreted as a geometry close to the linearized schwarzschild solution of einstein gravity . the more complex causal structure , however , leads to additional constants that modify the geodesic equations and in particular the shape of test particle orbits . this could be a means to explain the fly - by anomaly in the solar system . we emphasize that this consequence already at first order perturbation theory gives a glimpse on the potential of finsler gravity . finsler geometry is fundamentally based on the reparametrization invariant length integral ( [ eq : fact ] ) . in physics , this integral can be used as a very general clock postulate on the one hand , and on the other as an action for massive point particles which automatically guarantees the very precisely tested weak equivalence principle . in previous work @xcite we formulated a set of minimal requirements for the application of finsler geometry to the description of spacetime . this led us to definition 1 of finsler spacetimes @xmath233 which have sufficient structure to provide notions of causality and are mathematically controlled generalizations of lorentzian geometry . in this article , we constructed an action for finsler gravity from first principles . our theory of finsler gravity ( [ eq : compact ] ) fully includes the description of matter fields which are coupled to finsler spacetime by a lifting principle that generates the appropriate action from the standard lagrangian on lorentzian spacetime . we derived the gravitational field equation by variation with respect to the fundamental geometry function @xmath26 , and could show that it consistently becomes equivalent to the einstein field equations in the metric geometry limit . hence einstein gravity can be seen as special case of our gravity theory based on finsler geometry . we presented the geometric definition 2 of observers on finsler spacetimes along with a clear interpretation of how they measure physical fields . by theorem 2 we were able to characterize the class of transformations that relates two different observers at the same point of the spacetime manifold ; these transformations have the algebraic structure of a groupoid as proven in theorem 3 . any observer transformation can be understood as the composition of a usual lorentz transformation and an identification of the two observers four - velocities by a geometrically well - defined parallel transport . for the limiting case of metric geometries , this parallel transport trivializes so that the transformation groupoid becomes equivalent to the standard lorentz group . in this sense , finsler spacetimes generalize lorentz invariance instead of violating it . as a further formal development we presented definition 4 of symmetries of finsler spacetimes and theorem 4 that shows some of their basic properties . we applied this notion to determine the most general fundamental geometry functions consistent with spherical , cosmological , or maximal symmetry . maximally symmetric finsler spacetimes are in fact maximally symmetric metric spacetimes , and the only maximally symmetric source - free vacuum solution of finsler gravity is minkowski spacetime . as a concrete application of the results of this article we studied a simple spherically symmetric bimetric perturbation around this flat metric vacuum . we found a first order solution of the finsler gravity equation which is a refinement of the linearized schwarzschild solution of einstein gravity . with a special choice of the parameters in our solution the resulting geodesics are identical to those of linearized schwarzschild spacetime . this demonstrates that weak field gravitational experiments may not be sufficient to distinguish finsler spacetimes from lorentzian metric spacetimes . but we saw that the full set of parameters in our model solution could be capable to resolve the fly - by anomaly in the solar system . this is very promising , and creates a strong motivation for more intensive studies of our new theory of gravity . it is natural to ask whether the additional degrees of freedom of finsler spacetime solutions as compared to metric spacetimes may lead to new insights on the dark matter distributions in galaxies or dark energy in the universe . these could be effects of a fundamentally more complex spacetime geometry instead of being particle physics phenomena . it will be possible to study these questions once solutions of finsler gravity , especially for spherical symmetry and cosmology , become available . for cosmology , it will also be necessary to study perfect fluid sources for finsler spacetimes . we saw in @xcite that finsler spacetimes can provide a geometric explanation of the opera measurements of superluminal neutrinos @xcite . the velocity difference between the neutrinos and the speed of light recognized as the boundary velocity of observers depends not only on the energy and mass of the neutrino , but also on the underlying spacetime geometry . so also in this context , solutions for spherical symmetry and cosmology are needed in order to understand the size of the effect for the different observed neutrino sources . further important topics for future research are the coupling of spinor fields , the analysis of field theories on the generalized causal structure of finsler spacetimes and their quantization . cp and mnrw thank niklas hbel , claudio dappiaggi , andreas degner , klaus fredenhagen , manuel hohmann , matthias lange , falk lindner , gunnar preuss and felix tennie for inspiring discussions . they acknowledge full financial support from the german research foundation dfg under grant wo 1447/1 - 1 . this appendix presents technical details for several derivations in the main text . in particular , we show how to perform the variation of the finsler gravity and matter field actions on the unit tangent bundle ; for completeness we state the complete lifts of the cosmological symmetry generators ; and we display some additional material on the linearized finsler gravity equations . the finsler gravitational field equation presented in section [ sec : gravvac ] can be deduced from our new finsler gravity action as follows . before we consider the variation of the matter part with respect to the fundamental geometry function @xmath26 in the next section , we here take a look at the pure gravity action ( [ eq : gravact ] ) : @xmath479=\int d^4\hat xd^3u \left[\sqrt{g^fh^f}\ r^a{}_{ab}y^b\right]_{|\sigma}\,.\ ] ] the integrand is homogeneous of degree five ; to obtain the first intermediate step ( [ eq : calcgrav1 ] ) of the variation we use the facts that for @xmath254 homogeneous of degree @xmath388 holds @xmath480 and that @xmath481 . the second step ( [ eq : calcgrav2 ] ) is obtained by using the coordinate transformation formulae ( [ eq : invvv ] ) and the fact that @xmath482 to calculate @xmath483 which in turn is used to deduce @xmath484r^a{}_{ab}y^b+\delta r^a{}_{ab}\,y^b\big).\ ] ] the formulae ( [ eq : calcgrav ] ) used in the third step of the variation are basically obtained by means of integration by parts ( [ eq : intbp ] ) . for a function @xmath254 that is @xmath388-homogeneous in @xmath278 the following holds @xmath485 choosing @xmath486 which has @xmath487 proves formula ( [ eq : calcgrav31 ] ) . to show equation ( [ eq : calcgrav32 ] ) we first write @xmath488 and use @xmath489d}+2y^ps^a{}_{q[b}\gamma^{\delta q}{}_{c]p}\ ] ] to equate @xmath490q}-\frac{1}{2}s_{c}\delta\gamma^{\delta c}{}_{bq}\big)\nonumber\\ & = & -\nabla_a ( y^by^q\delta \gamma^{\delta a}{}_{bq})+y^b\nabla_b(y^q\delta \gamma^{\delta a}{}_{aq})+s_{c}\delta\gamma^{\delta c}{}_{bq}y^by^q\,.\end{aligned}\ ] ] the integration by parts formulae ( [ eq : intbp ] ) and @xmath491 then yield the desired equation @xmath492 combining these three steps as we did in section [ sec : gravvac ] finally produces the finsler gravity vacuum field equation ( [ eq : gravvac ] ) . in section [ sec : gravmat ] we presented a coupling principle of matter fields to finsler gravity . the crucial steps of the derivation of the constraints ( [ eq : cons ] ) , equations of motion ( [ eq : eom1 ] ) and ( [ eq : eom2 ] ) , and of the metric limit of the complete gravity equation ( [ eq : fgrav ] ) including the matter source terms shall be presented here . recall the matter action for a @xmath281-form field @xmath297 on finsler spacetime arises from a lift of the standard @xmath281-form action on lorentzian spacetime as @xmath493=\int_\sigma d^4xd^3u\ \big[\sqrt{g^fh^f } \big(\mathcal{l}(g,\phi , \mathrm{d}\phi)+\lambda ( 1-p^h)\phi\big ) \big]_{|\sigma}\,.\ ] ] in order to perform the variation we consider all objects in the horizontal / vertical basis of @xmath17 where @xmath494 is diagonal , see ( [ eq : sasaki ] ) . in the following the @xmath495 label both horizontal and vertical indices , @xmath496 label vertical indices , and @xmath497 label horizontal indices . then @xmath498 and the variation of this lagrangian can now be written as follows @xmath499 we can immediately read off the variation with respect to the lagrange multiplier components which produces ( [ eq : cons ] ) . hence the lagrange multiplier @xmath298 sets to zero all components of @xmath151 with at least one vertical index , so that only purely horizontal components remain on - shell . the expansion of @xmath500 in components with respect to the horizontal / vertical basis yields @xmath501 } - \frac{p(p+1)}{2}\gamma^q{}_{[nm_1}\phi_{|q|m_2 ... m_p]}\,,\ ] ] where we write @xmath502 , and @xmath503 denote the commutator coefficients of the horizontal / vertical basis . their only non - vanishing components are given by @xmath504^{\bar a}=r^{\bar a}{}_{bc}$ ] and @xmath505^{\bar a}=\bar\partial_b n^{\bar a}{}_c$ ] . one now uses the integration by parts formulae ( [ eq : intbp ] ) to obtain the variation of the matter action with respect to @xmath151 ; this produces the equations of motion ( [ eq : eom1 ] ) and ( [ eq : eom2 ] ) . finally the source term for the gravity field equation is obtained by variation of the matter action @xmath322 in ( [ eq : actm ] ) with respect to the fundamental geometry function @xmath26 . this not only includes the variation ( [ eq : vari ] ) but also that of the volume element which can be read off from ( [ eq : volu ] ) . we will now show that the metric limit of finsler gravity plus matter is consistent ; this can be done on - shell where we may use the lagrange multiplier constraints to set all explicitly appearing @xmath506 to zero . then the variation of @xmath322 with respect to @xmath26 becomes @xmath507&=&\int_\sigma d^4xd^3u\ \sqrt{g^fh^f}_{|\sigma } \big[\big(g^{fab}\bar\partial_a\bar\partial_b\mathcal{l}+4\mathcal{l}\big)\frac{\delta l}{nl}\nonumber\\ & & \qquad\qquad\qquad\qquad\qquad+\frac{\partial \mathcal{l}}{\partial g_{mn}}\delta g_{mn}+\frac{\partial \mathcal{l}}{\partial ( \mathrm{d}\phi_{nm_1 ... m_k})}\delta ( \mathrm{d}\phi_{nm_1m_k})\big]_{|\sigma}\nonumber\\ & = & \int_\sigma d^4xd^3u\ \sqrt{g^fh^f}_{|\sigma } \big[\big(g^{fab}\bar\partial_a\bar\partial_b\mathcal{l}+4\mathcal{l}\big)\frac{\delta l}{nl}\\ & & \qquad\qquad\qquad\qquad\qquad+\frac{\partial \mathcal{l}}{\partial g^f_{ab}}\delta g^f_{ab}+\frac{\partial \mathcal{l}}{\partial g^f_{\bar a\bar b}}\delta\big ( \frac{g^f_{\bar a \bar b}}{f^2}\big)+\frac{\partial \mathcal{l}}{\partial ( \mathrm{d}\phi_{ba_1 ... a_k})}\delta ( \delta_{[b}\phi_{a_1 .. a_k]})\big]_{|\sigma}\,.\nonumber\end{aligned}\ ] ] in order to determine the energy momentum scalar @xmath326 defined in ( [ eq : ems ] ) on a generic finsler spacetime one has to calculate all terms in the expression above carefully . however , in the metric geometry limit the last two terms vanish . indeed , @xmath508 is always composed from terms with vertical indices that must be either of the type @xmath509 or contain components of @xmath151 with at least one vertical index ; the last term is proportional to @xmath510 ; in the metric limit @xmath511 vanishes and the vertical index components of @xmath151 are zero on - shell . therefore the remaining terms that are relevant in the metric limit are @xmath512 \rightarrow \int_\sigma d^4xd^3u\ \sqrt{g^fh^f}_{|\sigma } \big[\big(g^{fab}\bar\partial_a\bar\partial_b\mathcal{l}+4\mathcal{l}\big)\frac{\delta l}{nl}+\frac{\partial \mathcal{l}}{\partial g^f_{ab}}\delta g^f_{ab}\big]_{|\sigma}\,.\ ] ] the rewriting @xmath513 and subsequent integration by parts yields @xmath514_{|\sigma}\\ & = & \int_\sigma d^4xd^3u\ \sqrt{g^fh^f}_{|\sigma } \big[-\bar\partial_c k^c+\big(-g^{fij}\bar\partial_cg^f_{ij}+4g^f_{ic}y^i\big)k^c\big]_{|\sigma}\frac{\delta l}{nl}\nonumber\,,\end{aligned}\ ] ] with @xmath515 applying the metric limit now means to consider @xmath23 with the consequence that @xmath516 for timelike @xmath278 . the expression for @xmath517 reduces to @xmath518 and @xmath519 . collecting all terms in the variation of the matter action in the metric geometry limit finally yields @xmath512 \rightarrow \int_\sigma d^4xd^3u\ \sqrt{g^fh^f}_{|\sigma } \big[4\mathcal{l}-4g_{cd}\frac{\partial \mathcal{l}}{\partial g_{cd}}-24y^cy^d\frac{\partial \mathcal{l}}{\partial g_{cd}}\big]_{|\sigma}\frac{\delta l}{nl}\,,\ ] ] from which we can read off the expression for the source term @xmath326 , @xmath520 the lift of this expression to @xmath7 requires making all terms zero homogeneous by multiplication with the appropriate powers of @xmath521 , which here means multiplication of the third term by @xmath522 . the result confirms equation ( [ eq : tsig ] ) that was used to prove the consistency of finsler gravity with einstein gravity in the metric geometry limit . we deduced the most general fundamental geometry function @xmath26 for finsler spacetimes with cosmological symmetries in section [ sec : fsym ] . the derivation requires the complete lifts of the symmetry - generating vector fields ( [ eq : csym ] ) which we display here explicitly : @xmath523 @xmath524 @xmath525 the complete lifts @xmath526 , @xmath527 and @xmath528 are stated in equations ( [ eq : ssymc ] ) . in the formulae above we use the abbreviation @xmath387 and primes denote differentiation with respect to the coordinate @xmath529 . in order to study the finsler gravitational field equation perturbatively we have considered a class of finsler spacetimes that are mild deviations from metric geometry in section [ subsec : modlgeom ] . here we list for completeness how to rewrite the appearing geometric objects in terms of the perturbation @xmath433 instead of the variable @xmath434 used in the main text . these identities can now be employed to determine the curvature scalar @xmath250 of the cartan non - linear connection which we use as the basic ingredient in our construction of finsler gravity : @xmath531-\frac{g^{1-k}}{2k}g^{ag}\big(\nabla_a\nabla \bar\partial_q h-\nabla_a\nabla_q h-\frac{1}{2}\nabla\nabla \bar\partial_a\bar\partial_q h\big)\\ & & { } -\frac{2(1-k)g^{-k}}{2k}g^{aq}\big(g_{qi}y^i\nabla_a\nabla h-\frac{1}{2}\nabla\nabla\big(g_{qi}y^i\bar\partial_a h+g_{ai}y^i\bar\partial_qh+(g_{aq}-\frac{2k}{g}g_{ai}y^ig_{qj}y^j ) h\big)\big)\nonumber\\ & = & -y^ay^br_{ab}[g(x)]-\frac{g^{1-k}}{2k}g^{ag}\big(\nabla_a\nabla \bar\partial_q h-\nabla_a\nabla_q h-\frac{1}{2}\nabla\nabla \bar\partial_a\bar\partial_q h\big)+\frac{(1-k)(1+k)g^{-k}}{k}\nabla\nabla h\,.\nonumber\end{aligned}\ ] ] c. m. will , living rev . rel . * 9 * ( 2005 ) 3 [ gr - qc/0510072 ] . s. baessler , b. r. heckel , e. g. adelberger , j. h. gundlach , u. schmidt and h. e. swanson , phys . * 83 * ( 1999 ) 3585 . d. bao , s.s . chern and z. shen , _ an introduction to riemann - finsler geometry _ , springer new york 2000 . z. chang and x. li , phys . b * 668 * ( 2008 ) 453 [ arxiv:0806.2184 [ gr - qc ] ] . z. chang and x. li , phys . b * 676 * ( 2009 ) 173 [ arxiv:0901.1023 [ gr - qc ] ] . c. pfeifer and m. n. r. wohlfarth , arxiv:1109.6005 [ gr - qc ] . c. lammerzahl , o. preuss and h. dittus , gr - qc/0604052 . j. c. baez , a. e. hoffnung and c. d. walker , arxiv:0812.4864 [ math.qa ] . x. li and z. chang , arxiv:1010.2020 [ gr - qc ] . k. yano and s. ishihara , tangent and cotangent bundles , marcel dekker , new york 1973 .

since the appearance of the first high - energy colliders in the 1980 s , the study of `` jets '' of particles produced in the final state has proved to be a powerful tool for probing the underlying elementary dynamics of the strong force as described by quantum chromodynamics ( qcd ) . jets have been extensively studied at @xmath2 ( lep , slc ) , @xmath3 ( hera ) and @xmath4/@xmath5 ( tevatron , rhic , lhc ) colliders , with a wide variety of jet algorithms , including infrared and collinear ( irc ) safe algorithms such as those of refs . @xcite . a topic of current interest is the use of jets in heavy - ion ( hi ) collisions , where , for example , they can be used to probe the hot , dense medium . in the past few years , the relativistic heavy ion collider ( rhic ) has amassed significant numbers of copper - copper and gold - gold collisions at nucleon - nucleon centre of mass energies up to @xmath6 , and the large hadron collider ( lhc ) should deliver high yields of lead - lead collisions at much higher energies in the near future . the main obstacle to studying jets in hi collisions is the presence of the huge background given by the underlying event ( ue ) produced simultaneously with the hard nucleon - nucleon collision that initiates the high - transverse momentum jet of interest . this ue needs to be properly subtracted from the momentum of a given jet in order to reconstruct its `` true '' momentum , _ i.e._the one it would have in the absence of the ue contribution . this problem is of course well known , and is also present , though to a much smaller extent , in jet studies in proton - proton collisions . various approaches to address it have been proposed ( see _ e.g._@xcite ) . in ref . @xcite two of us proposed a jet area - related technique to determine the transverse momentum density of a sufficiently uniformly distributed background and to subtract it from the jet momenta . the method of @xcite introduced several novel steps , such as the measurement of jet areas , and procedures to determine the transverse - momentum density of the underlying event and/or pileup . subsequent work of ours @xcite has sought to provide firmer foundations for these concepts and methods , as well as practical tests in @xmath5 jet reconstruction tasks with simulated events @xcite . preliminary experimental jet measurements from the star collaboration at rhic , whose analysis is partially based on the ideas of @xcite , have been presented in ref . @xcite . in this article we give a systematic examination of the performance of such methods for heavy - ion collisions , applying them to monte carlo simulations for rhic and the lhc . in particular , we determine the accuracy with which jet momenta can be effectively reconstructed , comparing the performance of several different jet algorithms . jets in hi collisions are produced in an environment that is far from conducive to their detection and accurate measurement . monte carlo simulations ( and real rhic data ) for gold - gold collisions at @xmath7 gev ( per nucleon - nucleon collision ) show that the transverse momentum density @xmath8 of final - state particles is about 100 gev per unit area ( in the rapidity - azimuth plane ) . for lead - lead collisions at @xmath9 tev at the lhc this figure is expected to increase by some factor @xmath10 . this means that jets returned by jet definitions with a radius parameter of , _ e.g. _ , @xmath11 , will contain background contamination of the order of @xmath12 50 and @xmath13 respectively . a related , and perhaps more challenging obstacle to accurate jet reconstruction is due to the fluctuations both of the background level ( from event to event , but also from point to point in a single event ) and of the jet area itself : knowing the contamination level ` on average ' is not sufficient to accurately reconstruct each individual jet . a wide range of strategies have already been advocated to reconstruct jets in hi collisions . just to mention them briefly , approaches to hi jet reconstruction may involve one or more of the following measures : choosing a jet algorithm returning jets with an area as constant as possible ; eliminating from the clustering all particles below a given transverse momentum threshold , say @xmath14 gev ; measuring the background level in a region of the detector thought to be not affected by the hard event ; or parametrising the average background level ( and fluctuations ) in terms of some other measured properties of the event , such as its centrality . our framework has the following characteristics : * we shall restrict ourselves to irc safe jet algorithms , to ensure that measurements can be meaningfully compared to higher - order perturbative qcd calculations . * the jet algorithms we use will not be limited to those yielding jets of regular shape . * our analysis will avoid excluding small transverse momentum particles from the clustering . doing so is collinear - unsafe and inevitably biases the reconstructed jet momenta , which must then be corrected using monte carlo simulations , which can have substantial uncertainties in their modelling of jet quenching and energy loss . instead , we shall try to achieve a bias - free reconstructed jet , working with all the particles in the event . cutoffs of their own . these detector artefacts should not have too large an impact on collinear safety as long as they appear at momenta of the order of the hadronisation scale of qcd , _ i.e._a few hundred mev . ] * the background level will be determined through adaptations of the jet - area / median procedure suggested in @xcite and analysed in detail in @xcite . this procedure is designed to give an estimate of the background that is minimally affected by the presence of hard jets . in order to test the effectiveness of subtracting the underlying event in a heavy - ion collision , and determine the quality of the jet reconstruction , we need access to the idealised hard jets , without the background , as a reference . this can be done by considering a simulated hard @xmath5 event first in isolation , and then embedded in a heavy - ion event . for clarity in the discussions , we shall adopt the following terminology : * the _ hard event _ refers to the hard @xmath5 event alone , without a heavy - ion background ; * the _ full event _ refers to combination of the hard event and the hi background ( which possibly includes many semi - hard events of its own ) ; * _ hard jets _ and _ full jets _ refers to jets from the hard and full events , respectively . note that though making the distinction between a hard and a full event is not possible with real data , it is feasible in certain monte carlo studies . one should nevertheless be aware that the extent to which such a distinction is physically meaningful is a question that remains open in view of the numerous issues related to the interaction between an energetic parton and the medium through which it travels . if the monte carlo program used to simulate heavy - ion events explicitly provides a separation between hard events and a soft part _ e.g._as does hydjet @xcite one can extract one of the former as the single hard event and take the complete event as the full one . alternatively , one can generate a hard event independently and embed it in a heavy - ion event ( obtained from a monte carlo or from real collisions ) to obtain the full event . this second approach , which we have adopted for the bulk of results presented here because of its greater computational efficiency , is sensible as long as the embedded hard event is much harder than any of the semi - hard events that tend to be present in the background . for the transverse momenta that we consider , this condition is typically fulfilled . note however that for studies in which the presence of a hard collision is not guaranteed , _ e.g._the evaluation of fake - jet rates , one should use the first approach . to measure the performance of our heavy - ion background subtraction procedure we apply it to both hard and full events . we then take the two hardest resulting jets in each hard event and compare them to the corresponding subtracted jets in the full event . to do this in practice , we need a method to match the jets obtained from the full event to those from the hard event alone , to make sure that we are comparing the same object . in other analyses , this matching was typically performed in terms of the position of the jet in the rapidity - azimuth plane , requiring that the two jet axes not differ by more than a given @xmath15 . the exact value of @xmath16 is an arbitrary choice , and values between 0.1 and 0.3 have been often used . in order to avoid the arbitrariness of the @xmath16 choice , we propose a different prescription : a jet reconstructed in the full event is considered matched to a hard jet _ if the constituents common to both the hard and the full jet make up at least 50% of the transverse momentum of the constituents of the hard jet_. of the @xmath1 of the hard jet after ue subtraction as in section [ sec : subtraction ] . in practice , this detail has negligible impact on matching efficiencies and other results . ] this definition has the advantage that for a given hard jet , at most one full jet can satisfy this criterion , therefore avoiding having to deal with multiple positive matchings . another advantage is that it automatically rejects fake matchings given by a soft jet that happens to be close to a hard one . for a matched pair of jets , we shall use the notation @xmath17 and @xmath18 for the transverse momentum of the jet in the hard and full event respectively , and @xmath19 and @xmath20 for their subtracted equivalents . the quantity that we shall mostly concentrate on is the @xmath1 offset , @xmath21 between the momentum of the background - subtracted jet in the full event and its equivalent in the hard event . large deviations from zero will indicate a poor subtraction . events without any matched jets are simply discarded and instead contribute to the evaluation of matching inefficiencies ( see section [ sec : efficiency ] ) . the average of the @xmath1 shift over many events , @xmath22 , is only one measure that one may examine to establish the quality of jet reconstruction ; its dispersion , @xmath23 can be another important one . this is especially true in the case of steeply - falling jet spectra , where this dispersion can have a large impact on the measured cross section , necessitating delicate deconvolutions , also known as `` unfolding '' . in this paper we shall concentrate on these two quality measures @xmath24 and @xmath25 , keeping only the pair of jets that have been matched to one of the two hardest ( subtracted ) jets in the hard event . small ( absolute ) values of both @xmath24 and @xmath25 will be the sign of a good subtraction . in practice a trade - off may exist in offset versus dispersion : which one to optimise may depend on the specific observable one wants to measure . and @xmath25 may not , in general , fully characterise the quality of the reconstruction , as the distribution of @xmath26 may be non - gaussian . ] quantifying the quality of background subtraction using monte carlo simulations has several advantages . besides providing a practical way of generating the hard `` signal '' separately from the soft background , one can easily check the robustness of one s conclusions by changing the hard jet or the background sample . one difficulty that arises in gauging the quality of jet reconstruction in heavy - ion collisions comes from the expectation that parton fragmentation in a hot medium will differ from that in a vacuum . this difference is often referred to as _ jet quenching _ @xcite ( for reviews , see @xcite ) . the details of jet quenching are far less well established than those of vacuum fragmentation and can have an effect on the quality of jet reconstruction . here we shall examine the reconstruction of both unquenched and quenched jets . for the latter it will be particularly important to be able to test more than one quenching model , in order to help build confidence in our conclusions about any reconstruction bias that may additionally exist in the presence of quenching . in practice , for this paper we have used both the fortran ( v1.6 ) @xcite and the c++ ( v2.1 ) @xcite versions of hydjet to generate the background . hard jets have been generated with pythia 6.4 @xcite , either running it standalone , or using the version embedded in hydjet v1.6 . the quenching effects have been studied using both qpythia @xcite and pyquen @xcite . as mentioned in the introduction , the last few years have seen many developments in the field of jet clustering , with the appearance of fast implementations @xcite of the @xmath0 @xcite and the cambridge / aachen @xcite sequential recombination algorithms , and the introduction of the siscone @xcite and anti-@xmath0 @xcite algorithms . besides these four main infrared - and - collinear - safe jet algorithms , a number of recent papers ( see _ e.g._@xcite ) , have introduced and studied jet - cleaning techniques in order to help the reconstruction of jets by reducing the ue contamination . `` filtering '' @xcite works by reclustering each jet with a radius @xmath27 smaller than the original radius @xmath28 and keeping only the @xmath29 hardest subjets ( background subtraction is applied to each of the subjets before deciding which ones to keep ) . other similar techniques , `` trimming '' and `` pruning '' , that exploit the substructure of jets have also been introduced recently @xcite and , in @xmath5 environments , have been found to give benefits comparable to those of filtering . note that the filtering that we use here is unrelated to the gaussian filter approach of @xcite ( not used in this paper as no public code is currently available ) . in analogy with the systematic analysis of the performance of various jet definitions in the kinematic reconstruction of dijet systems at the lhc @xcite , here we will study how different algorithms behave in the case of hi collisions . we will use the @xmath0 , cambridge / aachen ( c / a ) , and anti-@xmath0 algorithms , as well as filtering applied to cambridge / aachen ( c / a(filt ) ) with @xmath30 and @xmath31 . we have not used siscone , as its relatively slower speed compared to the sequential recombination algorithms makes it less suitable for a hi environment . all the algorithms have been used through their fast implementation available in the package @xcite , version 2.4.2 . additionally some features of a forthcoming fastjet release have been used to help simplify our analyses . , the _ circular _ range @xmath32 and the _ doughnut _ range @xmath33 . the last three are local ranges with a position depending on the jet being subtracted . see the text for detailed definitions.,scaledwidth=80.0% ] in order to subtract the hi background from the hard jets , we will mainly follow the method introduced in @xcite . when clustering the event we determine the 4-vector active area @xmath34 @xcite of each jet @xmath35 , as well as an estimate of the background density @xmath8 . then for each jet @xmath35 , we subtract from its four - momentum the expected background contamination : @xmath36 the background transverse - momentum density per unit area , @xmath8 , is determined , event - by - event , as proposed in @xcite . in that paper two ways of determining @xmath8 were explored . one was to take a global background - estimation range @xmath37 , covering the full rapidity - azimuth plane up to some rapidity @xmath38 ; one then considered the set of ratios of transverse components of the momentum and area 4-vectors , @xmath39 , for the jets within this range . the median of this set was used as an estimate for @xmath8 : @xmath40 where the subscript @xmath37 explicitly denotes that the background density has been calculated using only jets in the range @xmath37 . this method assumes that the background level is sufficiently constant within @xmath37 . this condition is known to be violated when going to sufficiently large rapidity , and the effect is particularly marked in heavy ion collisions . for this reason , it was alternatively proposed in @xcite to fit the mean @xmath39 as a function of rapidity @xmath41 with a quadratic functional form @xmath42 . while working on this paper , and partly inspired by discussions with members of the star collaboration ( c.f . also ref . @xcite ) , we came to the conclusion that neither of these two methods allows one to extract sufficiently accurate values of @xmath8 in the context of hi collisions . we propose therefore a variant of the first method , namely making the background - estimation range _ local _ and dependent on a given jet s position . graphical representations of a global range , as well as of three possible local ranges that we shall use hereafter , are given in fig . [ fig : ranges ] . more specifically , for a given jet @xmath35 , our three local ranges are defined as follows , * the _ strip _ range , @xmath43 , includes the jets @xmath44 satisfying @xmath45 , * the _ circular _ range , @xmath46 , includes the jets @xmath44 satisfying @xmath47 , * the _ doughnut _ range , @xmath33 , includes the jets @xmath44 satisfying @xmath48 . using @xmath49 to denote any local range around jet @xmath35 , the background density will then depend on the jet being subtracted and will be estimated using @xmath50 it is important to be aware that the estimate of the background density @xmath8 is just an input to eq . ( [ eq : subtraction ] ) and that the jet definition that is used for the computation of @xmath8 in eqs . ( [ eq : median ] ) or ( [ eq : local_median ] ) can be different from the jet definition used to obtain the `` physical '' jets and subtract them in eq . ( [ eq : subtraction ] ) . the only two recommended jet algorithms for the task of determining @xmath8 are the @xmath0 and c / a algorithms ( see @xcite ) ; others are not ideal because they return many jets with very small areas , which distorts the median procedure . the radius , @xmath51 , used in the jet definition for determining @xmath8 can also differ from that , @xmath28 , used to find the jets . when using a local range , the underlying idea is to limit the sensitivity to the long - range variations of the background density , by using only the jets in the vicinity of the jet we want to subtract . in practice , a compromise needs to be found between choosing a range small enough to get a valid local estimation , but also large enough to contain a sufficiently large number of background ( soft ) jets for the estimation of the median to be reliable . two effects need to be considered : ( 1 ) statistical fluctuations in the estimation of the background and ( 2 ) biases due to the presence of a hard jet in the region used to estimate the background . 1 . if we require that the dispersion in the reconstructed jet @xmath1 coming from the statistical fluctuations in the estimation of the background does not amount to more than a fraction @xmath52 of the overall @xmath25 , then as discussed in appendix [ app : minrange - fluct ] the range needs to cover an area @xmath53 such that @xmath54 taking @xmath55 and @xmath11 this corresponds to an area @xmath56 . for the applications below , when clustering with radius @xmath28 , we have used the rapidity - strip ranges @xmath57 and @xmath58 , the circular range @xmath59 and the doughnut range @xmath60 . one can check that their areas are compatible with the @xmath61 lower - limit estimated above . the bias in the estimate of @xmath8 due to the presence of @xmath62 hard jets in the range is given roughly by @xmath63 as discussed in @xcite and appendix [ app : minrange - bias ] . the ensuing bias on the @xmath1 can be estimated as @xmath64 ( for anti-@xmath0 jets ) . for @xmath11 , @xmath65 , @xmath66 and @xmath67 , the bias in the reconstructed jet @xmath1 is @xmath68 . given @xmath69 in the @xmath70 range ( as we will find in section [ sec : results ] ) this corresponds to a @xmath71 bias . in order to eliminate this small bias , we will often choose to exclude the two hardest jets in each event when determining of @xmath8 . a third potential bias discussed in ref . @xcite is that of underestimating the background when using too small a value for @xmath51 . given the high density of particles in hi collisions , this will generally not be an issue as long as @xmath72 . distribution of the background density @xmath8 per unit area ( left ) and its intra - event fluctuations @xmath69 ( right ) . it has been obtained from 5000 hydjet events with rhic ( auau , @xmath73 ) and lhc ( pbpb , @xmath74 ) kinematics . the background properties have been estimated using the techniques presented in section [ sec : subtraction ] , using the @xmath0 algorithm with @xmath75 , and keeping only the jets with @xmath76 ( excluding the two hardest).,title="fig:",scaledwidth=50.0% ] distribution of the background density @xmath8 per unit area ( left ) and its intra - event fluctuations @xmath69 ( right ) . it has been obtained from 5000 hydjet events with rhic ( auau , @xmath73 ) and lhc ( pbpb , @xmath74 ) kinematics . the background properties have been estimated using the techniques presented in section [ sec : subtraction ] , using the @xmath0 algorithm with @xmath75 , and keeping only the jets with @xmath76 ( excluding the two hardest).,title="fig:",scaledwidth=50.0% ] as mentioned already in section [ sec : mcs ] , for the simulation of the events , we have used pythia 6.4 @xcite to generate the unquenched hard jets ; the background events are generated using hydjet v1.6 @xcite with 0 - 10% centrality , @xmath77 , @xmath78 and ` ptmin`@xmath79 ; for lhc , @xmath80 , @xmath81 , @xmath82 and @xmath83 . in both cases quenching effects are turned on in hydjet , @xmath84 , even when they are not included for the embedded @xmath5 event . the corresponding pyquen parameters we have used are @xmath85 , @xmath86 , @xmath87 , @xmath88 fm and @xmath89 . [ foot : pyquen ] ] . our setup for the background leads , for rhic ( auau , @xmath90 ) , to an average background density per unit area at central rapidity of @xmath91 gev with average fluctuations in a single event of @xmath92 gev and event - to - event fluctuations @xmath93 gev . for the lhc ( pbpb , @xmath74 ) the corresponding values are @xmath94 gev , @xmath95 gev and event - to - event fluctuations @xmath96 gev . [ fig : bkg_distrib ] shows the distributions obtained from the simulations for @xmath8 and @xmath69 . in the case of the rhic simulation , the result for @xmath97 is somewhat higher than the experimental value of @xmath98 quoted by the star collaboration @xcite , however this is probably in part due to limited tracking efficiencies at low @xmath1 at star , and explicit star @xcite and phenix @xcite results for @xmath99 correspond to somewhat higher @xmath8 values , about @xmath100 . the multiplicity of charged particles ( @xmath101 for @xmath102 and 06% centrality ) and the pion @xmath1 spectrum in our simulation are sensible compared to experimental measurements at rhic @xcite . for the lhc our charged particle multiplicity is @xmath103 for @xmath102 and 010% centrality , which is comparable to many of the predictions reviewed in fig . 7 of @xcite . an independent control analysis has also been performed with hydjet++ 2.1 @xcite ( with default parameters ) for the background . the results at rhic are similar , while for lhc the comparison is difficult because the default tune of hydjet++ 2.1 predicts a much higher multiplicity , @xmath104 for @xmath102 and 010% centrality . most of the results of this section will be obtained without quenching , though in section [ sec : quenching ] we will also consider the impact on our conclusions of the pyquen @xcite and qpythia @xcite simulations of quenching effects . for the results presented below , we have employed a selection cut of @xmath105 on the jets with @xmath106 for rhic and @xmath107 for the lhc . in that case , in order for information to be derived from jets near the edge of the acceptance and thus bring the available area close to the ideal requirements set out in appendix [ sec : minimal - range ] . note that we also use particles beyond @xmath38 in the jet clustering and apply the acceptance cut only to the resulting jets . ] we only consider full jets that are matched to one of the two hardest jets in the hard event . the computation of the jet areas in , needed both for the subtraction and the background estimation , has been performed using active areas , with ghosts up to @xmath108 , a single repetition and a ghost area of 0.01 . the determination of the background density @xmath8 has been performed using the @xmath0 algorithm with @xmath75 . though the estimate of the background depends on @xmath109 @xcite , we have observed that choices between 0.3 and 0.5 lead to very similar results ( _ e.g._differing by at most a few hundred mev at rhic ) . with this setup , we have studied the various ranges presented in section [ sec : subtraction ] , with the jet algorithms from section [ sec : jetdef ] . in all cases , we have taken the radius parameter @xmath11 . we have adopted this value as it is the largest currently used at rhic . note that the effect of the background fluctuations of the jet energy resolution increases linearly with @xmath28 , disfavouring significantly larger choices . on the other hand , too small a choice of @xmath28 may lead to excessive sensitivity to the details of parton fragmentation , hadronisation and detector granularity . matching efficiency for reconstructed jets as a function of the jet @xmath1 . left : rhic , right : lhc . these results are independent of the choice of background - subtraction range in the heavy - ion events , since background subtraction does not enter into the matching criterion . here and in later figures , the label `` unquenched '' refers to the embedded @xmath5 event ; the background is always simulated including quenching . , title="fig:",scaledwidth=50.0% ] matching efficiency for reconstructed jets as a function of the jet @xmath1 . left : rhic , right : lhc . these results are independent of the choice of background - subtraction range in the heavy - ion events , since background subtraction does not enter into the matching criterion . here and in later figures , the label `` unquenched '' refers to the embedded @xmath5 event ; the background is always simulated including quenching . , title="fig:",scaledwidth=50.0% ] let us start the presentation of our results with a brief discussion of the efficiency of reconstructing jets in the medium . as explained in section [ sec : quality ] , the jets in the medium are matched to a `` bare '' hard jet when their common particle content accounts for at least 50% of the latter s transverse momentum . the matching efficiencies we observe depend to some extent on the details of the monte carlo used for the background so our intention is just to illustrate the typical behaviour we observe and highlight that these efficiencies tend to be large . we observe from fig . [ fig : efficiency ] that we successfully match at least 95% of the jets above @xmath110 gev at rhic , and at least 99% of the jets above @xmath111 gev at the lhc . it is also interesting to notice that the anti-@xmath0 algorithm performs best , likely as a consequence of its ` rigidity ' , namely the fact that anti-@xmath0 jets tend to have the same ( circular ) shape , independently of the soft - particles that are present . effect of the choice of range on the average @xmath1 shift , @xmath26 , as defined in eq . ( [ eq : deltapt ] ) . left : rhic , right : lhc . in this figure and those that follow , the yellow band corresponds to 1% of the @xmath1 of the hard jet.,title="fig:",scaledwidth=50.0% ] effect of the choice of range on the average @xmath1 shift , @xmath26 , as defined in eq . ( [ eq : deltapt ] ) . left : rhic , right : lhc . in this figure and those that follow , the yellow band corresponds to 1% of the @xmath1 of the hard jet.,title="fig:",scaledwidth=50.0% ] we now turn to the results concerning the measurement of the background density and the reconstruction of the jet transverse momentum . we first concentrate on the impact of the choice of a local range and/or of the exclusion of the two hardest jets when determining @xmath8 . for the determination of @xmath8 in the hard event , without exclusion of any jets . this ensures that the reference jet @xmath1 is always kept the same . the impact of subtraction in the hard event is in any case small , so the particular choice of range is not critical . ] in fig . [ fig : ptshift_ranges ] we show the average shift @xmath24 for the list of ranges mentioned in section [ sec : subtraction ] . the results presented here have been obtained with the anti-@xmath0 algorithm with @xmath11 , but the differences among the various range choices have been seen to be similar with other jet definitions . the label `` 2 excl '' means that the two hardest jets in the event have been excluded from the estimation of the background . we have found that this improves the precision of the subtractions whenever expected , _ i.e._for all choices of range except the doughnut range , where its central hole already acts similarly to the exclusion of the hardest jets . to keep the figure reasonably readable , we have only explicitly shown the effect of removing the two hardest jets for the global range . the change of 0.40.6 gev ( both for rhic and the lhc ) is in reasonable agreement with the analytic estimate of about 0.6 gev for rhic and the lhc obtained from @xmath112 with @xmath113 calculated using eq . ( [ eq : hard_median_offset ] ) . note that at lhc the exclusion of the two hardest jets for the global range appears to worsen the subtraction , however what is really happening is that the removal of the two hardest jets exacerbates a deficiency of the global range , namely the fact that its broad rapidity coverage causes it to underestimate @xmath8 , leading to a positive net @xmath113 . other features that can be understood qualitatively include for example the differences between the two strip and the global ( 2 excl ) range for rhic : while the rapidity width of the global range lies in between that of the two strip ranges , the global range gives a lower @xmath24 than both , corresponding to a larger @xmath8 estimate , which is reasonable because the global range is centred on @xmath114 , whereas the strip ranges are mostly centred at larger rapidities where the background is lower . the main result of the analysis of fig . [ fig : ptshift_ranges ] is the observation that all choices of a local range lead to a small residual @xmath26 offset : the background subtraction typically leaves a @xmath115 at both rhic and lhc , _ i.e._better than 1 - 2% accuracy over much of the @xmath1 range of interest . it is not clear , within this level of accuracy , if one range is to be preferred to another , nor is it always easy to identify the precise origins of the observed differences between various ranges . another way of viewing this is that the observed differences between the various choices give an estimate of the residual subtraction error due to possible misestimation of @xmath8 . for our particular analysis , at rhic this comment also applies to the choice of the global range ( with the exclusion of the two hardest jets in the event ) . this is a consequence of the limited rapidity acceptance , which effectively turns the global range into a local one , a situation that does not hold for larger rapidity acceptances , as we have seen for the lhc results . in what follows we will use the doughnut(@xmath116 ) choice , since it provides a good compromise between simplicity and effectiveness . the next potential systematic effect that we consider is the choice of the jet algorithm used for the clustering . algorithm with @xmath75 is used for the estimation of the background . [ fig : deltapt - dist ] shows the distribution of @xmath26 for each of our four choices of jet algorithm , @xmath0 , c / a , anti-@xmath0 and c / a(filt ) , given for rhic collisions and a specific bin of the hard jets transverse momenta , @xmath117 . one sees significant differences between the different algorithms . one also observes that gaussians with mean and dispersion set equal to @xmath24 and @xmath25 provide a fair description of the full histograms . this validates our decision to concentrate on @xmath24 and @xmath25 as quality measures . one should nevertheless be aware that in the region of high @xmath118 there are deviations from perfect gaussianity , which are more visible if one replicates fig . [ fig : deltapt - dist ] with a logarithmic vertical scale ( not shown , for brevity ) . average shift @xmath24 , as a function of @xmath119 , shown for rhic ( left ) and the lhc ( right).,title="fig:",scaledwidth=50.0% ] average shift @xmath24 , as a function of @xmath119 , shown for rhic ( left ) and the lhc ( right).,title="fig:",scaledwidth=50.0% ] the first observable we analyse is the average @xmath1 shift . we show in fig . [ fig : ptshift ] the @xmath120 results for the four algorithms listed in section [ sec : jetdef ] , as a function of @xmath119 . we use the doughnut range to estimate the background . the first observation is that , while the anti-@xmath0 and c / a(filt ) algorithms have a small residual @xmath24 , the c / a and @xmath0 algorithms display significant offsets . the reason for the large offsets of @xmath0 and c / a is well understood , related to an effect known as _ back - reaction _ @xcite . this is the fact that the addition of a soft background can alter the clustering of the particles of the hard event : some of the constituents of a jet in the hard event can be gained by or lost from the jet when clustering the event with the additional background of the full event . this happens , of course , on top of the simple background contamination that adds background particles to the hard jet . even if this latter contamination is subtracted exactly , the reconstructed @xmath1 will still differ from that of the original hard jet as a consequence of the back - reaction . contribution to @xmath24 due to back - reaction . note that these results are independent of the range used for estimating @xmath8 in the heavy - ion event . left : rhic , right : lhc.,title="fig:",scaledwidth=50.0% ] contribution to @xmath24 due to back - reaction . note that these results are independent of the range used for estimating @xmath8 in the heavy - ion event . left : rhic , right : lhc.,title="fig:",scaledwidth=50.0% ] the effect of the back - reaction can be studied in detail , since in monte carlo simulations it is possible to identify which hard - event constituents are present in a given jet before and after inclusion of the background particles in the clustering . the average @xmath1 shift due to back - reaction can be seen in fig . [ fig : ptshift_br ] for the different jet algorithms . as expected @xcite , it is largest for @xmath0 , and smallest ( almost zero , in fact ) for anti-@xmath0 . by comparing fig . [ fig : ptshift_br ] and fig . [ fig : ptshift ] one can readily explain the difference between the @xmath24 offsets of the various algorithms in terms of their back - reaction . the rigidity ( and hence small back - reaction ) of the anti-@xmath0 jets manifestly gives almost bias - free reconstructed jets , while the large back - reaction effects of the @xmath0 algorithm and , to a smaller extent , of the c / a algorithm translates into a worse performance in terms of average shift . the @xmath1 dependence of the back - reaction is weak . this is expected based on the interplay between the @xmath121 dependence found in @xcite and the evolution with @xmath1 of the relative fractions of quark and gluon jets . the case of the c / a(filt ) algorithm is more complex : its small net offset , comparable to that of the anti-@xmath0 algorithm , appears to be due to a fortuitous compensation between an under - subtraction of the background and a negative back - reaction . the negative back - reaction is very similar to that of c / a without filtering , while the under - subtraction is related to the fact that the selection of the hardest subjets introduces a bias towards positive fluctuations of the background . this effect is discussed in appendix [ app : filtbias ] , where we obtain the following estimate for the average @xmath1 shift ( specifically for @xmath122 ) : @xmath123 yielding an average bias of 2 gev for rhic and 4.5 gev at the lhc , which are both in good agreement with the differences observed between c / a with and without filtering in fig . [ fig : ptshift ] . note that while the bias in eq . ( [ eq : filt - bias - main - text ] ) is proportional to @xmath69 , the back - reaction bias is instead mainly proportional to @xmath8 @xcite . it is because of these different proportionalities that the cancellation between the two effects should be considered as fortuitous . since it also depends on the substructure of the jet , one may also expect that the cancellation that we see here could break down in the presence of quenching . dispersion @xmath25 . left : rhic , right : lhc.,title="fig:",scaledwidth=50.0% ] dispersion @xmath25 . left : rhic , right : lhc.,title="fig:",scaledwidth=50.0% ] our results for the @xmath26 dispersion , @xmath25 , are shown in fig . [ fig : dispersion ] , again using the doughnut range . ( our conclusions are essentially independent of the particular choice of range . ) we first discuss the case of rhic kinematics . for @xmath0 and anti-@xmath0 , the observed dispersions are similar to the result of @xmath124 quoted by star @xcite ( though the number from star includes detector resolution effects , so that the true physical @xmath25 may actually be somewhat lower ) . of note , the advantage enjoyed by anti-@xmath0 in terms of smallest @xmath24 does not hold at the level of the dispersion : c / a and @xmath0 tend to behave slightly better at small transverse momentum . the algorithm which performs best in terms of dispersion over all the @xmath1 range is now c / a with filtering , for which the result is smaller than that of the other algorithms by a factor of about @xmath125 . this reduction factor can be explained because the dispersion @xmath25 is expected to be proportional to the square - root of the jet area : the c / a(filt ) algorithm with @xmath122 and @xmath31 produces jets with an area of , roughly , half that obtained with c / a ; hence the observed reduction of @xmath25 . in the lhc setup , the conclusions are quite similar at the lowest transverse momenta shown . as @xmath1 increases , the dispersion of the anti-@xmath0 algorithm grows slowly , while that of the others grows more rapidly , so that at the highest @xmath1 s shown , the @xmath0 and c / a algorithms have noticeably larger dispersions than anti-@xmath0 , and c / a(filt ) becomes similar to anti-@xmath0 . the growth of the dispersions can be attributed to an increase of the back - reaction dispersion . the latter is dominated by rare occurrences , where a large fraction of the jet s @xmath1 is gained or lost to back - reaction , hence the noticeable @xmath1 dependence ( _ cf._appendix [ app : dispersion - contributions ] ) . an additional effect , especially for the @xmath0 algorithm , might come from the anomalous dimension of the jet areas , _ i.e._the growth with @xmath1 of the average jet area . note that the dispersion has some limited dependence on the choice of ghost area for example , reducing it from 0.01 to 0.0025 lowers the dispersions by about @xmath126 at rhic . this is discussed further in appendix [ app : ghost - area - choice ] . @xmath1 dependence dependence of the @xmath26 dispersion at rhic , for different centrality classes.,scaledwidth=60.0% ] so far , we have only considered central collisions . since it is known that non - central collisions give rise to elliptic flow @xcite , one might worry that this leads to an extra source of background fluctuations and/or non - uniformities , potentially spoiling the subtraction picture discussed so far . one can study this on azimuthally averaged jet samples ( as we have been doing so far ) or as a function of the azimuthal angle , @xmath127 , between the jet and the reaction plane . as above , we use hydjet v1.6 , whose underlying hydro component includes a simulation of elliptic flow @xcite . we have generated heavy - ion background events for rhic in four different centrality bins : 0 - 10% ( as above ) , 10 - 20% , 20 - 40% and 40 - 75% , with @xmath128 values respectively of @xmath129 , @xmath130 , @xmath131 and @xmath132 . was determined as the average of @xmath133 over all all particles with @xmath134 ( excluding the additional hard @xmath5 event ) . ] we first examine azimuthally averaged results , repeating the studies of the previous sections for each of the centrality bins . we find that the results for the average shift , @xmath24 , are largely independent of centrality , as expected if the elliptic flow effects disappear when averaged over @xmath135 . the results for the dispersion are shown in fig . [ fig : dispersion_centrality ] . we observe that the dispersion decreases with increasing non - centrality . even though one might expect adverse effects from elliptic flow , the heavy - ion background decreases rapidly when one moves from central to peripheral collisions , and this directly translates into a decrease of @xmath25 . the first conclusion from this centrality - dependence study is therefore that the subtraction methods presented in this paper appear to be applicable also for azimuthally averaged observables in non - central collisions . @xmath135 dependence of the @xmath26 shift at rhic , for the @xmath136 centrality bin . left : for three different ranges for the anti-@xmath0 algorithm ; right : for three different jet algorithms for the doughnut(@xmath116 ) range . the absolute size of the @xmath135 dependence is similar for centralities up to @xmath137 and then decreases beyond . , title="fig : " ] @xmath135 dependence of the @xmath26 shift at rhic , for the @xmath136 centrality bin . left : for three different ranges for the anti-@xmath0 algorithm ; right : for three different jet algorithms for the doughnut(@xmath116 ) range . the absolute size of the @xmath135 dependence is similar for centralities up to @xmath137 and then decreases beyond . , title="fig : " ] we next consider results as a function of @xmath138 , which is relevant if one wishes to examine the correlation between jet quenching and the reaction plane . an issue in real experimental studies is the determination of the reaction plane , and the extent to which it is affected by the presence of hard jets . in the hydjet simulations , this problem does not arise because the reaction plane always corresponds to @xmath139 . figure [ fig : ptshift - v - deltaphi ] ( left ) shows the average @xmath26 as a function of @xmath138 for the anti-@xmath0 algorithm and several different background - estimation ranges , for the @xmath136 centrality bin for rhic . the strip range shows significant @xmath138 dependence , which is because a determination of @xmath8 averaged over all @xmath135 can not possibly account for the local @xmath135-dependence induced by the elliptic flow . other ranges , such as the doughnut range , instead cover a more limited region in @xmath135 . they should therefore be able to provide information on the @xmath135-dependence of the background . however , since their extent in @xmath135 tends to be significantly larger than that of the jet , and @xmath8 varies relevantly over that extent , some residual @xmath135 dependence remains in @xmath24 after subtraction . the right - hand plot of figure [ fig : ptshift - v - deltaphi ] shows that the effect is reduced with filtering , as is to be expected since its initial background contamination is smaller . the conclusion from this part of the study is that residual @xmath135-dependent offsets may need to be corrected for explicitly in any studies of jets and their correlations with the reaction plane . the investigation of extensions to our background subtraction procedure to address this issue will be the subject of future work . the last issue we wish to investigate is how the phenomenon of jet quenching ( _ i.e._medium effects on parton fragmentation ) may affect the picture developed so far . the precise nature of jet quenching beyond its basic analytic properties ( see _ e.g._@xcite ) is certainly hard to estimate in detail , especially at the lhc , where experimental data from , say , flow or particle spectra measurements are not yet available for the tuning of the monte carlo simulations . additionally , the implementation of monte - carlo generators that incorporate the analytic features of jet quenching models is currently a very active field @xcite . we may therefore expect a more robust and complete picture of jet quenching in the near future , together with the awaited first pbpbcollisions at the lhc . in this section , we examine the robustness of our hi background subtraction in the presence of ( simulated ) jet quenching . for this purpose we have used two available models which allow one to simulate quenched hard jets , pyquen @xcite , which is used by hydjet v1.6 , and qpythia @xcite . pyquen has been run with the parameters listed in footnote [ foot : pyquen ] for the lhc , and with @xmath140 , @xmath141 fm and @xmath142 for rhic . for qpythia we have tested two options for the values of the transport coefficient and the medium length ( @xmath143 gev@xmath144/fm , @xmath145 fm and @xmath146 gev@xmath144/fm , @xmath147 fm ) , with similar results . no serious attempt has been made to tune the two codes with each other or with the experimental data , beyond what is already suggested by the code defaults : in the absence of strong experimental constraints on the details of the quenching effects , this allows us to verify the robustness of our results for a range of conditions . as in sections [ sec : choice_range ] and [ sec : choice_alg ] , we have embedded the hard pyquen or qpythia events in a hydjet v1.6 background and tested the effectiveness of the background subtraction for different choices of algorithm . for the unquenched case . we will therefore keep employing the doughnut range . ] we shall restrict our attention to the anti-@xmath0 and c / a(filt ) algorithms , as they appear to be the optimal choices from our analysis so far . average @xmath1 shift for background - subtracted jets with the anti-@xmath0 and c / a(filt ) jet algorithms . the dashed lines correspond to unquenched hard jets ( pythia ) and the solid ones to quenched hard jets ( pyquen ) . results are shown for rhic kinematics on the left plot and for the lhc on the right one.,title="fig:",scaledwidth=50.0% ] average @xmath1 shift for background - subtracted jets with the anti-@xmath0 and c / a(filt ) jet algorithms . the dashed lines correspond to unquenched hard jets ( pythia ) and the solid ones to quenched hard jets ( pyquen ) . results are shown for rhic kinematics on the left plot and for the lhc on the right one.,title="fig:",scaledwidth=50.0% ] we have found that the jet - matching efficiencies are still high , with essentially no changes at rhic , and at lhc a doubling of the ( small ) inefficiencies that we saw in fig . [ fig : efficiency](right ) . the dispersions @xmath25 are also not significantly affected within our sample of jet - quenching simulations . we therefore concentrate on the @xmath24 offset , which is plotted in fig . [ fig : ptshift_quenched ] for pyquen . the results are shown for both rhic and the lhc . in the case of rhic , and for the whole @xmath1 range up to about 50 gev , quenching can be seen not to significantly affect the subtraction offset @xmath148 ( within the usual uncertainty related to the choice of range , which was shown in fig . [ fig : ptshift_ranges ] ) . in the lhc case , instead , while the shift obtained using the anti-@xmath0 algorithm is largely similar to the unquenched case , the c / a(filt ) algorithm performance can be seen to deteriorate slightly when quenching is turned on , all the more so at very large transverse momentum . the @xmath0 and the c / a algorithms are not shown for clarity , but they share the behaviour of c / a(filt ) . this deterioration of the quality of the subtraction can be traced back to an increased back - reaction compared to the unquenched jets . anti-@xmath0 jets do not suffer from this effect as a consequence of the usual rigidity of this algorithm . in the case of c / a(filt ) , one should nevertheless emphasise that an error of ( at most ) 10 gev on the reconstruction of a 500 gev jet is still only a 2% effect . this is modest , both relative to the likely experimental precision and to the expected effect of quenching on the overall jet @xmath1 , predicted by pyquen to be at the level of @xmath149 at this @xmath1 . though for brevity we have not explicitly shown them , the results with qpythia are very similar . before closing this section , we reiterate that we have only investigated simple models for quenching and that our results are meant just to give a first estimate of the effects that one might have to deal with in the case of quenched jets . the expected future availability of new `` quenched '' monte carlo programs , together with specific measurements in the early days of heavy - ion collisions at the lhc , will certainly allow one to address this question more extensively . to close this section , we examine the relative importance of the average shift and its dispersion , taking the illustrative example of their impact on the inclusive jet cross - section as a function of @xmath1 . we start from a simple parametrisation of the inclusive - jet @xmath1 spectrum and see how its reconstruction is affected by the average shift and dispersion that remain for @xmath26 after subtraction . let us assume that the true @xmath1 spectrum decays exponentially _ i.e._@xmath150 while this expression does nt have the @xmath151 form that one expects to see , it is far easier to handle analytically , and not too poor an approximation to observed spectra over quite a broad range of @xmath1 . after embedding the hard events in a heavy - ion background and applying subtraction , the resulting spectrum will be the convolution of eq . ( [ eq : hard - spectrum ] ) with the @xmath26 distribution . assuming that the latter is a gaussian of average @xmath24 and dispersion @xmath25 ( _ cf._fig . [ fig : deltapt - dist ] ) , one obtains a reconstructed spectrum @xmath152 thus the average shift gives a bias by a multiplicative factor @xmath153 and the dispersion by a factor @xmath154 . the convolution works in such a way that for a given reconstructed @xmath1 , the most likely original true transverse momentum is : @xmath155 where we have neglected the small impact of subtraction on the @xmath5 jets . to illustrate these effects quantitatively , let us first take the example of rhic , where between 10 and 60 gev , the cross - section is well approximated by eq . ( [ eq : hard - spectrum ] ) with @xmath156 gev@xmath157 . both the anti-@xmath0 and c / a(filt ) have @xmath158 , leaving only the dispersion effect . in the case of the anti-@xmath0 ( respectively c / a(filt ) ) algorithm , we see from fig . [ fig : dispersion ] that @xmath159 ( @xmath160 ) , which gives a multiplicative factor of about 12 ( 3 ) . for a given reconstructed @xmath1 , the most likely true @xmath1 is about @xmath161 ( @xmath162 ) smaller . in comparison , for c / a ( @xmath0 ) , with @xmath163 ( @xmath164 ) ( fig . [ fig : ptshift ] ) and @xmath165 ( similar for @xmath0 ) there is a partial compensation between factors of @xmath166 ( @xmath167 ) and @xmath168 coming respectively from the shift and dispersion , yielding an overall factor of about @xmath169 ( @xmath170 ) , while the most likely true @xmath1 is about @xmath171 ( @xmath172 ) smaller than the reconstructed @xmath1 . at the lhc ( @xmath74 ) , ( [ eq : hard - spectrum ] ) is a less accurate approximation . nevertheless , for @xmath173 , it is not too unreasonable to take @xmath174 and examine the consequences . for anti-@xmath0 ( respectively c / a(filt ) ) , we have @xmath175 ( also for c / a(filt ) ) and @xmath176 ( @xmath177 ) , giving a multiplicative factor of @xmath178 ( @xmath179 ) , i.e. far smaller corrections than at rhic . for a given reconstructed @xmath1 , the most likely true @xmath1 is about @xmath180 ( @xmath181 ) smaller , rather similar to the values we found at rhic ( though smaller in relative terms , since the @xmath1 s are higher ) , with the increase in @xmath69 being compensated by the decrease in @xmath182 . in the lhc case , it is also worth commenting on the results for the @xmath0 algorithm , since this is what was used in ref . @xcite : we have @xmath183 and @xmath176 , giving a multiplicative factor of @xmath184 , which is consistent with the near perfect agreement that was seen there between the @xmath5 and subtracted @xmath185 spectra . that agreement does not however imply perfect reconstruction , since the most likely @xmath186 is about @xmath162 lower than @xmath187 . though the above numbers give an idea of the relative difficulties of using different algorithms at rhic and lhc , experimentally what matters most will be the systematic errors on the correction factors ( for example due to poorly understood non - gaussian tails of the @xmath26 distribution ) . note also that a compensation between shift and dispersion factors , as happens for example with the c / a algorithm , is unlikely to reduce the overall systematic errors . while the goal of this paper is not to discuss the issue of `` fake - jets '' in detail , it is a question that has been the subject of substantial debate recently ( see for example @xcite ) . here , therefore , we wish to devote a few words to it and discuss how it relates to our background - subtraction results so far . in a picture in which the soft background and the hard jets are independent of each other , one way of thinking about a fake jet is that it is a reconstructed jet ( with significant @xmath1 ) that is due not to the presence of an actual hard jet , but rather due to an upwards fluctuation of the soft background . the difficulty with this definition is that there is no uniquely - defined separation between `` hard '' jets and soft background . this can be illustrated with the example of how hydjet simulates rhic collisions : one event typically consists of a soft hydro background supplemented with @xmath188 @xmath5 collisions , each simulated with a minimum @xmath1 cut of @xmath189 on the @xmath190 scattering . to some approximation , the properties of the full heavy - ion events remain relatively unchanged if one modifies the number of @xmath5 collisions and corresponding @xmath1 cut and also retunes the soft background . the fact that this changes the number of hard jets provides one illustration of the issue that the soft / hard separation is ill - defined . additionally , while there are @xmath188 semi - hard @xmath5 collisions ( @xmath191 mostly central semi - hard jets event separately to obtain a long list of @xmath5 jets from all the separate hard events . ] ) in an event , there is only space within ( say ) the acceptance of rhic for @xmath192 jets . thus there is essentially no region in an event which does not have a semi - hard jet . from this point of view , every reconstructed jet corresponds to a hard @xmath5 jet and there are no fake jets at all . for inclusive analyses , such as a measurement of the inclusive jet spectrum , this last point is particularly relevant , because every jet in the event contributes to the measurement . then , the issue of fakes can be viewed as one of unfolding . in that respect it becomes instructive , for a given bin of the reconstructed heavy - ion @xmath1 , to ask what the corresponding matched @xmath5 jet transverse momenta were . specifically , we define a quantity @xmath193 , the distribution of the @xmath5 `` origin '' , @xmath194 , of a heavy - ion jet with subtracted transverse momentum @xmath187 . if the origin @xmath193 is dominated by a region of @xmath194 of the same order as @xmath187 , then that tells us that the jets being reconstructed are truly hard . if , on the other hand , it is dominated by @xmath194 near zero , then that is a sign that apparently hard heavy - ion jets are mostly due to upwards fluctuations of the background superimposed on low-@xmath1 @xmath5 jets , making the unfolding more delicate . in figure [ fig : origin ] we show the origins of heavy - ion jets as determined in our hydjet simulations . jets can match a single heavy - ion jet , i.e. have at least half their @xmath1 contained in the heavy - ion jet . in evaluating @xmath193 we take only the highest-@xmath1 matched jet . if there is no matched jet ( this occurs only rarely ) then we fill the bin at @xmath195 . note also that since we are not explicitly embedding hard jets , all @xmath5 jets in the events have undergone hydjet s quenching . ] the upper row provides the origin plots for anti-@xmath0 jets at rhic . each plot corresponds to one bin of @xmath187 , and shows @xmath193 as a function of @xmath194 . at moderate @xmath187 , the @xmath196 bin , the origin is dominated by low @xmath186 . this is perhaps not surprising , given the result in section [ eq : shift - dispersion - on - spectrum ] that the @xmath186 origin is expected to be @xmath197 lower than @xmath187 for anti-@xmath0 jets additionally , that result assumed an exponential spectrum for the inclusive jet distribution , whereas the distribution rises substantially faster towards low @xmath186 . as @xmath187 increases one sees that the contribution of high @xmath186 jets increases , in a manner not too inconsistent with the expected @xmath197 shift , though the @xmath186 distribution remains rather broad and a peak persists at small @xmath186 . these plots suggest that an inclusive jet distribution measurement with the anti-@xmath0 algorithm at rhic is not completely trivial since , up to rather large @xmath187 , one is still sensitive to the jet distribution at small values of @xmath186 where the separation between `` hard '' jets and the soft medium is less clear . nevertheless , two points should be kept in mind : firstly , the upper row of fig . [ fig : origin ] shows that different @xmath187 have complementary sensitivities to different parts of the @xmath186 spectrum . thus it should still be possible to `` unfold '' the @xmath187 distribution to obtain information about the @xmath186 , unfolding being in any case a standard part of the experimental correction procedure . secondly star quotes @xcite a @xmath149 smaller value for @xmath25 than the @xmath198 that we find in hydjet . such a reduction can make it noticeably easier to perform the unfolding . the impact of a reduction in @xmath25 is illustrated in the lower row of fig . [ fig : origin ] , which shows the result for c / a(filt ) . here , even the @xmath196 bin for @xmath187 shows a moderate-@xmath1 peak in the distribution of @xmath186 , and in the @xmath199 bin the low-@xmath1 `` fake '' peak has disappeared almost entirely . furthermore , the @xmath186 peak is centred about @xmath162 lower than the centre of the @xmath187 bin , remarkably consistent with the calculations of section [ eq : shift - dispersion - on - spectrum ] . overall , therefore , unfolding with c / a(filt ) will be easier than with anti-@xmath0 . corresponding plots for the lhc are shown in fig . [ fig : origin_lhc ] . while c / a(filt ) s lower dispersion still gives it an advantage over anti-@xmath0 , for @xmath200 , anti-@xmath0 does now reach a domain where the original @xmath5 jets are themselves always hard . procedures to reject fake jets have been proposed , in @xcite . they are based on a cut on ( collinear unsafe ) jet shape properties and it is thus unclear how they will be affected by quenching and in particular whether the expected benefit of cutting the low-@xmath186 peak in figs . [ fig : origin ] and [ fig : origin_lhc ] outweighs the disadvantage of potentially introducing extra sources of systematic uncertainty at moderate @xmath1 . one final comment is that experimental unfolding should provide enough information to produce origin plots like those shown here . as part of the broader discussion about fakes it would probably be instructive for such plots to be shown together with the inclusive - jet results . an example of an exclusive analysis might be a dijet study , in which one selects the two hardest jets in the event , with transverse momenta @xmath201 and @xmath202 , and plots the distribution of @xmath203 . here one can define `` fakes '' as corresponding to cases where one or other of the jets fails to match to one of the two hardest among all the jets from the individual @xmath5 events . this definition is insensitive to the soft / hard boundary in a simulation such as hydjet , because it naturally picks out hard @xmath5 jets that are far above that boundary . and , by concentrating on just two jets , it also evades the problem of high occupancy from the large multiplicity of semi - hard @xmath5 collisions . this simplification of the definition of fakes is common to many exclusive analyses , because they tend to share the feature of identifying just one or two hard reference jets . the specific case of the exclusive dijet analysis has the added advantage that it is amenable to a data - driven estimation of fakes . one divides the events into two groups , those for which the two hardest jets are on the same side ( in azimuth ) of the event and those in which they are on opposite sides ( a related analysis was presented by star in ref . @xcite ) . for events in which one of the two jets is `` fake , '' the two jets are just as likely to be on the same side as on the opposite side . this is not the case for non - fake jets , given that the two hardest `` true '' jets nearly always come from the same @xmath5 event and so have to be on opposite sides . , it is nearly always the case that the two hardest jets come from the same @xmath5 event . at the lhc , this happens above @xmath204 . ] thus by counting the number of same - side versus opposite - side dijets in a given @xmath205 bin , one immediately has an estimate of the fake rate . events , if one has only limited rapidity acceptance then the same - side / opposite - side separation is not infrared safe , because of events in which only one hard jet is within the acceptance and the other `` jet '' is given by a soft gluon emission . thus to examine the same - side / opposite - side separation in plain @xmath5 events with limited acceptance , one would need to impose a @xmath1 cut on the second jet , say @xmath206 . ] this is illustrated in fig . [ fig : dijets ] , which shows the distribution at rhic of the full , subtracted @xmath205 result , together with its separation into opposite - side and same - side components . one sees that in the peak region , the opposite and same - side distributions are very similar , indicating a predominantly `` fake '' origin for at least one of two hardest jets ( they are not quite identical , because there is less phase - space on the same side for a second jet than there is on the away side ) . however above a certain full , reconstructed @xmath205 value , about @xmath207 for anti-@xmath0 and @xmath208 for c / a(filt ) the same - side distribution starts to fall far more rapidly than the opposite - side one , indicating that the measurement is now dominated by `` true '' pairs of jets . the lhc results , fig . [ fig : dijets_lhc ] , are qualitatively similar , with the same - side spectrum starting to fall off more steeply than the opposite - side one around @xmath209 for the anti-@xmath0 algorithm and @xmath210 for c / a(filt ) . one can also examine origin plots for @xmath211 , in analogy with the monte carlo analysis of section [ sec : fakes : incl ] . for brevity , we refrain from showing them here , and restrict ourselves to the comment that in the region of @xmath211 where the result is dominated by opposite - side pairs , the origin plots are consistent with a purely hard origin for the dijets . in this article we have presented the results of a systematic study of heavy - ion jet reconstruction with jet - area based background subtraction , building on the brief initial proposal of @xcite . the questions we have examined include those of the choice of range for estimating the background , the choice of algorithm for the jet finding and the robustness of the reconstruction with respect to quenching effects and collision centrality . we have found that there is little difference between various ranges , as long as they are chosen to be localised to the vicinity of the jet of interest and of sufficient size ( at least 4 units of area for jets with @xmath11 ) . in comparing different algorithms we examined the systematic offset and the dispersion in the reconstructed jet @xmath1 . the offset can be brought close to zero by using the anti-@xmath0 algorithm , while the @xmath0 algorithm has the largest offset ; the cambridge / aachen ( c / a ) algorithm with filtering also gives a small offset , however this seems to have been due to a fortuitous cancellation between two only partially related effects . the dispersion is comparable for anti-@xmath0 , c / a and @xmath0 , but significantly smaller for c / a(filt ) ( except at high transverse momenta for lhc ) , as a consequence of its smaller jet area . among the different algorithms , anti-@xmath0 is the most robust with respect to quenching effects , and c / a(filt ) seems reasonably robust at rhic , though a little less so at the lhc . the precise numerical results for offset and dispersion can depend a little on the details of the simulation and the analysis , however the general pattern remains . overall our results indicate that the area - based subtraction method seems well suited for jet reconstruction in heavy - ion collisions . two jet - algorithm choices were found to perform particularly well : anti-@xmath0 , which has small offsets but larger fluctuations , and c / a with filtering , for which the offsets may be harder to control , but for which the fluctuations are significantly reduced , with consequent advantages for the unfolding of experimentally measured jet spectra . ultimately , we suspect that carrying out parallel analyses with these two choices may help maximise the reliability of jet results in hi collisions . we wish to thank n. armesto , h. caines , d. denterria , p. jacobs and c. salgado for numerous helpful comments on the manuscript . we are grateful to them and to b. cole , l. cunqueiro , y - s . lai , m. ploskon , j. puthschke , t. renk , c. roland , s. salur , u. wiedemann and many others , for useful and stimulating conversations on the topic of jets in heavy ion collisions over the past couple of years . we also thank j. andersen for helpful suggestions to obtain some of the computing resources used for this paper . gs would like to acknowledge his affiliation to the brookhaven national laboratory while a large part of this work was performed and jr would like to acknowledge his affiliation to the lpthe in paris during its initial stages . this work has been supported by the french agence nationale de la recherche , under grants anr-05-jcjc-0046 - 01 and anr-09-blan-0060 and the u.s . department of energy under contract de - ac02 - 98ch10886 . in section [ sec : subtraction ] , we gave an estimate of the minimum size of a range one should require for determining @xmath8 , given a requirement that fluctuations in the determination of @xmath8 should be moderate . we give the details of the computation in this appendix . we start from the fact that the error made on the estimation of the background density @xmath8 will translate into an increase of the dispersion @xmath25 : on one hand , the dominant contribution to @xmath25 comes from the intra - event fluctuations of the background _ i.e._is of order @xmath212 with @xmath213 the jet area ; on the other hand , the dispersion @xmath214 of the misestimation of @xmath8 leads to an additional dispersion on the reconstructed jet @xmath1 of @xmath215 . adding these two sources of dispersion in quadrature and using the result from section 3.4 of @xcite , @xmath216 with @xmath217 the area of the range under consideration , we get @xmath218 if we ask _ e.g._that the contribution to the total dispersion coming from the misestimation of the background be no more than a fraction @xmath52 of the total @xmath25 , then we obtain the requirement @xmath219 for anti-@xmath0 jets of radius @xmath28 , with @xmath220 , this translates to @xmath221 where the numerical result has been given for @xmath55 . for @xmath11 , it becomes @xmath222 . we can also cast this result in terms of the number of jets that must be present in @xmath37 . assuming that the jets used to estimate @xmath8 have a mean area of @xmath223 , or c / a algorithms for the background estimation @xcite . ] we find a minimal number of jets , @xmath224 where , as before , we have taken @xmath225 . taking the numbers quoted above and @xmath65 , as used in the main body of the article , this gives @xmath226 . from section 3.3 of @xcite , we know that the presence of hard jets and initial - state radiation leads to a bias in the extraction of @xmath8 of @xmath227 where @xmath228 is a numerical constant and @xmath229 is the average number of `` hard '' jets ( those above the scale of the background fluctuations , including initial - state radiation ) ; @xmath229 is given by @xmath230 where @xmath62 is the number of `` born '' partons from the underlying @xmath231 scattering that enter the region @xmath232 , while @xmath233 is the first coefficient of the qcd @xmath234-function . in the context of @xcite , principally directed towards a study of the ue in @xmath5 collisions , @xmath69 was rather small , causing the second term of eq . ( [ eq : nhard ] ) , associated with initial - state radiation above a scale @xmath235 , to be comparable in size to the first term . in our case , @xmath69 is significantly larger and this reduces the impact of the second term sufficiently that we can ignore it . we thus arrive at the result @xmath236 note that the presence of hard jets and initial - state radiation also affects the fluctuations in the misestimation of @xmath8 and this should in principle have been included in the estimates of appendix [ app : minrange - fluct ] . however , while the effect is not completely negligible , to within the accuracy that is relevant for us ( a few tens of percent in the estimation of a minimal @xmath217 ) it does not significantly alter the picture outlined there . we have seen from fig . [ fig : ptshift ] in section [ sec : choice_alg ] that the subtraction differs when we use the c / a algorithm with and without filtering . since this difference is not due to back - reaction ( see fig . [ fig : ptshift_br ] ) , it has to be due to the subtraction itself . the difference comes from a bias introduced by the selection of the two hardest subjets during filtering . the dominant contribution comes when only one subjet , that we shall assume harder than all the others , contains the hard radiation , all the other subjets being pure background . in that case , the selection of the hardest of these pure - background subjets as the second subjet to be kept tends to pick positive fluctuations of the background . this in turn results in a positive offset compared to pure c / a clustering , as observed in section [ sec : choice_alg ] . to compute the effect analytically , let us thus assume that we have one hard subjet and @xmath237 pure - background subjets of area @xmath238 @xcite . after subtraction , the momentum of each of the pure - background subjet can be approximated as having a gaussian distribution of average zero and dispersion @xmath239 . assuming that the `` hard '' subjet s transverse momentum remains larger than that of all the background jets , the 2 subjets that will be kept by the filter are the hard subjet ( subtracted ) and the hardest of all the subtracted background jets . the momentum distribution of the latter is given by the maximum of the @xmath237 gaussian distributions . gaussian - distributed random numbers , but with the result replaced by zero if all of them are negative . in the calculations here we ignore this subtlety , since we will have @xmath240 and only @xmath241 of the time are three gaussian - distributed random numbers all negative . ] we are only interested here in computing the average bias introduced by the filtering procedure , which is then given by @xmath242 for the typical case @xmath122 and @xmath243 , one finds @xmath244 if we insert in that expression the typical values for the fluctuations quoted in section [ sec : results ] and @xmath11 , we find average biases of 2 gev for rhic and 4.5 gev at the lhc , which are in good agreement with the differences observed between c / a with and without filtering in fig . [ fig : ptshift ] . we stated , in section [ sec : dispersion ] , that the increase of the dispersion at high @xmath1 seen in fig . [ fig : dispersion ] was mainly due to back - reaction . this is made explicit in fig . [ fig : br - v - full - dispersion ] , which decomposes the dispersion into its two components : that associated with the back - reaction , @xmath245 and that associated with background fluctuations and misestimation of @xmath8 ( defined as @xmath246^{\frac12}$ ] ) . one sees that the background - fluctuation component is essentially independent of @xmath1 , while the back - reaction dispersion has a noticeable @xmath1 dependence . this is the case because the back - reaction dispersion is dominated by rare events in which two similarly hard subjets are separated by a distance close to @xmath28 ( specifically by @xmath247 with @xmath248 ) . in such a configuration , the background s contribution to the two subjets can affect whether they recombine and so lead to a large , @xmath249 , change to the jet s momentum . in the limit of a uniform background , @xmath250 , this can be shown to occur with a probability of order @xmath251 . thus the contribution to the average shift @xmath24 is proportional to @xmath252 ( which in a full analysis is found to be enhanced by a logarithm @xcite for the @xmath0 and c / a algorithms ) , while the contribution to @xmath253 goes as @xmath254 , and so leads to a dispersion that should grow asymptotically as @xmath255 ( fig . [ fig : br - v - full - dispersion ] is , however , probably not yet in the asymptotic regime ) . it is worth keeping in mind that even though rare but large back - reaction dominates the overall dispersion , it will probably not be the main contributor in distorting the reconstructed jet spectrum . such distortions come from upwards @xmath26 fluctuations , whereas large back - reaction tends to be dominated by downwards fluctuations . the reason is simple : in order to have an upwards fluctuation from back - reaction , there must be extra @xmath1 near the jet in the original @xmath5 event . this implies the presence of a harder underlying @xmath231 scattering than would be deduced from the jet @xmath1 , with a corresponding significant price to pay in terms of more suppressed matrix elements and pdfs . figure [ fig : br - v - full - dispersion ] also shows that the non - back - reaction component is nearly independent of @xmath1 . this is expected since the anomalous dimension of the jet area is zero for anti-@xmath0 and small for c / a ( with or without filtering ) , and in any case leads to a weak scaling with @xmath1 , as @xmath121 . furthermore , there is roughly a factor of @xmath125 between anti-@xmath0 and c / a(filt ) , as expected based on a proportionality of the dispersion to the square - root of the jet area . one further source of @xmath26 fluctuations can come from imperfect estimation of the area of the jets . we recall that throughout this article we have used soft ghosts , each with area of @xmath256 , in order to establish the jet area . that implies a corresponding finite resolution on the jet area and related poor estimation of the exact edges of the jets , which can have an impact on the amount of background that one subtracts from each jet , and , consequently , on the final dispersion . it is therefore interesting to see , figure [ fig : ghost - area - impact ] , that the dispersion @xmath25 is reduced by about @xmath126 if one lowers the ghost area to @xmath257 . while this does nt affect any of the conclusions of our paper , it does suggest that for a full experimental analysis there are benefits to be had from using a ghost area that is smaller than the default fastjet setting of @xmath256 . s. catani , y. l. dokshitzer , m. h. seymour , and b. r. webber , _ nucl . * b406 * ( 1993 ) 187 ; s. catani , y. l. dokshitzer , m. olsson , g. turnock , and b. r. webber , _ phys . lett . _ * b269 * ( 1991 ) 432 ; s. d. ellis and d. e. soper , _ phys . rev . _ * d48 * ( 1993 ) 31603166 , [ http://arxiv.org/abs/hep-ph/9305266[hep-ph/9305266 ] ] . y. l. dokshitzer , g. d. leder , s. moretti , and b. r. webber , _ jhep _ * 08 * ( 1997 ) 001 , [ http://arxiv.org/abs/hep-ph/9707323[hep-ph/9707323 ] ] ; m. wobisch and t. wengler , http://arxiv.org/abs/hep-ph/9907280[hep-ph/9907280 ] . m. cacciari , g. p. salam and g. soyez , jhep * 0804 * ( 2008 ) 063 [ http://arxiv.org/abs/0802.1189[arxiv:0802.1189 ] ] . g. p. salam and g. soyez , _ jhep _ * 05 * ( 2007 ) 086 , [ http://arxiv.org/abs/0704.0292[arxiv:0704.0292 ] ] ; see also , http://projects.hepforge.org/siscone . o.l . kodolova _ et al _ , `` study of @xmath258+jet channel in heavy ion collisions with cms , '' cms note-1998/063 . v. gavrilov , a. oulianov , o. kodolova and i. vardanian , `` jet reconstruction with pileup subtraction , '' cms rn 2003/004 . n. grau for the atlas collaboration , j. phys . g * 35 * ( 2008 ) 104040 [ http://arxiv.org/abs/0805.4656[arxiv:0805.4656 ] ] . s. salur [ star collaboration ] , eur . j. c * 61 * ( 2009 ) 761 [ http://arxiv.org/abs/0809.1609[arxiv:0809.1609 ] ] . y. -s . lai , b. a. cole , . n. grau , b. a. cole , w. g. holzmann , m. spousta and p. steinberg [ atlas collaboration ] , eur . j. c * 62 * ( 2009 ) 191 [ http://arxiv.org/abs/0810.1219[arxiv:0810.1219 ] ] . y. s. lai for the phenix collaboration , http://arxiv.org/abs/0911.3399[arxiv:0911.3399 ] . m. estienne [ alice collaboration ] , http://arxiv.org/abs/0910.2482[arxiv:0910.2482 ] . j. kapitan , http://arxiv.org/abs/0911.4754[arxiv:0911.4754 ] . m. cacciari and g. p. salam , phys . b * 659 * ( 2008 ) 119 [ http://arxiv.org/abs/0707.1378[arxiv:0707.1378 ] ] . m. cacciari , g. p. salam and g. soyez , jhep * 0804 * ( 2008 ) 005 [ http://arxiv.org/abs/0802.1188[arxiv:0802.1188 ] ] . m. cacciari , g. p. salam and s. sapeta , jhep * 1004 * , 065 ( 2010 ) [ http://arxiv.org/abs/0912.4926[arxiv:0912.4926 ] ] . m. cacciari , j. rojo , g. p. salam and g. soyez , jhep * 0812 * ( 2008 ) 032 [ http://arxiv.org/abs/0810.1304[arxiv:0810.1304 ] ] ; see also http://quality.fastjet.fr . s. salur , nucl . phys . a * 830 * ( 2009 ) 139c [ http://arxiv.org/abs/0907.4536[arxiv:0907.4536 ] ] ; m. ploskon [ star collaboration ] , nucl . a * 830 * ( 2009 ) 255c [ http://arxiv.org/abs/0908.1799[arxiv:0908.1799 ] ] . i. p. lokhtin and a. m. snigirev , eur . j. c * 45 * ( 2006 ) 211 [ http://arxiv.org/abs/hep-ph/0506189[arxiv:hep-ph/0506189 ] ] . i. p. lokhtin , l. v. malinina , s. v. petrushanko , a. m. snigirev , i. arsene and k. tywoniuk , comput . commun . * 180 * ( 2009 ) 779 [ http://arxiv.org/abs/0809.2708[arxiv:0809.2708 ] ] . r. baier , y. l. dokshitzer , a. h. mueller , s. peigne and d. schiff , nucl . b * 484 * ( 1997 ) 265 [ http://arxiv.org/abs/hep-ph/9608322[hep-ph/9608322 ] ] ; b. g. zakharov , jetp lett . * 65 * ( 1997 ) 615 [ http://arxiv.org/abs/hep-ph/9704255[hep-ph/9704255 ] ] . j. casalderrey - solana and c. a. salgado , acta phys . b * 38 * ( 2007 ) 3731 [ http://arxiv.org/abs/0712.3443[arxiv:0712.3443 ] ] ; d. denterria , http://arxiv.org/abs/0902.2011[arxiv:0902.2011 ] ; u. a. wiedemann , http://arxiv.org/abs/0908.2306[arxiv:0908.2306 ] ; a. majumder and m. van leeuwen , http://arxiv.org/abs/1002.2206[arxiv:1002.2206 ] ; d. denterria and b. betz , lect . notes phys . * 785 * ( 2010 ) 285 . t. sjostrand , s. mrenna and p. skands , jhep05 ( 2006 ) 026 [ http://arxiv.org/abs/hep-ph/0603175[hep-ph/0603175 ] ] . n. armesto , l. cunqueiro and c. a. salgado , eur . j. c * 63 * ( 2009 ) 679 [ http://arxiv.org/abs/0907.1014[arxiv:0907.1014 ] ] ; n. armesto , l. cunqueiro , c. a. salgado and w. c. xiang , jhep * 0802 * ( 2008 ) 048 [ http://arxiv.org/abs/0710.3073[arxiv:0710.3073 ] ] . m. cacciari and g. p. salam , phys . b * 641 * ( 2006 ) 57 [ http://arxiv.org/abs/hep-ph/0512210[hep-ph/0512210 ] ] . j. m. butterworth , a. r. davison , m. rubin and g. p. salam , phys . * 100 * ( 2008 ) 242001 [ http://arxiv.org/abs/0802.2470[arxiv:0802.2470 ] ] . d. krohn , j. thaler and l. t. wang , jhep * 1002 * ( 2010 ) 084 [ http://arxiv.org/abs/0912.1342[arxiv:0912.1342 ] ] . s. d. ellis , c. k. vermilion and j. r. walsh , phys . d * 81 * ( 2010 ) 094023 . m. cacciari , g. p. salam and g. soyez , http://www.fastjet.fr . j. adams _ et al . _ [ star collaboration ] , phys . c * 70 * ( 2004 ) 054907 . s. s. adler _ et al . _ [ phenix collaboration ] , phys . c * 71 * ( 2005 ) 034908 [ erratum - ibid . c * 71 * ( 2005 ) 049901 ] . b. b. back _ et al . _ , phys . * 91 * ( 2003 ) 052303 [ http://arxiv.org/abs/nucl-ex/0210015[nucl-ex/0210015 ] ] . n. armesto , arxiv:0903.1330 [ hep - ph ] . s. s. adler _ et al . _ [ phenix collaboration ] , phys . * 91 * ( 2003 ) 182301 [ http://arxiv.org/abs/nucl-ex/0305013[nucl-ex/0305013 ] ] ; j. adams _ et al . _ [ star collaboration ] , phys . * 92 * ( 2004 ) 052302 [ http://arxiv.org/abs/nucl-ex/0306007[nucl-ex/0306007 ] ] ; c. alt _ et al . _ [ na49 collaboration ] , phys . c * 68 * ( 2003 ) 034903 [ http://arxiv.org/abs/nucl-ex/0303001[nucl-ex/0303001 ] ] ; b. b. back _ et al . _ [ phobos collaboration ] , phys . c * 72 * ( 2005 ) 051901 [ http://arxiv.org/abs/nucl-ex/0407012[nucl-ex/0407012 ] ] . j. y. ollitrault , phys . d * 46 * ( 1992 ) 229 ; a. m. poskanzer and s. a. voloshin , phys . c * 58 * ( 1998 ) 1671 [ http://arxiv.org/abs/nucl-ex/9805001[nucl-ex/9805001 ] ] . i. p. lokhtin and a. m. snigirev , http://arxiv.org/abs/hep-ph/0312204[hep-ph/0312204 ] . k. zapp , g. ingelman , j. rathsman , j. stachel and u. a. wiedemann , eur . j. c * 60 * ( 2009 ) 617 [ http://arxiv.org/abs/0804.3568[arxiv:0804.3568 ] ] . b. schenke , c. gale and s. jeon , phys . c * 80 * ( 2009 ) 054913 .

the study of the temperature anisotropies in the cosmic microwave background ( cmb ) can provide crucial information on the origin of the universe and help us to discriminate between different cosmological models . recent observations in many cosmological fields tend to support the standard inflationary model , which predicts the anisotropies to represent an isotropic gaussian random field . the wilkinson microwave anisotropy probe ( wmap ) 1-year data ( bennett et al . 2003 ) , measured the cmb fluctuations with high accuracy . in a first approach ( komatsu et al . 2003 ) , these were found to be consistent with the gaussian predictions . further analyses revealed asymmetries or non - gaussian features which have been confirmed in the 3-year wmap data ( hinshaw et al . 2007 ) . some of these anomalies are : low multipole alignments ( de oliveira - costa et al . 2006 , copi et al . 2007 , land & magueijo 2007 , chiang et al . 2007 ) , north - south asymmetries ( eriksen et al . 2007 ) , structure alignment ( vielva et al . 2007 ) , non - gaussian features found with steerable ( wiaux et al . 2008 ) and directional wavelets ( mcewen et al . 2006 ) , and a very cold spot in the southern hemisphere ( cruz et al . this cold spot ( cs ) is found at galactic coordinates ( @xmath1 ) , and has an angular radius in the sky of @xmath2 . the spherical mexican hat wavelet ( smhw ) was used to detect this structure . this spherical wavelet is an optimal tool to enhance non - gaussian features on the sphere . the first smhw analysis ( vielva et al . 2004 ) was performed on 15 different wavelet scales and two estimators , skewness and kurtosis , were used . the kurtosis deviated significantly from the range of values expected from gaussian simulations at scales around @xmath3 . the spot was found to be the main cause of the deviation from gaussianity , using the area estimator ( cruz et al . cruz et al . ( 2006 ) showed the incompatibility of the frequency independent signal of the cs with galactic foregrounds , and cayn , jin & treaster ( 2005 ) introduced two new estimators , namely @xmath4 and higher criticism , redetecting the cs . cruz et al . ( 2007a ) confirmed the detection in the 3-year wmap data , using all the mentioned estimators . in the latter paper , the probability of finding an at least as high deviation in gaussian simulations was calculated without any _ a posteriori _ assumption . a conservative estimate of this probability was 1.85% , and hence this spot is a very special feature whose origin deserves further investigation . rth et al.(2007 ) detect the cs in the 3-year wmap data using scalar indices instead of wavelets . several models were proposed in order to explain this anomaly : voids ( inoue & silk 2006 , 2007 , rudnick , brown & williams 2007 ) , second order gravitational effects ( tomita 2005 , tomita & inoue 2007 ) , anisotropic bianchi vii@xmath5 model ( jaffe et al . 2005 ) , finite cosmology model , ( adler , bjorken & overduin 2006 ) , asymptotically flat lemaitre - tolman - bondi model ( garcia - bellido & haugbolle 2008 ) and brane - world model ( cembranos et al . 2008 ) are some of them . also a large - scale , non - gaussian angular modulation ( naselsky et al . 2007 ) has been suggested although this hypothesis is not based on any physical model . the bianchi vii@xmath5 model was proven by jaffe et al . ( 2006 ) to be ruled out at high significance , whereas there is still no further evidence for the validity of the other explanations . a full bayesian evidence analysis of that bianchi model can be found in bridges et al . ( 2007a ) . bridges et al . ( 2007b ) show that the cold spot was likely to be driving any bianchi vii@xmath5 template detection . recently , cruz et al . ( 2007b ) showed that the cs could be caused by a cosmic texture . the amplitude and scale of the spot were consistent with that interpretation and the kurtosis of the data was compatible with the gaussian cmb plus texture model at all scales . the predicted number of spots at a scale of @xmath3 or larger , is @xmath6 which is consistent with the observed spot . in this paper we study two other possibilities , namely the rees - sciama ( rees & sciama 1968 ) effect caused by voids as proposed by inoue & silk ( 2006 , 2007 ) , rudnick , brown & williams ( 2007 ) and the sunyaev - zeldovich effect produced by clusters ( hereafter sz effect , sunyaev & zeldovich 1970 ) already considered in cruz et al . both of them could in principle produce temperature decrements as the cs . in addition a dip ( rudnick , brown & williams 2007 ) in the nrao vla sky survey data ( nvss , condon et al . 1998 ) and the eridanus group of galaxies ( brough et al . 2006 ) match approximately the size and position in the sky of the cs . however , the values required by the parameters of either model are not very realistic . a bayesian framework is very convenient to decide which model best describes the data . the prior theoretical and observational knowledge on the parameters is included in the bayesian analysis and can be decisive in some cases . our analysis is performed on the recently released 5-year wmap data ( hinshaw et al . the large - scale structure of the cmb anisotropies remains almost unchanged respect to the 3-year release ( komatsu et al . 2008 ) . we will first review the data and the template fitting method in section 2 , then we recall in section 3 the results obtained for the texture model in order to compare with the sz effect in section 4 and the void model in section 5 . finally we present our conclusions in section 6 . we use the combined , foreground - cleaned q - v - w map ( hereafter wcm ) of the 5-year wmap data release ( hinshaw et al . 2008 ) . the wcm is in the healpix pixelization scheme ( gorski et al . 2005 ) with resolution parameter nside = 64 . since the scale of the spot is around @xmath3 this resolution is sufficient and reduces the number of pixels used in the template fitting . the template fitting is performed in a circular area of @xmath7 radius centered at galactic coordinates ( @xmath1 ) . although the angular radius of the cs is about @xmath3 we have to consider at least a @xmath7 radius patch to take into account the whole neighborhood of the spot for the fit . the smhw convolves all the pixels in this region and they contribute in an important way to the detected structure . in our template fitting we apply bayesian statistics to the data , @xmath8 , in order to find the best hypothesis , @xmath9 , describing the data through a set of parameters , @xmath10 . bayes theorem states : @xmath11 where @xmath12 is the posterior distribution , @xmath13 the likelihood , @xmath14 the prior and @xmath15 the bayesian evidence . the null hypothesis , @xmath16 , describes the data as an isotropic gaussian random field ( cmb ) plus the noise of the wmap data . the alternative hypothesis , @xmath17 describes the data as cmb plus noise and an additional template , @xmath18 , given by a physical model as for instance a texture spot , the sz effect produced by a cluster or the rees - sciama effect generated by a local void . @xmath19 ( with @xmath20 ) depends on the parameter set @xmath21 . for @xmath19 , the probability density associated with the observed data is @xmath22 the evidence is hence the average of the likelihood with respect to the prior and has been extensively used for model selection in cosmology ( e.g. jaffe 1996 , hobson , bridle & lahav 2002 , hobson & mclachlan 2003 , mukherjee , parkinson & liddle 2005 , parkinson , mukherjee & liddle 2006 ) . a model having a large evidence shows a high likelihood for its allowed parameter space given the data . we redefine our notation , with the likelihood @xmath23 , and the prior @xmath24 so the previous integral for the evidence , @xmath25 , reads : @xmath26 the likelihood function is @xmath27 , where @xmath28 is defined as : @xmath29 and @xmath30 is the generalised noise matrix including cmb and noise . the template @xmath18 depends on the parameter set @xmath10 . the calculation of the noise contribution to @xmath30 is straightforward since the noise in the wcm is uncorrelated and well known . in order to obtain the cmb part , we calculate the correlation function for the wcm taking into account the pixel and beam effects . comparing the resulting evidence ratios obtained from the analytical expression for the wcm correlation function with those obtained using simulations , the errors are negligible . the posterior probability ratio : @xmath31 can be used to decide beween both hypotheses . the alternative hypothesis is the more probable one when @xmath32 . if @xmath33 the null hypothesis is the preferred one to describe the data . the a priori probability ratio for the two models , @xmath34 is set to the fraction of the sky , @xmath35 , covered by either texture , clusters or voids . in this manner we compensate the _ a posteriori _ selection of the pixel where the template is centred , namely the center of the cs , ( @xmath1 ) . unified theories of high energy physics predict the production of topological defects ( kibble 1976 , vilenkin & shellard 2000 ) during a symmetry - breaking phase transition in the early universe . depending on the order of the broken symmetry , different types of defects arise , such as domain walls , strings , monopoles or texture . each of these extremely energetic events leave a characteristic pattern in the cmb . for instance cosmic texture ( turok 1989 ) produce hot and cold spots in the cmb sky . cruz et al . ( 2007b ) showed recently that the cs could be due to a collapsing texture . a bayesian template fit was performed on the 3-year wmap data , using an approximated temperature profile based on the analytic result given in turok & spergel ( 1990 ) . the a priori probability ratio for the two models , @xmath34 , was estimated in this case by the fraction of sky covered by texture , @xmath36 . the evidence ratio was @xmath37 , hence the resulting posterior probability ratio was @xmath38 favoring the texture hypothesis versus the null hypothesis . the scale and amplitude of the spot ( @xmath39 and @xmath40 ) were consistent with having being caused by a texture at redshift @xmath41 and energy scale @xmath42 . the number of texture spots of scale @xmath39 or above is around @xmath43 consistent with the single observed spot . the amplitude of the texture spot was shown to be overestimated by a selection bias . texture spots with amplitude @xmath44 were simulated and added to cmb gaussian simulations . performing a template fit on these single extreme events , the inferred amplitude was in average @xmath45 , i.e. almost twice the true amplitude . this means that the amplitude required by the cs is not in conflict with the upper limits on the amplitude calculated from the power spectrum ( urrestilla et al . 2007 ) . moreover , bevis et al . ( 2008 ) and urrestilla et al . ( 2007 ) find a @xmath46 preference for a cosmological model including texture or strings , when considering only cmb data . however this preference is reduced when including hubble key project ( freedman et al . 2001 ) and big bang nuclesosynthesis ( kirkman et al . 2003 ) constraints in the bayesian power spectrum anlysis . in order to estimate the significance of the @xmath38 result , the same fitting procedure was performed on 10,000 gaussian cmb simulations . the most prominent spot was selected for each simulation and a texture profile was fitted to it . around @xmath47 of the simulations showed @xmath48 and the typical values where @xmath49 . this result was found to be almost independent on the priors or the exact shape of the approximated temperature profile . using the 5-year data release we obtain @xmath50 and @xmath51 . the lower noise level allows a slightly better detection of the texture template in the new data release . we have not repeated the frequentist estimation of the significance because the result was very similar to the 3-year one . the probability of having an at least as extreme value ( i.e. @xmath52-value ) of @xmath53 assuming that the null hypothesis is true , should be again around @xmath54 . the presence of texture would explain the excess of kurtosis observed at scales around @xmath3 . the @xmath55 texture predicted by the model would be confused with gaussian cmb spots since the latter have their maximum power at these scales . a number of follow - up tests can be performed to check the texture hypothesis . for instance an all - sky template fitting looking for several texture can confirm or refute the alternative hypothesis . also a polarization analysis can help to decide , since gaussian cmb spots have a correlated polarization pattern , whereas late texture do not show associated polarization . in addition the b - mode texture power spectrum is expected to be larger than the typical inflationary one , having also a different shape ( urrestilla et al . the hot gas trapped in clusters modifies the cmb spectrum through the sunyaev - zeldovich effect . the high - energetic electrons present in the gas collide with the cmb photons , changing their original frequency . two different components can be identified . the former is produced by random thermal motion of the electrons ( thermal effect ) , and the second one , by the peculiar motion of the cluster with respect to the cmb ( kinetic effect ) . the thermal sz effect is typically one order of magnitude higher than the kinetic one , and therefore we focus on the first one . the effect on the cmb photons is frequency dependent and at frequencies below 217 ghz it produces a temperature decrement . hence this effect is a possible candidate to explain an anomalously cold spot at the wmap frequencies which range from 23 ghz to 94 ghz . the temperature decrement is given by : @xmath56 where @xmath57 is the frequency in dimensionless units , with @xmath58 and @xmath59 planck and boltzmann constants respectively , and @xmath60 . the cluster comptonization parameter is defined as @xmath61 where @xmath62 and @xmath63 are the intracluster electron density and temperature , @xmath64 is the thomson cross section , and @xmath65 is the electron mass . the integral is performed along the line of sight through the cluster . in hot clusters ( @xmath66 kev ) , the relativistic electrons lead to a slightly modified spectral behaviour of the thermal sz effect , but this is negligible in our analysis . also neglected is the kinetic sz effect of the cluster . cruz et al . ( 2006 ) showed that the flat frequency dependence of the cs is incompatible with being caused by a sz signal alone . however a combination of cmb plus sz effect may explain the spot and could have a sufficiently flat frequency spectrum . cruz et al . ( 2005 ) already reported that a group of local galaxies were present in the direction of the cs . they argued that the amount of gas needed to explain the whole signal was too large for a small group of galaxies . however the group of galaxies in the direction of the spot has been shown recently to be larger than previously thought . brough et al . ( 2006 ) , define it as the eridanus _ supergroup _ with mass @xmath67 . the angular size and position in the sky of eridanus are similar to those of the cs , as can be seen in figure ( [ fig : patch ] ) . the top panel shows the wcm data and the bottom one , the nearby galaxies catalog ( tully , 1987 ) . both are @xmath68 patches centered at ( @xmath1 ) and were convolved with the smhw at scale @xmath3 . patch of the wmap combined map ( wcm , top ) and the nearby galaxies catalog ( bottom ) , centered at ( @xmath1 ) and convolved with the smhw at scale @xmath3 . the units are wavelet coefficients of temperature in mk and number of counts for the wcm and nearby galaxies catalog respectively.,width=317,height=566 ] however there are other , more prominent nearby groups which do not have an associated cmb spot , hence this could be a statistical coincidence , but it is worth testing through a template fit whether there is a hidden sz - like template in the wcm data . the temperature profile as a function of the angle in the sky , @xmath69 , can be calculated assuming a @xmath70-model ( with @xmath71 ) for the gas - density profile ( see e.g. atrio - barandela & mcket , 1999 ) : @xmath72 where @xmath73 is the central electron density and @xmath74 is the core radius of the cluster . expressing the virial radius of the cluster @xmath75 in terms of the core radius as @xmath76 , eq . [ [ yc ] ] can be rewritten as : @xmath77 where @xmath78r_c t n_c$ ] , and @xmath79 . \label{phi}\ ] ] the parameter @xmath80 is the angle subtended by the core radius of the cluster , whereas @xmath52 is not a parameter but a function of mass @xmath81 and redshift @xmath82 . typical values for clusters yield @xmath83 . summarising , our temperature profile with parameters @xmath80 and @xmath84 reads : @xmath85 . \label{dttfinal}\ ] ] eq . [ [ dttfinal ] ] gives the effect produced by a single cluster through a particular line of sight . as the wcm is a combination of maps at different frequencies , we choose the central frequency , @xmath86 ghz . we consider two alternative hypotheses in this context . in the first one the spot would be caused by a rich cluster and in the second one by a nearby group of the nearby galaxies catalog . we simulate the entire population of clusters with flux above 10 mjy ( at 353 ghz ) , using a monte carlo algorithm where masses and redshifts were assigned to each cluster according to the press - schechter prescription ( press & schechter 1974 ) . we assume @xmath87 , @xmath88 and @xmath89 ( where @xmath90 and @xmath91 are the matter and vacuum energy density parameters , and @xmath92 the dispersion of the density field at scales of 8 mpc ) . the cut in flux removes nearby small groups of galaxies and distant small clusters . the redshifts of the simulated clusters range from 0.005 to 3.4 . for each cluster with known mass and redshift the core radius is determined from the scaling relation : @xmath93 where @xmath94 is the mass of the cluster in @xmath95 @xmath96 . it is important to note that the cut in flux does not have any negative effect in the results of this paper since the fluxes we are trying to explain are much larger than 10 mjy . since the redshift and core radius are known for each cluster , we calculate the angle subtended by the core , @xmath80 . we find the maximum angle of the 300,000 clusters to be @xmath97 . therefore the cluster hypothesis has to be discarded for generating a @xmath3 spot . returning to the particular case of the eridanus supergroup found in the nearby galaxies catalog , we can consider larger values of @xmath80 , due to its proximity and lower values of @xmath84 since it is not a cluster but a big group or an ensemble of groups . in addition we assume lower values of @xmath52 because the gas could be rather dispersed since the gravitational attraction is lower than in big clusters . now the alternative hypothesis is that a big group from the nearby galaxies catalog is causing the cs . although the center of eridanus does not match exactly the center of the spot , we still center the template at ( @xmath1 ) , since the gas does not need to fit exactly the positions of the galaxies . this assumption is the most favorable for the alternative hypothesis . if the cs was due to the sz effect we would expect high x - ray emission in the considered region . we can use the rosat all - sky survey ( snowden et al . 1997 ) to estimate the prior on @xmath84 . the rosat survey excludes some small regions of the sky , and one of them crosses the considered patch . however the estimated average counts in the rosat r6 band are around @xmath98 cts s@xmath99 arcmin@xmath100 . the @xmath70-models which are compatible with this background and at the same time show an extended sz signal of several degrees , range between models with a core radius of @xmath101 mpc and bolometric luminosities of @xmath102 erg / s and models with @xmath103 mpc and bolometric luminosities of @xmath104 erg / s . these models would predict at most @xmath105 . the nearest groups in the nearby galaxies catalog are at @xmath106 mpc distance and as we have shown , the upper limit for the core radius is @xmath103 mpc . in this case we get @xmath107 rad , which we set as the upper limit for @xmath80 . we set the lower limit on @xmath80 to @xmath108 , since the spot has @xmath3 radius and the pixel size is around @xmath108 for the chosen resolution . hence we consider flat priors for parameter @xmath84 in the interval @xmath109 $ ] and for @xmath80 in @xmath110 $ ] radians . setting @xmath111 we obtain a relative evidence of @xmath112 . the fraction of the sky covered by nearby galaxies in the catalog , is @xmath113 therefore @xmath114 and the sz hypothesis is discarded again . the best fit parameters are the upper limits of the priors , @xmath115 and @xmath107 rad . the amplitude of the temperature decrement does not exceed @xmath116 and is hence negligible compared to the deep decrement . therefore the @xmath112 result is not suggesting that the sz hypothesis is viable for explaining the spot . as the template has a negligible effect on the data , @xmath117 , the likelihood is almost constant in the whole parameter range and as the priors integrate to unity we have @xmath118 . this is a well - known feature of bayesian model selection . in any problem , if @xmath17 has additional parameters compared to @xmath16 , and the likelihood is insensitive to these parameters , then @xmath119 . the results do not significantly depend on the value of @xmath52 , the chosen priors or the resolution . the gravitational effect due to a supercluster can not strongly affect large angular fluctuations because these massive structures are very concentrated in a particular region . on the contrary , voids are better candidates to generate large angular spots . the gravitational effect produced by cosmic voids can produce temperature decrements in the cmb and has been proposed as a possible source of the cs ( inoue & silk 2006 , 2007 ) . in analogy to the cluster case with the eridanus supergroup , there is a dip in the nvss radio source catalogue that approximately matches the position of the spot . rudnick , brown & williams ( 2007 ) , claim that the dip could be associated with a void at redshift @xmath120 which could be the origin of the cs , but affirm that the significance of this coincidence is unclear . as the fraction of the universe filled by voids is around @xmath121 in volume at redshift @xmath122 ( colberg et al . 2005 ) , the fraction of the sky covered in projection , @xmath35 , is close to one and hence such coincidences are not rare . the dip and the cs appear at different scales as can be seen in figure [ fig : wmap_nvss ] where we present a @xmath68 patch of the wmap and nvss data convolved with the smhw at scales @xmath3 and @xmath123 . at scale @xmath3 the cs reaches its maximum power in the wcm , whereas there is no significant dip visible in the nvss data . at scale @xmath123 where the cs is not an outstanding feature , a dip becomes visible but there are other similar dips with no associated cmb spots , therefore the association between dip and spot is unclear and could be a coincidence . moreover , there are several other dips with similar depth outside the selected region hence the dip does not seem to be an outstanding feature of the nvss data . mcewen et al . ( 2007 ) found several regions where both data sets were highly correlated and one of them was located close to the cs . this could be due to the very cold temperature of the cs which gives a high correlation with any underdense region in the nvss data at scale @xmath3 and not to the relation between dip and spot since they have different scales ( see figure [ fig : wmap_nvss ] ) . patch of the wcm ( left column ) and the nvss ( right column ) , centered at ( @xmath1 ) and convolved with the smhw at scale @xmath123 ( top row ) and scale @xmath3 ( bottom row ) . the units are wavelet coefficients of temperature in mk and radio source number counts for the wcm and nvss respectively.,width=317 ] rudnick , brown & williams ( 2007 ) also performed an order of magnitude calculation to estimate the radius of the void , @xmath75 , needed to generate such a cmb spot through the rees - sciama effect obtaining @xmath124 mpc in concordance with previous calculations ( inoue & silk 2007 ) . the probability of observing such a void in standard theory is very small . for instance , a void with @xmath125 mpc will occur with a probability of @xmath126 in an optimistic case ( rudnick , brown & williams 2007 ) . typical observed values are @xmath127 mpc for a matter density contrast of @xmath128 ( hoyle & vogeley 2004 , patiri et al . similar results are found in n - body simulations ( colberg et al . 2005 ) . however it is worth performing a bayesian template fit for the void hypothesis . martnez gonzlez et al . ( 1990 ) and martnez gonzlez & sanz ( 1990 ) , performed a calculation of the temperature anisotropy generated by a compensated void in an einstein - de sitter universe , @xmath129 where @xmath130 is the angle formed between the direction of observation and the direction defined by the centre of the structure , and @xmath131 is the distance to the centre of the structure . they adopt the thin shell approximation , where the void is an empty spherical region surrounded by a thin shell containing the matter which would have been inside the void . outside the void the universe is represented by an einstein - de sitter model and the shell propagates approximately as @xmath132 and @xmath133 . the units are @xmath134 , where @xmath135 is the horizon distance . [ [ eq : dt_t_void ] ] applies to photons crossing the void . the photons that do not cross the void are unaffected since there is no net mass excess . introducing the above expressions for @xmath131 and @xmath141 in eq . [ [ eq : dt_t_void ] ] we have our temperature template with parameters @xmath82 , @xmath75 in a flat @xmath136-cdm universe . we choose the prior on @xmath75 in concordance with the void radius distribution found in the two - degree field galaxy redshift survey data ( fig . 3 in hoyle & vogeley 2004 ) . we extrapolate these values for large @xmath75 through an exponential decay as shown in figure ( [ fig : pr_rv1 ] ) . this extrapolation is generous with big voids since the derived decay is faster than the exponential used . for the redshift , we choose a flat prior between 0 and 1 . this assumption is again rather generous with large radii , since according to n - body simulations ( colberg et al . 2005 ) the void sizes should be smaller at high , unobserved redshifts . we set the a priori probability ratio for the two models to unity since voids cover almost all the sky . the obtained evidence ratio is @xmath118 and therefore the posterior probability ratio is @xmath142 . this result is not suggesting that the void model is viable as the best fit template corresponds to @xmath143 mpc and @xmath144 which produces a temperature anisotropy of @xmath145 . as in the cluster case the template is completely negligible , the likelihood almost constant , and the priors integrate to unity , therefore @xmath118 . the contribution of voids with @xmath146mpc to the evidence integral is negligible and the results remain unchanged varying the upper limit of parameter @xmath75 for values above @xmath147mpc . the big voids ( @xmath148mpc and @xmath149 ) described in rudnick , brown & williams ( 2007 ) which do produce an important temperature decrement are strongly penalized by the prior , despite showing higher likelihood values . we redefine @xmath17 in order to test the big void hypotesis . now only voids with @xmath148mpc and @xmath150 are considered . we renormalise the priors in the new parameter ranges and recalculate the prior probability ratio since the big voids ( in case they exist ) are expected to cover only a small part of the sky . as shown above a void with @xmath125 mpc is expected to occur with a probability of @xmath126 ( rudnick , brown & williams 2007 ) . hence as a valid upper limit we can assume that there is at most one big void with @xmath125mpc and @xmath151 . the fraction of the sky covered by this void is @xmath152 . therefore we assume @xmath153 . recalculating the evidence for the new parameter range , we obtain @xmath154 which yields @xmath155 . this result discards the possibility of explaining the cs through such a void . hence the void hypothesis has to be rejected . inoue & silk ( 2006 , 2007 ) proposed slightly different voids with very high values for the radius of the void , @xmath156 mpc and quasi linear density contrast values @xmath157 , while rudnick , brown & williams ( 2007 ) used @xmath158 . inoue & silk ( 2007 ) recognize that the required values for @xmath75 are inconsistent with observations at low @xmath82 and n - body simulations . such a void is a @xmath159 result in the standard theory with @xmath160 . they consider that a percolation process of the typically observed small voids could form the large quasi linear voids . however these voids have not been observed yet and as already mentioned , according to n - body simulations ( colberg et al . 2005 ) the void sizes should be smaller at high , unobserved redshifts . in this paper we consider different possibilities to explain a non - gaussian cold spot found in the wmap data . this spot is located at galactic coordinates ( @xmath1 ) , covers around @xmath161 in the sky , has been shown to be incompatible with gaussian simulations and is unlikely to be produced by foregrounds ( vielva et al . 2004 , cruz et al . 2005 , cruz et al . 2006 , cruz et al . we compare three alternative hypotheses to explain the cs , namely texture , clusters and voids versus the null hypothesis . the latter describes the data as an isotropic gaussian random field ( cmb ) plus the noise of the wmap data . a bayesian analysis takes into account the prior observational and theoretical knowledge on the parameters and can be used to find the hypothesis best describing the data . the posterior probability ratio , @xmath53 , is the quantity we use to compare the different hypotheses . the alternative hypothesis is preferred if @xmath32 and disfavored otherwise . cruz et al . ( 2007b ) performed such an analysis for the texture model , finding that this model is favored versus the hypothesis of gaussian fluctuations with @xmath162 . the predicted size , amplitude , temperature profile , and number of spots generated by texture is consistent with the observed spot . repeating this analysis for the 5-year wmap data we find @xmath163 due to the lower noise level . the frequentist @xmath52-value associated to this result is around @xmath54 . the finding of a local supergroup ( brough et al . 2006 ) and a dip in the nvss survey ( rudnick , brown & williams 2007 ) , which match the position of the cs could be a hint for the cluster or void hypotheses , but our detailed bayesian analysis reveals that these are statistical coincidences . the fraction of the sky covered by clusters and voids is included in our analysis . the large angular size and temperature decrement of the cs are hard to explain with conventional clusters or voids . we would need either a local , rich cluster or a void with radius @xmath164 mpc located at redshift @xmath149 . in the cluster case only a local group with compton parameter @xmath165 is observed in the required direction . the angular scale is marginally consistent with the spot but the temperature decrement generated is only a few @xmath166 and hence can not account for the cs . our bayesian analysis gives @xmath167 , discarding the cluster hypothesis . for conventional voids with @xmath127 mpc we obtain @xmath142 because the likelihood is insensitive to the extra parameters . a detailed analysis reveals that these voids produce negligible temperature decrements below @xmath116 . renormalising the priors for voids with radius @xmath164 mpc , we obtain a posterior probability ratio of @xmath168 . therefore the void hypothesis is not viable either . explaining the cold spot by a void with @xmath148 mpc , would introduce a @xmath159 anomaly in the standard model to solve the @xmath169 anomaly of the spot . the solution would be worse than the problem . hence we can conclude that the texture hypothesis is the most plausible explanation for the cs . many scientists are uncomfortable with cosmic defects , because these failed to explain structure formation . however , cosmic defects are a generic consequence of symmetry breaking , which is an essential ingredient in unified theories of particle physics . hence it is very plausible that they exist . they might be formed at any scale below the grand unified scale , and would be consistent with current cosmological observations . they are even compatible with inflation . the only constraint is that they should make a subdominant contribution to structure formation ( see for instance urrestilla et al . 2007 ) . the virtue of the texture hypothesis is that it is fully testable . a texture would not lead to a correlated polarization pattern around the cold spot , but it would lead to a very specific gravitational lensing pattern . in addition there should be more cmb texture spots at scales around @xmath170 . although they are more difficult to detect because the gaussian cmb reaches its maximum power at these scales , we may be able to detect them through an all - sky bayesian analysis , which is already in progress . mc , emg and pv acknowledge financial support from the spanish mcyt project aya2007 - 68058-c03 - 02 . pv acknowledges financial support from the ramn y cajal project . we acknowledge the use of the legacy archive for microwave background data analysis ( lambda ) . support for lambda is provided by the nasa office of space science . this work has used the software package healpix ( hierarchical , equal area and iso - latitude pixelization of the sphere , http://www.eso.org/science/healpix ) , developed by k.m . grski , e. f. hivon , b. d. wandelt , j. banday , f. k. hansen and m. barthelmann ; the visualisation program univiewer , developed by s.m . mingaliev , m. ashdown and v. stolyarov ; and the camb and cmbfast software , developed by a. lewis and a. challinor and by u. seljak and m. zaldarriaga respectively . adler r. j. , bjorken j. d. , overduin j. m. , 2006 , gr - qc/0602102 . atrio - barandela f. , mcket j. p. , 1999 , apj 515 , 465 . garcia - bellido j. , haugbolle t. , 2008 , jcap , 04 , 003 . bennett c.l . , et al . , 2003 , apjs , 148 , 1 . bevis n. , hindmarsh m. , kunz m. , urrestilla j. , 2008 , phys . lett . , 100 , 021301 . bridges m. , mcewen j.d . , lasenby a.n . , hobson m.p . , 2007 , mnras , 377 , 1473 bridges m. , mcewen j.d . , cruz m. , hobson m.p . , lasenby a.n . , vielva p. , martnez gonzlez e. , 2008 , submitted to mnras ( arxiv:0712.1789 ) . brough , s. , forbes d. a. , kilborn v. a. , couch w. , colless m. , 2006 , mnras , 369 , 1365 . cayn l. , jin l. & treaster a. , 2005 , mnras , 362 , 826 . cembranos j. a. r. , de la cruz - dombriz a. , dobado a. , maroto , a. l. , 2008 , ( arxiv:0803.0694 ) condon j.j . , cotton w.d . , greisen e.w . , yin q.f . , perley r.a . , taylor g.b . , broderick j.j . , 1998 , aj , 115 , 1693 . chiang l .- y . , naselsky p. d. , coles p. , 2007 , apj , 664 , 8 . cruz m. , martnez gonzlez e. , vielva p. , cayn l. , 2005 , mnras , 356 , 29 . cruz m. , tucci m. , martnez gonzlez e. , vielva p. , 2006 , mnras , 369 , 57 . cruz m. , cayn l. , martnez gonzlez e. , vielva p. , jin j. , 2007a , apj , 655 , 11 . cruz m. , turok n. , vielva p. , martnez gonzlez e. , hobson m. , 2007b , science , 318 , 1612 . colberg j.m . , sheth r.k . , diaferio a. , liang g. , yoshida n. , 2005 mnras , 360 , 216 . copi c. j. , huterer d. , schwarz d. j. , starkman g. d. , 2007 , phys . rev . d 75 , 023507 . de oliveira - costa a. , tegmark m. , 2006 , phys . rev . d 74 , 023005 . eriksen h. k. , banday a. j. , grski k. m. , hansen f. k. , lilje p. b. , 2007 , apj , 660 , 81 . freedman et al . , 2001 , , apj , 553 , 47 . hinshaw et al . , 2007 , apjs , 170 , 288 . hinshaw et al . , 2008 , submitted to apj , ( arxiv:0803.0732 ) hobson m , bridle s. , lahav o. , 2002 , 335 , 377 . hobson m. , mclachlan , 2003 , mnras , 338 , 765 . m. s. , 2004 , apj , 607,751 . inoue k. t. , silk j. , 2006 , apj , 648 , 23 . inoue k. t. , silk j. , 2007 , apj , 664 , 650 . jaffe a. , 1996 , apj , 471 , 24 . jaffe t. r. , banday a. j. , eriksen h. k. , grski k. m. , hansen f. k. , 2005 , apj , 629 , 1 . jaffe t. r. , hervik s. , banday a. j. , grski k. m. , 2006 , apj 644 , 701 . kibble , j. phys . , 1976 , a9 , 1387 . kirkman d. , tytler d. , suzuki n. , omeara j.m . , lubin d. , 2003 , apjs , 149 , 1 . komatsu e. et al . , 2003 , apjs , 148 , 119 . komatsu e. et al . , 2008 , submitted to apj , ( arxiv:0803.0547 ) . land k. , magueijo j. , 2007 , mnras.378 , 153 . gonzlez e. & sanz j. l. , 1990 , mnras , 247 , 473 . martnez gonzlez e. , sanz j. l. & silk j. , 1990 , apj , 355 , 5 . mcewen j.d . , hobson m.p . , lasenby a.n . , mortlock d.j . , 2006 , mnras , 371 , l50 mcewen j.d . , vielva p. , hobson m.p . , martnez gonzlez e. , lasenby a.n . , 2007 , mnras , 376 , 1211 . mukherjee p. , parkinson d. , liddle a. , 2005 , apj , 635 , 51 . naselsky p. d. , christensen p. r. , coles p. , verkhodanov o. , novikov d. , kim j. , 2007 , apj submitted , arxiv:0712.1118 . parkinson d. , mukherjee p. , liddle a. , 2006 , phys rev . d 73 , 123523 . patiri s.g . , betancort - rijo j. , prada f. , klypin a. , gottlber s. 2006 , mnras , 369 , 335 . press w.h , schechter p. , 1974 , apj , 187 , 425 . rth c. , schuecker p. , banday a. j. , 2007 , mnras , 380 , 466 . rees m.j . , . , 1968 , nature , 217 , 511 . rudnick l. , brown s. , williams l. , 2007 , apj accepted , astro - ph/0704.0908 . tomita k. , 2005 , phys . d , 72 , 10 . tomita k. , inoue k.t . , 2007 , arxiv:0712.1291 . tully r. b. , 1988 , nearby galaxies catalog . cambridge university press , 1988 , 221 p. turok n. & spergel d. n. , 1990 , phys . letters , 64 , 2736 . urrestilla j. , bevis n. , hindmarsh m. , kunz m. , liddle a. , 2007 , arxiv:0711.1842 . vielva p. , martnez - gonzlez e. , barreiro r. b. , sanz j.l . , cayn l. , 2004 , apj , 609 , 22 . vielva p. , wiaux y. , martnez - gonzlez e. , vandergheynst p. , 2007 , mnras , 381 , 932 . vilenkin a. , shellard e.p.s . , 2000 , cosmic defects and other topological defects ( cambridge univ . press ) . wiaux y. , vielva p. , barreiro r. b. , martnez - gonzlez e. , vandergheynst p. , 2008 , mnras , 385 , 939 .

the influence of environmental degrees of freedom has been identified as the key concept in explaining the classical behavior of macroscopic systems in a quantum framework @xcite . according to this notion a preferred set of localized system states called the pointer basis @xcite is induced in the course of the interaction of the system with its surrounding . most characteristically , any initial superposition of these pointer states gets rapidly mixed , while the only states who retain their purity for a long time are the pointer states themselves . while the basic ideas behind the decoherence process seems to be settled , it still remains an open problem to understand the emergence , the dynamics , and the main properties of the pointer states for microscopic realistic environments . several strategies have been proposed so far for determining the pointer basis given the environmental coupling . in @xcite the suggestion was made to sort all pure states in the hilbert space according to their linear entropy production rate . the pointer states are then identified with the states having minimal loss of purity . similar results are obtained by the approach of @xcite which is based on a time evolution equation whose solitonic solutions are identified with the pointer states . so far , this concept has been applied to the damped harmonic oscillator @xcite and to a free quantum particle coupled linearly to a bath of harmonic oscillators @xcite . there , the solitonic solutions of the corresponding nonlinear equation are coherent states and gaussian wave packets , respectively . moreover , the decoherence to gaussian pointer states was proved to be generic for linear coupling models @xcite . in this paper , we extend the analysis from linear models to a non - perturbative treatment of the interaction . we focus on the model of collisional decoherence , which provides a realistic description of the decoherence process generated by an ideal gas environment . notably , experiments with interfering fullerene molecules displayed a reduction of interference visibility in agreement with this model @xcite . we derive the corresponding pointer states which are shown to form an overcomplete , exponentially localized set of basis states , and we prove decoherence to these states using the orthogonal unraveling of the master equation @xcite . this stochastic process on the one hand provides the statistical weights of the pointer basis , and on the other hand represents an efficient way of solving master equations which exhibit pointer states . moreover , it explains the emergence of classical hamiltonian dynamics . the main result was already announced in @xcite . here , we provide a more detailed explanation of the proofs and derivations . as an extension of @xcite , we prove the decoherence dynamics for a more general situation where the localization rate of collisional decoherence is in a non - saturated regime , and we utilize the relative entropy in order to illustrate the emergence of the statistical weights . while the above results are derived within the framework of decoherence theory , they can also be applied to dynamic reduction models which propose a modification of the schrdinger equation by means of nonlinear and stochastic terms . in fact , the observational consequences of the ghirardi - rimini - weber ( grw ) spontaneous localization model @xcite are equivalent to the ones of collisional decoherence , since they are described by the same master equation @xcite . the present work therefore applies to the grw model . in particular , it provides the corresponding pointer basis . the structure of the article is as follows . in sect . 2 , we briefly review the notion of pointer states , and we summarize the method for determining the pointer states discussed in @xcite . in order to motivate the approach , we consider a two level system and a dephasing process . this method is then applied to collisional decoherence in sect . 3 , which provides a set of solitonic states to be regarded as ` candidate ' pointer states . we show in sect . 4 that these solitons form an overcomplete basis of exponentially localized states , and we give an expression for their spatial extension . moreover , we demonstrate that the ` candidate ' pointer states move on classical phase space trajectories if they are sufficiently localized . in sect . 5 , we briefly review the method of quantum trajectories , focusing in particular on the orthogonal unraveling of the master equation . this stochastic process is then applied to collisional decoherence which allows us to show that the ` candidate ' states are indeed pointer states in the sense of the definition given in sect . 2 . furthermore , we use the orthogonal unraveling to show that the statistical weights of the pointer states are given by the overlap with the initial state . we present our conclusion in section 6 . to motivate the definition of pointer states , let us consider the quantum dynamics of the damped harmonic oscillator . its evolution can be described by a master equation in lindblad form defined by the standard hamiltonian @xmath0 , and a single lindblad operator @xmath1 , with associated rate @xmath2 @xcite . we take as initial state a superposition of two quasi - orthogonal coherent states @xmath3 which satisfy @xmath4 , with @xmath5 . it is then easy to show that for times larger than the decoherence time @xmath6 the solution of the master equation is well approximated by @xcite @xmath7 with @xmath8 . thus , any coherent state remains pure during the damped time evolution , while any superposition of distinct coherent states decays into a mixture ( with a decay rate @xmath9 ) whose statistical weights are determined by the initial overlaps @xmath10 and @xmath11 . due to this property , the coherent states are to be identified with the pointer states of the damped harmonic oscillator . the above observation serves as the starting point for the following definition of the pointer states for an open quantum system evolving according to a lindblad master equation @xmath12 . one says that the system exhibits a pointer basis if its dynamics involves a separation of time scales , characterized by a fast decoherence time @xmath13 , such that for any time much larger than @xmath13 , the evolved state is well approximated by a mixture of uniquely defined pure states @xmath14 which are independent of the initial state @xmath15 , @xmath16 with @xmath17 . following the above example , we further demand that for initial states @xmath15 , which are superpositions of mutually orthogonal pointer states @xmath18 , the probability distribution @xmath19 is given by the initial projections @xmath20 the pointer states @xmath21 initially form an ( overcomplete ) basis , and they may evolve in time , though slowly compared to @xmath13 . the name _ pointer state _ was coined in @xcite due to its relevance for the physical description of a measurement apparatus . a measurement device which probes an observable @xmath22 is constructed such that macroscopically distinct positions of the pointer or indicator are obtained for the different eigenstates of @xmath22 . for a quantum system initially prepared in an eigenstate of @xmath22 the read - out will display the corresponding eigenvalue with certainty provided these pointer states remain pure during the time evolution . on the other hand , if the quantum system is prepared in a superposition of eigenstates of @xmath22 , we expect the pointer not to end up in a superposition of different read - out states , but rather to be at a definite position , though probabilistically , with probabilities given by the overlap ( [ eq : bornrule ] ) . we emphasize that the importance of pointer states goes beyond the physics of measurement devices and the quantum - to - classical transition since they are also a practical tool for the solution of master equations . knowing the pointer states @xmath21 , their time evolution @xmath23 , and their probabilities @xmath24 , one can immediately specify the solution of the master equation for any initial state and times greater than the decoherence time . since the decoherence time is generically much shorter than the system and dissipation time scales of the pointer state motion , this allows one to capture a large part of the system evolution without solving the master equation . a practical way to obtain the pointer states @xmath21 , for a given environmental coupling , was discussed in @xcite . we will illustrate this method by means of a two level system , suspect to a dephasing environment . the corresponding master equation in interaction picture , @xmath25 is characterized by the lindblad jump operator @xmath26 @xmath27 . in bloch representation , @xmath28 with bloch vector @xmath29 and pauli matrices @xmath30 , the solution reads @xmath31 thus , the bloch sphere @xmath32 , which represents the set of all pure states , is projected onto the z - axis in the course of the dephasing process . this implies that the decohered state @xmath33 is a mixture of the eigenstates of @xmath34 ( denoted by @xmath35 and @xmath36 respectively ) , @xmath37 \mathsf{p_{\uparrow } } + { { \rm tr } } \left [ \mathsf{p_{\downarrow } } _ { } \rho_0 \right ] \mathsf{p_{\downarrow } \ , } . \label{eq : decoherencedephasing}\end{aligned}\ ] ] the comparison with ( [ eq : decoherence ] ) shows that the north and the south pole of the bloch sphere ( corresponding to @xmath38 and @xmath39 ) form the pointer basis of the pure dephasing process . since the solution of the master equation will not be at hand in general , one requires a method which yields the pointer states without the knowledge of @xmath40 . to motivate this , we notice that the north pole of the bloch sphere is the asymptotic end point of the trajectory illustrated by the thick line in fig . 1 . this trajectory is generated by an equation of motion with the following properties . ( _ i _ ) it is nonlinear because it must distinguish pointer states from their superpositions . ( _ ii _ ) it preserves the purity of pure initial states , i.e. an initial state which lies on the bloch sphere remains on the surface . ( _ iii _ ) the generated trajectory follows the exact solution of the master equation as close as possible . in order to find such an equation of motion , it is suggestive to minimize the distance of the initial increments , i.e. @xmath41 here , @xmath42 denotes the generator of the master equation ( [ eq : dephasing ] ) in bloch representation and @xmath43 is subject to the condition that the generated trajectory remains on the surface of the bloch sphere . the solution of the above optimization problem reads , in spherical coordinates , @xmath44 since the sine is positive for @xmath45 , the solutions of these equations tend asymptotically towards a pointer state of the system , see fig . 1 . the equator of the bloch sphere forms a set of unstable fixed points of ( [ eq : kugelkoordinaten ] ) . let us now generalize the above argument to general markovian master equations . replacing the euclidean norm by the hilbert - schmidt norm @xmath46 in the space of operators , the generalization of ( [ eq : minimum ] ) to higher dimensional systems reads @xmath47 where the minimization is with respect to all evolution equations @xmath48 which propagate @xmath49 within the set of pure states , such that @xmath50 . it can be shown that the general structure of an equation that evolves state vectors @xmath51 preserving their normalization has the structure @xmath52 , with @xmath53-dependent , hermitian and anti - hermitian mappings @xmath54 and @xmath55 . this implies that the equation of motion for the projector @xmath56 must be of the form @xmath57 \right ] \,,$]where @xmath58 $ ] . using this form , the optimization problem ( [ eq : minimum2 ] ) reduces to @xmath59 \right ] \|^{}_{\rm{hs } } \,_{}$ ] . as shown in @xcite the solution is determined by the generator of the master equation , @xmath60 . hence , the generalization of ( [ eq : kugelkoordinaten ] ) reads as @xcite @xmath61 \ , . \label{eq : ng } \right . \right.\end{aligned}\ ] ] motivated by the example in sect . 2.2 , one conjectures that the asymptotic solutions of ( [ eq : ng ] ) provide the pointer states in more complex systems as well . it will be important below that eq . ( [ eq : ng ] ) is known also in another context : it corresponds to the deterministic part of the orthogonal unraveling of the master equation . as we will demonstrate in section 5 , one can use this specific unraveling to prove for a specific model that the asymptotic solutions of ( [ eq : ng ] ) indeed provide the pointer states . in order to assess the nonlinear equation ( [ eq : ng ] ) in the context of a nontrivial environmental coupling we now apply it to the model of collisional decoherence @xcite . the latter describes the motion of a quantum test particle in an ideal gas environment and it accounts for the quantum effects of the scattering dynamics in a non - perturbative fashion . the corresponding master equation has lindblad form , @xmath62 + \int { \mathrm{d}}q { } \left ( \mathsf{l}_q \rho \mathsf{l}^{^{\dag}}_q - \frac{1}{2 } \mathsf{l}^{^{\dag}}_q \mathsf{l}_q \rho - \frac{1}{2 } \rho \mathsf{l}^{^{\dag}}_q \mathsf{l}_q \right ) \ , , \label{eq : lindblad}\end{aligned}\ ] ] where the jump operators are proportional to momentum kick operators , @xmath63 ( with position operator @xmath64 ) . the continuous label @xmath65 has the meaning of a momentum transfer experienced by the test particle with @xmath66 the corresponding distribution , @xmath67 ; @xmath2 is the collision rate of the gas environment . the 1d equation of motion thus reads @xmath68 + \gamma \int_{- \infty}^{\infty } { \mathrm{d}}q { } \ , g \left ( q \right ) e^{iq \mathsf{x } / \hbar } \rho e^{- iq \mathsf{x } / \hbar } - \gamma \rho \ , . \label{eq : cd}\end{aligned}\ ] ] it leads to a localization in position space , i.e. to a loss of spatial coherence , as can be seen by switching to the interaction picture , @xmath69 , and the position representation , @xmath70 the decay rate of the spatial coherences is thus characterized by a _ localization rate _ @xmath71 which is related to the momentum transfer distribution @xmath72 by @xmath73 since the fourier transform of the distribution @xmath74 tends to zero for large distances @xmath75 , the localization rate saturates for large @xmath75 at the maximum value given by the collision rate , @xmath76 , which can be interpreted as the limit where one collision is sufficient to reveal the particles ` which path ' information . this behavior is in sharp contrast to linear models , where the localization rate grows quadratically , and thus approaches infinity in the limit of a large separation @xmath75 . in order to apply the nonlinear equation ( [ eq : ng ] ) to collisional decoherence ( [ eq : cd ] ) of a free particle , that is @xmath77 , we rewrite the projector equation ( [ eq : ng ] ) in vector representation and choose for @xmath42 the lindblad form ( [ eq : lindblad ] ) , which gives @xcite @xmath78 the expectation value @xmath79 is disregarded in the following , since it contributes only an additional phase . now , we choose the jump operator of collisional decoherence , @xmath80 , and switch to position representation , which yields @xmath81 \left ( x \right ) \ , , \label{eq : ngcd}\\ \fl \lambda \left [ \left| \psi_t \right|^2 \right ] \left ( x \right ) & = & \gamma \left ( | \psi_t |^2 \ast \hat{g } \left ( x \right ) - \int^{\infty}_{- \infty } { \mathrm{d}}y| \psi_t |^2 \left ( y \right ) \left ( | \psi_t |^2 \ast \hat{g } \right ) \left ( y \right ) \right ) \ , . \label{eq : lambda}\end{aligned}\ ] ] here , @xmath82 denotes the convolution and @xmath83 is the fourier transform of @xmath72 , i.e. @xmath84 . the two summands in ( [ eq : ngcd ] ) have counteractive effects on the temporal evolution of the wave function : the coherent term leads to its dispersion , whereas the second , incoherent summand tends to localize the solution . in order to explain this localization , we note that the second summand in ( [ eq : lambda ] ) is independent of @xmath85 . this implies that the centered parts of the wave function , where the convolution @xmath86 exceeds the constant term in ( [ eq : lambda ] ) , get amplified , i.e. @xmath87 , whereas the tails of the wave function get damped , i.e. @xmath88 . as a consequence of these competing effects , solutions of ( [ eq : ngcd ] ) evolve towards solitonic states where both effects are in equilibrium , such that the state moves with fixed shape and constant velocity . as discussed above , these solitons are candidates for the pointer states of collisional decoherence . assuming the momentum transfer distribution @xmath74 to be a centered gaussian with variance @xmath89 , we can rewrite ( [ eq : ngcd ] ) in dimensionless form . @xmath90 here we use the dimensionless variables @xmath91 and @xmath92 to define the dimensionless wave function @xmath93 . notably , eq . ( [ eq : dimensionless ] ) depends only on the single dimensionless parameter @xmath94 . figure 2 shows a numerical solution of ( [ eq : dimensionless ] ) , where we choose a superposition of two counter propagating localized states @xmath95 as the initial state , @xmath96 . as expected from the above discussion , the ( modulus of the ) solution converges to a soliton . moreover , we find that the soliton inherits its initial position and momentum expectation value from that localized component @xmath97 of the initial state which has the greatest weight @xmath98 , @xmath99 . similar observations are found for various other initial states . we proceed to characterize the solitonic solutions of ( [ eq : ngcd ] ) . in sect . 4.1 , the consequences of the conservation of probability on the phase of the solitons are analyzed , allowing us to predict the asymptotic shape of the solitons in sect . 4.2 . in sect . 4.3 we estimate the spatial extension of the solitons , followed by the proof that they form a basis of the hilbert space in sect . 4.4 . finally , in sect . 4.5 , we discuss the dynamics of the solitonic solutions in the presence of an external potential . as demonstrated in the previous section , the nonlinear equation ( [ eq : ngcd ] ) exhibits solitonic solutions @xmath100 in the sense that the modulus of @xmath100 moves with constants shape and velocity , i.e. @xmath101 with @xmath102 and @xmath103 real . in this section , we analyze the general structure of the phase @xmath104 , which will be relevant subsequently . the time derivative of a solution @xmath105 of ( [ eq : ngcd ] ) , yields the continuity equation for @xmath106 , @xmath107 \left ( x \right ) \ , . \label{eq : continuity}\end{aligned}\ ] ] plugging the solitonic solution ( [ eq : shape ] ) into ( [ eq : continuity ] ) , gives @xmath108 \left ( x - vt \right ) - v \partial_x \log f^2 \left ( x - vt \right ) & = & - \frac{\hbar}{m } ( \partial_x^2 g \left ( x , t \right ) \nonumber\\ & & \left . + \partial_x^ { } g \left ( x , t \right ) \partial_x \log f^2 \left ( x - vt \right ) \right ) \ , . \label{eq : intermediate1}\end{aligned}\ ] ] here we have used that @xmath109 \left ( x \right ) = \lambda \left [ f^2 \right ] \left ( x - vt \right)$ ] , which follows from @xmath110 . the time dependence of the left hand side of ( [ eq : intermediate1 ] ) corresponds to a spatial shift . thus , also the right hand side of ( [ eq : intermediate1 ] ) must exhibit such a simple time dependence , which implies that @xmath111 where @xmath112 denotes the right hand side of ( [ eq : intermediate1 ] ) . it follows that @xmath113 since this equation must hold for all @xmath114 and @xmath115 , we may assume that the equality holds already for the summands , such that @xmath116 \ , . \label{eq : ableitung1}\end{aligned}\ ] ] therefore , the temporal and spatial dependence of the phase has the general structure @xmath117 with unknown functions @xmath118 and @xmath119 . to explore the tails of the solitonic states @xmath100 let us consider the form of ( [ eq : ngcd ] ) for asymptotically large positions . it reads @xmath120 with @xmath121 a @xmath53-dependent , positive constant . inserting the solitonic form ( [ eq : shape ] ) into ( [ eq : asymptotic ] ) yields @xmath122 + vf \left ( x - vt \right ) \nonumber \\ & & - \frac{\hbar}{m } \left [ \partial_x f \left ( x - vt \right ) \partial_x g \left ( x , t \right ) + f \left ( x - vt \right ) \partial_x^2 g \left ( x , t \right ) \right]\nonumber \\ & & - \gamma a_{\psi } f \left ( x - vt \right ) \ , . \label{eq : long}\end{aligned}\ ] ] using ( [ eq : phase ] ) , we find that both @xmath123 and @xmath124 are only a function of @xmath125 , and accordingly , that also the left hand side of ( [ eq : long ] ) must be a function of @xmath126 . if follows that @xmath127 is at most linear in @xmath128 ( that is @xmath129 , with unknown constants @xmath130 and @xmath131 ) . considering the real and imaginary part of ( [ eq : long ] ) separately , one obtains two coupled ( second order ) differential equations @xmath132 ^ 2 \right ) \ , , \label{eq : set2}\end{aligned}\ ] ] where @xmath133 and @xmath134 . this set of equations has two unique solutions @xmath135 where the constant @xmath136 depends on the boundary condition for ( [ eq : set1 ] ) ( which can be determined only by solving the full nonlinear equation ( [ eq : ngcd ] ) ) . the solution with the positive exponent in ( [ eq : f ] ) is irrelevant , since it is not normalizable . figure 3 confirms that the tails of the numerically obtained solitonic solutions of ( [ eq : dimensionless ] ) are in agreement with the functional form ( [ eq : f ] ) ; they are straight lines in the semi - logarithmic plot . this shows that , unlike in linear models @xcite where the pointer states are gaussian , the pointer states of collisional decoherence are exponentially localized . an important characteristic of the pointer states is their spatial extension . as explained in @xcite , the latter can be related to the experimentally accessible one - particle coherence length of a thermal gas . we will determine the pointer width in this section , and apply the result later , when studying the dynamics of pointer states in an external potential . as a first step , consider the standard deviation @xmath137 of the numerically obtained dimensionless solitonic solution @xmath138 of ( [ eq : dimensionless ] ) as a function the dimensionless parameter @xmath139 . as shown by the solid line in fig . 4 , the size @xmath137 increases linearly with @xmath140 . this observation can be reproduced by a simplified model which has the practical advantage that it can be applied to more complex systems , such as 3d gases with a microscopically realistic localization rate @xmath141 @xcite . the idea of the model goes as follows : the ideal gas environment consists of particles which collide with the system at a rate @xmath2 . at each collision , the ambient particles gain position information , such that the wave function gets spatially localized to a length scale @xmath142 determined by the localization rate @xmath143 , see ( [ eq : decaycoherences ] ) . after the scattering event , the particle disperses freely , until it gets localized again by a subsequent collision . the pointer width @xmath144 is then obtained by averaging the time - dependent width of the wave function over the waiting time distribution of a poisson process . more specifically , we assume that the length scale @xmath145 is characterized by the free parameter @xmath146 using ( [ eq : localization ] ) and taking the momentum transfer distribution @xmath147 to be a centered gaussian with variance @xmath148 , we obtain @xmath149 with @xmath150 . the free dispersion after the collision yields the time dependent size @xmath151 finally , the average over the waiting time distribution @xmath152 gives @xmath153 where we use a linearization of @xmath154 in the second line . the dimensionless version of ( [ eq : size ] ) reads @xmath155 the dashed line in figure 4 shows that the form of this equation agrees with the numerical solution of ( [ eq : dimensionless ] ) ; the fit yields a value of @xmath156 for the parameter characterizing the localization length scale . our next aim is to show that the solitonic solutions of ( [ eq : ngcd ] ) , which are interpreted below as the pointer states of collisional decoherence , form an overcomplete basis . for this purpose , we first present a general method to construct a whole manifold of solutions of ( [ eq : ng ] ) given a specific one . it relies on the symmetry properties of the corresponding master equation . since collisional decoherence exhibits galilean ( i.e. translation and boost ) invariance , it is then easy to show that the pointer states of this model form an overcomplete basis . suppose there is a family of unitary operators @xmath157 , satisfying @xmath158 , \label{eq : condition2}\end{aligned}\ ] ] where we denote by @xmath159 the incoherent part of the master equation , @xmath160 / \left ( i \hbar \right ) + \mathcal{d } \left ( \rho \right)$ ] . then , given a solution @xmath161 of the nonlinear equation ( [ eq : ng ] ) , also @xmath162 constitutes a solution of ( [ eq : ng ] ) . this can be verified easily : @xmath163 \right ] & = & \frac{1}{i \hbar } \left ( \mathsf{h } \mathsf{u } \mathsf{p } \mathsf{u}^{\dag } - \mathsf{u } \mathsf{p } \mathsf{u}^{\dag } \mathsf{h } \right ) + \mathsf{u } \left [ \mathsf{p } , \left [ \mathsf{p } , \mathcal{d } \left ( \mathsf{p } \right ) \right ] \right ] \mathsf{u}^{\dag } \nonumber\\ & = & { \frac{1}{i \hbar } ( \mathsf{h } \mathsf{u } \mathsf{p } \mathsf{u}^{\dag } - \mathsf{u } \mathsf{p } \mathsf{u}^{\dag } \mathsf{h } + \mathsf{u } \mathsf{h } \mathsf{p } \mathsf{u}^{\dag } - \mathsf{u } \mathsf{h } \mathsf{p } \mathsf{u}^{\dag } + \mathsf{u } \mathsf{p } \mathsf{h } \mathsf{u}^{\dag } } \nonumber\\ & & - \mathsf{u } \mathsf{p } \mathsf{h } \mathsf{u}^{\dag } ) + \mathsf{u } \left [ \mathsf{p } , \left [ \mathsf{p } , \mathcal{d } \left ( \mathsf{p } \right ) \right ] \right ] \mathsf{u}^{\dag } \nonumber\\ & = & \frac{1}{i \hbar } \left [ \mathsf{h } , \mathsf{u } \right ] \mathsf{p } _ { } \mathsf{u}^{\dag } + \mathsf{u } \left [ \mathsf{p } _ { } , \left [ \mathsf{p } _ { } , \frac{1}{i \hbar } \left [ \mathsf{h } , \mathsf{p } \right ] + \mathcal{d } \left ( \mathsf{p } _ { } \right ) \right ] \right ] \mathsf{u}^{\dag } \nonumber\\ & & - \frac{1}{i \hbar } \mathsf{u } \mathsf{p } _ { } \left [ \mathsf{u}^{\dag } , \mathsf{h } \right ] \nonumber\\ & = & \partial_t \left ( { \mbox{\rm $ \mathsf{u } \mathsf{p } \mathsf{u}^{\dag}$ } } \right ) \ , , \label{eq : upu}\end{aligned}\ ] ] where we dropped the time argument for brevity . here , the first equality makes use of ( [ eq : condition1 ] ) and the unitarity of @xmath164 , and the third equality is due to the relation @xmath165 = \left [ \mathsf{p } , \left [ \mathsf{p } , \left [ \mathsf{h } , \mathsf{p } \right ] \right ] \right]$ ] . in the last line , we use ( [ eq : ng ] ) and ( [ eq : condition2 ] ) . let us now apply this to the galilean invariance of master equation ( [ eq : cd ] ) . we will see that the phase space translations @xmath166 satisfy the symmetry conditions ( [ eq : condition1 ] ) and ( [ eq : condition2 ] ) provided the time dependence of @xmath167 and @xmath168 has the particular form @xmath169 the latter enact a phase space translation in accordance with the free shearing motion . let us first verify condition ( [ eq : condition1 ] ) using @xmath170 . @xmath171 in order to verify ( [ eq : condition2 ] ) , use the campbell - hausdorff formula to rewrite the translation operator as @xmath172 the time derivative thus yields @xmath173 \ , , \label{eq : timederivativeoft}\end{aligned}\ ] ] where the shearing transformation ( [ eq : time1 ] ) and ( [ eq : time2 ] ) is required in the second line . this confirms ( [ eq : condition2 ] ) for @xmath174 . we conclude that the nonlinear equation ( [ eq : ngcd ] ) exhibits a family of solitonic solutions @xmath175 , parameterized by the phase space coordinate @xmath176 . in order to verify that this family forms an overcomplete basis , let us consider a specific class of phase space representations . according to @xcite , any hilbert - schmidt operator @xmath22 can be represented as @xmath177 provided @xmath178 is a trace - class operator , i.e. @xmath179 . here , @xmath180 denotes a phase space integral and @xmath181 is a function of the phase space coordinate @xmath182 . choosing for @xmath22 the identity @xmath183 , and for @xmath178 the solitonic solution @xmath184 of ( [ eq : ngcd ] ) with vanishing position and momentum expectations , we obtain a resolution of the identity in terms of the solitons @xmath185 , @xmath186 this demonstrates that the pointer states of collisional decoherence form an overcomplete basis . * * * * * * so far , we have characterized the solitonic solutions of the nonlinear equation ( [ eq : ngcd ] ) which applies in the absence of an external force . if an additional potential is present the corresponding nonlinear equation contains an additional term @xmath187 on the right hand side of ( [ eq : ngcd ] ) . the numerical treatment shows that the solutions still converge to localized wave packets , which , however , change their shape and velocity in the course of the evolution . we find that the center of these wave packets moves on the corresponding classical phase space trajectory provided the collision rate is sufficiently large . we first summarize our numerical findings and then proceed with an analytic explanation . figure 5 shows the position and momentum expectation values of the numerical solution of the nonlinear equation , in case of an anharmonic external potential of the form @xmath188 ( starting from an gaussian initial state ) . the panel on the left hand side of fig . 5 was obtained in the limit of a vanishing collision rate @xmath189 ( i.e. @xmath190 ) , which turns ( [ eq : dimensionless ] ) into the schrdinger equation . the solution therefore disperses , and the solid line shows a typical evolution of the phase space expectation values . the dashed line , on the other hand , gives the classical trajectory of the phase space point where the initial state is localized . the result for a large collision rate @xmath2 ( or small @xmath191 ) is shown on the right hand side of fig . , the initial state turns rapidly into a soliton whose expectation values move on the corresponding classical trajectory . this illustrates that the temporal evolution turns from quantum to classical dynamics with increasing collision rate @xmath2 ( i.e. decreasing @xmath191 ) . we made similar observations with various other potentials . in order to explain the numerical observation , first consider a particle in a linear potential @xmath192 . the corresponding nonlinear equation reads as @xmath193 \ , . \label{eq : linear}\end{aligned}\ ] ] as discussed in sect . 4.1 , the field - free version of this equation @xmath194 exhibits uniformly moving solitonic solutions of the form @xmath195 \right ) \ , . \label{eq : movingsoliton}\end{aligned}\ ] ] this implies that ( [ eq : linear ] ) has solitonic solutions of the form @xmath196 with @xmath197 and @xmath198 . in order to verify this statement , we evaluate the left hand side of ( [ eq : linear ] ) with the ansatz given by ( [ eq : accelerated ] ) . this yields @xmath199 with @xmath200 , @xmath201 and @xmath202 . the expression can be further simplified by noting that the free soliton ( [ eq : movingsoliton ] ) is a solution of the field - free version of ( [ eq : linear ] ) , implying that @xmath203 \left ( x \right ) & = & if \partial_t \chi \left ( t \right ) - v \left ( \partial_x f + if \partial_x \phi \right ) \nonumber\\ & & + \frac{\hbar}{m } \left ( \partial_x f \partial_x \phi - \frac{i}{2 } \partial_x^2 f + \frac{i}{2 } f \left ( \partial_x \phi \right)^2 + \frac{1}{2 } f \partial_x^2 \phi \right ) \ , , \label{eq : freefieldcase}\end{aligned}\ ] ] with @xmath204 \left ( x \right)$ ] defined in ( [ eq : lambda ] ) . using ( [ eq : derivatives ] ) , ( [ eq : freefieldcase ] ) and the above definition of @xmath205 , one finds that @xmath206 \left ( x \right ) \,,\end{aligned}\ ] ] which confirms that @xmath106 evolves according to ( [ eq : linear ] ) . we conclude that in a linear potential the pointer states have the same shape as in the field - free case and they are uniformly accelerated like a classical particle . for general potentials , this implies that the pointer states follow the corresponding classical motion , provided the spatial width of the solitons is sufficiently small , such that the linearization of the potential is justified over their spatial extension . since the size of the pointer states decreases with the collision rate ( see section 4.3 ) , the pointer states must exhibit classical dynamics in the limit of large collision rates . thus far , we have calculated ` candidate ' pointer states as the solitonic solutions of ( [ eq : ngcd ] ) , and we have studied their properties and dynamics . next , it will be shown that these ` candidates ' are genuine pointer states @xmath21 in the sense of ( 1 ) . moreover , we find that the statistical weights @xmath207 of the pointer states are given by the overlap of the initial state @xmath15 with the initial pointer state @xmath208 , i.e. @xmath209 \ , . \label{eq : initaloverlap}\end{aligned}\ ] ] in order to verify the above conjectures , let us now make use of the formalism of quantum trajectories @xcite to solve the master equation ( [ eq : cd ] ) . more precisely , a specific quantum trajectory method , the _ orthogonal unraveling _ , is distinguished by the physics of pointer states because the deterministic part of the associated stochastic differential equation coincides with the nonlinear equation ( [ eq : ng ] ) . we start with a general description of the quantum trajectory approach and the orthogonal unraveling in section 5.1 . the latter will be applied to collisional decoherence in section 5.2 . this will allow us to evaluate the statistical weights of the pointer states in section 5.3 . in the quantum trajectory approach , the wave function corresponding to a pure initial state @xmath210 is propagated stochastically to generate pure state trajectories @xmath211 , whose ensemble average recovers the solution of the master equation ( [ eq : lindblad ] ) , i.e. @xmath212 such a stochastic process is called an unraveling of the master equation . a common unraveling is provided by the quantum monte carlo method @xcite which is based on a piecewise deterministic process . here , one realization of a trajectory consists of smooth deterministic pieces generated by an effective ( non - hermitian ) hamiltonian , in our case @xmath213 which are interrupted by random jumps . the jumps occur with the rate @xmath214 an expectation value with respect to @xmath215 , and they are effected by the operators @xmath216 the quantum monte carlo method is not the only stochastic process satisfying ( [ eq : average ] ) . in fact , there exists an infinite set of these stochastic processes , because the convex decomposition of the density matrix on the left hand side of ( [ eq : average ] ) is not unique . for instance , other unravelings can be obtained from the quantum monte carlo method , by noting that the generator @xmath42 does not uniquely fix the lindblad operators @xmath217 and the hamiltonian @xmath218 @xcite . this is due to the fact that the master equation ( [ eq : lindblad ] ) is invariant under certain transformations of the lindblad operators , such as the addition of a complex multiple @xmath219 of the identity @xmath220 , @xmath221 in the latter case , also the hamiltonian must be transformed as @xmath222 in order to assure the invariance of the master equation . to obtain a different , but again piecewise deterministic unraveling , we now make the choice @xmath223 . it then follows that the deterministic pieces of a sample path are given by the solution of the nonlinear equation ( [ eq : nonlinear ] ) ( or equivalently the corresponding projector equation ( [ eq : ng ] ) ) . the jumps occur with the rate @xmath224 and are caused by the nonlinear operators @xmath225 as a distinctive feature , the states @xmath226 into which the system may jump are orthogonal to the original state @xmath227 , thus justifying its naming . ( the states @xmath228 are not necessarily mutually orthogonal , though . ) to our knowledge , this unraveling was first noted by rigo and gisin @xcite , although , it has not been studied numerically so far . a related unraveling , which is also referred to as the ` orthogonal unraveling ' , was introduced by disi @xcite . here , the deterministic pieces of the evolution are as well generated by ( [ eq : nonlinear ] ) . however , the states @xmath228 into which the system may jump are obtained differently , as the eigenvectors of the hermitian operator @xmath229 where @xmath230 . as a consequence , these states are also _ mutually _ orthogonal ( in finite dimensional systems ) . since the orthogonal unraveling of @xcite requires the diagonalization of the operator ( [ eq : rateoperator ] ) , it is much more involved than the one defined by ( [ eq : nonlinear ] ) , ( [ eq : rate2 ] ) and ( [ eq : jump ] ) , which is why we will use the latter in the following . as mentioned above , all unravelings are equivalent in the same sense as the different convex decompositions of the density matrix @xmath231 . note , however , that a preferred set of pure states the pointer basis may be singled out through the environmental coupling . in that case , those unravelings are distinguished which generate for any initial state an ensemble of these state independent projectors . an unraveling will do this job if ( a ) its deterministic part exhibits stable fixed points or solitons @xmath232 , which ( b ) are characterized by a vanishing jump rate , @xmath233 . in that case , the sample paths of the process will end up in one of the states @xmath21 , by all means . hence , the ensemble mean is of the form ( [ eq : decoherence ] ) , such that the fixed points or solitons @xmath21 can be identified with the pointer states of the system . for the case of collisional decoherence the orthogonal unraveling fulfills the aforesaid conditions . ( a ) its deterministic part is given by ( [ eq : ngcd ] ) which exhibits the solitonic solutions @xmath234 shown in fig . 2 . we shall see explicitly in section 5.2.1 that these states are attractive fixed points . ( b ) we will show in section 5.2.2 that the jump rate ( [ eq : rate2 ] ) vanishes for the solitons @xmath234 , which finally demonstrates that our ` candidate ' pointer states @xmath235 are genuine pointer states @xmath21 in the sense of ( 1 ) . moreover , this shows that the orthogonal unraveling is a very efficient numerical scheme for the long time solution of the master equation ( [ eq : cd ] ) , since the state is no longer affected by the stochastic part , once it has turned into the soliton , and the trajectory is therefore more easy to integrate . we note that the orthogonal unraveling is not the only stochastic process which generates the ensemble of pointer states . in fact , there is a _ diffusive _ unraveling @xcite , which also involves ( [ eq : ng ] ) as its deterministic part . it was applied in @xcite to investigate pointer states in a linear model . we now apply the described orthogonal unraveling to collisional decoherence ( [ eq : cd ] ) , first evaluating the deterministic part ( [ eq : nonlinear ] ) of the stochastic process in sect . 5.2.1 and then the stochastic one ( [ eq : rate2 ] ) and ( [ eq : jump ] ) in sect . 5.2.2 . applying ( [ eq : nonlinear ] ) to the case of collisional decoherence yields the soliton equation ( [ eq : ngcd ] ) discussed in sect . 3.2 . we will now further simplify this equation , by considering initial states @xmath236 which are superpositions of non - overlapping wave functions @xmath237 , @xmath238 the latter are assumed to be localized in the sense that @xmath239 where @xmath240 and @xmath241 denote the variances of the distributions @xmath242 and @xmath243 , respectively . under this assumption , which will be justified at the end of this section , one can extract a system of evolution equations for the time evolution of the coefficients in ( [ eq : initialstate ] ) , @xmath244 here , the matrix @xmath245 is obtained from the localization rate ( [ eq : localization ] ) , where the @xmath246 denote the mean positions of the constituent wave functions @xmath247 . the latter evolve according to @xmath248 \left ( x , t \right ) \nonumber\\ & & + \phi_{^ { } i } ( x , t ) \sum^n_{j = 1 , j \neq i } { } \left| c_j \left ( t \right ) \right|^2 \tilde{\gamma}_{ij } \left ( x , t \right ) \ , , \label{eq : basis}\end{aligned}\ ] ] where @xmath249 is defined in ( [ eq : lambda ] ) and @xmath250 is a rate of the order of @xmath2 , @xmath251 let us now verify that @xmath252 , with @xmath253 and @xmath254 solutions of ( [ eq : coefficients ] ) and ( [ eq : basis ] ) , evolves according to ( [ eq : ngcd ] ) . first , we note that the assumption ( [ eq : a2a ] ) leads to the approximation @xmath255 for all @xmath65 contributing appreciably to integrals weighted with the momentum transfer distribution @xmath74 . this , in turn , implies @xmath256 hence , we find that @xmath257 , and one thus obtains @xmath258 \left ( x , t \right ) & = & - \left ( \left| \phi_i \right|^2 \ast f \right ) \left ( x , t \right ) \ , . \label{eq : lambdaa2}\end{aligned}\ ] ] now , consider the time derivative of @xmath259 , which gives @xmath260 where we dropped the arguments for brevity . in ( [ eq : beweis1 ] ) , we used ( [ eq : coefficients])-([eq : basisrate ] ) , ( [ eq : lambdaa2 ] ) , the fact that @xmath261 , and the normalization condition @xmath262 . now , we replace @xmath263 in ( [ eq : beweis1 ] ) by the right hand side of ( [ eq : fij ] ) , and use ( [ eq : a1 ] ) , which yields @xmath264 finally , by using ( [ eq : localization ] ) , we obtain @xmath265 which confirms that @xmath259 evolves according to ( [ eq : ngcd ] ) . let us now discuss the solution of eq . ( [ eq : coefficients ] ) for the time evolution of the coefficients . we first consider situations where the wave packets @xmath266 are sufficiently far apart such that the localization rate is saturated , i.e. @xmath267 under this assumption , ( [ eq : coefficients ] ) reduces to the equation @xmath268 which was already studied in @xcite in the context of a discrete model for quantum measurement . it is shown there that all stable fixed points of ( [ eq : coefficientssaturated ] ) have the form @xmath269 , and that the particular fixed point @xmath270 , with @xmath271 is approached monotonically , i.e. the component with the largest initial weight wins . this behavior is visualized in fig . 6 which was obtained by solving ( [ eq : coefficientssaturated ] ) numerically for the case @xmath272 . here , the @xmath85- and the @xmath273-axis indicate the weights @xmath274 and @xmath275 , respectively . the plot on the left hand side shows various trajectories , illustrating in particular the fixed points . the plot on the right displays the regions of attraction of the stable fixed points @xmath276 , in agreement with the criterion ( [ eq : argmax ] ) . for instance , area @xmath277 highlights the region of attraction of the fixed point @xmath278 . figure 7 , on the other hand , depicts a scenario where the wave packets @xmath266 are close together such that the localization rate is unsaturated , i.e. @xmath279 . here , we choose @xmath272 wave packets with non - equidistant position expectations , @xmath280 . similarly to the saturated case , we observe that all stable fixed points of ( [ eq : coefficients ] ) have the form @xmath281 . however , the regions of attraction are deformed such that criterion ( [ eq : argmax ] ) is no longer valid , and the fixed points are not necessarily approached monotonically . to see that @xmath281 are stable fixed points of ( [ eq : coefficients ] ) , assume that the coefficients are close by , i.e. @xmath282 with @xmath283 . it follows from ( [ eq : coefficients ] ) that @xmath284 and hence , @xmath285 . the knowledge of the fixed points of the coefficients allows one to discuss the asymptotic evolution of the initial state shown in eq . ( [ eq : initialstate ] ) . since the coefficients @xmath286 with @xmath287 tend to zero asymptotically , it follows that @xmath288 for a specific @xmath289 ( which is given by ( [ eq : argmax ] ) in the saturated case ) . the asymptotic behavior of the wave packets @xmath290 can , in turn , be predicted from eq . ( [ eq : basis ] ) . since the @xmath291 vanish for large times , the coupling term given by the last summand in * * ( [ eq : basis ] ) vanishes as well , implying that the time evolution ( [ eq : basis ] ) of @xmath290 is asymptotically equal to the soliton equation ( [ eq : ngcd ] ) . therefore , in the absence of stochastic jumps , the state @xmath292 evolves into that solitonic solution @xmath293 of ( [ eq : ngcd ] ) which is associated to the initial wave packet @xmath294 . it should be mentioned that eqs . ( [ eq : coefficients ] ) and * * ( [ eq : basis ] ) for the coefficients @xmath98 and the constituent wave packets @xmath97 are not completely decoupled , since ( [ eq : coefficients ] ) depends on the matrix @xmath295 which contains the position expectations @xmath296 of the wave packets @xmath97 . however , the position expectation follows the classical trajectory for sufficiently large @xmath191 s , such that ( [ eq : coefficients ] ) can be solved without knowing the solution of * * ( [ eq : basis ] ) . let us now discuss the validity of the assumption of small position variance ( [ eq : a2a ] ) , and the ensuing approximation ( [ eq : a2b ] ) . it can be justified by our observation in section 4.3 , that the dimensionless pointer width @xmath297 is a function of the parameter @xmath298 only , @xmath299 thus , for all @xmath300 we find that the position variance @xmath301 of a pointer state is one order of magnitude smaller than the reciprocal width of the momentum transfer distribution @xmath74 , @xmath302 the above relation for the width of the pointer state is sufficient to justify the approximation @xmath303 , as we checked numerically , by using the solitonic solution of ( 6 ) . the relative error is less than @xmath304 for @xmath305 $ ] and @xmath306 . upon inserting the lindblad operator @xmath307 into ( [ eq : jump ] ) we find that the jump operator takes the form @xmath308 with normalization @xmath309 . we again consider states of the form ( [ eq : initialstate ] ) which are superpositions of non - overlapping ( [ eq : a1 ] ) and localized ( [ eq : a2a ] ) wave packets @xmath310 . under this assumption , one can evaluate the expectation value in ( [ eq : jumpoperator ] ) @xmath311 such that the state @xmath312 into which the system may jump takes the form @xmath313 later we will choose the initial wave packets @xmath266 to be solitons @xmath314 . let us therefore assume that the @xmath97 s form a basis , such that @xmath315 can be represented as @xmath316 . then the transformed coefficients @xmath317 can be evaluated by the overlap @xmath318 similarly , one can evaluate the rate ( [ eq : rate2 ] ) associated to the jump operator of collisional decoherence , @xmath319 the above approximation ( [ eq : expectation ] ) further simplifies this expression , @xmath320 one observes that this rate vanishes for the stable fixed points @xmath321 , indicating that the quantum trajectories of the orthogonal unraveling evolve into the pointer states , the solitonic solutions of ( [ eq : ngcd ] ) . moreover , we note that the original stochastic process ( [ eq : nonlinear ] ) , ( [ eq : rate2 ] ) , ( [ eq : jump ] ) which is defined in the infinite dimensional hilbert space of the system has been reduced to a stochastic process in @xmath322 , demonstrating the efficiency of the orthogonal unraveling . however , due to the finite pointer width ( [ eq : sizedimensionless ] ) the _ exact _ expression for the jump rate ( [ eq : exactrate ] ) does not vanish identically , although it is very small compared to @xmath2 . for instance , the numerically obtained soliton displays a strongly suppressed total jump rate @xmath323 of @xmath324 for @xmath325 , while the superposition state decays with the rate @xmath326 . this implies that the solitons are not perfect pure state solutions of the master equation ( [ eq : cd ] ) , though the loss of purity is small . the previous section showed that the orthogonal unraveling of an initial superposition state subject to collisional decoherence can be reduced to a stochastic process with respect to the corresponding coefficients . in particular , this applies to the case where the initial state is a superposition of pointer states , @xmath327 thus we can now verify , by using the discrete process defined by eqs . ( [ eq : coefficients ] ) , ( [ eq : process2 ] ) and ( [ eq : process3 ] ) , that after decoherence the statistical weights of the pointer states are given by the overlap of the initial state with the initial pointer states . more specifically , this demonstrates that the initial state @xmath328 evolves into the mixture @xmath329 where the statistical weights are given by the overlap @xmath330 we first present an analytic proof of the above for @xmath331 . the general case , @xmath332 , is then treated numerically in the following section . we consider the expectation value for the coefficients after a jump , that is @xmath333 upon inserting ( [ eq : process2 ] ) one obtains @xmath334 , with the normalization constant @xmath335 , see ( [ eq : jumpoperator ] ) . using @xmath336 , we find @xmath337 which implies @xmath338 this shows that after an average jump the moduli of the coefficients are simply interchanged . this property ( which does not hold for @xmath339 ) makes the stochastic process analytically tractable , not least because the dynamics is independent of the phases of the coefficients . since the deterministic part ( [ eq : coefficients ] ) of the evolution is monotonic for @xmath340 , a trajectory starting from @xmath341 will end up in the state @xmath342 if and only if an odd number of jumps occurs in the process . this is demonstrated in fig . crucially , the jump rate @xmath343 , @xmath344 is unaffected by the jump ( [ eq : averagejump ] ) at all times , since it is invariant under interchanging the coefficients . hence , the time dependence of the jump rate is identical for all trajectories , which , in turn , means that the number of jumps follows an inhomogeneous poisson process . therefore , the probability for an odd number of jumps , which is equal to the statistical weight @xmath345 of the pointer state @xmath346 , is given by @xmath347 with @xmath348 the integrated jump rate . the latter can easily be evaluated by noting that ( [ eq : coefficients ] ) can be written for @xmath340 as @xmath349 upon inserting this result into ( [ eq : totalrate ] ) we obtain the integrated jump rate : @xmath350 noting ( [ eq : probodd ] ) we thus find the probability for an odd number of jumps @xmath351 this finally confirms that the statistical weights of the pointer states are indeed given by the expected overlap ( [ eq : distribution ] ) . the stochastic process is much more involved if the initial superposition consists of more than two pointer states . our numerical implementation of the stochastic process defined by ( [ eq : coefficients ] ) , ( [ eq : process2 ] ) and ( [ eq : process3 ] ) is based on a metropolis - hastings algorithm to draw the momentum transfer @xmath65 in accordance with the rate ( [ eq : process3 ] ) , with @xmath147 a gaussian . each of the generated trajectories ends asymptotically in one of the fixed points corresponding to a pointer state , and we thus obtain a numerical estimate of the statistical weights by means of the relative frequencies @xmath352 , of the asymptotic states . to confirm the expected probability distribution @xmath353 , we evaluate the relative entropy @xmath354 between these two distributions . figure 9 shows the result for a random initial state with @xmath355 as a function of the number of trajectories @xmath356 , indicating convergence to zero . i**n addition , we found for @xmath357 random initial states , with random @xmath358 , based on @xmath359 trajectories that the relative entropy was always less than @xmath360 . this holds both for cases where the initial wave packets @xmath361 are far apart such that the localization rate is saturated , @xmath362 , and for situations where the wave packets are located close together such that @xmath363 . we conclude that the asymptotic trajectories are indeed distributed according to the expected overlap ( [ eq : distribution ] ) . as an alternative confirmation of the statistical weights , we performed a @xmath364-test . similar to the treatment above , @xmath357 random initial states @xmath365 , with random @xmath366 , were drawn by the simplex picking method . for each random state , @xmath367 trajectories were generated , each of which ends asymptotically in one of the pointer states . using the observed relative frequencies @xmath368 , of the pointer states , we evaluate @xmath369 for each random state . in order to verify that the pointer states are distributed according to @xmath370 , the set @xmath371 must be shown to be sampled from a @xmath372-distribution with @xmath373 degrees of freedom . comparing the set @xmath371 with the @xmath374-quantiles * * quantile is the value such that @xmath375 of the samples lie below @xmath376 . ] ( denoted by @xmath377 ) of the corresponding @xmath364-distribution , a typical run shows ten cases where @xmath378 , one case where @xmath379 , but not a single case where @xmath380 , as one expects if the @xmath371 are @xmath364-distributed . like above , this confirms statistically that the asymptotic trajectories are distributed according to the expected overlap ( [ eq : distribution ] ) . in this article , we related the nonlinear pure state equation discussed in @xcite to a specific orthogonal unraveling of the collisional decoherence master equation . this puts into evidence that the dynamics of a particle in an ideal gas environment can be represented by an ensemble of pure state trajectories which evolve into spatially localized pointer states . for sufficiently strong collisions with the background gas , these solitonic wave packets move according to the classical equations of motion , thus explaining the emergence of classical dynamics within the quantum framework . once the pointer state is reached by an individual quantum trajectory , the latter is no longer affected by the stochastic part of the unraveling , such that the integration of the trajectory is reduced to the solution of the classical equations of motion . this suggests that the orthogonal unraveling is an efficient algorithm for the long time solution of master equations which exhibit a pointer basis . on the other hand , also the short time solution turns out to be efficient , since the orthogonal unraveling can be reduced ( under appropriate assumptions ) from an infinite dimensional unraveling to a stochastic process in @xmath322 . future studies might consider the emergence and dynamics of pointer states in dissipative quantum systems . we note that the present work relies on the model of pure collisional decoherence which does not describe long time effects such as dissipation or thermalization . it would certainly be worth to determine the pointer states of a more involved model such as the quantum linear boltzmann equation presented in @xcite . for large mass ratios between the test particle and the gas , one expects that the pointer states then evolve according to a langevin equation , thus explaining the emergence of classical brownian motion within the quantum framework . k. hornberger , _ introduction to decoherence theory _ , in : _ entanglement and decoherence . foundations and modern trends _ , ed . by a. buchleitner , c. viviescas , and m. tiersch , lect . notes phys . * 768 * , 221276 , springer , berlin , 2009 .

massive stars are rare in the universe , due both to the low mass favored imf @xcite and their short life times . however , their high luminosities and mass loss rates make them important to the ecology of the universe . underpinning this understanding is the observational determination of intrinsic stellar parameters for the most massive stars . however , this is not a simple proposition ; except in rare cases where the distance is known reliably , it requires a detached , non - interacting binary system . we are aided by the fact that the binary fraction of o stars is nearly unity at birth , though this degrades some as the stars evolve ( e.g. mergers and dynamical interactions : @xcite ) . still , there are only @xmath3 50 o - star binaries in the milky way and magellanic clouds which meet these criteria and for which individual stellar characteristics have been established @xcite . this number may increase in the future owing to advances in interferometry , enabling one to determine stellar masses in systems such as hd 150136 ( o3 - 3.5 v((f*))+o5.5 - 6 v((f))+o6.5 - 7 v((f ) ) , @xcite , @xcite ) . right now , the small number of systems spanning the o spectral type makes calibrating models difficult , but an excellent effort to determine stellar parameters as a function of spectral type was made by @xcite . unfortunately , the problem of determining fundamental parameters becomes significantly worse for evolved o stars . it is extremely important to understand what effect leaving the main sequence has on the fundamental parameters of the star . however , there are currently only three such systems for which well - constrained masses have been derived , through the use of long baseline interferometry : @xmath4 ori a ( o9.2 ibvar nwk+o9.5 ii - iii(n ) , @xcite , @xcite , @xcite ) , hd 193322 ( o9.5vnn+o8.5iii+b2.5v , @xcite ) and my ser ( o7.5 iii + o9.5 iii+(o9.5-b0)iii i , @xcite ) . while increasing the total number of massive star systems with well known parameters is worthwhile , it is difficult as the number of candidate o stars is quite small . a much more realistic goal is finding a non - interacting example system with minimal mass loss which can be used as a template for understanding the evolution of o stars . mintaka ( @xmath0 ori ) , the westernmost belt star of orion , by virtue of its proximity ( @xmath5 pc , @xcite ) and brightness ( @xmath6 , @xcite ) , as well as encompassing the detached binary system aa , is an excellent choice . the @xmath0 ori system has been known to contain multiple components for well over 100 years @xcite . in fact , it is currently known to have at least 5 stellar components ( a full diagrammed layout of this system and its relevant parameters can be found in fig . 1 of @xcite ) . the embedded triple system , especially the short period binary @xmath0 ori aa , has received a significant amount of attention in the last 15 years . an artist s rendering of the orbits and lay out of the @xmath0 ori a can be found in fig 1 . of @xcite . * hereafter h02 ) , used spectra obtained with the international ultraviolet explorer to determine a putative secondary radial velocity curve . this , in combination with hipparcos photometry , led to a controversial maximum mass for the o - star of 11.2 @xmath7 , less than half the expected value based on its spectral type ( see @xcite ) . the system was revisited by ( * ? ? ? * hereafter m10 ) , in an attempt to understand this anomalous result . they worked under the assumption that since the tertiary component contributed @xmath3 25% of the total light ( @xcite , @xcite , @xcite , @xcite , @xcite ) , significantly more than the secondary , its presence might be able to explain the apparent discrepancy . they found that motion of the primary in combination with a wide tertiary component with spectral lines broader than that of the primary gave the false impression of orbital motion of the secondary . while m10 did improve the known orbital elements , the actual masses and radii are still poorly constrained . this is the result of a relatively weak secondary component as well as relatively poor light curve coverage and precision . this paper is part of a series which will attempt to characterize as rigorously as possible the triple system @xmath0 ori a , focusing on the o9.5ii primary aa1 . this includes analysis of the @xmath8ray spectrum by ( * ? ? * hereafter paper i ) and @xmath8ray variability by ( * ? ? ? * hereafter paper ii ) , as well as spectral modeling by ( * ? ? ? * hereafter paper iv ) . in this paper we focus primarily on the optical variability . this includes an analysis of high precision space based photometry obtained with the _ most _ space telescope as well as simultaneous spectroscopic observations , described in detail in section 2 . in section 3 , we discuss the spectroscopic variability , including a derivation of the primary radial velocity ( rv ) curve . in section 4 , we use this rv curve in conjunction with the _ most _ light curve to provide three possible binary fits at varying values of the primary mass . in section 5 , we provide a fit to the primary mass using apsidal motion confirming our results from section 4 . finally , we discuss and characterize previously unknown photometric variability within this system and report frequency and period spacings consistent with non - linear interactions between tidally excited and stellar g - mode oscillations . while we are still are unable to disentangle the secondary spectrum and its kinematics , we provide a more complete model than has been achieved up to this point . our optical photometry was obtained with the _ most _ microsatellite that houses a 15-cm maksutov telescope through a custom broad - band filter covering 35007500 . the sun - synchronous polar orbit has a period of 101.4 minutes ( @xmath9 ) , which enables uninterrupted observations for up to eight weeks for targets in the continuous viewing zone . a pre - launch summary of the mission is given by @xcite . ori taken in mid december 2012early january 2013 . error bars are smaller than the size of the points . the red line is the combination of the binary fit ( see section 4 ) and that of the secondary variations ( see section 6 ) . ] @xmath0 ori a was observed for roughly half of each _ most _ orbit from dec . 2012 through 7 jan . 2013 with the fabry - mode ( see fig . [ fig : full - lc ] ) at a cadence of 40.4 s. these data were then extracted using the technique of @xcite , and show a point - to - point standard deviation of @xmath10 mmag . we initiated a professional - amateur campaign in order to obtain a large number of high - quality optical spectra simultaneous with our _ most _ and _ chandra _ campaigns . this resulted in more than 300 moderate resolution ( @xmath11 ) spectra obtained over the 3 week period . the spectra were reduced by standard techniques utilizing bias , dark , and flat field images . wavelength calibration was accomplished through comparison with emission - lamp spectra taken before and/or after the object spectra . our main goal was originally to monitor the h@xmath12 line for wind variability and correlate any such emission with the x - ray spectra ( paper i , paper ii ) while simultaneously obtaining a measure of the orbital motion from the 6678 transition . however , no significant h@xmath12 variability was seen , except for the radial velocity changes related to the orbital motion of aa1 . furthermore , h@xmath12 could be diluted by wind emission , making rv measurements questionable . due to both the telluric absorption lines around h@xmath12 and the potential for wind emission , we focused on the he i 6678 transition for our radial velocity measurements . details of the telescopes and spectrographs used are given in table [ table : speclog ] . a typical s / n for any observation is @xmath3 100150 . the spectroscopic treatment of @xmath0 ori a from m10 and h02 are not in agreement . the m10 analysis assumed a tertiary contribution that was subtracted off , and left a single - lined binary , while the h02 analysis found evidence of the tertiary after iterating to remove first the primary and then the secondary spectrum from the combined one . the resulting masses are highly discrepant , with m10 assuming a standard mass based on spectral type of @xmath13 , while h02 found @xmath14 . treatment of the tertiary may lead to severe errors in the inferred stellar parameters . paper iv presents a quantitative photospheric and wind model for each of the three stars in the @xmath0 ori a system . the results of @xcite and @xcite show that the tertiary orbits the binary aa very slowly with a period of several hundred years , although the uncertainties in the tertiary s orbit are very large . as the period is so long , we expect no measurable kinematic motion in the tertiary component during one month of observations . therefore , if we take the model of the tertiary spectrum from the models of paper iv , then we can subtract its contribution a priori from the spectra in order to best constrain the binary properties of aa1,2 . this is similar to what was done by m10 . if the tertiary is not taken into account , then our results are in agreement with h02 . however , the optical spectrum consists of @xmath3 25 % contribution from the tertiary ( @xcite , @xcite , @xcite ) , so even with our s / n @xmath3 100150 , this contribution should always be seen in our spectra . we convolved the calculated spectrum of the tertiary to the resolution of our data , and can immediately see that it is the likely source of excess absorption . in figure [ tert_oplot ] , we show two example profiles of the he i 6678 line ( with large red and blue shifts from the primary s orbital motion ) along with the weighted contribution of the tertiary s spectrum calculated by paper iv . for the redshifted spectrum , we see that the red edge of the profile has an excess of absorption that matches the theoretical spectrum of the tertiary quite well . on the blue edge of the profiles , blending with he ii 6683 becomes more problematic . while the tertiary seems to reproduce some of the absorption , most of blue excess comes from the primary star s he ii line . after we subtracted the weighted contribution from the tertiary star , we cross correlated our observations against the model of the primary calculated by paper iv . from these velocities we calculated a single - lined orbit with the period derived by m10 , as the length of our data sets was insufficient for determining a more precise period . most velocities had uncertainties on the order of @xmath15 km s@xmath16 from this determination . data for all the velocities obtained are given in table 4 . we examined the residuals after removing the primary star s spectrum ( both through shift - and - add and from model subtraction using the spectrum calculated by paper iv ) . we found no residuals due to the secondary in the data at this point . this is consistent with the observational results of m10 as well as the results of the modeling of paper iv . in particular , the modeling results predict that only @xmath17 of the light in this wavelength region comes from the secondary , so the noise associated with the weak features would dominate , as the individual spectra have a signal - to - noise on the order 100150 on average . in order to reliably see the secondary with a s / n of 50 , we suspect that the combined spectrum of the three component stars needs to have a s / n @xmath18 1000 . a complete solution for a binary system requires a substantial amount of information . in general , it requires an eclipsing light curve and two radial velocity curves . despite being studied numerous times , no one has been able to reliably determine all three of these quantities for @xmath0 ori aa . this is primarily due to the faintness of the secondary with respect to both the primary and the tertiary . despite the lack of a secondary rv curve , data have two significant advantages over the previous works of h02 and m10 : high precision photometry along with simultaneous spectroscopy . since there is known apsidal motion in this system ( see section 5 ) having all photometry and spectra taken all within a month timespan mitigates this issue . for simulation and fitting of the @xmath0 ori system we used the alpha release of phoebe 2.0 ( @xcite , @xcite ) , an engine capable of simulating stars , binaries , and concurrent observational data . unlike the original phoebe , this is not built directly on the work @xcite and has been rewritten to take into account many second - order effects . this is especially relevant when dealing with spaced - based data . before this system was fit , some initial processing was required . as shown in fig . [ fig : full - lc ] , the light curve of @xmath0 ori is composed of an eclipsing light curve with second order variations superimposed . these underlying variations can not be easily or completely removed ( see section 6 ) so we must lessen their effects in order to get a more accurate representation of the binary variability . this can be done by first phasing on the binary period , which requires a precise ephemeris for the system . since our data span only 3 weeks ( about 3.5 orbits ) , the period determined by m10 is much more precise than what we are able to obtain , so we adopt their value @xmath19 d. for @xmath20 and @xmath21 we adopted our own derived values . for @xmath21 we determined a value from fits to the radial velocity curve of @xmath22 hjd . for @xmath20 we are actually able to obtain a value from o - c calculations of @xmath23 hjd . we should note that these numbers offer no specific improvement in the associated error over those derived in m10 , and in the case of @xmath20 the error is slightly worse . however , unlike m10 our method for determining @xmath20 is independent of the binary fit . this is important as second order stellar effects ( unknown in previous works ) make determinations from the binary fit less reliable . in addition , apsidal motion is significant in this system , and we did not want it to affect our results . using the revised ephemeris , the data were phased and then binned to provide a mild smoothing , which yields a template for the binary variations . in this template , the error is much smaller than the size of the points . however , incomplete removal of the secondary variations leads to scatter at the milli - mag level ( see fig . [ phlc ] ) . ori phased on the orbital period . it is then binned to 0.0125 in phase to smooth out remaining secondary variations . the bottom x - axis is phased to time of minimum light and the top is phased to periastron . ] in addition , modeling of stars and binary systems requires the presence of realistic atmosphere models . unfortunately , the kurucz atmosphere tables are the only ones which have been incorporated into phoebe 2.0 at this stage and they are insufficient for o stars . therefore , our best option is to use blackbody atmosphere models . in some cases this would not be adequate , but for o stars we sample the rayleigh jeans tail at optical wavelengths , making the differences from a blackbody minimal . in our specific case , a comparison with non - lte models calculated for the components of the system indicates a deviation of @xmath24% in the integrated flux across the _ most _ bandpass ( paper iv ) . therefore , this assumption will not significantly affect our solution as all three stars have similar temperatures and deviations from a blackbody . ) . in the bottom panels are the o - c residuals for the corresponding light and radial velocity curves . the intrinsic stellar variations that we are unable to completely remove are the main source of error . only one model is plotted as the differences are nearly impossible to distinguish at this scale . ] phoebe 2.0 has several built in fitting routines , but for the purposes of this project we decided to use the monte carlo markov chain method ( mcmc ) . the specific implementation used by phoebe 2.0 is emcee @xcite , an affine invariant mcmc ensemble sampler based on @xcite . while this routine requires a good initial fit to converge to a solution in a reasonable number of iterations , it gives the most comprehensive analysis of both the fit and the parameter correlations . mcmc also has become a standard in the fitting of photometric time series ( e.g. @xcite and @xcite ) . for a more general overview of this technique , see @xcite . however , this fitting method is not particularly useful without a full complement of constraints . since we have no secondary rv curve , we must put an extra constraint on the mass . we know the spectral type of the primary star quite well and can make a reasonable assumption that the mass is @xmath25 @xcite , further supported by non - lte modeling of the system by paper iv . unfortunately , as there are several viable sets of parameters , the fit was unable to converge to a single minimum . therefore , we divided this mass range further into three separate ranges : low mass primary ( 21 - 24 @xmath7 ; lm ) , medium mass primary ( 24 - 27 @xmath7 ; mm ) , and high mass primary ( 27 - 29 @xmath7 ; hm ) . these mass ranges were chosen based on previous results , as the system would converge to one of three possible solutions , with one falling in each bin . for each possible solution we iterated at least 800 times . while this may seem insufficient to converge to a true solution , we were not starting from a random point in the parameter space . parameters were adjusted so that the overall shape and amplitude were consistent with the observed light curve and rv curve before beginning mcmc . indeed , after about 200 iterations the log probability function defining our goodness of fit is virtually constant . a representation of the quality of these fits for the lm model is shown in fig . [ fig : bin - fit ] with the upper panels showing the phoebe fit to the light curve ( left ) and rv curve ( right ) . the lower panels show the corresponding residuals . while there are differences between the three models , they are impossible to distinguish on this scale and so only the lm model is shown . the only motivation for this model over the others is that its primary mass agrees best with what is derived from apsidal motion ( see section 5 ) . one can see a comparison of the light curve residuals of each of these models in fig . [ fig : res - comp ] . this figure demonstrates the differences inherent in the residuals of each model . these differences are small but still significant , especially within the eclipses . however , it is impossible to determine quantitatively which solution is best as the standard deviation of the residuals for each fit are the same to within @xmath26 . it should also be noted that phoebe 2.0 is still undergoing rigorous testing , so while our models seem reasonable , we also simulated our solutions with the original phoebe @xcite to ensure that the results were consistent . table [ table : bin - fit ] shows the complete solutions for each model as well as comparisons to previous solutions by h02 and m10 . while the spurious detection of the secondary rv biases the h02 results , there is good agreement between our work and that of m10 . one point to note is the improvement made to the inclination angle . while each of the previous works had fits at substantially different inclinations from 67@xmath27 - 81@xmath27 , our work shows that this inclination is quite firmly centered at 76@xmath28 regardless of the fit . in addition all of the orbital parameters are well constrained . it is very difficult to say if the high , low , or medium mass fit is the best . the probability function is virtually identical for the three fits , and the only noticeable differences are those for the masses and radii of the primary and secondary . in both the rv and light curve solutions the residuals are due more to the accuracy of the points than in any differences in the model ( see fig . [ fig : bin - fit ] and fig . [ fig : res - comp ] ) . therefore , we must allow for a family of solutions until we are able to retrieve a secondary rv curve . with no clear way to select a best fit model , we adopt the lm fit throughout this paper when we need to apply a model . in close binaries such as @xmath0 ori aa , the deformations of the stars in the elliptical orbit cause the longitude of periastron ( @xmath29 ) to precess with time , which is referred to as apsidal motion . the amount of this precession depends on the mass , internal structure of the stars , and the orbital eccentricity . @xcite and @xcite made assumptions for structure constants of massive stars and then used the measurements of apsidal motion to calculate stellar masses . @xmath0 ori a has been studied extensively for more than a century and shows signs of significant apsidal motion . the work of @xcite compiled all determinations of the spectroscopic orbit . this is combined with more recent studies of h02 and m10 along with our own work to retrieve orbital information spanning over 100 years . we use these derived orbital elements to calculate the apsidal motion , @xmath30 , using a weighted linear least - squares fit to @xmath29 as a function of time . the results shown in figure [ apsidal ] present a reasonable fit to the measurements and provide a value of @xmath31 ori . a weighted linear fit to the published values ( @xcite , h02 , m10 ) yields a measurement of @xmath31 ] with this measurement of the apsidal advance , we can make some assumptions about the internal structure constants of the primary star . this relies heavily on the age of the system . the age of @xmath0 ori is likely best constrained by the study of @xcite who find @xmath0 ori aa1 to be the brightest star in a small cluster . their study involved deep imaging to understand low - mass stars , and their color - magnitude diagrams imply an age of @xmath24 myr . in order to continue with the calculation we adopt our orbital parameters from 4.1 . combining these and the models of @xcite , we obtained the results shown in fig . [ apsmass ] , which give the stellar mass as a function of age for our derived apsidal motion . with an age of 5 myr , we see that the derived mass of the primary is @xmath33 . this value agrees with the spectral type / mass @xcite and our low mass model ( table [ table : bin - fit ] ) . @xcite estimate the age of the cluster containing @xmath0 ori , to be 5 myr , corresponding to a mass of @xmath33 consistent with our low mass model . the different lines represent the model determinations with the extremes of the apsidal motion from the errors ( fig . [ apsidal ] ) . ] it is clear from fig . [ fig : full - lc ] that the dominant signal in the light curve comes from binarity . it is also apparent from the non - phase - locked variations that additional signals are present in @xmath0 ori a. the presence of such variation is not completely unexpected as @xcite reported spectral variability in this system . however , the variation timescales seen in fig.[fig : full - lc ] are much longer than the @xmath3 4h periodicities they report . before addressing this variation , we first bin our light curve with the _ most _ orbital period ( 0.074 d ) . this allows us to mitigate the problem of light scattered into the _ most _ optics ( mainly due to the bright earth ) . as a consequence , we ignore variability on time scales shorter than a single orbit . once this is done , we must remove the eclipsing binary signal . this variability has already been analyzed in the previous section , and hinders our ability to study the additional variations present . there are two options available for removal of the binary signal . we could use the binary template derived earlier ( see fig . [ phlc ] ) or the lm model fit to this template . the decision of which binary light curve to use is not obvious , so the analysis was done using both subtractions . the results are largely the same , and we chose the lm model fit subtraction as the amplitude of the significant peaks was higher and there were slightly more significant frequencies . this choice is supported by the agreement between the subsequent analysis and the full light curve ( see fig . [ fig : full - lc ] , red line ) . the residuals are shown in fig . [ fig : resids - lc ] ( blue dots ) . ori a ( black).the spectral window is plotted in green and shows no significant aliasing . the fourier transform after pre whitening is shown in blue . the red line at 1.49 millimag represents the @xmath34 significance . the inset is a zoom in of the fourier transform from 02 c / d . whitened frequencies are labeled with arrows ( see section 6 ) . note that some frequencies are not immediately apparent without removal of larger amplitude peaks . ] a common method for finding and classifying periodic signals is known as pre - whitening . in this method peaks are detected in the fourier transform and then a sinusoid with the corresponding period , amplitude , and phase is subtracted from the light curve . this is repeated , and a combination of sinusoids representing each frequency are fitted and removed until there are no more significant peaks present . this method only works if the signals are sinusoidal ( or mostly sinusoidal ) and in the case of @xmath0 ori a , this is not obvious . the fourier transform of the @xmath0 ori a light curve is shown in fig . [ fig : del - ft ] ( black ) using period04 @xcite . we apply pre whitening to the transform and the end product is shown in blue . in order to determine which frequencies to include as significant , we use an empirically determined amplitude - to - noise ratio of 4 @xcite . there are 12 frequencies ( black arrows ) above this significance threshold @xmath35 ( red line in fig . [ fig : del - ft ] ) . the parameters associated with each frequency are given in table [ table : ft - peaks ] . the fit of these frequencies to the residuals , which is shown in fig . [ fig : resids - lc ] ( black line ) , does a reasonable job of fitting the variations . while @xcite attribute the variability they find to non - radial pulsations , we see no significant variability in the spectrum around 4 hrs ( 6 c / d ) . one possible explanation of extra periods is the presence of rotational modulation through starspots . while this is rarely considered in massive stars , the models of @xcite lend some credibility to this idea . moreover , the work of @xcite , @xcite , and @xcite show evidence for the occurrence of this phenomenon among o stars . if rotational modulation is the root cause of the variation , then the largest frequencies would occur near the rotation period of the star . the vast majority of the remaining peaks could be explained as spurious signals due to the variable nature of star spots ( e.g. lifetimes and number ) as explored extensively by @xcite for the o7.5 iii(n)((f ) ) star @xmath36 per . in our case the period of the largest amplitude signal is 4.61 d. if we assume this this is the rotation period of @xmath0 ori aa1 , then combined with the @xmath37 ( paper iv ) and the assumption of alignment between the orbital and rotation axis , this yields a radius of @xmath38 for the primary aa1 . this is smaller than the value found in our models ( see tab . [ fig : bin - fit ] ) and closer to that of a luminosity class iii star @xcite as opposed to class ii which @xmath0 ori aa1 is known to be . however , it is unclear what the uncertainty in radius would be in the models of @xcite and so this peak remains roughly consistent with rotational modulation . indeed , the phase folded light curves of the two strongest frequencies ( see fig . [ fig : period - lcs ] ) show evidence of non - sinusoidal behavior . this could be a consequence of a highly non - sinusoidal signal such as rotational modulation . while this hypothesis is plausible , similar variations can also be caused by pulsations . while rotational modulation would result in a forest of peaks that show no clear relationship to each other @xcite , pulsations should show evidence of coherent spacings . all the periodicities we observe are several hours to days long , therefore they would fall in the g - mode regime for all three massive stars seen in this system @xcite . these modes have the unique quality that when they have the same @xmath39 number they are equally spaced in period . further explanation can be found in any asteroseismology review ( e.g. @xcite . now , we can check for period spacings by creating an echelle diagram where the period of each variation is phased modulo a given spacing and plotted . equally spaced periods will lie roughly along a vertical line . the echelle diagram in fig . [ fig : echelle ] shows that at least half of the frequencies share a common period spacing . the results of this analysis provide evidence for pulsations , but this evidence is no more convincing than the idea of rotational modulation . however , there also appear spacings in the frequency domain . curiously , 10 of the 13 frequencies show a spacing at two or 3 times the orbital frequency ( see tab . [ table : spacings ] ) . this spacing suggests that there are tidal interactions associated with these signals . this is not unprecented as the work of @xcite , @xcite and @xcite also show some evidence of effects caused by tidal forces in massive binaries . this variability , though , is unlikely to arise from tidally excited oscillations , which occur at integer harmonics of the orbital frequency ( @xcite , @xcite , @xcite ) , in contrast to the oscillations we observe . however , frequency spacing at multiples of the orbital frequency has been seen in several other close binary stars ( @xcite , @xcite , @xcite ) , and indicates that the pulsations are tidally influenced . " similar to the papers cited above we speculate that the frequency spacing at multiples of the orbital frequency may arise from non - linear interactions between stellar oscillation modes and tidally excited oscillations . moreover , we interpret the frequency spacing as evidence that the observed variability is generated by stellar oscillations rather than by magnetic activity or rotational modulation . the main caveat to this hypothesis is that our frequency resolution is limited to 0.047 d@xmath16 due to the short span ( 3 weeks ) of the observations . this means that there are several spacings that could have been deemed significant . for instance @xmath40 and @xmath41 would both , within error , fit the same spacing because of how close the two peaks are in frequency . when this happens , we choose the peak that resulted in a spacing closest to a multiple of the orbital frequency . additionally , the two frequencies which do not appear in tab . [ table : spacings ] , @xmath42 and @xmath43 , are actually split by the orbital frequency . however , since the uncertainty is half the actual spacing , we choose not to include it . despite these issues , this frequency spacing is pervasive , appearing strongly between nearly every significant peak . while there is still uncertainty in this claim , it can be lifted with the addition of longer time baseline observations . another problem is that even if these are tidally influenced " oscillations , they still do not completely account for all the variation in this system . it is possible that these discrepancies could be due to amplitude modulation , though the ratio of period versus observation length makes this unlikely . it is also possible that there is also rotation modulation in this system from one of the components , just at a lower level than the pulsations . it is worth noting that the nature of the significant periodicities found in the lowest frequency range of the _ most _ data set is uncertain and could be a result of red noise. normally , this noise could be modeled and removed using autoregressive moving average techniques . however , our data is not stationary which is a requirement for this type of modeling . attempts at making the data stationary disrupt the relevant signals ; not only in our data , but also in the synthetic data using combinations of purely sinusoidal signals . therefore , it is impossible to determine unequivocally the root cause of this variation without the addition of longer baseline photometric data ( see section 8) . the system @xmath44 ori a remains an ideal candidate to define the intrinsic parameters of evolved o stars . the fundamental parameters of the primary , @xmath0 ori aa1 , appear consistent with models from paper iv as well as the results from modeling of the apsidal motion . our analysis removes ambiguity in the derived stellar parameters and makes the primary more consistent with stellar models , without having to appeal to non - conservative mass loss or mass transfer that might play a role in the evolution of a close binary like @xmath0 ori aa . despite these consistencies , @xmath0 ori aa has a rather larger eccentricity of @xmath45 , which is unusual for a close o - star binary system with a moderate age of @xmath3 5 myrs . in such a situation , circularization of the orbit should already have occurred . we can check this using the procedure outlined in @xcite with one key assumption : knowledge of the stellar structure constant @xmath46 . this constant , which is dependent on the interior composition and structure of the star , requires extensive effort to calculate precisely , but can be approximated by the fractional core radius to the eighth power . we have no evolutionary models for @xmath0 ori aa so we have no way to calculate or estimate @xmath46 directly . we do know from @xcite that @xmath46 tends to increase as a function of mass , so we will take the value of @xmath47 which is that of a 15 @xmath7 luminosity class v star . we choose this value as this is the largest mass star for which a value of @xmath46 was derived . this is clearly an underestimate of @xmath46 which is ameliorated slightly by the fact that @xmath0 ori aa is a slightly expanded luminosity class ii star ( implying a smaller fractional core radius ) . it is also clear based on the inverse dependence of the circularization time with @xmath46 that the value we obtain will be a maximum . a quick calculation reveals a circularization timescale of @xmath48myr , which is well short of the age estimate that has been derived for this system . this result shows that our knowledge of this system is incomplete and could indicate that our assumed primary mass is incorrect . however , this discrepancy could also be caused by incorrect age calculations , strong interactions with the tertiary component , or issues with our knowledge of stellar structure in massive stars . at the same time , there is currently no evidence which strongly supports any of these hypotheses , so we are confident in the assumptions made . the most intriguing aspect of this system is the pronounced second order variations in its light curve . such variations have never been seen before in this system , and the frequency spacings seem suggestive of tidally induced pulsations . if confirmed , this would be an important discovery as these modes have been seen rarely if ever in massive stars . this could lead to asteroseismic modeling and increase our understanding of the interiors of both massive stars in binaries , and massive stars as a whole , especially considering the extremely high fraction of these stars in binaries @xcite . while we have made significant progress towards understanding this system , it is clear that we are not at the level we need to be . our ultimate goal is for @xmath0 ori to be a test for both evolutionary and structural models of massive stars . for this objective to be realized , we must determine fundamental parameters to high precision . for this reason it is imperative that we identify the presence of the secondary component within the spectrum . in addition , we must improve our frequency resolution within the fourier spectrum . this will require a long time - series of high - precision photometry . as a result , we could identify with confidence the source of the variation , which could possibly allow for asteroseismic modeling . a secondary consequence of this would be an improved binary light curve with significantly reduced scatter from which to model our system . this would also lead to improved values of fundamental parameters , and place constraints on the apsidal advance of the system . this analysis to understand @xmath0 ori , in conjunction with papers i , ii , and iv , has highlighted the need to obtain the spectral characteristics and rv curve of @xmath0 ori aa2 . we will use both the _ hubble space telescope _ and _ gemini - north _ to obtain high - resolution spectra of the binary at opposite quadratures while spatially separating the tertiary and observing its spectrum . these observations are scheduled for the upcoming observing season , and we also plan to obtain higher signal - to - noise optical spectroscopy from the observatoire de mont mgantic . these observations will provide excellent constraints on the modeling of the system and reveal any unusual spectroscopic variability seen in the tertiary star . this will provide more constraints on the photometric variability seen with _ most _ . a better characterization of this variability as well as the binary parameters requires access to long time baseline photometry , which _ most _ is not able to provide . the _ brite - constellation _ project , however , is designed to provide 6 months of continuous coverage in both blue and red filters @xcite . _ brite - constellation _ began initial data acquisition of the orion constellation in october of 2013 . since this was the initial data taken , the full 6 months of continuous coverage was not achieved . however , there are nearly three continuous months of data in both filters . these data should be released to the collaboration in early 2015 . in addition , a second run with _ brite - constellation _ began in oct . 2014 that should provide a full 6-month time - series . mfc , jsn , wlw , and kh are grateful for support via chandra grant go3 - 14015a and go3 - 14015e . yn acknowledges support from the fonds national de la recherche scientifique ( belgium ) , the communaut franaise de belgique , the prodex xmm and integral contracts , and the action de recherche concerte ( cfwb - acadmie wallonie europe ) . ndr gratefully acknowledges his craq ( centre de recherche en astrophysique du qubec ) fellowship . afjm , dbg , jmm and smr are grateful for financial aid to nserc ( canada ) . afjm and hp would also like to frqnt ( quebec ) and the canadian space agency . jma acknowledges support from [ a ] the spanish government ministerio de economa y competitividad ( mineco ) through grants aya2010 - 15081 , aya2010 - 17631 , and aya2013 - 40611-p and [ b ] the consejera de educacin of the junta de andaluca through grant p08-tic-4075 . rk and ww acknowledge support by the austrian science fund ( fwf ) . nre is grateful for support from the chandra x - ray center nasa contract nas8 - 03060 . jlh acknowledges support from nasa award nnx13af40 g and nsf award ast-0807477 . lcccccc christian buil & 4 & c9 ( 0.23 m ) & eshel & atik460ex & 11,000 + christian buil & 6 & c11 ( 0.28 m ) & eshel & atik460ex & 11,000 + thierry garrel & 70 & meade lx200 14 ( 0.35 m ) & eshel & sbig st10xme & 11,000 + keith graham & 2 & meade lx200 12 ( 0.30 m ) & lhires iii & sbig st8xme & 15,000 + bernard heathcote & 32 & c11 ( 0.28 m ) & lhires iii & atik314l+ & 13,000 + thierry lemoult & 5 & c14 ( 0.36 m ) & eshel & st8xme & 11,000 + dong li & 16 & c11 ( 0.28 m ) & lhires iii & qhyimg2pro & 15,000 + benjamin mauclaire & 8 & sct 12 ( 0.30 m ) & lhires iii & kaf-1603me & 15,000 + mike potter & 201 & c14 ( 0.36 m ) & lhires iii & sbig st8 & 11,000 + jose ribeiro & 13 & c14 ( 0.36 m ) & lhires iii & sbig st10 & 15,000 + asiago & 4 & 1.82 m & reosc echelle & & 22,000 + observatory uc santa martina & 33 & 0.5 m & pucheros echelle & fli pl1001e & 20,000 + mcdonald & 5 & 2.7 m & cross - dispersed & & 60,000 + nordic optical telescope & 4 & 2.5 m & fies & & 46,000 + calar alto & 3 & 2.2 m & caf echelle & & 65,000 + [ table : speclog ] @xmath49(@xmath7 ) & & 23.81 & & 24.20 & & & & 27.59 & & 25 & & & & 11.2@xmath501.8 + @xmath51(@xmath7 ) & & 8.54 & & 8.55 & & & & 9.27 & & 9.91 & & & & 5.6@xmath500.4 + @xmath52(@xmath53 ) & & @xmath54 & & @xmath55 & & & & @xmath56 & & 15.6 & & & & 13 + @xmath57(@xmath53 ) & & @xmath58 & & @xmath59 & & & & @xmath60 & & @xmath61 & & & & @xmath62 + @xmath63(@xmath64 ) & & @xmath65 & & @xmath65 & & & & @xmath65 & & @xmath65 & & & & @xmath66 + @xmath67(@xmath64 ) & & @xmath68 & & @xmath69 & & & & @xmath70 & & 23835 & & & & + @xmath71+@xmath72(hjd ) & & @xmath73 & & @xmath73 & & & & @xmath73 & & @xmath74 & & & & + @xmath20+@xmath72(hjd ) & & @xmath75 & & @xmath75 & & & & @xmath75 & & @xmath76 & & & & + @xmath77(days ) & & @xmath78 & & @xmath78 & & & & @xmath78 & & @xmath79 & & & & @xmath80 + @xmath29 & & @xmath81 & & @xmath82 & & & & @xmath83 & & 140@xmath501.8 & & & & + @xmath84(@xmath53 ) & & @xmath85 & & @xmath86 & & & & @xmath87 & & 44@xmath500.3 & & & & + @xmath88 & & @xmath89 & & @xmath89 & & & & @xmath90 & & 0.0955@xmath500.0069 & & & & 0.075@xmath500.06 + @xmath91 & & @xmath92 & & @xmath93 & & & & @xmath94 & & 73.6@xmath500.3 & & & & 77 + @xmath95(km s@xmath16 ) & & @xmath96 & & @xmath97 & & & & @xmath98 & & 21.7@xmath500.5 & & & & + @xmath99(km s@xmath16 ) & & 96.02@xmath500.60 & & 96.02@xmath500.60 & & & & 96.02@xmath500.60 & & 106.33@xmath500.71 & & & & 94.9@xmath500.6 + @xmath100 & @xmath101 & @xmath102 & @xmath103 & @xmath104 & + @xmath105 & @xmath106 & @xmath107 & @xmath108 & @xmath109 & + @xmath110 & @xmath111 & @xmath112 & @xmath113 & @xmath114 & + @xmath42 & @xmath115 & @xmath116 & @xmath117 & @xmath118 & + @xmath43 & @xmath119 & @xmath120 & @xmath121 & @xmath122 & + @xmath123 & @xmath124 & @xmath125 & @xmath126 & @xmath127 & + @xmath128 & @xmath129 & @xmath130 & @xmath131 & @xmath132 & + @xmath133 & @xmath134 & @xmath135 & @xmath136 & @xmath137 & + @xmath138 & @xmath139 & @xmath140 & @xmath141 & @xmath142 & + @xmath143 & @xmath144 & @xmath145 & @xmath146 & @xmath147 & + @xmath148 & @xmath149 & @xmath150 & @xmath141 & @xmath151 & + @xmath152 & @xmath153 & @xmath154 & @xmath146 & @xmath155 & + @xmath156 & @xmath157 & @xmath158 & @xmath159 & @xmath160 & + [ table : ft - peaks ] @xmath128@xmath148 & 0.337 & @xmath161 + @xmath133@xmath156 & 0.347 & @xmath161 + @xmath105@xmath138 & 0.347 & @xmath161 + @xmath138@xmath110 & 0.358 & @xmath161 + @xmath133@xmath143 & 0.522 & @xmath162 + @xmath100@xmath110 & 0.520 & @xmath162 + [ table : spacings ]

accelerating a thermonuclear flame to a large fraction of the speed of sound ( possibly supersonic ) is one of the main difficulties in modeling type ia supernovae ( see the recent review by @xcite and references therein ) . numerical results have shown that a prompt detonation does not yield the right abundance pattern to account for the observations @xcite ; consequently , the burning must begin as a subsonic flame . one dimensional simulations show that the flame needs to reach approximately one third of the speed of sound to properly account for the explosion energetics and nucleosynthetic yields @xcite , although , a transition from a deflagration to a detonation at some late stage of the explosion @xcite is not ruled out . laminar flame speeds are too slow by several orders of magnitude @xcite , so instabilities and the interaction with flame - generated turbulence are turned to in hopes of providing a mechanism to considerably accelerate the flame . in this paper , we focus on the effects of the rayleigh - taylor ( rt ) instability @xcite on the flame . after ignition , the subsonic flame moves outward from the center of the star . the pressure is essentially continuous across the flame . furthermore , since the star has time to expand in response to the burning , this pressure also remains constant in time . as the flame propagates , it leaves behind hot ash that is less dense than the cool fuel . gravity points toward the center of the star , so this flame front is rt unstable . bubbles of hot ash will rise and try to exchange places with spikes of cool fuel , increasing the surface area of the flame and leading to an enhancement of the flame speed . it is possible that this acceleration , operating on the stellar scale , can bring the flame speed up to a significant fraction of the speed of sound . furthermore , the shear layer between the fuel and ash which develops as the instability evolves is kelvin - helmholtz unstable , creating turbulence in the region of the flame . as the flame surface is wrinkled by the rt instability , it will interact with this turbulence . we look in detail at the interaction between the rt instability and the flame on small scales through 2-dimensional spatially resolved simulations in conditions appropriate to the late stages of a type ia supernova explosion . the difference in scale between the white dwarf and the flame width is enormous ( up to 12 orders of magnitude ) , making direct numerical simulations of the whole explosion process impossible . nevertheless , simulations of rt unstable flames have been attempted for some time , both on the full stellar scale and as smaller , constant density micro - physical studies . various techniques and approximations are used to follow an unresolved flame on these scales . some of the earliest such calculations @xcite used donor - cell advection to follow a combustion front through the star , burning most of it , and releasing enough energy to produce an explosion . in their model , the flame was propagated essentially through numerical diffusion . the rt instability was resolved only on the very largest scales of the star , missing the turbulence generated on the small scales . @xcite used a two - dimensional implicit lagrangian - remap hydrodynamics code with a mixing - length subgrid model to follow a rt unstable flame . the diffusion of reactants in the lagrangian - remap scheme is much lower than that used in @xcite , so a more physical flame speed could be used . the flame speed was chosen to be either the conductive ( laminar ) or turbulent speed which ever was larger . again , the whole star was modeled , with coarse resolution , showing that the rt instability sets up and greatly convolutes the flame front , but the resulting acceleration left the flame too slow to match observations . @xcite suggested that more resolution is needed to get the small scale physics right . the effects of turbulence on the burning were investigated by @xcite . two - dimensional , small scale rt simulations using a thickened - flame model , coupled to an explicit lagrange - remap ppm @xcite implementation to model a thin flame front were performed by @xcite and later extended to 3-d @xcite . in this model , a reaction - diffusion equation for the fuel concentration is solved along with the euler equations , with the conductivity and reaction rate modified to yield the desired laminar flame speed and a flame that was a few computational zones wide . the thermal structure of the true flame is stretched greatly , and , as a result , curvature effects of the flame are not properly included . energy is deposited behind the flame to couple this reaction - diffusion model to the hydrodynamics . similar techniques have also been used for terrestrial flames ( see for example @xcite ) . simulations of the rt instability at a density of @xmath1 lead the author to conclude that while the flame accelerates , it is too small of an acceleration ( when scaled up to the size of the star ) to account for the observations , and that a deflagration - detonation transition may be more promising . owing to the difference in behavior between two - dimensional and three - dimensional turbulence , @xcite concludes that the turbulence was more effective in increasing the flame speed in 3-d . @xcite demonstrated scaling relations for the 3-d rt unstable flame and concluded that the turbulent flame speed does not depend on the small scale burning or the laminar flame speed . the independence of the turbulent flame speed on the laminar speed was also postulated by @xcite , who performed a full - stellar model based on kolmogorov scaling of turbulence from the large stellar scales down to the flame scales . here , a turbulent scaling model using the gibson length ( the scale at which an eddy is completely burned during a single turnover ) as the small scale cutoff yields a turbulent flame velocity prediction for the grid scale . this was coupled with a subgrid model for the turbulent kinetic energy , and implemented within the ppm algorithm . their simulations of the flame propagating through the whole star showed the rt instability dominating the flow , but the flame did not accelerate enough to produce an explosion . more recently , large scale simulations of pure deflagration explosion have been performed using two different methods , level - sets @xcite and a thickened - flame model @xcite with a sharp - wheeler @xcite rt subgrid model . both groups using ppm as their underlying explicit hydrodynamics method . different numbers and locations of the ignition points are used , but the 3-d simulations by both groups release enough energy to explode the star , with the burning proceeding solely as a deflagration . these studies have demonstrated the tremendous effect the rt instability has as the dominant acceleration mechanism of the flame to the point where , in some simulations , a deflagration alone can unbind the star . however , in all the cases discussed above , some model for the flame or turbulence on scales smaller than the grid resolution was required . in this paper , we present fully resolved simulations of rt unstable flames at the densities appropriate to the late stages of the explosion no flame model is used . fully resolving the flame means that the effects of curvature and strain on the flame are implicitly accounted for . curvature effects on flames in conditions close to those studied here were found to be important in @xcite , where laminar flames were propagated in diverging and converging spherical geometries . here , the effects can be even larger as the flame will experience sharp kinks as it is distorted by the rt instability . to achieve converged steady - state laminar flames , approximately five grid point resolution in the thermal width is required . this limits the size of the domain that can be modeled significantly ( although the adaptive mesh refinement employed here helps greatly ) , but frees us from the need to deal with subgrid models . resolved simulations at the smallest scales complement the full star flame - model calculations , and it is hoped that progress from the small scales up will allow us in the future to develop and calibrate more accurate subgrid models . fully resolved multidimensional type ia - like flame simulations are rare @xcite , mainly because the flame moves so slowly , requiring excessive numbers of timesteps for compressible codes . resolved calculations of the reactive rt instability have been performed before @xcite , but using model flames ( a kpp - type reaction , @xcite ) and a boussinesq approximation , under conditions that are not directly applicable to type ia sne . when extrapolated to the astrophysical regime , they predict a flame speed independent of the laminar speed , similar to that found by @xcite . however , compressibility was neglected in these calculations . here we use a low mach number formulation of the equations of motion , which retains the compressibility , but frees us from the restrictive acoustic timestep constraint , and we use realistic input physics ( reaction rate , eos , etc . ) appropriate to the conditions in a type ia sne explosion . flame instabilities stretch and wrinkle the flame , increasing the bulk burning rate . the sne flame is subject to the landau - darrieus @xcite and rt instabilities , as well as the interaction with turbulence . in a previous paper , ( @xcite , henceforth referred to as paper i ) , we looked at landau - darrieus unstable flames , and validated the low mach number method for astrophysical flames , confirming the results of earlier studies that the acceleration of the laminar flame is quite small , while finding no evidence for the breakdown of the cusps in the non - linear regime . in the present paper , we look in detail at the interaction between the burning and the rt instability . a rt driven flame behaves very differently from its laminar counterpart . aside from looking at the increase in velocity over the laminar speed the wrinkling provides , we seek to determine whether there is a simple scaling relation that matches these results , and whether such a relation holds over the whole range of densities , where the character of the burning is expected to change dramatically . in [ sec : distributedburning ] , we discuss the effects of the burning on the rt instability when we enter the distributed burning regime . in [ sec : numerics ] , the low mach number numerical method and input physics are described , followed by the results of our simulations in [ sec : results ] . we conclude in [ sec : detonation ] by discussing the implications of the rt burning process on the explosion mechanism for type ia sne . the rt instability has been well studied analytically @xcite and numerically , both as an incompressible ( see recent results by @xcite ) and compressible ( see for example @xcite ) fluid . in the absence of burning , the growth of the multimode rt instability in the non - linear regime is characterized by two parameters , one describing the terminal velocity , and the other describing bubble merger @xcite . bubbles are assumed to merge when the difference in their heights exceeds some critical multiple of the smaller bubbles size . this leads to an expression for the average position of the interface , @xmath2 @xcite where @xmath3 is a function of the two parameters , and @xmath4 is the atwood number . much theoretical , computational , and experimental work aims to measure @xmath3 , and some evidence suggests that it may be constant for a wide range of initial conditions @xcite . the velocity of the merged bubbles can be written as @xmath5 where @xmath6 @xcite . these expressions were all formulated for the non - reactive rt instability and studies of rising bubbles . numerical experiments of reactive rt instabilities in sne conditions @xcite using a thickened - flame model found agreement with the above velocity expression , with @xmath7 . expressions of this form have been used as a subgrid model for the flame speed in full - star , pure deflagration calculations of type ia explosions in 1-d @xcite and 3-d @xcite , resulting in a successful explosion of the star . these results suggest that the rt instability is the dominant acceleration mechanism for the flame . @xcite and @xcite argued that , unlike a non - reactive rt instability , where all wavelengths can grow ( in the absence of surface tension ) , burning introduces a small scale cutoff to the instability . the growth rate for the pure rt instability is determined by the dispersion relation @xmath8 @xcite , where @xmath9 is the angular frequency and @xmath10 is the horizontal wave number for the dominant mode . we note that we have ignored contributions due to surface tension . by equating the period for the growth of the instability to the laminar flame propagation timescale , @xcite showed that there is a critical wavelength below which any perturbations will be washed out by burning . this wavelength is called the fire - polishing wavelength , @xmath11 : @xmath12 where @xmath13 is the effective gravitational acceleration . we note that this is a simple estimate , nevertheless , it provides a useful measure of the effectiveness of the rt instability on disrupting the flame . table [ table : flameproperties ] lists the flame properties , as computed by the code described in [ sec : numerics ] , including @xmath11 for the densities we are considering . stratification effects are negligible on the scales that we are considering , since we are modeling a tiny fraction of the star . the relative change in pressure over our domain is @xmath14 where we have expressed this ratio in terms of the mach number of the sharp - wheeler velocity for a rt flame encompassing our entire domain . since we are resolving the structure of the flame , the domain size restrictions limit the speed - ups we see to about 10 times the laminar speed , which gives a maximum @xmath15 of @xmath16 . as the density of the white dwarf decreases , the flame speed decreases sharply while the flame width , @xmath17 , increases . as table [ table : flameproperties ] shows , at the densities we are considering , we pass from the regime where @xmath18 ( low densities ) to @xmath19 ( high densities ) . when @xmath18 , the rt instability will dominate the flame , to the point where it may no longer be continuous , and one can no longer draw a simple curve from one side of the domain to the other marking the flame position . this increase in complexity of the flame surface signifies that we are leaving the flamelet regime . @xcite made estimates of the density where the burning transitions from the flamelet regime to the distributed burning regime , in which a distinguished local flame surface is not apparent . using the flame properties tabulated in @xcite and equating the gibson scale to the flame width , they found that for a 0.5 @xmath20c/0.5 @xmath21o flame , the burning will enter the distributed burning regime at a density of @xmath22@xmath23 . @xcite refined this estimate by suggesting that since the turbulence present in a type ia explosion is generated by the rt instability , it should follow bolgiano - obukhov rather than kolmogorov statistics . they found a slightly lower transition density of @xmath24 . at these densities , the turbulence is intense enough to disrupt the flame . we note that this transition density is right in the range where the fire polishing length becomes equal to the flame width , as discussed above . this density range is also interesting in the context of deflagration to detonation transitions ( ddt ) . arguments based on nucleosynthesis and one - dimensional modeling suggest that if the burning in a type ia supernova were to transition from a deflagration to a detonation , this transition would need to occur at a density around @xmath0 ( see for instance @xcite ) . the correlation between this ddt density and the density marking the transition to the distributed burning regime was pointed out by @xcite and @xcite . one - dimensional turbulence studies of the flame - turbulence interaction at these densities were performed by @xcite , where it was observed that the turbulence disrupts the flame sufficiently in some cases to create local explosions as the hot ash comes in contact with pockets of cool fuel . if these regions of fuel are large enough , these local explosions may be able to initiate a ddt . direct numerical simulations of thermonuclear flames at these densities do not exist . @xcite performed simplified flame model calculations in 3-d using a source term relevant to flames in type ia supernovae . they concluded that for densities larger than @xmath24 , scaling the flame speed in direct proportion to the increase in surface area is a valid subgrid model . for densities below @xmath0 however , they state that the effects of turbulence become important . as table [ table : flameproperties ] shows , the laminar flame speed in our density range is extremely subsonic our highest laminar flame mach number is @xmath25 . simulating a flame at these densities with a fully compressible code would be prohibitively expensive , requiring millions of timesteps . this in turn leads to a large accumulation of error . instead , a low mach number hydrodynamics code @xcite is employed for the present study . we discuss the numerical method in [ sec : numerics ] . together , the flame width and the fire - polishing length set strict limits on the densities and size of domains that can be addressed through fully resolved simulations . steady state laminar flame simulations show that we need about 510 zones inside the flames thermal width , @xmath26 we note that this width measurement is a factor of 23 narrower than the alternate measure defined as the width of the region where the temperature is 10% above the fuel temperature to 90% of the ash temperature that was used in @xcite . in order for the rt instability to develop , we need the computational domain to be at least one fire - polishing length wide , but in practice , we want 10 or more fire - polishing lengths in the domain for bubble merger to take place and the rt instability to become well developed . these restrictions limit us to consider densities @xmath27 in the present study . the lower limit of densities we can consider is set only by what is attained in the supernova explosion itself we stop at @xmath28 here . when the rt instability dominates the burning , it is the fluid motion mixing the fuel with the hot ash that controls the burning of the new fuel , rather than thermal conduction . thus , the flame has a very large thermal width , governed by the mixing , at the lowest densities . to initiate a detonation , the temperature in a sufficiently large region needs to rise in unison , such that the over - pressure generated by the burning is strong enough to drive a shock just ahead of the burning layer . the critical size of this region is called the matchhead size . if this rt thermal width becomes as large as the critical detonation matchhead size , as tabulated in @xcite , it is possible that a ddt can occur . furthermore , the size of the matchhead increases dramatically as the density decreases , and at low densities , it may be larger than the entire star . we note that recent simulations @xcite indicate that a pure deflagration can be successful in unbinding the star . we will explore the possible of a ddt , using the scaling behavior we find for flames in the distributed burning regime . the simulations were performed using an adaptive , low mach number hydrodynamics code , as described in @xcite . the state variables in the inviscid navier - stokes equations are expanded in powers of mach number , following @xcite . the result is that the pressure is decomposed into a dynamic and thermodynamic component , the ratio of which is of order mach number squared . only the dynamic component appears in the momentum equation . these equations are solved using an approximate projection formalism @xcite , breaking the time evolution into an advection and a projection step . the advection is computed using an unsplit , second - order godunov method that updates the species and enthalpy to the new time level and finds provisional velocities . a divergence constraint on the velocities is provided by the thermodynamics , and enforced in the projection step , where the provisional velocities are projected onto the space of vectors satisfying this constraint . burning is handled through operator splitting . the low mach number formulation allows us to follow the evolution of these slow moving flames in this density range , the laminar speeds are @xmath29without the constraint of the sound speed on the timestep . this affords us timesteps that are @xmath30 larger than a fully compressible code would take . we are still free however to have large density jumps as we cross the flame . this method was validated for astrophysical applications by comparing solutions of 1-d laminar flames to fully compressible results @xcite , and in our study of the landau - darrieus instability ( paper i ) , where the growth rate computed from our calculations matched the theoretical predictions across a range of wave numbers . the code is unchanged from the description in @xcite . we review the input physics used in the present simulations below . this method contrasts sharply with that used in the only previous study of the rt instability in type ia sne , as presented in @xcite . there , the fully compressible ppm algorithm was used , along with a model for a thickened flame to represent it on the small scales . the laminar flame speed matched the correct physical value , but the flame thickness was much larger than the physical value . in the present case , resolving the thermal structure frees us from the need for a flame model . the equation of state consists of a helmholtz free - energy tabular component for the degenerate / relativistic electrons / positrons , an ideal gas component for the ions , and a blackbody component for radiation , as described in @xcite . the conductivities contain contributions from electron - electron and electron - ion processes , and are described in @xcite . all of the flames are half carbon / half oxygen , but only the carbon is burned . a single reaction , @xmath20c(@xmath20c,@xmath31)@xmath32 mg , is followed , using the unscreened rate from @xcite . since we do not expect the burning to proceed up to the iron - group elements at these low densities , this reaction alone is sufficient to model the nuclear reactions . the reaction rate for oxygen burning is several orders of magnitude slower than the carbon rate at the ash temperatures we reach , so neglecting oxygen burning is valid on our scales . the calculations are initiated by mapping a one - dimensional steady state laminar flame ( computed with the same code ) onto our grid , shifting the zero point according to a random phase , 10 frequency sinusoidal perturbation . beginning with a steady - state flame in pressure equilibrium virtually eliminates any transients from disturbing the flow ahead of the flame . in all of the results presented , the flame is moving downward in our domain . the transverse boundaries are periodic , the upper boundary is outflow , and the lower boundary is inflow , with the inflow velocity set to the laminar flame speed . this keeps an unperturbed flame stationary on our grid . the solutions presented here were computed with an advective cfl number of @xmath33 and typically required 8000 timesteps . for all the runs we present , adaptive mesh refinement was used , with a base grid plus one finer level that has twice the resolution . this allows us to focus the resolution on the interesting parts of the flow , allowing larger domains to be studied . refinement triggered on the temperature gradient and the gradient of carbon mass fraction , ensuring that the flame s reaction zone was always at the finest resolution . table [ table : simparams ] lists the parameters used for all the simulations we present . for the lowest density case , @xmath34 , we actually present results for three different domain sizes and two resolutions and for the @xmath0 case , we present an additional simulation with the burning disabled . except where noted otherwise , all discussion of the @xmath28 results refer to the widest domain run . in the discussion below , we first concentrate on the change in character of the burning as the density is changed . next we look at integral quantities that help characterize the flame , followed by a measurement of the rt growth rate and a comparison of the rt instability with and without burning . finally , we consider resolution and domain size effects on the results we present . figures [ fig : rt_6.67e6_wide ] to [ fig : rt_1.5e7 ] show the results for the three densities we consider in this paper , @xmath34 , @xmath0 , and @xmath35 . the fuel is at the bottom of the domain ( and colored red ) , and the ash is at the top of the domain . gravity points up , toward increasing @xmath36 , in all simulations . for all runs , we set @xmath37 . this value is appropriate for the outer regions of the star , and assumes that some pre - expansion of the white dwarf has taken place . the domain width was chosen based on the flame width and fire - polishing length ( and considerations of how many zones we can reasonably compute ) to contain several unstable wavelengths . we see from the figures that as the density increases , the small scale structure is greatly diminished , as the burning is more effective in containing it . the calculations were run until the instability reached the top or bottom of the computational domain , or the size of the mixed region grew to be much larger than the width of the domain , leading to saturation of the instability which will be discussed later . in the descriptions below , ` bubble ' refers to the hot ash buoyantly rising into the cool fuel , and ` spike ' refers to the cool fuel falling into the hot ash . the results for @xmath38 , 768 cm wide simulation ( figure [ fig : rt_6.67e6_wide ] ) shows the flow quickly becoming mixed and the initial perturbations rapidly forgotten . a mixed region 5001000 cm wide forms , separating the fuel and ash . some smaller features begin to burn away as the flame evolves and the ash gets entrained by the fluid motions . at late times , it becomes impossible to draw a well defined flame surface separating the fuel and ash , as the rt mixing clearly dominates . in [ sec : res ] , we look at the effect of different domain widths on this growth . the @xmath0 run ( figure [ fig : rt_1.e7 ] ) shows a balance between the rt growth and the burning . spikes of fuel push well into the hot ash , forming well defined mushroom caps . before these caps can travel too far into the ash , they burn away . this is seen for several spikes in figure [ fig : rt_1.e7]the mushrooms gradually burn away , almost uniformly , turning yellow / green , and finally leaving a ephemeral outline in the ash as the burning completes . as the instability evolves , we see longer wavelength modes beginning to dominate , but they also begin to burn away . we are prevented from watching the final dominant mode that grows burn away because of it reaching the top of our domain . the highest density case , @xmath35 ( figure [ fig : rt_1.5e7 ] ) behaves qualitatively different than the two lower densities . a defining characteristic of this density is that the flame front is always sharply defined a trait that we expect to carry over to higher densities as well . at this density , the burning proceeds rapidly enough that it can suppress the rt instability significantly . this is as expected , since the fire - polishing length is larger than the flame width at this density ( see table [ table : flameproperties ] ) . as the spikes of fuel extend into the hot ash , we see them rapidly burning away from the outside inward . the rt mushroom caps hardly have time to form before their ends are burned away , giving them a hammerhead - like appearance . figure [ fig : carbon_destruction ] shows the carbon destruction rate , @xmath39 , at the midpoint of each calculation for the three densities . at the low density end , the burning is concentrated in small regions whereas at the high end , the burning is smoother . we would expect this trend to continue for densities outside the range we consider here . at higher densities , the burning will dominate even more , and we would expect the flame surface to be well defined on the small scales . at densities lower than @xmath28 , the mixed region will grow even larger , as the burning becomes even less effective at suppressing the rt instability . we will look at how the width of the reactive region grows in proportion to that of the mixed region at these low densities in [ sec : growth ] . having described the overall evolution of the flames , we now present some quantitative measures of flame behavior . we define the effective flame speed , @xmath40 , in terms of the carbon consumption in the domain ; namely , @xmath41 where @xmath42 is the spatial domain of the burning region , @xmath43 is the width of inflow face , @xmath44 is the inflow carbon mass fraction , and @xmath45 is the rate of consumption of carbon due to nuclear burning . accurate direct evaluation of this integral is problematic because of the operator split treatment of reactions in the numerical method . a robust and accurate estimate of @xmath40 can be obtain by integrating the conservation equation for carbon mass fraction @xmath46 over @xmath42 and a time interval @xmath47 $ ] to obtain @xmath48 @xmath49 is the inflow velocity . note that this formula is only valid when essentially all carbon is being consumed in the domain . another quantity of interest is the area of the flame , which for our two - dimensional studies is a length . computing a flame length for the lower density cases is problematic since there is often not a distinct flame surface . in this paper we define a flame length as the number of zone edges where the carbon mass fraction passes through 0.25 . we normalize the length to the initial flame length so we can measure the growth of the burning surface . this approach , while crude , is robust and defines a reasonable method for computing the length of a well mixed flame where other approaches are infeasible . in particular , this approach systematically overestimates the flame length ; however , this can be compensated for by normalizing by the initial flame length . we have validated that for well - defined flames , this method agrees well with the length of a contour as computed with the commercial idl package , after normalization . figures [ fig : rt_6.67e6_speeds ] to [ fig : rt_1.5e7_length ] show the speeds and lengths of the flames at the three densities as a function of time ( the additional curves on the @xmath28 plots will be discussed in [ sec : res ] ) . the maximum speedups we observe are 3 times the laminar speed for the @xmath28 run , 5 times for the @xmath0 run , and 2.5 times for the @xmath50 run . these speedups are much larger than any seen in our companion study on the landau - darrieus instability ( paper i ) , as expected , based on the outcome of the large scale simulations of the explosion . the flame length increases by a factor of 50 throughout the simulation for the two lowest densities , but only increases to by a factor of 8 for the @xmath35 case , due to the strong moderation of the rt instability by the burning . comparing the velocity curves to the length curves , we see that , in general , the velocity is not strictly proportional to the flame surface area . this is not unexpected , since the stretch and curvature that the flame experiences as well as the complicated motions of the rt instability pushing the fuel and ash together will modify the local burning rate . evidence of this variability is seen in figure [ fig : carbon_destruction ] where considerable variation in the carbon destruction rate is observed , even for the highest density case . to further quantify this effect , in figures [ fig : rt_6.67e6_v_vs_a ] to [ fig : rt_1.5e7_v_vs_a ] we plot the velocity divided by flame surface area normalized to @xmath51 , where @xmath43 is the width of the box . with the caveat that there are ambiguities in defining flame surface area for the `` distributed '' flames , we observe that , after an initial transient , the flame speed versus area relaxes to an essentially statistically stationary value . for the highest density case , the effective flame velocity is approximately 30% of the laminar flame speed @xmath52 the area enhancement . for the lower density cases , the value drops to approximately 10% . thus , at these densities , we find @xmath53 where @xmath54 is a proportionality constant not equal to 1 . this proportionality constant increases with density , reflecting the fact that the fire - polishing wavelength , which sets the smallest scale on which we can wrinkle the flame , grows as well . thus at higher densities , the localized curvature is smaller , and as a result , the localized velocity is less affected . this has strong implications for subgrid models , particularly in the flamelet regime , where it is normally assumed that the velocity scales directly as the increase in area of the flame surface . it appears that this is not the case , and is further complicated by the density dependence of the proportionality constant . @xcite introduced a fractal model to describe the growth of the flame surface in a supernova when the burning is in the flamelet regime . the self - similar range through which a fractal description applies is bounded by the fire - polishing length on the small scales and the sharp - wheeler description for the rt instability growth on the large scales . in two - dimensions , a fractal model for the growth of the flame surface would be @xmath55 where we expect @xmath56 in the self - similar scaling regime . the offset in time , @xmath57 , is chosen such that @xmath58 at @xmath59 . the value of @xmath3 is taken to be @xmath60 . as we will see in [ sec : growth ] , the sharp - wheeler model may not be the proper description of the reactive rt instability . with that caveat , figures [ fig : rt_1.e7_length ] and [ fig : rt_1.5e7_length ] , we plot the predictions of the fractal scaling model using @xmath61 and @xmath62 , which give area growing as @xmath63 and @xmath64 respectively . from these figures , it seems that , if this fractal model is correct , the growth in the flame surface is closely represented by a fractal dimension of 1.7 . this agrees with the value computed from two - dimensional rt studies by @xcite . some evidence suggests that the fractal dimension of rt induced turbulence may increase with time @xcite . we caution however that we only span about a decade between @xmath65 and @xmath66 in these simulations . if such a scaling holds , it could serve as a subgrid model for large scale simulations , however , the evolution of the proportionality constant in equation ( [ eq : v_vs_a ] ) needs to be understood . additional calculations that encompass a larger range of spatial scales are needed to further validate this model . not all of the energy release by the flame goes into driving the expansion of the star . the rt unstable flame and associated kelvin - helmholtz instabilities also generate turbulence . turbulence diagnostics can be difficult to define here we use an integral quantity , the favre average turbulent kinetic energy which plays a role in the @xmath67 subgrid model for turbulence . the favre average turbulent kinetic energy ( for a discussion of which see @xcite ) , can be written @xmath68 where @xmath69 denotes the horizontal spatial average of the quantity enclosed in the brackets . figure [ fig : rt_6.67e6_tke_y ] shows @xmath70 as a function of height at several instances in time for the @xmath28 flame . we see that , within the mixed region , the turbulent kinetic energy is roughly constant , and this plateau rises quadratically with time , as shown in figure [ fig : rt_6.67e6_tke_peak ] . the integral of @xmath71 over the vertical extent of the domain is the total turbulent kinetic energy @xmath72 . the generation of turbulent kinetic energy as a function of time is shown for the three densities in figures [ fig : rt_6.67e6_tke ] to [ fig : rt_1.5e7_tke ] . the data suggest that the increase in kinetic energy as a function of time is well approximated by a power law of the form @xmath73 are shown in figures [ fig : rt_6.67e6_tke ] to [ fig : rt_1.5e7_tke ] , and summarized in table [ table : tke ] . the power law fits , summarized in table [ table : tke ] match the data quite well , giving an exponent , @xmath74@xmath75 , depending on the density , with the power increasing with increasing density . in the lowest density run , there is considerable turbulent energy , and , as figure [ fig : rt_6.67e6_tke_y ] shows , @xmath76 is reached after @xmath77 s. this is on a scale of @xmath78 cm . @xcite presumed pre - existing turbulent kinetic energy of this level on scales of @xmath79 cm in their main calculation . assuming a one - third power scaling of the energy cascade , on their length scales the rt generate turbulence already overwhelms this pre - existing turbulence , and it is continuing to grow , apparently quadratically with time . these results suggest that the rt instability alone can provide the turbulent motions presumed to exist in the star . in the absence of any reactions , the extent of the mixed region should increase as @xmath80 , as predicted by the sharp - wheeler model , equation ( [ eq : sw ] ) . this relation is only expected to hold during the phase where bubbles are merging , which excludes the initial linear growth range . also , because we are resolving the flame structure , the interface between the fuel and ash is not infinitesimally thin , which complicates the definition of the atwood number . furthermore , at late times , once the bubble merging in our domain has stopped and the size of the mixed region rivals the width of the domain , this relation will also break down . this relation was derived for the purely hydrodynamical rt instability . burning can have large consequences here . the sharp - wheeler model describes the growth of this mixed region , which will contain fuel and ash , but that fuel is burning , and as it turns into ash , we would expect to find the mixed region smaller than that of the purely hydrodynamical case . for the present simulations , since we have a limited number of modes in our box ( typically 10 ) , we are only going to be able to do the fits on a small subset of the data . a future study will focus on scaling in the flamelet regime , using much larger domains . figures [ fig : rt_6.67e6_width ] to [ fig : rt_1.5e7_width ] show the extrema of the mixed region , computed by laterally averaging the carbon mass fractions and finding the positions where it first exceeds 0.05 and 0.45 . this definition is consistent with that used in @xcite . the curves are measured with respect to the initial position of the interface . the top curve measures the position the spikes of fuel pushing into the region of hot ash . the lower curve is the position of the bubbles of hot ash floating into the fuel . the sharp kinks in the spike curves represent the instances where the plume of fuel , having pushed far into the ash region , burned away . at the highest density ( figure [ fig : rt_1.5e7_width ] ) , the bubble curve is very smooth . we can make fits of the growth of the mixed region to time by fitting to a power law , @xmath81 we do not attempt to fit to the sharp - wheeler scaling and find a value of @xmath3 because of the limited range of wavelengths we follow , and the uncertainty as to whether that relation holds in the reactive case . we exclude the initial linear phase of the growth of the rt instability in our fits . figures [ fig : rt_6.67e6_width_fit ] to [ fig : rt_1.5e7_width_fit ] shows the results . in all cases , we find @xmath82 less than 2.0 , indicating that we are not in the sharp - wheeler regime . furthermore , @xmath82 increases with increasing density . we find @xmath83 for @xmath28 , @xmath84 for @xmath0 , and @xmath85 for @xmath35 . larger scale studies , with a greater range of wavelengths are needed to determine whether this is a general result for reactive rayleigh - taylor , or if this is because of the small size domains considered here . one way we can get some insight into this is to rerun one of these simulations without burning . ( note that in this case , we shift to a non - moving frame of reference whereas the reacting cases are performed in a reference frame moving at the laminar flame speed . ) the growth rate for the @xmath0 run with burning disabled is over - plotted in figure [ fig : rt_1.e7_width ] . we see that it is much smoother than the corresponding reactive case , since the spikes of fuel which push into the hot ash never burn away . the only wiggles in the mixing region growth curves result from bubble mergers . the fit to a power law for this non - reactive simulation is presented in figure [ fig : rt_1.e7_noburn_width_fit ] . the carbon mass fraction for several points in time are shown in figure [ fig : rt_1.e7_noburn ] , which compares directly to figure [ fig : rt_1.e7 ] . here , @xmath86 , higher than the reactive case , but still not equal to 2.0 . the difference between this value and the reactive case shows that the burning does have some influence . in the non - reactive case we really do expect @xmath87 , and therefore the difference we see is very likely because we also are probably not following enough modes to see the sharp - wheeler scaling . for the @xmath35 run , we were closest to growing as @xmath80 . the burning here is quite vigorous , compared to the rt growth , so as argued above , we would expect to see its influence on the extent of the mixed region . however , we would expect it to affect the position of the spikes of fuel only , since the bubbles of ash do not react . therefore , we can fit the growth of the bubbles and spikes from the initial interface separately . this is shown in figure [ fig : rt_1.5e7_bubble_spike_fit ] . again , we exclude the initial linear growth phase of the instability . we find the bubble position growing as @xmath88 and the spike region growing as @xmath89 . based on the trends we see here and above , we would expect that once we have enough unstable modes in the box , the width of the mixed region during the bubble merger phase will scale as slightly smaller @xmath80 , with the bubble position itself growing as @xmath80 . therefore , more studies , in larger domains with more unstable modes are needed . it is interesting to look at how the size of the reactive region scales with that of the mixed region , especially at the lowest density . as discussed in [ sec : rt ] , if the reactive region grows to the size of the detonation matchhead , then it may be possible for a deflagration - detonation transition to occur . measuring the scaling of the reactive region to the mixed region is not easy , because the flame is so wrinkled . we will use a volume weighted definition here . the collection of zones , @xmath90 , where the carbon mass fraction , @xmath91 , is between @xmath92 and @xmath93 is @xmath94 similarly , we can define the collection of zones , @xmath95 , where the density - weighted carbon destruction rate , is within @xmath31 of the peak , @xmath96 then , the ratio of the size of the reactive region to the mixed region is @xmath97 where @xmath98 is the volume of the zones in set @xmath90 . once the mixed region develops , we expect this quantity to be always less than unity . because of the freedom to choose the parameters @xmath92 , @xmath93 , and @xmath31 , the numerical value of @xmath99 itself does not have much meaning , but the trend with time does . we pick @xmath100 , @xmath101 , and two values of @xmath31 : 0.1 and 0.8 . we note that the trend is relatively insensitive to the choice of these parameters , as illustrated by figure [ fig:6.67e6_scale ] , which shows @xmath99 as a function of time for the @xmath28 rt unstable flame for the two choices of @xmath31 . after an initial transient , it appears that these curves level off , indicating that the reactive region continues to growth in size proportional to the mixed region as the flame evolves . we note that this shows that the reactive region is a small fraction of the mixed region , but this absolute scaling is dependent on the choice of the three parameters . this is illustrated in figure [ fig:6.67e6_scale_thresh ] , where the mixed and reactive regions are shown , midway through the calculation . at late times , the peak nuclear energy generation comes from regions where the carbon mass fraction is @xmath102 , as shown in figure [ fig : rt_6.67e6_yc12_dydt ] . this is also apparent by looking at the laterally averaged flame profile and comparing to the laminar state see figure [ fig : rt_6.67e6_wide_averages ] . if this reactive region can growth to a critical matchhead size , before the entire star is consumed , it may be possible for a deflagration to detonation transition to proceed . however , since at late times the peak burning is occurring at mass fractions of 0.15 rather than 0.5 , this will greatly increase the required matchhead size , further complicating the possibility of a deflagration - detonation transition . in this subsection , we examine the role of domain size and resolution on the computational results . figures [ fig : rt_6.67e6 ] and [ fig : rt_6.67e6_small2 ] show the evolution of the @xmath103 rt simulation in narrower domains ( 96 cm and 384 cm wide respectively ) . other than the domain size , the parameters are identical to those used in the results described shown in figure [ fig : rt_6.67e6_wide ] . in the narrowest case , once the size of the mixed region reaches a fixed size , comparable to the width of the domain , the mixing and reacting processes saturate and the flow enters a quasi - steady - state , with vortices that are remnants of the early mixing driving transverse shear flow in the mixed region . once this saturation is reached , the burning rate begins to slowly decay . this behavior is not seen in the corresponding wider domain calculation , although , we would expect a similar pattern if we use a taller domain and ran out to longer times . figure [ fig : rt_6.67e6_speeds ] shows three additional curves corresponding to these narrower domains , the narrowest at two different resolutions ( see table [ table : simparams ] ) . in the 96 cm wide domain , the peak velocity is much smaller , reaching only about 3 times the laminar speed . this is to be expected from equation ( [ eq : sw2])the wider domain allows longer wavelength modes to go unstable . the decrease in velocity at late times in the narrow domain begins when the size of the mixed layer becomes comparable to the width of the domain , preventing any further modes from growing . this is reflected in the later panels of figure [ fig : rt_6.67e6 ] . in the 384 cm wide domain , the velocity continues to grow , beyond 6 times the laminar speed we would expect our widest domain run to continue to accelerate as well , but it began to interact with the top of our computational domain and was not run out as long . the velocity of the low resolution run tracks that of the high resolution case very closely , suggesting that , for the large scale diagnostics , we have converged . this lowest resolution run has approximately 5 points in the thermal width ( equation [ eq : thermalwidth ] ) , a value found acceptable in the convergence study performed in @xcite for landau - darrieus unstable astrophysical flames . we note again that this definition of thermal width is smaller than some other commonly used definitions ( see , for example @xcite ) . finally , in all calculations , the initial perturbations are well resolved , with typically 50100 zones per wavelength ( with the only exception being the 96 cm wide , @xmath28 run , where there are only 20 zones per wavelength ) . this zoning is at or exceeds the resolution determined to give acceptable convergence of single - mode rt growth rates in the 3-d compressible study presented in @xcite . we also can look at how the scaling of the turbulent kinetic energy depends on the domain size . for the 384 cm wide , @xmath34 run , this is shown in figure [ fig : rt_6.67e6_tke_small2 ] . two fits are shown . when all of the data is included , the fit is quite poor , since it is biased by by data at the end of the run when the mixing has saturated and the turbulent kinetic energy stops increasing . excluding this data , the fit to the first @xmath104 s is quite good , and with the kinetic energy @xmath105 quite close to the 2.267 power obtained from the 768 cm wide run . thus , it appears that the scaling of the turbulent kinetic energy prior to saturation is insensitive to the domain size . based upon the observed systematics of flame propagation in the distributed regime , we can begin to speculate on the potential for a transition to detonation . as discussed by @xcite and @xcite , there are two possible modes for a delayed transition to detonation : a ) a `` local '' transition to detonation because a fluid element of critical mass burns with a supersonic phase velocity and b ) a `` macroscopic '' transition because some appreciable fraction of the white dwarf volume develops such complex topology that burning briefly consumes more fuel than could a spherical front encompassing that region and moving at supersonic speed . the former , also known as the zeldovich mechanism , is the basis for the `` delayed detonation model '' by @xcite ; the latter is the basis for a similar model by @xcite . @xcite has argued that the volume detonation model requires special preconditioning , but because we study only the small scales on which the flame is resolved , we can not comment in a meaningful way on this issue . also , because we have not included turbulence cascading down from scales larger than our grid , we can not conclusively argue about the local transition to detonation either ( a large amount of turbulence could , for example , induce a transition to distributed burning at a higher density ) . still some scaling relations are observed that , if they can be generalized by larger scale studies , argue against a transition to detonation . as is well recognized ( e.g. * ? ? ? * ) , transition to detonation can never occur so long as a well defined flame exists . the width of a steady flame is always thinner than the critical mass required for detonation . however , a mixture of hot fuel and cold ash is potentially explosive as calculations in the sn ia context @xcite have suggested . figure [ fig : carbon_destruction ] shows the existence , in the distributed regime , of localized hot spots of unsteady burning . the high temperature sensitivity of the @xmath20c + @xmath20c reaction itself makes it difficult for one of these hot spots to become supercritical in mass . for the conditions we consider , carbon fusion dominates the energetics , and the energy production is @xmath106 where @xmath91 is the mass fraction of carbon , and @xmath82 given by the coulomb barrier between two carbons nuclei , @xmath107 @xmath108 with @xmath109 the temperature in 10@xmath110 k. for our calculations with @xmath111 = @xmath112 , the temperature of the hot ash is @xmath113 . anticipating that the temperature of a combustible fuel - ash mixture will not be far from that , we find @xmath114 . the carbon mass fraction in the fuel is 0.5 and in the ash it is zero , and the temperature of the unburned fuel is negligible . in any mixed mass , @xmath115 , composed of fractions @xmath116 of ash and @xmath117 of fuel , the temperature will be @xmath118 and the carbon mass fraction @xmath119 the dependence of the heat capacity on temperature is important and mitigates somewhat the extreme sensitivity of the reaction rate . in the vicinity of @xmath109 = 2.5 , the heat capacity is predominantly due to the electron gas ( with @xmath120 ) , but with a non - negligible contribution from radiation ( @xmath121 ) @xcite . to good approximation , near @xmath109 = 2.5 , @xmath122 . the maximum in energy generation will then occur for @xmath123 \ = \ 0 \enskip , \ ] ] or for @xmath124that is , in the burning mixture , @xmath125 and @xmath126 . a factor of two variation in the energy generation occurs for @xmath127 , or @xmath91 from @xmath128 to @xmath129 . this agrees well with what is observed in the numerical simulation ( fig . [ fig : rt_6.67e6_yc12_dydt ] ) . coupled with the fact that burning to magnesium or silicon releases less energy than burning to the iron group , this low carbon mass fraction implies that the overpressure from burning will be small in the regions where ash and fuel are mixed . burning from @xmath130 to @xmath131 , the typical ash temperature , will only raise the pressure by 8% . since this overpressure determines the temperature reached in a detonation and the burning rate depends on this to a high power , the critical mass for initiating a detonation will be large . the size will certainly be greater than the distance a sound wave can travel during the time it takes the mixture to burn . at @xmath109 = 2.2 and @xmath132 , the nuclear time scale is @xmath133 or about 0.07 s. the speed of sound is approximately 5000 km s@xmath134 . hence a critical mass for these conditions must be at least 300 km in size . on the other hand , figure [ fig:6.67e6_scale ] shows that , in steady state , the size of the reactive region is approximately a small constant times the size of the mixed region . even allowing a range of burning time scales of a factor of 10 across the burning region , the explosive mixture only has dimensions about 10% that of the region that has been mixed by the rt instability . taking @xmath135 as an upper bound to that ( see [ sec : growth ] ) , and realizing that the temperature will decline still further in a few tenths of a second because of expansion , the largest mixed , potentially explosive region is less than 5 km . thus , though a transition to detonation is in principal possible , there simply may not be room , nor time enough to develop one in an exploding white dwarf . we presented direct numerical simulations of reactive rt instabilities in conditions appropriate to the late stages of a type ia supernova explosion . fully resolving the flame frees us from the need to specify any model parameters , such as the flame speed , markstein length , etc . in the density range we consider , the flame transitions from having a distinct , well defined interface at the high density end to being a chaotic , well mixed burning region at the lower density end , with a transition density of @xmath136 . this transition is expected based on the arguments presented in [ sec : rt ] . furthermore , this would suggest that at the even higher densities characteristic of the early stage of the explosion , the flame surface continues to be well defined , consistent with the conclusions of @xcite and the 2-d flame - model simulations at @xmath1 presented in @xcite . the effects of the burning were further made concrete by presenting the @xmath0 case both with and without reactions . at the lowest density , the mixed region becomes very large , although , still smaller than the critical carbon detonation matchhead size @xcite . it appears that the reactive region grows in direct proportion to the size of the mixed region , making it possible in theory that such a transition can occur , however , at these low densities , much of the star is already consumed , and , based on scaling predictions from the present simulations , the time required for the reactive region to grow to the matchhead size seems to be longer than the total explosion time . in effect , the flame runs out of star before a critical mass can be built up . future simulations will explore this in more depth . in all cases presented here , the flames accelerated considerably , reaching speeds between 2 and 6 times the laminar flame speed . this is significantly larger than accelerations seen in small - scale / resolved landau - darrieus instability studies ( paper i ) . the growth of the flame surface appeared to be well described by a fractal model , with a fractal dimension of @xmath137 . in the cases considered , the effective flame speed was asymptotically proportional to the increase in flame surface area but the constant of proportionality was about 0.1 for the low density cases and 0.3 for the higher density case . thus , the flame acceleration was considerably less than would be predicted by the increase in flame surface area alone . this proportionality constant may approach unity as the density is increased , because the smallest scale on which the rt instability can bend the flame ( the fire - polishing length ) increases dramatically with density , and therefore the magnitude of the curvature that the flame experiences locally decreases . we would expect that on larger domains the flame would continue to accelerate , further supporting the already widely held view that the rt instability provides most of the acceleration to the flame in a pure deflagration type ia sne explosion . the growth of the mixed region for the reactive rt instability appears to be slower than the non - reactive case . the present results do not seem to support the sharp - wheeler model , but larger scale studies are needed , with more bubbles merging over a longer period of time to better understand the growth of the mixed region . understanding the growth of the mixed region is critical to providing an accurate subgrid model . future studies will focus on the three - dimensional counterparts to the flames we discussed here . turbulence behaves very differently in two and three dimensions , so we expect to see some differences in the evolution of the instability . numerical results indicate that the growth rate of the pure rt instability is faster in 3-d than in 2-d @xcite . furthermore , in 3-d the surface to volume ratio of the spikes of fuel is larger than in 2-d , exposing more fuel to the hot ash , so we may expect them to burn away more quickly . we still expect , based on the arguments presented in [ sec : rt ] and the results of the simulations , that the flame will undergoing the same transition in behavior as the density decreases . larger two - dimensional studies in the flamelet regime are also needed , moving toward higher densities . capturing a greater range of scales on the grid will allow for validation of the fractal scaling model and a better understanding of the moderation of the sharp - wheeler prediction for the growth of the mixed region . finally , these fully resolved simulations can serve as the basis for testing various models to represent the flame on subgrid scales , which become necessary to address the larger scale type ia physics . the authors thank f. x. timmes for making his equation of state and conductivity routines available online . support for this work was provided by the doe grant no . de - fc02 - 01er41176 to the supernova science center / ucsc and the applied mathematics program of the doe office of mathematics , information , and computational sciences under the u.s . department of energy under contract no . de - ac03 - 76sf00098 . sew acknowledges nasa theory award nag5 - 12036 . some calculations were performed on the ibm sp ( seaborg ) at the national energy research scientific computing center , which is supported by the office of science of the doe under contract no . de - ac03 - 76sf00098 , the ibm power4 ( cheetah ) at ornl , sponsored by the mathematical , information , and computational sciences division ; office of advanced scientific computing research ; u.s . doe , under contract no . de - ac05 - 00or22725 with ut - battelle , llc , and the ucsc upsand cluster supported by an nsf mri grant ast-0079757 . rdrddr & & @xmath138 & & & + & & ( @xmath139 ) & & & + @xmath140 & 0.529 & @xmath141 & 5.6 & 0.026 & @xmath142 + @xmath143 & 0.482 & @xmath144 & 1.9 & 0.23 & @xmath145 + @xmath146 & 0.436 & @xmath147 & 0.54 & 1.8 & @xmath148 + rddddddl & & & & & & & comments + & & & + @xmath140 & 96.0 & 3072.0 & 1.0 & 17.1 & 3290 & 5.6 & + & 96.0 & 1536.0 & 0.5 & 17.1 & 3290 & 11.2 & + & 384.0 & 2304.0 & 1.0 & 68.6 & 14800 & 5.6 & + & 768.0 & 2304.0 & 1.0 & 137 & 29500 & 5.6 & + @xmath143 & 163.84 & 327.7 & 0.16 & 86.2 & 712 & 11.9 & + & 163.84 & 327.7 & 0.16 & 86.2 & 712 & 11.9 & no burning + @xmath146 & 53.5 & 107.0 & 0.0522 & 99.1 & 29.7 & 10.3 & + rrrl & & & comments + + & @xmath149 & 1.626 & 384 cm wide ; all data + & @xmath150 & 2.325 & 384 cm wide ; initial 0.005 s + & @xmath151 & 2.267 & 768 cm wide + @xmath143 & @xmath152 & 2.591 & + @xmath146 & @xmath153 & 2.867 & +

modelling and forecasting electricity loads is a problem well - known within both the academic and the applied statistics community ( see e.g. * ? ? ? the signals studied usually exhibit strong properties such as seasonalities or weekly and daily profiles , leading to some very accurate models that tend to perform rather well under normal forecasting conditions . the approaches used to model and forecast them vary a lot : we mention a couple of them in the lines below . some authors worked with univariate time series models : @xcite built a double seasonal exponential smoothing for the british electricity load , while @xcite and @xcite presented some comparative studies between univariate methods for different sets of data . @xcite opted for a non - parametric approach relying on the wavelet transform to forecast the load curve seen as a functional - valued autoregressive hilbertian process . others tried and modelled the load together with the use of exogenous variables : @xcite included the temperature in their model , which inspired the bayesian semi - parametric regression model found in @xcite . alternatives to univariate modelling were often considered too , such as building multiple - equation models : while the various hours of the day share the same equation , the associated parameters differ from one another . @xcite built an hourly independent seasonal auto - regressive model for their data , and @xcite also built an independent model for each hour of the day but took temperature effects into account . a state - space model is proposed in @xcite and @xcite where the parameters of the model are also allowed to vary along the time . the exogenous variables most commonly used to forecast the electricity load are weather - based , even though the decision to include one or more meteorological variables into a model may be open to discussion . the period of forecasting has to be taken into account , as well as the accuracy of the predictions for such variables . for temperate climates , the most important meteorological factor is the temperature ( see e.g. * ? ? ? for the french electricity load specifically , the importance of the temperature and cloud - cover was underlined in @xcite and @xcite . other weather - related models include the works of @xcite and @xcite within the bayesian framework . let us also mention machine learning work : @xcite for a neural networks implementation and @xcite for a detailed study of online mixing algorithms used on a set of various predictors . we are interested in the development of a methodology to improve the estimation and the predictions of a parametric multi - equation model ( similar to the one presented in @xcite ) over a short dataset . the limited size of the dataset coupled with the high dimensionality of the model leads to a typical overfitting situation when using the maximum likelihood approach : the fitted values are relatively close to the observations while the errors in prediction are an order of magnitude larger or more ( note that due to the very periodic nature of its regressors , the model typically requires 4 or 5 years of data to provide satisfactory predictions ) . this overfitting behaviour can be somewhat alleviated by the use of a bayesian estimation relying on an informative prior distribution , but the very fact that the data available is limited makes the posterior distribution all the more sensitive to the choice of that prior . although electricity load curves may largely differ from one population to another , they may also share some common features . the latter case is expected to happen when the global population studied is an aggregation of non - homogeneous subpopulations for which the estimations are made harder due to the relative lack of data . to design a sensible prior in such a situation , we consider the case where another long dataset is available , upon which the model performs equally well in both estimation and prediction . we assume the long and the short datasets are somehow similar in a non - obvious way . that the similarity between the parameters underlying the two datasets ( we will assume they are indeed coming from the model considered ) can not be easily guessed prevents us from trying to model the datasets simultaneously because it would require a rather precise knowledge of the link between the two . we propose a general way of building an informative hierarchical ( see * ? ? ? * for a general review on the subject of hierarchical models ) prior for the short dataset from the long one that goes as follows : 1 . we first estimate the posterior distribution on the long dataset using a non - informative prior , arguing that the design of an informative prior for this dataset is not necessary , since the data available is enough to estimate and predict the model in this case ; 2 . we extract key pieces of information from this estimation ( e.g. moments ) to design an informative prior for the short dataset which takes into account the prior information that the datasets are somehow similar , via the introduction of hyperparameters designed to model and estimate this similarity . the paper is organised as follows . in section [ sec : methodo ] we focus on the general methodology and describe the way we carried our experimentation , we also present the general regression model used for our tests and applications . in section [ sec : specpriors ] , we present the semi - conjugated priors ( informative and non - informative ) used on each of the datasets . the ad hoc mcmc algorithms we developed to estimate the mean and variance of the posterior distributions are push backed into the appendix so as not to obfuscate the main point of the paper by technical details . in section [ sec : simulations ] , we use these algorithms to illustrate and validate our approach in simulated situations : we show the contribution of the informative prior over the precision of both the estimated parameters and the forecasts in the case of a working electricity load forecasting model . in section [ sec : applications ] , we apply our method to french electricity datasets and compare the results with the outputs of 3 alternative standard methods to assess its competitivity . we also study the effect of the small dataset s sample size upon the predictive quality of the model and show that the informative prior provides reasonable forecasts even when the lack of data is severe . let us define here some notations that we shall keep throughout this paper . hereafter , we denote @xmath0 a short dataset over which we would like to estimate the model and we denote @xmath1 a long dataset known or thought to share some common features with @xmath0 . we will denote @xmath2 the parameters of the model and @xmath3 the observations from @xmath1 . we propose a method designed to help improve parameter estimations and model predictions over @xmath0 with the help of @xmath1 . let @xmath4 be the prior distribution used on @xmath1 and @xmath5 the associated posterior distribution . observe that the choice of @xmath4 is not crucial as long as it remains non - informative enough because the model can be correctly estimated from the data alone on @xmath1 . let @xmath6 denote the prior distribution to be used on @xmath0 . notice that the naive pick @xmath7 is not a viable solution as soon as the parameters of @xmath1 and @xmath0 differ since the variance of the posterior distribution @xmath5 is too small : in that case the data of @xmath0 will not be able to make up for the difference between the posterior mean on @xmath1 and the true value of the parameters of @xmath0 . assuming that the parameters corresponding to @xmath1 and @xmath0 are identical is too restrictive in practise . to allow for more flexibility we add hyperparameters accounting for the similarity between the datasets . we now described the informative hierarchical prior we designed . assume that the prior distribution @xmath6 is to be chosen within the parametric family @xmath8 since selecting @xmath9 is equivalent to picking @xmath10 , and since we want @xmath11 to retain some key - features of @xmath5 , we want to pick @xmath12 using some of the information contained inside the posterior distribution obtained on @xmath1 . we assume that there exists an operator @xmath13 , such that @xmath14 = \lambda,\ ] ] and choose @xmath12 proportional to @xmath15 $ ] , in the sense that @xmath16,\ ] ] where @xmath17 itself is an unknown linear operator that we assume diagonal for ease of use . the operator @xmath18 can be interpreted as a similarity operator between @xmath1 and @xmath0 , and its diagonal components as similarity coefficients measuring how close the two datasets really are when looked at through @xmath19 . the diagonal components of @xmath18 are hyperparameters of the prior we designed , and we give them a vague hierarchical prior distribution centred around @xmath20 , the prior on @xmath20 being vague and centred around @xmath21 . the hyperparameter @xmath20 may also be regarded as a more global similarity coefficient , since it represents the mean of all the similarity coefficients . the prior mean of @xmath20 is forced to @xmath21 to reflect the prior knowledge that the datasets are somehow similar . the variance of the prior distribution of @xmath20 could in theory be reduced , going from a vague prior to a more informative structure , depending on the confidence we have over the similarity between the datasets . we chose not to however , so as to keep the procedure we describe from requiring any delicate subjective adjustments . we present now two frequent situations where the above procedure can be written in a simpler way . we assume that the elements of @xmath22 can be identified via their @xmath23 first moments : the operator @xmath19 can then be reduced to a function @xmath24 of the @xmath23 first moments operators , i.e. @xmath25 = f({\mathbb{e}}(\theta ) , \ldots , { \mathbb{e}}(\theta^m))$ ] . the expression of @xmath12 then becomes @xmath26 note that , if the prior requires the specification of at least the two first moments , even though the priors from the upper layers of the model are vague , the correlation matrix estimated on the dataset @xmath1 remains untouched and is directly plugged into in the informative prior if we consider centred moments for orders greater than @xmath21 . we consider the case where @xmath22 is the family of priors conjugated for the model . if the prior @xmath4 belongs to @xmath22 then the associated distribution @xmath5 does too and there corresponds a parameter @xmath27 to it . the expression of @xmath12 thus reduces to @xmath28 modelling and forecasting the electricity load ( or demand ) on a day - to - day basis has long been a key activity for any company involved in the electricity industry . it is first and foremost needed to supply a fixed voltage at all ends of an electricity grid : to be able to do so , the amount of electricity produced has to match the demand very closely at any given time and experts usually make use of short - term forecasts with this aim in view as mentioned in @xcite . electricity load usually has a large predictable component due to its very strong daily , weekly and yearly periodic behaviour . it has also been noted in many regions that the weather usually affects the load too , the most important meteorological factor typically being the temperature ( see * ? ? ? * for an example ) . for each of the 48 instants of the day ( each instant lasts 30 minutes , starting from 00:00am ) , the non - linear regression model that we consider in this paper , first described in @xcite , is made of three components , which we explain briefly in the next paragraphs , and is usually formulated as follows : for @xmath29 , @xmath30 + \sum_{j=1}^{d_{12 } } \omega_j { \mathds{1}}_{\omega_j}(t ) , \nonumber\\ x^{(2)}_t & = \sum_{j=1}^{d_2 } \psi_j { \mathds{1}}_{\psi_j}(t ) , \nonumber\\ x^{(3)}_t & = g ( t_t - u){\mathds{1}}_{{[t_t,\,+\infty[}}(u ) , \nonumber\end{aligned}\ ] ] where @xmath31 is the load of day @xmath32 and where @xmath33 are assumed independent and identically distributed with common distribution @xmath34 . the @xmath35 component is meant to account for the average seasonal behaviour of the electricity load , with a truncated fourier series ( whose coefficients are @xmath36 and @xmath37 ) and gaps ( parameters @xmath38 ) which represent the average levels of electricity load over predetermined periods given by a partition @xmath39 of the calendar . this partition usually specifies holidays , or the period of time when daylight saving time is in effect i.e. major breaks in the electricity consumption behaviour . the left part of figure [ fig : illuconso ] shows a typical behaviour over two different periods of time ( summer vs. winter ) . the @xmath40 component allows for day - to - day adjustments of the seasonal behaviour @xmath35 through shapes ( parameters @xmath41 ) that depends on the so - called days types which are given by a second partition @xmath42 of the calendar . this partition usually separates weekdays from weekends , and bank holidays . the differences between two different daytypes are visible on the left part of figure [ fig : illuconso ] too . for obvious identifiability reasons , the vector @xmath43 is restricted to the positive quadrant of the @xmath44-unit sphere in @xmath45 , that we denote @xmath46 the @xmath47 component represents the non - linear heating effect that links the electricity load to the temperature ( see * ? ? ? * for a general presentation of non - linear models ) , with the help of 2 parameters . the heating threshold @xmath48}$ ] corresponds to the temperature above which the heating effect is considered null and is usually estimated to be roughly around 15c . the heating effect is supposed to be linear for temperatures below the threshold and null for temperatures above . the restriction on the support of the threshold @xmath49 simply expresses the fact that the threshold is sought within the range of the observed temperatures , i.e. @xmath48}$ ] with @xmath50 the heating gradient @xmath51 where @xmath52 represents the intensity of the heating effect , i.e. the slope ( assumed to be non - zero ) of the linear part that can be observed on the right part of figure [ fig : illuconso ] . using the notation @xmath53 for the @xmath54-th row of a matrix @xmath55 , the previous model can be re - written in the following condensed and more generic way : for @xmath56 @xmath57 the matrices @xmath58 of size @xmath59 , @xmath60 of size @xmath61 , @xmath62 of size @xmath63 , and @xmath19 of size @xmath63 are known exogenous variables while the parameters of the model to be estimated are @xmath64 } \times { \mathbb{r}}_+^*,\ ] ] where @xmath65 is the positive quadrant of the @xmath44-unit ball of dimension @xmath66 . one could , without too much difficulty , add a cooling effect to the model , whose definition would be similar to that of the heating effect . since the cooling effect remains far less important than the heating effect in france at the present time ( see the right part of figure [ fig : illuconso ] ) , and since the estimation of the associated cooling threshold is often unstable at best , we felt that adding such a part to the model was not as crucial as it would be in other countries where the cooling effect plays a much more important role . for the applications presented in section [ sec : applications ] of this paper , we thus went for a simpler implementation : given a cooling threshold @xmath67 , a regressor whose coordinates are @xmath68-\infty,\,t_t]}}(u^c)$ ] , for @xmath69 is added to the matrix @xmath58 . it models , in practise , a cooling effect with a fixed cooling threshold and an estimated cooling gradient that is multiplied by the daytype effect . considering the expression in , the model is quite general since the bulk of it could be thought of as the product of two linear regressions , with the added twist of a non - linearity introduced via the threshold parameter @xmath49 ( change - point of the model ) . even though the priors and algorithms constructed in the coming sections do depend on the model introduced here , they can be modify in a rather straightforward manner , should the reader want to tweak the model a bit ( for example to add another exogenous variable such as the wind or the cloud - cover ) . hereafter @xmath70 will denote the likelihood of the observations @xmath71 . we will often write the model for @xmath72 as @xmath73 and use the notation @xmath74 for short , where @xmath75 designate the parameters of interest . with these notations , since @xmath76 , the likelihood of the model described in reads : @xmath77 let us denote @xmath78 and @xmath79 the posterior mean and posterior variance of @xmath80 from a non - informative approach applied to the long dataset @xmath1 , that we assume have already been collected . for the sake of clarity , we drop the @xmath81 notation : when not explicitly specified , the dataset and observations as well as the prior and posterior distributions we refer to in this section will be those corresponding to @xmath0 . we now present the informative hierarchical prior for the model , and then prove that it leads to a proper posterior distribution ( see proposition [ hierarchicalproperpost ] ) . following the methodology exposed in section [ subsec : generalprinciple ] , the informative prior that we propose introduces new parameters to model the similarity between the two datasets , @xmath82 and @xmath83 such that @xmath84 where @xmath85 . the coordinates of the vector @xmath86 can be interpreted as similarity coefficients between parameters of @xmath87 and @xmath88 and the strictly positive scalar @xmath89 can be seen as a way to alternatively weaken or strengthen the covariance matrix as needed . hyperparameters @xmath20 and @xmath90 are more general indicators of how close @xmath1 and @xmath0 are , @xmath20 corresponding to the mean of the coordinates of @xmath86 and @xmath90 being their inverse - variance . @xmath89 , @xmath20 , @xmath90 and @xmath91 of course require a prior distribution too . for @xmath91 we use a non - informative prior ( we chose @xmath92 ) because we do not want to make any kind of assumptions about the noise around both datasets . this prior is non - informative in the sense that it matches jeffreys prior distribution on @xmath91 for a gaussian linear regression . for the three other parameters , based on semi - conjugacy considerations , we use : @xmath93 where @xmath94 , @xmath95 and @xmath96 are fixed positive real numbers such that the prior distribution on @xmath89 , @xmath20 and @xmath90 are vague . these prior distributions are chosen because of their conjugacy properties ( as will be seen in the mcmc algorithm ) . the vagueness requirement that we impose on these priors is motivated by the fact that we want to keep as general a framework as possible without having to tweak each and every prior coefficient for different applications . in the end , the informative prior is built as follows : @xmath97 recalling the notations introduced at the end of section [ sec : methodo ] , the posterior measure is given by @xmath98}\times{[\underline{u},\,\overline{u}]}\times { \mathbb{r}}_+^*}(\|\beta\|_1,u,\sigma^2 ) \nonumber \\ & \quad\quad\times l^{\frac{d}{2 } } \exp\left(-\frac{1}{2}(\eta - k\mu^{{\mathcal{a}}})^\prime l ( \sigma^{{\mathcal{a}}})^{-1}(\eta - k\mu^{{\mathcal{a}}})\right ) \nonumber \\ & \quad\quad\times |r|^{\frac{d}{2 } } \exp\left(-\frac{1}{2}r\sum_{i=1}^d ( k_i - q)^2\right ) l^{a_l-1 } \exp\left(-b_l l\right ) { \mathds{1}}_{{\mathbb{r}}_+^*}(l ) \nonumber \\ & \quad\quad\times |\sigma_q^{-2}|^{\frac{1}{2 } } \exp\left(-\frac{1}{2}\sigma_q^{-2 } ( q-1)^2\right ) r^{a_r-1 } \exp\left(-b_r r\right ) { \mathds{1}}_{{\mathbb{r}}_+^*}(r ) . \nonumber\end{aligned}\ ] ] [ hierarchicalproperpost ] for @xmath99}$ ] denote @xmath100 the matrix whose rows are @xmath101 , \quad t=1,\ldots , n,\ ] ] and suppose @xmath102 has full rank for every @xmath99}$ ] . assume furthermore that @xmath103 and that @xmath104 are observations coming from the model and the posterior measure is then a well - defined ( proper ) probability distribution . first notice that @xmath105 is proportional to @xmath106}}(\|\beta\|_1){\mathds{1}}_{{[\underline{u},\,\overline{u}]}}(u ) \pi(\eta|k , l)\pi(k | q , r)\pi(l)\pi(q)\pi(r),\end{aligned}\ ] ] for almost every @xmath107 and that the function @xmath108 is bounded , for almost every @xmath107 . the posterior integrability is hence trivial as long as @xmath109 itself is a proper distribution which is the case here . we now propose a non - informative prior to use with the long dataset @xmath1 . note that since the dataset @xmath1 is long enough , the choice of the prior distribution used in this situation does not matter much as long as it remains vague enough : the bayes estimator is expected to converge , as the number of observations grows , to the maximum likelihood estimator of the model we use . the main issue when studying the asymptotical behaviour of the posterior distribution or the maximum likelihood estimator is that the likelihood of the model is not continuously differentiable with regard to the heating threshold . isolating the heating part from the rest of the model , @xcite show that this issue can be dealt with and proove among other results that both the bayes estimator and the maximum likelihood estimator are consistent . the non - informative prior is thus to be considered hereafter as an equivalent to the maximum likelihood approach for all intents and purposes . for the sake of clarity again , we drop the @xmath110 notation : when not explicitly specified , the dataset and observations as well as the prior and posterior distributions we refer to in section will be those corresponding to @xmath1 . we show that the use of a non - informative prior distribution leads to a proper posterior distribution ( see proposition [ noninfoproperpost ] ) . we use the following non - informative prior @xmath111 this prior is non - informative in the sense that it matches jeffreys prior distribution on @xmath91 for a gaussian linear regression and matches laplace s flat prior on the other parameters . it leads to the following posterior distribution @xmath112}\times{[\underline{u},\,\overline{u}]}\times { \mathbb{r}}_+^*}(\|\beta\|_1,u,\sigma^2 ) . \nonumber\end{aligned}\ ] ] [ noninfoproperpost ] for @xmath99}$ ] denote @xmath100 the matrix whose rows are @xmath101 , \quad t=1,\ldots , n,\ ] ] and suppose @xmath102 has full rank for every @xmath99}$ ] . assume furthermore that @xmath103 and that @xmath104 are observations coming from the model , the posterior measure is then a well - defined ( proper ) probability distribution . notice first that @xmath113}}(\|b\|_1){\mathds{1}}_{{[\underline{u},\,\overline{u}]}}(u ) \quad \text{for almost every } y,\end{aligned}\ ] ] and observe then that @xmath114 ^ 2.\end{aligned}\ ] ] let @xmath115}$ ] and denote @xmath116 . we write @xmath117 ^ 2 \\ & = \|y - a_*(\beta_0 , u_0 ) \alpha_*\|_2 ^ 2,\end{aligned}\ ] ] and thus obtain the following equivalence , as @xmath118 and @xmath119 @xmath120 the triangular inequality applied to the right hand side of gives @xmath121 since @xmath122 has full rank , by straightforward algebra we get @xmath123 where @xmath124 is the smallest eigenvalue @xmath125 and is strictly positive . we can hence find an equivalent of the right hand side of as @xmath119 , which is @xmath126 combining , and together , we see that the integrability of the left hand side of as @xmath118 and @xmath119 is directly implied by that of @xmath127 . the latter is of course immediate for @xmath128 as can be seen via a quick cartesian to hyperspherical re - parametrisation . the previous paragraph thus ensures the integrability of @xmath129 over sets of the form @xmath130m(\beta_0,u_0),\,+\infty[}\ } , \quad \forall ( \beta_0 , u_0 ) \in b_+^{d_{\beta}}(0 , 1 ) \times { [ \underline{u},\,\overline{u}]}\ ] ] where the subset @xmath131 is an open neighbourhood of @xmath132 and @xmath133 is a real number depending on @xmath132 . by compacity of @xmath134}$ ] there exists a finite union of such @xmath135 that covers @xmath134}$ ] . denoting @xmath55 the maximum of @xmath136 over the corresponding finite subset of @xmath137 , we finally obtain the integrability of @xmath129 over @xmath138m,\,+\infty[}\}$ ] . the integrability of @xmath129 over @xmath139}\}$ ] is trivial , recalling that @xmath140 is continuous and does not vanish over this compact for almost every @xmath107 , meaning its inverse shares these same properties . the condition `` @xmath141 has full rank '' mentioned above is typically verified in our applications for the regressors used in our model . to see this , call `` vector of heating degrees '' the vector whose coordinates are @xmath142 , then not verifying the aforementioned condition is equivalent to saying that there exists an index @xmath54 and a threshold @xmath49 such that the family of vectors formed by the regressors @xmath58 and the vector of heating degrees is linearly dependent over the subset @xmath143 of the calendar . in this section we simulate a long dataset @xmath1 and a short dataset @xmath0 from the model to assess the performance of the informative prior as the similarity between the datasets varies . to measure the improvement brought by the informative prior we compare the estimation and prediction on dataset @xmath0 with a non - informative prior . for any estimation ( posterior mean and variance ) on a dataset ( be it @xmath1 or @xmath0 ) , the mcmc algorithms would typically run for 500,000 iterations after a small burn - in period . _ predictive distribution . _ the bayesian framework allows us to compute so - called predictive distributions , i.e. the distributions of future observations given past observations . given a prior distribution @xmath144 and the corresponding posterior distribution @xmath145 related to the past observations @xmath71 , the predictive distribution for the future observation @xmath146 is defined as @xmath147 \\ & = \int y_{n+k } g(y_{n+k}|y ) { \,\mathrm{d}}y_{n+k}. \label{optimalbayespred}\end{aligned}\ ] ] _ the comparison criterion . _ to assess the quality of the estimation of the model with our informative prior with regard to the estimation of the model with the non - informative prior , we compare both results based on the quality of the predictions . let @xmath148 be the upcoming observation , the prediction error can be written as @xmath149 + [ f_{n+1}(\eta_0)-\widehat{y}_{n+1}],\ ] ] which expresses the prediction error as a sum of a noise @xmath150 ( whose theoretical distribution is @xmath151 ) and a bias which can be seen as an estimation error over the prediction @xmath152 . we focus solely on the second part , since the first part ( the noise ) is unavoidable in real situation . given that we want to validate our model on simulated data , the quantity @xmath152 is indeed accessible here whereas it would not be in real situation . we thus choose to consider the quadratic distance between the real and the predicted model over a year as our quality criterion for a model , i.e. : @xmath153 ^ 2}.\label{comparisoncriterion}\end{aligned}\ ] ] both datasets @xmath1 and @xmath0 were simulated according to the model with @xmath154 ( 4 frequencies used for the truncated fourier series ) . the calendars and the partitions used for @xmath1 and @xmath0 were designed to include 7 daytypes ( @xmath155 , one daytype for each day of the week ) , but did not include any special days such as bank holidays . they also included 2 offsets ( @xmath156 ) to simulate the daylight saving time effect . in the end we thus had @xmath157 and @xmath158 i.e. @xmath159 using the expression of the model given in . _ dataset @xmath1 . _ we simulated 4 years of daily data for @xmath1 with parameters : @xmath160 these values were chosen to approximately mimic the typical electricity load of france up to a scaling factor . the temperatures we used for the estimation over @xmath1 are those measured from september 1996 to august 2000 at 10:00am . _ dataset @xmath0 . _ we simulated 1 year of daily data for @xmath0 with parameters : @xmath161 where the coordinates of the true hyperparameters @xmath86 were allowed to vary around @xmath21 . the temperatures we used for the estimation over @xmath0 are those measured from september 2000 to august 2001 at 10:00am . we also simulated an extra year of daily data @xmath0 for prediction , with the same parameters but using the so - called normal temperatures , meaning that for each day of this extra year the temperature is the mean of all the past temperatures at the same time of the year . we made such a choice to try and suppress any dependency between our simulated results and the chosen temperature for this fictive year of prediction , since we did not want to bias our results because of a rigorous winter or an excessively hot summer . we chose to use vague priors ( i.e. proper distributions with large variances ) for the uppermost layers of our informative hierarchical prior , and thus decided to use the values : @xmath162 a study of the bayesian hierarchical model s sensitivity to these values showed that changing these hyperparameters to achieve prior variances of greater magnitudes hardly influenced the posterior results ( means and variances ) at all . this is why we decided to stick to these values for the remainder of our experimentations . _ _ we benchmarked the bayesian model with its informative prior against its non - informative prior counterpart for different choices of true hyperparameters @xmath86 over 300 replications ( data being simulated anew for each replication ) , i.e. we simulated many different datasets @xmath0 looking more or less similar to @xmath1 and applied our method on them . figure [ fig : mean10 ] shows the posterior error of @xmath80 ( posterior mean minus the true value ) of @xmath80 , based on 300 replications that correspond to the case where @xmath163 i.e @xmath164 for both the informative ( leftmost ) and non - informative ( rightmost ) method . marginal confidence interval for the posterior means are much smaller when using the informative prior ( most of them hitting the true value ) . the marginal posterior standard deviations ( not shown here ) are also reduced when the informative prior is used instead of the non - informative prior . ( seasonal parameters ) , @xmath165 ( shape parameters ) , and @xmath166 and @xmath49 ( heating parameters ) , based on 300 replications . leftmost replications correspond to the informative method while the rightmost replications correspond to the non - informative method . here @xmath163 . ] except @xmath167 and @xmath168 . ] ( seasonal parameters ) , based on 300 replications . top row is for the case were @xmath163 and bottom row is for the case where @xmath167 and @xmath168 . leftmost replications correspond to the informative method while the rightmost replications correspond to the non - informative method . posterior errors of @xmath169 ( shape parameters ) , and @xmath170 and @xmath171 ( heating parameters ) are not shown here because no significant deviation from @xmath172 was found on either of these coordinates when the informative prior was used in either case ( the empirical variances on these coordinates were bigger in the non - ideal case though , in a similar fashion to what we observe here for @xmath173 ) . ] when the situation is far from being as ideal as the one mentioned above , the informative approach still shows improvement over the non - informative approach but to a lesser extent . figure [ fig : seasonheat_mean05 ] shows that the estimations of some of the parameters of the model are improved with the addition of the prior information ( @xmath174 and @xmath49 ) while some are not ( @xmath165 and @xmath166 ) in the case where @xmath167 and @xmath168 . situations such as @xmath175 and @xmath176 or @xmath177 and @xmath178 were studied too and yielded very similar results i.e. lesser improvements on the estimations of some parameters only . note that when some coordinates of @xmath86 are valued to @xmath179 while some are valued to @xmath21 , the `` similarity '' between @xmath1 and @xmath0 is very weak . the strength or weakness of the similarity between @xmath1 and @xmath0 can not be diagnosed directly from the posterior mean of @xmath86 itself but we will see that the estimations of the hyperparameters @xmath20 and @xmath90 may provide a partial answer to this question . we also estimated the hyperparameters ( see section [ subsec : informativeprior ] for the specifications of @xmath180 ) when the informative prior was used . let us first study the hyperparameter @xmath86 . its coordinates seem correctly estimated for the ideal situation where @xmath163 as illustrated in the top row of figure [ fig : mean_k10et05 ] which shows the posterior error of @xmath86 . when @xmath167 and @xmath168 , the estimations obtained are of lesser quality as demonstrated in the bottom row of figure [ fig : mean_k10et05 ] : most of the seasonal similarity coefficients appear to be biased ( while the posterior standard deviation on each coordinate , not shown here , are greater than in the ideal situation ) . these estimations may thus be used to quantify the closeness of the two datasets . the estimation of the hyperparameter @xmath89 itself does not seem to provide a lot of information about the data : during our simulations , its mean value exhibited a lot of variability around the same value over the 300 replications for each of the five simulated scenarios and no reasonable conclusion could be drawn from it . on the other hand , the estimation of the hyperparameter @xmath20 does reveal a bit of information about the two datasets @xmath1 and @xmath0 . it is the mean of the coordinates of @xmath86 on the real axis , as can be seen in the definition of the informative prior in on page . however its use remains somewhat limited in the sense that the parameters @xmath165 of the two datasets are most often very close ( meaning the coordinates of @xmath86 that correspond to them is likely close to 1 ) while other parameters may vary greatly . hence even though @xmath20 provides information about the similarity between @xmath1 and @xmath0 , it can not be interpreted alone and has to be considered jointly with @xmath90 . the left part of figure [ fig : seasonheat_qr ] shows the evolution of the posterior mean of @xmath20 as @xmath181 ranges over @xmath182}$ ] . the estimation of the hyperparameter @xmath90 ( inverse - variance of the prior distribution on @xmath86 , see again ) does in fact reveal some information about the two datasets too . it is a measure of dispersion of @xmath86 around @xmath20 , in the sense that the ( higher it is , the closer to @xmath20 the coordinates of @xmath86 should be . just like @xmath20 is the mean of the coordinates of @xmath86 , @xmath90 is in fact their inverse - variance . the right part of figure [ fig : seasonheat_qr ] shows a clear decline of @xmath90 when @xmath183 moves away from the ideal value 1 i.e. when the similarity between the datasets @xmath1 and @xmath0 decrease from strong to weak . as we previously stated , the similarity between the two datasets has to be assessed simultaneously with @xmath20 and @xmath90 and not @xmath20 only : the mean @xmath20 could be close to 1 , possibly hinting at a perfect similarity between the two datasets , while the variance @xmath184 could be great which would then indicate huge differences between the two estimated sets of parameters for the two datasets . ( left ) and @xmath90 ( right , on a log scale ) for the informative prior ( abscissas have been jittered a bit to prevent overlapping , and different shades of grey are used to indicate the level of the estimated density ) . 300 replications for each value of @xmath185 tested . in black : the circles correspond to the averages , while the squares correspond to the 5% and 95% empirical quantiles . here @xmath167.,title="fig : " ] ( left ) and @xmath90 ( right , on a log scale ) for the informative prior ( abscissas have been jittered a bit to prevent overlapping , and different shades of grey are used to indicate the level of the estimated density ) . 300 replications for each value of @xmath185 tested . in black : the circles correspond to the averages , while the squares correspond to the 5% and 95% empirical quantiles . here @xmath167.,title="fig : " ] ( left , where @xmath167 ) and @xmath171 ( right , where @xmath186 ) tested . in black : circles correspond to the averages , while squares and diamonds correspond to the 80% and 90% empirical quantiles of these ratios.,title="fig : " ] ( left , where @xmath167 ) and @xmath171 ( right , where @xmath186 ) tested . in black : circles correspond to the averages , while squares and diamonds correspond to the 80% and 90% empirical quantiles of these ratios.,title="fig : " ] _ prediction . _ we compared the informative and the non - informative models using our comparison criterion defined in and computing the ratio between the two models for different values of @xmath187 and @xmath170 , @xmath169 and @xmath171 being both set to 1 . the left part of figure [ fig : gain ] shows the results we obtained for @xmath173 and @xmath170 simultaneously set to the values @xmath188 . note that since the results appeared to be approximately symmetric with regard to @xmath21 ( i.e. for values @xmath189 and @xmath190 ) , we only included one side of the graph in the present article . on average , the bayesian informative model is a clear improvement over the bayesian non - informative one , its performances being maximised when the parameters @xmath191 and @xmath192 are identical ( which is the ideal situation ) . the performances in prediction are obviously somewhat weakened when the difference between the parameters @xmath191 and @xmath192 grows greater , but the use of the informative hierarchical prior still leads to an average improvement of 15% over the non - informative model , as can be seen on figure [ fig : gain ] . the results obtained when @xmath169 or @xmath171 are varying while the other coordinates of @xmath86 are fixed to 1 were very similar ( see for example the right part of figure [ fig : gain ] ) . the long dataset @xmath1 needed for the construction of the informative prior corresponds to a specific population in france frequently referred to as `` non - metered '' because their electricity consumption is not directly observed by edf but instead derived as the difference between the overall electricity consumption and the consumption of the `` metered '' population . for this population the data ranged from 07/01/2004 to 07/31/2010 . we illustrate the benefit of choosing our informative prior to predict electricity load on short datasets . we consider two short datasets : the first @xmath0 corresponds to the `` non - metered '' population for erdf , a wholly owned subsidiary of edf that manages the public electricity network for 95% of continental france . this population roughly covers the same people that @xmath1 does , but not exactly . the second dataset @xmath193 corresponds to a subpopulation of @xmath1 and represents around 50% of the total load of @xmath1 . for this application , only the days for which no special tariffs are enforced were considered : the so - called ejp ( effacement jour de pointe = peak tariff days ) were removed from the dataset beforehand to ensure the signal studied was consistent throughout time . bank holidays ( including the day before and the day after to avoid any neighbourhood contamination effects ) , the summer holiday break ( august ) and the winter holiday break ( late december ) were also removed from the dataset for this first application , as we wanted to benchmark our method against others on a smooth and rather easy - going signal , so as not to put any one method at a disadvantage due to the signal s specificity . the temperature considered in the model is the average temperature over france for the period of study , and the cooling threshold was chosen to be 16c throughout the 48 instants of the day . we benchmarked our bayesian method with informative prior against four alternative methods , comparing their predictions on dataset @xmath0 in two configurations . roughly speaking , for our first experiment we estimated the model for @xmath0 over the period ranging from 12/01/2009 to 06/30/2010 and predicted the next 30 days ( same as the application presented section [ subsec : applihyper ] ) , while for our second experiment , we estimated the model for @xmath0 over the period ranging from 01/01/10 to 07/31/10 and predicted the previous 30 days . we expect the first configuration to be the easy case and the second configuration to be the tough case , the signal being very smooth during summer and not so much during winter . the figures shown in table [ tab.datasize ] summarise the exact lengths of the various datasets for both experiments . the four alternative methods we benchmarked against our own bayesian informative method , relied on four different techniques : the first was the bayesian non - informative method that we exposed earlier in section [ subsec : noninformativeprior ] ( recall that it was meant to be an equivalent to the maximum likelihood approach ) , the second involved non - parametric estimation with kernels ( see * ? ? ? * ) , the third was a double exponential smoothing ( see * ? ? ? * ) and the fourth and last was an arima model . note that for the second experiment , the data available obviously had to be time - reversed in order to apply some the last three alternatives methods since time - dependence plays an important role for them . the arima model was automatically selected ( see * ? ? ? * ) and was the best model with regard to the aic criterion . .sample size ( in days ) of the datasets for both experiments . [ cols="^,^,^,^",options="header " , ] figure [ fig : mape_adelaide_mois ] shows the average error in prediction for each month while figures [ fig : mape_adelaide_instant_noninfo ] and [ fig : mape_adelaide_instant_info ] display the average error in prediction for each half hour ( from 00:00 to 23:30 ) . it is important to note that the use of the non - informative prior leads to overfitting the model : the results presented in table [ tab : qualite_global ] show that as the estimation window goes smaller , the estimation error decreases while the prediction error grows very quickly . a close inspection of the posterior densities of the different parameters of the model revealed that the bias induced by the increasing lack of data is mainly seasonal : this is due to the seasonality coefficients of the model being overfit . choosing the informative prior over the non - informative prior makes the estimation and prediction of the model more robust with regards to the lack of data . the informative prior especially improves the quality of the predictions when the lack of data is severe : it provides reasonable forecasts even in the worst scenario considered here , where only 6 months of estimation were used for 6 months of prediction . in this situation , estimation ( from 07/01/2010 to 12/31/2010 ) and prediction ( from 01/01/2011 to 06/30/2011 ) are performed on non - overlapping areas of the calendar : the informative prior makes up for the unavailable data and prevents the model from overfitting on the second half of the calendar , while the non - informative prior does not and consequently leads to heavily biased predictions over the first half of the calendar . the two mcmc algorithms presented below were developed because direct simulations from the posterior distribution were not possible . the justifications are given after the algorithms themselves . notice that the full conditional distributions of all the parameters but the threshold @xmath49 appear to be common distributions in both cases , due to the presence of multiple semi - conjugacy situations . we used a metropolis - within - gibbs algorithm ( see * ? ? ? * , for a quick description ) based on gibbs sampling steps for every parameter but @xmath49 for which we used a metropolis - hasting step based on a gaussian random walk proposal . the algorithm corresponding to the non - informative prior is detailed first since it is the simplest of the two . we define the ( commutative and associative ) operator @xmath194 as @xmath195^{-1}(\sigma_1^{-1 } \mu_1 + \sigma_2^{-1 } \mu_2 ) \\ ~[\sigma_1^{-1 } + \sigma_2^{-1}]^{-1 } \end{array}\right)\end{aligned}\ ] ] for any vectors @xmath196 and @xmath197 in @xmath198 , for any symmetric positive definite matrices @xmath199 and @xmath200 of size @xmath201 . [ lemmaconj ] let @xmath202 and @xmath203 be two random truncated gaussian vectors in @xmath198 @xmath204 and denote @xmath205 and @xmath206 their respective densities , then @xmath207 is integrable . let furthermore @xmath208 be a random variable with density @xmath209 , then @xmath208 has truncated gaussian distribution @xmath210 where @xmath211and this result easily extends to any finite number of random truncated ( or not ) gaussian vectors . [ lemmacond ] let @xmath212 be a random gaussian vector in @xmath198 @xmath213 } \sim { \mathcal{n}}\left({\left[\begin{array}{c}\mu_1 \\ \mu_2\end{array}\right ] } , { \left[\begin{array}{cc}r & s \\ s^\prime & t\end{array}\right]}^{-1}\right)\ ] ] and @xmath202 and @xmath203 the projections of @xmath212 over its @xmath214 first and @xmath215 last coordinates ( @xmath216 ) . the conditional distribution of @xmath202 with regard to @xmath203 is then gaussian @xmath217 [ lemmalinearreg ] let @xmath212 and @xmath208 be two random vectors respectively in @xmath198 and @xmath218 such as the conditional distribution of @xmath208 with regard to @xmath212 is gaussian @xmath219 with @xmath55 matrix of size @xmath220 that has full rank @xmath221 , and let @xmath222 be a fixed vector in @xmath218 . the conditional distribution of @xmath212 with regard to @xmath208 is then gaussian too @xmath223^{-1 } m^\prime ( y - z ) , \sigma^2 m^\prime m\right).\ ] ] denoting @xmath224 , straightforward algebra leads immediately to @xmath225^\prime \sigma^2 m^\prime m \big[(m^\prime m)^{-1}m^\prime w - x\big ] \\ & \quad\quad - \big[(m^\prime m)^{-1}m^\prime w \big]^\prime \sigma^2 m^\prime m \big[(m^\prime m)^{-1}m^\prime w\big ] \\ & \quad\quad + w^\prime\sigma^2 i_n w\end{aligned}\ ] ] where the two last terms on the right hand side of the equation do not depend on @xmath212 . in the lines below , we give the different steps of the mcmc algorithm we used to ( approximately ) simulate @xmath226 according to the posterior distribution @xmath227 corresponding to the non - informative prior we presented earlier . the algorithm goes as follows : 1 . initialise @xmath228 such that @xmath229 2 . for @xmath230 , repeat 3 . simulate @xmath231 i.e. @xmath232 4 . simulate @xmath233 i.e. @xmath234 5 . simulate @xmath235 i.e. @xmath236 6 . simulate @xmath237 i.e. @xmath238 7 . simulate @xmath239 , simulate @xmath240}$ ] and define @xmath241 * define @xmath242 if @xmath243 * or @xmath244 otherwise + where the covariance matrix @xmath245 used in this last metropolis - hastings step is first estimated over a burn - in phase ( the iterations coming from this phase are discarded ) , and then fixed to its estimated value `` asymptotically optimally rescaled '' for the final run by a factor @xmath246 ( as recommended for gaussian proposals in section 2 of * ? ? ? * ) . the justifications for each full conditional distribution used in the gibbs sampling steps , including the explicit expressions of @xmath247 , are now given . lemma [ lemmalinearreg ] is a key element to these justifications . _ full conditional distribution of @xmath174 . _ denote @xmath248 the size of the vector @xmath174 , and @xmath249 the vector @xmath2 from which the coordinates corresponding to @xmath174 have been removed . the full conditional distribution of @xmath174 can directly be deduced from both the prior and the likelihood contributions to it . let us first observe that , since the prior distribution we are using on @xmath174 is flat , the full conditional distribution of @xmath174 is in fact proportional to the likelihood function ( seen as a function of @xmath174 ) . now considering the likelihood contribution , we write @xmath250 let now @xmath251 be the diagonal matrix whose diagonal coefficients are given by @xmath252 let @xmath253 be the vector whose coordinates are given by @xmath254 and denote @xmath255 the matrix @xmath256 . we can now rewrite @xmath257 and get @xmath258 using lemma [ lemmalinearreg ] , it is then straightforward to see that the full conditional distribution of @xmath174 is gaussian @xmath259 where @xmath260^{-1 } m_\alpha^\prime ( y - z_\alpha)\\ \sigma^2 m_\alpha^\prime m_\alpha \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath165 . _ using similar arguments , we obtain the full conditional distribution of @xmath165 . namely , denoting @xmath261 the vector whose coordinates are given by @xmath262 and calling @xmath263 where @xmath264 is the diagonal matrix whose diagonal is @xmath265 , we obtain the truncated gaussian distribution @xmath266 where @xmath267^{-1 } m_\beta^\prime ( y - z_\beta)\\ \sigma^2 m_\beta^\prime m_\beta \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath166 . _ using again similar arguments , we obtain the full conditional distribution of @xmath166 . namely , denoting @xmath268 the vector whose coordinates are given by @xmath269 and calling @xmath270 the vector whose coordinates are @xmath142 we obtain the gaussian distribution @xmath271 where @xmath272^{-1 } m_\gamma^\prime ( y - z_\gamma)\\ \sigma^2 m_\gamma^\prime m_\gamma \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath91 . _ no calculations are required , as we immediately identify an inverse - gamma distribution from . in the lines below we give the different steps of the mcmc algorithm we used to ( approximately ) simulate @xmath226 according to the posterior distribution @xmath227 corresponding to the informative prior we presented earlier . the algorithm goes as follow : 1 . initialise @xmath228 such that @xmath229 2 . for @xmath230 , repeat 3 . simulate @xmath273 @xmath232 4 . simulate @xmath274 @xmath275 5 . simulate @xmath276 @xmath277^{-1}(\sigma_q^{-2 } + r \sum_{i=1}^d k_i ) , [ \sigma_q^{-2}+rd]^{-1}\right)\ ] ] 6 . simulate @xmath278 @xmath279 7 . simulate @xmath280 @xmath281 8 . simulate @xmath282 @xmath283 9 . simulate @xmath284 @xmath285 10 . simulate @xmath286 @xmath287 11 . simulate @xmath288}$ ] and define @xmath241 * define @xmath242 if @xmath289 * or @xmath244 otherwise where the covariance matrix @xmath245 used in the metropolis - hastings step is first estimated over a burn - in phase , and then fixed to its rescaled estimated value for the real run as in the non - informative approach . the justifications for each full conditional distribution used in the gibbs sampling steps , including the explicit expressions of @xmath290 and @xmath291 , are now given . to derive these full conditional distributions , we will make use of the technical lemmas [ lemmaconj ] , [ lemmacond ] and [ lemmalinearreg ] presented earlier . _ full conditional distribution of @xmath174 . _ denote @xmath248 the size of the vector @xmath174 , and @xmath249 the vector @xmath2 from which the coordinates corresponding to @xmath174 have been removed . the full conditional distribution of @xmath174 can directly be deduced from both the prior and the likelihood contributions to it . denote @xmath292 , and write the full conditional distribution of @xmath174 as @xmath293 where @xmath294 is the contribution of the likelihood ( seen as a function of @xmath174 to the full conditional distribution ) and @xmath295 is the contribution of the prior ( seen as a function of @xmath174 ) . we prove that @xmath296 and @xmath297 both correspond to gaussian distributions before using lemma [ lemmaconj ] to combine them into yet another gaussian distribution . 1 . let us first consider the prior contribution @xmath297 . recall first that @xmath174 only appears in the following component of the prior @xmath298 which directly implies that @xmath299 denote @xmath300 , @xmath301 and denote @xmath302 and @xmath303 the vectors resulting from the extractions of the coordinates corresponding to @xmath174 and @xmath304 from @xmath257 . finally denote @xmath305 the matrix resulting from the extraction of the rows and columns both corresponding to @xmath174 of @xmath306 and denote @xmath307 the one resulting from the extraction of the rows corresponding to @xmath174 and columns corresponding to @xmath304 of @xmath306 . using lemma [ lemmaconj ] ( and reordering indexes if necessary ) it is straightforward that @xmath295 is proportional to the density of a gaussian distribution @xmath308 2 . let us now consider the likelihood contribution . using exactly the same notations that we used for the full conditional distribution of @xmath174 for the algorithm associated with the non - informative approach we immediately find that @xmath294 is proportional to the density of a gaussian distribution @xmath309^{-1 } m_\alpha^\prime ( y - z_\alpha ) , \sigma^2 m_\alpha^\prime m_\alpha)\end{aligned}\ ] ] just as in . 3 . with the help of lemma [ lemmaconj ] and using the two results above , we can now deduce the posterior conditional distribution of @xmath174 and obtain the gaussian distribution @xmath310 where @xmath311^{-1 } m_\alpha^\prime ( y - z_\alpha)\\ \sigma^2 m_\alpha^\prime m_\alpha \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath165 . _ using similar arguments , we obtain the full conditional distribution of @xmath165 . namely , keeping the notation introduced to derive , and combining the prior and the likelihood contributions together with lemma [ lemmaconj ] we obtain the truncated gaussian distribution @xmath312 where @xmath313^{-1 } m_\beta^\prime ( y - z_\beta)\\ \sigma^2 m_\beta^\prime m_\beta \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath166 . _ using again similar arguments , we obtain the full conditional distribution of @xmath166 . namely , keeping the notation introduced to derive , and combining the prior and the likelihood contributions together with lemma [ lemmaconj ] we obtain the gaussian distribution @xmath314 where @xmath315^{-1 } m_\gamma^\prime ( y - z_\gamma)\\ \sigma^2 m_\gamma^\prime m_\gamma \end{array}\right).\end{aligned}\ ] ] _ full conditional distribution of @xmath86 . _ we denote @xmath316 and first notice that @xmath317 . using the definition of the informative prior and lemma [ lemmaconj ] , we then immediately derive @xmath318 where @xmath319 _ full conditional distribution of @xmath320 and @xmath91 . _ no calculations are required , as we respectively identify a gamma distribution , a gaussian distribution , a gamma distribution , and an inverse - gamma distribution from . the authors would like to thank adlade priou for collecting a part of the data as well as the corresponding results , and virginie dordonnat for the insightful discussions .

the elasto - viscoplastic deformation of solids is a complex process that couples widely separate time and length scales . in polycrystalline solids , plastic deformation is mainly mediated by dislocations , which form complex and strongly interacting networks that give rise to a rich macroscopic phenomenology . in disordered solids ( e.g. noncrystalline / amorphous solids below their glass temperature and granular materials ) , which lack long - range structural order , plastic deformation is mediated by localized shear - driven rearrangements involving a small number of basic elements ( molecules / atoms / grains ) . these rearrangements produce inhomogeneous mesoscopic deformation fields that affect their surroundings . in both cases , the small scale dynamics are highly heterogenous , intermittent and seemingly non - deterministic , yet the emergent macroscopic phenomena are quite robust . developing predictive continuum models that effectively bridge over the widely separate scales remains a major challenge of prime scientific and technological importance . the most successful methodology for bridging over widely separated scales is equilibrium statistical thermodynamics . however , the viscoplastic deformation of solids is an intrinsically out - of - equilibrium phenomenon , hence standard equilibrium thermodynamics is not directly applicable in this case . nevertheless , statistical concepts and the most general form of the laws of thermodynamics might still be applicable . our goal in this paper is to incorporate recently developed out - of - equilibrium statistical concepts into a continuum thermomechanical framework for the large elasto - viscoplastic deformation of amorphous solids . models for the elasto - viscoplastic deformation of amorphous solids have been quite extensively discussed in recent literature , see for example @xcite . the basic idea that has emerged in the statistical physics community in the last 15 years or so , is that the degrees of freedom of an amorphous / glassy material ( i.e. the various contributions to the internal energy and entropy ) can be approximately decomposed into two qualitatively different groups @xcite . one group includes the vibrational motions of particles around a mechanically stable configuration ( sometimes termed an inherent structure ) . these motions equilibrate with the thermal reservoir on a microscopic timescale ; i.e. the temperature apparent in the vibrational fluctuations quickly evolves to match that of the reservoir . the other group includes the mechanically stable configurations themselves and the slow configurational rearrangements that allow transitions between them . these are mainly driven by external mechanical forces , possibly assisted by thermal fluctuations . it is the configurational degrees of freedom that fall out of internal equilibrium with the vibrational degrees when a glass is formed ; and it is these degrees of freedom that undergo structural relaxation when a glass ages and evolve when a glass deforms plastically . numerous recent theoretical , numerical and experimental studies have provided quantitative evidence that the configurational degrees of freedom of either an aging or a deforming amorphous / glassy material attain a temperature of their own , which is markedly different from the temperature corresponding to the vibrational fluctuations of the microconstituents . the early works demonstrating the existence of a configurational temperature , usually termed an `` effective temperature '' in the statistical physics community , were rather exclusively based on deviations from the equilibrium fluctuation - dissipation relation ( for a very recent and comprehensive review see @xcite ) . that is , it was shown ( mainly using simple models and molecular dynamics simulations ) that the long time dynamics of an aging glassy material satisfy a generalized fluctuation - dissipation relation with a temperature that is different from the temperature associated with the vibrational motion of the microconstituents @xcite . of special importance for our present purposes are the colloidal glass experiments of @xcite , which explicitly demonstrate that the effective temperature of a sample slowly evolves toward the thermal ( i.e. vibrational ) temperature over very long times , and eventually approaches it . this suggests that the two temperatures interact thermodynamically through an equilibration process akin to fourier s relation to reach a common value . generalized fluctuation - dissipation relations were also tested for steadily deforming simulated glasses , again demonstrating the existence of a well - defined effective temperature @xcite . another line of investigation focused on `` athermal systems '' , i.e. systems where no ordinary thermal ( vibrational ) fluctuations take place either due to coupling to a zero temperature bath ( e.g. in computer simulations ) or when the basic elements are too big to experience thermal vibrations . realistic examples include foams ( made of bubbles ) and granular materials ( made of grains ) where the large size of the basic elements induces dissipation that rapidly quenches vibrations . in @xcite , durian s foam model @xcite was studied under slow , steady shear deformation conditions . five different definitions commonly used to define the ordinary thermal ( vibrational ) temperature pressure and energy fluctuation relations , the green - kubo relation for the shear viscosity , the stokes - einstein relation , and the thermodynamic definition from the rate of change of configurational energy with entropy were tested and it was discovered that all definitions yielded a similar effective temperature , though the system itself was strictly athermal . similar ideas were developed and tested for granular materials @xcite . both @xcite and @xcite showed by explicit calculation that the effective temperature is equivalently given by the derivative of the configurational energy with respect to the entropy of configurational states . this implies that the configurational subsystem behaves as a distinct thermodynamic entity , featuring a configurational temperature defined as the derivative of the configurational energy with respect to the configurational entropy , precisely as we formulate in detail below . experimental verification of the effective temperature concept is currently limited to systems in which the basic elements are sufficiently large to be directly tracked , such as grains in granular materials and colloidal particles in colloidal glasses @xcite . we finally note that the literature on the effective ( configurational ) temperature is immense and can not be exhaustively reviewed here . the interested reader is referred to @xcite for additional information and relevant references . two - temperature thermodynamic frameworks for spatially homogeneous , non - plastically deforming , glassy materials have been investigated by several authors @xcite , and extended models coupling the evolution of internal state variables to the new thermodynamic variables have begun to emerge . applications to amorphous plasticity recent shear - transformation - zone ( stz ) models @xcite , to soft glassy rheology @xcite , to dislocation - mediated crystal plasticity @xcite , to glassy memory effects @xcite and to hard - sphere materials @xcite have been proposed . very recently , this framework has been used to explain a ductile - to - brittle transition in bulk metallic glasses @xcite . with this background , we present the three major goals of this paper : * to determine the general features of a rational thermomechanical framework which permits , at the outset , the existence of a second primitive temperature . we view this as a question of basic , fundamental interest . + * to show that the essential two - temperature principles developed in the statistical physics community can be fit within a consistent thermomechanical framework . + * to highlight certain non - trivial phenomena captured by the resulting constitutive framework , when choosing forms for various constitutive functions . we proceed through several steps . in the first step , we describe the decomposition of the standard continuum laws of energy balance and entropy imbalance ( cf . @xcite ) into sub - balances for the two subsystems . at this point , we demonstrate basic thermomechanical consequences of these hypotheses ( including items 1 , 2 and 3 on the list below ) by presenting a simple example of a two - temperature plasticity model . to develop a micro - statistically motivated model we next revert back to our basic hypotheses and , borrowing certain notions from @xcite , introduce a minimal set of internal state variables , and permit the energy and entropy of the configuration to depend on them . we assume defect - like functional forms for the entropy and energy of the internal variables , and assume constitutive dependences for the other variables consistent with the subsystems decomposition . restrictions based on the second - law of thermodynamics then lead to a set of `` fourier - type relations '' for heat transfer between the subsystems and through space , which give rise to a pair of heat equations , one for each temperature . finally , we develop a specific plastic flow rule based on an stz - micromechanism @xcite . the resulting model is then shown to capture a number of non - trivial phenomena observed in glassy deformation , including 1 . structural rejuvenation ( i.e. increase in structural disorder ) under shearing . 2 . a finite configurational temperature in slow , steady shearing . slow aging of the structural degrees of freedom near the glass transition . newtonian flow near the glass transition . emergence of a flow stress for steady flows at low temperatures . our goal in this section is to mathematically express the basic hypotheses discussed in the introduction and to systematically explore the implications of them for thermodynamics . this will set the stage for the thermomechanical analysis to be presented later . by `` nonequilibrium '' processes we refer to thermodynamic processes that do not immediately maximize the entropy . as a motivational example , consider a closed system containing two identical blocks of a highly heat - conductive material , which make contact on either side of a thermally conducting layer . the layer conducts heat between the blocks ( labeled a and b ) at a rate @xmath0 , where @xmath1 and @xmath2 are the blocks temperatures . if @xmath3 , the system equilibrates on a microscopic timescale , i.e. a temperature difference can not exist for a finite time . if @xmath4 , we have a completely decoupled nonequilibrium system , for which the system is composed of two separate equilibrium subsystems that never approach a mutual equilibrium temperature ( if @xmath5 initially ) . for finite @xmath6 , we obtain a non - trivial , nonequilibrium system : any temperature difference between the blocks gradually approaches an equilibrium temperature and the joint system gradually approaches a maximum entropy state . as discussed above , this latter case is analogous to observations made in amorphous materials . instead of two blocks separated in space , the material is viewed as having two weakly coupled subsystems appearing within the internal energy and entropy functions . consider an idealized amorphous system containing randomly and densely packed interacting particles . focus first on the _ kinetic / vibrational _ ( k / v ) degrees of freedom , which describe random thermal vibrations around a mechanically stable microscopic configuration . in the amorphous setting , these degrees of freedom manifest as the part of the internal energy that arises due to the particles rattling about some mean positions within their cages of nearest neighbors . this subsystem rapidly equilibrates with an external bath . in the vicinity of the glass transition temperature and below it , the timescale of these vibrations is much faster than the slow timescale characterizing structural relaxation and typical rates of deformation . indeed , the subsystems separation is essentially a timescales separation . by averaging over the short vibrational timescale , the energy stored within the microscopic structure itself can be measured , which is classified as energy in the _ configurational _ subsystem . moreover , as the structure is deformed into new configurations , we should also account in this subsystem for the energy due to non - affine motion of the particles ( averaged over the vibrational timescale ) , resulting from the intermittent collective particle - level rearrangements . under most circumstances , the kinetic energy of this movement comprises a very small contribution to the configurational energy . the _ internally nonequilibrium _ nature of the material is characterized by the fact that for a given total internal energy , the distribution of energy into configurational and k / v subsystems does not generally maximize the total entropy . rather , a finite energy exchange rate @xmath7 can exist between the two subsystems , passing energy from the higher to the lower temperature subsystem , which plays the role of @xmath8 in the previous illustrative example . indeed , the internal heat flow @xmath7 accounts for entropy production in the material , together with more conventional sources of entropy production , such as spatial heat fluxes . the finiteness of @xmath7 in the amorphous setting can be distinguished from the example cases of weakly sheared newtonian fluids or elastically deformed crystals . in a weakly sheared fluid , thermal vibrations and stress - relaxing hopping motions occur on timescales that are not many orders of magnitude separated , and all of the degrees of freedom can be characterized by a single temperature . hence , equipartition is satisfied , i.e. the total configurational energy remains in equilibrium with the kinetic degrees of freedom , analogous to @xmath9 . in an elastically deformed crystal , both thermal vibrations and the external driving forces are , by the definition of the elastic regime , too weak to induce structural changes ( e.g. dislocations motion ) , and hence the configurational degrees of freedom are essentially decoupled from the vibrational ones , analogous to @xmath10 . nevertheless , the microscopic kinetic and potential energies due to the crystal vibrations do maintain equilibrium following the equipartition theorem , which is completely analogous to our caricature of the k / v subsystem as single `` block '' in the thermodynamic analysis . in the case of shearing an amorphous solid , as was discussed in detail above ( see , for example , @xcite ) , there exists ample evidence that the configurational degrees of freedom also form a single equilibrated `` block '' , which is distinguished from the k / v `` block '' , and hence two distinct temperatures may be sufficient to approximately describe the internal thermodynamics of the system . this physical picture is schematically sketched in fig . [ fig1 ] . let us denote by @xmath11 the temperature of the configurational subsystem and by @xmath12 the temperature of the kinetic / vibrational subsystem , which will both be formalized in detail in the following sections . the configurational temperature @xmath11 is a generalization of the relatively old notion of a fictive temperature @xcite . it is typically _ higher _ than the thermal temperature @xmath12 , because the structure typically retains memory of configurations characteristic of the higher temperature prior to a rapid quench and is also constantly being rejuvenated by the external driving forces . following our two - block analogy , one would expect a fourier - type relation of the form @xmath13 . we note in passing that thermo - elastic effects ( e.g. thermal expansion , temperature - sensitive elastic moduli ) provide another form of interaction between the two subsystems , which is not mediated by the heat flux @xmath7 . _ we neglect these phenomena in the present work _ to focus more heavily on the deviatoric ( non - volumetric ) plastic response . an important subtlety of the ensuing analysis pertains to the functional dependence of the conductivity @xmath14 . in the absence of plastic deformation , only ordinary thermal fluctuations can pull the system out of a local configuration and induce flow from the configurational subsystem to the kinetic / vibrational one . as the glass transition is approached , thermal fluctuations become increasingly ineffective in spontaneously inducing structural relaxation , and consequently @xmath14 becomes extremely small in the absence of imposed shearing ( it strictly vanishes if there exists an ideal glass transition ) . in the presence of plastic deformation , rearrangements are activated and energy can again flow between the two subsystems , i.e. @xmath15 , at a rate that is affected by the rate of plastic deformation itself . moreover , the rearrangements themselves can mediate diffusion of @xmath11 , when the latter is spatially inhomogeneous . this occurs microscopically through the long range eshelby - like fields that are generated by the localized shear rearrangements ( that can be viewed as eshelby - like inhomogeneities @xcite ) . the direct dependence of the inter - subsystem conductivity and intra - configurational subsystem diffusivity on the rate of plastic deformation is a unique feature of these systems and hence of the corresponding material models . we now turn to formulate these ideas mathematically . consider a deformation process for which @xmath16 is the reference coordinate , @xmath17 is the deformed coordinate and @xmath18 , where @xmath19 is the motion function . the spatial velocity field is then defined by @xmath20 , where @xmath21 denotes the lagrangian time derivative . the local first law of thermodynamics reads @xmath22 where @xmath23 is the mass density , @xmath24 is the internal energy per unit mass , @xmath25 is the cauchy stress tensor , @xmath26 is the spatial velocity gradient ( i.e @xmath27 ) , @xmath28 is the spatial heat flux and @xmath29 is the spatial divergence operator . in the above we have assumed no internal heat sources or sinks , which would otherwise add a scalar heating term to the right hand side . the second law of thermodynamics reads @xmath30 where @xmath31 is the internal entropy per unit mass and @xmath32 the entropy flux . in the spirit of the previous discussion , we now apply the assumption that the thermodynamical quantities can be approximately decomposed into their kinetic / vibrational and configurational contributions , denoted hereafter by the subscripts @xmath33 and @xmath34 respectively . first , we separate the internal energy and entropy into configurational and kinetic / vibrational contributions @xmath35 next , the heat flux is assumed to approximately decompose as @xmath36 where @xmath37 and @xmath38 are the fluxes of configurational energy ( @xmath39 ) and k / v energy ( @xmath40 , respectively . to gain physical insight into this decomposition , note that the heat flux @xmath28 into ( out of ) a part of a body represents the flux of energy due to the flow of randomized movement of the microscopic constituents movements that collectively have zero mean velocity relative to the continuum - scale motion . in our context , we can envision two different manifestations of the passage of randomized movement . the kinetic heat flux @xmath38 may be seen as the `` standard '' flux describing the passage of random vibrational motion into ( out of ) a part . the configurational contribution @xmath37 can be understood as the energy flux due to rearrangements of the structure , which pass energy across space by disturbing the neighboring structural assembly in a random fashion , and which does not count any k / v vibrational mechanisms in so doing . the entropy flux is similarly decomposed into two additive contributions @xmath41 in the classical clausius - duhem inequality , the temperature @xmath42 is positted as a scalar , non - negative field , which relates the heat and entropy fluxes by @xmath43 . we introduce the notion of a second primitive temperature the most noteworthy departure from common thermomechanics in this paper by assuming analogously that each subsystem contribution to the entropy flux is proportional to its corresponding heat flux via an as - yet - undefined temperature field , giving @xmath44 and @xmath45 . as a result @xmath46 the case of @xmath47 corresponds to internal nonequilibrium , whereas @xmath48 in internal equilibrium and we recover the standard entropy flux @xmath43 . we also assume each temperature is necessarily non - negative . we should also mention that our choice of @xmath44 and @xmath45 is somewhat analogous to multi - temperature thermomechanical models for immiscible mixtures @xcite ; instead of @xmath34 and @xmath33 subsystems , each phase of the mixture is permitted its own temperature and the relation between heat and entropy fluxes within each phase , and their additive appearance in the second law , appears in direct analogy to ours . adopting the conventional krner - lee decomposition of the total deformation gradient tensor @xmath49 @xcite , @xmath50 we can write the stress - power as @xmath51 with @xmath52 , @xmath53 , @xmath54 , @xmath55 , @xmath56 and @xmath57 @xcite . we note that @xmath58 maps the reference ( undeformed ) space into the so - called intermediate space and that @xmath59 maps the intermediate space into the deformed space . also note that @xmath60 and hence @xmath61 . in what follows we assume that plastic shear deformation involves no volumetric changes , i.e. we set @xmath62 . in doing so we neither imply that there exists no coupling between plastic shear deformation and dilation / contraction nor that this coupling is of no importance . such coupling is well documented for amorphous solids and might have consequences for their mechanics @xcite . we have developed a theory that includes pressure - sensitivity and shear dilation ; however , the added complexities that arise suit reporting in a subsequent publication . to proceed , we need separate energy balance equations for the two subsystems , which sum together to give the first law of thermodynamics in eq . ( [ 1st ] ) . we propose that these take the following form @xmath63 let us discuss in detail the physical meaning of the different terms on the right - hand - sides of these equations . the spatial contributions to the heat flow _ within _ each subsystem are @xmath64 and @xmath65 for the kinetic and configurational subsystems respectively . @xmath7 is the scalar internal heat flow that transfers energy _ between _ the configurational subsystem and the kinetic one . the rest of the terms represent the way the total stress - power @xmath66 is distributed among the subsystems . first , we note that elasticity resides in the configurational subsystem and hence reversible elastic energy appears in the @xmath34 subsystem energy balance ; recall that we neglect thermo - elastic and volumetric effects , which could conceivably admit an elastic term in the @xmath33 subsystem . the remaining terms give the plastic stress - power @xmath67 . to remain general , we permit the plastic stress - power to affect both subsystems , by splitting @xmath68 into the sum of two contributions @xmath69 . the _ stored plastic power _ describes the rate of stored energy of cold work , and thus contributes in the @xmath34 energy balance . plastic dissipation _ is the rate of energy dissipated as ordinary heat , which is in the form of vibrations and thereby affects the @xmath33 subsystem balance . applying the chain rule to eq . ( [ 2nd ] ) and substituting eqs . ( [ firstk])-([firstc ] ) into the result to eliminate @xmath70 and @xmath71 , we obtain @xmath72 where @xmath73 is the spatial gradient operator . letting @xmath74 be the mass density in the intermediate space , we redefine quantities relative to volumes , areas , and lengths in this space as @xmath75 multiplying eq . ( [ int_space_1b ] ) by @xmath76 and substituting the above quantities , we obtain an intermediate frame entropy imbalance ( for ease of notation , the @xmath77 are omitted hereafter ) @xmath78 similarly , the energy balances become @xmath79 where @xmath80 for any vector field @xmath81 . as was stated in the abstract , internal variables that carry information about the structural state of the deforming solid , and determine its subsequent response , play a central role in a theory of elasto - plasticity . yet , we present here first a simple two - temperature model for rate - dependent elasto - plasticity , which does not include internal variables . our goal is to illustrate general features of the two - temperature thermomechanical framework . we start by defining the configurational and kinetic helmholtz free - energy per unit intermediate volume @xmath82 let @xmath83 be the plastic rate - of - deformation tensor , @xmath84 the scalar plastic flow - rate , and @xmath85 the direction of plastic flow . let us assume plastic incompressibility , @xmath86 . we write the following functional dependences ( note that here , and throughout the paper , we use the @xmath87 symbol to denote constitutive functions of designated independent variables ) @xmath88 we use @xmath89 to denote the deviator , and note that the forms for @xmath90 and @xmath91 assume @xmath92 . these constitutive dependences are not unique . rather , they are informed by past experience with constitutive theories of thermal plasticity , and are properly extended for consistency with the two - subsystem decomposition . in addition , the @xmath93 and @xmath94 dependences are also motivated by the statistical - thermodynamics - based model to be described in the upcoming sections . the above dependences are frame - indifferent ( see appendix [ frame_indifference ] ) , and unlike standard heat transfer theories , we allow all heat flow variables associated with the configurational subsystem to explicitly depend on the plastic rate of deformation . the latter is a key point , as plastic deformation directly influences the time - scale for configurational energy transfer . finally , in the spirit of the subsystems separation , the above constitutive dependences aim at coupling configurational / kinetic dependent variables with configurational / kinetic independent variables . substituting the free - energies into the inequality in ( [ int_space_2 ] ) and the expanding time - derivatives , we obtain @xmath95 as the first three terms are independent of @xmath96 , @xmath97 and @xmath98 , they must vanish in order to ensure non - violation of the inequality @xcite . hence , @xmath99 which represent an elasticity relation and helmholtz relations for each subsystem , respectively . since @xmath100 is assumed independent of @xmath101 , the configurational helmholtz relation implies that @xmath102 . to ensure that the four last terms in ( [ simple_ineq ] ) do not violate the inequality , we assume as a sufficient , but not necessary condition that each is non - negative separately . for the last three terms , non - negativity is assured by the following set of fourier - type relations @xmath103 the term proportional to @xmath104 can be made non - negative under various selections of the functions for @xmath90 and @xmath91 . to obtain a simple , isotropic , incompressible bingham - like rheology , we choose @xmath105 which is valid only for @xmath92 . together , the plastic stress obeys @xmath106 , which makes it evident that @xmath107 is an effective plastic flow viscosity and @xmath108 acts as a yield strength . finally , we insert these results into eqs . ( [ firstki ] ) and ( [ firstci ] ) to obtain heat equations for the two subsystems . applying the chain rule and the definition of the free - energies , we can write @xmath109 and @xmath110 , to obtain @xmath111 \ , \label{conventional_heat_k } \\ c_c \dot\theta_c = & \ y^p_{store } d^p -{\kappa}_{ck}\ ( \theta_c-\theta_k ) + \text{div}_i \left[{\kappa}_c\ , { { \bm{g}}}_c \right ] \ . \label{conventional_heat_c}\end{aligned}\ ] ] to appreciate the implications of the two - temperature framework in general , and of the configurational heat equation in particular , we need to be more specific about how @xmath11 affects the flow rule . that is , we need to specify the @xmath11-dependence of @xmath112 . in the subsequent sections , where we will discuss an internal variables model within a statistical thermodynamic framework , we will show that @xmath104 is proportional to a boltzmann - like factor with @xmath11 appearing as the relevant temperature . for the present discussion this implies that @xmath113 where @xmath114 is boltzmann s constant and @xmath115 is a constant whose physical meaning will be explained later on . the most important thing for our purposes here is that eq . ( [ mu_p ] ) implies that the plastic rate of deformation @xmath104 in eq . ( [ conventional_flow ] ) is very ( i.e. exponentially ) sensitive to variations in @xmath11 . we therefore focus here on the behavior of @xmath11 , according to eq . ( [ conventional_heat_c ] ) , under various thermomechanical conditions : * _ thermodynamic aging _ let us consider a situation in which a homogeneous sample is held at a temperature @xmath12 near its glass temperature with no imposed deformation , @xmath116 . the evolution of the system under these conditions corresponds to `` structural relaxation '' or `` aging '' . structural relaxation is a central glassy phenomenon in which physical properties ( e.g. volume , enthalpy , elastic moduli etc . ) vary spontaneously ( i.e. in the absence of external driving forces ) and slowly with no characteristic timescale @xcite . in many cases , the time evolution is logarithmic . aging is one of the most basic manifestations of the nonequilibrium nature of the glassy state . under the stated conditions eq . ( [ conventional_heat_c ] ) predicts that @xmath117 hence , our model predicts that @xmath11 evolves from its initial value to @xmath12 ; the two subsystems gradually equilibrate , and the @xmath11 relaxation dynamics correspond to structural aging in our nonequilibrium thermodynamic framework ( typically the initial condition satisfies @xmath118 and consequently @xmath11 decreases during the aging process ) . since @xmath119 typically varies slowly and nonlinearly with @xmath11 , the relaxation of @xmath11 to equilibrium at @xmath12 will be slow and strongly non - exponential . + the thermodynamic picture of aging is partially supported by various generalized fluctuation - dissipation analyses of aging systems @xcite . as was mentioned in the introduction , the recent colloidal glass experiments of @xcite explicitly demonstrated that @xmath11 estimated through a generalized fluctuation - dissipation relation equilibrates with @xmath12 for sufficiently long aging times ( cf . 3b there ) . while these observations provide strong support for the existence and importance of @xmath11 in aging dynamics , it is important to note , however , that even if the function representing @xmath120 is chosen so as to quantitatively match experimental aging data , this single - variable description of aging remains incomplete , cf . _ low - temperature homogeneous steady state deformation _ consider then homogeneous steady state deformation at a low temperature @xmath12 . under these conditions we can set @xmath121 in eq . ( [ conventional_heat_c ] ) ( due to the steady state conditions ) , neglect @xmath12 as compared to @xmath11 ( low @xmath12 ) and omit the spatial term ( homogeneous deformation ) . with these we can isolate @xmath11 in eq . ( [ conventional_heat_c ] ) , obtaining @xmath122 this result shows that the configurational temperature @xmath11 attains a finite value ( obtained by solving the above equation if @xmath123 and @xmath14 depend on @xmath11 ) due to a competition between a `` rejuvenation '' term @xmath124 ( acting to increase @xmath11 ) and a `` relaxation '' term @xmath125 ( acting to decrease @xmath11 ) . as discussed above , at the low temperatures @xmath12 of interest here , it is sensible to assume that @xmath126 . this implies that the steady state value of @xmath11 in eq . ( [ conventional_thetac ] ) is _ independent _ of the deformation rate ( if @xmath123 is independent of @xmath127 ) , as is indeed widely observed ( see , for example , @xcite ) . * _ configurational temperature diffusion _ the configurational heat diffusion term @xmath128 $ ] in eq . ( [ conventional_heat_c ] ) is expected to play a role in the presence of inhomogeneous flows and localization . in particular , @xmath129 is inevitably related to a _ length - scale _ and is expected to give rise to , and potentially explain , shear - band diffusive widening . in this regard , it is possible that @xmath11 is related to the so - called `` fluidity '' state variable recently invoked in nonlocal flow theories @xcite . like @xmath11 , the fluidity is hypothesized to obey diffusional dynamics , which enable it to describe shear band widths in various amorphous media . + in more general terms , as the configurational heat conductivity @xmath129 introduces a length - scale , the two - temperature framework might provide a new mechanism for size effects in constitutive models . unlike other size - dependent plasticity models where scale - sensitive plastic flow properties are obtained by introducing gradient - based pairings in the internal / external power statements @xcite , the nonlocal behavior here emerges from the splitting of the usual first law into two pieces . the diffusion of the new temperature @xmath11 then affects the flow through @xmath112 . this coupling gives rise to a physically nonlocal constitutive mechanism , on time / space scales distinct from the usual ( i.e. k / v ) heat diffusion . this nonlocality bears a qualitative similarity to the implicit gradient model of @xcite , where an additional diffusive variable ( the `` fictitious plastic strain '' ) is introduced , and nonlocal plastic flow behaviors are induced by dependence of the flow rule on this variable . it is our goal now to go beyond the illustrative simple model discussed in the previous section and to derive a new constitutive model that integrates coarse - grained internal variables within the proposed framework . in so doing we hope to develop a more basic statistical thermodynamic approach , leading to models with improved predictive powers . internal variables implicitly carry , through their evolution , information about the history of deformation , and , as descriptors of the material state , directly influence the deformation behavior @xcite . the evolution of these variables , we posit , is driven by the exchange of energy and entropy with other parts of the deforming system and hence the internal variables which are naturally associated with the configurational subsystem must be explicitly associated with energy and entropy of their own @xcite . this is a rather significant deviation from conventional approaches , which traditionally do not associate entropy with internal variables @xcite . in an ideal scenario , one should be able to define macroscopic internal variables by systematically coarse - graining over microscopic dynamics . in practice , this is a daunting task and one usually resorts to various phenomenological descriptions , which are sometimes rationalized based on physical intuition , experimental observations , symmetries , computer simulations , etc . an alternative approach would be to first identify mesoscopic objects that are abstracted from the microscopic physics and only then to coarse - grain to obtain the internal variables and their evolution , as in shear - transformation - zone ( stz)-based theories @xcite to be discussed later . before we specify such a microscopic mechanism ( see section [ stz_sec ] ) , our goal in this section is to develop a broad theoretical framework that permits a minimal set of internal variables , to which a more specific micro - mechanism may later be attributed . in amorphous solids , different initial states of disorder would result from different cooling rates through which the glass transition was approached . the different initial structural states may determine whether the stress - strain curve exhibits a stress peak or not , and whether the material shear localizes or not ( see the review papers of @xcite ) . in addition , we know that solids ( both crystalline and amorphous ) carry orientational memory of previous plastic deformation , which affects subsequent deformation . this is demonstrated , for example , in the classical bauschinger effect in metals and glassy polymers , where specimens undergo a decrease in flow strength in compression after experiencing a certain amount of plastic extension . to address these observations in a minimal sense , in what follows we consider one scalar and one ( symmetric , deviatoric ) tensorial internal variable , denoted respectively by @xmath130 and @xmath131 . these variable types are common to isotropic and kinematic hardening models @xcite . in our setting , we assume that the internal variables @xmath130 and @xmath131 arise from dilute , weakly interacting flow defects ( whose precise nature is not specified yet ) that mediate plastic deformation . the internal variable @xmath130 stands for their density ( number per unit intermediate volume ) and the deviatoric tensor @xmath131 describes their ( dimensionless ) average orientation in the intermediate space . defining @xmath132 , we require @xmath133 where @xmath134 represents an isotropic orientational state and @xmath135 a completely anisotropic state where all defects align in the same direction . we use the set @xmath136 , its time derivatives , and the set @xmath137 as the independent variables , and assume the following constitutive dependences @xmath138 we stress that we use entropies and internal energies , rather than free - energies , as the basic quantities because we insist on directly associating entropy and energy with the internal variables . in our view , this constitutes an important deviation from conventional modelling . the above constitutive choices reflect the separation between subsystems by associating each entropy only to the variables naturally arising in that subsystem , and are also frame - indifferent ( see appendix [ frame_indifference ] ) . we now follow , nearly verbatim , the steps described in section [ sec : conventional ] . expanding time - derivatives in eq . ( [ int_space_2 ] ) , we obtain @xmath139 in spite of the fact that we do not yet specify explicit constitutive dependences for the other dependent variables in ( [ second2 ] ) , let us assume they are all independent of @xmath140 . then , since the three first terms in the above inequality appear linear in the independent variables @xmath141 , @xmath142 , and @xmath143 respectively , the corresponding coefficients must vanish to ensure non - violation of the inequality @xcite . consequently , @xmath144 which are the counterparts of eq . ( [ conventional_maxwell ] ) , just expressed in terms of entropies and internal energies , rather than free - energies . now we go beyond the simple model of section [ sec : conventional ] , and similar to @xcite we express the entropy as @xmath145 here , @xmath146 represents the part of the configurational energy corresponding to the energies of formation of the flow defects and @xmath147 is the energy from bulk elasticity . hence , @xmath148 and @xmath149 are the energy and entropy of the large background structure _ surrounding _ the defects , excluding the elastic energy . the @xmath150 degrees of freedom satisfy @xmath151 , and thus equivalently behave as the `` bath '' for @xmath152 . also , under persistent driving , @xmath152 is expected to equilibrate with @xmath150 relatively quickly compared to the ( extremely ) slow equilibration between the @xmath33 and @xmath34 subsystems . we choose @xmath153 such that eqs . ( [ good_cn])-([ini_eta_split ] ) give rise to the following elasticity relation @xmath154 consequently , @xmath155 is _ symmetric _ , and @xmath156 and @xmath157 are assumed symmetric as well . we propose that the energy and entropy of the flow defects take the following forms @xmath158 \ , \ ] ] where @xmath115 is the formation energy of an isolated defect and @xmath159 is the characteristic volume of a defect . the entropy density of defects , @xmath160 , can be obtained in the dilute limit by counting the number of ways defects can be placed on available sites and then given an orientation . the dimensionless entropic factor @xmath161 can be modeled as follows . clearly @xmath162 as there is only one way to align defects to obtain @xmath135 , modulo rigid rotation . it should also approach a maximum value at @xmath134 , the isotropic state , with @xmath163 at @xmath134 due to symmetry . we therefore assume @xmath164 where @xmath165 is the orientational entropy per flow defect when @xmath134 . we can now write @xmath166-\frac{e_z}{\theta_c}\ , \quad\quad \frac{{\partial}\hat\eta_c}{{\partial}{{\bm{m } } } } = -2 k_b \lambda s_0 { { { \bm{m}}}}\ .\ ] ] using these with eqs . ( [ good_cn ] ) in the second law ( [ second2 ] ) yields the following reduced second - law inequality @xmath167\right){\dot{\lambda}}\ge 0 \ .\nonumber\end{aligned}\ ] ] let us now switch to using @xmath168 and @xmath12 , instead of @xmath169 and @xmath170 , as independent variables for the remaining quantities and assume the following constitutive dependences @xmath171 which should be compared to eqs . ( [ conventional_cons2])-([conventional_cons3 ] ) . for consistency with the new independent variables , we invert eqs . ( [ ini_const_deps ] ) and ( [ ini_eta_split ] ) to obtain @xmath172 the above gives important perspective on how the different variables influence different contributions to the total internal energy . the configurational energy decomposes into the energy of the structure outside the defects , which depends on the state of structural disorder through @xmath11 , the defect energy , governed by @xmath130 and @xmath131 , and the reversible elastic energy , governed by @xmath173 . our strategy to prevent a violation of the second law inequality in ( [ second3 ] ) is the same as the one discussed in section [ sec : conventional ] . we constrain the last four terms to be non - negative under all circumstances and guarantee non - negativity of the other terms together ; it is a sufficient strategy for thermodynamic compatibility though not a unique one . non - negativity of the second - row terms is guaranteed by adopting the `` fourier - type relations '' in eqs . ( [ fourier_conventional ] ) , together with @xmath174\ , & \ \ \ \ \kappa_{cd}=\hat{\kappa}_{cd}(\theta_c,\theta_k , d^p ) \ge 0 , \quad\quad \lambda_0=\hat{\lambda}_0(m ) = \frac{e^{(1-m^2 ) s_0}}{v_0 } \ .\label{kcd}\end{aligned}\ ] ] note that at fixed @xmath175 , the condition for @xmath176 is an equilibrium number of defects _ determined by the configurational temperature _ through a boltzmann - like factor @xmath177 . this statistical thermodynamics result , alluded to in section [ sec : conventional ] , goes beyond conventional modelling and is one of the most important consequences of the proposed framework . the various modes of energy transfer in this model are visualized in fig . [ fig2]a . \a ) b ) our goal in this section is to complete the thermodynamic formulation by deriving dissipation inequalities and heat equations , similarly to the procedure discussed in section [ sec : conventional ] . to further exploit the second law inequality to derive additional constraints on the constitutive law , i.e. dissipation inequalities , we note that since all of the terms in the second row of the inequality in ( [ second3 ] ) were already made non - negative above , it is sufficient to satisfy the following reduced inequality @xmath178 where we have replaced @xmath179 with its symmetric part @xmath180 because it is contracted with symmetric tensors . to proceed , we need to discuss the evolution equation for @xmath131 , which represents the average orientation of the flow defects population . we propose a minimal evolution rule of the form @xmath181 where @xmath182 and @xmath183 are positive scalar functions . this equation is shown in appendix [ stz_app ] to emerge on its own from a mesoscopic model of stz dynamics . essentially , the first term produces orientational memory in accordance with the direction and rate of flow , and the latter is a relaxation term that pushes the system to lose orientational memory over time . using the last equation with eq . ( [ etacderivs ] ) , we obtain @xmath184 . substituting this expression into the inequality in ( [ second4 ] ) yields @xmath185 the first term is always non - negative so it is sufficient to ensure the remaining term is non - negative @xmath186 in the next section we will choose functional forms consistent with this inequality . the last step in the thermodynamic development is the derivation of heat equations using the first law relations in eqs . ( [ firstki])-([firstci ] ) . we focus first on eq . ( [ firstki ] ) . recalling that @xmath187 and taking into account the fourier - type relations in eq . ( [ fourier_conventional ] ) , we obtain @xmath188 \ , \ ] ] with @xmath189 . this equation is essentially the same as eq . ( [ conventional_heat_k ] ) , discussed in the context of the simple model in section [ sec : conventional ] . hats are included to remind us which terms have been assigned functional dependences and which are independent variables ( in the current stage of the analysis ) . this is an ordinary heat equation for @xmath12 in which the plastic dissipation @xmath190 and the heat flow between the kinetic / vibrational and configurational subsystems appear as source terms . more interestingly , for the configurational subsystem we apply the chain - rule to express @xmath191 in terms of @xmath98 and simplify the result using eq . ( [ firstci ] ) to obtain @xmath192 \ , \label{heatc}\end{aligned}\ ] ] with @xmath193 . this is an evolution equation for the temperature of the configurational degrees of freedom , @xmath11 , and hence describes the evolution of structural disorder . this is a central result within the present approach and is closely related to eq . ( [ conventional_heat_c ] ) , discussed in the context of the simple model in section [ sec : conventional ] . this section concludes the derivation of the thermodynamic framework . the constitutive model developed so far , along with the relevant balance laws , takes the form : + conservation of linear and angular momentum : @xmath194 . kinematic decomposition and definition of intermediate space variables : @xmath195 , @xmath196 . elasticity relation : @xmath197 . decomposition of the plastic stress : @xmath198 . evolution law for the defect concentration internal variable : @xmath199 $ ] with @xmath200 . kinetic / vibrational heat equation : @xmath201 $ ] . configurational heat equation : @xmath202 $ ] . + the not yet determined functions for the conduction coefficients @xmath14 , @xmath203 , @xmath129 , @xmath204 must all be non - negative . the functions for @xmath180 and @xmath156 , and the coefficients in the @xmath205 relation , have to be derived from physical considerations and shown to satisfy the constraint of eq . ( [ plast_dissip ] ) . while our focus in the present paper is on the thermomechanical framework , and by no means on specific constitutive models , we aim here at discussing a more elaborated internal variables model that goes beyond the simple example discussed in section [ sec : conventional ] . below we present a constitutive model based on an stz micro - mechanism , which rather naturally fits in the thermomechanical framework . the resulting complete set of equations will allow us in the next section to present a few illustrative applications of the model ( full details of the derivation can be found in appendix [ stz_app ] ) . based on extensive observations , we state that irreversible flow in amorphous solids is mediated by a population of flow defects in the form of oriented stz s sparsely distributed localized zones that are more susceptible to shear - driven transformations in a particular direction than their surrounding material ( see fig . [ fig2]b ) . we view stz s not as permanent flow defects , but rather as configurational fluctuations that can appear and disappear during flow and/or in response to thermal fluctuations . we adopt the main physical ingredients of the stz theory , which has been built on the flow - defect theories of turnbull , cohen , argon , spaepen and others @xcite , and incorporate them into the presently developed framework . we define the microscopic parameters : @xmath159 is the volume of an stz ( consistent with its prior usage ) , @xmath115 is the characteristic energy required to produce an stz , @xmath206 is the energy required to shear - transition an stz along its weak direction , @xmath207 is the half - amplitude of the shear strain of an stz transition , and @xmath208 is an attempt frequency for stz transitions . one version of the two - dimensional stz model leads the following equations ( see appendix [ stz_app ] ) : + evolution of the orientation internal variable : @xmath209 which is consistent with eq . ( [ assumed_dotm ] ) and allows a unique determination of @xmath182 and @xmath183 there . + flow rule : @xmath210 these two constitutive equations are then substituted into the second law inequality of ( [ plast_dissip ] ) , which can be satisfied by choosing the following form for @xmath156 . + stored plastic stress : @xmath211 our goal here is to discuss the properties of the internal variables constitutive model described above , focussing on features that go beyond those already discussed in the context of the simple model in section [ sec : conventional ] . in particular , the discussion of thermodynamic aging and configurational temperature diffusion applies here as is , since these are generic features of the proposed framework . the existence of a steady state configurational temperature is also generic , though we discuss it here again in order to gain a somewhat better physical understanding of its emergence . near the glass transition temperature , relaxation times are large ( macroscopic ) , but still fall within the experimental measurement window and phenomena such as newtonian flow and aging can be probed @xcite . our goal here is to explore the emergence of such phenomena within the proposed constitutive law . we first focus on newtonian flows and hence consider spatially homogeneous , pure shear situations . we reduce the analysis to two dimensions of space and assume the deformation is driven by a deviatoric stress of the form @xmath212 , where @xmath213 is the unit - magnitude deviatoric tensor with major principal direction @xmath214 . then we can write @xmath215 , @xmath216 , and @xmath217 for `` @xmath218 '' indicating any of our subscripts . the viscoplastic relation in eqs . ( [ mevolve1])-([d_1st1 ] ) gives @xmath219,\quad d^p \!=\ ! 2\,v_0\gamma_s\,\lambda_s\ , \nu\ , e^{-\delta / k_b\theta_k}\left[\frac{t^p_{dis}}{\tau_0}-m\right ] \ , \end{aligned}\ ] ] where we have assumed eq . ( [ kcd ] ) remains at its fixed - point @xmath220 . also , recall that @xmath221 . a crucial step in analyzing these equations is specifying the thermodynamic coupling coefficient @xmath222 that is responsible for the creation and annihilation of stz s . as was discussed in detail above , this ( and other ) thermodynamic coupling coefficient depends explicitly on the plastic rate of deformation @xmath127 . moreover , in the absence of plastic deformation , @xmath116 , @xmath203 may be finite due to thermal fluctuations alone . therefore , we set @xmath223 where the dimensionless function @xmath224 remains an unspecified function of the configurational temperature @xmath11 . it is reasonable that the thermal attempt frequency @xmath208 sets the time - scale for the @xmath12-dependent term , and the flow - rate plays the analogous role for the @xmath11-dependent term . the physical meaning of the dimensionless function @xmath225 will be discussed soon . we are now in a position to study eqs . ( [ simple_shear ] ) in various situations . first , consider the steady state linear response i.e. small variations about @xmath226 to a small applied stress @xmath227 . the newtonian ( extensional ) viscosity is defined by @xmath228 . to calculate @xmath229 , we first substitute eq . ( [ k_cd ] ) in eq . ( [ simple_shear ] ) to obtain @xmath230 \ .\ ] ] as we are interested in the linear response regime , the second - order term proportional to @xmath231 can be omitted . moreover , as we focus on steady state we can set @xmath232 and use @xmath233 of eq . ( [ simple_shear ] ) to obtain @xmath234 in the vicinity of the glass transition temperature and just below it , we expect spontaneous thermally - activated creation and annihilation of stz s to be strongly suppressed , i.e. @xmath225 becomes small . therefore , @xmath235 is a small parameter and to leading order in the smallness eq . ( [ m_eta ] ) reads @xmath236 by comparing the left side to the expression for @xmath233 in eq . ( [ simple_shear ] ) , we immediately obtain @xmath237 , which tells us that the material flows persistently only because @xmath225 is finite . the physics here is clear : when @xmath225 vanishes , stz s can not be created and annihilated by thermal fluctuations ( recall that mechanically - induced stz creation and annihilation do not contribute at linear order ) and hence existing ( properly oriented ) stz s can undergo transitions in the direction of the applied stress , but once they all do so , plastic flow can not persist . to complete the calculation of @xmath229 we use eq . ( [ tp_store_result1 ] ) for @xmath238 , which together with eq . ( [ m_eta ] ) implies @xmath239 and hence @xmath240 . therefore , we obtain @xmath241 where we used @xmath242 in the linear response regime and the definition of @xmath243 . we observe that @xmath244 , which shows that the newtonian viscosity increases as @xmath225 decreases with decreasing temperature @xmath12 and strictly diverges if @xmath225 vanishes . indeed , the essence of the glass transition puzzle , which is commonly probed by measuring the dramatic increase of @xmath229 with decreasing temperature @xmath12 , is encapsulated in @xmath225 , which at present can not be calculated from first principles . in the liquid state ( at high temperatures ) @xmath245 , while @xmath225 decreases dramatically in the supercooled and glassy regime ( for `` fragile '' glasses @xcite ) . we also note that @xmath229 strictly diverges only if there exists an ideal glass transition , @xmath246 , and that due to the appearance of the arrhenius factor @xmath247 in eq . ( [ stz_eta ] ) , @xmath225 can be identified as the non - arrhenius temperature dependence of the newtonian viscosity , similarly to @xcite . we would like now to explore the nonlinear response in the framework of the proposed model , and in particular the emergence of a yield stress . using eqs . ( [ tp_store_result1])-([simple_shear ] ) , we obtain @xmath248 m \equiv t^p - \bar\tau m \ .\ ] ] we focus on low bath temperatures @xmath12 for which @xmath225 is extremely small and can be neglected . in this case , eq . ( [ k_cd ] ) can be approximated as @xmath249 and the first equation in ( [ simple_shear ] ) takes the form @xmath250\left[\left(\lambda\,v_0\gamma_s\,\hat{f}(\theta_c ) \right)^{-1 } - m\ \text{sign}\left(t^p - \bar\tau m\right ) \right]\ , \ ] ] where @xmath251 and eq . ( [ dp_yielding ] ) were used . to understand how a flow resistance naturally emerges within the model in the nonlinear response regime , we need to analyze the fixed - points of eq . ( [ stability_exchange ] ) and their stability ( see fig . [ fig3 ] ) . the first bracketed term in eq . ( [ stability_exchange ] ) vanishes only at the fixed point @xmath252 . in the second bracketed term , @xmath130 reaches its fixed - point at @xmath253 , which depends itself on @xmath254 . the @xmath254-dependence of @xmath255 is gaussian ( cf . ( [ kcd ] ) ) , so we expect the second bracketed term to have zero or two positive solutions . we assume that the yet to be determined function @xmath224 is such that only one solution exists in the physical range @xmath256 . while this condition is not necessary on general grounds , it will be shown below to give rise to a physically appealing picture of yielding . this single physical solution is denoted as @xmath257 and satisfies @xmath258 while @xmath11 is indeed a dynamic variable in the above expression , it is expected to vary slowly compared to @xmath254 , so we treat it as a constant for the ensuing stability analysis . for stresses @xmath227 such that @xmath259 , @xmath260 is the only stable fixed - point , see fig . ( [ dp_yielding ] ) implies the system approaches @xmath261 . this is the sub - yielding regime in which after transient plastic deformation the flow arrests . as the applied stress @xmath227 further increases , we reach the point where @xmath262 . at this point an exchange of stability occurs , and @xmath263 becomes the stable fixed - point , see fig . [ fig3 ] . in this case , eq . ( [ dp_yielding ] ) shows that @xmath264 , i.e. that the material flows persistently . this happens at a threshold stress @xmath265 for @xmath266 which we interpret as a dynamic yield stress , responsible for the transition from non - flowing to flowing steady states . it is apparently a function of the state of disorder through @xmath11 , ultimately through the details of @xmath224 , which can not be determined through thermodynamics alone . it is clear , that eqs . ( [ stability_exchange])-([yield_stress ] ) make physical sense if @xmath224 leads to a solution with @xmath267 . in summary , we have shown that a steady - state flow resistance in the model emerges as a dynamic exchange of stability in the stz orientation variable , and need not be stipulated a priori as done , for example , in the simple model discussed in section [ sec : conventional ] . the steady state of configurational disorder , i.e. of @xmath11 , was discussed in section [ sec : conventional ] in the context of the simple phenomenological model , demonstrating its generic features . we briefly discuss this issue again here in the context of a the internal variables model . to that aim , we focus again on the configurational heat equation in ( [ heatc ] ) and observe that under rather generic situations in which @xmath268 , @xmath269 and @xmath118 , the only potentially positive term that can be responsible for rejuvenation is @xmath270 . when rejuvenation takes place , it can not persist forever and @xmath11 reaches a steady state under persistent plastic deformation , as is observed in numerous computer simulations of glass forming systems , e.g. those of @xcite . to see how it emerges , we specialize to pure shear isothermal ( in @xmath12 ) situations as above , which allows us to omit all time and space derivatives in eq . ( [ heatc ] ) to obtain @xmath271 @xmath179 was replaced by its symmetric part @xmath180 because it is contracted with a symmetric tensor and we have simplified @xmath272 to @xmath273 , since steady pure shearing under the prescribed driving guarantees positivity of @xmath238 and @xmath233 . to actually calculate the steady - state value , we recall that ( as discussed several times above ) the thermodynamic coupling coefficient @xmath14 includes a deformation - driven part that depends on @xmath127 and a thermally - driven part that depends on @xmath12 . at sufficiently small @xmath12 , we have @xmath274 , where @xmath275 is a positive dimensionless function . this leads to @xmath276 which is an implicit equation for the steady state value of @xmath11 . the most interesting aspect of this result , beyond the existence of a steady state ( which is expected on general grounds ) , is that the steady state value of @xmath11 is independent of the plastic deformation rate @xmath127 and the externally applied stress . for sufficiently low temperatures , @xmath12 can be also neglected . these results are consistent with many numerical simulations as long as @xmath277 is much larger than molecular timescales , see for example @xcite . in this paper we have developed systematically a finite - deformation , nonequilibrium , continuum thermomechanical framework for deforming solids wherein the configurational degrees of freedom evolve slowly compared to kinetic / vibrational degrees , and may be characterized by a temperature of their own . this description of a solid as composed of weakly interacting configurational and kinetic / vibrational subsystems is inspired by recent progress in statistical physics @xcite . within the proposed framework , internal variables play an essential role in characterizing the out - of - equilibrium state of a deforming solid and are explicitly associated with distinct energy and entropy contributions . an important deviation from standard thermodynamics is that coupling and transport coefficients associated with the configurational subsystem feature an explicit dependence on the plastic rate of deformation . we derived a configurational heat equation , in addition to an ordinary heat equation , and showed how a strain - rate independent steady state of the configurational temperature naturally emerges under persistent deformation . first , this general framework was applied to a simple phenomenological two - temperature plasticity model . then , building on the concept of shear - transformation - zones ( stz s ) , and invoking a minimal scenario involving only one scalar and one tensorial internal variables , an internal variables based constitutive law was developed . these constitutive models were shown to capture various amorphous plasticity phenomena , including isotropic and kinematic strain softening / hardening , newtonian viscosity and thermodynamic aging near the glass temperature , a sharp yield stress at low temperatures , flow driven structural rejuvenation and memory effects . in addition , as was discussed in section [ sec : conventional ] , the diffusion of the configurational temperature with the accompanying configurational heat conductivity @xmath129 might offer a new route to describing size effects in constitutive modelling . we hope that the present development can serve as a physically - motivated and self - consistent thermomechanical framework for formulating constitutive models for the elasto - viscoplastic deformation of solids that lack internal equilibrium . this should eventually allow to address a wide range of physical phenomena related to the shape deformations and failure modes of real amorphous materials ( e.g. bulk metallic glasses ) , which find an enormous range of modern engineering applications , but whose faithful description in the context of complex realistic problems is currently missing . in order to make progress in achieving this ambitious goal , we plan to follow up the present investigation in several directions . first , as was mentioned above , we plan to address pressure effects and the coupling between shear and volumetric deformation , which were not discussed here . this is an important aspect of the deformation of amorphous solids and hence should be systematically incorporated into the proposed thermomechanical framework and constitutive modeling . in addition , we need to perform a quantitative analysis of available experimental data to determine specific material parameters and functions , and to test the extent to which the proposed model can quantitatively account for various experimental protocols ( quasi - static loading , dynamic loading , strain - rate controlled experiment , stress controlled experiment , jumps in the applied strain - rate , creep , strain relaxation ) and control parameters ( strain rate , deformation temperature , thermal history etc . ) . we then plan to consider spatially inhomogeneous deformation situations in 2d and to specifically focus on shear band widening dynamics , which involve a lengthscale associated with the diffusion of the configurational temperature . finally , we plan to implement the proposed model in a finite element method code to allow testing and parameters calibration in various 3d problems . we hope to be able to report on some progress along these lines in the not too far future . * acknowledgements * + k.k . acknowledges funding from nsf cbet grant 1253228 . e.b . acknowledges support from the minerva foundation with funding from the federal german ministry for education and research , the israel science foundation ( grant no . 712/12 ) , the harold perlman family foundation and the william z. and eda bess novick young scientist fund . * appendix * our goal in this section is to augment our thermomechanical model with ideas from recent shear - transformation - zone ( stz ) theories @xcite . by doing this , we provide additional information to fill - in remaining details of the constitutive relations , culminating in a shear flow constitutive law . to motivate the connection , we first discuss the stz micromechanism and its ability to replicate the @xmath278 and @xmath205 functional forms deduced in the previous sections . we start by postulating the existence of a population of oriented stz s sparsely distributed localized zones that are more susceptible to shear - driven transformations in a given direction than their surrounding material ( see fig . [ fig2]b ) . we view stz s not as permanent flow defects , but rather as configurational fluctuations that can appear and disappear due to thermal and mechanically - generated noise ( i.e. the mesoscale stress fields that are generated by structural rearrangements ) . therefore , the density of stz s essentially describes the probability to find such a configurational fluctuation . the existence of stz s has been demonstrated in numerous computer simulations and experimental systems , see for example @xcite . below we will adopt the main physical ingredients of the stz theory , which has been built on the flow - defect theories of turnbull , cohen , argon , spaepen and others @xcite , and incorporate them into the presently developed framework . in two dimensions , each stz during its lifetime is described by its orientation @xmath279 with respect to a fixed reference direction . by definition , an stz oriented along @xmath279 is predisposed to undergo a shear ( volume - conserving ) transformation in the direction of the purely deviatoric , symmetric , unit - magnitude tensor @xmath280 whose major principal vector @xmath281 is oriented along the @xmath279 direction . the components of the tensor take the form @xmath282 . the population density of stz s at position @xmath17 at time @xmath283 oriented along @xmath279 is defined as @xmath284 , but we will generally write this without the @xmath17 and @xmath283 for notational simplicity . note that azimuthal periodicity implies @xmath285 . we define @xmath130 to be the ( intermediate ) volume density of stz s , expressed as an integral over all orientations @xmath286 similarly , we define @xmath131 to be a normalized tensorial moment of the stz distribution in the sense that @xmath287 where the minus sign is used for consistency with previous analyses @xcite . note also that by construction @xmath288 and that @xmath131 bears some resemblance to orientational order parameters in magnetic and liquid crystalline systems . the stz s are assumed to be two - state systems that can be driven to `` flip '' from their original orientation @xmath279 to a perpendicular orientation @xmath289 during their lifetime . while this is definitely an oversimplification as the stz s are multi - atom objects that can feature nontrivial internal structures , this assumption was shown to be sufficient to capture a range of glassy deformation phenomena @xcite . one might characterize an stz oriented along @xmath279 as an elliptical region whose semi - major axis aligns with @xmath281 . a `` flip '' is a pure shear transition , that deforms the stz inward along the @xmath281 direction and outward along @xmath290 , leaving the stz , post - transition , oriented in the @xmath290 direction . the rate at which stz s oriented along @xmath279 undergo a shear transformation into a state oriented along @xmath291 depends on the resolved stress , @xmath292 , and thermal fluctuations , as specified by a rate - factor @xmath293 . for simplicity we omit the arguments @xmath12 and @xmath11 in writing the rate - factor hereafter . note that we assume that stz transitions are driven by the dissipative stress @xmath157 . the evolution law for @xmath294 must account for the change in the stz s orientation when it undergoes shear transformations , as well as for the creation and annihilation of stz s due thermal and mechanically - generated structural fluctuations . therefore , we suggest that the evolution of @xmath294 satisfies the following master equation @xmath295 term by term , this equation represents the rate - of - change of @xmath294 in terms of the rate of stz s transitions _ into _ a state @xmath279 , minus those transitions _ out of _ a state @xmath279 , plus a production / annihiation term that tends to equilibrate the stz s population with @xmath11 in an isotropic manner . @xmath203 is a thermodynamic coupling coefficient that plays the role of a noise source that drives the creation and annihilation of stz s . we must emphasize that this form is consistent with the previous thermomechanical analysis ; it agrees exactly with our second - law constraint , eq . ( [ kcd ] ) , i.e. @xmath296 the integral simplifies as above because the transition terms do not affect the total number of stz s macroscopic plastic deformation arises in this model from the accumulated effect of all stz s transitions occurring at a given time . hence , we express the plastic shear ( deviatoric ) flow rate @xmath180 as @xmath297 where @xmath207 is ( half ) the shear strain magnitude of an stz transition . taking the time derivative of @xmath131 defined in eq . ( [ mdef ] ) and using eqs . ( [ lambdadef])-([int_flow ] ) , we obtain the following closed - form equation for @xmath205 @xmath298 it takes the form assumed in eq . ( [ assumed_dotm ] ) , allowing us to identify @xmath299 equation ( [ int_flow ] ) becomes an explicit flow rule when the function @xmath300 is specified . in the most general case this would lead to an integral representation of the shear flow rate , that is not necessarily reducible to a simple closed form relation in terms of @xmath130 and @xmath131 alone . in order to explore the properties of the emerging constitutive law , we proceed by choosing a prototypical form for the @xmath294 distribution compatible with some given @xmath130 and @xmath131 values : @xmath301 it is important at this point to reflect on the status of our analysis so far . we have at hand a flow rule ( eq . ( [ int_flow ] ) , together with eqs . ( [ levolve ] ) , ( [ mevolve ] ) and ( [ n_approx ] ) ) , which is not fully determined until the rate factor @xmath300 is specified . in addition , we still have to satisfy the second law inequality in ( [ plast_dissip ] ) , where @xmath302 is not yet specified ( note that @xmath303 is known , once @xmath302 is specified , through the relation @xmath304 ) . there are ( at least ) two possible routes to proceed ; either we model @xmath302 based on physical , but not purely thermodynamical , considerations and use the second law inequality in ( [ plast_dissip ] ) to constrain the rate factor @xmath300 or we model the rate factor @xmath300 based on physical considerations and use the second law inequality in ( [ plast_dissip ] ) to constrain @xmath302 . the two routes are expected to have different consequences and it is not a priori clear which one is more physically faithful . below we follow the second route by adopting a simple rate - factor @xmath300 and exploring the consequences through several approximations . we consider a simple thermally activated process for stz transitions , similar to the form for atomic jumps in @xcite , for which @xmath305 \ , \ ] ] where @xmath208 is a molecular attempt rate and @xmath206 is an stz transition barrier . using this simple rate factor in the flow rule in eq . ( [ int_flow ] ) , one can perform the integration analytically to obtain the following closed - form expression @xmath306,\ ] ] where @xmath307 and @xmath308 are the first- and second- order modified bessel functions of the first kind and @xmath309 and @xmath310 . expanding eq . ( [ simple_flow ] ) to first - order in @xmath311 , we obtain @xmath312 this simplified flow rule has been used in the body of the manuscript for illustrative purposes without insisting on satisfying the condition @xmath313 . next , we need to make sure that the flow rule in eq . ( [ d_1st ] ) does not violate the second law constraint in ( [ plast_dissip ] ) , which as explained previously , amounts to constraining @xmath302 . we then assume the following constitutive structure @xmath314 , i.e. that @xmath156 is linear in the orientation tensor @xmath131 , which is then substituted together with @xmath180 of eq . ( [ d_1st ] ) in the inequality in ( [ plast_dissip ] ) . the resulting inequality is satisfied if we set @xmath315 the expression for @xmath156 in eq . ( [ tp_store_result ] ) is interesting and deserves some discussion . the prefactor relating @xmath156 and @xmath131 is proportional to @xmath316 , which implies that @xmath156 in this analysis is of _ entropic _ nature . more specifically , it suggests that configurational fluctuations whose magnitude is determined by @xmath317 generate an entropic stress proportional to the orientation tensor @xmath131 . recalling the meaning of @xmath318 ( defined in eq . ( [ m_entropy ] ) ) , a lower bound value for @xmath318 is @xmath319 , obtainable by analyzing the minimal case where flow defects have two possible internal orientations . hence , @xmath320 , which gives positivity of the prefactor in eq . ( [ tp_store_result ] ) . here we show the constitutive dependences discussed in section [ constit_sys ] are objective or frame - indifferent . we suppose a rigid - body change in frame of the observer , @xmath321 where @xmath322 is an arbitrary time - varying rotation about some origin @xmath323 and @xmath324 is an arbitrary translation vector . hence , @xmath325 is the motion as observed by the new observer ; we will continue to use @xmath326 to describe quantities as seen in the new frame . frame - indifference is upheld if satisfaction of the constitutive functions in the original frame implies satisfaction in the new frame . to show this , we first need to transform the field variables into their corresponding forms in the @xmath326 frame and then check that the same constitutive functions correctly relate the transformed variables . we start with the fundamental transformations : @xmath327 under the krner - lee kinematic decomposition , the @xmath49 transformation above is satisfied by letting @xmath328 ( cf . @xcite ) . consequently , @xmath329 and @xmath330 combining the latter expression with the transformation for @xmath25 , the elastic and plastic stresses obey @xmath331 the rate of stored plastic power @xmath332 must , as a material scalar , be frame - invariant ( i.e. @xmath333 ) . since @xmath179 has just been shown to be frame - invariant as well , we deduce @xmath334 , and thus @xmath335 . all scalar constitutive quantities ( e.g. @xmath100 , @xmath338 , @xmath39 , @xmath339 , @xmath11 , @xmath12 , @xmath130 , @xmath7 ) are inherently frame - invariant . combining this fact with the transformation rules above , we have now shown that the dependent and independent variables related under the constitutive functions @xmath340 , @xmath341 , @xmath342 , @xmath343 , @xmath344 , @xmath345 , @xmath346 proposed in sections [ sec : conventional ] and [ constit_sys ] , are all frame - invariant . likewise , the functional dependences as proposed necessarily satisfy frame - indifference . lastly , we check that the heat - flux constitutive functions are also objective . let us temporarily bring back the sub- and superscript @xmath347 to denote intermediate space variables ( recall they were introduced and then dropped in section [ math ] ) . the heat flux vector in the current configuration @xmath38 transforms under the rotating new frame as @xmath348 . pulled back to the intermediate space , we have @xmath349 similarly the temperature gradient @xmath350 transforms as @xmath351 . likewise @xmath352 we thus have that @xmath353 and @xmath354 are both frame - invariant . the same argument gives that @xmath355 and @xmath356 are also frame - invariant . consequently , all dependent and independent variables of the functions @xmath357 and @xmath358 in section [ constit_sys ] are frame - invariant , implying that the functional dependences are themselves indifferent to frame . anand , l. , su , c. , 2005 . a theory for amorphous viscoplastic materials un- dergoing finite deformations , with application to metallic glasses . journal of the mechanics and physics of solids 53 ( 6 ) , 13621396 . anand , l. , aslan , o. , chester , s. a. , 2012 . a large - deformation gradient theory for elastic - plastic materials : strain softening and regularization of shear bands . international journal of plasticity 30 - 31 , 116143 . bouchbinder , e. , langer , j. s. , procaccia , i. , 2007 . athermal shear - transformation - zone theory of amorphous plastic deformation . ii . analysis of simulated amorphous silicon . physical review e 75 ( 3 ) , 036108 . gurtin , m. e. , anand , l. , 2005 . a theory of strain - gradient plasticity for isotropic , plastically irrotational materials . part i : small deformations . journal of the mechanics and physics of solids 53 ( 7 ) , 16241649 . henann , d. , anand , l. , 2008 . a constitutive theory for the mechanical response of amorphous metals at high temperatures spanning the glass transition temperature : application to microscale thermoplastic forming . acta materialia 56 , 3290-3305 . henann , d. , kamrin , k. , 2013 . a predicitive , size - sensitive continuum model for dense granular flows . proceedings of the national academy of sciences of the united states of america 110(17 ) , 67206735 . lu , j. , ravichandran , g. , johnson , w. , 2003 . deformation behavior of the zr41.2ti13.8cu12.5ni10be22.5 bulk metallic glass over a wide range of strain - rates and temperatures . acta materialia 51 ( 12 ) , 3429 3443 . pan , d. , inoue , a. , sakurai , t. , chen , m. w. , 2008 . experimental characterization of shear transformation zones for plastic flow of bulk metallic glasses . proceedings of the national academy of sciences of the united states of america 105 ( 39 ) , 1476914772 . picard , g. , ajdari , a. , lequeux , f. , bocquet , l. , 2004 . elastic consequences of a single plastic event : a step towards the microscopic modeling of the flow of yield stress fluids . european physical journal e 15 ( 4 ) , 371381 . shi , y. , katz , m. b. , li , h. , falk , m. l. , 2007 . evaluation of the disorder temperature and free - volume formalisms via simulations of shear banding in amorphous solids . physical review letters 98 ( 18 ) . song , c. m. , wang , p. , makse , h. a. , 2005 . experimental measurement of an effective temperature for jammed granular materials . proceedings of the national academy of sciences of the united states of america 102 ( 7 ) , 22992304 .

mixture models provide an intuitive statistical representation of datasets structured in groups , clusters or classes @xcite . a complex dataset is decomposed into the superposition of simpler datasets . the inverse problem consists in determining the group decomposition and the statistical parameters characterizing each group . for a fixed number of groups the expectation maximization ( em ) algorithm provides a recursive solution to the inverse problem @xcite . the estimation of the right number or groups has been , however , a great challenge . corrections such as the arkaike information criterion ( aic ) @xcite and the bayesian information criterion ( bic ) @xcite have been derived , penalizing model complexity and overfitting . yet , the number of groups estimated from these criteria is in general unsatisfactory . in contrast , a bayesian approach would not attempt to estimate what is the `` optimal '' number of groups , but instead average over models with a different number of groups @xcite . the bayesian approach is becoming a popular technique to solve problems in data analysis , model selection and hypothesis testing @xcite . many of the original ideas come from the early work of jeffreys @xcite , but it is just recently that they are starting to be used widely @xcite . the application of bayesian approaches to real problems can be , however , quite challenging . in most cases the solution is explored via monte carlo sampling @xcite or variational methods @xcite . the application of variational methods to bayesian problems results in the variational bayes ( vb ) algorithm @xcite . the vb algorithm is a set of self - consistent equations analog to the em algorithm . they can be solved recursively obtaining an approximate solution to the inverse inference problem . these methods have been applied , for example , to gaussian mixture models for real value data @xcite , dirichlet mixture models for categorical data @xcite and the problem of finding graph modules @xcite . here we further study the use of variational methods in the context of bayesian approaches , focusing on data clustering problems . i the first two sections we review the bayesian approach . in section [ variational ] we revisit the connection between the bayesian formulation and statistical mechanics . in section [ models ] we introduce the generalities of generative models with a hidden structure at the samples side and at both the samples and variables side . in section [ s : priors ] we extend the previous work by jaynes @xcite deriving prior distributions based on symmetry properties . we report a correction to his result for the model with a location and scale parameter and an extension of his result for the binomial model to the multinomial model . in the following sections we study the problem of two - sides clustering real value data and of clustering data represented by a hypergraph or bipartite graph . depending on our starting statistical model , we obtain a vb algorithm . because of its bayesian root , the vb algorithms have a built in correction for model complexity or bias and , therefore , they do not require the use of additional complexity criteria . the performance of the vb algorithms is tested in some examples , obtaining satisfactory results whenever there is a significant distinction between the groups . the _ bayesian approach _ is a systematic methodology to interpret complex datasets and to evaluate model hypothesis . its main ingredients or steps are : given a dataset @xmath0 , ( i ) introduce a statistical model with model parameters , @xmath1 , ( ii ) write down the likelihood to observe the data given the proposed model and parameters , @xmath2 , ( iii ) determine the prior distribution for the model parameters based on our current knowledge , @xmath3 , and , finally , ( iv ) invert the statistical model of the data given the likelihood and prior distribution to obtain the posterior distribution of the model parameters given the model and data , @xmath4 . the latter step is based on bayes rule @xmath5 where @xmath6 having obtained the distribution of the model parameters , at least formally , we can determine other magnitudes . for example , the average of a quantity @xmath7 is given by @xmath8 in practice calculating ( [ z ] ) or ( [ ave ] ) is a formidable task . a very powerful approximation scheme is the _ variational method _ the main idea of the variational method is to approximate the generally difficult to handle distribution @xmath4 by a distribution @xmath9 of a more tractable form . in the following we omit the dependency of @xmath10 on @xmath0 and just write @xmath11 . given @xmath11 we can obtain a bound for @xmath12 using jensen s inequality @xmath13 the latter equation can be rewritten as @xcite @xmath14 where @xmath15 , @xmath16 is minus the average log likelihood and @xmath17 is the kullback - leibler divergence of @xmath11 relative to the prior distribution @xmath3 @xcite . equation ( [ f ] ) resembles the usual free energy in statistical mechanics : @xmath18 , where @xmath19 , @xmath20 and @xmath21 are the internal energy , entropy and temperature of the system , the temperature being expressed in units of the boltzman constant @xmath22 . minus the average log likelihood plays the role of the internal energy , the kullback - leibler divergence of @xmath11 plays the role of the entropy and temperature equals one . equation ( [ f ] ) emphasizes the two components determining the best choice of variational distribution @xmath11 : better fit to the data and model bias . how well the data is fitted is quantified by the internal energy @xmath19 ( [ u ] ) . to achieve the best fit , or internal energy ground state , @xmath11 should be concentrated around the regions of the parameter space where @xmath2 is maximum . the best choice in this respect will by the maximum likelihood estimate ( mle ) @xmath23 where @xmath24 in the opposite extreme , when no data is presented to us , the best distribution is that maximizing the kullback - leibler divergence relative to the prior distribution . this maximum entropy ( me ) solution is the prior distribution itself @xmath25 in general , the drive to better fit the data is opposed by the tendency to obtain the least unbiased model . the variational solution is therefore in the middle between the one extreme of biased models fitting the data very well and completely unbiased models giving a bad fit to the data . it is obtained after minimizing ( [ f ] ) with respect to @xmath11 over a restricted class of functions . this variational solution @xmath11 represents the closest distribution to @xmath4 within the class of functions considered . in this section we present the generalities of statistical models with a first level population structure . similar models has been studied in @xcite . our working hypothesis is that there is a hidden population structure , characterized by the subdivision of the population samples into groups . we assume that we are given a dataset @xmath0 which , in some way to be determined , reflects the population structure . the problem consist in inferring this hidden structure and the associated model parameters from the data . to tackle this problem we introduce a statistical model with a built in population structure as a generative model of the data . the population structure and the model parameters are then inferred solving the inverse problem . more precisely * we consider a population composed of @xmath26 elements divided in @xmath27 groups . * the samples assignment to groups is generated by a multinomial model with probabilities @xmath28 , @xmath29 . denoting by @xmath30 the group to which the @xmath31-th sample belongs , we obtain + @xmath32 * given the group assignments @xmath30 , and depending on the dataset , we write down the likelihood @xmath33 to observe the data parametrized by the parameter set @xmath34 . * putting all this together we obtain the posterior distribution + @xmath35 + where @xmath36 and @xmath37 , @xmath38 and @xmath39 are the prior distributions of @xmath34 , @xmath40 and @xmath27 . the form of the prior distributions , except for @xmath39 , is the subject of the next section . the distribution @xmath39 is irrelevant for problems with large datasets . the difference between the log - likelihood of models with different values of @xmath27 is in general of the order of the dataset size and , as a consequence , the contribution of @xmath41 is negligible . thus , in the following sections we simply neglect the contribution given by @xmath39 . finally , we specify the likelihood @xmath33 when addressing specific problems . in some cases we are going to assume that the variables in our dataset are also divided in groups . here we consider a set of @xmath42 variables divided in @xmath43 groups . the variables assignment to groups is generated by a multinomial model with probabilities @xmath44 , @xmath45 . denoting by @xmath46 , @xmath47 , the variable group to which variable @xmath48 belongs we can then write @xmath49 after adding this variable group structure , the posterior distribution ( [ gm1 ] ) is replaced by @xmath50 where @xmath51 and @xmath52 is the prior distribution of @xmath53 . [ priors ] [ cols="<,<,<,<,<",options="header " , ] the choice of the prior distribution @xmath3 is probably one of the less obvious topics in bayesian analysis . currently the predominant choice is the use of conjugate priors . the form of conjugate priors is indicated by the likelihood , making the prior selection less ambiguous . for example , the binomial likelihood @xmath54 suggests a beta distribution for @xmath55 . furthermore , by choosing a beta distribution as a prior , @xmath56 , the posterior distribution remains a beta distribution , but with exponents @xmath57 and @xmath58 . in this sense , the beta distribution is the conjugate prior of the binomial likelihood . a list of conjugate priors relevant for this work is provided in table [ priors ] . yet , the fact that the form of conjugate priors is suggested by the likelihood does not demonstrate that they are the correct choice of priors . moreover , even if we accept their use , it is not clear what is the correct choice for the prior distribution parameters , e.g. @xmath59 and @xmath60 . different methods have been proposed to determine these parameters . in general they are based on _ a posteriori _ analyzes , e.g. calculations , making use of the data in some way or another . such methods violate , however , the concept of prior distribution , defined as the distribution of the model parameters in the absence of the data . an alternative approach is that by jaynes @xcite . according to jaynes , in the absence of any data , the priors should be solely determined based on the symmetries and constraints of the problem under consideration . in this work we make use of jaynes s approach to determine the prior distribution . below we derive jaynes s priors for the cases relevant for this work . consider a problem where the data consists of equally distributed random variables @xmath61 , @xmath62 , taking real values . furthermore let us assume that the likelihood has the form @xmath63 where @xmath64 is a probability density function in the real line and @xmath65 and @xmath66 are a location and scale parameter respectively . our task consist in determining the prior distribution of @xmath65 and @xmath66 . now , suppose @xmath61 represent positions , which could be measured from difference systems of reference and using different units . in this context the prior distribution should be the same regardless of our system of reference and units . more precisely , our system is invariant under the transformations @xmath67 where @xmath68 represents a translation and @xmath69 a change of scale or units . the likelihood is invariant under these transformations and so must be the prior distribution . therefore , @xmath70 the solution to this functional equation is @xmath71 this analysis was first reported by jaynes @xcite . he obtained , however , @xmath72 . this discrepancy is rooted in the fact that jaynes did not take into account that the location parameter @xmath65 follows the same rules than @xmath73 upon the translation and scale transformations . he assumed @xmath74 @xcite while the correct transformation is @xmath75 ( [ t1 ] ) . consider the multinomial model with @xmath27 states @xmath76 where @xmath77 is the number of times state @xmath78 was observed and @xmath28 is the probability to observe state @xmath78 in one trial , @xmath79 and @xmath80 . here we extend the approach followed by jaynes for the binomial model @xcite . the probabilities @xmath28 may be different depending on our believe , e.g. all states are equally probable . different investigators may have different believes , resulting in different choices of @xmath28 . the main assumption is that the prior distribution should be independent of what is our specific believe and , therefore , should be invariant under a believe transformation . _ believe transformation : _ let us represent by @xmath81 the state @xmath78 , and let @xmath82 and @xmath83 be the probabilities to observe state @xmath81 in one trial according to believe @xmath84 and @xmath85 , respectively . from bayes rule it follows that @xmath86 for @xmath29 . the latter equation can be rewritten as @xmath87 for @xmath88 and @xmath89 , where @xmath90 , @xmath91 , @xmath92 and @xmath93 equation ( [ t3 ] ) provides the transformation rules of the probabilities @xmath28 from one system of believe to another . the invariance under the above transformation lead to the functional equation @xmath94 to solve this equation we first need to compute the determinant of the transformation jacobian . the jacobian of the transformation ( [ t3 ] ) has the matrix elements @xmath95 @xmath96 . this matrix can be decomposed into the product @xmath97 , where @xmath98 is a diagonal matrix and @xmath99 has two eigenvalues , @xmath100 and a @xmath101-degenerate eigenvalue @xmath102 . putting all together we obtain @xmath103 ) , with @xmath104 , is given by @xmath105 note that for @xmath106 , @xmath107 and @xmath108 , we recover the result by jaynes for the binomial model @xmath109 the prior distributions ( [ p1 ] ) and ( [ p2 ] ) are improper , i.e. their integral over the parameter space is not finite . at first this may sound an unsuitable property for a prior distribution . nevertheless , the improper nature of these prior distributions is just indicating that the symmetries in our problem are not sufficient to fully determine them . data is required to obtain a proper distribution . the best example for an intuitive understanding of these arguments is the prior distribution of the location parameter . in the absence of any data and under the assumption of translational invariance , it is clear that every value in the real line is an equally probable value for the location parameter , resulting in an improper prior . from the operational point of view , the posterior distribution may be proper even when the prior is not . indeed , the integral @xmath110 may be improper , @xmath111 may be proper . the posterior distribution can be improper when the inference problem has not been correctly formulated or there is not sufficient data to determine the model parameters . to avoid dealing with improper distributions , we can renormalize improper priors to some limit of a proper distribution . since conjugate priors facilitate analytical calculations they are a good starting point . this is illustrated in table ( [ priors ] ) for selected examples . these are the prior distributions used herein . in particular , for the multinomial probabilities @xmath40 and @xmath53 we use the renormalized invariant priors @xmath112 @xmath113 with @xmath114 and @xmath115 . in this section we specify the form of the variational function @xmath11 . to allow for an analytical solution we neglect correlations between the group assignments and the remaining model parameters . we denote by @xmath116 the probability that sample @xmath31 belongs to sample group @xmath78 and by @xmath117 the probability that probe @xmath48 belongs to probe group @xmath118 . furthermore , given that @xmath34 , @xmath40 and @xmath53 always appear in different factors in ( [ gm1 ] ) or ( [ gm2 ] ) then their join distribution factorizes . within the mean - field approximation for the group assignments and the later factorization the variational function can be written as @xmath119 when dealing with the generative model ( [ gm1 ] ) and @xmath120 when dealing with the generative model ( [ gm2 ] ) , where @xmath121 denotes a generic probability density function of @xmath73 . summarizing , in the case studies below , we are going to solve the generative models ( [ gm1 ] ) or ( [ gm2 ] ) , making use of renormalized invariant priors ( table [ priors ] ) and the mf variational function ( [ mf1 ] ) or ( [ mf2 ] ) , respectively . this approach is based on the assumptions that : the population is divided in groups , the group assignments are generated by a multinomial model , the priors are renormalized invariant distributions , and a mf approximation of the variational solution with respect to the group assignments . quite often we deal with datasets consisting of a real value measurement @xmath122 over @xmath123 samples and @xmath47 variables , where the samples and variables are not necessarily independent . for simplicity , the particular kind of dependency we focus on is the existence of sample and variable groups . our problem is to infer the sample and variable groups and the statistical parameters characterizing them . to address this problem we consider the generative model ( [ gm2 ] ) with a normal likelihood , representing a two - sides gaussian mixture model . the two - sides gaussian mixture model is a natural extension of the gaussian mixture model @xcite to characterize datasets with a group structure for both the samples and variables . our contributions in this context are the use of prior distributions derived from symmetry arguments alone and the inclusion of a group structure at the variables side . the dataset , likelihood and priors associated with our statistical model are defined as follows : _ data : _ consider @xmath123 samples , @xmath47 variables , and the real value measurements @xmath122 . _ likelihood : _ we assume that @xmath122 are random variables with a normal distribution , with group dependent mean @xmath124 and group independent variance @xmath66 , resulting in the likelihood @xmath125 here we are assuming that the main difference between groups is given by the means while the variance is group independent . the latter is a good approximation when the source of noise is given by the measurement itself and it behaves the same independently of the sample and variable group . _ priors : _ for the prior @xmath126 we generalize the normal distribution prior in table [ priors ] . accounting for more than one location parameter we obtain @xmath127 and we work in the limit @xmath128 . to apply the variational method we consider the mf approximation ( [ mf2 ] ) . substituting the likelihood ( [ preal ] ) , the priors ( [ ppi ] ) , ( [ pkappa ] ) and ( [ pgm ] ) and the mf variational function ( [ mf2 ] ) into ( [ f ] ) , and integrating over @xmath1 ( summing over @xmath30 and @xmath46 and integrating over @xmath129 , @xmath66 , @xmath28 and @xmath44 ) we obtain @xmath130 \nonumber\\ & -&\sum_k\left(\sum_ip_{ik}+\tilde{\gamma}_k-1\right ) \langle\ln\pi_k\rangle \nonumber\\ & -&\sum_l\left(\sum_jq_{jl}+\tilde{\epsilon}_l-1\right ) \langle\ln\kappa_l\rangle \nonumber\\ & + & \int d\mu d\sigma r(\mu,\sigma)\ln r(\mu,\sigma ) \nonumber\\ & + & \int d\pi r(\pi)\ln r(\pi ) + \int d\kappa r(\kappa)\ln r(\kappa)\kappa \nonumber\\ & + & \sum_{ik } p_{ik}\ln p_{ik } + \sum_{jl } q_{jl}\ln q_{jl}\end{aligned}\ ] ] minimizing ( [ freal ] ) with respect to @xmath131 , @xmath117 , @xmath132 , @xmath133 and @xmath134 we obtain ( vb-1 ) : @xmath135 @xmath136 @xmath137 @xmath138 @xmath139 @xmath140 @xmath141\end{aligned}\ ] ] @xmath142 @xmath143 @xmath144 these are a set of self - consistent equations which can be solved recursively to determine the probabilistic group assignments and the @xmath65 , @xmath66 , @xmath40 and @xmath53 distributions . they are the same in spirit as those for the em algorithm @xcite . following @xcite we refer to them as _ variational bayes _ ( vb ) algorithm . the main difference between the em and vb algorithms is that in the former case we would take the average of the log likelihood over the group assignments but not over the distributions of @xmath65 , @xmath66 , @xmath40 and @xmath53 . by taking the average over @xmath65 and @xmath66 we obtain the additional @xmath145 term within the parenthesis in equations ( [ preal ] ) and ( [ qreal ] ) . according to ( [ alphakl ] ) @xmath146 is equal to @xmath59 plus the product of the average number of samples in sample group @xmath78 ( @xmath147 ) and the average number of variables in variable group @xmath118 ( @xmath148 ) . therefore , the @xmath149 term penalizes assignments to small size groups . and it balances the contribution of @xmath150 , which drives the estimates towards a better fit and consequently groups of minimal size . the actual implementation of the vb-1 algorithm in the context of real value data proceeds as follows . set sufficiently large values for @xmath27 and @xmath43 , larger than our expectation for the actual values of @xmath27 and @xmath43 . in the following test examples we use @xmath151 . set the parameters @xmath59 , @xmath152 , @xmath153 , @xmath154 and @xmath155 . we set @xmath156 , @xmath157 and @xmath158 . the choice of @xmath152 and @xmath153 is practically irrelevant provided we have chosen a sufficiently small @xmath59 . set random initial conditions for @xmath116 and @xmath117 . starting from these random initial conditions iterate equations ( [ preal])-([f_real ] ) until the solution converges up to some predefined accuracy . we use relative error of @xmath159 smaller than @xmath160 . in practice , compute @xmath161 , @xmath162 , @xmath163 , @xmath164 , @xmath165 , @xmath166 , @xmath167 , @xmath116 , @xmath117 and @xmath159 in that order . to explore different potential local minima use different initial conditions and select the solution with lowest @xmath159 . since this algorithm penalizes groups with few members it turns out that , for sufficiently large @xmath27 and @xmath43 , some sample and condition groups result empty . if this is not the case @xmath27 and/or @xmath43 should be increased until at least one sample group and one variable group results empty . between the original @xmath168 and estimated @xmath169 groups assignments , relative to its maximum value @xmath170 when @xmath171 . the original data was made of @xmath172 samples divided in @xmath27 groups and @xmath173 conditions divided in @xmath43 groups . the values of @xmath122 were extracted from a normal distribution with mean @xmath174 and variance @xmath66 . the figure shows the mutual information between the original groups and the group assignment , estimated by the vb-1 algorithm , as a function of the variance @xmath66 . the dashed - dotted , solid and dashed lines corresponds with the worst , average and best case on 100 test examples , respectively . in a ) @xmath175 and in b ) @xmath176 . in both cases the mutual information is approximately equal to its maximum @xmath170 for values of @xmath66 less than one , the minimum difference between the original means @xmath129.,width=307 ] to test the performance of the vb-1 algorithm , ( [ preal])-([f_real ] ) , we consider test examples generated by the likelihood ( [ preal ] ) itself . our aim is to test the variational result in the context of a relatively small number of samples and conditions . to quantify the goodness of the group assignment we consider the mutual information between the original @xmath168 ( @xmath177 ) and estimated @xmath169 sample group assignments , @xmath178 where @xmath179 @xmath180 @xmath181 note that @xmath182 takes its maximum value when @xmath171 , denoted by @xmath183 . off course , the same could be done for the condition group assignments as well . in our test examples the original data was made of @xmath172 samples divided in @xmath27 groups and @xmath173 conditions divided in @xmath43 groups . the values of @xmath122 were extracted from a normal distribution with mean @xmath174 and variance @xmath66 . we estimate the group assignment using the vb-1 algorithm , sampling one initial condition . figure [ fig_real ] shows the mutual information between the original and estimated groups as a function of the variance @xmath66 . in a ) @xmath175 and in b ) @xmath176 . in both cases the mutual information is approximately equal to its maximum @xmath170 for values of @xmath66 less than 1 . since 1 is the minimum difference between the original means @xmath129 , we conclude that the vb-1 algorithm performs well when there is a significant difference between the distributions associated with different groups . for larger values of @xmath66 the vb-1 algorithm performance starts to decrease . this is not , however , a deficiency of the algorithm but an unavoidable consequence of the mixing between the distributions coming from different groups . it is worth noticing that we obtain similar results for the case @xmath184 and @xmath185 , indicating that the method works when there is no group structure on one side , in this case the conditions . there are several datasets consisting of a certain number of properties and the information of whether or not each sample exhibits each of the properties . for example , the dataset in fig . [ fig_hg_bg ] describes a population of three animals characterized by two attributes , hair and legs . the attribute hair can take the value yes ( has hair ) or no ( does not have hair ) while the attribute legs takes the values 2 or 4 ( at least within this dataset ) . the mathematical treatment of this problem is significantly simplified if the variables are mapped onto boolean variables . to each @xmath20 states variable we associate @xmath20 boolean variables , each representing the occurrence or not of a specific letter of the alphabet . for example , the attribute hair is associated with hair - yes and hair - no and the attribute legs with legs-2 and legs-4 ( fig . [ fig_hg_bg]b ) . the outcome of this mapping is represented by the boolean matrix @xmath186 , taking the value 1 if the answer to the boolean variable @xmath48 is yes on sample @xmath31 and 0 otherwise . depending on our aim , the boolean matrix can be represented either by a hypergraph or a bipartite graph . when we aim to cluster the samples without attempting to cluster the boolean variables , @xmath186 is better interpreted as the adjacency matrix of a hypergraph . a hypergraph is an intuitive extension of the concept of graph to allow for connections between more than two elements . in our case , the hypergraph vertices represent samples and hyper - edges , one associated which each boolean variable , represent the set of all samples with the answer yes to the corresponding boolean variable ( fig . [ fig_hg_bg]c ) . on the other hand , when we aim to cluster both the samples and boolean variables then a bipartite graph interpretation is more appropriate , with one class of vertices for the samples and another one for the boolean variables , and an edge connecting sample @xmath31 and variable @xmath48 whenever @xmath187 ( [ fig_hg_bg]d ) . the differences between these two approaches will become clear below . in this case the samples are assumed to be divided in groups while the hypergraph edges are modeled as independent . here we follow the statistical model introduced in @xcite : _ data : _ consider a hypergraph with a vertex set representing @xmath26 samples and @xmath42 edges characterizing the relationships among them . the hypergraph is specified by its adjacency matrix @xmath69 , where @xmath187 if element @xmath31 belongs to edge @xmath48 and it is 0 otherwise . _ likelihood : _ the adjacency matrix elements are generated by a binomial model with sample group and variable dependent probabilities @xmath188 , @xmath29 and @xmath189 , resulting in @xmath190 _ priors : _ as priors we use the renormalized invariant prior of the binomial model ( table [ priors ] ) . taking into account that we have a binomial model for each pair of sample group and edge , we obtain @xmath191 with @xmath192 and @xmath193 . substitute the likelihood ( [ phg ] ) , the priors ( [ ppi ] ) and ( [ p_hg ] ) , and the mf variational function ( [ mf1 ] ) into ( [ f ] ) , and integrating over @xmath1 ( summing over @xmath30 and integrating over @xmath194 and @xmath28 ) we obtain @xmath195 minimizing ( [ f_hg ] ) with respect to @xmath131 , @xmath196 and @xmath133 we obtain ( vb-2 ) @xmath197 } } { \sum_s e^ { \langle\ln\pi_s\rangle + \sum_j \left [ a_{ij}\langle\ln\theta_{sj}\rangle + ( 1-a_{ij})\langle\ln(1-\theta_{sj})\rangle \right ] } } \ ] ] @xmath198 @xmath199 @xmath200 @xmath201 @xmath202 these equations represent the vb algorithm for the statistical model on hypergraphs . in this case we have not been able to disentangle the contributions weighting the fit to the data and the model bias , both being included in the averages @xmath203 and @xmath204 . the implementation of the vb algorithm for the statistical model on hypergraphs proceeds as follows . set sufficiently large values for @xmath27 , larger than our expectation for the actual values of @xmath27 . we use @xmath205 in the following test examples . set the parameters @xmath206 , @xmath207 and @xmath154 . we set the parameters @xmath208 . set random initial conditions for @xmath116 . starting from these initial conditions iterate equations ( [ p_hg])-([f_hg_min ] ) until the solution converges up to some predefined accuracy . we use relative error of @xmath159 smaller than @xmath160 . in practice , compute @xmath209 , @xmath210 , @xmath211 , @xmath204 , @xmath164 , @xmath165 , @xmath116 and @xmath159 in that order . to explore different potential local minima use different initial conditions and select the solution with lowest @xmath159 . since this algorithm penalizes groups with few members it turns out that , for sufficiently large @xmath27 , some sample and condition groups result empty . if this is not the case then increase @xmath27 until at least one group is empty . a matlab code implementing this algorithm can be found at http://www.sns.ias.edu/ vazquez / hgc.html . consider the animal population in fig . [ fig_zoo]a together with their attributes : habitat , nutrition behavior , etc . figure [ fig_zoo]b shows the mapping of this dataset onto a hypergraph . the hypergraph vertices represent animals and the edges represent the association between all animals with a given attribute : edge 1 , all non - airborne animals ; edge 2 , all airborne animals , and so on . the animal population stratification was already addressed in @xcite , finding the solution in fig . [ fig_zoo]c . although the starting statistical model is the same , the solution in @xcite was found assuming fixed the number of groups and estimating the group assignment using the em algorithm ( essentially a maximum likelihood estimate ) . then , in an an attempt to focus in the solution with better consensus , solutions for different number of groups were obtained and the most representative solution was selected . here we address the same problem using a bayesian approach and the variational solution . we start from the same statistical model on hypergraphs but now obtain a solution using the vb-2 algorithm ( [ p_hg])-([f_hg_min ] ) , sampling 10,000 initial conditions as in @xcite . the solution found by the vb-2 algorithm ( fig . [ fig_zoo]d ) is quite similar to that previously found in @xcite ( fig . [ fig_zoo]c ) . the main differences are the splitting of the terrestrial mammals , the exclusion of the platypus and the tortoise from the amphibia - reptiles group and the scorpion from the terrestrial arthropods . more important , in both cases the main groups represent terrestrial mammals , aquatic mammals , birds , fishes , amphibia - reptiles , terrestrial arthropods and aquatic arthropods . the vb-2 ( [ p_hg])-([f_hg ] ) algorithm represents , however , a significant improvement over the approach followed in @xcite . it finds the consensus solution in one run , because it has built in the balance between better fitting and less bias . between the original @xmath168 and estimated @xmath169 groups assignments , relative to its maximum value @xmath170 when @xmath171 . the original data was made of a graph with @xmath172 vertices divided in @xmath106 groups , with an intra- and inter - community connection probabilities @xmath212 and @xmath213 , respectively . the figure shows the mutual information , between the original groups and the group assignment estimated by the vb-2 algorithm ( [ p_hg])-([f_hg_min ] ) , as a function of the inter - community connectivity @xmath213 . the dashed - dotted , solid and dashed lines corresponds with the worst , average and best case on 100 test examples . in a ) we deal with dense communities ( @xmath214 ) and the algorithm performs well ( @xmath215 ) for small values of the inter - community connectivity probability @xmath213 . in b ) we deal with sparse communities ( @xmath216 ) and the algorithm performs well for large values of the inter - community connectivity probability @xmath213.,width=307 ] the work by newman and leicht @xcite provides a hint on how to apply the hypergraph clustering to the problem of finding modules or communities in a graph or network . a graph is made by a set of vertices and a set of edges , the latter being pairs of connected vertices the idea of leicht and newman is a `` guilty by association '' principle : vertices between the same module of a graph will tend to have connections to the same other vertices . this problem can be translated to a hypergraph problem , where the vertices are the graphs vertices , the hyper - edges are the set of nearest neighbors and the boolean variables characterize whether or not a vertex belongs to the a set of nearest neighbors @xcite ( fig . [ fig_hg_bg]e and f ) . more precisely , to each vertex we associate a hyper - edge , given by the set of its nearest neighbors . therefore , there are @xmath217 hyper - edges , one for every vertex in the original graph . the hypergraph adjacency matrix has the matrix element @xmath187 if vertex @xmath31 belongs to hyper - edge @xmath48 , i.e. if vertex @xmath31 belongs to the nearest - neighbor set of vertex @xmath48 , and @xmath218 otherwise . if we label the nearest - neighbor sets with the same label as the vertices then the hypergraph adjacency matrix coincides with the adjacency matrix of the original graph . thus , there is an exact mapping from the statistical model proposed by newman and leicht @xcite to the statistical model on hypergraphs . having specified this mapping we use the vb-2 algorithm ( [ p_hg])-([f_hg_min ] ) , sampling one initial condition , to find the graph modules in the original graph . to illustrate its performance we consider as a case study a graph composed by two communities , with probabilities @xmath212 and @xmath213 that two vertices within the same or different communities are connected , respectively . as already anticipated by newman and leicht @xcite , the nearest - neighbor approach can resolve both dense communities with lesser inter - community connections ( @xmath219 ) and sparse communities with more inter - community connections ( @xmath220 ) . figure [ fig_hg ] shows that the vb-2 algorithm performs quite well in those two regimes . we can face situations where there are groups of boolean variables as well , requiring the clustering of both samples and boolean variables . in this case the bipartite graph representation is more appropriate , with a class of vertices representing the samples and a class of vertices representing the boolean variables . more precisely , _ data : _ consider a bipartite graph with two vertex subsets , representing @xmath26 samples and @xmath42 boolean variables . the graph is specified by its adjacency matrix @xmath69 , where @xmath187 when sample @xmath31 is connected to boolean variable @xmath48 , i.e. if boolean variable @xmath48 is true for sample @xmath31 , and @xmath218 otherwise . _ likelihood : _ the adjacency matrix elements are generated by a binomial model with sample group and variable group dependent probabilities @xmath194 , @xmath29 and @xmath221 , resulting in @xmath222 _ priors : _ for @xmath37 we use the renormalized invariant prior of the binomial model . taking into account that we have one binomial model per each pair of sample and variable group we obtain @xmath223 with @xmath224 and @xmath225 . the likelihood ( [ pbg ] ) is quite similar to ( [ phg ] ) , the main difference being that now the statistical properties of the boolean variables appear through their corresponding group assignments @xmath46 . this increases the model complexity by considering a group structure for the boolean variables and , at the same time , reduces the number of @xmath34 parameters . furthermore , ( [ pbg ] ) contains ( [ phg ] ) as the particular case where @xmath226 and one group associated to each boolean variable . substituting the likelihood ( [ pbg ] ) , the priors ( [ p_bg ] ) , ( [ ppi ] ) and ( [ pkappa ] ) , and the mf variational function ( [ mf2 ] ) in ( [ f ] ) , and integrating over @xmath1 ( summing over @xmath30 and @xmath46 and integrating over @xmath194 , @xmath28 and @xmath44 ) we obtain @xmath227 minimizing ( [ f_bg ] ) with respect to @xmath131 , @xmath117 , @xmath196 , @xmath133 and @xmath134 we obtain ( vb-3 ) @xmath228 } } { \sum_s e^ { \langle\pi_s\rangle + \sum_{jl } q_{jl } \left [ a_{ij}\langle\ln\theta_{sl}\rangle + ( 1-a_{ij})\langle\ln(1-\theta_{sl})\rangle \right ] } } \ ] ] @xmath229 } } { \sum_s e^ { \langle\kappa_s\rangle + \sum_{ik } p_{ik } \left [ a_{ij}\langle\ln\theta_{ks}\rangle + ( 1-a_{ij})\langle\ln(1-\theta_{ks})\rangle \right ] } } \ ] ] @xmath230 @xmath231 @xmath232 @xmath233 @xmath234 @xmath235 equations ( [ p_bg])-([f_bg_min ] ) represent the vb algorithm for the statistical model on bipartite graphs . they can be used to found modules or communities in graphs with a bipartite structure , including those representing samples and boolean variables . the implementation of the vb-2 algorithm ( [ p_bg])-([f_bg_min ] ) for the statistical model on bipartite graphs proceeds as follows . set sufficiently large values for @xmath27 and @xmath43 , larger than our expectation for the actual values of @xmath27 and @xmath43 . set the parameters @xmath236 , @xmath237 , @xmath154 and @xmath155 . we set the parameters @xmath238 . set random initial conditions for @xmath116 and @xmath117 . starting from these initial conditions iterate equations ( [ p_bg])-([f_bg_min ] ) until the solution converges up to some predefined accuracy . we use relative error of @xmath159 smaller than @xmath160 . in practice , compute @xmath209 , @xmath210 , @xmath211 , @xmath204 , @xmath164 , @xmath165 , @xmath166 , @xmath167 , @xmath116 , @xmath117 and @xmath159 in that order . to explore different potential local minima use different initial conditions and select the solution with lowest @xmath159 . since this algorithm penalizes groups with few members it turns out that , for sufficiently large @xmath27 and @xmath43 some sample and/or variable groups result empty . if this is not the case , increase @xmath27 and/or @xmath43 until at least one sample group and one variable group results empty . let us go back to the zoo problem ( fig . [ fig_zoo]a ) . now we represent this dataset by a bipartitite graph , with one class of vertices representing the animals and the other class the boolean variables ( e.g. fig . [ fig_hg_bg]a , b and d ) using the vb-3 algorithm ( [ p_bg])-([f_bg_min ] ) , sampling 10,000 initial conditions as in @xcite , we perform a two - sides clustering of the bipartite graph obtaining the animal population stratification in fig . [ fig_zoo]e and the boolean variables stratification in fig . [ fig_zoo]f . the animal clusters are similar to those previously obtained using the statistical model on hypergraphs ( fig . [ fig_zoo]c and d ) . the main difference is the more refined subdivision of terrestrial mammals , now split in four groups ( 1 , 2 , 3 and 4 ) . in addition to the animal population stratification the two - sides clustering provides association groups between the boolean variables ( fig . [ fig_zoo]f ) . these associations reflect the fact that not all boolean variables are independent , some of them are linked . for example , group 2 cluster four typical attributes of terrestrial mammals , they have hair , do not put eggs , milk and have four legs . in the same way , group 3 clusters attributes of fishes and group four of birds . thus , in general , the bipartite graph model and the resulting two - sides clustering provides more information than the hypergraph approach . between the original @xmath168 and estimated @xmath169 groups assignments , relative to its maximum value @xmath170 when @xmath171 . the original data was made of a graph with @xmath172 vertices divided in @xmath106 groups , with an intra and inter - community connection probabilities @xmath212 and @xmath213 , respectively . the figure shows the mutual information , between the original groups and the group assignment estimate by the vem-3 algorithm ( [ p_bg])-([f_bg_min ] ) , as a function of the inter - community connectivity @xmath213 . the dashed - dotted , solid and dashed lines corresponds with the worst , average and best case on 100 test examples . in a ) we deal with dense communities ( @xmath214 ) and the algorithm performs well ( @xmath215 ) for small values of the inter - community connectivity probability @xmath213 . in b ) we deal with sparse communities ( @xmath216 ) and the algorithm performs well for large values of the inter - community connectivity probability @xmath213.,width=307 ] the bipartite graph model can be use to find network modules as well . in this case one class of vertices represents the original graph vertices and the other represents sets of nearest neighbors ( fig . [ fig_hg_bg]g ) . the two - sides clustering thus attempts to cluster both the original graph vertices and the sets of nearest neighbors . when the original graph is undirected the problem is symmetric ( e.g. see fig . [ fig_hg_bg]g ) . indeed , if vertex @xmath31 belongs to the nearest - neighbor set of vertex @xmath48 then vertex @xmath48 belongs to the nearest - neighbor set of vertex @xmath31 . as a consequence the clustering on the original vertices side can not be differentiated from the clustering of nearest - neighbor sets . intuitively this means that when two vertices belong to the same graph module we can say that their nearest - neighbor sets belong to the same nearest - neighbor set group . having specified this mapping we use the vb-3 algorithm ( [ p_bg])-([f_bg_min ] ) , sampling one initial condition , to find the graph modules in the original graph . to illustrate its performance we consider once again a graph composed by two communities , with probabilities @xmath212 and @xmath213 that two vertices within the same or different communities are connected , respectively . figure [ fig_bg ] shows that the vb-3 algorithm can resolve both dense communities with lesser inter - community connections ( @xmath219 ) and sparse communities with more inter - community connections ( @xmath220 ) . the comparison of fig . [ fig_bg ] and [ fig_hg ] indicates that the bipartite graph model performs slightly better than the hypergraph model . for example , focusing on the average performance , for @xmath214 the vb-3 algorithm performs almost perfectly till @xmath239 , while the vb-2 algorithm does till @xmath240 . this could be , however , specific to the tested set of examples . further research is required to determine which version performs better depending on the dataset under consideration . the bayesian approach allows for a systematic solution of data analysis problems . its starting point is a statistical model of the data under consideration . from there , using bayes rule , we can invert the statistical model to obtain the posterior distribution of the model parameters . the latter can be use , in principle , to calculate or compute averages or other magnitudes of interest . one of the main criticisms to the bayesian approach is the apparent ambiguity in selecting the prior distributions . here we have worked further on jaynes method @xcite , claiming that the prior distributions are given by the most general distribution dictated by the symmetries of the problem under consideration . one undesired consequence of this method is that when the symmetries are not sufficient constraints we obtain improper prior distributions . yet , the use of improper priors can be avoided by working with renormalized distributions that are proper , and approach the improper prior in a certain limit . using this approach we report here a correction to jaynes prior for a likelihood with translation and scale invariance and a generalization of jaynes prior for the binomial model to the multinomial model . having resolve the issue about the prior distributions , we can proceed to the application of the bayesian approach to resolve a population structured . taking inspiration from mixture models @xcite , in particular dirichlet mixture models @xcite , we introduce general statistical models with a built in population structure at the sample , and sample and variable , level . the model with a structure at the sample level aims one - side clustering problems , where the variables are assumed to be independent measurements . the model with a structure at both sample and variable level aims two - side clustering problems , where there are classes of variables . these statistical models are then postulated as generative models of some dataset . introducing a mf approximation as variational function , we then resolve the population structure by solving the inverse problem , i.e. determining the sample and/or variable groups and model parameters from the data . to illustrate the applicability and systematicity of the variational method , here we study the problem of data clustering , in the context of real value and boolean variables . the outcome is a variational bayes ( vb ) algorithm , a self - consistent set of equations to determine the group assignments and the model parameters . the vb algorithm is based on recursive equations similar to those for the em algorithm , but with some intrinsic penalization for model bias . in the case of real value data , and under the assumption of normal distributions , the contributions favoring fitting and penalizing model bias are clearly disentangled . the fitting is quantified , as it is expected for normally distributed variables , by the mean square deviation . the model bias is quantified by the inverse of the square root of the mean cluster sizes . the tendency to reduce the mean square deviation is thus balanced by a tendency to increase the cluster sizes . in the case of boolean variables our analysis is based on a mapping into a hypergraph or bipartite graph . when we cluster the samples but not the boolean variables the problem is mapped onto a statistical model on hypergraphs @xcite . on the other hand , when we perform a two - side clustering , clustering both the samples and the boolean variables , the problem is mapped onto a statistical model on bipartite graphs . the vb algorithms associated with the statistical model on hypergraphs and bipartite graphs can be used to find modules on a graph . starting on an idea by newman and leicht @xcite , we show that the problem of graph modules can be mapped onto the problem of finding hypergraph modules or bipartite graph modules , where the hypergraph edges and the augmented bipartite graph vertices represent nearest - neighbor sets in the original graph . the resulting vb algorithms represent a significant improvement over the maximum likelihood approaches followed in @xcite and @xcite , by including a self - consistent correction for model complexity and bias . it is worth mentioning that , depending on the starting statistical model , we could arrive to different versions of the vb algorithm . indeed , for the finding graph modules problem we could use both the hypergraph and bipartite graph models . furthermore , hofman and wiggins @xcite have obtained another version based on a statistical model with different intra and inter - community connection probabilities . these approaches differ in the definition of what constitutes a group , community or module . we use the definition by newman and leicht @xcite based on topological similarity , i.e. two vertixes are topologically identical if they are connected to the same other vertices in the graph . thus , we obtain group of vertices whose patterns of connectivity are similar . on the other hand , the definition used by hofman and wiggins @xcite is based on the existence of two edge densities , characterizing the tendency of having an edge between intra- and inter - group pairs of vertices . depending on the problem and the question we are asking we may adopt one or the other definition , and use the corresponding clustering method .

inflation @xcite has proven to be a fantastically successful paradigm for explaining why the universe is so big , so old , and so flat , and provides a mechanism for generating the primordial cosmological perturbations . however , a fundamental description of the physics responsible for inflation remains elusive : no convincing model for inflation yet exists . we can , however , discern the broad characteristics of the inflationary epoch from observational data . the absence of a detectable gravitational wave contribution to the cmb ( cosmic microwave background ) anisotropies tells us that the energy density of the universe during inflation was something less than @xmath2 . the fact that the primordial perturbation spectrum is close to scale invariant tells us that the potential function of the field or fields responsible for inflation ( the _ inflaton _ ) must be very flat . finally , the highly suppressed amplitude of the perturbation spectrum ( @xmath3 ) requires the introduction of a small parameter to the inflationary lagrangian , either as a small self - coupling or as a hierarchy of mass scales . probably the best guess as to the fundamental origin of the inflationary potential is that the inflaton ( if there is only one ) is an effective degree of freedom in the low - energy limit of quantum gravity @xcite . the last few years have seen substantial progress toward realizing this hypothesis within string theory @xcite , but no dominant and compelling stringy model of inflation has emerged , and string phenomenology is far from fully understood . in this context , effective field theory is a powerful tool . since the inflaton potential is tightly constrained by data , if the inflaton field takes on a vacuum expectation value ( vev ) at which higher order operators contribute significantly to the potential , it is very unlikely that the delicate balance between the height and slope of the potential will survive intact . therefore effective field theory arguments favor a vev for the inflaton which is small in planck units : @xmath4 . lyth @xcite derived a lower bound on the variation in the inflaton field during inflation in terms of the ratio @xmath5 between tensor and scalar perturbations generated during inflation , known as the _ lyth bound_. combined with a theoretical prejudice based on effective field theory , the lyth bound predicts a strong suppression of primordial gravitational waves relative to the observed amplitude of the scalar perturbations . the significance of the lyth bound has been the subject of much debate . in particular , linde has argued that only the _ energy density _ during inflation must be sub - planckian and that field values greater than @xmath1 are physically consistent @xcite , but this argument does not apply to generic supergravity or superstring inspired models @xcite . lyth s argument is restricted to single field models , and multifield models can evade it in two ways . firstly , one might imagine that several separate inflationary epochs are concatenated together . secondly , the large number of scalar fields introduced by the string landscape makes it natural to look for implementations of _ assisted inflation _ @xcite where a large number of similar ( if not identical ) scalar fields act cooperatively to drive inflation . the resulting cosmology mimics that produced by a single field with a large vev , but the individual fields never take on trans - planckian expectation values . a concrete example of this approach is provided by `` n - flation '' @xcite . questions surrounding the robustness and generality of the lyth bound will only grow in importance as measurements of the perturbation spectrum put ever tighter constraints on the amplitude of primordial tensor fluctuations @xcite . in this paper , we do not attempt to resolve the debate about the self - consistency of effective field theory in inflation , but simply pose the question : if we suppose that future observations show a strongly suppressed amplitude of primordial tensor fluctuations , what will that tell us about the inflaton potential ? to investigate this question , we derive an extended version of the lyth bound accurate to higher order in slow roll . this allows us to connect the field variation @xmath0 to the spectral index of scalar perturbations , as well as to the amplitude of tensor modes . we then investigate the implications of this bound for `` small field '' potentials , where the field rolls off a local maximum of the potential . we find that , in the absence of fine - tuning , the field variation during the inflationary period is _ generically _ of order @xmath1 , even for models with a very small tensor / scalar ratio . specifically , near the end of inflation , the field variation is rapid enough to ensure that @xmath6 during the last e - fold of expansion . this is true even when the field variation is small during the epoch in which the observable perturbations are generated . we show that this `` last e - fold problem '' can only be solved by suppressing the inflaton mass term either by fine - tuning or the imposition of a symmetry . the remainder of the paper is structured as follows : section [ sec : extendedlythbound ] derives the higher - order extension of the lyth bound . section [ sec : smallfieldmodels ] discusses the issue of field variation in generic small - field models and the `` last e - fold '' problem . section [ sec : conclusions ] contains a summary and conclusions . inflationary cosmology predicts the existence of a spectrum of large - scale tensor fluctuations ( gravitational waves ) in addition to scalar fluctuations ( adiabatic density perturbations ) . both are expected to contribute to the measured temperature anisotropy and polarization of the cosmic microwave background ( cmb ) , and different inflationary models predict varying contributions . the relative contributions are quantified in terms of the tensor - to - scalar ratio @xmath7 where @xmath8 are the amplitudes of tensor and scalar perturbations , respectively , and @xmath9 is the first hubble slow - roll parameter @xmath10 there is much interest in the possible detection of a tensor contribution to the cmb temperature anisotropy . upcoming cmb missions will have vastly increased sensitivity to the b - mode polarization signal characteristic of gravitational radiation . the planck satellite @xcite will have the sensitivity to detect @xmath11 and successors such as nasa s inflation probe @xcite may achieve a sensitivity of @xmath12 or better @xcite . direct detection of the primordial gravitational wave background with the envisioned big bang observer mission @xcite has a projected sensitivity of @xmath13 @xcite . together with the cobe normalization , @xmath14 , a detection of tensor fluctuations will ( assuming a nearly scale - invariant spectrum ) fix the energy scale of inflation via the relation @xmath15 where @xmath16 is the height of the inflationary potential when scales of cosmological interest exited the horizon . furthermore , since @xmath17 and @xmath18 to lowest order in slow roll , a measurement of tensor modes will also reveal information about @xmath19 when slow roll is valid . detection of the b - mode will therefore constrain the shape of the inflaton potential enough to eliminate many currently viable inflationary models . the discovery of primordial gravity waves is therefore a valuable and exciting prospect . however , there are theoretical arguments which suggest that the amplitude of the tensor component is expected to be well below current observational limits . lyth @xcite showed that the amplitude of the tensor component ( and therefore @xmath5 ) depends on the total inflaton field variation , @xmath20 , during the time that observable scales exit the horizon . from the expression @xmath21 and by using eq . ( [ r ] ) , it is possible to write the field variation as a function of @xmath5 , @xmath22 where @xmath23 is the number of e - folds before the end of inflation . it is assumed that there is negligible variation in @xmath5 over the period @xmath24 , which is a good approximation to lowest order in slow roll . in his original paper , lyth considered @xmath20 as scales corresponding to @xmath25 were exiting the horizon , during which time the universe expands by an amount @xmath26 . since this gives a lower limit on the total field variation , the result is an inequality known as the _ lyth bound _ : @xmath27 this indicates that for sizeable values of @xmath5 , the inflaton field must roll a distance of order @xmath1 while observable scales are exiting the horizon . field variations @xmath28 are , in fact , ubiquitous in inflationary model - building , and essential for the success of chaotic inflation and certain models based on symmetry breaking potentials . such a large field variation may , however , be problematic from the standpoint of effective field theory if the underlying inflationary model is embedded in either supergravity or string theory . it is widely believed that inflation can be appropriately described by the evolution of a fundamental scalar field : a field existing in the low energy limit of some more fundamental theory such as supergravity or string theory . it is possible to describe these fundamental theories at low energies by an effective field theory , obtained by integrating out high momentum degrees of freedom and rescaling fields and coupling constants . an effective field theory description of inflation thus involves an effective potential for the inflaton of the form @xmath29 the terms of order @xmath30 have couplings of negative mass dimension , and are therefore nonrenormalizable . however , provided @xmath31 , the nonrenormalizable terms remain under control and the series is at least asymptotically convergent . therefore , if we hope to be able to make observable predictions based on an effective field theory , we need @xmath32 as the scales relevant to these predictions are leaving the horizon . this is a conservative estimate : while we work with the full planck mass , the appropriate vertex factor for a loop expansion in gravity is actually the reduced planck mass , @xmath33 , about a factor of five smaller . ensuring that this relationship holds can easily lower the amplitude of the tensor modes to the point where they are effectively undetectable this strong theoretical prior on the inflationary model space is a stringent constraint on a key inflationary observable and we now ask : can the lyth bound can be extended to constrain other observables ? the power spectrum of superhorizon - sized perturbations ( @xmath34 ) is conventionally expressed in terms of the comoving wavenumber , @xmath35 , @xmath36 where @xmath37 is the spectral index . when slow roll is valid , the spectral index can be written @xmath38 where @xmath39 is the second hubble slow - roll parameter , @xmath40 we wish to obtain @xmath20 as a function of @xmath39 , so we will extend the lyth bound to next order in slow roll . consider the taylor expansion of the inflaton field @xmath41 , @xmath42 where the coefficients in the series are given by the flow equations , @xmath43 eq . ( [ exp ] ) then becomes @xmath44.\ ] ] the first term in eq . ( [ newbound ] ) is lyth s original result and the additional term provides an @xmath39-dependent correction . using eqs . ( [ r],[n ] ) and taking @xmath45 we obtain the extended lyth bound decreases as one goes forward in time . in this paper we adopt the convention @xmath46 so that @xmath47 . ] @xmath48.\ ] ] the effect of the extra term in eq . ( [ newbound ] ) depends on the sign of @xmath39 . we are particularly interested in models for which @xmath49 , as they often lead to a detectable value of @xmath5 @xcite . these consist of potentials that arise from spontaneous symmetry breaking , pseudo nambu - goldstone bosons ( natural inflation ) @xcite , and certain chaotic inflation scenarios . for @xmath50 , the bound tightens as @xmath51 becomes large . during slow roll , @xmath39 can be given in terms of the inflaton potential @xmath52 : @xmath53.\ ] ] for a given value of @xmath9 and vacuum energy density @xmath54 , @xmath39 grows more negative as the curvature of the potential increases over the region defined by @xmath55 . the field @xmath56 rolls more quickly as the concavity of the potential increases , and thus travels further as the universe expands by @xmath24 e - folds . for a given @xmath5 , the extended lyth bound eq . ( [ extended ] ) is tightened slightly for models with red spectra ( @xmath57 ) and loosened for those with blue spectra ( @xmath58 ) . equation ( [ extended ] ) determines the field variation @xmath20 as scales with @xmath25 leave the horizon . lyth @xcite pointed out that the lower order version of this result , eq . ( [ lythbound ] ) , is necessarily a lower limit on the total field variation . it is conservative , as @xmath9 can increase significantly in the remaining @xmath23 e - folds of inflation , resulting in @xmath59 . , they obtain the bound @xmath60 where @xmath20 is the variation over the remaining 55 e - folds of inflation . ] for example , if @xmath61 , @xmath62 over the last @xmath63 e - folds of inflation . these models predict @xmath64 , so @xmath65 over the first @xmath66 e - folds of inflation . we thus see that while this interval corresponds to around 7% of the total expansion of the universe , it provides only @xmath67 of the field variation . if inflation is driven by a single scalar field whose potential has the form of eq . ( [ veff ] ) , we need to ensure that @xmath20 does not exceed @xmath1 over the full course of inflation , and not just over the interval during which currently observable scales are leaving the horizon . for a given @xmath23 we can write down the equivalent field value , @xmath68 . for a specific potential , @xmath68 is a function of @xmath69 , the field value at the end of inflation which is determined from the slope of the potential . as we now show , field variations during the _ last _ e - fold of inflation are generically of order @xmath1 , making this constraint difficult to satisfy . we begin with an example . consider the potential @xmath70 where @xmath71 is the `` height '' of the potential and @xmath72 is the `` width '' . such potentials are characteristic of chaotic inflation scenarios @xcite and typically result in @xmath73 during inflation . to lowest order in slow roll , @xmath74 which for the case of eq . ( [ chaotic ] ) becomes @xmath75 from eq . ( [ phin ] ) , we obtain @xmath56 as a function @xmath23 , @xmath76 we find @xmath69 by setting @xmath77 in eq . ( [ chaoticeps ] ) written to lowest order in slow roll eq . ( [ epssr ] ) to obtain a field value when slow roll has clearly broken down . use of the exact expression for @xmath9 eq . ( [ eps ] ) actually yields a _ larger _ field variation than that obtained in eq . ( [ 27 ] ) , but agreement between the two curves is within 5% . ] @xmath78 we have now fully specified @xmath68 , and if @xmath79 the field variation over the _ last _ e - fold is @xmath80 this is much larger than the field variation over the period when cosmologically relevant perturbations are generated , but yields only a single e - fold of inflation . moreover , the field variation grows with @xmath81 . now consider @xmath82 giving @xmath83 for @xmath79 . we can understand the general case by noting that in eq . ( [ phin ] ) , if @xmath9 does not vary considerably as the universe expands by @xmath24 e - folds , we can write @xmath84 since inflation ends when @xmath85 , during the final e - fold @xmath86 , resulting in a field variation @xmath87 . as @xmath20 is the integral of @xmath88 over the last e - fold , the area under the curve in the figure is proportional to @xmath20 . when @xmath89 , @xmath9 is close to unity when @xmath90 , and is exactly unity ( by definition ) when @xmath91 . ( [ phin ] ) is a general relationship between @xmath56 and @xmath23 , so the conclusion that a slowly varying @xmath92 leads to @xmath20 of order @xmath1 over the last e - fold is actually model independent . consequently , there is a large class of inflationary potentials with this property , even though the explicit calculation here assumes @xmath93 . in order to keep the field variation small and solve this `` last e - fold '' problem , we need a potential for which @xmath9 rises very rapidly to unity as @xmath23 approaches zero . due to the vast number of inflation models consistent with current constraints @xcite , rather than investigating @xmath20 for the last e - fold of inflation on a case - by - case basis , we turn to _ monte carlo reconstruction _ @xcite , which relies on stochastic evaluation of the inflationary flow equations @xcite . by selecting initial conditions in the inflationary parameter space and integrating the flow equations , we can generate a vast collection of models , and then filter out those which satisfy any specific constraints . monte carlo reconstruction relies on a truncated slow - roll hierarchy , so it effectively assumes that the potential contains no sharp kinks or corners , since even very high order derivatives of @xmath94 and @xmath52 are non - zero in the vicinity of such a feature . however , we are excluding this class of potentials by hypothesis when we assume that the end of inflation is due to a smooth transition to @xmath95 from @xmath96 , rather than an almost instantaneous jump , so our analysis here is entirely self - consistent . over the last e - fold for 12,000 inflation models satisfying the constraint @xmath97 on the spectral index . red ( @xmath98 ) and blue ( @xmath99 ) spectra are plotted separately.,width=336 ] having solved the flow equations , one may then recover the inflaton potential for the models that survive the filtering process @xcite . the results of such an analysis are presented in figure [ figure2 ] . the histogram displays @xmath100 during the last e - fold of inflation for 12,000 models in a sixth - order flow reconstruction with spectral index within @xmath101 at astrophysical scales . the models at the left hand end of the histogram are those for which we might expect that higher order corrections to the inflaton potential are ignorable throughout the last 60 e - folds of inflation , and we see that they are are extremely rare . this result strongly suggests that @xmath102 over the last e - fold for a generic inflationary model . all of the models included in figure [ figure2 ] are `` nontrivial '' models , for which inflation ends dynamically by @xmath9 passing through unity . these have potentials of either the _ large field _ or _ small field _ type . large field models are characterized by a scalar field initially displaced from the minimum of the potential by an amount typically of order the planck mass . chaotic inflation scenarios @xcite arising from polynomial potentials , as well those arising from exponential potentials @xcite , are examples of large field models . since large field models generally require an initial field value @xmath103 , these are already problematic from an effective field theory perspective . in contrast , small field models are distinguished by a scalar field initially very close to a maximum of the potential , and roll off the `` hill - top '' as inflation progresses . ( we use the term `` small field '' in the original sense of ref . @xcite : the _ initial _ field value must be small , but the field variation @xmath0 need not be . ) examples include coleman - weinberg ( `` new '' inflation ) and pseudo nambu - goldstone bosons ( natural inflation ) potentials . we are interested in identifying which small field models lead to inflation with @xmath104 . as we will find , the behavior of @xmath92 during the last e - fold will be particularly important . in fig . [ figure2 ] we see models with both red ( @xmath98)and blue ( @xmath99 ) spectra . naively , a blue spectrum is associated with hybrid inflation , where inflation ends via an instability in a direction orthogonal to the inflaton trajectory . however , the `` blue '' models here are correlated with a significant running , and the spectral index is thus a function of scale . this blue to red transition is facilitated by including higher order terms in the slow roll expansion . this breakdown of the standard model - classification at higher order is discussed in detail in ref . for the analytic models we consider below , such higher - order effects are negligible , and expressions to first - order in slow roll are sufficient to describe the physics . scalar fields with spontaneous symmetry breaking potentials have long been seen as candidate inflationary models . such potentials possess a point of unstable equilibrium with non - vanishing vacuum energy ( the false vacuum ) , and a point of stable equilibrium with effectively zero vacuum energy ( the true vacuum ) . if @xmath56 is initially in ( or near ) the false vacuum , thermal and/or quantum fluctuations eventually cause the field to roll towards the true vacuum . these models are known as small field models , since @xmath105 initially . during inflation , the field is near the top of the potential , @xmath106 , so we can assume @xmath107.\ ] ] since inflation ends and slow roll breaks down before the field reaches the true vacuum @xmath108 , a realistic potential with @xmath4 during inflation must have @xmath109 , while satisfying observational constraints on the spectrum for scales exiting the horizon around @xmath110 . from the lyth bound , it is clear that the tensor - to - scalar ratio is negligible at these scales . therefore , a spectral index consistent with a nearly scale - invariant spectrum will be the key inflationary observable . models of inflation occurring at low scales have been investigated in refs . @xcite , among others . we will focus on the potential @xmath111,\ ] ] for @xmath112 and @xmath113 , and we will see that these two choices lead to significantly different inflationary phenomenology . working to first order in the slow roll parameters is sufficient for our purpose here , and we drop higher - order corrections . we start with @xmath112 , the so - called `` inverted quadratic '' potential . this is the generic case , since in the absence of a symmetry that suppresses the mass , the quadratic term will _ always _ dominate for small enough @xmath56 . many of the following results are derived in more detail in refs . @xcite . a similar , more recent analysis can be found in ref . the first slow - roll parameter , @xmath114 , is @xmath115 from eq . ( [ phin ] ) , the number of e - folds before the end of inflation can be found as a function of @xmath56 , @xmath116 next , we determine the field value at @xmath117 , @xmath118 , as this will be needed to calculate observables : @xmath119.\ ] ] using eq . ( [ n ] ) , we obtain the scalar spectral index at @xmath117 , @xmath120 where the approximation follows because the field is slowly rolling while scales corresponding to @xmath118 are leaving the horizon . we see that with @xmath121 , the spectral index @xmath122 , resulting in an extremely red spectrum . requiring @xmath123 yields a scale invariant spectrum , @xmath125 , is achieved in the limit @xmath126 . under what circumstances is it possible to satisfy the condition @xmath104 with an inverted quadratic potential ? in the limit @xmath106 the second slow - roll parameter becomes a constant , @xmath127 again , we find @xmath69 , the field value at the end of inflation , by solving @xmath128 @xmath129 = e^{-2 \left\vert\eta\right\vert n}.\ ] ] thus we can not naively take the small-@xmath39 limit with @xmath5 fixed . the rate of change of the field can be obtained from eq . ( [ phin ] ) , @xmath130 the change in the value of @xmath56 between @xmath131 and @xmath132 is @xmath133 taking @xmath123 , so that @xmath134 , the field variation from @xmath110 to @xmath135 is then @xmath136 which satisfies @xmath137 . however , the field variation during the last e - fold is large : @xmath138 as in our earlier flow analysis , virtually all of the field evolution happens near the _ end _ of inflation . this is not surprising , since the end of inflation is precisely where slow roll is breaking down . therefore , while a small value of @xmath5 is a necessary condition for @xmath0 to be small in planck units throughout inflation , it is not a _ sufficient _ condition . for any potential dominated by the inflaton mass term during inflation , the total field variation is always of order @xmath1 , even if the field variation is small early in inflation . potentials for which the mass term is suppressed , either through a fine - tuning or through the imposition of a symmetry @xcite , can avoid the `` last e - fold '' problem . consider eq . ( [ hill ] ) for @xmath113 . proceeding as in the @xmath112 case , we find @xmath139 the number of e - folds before the end of inflation is @xmath140,\ ] ] and from @xmath141 , @xmath118 is @xmath142^{1/(p-2)}.\ ] ] this approximation follows from eq . ( [ np ] ) since the first term in square brackets dominates when @xmath106 . from eq . ( [ n ] ) , the spectral index evaluated at @xmath118 is @xmath143 and depends only on the order of the dominant term in the potential eq . ( [ hill ] ) . this is in sharp contrast to the result for @xmath112 eq . ( [ 2n ] ) , where the spectral index is a function of @xmath72 . the origin of the difference in the form of @xmath37 in eqs . ( [ 2n ] ) and ( [ 40 ] ) is that @xmath144 is independent of @xmath56 when @xmath112 . since @xmath145 , as @xmath72 decreases , @xmath37 is driven smaller regardless of where on the potential it is measured ( _ i.e. _ the value of @xmath118 ) . potentials with @xmath113 have @xmath146 which , for a given @xmath56 , also becomes more negative as @xmath72 decreases . however , the point at which we measure the spectral index , @xmath118 , also scales with @xmath72 as in eq . ( [ 39 ] ) . the scaling behavior of @xmath147 and that of @xmath148 exactly cancel , so that as we go to smaller scales , the spectral index is being measured at point on the potential for which @xmath147 does not differ from one scale to the next . it is important to determine what additional constraints , if any , are imposed on the scale @xmath72 by the cobe normalization , @xmath14 , which is related to the amplitude of scalar perturbations eq . ( [ r ] ) by @xmath149 . at lowest order , the perturbation amplitude at @xmath118 is given in terms of the potential by @xmath150 the cobe normalization relates the vacuum energy density @xmath151 and the symmetry breaking scale @xmath72 at @xmath118 , @xmath152 when @xmath79 , the hierarchy @xmath153 drops out and @xmath54 scales linearly with @xmath72 . this is expected , since the @xmath154 field theory is scale - free . if one assumes perfectly efficient reheating , it is possible to derive a lower bound on @xmath151 . in order to preserve standard hot big bang cosmology , inflation must occur above the nucleosynthesis scale , or @xmath155 . from eq . ( [ 35 ] ) , this places a lower bound of @xmath156 on the symmetry breaking scale . therefore , the @xmath79 potential is capable of driving inflation at very low scales while providing an acceptable reheat temperature . in general , @xmath157 so that for a given scale @xmath158 , higher order potentials will admit larger @xmath54 . lastly , it remains to determine the behavior of @xmath92 during the last e - fold . using eqs . ( [ 38 ] ) and ( [ np ] ) , we find @xmath9 as a function of @xmath23 , @xmath159^{-\frac{2p-2}{p-2}}.\end{aligned}\ ] ] for an illustrative comparison , consider the plot of @xmath88 vs. @xmath23 for potentials of the form @xmath160 $ ] , fig . [ figure3 ] . vs. @xmath23 over the last e - fold for the potentials @xmath161 ( dotted line ) and @xmath162 $ ] ( solid line ) . the symmetry breaking scale for the inverted potential is @xmath163.,width=336 ] the symmetry breaking scale for the inverted potential is @xmath163 , a fairly large value considering that these potentials can lead to successful inflation at much lower scales . the rapid falloff of @xmath88 as a function of @xmath23 is exactly the behavior expected out of a potential which satisfies @xmath4 . we consider the consequences of requiring the change in the inflaton field during its evolution to be sub - planckian , or , @xmath164 . we derive a higher order version of lyth s relationship between the field variation and the tensor / scalar ratio , @xmath5 . this is valid to second order in slow roll , and is given by eq . ( [ extended ] ) . in addition to @xmath5 , the field variation also depends on the scalar spectral index @xmath37 . naively interpreted , this result appears to indicate that a spectrum close to scale invariance ( @xmath165 ) makes it easier to ensure that @xmath0 is small , but we find that this is a model - dependent statement . arguments based on effective field theory suggest that a self - consistent inflation model must have @xmath164 during the interval in which cosmological perturbations are laid down . to satisfy this constraint with a single field model , the tensor / scalar ratio is strongly suppressed : @xmath166 . this argument amounts to putting a strong theoretical prior on the inflationary model space namely that we are working with a model derived from supergravity or string theory . in this case , an effective field theory description of the inflaton potential is very unlikely to be reliable when the inflaton acquires a trans - planckian vev , as the delicate balance between the height and slope of the potential required for successful inflation will almost certainly be disrupted . in this analysis , we do not take a position on whether it is reasonable to restrict the inflationary model space in this way . rather , we take the view that observations will eventually either determine the value of @xmath5 or place a tight upper bound on its value , and we investigate the theoretical implications of these different possibilities . lyth s original discussion considered only the field variation during the interval during which perturbations relevant to the lowest multipole moments cmb were being laid down , relevant to the detectable signal from gravitational waves . we extend this result in two ways . firstly , we include next - to - leading order terms in the slow - roll expansion , which yields an expression for @xmath20 which depends on both @xmath5 , and the scalar spectral index @xmath37 . secondly , we consider the field variation over the final 60 e - foldings of inflation which gives a tighter constraint on the overall model space . in particular , we found that the field point generically evolves by an amount close to @xmath1 during the final moments of inflation , leading to what we have dubbed the `` last e - fold problem '' : almost all _ minimally coupled , single field _ inflationary models will have @xmath167 . the simplest way to build an inflationary model with @xmath137 is hybrid inflation @xcite . in a hybrid scenario , the field evolves toward a minimum with nonzero vacuum energy . on such a potential , the inflationary solution asymptotes to @xmath168 as @xmath56 approaches the minimum of the potential , after an _ infinite _ amount of inflation @xcite . consequently , in these models inflation ends because of an instability in an orthogonal direction , induced by an auxiliary field . clearly , this scenario yields an arbitrarily large amount of inflation with an arbitrarily small field variation . such models are therefore well - behaved from the standpoint of effective field theory . it might be argued that hybrid inflation is fine - tuned , as it relies on having an auxiliary field with an instability exponentially close to the origin . in practice , hybrid models exploit the presence of a hierarchy of mass - scales in nature , and so do not involve any _ extra _ tunings . in the case of `` inverted '' potentials characteristic of spontaneous symmetry breaking , inflation ends dynamically when the slow roll parameter @xmath9 passes through unity , and no extra parameters are required to fix the endpoint of inflation . near the origin , the potential is dominated by the lowest order term in its taylor series , so we consider `` small - field '' models where @xmath169 . without tuning or a symmetry , the inflaton mass term dominates , and @xmath170 . in this case , we derive an expression for the field variation over the entire inflationary period , eq ( [ phirun ] ) , or @xmath171 . therefore as we approach the scale - invariant limit , @xmath172 , the field variation becomes large in planck units . interestingly , most of the field evolution happens near the _ end _ of inflation , long after observable perturbations have been generated . the field variation can be small during the first four e - folds of inflation , @xmath136 thereby nominally satisfying @xmath137 . the field variation is , however , _ large _ during the last e - fold , @xmath138 therefore , such models still suffer from the `` last e - fold '' problem , and generically yield large field variations near the end of inflation . this behavior can be seen in the approximate flow relation @xmath173 valid when @xmath174 near the end of inflation , @xmath175 , and the field variation is generically of order @xmath1 , even for models which generate a negligible tensor amplitude on observable scales . this analytic result is consistent with conclusions based on a stochastic analysis using the flow equations , discussed here and in ref . whether or not a field variation @xmath176 is actually a problem is a model - dependent statement @xcite . one of us ( whk ) has argued that field variations of this magnitude are perfectly acceptable in the context of natural inflation @xcite . other models might incorporate such field variations in a natural way ( see , for example , ref . @xcite ) . in order for the field variation during inflation to be very small in planck units , the mass term for the inflaton must be suppressed . this might be accomplished by fine - tuning : for example a very flat potential with a sharp `` drop off '' , so the potential resembles a mesa rather than a rounded hilltop . in this case , a taylor expansion of the potential near the end of inflation is dominated by higher - order terms . ( a similar tuning problem was identified in ref . @xcite using a counting argument on a generic dimensionless paremeter set . ) a less contrived alternative is to introduce a symmetry which forces the mass term to vanish , so that the potential is dominated by quartic ( or higher - order ) interactions during the inflationary period . in this case , @xmath9 varies rapidly near the end of inflation , and @xmath137 over the entire inflationary period , without the need for further tuning . in ref . @xcite whk and mahanthappa constructed a detailed model based on an explicitly broken local symmetry , demonstrating that one can achieve the desired suppression of the inflaton mass term within a realistic lagrangian . we argue that this model satisfies the `` challenge '' issued by boyle , steinhardt , and turok in ref . @xcite to construct a deterministic , complete inflationary model which is consistent with @xmath177 . the model is technically natural : no fine - tuning is required even when loop corrections are included , and the potential @xmath178 - \ln\left(g^2\right)\right\rbrace \quad\end{aligned}\ ] ] has a smooth minimum , bounded below . the inflaton can be coupled to fermionic degrees of freedom , allowing reheating , without spoiling the form of the inflationary potential . a discussion of how to achieve a similar suppression of the mass term in extra - dimensional theories can be found in ref . we therefore conclude that tuning arguments _ can not _ be used as suggested by boyle , _ et al . _ to place a lower - bound on the tensor / scalar ratio for inflation models . conversely , there is no watertight argument that can be used to rule out inflationary models with @xmath73 , and the value of @xmath5 will ultimately be determined observationally . however , the analysis here shows that in typical models of inflation , the variation in the field value during the last e - fold of inflation can be significantly larger than the change during the epoch in which cosmological perturbations are laid down . consequently , if one is considering a class of models for which @xmath179 is disfavored , the analysis here can be used to put tight constraints on the models parameter space . observationally , we are making rapid strides . forthcoming observations will be capable of detecting a tiny deviation from scale - invariance . in the absence of running , a blue spectrum , ( @xmath99 ) will strongly favor hybrid - type inflation models . a red spectrum ( @xmath57 ) favors `` inverted '' potentials typical of small - field inflation . current upper bounds on the tensor amplitude are much weaker than the ultimate theoretical limits on their detectability . if tensor modes are detected by future measurements , we will acquire a wealth of phenomenological information about the inflationary epoch @xcite . if the tensor / scalar ratio turns out to be unobservably small , we can still learn much about inflation . however , despite recent progress , there is no guarantee that the theoretical questions surrounding the `` naturalness '' of inflation will be quickly resolved . re and whk thank the aspen center for physics , where this project was initiated , for its hospitality and for providing a stimulating environment . we thank eugene lim , david lyth and liam mcallister for comments on a draft version of this paper , and daniel chung for helpful conversations . re is supported in part by the united states department of energy , under contract doe - fc02 - 92er-40704 . whk is supported in part by the national science foundation under grant nsf - phy-0456777 . a. h. guth , phys . d * 23 * , 347 ( 1981 ) . d. h. lyth and a. riotto , phys . * 314 * , 1 ( 1999 ) [ arxiv : hep - ph/9807278 ] . g. r. dvali and s. h. h. tye , phys . b * 450 * , 72 ( 1999 ) [ arxiv : hep - ph/9812483 ] . s. kachru , r. kallosh , a. linde , j. maldacena , l. mcallister and s. p. trivedi , jcap * 0310 * , 013 ( 2003 ) [ arxiv : hep - th/0308055 ] . e. silverstein and d. tong , phys . d * 70 * , 103505 ( 2004 ) [ arxiv : hep - th/0310221 ] . c. p. burgess , j. m. cline , h. stoica and f. quevedo , jhep * 0409 * , 033 ( 2004 ) [ arxiv : hep - th/0403119 ] . m. alishahiha , e. silverstein and d. tong , phys . d * 70 * ( 2004 ) 123505 [ arxiv : hep - th/0404084 ] . c. p. burgess , r. easther , a. mazumdar , d. f. mota and t. multamaki , jhep * 0505 * , 067 ( 2005 ) [ arxiv : hep - th/0501125 ] . s. dimopoulos , s. kachru , j. mcgreevy and j. g. wacker , arxiv : hep - th/0507205 . a. aazami and r. easther , arxiv : hep - th/0512050 . r. easther and l. mcallister , arxiv : hep - th/0512102 . d. h. lyth , phys . * 78 * , 1861 ( 1997 ) [ arxiv : hep - ph/9606387 ] . a. linde , phys . scripta * t117 * , 40 ( 2005 ) [ arxiv : hep - th/0402051 ] . r. easther , arxiv : hep - th/0407042 . a. r. liddle , a. mazumdar and f. e. schunck , phys . d * 58 * , 061301 ( 1998 ) [ arxiv : astro - ph/9804177 ] . w. h. kinney , arxiv : astro - ph/0307005 . l. alabidi and d. h. lyth , arxiv : astro - ph/0510441 . see http://www.rssd.esa.int/index.php?project=planck . w. h. kinney , phys . d * 58 * , 123506 ( 1998 ) [ arxiv : astro - ph/9806259 ] . l. verde , h. peiris and r. jimenez , arxiv : astro - ph/0506036 . s. phinney , et al . , `` the big bang observer : direct detection of gravitational waves from the birth of the universe to the present , '' nasa mission concept study ( 2005 ) . c. ungarelli , p. corasaniti , r. a. mercer and a. vecchio , class . * 22 * , s955 ( 2005 ) [ arxiv : astro - ph/0504294 ] . w. h. kinney and a. riotto , arxiv : astro - ph/0511127 . l. boubekeur and d. h. lyth , jcap * 0507 * , 010 ( 2005 ) [ arxiv : hep - ph/0502047 ] . s. dodelson , w. h. kinney and e. w. kolb , phys . d * 56 * , 3207 ( 1997 ) [ arxiv : astro - ph/9702166 ] . k. freese , j. a. frieman and a. v. olinto , phys . lett . * 65 * , 3233 ( 1990 ) . f. c. adams , j. r. bond , k. freese , j. a. frieman and a. v. olinto , phys . d * 47 * , 426 ( 1993 ) [ arxiv : hep - ph/9207245 ] . n. arkani - hamed , h. c. cheng , p. creminelli and l. randall , phys . * 90 * , 221302 ( 2003 ) [ arxiv : hep - th/0301218 ] . n. arkani - hamed , h. c. cheng , p. creminelli and l. randall , jcap * 0307 * ( 2003 ) 003 [ arxiv : hep - th/0302034 ] . j. j. blanco - pillado _ et al . _ , jhep * 0411 * ( 2004 ) 063 [ arxiv : hep - th/0406230 ] . p. brax and j. martin , phys . d * 72 * , 023518 ( 2005 ) [ arxiv : hep - th/0504168 ] . k. freese and w. h. kinney , phys . d * 70 * , 083512 ( 2004 ) [ arxiv : hep - ph/0404012 ] . g. efstathiou and k. j. mack , jcap * 0505 * , 008 ( 2005 ) [ arxiv : astro - ph/0503360 ] . a. d. linde , phys . b * 129 * , 177 ( 1983 ) . h. v. peiris _ et al . _ , astrophys . j. suppl . * 148 * , 213 ( 2003 ) [ arxiv : astro - ph/0302225 ] . v. barger , h. s. lee and d. marfatia , phys . b * 565 * , 33 ( 2003 ) [ arxiv : hep - ph/0302150 ] . s. hannestad , s. h. hansen , f. l. villante and a. j. s. hamilton , astropart . * 17 * , 375 ( 2002 ) [ arxiv : astro - ph/0103047 ] . w. h. kinney , e. w. kolb , a. melchiorri and a. riotto , phys . d * 69 * , 103516 ( 2004 ) [ arxiv : hep - ph/0305130 ] . u. seljak _ et al . _ , phys . d * 71 * , 103515 ( 2005 ) [ arxiv : astro - ph/0407372 ] . r. easther and w. h. kinney , phys . d * 67 * , 043511 ( 2003 ) [ arxiv : astro - ph/0210345 ] . m. b. hoffman and m. s. turner , phys . d * 64 * , 023506 ( 2001 ) [ arxiv : astro - ph/0006321 ] . w. h. kinney , phys . d * 66 * , 083508 ( 2002 ) [ arxiv : astro - ph/0206032 ] . f. lucchin and s. matarrese , phys . d * 32 * , 1316 ( 1985 ) . l. knox and m. s. turner , phys . rev . lett . * 70 * , 371 ( 1993 ) [ arxiv : astro - ph/9209006 ] . w. h. kinney and k. t. mahanthappa , phys . d * 53 * , 5455 ( 1996 ) [ arxiv : hep - ph/9512241 ] . e. j. copeland , a. r. liddle , d. h. lyth , e. d. stewart and d. wands , phys . d * 49 * , 6410 ( 1994 ) [ arxiv : astro - ph/9401011 ] . a. d. linde , phys . d * 49 * , 748 ( 1994 ) [ arxiv : astro - ph/9307002 ] . w. h. kinney , phys . d * 72 * , 023515 ( 2005 ) [ arxiv : gr - qc/0503017 ] . d. boyanovsky , h. j. de vega and n. g. sanchez , arxiv : astro - ph/0507595 . m. berkooz , d. j. h. chung and t. volansky , arxiv : hep - ph/0507218 . l. a. boyle , p. j. steinhardt and n. turok , arxiv : astro - ph/0507455 . j. e. lidsey , a. r. liddle , e. w. kolb , e. j. copeland , t. barreiro and m. abney , rev . phys . * 69 * , 373 ( 1997 ) [ arxiv : astro - ph/9508078 ] . d. n. spergel _ et al . _ , arxiv : astro - ph/0603449 .

the two best studied variants of the massless gross - neveu ( gn ) model @xcite differ by their chiral symmetry groups . the first model has the lagrangian @xmath3 throughout this paper , we are always in 1 + 1 dimensions and suppress contracted color " labels on the fermion bilinears ( 1 ... @xmath0 ) . the lagrangian ( [ a1 ] ) is invariant under independent sign flips of left- and right - handed dirac fields , corresponding to the discrete chiral group z@xmath4z@xmath5 . the second model has a continuous u(1)@xmath6u(1)@xmath2 chiral symmetry and is often referred to as two - dimensional nambu jona - lasinio ( njl@xmath7 ) model @xcite , @xmath8 \qquad ( { \rm u(1)}{\text -}{\rm njl } ) . \label{a2}\ ] ] here , phases of left- and right - handed fermion fields can be rotated independently . as a phenomenological model , the njl model in 3 + 1 dimensions is in general considered with @xmath9 , but two additional flavors ( @xmath10 ) corresponding to isospin in strong interaction physics @xcite . this leads to the lagrangian with su(2)@xmath6su(2)@xmath2 chiral symmetry @xmath11 \qquad ( { \rm su(2)}{\text -}{\rm njl } ) . \label{a3}\ ] ] the same theory in 1 + 1 dimensions is also exactly solvable in the large @xmath0 limit , see the recent paper @xcite . as shown there and noticed before in the context of the phase diagram @xcite , the physics is closer to that of the gn model ( [ a1 ] ) than to that of the njl@xmath7 model ( [ a2 ] ) . this reflects the following group - theoretical fact : the su(2 ) chiral symmetry also entails the discrete symmetry of the one - flavor gn model through the center of the group su(2 ) , but obviously not the full u(1 ) symmetry . if lagrangian ( [ a3 ] ) is the generalization of lagrangian ( [ a1 ] ) to two flavors , it is not too hard to identify the generalization of lagrangian ( [ a2 ] ) to two flavors : it is the njl - type model with u(2)@xmath1u(2)@xmath2 chiral symmetry and lagrangian @xmath12 \qquad ( { \rm u(2)}{\text -}{\rm njl } ) . \label{a4}\ ] ] in which sense model ( [ a4 ] ) is the generalization of model ( [ a2 ] ) and model ( [ a3 ] ) the generalization of model ( [ a1 ] ) will be made more precise below , after having the solutions of all four models at our disposal . the u(2)-njl model has been considered in 3 + 1 dimensions before @xcite . in 1 + 1 dimensions , it has been explored using non - abelian bosonization at finite @xmath0 @xcite . when going through the four models ( [ a1])([a4 ] ) , the complexity increases significantly , as evidenced by the number of ( real ) bilinear condensates which is doubled at each step ( @xmath13 ) . nevertheless , according to the experience with the one - flavor models , the solution of the u(2)-njl model is expected to be simpler than that of the su(2)-njl model . model ( [ a4 ] ) is also interesting for another reason : recently , takahashi has generalized solutions of the bogoliubov - de gennes ( bdg ) equation to multicomponent form in the context of unconventional fermi superfluids @xcite . this work has been found useful for solving twisted kink dynamics in the su(2 ) symmetric model @xcite . it was noticed there that the elementary kink of takahashi , the simplest multi - fermion bound state , does not exist in the su(2 ) model , since it is not charge conjugation invariant . for the same reason the kink of shei in the njl@xmath7 model @xcite does not appear in isolation in the gn model , only as a constituent of kink - antikink bound states ( dashen - hasslacher - neveu baryon @xcite ) . it is then plausible , and will indeed be confirmed below , that takahashi s kink will appear in the model ( [ a4 ] ) as a physical state . the present paper deals mostly with lagrangian ( [ a4 ] ) , but cross references to the other variants of the gn model will be frequent . it is therefore necessary to adopt a simple naming of the four distinct models . since our emphasis is on the chiral symmetry group , we will refer in the present work to the models ( [ a1 ] ) ( [ a4 ] ) as z@xmath7-gn , u(1)-njl , su(2)-njl and u(2)-njl models , respectively , as already indicated in the equations . we also omit the subscript 2 for the number of dimensions from now on . this paper is organized as follows . [ sect2 ] develops the mean field theory of the u(2)-njl model , i.e. , the hartree - fock ( hf ) and time - dependent hartree - fock ( tdhf ) approaches . in sect . [ sect3 ] , spontaneous chiral symmetry breaking , the dynamical fermion mass and the vacuum manifold are discussed . the meson spectrum is the topic of sect . [ sect4 ] , where the relativistic random phase approximation ( rpa ) is invoked to analyze small fluctuations around the hf vacuum . massless hadrons , chiral spiral type condensates and the phase diagram as a function of @xmath14 are all related to the chiral anomaly and the fact that baryon number has a topological interpretation , as discussed in sect . [ sect5 ] . in sect . [ sect6 ] we generalize our previous solution of the tdhf equations from the u(1)-njl model to the two - flavor case , rederiving takahashi s recent work on the bdg equation in a slightly different language . this has been done to make sure that the formalism from condensed matter physics really applies to the relativistic quantum field case at hand , and to facilitate comparison with previous works on other variants of gn or njl models @xcite . a few simple applications to soliton problems with one and two bound states are given in sects . [ sect7 ] and [ sect8 ] . we finish with a short concluding section , sect . [ sect9 ] . the lagrangian ( [ a4 ] ) has the u(2)@xmath1u(2)@xmath2 chiral symmetry @xmath15 giving rise to 8 conserved noether currents @xmath16 in 1 + 1 dimensions , vector and axial vector currents are not independent , but satisfy @xmath17 adding and subtracting the conservation laws @xmath18 , and introducing light cone coordinates @xmath19 one finds @xmath20 if we take the expectation value of these equations in an arbitrary state , we conclude that the right - handed density @xmath21 depends only on @xmath22 , the left - handed density @xmath23 only on @xmath24 , i.e. , they can only move with the velocity of light to the right or to the left ( or be constant ) . in a localized , massive state like a solitonic bound state or breather , these densities must therefore vanish identically . the same argument goes through for the isovector currents and densities , since @xmath25 hence we anticipate that all densities and current densities must be zero inside an arbitrary soliton or multi - soliton state , a strong constraint on multi - fermion states following from chiral symmetry . the basic equation in the large @xmath0 limit is the relativistic version of the time dependent hartree - fock ( tdhf ) equation . in the present case of lagrangian ( [ a4 ] ) , it reads @xmath26 \psi = 0 \label{b7}\ ] ] together with the self - consistency conditions @xmath27 we choose a chiral representation of the dirac matrices ( diagonal @xmath28 ) @xmath29 the tdhf equation in hamiltonian form then assumes the form @xmath30 here , the hamiltonian has been written in 2@xmath312-block - form , @xmath32 and @xmath33 are 2d isospinors , and @xmath34 is the 2@xmath312 matrix @xmath35 the self - consistency conditions ( [ b8 ] ) are particularly concise in the 2@xmath312 matrix form @xmath36 in light cone coordinates ( [ b4 ] ) the covariant form of the tdhf equation takes on the form @xmath37 a 2-component generalization of the equations familiar from the one - flavor case . not only the dirac - tdhf equation ( [ b13 ] ) , but also the self - consistency condition ( [ b12 ] ) are manifestly preserved under chiral transformations @xmath38 that is to say that two solutions which differ only by a chiral transformation have to be considered as being physically indistinguishable . if chiral symmetry is spontaneously broken but lorentz invariance is preserved , the vacuum is characterized by a space - time independent mean field @xmath39 . in order to decide whether this happens , we diagonalize the hf hamiltonian ( a @xmath40 matrix ) with homogeneous @xmath39 , eq . ( [ b11 ] ) , in momentum space , @xmath41 the 4 eigenvalues are @xmath42 with dynamical fermion masses @xmath43 next we minimize the vacuum energy density , following closely the corresponding steps in the z@xmath7-gn model ( see e.g. @xcite ) , @xmath44 where @xmath45 is the single particle vacuum energy density of the z@xmath7-gn model , @xmath46 , \label{c5}\end{aligned}\ ] ] whereas @xmath47 denotes the double counting correction of the interaction energy density characteristic for the hf approach , @xmath48 minimizing @xmath49 with respect to @xmath50 yields the two conditions @xmath51 not surprisingly , the minimum is at @xmath52 , and we recover the gap equation of the z@xmath7-gn model with @xmath53 flavors . the renormalized vacuum energy density is accordingly @xmath54 what is the vacuum manifold ? the condition @xmath55 implies @xmath56 , eq . ( [ c3 ] ) , or @xmath57 the dynamical mass then becomes @xmath58 the vacuum potential matrix @xmath39 [ see ( [ b11 ] ) ] is given by @xmath59 and satisfies @xmath60 choosing units where @xmath61 , we first divide @xmath39 by @xmath62 from eq . ( [ c9 ] ) to get the u(2 ) matrix @xmath63 where the first factor belongs to u(1 ) , the second to su(2 ) . the condition @xmath61 implies furthermore that @xmath64 since a u(2 ) matrix has 4 real parameters , but @xmath34 in ( [ c12 ] ) is parametrized by 5 real numbers ( @xmath65 ) , one condition is indeed necessary . alternatively , it is obvious that expression ( [ c12 ] ) depends only on the 4 parameters @xmath66 . thus the vacuum manifold is u(2 ) or @xmath67 . in cases where only a single vacuum is involved , the simplest choice is @xmath68 , which can always be achieved by a chiral transformation . for twisted field configurations , we shall use @xmath68 for @xmath69 , but we then need the general expression ( [ c12 ] ) for @xmath70 . in order to derive the spectrum of fermion - antifermion bound states ( mesons ) , one has to go beyond the hf approximation and consider small fluctuations around the hf vacuum . the appropriate tool is the relativistic rpa . although the technicalities are fairly involved , we can be brief here since one can follow almost literally the corresponding caculation in the z@xmath7-gn model , described in more detail in refs . @xcite . one starts from the equation of motion for the color singlet bilinear operator @xmath71 for earlier applications of this method , mainly to two - dimensional quantum chromodynamics , see also refs . @xcite . in the present case , @xmath72 are combined dirac- and isospin indices ranging from 1 to 4 . the @xmath40 matrix operator @xmath73 satisfies the equation of motion @xmath74 o_n q(x , y ) - q(x , y ) o_n { \rm tr } \left [ o_n q(y , y)\right ] \right\ } \label{d2}\end{aligned}\ ] ] where @xmath75 we have dropped terms irrelevant in the large @xmath0 limit . in the case of isovector operators @xmath76 , scalar products are implied in eq . ( [ d2 ] ) . in the z@xmath7-gn model , only the @xmath77-term was present . the next steps will not be shown in detail , but only enumerated : 1 . expand the bilinear operator around the vacuum expectation value @xmath78 2 . linearize the equation of motion in the fluctuation part @xmath79 . 3 . choose the vacuum @xmath68 without loss of generality . 4 . transform @xmath80 and @xmath81 to momentum space . expand @xmath80 into vacuum spinors @xmath82 . in the large @xmath0 limit , only particle - hole components are important , @xmath83 this expansion is related to the dirac indices only , so that @xmath84 are still 2@xmath312 matrices consisting of isoscalar and isovector pieces , @xmath85 6 . sandwich the equation of motion between the vacuum and one - meson states with momentum @xmath86 , @xmath87 the rpa - amplitudes @xmath88 then satisfy coupled linear integral equations of the standard rpa form . what one finds is that the scalar and pseudoscalar mesons decouple , as do the 4 different isospin components ( isoscalar and isovector ) . the kernel of the integral equations is 1-term separable , so that there can be at most one meson bound state per channel . owing to the separability one can easily obtain an eigenvalue condition from the rpa equations . all 4 scalar mesons satisfy the same equation leading to the eigenvalue condition @xmath89 identical to the one of the z@xmath7-gn model with @xmath53 flavors . we use the notation @xmath90 and @xmath91 is the meson energy . for meson mass @xmath92 , the factor in curly brackets in eq . ( [ d8 ] ) becomes 1 and the whole equation reduces to the vacuum gap equation , hence the 4 scalar mesons all have the same mass @xmath92 . the 4 pseudoscalar mesons also satisfy identical equations , leading to the different eigenvalue condition @xmath93 from which we can read off a vanishing meson mass @xmath94 . these are the would - be goldstone bosons , reflecting the 4 flat directions ( i.e. , the dimension of the vacuum manifold ) . note that covariance is manifest in the relativistic rpa . since the meson spectrum has not been discussed in ref . @xcite , let us mention that a corresponding calculation for the su(2)-njl model would have yielded one massive scalar / isoscalar ( @xmath92 ) and three massless pseudoscalar / isovector mesons . for the present purpose , it is sufficient to consider the case where the internal rotation axis is frozen " , say in the 3-direction , @xmath95 then the two isospin channels decouple , each one corresponding to a u(1)-njl model with @xmath0 flavors and mean field @xmath96 in the same way as z@xmath7-gn solutions are particular solutions of the su(2)-njl model @xcite , we thus find that u(1)-njl solutions are particular solutions of the u(2)-njl model . the tdhf equations are evidently satisfied , as are the self - consistency conditions . the vacuum can be chosen as @xmath68 . to describe dense matter , we combine two u(1)-chiral spirals @xcite in the two isospin channels to the following transformation of the vacuum spinors @xmath97 the two resulting chiral spirals have same radius ( @xmath61 ) but different pitches , @xmath98 ( @xmath99 sign for isospin up , @xmath100 sign for isospin down ) . in this manner matter with arbitrary @xmath101 can be described most efficiently , @xmath102 in the low density limit we find massless , delocalized baryons with different isospin content . without detailed calculation , we can predict the phase diagram in ( @xmath103 space by using the known results for the u(1)-njl and su(2)-njl models @xcite . for temperatures above @xmath104 , chiral symmetry is restored ( @xmath105 ) and the fermions are massless for all ( @xmath106 ) . below @xmath107 the order parameter is expected to be @xmath108 ( @xmath109 is the dynamical fermion mass at @xmath110 ) with non - vanishing components @xmath111 and the densities and grand canonical potential density @xmath112 in the su(2)-njl@xmath7 model , we had a spatial modulation of the radius of the chiral spiral . here the radius is spatially constant , but the pitch gets modulated . notice the lack of periodicity of @xmath34 if @xmath113 and @xmath114 are incommensurate . thus the phase diagram is extremely simple and closely reminiscent of the u(1)-njl model , see fig . [ fig1 ] , in contrast to the more complicated phase diagram of the z@xmath7-gn and su(2)-njl models @xcite . it is worth noting that the type of order parameter ( [ e6 ] ) has appeared before in the literature as variational ansatz for the su(2)-njl model in 3 + 1 dimensions @xcite . in our case , the fact that the chiral group is u(2 ) rather than su(2 ) is crucial for this simple ansatz to work quantitatively , see also ref . @xcite for a comparison with the phase diagram of the su(2)-njl model in 1 + 1 dimensions . a framework which enables us to solve soliton dynamics in the u(2)-njl model is available from the theory of fermionic superfluidity and superconductivity in the form of a general solution of the multicomponent bdg equation @xcite . in a previous work , we have outlined how to adapt this formalism to the su(2)-njl model @xcite . since it is inconvenient to mix two formalisms with different conventions , and since details of the symmetries of @xmath34 and occupation fractions may be somewhat different in condensed matter and particle physics , we proceed in this section to generalize the one - flavor formalism of refs . @xcite to two flavors . we are strongly guided by takahashi s work and will arrive at results equivalent to his results , but in such a way that the equations resemble the previous ones from the one - flavor case and where covariance is more manifest . we hope that this will make applications to relativistic quantum field theories easier . after inspecting takahashi s formalism , one quickly discovers that the generalization of the framework of refs . @xcite is rather straightforward . due to the additional flavor index , the notation becomes more cumbersome , otherwise the whole scheme remains practically unchanged . this is true notably for the construction of transparent potentials where we follow closely the logic of ref . @xcite . in @xcite , the starting point for attacking @xmath115 soliton problems was a @xmath115-dimensional vector @xmath116 with components @xmath117 the @xmath118 are complex numbers ( i m @xmath119 ) characterizing the pole positions of the tdhf continuum wave functions in the complex @xmath120 plane , @xmath120 being the spectral parameter related to light cone momentum and energy ( uniformizing parameter in condensed matter language ) , @xmath121 note that @xmath122 so that @xmath123 is recognized as a plane wave evaluated at a complex spectral parameter corresponding to a bound state pole . the main generalization when going from one to two flavors consists in introducing two copies of each @xmath123 , @xmath124 here , @xmath125 is a 2-component , constant , complex vector with components @xmath126 . its precise meaning will be clarified later on when we investigate few - soliton problems in more detail . the generalization to more than 2 flavors is straightforward , but not needed here . without loss of generality , we can assume that the vectors @xmath125 are normalized ( @xmath127 ) . we shall use greek indices and the summation convention for flavor ( @xmath128 ) and suppress the indices @xmath129 referring to the bound state poles whenever possible . then we have to modify the equations of refs . @xcite as follows : continuum tdhf spinors are now 4-component objects represented as @xmath130 the following ansatz is inspired by the assumed pole structure of the continuum spinors ( @xmath115 poles , corresponding to @xmath115 bound states ) @xmath131 the @xmath132 are the amplitudes of the flavor components of the incoming plane wave @xmath133 when summing over all continuum states , they should be chosen as @xmath134 and @xmath135 to account for incoming waves in the two isospin channels ( at least , if the vacuum at @xmath136 is chosen as @xmath137 ) . the @xmath138 are closely related to bound state wave functions . they can be evaluated by linear algebra as follows : define a hermitean @xmath139 matrix @xmath140 , @xmath141 as in refs . @xcite , we suppress soliton indices ( @xmath142 ) to ease the notation . the @xmath143 satisfy the following system of linear , algebraic equations @xmath144 where @xmath145 . a constant , hermitean @xmath139 matrix @xmath146 encoding further information about the soliton configuration ( geometry , initial conditions , details about time dependence for breathers ) has been introduced . the dimension of the linear systems does not increase with the number of flavors , but depends only on the total number of bound state poles . what is new is the factor @xmath147 in @xmath148 and the fact that one needs to solve the linear equation for each flavor component @xmath149 in turn , a rather mild complication . notice that the pairs of exponentials ( @xmath150 ) provide us with @xmath151 non - normalizable solutions of the free , massive dirac equation ( @xmath61 ) , @xmath152 derivatives of @xmath140 are now 2-term separable , as opposed to 1-term separable before , @xmath153 differentiating ( [ f9 ] ) with respect to @xmath154 yields @xmath155 upon applying @xmath156 from the right , we arrive at @xmath157 this is just the 2-component version of the covariant tdhf equation with the following 2@xmath312 matrix potential @xmath158 turning to the continuum states , the dirac equation for @xmath159 introduced in ( [ f5 ] ) assumes the form @xmath160 we write down the three terms of each of these equations in detail . from the first line of ( [ f15 ] ) , @xmath161 as in the one - flavor case , terms containing @xmath162 in the first and 3rd line cancel . terms containing @xmath163 in the first and 2nd lines add up to @xmath164 and cancel the remaining terms in the 2nd and 3rd line . likewise , from the 2nd line of ( [ f15 ] ) , @xmath165 cancellations work exactly as for one flavor : terms containing @xmath163 in the 1st and 3rd lines cancel . terms containing @xmath162 in the 1st and 2nd lines add up to @xmath166 and cancel the remaining terms in 2nd and 3rd lines . next we consider the question of normalization and orthogonality of the bound state spinors . using @xmath167 we show that @xmath168 this yields the same overlap matrix as in the one flavor case , @xmath169 orthonormal bound states can be constructed as before by linear combinations of the @xmath170 , @xmath171 the resulting condition coincides with the one in the one - flavor case , @xmath172 we now turn to the self - consistency condition . the mean field @xmath173 receives contributions from the sea and the bound states , @xmath174 with @xmath175 to evaluate the continuum part , we insert the @xmath176 s and integrate over @xmath177 with a cutoff . the pole at @xmath178 yields the divergent contribution @xmath179 owing to the vacuum gap equation , this part gives self - consistency by itself , as usual in many variants of gn models . the convergent part of the sea contribution can be written down most easily if one introduces a diagonal matrix @xmath180 , @xmath181 we find @xmath182 the bound state contribution , expressed through the diagonal matrix @xmath115 with matrix elements proportional to bound state occupation fractions , @xmath183 is given by @xmath184 expressions ( [ f27 ] ) and ( [ f29 ] ) cancel provided that @xmath185 remarkably , the self - consistency condition is not affected at all when going from one to two flavors . as outlined in sect . [ sect2 ] , due to current conservation , the solitons described by this formalism should have identically vanishing baryon density ( @xmath186 ) and isospin densities ( @xmath187 ) . this should hold for left- and right - handed fermions separately , or , equivalently , for charge and current densities . it is a good test of the formalism to prove this in all generality . define @xmath188 introducing @xmath189 in addition to the ordinary su(2 ) generators @xmath190 , we can treat the baryon density and isovector densities on the same footing . the densities comprise a part from the dirac sea and a part from the bound states . the sea part has to be subtracted by the vacuum contribution for the case @xmath191 only ( baryon density ) . the sea contribution is evaluated as follows . we take into account the measure @xmath192 and write the total density as the integral @xmath193 here , @xmath194 are the contribution from the continuum state with spectral parameter @xmath120 to the densities , including the measure . inserting the continuum spinors and summing over the two isospin channels , we find @xmath195 we have introduced the vector @xmath196 with @xmath115 components @xmath197 generalizing the corresponding vector @xmath198 in the one - flavor case . consider the quadratic terms in @xmath199 in eq . ( [ f33 ] ) first . for @xmath200 , use the following identity : @xmath201 eq . ( [ f9 ] ) in the form @xmath202 then serves to eliminate @xmath148 from ( [ f35 ] ) , thereby generating two terms which cancel exactly the terms linear in @xmath162 in ( [ f33 ] ) . the remainder yields @xmath203 or , with the help of the diagonal matrix @xmath204 more compactly @xmath205 along the same lines , one finds for the right - handed density @xmath206 the bound state contribution in a notation similar to ( [ f31 ] ) is @xmath207 since @xmath208 with the matrix @xmath115 defined in ( [ f28 ] ) , the condition which guarantees vanishing of all 4 left- and right - handed densities is @xmath209 upon performing the integration over @xmath177 and noting that @xmath210 we recover the self - consistency condition ( [ f30 ] ) . this shows that the physical solutions indeed have identically vanishing densities . the result is non - trivial in the sense that it comes about as a cancellation between non - vanishing contributions from the dirac sea and the valence bound states and only holds for self - consistent solutions . hence it may be viewed as a complete shielding effect choose @xmath211 ( one soliton ) in the general formalism and work it out . after setting @xmath212 we find the mean field @xmath213 this is the most elementary type of kink interpolating between two vacua . in the rest frame , @xmath214 for @xmath136 , @xmath215 . for @xmath216 , @xmath217 with @xmath218 by a unitary transformation we can map @xmath219 onto the vector @xmath220 , in which case we recover the standard u(1)-njl kink in the isospin - up channel and the vacuum @xmath221 in the diagonal form @xmath222 hence we have a twisted kink in the isospin up channel and the vacuum for isospin down . the general expression ( [ g4 ] ) is nothing but the spectral representation of the unitary matrix @xmath221 . denoting the ( normalized ) eigenvectors corresponding to the eigenvalues @xmath223 and 1 by @xmath219 and @xmath224 , respectively , we have @xmath225 eliminating the second term with the help of the completeness relation @xmath226 then reproduces ( [ g4 ] ) . this shows that the vector @xmath219 can be interpreted as follows : @xmath219 is the eigenvector of the unitary matrix @xmath221 , the vacuum at @xmath216 , belonging to the twisted " eigenvalue @xmath223 . with one bound state , only one eigenvalue can be twisted . this is the fundamental twisted kink . since we can reduce the simplest kink problem to the kink of the u(1)-njl model , the spinors , self - consistency and vanishing baryon and isospin density follow from one - flavor results and need not be repeated here . this is exactly the kink of takahashi @xcite . it is special in the following sense : a generic twisted kink in the u(2)-njl model should connect the vacuum @xmath137 with the vacuum @xmath221 , a unitary matrix with eigenvalues @xmath227 . in the present framework this requires a bound state of two elementary kinks , see next section . this was already used in a previous work on the su(2)-njl model @xcite , where the eigenvalues have to be complex conjugates @xmath228 . it is then necessary to bind a kink with its charge conjugate . no such restriction exists in the u(2)-njl model . the elementary kink exists as a physical state in the u(2)-njl model , whereas it is confined in the su(2)-njl model . this is completely analoguous to what happens with shei s twisted kink @xcite in the u(1)-njl and z@xmath7-gn models , respectively @xcite . we start with the scattering of two elementary kinks of the type discussed in the preceding section . this requires a diagonal matrix @xmath146 . we set @xmath229 where @xmath230 will be interpreted below . then the result of using the general formalism for two poles ( @xmath231 ) follows the same pattern as kink - kink scattering in the u(1)-njl model or su(2)-njl model . the mean field can be cast into the form @xmath232 the interpretation of the u(2 ) matrices @xmath233 and of the factor @xmath230 follows upon considering the asymptotics of the scattering process , @xmath234 thus @xmath230 accounts for the time delay of the solitons during the collision , and we have made explicit the intrinsic form of the soliton by pulling out the factors @xmath235 ( for @xmath236 ) and @xmath237 ( for @xmath238 ) to the right . one recognizes @xmath239 and @xmath240 as the intrinsic twist factors of the participating solitons . the formalism yields the following results @xmath241 in this formulation , the normalized vectors @xmath242 are input and specify the twist of incoming kink 1 and outgoing kink 2 . the normalized vectors @xmath243 of the twist of outgoing kink 1 and incoming kink 2 can be expressed in terms of @xmath242 and the pole positions @xmath244 as follows , @xmath245 , \nonumber \\ \vec{q}_2 & = & { \cal n}_2 \left [ \sigma_{12 } \zeta_2^ * ( \zeta_1-\zeta_1^ * ) \vec{p}_1 - \zeta_1^ * ( \zeta_1-\zeta_2^ * ) \vec{p}_2 \right ] . \label{h5}\end{aligned}\ ] ] we have used the notation @xmath246 the normalization factors ensuring that @xmath247 are given by @xmath248 here , @xmath230 is the time delay factor also appearing in eqs . ( [ h2],[h3 ] ) for which we find @xmath249 let us consider some interesting limiting cases . obviously , the modulus of @xmath250 controls the strength of the kink - kink interaction between the two kinks . if @xmath242 are parallel , @xmath251 and everything can be rotated into the isospin up channel where we recover scattering of two standard kinks of the u(1)-njl model , with the known time delay factor @xmath252 if @xmath242 are orthogonal , @xmath253 and the two kinks live in the two isospin channels without any interaction ( @xmath254 ) . by varying the angle spanned by @xmath219 and @xmath224 , we can thus vary the strength of the interaction between these two extreme cases . a special case of the kink - kink scattering problem is the bound state . this may be viewed as the generic u(2)-twisted kink , since it enables us to connect an arbitrary vacuum @xmath221 at @xmath216 to @xmath137 at @xmath136 . in the rest frame of such a composite kink , we have to choose @xmath255 some simplifications occur , @xmath256 \nonumber \\ \vec{q}_2 & = & - { \cal n}_2 \left [ 2i \sigma_{12 } e^{i\phi_2 } \sin \phi_1 \vec{p}_1 - \left ( e^{i(\phi_1+\phi_2)}-1 \right ) \vec{p}_2 \right ] \nonumber \\ { \cal n}_1^{-2 } & = & { \cal n}_2^{-2 } = 2 \left ( 1- \cos ( \phi_1+\phi_2 ) - 2 \sigma_{12 } \sigma_{21 } \sin \phi_1 \sin \phi_2 \right ) \nonumber \\ \xi^{-1 } & = & 1 - \sigma_{12 } \sigma_{21 } \frac{2 \sin \phi_1 \sin \phi_2}{1-\cos(\phi_1+\phi_2 ) } \label{h11}\end{aligned}\ ] ] in the one flavor case , it was not possible to construct a bound state of two kinks with the same twist , @xmath257 . if one specializes the formalism to this case , one recovers a single kink . in the two flavor case , this restriction does not exist anymore and we can construct a non - trivial bound state out of two constituents with the same twist . in this case , eq . ( [ h11 ] ) reduces to @xmath258 finally , we turn to the twisted breather . a breather at rest can be generated by choosing @xmath259 and a non - diagonal matrix @xmath146 . the twisted breather is even more complicated here than in the single flavor case , so we refrain from discussing all possible parameter choices and illustrate the reult for a few simple special cases only . we choose the same matrix @xmath146 than for one flavor @xcite , @xmath260 and the vectors @xmath242 as @xmath261 thus @xmath262 is the angle between @xmath219 and @xmath224 , @xmath263 if we choose @xmath219 and @xmath224 to be parallel ( @xmath264 ) , we find that @xmath265 where @xmath266 is the twisted breather in the one - flavor case @xcite , @xmath267 choosing @xmath219 and @xmath224 to be orthogonal ( @xmath268 ) one finds @xmath269 with @xmath270 here , the diagonal components @xmath271 are static , whereas the off - diagonal components @xmath272 oscillate with the same frequency as the one - flavor breather , @xmath273 for any other choice of the angle @xmath262 , all components of @xmath34 start to oscillate with the same frequency but different phases . the corresponding more complicated expressions for @xmath34 can easily be generated using the general framework , but will not be given here . in this paper , we have studied a variant of the integrable gross - neveu model family which has not yet received much attention so far in 1 + 1 dimensions : the u(2)-njl model with u(2)@xmath1u(2)@xmath2 chiral symmetry . phenomenologically , in 3 + 1 dimensions , the su(2)@xmath1su(2)@xmath2 model is more relevant . however , from a theoretical point of view , it is quite instructive to include the u(2 ) model as well . after developing the mean field approach and setting up the tdhf equation for the u(2)-njl model , we have studied the vacuum and identified the vacuum manifold as u(2 ) . small fluctuations in the 4 flat directions give rise to 4 massless pseudoscalar mesons , whereas the other 4 directions yield massive scalars right at threshold @xmath274 . the chiral spiral construction is especially simple here , as it can be invoked both for baryonic and isospin charge . the phase diagram is extremely simple , generalizing the known u(1)-njl phase diagram to the ( @xmath275 ) space . perhaps the most interesting topic is that of soliton dynamics . here it turns out that a recently developed solution of the matrix bdg system in condensed matter physics fits perfectly the u(2)-njl model . we have rederived takahashi s results in the language of our previous work on the u(1)-njl model , and confirmed that the fundamental kink is a physical state in the u(2 ) model while appearing only as a confined constituent in the su(2 ) model . this is another striking example for the close relationship between relativistic quantum field theory toy models and sophisticated , realistic condensed matter problems . in the present work together with ref . @xcite , we have generalized the well - known z@xmath7-gn and u(1)-njl models with abelian chiral groups to two flavors and non - abelian chiral groups su(2 ) and u(2 ) . the relationship between these various models is summarized in table i. .relationship among four - fermion models as discussed in main text . [ cols="^,^",options="header " , ] the first row contains the original variants of the gn model @xcite with one flavor only and discrete or continuous chiral symmetry , respectively . the most conspicuous differences between these two models are the phase diagram in the ( @xmath276 ) plane and the role played by twisted kinks , which are free in the u(1 ) model but confined into bound states in the z@xmath7 model . this pattern repeats itself in the 2nd row , the two - flavor generalizations . in fact , models in the same column share identical phase diagrams in the ( @xmath276 ) plane , as first noticed in ref . @xcite for the first column . a further generalization to the groups su(@xmath115 ) and u(@xmath115 ) is straightforward , thereby extending the spectrum of integrable quantum field theories substantially . it is plausible ( but ought to be checked in future work ) that the marked differences in the phase diagram and in the role of twisted kinks show up for arbitrary numbers of flavors as well . our final remark concerns the relationship between the leading order large @xmath0 results and the method of non - abelian bosonization @xcite . it is well known that two - dimensional multi - color and -flavor njl models at any finite @xmath277 can be mapped onto decoupled bosonic field theories of wess - zumino - novikov - witten type @xcite and free bosonic fields . recently there has been a lot of progress in solving such models using conformal field theory techniques and numerical methods , even for the case when chiral symmetry is explicitly broken by a mass term and more than one coupling constants , see ref . @xcite and references therein . it would be interesting to compute space - time dependent condensates in these integrable models at finite @xmath278 and compare them with the results in the limit @xmath279 , as this limit is expected to be rather singular . the author thanks falk bruckmann for his interest in this work and useful discussions . 99 d. j. gross and a. neveu , phys . rev . d * 10 * , 3235 ( 1974 ) . y. nambu and g. jona - lasinio , phys . rev . * 122 * , 345 ( 1961 ) . y. nambu and g. jona - lasinio , phys . rev . * 124 * , 246 ( 1961 ) . m. thies , phys . d * 93 * , 085024 ( 2016 ) . a. heinz , f. giacosa , m. wagner , d. h. rischke , phys . d * 93 * , 024512 ( 2016 ) . d. ebert and m. k. volkov , z. phys . c * 16 * , 205 ( 1983 ) . d. ebert and h. reinhardt , nucl . b * 271 * , 188 ( 1986 ) . d. ebert , m. nagy , m. k. volkov , phys . atom . nucl . * 59 * , 140 ( 1996 ) . a. m. tsvelik , sov . jetp * 66 * , 754 ( 1987 ) . d. a. takahashi , phys . b * 93 * , 024512 ( 2016 ) . d. a. takahashi , prog . 2016 * , 043i01 . shei , phys . d * 14 * , 535 ( 1976 ) . r. f. dashen , b. hasslacher , a. neveu , phys . d * 12 * , 2443 ( 1975 ) . g. v. dunne and m. thies , phys . d * 89 * , 025008 ( 2014 ) . g. v. dunne and m. thies , phys . * 111 * , 121602 ( 2013 ) . g. v. dunne and m. thies , phys . a * 88 * , 062115 ( 2013 ) . v. schn and m. thies , _ at the frontiers of particle physics : handbook of qcd , boris ioffe festschrift _ , edited by m. shifman ( world scientific , singapore , 2001 ) , vol . 3 , p. 1945 . r. pausch , m. thies , v. l. dolman , z. phys . a * 338 * , 441 ( 1991 ) . w. r. gutierrez , nucl . b * 176 * , 185 ( 1980 ) . m. cavicchi , p. di vecchia , i. pesando , mod . lett . a * 8 * , 2427 ( 1993 ) ; erratum : mod . a * 8 * , 2909 ( 1993 ) . r. de mello koch , j. p. rodrigues , phys . d * 54 * , 7794 ( 1996 ) . v. schn and m. thies , phys . d * 62 * , 096002 ( 2000 ) . g. basar , g. v. dunne , m. thies , phys . d * 79 * , 105012 ( 2009 ) . m. thies and k. urlichs , phys . d * 67 * , 125015 ( 2003 ) . d. nowakowski , m. buballa , s. carignano , j. wambach , arxiv:1506.04260 [ hep - ph ] . e. witten , comm . phys . * 92 * , 455 ( 1984 ) . j. wess and b. zumino , phys . b * 37 * , 95 ( 1971 ) . s. p. novikov , sov . math * 24 * , 222 ( 1981 ) . p. azaria , r. m. konik , ph . lecheminant , t. plmai , g. takcs , a. m. tsvelik , arxiv:1601.02979 [ hep - th ]

the study of collisional relaxation in stellar nuclei around massive black holes and the associated rates of capture has a long history . the pioneering work of @xcite established a quasi - steady - state solution for the stellar distribution , now known as a bahcall - wolf cusp , which has a density @xmath6 inside the radius of influence @xmath7 , defined roughly as the radius enclosing a mass in stars equal to the mass @xmath8 of the hole . their solution was obtained from the steady - state , one - dimensional fokker - planck equation describing two - body relaxation and energy exchange between stars in the ( newtonian ) gravitational field of the massive object , and is characterized by zero ( or very small ) flux of stars with respect to energy into the central hole . a more refined treatment requires the concept of a `` loss cone '' , the region of phase space corresponding to stars with sufficienly low angular momenta to be captured by the black hole : @xmath9 , where the capture boundary @xmath10 is determined either by tidal disruption or by direct capture , at some radius @xmath11 @xcite . the latter paper also introduced the important distinction between empty- and full - loss - cone regimes , with the boundary between them defined as the energy at which the typical change of angular momentum in one radial period , @xmath12 , is equal to @xmath13 . @xcite derived the quasi - steady - state rate of consumption of stars as a function of energy and showed that the distribution function depends logarithmically on angular momentum near the loss cone . these authors , and subsequently @xcite , included capture from the loss cone via an energy - dependent sink term in the one - dimensional fokker - planck equation for @xmath14 . while the addition of such a term greatly increases the capture rate , it was found to have little effect on the form of the density profile at radii @xmath15 . @xcite solved two - dimensional fokker - planck equation in @xmath16 space and confirmed the results of the more approximate one - dimensional studies . these early studies were targeted toward massive black holes in globular clusters . the theory was subsequently applied to determine capture rates in galactic nuclei @xcite . studies based on the fokker - planck equation were also verified by other methods such as monte - carlo models @xcite , gaseous models @xcite and direct @xmath5-body simulation @xcite . relaxation times in galactic nuclei are much longer than in globular clusters , and in many cases much longer than galaxy lifetimes . partly as a consequence , galactic nuclei need not be spherically symmetric , and they may contain a substantial population of `` centrophilic '' orbits ( saucers , pyramids ) that dominate the capture rate , even in the absence of gravitational encounters @xcite . in the context of collisional loss - cone repopulation , by far the majority of studies have assumed spherical symmetry . this assumption is not crucial for @xmath5-body integrations ( except insofar as it can be difficult to construct nonspherical initial conditions ) , but it is an important ingredient of fokker - planck studies , because it ensures the conservation of angular momentum in the unperturbed motion . to date , all fokker - planck treatments that allowed for non - sphericity @xcite have assumed axisymmetry and have further restricted the allowed form of @xmath17 by writing @xmath18 , with @xmath2 the component of angular momentum parallel to the symmetry axis . two integrals of motion are not sufficient to specify regular motion in the axisymmetric geometry however , and numerical integrations in axisymmetric potentials typically reveal a third integral , @xmath19 ( with some orbits chaotic ) . of course , the reason for the neglect of @xmath19 is the absence of knowledge about its functional form . for mildly flattened systems , @xmath19 can be approximated by @xmath20 @xcite , and this approximation has been used as a basis for constructing steady - state models ( e.g. * ? ? ? the neglect of the third integral in the axisymmetric problem has several important consequences . instead of individual orbits populated by stars having the same values of their integrals of motion , one effectively considers ensembles of orbits composed of stars with different @xmath19 . moreover , by setting @xmath18 , the phase space density is forced to be uniform within this ensemble , which may lead to unphysical constraints on the possible evolution of the system . the diffusion coefficients must also be evaluated as if they did not depend on @xmath19 , or , more correctly , are averaged over all possible values of @xmath19 . finally , ignoring the third integral precludes the detailed study of regular orbits such as saucers , which might be expected to dominate the loss rate @xcite . however , sufficiently close to the black hole , the unperturbed motion is nearly keplerian , and standard `` planetary '' perturbation theory implies the existence of a third integral , which can sometimes even be expressed in terms of simple functions @xcite . in this approximation , the unperturbed stellar orbits are regular ( integrable ) and respect three independent integrals of the motion : @xmath0 , @xmath2 and @xmath3 , where @xmath3 is the `` secular , '' i.e. averaged , hamiltonian . given an analytic expression for the third integral in the vicinity of the black hole , fully three - dimensional fokker - planck studies become feasible , describing the time evolution of @xmath21 . for the present study , however , we choose to concentrate on evolution in the two - dimensional subspace @xmath22 with the gravitational potential , and the orbital energy @xmath0 , fixed . our justification for ignoring changes in @xmath0 is the same as in many previous studies of the loss - cone problem in galactic nuclei @xcite : energy relaxation time scales are typically very long in nuclei , too long for steady - state configurations like the bahcall - wolf cusp to be reached . instead , the dependence of @xmath17 on @xmath0 is inferred from the observed , radial density profile via eddington s formula . our goal is to generalize the well - known , one - dimensional solutions for @xmath23 in the spherical geometry to the two - dimensional case @xmath24 in the axisymmetric geometry . the paper is organized as follows . in [ sec_orbital_structure ] we use the averaging method to demonstrate the existence of a third integral of motion inside the supermassive black hole ( sbh ) sphere of influence and we use it to elucidate the behavior of orbits : the tube orbits that are generic to the axisymmetric geometry , and the saucer orbits that inhabit the low - angular - momentum parts of phase space . the fokker - planck equation is written down in [ sec_fp_definition ] , and a scheme for calculating diffusion coefficients is presented in the case of three integrals of motion . following this general treatment , we then restrict our attention to weakly flattened systems , which allows some simplification in the computations . we also concentrate on the case @xmath25 , which is both physically reasonable , and which results in analytic expressions for many of the diffusion coefficients . [ sec_boundary_cond ] discusses the proper boundary conditions for the fokker - planck equation , and [ sec_fp_solution ] is devoted to the solution of the two - dimensional equation and comparison with the one - dimensional ( spherical ) case . it turns out that the flux of stars into the sbh is enhanced with respect to the spherical case , but by a modest factor : at most a factor of a few . in [ sec_nbody ] we describe direct @xmath5-body simulations designed to test the predictions of the fokker - planck models ; sections [ sec_chaos ] and [ sec_triaxiality ] briefly discuss the role of chaotic orbits beyond the sbh influence sphere , and triaxiality of the stellar potential , respectively . finally , in [ sec_estimates ] we estimate capture rates in realistic galaxy models , using the fokker - planck models to access the range of parameters not presently accessible to @xmath5-body simulations . consider a stellar nucleus in which the density varies as a power of radius , @xmath26 , and in which the equidensity contours are flattened in the direction of the short ( @xmath1 ) axis ; in other words , an oblate system . let @xmath27 be the axis ratio , i.e. the ratio of radii along the minor and major axes at which densities are equal . in the first approximation , the stellar density and potential of a flattened system are described by the spherical part modified by an @xmath28 legendre polynomial : [ eq_modelrhophi ] @xmath29 \right ) , \label{eq_rho_cusp } \\ \phi_\star(\boldsymbol{x } ) & = & \phi_0 \left(\frac{r}{r_0}\right)^{2-\gamma}\ , \left(1 + { \epsilon_p}\left[\frac{z^2}{r^2}-\frac{1}{3}\right ] \right ) , \label{eq_phi_cusp } \\ \phi_0 & = & \frac{4\pi\,g\rho_0\,r_0 ^ 2}{(3-\gamma)(2-\gamma ) } \ ; , \label{eq_phi0 } \\ { \epsilon_p}&= & { \epsilon_d}\,\frac{(3-\gamma)(2-\gamma)}{\gamma\,(\gamma-5 ) } \ ; , \label{eq_epsilonp}\\ { \epsilon_d}&= & -\frac{3 ( p^{-\gamma}-1)}{2p^{-\gamma}+1 } \label{eq_epsilond}\end{aligned}\ ] ] where @xmath30 . the total gravitational potential is @xmath31 where @xmath32 is the mass of the supermassive black hole ( sbh ) located at @xmath33 . throughout this section , we restrict attention to motion that satisfies @xmath34^{1/(3-\gamma ) } , \ ] ] where @xmath35 is the gravitational radius of the sbhand @xmath7 its radius of influence ; the latter is conventionally defined as the radius of a sphere containing a mass in stars equal to @xmath36 . the first inequality permits us to ignore relativistic corrections to the equations of motion , and the second allows us to treat the force from the stars as a small perturbation to the inverse - square force from the sbh . the effects of relativity are discussed briefly in appendix [ sec_appendix_relativity ] , where more precise conditions for the validity of the newtonian approximation are derived . under these conditions , one expects the motion to be nearly keplerian on time scales comparable with the radial period , and we can employ the method of averaging @xcite : the forces acting on a star are averaged over the unperturbed motion , with the orbital elements the `` osculating elements '' fixed during the averaging . it is convenient to describe the motion using the delaunay variables which are action - angle variables in the unperturbed problem : @xmath37 ( actions ) and @xmath38 ( angles ) . here @xmath39 is the angular momentum of a circular orbit with given semimajor axis @xmath40 or total energy @xmath0 , @xmath41 is the magnitude and @xmath2 is the @xmath42component of the angular momentum ( so that @xmath43 gives the inclination of orbital plane with respect to the @xmath44 plane ) ; @xmath45 is the radial phase ( mean anomaly ) , @xmath46 is the argument of periastron ( @xmath47 corresponds to periapsis in @xmath44 plane ) , and @xmath48 is the longitude of ascending node . we further define dimensionless angular momentum variables as @xmath49 $ ] and @xmath50 $ ] . we will also have occasion to refer to their squared values , denoted , following @xcite , as @xmath51 and @xmath52 . these three pairs of canonically - conjugate variables evolve according to hamilton s equations of motion , with the hamiltonian given by @xmath53 and with @xmath54 expressed in terms of the delaunay variables . we assume that these variables with the exception of the radial phase @xmath45 are nearly constant over one radial period : @xmath55 the averaged equations of motion can then be defined as the equations of motion corresponding to the averaged hamiltonian @xmath56 where the variables @xmath57 are fixed in the averaging of @xmath54 over @xmath45 . after the averaging , @xmath58 is independent of @xmath45 and therefore @xmath59 is conserved , as is the semimajor axis @xmath40 and the energy @xmath0 . on the other hand , @xmath58 itself is a ( new ) integral of the motion . finally , @xmath2 is conserved due to axial symmetry , from which it follows that the motion is integrable . remarkably , even in the ( weakly ) triaxial case there can exist three integrals of motion , @xmath2 being replaced by another conserved quantity @xcite . exact expressions for the averaged hamiltonian are given in appendix [ sec_appendix_hamiltonian ] . a good approximation to the averaged perturbing potential is [ eq_happrox ] @xmath60 which is exact for @xmath61 ( for which @xmath62 ) and approximates the true value to within a few percent in other cases . this hamiltonian is very similar to the averaged hamiltonian of the hierarchical restricted three - body problem @xcite ; a detailed comparison is presented in appendix [ sec_appendix_kozai ] . expressed in terms of a dimensionless time @xmath63 , the equations of motion read @xmath64 the equations for @xmath65 and @xmath48 are not needed because @xmath65 is conserved and because nothing important depends on @xmath48 . the time @xmath66 is differs by a factor @xmath67 from @xmath68 defined in @xcite . ] @xmath69 where @xmath70 is approximately the mass in stars within radius @xmath71 . @xmath66 is the time associated with precession of the argument of periastron due to the spherically - distributed mass ( the `` mass - precession time '' ) . @xmath72{fig1_rminmax.eps}\ ] ] in what follows , it will be convenient to replace @xmath73 by @xmath74 , a linear combination of @xmath73 and @xmath75 : [ eq_calh ] @xmath76 to obtain the solution , we express @xmath46 from ( [ eq_calh ] ) and substitute it into the first of equations ( [ eq_motion ] ) : @xmath77 . \nonumber \\ { { \mathcal r}_\mathrm{max}}&\equiv & { \mathcal r}_1 , \label{eq_rminmax } \\ { { \mathcal r}_\mathrm{min}}&\equiv & \mathrm{max}({\mathcal r}_2,{\mathcal r}_3 ) , \nonumber \\ { { \mathcal r}_\mathrm{low}}&\equiv & \mathrm{min}({\mathcal r}_2,{\mathcal r}_3 ) . \nonumber\end{aligned}\ ] ] it is clear that @xmath78 is allowed to oscillate between @xmath79 and @xmath80 . we thus have two classes of orbit , depending on the relation between @xmath81 and @xmath82 . if @xmath83 , which occurs when @xmath84 , the orbit is an ordinary short - axis tube ( sat ) ; in the opposite case ( @xmath85 ) the orbit is a saucer , with @xmath86 . ( a similar class of orbits was described by @xcite for a @xmath87 scale - free potential ; see appendix [ sec_appendix_saucers ] for more discussion ) . the main distinction is that for saucer orbits the angle @xmath46 librates around @xmath88 , which means the apoapsis always lies above ( or below ) the @xmath44 plane ; the orbit resembles a conical saucer with inner hole ( e.g. * ? ? ? * figure 7 ) . for tubes , conversely , @xmath46 steadily decreases . saucer orbits only exist in the oblate , not prolate case ; the condition that the expression under the radical in ( [ eq_r12 ] ) is nonnegative requires that @xmath89 , hence the label `` separatrix '' . the solution to equation ( [ eq_dcalrdtau ] ) can be expressed exactly in terms of the elliptic cosine ( e.g. * ? ? ? * chapter 16 ; their parameter @xmath90 , where @xmath91 is the elliptic modulus used as the second parameter in our notation ) : @xmath92 we call the period of full oscillation in @xmath78 the `` precession time . '' it is given by the complete elliptic integral : @xmath93 for orbits not too close to the separatrix , @xmath94 . it is also useful to write down expressions relating @xmath79 and @xmath80 . for tube orbits , @xmath95 it is clear that if @xmath96 and @xmath97 , these two values are quite close to each other , justifying the practice of approximating the third integral as @xmath20 . for saucer orbits , the relation is simpler : @xmath98 in particular , the condition @xmath99 gives the fixed - point orbit , for which @xmath100 always . figure [ fig_rminmax ] shows the properties of a series of orbits with the same @xmath101 which start with @xmath100 and with different initial angular momenta @xmath102 ( left panel ) , and the time evolution of @xmath78 for several saucer orbits with different values of @xmath74 ( right panel ) . equation ( [ eq_calh ] ) can be inverted to express @xmath78 as a function of @xmath46 ( figure [ fig_poincare ] ) , which is more convenient for the averaging procedure in the next section : [ eq_r_of_omega ] @xmath103 for tubes only the upper root has physical meaning , while for saucers both roots are valid as long as @xmath104 is greater than the following threshold : @xmath105 this condition arises from nonnegativity of the expression under the radical in ( [ eq_r_of_omega_a ] ) . therefore , in a time @xmath106 , @xmath46 varies by @xmath107 for tube orbits and by an amount @xmath108 for saucers . @xmath72{fig_poincare.eps}\ ] ] finally , we outline the complete phase space in ( @xmath109 ) coordinates ( figure [ fig_phaseplane ] ) . the boundary between tubes and saucers is @xmath110 , and the other important boundaries are @xmath111 the location of circular orbits ( @xmath112 ) , and @xmath113 is the line of fixed - point saucer orbits , for which @xmath114 . a star is captured if it passes near periapsis while having @xmath115 , where @xmath116 is the absorption boundary ( @xmath11 is the distance to sbh at which a star is either tidally disrupted or captured directly , and @xmath40 is the semimajor axis ) . in the axisymmetric system , this condition corresponds to @xmath117 , although not every star satisfying this condition is immediately lost ( see [ sec_boundary_cond ] ) . the lines of constant @xmath118 , which mark the sbh capture boundary , are straight lines satisfying @xmath119 this boundary in the saucer region touches the fixed - point orbit curve ( [ eq_fpo ] ) at @xmath120 in what follows , we will often assume that @xmath121 ( or at least not too large ) , equivalent to assuming that the nuclear flattening is modest . an isotropic distribution of stars in velocity space corresponds to a distribution function which does nt depend on @xmath122 ; in these canonical action - angle variables , the phase - space volume element is constant and we can compute the proportion of phase space occupied by saucer orbits by sampling initial conditions from a uniform distribution in these variables and recording the fraction of cases for which @xmath85 . for small @xmath123 , the fraction of saucer orbits turns out to be approximately @xmath124 , i.e. proportional to the degree of flattening measured by @xmath125 ( [ eq_epsilonp ] , [ eq_calhb ] ) . @xmath72{fig2_phaseplane.eps}\ ] ] we begin by outlining the general method for deriving the covariant fokker - planck equation in arbitrary coordinates @xcite . the local ( position - dependent ) fokker - planck equation can be expressed in terms of generalized position- and velocity - space coordinates @xmath126 , @xmath127 as @xmath128}{{\partial}t } & = & - \frac{{\partial}}{{\partial}j^\alpha } ( \langle \delta j^\alpha\rangle { \mathcal g}f ) \;+ \\ & + & \frac{1}{2 } \frac{{\partial}^2}{{\partial}j^\alpha { \partial}j^\beta } ( \langle \delta j^\alpha\delta j^\beta \rangle { \mathcal g}f ) \nonumber\end{aligned}\ ] ] ( * ? ? ? * equation ( 5.121 ) ) . the position and velocity coordinates need not be related to each other in any particular way ( e.g. they need not be canonically conjugate ) : for instance , one can adopt the integrals of motion as the velocity - space coordinates . in equation ( [ eq_fplocal ] ) , coefficients in angled brackets represent average and mean square changes of the corresponding velocity variables per unit time . equation ( [ eq_fplocal ] ) is valid under the assumptions of ( i ) local encounters that change only the velocity but not the position of a star ; ( ii ) weak interactions , which allows the collision term to be expanded in powers of @xmath129 up to second order . later in this section we will use the orbit - averaged form of this equation , which additionally requires that the significant changes in @xmath130 occur on a time scale much longer than the orbital period . the quantity @xmath131 is the density of states , with @xmath40 the determinant of the velocity - space metric tensor @xmath132 , so that the squared distance between two points whose coordinates differ by @xmath133 is given by @xmath134 , and the number of stars in the phase - space volume @xmath135 is @xmath136 ( generalization to an arbitrary , non - trivial metric in coordinate space is straightforward ) . under a change of coordinates @xmath137 , the coefficients in equation ( [ eq_fplocal ] ) transform as [ eq_transform_angled_coefs ] @xmath138 ( * ? ? ? * equation ( 5.120 ) ) and @xmath139 to compute the diffusion coefficients @xmath140 , @xmath141 , one needs to know the distribution function describing the field stars . here we assume that the test and field stars are drawn from the same @xmath17 , but as is often done , we replace the exact @xmath17 by an approximation when computing the diffusion coefficients . namely , we neglect changes in the diffusion coefficients caused by the evolution , and compute them assuming a field star distribution of the form @xmath142 , where @xmath143 is the binding energy per unit mass . in other words , we neglect the anisotropy of the field star distribution . consistency with the ( spherically - symmetric part of the ) mass model ( [ eq_rho_cusp ] ) requires that @xmath144 the assumption of isotropy in the field - star distribution is obviously inconsistent with the density model being flattened ; however , for a small degree of flattening it is a reasonable approximation . the flattening may be due to streaming motions , to an elongated velocity ellipsoid or to both ; deviations from isotropy due to nuclear flattening are of order @xmath96 . for an isotropic field - star population , the diffusion coefficients expressed in terms of @xmath145 are well known ( e.g. * ? ? ? * equations 5.23 , 5.55 ) and may be transformed to any other coordinates using ( [ eq_transform_angled_coefs ] ) . we first adopt the velocity - space variables @xmath146 , then later replace @xmath78 by @xmath74 which is a true integral of the motion . the expressions for the local diffusion coefficients in these coordinates are given in appendix [ sec_appendix_difcoefs ] , equations ( [ eq_dif_coefs_local ] ) . the orbit - averaged fokker - planck equation is obtained by ( i ) selecting integrals of motion as the velocity - space coordinates ; ( ii ) integrating the local fokker - planck equation over the phase - space volume filled by an orbit , assuming that @xmath147 is a constant in this region ( jeans s theorem ) ; ( iii ) using the leibnitz - reynolds transport theorem to exchange the order of integration and differentiation . the result is @xmath148}{{\partial}t } & = & - \frac{{\partial}}{{\partial}j^\alpha } ( \langle \delta j^\alpha\rangle_\mathrm{av } { \mathcal g_\mathrm{av}}f ) \;+ \\ & + & \frac{1}{2 } \frac{{\partial}^2}{{\partial}j^\alpha { \partial}j^\beta } ( \langle \delta j^\alpha\delta j^\beta \rangle_\mathrm{av } { \mathcal g_\mathrm{av}}f ) \nonumber\end{aligned}\ ] ] ( * ? ? ? * equation ( 5.153 ) ) , which has the same form as the local equation ( [ eq_fplocal ] ) , but now @xmath17 is understood to be a function of the @xmath130 and of @xmath149 only , and the diffusion coefficients are averaged according to @xmath150 setting @xmath151 in this expression gives the `` density of states '' @xmath152 , which relates @xmath17 to the number of stars in a velocity space volume element @xmath153 : @xmath154 . in the spherically - symmetric case , orbit averaging reduces to a one - dimensional integration with respect to radius , @xmath155 , where @xmath156 and @xmath157 are peri- and apoapsis radii for a given @xmath0 and @xmath41 ; in other words , averaging reduces to weighting @xmath158 in proportion to the time an orbit spends near @xmath71 . in the axisymmetric case , what is traditionally done @xcite is to assume that the orbit fills the configuration - space region defined by the condition @xmath159 where ( @xmath160 ) are cylindrical coordinates ; in other words , existence of a third integral is ignored . then the average turns out to be proportional simply to @xmath161 , the integral being taken over this region . in our case , the distinction between saucer and tube orbits is critical for the loss - cone problem and we do not want to mix orbits having different values of the third integral . we must perform the averaging taking into account the shapes of the orbits as described in the previous section . this is most easily done by adopting delaunay angular variables ( @xmath38 ) as configuration - space coordinates @xmath162 . if the corresponding actions ( @xmath37 ) were taken as velocity - space coordinates @xmath130 , the jacobian @xmath163 of this coordinate system would be unity ; but since we are using a different set of @xmath130 , @xmath163 is determined by equation ( [ eq_transform_jacobian ] ) for the coordinates of choice . we split this transformation , and the averaging procedure , into two steps . initially we adopt ( @xmath164 ) as generalized velocities @xmath130 and carry out the averaging over the radial phase angle ( mean anomaly ) @xmath45 , obtaining the coefficients for the spherical problem . since these do not depend on @xmath46 or @xmath48 , averaging over these two angles means simply multiplying by @xmath165 : @xmath166 here all coefficients , including the jacobian ( explicitly written as a determinant ) , are understood to be functions of ( @xmath164 ) . in the case of a keplerian potential @xmath167 ( i.e. neglecting the contribution of the stars ) , and assuming a power - law density profile for the stars with index @xmath168 , the field star distribution function ( in our isotropic approximation ) is a constant @xmath169 ( equation [ eq_distr_function ] ) , and the diffusion coefficients can be expressed in terms of elementary functions as [ eq_difcoefsrrz ] @xmath170\big\}\qquad \phantom{. } \\ \langle(\delta{\mathcal e})^2\rangle_\mathrm{sph } & = & { \mathcal a}\,{\mathcal e}^2\;\frac{32 - 4\sqrt{{\mathcal r}}}{15\sqrt{{\mathcal r } } } , \\ \langle(\delta{\mathcal r})^2\rangle_\mathrm{sph } & = & { \mathcal a}\,{\mathcal r}\,\frac{58 + 32\sqrt{{\mathcal r}}-90{\mathcal r}}{15 } , \\ \langle(\delta{\mathcal r}_z)^2\rangle_\mathrm{sph}&= & \langle(\delta{\mathcal r})^2\rangle_\mathrm{sph } \left(\frac{{\mathcal r}_z}{{\mathcal r}}\right)^2 + \frac{2{\mathcal a}}{15 } ( { \mathcal r}-{\mathcal r}_z ) \nonumber \\ & \times & \frac{{\mathcal r}_z}{{\mathcal r } } \big [ 9{\mathcal r}+ 29(1-{\mathcal r } ) \cos^2\omega\big ] , \label{eq_difcoefrz2 } \ ] ] @xmath171 where @xmath172 here @xmath173 is the mass of a field star ( scatterer ) and @xmath174 is the mass of a test star ( whose evolution is described by the fokker - planck equation ) . we set @xmath175 in what follows . the density of states @xmath176 is split into two factors : @xmath177 so that @xmath178 is the phase space volume element that transforms @xmath17 to the number density of stars @xmath179 in the spherical geometry : @xmath180 these coefficients are based on the usual approximation of uncorrelated two - body encounters . the effect of resonant relaxation @xcite would be to enhance the diffusion coefficients for @xmath78 and @xmath75 ; we defer a discussion of this until [ sec_nbody ] . finally , the diffusion coefficients are expressed in terms of @xmath181 by substituting @xmath182 from equation ( [ eq_calh ] ) , transforming the above coefficients under this substitution according to equation ( [ eq_transform_angled_coefs ] ) , and averaging over the argument of periastron ( i.e. the precession phase ) @xmath46 : @xmath183 we have explicitly written the jacobian of transformation ( [ eq_transform_jacobian ] ) , and the coefficients with tildes are transformed from @xmath184 using ( [ eq_transform_angled_coefs ] ) . this is possible because the last transformation does not depend on @xmath45 or @xmath48 . the limits of integration in ( [ eq_coef_avg_omega ] ) are between 0 and @xmath107 for tubes and between @xmath185 and @xmath186 for saucers ( equation [ eq_omega_min ] ) . from here on , we omit the subscript for the averaged coefficients . the averaging must be done numerically , however , we can derive asymptotic expressions for the diffusion coefficients in the large- and small-@xmath78 regimes . in the former case , when @xmath187 , @xmath188 , so we trivially get @xmath189 , @xmath190 . in the limit of small @xmath78 , i.e. close to the capture boundary , the asymptotic behavior of the coefficients is different in the tube and saucer regions . in the case of tube orbits with @xmath191 , the asymptotic expressions are @xmath192 where @xmath193 ; the other two diffusion coefficients are smaller that @xmath194 by a factor @xmath195 . for saucer orbits with @xmath196 and not too close to either the separatrix or the fixed - point orbit , the asymptotic expressions for the diffusion coefficients are estimated to within @xmath197 as @xmath198 @xmath199 finally , it is convenient to cast the orbit - averaged fokker - planck equation ( [ eq_fporbitavg ] ) into flux - conservative form : @xmath200 the drift and diffusion coefficients are [ eq_difcoefs_definition ] @xmath201 this form is convenient from both from a computational and a physical point of view , since ( i ) it manifestly conserves the total number of stars , and ( ii ) drift coefficients @xmath202 and @xmath203 must be zero , which may be demonstrated explicitly , or inferred from the natural requirement that ( in the absence of capture ) collisions tend to isotropize @xmath17 , i.e. in the steady state @xmath17 should not depend on @xmath78 or @xmath75 . as is well known , the steady - state loss rate in the spherical geometry can only be inferred from the orbit - averaged fokker - planck equation if a certain condition is satisfied : the change in angular momentum over one radial period due to encounters must be small for orbits near the loss cone . far enough from the sbh , this condition is violated ( the `` pinhole '' or `` full - loss - cone '' regime ) and the orbit - averaged approximation is not valid . this problem is dealt with by returning to the local @xmath204 fokker - planck equation and allowing the phase - space density to vary along orbits , from zero at the edge of the capture sphere to some finite value at apoapsis @xcite . the result is a boundary condition for the fokker - planck equation that specifies the value and slope of @xmath205 at @xmath206 in terms of its average value at the ( fixed ) energy @xmath207 . no such exact analysis has ever been carried out in nonspherical ( axisymmetric , triaxial ) geometries , nor will we do so here . instead we will be satisfied with a more heuristic derivation of the relation between @xmath17 and @xmath208 at the loss cone boundary . our analysis will be similar in spirit to that of @xcite , but , as we will see , with some important differences . we begin by reviewing the boundary conditions in the spherical geometry . the evolution equation in terms of @xmath207 and @xmath78 can be derived by integrating equation ( [ eq_fpflux ] ) over @xmath75 from 0 to @xmath78 . we introduce two , two - dimensional quantities : the number density of stars , @xmath209 and the flux per unit energy in the @xmath78-direction : @xmath210 the relation between @xmath211 and the fluxes appearing in equation ( [ eq_fpflux ] ) is that the former is the integral , over the entire loss cone boundary , of the component of @xmath212 normal to that boundary , in the subspace @xmath213 . in the spherical case , the capture boundary in the @xmath214 plane is defined by @xmath206 , and so @xmath215 . as emphasized by @xcite , in the spherical geometry , changes in angular momentum are expected to dominate the capture rate . setting @xmath216 constant , we consider one - dimensional diffusion in @xmath78 , which obeys @xmath217 the final expression uses the result that for small @xmath78 , @xmath218 . also in this limit , @xmath219 , and we can write @xmath220 an inverse , orbit - averaged relaxation time . if we adopt the approximation @xmath221 over the entire range of @xmath78 , then equation ( [ eq_dif_1d ] ) is mathemathically equivalent to the heat conduction equation in a circular domain @xcite . the steady - state solution is @xmath222 . the natural boundary condition at @xmath112 is @xmath223 , which in this approximation translates to @xmath224 ( in the more exact treatment , @xmath225 itself tends to zero at @xmath112 ) . the boundary at @xmath206 is responsible for capture , however , simply setting @xmath226 is not always appropriate . as discussed by @xcite , there are two regimes characterizing the behavior of @xmath179 near loss cone boundary , depending on the ratio @xmath227 between the radial period and the time to random - walk out of the loss cone ( that is , to change @xmath78 by order of @xmath228 ) : @xmath229 the case @xmath230 is the `` empty loss cone '' regime , since a star scattered into the region @xmath231 is swallowed much faster than encounters can scatter it back out . the opposite situation , @xmath232 , is the `` full - loss - cone '' regime , because diffusion is so rapid that even capture of stars ( near periapsis ) does not substantially diminish the population of loss cone orbits away from periapsis . it turns out that much of the flux into the sbh comes from stars at energies where @xmath233 , so the behavior of the solution at the transition between the two regimes is important . let @xmath234 , and define @xmath235 , the flux through the boundary ( i.e. the number of stars captured per unit time per unit energy ) . in a steady state , equation ( [ eq_dif_1d ] ) implies that @xmath211 is independent of @xmath78 near @xmath228 . we can express the inner boundary condition in a general way as @xmath236 where all quantities are understood to be functions of @xmath207 , or equivalently of @xmath237 . after usinq equations ( [ equation : nofer ] ) and ( [ equation : fluxrofe ] ) to express @xmath179 and @xmath211 in terms of @xmath17 and @xmath238 , equation ( [ eq_boundary_cond_spher ] ) is seen to be a boundary condition of the robin type ( linear combination of function and its derivative ) @xcite . the dimensionless coefficient @xmath239 can be derived by returning to the local ( non - orbit - averaged ) fokker - planck equation and determining how @xmath17 varies with radial phase assuming @xmath240 at periapsis @xcite . @xcite produced a numerical solution in the spherical geometry and proposed the approximation @xmath241 for @xmath242 and @xmath243 for @xmath244 . an exact solution exists to the same set of equations solved by cohn & kulsrud : @xmath245 @xcite , where @xmath246 are consecutive zeros of the bessel function @xmath247 . equation ( [ equation : besselalpha ] ) is unwieldy ; a good approximation is @xmath248 which has the asymptotic forms @xmath249 these expressions differ most strongly near @xmath250 , where the exact solution ( [ equation : besselalpha ] ) gives @xmath251 . cohn & kulsrud s approximation is @xmath252 while equation ( [ eq_alpha_q ] ) gives @xmath253 . in terms of @xmath239 , the variation of @xmath179 with @xmath78 near the loss cone boundary is [ eq_f_near_lc_spher ] @xmath254 here @xmath255 plays the role of the effective absorbing boundary at which @xmath256 . it is convenient to introduce the `` draining rate '' of a uniformly - populated loss cone : @xmath257 this is the capture rate that results if the following two conditions are satisfied : ( i ) the distribution function is constant inside the loss cone , with value @xmath258 ; ( ii ) the volume of the loss cone ( per unit energy ) , @xmath259 , is emptied every radial period . as is apparent from its definition , @xmath260 does not depend on the diffusion coefficient @xmath261 . we can rewrite the boundary condition ( [ eq_boundary_cond_spher ] ) in terms of @xmath260 as @xmath262 in the full - loss - cone regime , @xmath263 and the capture rate is @xmath264 ; otherwise @xmath265 , reflecting the fact that the phase space density decreases to zero at some @xmath266 ( [ eq_r0 ] ) . define @xmath267 to be the integral of @xmath268 over angular momenta : @xmath269 roughly speaking , @xmath270 is the quantity that would be inferred from an observed radial density profile , for instance , via eddington s formula ( e.g. * ? ? ? * equation 3.47 ) . if we extrapolate the logarithmic solution ( [ eq_f_near_lc_spher ] ) to all @xmath78 , we can relate @xmath271 and the capture rate @xmath272 to @xmath273 : @xmath274 in the full - loss - cone regime , @xmath232 , the distribution function is almost isotropic ( @xmath275 ) , and @xmath276 that is : the full volume of the loss cone is consumed every radial period , and @xmath260 and @xmath272 are equivalent . this also means that the exact value of the diffusion coefficient or even the very process responsible for shuffling orbits in @xmath78 does not affect the capture rate , as long as it is efficient enough to keep the loss cone full . in the opposite limit of @xmath277 , @xmath278 now the capture rate is limited by diffusion from larger @xmath78 , that is , by the gradient of @xmath179 near @xmath228 . the distribution function is depleted at small @xmath78 . here we note a distinction that will be important when discussing the axisymmetric problem . equation ( [ eq_fdrain ] ) expresses @xmath260 in terms of the value of @xmath179 at the loss cone boundary . in the spherical case , the full - loss - cone boundary condition ( @xmath279 ) implies @xmath280 ( equation [ eq_f_global_spher ] ) . this is because the same process gravitational scattering is responsible both for populating loss - cone orbits uniformly with respect to radial phase and for driving the global shape of @xmath205 towards isotropy . as we will see , the same is not necessarily true in the axisymmetric geometry , because these two actions are driven by different processes : the latter is always attributed to two - body relaxation ( scattering ) , but the former may also be driven by regular precession . in what follows , we use the terms `` empty '' and `` full '' loss - cone regimes to distinguish between the cases when the flux @xmath281 and @xmath282 , respectively , whatever the global shape of the solution . in the spherical problem , the transition between the two regimes is naturally defined as @xmath283 ( so that expressions ( [ eq_fflc ] ) and ( [ eq_felc ] ) are equal ) . sometimes another definition is used , based on the requirement that the draining rate equals the repopulation rate from nearby regions in phase space , implying @xmath250 . in [ sec_nbody ] we denote the energy of the former transition as @xmath284 and the latter as @xmath285 , in reference to the fact that these conditions are based either on the global shape of the solution or on its local properties near the capture boundary . returning to the time - dependent equation ( [ eq_dif_1d ] ) , if @xmath286 , an analytic solution exists in terms of bessel functions @xcite . if we take @xmath287 as the initial condition , where @xmath288 is the heaviside step function , then the logarithmic profile is established after @xmath289 , and the flux @xmath272 , after the initial transient , is well described by the quasi - steady - state value ( [ eq_capt_rate_spher ] ) . numerical solutions to equation ( [ eq_dif_1d ] ) without the simplifying assumption @xmath290 are found to match the analytical solution very well ( to within a few percent ) . in [ sec_nbody ] we will refer to both the quasi - stationary flux value and the time - dependent solution . one should keep in mind that the capture rate in the time - dependent case may be substantially higher than the stationary flux for @xmath291 , at least in the empty - loss - cone regime ; sometimes the ratio can be more than an order of magnitude ( e.g. * ? ? ? * ) . this , however , depends critically on the details of the initial distribution . if instead of a step - function at @xmath228 one starts from a profile with a larger area @xmath292 where the distribution function has been depleted , for example , as a result of a binary smbh scattering away stars with periapsides smaller than the binary separation , then initially the capture rate is , conversely , much lower than the stationary flux @xcite . we first summarize the various time scales in the axisymmetric geometry . the three times defined above that characterize motion in the smooth potential satisfy @xmath293 where @xmath294 ( equation [ eq_trad ] ) is the radial ( keplerian ) period , @xmath66 ( equation [ eq_tprec ] ) is the approximate time for @xmath46 to change by @xmath295 due to the spherically - distributed mass , and @xmath106 ( equation [ eq_tosc ] ) is the oscillation time for @xmath78 due to torquing from the axisymmetric component of the potential . the latter inequality is strongest for saucer orbits that are near the separatrix and which precess very slowly ( figure [ fig_rminmax ] ) . the two - body relaxation time can be estimated from the diffusion coefficients by taking the inverse of the common dimensional factor @xmath296 , equation ( [ eq_difcoefa ] ) : @xmath297 this time is smaller by a factor 0.87 than the more standard definition of the relaxation time in terms of local density and velocity dispersion ( cf . @xcite , equation 5.61 ) . clearly @xmath298 . @xcite used the term `` loss wedge '' to describe the set of orbits which can be captured by the sbh in the absense of relaxation , i.e. if their angular momentum falls below @xmath228 at some phase of the precession cycle , or , equivalently , if @xmath299 . the name reflects the fact that this region is elongated in the @xmath74 direction much more than in @xmath75 ( figure [ fig_phaseplane ] ) : its boundary is defined by setting @xmath300 in equation ( [ eq_capture_boundary ] ) . in what follows , we define @xmath301 as the value of the distribution function at the loss wedge boundary , which we approximate to be constant throughout the loss wedge , and @xmath302 as the capture rate of stars per unit energy : @xmath303 here @xmath304 and @xmath305 are fluxes defined in equation ( [ eq_fpflux ] ) , @xmath306 is the lowest possible value of @xmath74 ( [ eq_hlcfpo ] ) for orbits outside the loss wedge , and the two terms in the last integral give the contributions to the capture rate from the `` downward '' flux in the @xmath74 direction in the tube region of phase plane and the `` upward '' flux in the saucer region ( see figure [ fig_phaseplane ] ) . signs are chosen so that @xmath302 is the positive rate of capture . we now argue that it is the flux in the @xmath75 direction in the saucer region that provides the main contribution to the total capture rate , in the case @xmath307 . from equations ( [ eq_coef_asympt_sau ] ) , ( [ eq_difcoefs_definition_b ] ) we see that @xmath308 . the capture boundary ( [ eq_capture_boundary ] ) is almost parallel to the vertical ( @xmath74 ) axis in this region , with the slope @xmath309 . if we assume that @xmath310 has a certain gradient perpendicular to the capture boundary line , then its derivatives are in a similar relation : @xmath311 . the fluxes @xmath305 and @xmath304 in ( [ eq_fpflux ] ) are then comparable in magnitude ; however , in equation ( [ eq_capt_rate_total_flux ] ) the former flux is integrated in @xmath312 on an interval of length @xmath313 , while the latter is integrated in @xmath314 on an interval of @xmath228 . therefore , the contribution from the flux in the @xmath75 direction is the largest in the saucer region . from similar arguments we estimate that in the tube region the flux in the @xmath74 direction is dominant . finally , if @xmath315 , most of the loss wedge lies in the saucer orbit region , so it gives the largest overall contribution to the total capture rate @xmath302 . ( this is true only asymptotically ; as shown in the next section , even for @xmath316 the tube and saucer regions give roughly equal contribution ) . the number of stars ( per unit energy ) inside the loss wedge is given by denotes the _ integral _ of the distribution function over the loss region , not its _ value _ at the boundary as in ( [ equation : nofer ] ) . ] @xmath317 where the boundaries of the region of integration are given by @xmath318 and equation ( [ eq_capture_boundary ] ) , and we have taken @xmath17 to be constant ( @xmath301 ) within this region and used the asymptotic expression ( [ eq_coef_asympt_sau ] ) for @xmath152 in the saucer region ( which gives the main contribution to the integral in the case @xmath315 ) . the number of stars which are instantaneously inside the loss cone ( @xmath115 ) is essentially the same as in the spherical case : @xmath319 . this is an expected result : in the axisymmetric case , the effective volume of the loss region increases by a factor @xmath320 , but the probability for any star inside this loss region having an angular momentum less than the capture threshold decreases by the same factor . we are now in a position to derive the boundary condition , i.e. , to relate the capture rate @xmath302 to the value of @xmath17 at the boundary of the loss wedge , @xmath301 . an orbit inside the loss _ wedge _ is captured if its instantaneous value of @xmath78 is less than @xmath228 , i.e. , if it is in the loss _ cone _ , during periapsis passage . here , as in the spherical case , there are two possible regimes . if the radial period is short compared with the time required for an orbit to pass the minimum of its precession cycle while having @xmath115 , then every orbit in the loss wedge will be captured in a time no longer than one precession period . we call this the `` empty loss wedge '' regime . the rate of consumption of stars per unit energy , @xmath302 , is then given by the number of stars inside the loss wedge , @xmath321 , divided by their lifetime on these orbits , @xmath106 : @xmath322 in the opposite limit , a star that achieves @xmath115 while being far from the sbh may precess out of the loss cone before reaching periapsis , similar to what happens in the full - loss - cone case of the spherical problem . then the capture rate is less than given by the above equation , because not all stars in the loss wedge are captured after one precession period . it is easy to see that in this case , which can be called the `` full loss wedge regime '' , the rate of consumption is equivalent to the draining rate of the full loss cone ( [ eq_fdrain ] ) . in other words , the precession is fast and shuffles stars in angular momentum quickly enough that the loss cone stays full , hence the capture rate is just the instantaneous number of stars inside the loss cone divided by their radial period . by ignoring the effects of a finite precessional time , @xcite were essentially in this regime . by analogy with the spherical case , we introduce the quantity @xmath323 separating the two regimes : @xmath324 it is easy to see that @xmath325 at the radius of influence , where @xmath326 , as long as the flattening is not too small ( @xmath327 ) . unlike the spherical problem , the transition from empty- to full - loss - wedge regimes always occurs well within the radius of influence , and therefore the main contribution to the total capture rate comes from the full - loss - wedge regime . moreover , for most realistic cases @xmath328 for the entire range of radii . indeed , combining expresssions ( [ eq_tprec ] , [ eq_q_spher ] , [ eq_trel ] , [ eq_q_axi ] ) and substituting @xmath329 , where @xmath40 is the orbit semimajor axis , we obtain @xmath330 this ratio decreases with radius ; evaluating it at @xmath331 and substituting @xmath332 ( see [ sec_estimates ] ) , @xmath333 , where @xmath334 is the velocity dispersion of stars outside @xmath7 , and @xmath335 , we may rewrite the above expression as @xmath336 which is likely to be @xmath337 if the flattening is not too small . in other words : changes in angular momentum near the loss cone boundary are determined by the regular precession ( @xmath323 ) , and not by relaxation ( @xmath227 ) . to summarize , the boundary condition in the axisymmetric problem is @xmath338 and the relation between the flux and the value of @xmath17 at the boundary , expressed in the same way as in the spherical problem ( [ eq_boundary_cond_spher ] ) , reads @xmath339 the derivation above only gives this relation in `` integrated '' form , that is , one coefficient @xmath340 for the entire @xmath341 plane at a given energy . while this is certainly an oversimplification , we argue below that it does not greatly affect the capture rate . the distinction between empty and full loss cones in the spherical problem depends on whether @xmath281 ( equation [ eq_flux_sph ] ) or not , or equivalently whether @xmath342 . in the spherical case , @xmath343 , and the transition occurs at @xmath233 . in the axisymmetric problem , the distinction between empty and full loss _ wedges _ is whether @xmath281 or not , and the transition is at @xmath344 . in most realistic cases , @xmath328 , although it is not necessarily true that @xmath345 . therefore , the coefficient @xmath340 in the boundary condition ( [ eq_boundary_cond_axi ] ) can be both greater or less than its spherical counterpart @xmath239 ( equation [ eq_alpha_q ] ) for the same @xmath207 and @xmath228 . in the case @xmath346 ( full loss cone of the spherical problem ) there is essentially no difference in boundary conditions since @xmath323 is also greater than @xmath347 . in the opposite case ( @xmath348 ) , @xmath349 regardless of the value of @xmath323 , and the capture rate turns out to depend only weakly on it , as argued in the next section . in this latter case @xmath323 may be both greater or less than unity , i.e. the loss wedge may be either empty or full . however , as we noted above , the relation between the capture rate and @xmath350 in the axisymmetric problem depends not only on the boundary condition , but also on the structure of the entire solution , as addressed in the next section . we are primarily interested here in the capture rate , i.e. the flux of stars into the sbh , and not in the evolution of the mass distribution ( density profile , flattening ) which we assume to be fixed . the flux is determined mainly by diffusion in angular momentum . accordingly , we consider the two - dimensional fokker - planck equation describing evolution in ( @xmath109 ) and neglect diffusion in energy . we note that most of our results are quoted for a density profile @xmath351 which is reasonably close to the bahcall - wolf stationary solution , @xmath352 , further justifying the neglect of energy evolution . to date , all studies of axisymmetric systems assumed that the distribution function depends only on the two classical integrals of motion , @xmath207 and @xmath75 . then it is easy to show that @xmath353 for @xmath354 . indeed , from ( [ eq_difcoefrz2 ] ) we see that @xmath355 for small @xmath75 , and from ( [ eq_densityofstates ] ) that the density of states @xmath356 . then the total capture rate per unit energy is @xmath357 , and should be independent of @xmath75 , which leads to the square - root profile of @xmath358 . @xcite used this argument to derive the relation between the capture rate @xmath208 and the average value of the distribution function at a given energy @xmath350 , as follows . start by writing the relation between @xmath208 and the value of @xmath17 at the loss wedge boundary @xmath359 , which corresponds essentially to the full loss wedge regime ( equation [ eq_boundary_cond_axi ] with @xmath360 ) . then express the integrated flux in the @xmath75 direction as @xmath361 independent of @xmath75 . the numerical factor @xmath362 is related to the `` area of the loss wedge '' , @xmath363 , where @xmath364 is the peak angular momentum of saucer orbits and may be associated with our definition of @xmath123 . the distribution function is then @xmath365 the average distribution function is @xmath366 where @xmath367 is the fraction of the phase space volume at a given @xmath75 . @xcite took @xmath368 ; a more correct value is @xmath369 . using their value , one finds @xmath370 the correct expression would contain @xmath371 instead of @xmath372 in the brackets . the relation between the steady - state capture rate per unit energy and the average ( isotropized ) value of the distribution function is @xmath373 where @xmath374 . comparing of this expression with the analogous one in the spherical case ( [ eq_capt_rate_spher ] ) , we see that when @xmath375 , the capture rate is essentially the same as in the spherical case ( full - loss - cone regime ) , while in the opposite limit it is determined by the diffusion coefficient @xmath261 and the value of @xmath123 , rather than by the size of the loss cone @xmath228 . this is a consequence of the geometry of loss wedge boundary , which stretches in one direction to a fixed fraction @xmath123 of the phase space . @xmath376 @xmath72{fig4_f_r.eps}\ ] ] @xmath72{fig5_reff.eps}\ ] ] we set up an initially uniform ( isotropic ) distribution ( @xmath377 ) in the @xmath341 plane outside the capture boundary , defined by equating @xmath228 and @xmath79 found from equation([eq_rminmax ] ) ( @xmath81 for saucers and @xmath82 for tubes ) ; a series of isolines of constant @xmath79 is plotted in figure [ fig_phaseplane ] . we studied a range of values for both @xmath123 and @xmath315 , as well as various parameters @xmath340 in the boundary condition ( [ eq_alpha_axi ] ) . the numerical solution of equation ( [ eq_fpflux ] ) was obtained on a non - uniform rectangular grid using two different sets of coordinates , defined such that the capture boundaries are parallel to the coordinate axes ( see appendix [ sec_appendix_coords ] for details ) ; grid sizes were typically @xmath378 in each dimension and the loss region was resolved by 10 - 20% of the grid cells . we advanced the solutions until time @xmath379 to achieve a steady - state profile , from which we could extract the relation between the capture rate and the average value of @xmath17 . in the spherical case , the solution is controlled by two parameters ( aside from @xmath261 which scales the time ) : the capture boundary @xmath228 and the boundary coefficient @xmath239 ( or @xmath227 ) . as regards the steady - state profile , these two parameters are not independent , since one can always transform the problem to another ( primed ) one with @xmath380 and @xmath381 ( equation [ eq_r0 ] ) , so the family of solution is effectively one - parametric . to compare the time - dependent solution of the axisymmetric problem to the spherical case , we introduce the concept of `` equivalent spherical problem '' , that is , the one - dimensional problem with the same coefficient @xmath239 in the boundary condition as @xmath340 in equation ( [ eq_alpha_axi ] ) , and with some effective capture boundary @xmath382 chosen such that the time - dependend capture rate closely follows that of the axisymmetric problem . our goal is then to find @xmath382 as a function of the loss cone size @xmath228 and degree of flattening , the latter parametrized by @xmath123 . in the remainder of this section we present simple analytical arguments that give a qualitatively correct description of the two - dimensional numerical solution of the axisymmetric problem , and provide a fitting formula for @xmath382 . figure [ fig_streamlines ] shows stream lines of flux and isocontours of @xmath17 in a quasi - stationary 2d solution for a rather exaggerated value of @xmath383 . even in this case , most of the stream lines end inside the saucer region , and that is definitely so for more realistic ( smaller ) values of @xmath228 . it is also clear that in the saucer region , @xmath17 depends mainly on @xmath75 and is almost independent of the second coordinate , which justifies the square - root profile of @xmath358 as in equation ( [ eq_f_rz ] ) , but only in this region . in the tube region , for the greater part of the phase space ( @xmath384 ) , the solution is close to that of the spherical problem , that is , @xmath385 , with @xmath386 experiencing only small oscillations . we can build an approximate solution by joining the two asymptotic forms at @xmath123 . a better description for the saucer region accounts for the fact that the flux in the @xmath75 direction gradually decreases from @xmath208 at the capture boundary to zero at @xmath123 : @xmath387 @xmath388 is the minimum value of @xmath74 , corresponding to the fixed - point saucer orbit ( [ eq_fpo ] ) . the second , approximate equality in equation ( [ eq_fluxrz_corr ] ) is an empirical fit to the numerical 2d solution . using the asymptotic expressions ( [ eq_coef_asympt_sau ] ) , it is easy to show that for @xmath389 the flux has the form ( [ eq_fluxrz ] ) , with the numerical coefficient @xmath390 ( which is @xmath391 larger than the value used by @xcite ) . the solution in the saucer region is obtained by solving the differential equation ( [ eq_fluxrz_corr ] ) : @xmath392 which is a somewhat improved form of equation ( [ eq_f_rz ] ) . the solution in the tube region outside @xmath123 is approximated by @xmath393 the coefficient of the logarithmic term gives the same flux in the @xmath78 direction as in the spherical problem , and @xmath394 . figure [ fig_rprofile ] shows the two asymptotic expressions along with the actual numerical solution . we compute the isotropized value @xmath350 taking into account only the contribution from @xmath395 , which introduces a fractional error of at most @xmath123 : @xmath396 putting all this together and expressing the relation between @xmath208 and @xmath273 in terms of the coefficient @xmath340 ( [ eq_boundary_cond_axi ] ) , we obtain @xmath397 equation ( [ eq_capt_rate_axi_approx ] ) can be compared with equation ( [ eq_fluxmt ] ) of @xcite : both share the property of being independent of @xmath228 , replacing it with some effective capture boundary @xmath382 for the empty - loss - cone regime , although this effective value is different . by comparing ( [ eq_capt_rate_axi_approx ] ) with the spherical analog ( [ eq_capt_rate_spher ] ) , we see that in our approximation , @xmath398 . figure [ fig_reff ] shows that equation ( [ eq_capt_rate_axi_approx ] ) predicts well the flux in the numerical steady - state solutions ; a better approximation to the effective capture boundary and the capture rate is @xmath399 overall , the capture rates in the axisymmetric geometry are higher than in the spherical case with the same boundary condition @xmath400 , but not by a large factor : in the full - loss - cone regime ( @xmath401 ) they are essentially the same , while in the empty - loss - cone regime the effective boundary @xmath382 is higher than @xmath228 , but since the flux depends on it only logarithmically , the difference is not likely to be more than a factor of a few . of course , more relevant is a comparison that takes into account that @xmath239 in the spherical case may be different from @xmath340 for the same values of @xmath207 and @xmath228 , as noted near the end of the previous section . for the least bound stars which are in the full - loss - cone regime ( @xmath402 ) , @xmath325 and @xmath403 . in this case , the capture rate does not depend on the diffusion coefficient @xmath404 but only on the value of @xmath228 . the boundary condition ( [ eq_boundary_cond_axi ] ) states that the flux @xmath302 is proportional to the average , isotropized value @xmath405 , and @xmath261 cancels out . in the opposite case @xmath406 , the capture rate is limited by diffusion , and @xmath407 is no longer close to isotropic . if @xmath408 , @xmath340 is also @xmath409 , and the denominator in the expression for the capture rate ( [ eq_capt_rate_axi ] ) tends to some constant value which depends only on @xmath123 , but not on @xmath227 or @xmath228 ( provided that @xmath315 ) . it is largely irrelevant whether the boundary condition itself corresponds to the empty ( @xmath410 ) or full - loss - wedge regime . in other words , in this _ diffusion - limited regime _ ( both in spherical and axisymmetric cases ) the boundary conditions ( [ eq_boundary_cond_spher ] ) or ( [ eq_boundary_cond_axi ] ) determine @xmath258 for a given flux @xmath272 , which itself is set by the gradient of the overall steady - state profile of solution . by comparison , in the spherical case the denominator in the expression for the capture rate ( [ eq_capt_rate_spher ] ) also depends only weakly ( logarithmically ) on @xmath228 and is almost independent of @xmath239 . therefore , the difference in capture rates between the axisymmetric and spherical problems , which results from the difference between @xmath411 and @xmath412 , is at most a factor of few in the case @xmath406 . the fokker - planck formalism developed in the previous sections relied on the existence of three integrals of motion deep inside the sbh influence region . apart from some special , fully integrable cases ( e.g. * ? ? ? * ) , most axisymmetric potentials containing central point masses are characterized by chaotic motion in the low - angular - momentum parts of phase space beyond the influence radius . chaotic orbits still respect two integrals of the motion , @xmath207 and @xmath75 , but in the absence of a third integral , they can in principle fill the accessible region in the meridional plane , allowing them to be captured as long as @xmath413 . we estimate the capture rate from these orbits by the following argument , similar to an argument of @xcite . first we introduce the concept of draining of the loss region , arising from non - conservaton of angular momentum without any relaxation effects , then estimate the role of relaxation , and finally discuss the combined effects of draining and relaxation . assume that the chaotic orbits occupy a region in the @xmath214 plane with @xmath414 . the value of @xmath415 plays the same role as @xmath123 inside the radius of influence , and is comparable to it for the same degree of flattening . furthermore , we assume that every chaotic orbit with a given @xmath75 can attain values of @xmath416 $ ] with equal probability ( numerical tests verify that this is a reasonable assumption ) . recall that the number of stars with given @xmath417 , which may be identified with the probability of finding a star in a given interval of @xmath418 , is @xmath419 , with @xmath420 given by equation ( [ eq_densityofstates ] ) . then the fraction of time such an orbit spends below the capture boundary is @xmath421 , and this is essentially the probability of being captured during one radial period ( assuming the full - loss - cone boundary condition , i.e. that the change in angular momentum during one period is much larger than the capture boundary , which is reasonable for chaotic orbits ) . next we evaluate the time - dependent rate of capture of stars from chaotic orbits . from the above argument it follows that we need to consider @xmath358 decaying exponentially at every value of @xmath75 from its initial value @xmath422 , but with a different rate : @xmath423 .\ ] ] the total number of chaotic orbits with @xmath413 and their capture rate is then given by @xmath424 where @xmath425 from here it is clear that if we identify @xmath422 with the initial value of the distribution function in the loss cone @xmath258 , then the capture rate is initially equal to the draining rate of a uniformly populated loss cone ( [ eq_fdrain ] ) . in particular , when @xmath426 , we recover the standard , full - loss - cone draining rate , regardless of the value of @xmath415 . on the other hand , the draining time does depend on @xmath415 , since the number of stars in the chaotic region to be drained is @xmath427 times larger than the number of stars in the loss cone , therefore the draining time is longer than the radial period by the same factor . at times much longer than the draining time , the capture rate declines as @xmath428 , not exponentially , since it is dominated by the draining of chaotic orbits with @xmath429 . there is a great deal of similarity between the capture rates from the loss wedge of the saucer region of phase space for regular orbits within the radius of influence , and the chaotic region outside it . the details of draining are somewhat different ( in particular , for the regular orbits the draining rate declines as @xmath430 , as noted by @xcite ) , but since the draining time for saucer orbits is usually much shorter than a hubble time , we ignore that distinction and adopt the same expressions for them as for chaotic orbits . in both cases , the _ local _ boundary condition for the loss region corresponds to the full loss cone ( [ eq_boundary_cond_axi ] ) , at least for the case @xmath431 relevant for all but the most tightly bound orbits . however , the _ global _ shape of the steady - state solution depends on whether the overall flux into the low angular momentum region is limited by diffusion ( @xmath432 ) or not . in the first case , the steady - state solution will still have a logarithmic form for @xmath433 or @xmath123 , corresponding to some effective capture boundary @xmath382 , and the capture rate depends on this effective boundary only logarithmically . in the latter case , the capture rate is essentially the full loss cone rate for an isotropic distribution function . the latter case , however , is rarely attained because @xmath237 rapidly drops with decreasing binding energy . on the other hand , if the draining time for chaotic orbits is comparable to the hubble time , then their capture rate may still be quite high even in the absence of relaxation , provided that the initial value @xmath422 of the distribution function inside the chaotic region was not much different from the isotropic value @xmath350 . the combined effect of draining and relaxation may be approximately accounted for by the following recipe . let @xmath434 be the capture rate per unit energy from the fokker - planck equation with initial conditions corresponding to the loss region being initially empty ( i.e. the solution considered in [ sec_fp_solution ] ) . since the loss region initially may have some nonzero value of @xmath17 , @xmath435 , the phase - space gradient of @xmath17 near the loss region boundary will be less than arising in the fokker - planck solution , and the capture rate from relaxation alone may also be lower . we approximate the total capture rate by the sum of the draining rate @xmath436 and the collisional flux @xmath437 multiplied by @xmath438 , where @xmath439 is the number of chaotic orbits at @xmath440 with @xmath426 . this expression is used to compare fokker - planck models against @xmath5-body simulations in [ sec_nbody ] and to compute the capture rates for real galaxies in [ sec_estimates ] . it is important to note that the effective capture boundary @xmath382 defined in ( [ eq_reff_axi ] ) , as the parameter controlling the overall shape of the steady - state solution and the gradient of the distribution function ( and hence the capture rate due to diffusion ) , is not the same as the size of the loss region ( @xmath441 for saucer orbits , @xmath442 for chaotic orbits ) which determines the draining time . the former , being a fixed fraction of @xmath123 ( @xmath415 ) , is usually much larger than the latter . these results will be used in [ sec_nbody ] when we compare the model predictions with the results of @xmath5-body simulations . for completeness , and to put our results in a broader context , we briefly discuss the case when the stellar cusp around the sbh is triaxial . triaxial potentials support two distinct familes of tube orbits , circulating about the long and short axes of the triaxial figure @xcite . in addition , there is a new class of centrophilic regular orbits , the pyramids ( * ? ? ? * figure 11 ) . the defining feature of pyramids is that @xmath443 for all of them , and a star on such an orbit will eventually find its way into the sbh even without the assistance of collisional relaxation . the fraction of phase space occupied by pyramids is comparable to that of saucer orbits , i.e. @xmath313 . outside the radius of influence , the regular pyramid orbits are mostly replaced by chaotic orbits , which are however still centrophilic @xcite . .comparison of three geometries[tab_geoms ] [ cols="<,^,^,^ " , ] we scale the velocity unit of the dehnen model by identifying the line - of - sight velocity dispersion , @xmath444 , at the influence radius with the quantity @xmath334 in equation ( [ eq_mbhsigma ] ) ; this is reasonable given that @xmath445 is a weak function of projected radius @xmath446 for @xmath447 , with @xmath448 the half - light radius . ( in the case @xmath87 the central @xmath445 in the model lacking a sbhwas used . ) our models are thus defined by the two parameters @xmath449 . setting @xmath450 , we have @xmath451 pc for @xmath452 . the radius @xmath11 that defines the loss sphere around the sbhis the larger of the radius of tidal disruption , @xmath453 , and the ( newtonian ) periapsis of an orbit that just continues inside the event horizon . for the eccentric orbits that dominate the flux into the sbh , the latter quantity is @xmath454 for a schwarzschild ( nonrotating ) sbh @xcite ; this is the radius of periapsis of a keplerian orbit having the critical angular momentum . a star is tidally disrupted if the periapsis radius is less than @xmath455 ( e.g. * ? ? ? * equation 6.3 ) . here @xmath456 depends on the stellar equation of state and is @xmath457 for a solar - type main - sequence star . these disruption events , as opposed to direct captures , may be observed as optical and x - ray flares in otherwise quiescent galactic nuclei @xcite . from the condition @xmath458 we find that solar - type stars on eccentric orbits are disrupted ( not swallowed ) if @xmath459 ; disruption can occur for any @xmath460 if the star is on a less eccentric orbit , or for kerr sbhs even more massive than @xmath461 @xcite . red giants or agb stars can also be disrupted ( or at least tidally limited ) by sbhs more massive than @xmath461 @xcite . in what follows , we compute the total number of events associated with a given @xmath11 ; the ratio of number of tidal disruption flares to the total number of capture events is well studied in the literature and we do not consider it separately here . we used fokker - planck models to evaluate steady - state and time - dependent capture rates @xmath462 for galaxies after @xmath463 years starting from an initially isotropic distribution function , in both the spherical and axisymmetric geometries ( using our one - dimensional approximation of section [ sec_fp_solution ] ) , for @xmath464 . the latter value is meant to represent a `` typical , '' moderately - flattened system ; the results do not strongly depend on @xmath123 . in the time - dependent calculations , the initial conditions consisted of the isotropic models with loss - cone orbits removed ; as a result , these initial models are characterized by strong gradients of @xmath17 with respect to @xmath41 near the loss cone . overall , the calculation of total capture rate is done in the same way as in the previous section . the left panel of figure [ fig_rates ] shows results for two series of models : models with a steep ( @xmath465 ) central cusp and @xmath466 ; and models with shallow ( @xmath467 ) cores and @xmath468 . over the entire range of @xmath32 , the steady - state capture rates differ by only a factor of @xmath469 between spherical and axisymmetric geometries , consistent with the discussion near the end of [ sec_fp_solution ] . this result holds for any galactic model and depends only weakly on @xmath123 . the time - dependent rates are generally higher than in the steady state , due to the strong gradients in the initial conditions . especially for massive galaxies with long relaxation times , the approach to a steady state is slow and the flux at early stages is much higher than in equilibrium . for the most massive sbhs ( @xmath470 ) the draining time of the loss region becomes comparable to the hubble time and the capture rate is dominated by draining of chaotic orbits ( double - dashed line ) , reaching values up to @xmath471 yr@xmath472 at the upper end of the @xmath32 range . however , this should be regarded as a strong upper limit since we do not know the initial state : for instance , if the sbh formed as the result of a merger of a binary sbh , it is very likely that the low angular momentum region of phase space will have been depleted in the course of the binary s evolution , and the capture rates could initially be much _ lower _ than the steady - state values @xcite . the dependence of @xmath462 on @xmath32 for massive sbhscan be simply estimated as follows ( e.g. * ? ? ? the nuclei of massive galaxies are in the empty - loss - cone regime , so the flux per unit energy is roughly @xmath473 $ ] . the capture rate peaks at @xmath474 @xcite so the total flux is estimated as @xmath475 . assuming the @xmath476 relation , we obtain @xmath477 , and even the normalization constant evaluates to a reasonable @xmath478 yr@xmath472 for @xmath479 , despite the crudeness of the estimate . the right panel of figure [ fig_rates ] shows the uncertainties in the capture rate associated with the parameters @xmath239 and @xmath480 in the @xmath476 relation ( [ eq_mbhsigma ] ) and the slope of the density cusp @xmath481 . plotted are stationary capture rates for the spherical case ; other values scale roughly in the same proportion . the capture rates evaluated for a selection of individual galaxies from several previous studies are also plotted for comparison . it is clear that the scatter in the derived values is fairly large , about two orders of magnitude , although a general trend of decreasing rate with increasing @xmath32 is clear . comparing the inverted triangles in figure [ fig_rates]b with the double - dashed curve in figure [ fig_rates]a , we see that our estimates for the capture rate due to draining of chaotic orbits are substantially higher than those of @xcite . as argued above , this is most likely an overestimate resulting from simplistic initial conditions . it is also due partly to our selection of a different relation between the capture rate and the isotropized distribution function ; had we used their equation ( [ eq_fluxmt ] ) instead of our equation ( [ eq_capt_rate_axi_approx ] ) , the steady - state flux in the axisymmetric case would be factor of a few lower , although it should not affect the draining rate of centrophilic orbits which starts to dominate the capture rate at @xmath470 . overall , it is fair to say that our estimates predict capture rates in the range @xmath482 yr@xmath472 for less massive galaxies , and a few@xmath483 yr@xmath472 for giant galaxies with sbh masses in excess of @xmath461 . this is roughly consistent with the observationally derived estimates of rates of tidal disruption flares @xcite . ( axisymmetric ) nuclear flattening may increase these numbers by a factor of few , and triaxiality may have a more dramatic effect on the consumption rate of the most massive sbhs provided there are enough stars on centrophilic orbits . one potentially important feature of axisymmetric ( and triaxial ) systems is that most stars are consumed in the full - loss - cone regime of boundary conditions ( in spite of the fact that the angular momentum need not be isotropic ) . this means that stars approach the sbh with a wide distribution in periapsis radii , as opposed to `` barely touching '' the disruption sphere in the empty - loss - cone regime . as a consequence , many stars will be strongly tidally distorted before disruption , which may result in a distinct observational signature @xcite , although more recent studies show that the difference may not be so pronounced @xcite . we have considered collisional ( gravitational - encounter - drive ) relaxation processes near supermassive black holes ( sbhs ) in spherical and axisymmetric models of galactic nuclei . we derived a fokker - planck formalism and compared its predictions with direct @xmath5-body simulations of capture . inside the sbh radius of influence , the unperturbed motion of stars is regular , admitting three integrals of motion ( energy , @xmath1-component of the angular momentum @xmath2 , and secular hamiltonian @xmath74 ) . there are two families of orbits , tubes and saucers ; the latter exhibit large angular momentum variations and stars on saucer orbits approach much more closely to the sbh than would be expected based on their average angular momentum . regularity of the motion allowed us to write down the orbit - averaged fokker - planck equation and to calculate the diffusion coefficients based on the standard formalism . we discussed the appropriate boundary conditions for capture by the sbh , and numerically solved the two - dimensional ( @xmath484 ) fokker - planck equation . we showed that its solution can be well approximated by an equivalent one - dimensional solution for diffusion in angular momentum , given appropriate boundary conditions . an important difference with the spherical case is that the boundary condition at the loss region typically corresponds to the full - loss - cone regime , in the sense that a change in angular momentum per one orbital period is larger than the size of the loss cone . nevertheless , the global shape of the solution , and , consequently , the capture rate of stars , is determined by the diffusion coefficient ( inversely proportional to the relaxation time ) , and depends only weakly on the effective size of the loss region . this treatment was not entirely self - consistent , as it did not account for resonant relaxation or for the breakdown of the orbit - averaged approximation near or beyond the radius of influence , but it nevertheless suggests a conclusion which turns out to be robust : compared with the spherical case having the same _ real _ ( not _ effective _ ) capture boundary , the flux into the loss cone is higher in the axisymmetric case , but not by a large factor , and only in the regime where the loss cone would be empty in the spherical system . in the axisymmetric case , most of the stars find their way into the sbh while on saucer orbits . we also carried out a number of @xmath5-body integrations which were found to agree remarkably well with the corresponding fokker - planck models . the agreement was not limited to the number of captured stars ; other quantities like the distribution of energy of the captured stars , the loss cone population , the diffusion coefficients , and the change in angular momentum during one radial period prior to capture were also found to be well reproduced . we note , however , that the angular momentum diffusion is larger in @xmath5-body models for stars with high binding energies ( close to the sbh ) , which may be an indication of resonant relaxation not accounted for in the fokker - planck models ; nevertheless , its influence on the capture rate is likely to be small . we applied our fokker - planck formalism to realistic galaxies with sbh masses @xmath485 . we found that stationary capture rates are in the range @xmath486 yr@xmath472 in spherical galaxies and a factor @xmath469 higher in flattened systems , with an overall trend of decreasing event rate with increasing @xmath32 . time - dependent solutions were found to give generally higher estimated capture rates ; however , that result is likely to depend strongly on the assumed initial conditions @xcite . in particular , for massive ( @xmath470 ) sbhs in axisymmetric nuclei , the draining time of chaotic orbits just outside the radius of influence can be comparable to the hubble time ; if such orbits were not depleted by ( for instance ) the binary sbh that preceded the single sbh , capture rates might reach @xmath471yr@xmath472 . such high rates may be more relevant to triaxial galaxies in which the number of centrophilic orbits is likely to be large enough to maintain a full - loss - cone capture rate for a hubble time . the work was supported by the national science foundation via grant no . ast 1211602 and by the national aeronautics and space administration via grant no . nnx10af84 g . ev acknowledges the hospitality of the aspen center for physics . codes for computing capture rates in the spherical and axisymmetric geometries can be downloaded at http:@xmath487td.lpi.ru@xmath488@xmath489eugvas@xmath488losscone@xmath488 . abramovitz , m. & stegun , i. 1972 , handbook of mathematical functions ( new york : dover publications ) amaro - seoane p. , freitag m. , & spurzem r. 2004 , mnras , 352 , 655 bahcall , j. & wolf , r. 1976 , apj , 209 , 214 bahcall , j. & wolf , r. 1977 , apj , 216 , 883 baldwin , d. e. , cordey , j. g. , & watson , c. j. h. 1972 , nucl . fusion , 12 , 307 bar - or , b. , kupi , g. , alexander , t. 2013 , apj , 764 , 52 baumgardt h. , makino j. , ebisuzaki t. 2004 , apj , 613 , 1133 brockamp , m. , baumgardt , h. , & kroupa , p. 2011 , mnras , 418 , 1308 cohn , h. , & kulsrud , r. 1978 , apj , 226 , 1087 dehnen , w. 1993 , mnras , 265 , 250 dehnen , w. , & gerhard , o. 1993 , mnras , 261 , 311 donley , j. , brandt , w. , eracleous , m. , & boller , th . 2002 , aj , 124 , 1308 eilon , e. , kupi , g. & alexander , t. 2009 , apj , 698 , 641 einsel , ch . , & spurzem , r. 1999 , mnras , 302 , 81 eriksson , k. , estep , d. , hansbo , p. , & johnson , c. , 1996 , `` computational differential equations '' , cambridge university press , cambridge , uk evans , n. 1994 , mnras , 267 , 333 ferrarese , l. , & merritt , d. 2000 , apj , 539 , l9 fiestas , j. , spurzem , r. , & kim , e. 2006 , mnras , 373 , 677 fiestas , j. , & spurzem , r. 2010 , mnras , 405 , 194 fiestas , j. , porth , o. , berczik , p. , & spurzem , r. 2012 , mnras , 419 , 57 frank , j. , & rees , m. 1976 , mnras , 176 , 633 freitag , m. , benz , w. 2002 , a&a , 394 , 345 gaburov , e. , harfst , s. , portegies zwart , s. 2009 , new astron . , 14 , 630 gebhardt , k. , bender , r. , bower , g. , et al . 2000 , apj , 539 , l13 gerhard , o. , & binney , j. 1985 , mnras , 216 , 467 gezari , s. , basa , s. , martin , d. , et al . 2008 , apj , 676 , 944 goodman , j. 1983 , phd thesis , princeton univ . graham , a. , onken , c. , athanassoula , e. , & combes , f. 2011 , mnras , 412 , 2211 gltekin , k. , richstone , d. , gebhardt , k. , et al . 2009 , apj , 698 , 198 harfst , s. , gualandris , a. , merritt , d. , & mikkola , s. 2008 , mnras , 389 , 2 holley - bockelmann , k. , & sigurdsson , s. 2006 , arxiv : astro - ph/0601520 hopman , c. , & alexander , t. 2006 , apj , 645 , 1152 hunter , c. , & qian , e. 1993 , mnras , 262 , 401 kesden , m. 2012 , phys.rev.d , 85 , 024037 komossa , s. , & merritt , d. 2008 , apjl , 683 , l21 kozai , y. 1962 , aj , 67 , 591 lees , j. f. , & schwarzschild , m. 1992 , apj , 384 , 491 lidov , m. 1962 , planetary and space science , 9 , 719 lightman , a. , & shapiro , s. 1977 , apj , 211 , 244 lupton , r. , & gunn , j. 1987 , aj , 93 , 1106 lynden - bell , d. 1962 , mnras , 123 , 447 macleod , m. , guillochon , j. , & ramirez - ruiz , e. , 2012 , apj , 757 , 134 magorrian j. , & tremaine s. 1999 , mnras , 309 , 447 marconi , a. , & hunt , l. k. 2003 , , 589 , l21 mcconnell , n. , ma , c .- p . , gebhardt , k. , et al . 2011 , nature , 480 , 215 merritt , d. 1999 , in asp conf . 197 , dynamics of galaxies : from the early universe to the present , ed . f. combes , g. a. mamon , & v. charmandaris ( san francisco : asp ) , 221 merritt , d. , 2013 , dynamics and evolution of galactic nuclei , ( princeton : princeton university press ) merritt , d. , berczik , p. , & laun , f. 2007 , aj , 133 , 553 merritt , d. , & ferrarese , l. 2001 , apj , 547 , 140 merritt , d. , & poon , m. 2004 , apj , 606 , 788 merritt , d. , & quinlan , g. d. 1998 , apj , 498 , 625 merritt , d. , & valluri , m. 1999 , aj , 118 , 1177 merritt , d. , & vasiliev , e. 2011 , apj , 726 , 61 merritt , d. , & wang , j. 2005 , apjl , 621 , l101 mikkola , s. , & merritt , d. 2006 , mnras , 372 , 219 mikkola , s. , & merritt , d. 2008 , aj , 135 , 2398 milosavljevi , m. & merritt , d. 2003 , apj , 596 , 860 norman c. , & silk j. 1983 , apj , 266 , 502 zisik , m. n. 1993 , heat conduction ( new york : wiley ) poon , m. y. , & merritt , d. 2001 , apj , 549 , 192 preto , m. , merritt , d. , & spurzem , r. 2004 , apjl , 613 , l109 rauch , k. & tremaine , s. 1996 , new astron . , 1 , 149 richstone , d. 1982 , apj , 252 , 496 rosenbluth , m. , macdonald , w. , & judd , d. 1957 , phys.rev . , 107 , 1 saaf , a. f. 1968 , apj , 154 , 483 sambhus , n. & sridhar , s. 2000 , apj , 542 , 143 schwarzschild , m. 1979 , apj , 232 , 236 shapiro , s. , & marchant , a. 1978 , apj , 225 , 603 sirota , v. , ilyin , a. , zybin , k. , & gurevich , a. 2005 , jetp , 100 , 294 sridhar , s. & touma , j. 1997 , mnras , 287 , l1 sridhar , s. & touma , j. 1999 , mnras , 303 , 483 stone , n. , sari , r. , & loeb , a. , 2013 , mnras ( in press ) strubbe , l. & quataert , e. 2009 , mnras , 400 , 2070 syer , d. , & ulmer , a. 1999 , mnras , 306 , 35 tremaine , s. , gebhardt , k. , bender , r. , et al . 2002 , apj , 574 , 740 van velzen , s. , farrar , g. 2012 , epj web of conferences , 39 , 08002 vasiliev , e. 2013 , mnras , accepted wang , j. , & merritt , d. 2004 , apj , 600 , 149 will , c. 2012 , classical and quantum gravity , 29 , 217001 we discuss briefly the character of the motion in axisymmetric nuclei when the lowest - order post - newtonian ( pn ) corrections are included in the equations of motion . the 1pn accelerations imply an orbit - averaged rate of of periapsis advance @xmath490 ( * ? ? ? * equation ( 4.205 ) ) . the other elements of the osculating orbit exibit no secular variations at this pn order , and we ignore rotation of the sbh . expressed in terms of the dimensionless time variable @xmath63 defined in equation ( [ eq_motion ] ) , the relativistic precession rate becomes @xmath491 which can be added to the orbit - averaged equation of motion for @xmath46 , equation ( [ eq_motion ] ) . setting @xmath467 ( @xmath492 ) in the hamiltonian ( [ eq_happrox ] ) , the two nontrivial equations of motion become [ eq : dwdldtgr ] @xmath493 . \label{eq_dwdldtgrb}\end{aligned}\ ] ] if there is a fixed point , its angular momentum can be found by setting @xmath494 when @xmath100 , or @xmath495 as @xmath496 is increased from zero , the value of @xmath497 at the fixed point increases , reaching @xmath498 when @xmath499 for @xmath500 all orbits are tubes . for @xmath501 , where @xmath502 there is one family of tubes , passing below the fixed point in the @xmath503 plane , and a family of saucer - like orbits that librate around the fixed point . for @xmath504 , the separatrix encloses the fixed point and there are two families of tubes , at low and high angular momenta . in the latter case , the angular momentum associated with the separatrix at @xmath47 can be found by setting @xmath494 in equation ( [ eq_dwdldtgrb ] ) : @xmath505 @xmath506{fig_grorbits.eps}\ ] ] since @xmath507 , saucer - like orbits are present only when @xmath508 even when saucers are present , their angular momentum variations are limited by the relativistic term . a rough lower limit on the attainable angular momentum for saucers can be derived by equating the change in @xmath497 due to the torques over one gr precessional period to @xmath497 : @xmath509 which yields @xmath510 figure [ fig_grorbits ] shows numerical solutions of the equations of motion ( [ eq : dwdldtgr ] ) for @xmath511 and @xmath512 ; the nuclear flattening parameter has the same value as in figure 2 . the neglect of gr precession on the evolution of saucer orbits is justified if @xmath513 where the latter inequality expresses the fact that @xmath514 for direct captures and larger than that for tidal disruptions . requiring that @xmath515 and using equations ( [ eq_rinfl ] , [ eq_kappa ] , [ eq_ellmingr ] ) , we obtain [ eq_grcondition2 ] @xmath516 where the second line approximates @xmath333 . this is essentially the same condition that was obtained in @xcite for capture of pyramid orbits . since the quantity in the brackets is likely to be small ( unless @xmath517 is tiny ) , and because most of the flux into the sbh comes from orbits with @xmath518 , equation ( [ eq_grcondition2 ] ) suggests that relativity is not likely to be important for the total capture rate . here we give the exact expression for the orbit - averaged hamiltonian corresponding to the potential of equation ( [ eq_modelrhophi ] ) . it is convenient to express @xmath71 and @xmath1 in terms of eccentric anomaly @xmath456 rather than mean anomaly @xmath45 ( e.g. * ? ? ? * eq.1 ) : @xmath519 \;,\;\ ; e \equiv \sqrt{1-\ell^2 } .\ ] ] then @xmath520 where @xmath521 is expressed in terms of gauss hypergeometric function : @xmath522 @xmath523 \frac{\ell^2}{2 } \right\ } \left(1-\frac{\ell_z^2}{\ell^2}\right ) .\ ] ] motion in the hierarchical three - body problem is often derived from a doubly - averaged hamiltonian after expressing the equations of motion in jacobi coordinates and retaining only the lowest - order ( quadrupole ) term in the perturbation potential . the `` inner restricted problem '' @xcite assumes furthermore that the test mass orbits well inside the perturber mass . the averaged hamiltonian describing the test particle , with an appropriately chosen unit of time , is @xmath524\ ] ] ( merritt 2013 , equation 4.315 ) which may be obtained from equation [ eq_hamiltonian_approx ] by setting @xmath525 and eliminating unity in the first bracket of the first term ( equivalent to taking the limit @xmath526 , eliminating terms that do not contain @xmath125 and normalizing the unit of time to @xmath125 ) . equation ( [ eq_calhb ] ) then yields a value of @xmath527 for the parameter @xmath123 . as in the axisymmetric problem , motion in the lidov - kozai problem also exhibits two regimes : circulation in @xmath46 ( corresponding to tube orbits ) or libration about @xmath100 ( analogous to saucers ) , the latter appearing for @xmath528 . the equation of motion for @xmath78 ( [ eq_dcalrdtau ] ) is @xmath529 with the same relation between @xmath530 as in equation ( [ eq_rminmax ] ) : @xmath531 the case of circulation corresponds to @xmath532 , and libration to @xmath533 ; in the latter case the relation between minimum and maximum values of @xmath78 is the same as in equation ( [ eq_rminmax_saucer ] ) , from which it follows that the librating regime exists for @xmath534 . the separatrix between `` tubes '' and `` saucers '' is at @xmath535 and the fixed - point saucer has @xmath536 . for comparison , in the oblate axisymmetric potential considered throughout this paper , the first regime ( tube orbits ) has @xmath537 and the second ( saucers ) has @xmath538 . beyond the influence sphere , orbits similar to the saucers can still exist in axisymmetric potentials @xcite , but they are not describable in terms of osculating keplerian elements . typically such orbits are described as tube orbits that lie close to a resonance between the radial and vertical motions . to make the correspondence with our work more clear , we recast the fixed - point saucer orbit near a sbh in terms of the cartesian variables ( @xmath539 ) , i.e. cylindrical coordinates in the meridional plane . setting @xmath100 in equation ( a1 ) yields for the fixed - point orbit that generates the saucers @xmath540 where it is understood that @xmath541 and @xmath542 have their fixed - point values : @xmath543 equation ( [ eq_meridional_fp ] ) is a hyperbola in the ( @xmath160 ) plane , between the points @xmath544 the curve crosses the equatorial plane at @xmath545 , the semi - latus rectum . when the fixed - point saucer orbit first appears , at @xmath546 , it lies in the equatorial plane ( @xmath547 ) and has zero vertical thickness . for @xmath75 values smaller than this maximum , the trajectory ( [ eq_meridional_fp ] ) can be interpreted as a @xmath548 resonance between motions in the @xmath446 and @xmath1 directions . described in this way , saucer orbits near a sbh are seen to have very similar properties to orbits described by other authors in more general axisymmetric potentials . @xcite studied orbits in scale - free axisymmetric models with logarithmic potentials , @xmath549 and no central sbh . for models with density axis ratio @xmath550 , they found that saucers first appear at @xmath551 ; for smaller @xmath75 the fixed - point orbit ( which they called a `` reflected banana '' ) traces a path in the ( @xmath160 ) plane similar to equation ( [ eq_meridional_fp ] ) . they noted that motion near the fixed - point orbit is regular , i.e. non - chaotic . similar orbits were described by @xcite ( who called them `` pipe orbits '' ) and @xcite in surveys of orbits in other scale - free families of oblate models . here we present the local ( position - dependent ) diffusion coefficients appearing in equation ( [ eq_fplocal ] ) , expressed in the spherical coordinates @xmath552 and generalized velocities @xmath553 , @xmath554 , @xmath555 . we denote @xmath556 , @xmath557 , @xmath558 . @xmath559 these coefficients have been expressed in terms of velocity diffusion coefficients for @xmath145 via integrals of the distribution function @xmath560 describing the field stars ( of mass @xmath173 ) using the relative potential @xmath561 : @xmath562 where @xmath563 when making comparisons with the relaxation rates measured from @xmath5-body simulations we averaged the diffusion coefficients over the subspace @xmath564 : @xmath565 although we discuss the two - dimensional fokker - planck equation in the @xmath341 plane throughout the paper , it is more convenient to obtain the numerical solution using a different set of coordinates , for which we have implemented two variants : @xmath566 and @xmath567 , where @xmath568 the lines of constant @xmath569 are diagonal lines in the @xmath341 plane , parallel to the capture boundary in the tube ( @xmath570 ) region , and lines of constant @xmath571 or @xmath572 are also straight lines in that plane , designed in such a way that the capture boundary in the saucer region ( @xmath570 ) has @xmath573 or @xmath574 . this facilitates setting boundary conditions on these capture boundaries which are parallel to coordinate axes . we used both coordinate sets to cross - check the results .

markov chain monte carlo ( mcmc ) has become a standard tool in bayesian analysis . the greatest benefit of mcmc is its generality it is guaranteed to be consistent with virtually no assumptions on the underlying model . however , the practical applicability of mcmc generally depends on the dimension of the unknown variables , the number of data , and the computational resources available . furthermore , because mcmc is sequential in nature , it can be difficult to implement efficiently with modern parallel and distributed computing architectures ; see @xcite for general discussion about mcmc in challenging scenarios , and with parallel computing architectures . in recent years , a number of generic approximation methods have been developed for complex bayesian inference , for instance variational bayes ( cf . reviews * ? ? ? * ; * ? ? ? * ) , expectation propagation @xcite and laplace approximations @xcite . these methods have been reported to provide accurate enough inference for practical purposes in many scenarios , with a fraction of computing time compared with ( asymptotically ) exact approaches such as mcmc . one problem with approximate methods is that the related bias is typically hard or impossible to quantify , and in practice the only way to ensure the validity of the method is to compare the outcome with an exact method , such as mcmc . we promote , in broad terms , a well - known principle : the combination of an approximate method for ` preliminary ' inference , and importance sampling type correction of the output . our primary focus is on the approximation of a marginal distribution ( or marginal likelihood ) and the use of mcmc for approximate marginal inference . the outcome of the mcmc is corrected with importance sampling type estimates . we detail minimal requirements which lead to strongly consistent estimators and give general conditions under which central limit theorems hold . our setting highlights explicitly the connection of importance sampling type correction and recently developed pseudo - marginal type mcmc @xcite , such as particle mcmc @xcite , grouped independence metropolis - hastings ( gimh ) @xcite , approximate bayesian computation mcmc @xcite , the algorithm for estimation of discretely observed diffusions suggested in @xcite , and using annealed importance sampling @xcite . it also enables the use of debiased estimators as suggested in @xcite . all these methods are often computationally expensive , but our approach is based on independent realisations of such importance sampling type estimates , which can be calculated efficiently in parallel . we focus on bayesian state space models , where importance sampling and particle filters are used for correction . we believe that the setup is useful much more generally , for instance in latent variable models or discretisations of continuous models , such as diffusions or inverse problems . importance sampling correction of mcmc has been suggested early in the mcmc literature ; see at least @xcite . it has been used , for instance , when estimating bayes factors using a single mcmc output @xcite . bhattacharya @xcite has suggested a consistent estimator based on regeneration and tan , doss and hobert @xcite give such estimators in case of using multiple markov chains . using unbiased estimators of importance weights in this context has been suggested at least in @xcite , who consider a generalisation of the pseudo - marginal method , allowing for likelihood estimators that occasionally take negative values . this allows , for instance , using the recently discovered ` debiasing tricks ' @xcite in order to construct an unbiased estimator . similar tricks , which would lead into non - negative unbiased estimators , as required by the pseudo - marginal method , may be unavailable ( cf . * ? ? ? quiroz , villani and kohn @xcite applied the is correction in the data sub - sampling context . nested sampling has also appeared in many forms in the monte carlo literature . the smc^2^ algorithm @xcite is based on an application of nested sequential monte carlo steps , which has similarities with our framework , and the is^2^ method @xcite focuses on the case where the preliminary inference is based on importance sampling . we focus on the mcmc approximation of the marginal distribution , which we believe often to be easily implementable in practice . the markov chain nature of the marginal monte carlo approximation comes also with some extra theoretical issues which we address . another specific feature of our work is that we connect is type correction explicitly with recently discovered ` exact approximation mcmc ' techniques @xcite . theoretical advances @xcite have already led to more efficient implementation of such methods , but has also revealed fundamental limitations . for instance , the methods may suffer from slow ( non - geometric ) convergence in practically interesting scenarios @xcite . recently proposed correlated version of the pseudo - marginal mcmc @xcite may help in more efficient implementation of pseudo - marginal type methods , but the question of their efficient parallelisability remains a challenge . the blocked parallelisable particle gibbs @xcite has appealing limiting properties , but its implementation still requires synchronisation between every update cycle , which may be costly in some computing environments . we recognise that our is approach comes with its own limitations ( see section [ sec : discussion ] ) , but we believe that it can often be very useful with parallel computing facilities . after preliminaries and notation in section [ sec : notation ] , we discuss a general importance sampling ( is ) type correction of mcmc outputs and we state related consistency results in section [ sec : simple ] . we detail the general case where unbiased estimators are used ( proposition [ prop : simple - is ] ) and then extend the setting in order to allow inference in an extended state space ( proposition [ prop : proper - consistency ] ) . the latter result is based on a general concept ( assumption [ a : proper ] ) , which we call ` proper weighting scheme ' following the terminology of liu @xcite . it accommodates naturally estimators from importance sampling and sequential monte carlo . we believe that assumption [ a : proper ] , or the more general assumption [ a : super - general ] , could be satisfied relatively easily in a wide variety of scenarios . in section [ sec : ssm ] , we discuss how our importance sampling correction method applies in general state space models ( ssm ) , and observe that sequential importance sampling and particle filters both lead to proper weighting schemes ( corollary [ cor : proper - ssm ] ) . the unbiasedness property of particle filters observed in @xcite suggests a method of its own interest : a simple , general and parallelisable general particle ssm smoothing algorithm , with naturally available consistent confidence intervals ( proposition [ prop : particle - smooth - confidence ] ) . we are unaware of earlier works which explicitly state this potentially useful result , but we are mindful that it may be widely known by particle filter experts . section [ sec : asvar - clt ] discusses conditions under which our estimators admit a central limit theorem . the expressions for the asymptotic variance may be useful , for instance , when optimising the computational resources . section [ sec : block ] focuses on a so - called jump chain representation @xcite , where we propose block estimators that use the jump chain structure . in this context , it is possible to employ variance reduction techniques , such as stratification or rao - blackwellisation using the scheme of @xcite . section [ sec : delayed ] starts by making an explicit connection between the proper weighting schemes and the pseudo - marginal type mcmc , and then between the is type correction of mcmc and a two - stage delayed acceptance scheme @xcite . in section [ sec : exp ] , we detail a more specific class of ssms with linear - gaussian state dynamics and explain a generic laplace approximation method that fits this class of models . we compare different algorithmic variations with a model having poisson observations and with a stochastic volatility model . we conclude in section [ sec : discussion ] with a discussion of the possible implications of our findings . throughout the paper , we consider general state spaces while using standard integral notation . if the model at hand is defined on euclidean space using standard probability densities , the following paragraph can be skipped . each space @xmath0 is assumed to be equipped with a @xmath1-finite dominating measure ` @xmath2 ' on a @xmath1-algebra denoted with a corresponding calligraphic letter , such as @xmath3 . product spaces are equipped with the related product @xmath1-algebras and product dominating measures . if @xmath0 is a subset of an euclidean space @xmath4 , @xmath2 is taken by default as the lebesgue measure and @xmath3 as the borel subsets of @xmath0 . if @xmath5 is a probability density on @xmath0 , we denote the support of @xmath5 as @xmath6 , and the probability measure corresponding to @xmath5 with the same symbol @xmath7 . if @xmath8 , we denote @xmath9 , whenever well - defined . for a probability density or measure @xmath5 on @xmath0 and @xmath10 , we denote by @xmath11 the set of measurable @xmath8 with @xmath12 , and by @xmath13 the corresponding set of zero - mean functions . if @xmath14 is a markov transition probability , we denote the probability measure @xmath15 , and the function @xmath16 . iterates of transition probabilities are defined recursively through @xmath17 for @xmath18 . we follow the convention @xmath19 , and if @xmath20 , we denote by @xmath21 the integers within the interval @xmath22 $ ] . we use this notation in indexing , so that @xmath23 , @xmath24 . if @xmath25 , then @xmath26 or @xmath27 is void , so that for example @xmath28 means @xmath29 . similarly , if @xmath30 is a vector , then @xmath31 and @xmath32 . we also use double - indexing , and write @xmath33 . the markov chain @xmath34 on @xmath0 is harris recurrent with invariant probability @xmath5 , if for every @xmath35 , and every initial distribution , the ergodic averages are strongly consistent , @xmath36 we call such a chain _ harris ergodic ( with respect to @xmath5)_. recall that virtually all mcmc schemes are harris ergodic ( cf . * ) , although in some cases careless implementation could lead to non - harris chains ( cf . hereafter , @xmath37 is a probability density on @xmath0 and represents an approximation of a probability density @xmath38 of interest . [ a : mcmc - is ] the density @xmath37 and the related markov chain @xmath39 satisfy a. @xmath34 is harris ergodic with respect to @xmath37 . b. [ item : support ] @xmath40 . c. @xmath41 , where @xmath42 is a constant . assumption [ a : mcmc - is ] is equivalent to a seemingly more general statement , where @xmath38 and @xmath37 are probability measures with @xmath43 . in particular , taking @xmath37 as the dominating measure on @xmath0 , then one can define @xmath44 , and the density of @xmath38 is the corresponding radon - nikodym derivative @xmath45 . if assumption [ a : mcmc - is ] holds and it is possible to calculate the unnormalised importance weight @xmath46 pointwise , the chain @xmath39 can be weighted in order to approximate @xmath47 for every @xmath48 , using ( self - normalised ) importance sampling ( e.g. * ? ? ? * ; * ? ? ? * ) @xmath49{n\to\infty } \frac{\pi_a(w_u f)}{\pi_a(w_u ) } = \pi(f ) , \label{eq : mcmc - is}\ ] ] as the harris ergodicity guarantees the almost sure convergence of both the numerator and the denominator . in many cases of interest , the importance weights @xmath46 can not be calculated or they are extremely costly to evaluate . instead , it is often possible to construct unbiased estimators of @xmath50 , which may be used in place of @xmath50 under mild conditions . in order to formalise such a setting , we formulate the following abstract setting , which accommodates further generalisations of importance sampling type corrections . [ a : aug ] suppose @xmath34 is a harris ergodic markov chain on @xmath0 with invariant probability @xmath37 , and @xmath51 are conditionally independent given @xmath34 taking values on a space @xmath52 , such that the distribution of @xmath53 depends only upon the value of @xmath54 : @xmath55 in what follows , we consider several such augmented chains . note that the mapping @xmath56 in is never needed in practice , but only introduced for theoretic purposes . technically , @xmath56 is required to be a regular conditional distribution of @xmath53 given @xmath54 , but in practice , the concerns about the regularity of @xmath56 are usually unnecessary . [ thm : aug - consistency ] suppose assumption [ a : aug ] holds . then , @xmath57 is a harris ergodic markov chain on @xmath58 with invariant probability @xmath59 theorem [ thm : aug - consistency ] follows from lemma [ lem : aug - prop ] in appendix [ app : aug ] . let us now consider the simple importance sampling type correction where @xmath50 are replaced with unbiased estimates @xmath60 . this idea has been suggested in more specific contexts earlier at least in @xcite . [ a : unbiased ] suppose assumption [ a : mcmc - is ] holds , @xmath61 are independent random variables on @xmath62 conditional on @xmath34 , the distribution of @xmath60 depends only on the value of @xmath54 , and satisfies @xmath63 { \overset{\mathrm{a.s.}}{=}}w_u(x)$ ] for @xmath37-almost every @xmath64 . [ prop : simple - is ] suppose assumption [ a : unbiased ] holds and denote @xmath65 the estimator @xmath66 is strongly consistent , @xmath67 almost surely , if either of the following hold : a. @xmath68 almost surely and @xmath48 . b. @xmath8 is such that @xmath69 , where @xmath70 $ ] . the results follow from theorem [ thm : aug - consistency ] , once we check that the functions @xmath71 and @xmath72 . in proposition [ prop : simple - is ] : a. the case @xmath68 coincides with the case where pseudo - marginal algorithms can be implemented , and provide similarly consistent estimators for all @xmath48 ; see proposition [ prop : proper - pseudo ] . b. when @xmath60 are allowed to take also negative values , we believe that @xmath66 might fail to be consistent for all @xmath48 , and an extra condition such as stated might be indeed necessary . c. if the relative error in @xmath60 is bounded : there exists @xmath73 such that @xmath74 for all @xmath64 , then @xmath66 is guaranteed to be strongly consistent for all @xmath48 . we then focus on our main scenario , where both @xmath50 and @xmath75 may be estimated . such a scenario occurs when @xmath37 is approximation of a marginal density , which we discuss in detail after we state the following abstract assumption . [ a : super - general ] suppose assumption [ a : mcmc - is ] holds , let @xmath48 and let @xmath76 be @xmath77-valued random variables conditionally independent given @xmath34 , such that the distribution of @xmath53 depends only on the value of @xmath54 , and satisfies @xmath78 = f(x)w_u(x)\qquad\text{and}\qquad { \mathbb{e}}[w_k\mid x_k = x ] = w_u(x),\ ] ] for @xmath37-almost every @xmath64 , and such that @xmath79\ ] ] satisfies @xmath80 . [ prop : super - general ] if assumption [ a : super - general ] holds , then @xmath81{n\to\infty } \pi(f).\ ] ] follows from theorem [ thm : aug - consistency ] , because both @xmath82 and @xmath83 are in @xmath84 . in the specific cases we consider later , we define @xmath85 , where @xmath86 are random variables which satisfy @xmath87 = c_w \pi(x ) f(x ) ] \qquad \text{and}\qquad { \mathbb{e}}[u_k\mid x_k = x ] = c_w \pi(x).\ ] ] in some cases , however , it may be that @xmath88 is unavailable , for instance if @xmath89 is a pseudo - marginal algorithm targeting @xmath37 . it may still be possible to estimate the ratio @xmath46 in an unbiased fashion . in some scenarios , it may be possible to use simply @xmath90 which are independent of @xmath60 ( or @xmath91 ) , with @xmath92=f(x)$ ] . however , such @xmath90 may be more difficult to construct than dependent @xmath93 which satisfy assumption [ a : super - general ] . it is clear that assumption [ a : unbiased ] and the assumptions in proposition [ prop : simple - is ] form a special case of assumption [ a : super - general ] , by taking @xmath94 . the following scheme formulates a case where @xmath34 targets @xmath88 , which is an approximation of a marginal density @xmath95 , and the aim is inference over a joint target density @xmath96 on an extended state space @xmath97 . the following assumption is similar to ( * ? ? ? * definition 2.5.1 ) : [ a : proper ] suppose assumption [ a : mcmc - is ] holds , @xmath98 and @xmath99 are conditionally independent given @xmath34 , and take values on @xmath100^m\times{\mathbb{r}}$ ] with @xmath101 for all @xmath102 , and such that for every @xmath103 , the random variables @xmath104 satisfy @xmath105 { \overset{\mathrm{a.s.}}{=}}w_u(x ) f^*(x)$ ] , where @xmath106 for @xmath37-almost every @xmath64 . [ prop : proper - consistency ] suppose that assumption [ a : proper ] holds . then , the estimator @xmath107 satisfies @xmath108 almost surely , if either of the following hold : a. [ item : proper - positive ] @xmath68 almost surely for all @xmath109 and @xmath103 . b. [ item : proper - general ] @xmath103 and @xmath110\ ] ] satisfies @xmath80 . this is a special case of proposition [ prop : super - general ] , because assumption [ a : proper ] together with , which is also implied by , implies assumption [ a : super - general ] . the following immediate result records that importance sampling estimators fit assumption [ a : proper ] naturally . [ prop : augmented - is ] suppose @xmath111 such that @xmath112 is a probability density for each @xmath64 and such that @xmath113 . let @xmath114 and denote @xmath115 where @xmath42 is some constant , and let @xmath116 . then , @xmath117 form a proper weighting scheme ( assumption [ a : proper ] ) , and therefore , for any @xmath103 , @xmath118{n\to\infty } \pi^*(f).\ ] ] the next result shows that sub - sampling within a proper weighting scheme leads to a proper weighting scheme . [ prop : randomise - proper ] suppose that @xmath117 forms a proper weighting scheme . let @xmath119 be random variables conditionally independent of @xmath120 such that @xmath121 . then , @xmath122 forms a proper weighting scheme . proposition [ prop : randomise - proper ] follows by observing that @xmath123 is a proper weighting scheme , with @xmath124 . proposition [ prop : randomise - proper ] implies that the estimator @xmath125{n\to\infty } \pi^*(f)\ ] ] under the conditions stated in proposition [ prop : proper - consistency ] . such extra randomisation generally leads to a higher asymptotic variance , but may be preferable if the weighted sample @xmath126 needs to be stored for further analysis . in case of non - negative weights , it is easy to see that the following convex combinations of proper weighting schemes remain proper . [ prop : convex - proper ] suppose @xmath127 and @xmath128 forms a proper weighting scheme for each @xmath129 , @xmath130 are conditionally independent given @xmath34 in the sense of assumption [ a : aug ] , and @xmath131 almost surely for all @xmath132 . then , the random variables @xmath133 form a proper weighting scheme @xmath134 for every @xmath135^n$ ] with @xmath136 . we discuss the implication of proposition [ prop : augmented - is ] in case of state space models in section [ sec : ssm ] . in such a setting , @xmath89 stands for the ( static ) parameters of the model , @xmath137 for the sampled latent state trajectories , and @xmath138 for the corresponding ( unnormalised ) importance weights . we recall also the unbiasedness property of particle filters which leads to proper weighting schemes . state space model ( ssm ) techniques are applied in a wide variety of applications ( cf . * ? ? ? we assume the existence of a family of ssms indexed by a static parameters @xmath139 taking values on a space @xmath140 . the model consists of a sequence of latent ( unobserved ) states @xmath141 taking values in @xmath142 , and observations @xmath143 taking values in @xmath144 . more specifically , the latent states @xmath141 form a markov chain with initial density @xmath145 and state transition densities @xmath146 for @xmath147 . the observations @xmath143 are conditionally independent given @xmath141 and follow the observation model @xmath148 . this model defines the following probability density on @xmath149 : @xmath150 where , by convention , @xmath151 . the inference is based on the so - called smoothing distribution @xmath152 where @xmath153 is a normalising constant . in a bayesian state space model , the parameters @xmath139 are regarded random as well , with a prior density @xmath154 , which leads to the following joint density on @xmath155 : @xmath156 the ( full ) bayesian inference is based on the posterior distribution @xmath157 with overall normalising constant @xmath158 . let us start with a description of a sequential importance sampling procedure , which requires user - supplied ` proposal ' distributions @xmath159 for @xmath160 . [ alg : sis ] input : @xmath143 and @xmath161 . a. for @xmath160 , sample @xmath162 and calculate @xmath163 output : @xmath164 , where @xmath165 and @xmath166 . if the following well - known support condition holds , we recall that algorithm [ alg : sis ] provides a weighting which is shown to lead to proper weighting in corollary [ cor : proper - ssm ] . [ a : support ] suppose that @xmath167 for all @xmath160 , where @xmath168 [ prop : sis - proper ] suppose assumption [ a : support ] holds , then the random variables @xmath169 , @xmath170 and @xmath171 produced by algorithm [ alg : sis ] satisfy @xmath172 & = p^{(\theta)}(y_{1:t } ) \int p^{(\theta)}(\alpha_{1:t}\mid y_{1:t } ) f(\alpha_{1:t } ) { \mathrm{d}}\alpha_{1:t}. \label{eq : ssm - proper}\end{aligned}\ ] ] for any function @xmath173 with the above integral well - defined . proposition [ prop : sis - proper ] follows from proposition [ prop : augmented - is ] , because @xmath174 are ratios of @xmath175 and @xmath176 , which is the law of the independent trajectories @xmath177 . let us consider next a particle filter algorithm @xcite ; see also the monographs @xcite , which is often much more efficient in practice . the algorithm is based on same user - supplied proposal distributions @xmath178 as sis , but requires also a ` resampling ' distribution @xmath179 , which draws a sample @xmath180 from a discrete probability mass @xmath181 . whenever the index ` @xmath182 ' appears in the algorithm , it takes values @xmath183 . [ alg : pf ] input : @xmath143 and @xmath161 . a. sample @xmath184 , set @xmath185 and calculate @xmath186 for @xmath147 , do a. sample @xmath187 , where @xmath188 b. sample @xmath189 and set @xmath190 . c. calculate @xmath191 if @xmath192 in algorithm [ alg : pf ] , the algorithm may be terminated immediately all the estimators considered in proposition [ prop : particle - proper ] below equal to zero then . we prefer to avoid such explicit stopping mechanism for mathematical convenience . we require the following well - known unbiasedness condition on the resampling procedure which is satisfied by the common multinomial , stratified , residual and systematic resampling methods ( cf . * ? ? ? [ a : resampling ] the resampling procedure satisfies @xmath193=m\bar{\omega}^{(j)}.\ ] ] for any @xmath194 and any probability mass vector @xmath181 . [ prop : particle - proper ] suppose that assumptions [ a : support ] and [ a : resampling ] hold , let @xmath161 and suppose @xmath173 is such that the integral in well - defined . consider the random variables generated in algorithm [ alg : pf ] , and let @xmath195 . a. [ item : filter - smoother ] the random variables @xmath164 where @xmath196 and @xmath197 satisfy . define for @xmath147 , and any @xmath198 , the backwards sampling probabilities @xmath199 and in case the sum is zero , set @xmath200 . a. [ item : backwards - sampling ] let @xmath201 be a random variable generated recursively by @xmath202 , and @xmath203 for @xmath204 . for any @xmath103 , the random variables @xmath205 satisfy , where @xmath206 and @xmath207 . b. [ item : fwd - bwd - smoothing ] if @xmath103 and @xmath208 for some @xmath209 , that is , @xmath210 is constant in all coordinates except @xmath211 and @xmath212 , then , the random variables @xmath213 satisfy ( with @xmath214 ) , where a. @xmath215 , b. @xmath216 , and where c. [ item : smoothing - weights ] @xmath217 and @xmath218 for @xmath219 . c. [ item : fwd - bwd - smoothing2 ] if @xmath103 and @xmath220 for some @xmath221 , then the random variables @xmath222 satisfy ( with @xmath214 ) , where @xmath223 and @xmath224 are defined in . we provide a self - contained but concise proof of proposition [ prop : particle - proper ] in appendix [ app : particle ] . proposition [ prop : particle - proper ] * corresponds to so - called filter - smoother @xcite . this property was shown in ( * ? ? ? * theorem 7.4.2 ) , although only in case of multinomial resampling . the same property is behind the original version of the particle marginal metropolis - hastings @xcite ; see also remark [ rem : proper - mcmc ] . + the statement holds also when the particle filter is applied with a general sequence of distributions @xmath225 rather than the ssm @xcite . * corresponds to backwards simulation , is analogous to the developments in conditional sequential monte carlo with backwards sampling @xcite . this allows us to draw any number of independent trajectories ( indices ) , and average them , in order to satisfy . * and its special case correspond to the forward - backward smoother @xcite ; see also @xcite . it can be seen as rao - blackwellised version of , but is applicable only when @xmath210 is a specific form . this scheme can lead to lower variance , but it also has square complexity in @xmath226 , rendering it inapplicable with large @xmath226 . we next note how propositions [ prop : sis - proper ] and [ prop : particle - proper ] ensure that both algorithms [ alg : sis ] and [ alg : pf ] yield a proper weighting scheme . [ cor : proper - ssm ] let @xmath227 be a markov chain which is harris ergodic with respect to @xmath37 , and let @xmath228 correspond to a. independent runs of algorithm [ alg : sis ] with parameters @xmath227 , respectively , or b. independent runs of a algorithm [ alg : pf ] with parameters @xmath227 , and the random variables defined in proposition [ prop : particle - proper ] or then , the random variables @xmath229 with @xmath230 provide a proper weighting scheme for target distribution @xmath231 ( assumption [ a : proper ] ) . similarly , if @xmath228 correspond to a. [ item : pf - marginal - inference ] independent runs of a algorithm [ alg : pf ] with parameters @xmath227 , and the random variables defined in proposition [ prop : particle - proper ] or , then @xmath229 provide a proper weighting scheme for the marginal distribution @xmath232 or @xmath233 , respectively . corollary [ cor : proper - ssm ] is stated for a single marginal ( pair ) , but it is clear that we may estimate simultaneously several marginal ( pairs ) , so that proposition [ prop : particle - proper ] applies for every function which is of the form @xmath234 and proposition [ prop : particle - proper ] whenever @xmath235 . it is possible to consider also more general classes of functions ; see for instance the discussion in @xcite . we next discuss one implication of proposition [ prop : particle - proper ] outside the main focus of this paper . namely , in the general state space smoothing context , that is , where @xmath161 is constant , proposition [ prop : particle - proper ] yields easily parallelisable particle smoothing algorithms , which also admit easily calculated consistent confidence interval estimates . [ prop : particle - smooth - confidence ] suppose @xmath161 is fixed , and let @xmath228 correspond to independent random variables as defined in proposition [ prop : particle - proper ] . a. if the corresponding conditions of proposition [ prop : particle - proper ] are satisfied , then the estimator @xmath236{n\to\infty } \mu_f { \mathrel{\mathop:}=}\int p^{(\theta)}(\alpha_{1:t}\mid y_{1:t } ) f(\alpha_{1:t}){\mathrm{d}}\alpha_{1:t},\ ] ] where @xmath237 , whenever the integral is well - defined . b. if also @xmath238\big)^2\big]<\infty$ ] , then @xmath239 \xrightarrow[\mathrm{d}]{n\to\infty } n(0,\sigma^2 ) , \qquad\text{where}\qquad \sigma^2 { \mathrel{\mathop:}=}\frac{\sigma_*^2}{p^{(\theta)}(y_{1:t})^2}.\ ] ] c. if in addition @xmath240<\infty$ ] , then @xmath241 , almost surely , where @xmath242\bigg)^2.\ ] ] the result of proposition [ prop : particle - smooth - confidence ] is straightforward and probably widely known , but we highlight it here because of the practical consequences it has for parallel particle smoothing . a. proposition [ prop : particle - smooth - confidence ] ensures that , under the given conditions , the confidence intervals @xmath243 $ ] are asymptotically consistent , where @xmath244 corresponds to the desirable standard gaussian quantile . b. calculation of consistent confidence intervals for one realisation of a particle smoothing algorithm is not a straightforward task in general . another recently suggested method @xcite relies on unbiased estimators obtained by coupling of conditional sequential monte carlo chains and randomisation as suggested by @xcite . c. if the particle filter estimates satisfy a variance asymptotic such as @xmath245 where @xmath246 correspond to a particle filter run with @xmath226 particles , then if @xmath226 is taken sufficiently large , the estimate as in proposition [ prop : particle - smooth - confidence ] is nearly as efficient than the single particle filter estimate with @xmath247 particles , but benefits from easy parallel implementation . d. even estimates which have quadratic cost ( in @xmath226 ) , as in proposition [ prop : particle - proper ] and , may be of practical interest in some scenarios . then , choosing an optimal @xmath226 leads to a non - trivial optimisation task . the asymptotic variance is a common efficiency criterion for markov chains . when a central limit theorem ( clt ) holds , the limiting variance coincides with the asymptotic variance . [ def : asvar ] suppose the markov chain @xmath34 on @xmath0 has transition probability @xmath14 which is harris ergodic with respect to invariant probability @xmath37 . for @xmath248 , the asymptotic variance of @xmath210 with respect to @xmath14 is @xmath249\bigg)^2,\ ] ] whenever the limit exists in @xmath250 $ ] , where @xmath251 stands for the _ stationary markov chain _ with transition probability @xmath14 , that is , with @xmath252 . [ def : clt ] the markov chain @xmath34 satisfies a clt for @xmath210 , if @xmath253 , and for any initial distribution , @xmath254 \xrightarrow{n\to\infty } n\big(0,{\mathrm{var}}(f , p)\big),\ ] ] in distribution . in case @xmath255 , the limit is a unit mass at zero . in order to guarantee a clt for a markov chain , some form of rate of convergence is necessary . we follow jones @xcite and define the following : [ def : convergence - rate ] suppose that a markov transition probability @xmath14 is harris ergodic with respect to a probability measure @xmath5 , and @xmath256 and @xmath257 are such that @xmath258 if @xmath259 , we say that @xmath14 admits @xmath260 convergence rate . different combinations of @xmath261 and @xmath262 imply a clt for different classes of functions ; see @xcite . we record the following summary of conditions which guarantee a clt for the augmented markov chain . the result follows easily from ( * ? ? ? * theorem 9 ) , given the following simple result whose proof can be found in appendix [ app : clt ] . [ lem : inherited - rate ] suppose that assumption [ a : aug ] holds , and the chain @xmath263 admits @xmath260 convergence rate . then , the chain @xmath57 admits a @xmath264 convergence rate , where @xmath265 . hereafter , we denote by @xmath266 the transition probability of @xmath267 and recall that @xmath268 is given in [ prop : aug - clt ] suppose that assumption [ a : aug ] holds , and @xmath34 satisfies definition [ def : convergence - rate ] with some @xmath260 . if @xmath269 for some @xmath270 and if one of the following hold : a. [ item : uniform ] @xmath271 and @xmath34 is uniformly ergodic , that is , @xmath272 , b. [ item : geometric ] @xmath273 for some @xmath274 and @xmath34 is geometrically ergodic , that is , definition [ def : convergence - rate ] holds with @xmath275 for some constants @xmath276 , c. [ item : geometric - rev ] @xmath271 and @xmath34 is geometrically ergodic and reversible , d. [ item : polynomial ] @xmath273 for some @xmath274 and @xmath34 is polynomially ergodic , @xmath277 with @xmath278 , and @xmath279 , then @xmath57 satisfies a clt for @xmath210 , with a finite limiting asymptotic variance @xmath280 where @xmath281 and @xmath282 are the conditional mean and variance , respectively : @xmath283 proposition [ prop : aug - clt ] is a consequence of lemma [ lem : inherited - rate ] , ( * ? ? ? * theorem 9 ) , and lemma [ lem : aug - prop ] and proposition [ prop : aug - asvar ] in appendix [ app : aug ] , because also reversibility of @xmath34 is inherited by @xmath57 . if @xmath284 , then @xmath285 for all @xmath274 . it follows that assuming @xmath286 in is enough to guarantee a clt for @xmath287 . note that the @xmath288-geometric ergodicity can be shown to be inherited easily as well ( * ? ? ? * lemma 45 ) . let us finally apply these results to the estimator introduced in proposition [ prop : proper - consistency ] . [ prop : proper - clt ] suppose conditions of proposition [ prop : proper - consistency ] are satisfied , which guarantee the strong consistency of @xmath66 given in , and suppose @xmath34 satisfies one of the conditions in proposition [ prop : aug - clt ] with @xmath289 . denote @xmath290.\ ] ] if @xmath291 , then @xmath292 \xrightarrow{n\to\infty } n(0,\sigma_f^2)\ ] ] in distribution , where @xmath293 and where @xmath294 and @xmath295 . proof of proposition [ prop : proper - clt ] is given in appendix [ app : clt ] . the latter term in the asymptotic variance expression of proposition [ prop : proper - clt ] contains the contribution of the importance sampling noise . if the estimators @xmath296 are made increasingly accurate , in the sense that @xmath297 becomes negligible , the limiting case corresponds to an importance sampling corrected approximate mcmc , and evaluating the conditional expectation @xmath298 . we conclude this section by noticing that the contribution of the importance sampling correction in asymptotic variance as in proposition [ prop : proper - clt ] can be estimated in a straightforward manner . [ prop : importance - var ] suppose conditions of proposition [ prop : proper - consistency ] are satisfied , which guarantee the strong consistency of @xmath66 given in . suppose also that @xmath299 where @xmath300 is defined in proposition [ prop : proper - clt ] , and @xmath301 , where @xmath302 $ ] . then , the estimator @xmath303 satisfies @xmath304 almost surely as @xmath305 . proof of proposition [ prop : importance - var ] is given in appendix [ app : clt ] . the estimator @xmath306 in proposition [ prop : importance - var ] provides a _ lower bound _ estimate for the clt variance @xmath307 of @xmath66 . it can provide useful information about the importance sampling noise contribution , and may be used as an optimisation criteria for the accuracy of the related estimators . estimation of the full asymptotic variance @xmath308 , which includes the marginal chain contribution @xmath309 , is not a straightforward task in general ; see for example @xcite and references therein . many mcmc algorithms include an accept - reject mechanism , which results in blocks of repeating values @xmath310 . when @xmath50 or @xmath311 are estimated , an obvious approach is to construct a single estimator which is used for the whole block . this may allow for variance reduction , for instance when replacing simple independent importance sampling with block - wide stratified estimator , or using particle filter estimators as in proposition [ prop : particle - proper ] , and choosing the number of particles proportional to the block length . to formalise such an algorithm we consider the `` jump chain '' representation of @xmath34 ( cf . [ def : jump ] suppose that @xmath34 is a markov chain with non - degenerate transition probability @xmath14 , that is , @xmath312 for all @xmath64 . the jump chain @xmath313 with corresponding durations @xmath314 of @xmath34 is defined as follows : @xmath315 and @xmath316 , and then recursively @xmath317 where @xmath318 . we note the following observations regarding definition [ def : jump ] : a. if @xmath14 is harris ergodic with respect to @xmath37 which is non - trivial ( that is @xmath319 for any @xmath64 ) , then necessarily @xmath312 for all @xmath64 , so definition [ def : jump ] is consistent . b. if @xmath34 corresponds to a metropolis - hastings chain , with non - diagonal proposal distributions @xmath320 ( that is , @xmath321 for every @xmath64 ) , then the jump chain @xmath322 is simply the accepted states , and @xmath323 is the number of rejections occurred at state @xmath322 . the importance sampling type estimator which we consider is of the following general form : [ a : block - super - general ] suppose that assumption [ a : mcmc - is ] holds , let @xmath313 denote the corresponding jump chain ( definition [ def : jump ] ) , and assume @xmath48 . suppose that @xmath324 are conditionally independent random variables on @xmath77 given @xmath313 , whose distribution depends only on the value of @xmath325 , and which satisfy @xmath326 = \frac{w_u(x)f(x)}{\alpha(x)}\qquad\text{and}\qquad { \mathbb{e}}[\tilde{w}_k\mid \tilde{x}_k = x ] = \frac{w_u(x)}{\alpha(x)},\ ] ] for @xmath37-almost every @xmath64 , and @xmath327\ ] ] satisfies @xmath328 . [ prop : block - super - general ] suppose assumption [ a : block - super - general ] holds , then @xmath329{n\to\infty } \pi(f).\ ] ] proposition [ prop : jump - properties ] implies that the jump chain @xmath322 corresponding @xmath89 is harris ergodic with invariant probability @xmath330 . the rest follows from proposition [ prop : super - general ] . we consider next the most obvious block estimator , based on @xmath314 , corresponding to the jump chain and assumption [ a : super - general ] . [ a : block - natural - general ] suppose that assumption [ a : mcmc - is ] holds , and let @xmath331 denote the corresponding jump chain ( definition [ def : jump ] ) . let @xmath332 correspond to a proper weighting scheme ( assumption [ a : proper ] ) , such that the variables in the scheme may depend on both @xmath325 and @xmath333 , and such that @xmath334 = w_u(x)f^*(x),\ ] ] for all @xmath103 and @xmath37-almost every @xmath64 . [ prop : block - natural - general ] suppose assumption [ a : block - natural - general ] holds , and @xmath335 \quad \text{satisfies } \quad \pi_a(b)<\infty . \label{eq : block - consistency - required}\ ] ] then , @xmath336{n\to\infty } \pi^*(f ) . \label{eq : block - natural}\ ] ] this follows from proposition [ prop : block - super - general ] , because the holding times @xmath337 are conditionally independent on @xmath322 and geometric with parameter @xmath338 , with expectation @xmath339 = 1/\alpha(x)$ ] . therefore , @xmath340 and @xmath341 satisfy assumption [ a : block - super - general ] . regarding assumption [ a : block - natural - general ] : a. the estimators @xmath342 and @xmath60 which satisfy assumption [ a : block - natural - general ] may often come from a proper weighting scheme as in assumption [ a : proper ] , but here the parameter @xmath226 is allowed to depend on @xmath343 . b. the condition on @xmath344 $ ] is not optimal . for instance , it is easy to find examples where the estimator is strongly consistent , even though the conditional expectation is unbounded in @xmath345 . in practice , however , the estimators are usually chosen either as independent of @xmath345 , or increasingly accurate in @xmath345 ; see remark [ rem : block - varying - or - not ] . c. the estimator of proposition [ prop : augmented - is ] coincides with a block estimator where @xmath346 . however , the block estimator offers more flexibility . for instance , @xmath347 could be stratified , leading to a smaller variance . d. even though we believe that the estimators of the form are often appropriate , we note that general form of the estimator as given in assumption [ a : block - super - general ] may be useful in some cases , in order to further reduce variance . for instance , assumption [ a : block - super - general ] accommodates the case where rao - blackwellised lower - variance estimators of @xmath348 are used instead of @xmath349 , as suggested in @xcite . let us then consider the asymptotic variance of the estimator . [ thm : block - clt ] suppose assumption [ a : block - natural - general ] holds , consider the estimator @xmath66 in and assume holds , @xmath350 , where @xmath351 , and @xmath352 , where @xmath353.\ ] ] if @xmath34 is @xmath288-geometrically ergodic , then @xmath354 \xrightarrow{n\to\infty } n(0,\sigma^2 ) \qquad\text{in distribution},\ ] ] where the limiting variance can be given as : @xmath355 , \label{eq : block - clt - var}\ ] ] where @xmath356,\ ] ] and @xmath14 stands for the markov transition probability of @xmath34 . proof of theorem [ thm : block - clt ] is given in appendix [ app : jump ] . [ rem : block - varying - or - not ] in theorem [ thm : block - clt ] asymptotic variance expression , the latter term corresponds to the contribution of noise in importance sampling estimates . a. when @xmath60 and @xmath357 correspond to @xmath333 independent estimates and @xmath358 , we might have @xmath359 , which leads to @xmath360 , so the contribution of the is would be upper bounded by @xmath361 . b. similarly , if @xmath362 , which could occur when @xmath93 are independent of @xmath333 , then , @xmath363 and the contribution of is would be upper bounded by @xmath364 . the acceptance probability is lower bounded for geometrically ergodic chains @xcite , rendering the estimator a safe choice . about the assumption of @xmath288-geometric ergodicity in theorem [ thm : block - clt ] : a. the asymptotic variance expression we derive for the ` numerator ' part of our estimator holds whenever @xmath34 is reversible , due to a result of deligiannidis and lee @xcite . this does not , however , imply a clt . b. our proof relies on the existence of a solution @xmath365 to the poisson equation @xmath366 . we believe that the jump chain @xmath313 inherits a central limit theorem from the base chain @xmath34 under general conditions , similarly as in section [ sec : asvar - clt ] . however , the jump chain @xmath313 does not inherit a convergence rate from @xmath34 in general , so a simple approach such as in section [ sec : asvar - clt ] does not apply . we start by making an explicit connection of proper weighting schemes with positive weights and pseudo - marginal type mcmc algorithms . [ a : proper - pseudo ] suppose that @xmath367 is a probability density on @xmath97 , @xmath368 is a family of proposal densities on @xmath0 and @xmath369 is a family of probability distributions on @xmath370^m\times[0,\infty)$ ] such that for almost every @xmath64 and all @xmath103 , the variables @xmath371 satisfy @xmath372 = \varpi(x ) \int \pi^*(z\mid x ) f(x , z ) { \mathrm{d}}z,\ ] ] where @xmath373 . [ alg : proper - pseudo ] suppose assumption [ a : proper - pseudo ] holds , let @xmath374 and @xmath375 and for @xmath102 iterate : a. draw @xmath376 and @xmath377 . b. with probability @xmath378 accept and set @xmath379 ; otherwise reject and set @xmath380 . [ prop : proper - pseudo ] consider algorithm [ alg : proper - pseudo ] . if @xmath381 is integrable and the markov chain @xmath57 is harris recurrent , then @xmath382{n\to\infty } \pi^*(f)\qquad\text{for every $ f\in l^1(\pi^*)$,}\ ] ] if and only if assumption [ a : proper - pseudo ] holds with @xmath383 . if @xmath381 is integrable , then @xmath384 defines an unnormalised probability on @xmath385 . define a composite proposal distribution @xmath386 , then it is straightforward to check that @xmath387 that is , algorithm [ alg : proper - pseudo ] is a metropolis - hastings with target @xmath388 , and therefore reversible with respect to @xmath389 . if the chain is harris ergodic , then for every @xmath390 , the strong law of large numbers holds . [ rem : proper - mcmc ] proposition [ prop : proper - pseudo ] connects proper weighting explicitly with the validity of pseudo - marginal type algorithms . this analogue also suggests that our is type correction may be applied instead of any pseudo - marginal method . the proper weighting formulation may also suggest variations of pseudo - marginal schemes . for example , proposition [ prop : convex - proper ] suggests about the possibility of using several independent particle filters in a particle marginal metropolis - hastings update , which may be of interest in parallel computing . let us then focus on delayed acceptance ( da ) mcmc @xcite , which is a multiple - stage acceptance method in metropolis - hastings algorithm , which is often based on an approximate target distributions in order to reduce computational effort of a metropolis - hastings algorithm . we focus in particular on the two - stage da scheme based on an approximation @xmath37 of the true probability density @xmath38 , which has obvious similarities with the importance sampling scheme suggested earlier . such da scheme with proposal densities @xmath368 can be summarised as follows : [ alg : da ] suppose @xmath391 and for @xmath102 , a. [ item : proposal ] generate a proposal @xmath392 b. [ item : first - stage ] with probability @xmath393 continue to step , otherwise reject and set @xmath394 . here , @xmath395 c. [ item : second - stage ] with probability @xmath396 set @xmath397 , otherwise reject and set @xmath398 . the steps and alone would implement a regular metropolis - hastings step targeting @xmath37 , whereas the second - stage acceptance step alone would implement the accept - reject step of independence metropolis - hastings algorithm with proposal @xmath37 . the rationale behind the da scheme is based on the same two assumptions which make the importance sampling corrected mcmc appealing : the evaluation of @xmath37 costs less than that of @xmath38 , and @xmath37 approximates @xmath38 reasonably well . the former directly affects the expected cost of the algorithm : if @xmath399 is rejected at step , there is no need to calculate @xmath400 , implying lower expected cost . the latter assumption guarantees that the second - stage acceptance rate is typically close to one . even though the da version is guaranteed to be worse than direct metropolis - hastings targeting @xmath38 in the peskun sense @xcite , the lower cost often outweighs the difference by allowing a greater number of samples to be drawn with da . we note that a proper weighting scheme ( assumption [ a : proper ] ) with non - negative weights leads to a valid da scheme . [ a : da - proper ] suppose assumption [ a : proper - pseudo ] holds with @xmath401 and a @xmath37 is a probability density on @xmath0 with @xmath40 . [ alg : da - proper ] suppose assumption [ a : da - proper ] holds , and @xmath402 are proposal densities as in algorithm [ alg : da ] . define the markov chain @xmath267 with initial values @xmath403 and @xmath404 , and for @xmath102 as follows : a. generate a proposal @xmath392 . b. [ item : first - stage2 ] with probability @xmath393 continue to step , otherwise reject and set @xmath405 . c. [ item : second - stage2 ] generate @xmath406 , then with probability @xmath407 accept and set @xmath408 , otherwise reject and set @xmath380 . [ prop : da - valid ] if assumption [ a : da - proper ] holds and @xmath402 are such that algorithm [ alg : da - proper ] defines a harris recurrent markov chain @xmath57 , then algorithm [ alg : da - proper ] satisfies @xmath409{n\to\infty } \pi^*(f ) , \qquad\text{for every $ f\in l^1(\pi^*)$.}\ ] ] the proof of proposition [ prop : da - valid ] follows from @xcite , because algorithm [ alg : da - proper ] is a valid da version of algorithm [ alg : proper - pseudo ] . proposition [ prop : da - valid ] highlights that when a da - mcmc scheme is of the pseudo - marginal form as algorithm [ alg : da - proper ] , the importance sampling type correction as we suggest is applicable , and a natural alternative for da - mcmc . the efficiency of da - mcmc was studied by banterle , grazian , lee and robert @xcite . sherlock , thiery and golithly @xcite studied the efficiency of pseudo - marginal with da as in algorithm [ alg : da - proper ] , with diffusion limit techniques . contrary to the is scheme , it is much harder to study the effect of noisy importance weights with da in general , as the noise affects the markov chain convergence rate . however , the authors @xcite are able to provide useful heuristic how to tune the accuracy of the pseudo - marginal da method . we now turn into an illustration of our generic framework with more specific state space models . section [ sec : lin - gauss - ssms ] discusses an important sub - class of state space models , where the state dynamics are linear - gaussian . we describe a general method which is based on a gaussian approximation of the conditional smoothing distribution . a special case of the general state space model introduced in section [ sec : ssm ] is a model where both @xmath142 and @xmath144 are euclidean , and the state dynamics @xmath410 are linear - gaussian , but the observational distribution @xmath411 may be non - gaussian and/or non - linear . more specifically , @xmath412 are defined through @xmath413 and @xmath414 our approach allows , for instance , exponential family observation models with gaussian , poisson , binomial , negative binomial , and gamma distributions , and a stochastic volatility model . we discuss the specific models later in sections [ poisson ] and [ sv - model ] . compared to the general state space form , this class of models may seem restrictive , but it still contains a large number of commonly used statistical models , such as ( generalised ) structural time series models , cubic splines , generalised linear models , and autoregressive integrated moving average ( arima ) models . the benefit of working within this subset of general state space models is that we have access to approximations taking advantage of the specific structure of the model . our approximation scheme , based on @xcite , relies on a laplace approximation @xmath415 of the conditional densities @xmath416 . the approximating gaussian model is found by using an iterative process closely related to iterative reweighted least squares algorithm used in generalised linear models context @xcite . in the approximating gaussian model , the observation equation is replaced by a linear - gaussian one @xmath417 where the pseudo - observations @xmath418 and their variances @xmath419 are based on the first and second derivatives of @xmath420 with respect to @xmath421 at @xmath422 , the conditional mode estimate of the approximate model @xcite . our approximation of the marginal density @xmath37 is based on the marginal likelihood decomposition @xmath423,\ ] ] where @xmath424 , the terms @xmath425 correspond to the product of pseudo - observation densities , and the expectation is taken with respect to @xmath426 . if the approximation @xmath427 is close to @xmath428 in the high - probability regions of @xmath426 , the expectation term is close to one , so we take @xmath429 our approximation defines , in fact , a joint approximate distribution @xmath430 , and we use @xmath431 as a proposal distribution in our importance sampling correction . we use the laplace approximation @xmath432 described in section [ sec : lin - gauss - ssms ] as the importance distribution , and we use the simulation smoothing algorithm @xcite to simulate antithetic pairs of @xmath141 from @xmath432 . we implemented the following inference algorithms in our experiments : amcmc : : approximate mcmc targeting @xmath433 , and for each accepted @xmath139 , sampling one realisation from @xmath434 with the simulation smoother . mcmc : : exact pseudo - marginal mcmc ( algorithm [ alg : proper - pseudo ] ) algorithm using importance sampling estimates with @xmath226 samples . da : : delayed acceptance version of the above ( algorithm [ alg : da - proper ] ) . is1 : : simple importance sampling estimator given in with @xmath226 samples . is2 : : blocked estimator as in , with is estimators based on @xmath226 samples , independent of @xmath435 . is - pf : : this is similar to is2 but uses the bootstrap particle filter @xcite ( algorithm [ alg : pf ] ) with @xmath226 samples , and with stratified sampling @xcite . the filter - smoother estimate ( proposition [ prop : particle - proper ] ) was used . in all the mcmc methods , the adaptive random walk metropolis algorithm of @xcite with gaussian proposal distribution was used . the target mean acceptance rate was set to 0.234 . in the delayed acceptance , the first - stage acceptance rate was optimised , and in the pseudo - marginal , the total acceptance rate was optimised . in all cases , the adaptive phase was terminated after the burn - in phase . all the experiments were conducted using the ` bssm ` @xcite package in ` r ` @xcite , which will be made available . we also experimented with the following algorithmic modifications : global approximation : : in the standard version , a laplace approximation is constructed for every new proposed value of @xmath139 . we experimented also faster but less accurate alternative , where we consider only one series of pseudo - observations constructed at the maximum likelihood estimate of @xmath139 . parallel computation : : we tested our implementation with different number of cores . the number of cores used in the experiment is indicated in parenthesis such as is2 ( 8) . sub - sampling for memory constraints : : depending on the problem , we might be interested in either computing some given summary statistics of @xmath141 , which can be performed recursively , or constructing a weighted sample from @xmath436 . in order to save in storage , we used sub - sampling as in proposition [ prop : randomise - proper ] for the latter . our main aim is to compare the performance of the da - mcmc with our is type correction schemes , using the same approximate marginal target distribution in both . note that , if parallel aspects are not taken into account , the is2 estimator has typically lower cost than da with same parameter @xmath226 . indeed , if the second - stage proposal would always be accepted , then da would require approximately the same number of computations as is2 . on the other hand , is1 has typically higher cost than da . because da usually has non - zero second - stage acceptance probability , da with @xmath226 samples has roughly the same cost as is1 using @xmath437 samples , where @xmath438 accounts for rejected second - stage proposals . because the noise in the weighting scheme affects the mixing of the mcmc and da , whereas is1 and is2 are based solely on the approximate marginal mcmc chain plus is type correction , the efficiency of the estimators is likely to be problem dependent . note also that mcmc and da methods require is steps also during the burn - in , whereas is1 , is2 and is - pf do not . we consider first the following special case of the model discussed in section [ sec : lin - gauss - ssms ] : @xmath439 with @xmath440 . for testing our algorithms , we simulated @xmath441 observations from this model with @xmath442 fixed to @xmath443 , and @xmath444 fixed to @xmath445 . based on our pilot experiments , the pseudo - marginal type mcmc performed well with this model using global marginal likelihood approximation , so we employed it for the experiments . we also observed that the number of importance samples @xmath226 could be set relatively low without affecting the mixing or accuracy much . we ran amcmc , mcmc , da , is1 , and is2 algorithms with @xmath446 , using 20,000 mcmc iterations with the first half discarded as burn - in . we used a uniform prior @xmath447 for the standard deviation parameters , where the cut - off parameter @xmath448 were set to @xmath449 based on the sample standard deviation of @xmath450 , where zeros were replaced with 0.1 . results were not sensitive to this upper bound . we performed two experiments . in the first , we stored only one realisation of @xmath141 at each iteration , and the resulting samples was used to calculate the posterior mean and variance estimates of @xmath139 , @xmath451 , and @xmath452 . in the second experiment , we used all realisations from the simulation smoother and computed the posterior mean and variance estimates of @xmath451 and @xmath452 recursively . tables [ table : poisson ] and [ table : poisson_summary ] summarise the results over 1000 repeated experiments as discussed above . in table [ table : poisson ] , the approximate method gave slightly biased estimates for all variables , but with much less cost than exact mcmc . delayed acceptance decreased the computation costs by about two thirds , while the is2 ( with same @xmath226 and no parallelisation ) almost halves the costs of da . parallel computing with eight cores further decreased the computation time of is - corrected methods , but because of small @xmath226 , the speed - up was limited , as it is impossible to get much lower than with amcmc . using @xmath226 proportional to the block size ( is1 ) does not seem to bring significant benefits in this experiment . overall , the main differences in da and is corrected methods are related to computational time . the parameter estimates and their standard errors are comparable . the pseudo - marginal mcmc provided perhaps slightly smaller standard errors , but due to the much higher computational time , the da and is2 appear preferable in this scenario . .means and the corresponding standard errors from 1000 mcmc runs for poisson model with 10 runs of simulation smoother per iteration , with one stored sample path . [ cols=">,>,>,>,>,>,>",options="header " , ] as an illustration of the estimators in proposition [ prop : particle - smooth - confidence ] , we simulated one realisation @xmath441 from a linear - gaussian ssm of the form @xmath453 with @xmath454 and @xmath455 . given @xmath441 , we estimated @xmath139 via maximum likelihood , resulting @xmath456 and calculated the exact smoothing means @xmath457 $ ] for @xmath458 . we performed particle smoothing using the filter - smoother ( fs ) and forward - backward smoother ( fbs ) ( proposition [ prop : particle - proper ] and , respectively ) , and repeated these computations @xmath459 times , from which we obtained 95% confidence intervals as in proposition [ prop : particle - smooth - confidence ] . figure [ fig : ci_plot ] shows confidence intervals from both methods with varying number of particles . for fbs , we used @xmath460 , and for fs , the corresponding squares @xmath461 , making the computational times roughly comparable . we note that in our implementation , fbs was about 1.4 times slower than fs on average . it appears that in this experiment , fs outperformed fbs with comparable computational cost . $ ] with @xmath462 , using forward - backward smoother algorithm ( fbs ) , and filter - smoother algorithm ( fs ) with varying number of particles ( note that algorithms use different amount of particles ) . red line corresponds to the exact value computed by kalman smoothing . ] because fbs may be of interest in some other scenarios , we attempted to optimise the choice of @xmath226 for the fbs . we ran the fbs with varying choices of @xmath226 as above , and using @xmath463 , respectively . figure [ fig : fbs_time ] shows the square root of average variance estimates @xmath306 of proposition [ prop : particle - smooth - confidence ] over time points @xmath160 , multiplied by the square root of computation time in minutes for different @xmath226 . in this experiment , @xmath464 or @xmath465 appeared good choices , but more importantly , the experiment suggests that the choice of @xmath226 in fbs may be a non - trivial optimisation problem . . we suggested to use mcmc targeting a marginal of the posterior distribution and consequent importance sampling type estimation in order to perform easily parallelisable consistent inference over the full posterior distribution . we focused on state space models ( ssms ) , where importance sampling and particle filters provide natural and efficient correction mechanisms . our specific application was based on a laplace approximation of ssms with linear - gaussian state dynamics , but arbitrary observation distributions . our experiments are promising , and we believe that our methodology has potential to make exact inference feasible with a large class of models where only approximate inference is currently possible . alternative approximation methods in the state space context include simple linearisation - based methods such as the extended kalman filter , moment matching , or variational approximations ; see the recent monograph @xcite . gaussian latent variable models beyond the state space context could be handled with the laplace approximations , and more refined approaches such as the integrated nested laplace approximation @xcite could provide a useful approximate marginal . we note that , unlike with purely approximate inference , the approximate posterior @xmath37 does not need to be very accurate . we suggested importance ( is ) sampling type correction with sequential is and sequential monte carlo ( smc ) . we note that various smc variations , such as rao - blackwellised smc , and alternative resampling strategies , apply directly @xcite . randomised sequential quasi - monte carlo @xcite can be useful , when applicable . smc samplers can be useful beyond the state space context @xcite . the correction could be based on exact sampling for diffusions @xcite , and debiasing tricks and multilevel monte carlo @xcite can be useful for instance when the true target distribution is based on a continuum model , and only discretisations are accessible . importance sampling is known to be generally hard to implement efficiently in higher dimensions , and our setting is no exception . the fact that the approximation @xmath37 is required only on a marginal distribution may help , if the model admits a natural smaller dimensional parameter space , as our state space application . indeed , in the state space context our approach relies on similar assumptions which make particle mcmc @xcite appealing , with the added assumption of the existence of suitable marginal approximation @xmath37 . if such approximation is available , we suggest to take advantage of it , as an alternative to particle mcmc . one potential issue of importance sampling correction , in general , is when the importance weight @xmath46 is unbounded . in our experiments , we did not observe large values of importance weights , but this is likely due to the fact that our space @xmath466 was bounded . in case @xmath0 is unbounded , it may be that the ratio @xmath467 is unbounded , even if @xmath88 is a reasonable approximation in the ` centre ' of @xmath0 . because bounded @xmath468 is generally desirable , we suggest tempering as a possible remedy . that is , if @xmath37 is our original approximate marginal , the chain @xmath34 may be designed to target @xmath469 instead , where @xmath470 is a tempering parameter . it may be of practical interest to construct the marginal chain @xmath34 using adaptive mcmc techniques such as @xcite or as discussed in the review @xcite . we note that our theory does not apply directly with adaptive mcmc , unless the adaptation is stopped after suitable burn - in , like we do in our experiments . an obvious alternative to our approach is that , instead of running one markov chain targeting @xmath37 and then calculating conditional importance sampling type estimates , one may consider running multiple independent pseudo - marginal type mcmc methods in parallel . we note that while valid , this approach may suffer from bias if the mcmc chains are not run long enough , and burn - in ( and possible adaptation ) of each chain could be seen wasteful . it is well - known that pseudo - marginal methods may be slow mixing when the importance weights have large variation in the extreme case they may lose desirable rates of convergence such as geometric ergodicity @xcite . large variation of importance weights affects also our method , but we believe that mcmc targeting the approximate target @xmath37 may be often easier to design in an efficient manner . the authors were supported by an academy of finland research fellowship ( grants 274740 and 284513 ) . we thank christophe andrieu for insightful remarks . throughout this section , suppose that @xmath471 is a markov transition probability on @xmath0 and @xmath472 is a markov transition probability from @xmath0 to a space @xmath52 . we consider here properties of an augmented markov transition probability @xmath473 defined on @xmath58 as follows : @xmath474 we first state the following basic result . [ lem : aug - prop ] the properties of @xmath471 and the augmented chain @xmath473 are related as follows : a. [ lem : aug - mim - irr ] let @xmath475 denote the set of @xmath476-irreducibility measures of a markov transition probability @xmath471 , then @xmath477 b. [ lem : aug - mim - mim ] the implications in hold when @xmath475 and @xmath478 are replaced with sets of maximal irreducibility measures of @xmath471 and @xmath473 , respectively . c. [ item : aug - invar ] the invariant probabilities of @xmath471 and @xmath473 satisfy : @xmath479 the above implications hold also with invariance replaced by reversibility . d. [ item : aug - harris ] @xmath471 is harris recurrent if and only if @xmath473 is harris recurrent . e. [ item : aug - iterates ] suppose @xmath480 is measurable and such that @xmath481 and @xmath482 are well - defined . then , for any @xmath483 , @xmath484 the inheritance of irreducibility measures ( [ lem : aug - mim - irr ] ) , maximal irreducibility measures ( [ lem : aug - mim - mim ] ) , invariant measures , and reversibility are straightforward . for harris recurrence , let the probability @xmath485 be a maximal irreducibility measure for @xmath471 , then @xmath486 is the maximal irreducibility measure of @xmath473 . let @xmath487 with @xmath488 , and choose @xmath489 such that @xmath490 , where @xmath491 with @xmath492 . notice that @xmath493 where @xmath494 are the hitting times of @xmath89 to @xmath495 . this concludes the proof because @xmath496 are independent bernoulli random variables with success probability at least @xmath497 . the converse statement is similar . for , it is enough to notice that for any @xmath498 and @xmath483 , it holds that @xmath499 . we next state the following generic result which gives an expression for the asymptotic variance of an augmented markov chain . [ prop : aug - asvar ] let @xmath500 and define the conditional mean @xmath501 and the conditional variance @xmath502 , then @xmath503 whenever @xmath504 is well - defined . let @xmath267 be the stationary markov chain with transition probability @xmath473 . @xmath505\bigg ) \\ & = \check{\mu}(h^2 ) + \frac{2}{n } \sum_{i=1}^{n-1 } \sum_{\ell=1}^{n - i } { \mathbb{e}}[h(x_0,s_0)h(x_\ell , s_\ell)],\end{aligned}\ ] ] by stationarity . we may write by lemma [ lem : aug - prop ] , @xmath506 & = \int \check{\mu}({\mathrm{d}}x\times { \mathrm{d}}s ) h(x , s ) ( k^\ell m_h)(x ) \\ & = { \mathbb{e}}[m_h(x_0 ) m_h(x_\ell)].\end{aligned}\ ] ] we deduce that @xmath507 where @xmath508 . the claim follows by taking limit @xmath305 . we finally state a simple result about the relationship of the solutions of the poisson equation for augmented chains , used in appendix [ app : jump ] . [ lem : aug - poisson ] suppose that @xmath500 . if there exists @xmath509 such that @xmath510 , where @xmath511 , then @xmath512 it is clear that @xmath513 and @xmath514 , so @xmath515 we follow @xcite and prove the statement assuming that the resampling scheme @xmath516 satisfies the following property : @xmath517 for all @xmath518 . we can do this without loss of generality , because this condition would be satisfied if we applied an independent random permutation of indices to the resampling scheme satisfying assumption [ a : resampling ] . the statement is valid without such randomisation , because both the algorithm and the statements are independent of ordering . for , define the filtrations @xmath519 and @xmath520 , and functions @xmath521 , and for @xmath522 @xmath523 clearly @xmath524 , and all @xmath525 are ( almost everywhere ) well - defined if the latter integral is well - defined . let us first observe that for @xmath147 , @xmath526 & = { \mathbb{e}}\big [ { \mathbb{e}}\big [ \omega_t^{(i ) } f_t(\bar{\alpha}_{t-1}^{(a_{t-1}^{(i)})},\alpha_t^{(i ) } ) { \mathrel{\big|}}{\mathcal{g}}_{t-1 } \big ] { \mathrel{\big|}}{\mathcal{f}}_{t-1 } \big ] \label{eq : unbiased - step } \\ & = { \mathbb{e}}\big [ f_{t-1}(\bar{\alpha}_{t-1}^{(a_{t-1}^{(i ) } ) } ) { \mathrel{\big|}}{\mathcal{f}}_{t-1 } \big ] \notag \\ & = \sum_{i=1}^m \bar{\omega}_{t-1}^{(i ) } f_{t-1}(\bar{\alpha}_{t-1}^{(i ) } ) . \notag\end{aligned}\ ] ] by applying recursively , we obtain @xmath527 & = { \mathbb{e}}\bigg[\bigg(\prod_{t=1}^{t-1 } \frac{1}{m}\omega_t^ * \bigg ) \frac{1}{m } \sum_{i=1}^m { \mathbb{e}}\big [ \omega_t^ * \bar{\omega}_t^{(i ) } f_t(\bar{\alpha}_{t-1}^{(a_{t-1}^{(i)})},\alpha_t^{(i)}){\mathrel{\big|}}{\mathcal{f}}_{t-1}\big]\bigg ] \\ & = { \mathbb{e}}\bigg[\bigg(\prod_{t=1}^{t-1 } \frac{1}{m } \omega_t^*\bigg ) \sum_{i=1}^m \bar{\omega}_{t-1}^{(i ) } f_{t-1}(\bar{\alpha}_{t-1}^{(i ) } ) \bigg ] \\ & = \frac{1}{m}\sum_{i=1}^m { \mathbb{e}}\big [ \omega_1^ * \bar{\omega}_t^{(i ) } f_1(\alpha_1^{(i)})\big],\end{aligned}\ ] ] which equals to @xmath528 by a similar calculation as in . the statement is equivalent with @xmath529 = \int p^{(\theta)}(\alpha_{1:t},y_{1:t } ) f(\alpha_{1:t } ) { \mathrm{d}}\alpha_{1:t}. \label{eq : all - trajectories - proper}\ ] ] similarly as above , we have for @xmath522 , @xmath530 \\ & = \int \sum_{j=1}^n \bar{\omega}_{t-1}^{(j ) } \mu_t(\alpha_t\mid \alpha_{t-1}^{(j ) } ) g_t(y_t\mid \alpha_t ) \frac { \bar{\omega}_{t-1}^{(i_{t-1 } ) } \mu_t(\alpha_t\mid \alpha_{t-1}^{(i_{t-1 } ) } ) } { \sum_{\ell=1}^m \bar{\omega}_{t-1}^{(\ell ) } \mu_t(\alpha_t\mid\alpha_{t-1}^{(\ell ) } ) } f_t(\alpha_{1:t}^{(i_{1:t } ) } ) { \mathrm{d}}\alpha_t \\ & = \bar{\omega}_{t-1}^{(i_{t-1 } ) } f_{t-1}(\alpha_{1:t-1}^{(i_{1:t-1})}),\end{aligned}\ ] ] and application of this recursively leads to . the result follows also from , by first summing over the indices @xmath531 . the last result follows as a special case of . consider the artificial trivial markov chain @xmath89 with @xmath532 for @xmath102 which is harris ergodic with respect to @xmath533 , with @xmath534 , and let @xmath535 . then , @xmath536 and @xmath537 with @xmath538 forms a proper weighting scheme for @xmath539 . the strong consistency follows from proposition [ prop : proper - consistency ] and the central limit theorem from proposition [ prop : proper - clt ] , because @xmath34 is trivially uniformly ergodic and @xmath540 . note that @xmath541 implying @xmath542 . similarly , the convergence of @xmath543 follows from proposition [ prop : importance - var ] . let @xmath544 $ ] , then @xmath545 where @xmath546 $ ] . this implies that @xmath547 whenever @xmath548 , we may write @xmath354 = \frac{n^{-1/2 } \sum_{k=1}^n w_k \sum_{i=1}^n v_k^{(i ) } \bar{f}(x_k , z_k^{(i)})}{n^{-1 } \sum_{j=1}^n w_j}. \label{eq : clt - form}\ ] ] the denominator converges to @xmath42 almost surely , so by slutsky s lemma it is enough to show that the numerator converges in distribution to @xmath549 . this follows from proposition [ prop : aug - clt ] , because @xmath291 . for @xmath345 large enough such that @xmath550 for @xmath551 , we may write @xmath552 the denominator converges to @xmath553 , and the numerator can be written as @xmath554,\ ] ] and @xmath555 . the term @xmath556 , and because @xmath557 , the remainder terms @xmath558 and @xmath559 , as @xmath560 . the following proposition complements ( * ? ? ? * lemma 1 ) and @xcite , which are stated for more specific cases . [ prop : jump - properties ] suppose @xmath89 is a markov chain with non - degenerate transition probability @xmath471 and @xmath322 its jump chain with corresponding holding times @xmath561 ( definition [ def : jump ] ) . then , the following hold : a. [ item : jump - trans ] @xmath322 is a markov chain with transition probability @xmath562 b. [ item : jump - joint - trans ] the holding times @xmath561 are conditionally independent given @xmath322 , and each @xmath333 has geometric distribution with parameter @xmath563 . c. [ item : jump - invar ] if @xmath471 admits invariant probability @xmath564 , then @xmath565 admits the invariant probability @xmath566 in addition , if @xmath471 is reversible with respect to @xmath5 , then @xmath565 is reversible with respect to @xmath567 . d. [ item : psi - irreducibility ] @xmath89 is @xmath568-irreducible if and only if @xmath322 is @xmath568-irreducible , with the same maximal irreducibility measure . e. [ item : harris - recurrence ] @xmath89 is harris recurrent if and only if @xmath322 is harris recurrent . the expression of the transition probability is due to straightforward conditioning , and was observed in @xcite . the invariance follows from @xmath569 \\ & = \frac{1}{\mu(a)}\bigg [ \mu(a ) - \int_a \mu({\mathrm{d}}x)\big(1-a(x)\big)\bigg ] = \tilde{\mu}(a ) , \end{aligned}\ ] ] and the reversibility is shown in @xcite . for it is sufficient to observe that @xmath570 which holds because the sets @xmath571 and @xmath572 coincide . similarly , holds because @xmath573 where @xmath574 and @xmath575 . from the expression of the jump chain transition probability in proposition [ prop : jump - properties ] , we obtain for any @xmath582 with @xmath583 well - defined , @xmath584 consequently , we observe that @xmath585 implying . because @xmath579 , lemma [ lem : poisson - variance ] and a straightforward calculation yield @xmath586 , \end{aligned}\ ] ] where @xmath587 . similarly , by lemma [ lem : poisson - variance ] , @xmath588 which allows us to conclude . [ lem : poisson - variance ] suppose the markov transition probability @xmath471 on @xmath0 is harris ergodic with respect to @xmath5 and @xmath589 . if there exists @xmath509 which solves the poisson equation , @xmath590 then @xmath591 and the markov chain @xmath34 following @xmath471 satisfies a clt . let @xmath14 , @xmath592 and @xmath266 be the markov transition probabilities of @xmath34 , @xmath313 and @xmath593 , respectively , and let @xmath37 , @xmath594 and @xmath268 be the corresponding invariant probabilities . because @xmath34 and is @xmath288-geometric and @xmath595 , there exists a solution of poisson s equation @xmath596 such that @xmath597 . geometric ergodicity of @xmath34 implies also @xmath598 for @xmath37-almost every @xmath64 @xcite , so @xmath599 . proposition [ prop : jump - variance ] implies that @xmath600 , and @xmath601 . denote then @xmath602 which has conditional mean @xmath603 , and @xmath604 is finite because @xmath605 \le \frac{2-\alpha(x)}{\alpha^2(x ) } m_\xi^{(2)}(x).\ ] ] lemma [ lem : aug - poisson ] therefore implies that there exists @xmath606 such that @xmath607 , and consequently lemma [ lem : poisson - variance ] implies that @xmath593 satisfies a clt for @xmath608 . for the expression of the asymptotic variance , proposition [ prop : aug - asvar ] applied with @xmath609 , @xmath610 , @xmath611 and @xmath612 , implies @xmath613 where , by the variance decomposition formula , @xmath614{\mathrel{\big|}}\tilde{x}_k = x\big ) \\ & = v_{n\xi}(x ) + w_u^2(x ) \bar{f}^{*2}(x ) \big(1-\alpha(x)\big)/\alpha^2(x).\end{aligned}\ ] ] proposition [ prop : jump - variance ] implies that @xmath615.\ ] ] if we write the scaled estimator analogously as in , we notice that the clt holds for the numerator , with asymptotic variance @xmath616/\pi_a(\alpha).\ ] ] the denominator converges to @xmath617 , so we conclude with slutsky s lemma . a. beskos , o. papaspiliopoulos , g. o. roberts , and p. fearnhead . exact and computationally efficient likelihood - based estimation for discretely observed diffusion processes . , 68(3):333382 , 2006 . ( with discussion ) .

it is by now well established that baryonic matter represents only about @xmath1 of the total matter density of the universe ( e.g. * ? ? ? * ) , while the majority is in the form of non - baryonic dark matter ( dm ) . since galaxies form through the accretion of baryonic material onto dynamically dominant dm structures ( halos ) , it would be reasonable to assume that the baryon mass fraction of present day galaxies approximately equals the cosmic value ( @xmath2 ) . despite this expectation , observations point to the fact that galaxies are not able to retain their cosmic `` fair share '' of baryons , and that the resulting baryon deficit depends strongly on the mass of their host halo . the first line of evidence is provided by observational estimates of the abundance of galaxies as a function of their total stellar mass , a distribution referred to as the galactic stellar mass function ( smf ) . thanks to the advent of wide area optical surveys with multiband photometric and spectroscopic information , such as the two degree field galaxy redshift survey ( 2dfgrs ) and the sloan digital sky survey ( sdss ) , the smf has been measured over the mass range @xmath3 , using statistical samples of tens of thousands of galaxies and a variety of stellar mass estimation techniques ( @xcite , @xcite , @xcite , @xcite , @xcite , @xcite , @xcite , to name a few ) . the smf displays an exponential cutoff at masses @xmath4 and an approximate power - law behavior at low masses ( @xmath5 ) , with a `` shallow '' exponent of @xmath6 . on the other hand , the halo mass function ( hmf ) predicted in the lambda cold dark matter ( @xmath0cdm ) model , follows a much `` steeper '' power - law ( @xmath7 ) over the mass range of interest . this observation alone excludes the possibility that the stellar mass of a galaxy is simply a fixed fraction of the host halo mass . one can furthermore statistically derive an average relation between the stellar mass of a galaxy ( @xmath8 ) and the mass of its host halo ( @xmath9 ) , through the technique of abundance matching ( see [ sec : am ] for details ) . @xmath8 - @xmath9 relations based on abundance matching ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) have shown that the `` stellar conversion efficiency '' , @xmath10 , never exceeds 25 - 30% . furthermore , @xmath11 peaks for milky way - sized galaxies ( @xmath12 ) , and declines rapidly on either side of the peak ( e.g. figure 2 in @xcite ) . the second line of evidence comes from direct halo mass measurements , obtained through weak lensing or kinematics studies ( e.g. * ? ? ? * ; * ? ? ? for example , @xcite used stacked weak lensing measurements to estimate the average host halo mass of galaxies in different stellar mass bins , and found that @xmath11 never exceeds @xmath13% . direct halo mass measurements can circumvent a number of assumptions inherent in the application of abundance matching , but such techniques can presently only be applied to a restricted range of stellar mass ( @xmath14 ) , and are affected by their own set of systematics . stellar mass is not always the dominant baryonic component in a galaxy . in fact , the hi to stellar mass ratio ( `` hi fraction '' ; @xmath15 ) tends to increase with decreasing stellar mass , and hi often dominates the baryonic content of low - mass galaxies . the transition from stellar - mass dominated to hi dominated systems takes place at @xmath16 for hi - selected samples ( e.g. * ? ? ? * see also fig . [ fig : gsmr ] in this work ) , or at @xmath17 for optically - selected samples ( e.g. * ? ? ? as a result , it is presently not clear what is the behavior of the `` baryon retention fraction '' @xmath18 in low - mass galaxies , when both stars and cold gas are taken into account . in particular it is not well understood whether the very low average value of @xmath11 inferred for low - mass halos is a result of poor retention of baryonic material , of the low efficiency of gas - to - stars conversion , or of a combination of both . for example , @xcite argue that the increasing gas fraction in low - mass galaxies should approximately offset the decreasing stellar - to - halo mass ratio , and result in a roughly constant @xmath19% . this conclusion was based on an indirect estimate of the cold gas content of galaxies , based on the average @xmath20 relation observed in a set of samples in the literature . an early work by @xcite , based on the same indirect method , also reached a qualitatively similar conclusion . @xcite found an approximately constant @xmath21 at the low - mass end using a different indirect method , which involves the comparison of the stellar and hi mass distributions of two different galaxy samples . these results would imply that low - mass galaxies are relatively efficient at retaining baryonic mass , but very inefficient in converting their gas into stars . this conclusion , however , would require a `` steep '' hi mass function ( himf ) in the local universe , in contrast to what is measured @xcite . moreover , the recent work of @xcite , also based on using the average @xmath20 relations for blue and red galaxies separately , found no signs for a flat @xmath21 at low masses . in this article we _ directly _ measure the abundance of galaxies as a function of their `` baryonic mass '' ( throughout this article the term baryonic refers to the combined stellar and atomic gas components of galaxies , and baryonic mass is calculated as @xmath22 , where the 1.4 factor accounts for the presence of helium ) . we use optical data from the seventh data release of the sdss ( sdss dr7 ) to estimate stellar masses , and hi - line flux measurements from the arecibo legacy fast alfa ( alfalfa ) survey to measure atomic gas masses . the resulting distribution , referred to hereafter as the baryon mass function ( bmf ) of galaxies , can be used in abundance matching to derive a robust @xmath21 - @xmath9 relation . in order to investigate sample selection effects , we employ both an hi - selected and an optically - selected sample drawn from the _ same _ volume to derive the mass distributions for the stellar , atomic hydrogen and baryonic components . the paper is organized as follows : in section [ sec : data ] , we introduce the datasets used to measure the stellar , hi and baryon mass distributions . we describe the methodology used to measure atomic hydrogen masses and we estimate stellar masses for our galaxy samples . in section [ sec : bmf ] , we present our measurements of the smf , himf & bmf from both the hi - selected and the optically - selected samples , and compare them against one another as well as against other published results . in section [ sec : biases ] we consider the impact of possible systematics on our measurements , such as stellar mass estimation method , distance uncertainties and the exclusion of some baryonic components ( e.g. molecular gas ) in the calculation of the bmf . in section [ sec : bar_frac ] , we present the @xmath11 - @xmath9 and @xmath21 - @xmath9 relation in a @xmath0cdm universe . in section [ sec : conclusion ] , we discuss the implications of the result and summarize our main conclusions . throughout this paper , we use a hubble constant of @xmath23 . we select galaxies from the current data release of the alfalfa survey , which covers 40% of the planned final survey area ( `` @xmath24.40 '' catalog ; @xcite ) . we restrict ourselves to two rectangular areas of the `` spring '' coverage of @xmath24.40 ( @xmath25 , @xmath26 & @xmath27 ) , which encompass the virgo cluster as well as the supergalactic plane at low velocities . we restrict ourselves to galaxies with @xmath28 15000 km s@xmath29 ( @xmath30 ) , in order to avoid the strong radio frequency interference ( rfi ) present at frequencies that correspond to @xmath31 km s@xmath29 . we discard the nearest extragalactic sources with @xmath32 mpc , because they can carry extreme fractional uncertainties on their distances ( see [ sec : derived ] for details on the distance assignment method ) . we furthermore select only hi sources designated as `` code 1 '' in @xmath24.40 , i.e. extragalactic sources detected at high significance ( @xmath33 ) . in addition , we exclude sources with integrated fluxes below the 50% completeness limit of the alfalfa survey ( see @xcite , section 6 for the derivation of the alfalfa completeness limits ) . the above requirements are satisfied by 7618 galaxies . we remove from our sample 204 @xmath24.40 sources which are not crossmatched with an optical source in sdss , as well as 208 additional sources which have been flagged as having problematic sdss photometry ( crossmatch code `` p '' in @xmath24.40 , for details see section 4 in @xcite ) . this quality cut on the sdss photometry introduces some bias against faint , low surface brightness galaxies of irregular morphology ; such sources are often `` shredded '' ( i.e. assigned multiple photometric objects ) by the sdss magnitude extraction process , and are usually assigned a `` p '' ( `` photometry suspect '' ) crossmatch code in @xmath24.40 . lastly , 11 additional objects were discarded , in cases where the stellar mass computation method described in [ sec : derived ] failed to produce physically plausible results . our final sample thus consists of 7195 extragalactic objects , detected over @xmath342000 deg@xmath35 of high galactic latitude sky and out to @xmath36 mpc . the upper panel of figure [ fig : coneplot ] displays the spatial distribution of our hi - selected galaxies , and puts in evidence the complex large scale structure in the survey volume . note that all objects in our hi - selected sample have 21 cm redshifts and line fluxes as well as multi - band optical photometry , and hence estimates of both their stellar and atomic hydrogen masses . we draw an optically - selected sample from the sdss dr7 @xcite spectroscopic database , in the same volume used to define our hi - selected sample . specifically , we select galaxies that lie within the same sky area ( @xmath25 , @xmath26 & @xmath27 ) and satisfy the same velocity and distance restrictions ( @xmath37 15000 km s@xmath29 , @xmath38 mpc ) . cmb velocities for our optically - selected galaxies are calculated based on their sdss spectroscopic redshifts ( @xmath39 ) . we restrict ourselves to objects spectroscopically classified as galaxies in sdss ( ` specclass = 2 ` ) that also have an apparent petrosian magnitude brighter than 17.5 in the @xmath40-band ( @xmath41 ) . this initial cut results in 22707 galaxies . due to their large number , it is not practical to inspect all galaxies individually for the quality of their sdss photometry / spectroscopy . as a result , we expect a fraction of our sources to have issues with their sdss photometry , most often related to `` shredding '' ( i.e. assignment of multiple photometric objects to a single galaxy ) . this issue affects mostly extended sources with structure in their light distribution , such as low surface brightness ( lsb ) galaxies with irregular morphology . in such cases , the sdss magnitude will underestimate the true flux of the galaxy , which in turn will result in an underestimate of its stellar mass . a second issue related to shredding , is that bright star forming knots in the disks of nearby spiral galaxies can sometimes be cataloged as separate spectroscopic objects , and hence interpreted as low - mass satellites of the main spiral . we find that applying a color cut on our sample , @xmath42 , removes a fair fraction of these unwanted cases . on the other hand , cuts based on the quality of the sdss spectrum ( such as cuts on ` zconf ` , ` zstatus ` or ` zwarning ` ) are ineffective , since they exclude mostly legitimate faint or lsb dwarf galaxies with noisy spectra . lastly , we exclude objects for which the stellar mass computation described in [ sec : derived ] failed to produce physically plausible results . our final optically - selected sample consists of 22587 galaxies , occupying the same volume as our hi - selected sample . we crossmatch the optical sample with the full @xmath24.40 catalog ( including code 1 & 2 sources ) , and find 7551 hi source counterparts . the crossmatch rate is thus approximately @xmath43 , as reported in @xcite . note that alfalfa non - detected galaxies are not necessarily hi - poor objects ; due to the low emissivity of atomic hydrogen in the 21 cm line , even moderately gas - rich galaxies in the outer portion of the survey volume can be missed by alfalfa . this point is illustrated by the lower panel of figure [ fig : coneplot ] , which compares the spatial distribution of galaxies in the hi - selected and optically - selected samples . note that all optically - selected galaxies have multiband optical photometry as well as optical redshifts , and hence an estimate of their stellar mass . however , only galaxies crossmatched with an @xmath24.40 source have a 21 cm flux measurement , and hence a value for their atomic hydrogen mass . -band absolute magnitude ( _ panel a _ ) , @xmath44 color ( _ panel b _ ) , stellar mass ( _ panel c _ ) and hi mass ( _ panel d _ ) , for the optically - selected ( _ red solid line _ ) and hi - selected ( _ blue dashed line _ ) samples . as evident in panel _ b _ , the hi - selected sample is strongly biased against the red galaxy population and as a result it is skewed towards lower luminosity and stellar mass systems . conversely , the fractional contribution of bright and massive galaxies ( @xmath45 ) is larger for the optically - selected sample ( panels _ a _ and _ c _ ) . ] we calculate hi masses from the measured 21 cm integrated flux reported in @xmath24.40 . assuming optically thin emission @xmath46 where @xmath47 is the hi mass in units of the solar mass ( @xmath48 ) , @xmath49 is the integrated flux in jy@xmath50km@xmath51s@xmath29 and @xmath52 is the distance in mpc . distances in this article are calculated according to the method used in @xmath24.40 @xcite : nearby galaxies ( @xmath53 km s@xmath29 ) are assigned distances through the use of a peculiar velocity flow model developed by @xcite , while for more distant galaxies simple hubble distances are used ( @xmath54 , with @xmath55 km@xmath51s@xmath29mpc@xmath29 ) . moreover , group and cluster membership information is taken into account when available , as well as primary distance measurements published in the literature . we would like to point out that most of the galaxies in our optically - selected sample are not included in @xmath24.40 , and hence lack the systematic group assignments and primary distance information contained in the catalog . nevertheless , optically - selected galaxies that lie within the sky area and redshift range of the virgo cluster are placed collectively at the virgo distance ( @xmath56 mpc ) , in order to minimize the effects of peculiar motions on the inferred distances of galaxies in the region . we would also like to note that the distance assignment method can have a large impact on the determination of mass functions , especially at the low - mass end . we illustrate this issue in [ sec : dist_unc ] , where we consider the effect on the himf of using uniformly hubble distances for all galaxies . we compute stellar masses for our galaxies based on fitting all 5 sdss photometric bands ( @xmath57 ) , with model spectral energy distributions ( seds ) . the full details of the method can be found in @xcite , but here we summarize the main points : a library of model seds are generated , using the @xcite stellar population synthesis code and assuming a @xcite stellar initial mass function ( imf ) . models with an extensive range of internal extinction , metallicity and star formation histories are considered . in particular , star formation history templates include both an exponentially declining component as well as random starburst episodes . the final physical properties ( e.g. stellar mass , star formation rate , internal extinction etc . ) are computed as the average of all model values , where each model is weighted according to its fit likelihood . in addition to mean values , `` @xmath58 '' uncertainties of the physical properties can also be derived , as one quarter of the 2.5 - 97.5 percentile range of model values . the median @xmath58 uncertainty in @xmath59 is 0.086 dex , or about 22% ( excluding uncertainties on the distance ) . it is important to note that stellar mass estimates of the same galaxy obtained with different methods can have systematic offsets of up to factors of a few . in [ sec : stellar_est ] we address issues related to stellar mass estimation , and consider alternative methods for calculating stellar masses @xcite . figure [ fig : histogram ] compares the distributions of @xmath60-band absolute magnitude ( @xmath61 ) and @xmath44 color ( both corrected for milky way extinction ) , stellar mass ( @xmath8 ) and hi mass ( @xmath47 ) for the hi - selected and optically - selected samples . the most notable difference is in the @xmath44 color distribution , where the hi - selected sample shows a strong bias against the red galaxy population ( see also @xcite ) . as a result , the optically - selected sample contains a larger proportion of high luminosity and stellar mass systems compared to the hi - selected sample . by contrast , the @xmath47 distribution of the two samples is very similar , but remember that only those optically - selected galaxies that are detected in alfalfa are included in the histogram . stellar masses for all galaxies , and hi masses for all alfalfa - detected galaxies are calculated as described in [ sec : derived ] . for the @xmath3415000 galaxies in the optically - selected sample that lack an alfalfa detection , we assign a lower and an upper limit on their atomic hydrogen content ( @xmath62 ) . the lower limit is simply @xmath63 , which corresponds to an hi - devoid galaxy . the upper limit is calculated by assuming that the hi flux of the non - detected galaxy lies just below the alfalfa `` detection limit '' , as defined by the 25% completeness limit of the @xmath24.40 catalog when both code 1 & 2 sources are considered . more precisely @xmath64 where @xmath52 is the galaxy distance in mpc determined by its sdss optical redshift , and @xmath65 is the flux level at which the completeness of the @xmath24.40 catalog falls to 25% , in jy km s@xmath29 . according to eqns . 6 & 7 of @xcite @xmath66 where @xmath67 is the hi - line profile width in km s@xmath29 , measured at the 50% flux level of the profile peak . since alfalfa non - detected galaxies lack a measurement of @xmath67 , we assign a value based on the average @xmath8 @xmath68 relation ( i.e. the stellar mass tully - fisher relation ) of @xmath24.40 galaxies . we then project the @xmath68 value on the line - of - sight according to the sdss @xmath40-band axial ratio , and assuming an intrinsic axial ratio of @xmath69 for all galaxies . baryonic masses ( i.e. stellar mass + atomic gas mass ) for all galaxies are calculated as @xmath22 , where the 1.4 factor is used to account for the cosmic abundance of helium . note that alfalfa non - detected galaxies have two assigned values for their hi mass , and consequently two values for their baryonic mass , @xmath70 and @xmath71 . we calculate cumulative mass functions in logarithmic mass bins for all three components ( i.e. stellar mass , atomic hydrogen mass , baryonic mass ) , separately for the hi - selected and optically - selected samples . since neither sample is volume - limited , mass functions have to take into account the sample selection criteria as well as the large - scale structure in the survey volume . hi selection is based on a combination of galactic hi integrated flux , @xmath49 , and profile width , @xmath67 ( see [ sec : hi_selected ] & discussion in section 6 of @xcite ) ; as a result , galaxies of different hi masses and linewidths are detected out to different distances . similarly , our optically - selected sample is a flux - limited sample , which results in galaxies with different @xmath40-band absolute magnitudes being detected in different volumes . as a result , mass distributions are calculated by summing up the number of detections in a given mass bin ( see fig . [ fig : histogram ] ) , with each detection weighted by an appropriate volume factor . individual weighting factors are calculated via the `` @xmath72 '' method , as implemented in @xcite . this is a non - parametric , maximum - likelihood method , which reduces to the standard @xmath73 method @xcite when applied to a spatially homogeneous galactic sample . the advantage of @xmath72 consists in the fact that it is insensitive to local density fluctuations , and hence mostly immune to structure - induced bias . we refer the interested reader to the literature for the definition , implementation and details of the method . the method definition and basic setup is described in @xcite ; details of the implementation of the method for hi data can be found in @xcite ; the specifics of the application of the method on alfalfa data can be found in @xcite , while a shorter qualitative description can be found in @xcite . lastly , a fraction of galaxies that satisfy all criteria for sdss spectroscopic followup can not be observed for technical reasons ( mostly fiber collisions ) , and are therefore not included in the sdss spectroscopic database . we therefore correct the normalization of all optically - selected distributions by @xmath74 , using the average spectroscopic completeness value reported in @xcite , @xmath75 . similarly , a fraction of the alfalfa volume is `` lost '' due to rfi contamination of certain frequency bands in the alfalfa passband . we correct the normalization of all hi - selected distributions by @xmath76 , where @xmath77 . we would also like to note that , due to the @xmath78 beam size of the alfa receiver , a number of hi sources are expected to be blended . we do not attempt to correct for blending but , given that hi - selected galaxies are a weakly clustered population ( e.g. * ? ? ? * ) , we anticipate the effect on the hi - selected distributions to be small . ( bmf , _ black symbols _ ) , derived from the hi - selected galaxy sample . error bars represent just the poisson counting error assuming independent errors among different mass bins . ] ( _ black lines _ ) , derived from the optically - selected galaxy sample . the atomic hydrogen and baryonic distributions are represented as allowed ranges , based on estimates of the minimum and maximum hi mass for galaxies that are not detected by alfalfa ( \{@xmath62 } , see [ sec : derived ] for details ) . error bars again represent just the poisson counting error assuming independent errors among different mass bins . ] figure [ fig : mfs_hi ] shows the cumulative distribution of stellar mass ( smf ; _ gold symbols _ ) , hi mass ( himf ; _ cyan symbols _ ) and baryonic mass ( bmf ; _ black symbols _ ) , derived from the hi - selected galaxy sample . the hi - selected bmf follows closely the hi - selected smf at high masses , while at the low - mass end the contribution of the himf becomes dominant ; this is because hi - selected galaxies become more gas - rich as their stellar mass decreases . figure [ fig : mfs_opt ] shows the same distributions ( smf , _ gold line _ ; himf , _ cyan lines _ ; bmf , _ black lines _ ) derived from the optically - selected sample . recall that , in the case of the optically - selected sample , a lower and upper limit of the himf and bmf are shown , since sdss galaxies that are undetected by alfalfa are assigned an upper and lower limit on their hi content , and therefore also on their baryonic mass ( see [ sec : derived ] ) . ) and low - mass ( @xmath79 ) ends . this difference is mostly due to the bias of the hi - selected sample against the red galaxy population ( see [ sec : results ] for a detailed discussion ) . ] for a detailed discussion . ] ) . [ sec : results ] for a detailed discussion . ] in figure [ fig : opt_vs_hi_gsmf ] we compare the smfs derived from the optically - selected and hi - selected samples . the two smfs are consistent at intermediate masses , but the optically - selected smf ( gold line ) is systematically higher at the high - mass and low - mass ends . at high masses the discrepancy is due to the bias of the hi - selected sample against the most massive galaxies , which are usually red passive systems . the discrepancy at the low - mass end is mostly due to the population of red - sequence dwarf galaxies in the nearby virgo cluster that are undetected by alfalfa ; these are mostly dwarfs with early - type morphologies and very low hi content ( see @xcite , for example ) . on the other hand , the hi - selected and optically - selected himfs ( fig . [ fig : opt_vs_hi_himf ] ) are in good agreement with one another , with the hi - selected himf having a slightly steeper low - mass end slope than what suggested by the range of the optically - selected distribution . the two bmfs are mostly in agreement with one another , except at the high - mass end ( a factor of @xmath344 at @xmath80 ) . this is a direct consequence of the discrepancy between the optically - selected and hi - selected stellar mass functions at high masses . note that there is little difference between the two bmfs at low masses , which suggests that the low - mass end of the bmf has been measured robustly . ) . the red diamonds correspond to the @xcite smf in the local universe ( @xmath81 ) , while the purple triangles correspond to the smf of @xcite , extracted over a larger volume ( @xmath82 ) . both the baldry at el . and yang et al . smfs are based on the nyu - vagc galaxy catalog . [ sec : comparison ] for a discussion of the comparison . ] figure [ fig : gsmf_vs_other ] compares the optically - selected smf presented in this work ( same as gold line in fig . [ fig : opt_vs_hi_gsmf ] ) with the local - universe smf of @xcite and the @xcite smf , which are both based on the new york university value - added galaxy catalog ( nyu - vagc ; * ? ? ? there is excellent agreement between the baldry et al . smf and our optically - selected smf , especially at intermediate and low stellar masses ( @xmath83 ) . the deviations at high masses are due to the fact that stellar masses in baldry et al . are calculated differently than in this work ( see sec . 3 of @xcite for details ) ; note that a systematic difference of just 0.1 dex ( 26% ) in stellar mass would be enough to explain the observed difference in abundance at the high - mass end . the @xcite smf is systematically higher than our optically - selected smf at high masses , and displays a more pronounced `` plateau '' at intermediate masses . it is not clear what the cause of the difference at the high - mass end of the distributions is , but it may relate to the fact that the yang et al . smf is extracted from a significantly larger volume than our measurement ( the maximum redshift is @xmath84 for the yang et al . sample and @xmath85 for the sample used in this work ) . moreover , yang et al . use the prescription of @xcite to estimate stellar masses , which is based on the galactic @xmath86 color . as discussed in more detail in [ sec : stellar_est ] , the use of different stellar mass estimators can significantly affect the shape of the measured smf . ) . the solid purple line corresponds to the schechter function fit to the @xcite himf , which is based on the full @xmath24.40 catalog of alfalfa sources and without any optical selection cuts . the green dashed line corresponds to the schechter fit to the @xcite himf , based on 4315 galaxies detected by the hipass survey . [ sec : comparison ] for a discussion of the comparison . ] figure [ fig : himf_vs_other ] compares the hi - selected himf presented in this work ( same as the cyan symbols in fig . [ fig : opt_vs_hi_himf ] ) with the himf of @xcite derived from 10119 galaxies detected by alfalfa over @xmath342600 deg@xmath35 of sky ( purple solid line ) . there is excellent agreement at intermediate and large hi masses ( @xmath87 ) between the @xcite himf and the himf derived in this work , while at lower masses the @xcite himf is slightly steeper than ours . this disagreement may be due to the set of additional optical requirements imposed on our hi - selected sample . as argued in [ sec : hi_selected ] , these requirements are expected to reduce the number of low - mass systems in the sample and therefore decrease the inferred space density at the low - mass end . the dashed green line represents the himf of @xcite based on 4315 sources detected by the hi parkes all - sky survey ( hipass ) over the whole southern celestial hemisphere ( @xmath34 29000 deg@xmath35 ) . there is disagreement between both alfalfa - based himfs and the hipass - based himf of @xcite , at the high - mass end . as argued in @xcite , the higher sensitivity of the alfalfa survey compared to hipass , which enables alfalfa to detect hi - massive systems over a larger volume , should give a statistical advantage to the alfalfa survey in determining the high - mass end of the himf . however , the difference is too large to be explained by counting statistics or cosmic variance ( e.g. according to the estimates of @xcite or @xcite ) . on the other hand , due to the exponential drop - off of the himf at high masses , a flux calibration difference of as low as 0.1 dex could give rise to a similar discrepancy . color and the @xmath60-band luminosity , according to the widely used @xcite calibration . the gold dotted line represents the smf computed using stellar masses derived from the galactic @xmath44 color and the @xmath60-band luminosity , according to the the more recent calibration of @xcite . [ sec : stellar_est ] for a discussion of the comparison . ] a variety of methods exist to estimate stellar masses from spectra or broadband photometric measurements of galaxies ( * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * to name a few ) . most methods rely on comparing the actual galactic emission to the light output of a set of galactic stellar population models . the models that most closely reproduce the observed data are then used to estimate the galactic properties of interest ( e.g. stellar mass , present star formation rate , internal extinction etc . ) ; it is therefore very important to consider model stellar populations which span as large a range of physical parameters as the galaxies in the sample being studied . inferred galactic properties depend not only on the particular type of data employed by each method ( e.g. spectroscopy vs. broadband photometry or optical vs. near - infrared photometry ) , but also by differences in the way in which the model stellar populations are constructed . this means that different methods can yield different estimates of a galactic property even when the _ same _ observational measurements are used . for example , @xcite find that unbiased stellar masses can only be recovered when the true star formation history ( sfh ) of a galaxy is known . in practice however a restricted set of sfhs is considered ( often in the form of a parametrized function ) , which may introduce systematics for galaxies with sfhs that are not well described by the assumed general form . additional complications can be introduced by the different treatment of dust reddening among different models . in general , stellar mass estimates can differ systematically by as much as 0.3 dex , while for individual galaxies the scatter can be as large as 0.6 dex @xcite . bands used in this work ( see [ sec : derived ] ) , and those derived from the galactic @xmath86 color and the @xmath60-band luminosity according to the widely used @xcite calibration . each datapoint corresponds to a galaxy in the optically - selected sample , and the symbol color represents the g - r color of the galaxy . the two stellar mass estimates agree fairly well for red passive galaxies , while for blue star - forming galaxies bell et al . masses are systematically larger by up to a factor of @xmath342 ( see [ sec : stellar_est ] for a detailed discussion ) . ] here we compare the optically - selected smf presented in this work ( computed from stellar masses derived from sed - fitting , see [ sec : derived ] ) with the smf obtained using stellar masses derived from a single galactic color , according to the widely - used @xcite calibration as well as the more recent calibration of @xcite . more specifically , we compute bell et al . masses by multiplying the @xmath60-band luminosity of each galaxy by a mass - to - light ratio inferred from its @xmath86 color . we choose this particular combination of bands because it is relatively immune to contamination of galactic spectra by bright nebular emission lines @xcite . we use a similar procedure to calculate taylor et al . masses , by using their calibration of @xmath60-band mass - to - light ratio versus @xmath44 color . @xcite argue that using @xmath44 colors best constrains the galactic stellar mass estimates . figure [ fig : sed_vs_color ] shows the impact that different methods of estimating stellar mass have on the measurement of the smf . when bell et al . stellar masses are used ( _ gold dashed line _ ) , the smf becomes systematically higher at low and intermediate masses ( @xmath88 ) , while it remains mostly unchanged at the high - mass end . the reason for this pattern becomes evident in fig . [ fig : bell_vs_sed ] , where we see that bell et al . masses agree fairly well with the masses derived in this work for red passive galaxies , but are systematically larger ( by up to a factor of @xmath342 ) for blue star - forming galaxies . @xcite argue that the difference can be primarily attributed to the fact that the stellar population models used for the bell et al . calibration do not consider `` bursty '' star formation histories which are typical of low - mass galaxies with blue colors . this leads to systematically older stellar populations for blue galaxies according to the bell et al . method , which in turn results in systematically higher stellar mass estimates . note that including models with bursty sfhs in a stellar population library does not by itself guarantee a correct estimate of stellar mass ; overestimating the effect of bursts would result in systematically low stellar masses for blue galaxies . conversely , when taylor et al . masses are used ( gold dotted line ) , the smf becomes systematically lower at the high - mass end , while it is mostly unchanged at low and intermediate masses . again , this is a result of the fact that taylor et al . masses agree well with the sed - fitting masses used in this work for blue star - forming galaxies , but are systematically lower ( by up to a factor of @xmath341.4 ) for red passive galaxies . for further discussion ) . ] stellar , hi and baryonic masses are distance - dependent quantities and hence their statistical distributions are affected by distance uncertainties . this is particularly true for the low - mass end of the distributions , which is determined by the properties of nearby galaxies ; neglecting the peculiar velocity of some of these objects can cause fractional distance and mass errors of order @xmath89 , especially in a volume with complex large - scale structure such as the one surveyed by alfalfa ( see fig . [ fig : coneplot ] ) . for this reason , nearby galaxies ( @xmath90 km s@xmath29 ) in the @xmath24.40 catalog are assigned distances based on a parametric peculiar velocity flow model @xcite , and only more distant galaxies ( @xmath91 km s@xmath29 ) are assigned simple hubble distances according to their cmb recessional velocity . the @xcite flow model includes a dipole and a quadrupole component ( local group bulk motion & asymmetric expansion ) and two local attractors ( virgo cluster & great attractor ) , and is calibrated against the sfi++ catalog of galaxies with redshift and independent distance distance information ( from tully - fisher ) . the residuals are then attributed to random thermal motions , which are estimated to have a magnitude of @xmath92 km s@xmath29 . in addition , distances reported in the @xmath24.40 catalog take into account known group and cluster membership as well as primary distance information published in the literature . this latter information is not available for the majority of the galaxies in our optically - selected sample ( which are not included in @xmath24.40 ) , and we only make an attempt to assign all probable virgo members to the virgo cluster distance ( @xmath93 ) . here we re - evaluate the himf for our hi - selected sample using uniformly hubble distances for all galaxies , in order to illustrate the impact of the distance assignment scheme on the derived distributions . figure [ fig : hubble_vs_flow ] shows that the himf computed using hubble distances ( cyan dashed line ) has a much shallower low - mass end slope compared to the himf presented in this work , which uses flow model distances for most nearby galaxies ( cyan solid line ) . this result is in agreement with the work of @xcite , who find that neglecting the local peculiar velocity field when estimating distances to nearby galaxies in the alfalfa volume will lead to a systematically shallower low - mass end slope for the himf . . cyan diamonds represent 14 galaxies of the heracles survey @xcite detected in co line emission . the solid and dotted blue lines represent the average and @xmath94 scatter of the distribution found for 125 co detected galaxies in the cold gass survey @xcite . both samples show that , at least for @xmath95 , molecular gas is almost always a subdominant mass component ( see [ sec : h2 ] for details ) . ] throughout this article we use the term `` baryonic mass '' to refer to the sum of the stellar and atomic gas mass ( @xmath22 ) , a convention that is common in the literature . this definition however excludes a number of baryonic components that are definitely present in galaxies , most notably molecular and ionized -warm or hot- gas . figure [ fig : h2_mass_frac ] displays the fraction of molecular hydrogen ( h@xmath96 ) mass ( accounting for helium ) to `` baryonic mass '' as defined above , as a function of stellar mass . the cyan diamonds represent 14 galaxies from the heracles survey @xcite with h@xmath96 masses measured from interferometric co line observations ( using a fixed @xmath97 conversion factor ) . over the probed stellar mass range ( @xmath98 ) , molecular gas rarely contributes more than 10% of the `` baryonic mass '' . the same conclusion is reached when h@xmath96 mass measurements from the cold gass survey are considered @xcite . the blue solid and dotted lines show the mean and @xmath94 scatter of the relation between the molecular and `` baryonic '' mass components , based on 125 galaxies detected in co emission with the iram 30 m telescope . in this latter case we have estimated the atomic hydrogen mass of galaxies indirectly , using the average @xmath99 vs. @xmath8 relation of the cold gass parent sample @xcite . again , over the stellar mass range probed by the survey ( @xmath100 ) molecular gas is always a subdominant mass component . at lower stellar masses there is large uncertainty on the fractional contribution of h@xmath96 , as it is not precisely known how well the galactic co emission traces molecular hydrogen mass . in particular , the @xmath97 conversion factor may vary by about an order of magnitude as we consider less luminous and more metal poor late - type galaxies ( e.g. * ? ? ? determining the contribution of ionized gas to the total baryonic mass budget of galaxies is much more challenging . for example @xcite argue that warm ionized hydrogen ( hii ) may amount to about 1/3 of the mass of atomic hydrogen ( hi ) in the disk of the milky way . if the ratio of ionized - to - neutral hydrogen mass ( @xmath101 ) were fairly constant among galaxies , then the baryonic mass of a galaxy would be given by the expression @xmath102 . if @xmath103 , then the peak value of the @xmath21 - @xmath9 relation ( see fig . [ fig : am2 ] ) would shift to lower halo mass and the peak value would slightly increase . however , since the precise value and scatter of @xmath101 is not well constrained -and its dependence on galaxy size is not known- we choose not to include the contribution of warm ionized gas in the calculation of @xmath104 . assessing the contribution of the hot ionized medium ( him ) to the total baryonic mass of a galaxy is even more challenging . the coronal him may be the dominant baryonic mass component in galactic halos , especially in massive ellipticals . determining the mass contribution of the him for less massive galaxies however is observationally challenging . in any case , the tightness of the `` baryonic tully - fisher relation '' when computed just from the stellar and hi mass ( e.g. * ? ? ? * ; * ? ? ? * ) implies that the him never dominates the total baryonic mass budget of late - type galaxies , at least within the extent of the galactic hi disk . let @xmath105 be the cosmic number density of galaxies with baryon mass greater than @xmath104 and let @xmath106 be the cosmic number density of haloes with mass greater than @xmath9 . the fundamental assumption of the abundance matching method ( @xcite ; @xcite ; also see @xcite for a review ) is that @xmath104 is a monotonically increasing function of @xmath9 . with this assumption , @xmath107 can be determined by solving the equation @xmath108 in reality , the baryon content of a halo will depend not only on its mass but also on other parameters , such as its formation history . as a result , a scatter in the distribution of @xmath104 at a given @xmath9 is expected . neglecting the scatter is , nevertheless , justifiable because the aim of abundance matching is precisely to determine the average value of @xmath104 within a halo of mass @xmath9 . we evaluate the right hand side of equation [ eq : am ] , using a halo mass function extracted from one of the cosmological n - body simulations of the horizon project . the simulation was run with a public version of the gadget code @xcite , and uses @xmath109 particles of mass @xmath110 to simulate the formation and evolution of dm structures in a comoving volume of 100 @xmath111 mpc on a side . it assumes a cosmology and initial conditions which are consistent with wilkinson microwave anisotropy probe ( wmap ) third year results @xcite , namely @xmath112 , @xmath113 , @xmath114 , and @xmath115 . the identification of dm haloes and sub - haloes was done with the adaptahop algorithm presented in @xcite . haloes are identified as groups of particles above a threshold over - density of 80 times the mean density of the universe , which corresponds to a mean overdensity contrast of about 200 . the identification of subhaloes within haloes is done using the method described in @xcite . we only keep haloes and sub - haloes with more than 20 particles , i.e. we introduce a minimum halo mass of @xmath116 . - @xmath9 relation . the black solid line shows the @xmath8 - @xmath9 relation obtained by matching the stellar mass function of central galaxies to the mass function of haloes excluding subhaloes . this matching should reproduce the `` true '' relation for central galaxies . the blue dot - dashed line is the relation obtained by matching the total galaxy stellar mass function , including central and satellite galaxies , to the mass function of haloes excluding subhaloes . the red dashed line is the relation obtained by matching the total galaxy stellar mass function with the mass function of haloes including subhaloes . note that , unlike all other figures , this figure uses the stellar mass functions of @xcite , who have separately measured the smf for central and satellite galaxies . ] an important issue for our analysis is determining whether subhaloes should be included or excluded in the calculation of the hmf used in the abundance matching procedure . let @xmath117 be the stellar mass of the central galaxy in a halo of mass @xmath9 and let @xmath118 be the stellar mass of the satellite galaxy in a subhalo of mass @xmath9 . if subhaloes are excluded , then we implicitly assume that all galaxies in our samples are central galaxies , that is , @xmath119 . on the other hand , if subhaloes are included , we implicitly assume that the @xmath104 - @xmath9 relation is the same for both central and satellite galaxies , that is , @xmath120 . in order to understand how much this choice could affect our results , we consider the stellar mass functions of @xcite , for which separate distributions for central and satellite galaxies have been presented . this allows us to calculate the @xmath8 - @xmath9 relation in three ways : firstly , we can match the smf of central galaxies with the halo mass function excluding subhaloes ; this method will give us the correct @xmath8 - @xmath9 for central galaxies , shown by the black solid line in figure [ censat ] . then , we consider the total smf of central plus satellite galaxies and match it with the halo mass function excluding subhaloes . this is equivalent to assuming that all galaxies are central and overestimates the @xmath8 - @xmath9 for central galaxies ( blue dot - dashed line in fig . [ censat ] ) . finally , we can match the total smf of central plus satellite galaxies with the total halo mass function including subhaloes . the result , shown by the red dashed line in figure [ censat ] , lies below the black solid line because it is effectively a weighted average of the relation for the dominant central galaxy population ( the black solid line ) and the relation for the satellite population , which has lower @xmath8 for a given @xmath9 . quantitative comparisons of @xmath118 and @xmath117 have received the attention of much recent literature ( see e.g. cattaneo et al . 2012 , submitted , @xcite , @xcite ) . as we do not distinguish between central and satellite galaxies in the samples used in this work , we shall choose the third method ( the one that corresponds to the red dashed line ) as our best estimator of the @xmath8 - @xmath9 relation and the @xmath104 - @xmath9 relation . this choice will introduce some systematic bias which is , nonetheless , smaller than the typical uncertainty involved in the determination of the @xmath8 - @xmath9 relation . we also considered whether abundance matching our hi - selected smf with the hmf excluding subhaloes would give consistent results with our fiducial abundance matching result , obtained by matching our optically - selected smf and the hmf including subhaloes . physically , this consideration was motivated by the fact that satellite galaxies tend to be redder than central galaxies of the same mass ( e.g. * ? ? ? * ) , and so hi - selection may be equivalent to the exclusion of satellite galaxies from an observational sample . however , the comparison in fig . [ censatbis ] of our hi - selected smf with the smf for central galaxies of @xcite shows that this argument may not be valid . we note that the comparison between the yang et al . smf for central galaxies and our hi - selected smf is subject to observational systematics , such as the distance assignment scheme or the stellar mass estimator ; for example using hubble distances for the galaxies in our hi - selected sample ( see fig . [ fig : hubble_vs_flow ] ) would bring the two distributions in fair agreement . relation ) . the thick yellow line shows our main result , obtained from abundance matching the stellar mass function of our optically - selected sample with the halo mass function including subhaloes . the yellow dashed and dash - dotted lines correspond to variations of our main result , obtained by considering the @xcite smf and excluding subhaloes from the hmf respectively , and are shown to illustrate uncertainties . the magenta dotted , cyan dashed , blue dot - dashed , green solid and red solid lines correspond to the abundance matching results of @xcite , @xcite , @xcite , @xcite and @xcite , respectively . the big green circles are the results of a stacked weak lensing study of sdss galaxies by @xcite . all other data points refer to measurements for individual galaxies : the small circles correspond to galaxies with halo mass measurements from weak lensing studies ( @xcite : cyan circles ; @xcite : red circles ; @xcite : blue circles ) . the star symbols show galaxies in which the halo mass was determined from studies of stellar dynamics ( @xcite : cyan stars ; @xcite : red stars ) . the triangles show galaxies with halo masses determined from the disc rotation speed ( @xcite : cyan triangles ; @xcite : red triangles ; @xcite : blue triangles ) . the dotted - dashed horizontal line shows the cosmic baryon fraction @xmath121 . ] our abundance matching analysis produces two main results : 1 . we determine @xmath122 as a function of @xmath9 by comparing the galaxy stellar mass function ( smf ) from our optical sample to the total halo mass function ( including subhaloes ) . [ item1 ] 2 . we determine @xmath123 as a function of @xmath9 by comparing the baryonic mass function ( bmf ) from our optically selected sample to the total halo mass function ( including subhaloes ) . [ item2 ] the latter relation is the focus of this paper , but we first discuss the former because the results can be compared to an extensive literature of previous studies . the consistency of our findings with previous work on point [ item1 ] boosts our confidence that our conclusions on point [ item2 ] are robust . figure [ fig : am1 ] shows our result for @xmath122 as a function of @xmath9 , obtained from our optically selected sample ( gold thick solid line ) . our analysis extends to halo masses as low as @xmath124 , since both our optically - selected and hi - selected samples probe stellar masses down to @xmath125 . at the high mass end , our relation stops at @xmath126 because our galactic samples are drawn from a relatively small volume , and are not appropriate to measure the abundance of the most massive galaxies and clusters . we find good agreement with previous estimates from abundance matching obtained in the same manner ( @xcite : cyan dashed line ; @xcite : blue dotted - dashed line ; @xcite : green solid line ; @xcite : red solid line ; @xcite : purple dotted line ) . the thick yellow dash - dotted and dashed lines are shown to illustrate the systematics introduced by the choice of hmf and smf : the former represents the abundance matching result when our fiducial smf is matched the hmf excluding subhaloes ; the latter is the result of matching the yang et al . smf with our fiducial hmf , which includes subhaloes . we also compare our average @xmath8 - @xmath9 relation with values measured for individual galaxies . the small cyan , red and blue circles correspond to galaxies with measurements of their halo mass @xmath9 from weak lensing @xcite . the star symbols correspond to galaxies for which @xmath9 was estimated form stellar dynamics @xcite . the triangles correspond to disc galaxies for which @xmath9 was determined from their rotation speed @xcite . we remark that , while results for individual galaxies have large scatter , they seem to be systematically lower than any of the relations inferred from abundance matching ( at least over the halo mass range @xmath127 ) . furthermore , the halo mass for which @xmath128 has a maximum appears to be higher when inferred from measurements of individual galaxies compared to the value derived from abundance matching : in the first case the peak is at @xmath129 , while in the second case it is at @xmath130 . there is also slight disagreement of all abundance matching results with the results of @xcite ( large green circles ) , who used a stacked weak lensing analysis of over a hundred thousand disk galaxies in sdss separated in three bins of stellar mass . regardless of the method used , however , there is a clear consensus that @xmath122 is much lower than the universal baryon fraction @xmath121 ( horizontal black dash - dotted line in figure [ fig : am1 ] ) , at all halo masses . relation ) . the gray shaded area shows the results of an abundance matching analysis of our optically - selected sample . its boundaries correspond to two extreme assumptions for the gas content of galaxies detected optically but not in hi : i ) galaxies that are not detected in hi contain no gas ( lower boundary ) and ii ) galaxies that are not detected in hi contain the largest amount of gas that could have escaped detection from alfalfa ( upper boundary ) . the thick solid yellow line is the @xmath128-@xmath9 relation ( same as in figure [ fig : am1 ] ) , and is shown for comparison . the dotted - dashed line shows the baryon fraction that @xcite predict based on hydrodynamic simulations that include cosmic reionisation . ] let us now examine the results for the @xmath123 - @xmath9 relation ( fig . [ fig : am2 ] ) . the gray shaded area shows the relation derived from our optically - selected sample , matched to a halo mass function that includes both haloes and subhaloes : its upper and lower envelopes correspond to the distribution for @xmath131 and @xmath132 , respectively , as defined in [ sec : derived ] . the thick gold solid line represents the relation for the stellar mass ( same as in figure [ fig : am1 ] ) and has been added for reference . figure [ fig : am2 ] puts in evidence the fact that @xmath21 decreases monotonically with decreasing halo mass , despite the fact that atomic gas contributes progressively more to the baryonic mass budget of galaxies . - @xmath9 relation derived in this work from our optically - selected sample ( same as fig . [ fig : am2 ] ) . the magenta dotted line and the green solid lines show the @xmath123 - @xmath9 relations that @xcite and @xcite derived from their @xmath122 - @xmath9 relations ( fig . [ fig : am1 ] ) , using the mean gas - to - stellar mass ratio as a function of stellar mass to account for the gas content of galaxies . the black solid line correspond to the results obtained from our optically - selected sample when we use the same method . the red solid line represents the result of @xcite , who used separate @xmath133 - @xmath8 relations for blue and red central galaxies . lastly , the thin black dotted - dashed line is the same as fig . [ fig : am3 ] and shows the baryon fraction that @xcite predict based on hydrodynamic simulations that include cosmic reionisation . ] in fig . [ fig : am3 ] we compare our results to those by @xcite and @xcite , who derived their @xmath123 - @xmath9 relations from the equation @xmath134 in eqn . [ eqn : mbmh ] the baryonic mass is computed from the stellar mass , using the mean gas - to - stellar mass ratio for galaxies as a function of stellar mass . to enable a cleaner comparison , we have made the exercise of re - deriving the @xmath123 - @xmath9 relation from our @xmath122 - @xmath9 relation , using the same procedure followed by @xcite ( as in eq . [ eqn : mbmh ] ) . the result is shown by the black solid line in fig . [ fig : am3 ] . the main difference between the results obtained by using individual galaxy gas masses ( gray shaded region ) and by adopting a mean gas - to - stellar mass ratio ( black solid line ) seems to be an artificial flattening of the @xmath135 relation at low masses . this is probably due to the fact that the latter method ignores the large scatter of galactic @xmath136 values from the mean , and leads to the incorrect interpretation that the baryon retention fraction ( @xmath21 ) of low - mass halos asymptotes to some fixed value . the red solid line in fig . [ fig : am3 ] corresponds to the result of @xcite , who used a separate mean @xmath136 - @xmath8 relation for blue and red galaxies . their result shows no signs of flattening at low masses , however the relation only extends down to @xmath137 . independently of the used method however , all results point to values of @xmath21 that are well below unity , and can not be explained by the effects of cosmic reionization alone ( black dotted - dashed line in fig . [ fig : am3 ] ; see sec . [ sec : discussion ] for details ) . ) for the galaxies in our optically - selected sample that are detected by alfalfa , while the red inverted triangles are maximum hi - fractions ( @xmath138 ) for a representative subset of our optically - selected galaxies that are not detected by alfalfa . ] the @xmath139 relation that we use to evaluate the right hand side of eqn . [ eqn : mbmh ] ( black solid line in fig . [ fig : gsmr ] ) is a power - law fit to the @xmath99 data by @xcite , @xcite , @xcite and @xcite . the best fit relation , plotted as a thin solid black line in fig . [ fig : gsmr ] , is given by @xmath140 . the blue contours in fig . [ fig : gsmr ] represent the distribution of @xmath99 values for the galaxies in our optically - selected sample that are detected by alfalfa . since the alfalfa survey is a blind hi survey with a fixed integration time per pointing , the alfalfa distribution is expected to lie above the data obtained by pointed observations of optically - selected galaxies ( as in @xcite ) . the inverted red triangles correspond to the maximum hi - fraction ( @xmath138 ) for a representative subsample of our optically - selected galaxies that lack an alfalfa detection . note that these upper limits are also systematically higher than the relationship indicated by the power - law fit . the main result of this paper is the large `` gap '' between the present - day baryon fraction of galaxies in low - mass halos and the cosmic value ( @xmath121 ) , which is present even when the atomic gas content of the galaxies is taken into account . this result is not contrived , given that atomic gas dominates the baryonic mass budget of galaxies with @xmath141 . moreover , the low - mass behavior is in disagreement with previous studies ( e.g. * ? ? ? * ; * ? ? ? * ) , who find that the @xmath21 - @xmath9 relation flattens out at low masses , and approaches a roughly constant value of @xmath142 . these previous results would then require an exceptionally low efficiency of gas - to - stars conversion in low mass systems to explain the observed values of @xmath11 , which decrease monotonically . in fig . [ fig : am2 ] we compare our result for the baryon fraction of halos to the predictions at @xmath143 of a cosmological hydrodynamic simulation that includes heating from a photoionizing uv background ( * ? ? ? * black dotted - dashed curve ) . at the high - mass end ( @xmath144 ) , the low values of @xmath21 can be explained by the fact that infalling gas is shock - heated @xcite , and is kept hot by feedback from active galactic nuclei @xcite . this picture is supported by considerable observational evidence in the case of x - ray groups and clusters , but the situation within individual galaxies is not so well understood ( see @xcite for a review ) . at the low mass end , photoionization heating is expected to become important , since the intergalactic medium is too hot to fall into the shallow potential wells of haloes with @xmath145 , and their baryon fraction is heavily suppressed . however , this process alone can not account for the shape of the @xmath21 - @xmath9 relation at low masses , especially for the onset of a sharp decrease in the @xmath123 ratio at relatively large halo masses ( @xmath146 ) . additional feedback is therefore needed , usually attributed to the ejection of baryons by stellar - driven winds . semianalytic models of galaxy formation based on this assumption reproduce a good fit to luminosity functions in the local universe @xcite but the implied outflow rates are enormous . to reproduce the result presented in figure [ fig : am2 ] , the outflow rate in a halo with @xmath147 must be of order a hundred times higher than the star formation rate . it is difficult at present to reproduce such outflow rates in hydrodynamic simulations . moreover , observational estimates place the total mass of outflows in normal star - forming galaxies at approximately the same level as the galaxies final stellar mass @xcite . even in the most extreme observational cases , the `` mass - loading factor '' ( the ratio of mass loss rate due to outflows over the star formation rate ) is estimated to be @xmath148 @xcite . therefore , explaining in detail the mechanisms responsible for the very low @xmath21 values found in low - mass galaxies seems to be a fundamental challenge for models of galaxy formation in a @xmath0cdm cosmological context . recent cosmological hydrodynamic simulations have challenged this statement , by managing to reproduce `` realistic '' galaxies whose properties satisfy a number of observational constraints . for example , the high - resolution eris simulation @xcite has managed to produce a milky - way type object with values of @xmath11 in agreement with those that we see in fig . [ fig : am2 ] . more recently , @xcite has managed to reproduce a population of @xmath1491000 simulated galaxies with low stellar - to - halo mass ratios ( @xmath150 at @xmath151 ) , in accordance to observations . notice , however , that this simulation uses a kinetic rather than thermal wind model , in which the initial wind speed is @xmath152 and the initial mass - loading factor is a factor of four , by construction . lastly , the work of @xcite has been extended to lower masses by @xcite , who managed to produce a pair of dwarf galaxies ( @xmath153 ) , which obey the observed `` baryonic tully - fisher '' relation and are therefore expected to have the correct baryon - to - halo mass fractions . while these studies indicate that we may be heading toward a solution of the discrepancy between the observed and the expected baryon content of dark matter halos , many questions ( e.g. the expected outflow rates and re - accretion timescales ) remain unanswered . therefore , explaining in detail the mechanisms responsible for the very low @xmath21 values found in low - mass galaxies remains an open problem for studies of galaxy formation in a @xmath0cdm universe . furthermore , our measurement of @xmath154 provides an additional constraint with which cosmological hydrodynamical simulations and semianalytic models will have to confront . we use optical data from the seventh data release of the sloan digital sky survey ( sdss dr7 ) and 21 cm emission - line data from the arecibo legacy fast alfa ( alfalfa ) survey to measure the `` baryonic mass '' ( defined as @xmath22 ) of galaxies in the local universe , and determine the @xmath143 baryon mass function ( bmf ) . we use both an hi - selected and an optically - selected sample ( 7195 and 22587 galaxies respectively ) drawn from the same volume , in order to address the effects of sample selection on the derived distributions . we find that the main difference consists of the optically - selected stellar mass function ( smf ) being systematically larger at high - masses than the hi - selected smf , and find that this difference carries over to the high - mass end of the bmf ( see fig . [ fig : opt_vs_hi_gsmf ] & [ fig : opt_vs_hi_bmf ] ) . we combine the obtained mass distributions with the halo mass function in a wmap3 @xmath0cdm cosmology , to obtain average values of @xmath128 and @xmath135 as a function of halo mass ( fig . [ fig : am1 ] & [ fig : am2 ] ) . our most important result is that low - mass halos seem to have very low galactic baryon fractions compared to the cosmic value ( @xmath2 ) , even when their atomic gas content is taken into account ; for example , the average baryon fraction of halos with @xmath155 is just 2% of the cosmic value ( @xmath156 ) , and displays a monotonically decreasing trend . this result contrasts with previous indirect measurements of the bmf @xcite , which pointed to an approximately constant value of @xmath157 at the low halo - mass end . such very low values of @xmath21 are difficult to reconcile with current models of galaxy formation . photoionization heating in the early universe suppresses the baryonic content of halos only at @xmath158 @xcite , but this mass is more than an order of magnitude smaller than what is required by our result . therefore , additional feedback mechanisms , such as baryon blowout by supernova explosions , must be present and must be extremely efficient . it is not yet clear whether hydrodynamic simulations or observational results can accommodate such intense galactic outflows in low mass halos . as a result , the observed @xmath21 - @xmath9 relation remains difficult to explain , and may represent a challenge to our understanding of galaxy formation and/or the properties of dark matter . the authors acknowledge the work of the entire alfalfa collaboration team in observing , flagging , and extracting the catalog of galaxies used in this work . the alfalfa team at cornell is supported by nsf grants ast-0607007 and ast-1107390 to rg and mph and by grants from the brinson foundation . ep would also like to thank the onassis foundation for partial support . ac thanks alexie leauthaud for providing many of the data points used to make fig . [ fig : am1 ] . funding for the sdss and sdss - ii has been provided by the alfred p. sloan foundation , the participating institutions , the national science foundation , the us department of energy , the nasa , the japanese monbukagakusho , the max planck society and the higher education funding council for england . the sdss web site is ` http://www.sdss.org ` . the sdss is managed by the astrophysical research consortium for the participating institutions .

in this short introduction we would like to introduce the two players in the title , random matrix theory ( rmt ) and quantum chromodynamics ( qcd ) , on a superficial level . this gives a motivation why and where it will be beneficial to relate these two seemingly unrelated theories , and what the reader may expect to learn from these notes . let us begin with qcd , the theory of the strong interactions . it is part of the standard model of elementary particles that also describes the weak and electromagnetic interactions among all elementary particles . the latter two interactions will however not play any role here . the particle content of qcd are the @xmath0 quarks with masses @xmath1 , @xmath2 , and the gluons . in nature @xmath3 flavours have been observed named up , down , strange , charm , top and bottom quark . in the following we will keep @xmath0 as a free parameter , and often consider only the lightest up and down quark ( @xmath4 ) as they constitute the most common particles as proton ( p ) , neutron ( n ) and pions ( @xmath5 ) . the quarks interact through the gluons , the carriers of force , with a coupling constant @xmath6 . both come in 3 colours and the interaction is described through a field strength and covariant derivative carrying an @xmath7 lie group structure . qcd is a strongly interacting , relativistic quantum field theory ( qft ) which is very difficult . fortunately we will only need to know a few features and some of its global symmetries that will be described in section [ symm ] . for a standard textbook on perturbative qcd we refer for example to @xcite . roughly speaking qcd has two different phases that are schematically depicted in figure [ fig : phasediag ] . they are characterised by an order parameter , the chiral condensate @xmath8 . at high energies corresponding to high temperatures t , @xmath9 and quarks and gluons form a plasma , that has been observed in collision experiments e.g. at the relativistic heavy ion collider rhic in brookhaven . this phase also existed in the early universe , with the cooling down happening close to the temperature axis at low baryon density parametrised by @xmath10 . in these situations as well as in single particle collisions produced at collider experiments a perturbative expansion in powers of the coupling @xmath6 typically applies , as described e.g. in @xcite . the second phase with @xmath11 at low temperature and density is the one in which we live . here quarks condense into colourless objects , that is into baryons made of 3 quarks of different colour adding up to white ( like p or n ) , or mesons made of a quark and an anti - quark with its anti - colour ( like the @xmath12 s ) . gluons are also confined to these objects , and a proof of confinement of quarks and gluons from qcd is still considered to be an open millennium problem . in qcd it happens that confinement goes along with the spontaneous breaking of chiral symmetry - hence the name of @xmath8 . this global symmetry will be described in detail later , and eventually lead us to a rmt description . ( 12.2,5.5 ) ( 1.0,0.3)(1,0)8 ( 9.2,0.2)@xmath10 ( 1.0,0.3)(0,1)5 ( 0.1,5.1)@xmath13 ( 0.1,4.3)@xmath14 ( 7.1,4.0)@xmath9 ( 1.0,-0.1)@xmath15 ( 2.9,1.8)@xmath11 ( 3.9,4.25)@xmath16 ( 8.0,-0.1)@xmath17 ( 1.0,4.4)(4.7,4.3)(4.0,4.3 ) ( 4.0,4.3)(7.0,4.2)(8.0,0.3 ) in the phase with @xmath11 perturbation theory breaks down and one has to apply other methods . if one wants to continue to work with first principles and the qcd action , one possibility is to study qcd on a finite space - time lattice of volume @xmath18 and lattice spacing @xmath19 , equipped with a euclidean metric . the numerical solution of this theory nowadays reproduces the masses of particles that are composed of elementary ones to a very high precision , as testified in the particle data booklet @xcite . to that aim two limits have to be made , the continuum limit sending @xmath20 and the thermodynamical limit sending @xmath21 . for a standard textbook on lattice qcd we refer to @xcite . a second possibility is to approximate qcd in the confined phase by effective theories , that describe for example only the low energy excitation . one of these is chiral perturbation theory ( chpt ) , see @xcite for a review , that describes the low momentum modes that appear after the spontaneous breaking of chiral symmetry , the so - called goldstone bosons . another such theory is rmt . it should be clear by now that rmt will not solve qcd ( nor chpt ) , in the continuum or on the lattice . however , it will describe certain aspects in an analytic fashion , namely the spectral properties of the small eigenvalues of the qcd dirac operator @xmath22 . in a finite volume the density of eigenvalues of @xmath22 satisfies the banks - casher relation @xcite @xmath23 which relates it to the described setup . in particular rmt will predict the detailed dependence on the parameters @xmath1 , @xmath10 , @xmath19 , and @xmath18 after being appropriately rescaled , as well as on a the zero - eigenvalues of @xmath22 which relate to a topological index @xmath24 . and most remarkably these predictions have been verified in comparison to lattice qcd by many groups . we will not repeat these findings here and refer to the literature at the end of each section . the idea to apply rmt to qcd goes back to the seminal works @xcite that started from the simplest case with @xmath25 which is called quenched approximation , at @xmath26 . the field has since developed enormously , leading to the detailed analytical knowledge that we will describe . there exist a number of excellent reviews already , @xcite , notably the lecture notes from the les houches session in 2004 by j. verbaarschot @xcite . this brings me to the main goals of these lecture notes , to be addressed in the order of the subsequent sections . what can we predict from rmt that can be compared to lattice qcd ? in section [ opdirac ] we will start with the solution of the corresponding rmt using the theory of orthogonal polynomials . this section is rather mathematical and contains no further physics input . the second question , what the limit is , in which qcd reduces to rmt , is addressed in section [ symm ] . here the global symmetries of qcd are explained , leading to chpt and eventually to a rmt description . the more physics inclined reader may jump to this section first . one of my personal motivations why to add another review on this topic is answered in section [ recent ] where the development in the last 10 years is reviewed , including some mathematical aspects . this covers the dependence on finite lattice spacing @xmath19 and on chemical potential @xmath10 , for an earlier review on the latter see @xcite . let us move to rmt , where i assume that the reader has already some working knowledge . rmt is a much older topic , beginning in the late 1920 s with wishart in mathematical statistics and in the late 1950 s with wigner and dyson in nuclear physics . the number of its applications is huge and still increasing , and we refer to @xcite for applications to quantum physics and to @xcite for a recent compilation containing physics ( including qcd in chapter 32 ) , mathematics and more . in these lecture notes we will focus only on the theory of orthogonal polynomials . for the standard gaussian ensembles they are covered in the last edition of mehta s classical book @xcite , for more recent monographs see @xcite . the symmetry class we will be interested in for the application to qcd is the chiral gaussian unitary ensemble ( chgue ) , which is also known as wishart or laguerre unitary ensemble , and extensions thereof . in the simplest case of @xmath25 and when all the abovementioned parameters from qcd are absent it is defined by the probability space of complex @xmath27 matrices @xmath28 . all its matrix elements are independent and share the same complex normal distribution . the mean density @xmath29 of the positive eigenvalues of @xmath30 , where @xmath31 denotes the hermitian conjugate , is well known in the limit @xmath32 . after an appropriate rescaling it is given by the marchenko - pastur law depicted in figure [ fig : mpdensity ] , cf . the lectures by bouchaud where this was derived . ( 12.2,4.5 ) which is the limiting global spectral density of the chgue . we also indicate the locations where different local statistics applies : they are given by i ) the bessel - kernel close to the hard edge , ii ) the sine - kernel in the bulk and iii ) the airy - kernel at the soft edge . [ fig : mpdensity],title="fig:",width=264 ] ( 0.2,-0.2)@xmath33 ( -7.8,4.0)@xmath29 ( -1.9,-0.2)iii ) ( -6.6,-0.2)i ) ( -4.5,-0.2)ii ) the same law is obtained when taking real or quaternion valued matrix elements , constituting the chiral gaussian orthogonal and symplectic ensemble ( chgoe and chgse ) , respectively . however , the local statistics of the 3 ensembles differs , and depends on the location in the spectrum we consider . for finite @xmath34 the eigenvalue statistics follows a determinantal ( chgue ) or pfaffian point process ( chgoe , chgse ) , and the limiting kernels in the various locations of the spectrum shown in figure [ fig : mpdensity ] are given in terms of bessel , sine or airy functions . in view of the relation ( [ bcrel ] ) we will be interested in the local statistics for the chgue at the origin also called hard edge . the other two symmetry classes at the hard edge will also find applications in qft as pointed out in @xcite , cf . @xcite , but will not be discussed here . in the limit where rmt applies the eigenvalues @xmath35 of the qcd dirac operator @xmath22 will be given by @xmath36 and thus come in pairs , where @xmath37 are the eigenvalues of @xmath30 ( the positive part @xmath38 are the singular values of @xmath28 ) . after this change of variables the marchenko - pastur law becomes a semi - circle , which is why the global density of @xmath22 at the origin will become constant ( and not divergent as in figure [ fig : mpdensity ] ) . for finite @xmath34 the chgue can be solved in terms of the classical laguerre polynomials . it will be the content of sections [ opdirac ] and [ recent ] to include more structure from qcd into this rmt , while maintaining its exact solvability . in this section we go directly to the rmt of qcd in an eigenvalue representation that generalises the chgue , by including the mass terms of @xmath0 quark flavours . in subsection [ defs ] we briefly recall the orthogonal polynomial ( op ) formalism and define the quantities of our interest . we then attack the mathematical problem of computing these by deriving properties of op with a general weight , in particular including averages of characteristic polynomials in subsection [ genweight ] . this allows us to determine all quantities explicitly in terms of standard laguerre polynomials in subsection [ alllaguerre ] , which facilitates the large-@xmath34 limit to be taken in subsection [ lim ] . we begin by stating the rmt that we want to solve in its eigenvalue representation , defined by the following partition function @xmath39 it gives the normalisation constant of the unnormalised joint probability density function ( jpdf ) @xmath40 of all eigenvalues . here we have included the dependence on the following parameters to be related to qcd later : @xmath41 called the topological index taking fixed values , the masses @xmath42 , @xmath43 , with @xmath0 counting the number of inserted characteristic polynomials . in the quenched theory with @xmath25 their product is absent . these parameters are all collected in the weight function @xmath44 and finally we also have introduced the vandermonde determinant @xmath45=\prod_{1\leq i < j\leq n}(x_j - x_i)\ . \label{delta}\ ] ] it depends on the set of all eigenvalues @xmath46 . the partition function eq . ( [ partitionfunct ] ) is relevant for qcd for @xmath47 , representing the chgue in the presence of @xmath0 characteristic polynomials ( mass terms ) . for completeness and in order to state some open problems later we have also introduced the chgoe and chgse with @xmath48 and 4 , respectively . all three rmt are also called wishart - laguerre ensembles and can be written in matrix representation , @xmath49\prod_{f=1}^{n_f}\det \left [ \begin{array}{cc } m_f 1_n&i w\\ iw^\dag & m_f 1_{n+\nu}\\ \end{array } \right ] \exp[-\mbox{tr}(ww^\dag)]\ . \label{matrixrep}\ ] ] here @xmath28 is an @xmath50 matrix taking values @xmath51 for @xmath52 respectively , and @xmath53 is its hermitian conjugate . the integration @xmath54 $ ] is over the flat lebesgue measure of all independent matrix elements . the identity matrices in front of the scalar mass variables @xmath55 of dimensions @xmath34 and @xmath56 are denoted by @xmath57 and @xmath58 , respectively . here we see that @xmath24 counts the number of zero eigenvalues of the wishart matrix @xmath59 , and the @xmath37 , @xmath60 denote its non - zero positive eigenvalues we always have an even number @xmath61 of doubly degenerate masses in eq . ( [ partitionfunct ] ) . we will come back to this at the end of subsection [ alllaguerre ] . ] . the change of variables from matrix elements to eigenvalues leads to the vandermonde determinant to the power @xmath52 times the @xmath24-dependent part as the jacobian , and we refer to standard text books on rmt for the derivation @xcite . in the lectures of j .- bouchaud these ensembles were discussed for @xmath25 , in particular when @xmath24 is of the order of @xmath34 . this matrix representation eq . ( [ matrixrep ] ) will play an important role later in section [ symm ] when relating the rmt to qcd based on symmetries . in the following we will mainly focus on the chgue with @xmath47 which is the rmt relevant for qcd . the values @xmath48 and 4 apply to other qfts and are very briefly discussed at the end of subsection [ alllaguerre ] . what are the quantities we would like to calculate as functions of the masses @xmath55 and @xmath24 , apart from the normalising partition function ? the jpdf represents a determinantal ( or pfaffian ) point process introduced in the lectures by a. borodin . therefore we can determine all @xmath62-point eigenvalue density correlations defined for all @xmath63 as @xmath64\ , \ \ \mbox{for } \ \ \beta=2\ . \label{rkb2}\end{aligned}\ ] ] in the second line valid for @xmath47 only they can be expressed in terms of the determinant of the kernel @xmath65 that does not contain the weights in our convention . it contains op with respect to the weight eq . ( [ weight ] ) , @xmath66 for @xmath67 similar expressions exist in terms of pfaffian determinants of a matrix kernel of op with respect to a skew symmetric inner product replacing eq . ( [ op ] ) . these are called skew op and we refer to @xcite for details . we choose the op in eq . ( [ op ] ) to be monic , @xmath68 , with squared norms @xmath69 . in particular eq . ( [ rk ] ) leads to the spectral density @xmath70 ( normalised to @xmath34 ) and to the determinantal expression for the jpdf , @xmath71 . two further quantities can be defined that are of our interest . the @xmath62-th gap probability defined for all @xmath63 reads @xmath72 it is the probability that @xmath62 eigenvalues are in @xmath73 $ ] and @xmath74 in @xmath75 . for example @xmath76 gives the gap probability that the interval @xmath73 $ ] is empty . if we order the eigenvalues @xmath77 the @xmath62-th eigenvalue distribution defined as @xmath78 gives the probability to find the @xmath62-th eigenvalue at @xmath79 , @xmath80 eigenvalues in @xmath73 $ ] and @xmath74 in @xmath75 . these individual eigenvalue distributions are normalised @xmath81 , as can be easily checked ( see e.g. the appendix of @xcite ) . we will be mostly interested in the example of the distribution of the smallest non - zero eigenvalue , @xmath82 . equations ( [ gapk ] ) and ( [ pk ] ) are not independent , in fact the latter follow by differentiation : @xmath83 which can be solved for @xmath84 as follows : @xmath85 the gap probabilities can also be expressed in terms of the kernel @xmath86 as a fredholm determinant , cf . the lectures by a. borodin and eqs . ( [ e0andreiev ] ) and ( [ expand ] ) , but we will use a different route to compute them explicitly . our goal is now to determine all the quantities that we have introduced in this subsection , including the partition function . for @xmath25 this would be an easy task as the polynomials orthogonal with respect to @xmath87 $ ] are the well known laguerre polynomials . it turns out that also for @xmath88 we can express all quantities in terms of laguerre polynomials , including the dependence on the parameter @xmath55 and @xmath24 that we are after . this will be the goal of the following 2 subsections . the first part of this subsection is standard material about op on the real line and can be found in any standard reference , e.g. in @xcite including also op in the complex plane needed later . for the second part with weights containing characteristic polynomials we refer to @xcite for more details . let us consider a general measurable weight function on the positive real half line with the condition that all moments exist , @xmath89 . the monic polynomials @xmath90 orthogonal with respect to @xmath91 can then be recursively constructed using gram - schmidt , expressing them as the ratio of two determinants : @xmath92\right)^{-1},\ ] ] see @xcite for details . it obviously satisfies @xmath93 . these polynomials obey a three - step recurrence relation , @xmath94 which can be see as follows . with the @xmath90 forming a complete set we can expand @xmath95 using that in the last integral @xmath96 is a polynomial of degree @xmath97 and the orthogonality eq . ( [ op ] ) it follows that @xmath98 for @xmath99 and @xmath100 . for the orthonormal polynomials @xmath101 the recursion eq . ( [ recurrence ] ) becomes more symmetric , @xmath102 and we obtain the christoffel - darboux formula for the kernel @xmath103 this simply follows by multiplying the sum by @xmath104 with @xmath105 , using the recursion and the fact that this is a telescopic sum . from lhpital s rule for the kernel at equal arguments we see that in the asymptotic limit @xmath32 we only need to evaluate the asymptotic of @xmath106 , @xmath107 , their derivatives , and the ratio of norms @xmath108 to determine the spectral density : @xmath109 let us give two examples . for the gue with weight @xmath110 $ ] on @xmath111 instead of @xmath112 we have @xmath113 the probabilist s hermite polynomials . second , for the laguerre weight @xmath87 $ ] , eq . ( [ weight ] ) with @xmath25 on @xmath112 , the monic op are given by the appropriately rescaled generalised laguerre polynomials @xmath114 : @xmath115 where @xmath24 can also be real here . the asymptotic analysis of the kernel ( and density ) can now be made , and we will come back to this in subsection [ lim ] . in the following we will express the partition function , corresponding ops , kernel and gap probability @xmath76 in terms of expectation values of characteristic polynomials in a first step , and then reduce them to determinants of laguerre polynomials in a second step . in order to prepare this the following exercise is very useful . consider the 2 vandermode determinants eq . ( [ delta ] ) inside the partition function ( [ partitionfunct ] ) for @xmath47 . by adding columns inside the vandermonde determinant we do not change its value . starting with the last column of highest degree we can thus express @xmath116 in terms of arbitrary monic polynomials @xmath117 : @xmath118 \ . \label{delta - id1}\ ] ] choosing those monic polynomials @xmath119 that precisely satisfy the orthogonality relation ( [ op ] ) we obtain for the partition function : @xmath120\right)^2\nonumber\\ & = & \sum_{\sigma,\sigma^\prime\in s_n}(-1)^{\sigma+\sigma^\prime } \prod_{l=1}^n\int_0^\infty dx_l w(x_l){p}_{\sigma(l)-1}(x_l){p}_{\sigma^\prime(l)-1}(x_l ) \ = \ \label{znorm}\end{aligned}\ ] ] in the second step we have laplace expanded both determinants into sums over permutations @xmath121 and used the orthogonality ( [ op ] ) . the computation of the partition function thus amounts to determine the squared norms of the ops with respect to the weight eq . ( [ weight ] ) . let us define the expectation value of an operator @xmath122 that only depends on the set of eigenvalues @xmath123 : @xmath124 then the following identity holds for the monic ops : @xmath125 it is given by the expectation value of a characteristic polynomial with respect to @xmath126 eigenvalues for an arbitrary degree , where @xmath127 . it can also be expressed in terms of matrices of size @xmath126 , choosing @xmath128 $ ] . this relation is known as the heine formula , dating back to the 19th century . the proof uses the simple identity for the vandermonde determinant , @xmath129 \ . \label{delta - id2}\ ] ] using this identity in the expectation value we arrive at the same situation as in eq . ( [ znorm ] ) , with one larger determinant : @xmath130 here the orthogonality of the polynomials enforces @xmath131 , and the remaining norms cancel due to eq . ( [ znorm ] ) . as a check also the leading coefficient of eq . ( [ opid ] ) can be compared , taking the asymptotic limit @xmath132 which leads to @xmath133 on both sides . note that eq . ( [ opid ] ) provides an @xmath126-fold integral representation of the @xmath134 for an arbitrary weight function . compared to the known single integral representation , e.g. for the hermite polynomials in the example in eq . ( [ hermite ] ) , we thus have the following duality relation @xmath135\ . % = \langle\det[x - ih]^n\rangle_1 .\ ] ] here we take a single average over @xmath136 copies of characteristic polynomials . the existence of such relations between 1- and @xmath136-fold averages of characteristic polynomials are known only for gaussian weights . they can also be derived using the supersymmetric method as reviewed in chapter 7 of @xcite by thomas guhr , see also @xcite for further dualities using op techniques . in the next step we will express the kernel itself as an expectation value of two characteristic polynomials , following the work of p. zinn - justin @xcite . namely it holds that @xmath137 the proof uses the same idea as before , now including each product into a different vandermonde determinant of size @xmath138 . we have @xmath139 compared to eq . ( [ proofpn ] ) the orthogonality only implies @xmath140 , and its value can still run over all possible values from 1 to @xmath138 . in the last line we give the expression for the kernel @xmath141 in terms of the christoffel - darboux formula ( [ cd ] ) , using monic polynomials instead . once again the leading coefficients of left and right hand side can be see to agree @xmath142 in the limit @xmath143 . an immediate question arises : what is the expectation value for more than two products of characteristic polynomials ? for that purpose let us introduce some more notation following baik , deift and strahov @xcite , to where we refer for more details . denote by @xmath144}(x)$ ] the op with respect to weight @xmath145}(x)=(\prod_{j=1}^l(y_j - x))w(x)$ ] for @xmath146 , with @xmath147}(x)=p_n(x)$ ] . the following relation holds which is called christoffel formula : @xmath148}(x)=\frac{1}{(x - y_1)\cdots(x - y_l ) } \left| \begin{array}{lll } p_n(y_1)&\ldots&p_{n+l}(y_1)\\ \vdots & & \vdots\\ p_n(y_l)&\ldots&p_{n+l}(y_l)\\ p_n(x)&\ldots&p_{n+l}(x)\\ \end{array } \right|\left(\det_{1\leq i , j\leq l}[p_{n+j-1}(y_i)]\right)^{-1}. \label{christoffel}\ ] ] as explained in @xcite this can be seen as follows . it is clear that the determinant in the middle , which we call @xmath149}(x)$ ] , is a polynomial of degree @xmath150 in @xmath33 . it has zeros at @xmath151 , and thus @xmath149}(x)/(x - y_1)\cdots(x - y_l)$ ] is a polynomial of degree @xmath136 in @xmath33 . thus we have that @xmath152}(x)}{(x - y_1)\cdots(x - y_l)}w^{[l]}(x)=0\ , \mbox{for}\ \ j=0,1,\ldots , n-1\ , \ ] ] due to @xmath153 being op with respect to weight @xmath91 , after cancelling the factors from @xmath145}(x)$ ] . the last factor in eq . ( [ christoffel ] ) ensures that @xmath144}(x)$ ] is monic , as can bee seen from taking the limit @xmath132 . the repeated product of the christoffel formula ( [ christoffel ] ) leads to the following theorem ( see theorem 2.3 @xcite ) for the average of products of @xmath97 characteristic polynomials with @xmath146 , which is due to brzin and hikami @xcite : @xmath154 = \prod_{j=0}^lp_n^{[j]}(y_{j+1})= \left\langle \prod_{i=1}^{n}\left(\prod_{j=1}^{l+1}(y_j - x_i ) \right)\right\rangle_n . \label{prodchar}\ ] ] for the last equality we have used eq . ( [ opid ] ) that leads to @xmath155}(y_{j+1})=\frac{\left\langle \prod_{i=1}^{n}\left((y_{j+1}-x_i)\prod_{p=1}^{j}(y_p - x_i ) \right)\right\rangle_n}{\left\langle \prod_{i=1}^{n}\left(\prod_{p=1}^{j}(y_p - x_i ) \right)\right\rangle_n}. \label{pnvev}\ ] ] in the product almost all factors cancel out . the case for a single characteristic polynomial with @xmath156 in eq . ( [ prodchar ] ) was already stated eq . ( [ opid ] ) . in @xcite in theorem 3.2 a further identity was derived for the expectation value of the product of an even number of characteristic polynomials , expressing it through a determinant of kernels divided by two vandermonde determinants , @xmath157= \left\langle \prod_{i=1}^{n}\left(\prod_{j=1}^{k}(\lambda_j - x_i ) ( \mu_j - x_i)\right)\right\rangle_n . \label{prod2char}\ ] ] the simplest example for this relation was derived in eq . ( [ kernelid ] ) . in @xcite this set of identities was further generalised , expressing the expectation value of arbitrary products of characteristic polynomials by a determinant containing both kernels and polynomials divided by two vandermonde determinants in many different and equivalent ways , cf . ( [ cprodchar ] ) . the simplest example with 3 products reads @xcite @xmath158 let us mention that in @xcite also expectation values of ratios of characteristic polynomials were determined . they can be expressed in terms of the cauchy transforms @xmath159 of the monic polynomials @xmath90 , @xmath160 as well as through mixed chistoffel - darboux kernels containing both cauchy transforms and polynomials . the simplest example is given by @xmath161 the fact that both sides agree @xmath162 for @xmath132 can be easily seen from the definition ( [ cauchy ] ) , expanding the geometric series and using the orthogonality of the polynomial @xmath163 to monic powers less than @xmath62 . a second particularly important example is the expectation value of the ratio of two characteristic polynomials , due to the following relation to the resolvent or stieltjes transform @xmath164 : @xmath165 the resolvent can be used to obtain the spectral density through the following relation @xmath166 \ . \label{inversion}\ ] ] there are many examples where the op technique is not available , but where the expectation value of ratios of characteristic polynomials can be found by other means , e.g. by using supersymmetry , replicas or loop equations . this then leads to an alternative way to determine the spectral density , or higher @xmath62-point correlation functions , by taking the average over @xmath62 resolvents and then taking the imaginary parts with respect to each of the @xmath62 arguments . we refer to chapters 7 , 8 and 16 in @xcite for further details and references . using the results from the previous subsection we are now ready to express the partition function , ops , kernel , and consequently all @xmath62-point correlation functions with @xmath0 characteristic polynomials , as well as the corresponding gap probabilities through expectation values of characteristic polynomials with respect to the quenched weight @xmath87 $ ] with @xmath25 . because the ops of this quenched weight are laguerre polynomials , everything will be finally expressed through these , which yields explicit and exact expressions for any finite @xmath34 and @xmath0 . in the next subsection [ lim ] will use them to take the large-@xmath34 limit based on the known asymptotic of the laguerre polynomials . we shall adopt the notation from the previous section , labelling all the above listed quantities by superscript @xmath167 $ ] compared to the quenched weight without superscript . we have for the partition function eq . ( [ partitionfunct ] ) with masses @xmath168}(\{m\})&=&\left(\prod_{j=1}^n\int_0^\infty dx_j\ x_j^\nu e^{-x_j}\prod_{f=1}^{n_f}(x_j+m_f^2)\right)\delta_n(\{x\})^2 = z_n\left\langle \prod_{j=1}^n\prod_{f=1}^{n_f}(x_j+m_f^2)\right\rangle_n \nonumber\\ & = & n!(-1)^{n_f(n+(n_f-1)/2)}\prod_{j=1}^{n}\gamma(j+\nu)\prod_{j=1}^{n+n_f}\gamma(j)\ \frac{\det_{1\leq i , j \leq n_f } \left[l_{n+j-1}^\nu(-m_i^2)\right]}{\delta_{n_f}(\{-m^2\})}. \label{zmassive}\end{aligned}\ ] ] this result simply follows from choosing @xmath169 in eq . ( [ prodchar ] ) , together with eqs . ( [ znorm ] ) and ( [ laguerre ] ) for the quenched polynomials . it is clear that if some of the masses become zero , say @xmath126 out of @xmath0 , from looking at eq . ( [ partitionfunct ] ) for @xmath47 without further calculation this leads to the shift @xmath170 in the remaining determinant of size @xmath171 . this property is called flavour - topology duality . in the case when all masses are degenerate , @xmath172 , lhopital s rule eventually leads to @xcite @xmath173}(m)=\frac{(-1)^{nn_f}}{\prod_{l=0}^{n_h-1}l!}\det_{0\leq i , j \leq n_f-1 } \left[(-1)^{n+i}(n+i)!(l_{n+i}^\nu(-m^2))^{(j)}\right]\ , \label{zmassivedeg}\ ] ] where the superscript @xmath174 denotes the @xmath175-th derivative with respect to the argument @xmath176 . the unquenched op @xmath177}(x)$ ] follow directly from eq . ( [ pnvev ] ) at @xmath178 , with @xmath179 . for example using eq . ( [ laguerre ] ) the op for the weight @xmath180}(x)=(x+m^2)x^\nu\exp[-x]$ ] the corresponding op read @xcite @xmath181}(x)=\frac{h_nk_{n+1}(x ,- m^2)}{p_n(-m^2)}=\frac{(-1)^{n+1}(n+1)!}{(x+m^2)l_n^\nu(-m^2)}(l_{n+1}^\nu(x)l_{n}^\nu(-m^2)-l_{n+1}^\nu(-m^2)l_{n}^\nu(x))\ . \label{pmassive}\ ] ] however , it is much simpler to directly use the kernel from eq . ( [ kernelid ] ) , rather than expressing it through the polynomials @xmath182}(x)$ ] using the christoffel - darboux identity ( [ cd ] ) . indeed the kernel is given by eq . ( [ kernelid ] ) as @xmath183}(x , y)&= & \frac{1}{h_{n-1}^{[n_f]}}\left\langle \prod_{i=1}^{n-1}(x - x_i)(y - x_i ) \right\rangle_{n-1}^{[n_f ] } = \frac{1}{h_{n-1}}\frac { \left\langle\prod_{i=1}^{n-1}\left((x - x_i)(y - x_i)\prod_{f=1}^{n_f}(m_f^2+x_i)\right ) \right\rangle_{n-1 } } { \left\langle\prod_{i=1}^{n}\prod_{f=1}^{n_f}(m_f^2+x_i ) \right\rangle_{n } } \nonumber\\ & = & \frac{(-1)^{n_f-1}(n+n_f)!}{\gamma(n+\nu)(y - x)\prod_{f=1}^{n_f}(y+m_f^2)(x+m_f^2 ) } \frac { \left| \begin{array}{lll } l_{n-1}^\nu(-m_1 ^ 2)&\ldots&l_{n+n_f}^\nu(-m_1 ^ 2)\\ \vdots & & \vdots\\ l_{n-1}^\nu(-m_{n_f}^2)&\ldots&l_{n+n_f}^\nu(-m_{n_f}^2)\\ l_{n-1}^\nu(x)&\ldots&l_{n+n_f}^\nu(x)\\ l_{n-1}^\nu(y)&\ldots&l_{n+n_f}^\nu(y)\\ \end{array } \right| } { \det_{1\leq f , g\leq n_f}[l_{n+g-1}^\nu(-m_f^2)]}. \label{kernelmassive}\end{aligned}\ ] ] in the first step after inserting eq . ( [ kernelid ] ) we have included the mass dependent norm @xmath184}$ ] into the massive partition function in the denominator . it is given by the product of its norms times @xmath185 . thus we have increasing the average from @xmath186 to @xmath34 . in the final result in the second line we have already taken out all signs and factorials from the two determinants and cancelled them partly . also the ratio of the two vandermonde determinants , one of which containing the arguments @xmath33 and @xmath187 has been simplified using the definition ( [ delta ] ) . the kernel eq . ( [ kernelmassive ] ) determines all @xmath62-point eigenvalue correlation functions from eq . ( [ rkb2 ] ) including their mass dependence . in particular for @xmath188 we obtain the spectral density with @xmath0 masses : @xmath189}(x)=x^\nu\exp[-x]\prod_{f=1}^{n_f}(x+m_f^2)\ k_n^{[n_f]}(x , x)\ , \label{r1massive}\ ] ] after applying lhpital s rule once . this leads to @xmath190 in the last row in the numerator . note that the product from the weight in front of the kernel cancels part of its denominator . let us now turn to the gap probabilities and distribution of smallest eigenvalues , starting with @xmath76 from eq . ( [ gapk ] ) . the following andriev integral identity ( cf . borodin s lectures ) can be used to get a first expression : @xmath191\det_{1\leq i , j \leq n}[\psi_i(x_j ) ] = n!\det_{1\leq i , j \leq n}\left[\int dx\phi_i(x)\psi_j(x)\right]\ . \label{andreiev}\ ] ] the only condition to hold is that all integrals of the functions @xmath192 and @xmath193 exist . the proof merely uses the laplace expansion of the left hand side : @xmath194 and integration over common arguments of @xmath195 and @xmath196 . this implies @xmath197 , which is yet another permutation . following eq . ( [ gapk ] ) we can thus write @xmath198}(s ) = \frac{1}{n!%\prod_{l=0}^{n-1}h_l } \prod_{p=1}^n\int_s^\infty dx_p \det_{1\leq i , j \leq n}\left[\sqrt{w^{[n_f]}(x_j)}\hat{p}_{i-1}^{[n_f]}(x_j)\right]^2 = \det_{1\leq i , j \leq n}\left[\delta_{ij}-\int_0^s dx w^{[n_f]}(x)\hat{p}_{i-1}^{[n_f]}(x)\hat{p}_{j-1}^{[n_f]}(x)\right]\ ! , \label{e0andreiev}\ ] ] which can be interpreted as a fredholm determinant . in the first step we have replaced the vandermonde determinants by determinants of monic polynomials , and then included the square root of the weight functions and of the norms stemming from the normalising partition function into the determinants , in order to make the polynomials orthonormal , @xmath199}(x_j)$ ] . in the next step we have used the andrief formula and the orthonormality on @xmath200 . knowing both the monic polynomials from eq . ( [ pnvev ] ) and their squared norms from @xmath201}=z_j^{[n_f]}/(j z_{j-1}^{[n_f]})$ ] in terms of expectation values of laguerre polynomials , eq . ( [ e0andreiev ] ) is a valid expression for the gap @xmath76 . however , we will use a more direct expression to be derived below . as a further remark we note that replacing @xmath202 in the first equality in ( [ e0andreiev ] ) we obtain an equivalent expansion of the fredholm determinant in the second equation : @xmath198}(s ) = 1+\sum_{k=1}^n\frac{(-1)^k}{k!}\int_0^s dx_1\ldots \int_0^s dx_k r_k^{[n_f]}(x_1\ldots , x_k ) . \label{expand}\ ] ] in order to express the gap probability directly as an expectation value let us go back to its definition ( [ gapk ] ) , which we state for general @xmath63 : @xmath203}(s)&=&\frac{1}{z_n^{(\beta , n_f,\nu)}}\left(\prod_{j=1}^n\int_s^\infty dx_j\ x_j^{\frac\beta2(\nu+1)-1}e^{-x_j}\prod_{f=1}^{n_f}(x_j+m_f^2)\right ) |\delta_n(\{x\})|^\beta\nonumber\\ & = & \frac{e^{-n s}}{z_n^{(\beta , n_f,\nu)}}\left(\prod_{j=1}^n\int_0^\infty dy_j\ y_j^0 ( y_j+s)^{\frac\beta2(\nu+1)-1}e^{-y_j}\prod_{f=1}^{n_f}(y_j+s+m_f^2)\right ) |\delta_n(\{y\})|^\beta . \label{gap - shift}\end{aligned}\ ] ] here we have shifted all integration domains , substituting @xmath204 , @xmath60 . the shift can be simply worked out , in particular it leaves the vandermonde determinant invariant , as a consequence the numerator is again given by a partition of @xmath205 masses with values @xmath206 for the first @xmath0 , and @xmath207 for the remaining ones , with an effective topological charge @xmath208 parameter satisfying @xmath209 . this idea has been used in several papers @xcite to give closed form expressions as an alternative to the fredholm determinant above . let us first consider @xmath47 , with @xmath210 . consequently eq . ( [ gap - shift ] ) can be written as @xmath211}(s)= e^{-ns } \frac{z_n^{(2,n_f+\nu,0)}}{z_n^{(2,n_f,\nu ) } } \left\langle \prod_{i=1}^n\left ( ( s+y_j)^\nu\prod_{f=1}^{n_f}(m_f^2+s+y_j)\right)\right\rangle_{n,\,\nu_{ef\!f}=0}. \label{gap - vevb2}\ ] ] the simplest examples with @xmath25 and @xmath212 thus read @xmath213 in the second line we have replaced the expectation value of a single characteristic polynomial eq . ( [ opid ] ) by the laguerre polynomial eq . ( [ laguerre ] ) and fixed the normalisation by the requirement @xmath214 . the quenched gap probability ( @xmath25 ) with arbitrary @xmath24 can be computed from the massive partition functions eq . ( [ zmassivedeg ] ) at complete degeneracy . the smallest eigenvalue distribution @xmath82 easily follows for the examples we have just given by differentiation of the gap probability , see eq . ( [ p1 ] ) . for higher gap probabilities and for computing @xmath84 directly the same trick from eq . ( [ gap - shift ] ) can be used . first consider only the integrals @xmath215 in the definitions ( [ gapk ] ) and ( [ pk ] ) . then do the shift @xmath204 , and perform the remaining integrals @xmath216 over the obtained expectation value of characteristic polynomials in the end . we refer to @xcite for more details . let us now briefly comment on @xmath67 , mainly because of recent developments and open questions in these symmetry classes . we begin with @xmath48 . here the effective topological charge is @xmath217 to satisfy @xmath218 in eq . ( [ gap - shift ] ) . the additional mass terms originating from the shift by @xmath219 in eq . ( [ gap - shift ] ) appear with multiplicity @xmath220 : @xmath221}(s)= e^{-ns } \frac{z_n^{(1,n_f+(\nu-1)/2,1)}}{z_n^{(1,n_f,\nu ) } } \left\langle \prod_{i=1}^n\left ( ( s+y_j)^{(\nu-1)/2}\prod_{f=1}^{n_f}(m_f^2+s+y_j)\right)\right\rangle_{n,\,\nu_{ef\!f}=1}. \label{gap - vevb1}\ ] ] that is for @xmath222 odd , @xmath223 we have @xmath224 extra masses @xmath207 compared to the @xmath0 shifted masses . here the simplest case is @xmath225 with @xmath25 , for which we obtain @xmath226 for arbitrary @xmath34 . it agrees with the result for @xmath47 at @xmath227 , eq . ( [ e0nu0 ] ) . for general odd @xmath24 the gap probabilities have been computed and we refer to @xcite for results . however , for @xmath228 even we obtain a half integer number @xmath229 of additional mass terms . this leads to the question of calculating expectation values including products of square roots of characteristic polynomials in eq . ( [ gap - vevb1 ] ) . in @xcite the expectation values needed to determine @xmath76 for arbitrary even @xmath228 and general @xmath0 have been computed for the chgoe , answering at least the question for the gap probability . the cases @xmath227 @xcite and @xmath230 @xcite were previously known from different considerations . namely in @xcite a recursive construction was made for the smallest eigenvalue for @xmath48 valid for all @xmath24 . however , in this approach the pfaffian structure appearing for @xmath231 is not at all apparent . independently in @xcite the question about expectation values of square roots of determinants has been asked and answered for special cases for the goe in the context of scattering in chaotic quantum systems . it is an open problem if the most general expectation value of such products ( or ratios ) of square roots has a determinantal or pfaffian structure as in eq . ( [ prodchar ] ) . let us turn to @xmath232 . the effective topological charge we obtain is @xmath233 to satisfy @xmath234 in eq . ( [ gap - shift ] ) . the fact that it is not integer does not pose a problem as the index of the generalised laguerre polynomials eq . ( [ laguerre ] ) in terms of which we have expressed expectation values of characteristic polynomials can be chosen accordingly . the more sever problem is the following . due to kramer s degeneracy the eigenvalues @xmath37 of a self - dual quaternion valued @xmath27 matrix @xmath235 always come in pairs , implying @xmath236=\prod_{j=1}^n(x - x_j)^2 $ ] . consequently the mass terms generated from eq . ( [ matrixrep ] ) always occur with an _ even _ power @xmath0 of two - fold degenerate masses in eq . ( [ partitionfunct ] ) . in contrast in eq . ( [ gap - shift ] ) we always need to evaluate an odd number of powers @xmath237 . thus as for @xmath48 with even @xmath24 we need to evaluate square roots of determinants in order to compute the gap probability for @xmath232 . this is an open problem so far and only the gap probability ( and smallest eigenvalue distribution ) with @xmath227 is known explicitly @xcite . a taylor series expansion exists though for @xmath238 , following from group integrals of kaneko - type @xcite . after having exhaustively presented the solution of the eigenvalues model eq . ( [ partitionfunct ] ) and its correlation functions for a finite number of eigenvalues @xmath34 at arbitrary @xmath0 for @xmath47 , we will now turn to the large-@xmath34 limit . as it is true in general in rmt one has to distinguish between different large-@xmath34 scaling regimes . the global spectral statistics , which includes the global macroscopic density given by the semi - circle law for the gue , or the marchenko - pastur law for the chgue , see figure [ fig : mpdensity ] , is concerned with correlations between eigenvalues that have many other eigenvalues ( in fact a finite fraction of all ) in between them . for a discussion of this large-@xmath34 limit alternative techniques are available such as loop - equations , and we refer to @xcite for a standard reference where all correlation functions including subleading contributions are computed recursively for @xmath47 . in this limit the fluctuations of the eigenvalues on a local scale ( of a few eigenvalues ) are averaged out , and typically expectation values factorise . in contrast when magnifying the fluctuations among eigenvalues at a distance of @xmath239 one speaks of microscopic limits . the value of @xmath240 and the form of the limiting kernel depend on the location in the spectrum , see figure [ fig : mpdensity ] . at the so - called soft edge of the marchenko - pastur law one finds the airy - kernel , whereas in the bulk of the spectrum the sine - kernel is found , see e.g. @xcite . here we will be interested in the eigenvalues in the vicinity of the origin presenting a hard edge with @xmath241 and the limiting kernel to be computed below is the bessel - kernel . why this limit is relevant for the application of rmt to qcd has been indicated already in the introduction after eq . ( [ bcrel ] ) . as a further consequence of this comparison to qcd we will change variables from wishart eigenvalues @xmath37 of @xmath30 to dirac operator eigenvalues @xmath36 ( more details follow in section [ symm ] ) . looking at the partition function eq . ( [ partitionfunct ] ) the dirac eigenvalues and masses have to be rescaled with the same power in @xmath34 . otherwise we would immediately loose the dependence on the masses . on the rmt side the hard - edge scaling limit zooming into the vicinity of the origin is conventionally defined including factors by taking @xmath32 and @xmath242 , such that the following product remains finite - independent here in contrast to other conventions . ] @xmath243 in the last subsection we have eventually expressed all quantities of interest in terms of the laguerre polynomials of the quenched weight . therefore the only piece of information we need to take the asymptotic limit of all these quantities is the well known limit @xcite : @xmath244 consequently all laguerre polynomials of positive argument turn into @xmath245-bessel functions . those polynomials containing negative arguments will turn into @xmath246-bessel functions , due to the relation @xmath247 for integer @xmath24 . note that both types of bessel functions also appear together , see e.g. in eqs . ( [ pmassive ] ) and ( [ kernelmassive ] ) . the asymptotics of the laguerre op eq . ( [ laguerrelim ] ) in fact holds for op with respect to much more general weight functions . this phenomenon is called universality . after first results in @xcite more sophisticated rigorous mathematical methods including the riemann - hilbert approach were developed , see chapter 6 by a. kuijlaars in @xcite for a detailed discussion and for references . in the following we will give a few examples for the hard edge limit , starting with the partition function . looking at eq . ( [ zmassive ] ) it is clear that we have to normalise it differently , such that the limit exists and gives a function of the limiting masses from eq . ( [ rmtmass ] ) . for @xmath248 this is very easy to do by dividing out the quenched partition function , and we define : @xmath249}(m)=m^{\nu}n^{-\frac{\nu}{2}}\frac{z_n^{[n_f=1]}(m)}{z_n n!}= m^{\nu}n^{-\frac{\nu}{2 } } l_n^\nu(-m^2)\ , \label{zrmtredef}\ ] ] leading to the limit @xmath250}(m)={\cal z}^{[n_f=1]}(\tilde{m})= i_\nu(\tilde{m})\ . \label{zrmtnf1}\ ] ] note that our conventions are chosen such that for small mass the limiting partition function behaves as @xmath251}(\tilde{m}\approx0)\sim \tilde{m}^\nu$ ] , rather than unity . the reason for this choice will also become transparent in the next section . for general @xmath0 we do not spell out the normalisation constant explicitly and directly give the limiting result for non - degenerate masses @xmath252}(\{m\})={\cal z}^{[n_f]}(\{\tilde{m}\})= 2^{\frac{n_f(n_f-1)}{2}}\prod_{j=0}^{n_f-1}\gamma(j+1)\ \frac{\det_{1\leq f , g\leq n_f}\left[\tilde{m}_f^{g-1}i_{\nu+g-1}(\tilde{m}_f)\right]}{\delta_{n_f}(\{\tilde{m}^2\})}\ , \label{zrmtnf}\ ] ] where we adopted the normalisation from @xcite . in the completely degenerate limit this leads to : @xmath253}\right|_{m_f = m } { \cal z}^{[n_f]}(\tilde{m } ) = { \det_{1\leq f , g\leq n_f}\left[i_{\nu+g - f}(\tilde{m})\right]}\ . \label{zrmtnfdeg}\ ] ] the direct limit @xmath32 of eq . ( [ zmassivedeg ] ) ( or lhopital s ) rule from eq . ( [ zrmtnf ] ) ) leading to a determinant of derivatives of @xmath246-bessel functions can be brought to this form of a toeplitz determinant using identities for bessel functions . as the next step we will directly jump to the limiting kernel , the reason being two - fold . first , it is shorter to directly construct all correlation function form the kernel , eq . ( [ kernelmassive ] ) , rather than constructing the kernel through the massive op s first . second and more importantly , the limit of the individual op s does not necessarily exist . while here in the hard edge limit this is not the case , the well - known sine- and cosine - asymptotic of the hermite polynomials in the bulk of the spectrum is only achieved after multiplying them with the square root of the weight function . we obtain for the limiting kernel from eq . ( [ kernelmassive ] ) , after appropriately normalising and changing to dirac eigenvalues , @xmath254}(\tilde{y}_1,\tilde{y}_2)&\sim & \lim_{\stackrel{\mbox{\small $ n\to\infty$}}{\mbox{\small $ x_1,x_2,m_f\to0 $ } } } ( w(x_1)w(x_2))^{\frac12}k_n^{[n_f]}(x_1,x_2)\nonumber\\ & = & \frac{1}{(\tilde{y}_1 ^ 2-\tilde{y}_2 ^ 2)\prod_{f=1}^{n_f}\sqrt{(\tilde{y}_1 ^ 2+\tilde{m}_f^2)(\tilde{y}_2 ^ 2+\tilde{m}_f^2 ) } } \frac { \left| \begin{array}{lll } i_\nu(\tilde{m}_1)&\ldots&\tilde{m}_1^{n_f+1}i_{n_f+\nu+1}(\tilde{m}_1)\\ \vdots & & \vdots\\ i_\nu(\tilde{m}_{n_f})&\ldots&\tilde{m}_{n_f}^{n_f+1}i_{n_f+\nu+1}(\tilde{m}_{n_f})\\ j_\nu(\tilde{y}_1)&\ldots&(-\tilde{y}_1)^{n_f+1}j_{\nu+n_f+1}(\tilde{y}_1)\\ j_\nu(\tilde{y}_2)&\ldots&(-\tilde{y}_2)^{n_f+1}j_{\nu+n_f+1}(\tilde{y}_2)\\ \end{array } \right| } { \det_{1\leq f , g\leq n_f}[m_f^{g-1}i_{\nu+g-1}(\tilde{m}_f)]}.\ \ \ \ \label{kernelmassivelim}\end{aligned}\ ] ] at @xmath25 it reduces to the standard bessel - kernel : @xmath255 here we have given two equivalent forms of the same kernel using an identity @xcite for @xmath245-bessel functions , @xmath256 . following eq . ( [ rkb2 ] ) we have already included the weight function into the limiting kernel above . while the exponential part cancels out due to @xmath257\to1 $ ] the remaining @xmath24-dependent part is contributing . we can thus immediately write down an example , the quenched microscopic density @xmath258 we have omitted here the @xmath24 delta - functions of the exact zero - eigenvalues . they will enter the discussion later in subsection [ wrmt ] where they can be seen . the density is plotted in figure [ fig : besseldensity ] for @xmath227 and 1 . the next step is to obtain the limiting distribution of the smallest eigenvalue . because we change from wishart to dirac eigenvalues @xmath259 , we obtain the following scaling limit for the examples of the gap probability ( [ e0nu0 ] ) and ( [ e0nu1 ] ) at @xmath25 , using the same scaling for the dirac eigenvalues as in eq . ( [ rmtmass ] ) : @xmath260 note that the extra power of @xmath34 from the shift ( [ gap - shift ] ) leads to the fact that here the exponential is not vanishing . the liming quenched gap probability for general @xmath24 follows from the completely degenerate massive partition function ( [ zrmtnfdeg ] ) : @xmath261\ . \label{equenched}\ ] ] the limiting distribution of the corresponding smallest dirac eigenvalue follows according to eq . ( [ p1 ] ) , after changing to dirac eigenvalues : @xmath262 where for the last step we have used an identity for bessel functions . the last two formulas are compared to the quenched bessel density eq . ( [ besseldensity ] ) in figure [ fig : besseldensity ] . it turns out that for general @xmath24 the derivative of the determinant in eq . ( [ equenched ] ) can be written again as a determinant , following the ideas sketched after eq . ( [ e0nu1 ] ) that the distribution of the smallest eigenvalue can itself be expressed in term of an expectation value of characteristic polynomials , @xmath263\ , \label{p1quenched}\ ] ] see @xcite for a detailed derivation , including masses . ( 12.2,5.5 ) ) and the distribution of the smallest eigenvalue for @xmath227 ( left ) and @xmath264 ( right ) using eqs . ( [ p1nu0lim ] ) and ( [ p1nu1lim ] ) . it can be nicely seen how the distribution of the smallest eigenvalues follows the microscopic density for small @xmath265 . the local maxima of the density further to the right mark the locations of the second , third etc . eigenvalues , cf . the limiting value @xmath266 that equals the macroscopic density at the origin is marked as a horizontal line . [ fig : besseldensity],title="fig:",width=264 ] ) and the distribution of the smallest eigenvalue for @xmath227 ( left ) and @xmath264 ( right ) using eqs . ( [ p1nu0lim ] ) and ( [ p1nu1lim ] ) . it can be nicely seen how the distribution of the smallest eigenvalues follows the microscopic density for small @xmath265 . the local maxima of the density further to the right mark the locations of the second , third etc . eigenvalues , cf . the limiting value @xmath266 that equals the macroscopic density at the origin is marked as a horizontal line . [ fig : besseldensity],title="fig:",width=264 ] ( 0.1,0.2)@xmath265 ( -7.1,0.2)@xmath265 ( -2.5,3.5)@xmath267 ( -10.8,3.5)@xmath267 ( -3.5,1.4)@xmath268 ( -11.5,1.4)@xmath268 it should be mentioned that alternative representations exist for the gap probability eq . ( [ p1quenched ] ) that are based on fredholm determinant analysis @xcite . they are given in terms of an exponential of an integral of the solution of a painlev v equation , which is valid for any real @xmath269 . for integer values of @xmath24 this expression reduces to the simpler forms that we have given here . furthermore , in @xcite representations in terms of hypergeometric functions with matrix argument were given that are valid for all @xmath270 . in @xcite the equivalence to the above expressions for @xmath47 was proved . the distributions that we have computed so far , as for example the density and smallest eigenvalue in figure [ fig : besseldensity ] have been extensively compared to numerical solutions from lattice qcd . for example the quenched densitiy in figure [ fig : besseldensity ] was matched with qcd in @xcite , whereas the quenched distribution of the smallest eigenvalue also given in figure [ fig : besseldensity ] was compared in @xcite to qcd for different values of @xmath24 ( and also to other qfts corresponding to @xmath67 ) . a matching to unquenched results for the smallest eigenvalues can be found in @xcite , and we refer to @xcite for a more exhaustive discussion of the lattice results . apart from giving non - specialists some idea about the theory of quantum chromodynamics ( qcd ) we will have to answer three major questions in this section : * how is the dirac operator defined the spectrum of which can be described by rmt ? * what are the global symmetries of qcd that determine the rmt to be used ? * what is the limit that leads from qcd to this rmt ? let us give first short answers that also indicate how this section will be organised . the eigenvalues of the qcd dirac operator will be partly described by rmt , actually only its smallest eigenvalues , and we will introduce this operator first in subsection [ qcd ] . the anti - hermiticity properties of the dirac operator and the concept of chiral symmetry will allow us to introduce some formal aspects and the definition of the qcd partition function . the breaking of chiral symmetry in subsection [ xsb ] then leads us to the first approximation of qcd , to chiral perturbation theory that only contains the lightest excitations , the goldstone bosons . in a last step in subsection [ echpt ] this interacting non - linear , effective qft is simplified in the epsilon - regime , that agrees with rmt to leading order . in what follows it is important that we will talk about space - time coordinates @xmath271 with components @xmath272 , @xmath273 equipped with the euclidean metric @xmath274 given by the kronecker symbol . this so - called wick rotation from minkowski to euclidean metric is necessary for two reasons : only in this setting does the dirac operator have defined anti - hermiticity properties , and second only for qcd on a euclidean space - time lattice numerical solutions of qcd can be easily performed to which we can compare our rmt . we use einstein s summation conventions meaning that greek indices appearing twice are automatically summed over from 1 to 4 . the qcd dirac operator is a linear , matrix valued differential operator defined as follows : @xmath275 where @xmath276 . the @xmath277 are the dirac - gamma matrices in euclidean space - time , satisfying the clifford algebra @xmath278 the curly brackets denote the anti - commutator . a standard representation as @xmath279 matrices is given as follows in terms of the pauli matrices @xmath280 , @xmath281 : @xmath282 all matrices are hermitian , @xmath283 , including @xmath284 . this is only true in the euclidean setting . obviously one can construct projection operators as @xmath285 the real parameter @xmath6 in the dirac operator is the coupling constant of the strong interactions , thus @xmath286 corresponds to the free dirac operator without interactions . it acts on 4-vectors @xmath287 called spinors in minkowski space , due to their transformation properties there which we do nt need to specify here . the real vector potential @xmath288 is an element of this lie group for the three colours ( the subscript @xmath289 stands for colour ) . it is a scalar with respect to the @xmath277 s , just as the partial derivatives in eq . ( [ dirac ] ) . the @xmath287 come in 3 copies with colours blue @xmath290 , green @xmath291 and red @xmath292 , and so the interacting dirac operator acts on the full vector @xmath293 . for each quark flavour up , down etc . we have such a vector @xmath294 with @xmath295 . in the standard model of elementary particle physics we have @xmath3 , but often we will keep @xmath0 fixed as a free parameter . we could have written the dirac operator with tensor product notation to underline these structures , but the precise @xmath296 structure will not be crucial in the following . we are now ready to state some important global symmetries of the dirac operator . first , it follows from the algebra ( [ clifford ] ) that @xmath297 the dirac operator is block off - diagonal also called chiral , where the symbol @xmath298 still denotes a differential operator that depends on @xmath33 and @xmath299 . suppose we can find the eigenfunctions @xmath300 of the dirac operator labelled by @xmath62 : @xmath301 in a suitably regularised setting ( e.g. a finite box ) . because of the anti - hermiticity the eigenvalues @xmath302 are purely imaginary . furthermore , because of eq . ( [ dchiral ] ) multiplying equation ( [ eigenfunct ] ) by @xmath284 leads to different eigenfunctions @xmath303 with eigenvalues @xmath304 , provided that @xmath305 . thus the non - zero eigenvalues of the dirac operator come in pairs @xmath306 . the euclidean qcd partition function or path integral as it is called in qft can be formally written down as follows : @xmath307\int[d\psi]\exp\left [ -\int d^4x \sum_{q=1}^{n_f}\overline{\psi}^q(x)(d\!\!\!\!/+m_q)\psi^q(x)-\int d^4x \frac12\mbox{tr}(f_{\mu\nu}f_{\mu\nu } ) \right]\nonumber\\ & = & \int[da_\mu]\prod_{q=1}^{n_f}\det[d\!\!\!\!/+m_q]\exp\left [ -\int d^4x \frac12\mbox{tr}(f_{\mu\nu}f_{\mu\nu})\right]\ . \label{zqcd}\end{aligned}\ ] ] here @xmath308 is the dirac conjugate and the @xmath296 field strength tensor is defined by the commutator @xmath309=-ig_sf_{\mu\nu}$ ] . the integration over @xmath310 and @xmath311 are formal ( path ) integrals that we will not specify in the following , see standard textbooks on qcd for a discussion as @xcite . in the second equation however we have formally integrated out the quarks due to the following observation . for a complex @xmath34-vector @xmath312 and @xmath27 matrix @xmath313 we have @xmath314\sim \frac{1}{\det[b]}\ , \ \ \int d^{2n}\psi\exp[-\psi^\dag b \psi]\sim { \det[b]}\ , \label{detintegrals}\ ] ] whereas in the second identity we have integrated over a complex vector @xmath315 of anticommuting variables . integration and differentiation over such fermionic variables can be defined in a precise way , and we refer to chapter 7 in @xcite for details . the reason for fermions to be represented by anti - commuting variables is that in the canonical quantisation of free dirac fields one has to use anti - commutators for these fields . comparing eq . ( [ dchiral ] ) and the second line of eq . ( [ zqcd ] ) with eq . ( [ matrixrep ] ) we already get a first idea about the random matrix approximation of qcd , as it was suggested initially in @xcite : the dirac operator from eq . ( [ dchiral ] ) is replaced by a constant matrix with the same block structure , @xmath316 , and the average over the fields strength tensor is replaced by a gaussian average over @xmath28 . we further learn in this approximation that the eigenvalues @xmath317 of the gaussian random matrix @xmath30 correspond to the squared eigenvalues of the dirac operator rotated to the real line , and that the number of eigenvalues @xmath34 is proportional to the dimension of the qcd dirac operator truncated in this way . however , as we will explain in the next subsection the precise form of the limit from qcd to rmt was better understood later , representing a controlled approximation . furthermore , in the case when all masses vanish @xmath318 the action in the exponent in the first line of eq . ( [ zqcd ] ) has an additional symmetry . defining the vector @xmath319 as well as the projections @xmath320 onto left ( l ) and right ( r ) chirality we can write the fermionic part of the action as @xmath321 it is invariant under the global rotations @xmath322 with @xmath323 and likewise for @xmath324 . in contrast , even when all masses are equal , @xmath325 , the mass term @xmath326 is only invariant under the diagonal transformation with @xmath327 . in qcd the resulting global symmetry is @xmath328 , with the @xmath329-factors called @xmath330 for vector and @xmath331 for axial - vector split off . the latter is broken through quantum effects also called an anomaly , whereas the former remains unbroken through the mass term . this leads to the explicit chiral symmetry breaking pattern in qcd : @xmath332 in this subsection we will describe the first step of simplifying qcd to what is called its low - energy effective theory . we begin with the following observation . it is found that the qcd vacuum @xmath333 , the ground state with the lowest energy , breaks chiral symmetry _ spontaneously _ , in precisely the same way as the theory with equal masses in eq . ( [ xsbpattern ] ) breaks it explicitly : @xmath334 the parameter @xmath8 is called the chiral condensate and it is the order parameter for the spontaneous breaking of the chiral symmetry ( [ xsbpattern ] ) . it turns out that for qcd the low energy phase with non - vanishing @xmath11 coincides with the phase or different representations this is not necessarily the case . ] where the quarks and the gluons , the carriers of the strong interaction represented by @xmath299 , are confined to colourless objects , see the phase diagram in figure [ fig : phasediag ] . in the phase with @xmath11 we need lattice qcd or other effective field theory description ( such as rmt ) because a perturbative expansion in the coupling constant @xmath6 will not work here , because the coupling is large . an important consequence was drawn by banks and casher @xcite , relating the global continuum density @xmath335 of the dirac operator eigenvalues in eq . ( [ eigenfunct ] ) at the origin , suitably regularised in a finite volume @xmath18 , to @xmath8 : @xmath336 here we added the second order term to eq . ( [ bcrel ] ) derived by stern and smilga @xcite , containing the pion decay constant @xmath337 as an additional parameter . the first term leads to the following interpretation : at the origin the global density is constant , and thus the average distance between eigenvalues there is @xmath338 . this is very different from free particles in a box where @xmath339 with @xmath340 in 4 euclidean dimensions . one can thus say that the smallest dirac operator eigenvalues make an important contribution to build up the chiral condensate @xmath11 , by piling up very closely spaced at the origin . we are thus lead to define rescaled , dimensionless eigenvalues and masses , as they appear on the same footing inside the determinant in eq . ( [ zqcd ] ) , together with a rescaled microscopic density of the dirac operator @xmath341 : @xmath342 this is called the microscopic limit of qcd . it has to be compared e.g. to eq . ( [ besseldensity ] ) for @xmath25 , after identifying rescaled rmt eigenvalues @xmath343 and masses @xmath344 from eq . ( [ rmtmass ] ) . the breaking of a global continuous symmetry in any theory leads to goldstone bosons . this can be visualised by analogy in figure [ fig : goldstone ] , showing the classical potential of a particle before ( left ) and after symmetry breaking ( right ) . in the left picture with a convex potential any excitation to make a particle move in the potential costs energy . after symmetry breaking the curvature at the origin has changed . consequently there now exists the possibility to excite a mode in angular direction along the valley of the potential , without costing potential energy . modes that do need energy to be excited still exist in radial direction . ( 12.2,5.5 ) before symmetry breaking ( @xmath345 , left ) and after breaking ( @xmath346 , right ) , obtained simply by switching the sign of the quadratic term . [ fig : goldstone],title="fig:",width=264 ] before symmetry breaking ( @xmath345 , left ) and after breaking ( @xmath346 , right ) , obtained simply by switching the sign of the quadratic term . [ fig : goldstone],title="fig:",width=264 ] ( -5.0,0.7)@xmath33 ( -12.0,0.7)@xmath33 ( -7.0,4.8)@xmath347 ( -0.9,1.4)@xmath187 ( -7.9,1.4)@xmath187 ( -14.1,4.8)@xmath348 in our analogy particles that require energy to be excited correspond to massive particles . consequently after symmetry breaking there appear massless modes - the goldstone bosons - and modes that remain massive . if in qcd chiral symmetry was an exact symmetry we would expect to observe exactly massless goldstone bosons in the phase with broken symmetry . however , it is only an approximate symmetry due to @xmath349 , leading to approximately massless bosons after symmetry breaking . these are the 3 light pions observed in nature , @xmath350 and @xmath351 , with masses of about 135 mev compared to the heavier non - goldstone particles with masses of 800 -1000 mev . recall that in the confined phase we only observe colourless bound states made from two quarks , mesons , which are bosons , from three quarks , baryons , which are fermions , or more . the pions are unstable and decay mostly into a muon and its anti - neutrino for @xmath352 , the corresponding anti - particles for @xmath353 , or 2 photons for @xmath351 . the number of goldstone bosons equals the dimension of the coset @xmath354 of the groups before and after breaking , which is @xmath355 , the dimension of @xmath356 in our case . to consider the pions as goldstone bosons thus means that we only keep the lightest quarks up and down , so that @xmath4 in our theory . for many purposes at low energy this is a good approximation . sometimes also the strange quark is included , with @xmath357 leading to consider the 8 lightest particles as goldstone bosons . if we neglect all non - goldstone bosons in our theory - as they will not be relevant at low enough energies - and formally integrate them out we arrive at the low - energy effective theory of the goldstone bosons , @xmath358\exp\left[-\int d^4x \frac{f_\pi^2}{4}\mbox{tr}\left({\partial_\mu u(x)}{\partial_\mu u(x)^\dag}\right ) + \int d^4x\frac{\sigma}{2}\mbox{tr}\left(m(u(x)+u^\dag(x))\right ) \right ] \label{zchpt}\ ] ] where @xmath359 , \ m=\mbox{diag}(m_1,\ldots , m_{n_f } ) . \label{u(x)}\ ] ] here the implicit sum over @xmath290 runs from 1 to @xmath355 . for @xmath4 the 3 fields @xmath360 actually denote the pion fields , with @xmath361 and @xmath362 . in eq . ( [ u(x ) ] ) we have already split off the constant , @xmath33-independent modes @xmath363 of the fields @xmath360 for later convenience . usually at this stage @xmath363 is set to the identity . ( [ zchpt ] ) represents the partition function or path integral of chiral perturbation theory ( chpt ) , which replaces that of qcd eq . ( [ zqcd ] ) in the low - energy regime . in eq . ( [ zchpt ] ) we only give the leading order ( lo ) lagrangian , all higher order terms in powers of @xmath364 and its derivatives that are invariant exist and we will come back to the question when they contribute in the next section . for more details about chpt we refer to @xcite . chpt is an interacting , non - linear qft . when expanding the exponential in eq . ( [ u(x ) ] ) in powers of @xmath360 and considering only quadratic terms we obtain the standard lagrangian for scalar fields that leads to the klein - gordon equation . from that expansion we can make the following identification for the masses @xmath365 of the pion fields @xmath360 , the gell - mann oakes renner ( gor ) relation : @xmath366 the qft in eq . ( [ zchpt ] ) is not yet related to rmt , and we turn to this limit in the next subsection . the so - called epsilon - regime of chpt ( echpt ) was introduced by gasser and leutwyler @xcite before a rmt of qcd was constructed . their motivation was to introduce a regime where analytic computations in chpt were feasible , in particular to understand the mass dependence of the partition function and the role of topology . in particular leutwyler and smilga later computed the partition function analytically in this regime and deduced so - called sum rules for the dirac operator eigenvalues @xcite . working in a finite volume @xmath340 of linear size @xmath126 , the authors of @xcite introduced the following scaling : @xmath367 this scaling implies in particular @xmath368 where the former follows from the microscopic scaling limit of the banks - casher relation ( [ microlim ] ) , and the latter ensures that the kinetic term of the lo action is dimensionless . the physical interpretation of the epsilon - regime follows from the scaling of the pion mass , which is implied from multiplying the gor relation eq . ( [ gor ] ) with @xmath18 . it has to hold that @xmath369 , such that both sides are of the order @xmath370 . in other words , the compton wavelength of the pion fields is much larger than the size of the box : @xmath371 which is clearly an unphysical regime . normally one would consider this to be undesirable , because in order to observe physical , propagating pions one would choose a scaling such that the pions do fit into the box @xmath18 , with @xmath372 . in the so - called @xmath373-regime both momenta and pion mass are scaled in the same way , @xmath374 , where such an inequality can be satisfied . we will come back to this issue very briefly below . in the epsilon - regime introduced above we obtain the following limit @xcite @xmath375 \exp\left[-\int d^4x \frac12\partial_\mu\pi^b(x)\partial_\mu\pi^b(x ) \right]\int_{su(n_f)}du_0\exp\left[\frac12 \sigma v\mbox{tr}(m(u_0+u_0^\dag))+{\cal o}(\epsilon^2)\right ] . \label{zechpt}\ ] ] here the propagating modes @xmath360 completely factorise from the zero - modes @xmath363 and contribute to the partition function as a free gaussian field theory - which is just a constant factor . the reason that we do not expand in the constant modes of the pions as well , but keep them as a full matrix @xmath363 as in eq . ( [ u(x ) ] ) is that their quantum fluctuations are of @xmath376 , see @xcite for a more detailed discussion . all higher order terms including those not given in the chiral lagrangian ( [ zchpt ] ) are subleading in this limit . it is in this limit that the mass - dependence of the limiting partition function eq . ( [ zechpt ] ) and of rmt agree , and in fact also all dirac operator eigenvalue correlation functions . before we make this statement more precise let us add further comments . first , it is not only possible to make analytical computations in the epsilon - regime , but also to compare them to lattice qcd . on the lattice the simulation parameters can be freely chosen , such that the epsilon - regime is reached and a quantitative comparison is possible . second , it is possible to compute so - called one or higher loop corrections to the partition function ( and correlation functions ) eq . ( [ zechpt ] ) . for example writing down the @xmath377 term explicitly , expanding the exponential and using wick s theorem with the finite volume propagator corresponds to computing the one - loop corrections . third , one may perform also perturbative computations in the physical @xmath373-regime , or even chose an extrapolation between the two regimes , cf . @xcite for a discussion and references . let us now make the map between rmt and echpt more precise . for that purpose it is convenient to introduce an extra term in the qcd lagrangian , the so - called theta - vacuum term containing the dual field strength tensor @xmath378 . it is defined using the totally antisymmetric epsilon - tensor @xmath379 in 4 dimensions . the extra term reads in appropriate normalisation : @xmath380 it is a topological term , and here @xmath24 is the number of left handed @xmath381 minus right handed @xmath382 zero modes of the dirac operator eq . ( [ dirac ] ) - which so far has not made an appearance in this section about qcd . due to the atiah - singer index theorem @xmath24 is equal to the winding number of the gauge fields @xmath299 which is also a topological quantity . in the partition function eq . ( [ zqcd ] ) we have thus implicitly summed over all the topologies , which can now be written as follows : @xmath383 here @xmath384 denotes the qcd partition function at fixed topology . inverting this equation it can be written as @xmath385 what is the advantage ? in rmt in the previous section we always had a fixed @xmath386 - to which we should now compare for non - negative values . ] . furthermore , the unitary group integral over @xmath356 that we found in the epsilon - regime eq . ( [ zechpt ] ) can now be promoted to an integral over the full unitary group , according to @xmath387^\nu \label{groupint}\ ] ] defining the rescaled quark masses in the epsilon - regime as follows , @xmath388 we obtain the limiting relation between the mass - dependent part of the qcd partition function in the epsilon - regime at fixed topology and rmt : @xmath389}(\{\hat{m}\})&=&\int_{u(n_f ) } du_0\det[u_0]^\nu\exp\left[\frac12 \mbox{tr}(\hat{m}(u_0+u_0^\dag))\right]\nonumber\\ & = & 2^{\frac{n_f(n_f-1)}{2}}\prod_{j=0}^{n_f-1}\gamma(j+1 ) \frac{\det_{1\leq i , j\leq n_f}[\hat{m}_i^{j-1}i_\nu(\hat{m}_i)]}{\delta_{n_f}(\{\hat{m}^2\})}= % \lim_{\stackrel{\mbox{\small $ n\to\infty$}}{\mbox{\small $ m_f\to0 $ } } } { \cal z}^{[n_f]}(\{\hat{m}\})\ . \label{lsint}\end{aligned}\ ] ] in the last equality with the limiting rmt partition function from eq . ( [ zrmtnf ] ) we have identified rescaled random matrix and qcd quark masses @xmath344 for all flavours . the computation of the integral over the appropriately normalised haar measure of the unitary group @xmath390 in eq . ( [ lsint ] ) goes back to @xcite , see also @xcite and @xcite for our normalisation . a concise derivation in terms of group characters was presented in @xcite . in the simplest case with @xmath248 it is easy to see that : @xmath391}(\hat{m})=\int_{-\pi}^\pi\frac{d\theta}{2\pi } e^{i\nu\theta}\exp\left [ \frac12 \hat{m}(e^{i\theta}+e^{-i\theta})\right]=i_\nu(\hat{m})\ , \label{zechptnf1}\ ] ] which is one of the standard definitions of the modified bessel - function . furthermore , we can also verify the completely degenerate case @xmath392 , by applying the andrief formula ( [ andreiev ] ) . diagonalising @xmath363 by a unitary transformation @xmath393 , @xmath394 , we obtain again the squared vandermonde determinant as a jacobian : @xmath395^\nu\exp\left[\frac{\hat{m}}{2}\mbox{tr}(u_0+u_0^\dag)\right]=const . \prod_{j=1}^{n_f}\left(\int_{-\pi}^\pi\frac{d\theta_j}{2\pi}\exp[i\nu\theta_j+\hat{m}\cos(\theta_j)]\right ) |\delta_{n_f}(\{e^{i\theta}\})|^2\nonumber\\ & = & n_f!\ const.\det_{1\leq j , k\leq n_f}\left [ \int_{-\pi}^\pi\frac{d\theta}{2\pi}\exp[i(\nu+k - j)\theta+\hat{m}\cos(\theta)]\right]=\det_{1\leq j , k\leq n_f}[i_{\nu+k - j}(\hat{m})]\ . \label{zdeg}\end{aligned}\ ] ] it agrees with eq . ( [ zrmtnfdeg ] ) from rmt , after appropriately normalising . we observe that both for degenerate and non - degenerate masses the partition functions with @xmath0 flavours eqs . ( [ lsint ] ) and ( [ zdeg ] ) are determinants of single flavour partition functions with @xmath248 . the matching of the limiting rmt partition function and the group integral from the epsilon - regime of chpt at fixed topology is only part of their equivalence . in particular this matching does not imply that all dirac operator eigenvalue correlation functions agree . for example at @xmath248 the echpt partition function agrees for all three symmetry classes @xmath52 @xcite , while their densities differ . at first sight it is not clear how to access the dirac operator spectrum , once we have moved from the quarks to the goldstone bosons as fundamental low energy degrees of freedom . however , looking at the alternative construction of the spectral density via the resolvent eq . ( [ resolvent ] ) helps . consider adding an auxiliary pair of quarks with masses @xmath33 and @xmath187 respectively to the qcd lagrangian eq . ( [ zqcd ] ) . if we arrange for the second quark to be bosonic , we obtain the following average with respect to the unperturbed qcd partition function , at fixed topology : @xmath396}{\det[y - d\!\!\!\!/\ ] } \right\rangle_\nu^{\rm qcd}= \frac{\int_{u(n_f+1|1 ) } du_0\ \mbox{sdet}[u_0]^\nu\exp\left[\frac12 \mbox{str}(\hat{{\cal m}}(u_0+u_0^\dag))\right ] } { { \cal z}_{\rm echpt}^{[n_f]}(\{\hat{m}\})}\ , \label{qcdresolvent}\ ] ] with @xmath397 , in the limit of the epsilon - regime . because of the bosonic nature of the second auxiliary quark in the denominator the integral is now over the supergroup @xmath398 with the corresponding supersymmetric trace and determinant , str and sdet respectively , see e.g. chapter 7 of @xcite . its evaluation and differentiation to obtain the resolvent has to be done carefully . in @xcite to where we refer for details it was shown in this fashion that the microscopic density eq . ( [ besseldensity ] ) and its extension to arbitrary @xmath0 follows from eq . ( [ qcdresolvent ] ) , thus establishing the equivalence to rmt on the level of the microscopic density . in order to know the distribution of the smallest eigenvalues the knowledge of all density correlation functions is necessary , see eq . ( [ expand ] ) , as was pointed out in @xcite . the equivalence proof for all @xmath62-point density correlation functions was achieved in @xcite by directly matching the @xmath62-point resolvent generating functions with @xmath399 additional ratios of quark determinants of both theories , using superbosonisation techniques . a few comments regarding universality and corrections to the rmt regime are in place here . already the fact that two apparently different theories as echpt and rmt yield the same dirac operator correlation functions is a striking result which is based on symmetries and universality . it is thus surprising that even when taking into account one - loop corrections to echpt the same universal correlation functions were found in @xcite . the only corrections to be made are subsumed in a modification @xmath400 , and @xmath401 , where @xmath402 and @xmath403 are constants that depend on @xmath0 , @xmath337 and the geometry of the discretisation , see e.g. @xcite for details and references . so far @xmath337 did not appear in eigenvalue correlation functions , but later when including a chemical potential it will . this modification means that the chiral condensate ( and pion decay constant ) and thus the rescaled masses in eq . ( [ hatm ] ) and eigenvalues simply get renormalised accordingly . only two - loop corrections within echpt contain non - universal terms that lead out of the rmt universality class @xcite . however , even at the lo the agreement between the rmt and echpt rapidly breaks down when going to higher energies , meaning to higher rescaled dirac operator eigenvalues . at some point the fluctuating modes will start to contribute , and eventually also the chpt approach breaks down as contributions from full qcd will appear . the scale where this happens has been called thouless energy @xcite in analogy to applications in condensed matter physics , to where we refer for details . the point where this happens scales with @xmath404 and @xmath405 . beyond this scale one sees a rise of the global spectral density as in eq . ( [ bcss ] ) , away from the rmt prediction . the scaling of the thouless energy has been confirmed from lattice data in a different theory , in qft with two colours @xcite . for the same theory deep in the bulk regime the local statistics was seen to agree again with rmt bulk statistics of the gse @xcite , which is the non - chiral @xmath232 symmetry class . here we have already mentioned one of the two different symmetry classes corresponding to @xmath48 and 4 in eq . ( [ partitionfunct ] ) . they are also relevant but for different qfts other than qcd , e.g. with only two colours @xmath406 . these have chiral symmetry patters different from eq . ( [ xsbpattern ] ) and we refer to @xcite for a detailed discussion . the matching with the 3 chiral rmts in eq . ( [ partitionfunct ] ) was initially proposed in @xcite . while the agreement with lattice simulations of the respective theories ( cf . @xcite ) leave little doubt that these rmt are equivalent to the epsilon - regime of the corresponding chpt , such an equivalence has not been established even on the level of the partition function for general @xmath0 . this is due to the more complicated structure of the corresponding group integral , see however @xcite and @xcite for a discussion . let us briefly comment also on the influence of the dimension . in 3 space - time dimensions there is no chiral symmetry , and the rmt for qcd is given by the gue instead @xcite , including mass terms . in 2 dimensions which again has chiral symmetry the classification is much richer , and we refer to @xcite for details . this makes contact with the symmetry classification of topological insulators @xcite and the so - called bott - periodicity in any dimension . in this section we will cover some recent developments in the application of rmt to qcd , where the detailed dependence of the partition function and dirac operator eigenvalue correlation functions on finite lattice spacing @xmath19 or chemical potential @xmath10 are computed . in the first subsection [ wrmt ] we will study the influence of a finite space - time lattice on the spectrum as it is expected to be seen in numerical simulation of qcd on a finite lattice far enough from the continuum . the corresponding rmt and effective field theory that we will consider go under the name of wilson ( w)rmt as introduced in @xcite , and wilson ( w)chpt , see @xcite for a standard reference . we will only sketch the ideas beginning with the symmetries and wchpt , and then write down the joint density of eigenvalues of the hermitian wilson dirac operator . its solution and its non - hermitian part will be referred to the literature . the addition of a chemical potential term for the quarks to the dirac operator renders its eigenvalues complex . the solution of the corresponding rmt using op in the complex plane is sketched in the second subsection [ rmtmu ] where we will give some details . the first application of such a rmt was already introduced by stephanov in 1996 @xcite to explain the difference between the quenched and unquenched theory . the development of the corresponding theory of op in the complex plane started later in @xcite and @xcite . it is already partly covered in the les houches lecture notes from 2005 @xcite , for a slightly more recent review on rmt of qcd with chemical potential see @xcite . here we will comment on some recent developments relating to products of random matrices , see @xcite for a detailed review ( and the lectures of y. tourigny in this volume ) . the discretisation of derivatives on a space - time lattice is not a unique procedure , which also applies to the dirac operator in eq . ( [ dirac ] ) . we would thus first like to motivate why in this subsection we will study the modification proposed by wilson , the wilson dirac operator @xmath407 which is no longer anti - hermitian . here @xmath19 is the lattice spacing and @xmath408 denotes the laplace operator which is hermitian . of course both operators still have to be discretised . we can not possibly give justice to the vast literature on this subject and will only focus on aspects relevant for the application of rmt . the reason why wilson proposed to add the laplacian is the so - called doubler problem . in the continuum with @xmath409 the relativistic energy - momentum relation for a particle with mass @xmath410 , energy @xmath411 , and 4-momentum @xmath412 reads @xmath413 here we have spelled out the sum explicitly . particles satisfying this relation are called on - shell ( on the energy shell ) . after standard discretisation this relation becomes @xmath414 which in the limit @xmath20 leads back to eq . ( [ emom ] ) . in contrast to the continuum relation , here with any @xmath412 satisfying this equation also @xmath415 , @xmath416 etc . fulfil eq . ( [ emoma ] ) . in total we have @xmath417 possibilities to add @xmath418 to the components of @xmath412 , which all correspond to the same on - shell particle . this is the doubler problem . wilson s modification to add the laplacian to the dirac operator amounts to adding the term @xmath419 to the right hand side of eq . ( [ emoma ] ) . in the limit @xmath20 this makes the extra 15 particles heavy and thus removes the doublers . the modification comes with a prices , as the laplacian explicitly breaks chiral symmetry . we will not discuss other possibilities here , for example in @xcite for a rmt of the so - called staggered dirac operator , but rather stick to wilson s choice . despite the non - hermiticity of @xmath420 eq . ( [ wdirac ] ) it satisfies the so - called @xmath284-hermiticity : @xmath421 here we have simply used that @xmath422 and the anti - commutator from eq . ( [ dchiral ] ) . consequently one can define the following hermitian operator called @xmath423 @xmath424 where traditionally the quark mass @xmath425 is added to the definition . looking at eq . ( [ dchiral ] ) for the dirac operator and eq . ( [ gammas ] ) for @xmath284 we can immediately give the chiral block structure of @xmath420 and @xmath423 : @xmath426 here @xmath427 and @xmath428 are hermitian operators . note the change in sign in the mass term in the lower right block of @xmath423 . as previously for @xmath409 this immediately leads to a good ansatz to make for a wrmt , as it was made in @xcite . before we turn to analyse this in some detail let us directly go to the epsilon - regime of qcd in the wilson formulation at fixed topology , wechpt . it turns out that at lo 3 extra terms appear in the chiral lagrangian as was discussed in @xcite ( note the different sign convention ) . they contain 3 new low energy constants ( lec ) @xmath429 , @xmath430 and @xmath431 , and all terms are proportionally to @xmath432 : @xmath433}(\{\hat{m}\})&= & \int_{u(n_f ) } du_0\det[u_0]^\nu\exp\big[+\frac12 \sigma v\mbox{tr}({m}(u_0+u_0^\dag ) ) -a^2v w_8\mbox{tr}(u_0 ^ 2+u_0^{\dag\,2})\nonumber\\ & & \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad -a^2 vw_6(\mbox{tr}(u_0+u_0^\dag))^2-a^2 vw_7(\mbox{tr}(u_0-u_0^\dag))^2 \big]\ . \label{zwechpt}\end{aligned}\ ] ] it is clear that this implies the scaling @xmath434 or @xmath435 , in addition to the scaling of the quark masses with the volume in eq . ( [ exponenteps ] ) . let us discuss some simplifications first . without fixing topology in the special case of @xmath436 the term tr@xmath437 vanishes , and the terms proportional to @xmath431 and @xmath429 are equivalent . in general , in the second line of eq . ( [ zwechpt ] ) the 2 squares of the traces can be linearised by so - called hubbart - stratonovich transformations , at the expense of 2 extra gaussian integrals . the linearised terms can be included into a shift of the mass for @xmath429 , and into so - called axial mass terms @xmath438 for @xmath430 , see @xcite for more details . also for the latter term the corresponding group integral generalising eq . ( [ lsint ] ) is known , see @xcite . therefore in the following we will set @xmath439 , having in mind that we need to perform 2 extra integrals for the full partition function ( [ zwechpt ] ) for non - zero values of these lecs . let us define @xmath440 and compute the partition function only including this term . for @xmath248 we obtain @xmath441}(\hat{m})=\int_{-\pi}^\pi\frac{d\theta}{2\pi } e^{i\nu\theta+\hat{m}\cos(\theta)-4\hat{a}_8 ^ 2\cos(\theta)^2 + 2\hat{a}_8 ^ 2 } = e^{2\hat{a}_8 ^ 2}\int_{-\infty}^\infty \frac{dx}{\sqrt{\pi}}e^{-x^2}\left(\frac{\hat{m}-4ix\hat{a}_8}{\hat{m}+4ix\hat{a}_8}\right)^{-\frac{\nu}{2 } } i_\nu\left(\sqrt{\hat{m}^2 + 16x^2\hat{a}_8 ^ 2}\right ) . \label{zwechptnf1}\ ] ] it is no longer elementary , but after linearising the @xmath442 term it can be written as a gaussian integral over the one - flavour partition function @xmath443 in the continuum at shifted mass . for degenerate masses and general @xmath0 we can again diagonalise @xmath363 and use andrief s integral formula , as in the derivation of eq . ( [ zdeg ] ) , leading to : @xmath444}(\hat{m})=\det_{1\leq j , k\leq n_f}\left[e^{2\hat{a}_8 ^ 2 } \int_{-\pi}^\pi\frac{d\theta}{2\pi}\exp\left[i(\nu+k - j)\theta+\hat{m}\cos(\theta)-4\hat{a}_8 ^ 2\cos(\theta)^2\right]\right ] . \label{zwechptnf}\ ] ] once again it is given by the determinant of @xmath248 flavour partition functions . in order to determine the spectral density of @xmath423 or @xmath420 one then has to introduce additional auxiliary quark pairs as described in eq . ( [ qcdresolvent ] ) ( or use replicas ) . prior to the rmt calculation this was done for the spectral density from wechpt in @xcite , for both @xmath423 or @xmath420 . let us now turn to the rmt side as introduced in @xcite . for simplicity we will only consider the hermitian operator @xmath423 , for the discussion of the complex eigenvalue spectrum of @xmath420 we refer to @xcite and to @xcite for its complete solution . actually two slightly different rmt have been proposed , which have turned out to be equivalent on the level of the jpdf . in @xcite the partition function ( [ matrixrep ] ) was generalised using eq . ( [ dwchiral ] ) in an obvious way , by adding two hermitian gaussian matrices @xmath445 and @xmath313 of dimensions @xmath34 and @xmath56 , respectively : @xmath446}\sim\int[dw][da][db]\prod_{f=1}^{n_f}\det \left [ \begin{array}{cc } m_f 1_n+aa&i w\\ -iw^\dag & -m_f 1_{n+\nu}-ab\\ \end{array } \right ] \exp\left[-\frac14\mbox{tr}(a^2+b^2 + 2ww^\dag)\right ] . \label{wimatrixrep}\ ] ] in @xcite the following modification was made : rather than adding a block - diagonal hermitian matrix to @xmath423 as the last term in eq . ( [ dwchiral ] ) , a full hermitian matrix @xmath447 was added to it , @xmath448 , filling the off - diagonal blocks with a rectangular complex matrix @xmath449 ( and trivially changing the sign of @xmath450 ) : @xmath451}&=&\int[dw][dh]\prod_{f=1}^{n_f}\det \left [ \begin{array}{cc } m_f 1_n+a&i w+\omega\\ -iw^\dag + \omega^\dag&-m_f 1_{n+\nu}+b\\ \end{array } \right ] \exp\left[-\frac{1}{2(1-a^2)}\mbox{tr}ww^\dag-\frac{1}{4a^2}\mbox{tr}h^2\right],\nonumber\\ & & h=\left(\begin{array}{cc } a&\omega\\ \omega^\dag&b\\ \end{array}\right),\ \ a\in[0,1]\ . \label{wiimatrixrep}\end{aligned}\ ] ] the different rescaling of the variances with @xmath19 was made in @xcite to underline that on the level of rmt this corresponds to a two - matrix model that interpolates between the chgue and the gue , in the limits @xmath20 and @xmath452 , respectively . it was shown in @xcite and @xcite that in the limit @xmath32 , identifying @xmath453 in addition to the rescaled mass terms the two rmt partition functions and eq . ( [ zwechpt ] ) with @xmath454 agree : @xmath444}=\lim_{\stackrel{\mbox{\small $ n\to\infty$}}{\mbox{\small $ a , m_f\to0 $ } } } { \cal z}_{wi,\,n}^{[n_f]}= \lim_{\stackrel{\mbox{\small $ n\to\infty$}}{\mbox{\small $ a , m_f\to0 $ } } } { \cal z}_{wi\!i,\,n}^{[n_f]}\ . \label{wequiv}\ ] ] as we have said before the presence of the @xmath429- and @xmath430-terms can be achieved by 2 extra gaussian integrals , see @xcite . it is quite remarkable that rmt allows to include such a detailed structure of wchpt . the fact that we have `` only '' considered rmts for @xmath423 does not affect the equivalence argument of the partition functions . because of eq . ( [ d5 ] ) and @xmath455=1 $ ] the partition functions of @xmath420 and @xmath423 agree . let us just state the jpdf of the eigenvalues @xmath456 of @xmath423 for degenerate masses @xmath425 without derivation , see @xcite and @xcite for details , @xmath457}(m ) & \sim & \exp\big[\frac{nm^{2}}{2a^2(1-a^2)}\big ] \int_{-\infty}^\infty\prod_{j=1}^{2n+\nu } dd_jd_j^{n_f } \exp\big[{-\frac{d_j^2}{4a^2}}\big ] \delta_{2n+\nu}(\{d\})\ \ \ \ \nonumber\\ & & \times\mbox{pf}_{1\leq i , j\leq 2n+\nu;\,1\leq q\leq \nu } \left [ \begin{array}{cc } f(d_j - d_i ) & \!d_i^{\ , q-1}\exp\left[-\frac{d_im}{2a^2}\right]\\ -d_j^{\ , q-1}\exp\left[-\frac{d_jm}{2a^2}\right ] & { 0}\\ \end{array } \right]\ , \label{zdevpf}\end{aligned}\ ] ] where we have defined the antisymmetric weight @xmath458 \sinh\big[\frac{xu}{2a^2}\big].\ ] ] this jpdf has two new features . first , there appears a vandermonde times a pfaffian ( pf ) defined as @xmath459=\sqrt{\det[a]}$ ] for antisymmetric matrices @xmath445 of even dimension , see e.g. @xcite for a definition in terms of permutations . this is quite common for such interpolating two - matrix models , see @xcite for classical works regarding the gue - goe and the gue - gse transitions . second , and most importantly the solution for the correlation functions involves both skew op and op at the same time , where the op are hermite polynomials of a gue of size @xmath460 . we refer to @xcite for details , in particular that this sub - structure of a finite size gue remains valid in the large-@xmath34 limit . in @xcite where only the cases @xmath227 and 1 were solved this structure was not observed , because only hermite polynomials of degree 0 and 1 occur . instead of giving any further details of the solution of wrmt we show plots in figure [ fig : wbessel ] below for the limiting quenched microscopic density @xmath461 of @xmath423 for @xmath212 , illustrating the effect of the rescaled finite lattice spacing @xmath462 . in order to compare to the results from the chgue in figure [ fig : besseldensity ] let us first discuss the difference between @xmath423 and @xmath22 at @xmath409 , where @xmath463 , see eqs . ( [ wdirac ] ) and ( [ d5 ] ) . the rotation with @xmath284 merely makes the dirac operator hermitian rather than anti - hermitian , which corresponds to rotating the eigenvalues from the imaginary to the real axis ( as we had already done in writing the jpdf in eq . ( [ partitionfunct ] ) of real eigenvalues ) . the shift by @xmath464 however introduces a gap such that there are no eigenvalues of @xmath423 in @xmath465 $ ] . furthermore , the @xmath24 exact zero eigenvalues which are not shown in figure [ fig : besseldensity ] get moved away from the origin to one of the edges of this gap , which in our convention is at @xmath466 . therefore we now have to give the density on @xmath111 instead of @xmath112 , and it holds for @xmath25 : @xmath467 the effect of @xmath468 is both to broaden the @xmath24 delta - functions and to eventually fill the gap @xmath469 $ ] . because eigenvalues repel each other the @xmath24 zero - modes will now spread , and form approximately a gue of finite size @xmath24 , see @xcite for a further discussion . in order to make the comparison more transparent let us also choose @xmath470 , so that we can plot the density on @xmath112 only , as is shown in figure [ fig : wbessel ] . the effect of @xmath468 is particularly striking on the zero - modes . ( 12.2,5.5 ) for @xmath227 ( left ) and @xmath264 ( right ) as a function of @xmath471 , taken from @xcite . because of choosing @xmath470 it is still symmetric around the origin . for @xmath472 we are back to the microscopic density of bessel functions as in figure [ fig : besseldensity ] , whereas for @xmath473 ( blue ) and @xmath474 ( red ) the density starts to spread into the region around the origin . for @xmath264 the delta - function of the one zero - eigenvalue is broadened by increasing @xmath471 - which is why this is plotted on a different scale compared to @xmath227 . at these small values of @xmath471 the effect on the density is remarkably localised to the origin . only in the limit @xmath475 we will arrive at the microscopic density of the gue @xcite which is completely flat ( and normalised to @xmath476 here ) . [ fig : wbessel],title="fig:",width=264 ] for @xmath227 ( left ) and @xmath264 ( right ) as a function of @xmath471 , taken from @xcite . because of choosing @xmath470 it is still symmetric around the origin . for @xmath472 we are back to the microscopic density of bessel functions as in figure [ fig : besseldensity ] , whereas for @xmath473 ( blue ) and @xmath474 ( red ) the density starts to spread into the region around the origin . for @xmath264 the delta - function of the one zero - eigenvalue is broadened by increasing @xmath471 - which is why this is plotted on a different scale compared to @xmath227 . at these small values of @xmath471 the effect on the density is remarkably localised to the origin . only in the limit @xmath475 we will arrive at the microscopic density of the gue @xcite which is completely flat ( and normalised to @xmath476 here ) . [ fig : wbessel],title="fig:",width=264 ] ( 0.1,-0.2)@xmath265 ( -7.1,-0.2)@xmath265 ( -10.8,3.5)@xmath461 ( -3.5,1.4)@xmath461 the density of the hermitian wilson dirac operator @xmath423 has been compared to quenched lattice simulations in @xcite . because this involves multi - parameter fits several quantities have been proposed that are simple to measure @xcite . a comparison to unquenched data remains a difficult question , see however @xcite . the rmt approach presented here has also lead to insights about the sign and constraints amongst the lecs @xmath477 that were previously controversial @xcite . when the lattice artefacts become very strong additional unphysical phases appear , and the possibility to access these as a function of the lec and quark content @xmath0 was clarified in @xcite . the derivation of all @xmath62-point correlation functions including @xmath478 based on wechpt has been pushed forward in @xcite , leading to alternative representations compared to the op approach @xcite . for the density of @xmath423 in the @xmath373-regime see @xcite . the computation of individual eigenvalue distribution is not known beyond the expansion of the fredholm expansion proposed in @xcite . in this subsection we will study the influence of a chemical potential @xmath10 on the dirac operator spectrum . we will first state its global symmetry in qcd , then construct the rmt and at the end link to the epsilon - regime of the corresponding chiral lagrangian . the addition of a chemical potential @xmath479 for each quark flavour to the qcd action in eq . ( [ zqcd ] ) amounts to add the terms @xmath480 . in classical statistical mechanics the addition of a chemical potential allows the number of the corresponding particles to fluctuate . here we will add the same chemical potential @xmath481 to all flavours ( which makes it the baryon chemical potential ) . in analogy to eq . ( [ dchiral ] ) the global symmetry of the dirac operator plus @xmath482 changes to @xmath483 and thus becomes complex non - hermitian . it holds that @xmath484 , and thus for purely imaginary chemical potential @xmath485 it is still anti - hermitian . because the operator remains chiral @xmath486 the non - zero complex eigenvalues continue to come in pairs @xmath487 , using the same argument as after eq . ( [ eigenfunct ] ) . the fact that the spectrum of the dirac operator with chemical potential is complex ( and thus the qcd action too ) has severe repercussions for lattice qcd . because the action can no longer be interpreted as a probability weight , the standard method on the lattice , weighting a given configuration by its action breaks down . several methods have been invented to circumvent this so - called sign problem , and we refer to @xcite for a review . in contrast the rmt for qcd with chemical potential remains analytically solvable with op in the complex plane , as we will see . in @xcite a rmt was constructed by stephanov where @xmath298 was replaced by a complex gaussian random matrix @xmath488 , cf . ( [ matrixrep ] ) for @xmath47 compared to ( [ dchiral ] ) . while this rmt was very useful to illuminate the difference between @xmath25 and @xmath88 , which is more dramatic than for @xmath489 , this model was not analytically tractable with op at finite-@xmath34 . therefore osborn @xcite introduced a two - matrix model where @xmath10 is multiplied by a second random matrix @xmath490 , see eq . ( [ 2matrixrep ] ) below , implying that the chemical potential term is not diagonal in that basis . while at first sight more complicated this model turns out to be exactly solvable using op in the complex plane , as we will demonstrate . a further equivalent representation was chosen in @xcite , multiplying @xmath491 by @xmath492 $ ] instead . this parametrisation allowed to understand , why the sign problem in the numerical solution of this rmt for qcd with @xmath493 can be circumvented @xcite . prior to introducing an rmt with @xmath10 as in eq . ( [ dchiralmu ] ) a similar model was introduced @xcite describing the effect of temperature @xmath13 . such a model can only be solved using the supersymmetric method , see @xcite , with the effect of renormalising @xmath494 to be temperature dependent , keeping the correlation functions otherwise unchanged . a further modification of stephanov s model by including both parameters @xmath495 in eq . ( [ dchiralmu ] ) was very successful as a schematic model for the phase diagram of qcd @xcite , predicting the qualitative features depicted in figure [ fig : phasediag ] . let us turn to the rmt @xcite defined as @xmath496[dw_2]\prod_{f=1}^{n_f}\det \left [ \begin{array}{cc } m_f 1_n&i w_1+\mu w_2\\ iw_1^\dag + \mu w_2^\dag & m_f 1_{n+\nu}\\ \end{array } \right ] \exp[-\mbox{tr}(w_1w_1^\dag + w_2w_2^\dag)]\nonumber\\ & \sim & \int[dx_1][dx_2]\prod_{f=1}^{n_f}\det \left [ \begin{array}{cc } m_f 1_n&x_1\\ x_2&m_f 1_{n+\nu}\\ \end{array } \right ] \exp\left[-a(\mu)\mbox{tr}(x_1x_1^\dag + x_2x_2^\dag)\right]\nonumber\\ & & \quad\quad\quad \times\exp\left[b(\mu)\mbox{tr}(x_1x_2+x_2^\dag x_1^\dag)\right]\ , \ \ a(\mu)=\frac{1+\mu^2}{4\mu^2}\ , \ \ b(\mu)=\frac{1-\mu^2}{4\mu^2}\ . \label{2matrixrep}\end{aligned}\ ] ] in the second step we have simply changed variables to the matrices @xmath497 and @xmath498 which are now coupled . while at @xmath499 the matrices @xmath500 and @xmath501 are again independent gaussian random matrices - this is called maximal non - hermiticity - at @xmath489 we have @xmath502 with total correlation ( and we are back to the chgue ) . a look at the characteristic equation for the massless rmt dirac operator , @xmath503=\det[z^2-x_1x_2]\ , \label{muchar}\ ] ] reveals that we are after the complex eigenvalues in this setting see @xcite . ] of the product of the two correlated matrices @xmath504 . as in eq . ( [ matrixrep ] ) we could have allowed for real or quaternion valued matrix elements with @xmath67 . however , the joint density of complex eigenvalues does not enjoy a closed form for all 3 @xmath63 s as in eq . ( [ partitionfunct ] ) . in particular real matrices are special as they can have real eigenvalues and complex conjugated eigenvalues pairs , as the characteristic equation ( [ muchar ] ) is real . for the solution of the respective rmts see @xcite and @xcite , respectively . the parametrisation of a generic complex non - hermitian matrix @xmath505 in terms of its complex eigenvalues is usually done using the schur decomposition @xmath506 , where @xmath507 is unitary , @xmath508 contains the complex eigenvalues , and @xmath13 is a strictly upper triangular complex matrix ( it can be chosen to be strictly lower triangular instead ) . for two matrices however , in order to compute the jacobian we have to use the following generalised schur decomposition @xcite @xmath509 with @xmath510 and @xmath511 . here are unitary , @xmath513 are diagonal matrices with complex entries ( which are not the complex eigenvalues of @xmath500 and @xmath501 ) , and @xmath514 are strictly upper triangular . the complex eigenvalues of the product matrix @xmath505 are thus given by the elements of the diagonal matrix @xmath515 . let us directly give the result for the joint density from @xcite ( see @xcite for an alternative derivation ) : @xmath516}&= & \left(\prod_{j=1}^n\int d^2z_j|z_j|^{\nu}k_\nu(a(\mu)|z_j|)\exp\left[-\frac{b(\mu)}{2}(z_j+z_j^*)\right]\prod_{f=1}^{n_f}(z_j+m_f^2)\right ) |\delta_n(\{z\})|^2 \label{zrmtmu } \\ & = & \left(\prod_{j=1}^n\int d^2z_j\right ) { \cal p}_{jpdf}(z_1,\ldots , z_n)\ . \label{jpdfmu}\end{aligned}\ ] ] note that in view of eq . ( [ muchar ] ) we should later take the square root of the eigenvalues @xmath517 of @xmath504 to change to dirac operator eigenvalues . in the complex plane this is not unique and without loss of generality one can then restrict to the half - plane , also in view of the @xmath518 pairs of dirac eigenvalues . before we discuss the properties of eq . ( [ zrmtmu ] ) and its correlation functions let us try to understand why the jacobian of the transformation ( [ schur ] ) does not only include the modulus squared vandermonde determinant ( as in the complex ginibre ensemble , see @xcite ) and the zero - modes @xmath519 , but also a modified @xmath520-bessel function , even though we started from ( coupled ) gaussian matrices . the same phenomenon happens when looking at products of complex gaussian random variables @xmath521 , and asking for the distribution @xmath522 of the product random variable @xmath523 : @xmath524\delta^{(2)}(z - z_1z_2)=(2\pi)^2 \int_0^\infty \frac{drr}{r^2}\exp[-r^2-|z|^2/r^2]= 2\pi k_0(2|z|)\ . \label{kweight}\ ] ] obviously we are dealing with the analogue of quadratic matrices here . multiplying @xmath525 such random variables we obtain an @xmath526-fold integral representation for the distribution of @xmath527 now given by a meijer g - function : @xmath528 see @xcite for its standard definition . these weights appear when studying the complex eigenvalues of the product of @xmath136 independent gaussian matrices @xmath529 and we refer to @xcite for the solution of this model . the complex eigenvalue correlation functions of the model ( [ zrmtmu ] ) are defined as in eq . ( [ rk ] ) and it holds @xmath530\ . \label{rkc}\ ] ] they are again given in terms of the kernel @xmath531 of polynomials @xmath532 and @xmath533 which are bi - orthogonal in the complex plane with respect to the weight @xmath534\prod_{f=1}^{n_f}(z_j+m_f^2)\ , \label{muweight}\\ & & \ \ \int d^2z\ w(z ) p_k(z)q_l(z)=h_k\delta_{kl}\ . \label{opc}\end{aligned}\ ] ] we will not discuss gap probabilities or individual eigenvalue distributions here and refer to @xcite . unlike in the real eigenvalue case they will now depend on the choice of the geometry of the region containing @xmath535 eigenvalues , and only in the rotationally invariant case @xmath499 a radial ordering of eigenvalues according to their modulus provides a natural choice . before we solve the rmt eq . ( [ zrmtmu ] ) a word of caution is in order . in standard op theory in the complex plane only real positive weights are considered , which is only true for @xmath25 in eq . ( [ muweight ] ) . in this case @xmath536 and standard techniques as gram - schmidt apply to construct these polynomials , see e.g. @xcite . when @xmath88 in our case the weight is complex , and thus a priori the partition function is as well as the @xmath537 are . therefore the density no longer has a probabilistic interpretation , it has a real and imaginary part , and we refer to @xcite for a more detailed discussion . we will first provide the solution of the simplest quenched case with @xmath25 . then the partition function and general eigenvalue correlation functions including @xmath0 mass terms can be obtained along the same lines as in section [ opdirac ] , expressing them in terms of the op at @xmath25 . surprisingly , despite the complicated non - gaussian weight function in the complex plane the op are again given by laguerre polynomials of complex arguments @xcite . ) was made in @xcite . ] the following holds , where we drop the argument @xmath10 of the parameters @xmath538 in eq . ( [ 2matrixrep ] ) : @xmath539 l_j^\nu(cz)l_k^\nu(cz)=\delta_{jk}h_k\ , \ \mbox{with } \ \ c=\frac{a^2-b^2}{2b}\ , \ \ \label{opmu}\ ] ] here @xmath540 is an integer , and the laguerre polynomials can be made monic as in eq . ( [ laguerre ] ) . the proof we sketch follows @xcite and uses induction of depth 2 , for an earlier longer proof see @xcite . insert the following integral representations into the orthogonality relation ( [ opmu ] ) for @xmath212 , @xmath541\ , \nonumber\\ l_j^{\nu}(z)&=&\oint_\gamma \frac{du}{2\pi } \frac{\exp\left[-\frac{zu}{1-u}\right]}{(1-u)^{\nu+1}u^{j+1}}\ , \label{cintrep}\end{aligned}\ ] ] where @xmath542 is a contour enclosing the origin in positive direction , but not the point @xmath543 . then the gaussian integrals over the real and imaginary parts of @xmath544 can be solved , and subsequently the integral over @xmath545 from the @xmath520-bessel function . the orthogonality then easily follows applying the residue theorem twice . for the induction we use the 3-step recurrence relation for the laguerre polynomials and @xmath520-bessel function @xcite , see @xcite for details . we thus have solved the quenched case as we know all correlation functions from the kernel @xmath546 in terms of the quantities in eq . ( [ opmu ] ) , by using eq . ( [ rkc ] ) . a similar solution exists @xcite for the so - called elliptic complex ginibre ensemble of a single complex matrix @xmath547 , where the hermitian part @xmath548 and anti - hermitian part @xmath549 of the matrix have different variances that depend on a parameter . the solution is given in term of hermite polynomials in the complex plane and allows to interpolate between the gue and the complex ginibre ensemble at maximal non - hermiticity , where @xmath447 and @xmath445 have the same variance . we refer to @xcite for a detailed discussion . let us now turn to the unquenched solution with @xmath88 described in great detail in @xcite , expressing everything in terms of the laguerre polynomials of the quenched solution from eq . ( [ opmu ] ) . because we follow the same strategy as in subsections [ genweight ] and [ alllaguerre ] using expectation values of characteristic polynomials we can be very brief . the corresponding proofs for manipulating expectation values of characteristic polynomials and their complex conjugates defined as in eq . ( [ ave ] ) can be found in @xcite . in particular for the monic op in the complex plane satisfying eq . ( [ opc ] ) for a weight @xmath522 that can be complex we have @xcite @xmath550 the proof goes exactly the same way as in eq . ( [ proofpn ] ) , where for @xmath551 we multiply the product into @xmath552 , increasing it to @xmath553 . for @xmath554 we simply choose @xmath555 , increasing it to @xmath556 . likewise for the kernel we obtain @xmath557 modifying the proof in eq . ( [ proofkernel ] ) accordingly by multiplying the first product into @xmath558 , and the second into @xmath559 . the only difference here is that in general no christoffel - darboux formula is available . this is despite the fact that the laguerre polynomials in our example eq . ( [ opmu ] ) still satisfy the standard 3-step recurrence ( which is not true for op in the complex plane in general ) . the step in the proof of eq . ( [ cd ] ) that breaks down is the cancellation of the summands after multiplying eq . ( [ clagkernel ] ) by @xmath560 . the most general result corresponding to eq . ( [ prodchar ] ) now contains both products of @xmath520 characteristic polynomials and of @xmath126 products of complex conjugated ones . without loss of generality let us choose @xmath561 ( @xmath562 can be obtained by complex conjugation ) , to state the result from @xcite : @xmath563 . \label{cprodchar}\ ] ] here the left block of kernels inside the determinant is of size @xmath564 . the proof is straightforward and uses induction . using this result we are now ready to compute the partition function eq . ( [ zrmtmu ] ) for general @xmath565 . in the quenched case it is given again by @xmath185 times the products of the norms from eq . ( [ opmu ] ) , as derived in eq . ( [ znorm ] ) . for @xmath88 we can express it in terms of the laguerre polynomials from eq . ( [ opmu ] ) , and obtain an almost identical answer to eq . ( [ zmassive ] ) : @xmath173}(\{m\})=%z_n^{[n_f=0]}\left\langle \prod_{l=1}^n\prod_{f=1}^{n_f}(z_l+m_f^2)\right\rangle_n&= & \frac{(-1)^{nn_f}n!\prod_{j=0}^{n-1}h_j}{\delta_{n_f}(\{m\ } ) } \det_{1\leq j , g\leq n_f}\left[(-1/c)^{n+g-1}(n+g-1)!l^\nu_{n+g-1}(-cm_f^2)\right ] . \label{zmassivec}\ ] ] as in eq . ( [ zmassive ] ) it is given by a determinant of laguerre polynomials of real arguments which are real . the only difference to @xmath489 is the scale factor @xmath566 and the @xmath10-dependence inside the norms @xmath567 from eq . ( [ opmu ] ) . we will not give further details regarding the kernel @xmath568}(z , u^*)$ ] constituting the @xmath62-point correlation functions in eq . ( [ rkc ] ) . following the first line of eq . ( [ kernelmassive ] ) it can be expressed through the quenched average of @xmath569 characteristic polynomials times one conjugated characteristic polynomial , divided by the partition function from eq . ( [ zmassivec ] ) . the spectral density with @xmath0 masses then follows as in eq . ( [ r1massive ] ) , by multiplying @xmath568}(z , z^*)$ ] with the weight eq . ( [ muweight ] ) , see @xcite for more details . clearly it is not real in general . we will now very briefly discuss the large-@xmath34 limit and matching with the corresponding echpt . it turns out that for non - hermitian matrices that are coupled as in eq . ( [ 2matrixrep ] ) the possibility of taking such limits is much richer than in the hermitian ( or maximally non - hermitian ) case . in addition to the standard macroscopic and microscopic statistics a further class of limits exist , where the matrix @xmath504 becomes almost hermitian , taking the parameter @xmath570 at such a rate that that the following products @xmath571 are constant . here the scaling of the complex eigenvalues ( and masses ) is unchanged compared to eq . ( [ rmtmass ] ) . this limit was first observed in the elliptic complex ginibre ensemble already mentioned @xcite and called weakly non - hermitian @xcite . this is in contrast to the limit at strong non - hermiticity where @xmath572 $ ] is kept fixed and a different scaling of @xmath527 is chosen . here the statistics of the product of two independent complex matrices with @xmath499 is found , see @xcite and @xcite for details . looking at our result eq . ( [ zmassivec ] ) it is clear that the limiting partition function is the same as for @xmath489 in eq . ( [ zrmtnf ] ) and is thus independent of the rescaled @xmath573 . in contrast already the limiting quenched density does depend on @xmath573 . it can be easily obtained from eq . ( [ clagkernel ] ) , replacing the sum by an integral and using eq . ( [ laguerrelim ] ) for the laguerre polynomials , @xmath574\int_0 ^ 1 dtt \exp[-2 t^2\hat{\mu}^2]|j_\nu(\tilde{z}t)|^2\ . \label{cdensity}\ ] ] it is shown in figure [ fig : besselc ] . here we have again changed to dirac eigenvalues following @xcite . this result was obtained earlier using replicas and the toda equation in @xcite . for @xmath575 the integral becomes elementary and matches eq . ( [ besseldensity ] ) , while the prefactors turn into a delta - function . ( 12.2,5.5 ) ) as a function of @xmath576 for @xmath227 ( left ) and @xmath264 ( right ) , at @xmath577 . the density of real eigenvalues from figure [ fig : besseldensity ] can still be recognised , it now spreads into the complex plane . because in the limit @xmath578 it will become proportionally to a delta - function @xmath579 times the bessel density the vertical scale is different . [ fig : besselc],title="fig:",width=264 ] ) as a function of @xmath576 for @xmath227 ( left ) and @xmath264 ( right ) , at @xmath577 . the density of real eigenvalues from figure [ fig : besseldensity ] can still be recognised , it now spreads into the complex plane . because in the limit @xmath578 it will become proportionally to a delta - function @xmath579 times the bessel density the vertical scale is different . [ fig : besselc],title="fig:",width=264 ] ( -5.1,0.8)@xmath265 ( -1.0,1.3)@xmath580 ( -12.1,0.8)@xmath265 ( -8.0,1.3)@xmath580 ( -7.0,4.4)@xmath581 ( -14.0,4.4)@xmath581 the fact that the partition function is @xmath573-independent and the density is not has been called the `` silver blaze problem '' @xcite , alluding to a novel by a.c . doyle . how can we resolve this puzzle ? let us look again at chpt , the theory of goldstone bosons . because these are mesons they do nt carry baryon charge and thus do not feel the effect of adding @xmath10 to the action ( mesons consist of a quark and anti - quark whereas baryons , as the proton , of 3 quarks ) . we can still make the chpt partition function @xmath10-dependent by adding an integer number @xmath582 of quarks with the opposite chemical potential @xmath583 . as discussed after eq . ( [ dchiralmu ] ) this amounts to add complex conjugated determinants of the dirac operator to the action , leading to @xmath584}(\{m\},\{n\})= z_n \left\langle \prod_{l=1}^n\left(\prod_{i=1}^{n_f}(z_l+m_f)\right)\left(\prod_{j=1}^{n_f^*}(z_l^*+n_f^2)\right)\right\rangle_n\ , \label{zrmtmix}\ ] ] which we can also compute using eq . ( [ cprodchar ] ) . the corresponding limiting partion function of echpt now does depend on @xmath10 as follows : @xmath585}(\{m\},\{n\})=\int_{u(n_f+n_f^ * ) } du_0\det[u_0]^\nu\exp\left[\frac12 \sigma v\mbox{tr}({m}(u_0+u_0^\dag))-\frac14\mu^2 vf_\pi^2[u_0,b][u_0,b]\right ] . \label{zechptmu}\ ] ] here @xmath586 contains both kinds of masses and @xmath587 is the metric with the signature of the chemical potentials . the commutator comes from the fact that in the dirac operator @xmath10 acts as an additional vector potential . it obviously vanishes at @xmath588 . once again in the limit the partition functions from rmt and echpt agree , when identifying the rescaled masses and chemical potentials as @xmath589 : @xmath590}(\{m\},\{n\})={\cal z}_{\rm echpt}^{[n_f+n_f^*]}(\{\hat{m}\},\{\hat{n}\}).\ ] ] they are given by the determinant from eq . ( [ cprodchar ] ) in terms of the limiting kernel and the limiting polynomials which are again @xmath443 . the corresponding group integral eq . ( [ zechptmu ] ) was computed independently in @xcite using an explicit parametrisation of @xmath591 , see @xcite for the general @xmath0 case . what about the equivalence of rmt and echpt for the density , in other words how can we generate the spectral density eq . ( [ cdensity ] ) from a chiral lagrangian ? although one may still define and generate the resolvent as in eq . ( [ resolvent ] ) , the relation to the spectral density changes compared to eq . ( [ inversion ] ) : @xmath592 due to @xmath593 on @xmath594 . while it is subtle to extract the non - holomorphic part from the expectation value of the ratio of characteristic polynomials two alternative ways exist to formulate the generation of a two - dimensional delta - function . one way is to use so - called replicas which amounts to introduce @xmath62 extra pairs of complex conjugated quarks in the chiral lagrangian . using the relation to the toda lattice equation , the replica limit can be made well defined as was developed in @xcite . the generating function for the density with @xmath595 , @xmath596}$ ] on the echpt side thus depends on @xmath573 , which is why the resulting density also depends on it . the second possibility is to use supersymmetry , which will also involve pairs of complex conjugated bosons @xcite , @xmath597}{\det[(u - d\!\!\!\!/(\mu))(u - d\!\!\!\!/(\mu))^\dag+\kappa^2]}\right\rangle_\nu^{\rm qcd}\right|_{u = z}\right)\ , \label{susyresmu}\ ] ] and both methods have been shown to be equivalent @xcite . in either case the silver blaze puzzle is resolved . we would also like to mention the papers @xcite where averages of ratios of characteristic polynomials and their complex conjugates were evaluated in a purely complex op framework . the most prominent example for such an average is the distribution of the phase of @xmath598 @xmath599}\ , \label{phase}\ ] ] and we refer to @xcite and @xcite for a detailed discussion including its physical interpretation . when trying to compare the rmt predictions to qcd lattice simulations the presence of the chemical potential also has a virtue , despite the sign problem : it couples to the second lec in the chiral lagrangian , @xmath337 , in eq . ( [ zechptmu ] ) . just as for @xmath26 a fit of the density to lattice data helps to determine @xmath8 , a two - parameter fit at @xmath600 allows to measure @xmath337 in this way . because for imaginary chemical potential @xmath601 the dirac operator spectrum remains real as discussed after eq . ( [ dchiralmu ] ) , this was proposed to determine @xmath337 using standard lattice simulations @xcite . there is also a corresponding two - matrix model @xcite which in the large-@xmath34 with rescaled @xmath601 as in eq . ( [ weaklim ] ) becomes equal to eq . ( [ zechptmu ] ) , with @xmath602 ( which is of course still convergent ) . turning back to real chemical potential most comparisons were done within the quenched approximation , starting with @xcite , cf . @xcite for the influence of topology . despite the difficulty of defining individual eigenvalue distributions these were compared with in @xcite . there are many issues that we have not discussed about the application to qcd with chemical potential , apart from methods different from the op approach mentioned above . this includes the different origin of chiral symmetry breaking at @xmath603 that is strongly related to the oscillatory behaviour of the unquenched density and we refer to @xcite and to @xcite for a review . thanks to its analytical solution the rmt eq . ( [ zrmtmu ] ) has been used as a testing ground for numerical ways to solve the sign problem . this has been successful for the rmt itself and is reported in @xcite . furthermore , the study of the singular values of the qcd dirac operator with chemical potential has been proposed in @xcite , that could also be used to analyse different phases ( not depicted in figure [ fig : phasediag ] ) for larger values of @xmath10 . here both the lattice realisation and the rmt have many open questions , and a first step in rmt has been taken in @xcite . here a different kernel is found at the origin , the so - called meijer g - kernel @xcite that generalises the bessel kernel . the mathematical question of universality in the limit of weak non - hermiticity for non - gaussian rmts is still open at the moment , despite first heuristic attempts @xcite . for a further discussion including all known limiting kernels in this weak non - hermiticity limit resulting from gaussian models we refer to @xcite and references therein . let us briefly summarise the lessons we have learned in these lectures . the theory of orthogonal polynomials on the real line and in the complex plane and their properties has been extensively discussed . a particular focus was put on weights including characteristic polynomials , as this allowed us to determine the dirac operator spectrum of qcd including the contribution of @xmath0 quarks flavours in a random matrix approach . based on the global symmetries of qcd we have taken a limit to the low energy regime that eventually leads to a random matrix description in a controlled approximation , to which corrections can be computed . even though this is an unphysical limit it can be compared to numerical solutions of qcd on the lattice . this leads to a determination of the low energy constants of the theory and tests of lattice algorithms to be used then in the physical regime . we have shown how the introduction of more details in the theory of qcd such as the effect of finite lattice spacing or of chemical potential , that make the eigenvalues complex , modify the spectrum . and we have mentioned several open problems relevant for rmt and for its application like the computation of new correlation functions or the proof of universality for existing results . they hopefully contribute to keep this an interesting and lively area of research for both mathematicians and theoretical physicists . + acknowledgements : it has been a great pleasure to be back in les houches , now as a lecturer , and i would like to express my sincere gratitude to the organisers for inviting me . my understanding of this subject would not have been possible without the many collaborations on this topic , and i would like to thank all my coworkers for sharing their insights and compassion . i am particularly thankful to jac verbaarschot for many valuable comments on a first draft of this manuscript . furthermore i would like to thank the simons center for geometry and physics at stony brook university ( programme on statistical mechanics and combinatorics 2016 ) for the kind hospitality where part of these lecture notes were written up . this work was partly supported by dfg grant ak 35/2 - 1 . f. basile and g. akemann , jhep * 0712 * ( 2007 ) 043 [ arxiv:0710.0376 [ hep - th ] ] . c. lehner and t. wettig , jhep * 0911 * ( 2009 ) 005 [ arxiv:0909.1489 [ hep - lat ] ] . g. akemann , f. basile and l. lellouch , jhep * 0812 * ( 2008 ) 069 [ arxiv:0804.3809 [ hep - lat ] ] .

this article will be about discussing the possibilities to characterize _ randomness _ and _ dynamical complexity _ in quantum mechanics and relating this issue to the questions of non - equilibrium statistical mechanics . we shall try to illustrate , mainly by presenting various numerical examples , a possible ` cyclist approach ' @xcite towards the _ quantum many - body problem _ which is inspired by experiences gained in studies of quantum and classical chaos of one or few particles ( see e.g. @xcite ) . solving the many body problem in quantum mechanics presents a major challenge from its very beginnings . and along this way , many ingenious important analytical and efficient numerical methods of solution have been developed , for example bethe ansatz @xcite , later generalized and interpreted as the quantum inverse scattering problem @xcite , real space renormalization group methods @xcite , quantum monte carlo @xcite , density matrix renormalization group ( dmrg ) @xcite , and more recently various quantum information - theoretic based time - dependent dmrg ( tdmrg ) @xcite . the ultimate aim of any of these methods is to efficiently find analytical or numerical approximation to the solution for some of the physical observables in the quantum many body problem , however many methods work only under some specific conditions which are not always well understood . for example , bethe ansatz and quantum inverse scattering work only for a small subset of problems which are _ completely integrable _ , and which may often have very non - generic non - equilibrium properties , such as e.g. ballistic transport at high temperatures @xcite . quantum monte carlo techniques represent a very successful set of numerical methods which can yield thermodynamic equilibrium averages for generic ( non - integrable ) systems , however they are practically useless for computing non - equilibrium , or time dependent quantities , such as transport coefficients . traditional dmrg method @xcite is provably very successful for computing accurate ground state expectation values of almost any physical observables of one - dimensional interacting systems . and tdmrg @xcite promised to extend the success of dmrg to time dependent physics . indeed , the first numerical experiments looked very promising , but after a closer look one may realize that they have all been applied to a rather special subset of interacting systems and to a rather special subset of initial states . for generic ( non - integrable ) interacting systems or for sufficiently complex initial states , tdmrg should fail to provide an efficient computation as shall be discussed below . in our view this is to be expected , and represents an intrinsic characteristic of quantum complexity of such system and should correspond to many body extensions of the phenomena of quantum chaos @xcite where spectra and eigenvector coefficients can be described by statistical ensembles of random matrices . in the past two , almost three decades there has been a lot of activity in the so called field of quantum chaos , or quantum chaology @xcite , where people were trying to understand the essential and significant features of quantum systems which behave chaotically in the classical limit . classical chaos can be defined in terms of positive ( algorithmic , kolmogorov - chaitin ) complexity of systems orbits . still , the question whether such definition of complexity can be in an intuitively sensible way translated to quantum mechanics failed to be answered in spite of many efforts . it seems that quantum systems of finite ( one or few body ) chaotic systems are generically more robust against imperfections @xcite than their classical counterparts , which is consistent with a simple illustration based on wave - stability of unitary quantum time evolution @xcite . nevertheless , it has been suggested that exponential sensitivity to initial conditions , the essential characteristics of classically chaotic systems , has many fingerprints in quantum mechanics . for example , in one of the pioneering works on quantum loschmidt echoes ( or fidelity decay ) , jalabert and pastawski @xcite have shown that in the semi - classical regime , decay of system s sensitivity to external perturbation , as defined by fidelity , is exponential with the rate which can be related to classical lyapunov exponents . however , this is only true in sufficiently semi - classical regime , where effective planck constant is smaller than effective strength of perturbation , and where fidelity can be essentially explained classically @xcite . on the other hand , in purely quantum regimes , quantum fidelity decays in completely different manner than the classical fidelity , and quite surprisingly displays slower decay for systems with stronger decay of temporal correlations @xcite . throughout this paper we shall discuss various ergodic properties of a simple toy model of non - integrable interacting quantum many body system , namely kicked ising ( ki ) chain @xcite , and its time - independent version , which ( both ) undergo a transition from integrable to quantum chaotic regimes when the direction of the ( kicking ) magnetic field is changed . by simulating the dynamics of the model in terms of tdmrg we find that it can be performed efficiently only in the integrable regimes . being interested in statistical mechanics of the model we shall describe numerical experiments addressing the questions on the relationship between the onset of quantum chaos in ki model and its quantum mixing property , namely the nature of relaxation to equilibrium , and the properties of non - equilibrium steady state . on one hand we argue that the regime of quantum chaos essentially corresponds to the regime of quantum ergodicity and quantum mixing where diffusive transport laws are valid . on the other hand , we conjecture that the transition to non - ergodic regime may occur before the system parameters reach the integrable point ( even in thermodynamic limit ) , and that non - ergodic to ergodic transition can be characterized with order parameters which change discontinuously at a critical value of system parameter . we argue that the process of relaxation to equilibrium in quantum chaotic ( mixing ) case can be described in terms of quantum analogues of ruelle resonances @xcite . furthermore , we conjecture that the eigenvectors of this process possess a certain scale invariance which can be described by simple power laws . we also discuss a possibility of numerical calculation of quantum dynamical entropies @xcite in a non - trivial setting of ki model , and find , quite remarkably , that positivity of such dynamical entropies _ does not correspond _ to any other measures of quantum chaos , namely quantum dynamical entropies appear to be positive even in the integrable regimes . about two thirds of the material presented in this paper constitutes a review of a selection of recent results related to quantum chaos in many - body systems , with a flavor of quantum information . however , about one third of the material , constituting a major part of section [ sect : statmech ] , is new and original and has not yet been published before . the paper is organized as follows . in short section [ sect : definitions ] we review basic definitions of complete integrability and quantum chaos , in particular in the context of many - body systems . in section [ sect : model ] we introduce ki model which will serve us as a very convenient and efficient toy model to numerically demonstrate all the phenomena discussed in this paper . in section [ sect : losch ] we discuss the robustness of quantum systems to external perturbations and decoherence in open quantum systems , mainly as characterized by quantum loschmidt echoes and entanglement between the system and the environment . in section [ sect : mps ] we discuss the time efficiency of best ( known ) classical simulation , namely using tdmrg , of locally interacting 1d quantum systems and its possible dependence on integrability of the model . in section [ sect : transport ] we relate standard criteria of quantum chaos to normal ( diffusive ) versus anomalous ( ballistic ) transport and discuss a simulation of a quantum heat current in non - equilibrium steady state . in section [ sect : statmech ] we discuss a problem of quantum relaxation to equilibrium , i.e. the quantum mixing property , and quantitative characterization of quantum dynamical complexity . this section contains a large portion of original intriguing numerical results which support few interesting and perhaps surprising conjectures . in section [ sect : conc ] we conclude and discuss some open problems . the central issue of this paper is to verify and demonstrate to what extend the _ complete integrability _ affects non - equilibrium properties of quantum many - body systems and their dynamical complexities , and conversely , whether _ ( strong ) non - integrability _ generically coincides with the established criteria of _ quantum chaos_. let us start by giving some established definitions ( see e.g. @xcite ) of the basic terms needed to understand the issue . a classical hamiltonian many - body system with @xmath0 degrees of freedom is _ completely integrable _ , if there exist @xmath0 functionally independent global smooth phase space functions ( integrals of motion ) which are mutually in involution , i.e. all poisson brackets among them vanish . in such a case global canonical transformation to canonical _ action - angle _ variables can be constructed , and dynamics can be explicitly solved - at least in principle . for locally interacting _ infinite _ systems ( @xmath1 ) analytic methods for an explicit construction of integrals of motion and canonical action - angle variables are known which usually go under a common name of _ inverse scattering method_. explicit solution by a mapping to an iso - spectral sch " rodinger problem in terms of inverse scattering technique is usually understood as a definition of _ complete integrability _ in such context . definition of _ quantum complete integrability _ is less unique . nevertheless , algebraic , non - commutative versions of the inverse scattering technique exist and can be applied to some quantum interacting lattices in one - dimension , generalizing the famous bethe ansatz solution of the heisenberg spin @xmath2 chain , and this is used as the most general known definition of _ quantum complete integrability_. other versions of integrability for finite @xmath0 quantum systems have been proposed but they do not correspond to the integrability of the underlying classical limit , if the latter exists , so these ideas will not be considered in this paper . it has not been generally accepted yet , though demonstrated in many occasions , that completely integrable quantum systems constitute only a small subset of physical models and posses many exceptional ( non - generic ) non - equilibrium dynamical properties , like for example anomalous transport at finite temperatures ( see e.g.@xcite . ) on the other extreme of ergodic hierarchy we have chaotic systems . in classical hamiltonian dynamics of few particles , chaos is best defined in terms of positive algorithmic ( kolmogorov ) complexity of systems trajectories , or equivalently , by exponential sensitivity to initial conditions . however , bounded quantum systems of finite number of particles can not be dynamically complex as their excitation spectrum is discrete , and hence the evolution is necessarily quasi - periodic ( or almost periodic ) . still , quite surprisingly , even for such systems certain dynamical properties are _ random _ and _ universal _ , if the underlying classical limit is sufficiently strongly chaotic @xcite . but genuine dynamical complexity may emerge in thermodynamic limit . however , there is still no completely satisfactory definition of dynamical chaos for infinite quantum systems . the commonly used working definition is the reference with the _ random matrix theory _ @xcite , namely the many - body quantum system is said to be quantum chaotic if its excitation spectrum or some other dynamical properties can be ( on certain energy , or time scale ) well described by ensembles of random hermitian matrices with appropriate symmetry properties . throughout this paper we shall use , either for illustration of theoretical results , or for numerical experimental studies , the following 1d locally interacting quantum lattice system , namely a chain of @xmath0 qubits , or spin @xmath2 particles , coupled with nearest neighbour ising interaction and periodically kicked with spatially homogeneous , but _ arbitrarily oriented _ magnetic field . in a suitable dimensionless units our model can be written in terms of a three - parameter periodic time - dependent hamiltonian @xmath3 @xmath4 is a unit - periodic dirac delta and @xmath5 are the usual pauli spin variables on a finite lattice @xmath6 , satisfying the commutation ( lie ) algebra @xmath7=\sum_\gamma 2\ii\varepsilon_{\alpha \beta\gamma } \delta_{jk } { \sigma^{\gamma}_{j}}$ ] . sometimes it will turn fruitful if we extend the set of pauli matrices by identity matrix and assign them a numerical superscript @xmath8 , namely @xmath9 . the finite chain will often be treated with _ periodic boundary conditions _ , @xmath10 , and sometimes the thermodynamic limit ( tl ) @xmath1 will be considered , in particular when we shall study the statistical mechanics of ( [ eq : hamiltonian ] ) in section [ sect : statmech ] . although kicked hamiltonian one - particle models have been very popular in the field of nonlinear dynamics and quantum chaos for decades , for example the chirikov s kicked rotator model @xcite , the use of kicked systems in quantum many body physics has been so far very limited . if one is not only interested in zero temperature ( ground state ) or low temperature physics , then as we shall try to demonstrate in this review , kicked many - body models like ( [ eq : hamiltonian ] ) provide simpler and clearer phenomenological picture of global dynamics than time - independent models . the main reason is that since energy is not a conserved quantity , the entire hilbert space of many - body configurations is accessible for non - integrable dynamics , and the notions of quantum ergodicity and mixing ( see e.g. ref.@xcite for definitions and further references ) can be defined more clearly than in the time - independent , autonomous setting . perhaps the first kicked interacting quantum lattice has been introduced in ref.@xcite . even if in traditional solid state physics such kicked dynamics would represent un - physically strong excitations , one has to realize that kicked quantum chains could be attractive options as benchmark models of quantum state manipulation and quantum computation in optical lattices . the _ kicked ising _ ( ki ) model ( [ eq : hamiltonian ] ) has been defined for the first time in ref.@xcite , generalizing the integrable ki model with transverse field introduced in ref.@xcite . clearly , as shown there @xcite for the case of _ transverse _ field , @xmath11 , the time - dependent model ( [ eq : hamiltonian ] ) can be considered _ completely integrable _ since it can be solved explicitly , for example by wigner - jordan - bogoliubov transformation to non - interacting spinless fermions , and a large class of its time correlation functions can be calculated explicitly . in addition , an infinite sequence of local conserved quantities ( integrals of motion ) can be constructed in such a case . there is another , trivial _ completely integrable _ regime of ki model , namely the case of _ longitudinal _ field , @xmath12 yet another , more non - trivial _ completely integrable _ regime of ki model is found when the magnitude of dimensionless field is a multiple of @xmath13 , namely @xmath14 , since then the magnetic kick can be considered as a multiple of @xmath15 rotation , and generated by a slightly obscure set of non - interacting spinless fermions . however , in a general case of titled magnetic field when both components @xmath16 and @xmath17 are non - vanishing , and @xmath18 is non - integer , the model is non - integrable , and is conjectured to be not amenable to exact analytical methods . as discussed in the following sections , non - integrable ki model can display a variety of regimes according to the criteria of quantum chaos , quantum ergodicity and quantum mixing . in fact , recently the spectral statistics of ki model in strongly non - integrable regime has been studied in detail and random matrix behaviour has been clearly confirmed at short energy ranges @xcite . due to kicked nature of interaction , the evolution propagator of ki model for one unit period of time ( the so - called floquet operator ) , starting just before the kick , can be constructed explicitly in terms of a time - ordered product @xmath19 the last line suggests a simple efficient quantum protocol to simulate ki model in terms of simple 1-qubit @xmath20 and 2-qubit @xmath21 quantum gates . if we write a compact kicked ising 2-qubit gate @xmath22 and introduce a left - to - right ordered operator product , namely @xmath23 , then we can write ki protocol as a simple string of @xmath24gates @xmath25 the ki model has a better known autonomous limit , namely time - independent ising chain in a tilted magnetic field @xmath26 which is again non - integrable unless the field is transverse @xmath27 , or longitudinal @xmath28 . one of the central questions about the dynamics of complex quantum systems is their robustness against small imperfections in the hamiltonian . while it is clear that due to unitarity the quantum evolution is stable against variation of initial states @xcite , it is not so clear how stable it is against variation of the hamiltonian , either being static or time - dependent - perhaps even noisy . let us write the hamiltonian as @xmath29 , where @xmath30 is the unperturbed hamiltonian , @xmath31 is a hermitian perturbation operator and @xmath32 is a small perturbation parameter . peres @xcite proposed the following measure of stability of quantum evolution : let us start from some fixed initial state @xmath33 , and write the time evolutions of this state generated with the unperturbed and perturbed hamiltonian , as @xmath34 and @xmath35 , respectively . then the stability is characterized by fidelity , _ i.e. _ the squared overlap between these two states @xmath36 there are two alternative interpretations of quantum fidelity : ( i ) one can interpret ( [ eq : fidelity ] ) as a _ quantum loschmidt echo _ , namely the probability that the initial state @xmath33 , which is propagated forward with perturbed evolution @xmath37 , and after that propagated backwards in time with unperturbed evolution @xmath38 , is again measured in the same ( initial ) state . alternatively ( ii ) is just the square modulus of expectation value of the unitary echo operator @xmath39 which is the quantum propagator in the _ interaction picture_. there have been three main theoretical approaches to understanding of fidelity decay in quantum dynamical systems : in a seminal work jalabert and pastawski derived quantum loschmidt echo for systems which posses well defined classical limit . they have shown that under certain conditions , namely that the perturbation strength is large enough - typically larger than appropriately scaled planck constant , and that initial states have certain classical interpretations - like coherent states , position states , etc , the quantum loschmidt echo decays exponentially @xmath40 with the rate @xmath41 which is _ perturbation independent _ and typically equals the lyapunov exponent of the underlying classical dynamics . more recently a completely classical interpretation of this so - called lyapunov decay has been given @xcite in terms of _ the classical loschmidt echo _ and a theory for exponents @xmath41 has been developed in terms of the full spectrum of lyapunov exponents for classical dynamical systems with few @xcite or many @xcite degrees of freedom . in the opposite regime , where the scaled planck constant is bigger than the perturbation strength @xmath32 , one can use time - dependent quantum perturbation theory to second order to derive a simple linear response formula for fidelity decay @xcite @xmath42 in terms of 2-point time correlation function of the perturbation @xmath43 where @xmath44 is the perturbation operator in the interaction picture and @xmath45 is an expectation value in the initial state . from this formula - which can be viewed as a kind of kubo - like linear response theory for dissipation of quantum information , an interesting conclusion can be drawn : fidelity decays faster for systems with slower decay of temporal correlations , or alternatively phrased , quantum system is more robust against external perturbations if it relaxes to equilibrium faster . one can actually go beyond the second order perturbation theory , and sum up the born series for fidelity to all orders in many specific situations @xcite . since we are here mainly interested in qubit ( spin 1/2 ) chains , we shall only review specific results for kicked ising chain @xcite . we shall discuss three different specific values of system parameters , in all three we take @xmath46 : ( a ) integrable regime of transverse field @xmath27 , ( b ) weakly non - integrable regime with @xmath47 , and ( c ) strongly non - integrable regime with @xmath48 . all the results , correlation functions and fidelity , are for random initial states , which can be interpreted as pure states of maximum quantum information , and averaging over random states is equivalent to a tracial state @xmath49 . we consider the evolution operator ( [ eq : floquet ] ) and perturb it by changing the magnetic field such that the perturbation is generated by the transverse component of the magnetization @xmath50 , namely @xmath51^t$ ] . in the integrable case , or in general , in non - ergodic case , where the correlation function @xmath52 ( note that @xmath53 ) approaches a non - vanishing plateau value as @xmath54 , namely time averaged fluctuation of magnetization @xmath55 is non - vanishing @xmath56 , one can sum up born series to all orders , yielding a gaussian decay of fidelity @xmath57 the only assumptions are that @xmath58 is long enough for the time average in the definition of @xmath59 to converge , and short enough that fidelity is still above the finite size plateau value @xmath60 . note that @xmath59 can be considered as an analog of drude weight or charge stiffness in the linear response solid state transport theory ( see e.g. @xcite ) . in figs.([fig : corr],[fig : fid ] ) ( a , b ) one can observe the correlation plateaus of correlation functions for integrable and weakly non - integrable cases and respective good agreement with a gaussian decay of fidelity ( [ eq : gauss ] ) . note that in the integrable case the correlation function and the plateau @xmath59 have been calculated also analytically @xcite . note that for comparing different lattice sizes @xmath0 a size - scaled value of the perturbation strength @xmath61 , with @xmath62 , has been fixed rather than @xmath32 itself . one the other hand , for sufficiently strong integrability breaking , say in case ( c ) of ki model , the correlation function @xmath63 decays to zero in tl , which can be interpreted as quantum mixing behviour . it has been found that quantum mixing behaviour typically corresponds to random matrix ( quasi)energy level statistics , see e.g. ref.@xcite and the subsequent sections of the present paper . we shall call such behaviour the regime of quantum chaos . here again , born series for fidelity can be summed up to all orders , yielding an exponential decay @xcite @xmath64 where @xmath65 is a transport coefficient . the assumptions for validity of ( [ eq : ne ] ) are that @xmath58 is larger than a characteristic mixing time scale on which @xmath66 decays and short enough so that finite size effects in quantum correlation function @xmath66 does not yet affect the transport coefficient @xmath67 . this regime of fidelity decay is sometimes referred to as the fermi golden rule regime . again , as demonstrated in figs . ( [ fig : corr],[fig : fid])c the agreement of numerical data for ki model with the theory is excellent . note that decay time of fidelity scales as @xmath68 for ergodic and mixing dynamics , and as @xmath69 for non - ergodic dynamics , making the latter much more sensitive to small perturbations . complex quantum systems can often be well described by statistical ensembles of hamiltonians @xcite . assuming that @xmath30 and @xmath31 both belong to canonical ensembles of gaussian random matrices , one can evaluate ensemble averages of the fidelity and relate @xmath66 to the spectral form factor of the underlying random matrix ensemble . the perturbative ( linear response ) theory has been developed in ref . @xcite , see also ref . @xcite . furthermore a non - perturbative ( super - symmetric ) averaging has been successfully applied to obtain exact expressions for average fidelity amplitude in the most interesting cases @xcite . rmt theory of fidelity has been successfully applied to chaotic quantum systems and even to several experimental situations @xcite . following more applied philosophy , groups from toulouse and como have performed a series of numerical experiments @xcite analyzing the robustness of several reasonable models of quantum computer hardware under small imperfections , being either due to ( static ) unwanted inter - qubit coupling or due to stochastic ( noisy ) unwanted coupling to the environment , when performing quantum algorithms simulating various toy models of classical and quantum single - particle chaos @xcite , like for example quantum kicked rotator @xcite . the numerical results are in line with a general linear response theory , stating that static perturbations are in general more dangerous than noisy ones . similar thinking as in the previous subsection can be applied to perhaps even more fundamental problem of quantum physics , to the problem of decoherence @xcite . here we shall limit ourselves to an abstract unitary model of decoherence , where we treat a complete unitary evolution over two subsystems , a central system c , and an environment e , and then address a relevant information about the central system ( i.e. the part of the system which is of physical interest ) by partially tracing over the environment . such a discussion can indeed be followed with a close link to the problem of loschmidt echoes , by writing the total hamiltonian @xmath29 as an ideal ( unperturbed ) separable evolution @xmath70 perturbed by a small coupling @xmath31 between the system and the environment . we shall also assume that we start from initial pure state which is a product state @xmath71 we are interested in the properties of a generally mixed state of the central system obtained by partial tracing over the environment @xmath72\ ] ] then , under an ideal evolution , the state of the system remains a product state at all times , so the state of the central system remains pure and there is no decoherence . decoherence is usually characterized in terms of decaying off - diagonal matrix elements of @xmath73 in a suitable pointer state basis , for example in the eigenbasis of the central system hamiltonian @xmath74 . in fact for certain special forms of the perturbation operator @xmath31 , the magnitudes of off - diagonal matrix elements of @xmath73 can be literally written as fidelities , or quantum loschmidt echoes , in the evolution of the environment perturbed by system - environment coupling @xcite . in such framework , the lyapunov regime of fidelity decay corresponds exactly to the lyapunov growth of decoherence discussed by zurek and paz @xcite . however , another interesting indicator of decoherence is the growth of _ bi - partite entanglement _ between the system and the environment . perhaps this notion is even more general since it does not depend on a particular choice of the pointer state basis . since the state of the universe ( central system + environment ) is always pure , the characterization of entanglement is easy , say in terms of von neuman entropy of @xmath73 , @xmath75 , or linear entropy @xmath76 which is a negative logarithm of purity @xmath77 . usually , the quantities @xmath78 and @xmath79 essentially give equivalent results - namely they can both be interpreted as the logarithm of an effective rank of the state @xmath73 , but the linear entropy , or purity , is more amenable to analytical calculations . again considering chaotic models for the central system and the environment , miller and sarkar @xcite ( see also @xcite ) have been able to observe the lyapunov regime of entanglement growth , namely @xmath78 is for sufficiently strong perturbations @xmath32 found to grow linearly @xmath80 with the rate @xmath81 which is perturbation independent and given by ( the sum of positive ) classical lyapunov exponents of the unperturbed , separated ( sub)systems . on the other hand , in the purely quantum regime of small perturbation @xmath32 , again typically smaller than effective planck constant , one can use time - dependent perturbation theory and derive perturbation dependent entanglement entropy @xcite , namely the purity can be explicitly expressed as @xmath82 where @xmath83 is a particular integrated correlation function of the perturbation . in this regime we again find that quantum systems , and quantum environments , which display faster relaxation to equilibrium , are more robust against decoherence due to typical couplings . for studies of bi - partite entanglement , in particular in ki model system , see refs . @xcite . there exists even closer connection to fidelity theory , namely one can prove a general inequality @xcite stating that purity is always bounded by the square of fidelity @xmath84 in other words : @xmath85 gives an upper bound for the growth of the linear entanglement entropy , @xmath86 . in a slightly different context , _ bit - wise _ entanglement between a pair of qubits of a quantum register representing a time dependent quantum state , where the rest of the register is considered as an environment , has been demonstrated to be an indicator of quantum chaos @xcite and even a signature of classical chaos @xcite . in the theory of classical dynamical systems there is a fundamental difference between integrable and chaotic systems as outlined in section [ sect : definitions ] . chaotic systems , having positive algorithmic complexity , unlike the integrable ones , can not be simulated for arbitrary times with a finite amount of information about their initial states . computational complexity of individual chaotic trajectories is linear in time , however , if one wants to describe statistical states ( phase space distributions ) or observables of chaotic classical systems , up to time @xmath58 , exponential amount of computational resources is needed , typically @xmath87 where @xmath81 is the kolmogorov s dynamical entropy related to exponential sensitivity to initial conditions . but on the other hand , how difficult is it to simulate isolated and bounded quantum systems of many interacting particles using classical resources ? in analogy with the classical ( chaotic ) case , we might expect that the best classical simulation of typical quantum systems ( in tl ) is exponentially hard , i.e. the amount of computing resources is expected to grow exponentially with time . even though there is no exponential sensitivity to initial conditions in quantum mechanics , there is a tensor - product structure of the many - body quantum state space which makes its dimension to scale exponentially with the number of particles , as opposed to linear scaling in the classical case , and due to presence of entanglement generic quantum time evolution can not be reduced to ( efficient ) classical computation in terms of non - entangled ( classical like ) states . here we propose the idea to use _ the computation complexity of best possible classical simulation of quantum dynamics , as a measure of quantum algorithmic complexity_. this section essentially reviews the article @xcite . we note that our proposal is different from existing proposals of quantum algorithmic complexity @xcite , namely we consider merely the classical complexity of ( best ) classical simulations of quantum states . in the sense of mora and briegel @xcite , quantum algorithmic complexity _ per unit time _ of initially simple time - evolving quantum states propagated by locally interacting hamiltonian @xmath88 is clearly vanishing , since an approximate quantum circuit which reproduces the state after time @xmath58 is a simple repetition of trotter - suzuki decomposition of @xmath89 . recently , a family of numerical methods for the simulation of interacting many - body systems has been developed @xcite which is usually referred to as time - dependent density - matrix - renormalization group ( tdmrg ) , and which has been shown to provide an efficient classical simulation of some interacting quantum systems . of course , it can not be proven that tdmrg provides the best classical simulation of quantum systems , but it seems that it is by now the best method available . simulations of locally interacting one - dimensional quantum lattices were actually shown rigorously to be efficient in the number @xmath0 of particles @xcite ( i.e. , computation time and memory resources scale as polynomial functions of @xmath0 at fixed @xmath58 , whereas here we are interested in the scaling of computation time and memory with physical time @xmath58 ( in tl , @xmath90 ) , referred to as _ time efficiency_. in this section we address the question of time efficiency implementing tdmrg for a family of ising hamiltonian ( [ eq : hamaut ] ) which undergoes a transition from integrable ( transverse ising ) to non - integrable quantum chaotic regime as the magnetic field is varied . we mainly consider time evolution in operator spaces @xcite , say of density matrices of quantum states , or ( quasi ) local operator algebras . note that time evolution of pure states is often ill defined in tl @xcite . as a quantitative measure of time efficiency we define and compute the minimal dimension @xmath91 of matrix product operator ( mpo ) representation of quantum states / observables which describes time evolution up to time @xmath58 within fidelity @xmath92 . our main question concerns possible scaling of @xmath91 for different types of dynamics , and indeed we shall demonstrate a correspondence between , respectively , quantum chaos or integrability , and exponential or polynomial growth of @xmath91 . the key idea of operator - valued tdmrg @xcite is to represent any operator in a matrix product form , @xmath93 in terms of @xmath94 matrices @xmath95 of fixed dimension @xmath96 . the number of parameters in the mpo representation is @xmath97 and for sufficiently large @xmath96 it can describe any operator . in fact , the minimal @xmath96 required equals to the maximal rank of the reduced super - density - matrix over all bi - partitions of the chain . the advantage of mpo representation lies in the fact that doing an elementary local one or two qubit unitary transformation @xmath98 can be done locally , affecting only a pair of neighboring matrices @xmath95 . we will illustrate evolution of ising chain ( [ eq : hamaut ] ) , with _ open _ boundary conditions , for two different magnetic field values : ( i ) an integrable ( regular ) case @xmath99 with transverse magnetic field , and ( ii ) non - integrable ( quantum chaotic ) case @xmath100 with@xmath101 tilted magnetic field . to confirm that @xmath102 , and @xmath103 , indeed represent generic quantum chaotic , and regular , system , respectively , we calculated level spacing distribution ( lsd ) of their spectra ( shown in fig . [ fig : lsd ] ) . lsd is a standard indicator of quantum chaos @xcite . it displays characteristic level repulsion for strongly non - integrable quantum systems , whereas for integrable systems there is no repulsion due to existence of conservation laws and quantum numbers . evolution by tdmrg proceeds as described in @xcite using an approximate trotter - suziki factorization , for some time step @xmath104 , of the evolution operator @xmath105 in terms of 2-qubit gates . and for each two qubit gates , the matrices @xmath95 can be updates using a singular value decomposition with some truncation error @xmath106 . interestingly , it has been found @xcite that the sum of all truncation errors up to time @xmath58 , denoted by @xmath107 ( provided @xmath104 is small enough so that the trotter - suzuki factorization error is negligible ) is simply proportional to _ fidelity error _ , namely @xmath108 where @xmath109 and numerical constant @xmath110 does not dependent on either @xmath104 , @xmath96 or @xmath0 . the central quantity we are going to study is @xmath111 which is the minimal dimension @xmath96 of matrices @xmath112 in order for the total truncation error @xmath107 to be less than some error tolerance @xmath113 , for evolution to time @xmath58 . in numerical experiments shown we take @xmath114 except for simulating thermal states of quenched hamiltonians where @xmath115 . for a reference we start by investigating the evolution of pure states following the basic tdmrg algorithm @xcite . we studied time efficiency of simulation of _ pure states _ in schr " odinger picture , for which many examples of efficient applications exist , however all for initial states of rather particular structure , typically corresponding to low energy sectors of few quasi - particle excitations or to low dimensional invariant subspaces . treating other , typical states , e.g. eigenstates of unrelated hamiltonians , linear combinations of highly excited states , or states chosen randomly in the many - particle hilbert space , we found that , irrespectively of integrability of dynamics , tdmrg is _ not _ time - efficient , i.e. @xmath91 typically grows exponentially in time even in the integrable case of transverse field ( consistently with a linear growth of entanglement entropy @xcite ) . numerical results are summarized in fig . [ fig : pure ] . we continue by discussing the time efficiency of operator - valued tdmrg method using mpos ( [ eq : mpo ] ) . let us first study the case where the initial operator is a local operator in the center of the lattice @xmath116 . in the integrable case time evolution @xmath117 can be computed exactly in terms of jordan - wigner transformation and toeplitz determinants @xcite , however for initial operators with infinite _ index _ [ sect.2 , 1st of refs.@xcite ] is half the number of fermi operators in jordan - wigner transformation of @xmath118 and is a conserved quantity for @xmath103 . ] like e.g for @xmath119 , @xmath1 , the evolution is rather complex and the effective number of terms ( pauli group elements ) needed to span @xmath117 grows exponentially in @xmath58 . in spite of that , our numerical simulations shown in fig . [ fig : local ] strongly suggest the linear growth @xmath120 for initial operators with infinite index . quite interestingly , for initial operators with _ finite _ index , @xmath121 saturates to a finite value , for example @xmath122 for @xmath123 , or @xmath124 for @xmath125 . in non - integrable cases the rank has been found to grow exponentially , @xmath126 with exponent @xmath127 which does _ not _ depend on @xmath113 , properties of @xmath128 or @xmath0 , for large @xmath0 . for @xmath129 we find @xmath130 . in physics it is often useful to consider extensive observables , for instance translational sums of local operators , e.g. the hamiltonian @xmath88 or the total magnetization @xmath131 . as opposed to local operators , extensive initial operators , interpreted as w - like states in operator space , contain some long - range ` entanglement ' so one may expect that tdmrg should be somewhat less efficient than for local operators . indeed , in the integrable case we find for extensive operators with _ finite _ index that @xmath91 does no longer saturate but now grows linearly , @xmath120 , whereas for extensive operators with _ infinite _ index the growth may be even somewhat faster , most likely quadratic @xmath132 but clearly slower than exponential . in the non - integrable case , we again find exponential growth @xmath133 with the same exponent @xmath134 as for local initial observables . the results are summarized in fig . [ fig : transl ] . in the last set of numerical experiments we consider time efficiency of the evolution of a thermal initial state @xmath135 under a sudden change of the hamiltonian at @xmath136 , namely @xmath137 . again , we treat two situations : in the first case we consider change after which the hamiltonian remains integrable , @xmath138 , while in the other case the change breaks integrability of the hamiltonian , @xmath139 . the initial state is prepared by means of imaginary time tdmrg from initial identity operator using the same mpo rank @xmath96 as it is later used for real time dynamics . we find , consistently with previous results , that at high temperature ( @xmath140 ) the rank @xmath91 grows very slowly , perhaps slower than linear , in the integrable case , and exponentially @xmath133 , in the non - integrable case . interestingly , at lower temperatures we find exponential growth in both cases , even in the integrable one . this is not unreasonable as the initial ( thermal ) state can be expanded in a power series in @xmath141 and the higher orders @xmath142 become less local with longer entanglement range as we increase the power @xmath143 . these results are summarized in fig . [ fig : rhot ] . in this section we would like to demonstrate the connection between quantum chaos , ( non)integrability and transport in non - equilibrium steady states of interacting quantum chains @xcite . within the linear response theory , the property of quantum mixing ( which is typically implied by quantum chaos , i.e. by the validity of random matrix level statistics @xcite , see also correlation functions in sect.[sect : losch ] ) , is typically synonymous for normal quantum transport since it implies _ finite _ kubo transport coefficients @xcite ( provided temporal correlation functions decay fast enough ) . however , here we would like to address the connection between quantum chaos and transport in far - from - equilibrium steady state , which may be beyond applicability of linear response . in particular , we are interested in the validity @xcite of the fourier law @xmath144 in quantum chains , relating the macroscopic heat flux @xmath145 to the temperature gradient @xmath146 . to investigate this problem one has to deal with a finite open system connected to heat baths . here we consider an interacting quantum spin-@xmath2 chain ( [ eq : hamaut ] ) which exhibits the transition from integrability to quantum chaos as a parameter , e.g. the magnetic field , is varied . the standard treatment of this problem is based on the master equation , thus limiting numerical investigations to relatively small system sizes . by using this method , in an interesting paper @xcite , the decay of current correlation function in a model of non - integrable chain of quantum spins is computed . however , these results were not fully conclusive and the conclusions rely on linear response theory . also , in @xcite lindblad formalism was used to study the validity of fourier s law for different type of spin - spin interaction . here we describe a different approach ( see ref.@xcite for details ) namely we follow the evolution of the system described by a _ pure _ state , which is _ stochastically _ coupled to an idealized model of heat baths . stochastic coupling is realized in terms of a local measurement at the boundary of the system and _ stochastic _ but _ unitary _ exchange of energy between the system and the bath . by this method we have been able to perform very effective numerical simulations which allow to observe a clear energy / temperature profile and to measure the heat current @xmath145 . again we consider ising spin chain in the magnetic field ( [ eq : hamaut ] ) of size @xmath0 , where the first and the last spin are coupled to thermal baths at temperatures @xmath147 and @xmath148 , respectively . in the non - integrable regime where the spectral statistics is described by rmt - the regime of quantum chaos - we found very accurate fourier law scaling @xmath149 , where @xmath0 is the size of the chain . in the integrable and near - integrable ( _ non - ergodic _ ) regimes instead , we found that the heat transport is ballistic @xmath150 . let us describe the numerical simulations . again we consider the autonomous model ( [ eq : hamaut ] ) at three particular cases : ( _ i _ ) _ quantum chaotic case _ @xmath151 at which lsd agrees with rmt and thus corresponds to the regime of quantum chaos , ( _ ii _ ) _ integrable case _ @xmath152 , at which lsd is close to the poisson distribution , and ( _ iii _ ) _ intermediate case _ @xmath153 at which the distribution lsd shows and intermediate character of weak level repulsion and exponential tail @xcite . let us now turn to study the energy transport in this model system . to this end we need to couple both ends of the chain of spins to thermal baths at temperatures @xmath154 . we have devised a simple way to simulate this coupling , namely the state of the spin in contact with the bath is statistically determined by a boltzmann distribution with parameter @xmath155 . our model for the baths is analogous to the stochastic thermal baths used in classical simulations @xcite and we thus call it a _ quantum stochastic thermal bath_. in the representation basis of @xmath156 the wave function at time @xmath58 can be written as @xmath157 where @xmath158 represents the _ up _ , _ down _ state of the @xmath159-th spin , respectively . the wave function at time @xmath58 is obtained from the unitary evolution operator @xmath160 . the interaction with the bath is not included in the unitary evolution . instead , we assume that the spin chain and the bath interact only at discrete times with period @xmath161 at which the states of the leftmost ( @xmath162 ) and the rightmost ( @xmath163 ) spins are stochastically reset . thus , the evolution of the wave function from time @xmath58 to time @xmath164 can be represented as @xmath165 where @xmath166 represents the _ stochastic _ action of the interaction with the left and right baths at temperatures @xmath167 and @xmath168 respectively . the action of @xmath169 takes place in several steps : * ( _ i _ ) the wave function is first rotated by the angle to the eigenbasis of the components @xmath170 , @xmath171 of the edge spins along the field @xmath172 , that is @xmath173 . here , @xmath174 stands for the magnitude of the magnetic field . * ( _ ii _ ) a local measurement of the observables @xmath175 , @xmath176 is performed . then the state of the spins at the borders collapses to a state ( @xmath177,@xmath178 ) with probability @xmath179 , we put all coefficients @xmath180 with @xmath181 to zero . * ( _ iii _ ) the new state of the edge spins @xmath182 is stochastically chosen . after this action simulating the thermal interaction with the baths each of the edge spins is set to _ down _ , ( _ up _ ) state with probability @xmath183,(@xmath184 ) . if the new state of @xmath162 , or @xmath163 , is different than the one after step ( ii ) , then a simple unitary spin flip is performed to the wave - function . the probability @xmath185 depends on the canonical temperature of each of the thermal baths : @xmath186 * ( _ iv _ ) finally , the wave function is rotated back to the @xmath187 basis , @xmath188 . this completes the description of the interaction with the quantum stochastic bath . this interaction thus ( periodically ) resets the value of the local energy @xmath189 of the spins in contact with the baths . therefore , the value of @xmath161 controls the strength of the coupling to the bath . we have found that , in our units , @xmath190 provides an optimal choice . we have nevertheless performed simulations for other values of @xmath161 with qualitatively similar results . in particular , for weak couplings ( @xmath191 ) the heat conductivity does not depend on the coupling strength . note that our method does not correspond to a standard stochastic unraveling of a master equation for the density operator ( _ e.g. _ , in the lindblad form ) @xcite . however , we have tested that , using our prescription , averages over the ensemble of quantum trajectories or time averages of one given quantum trajectory are sufficient to reconstruct a density matrix operator @xmath192 that correctly describes the internal thermal state of the system in and out of equilibrium . for each run the initial wave - function @xmath193 of the system is chosen at random . the system is then evolved for some relaxation time @xmath194 after which it is assumed to fluctuate around a unique steady state . measurements are then performed as time averages of the expectation values of suitable observables . we further average these quantities over different random realizations of `` quantum trajectories '' . out of equilibrium energy profile @xmath195 for the chaotic chain . the temperatures of the baths @xmath196 and @xmath197 , are both in the high temperature regime . results for chains of size @xmath198 ( crosses ) , @xmath199 ( open circles ) and @xmath200 ( solid circles ) are shown . the dashed line was obtained from a linear fit of the data for @xmath200 for the @xmath201 central spins . insets ( i ) and ( ii ) show the energy profile for the integrable and intermediate cases respectively , for @xmath202 . ] in order to compute the energy profile we write the hamiltonian ( [ eq : hamaut ] ) in terms of local energy density operators @xmath203 : @xmath204 such that the total hamiltonian ( [ eq : hamaut ] ) can be written as @xmath205 apart from the boundary corrections . first we have performed equilibrium simulations in order to show that time averaged expectation values of the local energy density can be used to determine canonical local temperature . to this end we set the left and right baths to the same temperature @xmath155 . for low @xmath155 , the energy per site @xmath206 saturates to a constant which , together with the entire energy profile @xmath195 , is determined by the ground state . however , for larger @xmath207 , the energy profile is constant within numerical accuracy , and numerical simulations give @xmath208 , all results being almost independent of @xmath0 for @xmath209 . the numerical data for @xmath210 can be well approximated with a simple calculation of energy density for a two - spin chain ( @xmath211 ) in a canonical state at temperature @xmath155 , namely @xmath212 . therefore , if the temperatures of both baths are in high @xmath155 regime , then we can define the local temperature via the relation @xmath213 . we stress that equilibrium numerical data shown are _ insensitive _ to the nature of dynamics ( consistent with results of ref.@xcite ) , whether being chaotic , regular or intermediate . size dependence of the energy current in the chaotic chain with @xmath214 and @xmath197 . we show @xmath215 ( open circles ) and @xmath216 ( open squares ) . the dashed lines corresponds to @xmath217 scaling . in the inset , the size dependence of the energy current is shown for the integrable ( solid circles ) and the intermediate ( solid squares ) cases . ] in fig . [ fig:3 ] we show the energy profile @xmath195 for an out of equilibrium simulation of the chaotic chain . in all non - equilibrium simulations , the temperatures of the baths were set to @xmath218 and @xmath219 . after an appropriate scaling the profiles for different sizes @xmath0 collapse to the same curve . more interesting , in the bulk of the chain the energy profile is in very good approximation linear . in contrast , we show that in the case of the integrable ( inset i ) and intermediate ( inset ii ) chains , no energy gradient is created which is a characteristic of ballistic transport . we now define the local current operators through the equation of continuity : @xmath220 = - ( j_{n+1 } - j_n)$ ] , requiring that @xmath221 $ ] . using eqs . ( [ eq : h_local ] ) and ( [ eq : hamaut ] ) the local heat current operators are explicitly given by @xmath222 . in fig . [ fig:4 ] we plot @xmath215 as a function of the size @xmath0 of the system for sizes up to @xmath200 . the mean current @xmath145 is calculated as an average of @xmath223 over time and space @xmath159 . the energy difference was obtained from the energy profile as @xmath224 . two spins near each bath have been discarded in order to be in the bulk regime . since @xmath225 is an effective size of the truncated system , the observed @xmath226 dependence confirms that the transport is normal ( diffusive ) . moreover , also the quantity @xmath227 , where @xmath228 , shows the correct scaling with the size @xmath0 . on the other hand , in integrable and intermediate chains we have observed that the average heat current does not depend on the size @xmath229 , clearly indicating the ballistic transport . let us now come back to the kicked ising model ( [ eq : hamiltonian ] ) and try to consider some very elementary but fundamental questions considering its dynamics and non - equilibrium statistical mechanics . observing data of fig . [ fig : corr ] in section ( [ sect : losch ] ) one can conclude that the model perhaps displays an interesting order to chaos , or non - ergodicity to ergodicity transition when the integrability breaking parameter is increased . now we would like to inspect this transition more closely , and in particular understand the rate of relaxation to equilibrium in the ergodic and mixing case . we should stress right from the start that we are unable to prove any non - trivial statements about the model , but we can provide many suggestive numerical experiments which can be performed in a rather efficient way . we have learned in section [ sect : mps ] that it may be more fruitful to consider time evolution in the operator algebra spaces instead of in the spaces of pure states . let us go now a bit deeper into this subject . for some related results on high - temperature relaxation in isolated conservative many - body quantum systems see e.g. refs.@xcite . time automorphism on unital quasi - local @xmath230 algebra ( see e.g. @xcite for introduction into the subject ) @xmath231 , @xmath232 of an infinite ki lattice , for one period of the kick , can be explicitly constructed by the following observations . formally , for any @xmath233 , @xmath234 , where @xmath235 is given by either ( [ eq : floquet],[eq : protocol],[eq : protocol2 ] ) . let @xmath236}$ ] , with @xmath237 , denote a finite , local @xmath238 dimensional algebra on a sub - lattice @xmath239 \subset \z$ ] , which is spanned by operators @xmath240 . it is straightforward to prove that dynamics is _ strictly local _ @xmath241 } \to { \cal a}_{[m-1,n+1]}. \label{eq : locality}\ ] ] in other words , the homomorphism @xmath242 ( [ eq : locality ] ) is a simple nontrivial example of a quantum cellular automaton as defined by schumacher and werner @xcite . @xmath231 can also be treated as a hilbert space with respect to the following inner product @xmath243 , where @xmath244 is a tracial state , @xmath245 for @xmath246}$ ] . this hilbert space can be in fact considered as a 1d lattice of 4-level quantum systems ( qudits with @xmath247 ) with the orthonormal basis @xmath248 labeled by an infinite sequence of 4-digits @xmath249 . restricting for a moment to a dimer lattice @xmath250}$ ] we can write the adjoint action of a 2-qubit gate @xmath251 ( [ eq : wgate ] ) in terms of a @xmath252 unitary matrix @xmath253 where , very explicitly @xmath254 . \quad\quad\ ] ] we should note that the map is unital @xmath255 , and that due to anti - unitary symmetry of ki model , the matrix @xmath256 is _ real_. we extend the map @xmath257 to entire algebra @xmath231 by @xmath258 , for any @xmath259}$ ] , @xmath260}$ ] . now , following the protocol ( [ eq : protocol2 ] ) we finish the construction of the time automorphism as a string of right - to - left ordered 2-qudit ( with @xmath247 ) gates @xmath261 explicitly , for any local observable @xmath262}$ ] , we have the following * algorithm 1 * : 1 . set an initial vector : @xmath263 . 2 . for @xmath264 : @xmath265 3 . the result is @xmath266 the algorithm produces exact result in @xmath267 multiplications and about the same number of additions . let @xmath268 } : { \cal a}_{\z } \to { \cal a}_{[m , n]}$ ] denote a _ linear orthogonal projector _ , satisfying @xmath268}(a\otimes e ) = ( \one|e)a$ ] if @xmath269}$],@xmath270}$ ] . let us define a truncated time evolution operator @xmath271 } = \ap_{[m , n ] } \at : { \cal a}_{[m , n ] } \to { \cal a}_{[m , n]}$ ] which we can actually implement on a computer with a finite memory register . the `` infinite - temperature '' time correlation function of ( traceless ) local quantum observables @xmath272}$ ] can be written as @xmath273 , where @xmath274 . an interesting question is , up to what time @xmath58 the @xmath275 can be computed numerically exactly with a finite computer register @xmath276 $ ] of @xmath277 qudits of size @xmath278 ? due to locality ( [ eq : locality ] ) of time homomorphism one can easily prove that @xmath279}^t a } , \quad { \rm for } \quad t \le 2l \label{eq : excorloc}\ ] ] hence the correlation functions are computable exactly up to time @xmath280 , and as we shall see later , truncated correlation function @xmath281 } a}$ ] often well approximates @xmath275 even at later times , or even its asymptotic decay . let us continue our discussion by considering time evolution for _ translationally invariant extensive _ ( tie ) observables . given some quasi - local observable @xmath282 we shall construct the corresponding tie observable by a formal mapping , @xmath283 in terms of lattice translation automorphisms @xmath284 , @xmath285 . the image of the entire quasi - local algebra under this mapping @xmath286 , is not a @xmath230 algebra , but it is a linear space which can be again turned into a hilbert space with the following inner product @xmath287}(x)}{\ap_{[-n , n]}(y)}\ ] ] where the domain of projector @xmath288}$ ] is extended to @xmath289 by continuity . orthonormal basis of @xmath289 is given by tie observables @xmath290 , for _ orders _ @xmath291 , and for uniqueness of notation , requiring @xmath292 . we shall interchangeably represent finite sequences of 4-digits with integers , @xmath293 . let @xmath294 be @xmath295 dimensional subspace spanned by tie observables @xmath296 with order @xmath297 , i.e. for @xmath298 having at most @xmath299 base-4 digits , so we have an inclusion sequence @xmath300 . since time and space automorphisms commute @xmath301 , one can immediately extend the time map onto the space of tie observables , @xmath302 by continuity . formally , we have @xmath303 . furthermore , locality ( [ eq : locality ] ) implies @xmath304 so we can write a simple adaptation of algorithm 1 for explicit construction of a time map of an arbitrary finite - order tie observable @xmath305 : * algorithm 2 * : 1 . take the following pre - image of the tie observable @xmath306}$ ] , namely @xmath307 . 2 . compute @xmath308 of @xmath309}$ ] according to * algorithm 1*. 3 . transforming back to @xmath289 the result reads @xmath310 let us further define the _ natural truncations _ of tie space to order @xmath299 , @xmath311 as orthogonal projections @xmath312 , and truncated time evolution operators @xmath313 , which are naturally implemented on a computer by simply truncating overflowing coefficients @xmath314 . physically interesting question now concerns computation of time correlation functions between a pair of finite order ( say @xmath315 ) tie observables @xmath316 , namely @xmath317 , for example in fig . [ fig : corr ] we have shown the case of @xmath318 . as a consequence of locality ( [ eq : locality ] ) , and translational invariance , we find that the truncated evolution on @xmath294 , reproduces correlation functions exactly @xmath319 few remarks are in order : ( i ) truncated translationally invariant time evolution @xmath320 is perhaps more natural object to study than truncated local time evolution @xmath271}$ ] , for a simple reason that the truncation @xmath321 commutes with a shift @xmath322 , while @xmath268}$ ] does not . ( ii ) a space @xmath289 can be identified with translationally invariant linear functionals over @xmath231 , namely @xmath323 , @xmath324 , @xmath282 . we have @xmath325 and @xmath326 where hermitian adjoint maps @xmath327 and @xmath328 simply correspond to _ time reversed _ dynamics . ( iii ) convex subspace of positive translationally invariant functionals @xmath329 is an interesting invariant subspace of physical states , @xmath330 . in classical mechanics of chaotic systems one typically observes that states ( phase - space densities ) develop small details in the course of time evolution at an exponential average rate . consequently , introducing a small stochastic noise of strength @xmath113 to perron - frobenius operator ( pfo , i.e. liouvillian propagator for discrete time dynamics ) , makes it non - unitary and shifts its spectrum inside the unit circle . typically , the effect of noise is equivalent to an ultraviolet cutoff - truncation of pfo - at the fourier scale @xmath331 , and often the leading eigenvalues of truncated pfo - the so - called ruelle resonances - remain frozen inside the unit circle in the limit @xmath332 ( or @xmath333 ) @xcite . for a general introduction to relaxation phenomena in classical hamiltonian dynamics see e.g. @xcite . let us now draw some some analogies with our quantum setting . we have seen that that the evolution @xmath334 somehow most closely resembles liouvillian evolution of classical hamiltonian dynamics . in hilbert space topology , operator @xmath334 is unitary and its spectrum lies on a unit circle , just like in the case of classical pfo . however , truncated @xmath335 matrices @xmath320 represent natural `` ultraviolet '' cutoff truncations for increasing orders @xmath299 . let us check numerically if some eigenvalues of these matrices remain frozen when @xmath336 . indeed , as we demonstrate in fig . [ fig : s ] , we find several eigenvalues which converge as @xmath299 increases in the case of strongly non - integrable ( quantum chaotic ) case , with a gap between an eigenvalue of maximal modulus and the unit circle , whereas in the integrable case a set of @xmath299 eigenvalues touch the unit circle ( actually eigenvalue 1 is @xmath299-fold degenerate ) . numerical results suggest the following speculative conclusions . let @xmath337 be the converged ( frozen ) eigenvalues of @xmath320 , and @xmath338 , @xmath339 the corresponding _ right _ and _ left _ eigenvectors , respectively . then for arbitrary pair @xmath340 , the time correlation function can be expressed in terms of spectral decomposition ( see e.g. @xcite ) @xmath341 the above relation is the contribution of the point spectrum and is exact if the spectrum is pure - point . however , in classical cases one may quite typically have various singular components and branch cuts @xcite . note that the denominator @xmath342 is finite although both vectors should have infinite @xmath343 norm @xmath344 , @xmath345 , for any eigenvalue away from the unit circle , @xmath346 . there is a simple relation between the spectrum of pfo and the ergodic properties of dynamics : ( i ) if there is a _ spectral gap _ , i.e. there exists @xmath347 such that for all @xmath159 , @xmath348 , then dynamics is _ exponentially mixing _ , @xmath349 ( ii ) if some eigenvalues are on the unit circle , meaning that the corresponding eigenvector coefficients should be in @xmath343 , then the system is _ non - mixing _ since there are correlation functions which do not decay . ( iii ) if some eigenvalues are at @xmath350 then the system is _ non - ergodic _ since the correlation functions may have non - vanishing time - averages . if @xmath351 is a complete set of orthonormalized eigenvectors corresponding to eigenvalue @xmath350 , @xmath352 ( and note that since we are on the unit circle : @xmath353 ) then @xmath354 the latter ( iii ) happens in generic completely integrable quantum lattices , where @xmath351 correspond to an infinite sequence of conservation laws @xcite . furthermore , we have a strong numerical evidence that also in certain non - integrable quantum lattices @xcite , and also in ki model @xcite , one has a regime where few normalizable ( pseudo - local ) but not local ( like in integrable models ) conservation laws exist . this situation we call the regime of _ intermediate _ dynamics and is characterized by a non - vanishing _ stiffness _ @xmath355 signalling _ ballistic transport_. in figure [ fig : c ] we compare the time autocorrelation function of the transverse magnetization @xmath356 computed in three different ways : ( 1 ) from exact time evolution @xmath357 on a finite lattice of length @xmath0 with periodic boundary conditions , ( 2 ) iteration of truncated ta matrix on infinite lattice @xmath358 and ( 3 ) asymptotics based on ( few ) leading eigenvalue resonance(s ) ( using formula ( [ eq : corr ] ) in terms of @xmath359 and @xmath360 . ) . . gray level indicates the spectral gap @xmath361 of @xmath320 as a function of @xmath16 and @xmath17 at fixed @xmath362 . white regions correspond to @xmath363.,width=302 ] we note that the leading eigenvalue and eigenvector of truncated ta @xmath320 can most efficiently be computed using our algorithm 2 as a key step of an iterative _ power - method_. in this way we were able to perform calculations of the leading ruelle resonances up to @xmath364 in contrast to full diagonalization of truncated matrices @xmath320 which were feasible only up to @xmath365 . let us observe the structure of the eigenvector coefficients @xmath366 corresponding to the leading eigenvalue . numerical results ( see fig . [ fig : v ] , see also subsect.[sect : scaling ] later ) strongly suggest self - similar behaviour upon multiplying the code @xmath298 by @xmath367 which is a consequence of the fractal structure of the transfer matrix ( see illustration in ref.@xcite ) . as a function of @xmath16 at @xmath368 at different truncation orders @xmath299 . in the inset we indicate a line of transition in 2d phase diagram . , width=151 ] as a function of @xmath369 at @xmath370 at different truncation orders @xmath299 . in the inset we indicate a line of transition in 2d phase diagram . , width=151 ] the stiffness @xmath371 and the spectral gap @xmath372 may be considered as _ order parameters _ , characterizing a particular kind of dynamical phase transition , namely the transition from non - ergodic dynamics - ordered phase , where @xmath355 for a typical @xmath373 which is not orthogonal to all conservation laws @xmath351 , i.e. normalizable eigenvectors of @xmath334 with eigenvalue @xmath350 . ] , or @xmath374 , to an ergodic and mixing dynamics - disordered phase , where @xmath375 , for all traceless @xmath324 , or @xmath376 . we know that ki model is non - ergodic in integrable regimes let us consider a fixed transverse field case @xmath11 , and start to switch on a small amount of longitudinal field @xmath16 . the interesting question is whether the transition happens for infinitesimal integrability breaking parameter @xmath16 in tl , or at a finite - _ critical _ field . in fig . [ fig : phasediagram ] , we fix @xmath362 and plot a two dimensional phase diagram of the spectral gap @xmath377 . it is clear that we have different behaviours in different regions of parameter space , for example we identify two : ( i ) type i transition : if the transverse field is roughly on the interval @xmath378 $ ] , then the spectral gap opens in the fastest possible manner which is allowed by a @xmath379 symmetry and the analyticity of the problem , namely @xmath380 . see fig . [ fig : typei ] . ( ii ) type ii transition : if the initial transverse field @xmath381 , or @xmath382 , then the gap opens up in a much more abrupt - perhaps a discontinuous way . we give an example by scanning the diagonal transition , i.e. putting @xmath369 and increasing @xmath16 from zero . numerical results , shown in fig.[fig : typeii ] can not be made fully conclusive , but they are not inconsistent with a conclusion that an abrupt transition to ergodic behaviour takes place at @xmath383 . there is another possibly interesting way of characterizing the transition , i.e. in terms of pseudo - local translationally invariant conservation laws @xcite . such conservation laws are the square normalizable elements @xmath384 , which are mapped onto themselves under the dynamics @xmath385 . we shall first make a non - trivial variational mpo ansatz for elements of @xmath289 , namely let us take an auxiliary vector space @xmath386 , where @xmath387 . then any operator @xmath388 , which is formally written in terms of mpo on the _ infinite _ spin chain @xmath389 where @xmath390 have the block matrix form ( for @xmath391 ) : @xmath392 represents a translationally invariant pseudo - local operator , i.e. an element of @xmath289 , provided that @xmath393 and @xmath394 , where @xmath395 is a spectral matrix norm and @xmath396 s stand for arbitrary matrices / vectors . of course , converse can not be generally true , not any element of @xmath289 can be written as mpo ( [ eq : timpo ] ) with finite @xmath96 , but still there are elements of the form ( [ eq : timpo ] , [ eq : mpodata ] ) which are not in @xmath294 for _ any _ finite @xmath299 . there exist a straightforward algorithm which performs time evolution on mpo data ( [ eq : mpodata ] ) , namely @xmath397 is also of the form ( [ eq : timpo],[eq : mpodata ] ) with dimension @xmath398 . we shall now make the following simple numerical experiment . let us fix @xmath96 , setting @xmath399 representing the simplest @xmath334-invariant subclass of ( [ eq : timpo],[eq : mpodata ] ) , and optimize ( maximize ) the fidelity - like quantity @xmath400 within this class of operators . let us write an operator which maximizes @xmath401 for a given @xmath96 as @xmath402 . note that @xmath403 gives a strong - topology measure of conservation of observable @xmath402 in one step of time evolution , so @xmath404 only for exact conservation laws . increasing @xmath96 may improve fidelity , if pseudo local conservation laws exist to which @xmath402 may converge , however in ergodic and mixing situation where no exact pseudo - local conservation laws exist , increasing @xmath96 should have no significant effect to fidelity @xmath405 . this is exactly what we observe in ki model following a line of type ii transition ( see fig . [ fig : conslaw ] ) . numerical results of section [ sect : mps ] suggested that operator space entanglement measures can be used to characterize the complexity of time evolution , namely the minimal required rank @xmath406 of mpo ansatz is simply related to entanglement entropy of a time - evolving local observable , which is interpreted as a hilbert space vector . in order to make things as simple and precise as possible we go back to the time evolution automorphism @xmath242 over the quasi - local spin algebra @xmath231 . let us take some truncation order @xmath407 and consider the truncated map @xmath408}$ ] . starting with local operators on a single site @xmath409}$ ] this truncated map is exact up to time @xmath410 . writing time evolved observable at any instant of time as @xmath411 } a(0 ) = \sum_{\ve{s } } a^{(t)}_{(s_{-n},\ldots s_n ) } { { | s_{-n},\ldots , s_n ) } } $ ] in terms of a ` wave - function ' @xmath412 , and partitioning a sub - lattice at @xmath413 , as @xmath414 = [ -n , m-1 ] \cup [ m , n]$ ] , we can define @xmath415 reduced super - density matrix as @xmath416 since @xmath417 is interpreted as a ` _ pure state _ ' , namely a vector from @xmath418}\otimes{\cal a}_{[m , n]}$ ] , the operator space entanglement is most simply characterized either by von neuman entropy @xmath419 or linear entropy @xmath420 where @xmath421 ^ 2 $ ] is a purity of reduced super - density matrix . in fig . [ fig : pursuperrho ] we plot the entanglement purity @xmath422 for close to symmetric bipartition @xmath423 where the entanglement is expected to be maximal for three different characteristic cases of ki model , which will be in the following referred to as : _ quantum chaotic _ ( qc ) , @xmath424 , _ integrable _ ( in ) , @xmath425 , and _ non - ergodic _ ( ne ) non - integrable case , @xmath426 . we find that in qc case purity decreases exponentially @xmath427 , meaning @xmath428 , where the exponent @xmath127 is independent of the initial observable @xmath429 and asymptotically independent of @xmath299 . on the other hand , in in case @xmath422 does not decay at all so the resulting dynamical entropy @xmath430 , whereas in ne case @xmath422 decays slowly , likely slower than exponentially . the eigenvectors of @xmath334 corresponding non - unimodular eigenvalues seem to exhibit a certain scaling invariance ( fig . [ fig : v ] ) . here we would like to explore this property in a little bit more detail . for that purpose we again explore the map @xmath408}$ ] on the quasi - local algebra @xmath431}$ ] since the representation of dynamics is conceptually simpler ( algorithm 1 ) than dynamics @xmath432 on @xmath433 . in particular , it is worth to mention that if one traces out an additional qudit after each time step , then the dynamics is exact on a _ closed _ set of @xmath434 super - density matrices and has a simple explicit form in terms of a completely positive matrix map . ] @xmath435 t^\dagger_{n - m}\right\ } , \label{eq : rdynamics1 } \\ & & t_{r } = \prod_{0\le k \le r}^- \one_4^{\otimes k}\otimes w\otimes \one_4^{\otimes ( r - k ) } , \label{eq : rdynamics2}\end{aligned}\ ] ] namely no truncation is needed since at time @xmath436 we are describing an exact observable on @xmath437}$ ] . we write @xmath438 for tracing out the least significant qudit , and @xmath439 is an elementary @xmath440 projector . note that @xmath441 is just a matrix of @xmath271}$ ] in the canonical basis @xmath442 . it is rather trivial to exactly solve this dynamics for a small finite @xmath443 , however this does not yield physically very useful information about the ki dynamics . one would wish to study the correct tl by first taking @xmath444 , and only after that @xmath445 , however this task seems almost computationally intractable . still , we were able to make some modest numerical experiments exploring this question , suggesting that for sufficiently strong integrability breaking ( say qc case ) the asymptotic matrix @xmath446 has a remarkable scale invariance if we coarse - grain it by tracing over its @xmath440 blocks : @xmath447 where @xmath448 is some scaling factor . this seems to be true for both orders of the limits @xmath449 , @xmath444 , although better numerical results have been obtained for the ` incorrect ' limit , namely letting the number of iterations @xmath449 for a finite register size @xmath159 , and then checking the convergence of results with increasing @xmath159 . before discussing numerical results we note another useful observation let us define and briefly study ki chain on a _ semi - infinite _ lattice @xmath450 $ ] , with the hamiltonian ( [ eq : hamiltonian ] ) for @xmath90 and open boundary condition on the left edge . now we consider a quasi - local algebra @xmath451 . the time automorphism is again strictly local @xmath452 } \to { \cal a}_{[0,n+1]}$ ] , and can be written as a semi - infinite product @xmath453 . in the definition of the truncated time map @xmath454 } = \ap_{[0,n ] } \at^+ : { \cal a}_{[0,n ] } \to { \cal a}_{[0,n]}$ ] truncation is needed only on the right edge . note that due to this property , simulation of local observables ( localized near the edge of the lattice ) is twice as efficient than on doubly - infinite lattice , meaning that with the same size of computer register one can exactly simulate for twice longer times . for example , computation of correlation functions @xmath455^t a}$ ] is exact @xmath456^t a } = \braket{b}{[\at^+_{[0,n]}]^t a},\quad { \rm for}\;\ ; t \le 2(n - q ) , \quad { \rm if}\;\ ; a , b\in { \cal a}_{[0,q]}.\ ] ] numerically inspecting the correlation functions of a simple local spin @xmath457 in fig . [ fig : cfsc ] we find clear asymptotic exponential decay for qc case , whereas for in and ne cases we find non - vanishing plateaus in the correlation function , i.e. non - vanishing stiffness @xmath458 which signals non - ergodicity and existence of local ( for integrable cases ) and pseudo - local ( for non - ergodic and non - integrable cases , e.g. ne ) conservation laws of @xmath459 . we note that the asymptotic correlation decay in non - integrable cases , say qc and ne , seems quite insensitive to increasing truncation order @xmath159 - indicating that the leading eigenvalues of @xmath454}$ ] remain frozen when increasing @xmath159 . note that the asymptotic exponents of correlation decay for a semi - infinite chain are not the same as for an infinite one , i.e. the point spectra of @xmath454}$ ] and @xmath460 are in general different , however we have some indications to believe that their phase diagrams should agree , namely ergodic regimes in the ki model on semi - infinite chain model are in one - to - one correspondence with ergodic regimes of the model on infinite chain . however , the most remarkable feature of dynamics @xmath459 is the following . splitting the truncated semi - lattice as @xmath461=[0,m-1]\cup[m , n]$ ] and following the time evolution of an observable @xmath462}]^t a(0 ) = \sum_{\ve{s } } a^{(t)}_{(s_{0},\ldots s_n ) } { { | s_0,\ldots , s_n ) } } $ ] in terms of a ` super - wave - function ' @xmath463 , one can again define the reduced super - density matrix as @xmath464 the dynamical equation for @xmath465 in terms of @xmath466 is _ exactly the same _ as for doubly - infinite lattice , namely eqs . ( [ eq : rdynamics1],[eq : rdynamics2 ] ) . hence also the conjecture ( [ eq : scaling ] ) on scaling invariance of @xmath446 should be the same for the two lattice topologies . however , numerical computations are much easier and thus the results are more suggestive for the semi - infinite case . in log scale ( units are arbitrary ) versus the partial tracing index @xmath143 , for cases qc ( a ) , ne ( b ) and in ( c ) . diamonds , stars , and squares represent data for the semi - infinite ki lattice with truncation size @xmath467 , ( diamonds , stars ) and @xmath468 ( squares ) , all for @xmath469 . finite number of time steps @xmath470 ( diamonds ) , just before the absorbing boundary affects any of the data shown , is compared to steady state observable @xmath471 ( stars ) . in the non - integrable cases ( a , b ) , data are compared also with steady state @xmath471 simulation of the same local initial observable @xmath472 on the two - sided ( doubly - infinite ) ki lattice with truncation size @xmath473 ( @xmath364 ) , and @xmath474 ( triangles ) . , width=604 ] let us now discuss some numerical results . we have always started from local initial operator @xmath472 . we took @xmath159 as large as allowed by existing computing resources , namely @xmath467 for the semi - infinite chain and @xmath473 for the infinite chain , and compared the data for asymptotic matrices @xmath475 ( in numerics @xmath58 has been chosen such that the results converged , typically @xmath476 ) with finite time data @xmath477 where time @xmath478 was set as large as allowed so that the data were still exact and no truncation was needed , typically @xmath479 . in all cases , numerical results were quite insensitive to small changes in truncation order @xmath159 . first we have computed the scaling of the principal matrix element , or the _ partial norms _ @xmath480 which , assuming ( [ eq : scaling ] ) , should asymptotically scale as @xmath481 ( see fig.[fig : scalsinorm ] ) . if asymptotic dynamics @xmath449 is determined by normalizable eigenvectors of @xmath459 , which necessarily correspond to uni - modular eigenvalues , then we should have @xmath482 . in qc case a clear scaling was observed for both topologies ( @xmath483 and @xmath484 ) with the same exponent @xmath448 , however the exponent was slightly different for @xmath475 and @xmath477 . in the cases of non - ergodic dynamics ( ne and in ) the results for two topologies were quite different . for semi - infinite topology we find very clearly that @xmath485 indicating that @xmath486 and even @xmath487 can be written as @xmath343 convergent sums of local operators . for initial tracing index @xmath488 , and @xmath489 ( diamonds ) , @xmath490 ( stars ) , @xmath491 ( squares ) , @xmath492 ( triangles ) , for non - integrable cases of ki chain on semi - infinite lattice truncated at @xmath467 , namely for case qc ( a , b ) , and case ne ( d , e ) . finite time data at @xmath470 kicks are shown in ( a , d ) , while steady state observables @xmath471 are analyzed in ( b , e ) , all starting from initial observable @xmath493 . for comparison , asymptotic steady state @xmath471 data for two - sided , doubly infinite ki chain , truncated at @xmath494 , and with initial tracing index @xmath474 , are shown in ( c , f ) , namely for qc case ( c ) and ne case ( f ) . note that data in plot ( b ) are practically exactly overlapping.,width=604 ] as a more quantitative test of conjecture ( [ eq : scaling ] ) we compare the upper - left ( ` most important ' ) @xmath252 block of the super density matrix scaled to a unit principal element @xmath495 . in fig.[fig : scalsi1d ] we plot the diagonal elements @xmath496 for different @xmath143 , while in fig.[fig : scalsi2d ] we plot 2d charts of the entire scaled density matrices @xmath497 . indeed we find for qc case that the matrix @xmath497 is practically insensitive to increasing truncation ( @xmath143 ) and to topology of the lattice ( semi - infinite versus infinite ) , both for finite time @xmath498 and ` steady - state ' @xmath471 . on the other hand , in non - ergodic cases , the scaling ( [ eq : scaling ] ) is typically broken . however , it seems to be observed in the steady state ( @xmath498 ) of ne case , which is ( in our setting ) probably a non - physical but still quite robust effect due to truncation ( a kind of absorbing boundary condition ) . - the grayness level is proportional to @xmath499 - for initial tracing index @xmath469 , and @xmath500 ( columns plots ) , for six different cases , qc , ne , and in of truncated semi - infinite ki lattice ( @xmath467 ) , at finite time @xmath470 , and asymptotic steady state @xmath471 ( row plots ) , always starting form initial observable @xmath493 . @xmath252 ` most important ' matrix elements are plotted with the matrix site @xmath501 at lower - left corner.,width=604 ] summarizing this subsection , we conjectured that in the regime of quantum chaos reduced super - density matrices of time - evolved observables typically obey the scaling law ( [ eq : scaling ] ) with the exponent @xmath448 which only depends on global dynamics on quasi - local algebra of observables and not on a particular choice of initial observable . we expect that accurate numerical calculations of exponent @xmath448 would be possible within a certain quantum dynamic renormalization group scheme , however its precise formulation at present remains an open problem . in section [ sect : mps ] , further elaborated in subsection [ sect : opent ] , we have proposed an entanglement in operator space of quasi - local algebra as a possible new measure of quantum algorithmic complexity . here we would like to compare this briefly to some established proposals of quantum dynamical entropy , such as for example cnt entropy @xcite or af entropy @xcite , both being ideally suited for a quantum dynamical system formulated in terms of time automorphism @xmath242 , tracial invariant state @xmath502 , and quasi - local @xmath230 algebra @xmath231 . we shall here briefly review af entropy which is conceptually simpler . let us start by taking a set of say @xmath503 elements @xmath504 of @xmath231 which form a partition of unity , namely @xmath505 . following the dynamics up to integer time @xmath58 , @xmath506 , we construct a set of @xmath507 elements depending on a multi - index @xmath508 , namely @xmath509 from the homomorphism property and unitality of dynamics it follows that @xmath510 is also a partition of unity , @xmath511^ * a^{(t)}_{\ve{\alpha } } = \one $ ] . hence using an invariant state @xmath502 one can form a _ positive _ , _ hermitian _ , _ trace - one _ , @xmath512 matrix @xmath513^ * a^{(t)}_{\ve{\alpha}}\right).\ ] ] @xmath514 can clearly be interpreted as a density matrix pertaining to dynamically generated partition , and its von neumann entropy generated per unit time defines the af entropy @xmath515,\ ] ] where , strictly speaking , supremum over all possible generating partitions has to be taken . it is not surprising that practical evaluation of @xmath516 is impossible except for rather trivial cases , such as dynamics generated by shift automorphism @xmath517 @xcite . however , one can easily show that a related _ linear af entropy _ @xmath518 ^ 2 , \label{eq : aflin}\ ] ] is tractable much more easily , while the behaviour of @xmath519 and @xmath520 is likely to be similar in practice . the key observation is to write the purity @xmath521 ^ 2 = \sum_{\ve{\alpha},\ve{\beta } } \rho^{(t)}_{\ve{\beta},\ve{\alpha } } \rho^{(t)}_{\ve{\alpha},\ve{\beta}}$ ] in terms of dynamics over a product algebra @xmath522 , with time automorphism @xmath523 and an invariant state @xmath524 . to this product structure we add a linear map @xmath525 depending on a generating partition @xmath526 @xmath527 if @xmath528 is a unit element in @xmath231 then @xmath529 is a unit element in @xmath530 , and purity can be written using a transfer - matrix - like approach as @xmath531^t ( \one \times \one)\right\ } .\ ] ] the idea can be worked out in detail for a _ complete _ generating partition of a local sub - algebra @xmath532}$ ] of size @xmath533 . there it turns out that , due to locality of dynamics ( [ eq : locality ] ) , the resulting purity is _ independent _ of @xmath534 , i.e. the supremum ( [ eq : aflin ] ) is already achieved by a rather modest partition of @xmath535 elements , and the map @xmath536 can be _ factored _ into a direct product of two maps acting separately on _ two independent copies _ of a _ semi - infinite lattice_. leaving out some technical details of derivation , the final result reads as follows . let us consider a time dependent @xmath537 matrix @xmath538 , representing a state on @xmath539 , with initial value @xmath540 , and dynamics given by the following completely positive matrix map @xmath541}\left\{t_{t+1 } \left [ e_{00}\otimes \one_4\otimes q(t ) \otimes e_{00 } \right ] t^\dagger_{t+1}\right\ } , \label{eq : qdynamics}\ ] ] where @xmath542 } r)_{j , k } = \sum_{s=0}^{15 } r_{s + 16j , s+16k } $ ] traces out two least important qudits , and unitary time evolution matrix @xmath543 is given in ( [ eq : rdynamics2 ] ) . then the purity , and linear dynamical entropy ( lde ) , are simply given as @xmath544 ^ 2 $ ] , @xmath545 , respectively . the asymptotic increase of @xmath79 per unit time yields the linear af entropy ( [ eq : aflin ] ) . again , for practical calculations it is convenient to consider truncation of matrices @xmath538 after each iteration ( [ eq : qdynamics ] ) , say at dimension @xmath278 . in fig . [ fig : smb ] we plot lde for different cases of ki dynamics , and we observe that lde is always clearly growing linearly @xmath546 , so the af entropy is always strictly positive , even in non - ergodic ( ne ) and integrable ( in ) cases , and that the results are robust and stable against changing the truncation order @xmath299 . perhaps this finding appears surprising , but one has to bear in mind that af and cnt entropies can be positive even for simpler dynamics , such as quasi - free flows on @xmath230 algebras . for _ ergodic _ dynamical systems on @xmath230 algebras one can use shannon - mcmillan - breiman ( smb ) theorem ( for classical smb theorem see e.g. @xcite , and @xcite for a possible quantum extension ) , which states that for _ typical _ sequences @xmath547 , _ multi - time correlation function _ ( mtcf ) should decay exponentially @xmath548 where the exponent @xmath81 , which should be essentially independent of @xmath549 , is equal to a dynamical entropy of the map @xmath242 with respect to an invariant state @xmath502 . for an interesting application of smb theorem in the context of quantum dynamical chaos see ref . @xcite . in our numerical experiments we considered truncated dynamics @xmath408}$ ] , writing truncation order as @xmath550 , and computed two kinds of mtcf : ( i ) for a uniform sequence @xmath551 , where @xmath552 ( or @xmath553 in which case the truncated lattice was placed as @xmath554 $ ] , hence @xmath555 ) , we estimated mtcf ( [ eq : mctf ] ) as @xmath556 } a_1 \at_{[-n , n ] } \cdots \at_{[-n , n ] } a_1}$ ] . ( ii ) for a random sequence @xmath547 , we took a complete generating partition @xmath557 on @xmath558}$ ] with @xmath559 elements , for @xmath560 or @xmath561 , and computed an average square modulus of mtcf @xmath562 } a_{\alpha_1 } \at_{[-n , n ] } \cdots \at_{[-n , n ] } a_{\alpha_{t-1}}}|^2 \rangle}}$ ] over 20 randomly sampled sequences @xmath547 . in both cases ( i , ii ) we have found a rather good agreement between @xmath563 and lde @xmath79 for the case of quantum chaotic dynamics ( fig . [ fig : smb ] ) . for average random mtcf we found rather good agreement even for non - ergodic cases ( ne and in ) , however mtcf for the uniform sequence exhibited big oscillating fluctuations there , indicating simply that smb theorem does hold for non - ergodic dynamics of ki model . summarizing , it seems that af entropy , even though it very cleanly generalizes the concept of kolmogorov - sinai classical dynamical entropy to quantum dynamical systems , can not be used as an indicator of quantum chaos or an indicator of computational complexity of quantum dynamical systems . nevertheless , it has to be stressed that quantum dynamical entropies can be related to the notion of quantum algorithmic complexity , mentioned in section [ sect : mps ] , by a quantum version of brudno theorem established in ref . in the present article we have reviewed several possible approaches to describe dynamic instabilities , relaxation phenomena , and computation complexity in the simple model of one - dimensional non - integrable locally interacting quantum many body dynamics . we have argued that the essential non - equilibrium statistical properties of quantum dynamical systems - in the absence of idealized external baths - may be crucially related to the integrability of the system , or in the complementary case , to the existence of the regime of quantum chaos . the general flavor which remains after such studies is that the non - integrable quantum many body problem at high temperature will preclude any exact and complete solution by its very nature . still it is hoped that a more complete ergodic theory of such systems could be developed , allowing for example for exact calculation of relaxation rates , scaling exponents of resonance eigenvectors , etc . it has been shown in many cases that integrable quantum systems have anomalous non - equilibrium statistical mechanics at high temperature , for example they exhibit ballistic transport . this can easily be understood as being a consequence of existence of ( an infinite sequence of ) exact local conservation laws which prevents quantum ergodicity , similarly as existence of canonical action variables in classical integrable systems prevents classical ergodicity . an interesting open question is the following : how strong integrability breaking perturbation is needed , in generic cases , to break all the exact conservation laws and yield normal ( diffusive ) transport ? in other words : a quantum kam theory is needed ! numerical experiments shown in this paper suggest an interesting possibility , namely that in some cases a finite , critical perturbation strength is required . i would like to thank g. casati , t. gorin , c. mejia - monasterio , t. h. seligman , and m. nidari , for collaborations on related projects and acknowledge stimulating discussions with f. benatti , j. eisert , s. fishman and p. zoller . this work has been supported by the grants p1 - 0044 , and j1 - 7347 of slovenian research agency . 9 g. vidal , phys . rev . lett . * 91 * , 147902 ( 2003 ) ; _ ibid . _ * 93 * , 040502 ( 2004 ) ; s. r. white and a. e. feiguin , phys . rev . lett . * 93 * , 076401 ( 2004 ) ; a. j. daley , c. kollath , u. schollw " ock and g. vidal , j. stat . mech . * 4 * , p04005 ( 2005 ) ; g. vidal , cond - mat/0605597 . d. gobert , c. kollath , u. schollw " ock and g. sch " utz , phys . e * 71 * , 036102 ( 2005 ) ; u. schollw " ock and s. r. white , in g. g. batrouni , and d. poilblanc ( eds . ) : effective models for low - dimensional strongly correlated systems , p.155 , aip , melville , new york ( 2006 ) . f. verstraete , j.j . garc ' ia - ripoll and j.i . cirac , phys.rev.lett . * 93 * , 207204 ( 2004 ) ; m. zwolak and g. vidal , phys.rev.lett . * 93 * , 207205 ( 2004 ) ; a.e . feiguin and s.r . white , phys.rev.b * 72 * , 220401(r ) ( 2005 ) . r. sch " affer , h .- j . st " ockmann , t. gorin , t. h. seligman , phys . lett . * 95 * , 184102 ( 2005 ) ; r. sch " affer , t. gorin , h .- j . st " ockmann , t. h. seligman , new j. phys . * 7 * , 152 ( 2005 ) ; t. gorin , t. h. seligman and r. l. weaver , phys . rev . e * 73 * , 015202(r ) ( 2006 ) . o. bratelli and d. w. robinson , _ operator algebras and quantum statistical mechanics i , ii _ , ( springer - verlag , 1979 , 1981 ) ; b. simon , _ the statistical mechanics of lattice gases _ , ( princeton university press , 1993 ) .

the ability to perform quantum error correction ( qec ) by feedback is a crucial step towards fault - tolerant quantum computation @xcite . an open challenge , that has drawn considerable interest recently@xcite , is to find the best strategy for this task . two nominally distinct feedback strategies for qec are the measurement - based and driven - dissipative approaches . the former has been more well - understood@xcite , owing to an existing foundation in classical control and feedback in engineering . in the measurement - based ( mb ) approach , a classical controller performs projective measurements of a set of multi - qubit stabilizer operators that encode the logical qubit@xcite in order to track errors and/or perform any necessary correction . thus for good performance , this approach requires both high - fidelity projective measurements and low - latency control electronics to process the measurement result within the relevant coherence times of the quantum system . the elements required for this mb strategy have been demonstrated for small quantum systems on various physical platforms such as rydberg atoms@xcite , trapped ions@xcite , photons@xcite , spin@xcite and superconducting qubits@xcite . however , a steady - state multi - qubit qec capability has yet to be achieved and one of the key questions for this development is whether the mb strategy is scalable to larger systems or whether an alternative approach is more optimal . one such alternative , driven - dissipative ( dd ) schemes@xcite , also called reservoir / bath engineering or coherent feedback ( as discussed below ) , utilizes coupling between the quantum system of interest and a dissipative environment to transfer the entropy caused by decoherence - induced errors out of the quantum system . they have been demonstrated on a variety of physical systems including atomic ensembles@xcite , trapped ions@xcite , mechanical resonators@xcite and superconducting qubits@xcite . moreover , experiments with trapped ions@xcite and superconducting qubits@xcite have demonstrated some of the basic elements of autonomous qec . these schemes do not require high - fidelity projective measurements , external control and its associated latency . they can also be described as autonomous or coherent feedback@xcite , where the reservoir coupled to the target quantum system can be considered as an effective `` quantum controller '' that reacts with quantum degrees of freedom@xcite . adjusting the feedback by changing the `` quantum controller '' , however , can be more challenging than re - programming a classical controller , built with conventional electronics . thus , a further question is whether one can combine this dd approach and the conventional mb approach with minimal negative consequences from their respective drawbacks . here we report an experiment in which we built a feedback platform utilizing a nearly quantum - limited measurement chain and a customized field - programmable gate array ( fpga ) system to perform mb and dd schemes within the same setup . the task of this platform was to stabilize an entangled bell state of two superconducting transmon qubits@xcite . this particular task of stabilizing a single state is a proxy for more general qec experiments where a manifold of states is protected . we realize , for the first time , an mb _ stabilization _ of a bell state by repeated active correction through conditional parity measurements@xcite . we compare this scheme to a dd entanglement stabilization scheme@xcite in which the conditional parity switch is autonomous . by performing both schemes on the same hardware setup and circuit qed ( cqed ) system@xcite , we shed light on their close connection and compare them on a level playing field . previous theoretical works have compared dd ( under the name of `` coherent feedback '' ) and mb for linear quantum control problems@xcite , such as for minimizing the time required for qubit state purification@xcite or for cooling a quantum oscillator@xcite . these comparisons showed coherent feedback to be significantly superior . in our particular setup , we find that distinguishing the superior approach among dd and mb is a more subtle task . the subtlety is two - fold . first , the performance difference depends on which process can be better optimized : the design of the cqed hamiltonian or the efficiency of quantum measurement and classical control . in the current experiment , we show that dd has better steady - state performance as the cqed hamiltonian parameters are engineered such that dd has a shorter feedback latency . but dd s advantage over mb is not immutable . as certain experimental parameters are improved , such as coherence times and measurement efficiency , mb s performance can catch up with dd . secondly , in the current situation in which neither the cqed hamiltonian parameters nor the measurement and control parameters are ideal , we can obtain a boosted performance by combining dd and mb to get the best of both worlds . we explored this by devising a heralding method to improve the performance of both stabilization approaches . this protocol exploits the high - fidelity measurement capability and the programmability of the feedback platform . the protocol is termed `` nested feedback '' since it has an inner feedback loop based on either the dd or mb scheme , and an outer loop that heralds the presence of a high - fidelity entangled state in real - time . previously , heralding schemes have been demonstrated for state preparation to combat photon loss or decoherence@xcite . extending such heralding capability to state stabilization will be a valuable addition to the qec toolbox . furthermore , the ability to herald in real time as opposed to post - selection is important for on - demand and deterministic quantum information processing since only successful events lead to subsequent processing . real - time heralding for entanglement stabilization is particularly challenging for superconducting qubits due to their shorter coherence times compared to other systems . in this article , we implement this real - time heralding capability on a time scale faster than the few microsecond coherence time of our qubit - cavity system . by extending the feedback platform developed primary for the mb approach to the dd approach , our results bring to light a new application of mb . adding a level of mb feedback can significantly improve performance beyond what a single layer of feedback , whether dd or mb , can achieve . the simplified schematic of our experimental setup is shown in fig . [ fig : schematic]a . two transmon qubits@xcite , alice and bob , are dispersively coupled to a three - dimensional aluminum cavity @xcite , with linewidth @xmath0 mhz ( see supp i , for other parameters ) . the qubit - cavity dispersive shifts are nearly equal and in the strong dispersive regime ( @xmath1 mhz , @xmath2 mhz ) with photon - number resolved qubit transition frequencies@xcite . the cavity output is amplified by a josephson parametric converter ( jpc ) operated as a nearly quantum - limited phase - preserving amplifier @xcite enabling rapid , single - shot readout@xcite and thus real - time feedback . the key component of the experiment is a controller realized with two fpga boards that both measure and actively control the cavity - qubit system . an essential operation for our experiment is a two - qubit joint quasi - parity measurement using the common readout cavity @xcite . as shown in fig . [ fig : schematic]b , the cavity is driven at @xmath3 ( both qubits in ground state ) and at @xmath4 ( both in the excited state ) at the same time . the output at @xmath3 and @xmath4 together distinguishes the even parity manifold @xmath5 from the odd parity manifold @xmath6 . when the two cavity output responses _ both _ have an amplitude below a certain threshold , the qubits are declared to be in odd parity ; when either one has amplitude above the threshold , the qubits are declared to be in even parity . we note that , unlike a true parity measurement , this readout actually distinguishes the two even parity states @xmath7 and @xmath8 , hence we refer to it as a `` quasi '' parity measurement . however , the feedback schemes described below apply the same operation on both even states , and thus we need only record the parity of the measured state . the choice of driving at the `` even '' cavity resonances rather than between the `` odd '' resonances ( @xmath9 and @xmath10 ) mitigates the effect of the @xmath11 mismatch , reducing associated measurement - induced dephasing of the odd manifold@xcite . the controller fpga a ( b ) modulates the @xmath3 ( @xmath4 ) drive to the cavity and also demodulates the response . the two fpgas share their measurements of the cavity response to jointly determine the parity . in addition , fpga a and b generate the qubit pulses to alice and bob , respectively , which are conditioned on the joint state estimation during real - time feedback . we first briefly outline the dd stabilization of entanglement , described in detail in ref . and . this stabilization targets the two - qubit bell state @xmath12 . figure [ fig : comparison]a displays the states coupled by the autonomous feedback loop . two rabi drives on alice and bob at their zero - photon qubit frequencies ( @xmath13 and @xmath14 , see supp . mat . sec . i ) couple the wrong bell state @xmath15 to the even states , @xmath7 , @xmath8 , in the energy manifold with zero cavity photons . a second pair of rabi drives at the @xmath16-photon qubit frequencies ( @xmath17 and @xmath18 , @xmath19 ) , with their relative phase opposite to the first pair , couple @xmath20 , @xmath21 to the bell state @xmath22 . the two cavity drives , at @xmath3 and @xmath4 connect the two manifolds and hence the combined action of the six drives transfers the population from @xmath7 , @xmath8 and @xmath23 to @xmath22 . finally , cavity photon decay brings @xmath22 back to @xmath24 . in effect , the cavity drives separate qubit states based on their parity , allowing one pair of rabi drives to move the erroneous odd population to the even states while the other pair transfers the even states population to @xmath25 . counterparts to these elements of the dd feedback loop can be found in the corresponding mb feedback scheme . the action of our mb algorithm is shown as a state machine in fig . [ fig : comparison ] . we describe the quasi - parity measurement @xmath26 by the projectors @xmath27 , @xmath28 and @xmath29 . we assign the outcomes @xmath30 to the even projectors , @xmath31 and @xmath32 and @xmath33 to @xmath34 . the mb algorithm is built with a sequence of correction steps , each of which consists of a conditional unitary and a quasi - parity measurement . the two possible states of the state machine correspond to whether we apply the unitary @xmath35 or @xmath36 , followed by the quasi - parity measurement . specifically , @xmath37 where a ( b ) denotes alice ( bob ) , and @xmath38 . in a correction step @xmath39 , the qubits are initially in either @xmath7 , @xmath8 or in the odd manifold , due to the projective quasi - parity measurement in step @xmath40 ; the controller then applies @xmath35 ( @xmath36 ) if @xmath41 in previous step reported @xmath42 ( @xmath43 ) . the effect of the state machine on the two - qubit states is shown in tab . [ tabu : mb ] , where the action of the controller during one correction step is described in terms of the four basis states , @xmath25 , @xmath15 , @xmath7 and @xmath8 ( the latter two are grouped in the `` even '' column ) . the quasi - parity measurement infidelity , labeled by @xmath44 ( @xmath45 ) , gives the error probability of obtaining an even ( odd ) parity outcome after generating an odd ( even ) state . because these measurement infidelities are small , the dominant events are those that occur without measurement errors . at each step , @xmath35 on either @xmath7 or @xmath8 followed by the quasi - parity measurement @xmath26 transfers the states to @xmath46 with 50% probability . since @xmath25 is an eigenstate of @xmath36 and @xmath26 ( modulo a deterministic phase shift that can be undone , see later discussion ) , these operations leave it unaffected . on the other hand , @xmath36 and @xmath26 transform @xmath47 into @xmath5 ; more generally , they take population in any other odd state ( i.e. , a superposition of @xmath25 and @xmath47 ) into @xmath25 and the even states . l4 cm | p1.8 cm | p1.8 cm | p1.8 cm | p1.8cm| p0.9 cm | p0.9 cm | p0.9 cm | p0.9 cm previous state & & & + @xmath48 & @xmath49 & @xmath50 & @xmath49 & @xmath50 & & + outcome probability & @xmath44 & @xmath51 & @xmath44 & @xmath51 & & + unitary & @xmath35 & @xmath36 & @xmath35 & @xmath36 & & + next state & even & @xmath52 & @xmath47 & even & & + by repeating a sufficient number of these correction steps in sequence , the controller stabilizes the target bell state irrespective of the initial two - qubit state . the similarity between this active feedback and dd is that mb also transfers population between different parity states by conditional rabi drives . however , while the rabi drives in dd are conditioned autonomously by the photon number in the cavity , the unitary rabi pulses in mb are conditioned by real - time parity measurement performed by active monitoring of cavity outputs . the pulse sequences for dd and mb are shown in fig . [ fig : comparison]b and e. in dd , a set of continuous - wave drives are applied for a fixed time @xmath53 and after some delay @xmath54 to allow remaining cavity photons to decay , a two - qubit state tomography is performed @xcite . the cavity and rabi drive amplitudes and phases were tuned for maximum entanglement fidelity , following the procedure described in ref . . in particular , the optimal cavity drive amplitudes were found to be @xmath55 . for mb , the continuous drives are replaced by a pre - defined number of correction steps @xmath56 , resulting in a stabilization duration of @xmath57 where @xmath58 @xmath59s . there is no extra delay before tomography since each correction step already contains a delay after the quasi - parity measurement due to feedback decision latency . the strength and duration of the quasi - parity measurement @xmath26 were optimized as discussed in supp . mat . the optimization achieved low parity measurement infidelities @xmath44 and @xmath45 while keeping the measurement - induced dephasing arising from the @xmath11 mismatch@xcite small compared to the natural decoherence in the same duration . we experimentally determined the infidelity of the quasi - parity measurement to be @xmath44 and @xmath45 of @xmath60 and @xmath61 , respectively . the quasi - parity measurement also causes a deterministic qubit rotation about the respective @xmath62 axis due to an ac stark shift @xcite ; this rotation was corrected within the unitary gate @xmath36 as discussed in supp . [ fig : comparison]c , f show the fidelity to the target bell state @xmath25 as a function of stabilization time for dd and mb , respectively . the fidelity rises exponentially with a characteristic time constant of @xmath63 @xmath59s ( @xmath64 @xmath59s ) and a steady - state fidelity of @xmath65 ( @xmath66 ) for dd ( mb ) . both fidelity values agree with numerical modeling based on master equation simulation , which gives @xmath65 and @xmath67 for dd and mb , respectively ( see supp . mat . iv ) . the exponential dependence is a signature of feedback and arises from the characteristic loop time . the experimentally determined time constants are in reasonable agreement with their simulated values of @xmath68 @xmath59s ( @xmath64 @xmath59s ) for dd ( mb ) . in mb , this loop time is related to the step length ( @xmath69 @xmath59s ) , which is given by the sum of the quasi - parity measurement duration ( @xmath70 @xmath59s ) , the cable , instrument and fpga latencies ( @xmath71 @xmath59s ) , and the duration of unitary pulses ( @xmath72 @xmath59s ) . on the other hand for dd , the measured loop time is close to @xmath73 cavity lifetimes , the expected time as shown in ref . . the superior performance of dd over mb for the steady - state fidelity is due to the difference in correction loop time , which needs to be shorter than the coherence times of the two qubits for high fidelity entanglement . for the current experimental setup , the latency of the controller and quantum efficiency of the measurement chain , which affects the fidelity of the single - shot readout , result in a longer loop time in mb . a source of the longer feedback loop time is the quasi - parity measurement duration . this measurement duration , which was optimized as discussed in supp . ii , is limited by dephasing induced by the mismatch in @xmath11 ( @xmath74 ) and the measurement efficiency of the output chain ( @xmath75 ) , which can both be improved in future experiments . our simulations ( see supp . mat . sec . iv ) suggest that with current state - of - the - art measurement efficiency value and optimization of the fpga / cable latency , the mb steady state fidelity can be improved to @xmath76 . the limited measurement efficiency does not affect the performance of dd because the parity measurement and correction take place autonomously within the qubit - cavity system , indicative of its robustness against this hardware limitation . on the other hand , both dd and mb schemes benefit from longer intrinsic coherence times and reduction of the @xmath11 mismatch . for example , simulations show that if the coherence times are improved to one hundred microseconds ( achieved in other state - of - the - art cqed setups ) , both dd and mb fidelities can increase to above 85% . for the rest of the article , however , we consider boosting the fidelity in a different manner , without making any physical changes to the qubit - cavity system . the dd and mb schemes described so far are synchronous in the sense that the stabilization always ends after a pre - determined duration and the tomography follows . decoherence causes the qubits to have a finite probability of jumping out of the target state immediately before the stabilization terminates . hence , this `` fixed time '' protocol does not always output the target state with maximum fidelity . an optimum protocol would rather utilize all available information . for both dd and mb , we can measure the cavity output at the end of the stabilization period . the outcomes of these measurements , @xmath77 and @xmath78 , give real - time information on the state of the two qubits , and thus can herald a successful stabilization sequence . in fig . [ fig : thresh_sweep ] , we describe how monitoring the cavity outputs improves target state fidelity . we introduce two thresholds @xmath79 ( see supplementary material for details ) to post - select the measurement outcomes of @xmath77 and @xmath78 respectively , and identify successful stabilization runs @xcite . the results of varying @xmath79 are shown in fig . [ fig : thresh_sweep]b , d for dd and mb , respectively . the color plots show fidelity improving as the thresholds become more stringent . the success probability defined as the percentage of stabilization runs kept for tomography given a set of thresholds , is also plotted as contours for both dd and mb . there is a clear trade - off between success probability and fidelity . to reach the maximum fidelity in dd of 82% , at least 75% of experiment runs need to be discarded . the trade - off is less severe in mb , where only 50% of runs need to be discarded to reach the maximum fidelity of 75% . however we aim to eliminate this trade - off all together , i.e. , to improve the fidelity while maintaining a high success probability . this goal is achieved by introducing a nested feedback protocol ( nfp ) , in which the stabilization feedback loop enters into a higher layer of feedback for `` fidelity boosting '' instead of proceeding to state tomography directly . in contrast to the `` fixed time '' protocol , nfp conditions the termination of stabilization on the quality of the entanglement , i.e. , it heralds a successful stabilization run in real - time , as illustrated by the state machine diagram in fig . [ fig : real_time]a . the control variable @xmath80 is given by @xmath81 , where @xmath79 are determined by the same post - selection experiment discussed previously to optimize the fidelity ( black square in fig . 3b and d ) . if the controller determines that the entanglement quality is not sufficient ( @xmath82 ) , a boost phase is attempted which comprises exactly one correction step for mb or a stabilization period of similar duration for dd ( @xmath64 @xmath59s ) . during the boost phase , the cavity outputs are integrated to give @xmath83 which enables the next real - time assessment of @xmath80 . in dd , the parity measurement and first layer of feedback is accomplished autonomously , therefore the fpga only needs to check @xmath80 . however in the mb scheme both layers of feedback are performed solely by the fpga . it therefore checks if @xmath84 to herald that the entanglement meets the desired quality . if not , it uses the quasi - parity thresholds ( grey circles in fig . 3d ) to decide whether the qubits are in even or odd state in order to continue stabilization . this asynchronous pulse sequencing and conditioning by multiple thresholds exploit the programmable nature of the fpga - based platform . the asynchronous behavior of nfp is displayed in fig . [ fig : real_time]b(e ) for dd ( mb ) , which demonstrates 200 single - shot runs . the dd ( mb ) fidelity boosting sequence continues until either success or a maximum limit on boost attempts ( set to 11 in the experiment ) is reached . for the mb protocol , the trajectory of the qubits parity can be tracked by the conditioning outcomes of the inner - loop control variable @xmath41 and the outer - loop control variable @xmath80 , which are independent . through repeated boost attempts until success , nfp significantly improves the overall success probability . within 11 attempts , 95% ( 99.8% ) of dd ( mb ) runs satisfy the success condition compared to just 25% ( 50% ) with simple post - selection . this is assessed by the cumulative probability , the integral of the probability of having completed a certain number of boost attempts before tomography , as plotted in fig . [ fig : real_time]c , f . since mb requires a less stringent threshold than dd to gain fidelity improvement , the mb success probability converges to unity much faster than that of dd . finally , we show that the high success probability does not come at the cost of reduced fidelity . the fidelity to @xmath25 for dd improves from an unconditioned value of 76% to 82% ( averaged over all successful attempts ) . for mb , the improvement is more pronounced : fidelity rises from an unconditioned value of 57% to 74% . thus for both dd and mb , nfp attain close to the fidelity achieved via stringent post - selection . these results for nfp also agree well with a numerical simulation ( see supp . mat . sec . one will note , however , a continuous downward trend of the fidelity in both dd and mb schemes as the number of attempts increases . this is due to the non - negligible population in the @xmath85 states of the two qubits in the experiment , which escape correction by the stabilization feedback loops . after each further boost attempt of stabilization , the probability of the population escaping outside the correction space thus increases , diminishing the fidelity ( see supp . mat . vii ) . also note that the error bars on the fidelity of mb are bigger than those in dd for large attempt numbers simply because the probability of needing many attempts is lower in mb than in dd . while real - time heralding by nfp removes the trade - off between fidelity and success probability , it does so by introducing a different trade - off high fidelity and success probability are achieved but the protocol length now varies from run to run . if nfp is a module within a larger quantum information processing ( qip ) algorithm , then this asynchronous nature must be accommodated by the controller . for our fpga - based control , nfp is easily accommodated because it is a natural extension to `` fixed - time '' or synchronous operation . in `` fixed time '' operation , the controller conditions its state by the protocol length which is pre - determined and stored in an internal counter by the experimenter . on the other hand in nfp , the controller conditions its state on a pre - determined logical function of its real - time inputs . in conclusion , we have implemented a new mb stabilization of an entangled state of two qubits , which parallels a previous dd stabilization scheme . instead of coherent feedback by reservoir engineering , mb relies on actively controlled feedback by classical high - speed electronics external to the quantum system . when comparing both schemes in the `` fixed - time '' protocol , we observe that dd gives a higher fidelity to the target state due to lower feedback latency . furthermore , we have improved the fidelity of both schemes by a nested feedback protocol which heralds stabilization runs with high - quality entanglement in real time . the real time heralding brings about the fidelity improvement without a common trade - off in qip : it does not sacrifice the experiment success probability . it eliminates this trade - off by allowing asynchronicity in the experiment . our experiment shows some of the key advantages of mb platforms that have not been previously explored . typically , the performance of mb feedback has not been at par with methods based on post - selection , due to the latency of the controller . however it is widely recognized that the trade - off of success probability for fidelity in the case of post - selection is untenable for large scale systems . therefore , existing digital feedback@xcite have focused on achieving nearly perfect success probability . here , we are exploring another direction of feedback which achieves high fidelity with high success probability . our nested feedback strategy maximizes the use of the information coming out of the qubit - cavity system in order to make the correction process as efficient as possible . we find that our feedback platform , comprised of a nearly - quantum - limited measurement chain and a real - time classical controller , provides the necessary tool - set to implement such a strategy . we show that this technology can be extended to improve the performance of dd approaches as well as single - layer mb approaches themselves . this strategy could be carried out further in the future . for example , the fpga state estimator could perform a more sophisticated quantum filter of the microwave output of the dd stabilization to herald successful events with better accuracy , significantly improving the success probability convergence rate . similar ideas can be applied in the future towards other forms of stabilization , such as for stabilizing schrdinger cat states of a cavity mode@xcite , a proposed logical qubit . initial experiments on such logical qubits with high fidelity - measurement@xcite or dissipation engineering@xcite have been performed and could now be combined . likewise , future logical qubits based on the surface code@xcite could also be stabilized by either active stabilizer measurements@xcite or as recently proposed by dissipation engineering@xcite . our experiment demonstrates that measurement - based and driven - dissipative approaches , far from being antagonistic , can be merged to perform better than either approach on its own . we thank zaki leghtas , mazyar mirrahimi , matti silveri and shantanu mundhada for helpful discussions . this research was supported by the u.s . army research office ( w911nf-14 - 1 - 0011 and w911nf-14 - 1 - 0563 ) . facilities use was supported by the yale institute for nanoscience and quantum engineering and nsf mrsec dmr 1119826 . all statements of fact , opinion or conclusions contained herein are those of the authors and should not be construed as representing the official views or policies of the u.s . government . 62ifxundefined [ 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 * * , ( ) link:\doibase 10.1103/physrevlett.105.040502 [ * * , ( ) ] link:\doibase 10.1103/physreva.86.032324 [ * * , ( ) ] link:\doibase 10.1103/physrevx.4.041039 [ * * , ( ) ] @noop _ _ ( , ) http://dx.doi.org/10.1038/nature10376 [ * * , ( ) ] @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) link:\doibase 10.1103/physrevlett.109.240502 [ * * , ( ) ] http://dx.doi.org/10.1038/nature11505 [ * * , ( ) ] link:\doibase 10.1103/physrevx.3.021008 [ * * , ( ) ] @noop * * , ( ) http://dx.doi.org/10.1038/nature12422 [ * * , ( ) ] http://dx.doi.org/10.1038/nature13171 [ * * , ( ) ] http://dx.doi.org/10.1038/ncomms5015 [ * * ( ) ] http://arxiv.org/abs/1411.5542 [ * * ( ) ] @noop * * , ( ) @noop * * ( ) link:\doibase 10.1103/physrevlett.77.4728 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.107.080503 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1103/physrevx.3.021013 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1103/physrevlett.109.183602 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.110.120501 [ * * , ( ) ] link:\doibase 10.1103/physreva.88.023849 [ * * , ( ) ] http://dx.doi.org/10.1038/nature12802 [ * * , ( ) ] @noop * * , ( ) @noop ( ) @noop * * , ( ) http://dx.doi.org/10.1038/nature10786 [ * * , ( ) ] @noop * * , ( ) \doibase http://dx.doi.org/10.1016/j.automatica.2009.04.018 [ * * , ( ) ] link:\doibase 10.1103/physrevb.77.180502 [ * * , ( ) ] link:\doibase 10.1103/physreva.81.040301 [ * * , ( ) ] link:\doibase 10.1103/physreva.82.012329 [ * * , ( ) ] link:\doibase 10.1103/physreva.69.062320 [ * * , ( ) ] @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) link:\doibase 10.1103/physrevlett.109.050506 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.109.050507 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1103/physrevlett.107.240501 [ * * , ( ) ] http://dx.doi.org/10.1038/nature05461 [ * * , ( ) ] http://dx.doi.org/10.1038/nature09035 [ * * , ( ) ] http://dx.doi.org/10.1126/science.1226897 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.102.200402 [ * * , ( ) ] link:\doibase 10.1103/physreva.81.062325 [ * * , ( ) ] http://stacks.iop.org/1367-2630/16/i=4/a=045014 [ * * , ( ) ] http://dx.doi.org/10.1038/nature13436 [ * * , ( ) ] link:\doibase 10.1103/physreva.91.062324 [ * * , ( ) ] ( a ) schematic of the experimental set - up . two independently addressable transmon qubits , alice and bob , are dispersively coupled to a three - dimensional microwave cavity . the cavity output is directed to a nearly quantum - limited measurement chain consisting of a josephson amplifier ( jpc ) followed by a semiconductor amplifier ( hemt ) . a pair of custom field - programmable gate array boards ( fpga a , b ) monitor the amplified output and generate real - time modulated microwave drives to control the cavity - qubit system . ( b ) transmission spectra of the cavity . cavity outputs at @xmath3 and @xmath4 are fed to fpga a and b , respectively . ] ] fig . comparison between the driven - dissipative ( dd ) and the measurement - based ( mb ) entanglement stabilization , shown in the left ( dd ) and right ( mb ) panels . ( a ) diagram of qubit / cavity state evolution in the dd feedback loop . two - qubit state manifolds are laddered for different photon numbers in the cavity ( labeled by @xmath86 ) . the green and pink sinusoidal double arrows represent cavity drives , while the straight double red / blue and cyan / yellow arrows are rabi drives on the qubits . these six drives and the cavity dissipation ( black decaying arrow ) couple the different states of the system such that the target state @xmath25 is stabilized . ( b ) functional pulse sequence for dd . the duration needed to empty the cavity of residual photons ( see text ) before tomography is indicated by @xmath54 . ( c ) fidelity to the target as a function of stabilization duration ( @xmath53 ) . dashed line at @xmath87 denotes the threshold for entanglement . the time given in the white box , @xmath88 , is the characteristic time constant of the exponential rise of fidelity . ( d ) state machine representation of the mb feedback loop . the quasi - parity measurement reports @xmath7 , @xmath8 as @xmath89 ( even ) and @xmath25 , @xmath15 as @xmath90 ( odd ) . for odd ( even ) parity , two @xmath91 pulses , with identical ( opposite ) phases , are applied to alice and bob , respectively . ( e ) sequence of correction steps conditioned by the quasi - parity measurement and leading into tomography . counter @xmath39 limits the number of steps to @xmath56 . ( f ) fidelity to the target bell state as a function of stabilization duration ( @xmath92 ) or number of correction steps ( @xmath56 ) . trade - off between success probability and fidelity for both dd and mb schemes . ( a ) pulse sequence for heralding dd stabilization by post - selection . at the end of the stabilization period , the cavity outputs , @xmath77 and @xmath78 are measured at their respective frequencies . ( b ) color plot of fidelity to the target state for dd as a function of thresholds chosen for @xmath77 and @xmath78 ( see supplementary material for details of the thresholds ) . also plotted as white dashes are contour lines of the success probability associated with each choice . solid black square indicates the thresholds chosen for the condition @xmath80 in the nested feedback protocol described in fig . ( c ) and ( d ) same as above for mb , including the corresponding thresholds for @xmath80 . gray circle indicates the thresholds chosen for the quasi - parity measurement @xmath41 used to condition mb correction steps . ] nested feedback protocol for boosting fidelity and heralding successful stabilization run in real - time . ( a ) state machine representation . the control variable @xmath80 ( see main text and fig . 3 ) determines the repetition of a boost cycle or the heralding of a successful run . the maximum number of boost attempts allowed is set to 11 in the experiment . ( b ) , ( e ) 200 single - shot sequence trajectories of nested feedback for dd and mb , respectively . the trajectories are colored yellow during boost attempts and black after @xmath80 is satisfied . trajectories that are entirely black satisfied the success criterion without the need for boost . the inset in ( e ) shows an example of an mb trajectory consisting of both @xmath41 and @xmath80 outcomes . ( c),(f ) cumulative success probability ( gray shade ) of having completed at most a given number of boost attempts before tomography for dd and mb , respectively . yellow bars indicate differential success probability . ( d),(g ) fidelity to target bell state for dd and mb , respectively . green squares show the corresponding fidelity as a function of the number of boost attempts . the blue dashed line denotes fidelity without any boosting ( i.e. unconditioned by @xmath80 ) . blue arrows denote the improvement in fidelity due to nested feedback . while the absolute fidelity of mb is worse than dd due to latency , the fidelity boost is higher . ] * supplementary materials * the cavity - qubit system was housed in an oxford triton 200 dilution refrigerator , at a base temperature below 20 mk . the setup of the experiment is described in detail in fig . [ fig : exp_setup ] . the key feature is that both data acquisition and arbitrary waveform / digital marker generation were accomplished by two fpga boards programmed with a customized logic . the cavity frequency is @xmath93 ghz when both qubits are in the ground state and its linewidth @xmath94 is @xmath95 mhz . the qubit frequencies with no cavity photons are @xmath96 ghz and @xmath97 ghz , for alice and bob respectively . the two dispersive shifts are @xmath1 mhz , @xmath2 mhz . the anharmonicity for the two qubits are @xmath98 mhz , @xmath99 mhz . alice ( bob ) has a @xmath100 of @xmath101 @xmath59s ( @xmath102 @xmath59s ) , @xmath103 of @xmath104 @xmath59s ( @xmath73 @xmath59s ) and excited state population in @xmath105 and @xmath106 of 5% each . as mentioned in the main text , the fast dispersive single - shot readout used in the experiment is made possible by a nearly quantum - limited phase - preserving measurement chain based on the josephson parametric converter(jpc ) . the jpc was operated with a gain of @xmath107 db and bandwidth of 6.2 mhz , with the frequency for maximum gain centered between the two readout frequencies , @xmath3 and @xmath4 . the gain and noise visibility ratio at both @xmath3 and @xmath4 were about 17 db and 6 db , respectively . [ sec : msmt time ] the quasi - parity measurement strength and duration were optimized in order to maximize the fidelity of mb . this optimization was done by maximizing the fidelity of the bell state created in a calibration experiment , similar to ref . . the qubits are prepared in ground states ( by post - selection ) and then two @xmath108 pulses , are applied to alice and bob , producing the state @xmath109 . the quasi - parity measurement , consisting of the two cavity drives on @xmath4 and @xmath3 respectively , projects the qubits into one of the two even states or entangles the qubits into a bell state with odd parity . we varied the duration of this parity measurement and its strength in terms of photon number ( set to be identical ) for each readout frequency to find the parameters that maximize the fidelity of the entangled state to the closest bell state ( fig . [ fig : measurement_calibration ] ) . the bell state fidelity would ideally increase and asymptotically approach one with increasing measurement time as the parity measurement better distinguishes the odd bell state from the even states . on the other hand , at long measurement duration , the coherence of the entangled state decreases due to both natural and measurement - induced dephasing , the latter of which is caused by the @xmath11 mismatch between the qubits and is proportional to the average number of photons used for the measurement @xcite . therefore , there is an optimal measurement strength and duration . for our experiment , @xmath110 for each readout frequency and a measurement duration of @xmath111 ns are found to be close to the optimal values and are chosen to attain a bell state fidelity of @xmath112% . the value of @xmath112% sets the upper bound on the fidelity that we should expect for heralding mb . in the actual mb experiment , an extra @xmath113 ns delay was introduced after the quasi - parity measurement in a correction step , which does not occur in the sequence described in this section for optimizing the parity measurement parameters . this extra delay was required to accommodate the feedback latency in mb . the conditioned fidelity we obtained for heralded mb is about 6% lower . [ sec : thresholds ] the cavity outputs at @xmath3 and @xmath4 for both dd and mb can be used to monitor the state of the qubits during stabilization ( fig . [ fig : thresh_hist]a ) . histograms of the measurement outcomes @xmath77 and @xmath78 , recorded by integrating the cavity output at @xmath3 and @xmath4 , respectively , are shown in fig . [ fig : thresh_hist]b , c , d , e for dd and mb respectively . in dd , the cavity output signals are captured while all the cw drives are still on , i.e. , the qubits are being driven while their states are being monitored , whereas in mb the outputs are a result of the quasi - parity measurements which occur after the qubit pulses and thus when the qubits are not driven . we observe that the measurement outcome distribution of dd lacks the separation seen in mb which has a clear parity separatrix @xmath114 . this feature also appears in numerical simulations of dd by the stochastic master equation @xcite . the state estimation used in both dd and mb uses the `` box car '' filtering @xcite which simply sums up the recorded cavity output signals over time to obtain the measurement outcomes . this method , while appropriate for mb , is not suited for dd since in the latter , the qubits are undergoing actively - driven dynamics when the measurement is taking place . a more advanced filter , such as a non - linear quantum filter can be designed from either first - principles or machine learning @xcite in the future to improve the state discrimination accuracy in dd . the measurement outcomes to the left of both the @xmath115 ( shown in figure ) and @xmath116 thresholds are much less likely to come from even states than those to the right . therefore the experiment runs with these outcomes are selected for state tomography , giving the results plotted as a color map in fig . 3 ( main text ) . moving the threshold further to the left increases the stringency of the threshold as fewer measurement outcomes are included . the success probability for each threshold choice ( plotted as contours in fig . 3 ) is calculated by the ratio of included outcomes to the total number of experiment runs . [ sec : supp model ] the steady state behavior of both dd and mb is simulated by a lindblad master equation , given by @xmath117+\kappa d[a]\rho(t)+\sum\limits_{j = a , b}\left(\frac{1}{t_{\downarrow}^j}d[\sigma_-^j]\rho(t)+\frac{1}{t_{\uparrow}^j}d[\sigma_+^j]\rho(t)+\frac{1}{2t_{\phi}^j}d[\sigma_z^j]\rho(t)\right)\ ] ] @xmath118 is the lindblad super - operator , defined for an operator @xmath119 as @xmath120\rho = o\rho o^{\dagger}-(1/2)o^{\dagger}o\rho-(1/2)\rho o^{\dagger}o$ ] . the pure dephasing rate for alice and bob , respectively , is given by @xmath121 , where @xmath122 . the hamiltonian @xmath123 is treated differently in dd and mb . for dd , the hamiltonian is described in detail in the theory proposal @xcite and parameters in the hamiltonian , such as the cavity and qubit drive amplitudes , are swept in simulation to find the optimal values . the optimal value for the cavity drive amplitude is found to be @xmath124 with @xmath125 , and @xmath126 for the qubit drive amplitudes at both zero - photon and n - photon qubit frequencies . the dd simulation predicts a characteristic time constant of 1 @xmath59s and a steady state fidelity of 76% ( accounting for the delay between stabilization and state tomography to allow remaining cavity photons to decay ) . for mb , a correction step is broken into four segments for effectively piecewise master equation simulation . the first part contains the conditional rabi pulses which are simulated as perfect instantaneous unitary operations on the qubits . the second part is the decay during the pulses ( 154 ns total ) . the hamiltonian during this part is just the the dispersive interaction between the qubits and the cavity , @xmath127 , in the rotating frame of the two qubits ( @xmath128 , @xmath129 ) and the cavity mode ( @xmath130 ) . the third part is the quasi - parity measurement during which the cavity drives at the @xmath3 and @xmath4 resonances are on and the hamiltonian is given by @xmath131 , where @xmath132 is the amplitude of the cavity drive ( 660 ns total ) . the last part is the remainder of the correction step , incurred by the latency of the feedback during which all drives are off and the qubit - cavity system is in free - decay . the dynamics in this part is again simulated by the dispersive interaction , @xmath133 ( 686 ns total ) . for the piecewise master equation simulation of a complete correction step as four segments , the density matrix at the end of a segment is used as the initial density matrix for the next segment . since in mb , the state at the end of a correction step depends only on the initial state at the beginning of the step , we can model the mb scheme as a markov chain ( fig . [ fig : mb_markov ] ) . in the rest of this section , we show how we derive the transition matrix that describes this markov chain . as discussed in the main text , we can describe the qubits by the density matrix @xmath134 . therefore , in terms of probability distributions in the four basis states , the qubits s state , @xmath135 , can be represented by a vector , @xmath136 if the qubits are prepared in @xmath25 , that is @xmath137 , we can calculate @xmath138 after a correction step by applying the master equation simulation method described above . we need to consider the two possible cases where the conditional unitary applied is @xmath36 or @xmath35 , respectively . the @xmath138 is a weighted average of the two cases . @xmath139 @xmath44 and @xmath45 are the quasi - parity measurement infidelities due to limited measurement efficiency , introduced in the main text . in a similar manner , we can obtain @xmath140 . in the case of an even initial state , for example , @xmath141 , we have @xmath142 and similarly for @xmath143 . given @xmath138 , @xmath140 , @xmath144 and @xmath143 , we can construct the transition matrix @xmath145 of a correction step , @xmath146 where the @xmath147 s are the columns of the 4 by 4 matrix . now applying this transition matrix on any arbitrary initial state gives the final state after a correction step , @xmath148 the transition matrix @xmath145 is also called the stochastic matrix , with the property that each column sums to 1 . one of @xmath145 s eigenvalues is guaranteed to be 1 and the corresponding eigenvector , @xmath149 , is the steady state of the markov chain . it can easily be shown that for any arbitrary initial state , @xmath150 @xmath151 for the given experimental parameters in mb , the @xmath145 matrix is displayed in fig . [ fig : mb_markov ] and we find the steady state eigenvector to be @xmath152 thus , the markov model predicts a steady state fidelity of 58% to @xmath25 . taking into account of the duration of a correction step ( 1.5 @xmath59s ) , we can also calculate the characteristic time constant of the mb scheme from the model , which gives 1.4 @xmath59s . both values agree very well with experimental results . we can calculate the expected fidelity when some of the experimental parameters are improved in the near future . if the measurement efficiency is improved from 30% to the current state - of - the - art value of 60% , the measurement duration can be reduced by half while maintaining the quasi - parity measurement infidelities@xcite . the instrument and fpga latencies incurred in the experiment can also be reduced by 100 ns , in the latest hardware setup and fpga logic design in operation while this article was being prepared . the measurement duration and control latency reduction can shorten the correction step length to 1 @xmath59s , which can improve the steady state fidelity to 66% . furthermore , if the coherence times are also improved to the state - of - the - art values in the hundred of microseconds for superconducting qubits , both dd and mb fidelities in the `` fixed time '' protocol can be above 85% , limited by the @xmath11 mismatch ( assumed to be 10% , as in the current experiment ) . the prospects of both dd and mb schemes are summarized in tab . [ tab : fidelity_prospects ] l | l p2 cm p2 cm & + & dd & mb + experiment & current parameters & 76% & 57% + & @xmath153 , latency@xmath154 ns & 76% & 66% + & @xmath100,@xmath155 @xmath59s & 86% & 86% + [ sec : nfp_simulation ] the markov chain model introduced in sec . [ sec : supp model ] can be extended to simulate nfp ( nested feedback protocol ) . we can construct a `` nested '' markov chain ( fig . [ fig : nfp_markov ] ) for each boost attempt of nfp . at the outer - most level , there are two nodes . one node denotes the trajectories that have just been heralded as successful ; the other node denotes the trajectories that require at least another boost attempt . building on the vector description established in the previous section , we represent the heralding of trajectories before a boost attempt by a diagonal matrix @xmath156 , @xmath157 after stabilization of some pre - determined duration , if the qubits are ( on average ) in state @xmath158 , then the average state of the qubits that are heralded is then given by , @xmath159 where @xmath160 is the @xmath161 norm of the vector ( the sum of the entries in the vector ) . the normalization is required since heralding selects only a subset of the trajectories , i.e. , the matrix , @xmath156 , does not preserve the norm of @xmath135 . each entry on the diagonal of @xmath156 gives the fraction of trajectories that get heralded ( selected ) from a particular state . their values depend on the particular heralding thresholds used , @xmath115 and @xmath116 ( supp . [ sec : thresholds ] ) . this matrix can be determined phenomenologically by the post - selection experiment shown in fig . 3 ( main text ) . from the experiment , we can find the average state without any conditioning ( no heralding ) , @xmath162 and the average state of the heralded trajectories , @xmath163 . given the success probability @xmath164 of using the thresholds ( the white dashed contour line of fig . 3b , d in the main text ) , the diagonal elements can be calculated as , @xmath165 for the specific heralding thresholds used in the experiment ( represented by the black squares in fig . 3b , d in the main text ) , @xmath166 and @xmath167 are explicitly given by , @xmath168 ideally for @xmath156 , only @xmath169 should be non - zero . but in practice , since we can not distinguish @xmath46 and @xmath47 , @xmath170 is comparable to @xmath169 . furthermore , for both dd and mb , @xmath171 and @xmath172 are also non - neglibile . in dd , this is predominantly due to the lack of separation between the even and odd measurement outcomes as discussed in supp . sec . [ sec : thresholds ] . in mb , the qubits can jump during the delay between the completion of the quasi - parity measurement and the end of a correction step due to @xmath100 events . thus for mb , trajectories that are heralded by very stringent thresholds still have a non - zero probability of being in the even parity states . given the heralding matrices @xmath166 and @xmath167 , we can now calculate the average state of the trajectories that are not heralded and thus require a boost attempt as @xmath173 where we introduce the lowercase @xmath174 as the unnormalized population distribution vector . @xmath160 denotes the @xmath161 norm of the numerator . after this boost attempt , the qubits are in state @xmath175 , @xmath176 where @xmath145 is the stochastic transition matrix that models the stabilization during a boosting attempt . in supp . [ sec : supp model ] , we have already found @xmath145 for mb . by the same method , we can also derive the effective transition matrix of a boost attempt for dd . the calculation of @xmath145 for dd is an approximation : due to the continuous cavity drives , the state of the qubits at the beginning of a boost attempt is entangled with a qubit - state dependent cavity state , which we approximate unconditionally by the average steady - state cavity state in dd . nonetheless , as we shall show , the model still produces a quantitative behavior that agrees very well with the experimental results . from the above equations , it is easy to show that the average qubits state of the heralded trajectories after @xmath39 boost attempts is given by @xmath177 in the case of a pre - stabilization of sufficient number of correction steps ( or duration ) , @xmath158 is given by the steady - state , @xmath149 , introduced in supp . [ sec : supp model ] . for each of the @xmath39 boost attempts , the first entry in the vector @xmath178 , i.e. , @xmath179 , gives the fidelity to @xmath180 . the @xmath161 norm of the numerator in the expression for @xmath178 gives the percentage of trajectories that have completed @xmath39 boost attempts . summing the percentages over all values of @xmath39 from 0 to the maximum limit gives the overall success probability . [ fig : nfp_sim_results ] , we show the results of the simulation using the model described here and compare them to the experimental results presented in fig . 4 of the main text . furthermore , from the simulation , we find that with the realistic system parameter improvement as specified in supp . [ sec : supp model ] , the fidelity of heralded trajectories can be 90% and 95% for dd and mb , respectively , with order unity success probability . [ sec : ac stark shift ] the quasi - parity measurement induces a deterministic phase shift between the two qubits due to measurement - induced ac stark shift @xcite . this is evidenced by examining the phase of the bell state created in the measurement optimization experiment described in supp . [ sec : msmt time ] , in which we varied the measurement duration . as the measurement duration is increased , the bell angle of the final bell state changes linearly ( fig . [ fig : measurement_dephasing]a ) . in order for mb to work , we need to account for the deterministic phase shift induced by the measurement . this correction is accomplished by a `` _ _ z _ _ '' rotation on bob before the unitary @xmath36 . [ fig : measurement_dephasing]b gives one example of a sequence trajectory to illustrate how the correction works . with no loss of generality ( and the reason will become clear soon in the discussion that follows ) , we construct @xmath181 and @xmath182 ( where @xmath183 = 0 corresponds to the x axis ) such that @xmath184 is the eigenstate of @xmath36 and applying @xmath35 on the even states results in @xmath185 with 50% probability after the quasi - parity measurement . suppose that the qubits are in the ground states , @xmath35 is applied and the subsequent quasi - parity measurement gives @xmath33 . during the quasi - parity measurement , a deterministic phase shift of @xmath186 since the measurement reports odd , the next conditional unitary is @xmath36 . the _ z _ gate before @xmath36 undoes the phase shift and recovers the eigenstate which @xmath36 leaves unchanged . after another parity measurement , the qubits are in the state @xmath187 , with the phase shift added again . consequently , we can see that the mb sequence actually stabilizes @xmath187 state . in practice , for our experiment , the _ z _ rotation for bob is constructed from a composite of _ x _ and _ y _ rotations , such that @xmath188 , and the effective correction angle @xmath189 is swept ( fig . [ fig : measurement_dephasing]c ) to find the optimal value that cancels the deterministic phase shift and thus maximizes the fidelity . furthermore , to make the target state of the stabilization @xmath180 , the rotation axis of the pulses on bob in @xmath36 and @xmath35 is chosen such that @xmath190 . this correction for measurement - induced ac stark shift is also done in the simulation for mb . [ sec : f state ] while we have been treating our two qubits as purely two - level systems , in reality there are higher energy levels , in particular the second excited level is expected to play a non - negligible role in the dynamics . we find that the equilibrium qubit population not in the @xmath7 state was about @xmath191% , and the @xmath192 state was also populated . to investigate whether the decrease of the fidelity in nfp as a function of boost attempts number is due to the role played by the @xmath192-state population , we measured the populations in @xmath193 , @xmath194 , @xmath195 , @xmath196 , @xmath197 after a given number of boost attempts in dd . the population in @xmath193 was measured by applying a @xmath198 pulse on alice s _ e - f _ transition , then a @xmath198 pulse on its _ e - g _ transition , followed by a measurement of the population in @xmath7 . the other @xmath192-state populations are similarly obtained . the sum of these 5 populations gives the total @xmath192-state population plotted in fig . [ fig : f_pop ] , which indeed increases as a function of boost attempt number . it is therefore plausible that the decrease in fidelity with respect to number of boost attempts is due to the system being trapped in other levels outside the correction space . this problem , which is common in similar atomic physics experiments , could be addressed by additional drives . 7ifxundefined [ 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 * * , ( ) link:\doibase 10.1103/physreva.82.012329 [ * * , ( ) ] @noop ( ) link:\doibase 10.1103/physreva.76.012325 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.114.200501 [ * * , ( ) ] link:\doibase 10.1103/physreva.88.023849 [ * * , ( ) ] http://dx.doi.org/10.1126/science.1226897 [ * * , ( ) ] fig . s.1 . setup of the experiment . the qubit - cavity system was placed at the base stage of a dilution refrigerator ( oxford triton200 ) below 20 mk . on the input side : two fpgas ( innovative integration x6 - 1000 m ) generated the pulse envelopes in i and q quadratures to modulate the qubit drives at @xmath199 and @xmath200 frequencies ( vaunix labbrick lms-802 ) for alice and bob , respectively . the i / q modulations were output by the fpgas with one and two single - sideband modulations in mb and dd ( so that both zero - photon and n - photon qubit frequencies were addressed ) , respectively . the microwave frequency drives and i / q modulation were mixed by iq mixers . two agilent n5183 microwave generators produced the cavity drives at @xmath4 and @xmath3 respectively . the cavity drives were also pulsed by the fpgas . on the output side : the transmitted signal through the cavity was directed by two circulators to the jpc for nearly quantum - limited amplification . it was further amplified at the 3 k stage by a cryogenic hemt amplifier . after additional room temperature amplification , the signal was split into two interferometric setups for readouts at the @xmath3 and @xmath4 frequencies , respectively . in each of the interferometers , the signal arriving from the fridge was mixed with a local oscillator set @xmath201 mhz away to produce a down - converted signal at 50 mhz . a copy of the cavity drive that did not go through the fridge was also down - converted in the same manner to produce a reference . finally , the two signals with their respective references were sent to the analog - to - digital - converters on the fpga boards for digitization and further demodulated inside the fpgas . the two fpgas jointly estimate the qubits state by communicating their results with each other . along the input and output lines , attenuators , low pass filters and homemade eccosorb filters were placed at various stages to protect the qubits from thermal noise and undesired microwave and optical frequency radiation . moreover , the cavity and jpc were shielded from stray magnetic fields by aluminum and cryogenic @xmath59-metal ( amumetal a4k ) shields . experimental calibration of measurement strength and duration for mb . fidelity to the state @xmath202 after preparing the qubits in a maximally superposed state followed by a quasi - parity measurement with post - selection . the fidelity is plotted as a function of measurement duration and shown for a set of measurement strengths . the optimal values chosen for the experiment are 660 ns and @xmath110 . ] ( a ) the cavity outputs at @xmath3 and @xmath4 are integrated to obtain measurement outcomes @xmath77 and @xmath78 , respectively . ( b),(c ) , ( d ) , ( e ) histograms of the measurement outcomes @xmath77 and @xmath78 for dd and mb , respectively . the leftmost dashed line ( labeled @xmath115 and @xmath116 ) indicates a particular choice of threshold for heralding the measurement outcomes . counts to the left of both thresholds are declared success . the right dashed line ( @xmath203 and @xmath204 ) indicates the threshold for quasi - parity measurement in mb . ] markov model of a correction step in mb . transition between any two nodes is possible and is represented as a directional edge . the 16 possible transitions make up the 16 matrix elements in the stochastic transition matrix , @xmath145 , corresponding to a correction step in mb . numbers next to the edges represent the elements of @xmath145 calculated for the current experiment parameters by a master equation simulation . ] markov model of the nfp ( nested feedback protocol ) for real - time heralding . this model is an extension of the model introduced for mb . during a boost attempt , transitions occur inside the `` boosting '' node for further stabilization . the edge leading from the `` boosting '' to the `` heralded '' node represents the real - time heralding that selects some fraction of trajectories by the threshold - dependent matrix @xmath156 ( see text ) at the end of a boost attempt . those that are not selected ( @xmath205 ) enter into another boost attempt . ] simulation of real - time heralding by nfp . the left ( right ) panel presents the results for dd ( mb ) . ( a ) , ( c ) cumulative success probability of having completed at most a given number of boost attempts before tomography for dd and mb , respectively . green curve shows the experimental result as presented in the main text . blue curve is the simulation result using the same experimental parameters . ( b ) , ( d ) fidelity to @xmath46 for dd and mb , respectively . cyan dashed line denotes the unconditioned steady state fidelity obtained in the experiment . green squares ( blue circles ) show the corresponding fidelity as a function of the number of boost attempts during nfp obtained in the experiment ( simulation ) . ] correcting for deterministic measurement - induced ac stark shift ( a ) deterministic phase shift induced by measurement as a function of measurement duration . ( b ) one example of a sequence trajectory illustrating the phase correction @xmath206 at work . see detailed explanation in accompanying text . ( c ) fidelity of the steady state to the target bell state as a function of the correction angle of the effective _ z _ gate on bob . ]

clusters of galaxies are usually located at the knots of the filamentary structures in the universe . they are the most massive bound systems to trace the large - scale structure @xcite . statistical studies of clusters provide very powerful constraint on the cosmological parameters ( see a review in * ? ? ? * ) by using , e.g. , cluster mass function @xcite and gas fraction in massive clusters @xcite . clusters are also important laboratories to investigate the evolution of galaxies in dense environment @xcite and act as natural telescope to study lensed high - redshift faint background galaxies @xcite . as a large number of clusters have been detected @xcite , the distributions of galaxy clusters are used to detect the baryon acoustic oscillation of the universe @xcite . correlation of background objects with a large sample of clusters shows the effects of weak lensing and spectra line absorption @xcite . galaxy clusters have been found from single - band optical imaging data ( e.g. , * ? ? ? * ; * ? ? ? * ) , multicolor photometric data ( e.g. , * ? ? ? * ; * ? ? ? * ) and spectroscopic redshift surveys ( e.g. , * ? ? ? * ; * ? ? ? in addition , clusters have also been found from x - ray surveys ( e.g. , * ? ? ? * ; * ? ? ? * ) . at millimeter wavelength , the sunyaev zeldovich ( sz ) effect has recently used to find clusters , which is insensitive to cluster redshift @xcite . hundreds of sz clusters have been identified @xcite . the sloan digital sky survey ( sdss ; * ? ? ? * ) offers an opportunity to produce the largest and most complete cluster sample . it provides photometry in five broad bands ( @xmath6 , @xmath7 , @xmath8 , @xmath9 , and @xmath10 ) covering 14,000 deg@xmath11 and the follow - up spectroscopic observations . the photometric data reach a limit of @xmath12 @xcite with the star galaxy separation reliable to a limit of @xmath13 @xcite . the spectroscopic survey observes galaxies with an extinction - corrected petrosian magnitude of @xmath14 for the main galaxy sample @xcite and @xmath15 for the luminous red galaxy ( lrg ) sample @xcite . galaxy clusters or groups have been found by using the sdss spectroscopic data ( e.g. , * ? ? ? * ; * ? ? ? * ) and the photometric data ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . for large samples of galaxy clusters , the determination of optical richness still has a large uncertainty because of difficulties in the discrimination of cluster member galaxies from background galaxies using photometric data . this causes a large scatter when the optical richness is used to represent the cluster mass for the constraint of cosmological parameters @xcite . in this paper , we improve the method of @xcite using photometric redshifts to identify a large sample of galaxy clusters up to @xmath16 . in section 2 , we identify clusters using a new cluster detection algorithm and determine a cluster richness that is closely related to cluster mass . in section 3 , we compare our sample with the previous cluster samples from the sdss , and correlate the cluster richness with x - ray and sz measurements . in section 4 , we study the brightest cluster galaxies ( bcgs ) and cross - identify them with the sdss lrgs . a summary is presented in section 5 . throughout this paper , we assume a @xmath17cdm cosmology , taking @xmath18100 @xmath19 @xmath20 @xmath21 , with @xmath22 , @xmath23 and @xmath24 . @xcite identified 39,668 galaxy clusters from the sdss dr6 by the discrimination of member galaxies of clusters using photometric redshifts ( hereafter photo-@xmath10s ) of galaxies . @xcite improved the method and successfully identified the high - redshift clusters from the deep fields of the canada france hawaii telescope wide ( cfht ) survey , the chft deep survey , the cosmic evolution survey , and the _ spitzer _ wide - area infrared extragalactic survey . here , we follow and improve the algorithm to identify clusters from sdss - iii . we first use the sdss data to determine scaling relations for cluster mass and radius which will be used in the cluster detection . the galaxy data of 14,000 deg@xmath11 are downloaded from the sdss - iii database . the photo-@xmath10s , the @xmath25-corrections , and the absolute magnitudes are obtained from the table _ photoz _ in which photo-@xmath10s are estimated based on the method of @xcite . we remove objects with deblending problems and saturated objects using the flags20 ) = 0 and ( flags & 0@xmath2680000 ) = 0 and ( ( flags & 0@xmath26400000000000 ) = 0 or psfmagerr@xmath27 ) and ( ( flags & 0@xmath2640000 ) = 0 ) . ] . as shown in figure [ photozg ] , the uncertainties of photo-@xmath10s are @xmath280.0250.030 in the redshift range @xmath29 . they become larger at higher redshifts . in the following analysis , we assume that the uncertainty of photo-@xmath10 , @xmath30 , increases with redshift in the form of @xmath31 for all galaxies . here , we discard those galaxies with a large photo-@xmath10 error @xmath32 , i.e. , about @xmath33 , which suffer bad photometry or contamination of stars . this procedure removes 20% of objects , most of which are faint objects ( @xmath34 ) with large photometric errors . bcgs are the most luminous members of clusters . generally , the bcgs are elliptical galaxies and have smaller photo-@xmath10 error than other member galaxies . proper selection of bcgs can be helpful to identify clusters and estimate the cluster parameters . to get right bcgs , we select those galaxies as bcg candidates which have a photo-@xmath10 error @xmath35 and a galaxy ellipticity in the @xmath8 band less than 0.7 . + the cluster mass and radius are two fundamental parameters for clusters . the widely used are @xmath36 , the radius within which the mean density of a cluster is 200 times of the critical density of the universe , and @xmath37 , the cluster mass within @xmath36 . we use the sdss data of known clusters to get the scaling relations for @xmath36 and @xmath37 . here , we take the clusters whose masses or radii have been estimated by x - ray or weak - lensing methods as compiled by @xcite . for each cluster , we calculate the total luminosities of cluster member candidates in the sdss @xmath8-band within a radius of 1 mpc from its bcg by summing luminosities of member galaxy candidates brighter than @xmath38 but fainter than the bcg within a photo-@xmath10 gap of @xmath39 , with a local background subtraction . here , @xmath40 is the evolution - corrected absolute magnitude in the @xmath8 band , @xmath41 , where we adopt a passive evolution of @xmath42 @xcite . to estimate local background , we follow the method analogous to @xcite . for each cluster , we divide the annuals between 2 and 4 mpc from its bcg into 48 sections with equal area . within the same magnitude and photo-@xmath10 range , we calculate the total luminosity in each sector and estimate the mean value and its root mean square . the regions with luminosities larger than 3@xmath43 are discarded and the mean is recalculated to be the local background . we measure the total luminosity of cluster member candidates in units of @xmath44 , here @xmath44 is the evolved characteristic luminosity of galaxies in the @xmath8 band , defined as @xmath45 @xcite . in figure [ r200m200 ] , we show the correlation between cluster radius @xmath36 and total luminosity within a radius of 1 mpc , @xmath46 . the best fit gives @xmath47 where @xmath36 is in units of mpc and @xmath46 is in units of @xmath44 . similarly , we also get the total @xmath8-band luminosity within the radius of @xmath36 , @xmath48 , in units of @xmath44 , with a background subtraction . in this paper , we define the cluster richness as @xmath49 . we find that cluster mass @xmath37 and cluster richness are closely correlated as ( lower panel of figure [ r200m200 ] ) @xmath50 where @xmath37 is in units of @xmath51 . using these scaling relations , we can estimate the cluster radius and mass from observables of @xmath46 and @xmath48 or @xmath52 . we have shown in our previous work that the measurements of the total luminosity or richness are robust for adopting different widths of the photo-@xmath10 gap @xcite . in the following , we extrapolate these relations to lower richness for cluster identification . based on the estimated richness , we identify clusters with a richness threshold of @xmath53 . rrrrrccrrl[t ] whl j000000.6@xmath54321233 & 0.00236 & 32.20925 & 0.1274 & @xmath55 & 14.92 & 1.72 & 70.63 & 24 & abell + whl j000002.3@xmath54051718 & 0.00957 & 5.28827 & 0.1696 & @xmath55 & 16.20 & 0.94 & 17.48 & 9 & + whl j000003.3@xmath54311354 & 0.01377 & 31.23175 & 0.5428 & @xmath55 & 20.17 & 0.87 & 14.27 & 8 & + whl j000003.5@xmath54314708 & 0.01475 & 31.78564 & 0.0932 & @xmath55 & 15.18 & 0.94 & 16.97 & 9 & + whl j000004.7@xmath54022826 & 0.01945 & 2.47386 & 0.4179 & @xmath55 & 19.32 & 0.95 & 13.71 & 10 & + whl j000004.9@xmath56033248 & 0.02024&@xmath57 & 0.5968 & @xmath55 & 20.67 & 1.00 & 19.19 & 11 & + whl j000005.5@xmath54354610 & 0.02303 & 35.76957 & 0.4762 & @xmath55 & 19.59 & 0.91 & 15.58 & 9 & + whl j000006.0@xmath54152548 & 0.02482 & 15.42990 & 0.1656 & @xmath55 & 16.60 & 1.13 & 23.53 & 19 & maxbcg , whl09,gmbcg + whl j000006.3@xmath54221220 & 0.02643 & 22.20558 & 0.3985 & @xmath55 & 19.36 & 0.84 & 12.73 & 11 & + whl j000006.6@xmath54100648 & 0.02755 & 10.11333 & 0.3676 & @xmath55 & 19.07 & 0.93 & 16.73 & 13 & + whl j000006.6@xmath54315235 & 0.02762 & 31.87626 & 0.2134 & @xmath55 & 17.11 & 1.18 & 28.58 & 15 & + whl j000006.6@xmath54292129 & 0.02765 & 29.35813 & 0.2489 & @xmath55 & 18.13 & 0.92 & 15.11 & 16 & + whl j000007.1@xmath56092910 & 0.02957&@xmath58 & 0.3332 & @xmath55 & 19.11 & 0.80 & 12.81 & 13 & amf + whl j000007.6@xmath54155003 & 0.03177 & 15.83424 & 0.1436 & 0.1528 & 15.99 & 1.17 & 34.11 & 27 & abell , maxbcg , whl09,gmbcg , amf + whl j000007.7@xmath54185245 & 0.03208 & 18.87909 & 0.4347 & @xmath55 & 18.95 & 1.03 & 20.47 & 12 & + in @xcite , we identified a cluster if more than eight member galaxies of @xmath59 ( without evolution correction and different from @xmath60 in this paper ) are found within a radius of 0.5 mpc and a photo-@xmath10 gap of @xmath61 . the richness was estimated to be the number of galaxies within a radius of 1 mpc which may be systematically smaller than the true radius for rich clusters but larger than that for poor clusters . note also that the radius of 0.5 mpc for cluster identification is not the same as the radius of 1 mpc for richness estimate . with the tight scaling relations shown in section [ scaling ] , we can improve the method of @xcite and @xcite to identify clusters from the sdss - iii . our new algorithm includes following steps . \1 . for each galaxy at a given photometric redshift , @xmath10 , we count the number of luminous member galaxies of @xmath38 within a radius of 0.5 mpc and a photo-@xmath10 gap of @xmath61 , @xmath62 . to get cluster candidates , we apply the friend - of - friend algorithm @xcite to the luminous galaxies using a linking length of 0.5 mpc in the transverse direction and a photo-@xmath10 difference of @xmath63 . the linked galaxy with the maximum @xmath62 is taken as the temporary center of a cluster candidate . if two or more galaxies have the same maximum number , the brightest one is taken as the temporary central cluster galaxy . then , we obtain a temporary list of cluster candidates . \3 . for each cluster candidate at @xmath10 , we assume that the galaxies of @xmath38 within a radius of 1 mpc from the temporary central galaxy and the photo-@xmath10 gap of @xmath61 are member galaxies . the cluster redshift @xmath64 is then defined to be the median value of the photometric redshifts of the recognized `` members '' . we recognize the bcg of a cluster candidate as the brightest galaxy from the bcg candidate sample within a radius of 0.5 mpc from the temporary center and the photo-@xmath10 gap of @xmath61 . as a new development in this paper , we use the bcg as the center of a cluster and calculate @xmath46 from the estimated redshift . subsequently , we get @xmath36 using equation ( [ r200 ] ) and measure @xmath52 from @xmath48 within the estimated @xmath36 for each cluster candidate . note that the local background has been subtracted already here . we define a galaxy cluster if @xmath65 , which corresponds to @xmath66 using equation ( [ m200 ] ) . note that a cluster candidate at low richness may be contaminated by a small number of bright field galaxies with their photo-@xmath10s seriously overestimated . to avoid such contamination , we require the number of galaxies within a radius of @xmath36 and a photo-@xmath10 gap of @xmath61 , @xmath67 . this assures that the cluster sample above the detection criteria has a high completeness and a low false detection rate ( see sections [ complete ] and [ false ] ) . after we get the cluster candidate list , we merge possible repeatedly identified clusters which may survive in the previous procedure by using the friend - of - friend algorithm . if two cluster candidates have photo-@xmath10 difference less than @xmath63 and a projection separation less than @xmath36 , we merge the cluster candidates as one cluster , and the poorer one is then removed from the list . finally , we visually inspect the color images on the sdss web site for all cluster candidates . about 5000 ( 3.6% ) cluster candidates are obvious contaminations from bad photometries and are removed . by this procedure , we discovered 68 lensing systems from the color images @xcite . after above procedures , we find 132,684 clusters in the redshift of @xmath68 from the sdss - iii data . all clusters are listed in table [ cat ] . figure [ histcz ] show the redshift distribution of the identified clusters in the sdss - iii which is compared with those of the maxbcg @xcite , whl09 @xcite , gaussian mixture brightest cluster galaxy ( gmbcg ; * ? ? ? * ) and adaptive matched - filter ( amf ; * ? ? ? * ) samples . our new sample has a peak at @xmath69 and the number of our clusters is nearly two times as the previously largest amf sample by @xcite . the clusters of @xmath70 seem to be a complete sample . because of the magnitude limit of the sdss photometric data , the member galaxies of @xmath3 are incomplete for the threshold of @xmath71 , so that clusters of @xmath3 have a biased smaller richness than the real one because of missing members . in the following , we only consider clusters of @xmath70 when we compare our richness with other richness estimates and x - ray measurements . using the sdss spectroscopic data , we can get the spectroscopic redshift of identified clusters and verify the accuracy of their photometric redshifts . the spectroscopic redshift of a cluster is taken to be that of its bcg . we find that the bcgs of 38,116 clusters have spectroscopic redshifts . figure [ phzc ] shows the distribution of the difference between photometric and spectroscopic redshifts , @xmath72 . the standard deviations roughly indicate the accuracy of redshift estimate for clusters . in each panel , we fit the distribution of @xmath73 with a gaussian function . the systematic offset @xmath74 of the fitting is less than @xmath75 , and the standard deviation @xmath76 is less than 0.018 . + we use monte carlo simulation to test the completeness of our sample . mock clusters are simulated with assumptions for their distributions and then added to the real data of the sdss . the cluster detection algorithm is applied to the combined data to find the input mock clusters . the detection rate of mock clusters is considered as the indication of completeness of our detection procedure . first , we generate a population of halos with masses @xmath77 following a mass function of @xcite in a cosmology with @xmath23 and @xmath78 in the redshift range of @xmath79 . then , we relate the cluster mass to cluster richness ( i.e. , total luminosity in units of @xmath44 ) by @xmath80 where @xmath81 is a scaling factor between input richness @xmath82 and output richness @xmath83 of our detection procedure . the discrepancy between @xmath84 and @xmath83 is caused by the uncertainty of galaxy photo-@xmath10 . because the equation ( [ m200 ] ) is obtained for the estimated richness based on the sdss data , we adjust this factor @xmath81 to make output richness consistent with equation ( [ m200 ] ) . we first set @xmath85 for equation ( [ m200s ] ) to get input richness . member galaxies are simulated for each halo within a radius of @xmath36 . the luminosity function and density profile of member galaxies are taken as described in @xcite . here , the luminosities of member galaxies are evolution - corrected , as mentioned in section [ scaling ] . we assume that the uncertainty of photometric redshift of member galaxies follows a gaussian probability function with a standard deviation of @xmath30 , but varies with redshift in the form of @xmath86 . we add the mock clusters into the real sdss data and apply our cluster detection algorithm to measure the output richness . figure [ outrich ] shows the correlation between the input richness and the output richness . the best fit gives @xmath87 therefore , we get the factor @xmath88 for equation ( [ m200s ] ) so that the output richness @xmath83 can be related to cluster mass by equation ( [ m200 ] ) . subsequently , we use equation ( [ m200s ] ) to get a new population of input richness . again , we apply our method to detect the mock clusters . a cluster is detected if the output richness @xmath89 and @xmath67 . the detection rate is given to be the number of detected mock clusters divided by the total number of added mock clusters . figure [ detrate ] shows the detection rate as a function of redshift . the detection rate is nearly 100% for clusters with masses @xmath90 ( richness @xmath91 ) up to redshift of @xmath92 , and @xmath2875% for clusters of @xmath93 ( richness @xmath94 ) up to redshift of @xmath95 . the detection rate drops at higher redshifts , because the member galaxy selection of our algorithm is consistent up to @xmath95 , as mentioned in section [ algorithm ] . in figure [ detrate ] , we also show the detection rate as a function of cluster mass for clusters of @xmath70 . more than 95% of clusters of @xmath96 are detected . the presence of the large - scale structures makes it possible to detect false clusters because of projection effect . we perform a monte carlo simulation with the real sdss data to estimate the false detection rate . first , we discard the member candidates of identified clusters within a radius of @xmath36 and a photo-@xmath10 gap of @xmath97 . second , we shuffle the data following @xcite . our new cluster detection algorithm is applied to the shuffled data . after above procedures , we get the `` false clusters '' which satisfy the criteria of @xmath65 and @xmath67 . the false detection rate is calculated to be the total number of `` false clusters '' identified from the shuffled data divided by the number of clusters identified from the original sdss data . we show that the false detection rate is 6.0% for a cluster richness of 12 , but decreases to less than 1% for a richness of 23 ( see figure [ falrich ] ) . note that we have removed some obvious contaminations from cluster list in the cluster identification ( section [ algorithm ] ) . here , most of the contaminations can not be found when shuffling the cataloged galaxy data . hence , removing the contaminations slightly decreases the false detection rate . we compare our cluster sample in this paper ( hereafter whl12 ) with previous ones , including the classical abell sample from the palomar sky survey @xcite , the maxbcg @xcite , whl09 @xcite , gmbcg @xcite and amf @xcite cluster samples from the sdss . we also compare our clusters with x - ray and sz cluster samples . we get the abell cluster sample @xcite from the nasa / ipac extragalactic database . for the abell clusters without redshift measurements , we take their redshifts from the sdss photo-@xmath10 data . there are 1844 abell clusters with redshifts @xmath98 in the sky coverage of sdss - iii , of which 1688 ( 92% ) clusters are within a projected separation of @xmath36 and redshift difference of @xmath99 ( about 2@xmath76 of the uncertainty of the redshift difference ) from the whl12 clusters . the non - matched abell clusters are relatively poor clusters or projected to neighbors of the large scale structure @xcite . we mark the abell clusters in column 10 of table [ cat ] . + @xcite developed a `` red - sequence cluster finder '' , maxbcg , to detect clusters dominated by red galaxies . they identify clusters with a richness of @xmath100 , where @xmath101 is the number of galaxies brighter than @xmath102 in the @xmath9 band within @xmath36 and @xmath103 of the ridgeline colors . from the sky coverage of the sdss dr4 , they obtained a complete volume - limited sample containing 13,823 clusters in the redshift range of @xmath104 . the maxbcg sample is approximately 85% complete for the clusters with masses @xmath105 . we cross - match the whl12 clusters with the maxbcg clusters in the sky coverage of the sdss dr4 . figure [ hist_dz ] shows the redshift difference ( @xmath106 ) between the whl12 clusters and the maxbcg clusters ( also the whl09 , gmbcg , and amf clusters for discussion below ) within a projected separation of @xmath36 . it shows that a value of @xmath107 is an appropriate selection threshold for cluster matching in redshift difference ( top left panel of figure [ hist_dz ] ) . we find that 10,301 ( 75% ) of 13,823 maxbcg clusters are matched with the whl12 clusters within a separation of @xmath36 and a redshift difference of @xmath107 . figure [ rate_max ] shows the matching rates between the whl12 clusters and the maxbcg clusters as a function of redshift . we find that 70%80% maxbcg clusters are matched with the whl12 clusters in the redshift range of @xmath104 , while only 30%40% of the whl12 clusters are matched with the maxbcg clusters because our sample is more complete for poor clusters . the matching rates increase with cluster richness for both samples . as shown in figure [ rate_max ] , 95% of the maxbcg clusters of @xmath108 are matched with the whl12 clusters , and 85% of the whl12 clusters of @xmath109 are matched with the maxbcg clusters . those non - matched maxbcg clusters mostly have a richness smaller than the criteria of our cluster finding algorithm . a small number of the maxbcg clusters of @xmath108 are missing in our sample . as we have checked , some of them are merged in the friend - of - friend algorithm of our procedure ( steps 2 and 5 ) , and some of them can be found with a larger separation . if we match the clusters within a separation of @xmath110 , @xmath111 of the maxbcg clusters of @xmath108 can be found in our sample . figure [ rich_max ] shows the correlation between the maxbcg richness , @xmath101 , and the richness of this paper , @xmath52 , for the matched clusters . from the sdss dr6 , @xcite identified 39,668 clusters of galaxies in the redshift range @xmath112 . monte carlo simulations show that the sample is complete up to redshift @xmath113 for clusters with a richness @xmath114 . the cluster richness is the number of member candidates of @xmath115 within a radius of 1 mpc and a photo-@xmath10 of @xmath116 after background subtraction . comparing to @xcite , we modify the richness estimate and cluster identification criteria . a cluster is identified in this paper when the richness @xmath53 and the number of member candidates @xmath117 within @xmath36 , which is looser than the criteria of @xmath118 within a radius of 0.5 mpc in @xcite . the cluster sample in this paper is more complete for clusters of low richness . the completeness is 40% for clusters with a mass of @xmath51 ( richness of @xmath2812 ) in @xcite , while it is @xmath2890% in this paper . + matching is applied between two cluster samples in sky coverage of the sdss dr6 . we find that 25,929 ( 65% ) of 39,668 whl09 clusters are matched with the whl12 clusters within a separation of @xmath36 and a redshift difference of @xmath107 . figure [ rate_whl ] shows the matching rate as a function of redshift between the whl12 clusters and whl09 clusters . the matching rates of the whl12 clusters to whl09 clusters decrease from 85% at @xmath119 to 55% at @xmath120 . the matching rates of the whl09 clusters to the whl12 clusters increase from 20% at @xmath119 to 40% at @xmath120 . the matching rates vary as a function of cluster richness within @xmath121 , 90% of the whl09 clusters of @xmath122 are matched with the whl12 clusters , 86% of the whl12 clusters of @xmath109 are matched with the whl09 clusters . if we match the clusters within a separation of @xmath110 , @xmath123 of the whl09 clusters of @xmath122 are matched . the non - matched whl09 clusters are probably merged in the friend - of - friend algorithm of our procedure , or are poorer than the criteria of cluster finding in this work . figure [ rich_whl ] shows a tight correlation between the whl09 richness , @xmath124 , and the richness of this paper , @xmath52 , for the matched clusters . + @xcite presented a gmbcg algorithm , which detects clusters by identifying the red sequence and the bcg feature . from the sdss dr7 , they found 55,424 clusters in the redshift range of @xmath125 . they provide two richnesses , the weighted richness which is the total number of galaxies brighter than @xmath102 within @xmath36 weighted by a factor from the fitting of color distribution , and the scaled richness which is the number of galaxies within @xmath36 and @xmath103 of the ridgeline colors . we adopt the weighted richness as gmbcg richness @xmath126 if available , otherwise the scaled richness . we cross - match the whl12 clusters with the gmbcg clusters in the sky coverage of the sdss dr7 . we find that 25,370 ( 46% ) of 55,424 gmbcg clusters are matched with the whl12 clusters within a separation of @xmath36 and a redshift difference of @xmath107 ( see figure [ rate_gmbcg ] ) . the matching rates of the whl12 clusters to the gmbcg clusters decrease from 80% at @xmath119 to 40% at @xmath120 . the matching rates of the gmbcg clusters to the whl12 clusters are 30%40% within @xmath127 . within the redshift range of @xmath121 , the matching rates increase with cluster richness for both samples . we find that 72% of the gmbcg clusters of @xmath128 are matched with the whl12 clusters , 83% of the whl12 clusters of @xmath109 are matched with the gmbcg clusters . the fraction of missing rich gmbcg clusters is larger than those of the maxbcg and whl09 clusters . this may be due to a broader distribution of redshift difference ( see figure [ hist_dz ] ) . if we match clusters within a separation of @xmath110 and a redshift difference of @xmath129 , 55% of the gmbcg clusters are matched with the whl12 clusters , 85% of the gmbcg rich clusters of @xmath130 are matched ( see dashed lines in figure [ rate_gmbcg ] ) . we note that a few rich gmbcg clusters are still missing . some of them are merged in the friend - of - friend algorithm of our procedure , or have a redshift difference of @xmath131 from the whl12 clusters . most of other missing gmbcg clusters are poorer than the criteria of cluster finding in this work . we note a large scatter between the gmbcg richness @xmath126 and @xmath52 ( see figure [ rich_gmbcg ] ) . + @xcite used a modified amf algorithm to identify clusters using photometric redshifts of galaxies . from the sdss dr6 , 69,173 clusters are identified in the redshift range @xmath132 with a richness of @xmath133 . the richness of the amf clusters is defined as the total luminosity within @xmath36 in units of @xmath44 , the same as that in this work . we cross - match the whl12 clusters with the amf clusters in the sky coverage of the sdss dr6 . within a separation of @xmath36 and a redshift difference of @xmath107 , 27,966 ( 40% ) of 69,173 amf clusters are matched with the whl12 clusters . figure [ rate_amf ] shows the matching rate as a function of redshift . the matching rates of the whl12 clusters to the amf clusters decrease from 80% at @xmath119 to 30% at @xmath92 . the matching rates of the amf clusters to the whl12 clusters are constantly @xmath2830% . the matching rates vary with cluster richness within @xmath121 , 72% of the amf clusters of @xmath134 are matched with the whl12 clusters , 74% of the whl12 clusters of @xmath109 are matched with the amf clusters . we note that many amf clusters have a large redshift difference of @xmath135 ( see figure [ hist_dz ] ) . if we match clusters within a separation of @xmath110 and a redshift difference of @xmath129 , 91% of the amf rich clusters of @xmath134 are matched with the whl12 clusters ( see dashed lines in figure [ rate_amf ] ) . those non - matched amf clusters are missing due to friend - of - friend merging of our algorithm , large redshift difference from the whl12 clusters and large scatter between two richness estimates . figure [ rich_amf ] shows the correlation between the amf richness @xmath136 and the @xmath52 for the matched clusters . + with luminous member galaxies well discriminated in the identification procedure , good correlations between the cluster richness with x - ray measurements are expected @xcite . we get x - ray clusters from the northern rosat all - sky ( noras ) galaxy cluster survey @xcite and the rosat - eso flux limited x - ray ( reflex ) galaxy cluster survey @xcite . the noras sample contains 378 clusters and the reflex contains 447 clusters . survey provides x - ray luminosity for detected clusters . we also get x - ray clusters from the query of the x - ray cluster database ( bax ) , where luminosities for 1028 clusters and temperatures for 210 clusters are available compiled from reference . the _ rosat _ clusters mostly are massive with a luminosity greater than @xmath137 in the 0.12.4 kev band , while the bax sample contains many less massive clusters with a low luminosity less than @xmath137 . previous work shows that the luminosity in both samples is statistically consistent @xcite . here , we combine two sample for the correlation with our cluster richness . for the clusters in both _ rosat _ and bax samples , we adopt the x - ray luminosity from _ rosat _ data . x - ray clusters are matched with the whl12 clusters within a separation of @xmath36 and a redshift difference of @xmath107 . we find 611 matched x - ray clusters with luminosities measured , 84 matched x - ray clusters with temperatures measured . figure [ rich_xray ] shows the correlations between cluster richness with x - ray luminosity and temperature for the matched clusters . the best fits give @xmath138 and @xmath139 where @xmath140 refers to x - ray luminosity in the 0.12.4 kev band in units of @xmath137 , @xmath141 refers to x - ray temperature in units of kev . the tight correlations suggest that the richness we estimate is reasonable and statistically reliable . the slopes of the @xmath142@xmath52 and the @xmath141@xmath52 relations are in agreement with those of @xcite . we measure the scatter in x - ray luminosity to the best - fit relation and get a value of @xmath143 , i.e. , @xmath144 . the scatter is in agreement with that for the correlation of @xmath142 with the maxbcg richness @xcite , but larger than those for the correlations with the improved maxbcg richness @xcite . for a similar comparison , we measure the scatter in cluster mass to the best - fit @xmath37@xmath52 relation given in section [ scaling ] and get @xmath145 , i.e. , @xmath146 . the scatter in mass based on our richness is also in agreement with that based on the maxbcg richness @xcite . the sz effect is the result of high energy electrons interacting with the cosmic microwave background radiation through inverse compton scattering . it has been detected around rich galaxy clusters due to a high temperature of host ionized gas @xcite . the sz signal is characterized by the quantity @xmath147 , where @xmath148 is the angular distance to a cluster at redshift @xmath10 , @xmath149 is the thomson cross - section , @xmath150 is the speed of light , @xmath151 is the electron rest mass , @xmath152 is the integration of the pressure of hot ionized gas over the cluster volume . as the pressure is related to gravitational potential , @xmath153 is expected to be a proxy of cluster mass @xcite . recently , the planck releases an sz cluster sample including 189 clusters @xcite . we find that 71 sz clusters are located at the sky coverage of the sdss - iii . six of them have redshifts less than 0.05 . other 65 clusters are matched with the whl12 clusters , of which 61 have x - ray measurements . figure [ rich_sz ] shows the correlations between the richness and the sz measurements by the planck for the 61 clusters , where @xmath154 is the sz signal measured within a radius of @xmath155 and @xmath156 . the matched clusters mostly have a richness of @xmath157 . clearly , there is a positive correlation between cluster richness and sz measurement . bcgs are luminous and usually located at the centers of clusters . their properties provide the information of galaxy formation in the extreme dense environment . using our sample , we can study the evolution of the bcgs . moreover , we cross - identity the bcgs with the sdss lrg sample . + + the bcgs in rich clusters are formed at redshift @xmath158 and evolved passively ( e.g. , * ? ? ? the properties of bcgs are related to their host clusters @xcite . richer clusters tend to have brighter bcgs . here , we use our sample to address the magnitude evolution of the bcgs with redshift and richness . figure [ bcg1 ] shows the evolution of absolute magnitudes of the bcgs with cluster redshift in the range of @xmath121 for six richness bins and with cluster richness for six redshift bins . we get the mean values and fit the correlation ( solid lines ) for each bin . here , the variations of completeness and false detection rate are considered by a weighting factor in the fitting . the weight is calculated as the percentage of true cluster ( i.e. , @xmath159false detection rate ) divided by completeness for each richness and redshift bin . there are very small changes of the average richnesses within the redshift ranges ( upper panels ) or the average redshifts within the richness ranges ( lower panels ) , indicating that the changes of bcg magnitudes have been well decoupled for redshift and richness . the best fits between absolute magnitude and redshift give @xmath160 clearly , the bcg are brighter in clusters of higher redshifts . as the stellar population in the bcgs was formed at redshift @xmath158 and becomes old with comic time , the bcgs become fainter at lower redshifts . to verify the evolution of bcg magnitude with redshift , using the amf clusters @xcite and the brightest one of three bcg candidates for each cluster , we get @xmath161 which is roughly consistent with the results derived from our cluster sample . the bcg magnitude evolution is in agreement with that of @xmath44 found by @xcite . in addition , we find that bcgs are brighter in richer clusters , consistent with conclusions in previous works @xcite . from the fitting , the bcgs have an average absolute magnitude of @xmath162 for cluster richness of @xmath16315 at redshift @xmath164 , but a brighter magnitude of @xmath165 for cluster richness of @xmath166 . this tendency can be clearly shown from the correlation between bcg absolute magnitude with cluster richness for six redshift bins ( the lower panel of figure [ bcg1 ] ) . the evolution of the bcgs magnitudes with redshift and the dependence of richness can be formulated together as , @xmath167 the sdss lrgs are selected to produce a volume - limited sample of massive galaxies to redshift of 0.5 . they are efficiently selected by color cuts based on passive evolution model at redshifts @xmath168 @xcite . generally , lrgs are more massive and located at denser environment than general galaxies . they are most likely bcgs or brightest group galaxies . bcgs are luminous and mostly have red colors . they are likely to be lrgs . however , this may not always be true . @xcite showed that about 27% bcgs in x - ray bright clusters have optical emission lines . they may have blue colors due to recent star formation @xcite . therefore , many bcgs are not lrgs . first , we cross - identify the clusters in our sample with the sdss lrgs to address how many lrgs are member galaxies . we get 112,191 lrgs of @xmath168 from the sdss spectroscopic data . cross - matching between lrgs with bcgs in our sample , we find that 28,336 ( 25% ) lrgs are bcgs and 40,039 ( 36% ) lrgs are located within a projected separation of @xmath36 and a redshift difference of 0.05 . this suggests that most lrgs are located not in our clusters but in an environment poorer than our clusters . on the other hand , we estimate the fraction of the bcgs as the sdss lrgs . follow @xcite , we match the colors of the bcgs with the lrg selection criteria . figure [ sdsslrg ] shows the color magnitude and color color diagrams for the bcgs in our clusters . the color @xmath169 is defined as @xmath170.\ ] ] if we only consider the color cuts and do not limit the petrosian magnitude @xmath171 , the bcgs can be classified as lrgs for those with the color @xmath169 above the line in the left panel of figure [ sdsslrg ] and the colors , @xmath172 and @xmath173 , between the lines in the right panel . there are 83,740 ( 66% ) of 126,041 bcgs in the redshift range @xmath168 satisfying the color cuts of the sdss lrg selection criteria . when the petrosian magnitude limits of @xmath174 for cut i and @xmath5 for cut ii are considered , we find that 63% of bcgs of @xmath5 are lrgs . we identify 132,684 clusters of galaxies in the redshift range @xmath68 using photometric redshifts of galaxies from the sdss - iii . the clusters are recognized for those with a richness @xmath53 and a number of member galaxies candidates @xmath117 within @xmath36 . monte carlo simulations show that the false detection rate is less than 6% for the whole sample . the cluster detection rate is more than 95% for clusters of @xmath96 in the redshift range of @xmath2 . we cross - match our cluster sample with previous cluster samples . the matching rates are 5070% for previously known clusters but depend on cluster richness . clusters in our sample match more than 90% of previously known rich clusters . comparing with x - ray and the planck data , we find that the determined cluster richness is related to the x - ray luminosity , temperature and sz measurements . the richer clusters have brighter bcgs . the bcgs are brighter in higher redshift clusters . cross - matching the bcgs with the sdss lrgs , we find that 25% lrgs are bcgs of our clusters , 36% lrgs are member galaxies , and 63% of the bcgs of @xmath5 in our clusters satisfy the sdss lrg selection criteria . we thank the referee for valuable comments that helped to improve the paper . the authors are supported by the national natural science foundation ( nnsf ) of china ( 10833003 and 11103032 ) . funding for sdss - iii has been provided by thealfred p. sloan foundation , the participating institutions , the national science foundation , and theusdepartment of energy . the sdss - iii web site is http://www.sdss3.org/. sdss - iii is managed by the astrophysical research consortium for the participating institutions of the sdss - iii collaboration including the university of arizona , the brazilian participation group , brookhaven national laboratory , university of cambridge , university of florida , the french participation group , the german participation group , the instituto de astrofisica de canarias , the michigan state / notre dame / jina participation group , johns hopkins university , lawrence berkeley national laboratory , max planck institute for astrophysics , newmexico state university , new york university , ohio stateuniversity , pennsylvania state university , university of portsmouth , princeton university , the spanish participation group , university of tokyo , university of utah , vanderbilt university , university of virginia , university of washington , and yale university . , r. , gunn , j. e. , ivezi , z. , knapp , g. r. , & kent , s. 2001 , in asp conf . 238 , astronomical data analysis software and systems x , ed . f. r. harnden , jr . , f. a. primini , & h. e. payne(san francisco , ca : asp ) , 269

the problem of violations of the axioms of probability go back to the early 60 s . @xcite published a work that influenced modern psychology by showing that humans violate the laws of probability theory when making decisions under risk . the principle that humans were constantly violating is defined by _ the sure thing principle_. it is a concept widely used in game theory and was originally introduced by @xcite . this principle is fundamental in bayesian probability theory and states that if one prefers action @xmath0 over @xmath1 under state of the world @xmath2 , and if one also prefers @xmath0 over @xmath1 under the complementary state of the world @xmath2 , then one should always prefer action @xmath0 over @xmath1 even when the state of the world is unspecified . cognitive psychologists a. tversky and d. khamenman also explored more situations where classical probability theory could not be accommodated in human decisions . in their pioneering work , @xcite realised that the beliefs expressed by humans could not follow the rules of boolean logic or classical probability theory , because humans can not process large amounts of data in order to make estimations or judgements . consequently , the inferences performed are based on limited data coupled with several heuristics , leading to a violation on one of the most important laws in bayesian theory : the law of total probability . one of the key differences between classical and quantum theories is the way how information is processed . according to classical decision making , a person changes beliefs at each moment in time , but it can only be in one precise state with respect to some judgement . so , at each moment , a person is favouring a specific belief . the process of human inference deterministically either jumps between definite states or stays in a single definite state across time @xcite . most computer science , cognitive and decision systems are modelled according to this single path trajectory principle . figure [ fig : definite_states ] illustrates this idea . in quantum information processing , on the other hand , information ( and consequently beliefs ) are modelled via wave functions and therefore they can not be in definite states . instead , they are in an indefinite quantum state called the @xmath3 state . that is , all beliefs are occurring on the human mind at the same time . according to cognitive scientists , this effect is responsible for making people experience uncertainties , ambiguities or even confusion before making a decision . at each moment , one belief can be more favoured than another , but all beliefs are available at the same time . in this sense , quantum theory enables the modelling of the cognitive system as it was a wave moving across time over a state space until a final decision is made . from this superposed state , uncertainty can produce different waves coming from opposite directions that can crash into each other , causing an interference distribution . this phenomena can never be obtained in a classical setting . figure [ fig : interference ] exemplifies this . when the final decision is made , then there is no more uncertainty . the wave collapses into a definite state . thus , quantum information processing deals with both definite and indefinite states @xcite . @xcite were one of the first researchers to test the veracity of savage s principle under human cognition in a gambling game . in their experiment , participants were asked at each stage to make the decision of whether or not to play a gamble that has an equal chance of winning $ 200 or losing $ 100 . figure 3 illustrates the experiment . three conditions were verified : 1 . participants were informed if they had won the first gamble ; 2 . participants were informed if they had lost the first gamble ; 3 . participants did not know the outcome of the first gamble ; [ fig : two_stage_gambling ] the two - stage gambling game was one of the first experiments used in order to determine if the sure thing principle would be verified even with people that did not know about the existence of this principle . the results obtained in @xcite experiment showed that this principle is constantly being violated and consequently humans do not perform inferences according to the laws of probability theory and boolean logic . the overall results revealed that participants who knew that they won the first gamble , decided to play again . participants who knew that they lost the first gamble , also decided to play again . through savage s sure thing principle , it was expected that the participants would choose to play again , even if they did not know the outcome of the first gamble . however , the results obtained revealed something different . if the participants did not know the outcome of the first gamble , then many of them decided not to play the second one . several researchers replicated this experiment . the overall results are specified in table [ tab : game_results ] . why did the findings reported in table [ tab : game_results ] generate so much controversy in the scientific community ? because , the data observed is not in accordance with the classical law of total probability . in tversky and shafir s experiment @xcite , the probability of a participant playing the second gamble , given that the outcome of the first gamble is unknown , @xmath4 , can be computed through the law of total probability : @xmath5 in equation [ eq : law_total_prob ] , @xmath6 corresponds to the probability of a player winning the first gamble , given that ( s)he participated on the game in the first place . @xmath7 is the probability of playing the second gamble , given that it is known that the player won the first one . @xmath8 corresponds to the probability of losing the first gamble , given that the participant decided to play the game in the first place . and finally , @xmath9 is the probability of a participant playing the second gamble , given that it is known that ( s)he lost the first one . following the law of total probability in equation [ eq : law_total_prob ] , the probability of playing the second gamble , given that the player did not know the outcome of the first one , should be between the following values @xcite : @xmath10 the findings reported by @xcite , however , revealed a different relation . equation [ eq : relation_observed ] demonstrates that this relation is violating one of the most fundamental laws of bayesian probability theory : @xmath11 @xcite explained these findings in the following way : when the participants knew that they won , then they had extra house money to play with and decided to play the second round . if the participants knew that they lost , then they chose to play again with the hope of recovering the lost money . but , when the participants did not know if they had won or lost the first gamble , then these thoughts , for some reason , did not emerge in their minds and consequently they decided not to play the second gamble . other works in the literature also replicated this two - stage gambling experiment @xcite , also reporting similar results to @xcite . their results are summarised in table [ tab : game_results ] . there have been different works in the literature trying to explain and model this phenomena @xcite . although the models in the literature diverge , they all agree in one thing : one can not use classical probability theory to model this phenomena , since the most important rules are being violated . this two stage gambling game experiment was one of the most important works that motivated the use of different theories outside of classical bayesian theory and boolean logic , more specifically the usage of quantum probability theory . recent findings in the cognitive psychology literature revealed that humans are constantly violating the law of total probability when making decisions under risk @xcite . these researchers also showed that quantum probability theory enables the development of decision models that are able to simulate human decisions . given that most of the systems that are used nowadays are based on bayesian probability theory , is it possible to achieve better inference mechanisms in these systems using quantum probability theory ? for instance , many medical diagnosing systems are based in classical probabilistic graphical models such as bayesian networks . can one achieve better performances in diagnosing patients using quantum probability ? generally speaking , a bayesian network is a probabilistic graphical model that represents a set of random variables and their conditional dependencies via a directed acyclic graph . there are two main works in the literature that have contributed to the development and understanding of quantum bayesian networks . one belongs to @xcite and the other to @xcite . in the work of @xcite , it is argued that any classical bayesian network can be extended to a quantum one by replacing real probabilities with quantum complex amplitudes . this means that the factorisation should be performed in the same way as in a classical bayesian network . one big problem with tucci s work is concerned with the inexistence of any methods to set the phase parameters . the author states that , one could have infinite quantum bayesian networks representing the same classical bayesian network depending on the values that one chooses to set the parameters . this requires that one knows _ a priori _ which parameters would lead to the desired solution for each node queried in the network ( which we never know ) . in the work of @xcite , the authors argue that , in order to develop a quantum bayesian network , it is required a quantum version of probability distributions , quantum marginal probabilities and quantum conditional probabilities . the proposed model fails to provide any advantage relatively to the classical models , because it can not take into account interference effects between unobserved random variables . in the end , both models provide no advantages in modelling decision making problems that try to predict decisions that violate the laws of total probability . [ fig : qbn_example1 ] in this paper , the core of the proposed bayesian network is based on the psychological findings uncovered in the works of @xcite and on quantum information processing . these authors show that , before taking a decision , human thoughts are seen as superposed waves that can interfere with each other , influencing the final decision . in bayesian networks , nodes can either be query variables , evidences or simply unknown . given that we do not observe the unknown nodes of a bayesian network , since we do not know for sure what values they can take , then what would happen to the inference process if these nodes are put in a representation of a quantum superposition and interfere with each other ( figure 4 ) ? can a better inference process be achieved ? these are the main research questions that this paper aims at answering . so far , to the best of our knowledge , there is no previous work in the computer science community that attempts to map these psychological findings into computer science decision making systems , such as bayesian networks . in order to validate our hypothesis , we performed experiments with well known classical bayesian networks from the literature . we first create a quantum bayesian network that can accommodate the paradoxical findings in the two - stage gambling game . we then generalise our quantum bayesian network in order to deal with larger and more complex datasets that are used in the literature : the well known burglar / alarm bayesian network from @xcite and the lung cancer bayesian network from @xcite . before describing the proposed model , we first need to introduce some quantum probability concepts for the understanding of this work . sections [ sec : probability ] and [ sec : traj ] present the main differences between classical and quantum probability theory . instead of just presenting a set of formulas , we show this difference by means of an illustrative example , just like proposed in @xcite . in section [ sec : interference ] , we describe how beliefs can act like waves and interfere with each other . we show mathematically how this interference term can be derived by using well known rules of complex numbers . section [ sec : theta_lit ] addresses the main works of the literature that contributed for the development of the interference term . it also introduces a new interference formula that will be applied in the proposed quantum probabilistic graphical models . section [ sec : classic_vs_quantum_gamble ] presents a comparison between a classical bayesian network model against the proposed quantum interference bayesian network applied to the problem of two - stage gambles . section [ sec : classic_vs_quantum_burglar ] presents another comparison between the classical and quantum bayesian networks , but for a more complex network from the literature . in section [ sec : discussion ] , it is made a discussion about the results obtained in the experiments performed in section [ sec : classic_vs_quantum_burglar ] . section [ sec : opt ] presents an additional experiment over another bayesian network in order to study the impact of the quantum interference parameters in different scenarios . section [ sec : rel_work ] presents the most relevant works of the literature . finally , section [ sec : conclusions ] presents the main conclusions of this work . in this section , we describe the main differences between classical theory and quantum probability theory through examples . the example analyzed concerns jury duty . suppose you are a juror and you must decide whether a defendant is guilty or innocent . the following sections describe how the classical and quantum theory evolve in the inference process . all this analysis is based on the book of @xcite . in classical probability theory , events are contained in sample spaces . a sample space @xmath12 corresponds to the set of all possible outcomes of an experiment or random trial @xcite . for example , when judging whether a defendant is guilty or innocent , the sample space is given by @xmath13 . figure [ fig : sample_space ] presents a diagram showing the sample space of a defendant being guilty or innocent . + in quantum probability theory , events are contained in the so called hilbert spaces . a hilbert space @xmath14 can be viewed as a generalisation and extension of the euclidean space into spaces with any finite or infinite number or dimensions . it can be see as a vector space of complex numbers and offers the structure of an inner product to enable the measurement of angles and lengths @xcite . the space is spanned by a set of orthonormal basis vectors @xmath15 . together , these vectors form a basis for the space . figure [ fig : hilbert_space ] presents a diagram showing the hilbert space of a defendant being guilty or innocent @xcite . since a hilbert space enables the usage of complex numbers , then , in order to represent the events @xmath16 and @xmath17 , one would need two dimensions for each event ( one for the real part and another for the imaginary part ) . in quantum theory , one usually ignores the imaginary component in order to be able to visualise geometrically all vectors in a 2-dimensional space . in classical probability theory , events can be defined by a set of outcomes to which a probability is assigned . they correspond to a subset of the sample space @xmath12 from which they are contained in . events can be mutually exclusive and they obey to set theory . this means that operations such as intersection or union of events are well defined . since they respect set theory , the distributive axiom is also defined between sets . in our example , @xmath16 or @xmath17 can be seen as two mutually exclusive events . . ] in figure [ fig : quantum_events ] , the superposition state @xmath18 can be defined as follows . @xmath20 in equation [ eq : ampl ] , one might be wondering what the @xmath21 values mean . they are called @xmath22 @xmath23 . they correspond to the amplitudes of a wave and are described by complex numbers . the @xmath24 term is defined as the phase of the amplitude . it can be seen as a shift of the wave . these amplitudes are related to classical probability by taking the squared magnitude of these amplitudes . this is achieved by multiplying the amplitude with its complex conjugate ( represented by the symbol @xmath25 ) . @xmath26 in quantum theory , it is required that the sum of the squared magnitudes of each amplitude equals @xmath27 . this axiom is called the normalization axiom and corresponds to the classical theory constraint that the probability of all events in a sample space should sum to one . [ sec : system_state ] a system state is nothing more than a probability function @xmath29 which maps events into probability numbers , i.e. , positive real numbers between @xmath30 and @xmath27 . in classical theory , the system state corresponds to exactly its definition . there is a function that is responsible to assign a probability value to the outcome of an event . if the event corresponds to the sample space , then the system state assigns a probability value of @xmath27 to the event . if the event is empty , then it assigns a probability of @xmath30 . in our example , if nothing else is told to the juror , then the probability of the defendant being guilty is @xmath31 . in quantum theory , the probability of a defendant being @xmath16 is given by the squared magnitude of the projection from the superposition state @xmath18 to the subspace containing the observed event @xmath16 . figure [ fig : hilbert_system_state ] shows an example . if nothing is told to the juror about the guiltiness of a defendant , then according to quantum theory , we start with a superposition state @xmath18 . @xmath32 when someone asks whether the defendant is guilty , then we project the superposition state @xmath18 into the relevant subspace , in this case the @xmath16 subspace @xmath33 , just like shown in figure [ fig : hilbert_system_state ] . the probability is simply the squared magnitude of the projection , that is : @xmath34 which has exactly the same outcome as in the classical theory . . then , this superposition state is projected into the @xmath35 subspace @xmath36 . the final probability corresponds to the squared magnitude of the projection . ] state revision corresponds to the situation where after observing an event , we are interested in observing other events given that the previous one has occurred . in classical theory , this is addressed through the conditional probability formula @xmath37 . so , returning to our example , suppose that some evidence has been given to the juror proving that the defendant is actually guilty , then what is the probability of him being innocent ? this is computed in the following way . @xmath38 since the events @xmath16 and @xmath17 are mutually exclusive , then their intersection is empty , leading to a zero probability value . in quantum theory , the state revision is given by first projecting the superposition state @xmath18 into the subspace representing the observed event . then , the projection is normalised such that the resulting vector is unit length . again , if we want to determine the probability of a defendant being innocent , given he was found guilty , the calculations are performed as follows . we first start in the superposition state vector @xmath18 . @xmath39 then , we observe that the defendant is guilty , so we project the state vector @xmath18 into the @xmath16 subspace and normalise the resulting projection . @xmath40 @xmath41 from the resulting state , we just extract the probability of being innocent by simply squaring the respective probability amplitude . again , we obtain the same results as the classical theory . @xmath42 subspace . ] in order to describe direct dependencies between a set of variables , path diagrams are generally used . this section shows how to compute quantum probabilities in a markov model using feynman s path rules , just like presented in the work of @xcite . consider the diagram represented in figure [ fig : single_trajectory ] . the computation of the probability of transiting from an initial state @xmath0 to a final state @xmath43 , transiting from an intermediate state @xmath1 , can be achieved through a classical markov model . the probability can be computed by making the product of the individual probabilities for each transition , from one state to another , through the usage of conditional probabilities . according to figure [ fig : single_trajectory ] , the probability of transiting from state @xmath0 , followed by state @xmath1 and ending in state @xmath43 , that is , @xmath44 , is given by : @xmath45 in a quantum path diagram model , the computation of the probabilities for a single path trajectory is similar to the classical markov model . the calculation can be performed using feynman s first rule , which asserts that the probability of a single path trajectory consists in the product of the squared magnitudes of the amplitudes for each transition from one state to the next along the path . this means that the quantum probability value of a single path trajectory is the same as the classical markov probability for the same path . in equation [ eq : feynman1 ] and through the rest of this work , complex probability amplitudes will be represented by the symbol @xmath46 . @xmath47 an indistinguishable path consists in transiting from an initial state @xmath0 to a final state @xmath48 by transiting from multiple possible paths without knowing for certain which path was taken to reach the goal state . figure [ fig : multiple_trajectory_unobs ] shows an example of multiple indistinguishable trajectories . in a classical markov model , if one does not observe which path was taken to reach the final state @xmath48 , then one simply computes this probability by summing the individual probabilities of each path . this is known as the single path principle and is in accordance with the law of total probability . so , following figure [ fig : multiple_trajectory_unobs ] , in order to reach state @xmath48 starting in state @xmath0 , one can take the path @xmath49 or the path @xmath50 . the final probability is given by : @xmath51 quantum probability theory rejects the single path trajectory principle . if one does not observe which path was taken to reach the goal state , then one can not assume that one of each possible paths was used . instead , quantum probability argues that , when the path is unobserved , then the goal state can be reached through a superposition of path trajectories . this is known as feynman s second rule , which states that the amplitude of transiting from an initial state @xmath0 to a final state @xmath48 , taking multiple indistinguishable paths , is given by the sum of all amplitudes for each path . this rule is in accordance with the law of total amplitude and the probability is computed by taking the squared magnitude of this sum . this probability is not equal to the classical markov model . @xmath52 @xmath53 the term @xmath54 is the inner product between the vectors formed by @xmath55 and @xmath56 . it comes from euler s rule : @xmath57 and corresponds to a quantum interference term that does not exist in classical probability theory . section [ sec : interference ] details how this term is derived . since the interference term can lead to estimations with values higher than 1 , then it is necessary to normalize this value in order to obtain a probability value . when the path is observed , quantum probability theory collapses to the classical markov model . this is know as feynman s third rule and states that the probability amplitude of observed multiple path trajectories corresponds to the sum of the amplitudes of each individual path . the probabilities are then taken by making the squared magnitude of each individual path . figure [ fig : multiple_trajectory_obs ] illustrates this example : @xmath58 @xmath59 quantum theory enables the modeling of the decision system as a wave moving across time over a state space until a final decision is made . under this perspective , interference can be regarded as a chain of waves in a superposition state , coming from different directions . when these waves crash , one can experience a destructive effect ( one wave destroys the other ) or a constructive effect ( one wave merges with another ) . in either case , the final probabilities of each wave is affected . psychological findings showed that this not only occurs in a microscopic scale ( such as electrons ) , but also occurs at a macroscopic setting @xcite . in this section we show how to derive the interference term by just taking into account well known properties of complex numbers @xcite . the interference term can be derived in two different ways : ( 1 ) from the general probability formula of the union of @xmath60 mutually exclusive events and ( 2 ) through the total law of probability . for simplicity , we will start by deriving the interference term for 2 events and then we will generalize for @xmath60 events . the classical probability formula of the union of two mutual exclusive events is given by : @xmath61 and the relation between a classical probability density function and a quantum probability amplitude is given by born s rule , that is : @xmath62 again , @xmath63 corresponds to the square magnitude of a complex amplitude . it is obtained by multiplying the probability amplitude with its complex conjugate . that is , @xmath64 . taking into account equation [ eq : relation ] , one can write a superposition state between two mutual exclusive events @xmath0 and @xmath1 in the following way : @xmath65 the relation of this superposed state with classical probability theory remains : @xmath66 the quantum counterpart of the classical probability of the union of two mutually exclusive events , when we do not observe them , collapses to feynmann s second rule , that is : @xmath67 @xmath68 @xmath69 the classical probability of the union of two mutual exclusive events is @xmath70 . in equation [ eq : int_1 ] , the amplitude @xmath71 corresponds to @xmath72 and @xmath73 corresponds to @xmath74 . so , what is the additional term , @xmath75 , that we derived in equation [ eq : int_1 ] ? this term does not exist in classical probability theory and is called _ the interference term_. the interference term can be rewritten like in equation [ eq : int_2 ] : @xmath76 knowing that , @xmath77 then equation [ eq : int_2 ] can be rewritten as : @xmath78 so , the complete quantum probability formula for the union of 2 mutually exclusive events is given by : @xmath79 in the above formula , the angle @xmath80 corresponds to the phase of the inner product between @xmath81 and @xmath82 . equation [ eq : int_4 ] only computes the probability of the union of 2 mutually exclusive events . in classical probability theory , if we want to compute the probability of the union of @xmath60 mutually exclusive events , then we use a generalization of equation [ eq : union ] , that is : @xmath83 in quantum theory , we make an analogous calculation . in order to compute the probability of the union of @xmath60 mutually exclusive events , then one needs to generalize equation [ eq : qn_union ] . the outcome is given by the following formula : @xmath84 @xmath85 this interference term can also be derived directly from the law of total probability . suppose that events @xmath86 form a set of mutually disjoint events , such that their union is all in the sample space , @xmath12 , for any other event @xmath1 . then , the classical law of total probability can be formulated like in equation [ eq : law_total_prob_c ] . @xmath87 the quantum interference law of total probability can be derived through equation [ eq : law_total_prob_c ] by applying born s rule ( equation [ eq : relation ] ) . that is : @xmath88 for simplicity , we will expand equation [ eq : total_law_q_1 ] for @xmath89 and only later we will find the general formula for @xmath60 events : @xmath90 @xmath91 @xmath92 @xmath93 @xmath94 @xmath95 simplifying equation [ eq : total_law_q_4 ] , we obtain : @xmath96 in equation [ eq : total_law_q_5 ] , one can see that it is composed by the classical law of total probability and by an interference term . this interference term comes from equation [ eq : total_law_q_4 ] and corresponds to : @xmath97 @xmath98 @xmath99 knowing that @xmath100 then , equation [ eq : total_law_q_5 ] becomes @xmath101 generalizing equation [ eq : total_law_q_7 ] for @xmath60 events , the final probabilistic interference formula , derived from the law of total probability , is given by : @xmath102 following equation [ eq : total_law_q_final ] , when @xmath103 equals zero , then it is straightforward that quantum probability theory converges to its classical counterpart , because the interference term will be zero . for non - zero values , equation [ eq : total_law_q_final ] will produce interference effects that can affect destructively the classical probability ( when interference term in smaller than zero ) or constructively ( when it is bigger than zero ) additionally , equation [ eq : total_law_q_final ] will lead to a large amount of @xmath104 parameters when the number of events increases . for @xmath60 binary random variables , we will end up with @xmath105 parameters to tune . in the 20th century , the physicist max born proposed a problem related to quantum probabilities , which was later known as the _ the inverse born problem_. the problem consisted in constructing a probabilistic representation of data from different sources ( physics , psychology , economy , etc ) , by a complex probability amplitude , which could match born s rule ( already presented in equation [ eq : relation ] ) . the probabilistic interference formula for the law of total probability , derived in the previous section ( equation [ eq : total_law_q_final ] ) , can be seen as an answer to the inverse born problem . the most important works in the literature that contributed for the derivation of this interference term , through the law of total probability , correspond to the works of a. khrennikov @xcite . these authors address the interference term as @xmath106 . in the situations where @xmath107 , then one can apply the trigonometric formula derived in equation [ eq : total_law_q_final ] . however , there are some data where this condition is not verified . therefore , when @xmath108 , the authors propose the usage of a hyperbolic interference term , to act like an upper boundary in order to constraint the probability value to a maximum value of @xmath27 . this would require the usage of hyperbolic hilbert spaces instead of the complex ones . in this paper , we argue that there is no need to represent probabilities in a hyperbolic hilbert space in order to avoid non - probability values . since we will be dealing with probabilistic graphical models , we will always be required to normalise the probability amplitudes when performing probabilistic inferences . for this work , we follow the same probabilistic paradigm used in traditional bayesian networks . thus , we constrain equation [ eq : total_law_q_final_norm ] and equation [ eq : union_qn_norm ] by a normalisation factor @xmath109 that will guarantee that the computed values will always be probabilities lesser or equal than one . this normalisation factor corresponds to feynman s conjecture @xcite that an electron can follow any path . thus , in order to compute the probability @xmath74 that a particle ends up at a point @xmath1 , one must sum over all possible paths that the particle can go through . since the interference term can lead to estimations with values higher than @xmath27 , then it is necessary to normalise in order to obtain a probability value . @xmath111 \label{eq : union_qn_norm}\ ] ] @xmath112 through the triangular inequality , we can also prove that the proposed interference term is also always positive . this way , the axioms of probability theory that state that @xmath113 will always be satisfied . by applying the triangular inequality , one can easily demonstrate that equation [ eq : union_qn_norm ] and , consequently , equation [ eq : total_law_q_final_norm ] have always to be positive . through the triangular inequality , equation [ eq : union_qn_norm ] can be related to : @xmath114 @xmath115 @xmath116 @xmath117 @xmath118 \ge 0\ ] ] the maximum value that @xmath119 can have is @xmath27 and the minimum value is @xmath120 . so , the minimum value that the term @xmath121 can have is @xmath30 ( when the cosine achieves its maximum value ) . given that @xmath122 are always positive numbers , then it is straightforward that equation [ eq : total_law_q_final_norm ] has to be always positive . a classical bayesian network can be defined by a directed acyclic graph structure in which each node represents a different random variable from a specific domain and each edge represents a direct influence from the source node to the target node . the graph represents independence relationships between variables and each node is associated with a conditional probability table which specifies a distribution over the values of a node given each possible joint assignment of values of its parents . this idea of a node depending directly from its parent nodes is the core of bayesian networks . once the values of the parents are known , no information relating directly or indirectly to its parents or other ancestors can influence the beliefs about it @xcite . associated to bayesian networks there is always the concept of conditional independence . two random variables @xmath2 and @xmath123 are conditionally independent given a third random variable @xmath124 if and only if they are independent in their conditional probability distribution given @xmath124 . in other words , @xmath2 and @xmath123 are conditionally independent given @xmath124 , @xmath125 , if and only if , given any value of @xmath124 , the probability distribution of @xmath2 is the same for all values of @xmath123 and the probability distribution of @xmath123 is the same for all values of @xmath2 . this means that an independence statement over random variables is a universal quantification over all possible values of random variables @xcite . therefore , a probability distribution @xmath29 satisfies @xmath126 if and only if : @xmath127 in classical probability theory , a random variable @xmath2 is defined by a function that associates a value to each outcome in the sample space @xmath12 , @xmath128 . in the two - stage gambling game , the random variables correspond to the nodes and their respective conditional probability tables of the bayesian network in figure [ fig : classical ] . that is , the variable @xmath129 corresponds to a player willing or not to participate in the game . variable @xmath130 corresponds to the variable winning or losing the first gamble . variable @xmath131 corresponds to a player playing or not the second gamble . in classical probability theory , the full joint distribution over a set of @xmath132 random variables @xmath133 @xmath134 @xmath135 @xmath136 @xmath137 @xmath138 @xmath139 defined over the same sample space , @xmath140 , is the distribution that assigns probabilities to events that are specified in terms of these random variable @xcite.then , the full joint distribution of a bayesian network , where @xmath2 is the list of variables , is given by @xcite : @xmath141 using equation [ eq : joint ] , the full joint distribution of figure [ fig : classical ] corresponds to the calculations discriminated in tables [ tab : full_joint ] . .fulll joint distribution of the bayesian newtwork in figure [ fig : classical ] [ cols="<,<,<,^",options="header " , ] figures [ fig : dyspnea]-[fig : cancer ] also show all the possible probability values that the variables from the lung cancer bayesian network can achieve , when nothing is known . when we are at a maximum level of uncertainty , classical probabilities tend to assume that events are equiprobable , that is , the probability of their outcome is always the same no matter their context . quantum theory , on the other hand , provides a more relaxed framework . when nothing is known , then there are more degrees of freedom that enable the outcome of any event to be any possible value . this way , by analysing the context of certain events ( for example , a medical decision scenario , a gambling game , etc ) , the quantum parameters @xmath104 can be modelled in such a way that they can simulate reality more accurately and more precisely than its classical counterpart . in this sense , in order to develop more accurate quantum bayesian networks , a study of the context of the problem is required in order to start the search for the optimum @xmath104 parameters . techniques to search for these optimum quantum parameters is still an open research question and an unexplored field in the literature of quantum cognition and quantum decision models . since the preliminary findings of @xcite , cognitive scientists began to use alternative theories , such as quantum probability theory , in an attempt to explain the paradoxical findings that classical probability theory could not explain . given the flexibility of the quantum probabilistic framework , several researchers applied these concepts in several fields outside of physics . in this section , we present some of the topics in which quantum probability has made some impacts in the literature . in what concerns violations of probability theory , there are many paradoxical situations where classical probability is unable to model . the most important paradoxes consist in violations in the sure thing principle . examples of such violations correspond to the ellsberg paradox @xcite , the two - stage gambling game @xcite and the prisoner s dilemma game @xcite . in what concerns the prisoner s dilemma game , many works in the literature have been proposed , which formalise the problem in a quantum approach . for instance , @xcite proposed a quantum dynamic markov model for the prisoner s dilemma game . in their model , the authors represented the players beliefs and actions as a superposition in a quantum state vector . when the player knew the information about the opponent s action , the superposition state collapsed to a new state that was compatible with the action chosen by the opponent . when the player did not know about the opponent s move , then the system involved through the application of unitary transformations . in @xcite , the authors focus on irrational choices and developed a model based on the quantum superposition and interference principles for the prisoner dilemma game . after each play , the state is updated until it reaches an equilibrium . in @xcite , the authors analysed quantum conditional probabilities in the prisoner s dilemma game . there are also works that consider that the probabilities in the prisoner s dilemma game correspond to different contexts and these contexts can be incompatible . the authors argue that the probabilities observed in the prisoner s dilemma game were not classical , but were not quantum either in the sense of traditional quantum mechanics theory . so the author introduced the quantum - like approach and suggested a new trigonometric interference term @xcite . @xcite proposed quantum model which incorporated entangled decisions and explained the violations of the sure thing principle through interference effects between intentions . in @xcite , the authors explored the impact of these violations in economics and analyse the two - stage gambling game and the prisoner s dilemma game under a quantum probabilistic point of view , by proposing a quantum markov model to explain the paradoxical observations in these games . finally , @xcite proposed a quantum bayesian updating scheme based on the formalisms of quantum mechanics that represents mental states on a hilbert state . through the usage of projections , they were able to introduce a model that could explain the paradoxical findings of the two - stage gambling game @xcite . another similar work that compares bayes rule to its quantum counterpart corresponds to @xcite . other situations , in which the laws of probability are being violated , correspond to disjunction or conjunction fallacies . a conjunction error occurs when it is assumed that specific conditions are more probable than a single general one . a disjunction error , on the other hand , consists in assuming that the disjunction of two events is at least as likely as either of the events individually . several works in the literature used quantum probabilistic models to address these fallacies . for instance , @xcite developed a quantum probabilistic model to explain fallacies when making preferences over a set of events , more specifically , the conjunction fallacy . @xcite also focused on quantum models to explain conjunction and @xcite focused on disjunction fallacies . the first experiments where these fallacies were observed to occur were performed by @xcite . an alternative model for conjunction and disjunction errors correspond to @xcite . the authors used support theory as a subjective probabilistic model that describes the way people make probability judgments . the main problem of their model is that support theory is able to successfully describe conjunction errors , but fails at explaining disjunction errors . other models correspond to @xcite . @xcite also addressed the problem of conjunction and disjunction fallacies . they proposed quantum model which incorporated entangled decisions . @xcite also modelled mental processes through quantum probabilities , where the interference process plays an important role in the process of recognising images@xcite . other interesting works of this author applying similar quantum formalisms correspond to @xcite . in what concerns quantum probabilistic graphical models , there are various contributions . for example , in @xcite the model proposed by the author is exactly the same as a classical bayesian network . the only difference is that it uses complex numbers in the conditional probability tables instead of real values . a similar model has been proposed by @xcite , but for markov networks . in @xcite , the author proposed a quantum bayesian network by replacing the classical formulas used to perform the inference process by their quantum counterpart . @xcite , @xcite and @xcite proposed a quantum dynamic markov model based on the findings of cognitive psychologists and interference terms . one of the first and most influential models that applied quantum probability for decision making belongs to @xcite . in their work , the authors proposed the @xmath142-model , which can be defined as a quantum machine that corresponds to a two dimensional hilbert space . given an event in this quantum machine one can compute quantum like probabilities . the model also makes use of a parameter @xmath142 that measures the total amount of uncertainty when computing the probabilities . other works related to this model can be found in @xcite . @xcite proposed a quantum bayesian updating scheme based on the formalisms of quantum mechanics that represents mental states on a hilbert state . through the usage of projections , they were able to introduce a model that could explain the paradoxical findings of the two - stage gambling game @xcite . another similar work that compares bayes rule to its quantum counterpart corresponds to @xcite . there is also an increasing interest in applying quantum like models to subjects outside of psychology . in game theory there have been works exploring the uncertainties of a player towards another through quantum probabilistic models . for instance , @xcite focus on irrational choices and developed a model based on the quantum superposition and interference principles for the prisoner dilemma game . after each play , the state is updated until it reaches an equilibrium . @xcite used the data collected by @xcite and analysed quantum conditional probabilities . @xcite generalised the classical expected utility and proposed a quantum projective expected utility function that does not violate the laws of classical probability theory and can accommodate allais and ellsberg paradoxes . @xcite made a formalisation of the structure of quantum probability theory and @xcite also applied these formalists to expected utilities . @xcite applied quantum theory to the fields of economics and game theory and developed a principle to minimise financial risks @xcite modelled how a person updates its beliefs in the liar s paradox . @xcite also modelled mental processes through quantum probabilities , where the interference process plays an important role in the process of recognising images@xcite . other interesting works of this author applying similar quantum formalisms correspond to @xcite . @xcite developed a quantum mechanics based framework in order to model agents preferences under uncertainty . more recently , @xcite published a work that generated discussions over the scientific literature , about whether or not , quantum probability can provide a new direction to compute quantum probabilistic inferences . in their work , they summarised the main areas where quantum probability has made some impact ( conjunction / disjunction errors , sure thing principle , etc ) . a deep analysis of the differences between classical and quantum probability is also made in this work . @xcite also shares a similar opinion that quantum probability can , in fact , provide a new framework to explain several situations in which the laws of probability are violated . however , they defend that the usage of quantum probability should be more close to its real meaning in quantum mechanics and proposes to explain the phenomena , such as contextually , entanglement , observables and flock spaces . the work of , @xcite has stimulated some discussions on the implications of quantum probability in functional brain networks . for instance , @xcite have discussed the similarities between the vector representations of quantum states with a vector symbolic architecture , which is used to model realistic biological neural models . in this sense , biological neural models can also incorporate the quantum probabilistic framework and also introduce quantum interference effects , as a mathematical and geometric alternative framework , without taking into account the true meaning that is given these effects under quantum physics . other authors also support this idea of representing neuronal model using the quantum probabilistic framework @xcite . this work was motivated by the preliminary experiments of @xcite about violations of the classical probability theory on the sure thing principle . this principle states that if one chooses action @xmath0 over @xmath1 in a state of the world @xmath2 , and if one also chooses action @xmath0 over @xmath1 under the complementary state of the world @xmath2 , then one should always choose action @xmath0 over @xmath1 , even when the state of the world in unspecified . when humans need to make decisions under risk , several heuristics are used , since humans can not process large amounts of data . these decisions coupled with heuristics lead to violations on the law of total probability . recent work in cognitive psychology revealed that quantum probability theory provides another method of computing probabilities without falling into the restrictions that classical probability have in modelling cognitive systems of decision making . quantum probability theory can also be seen as a generalisation of classical probability theory , since it also includes the classical probabilities as a special case ( when the interference term is zero ) . the main difference between quantum and classical probability lies in the fact that on quantum probability we are constantly updating some beliefs when making a decision , while in classical probability all beliefs are assumed to have a definite value before a decision is made , and this value is the outcome of the decision @xcite . the main research question for this work was how could these quantum probabilities affect probabilistic graphical models , such as bayesian networks , since many of nowadays decision making systems are based on such structures ( medical diagnosis , spam filtering , image segmentation , etc ) . in this work , we proposed a novel bayesian network for the computer science community based on quantum probabilities . our method can accommodate puzzling observations that the classical probability failed to explain ( for instance , the two - step gambling game ) . when the nodes of the proposed bayesian network are represented as a superposition state , then one can look at this state as many waves moving across in different directions . these waves can crash into each other causing waves to be bigger or to cancel each other . this is the interference phenomena that the proposed bayesian network offers and that has direct implications when making inferences . therefore , the proposed network represents and simulates quantum aspects motivated by feynman s path integrals . experimental results revealed that the proposed quantum bayesian network enables many degrees of freedom in choosing the final outcome of the probabilities . if we had a real scenario , with real observations , one could use the present model to fit it to the observed data , by simply tuning the parameter @xmath104 . this parameter can open a door into machine learning approaches . learning algorithms using the proposed method might produce better prediction models , since the quantum probability amplitudes are able to fully represent real word data . for future work , we intend to explore machine learning algorithms under quantum probabilistic graphical model formalisms . the overall results also suggested that when the classical probability of some variable is already high , then the quantum probability tends to increase it even more . when the classical probability is very low , then the proposed model tends to lower it . when there are many unobserved nodes in the network then the levels of uncertainty are very high . but , in the opposite scenario , when there are very few unobserved nodes , then the proposed quantum model tends to collapse into its classical counterpart , since the uncertainty levels are very low . the proposed bayesian network can integrate human thoughts by representing a person s beliefs in an n - dimensional unit length quantum state vector . in the same way , the proposed quantum structure is general enough to represent any other context in which there is a need to formalise uncertainty , including prediction problems in data fusion . in the context of bayesian networks , data fusion is introduced in the work of @xcite . the author argues that , just like people , bayesian networks are structures that integrate data from multiple sources of evidence and enable the generation of a coherent interpretation of that data through a reasoning process . the fusion of all these multiple data sources can be done using bayes theorem . when a data source is unknown , then the bayes rule is extended in order to sum out all possible values of the probability distribution representing the unknown data source . the proposed quantum bayesian network takes advantage of these uncertainties by representing them in a superposition state , which enables the fusion of the data sources through quantum interference effects . these effects produce changes in the final likelihoods of the outcomes and provide a promising way to perform predictions more accurately according to reality . so , the quantum bayesian network that is proposed in this work is potentially relevant and applicable in any behavioural situation in which uncertainty is involved . 72 luigi accardi , andrei khrennikov , and masanori ohya . quantum markov model for data from shafir - tversky experiments in cognitive psychology . , 14:371385 , 2009 . diederik aerts . quantum structures : an attempt to explain the origin of their appearence in nature . , 34:122 , 1995 . diederik aerts . quantum structure in cognition . , 53:314348 , 2009 . diederik aerts , jan broekaert , and liane gabora . a case for applying an abstracted quantum formalism to cognition . , 29:136146 , 2011 . diederik aerts , jan broekaert , liane gabora , and sandro sozzo . quantum structure and human thought . , 36:274276 , 2013 . diederik aerts , jan broekaert , and sonja smets . a quantum structure description of the liar paradox . , 38:32313239 , 2004 . diederik aerts and thomas durt . quantum , classical and intermediate , an illustrative example . , 24:13531369 , 1994 . diederik aerts , thomas durt , and bruno van bogaert . a physical example of quantum fuzzy sets and the classical limit . , 1:515 , 1993 . diederik aerts , thomas durt , and bruno van bogaert . quantum probability , the classical limit and nonlocality . in _ proceedings of the international symposium on the foundations of modern physics _ , 1993 . diederik aerts and liane gabora . a theory of concepts and their combinations i : the structure of the sets of contexts and properties . , 34:167191 , 2005 . diederik aerts and liane gabora . a theory of concepts and their combinations ii : a hilbert space representation . , 34:192221 , 2005 . sven aerts . conditinal probabilities with a quantal and kolmogorovian limit . , 35:2245 , 1996 . sven aerts . interactive probability models : inverse problems on the sphere . , 37:305309 , 1998 . masanari asano , irina basieva , andrei khrennikov , masanori ohya , and yoshiharu tanaka . quantum - like generalization of the bayesian updating scheme for objective and subjective mental uncertainties . , 56:166175 , 2012 . masanari asano , masanori ohya , and andrei khrennikov . quantum - like model for decision making process in two players game , a non - kolmogorovian model . , 41:538548 , 2010 . arpan banerjee and barry horwitz . can quantum probability help analyze the behaviour of functional brain networks ? , 36:278279 , 2013 . jerome busemeyer and peter bruza . . cambridge university press , 2012 . jerome busemeyer , marvin matthew , and zheng wang . a quantum information processing explanation of disjunction effects . in _ proceedings of the 28th annual conference of the cognitive science society _ , 2006 . jerome busemeyer , emmanuel pothos , riccardo franco , and jennifer trueblood . a quantum theoretical explanation for probability judgment errors . , 118:193218 , 2011 . jerome busemeyer and jennifer trueblood . comparison of quantum and bayesian inference models . in _ proceedings of the 3rd international symposium on quantum interaction _ , 2009 . jerome busemeyer , zheng wang , and ariane lambert - mogiliansky . empirical comparison of markov and quantum models of decision making . , 53:423433 , 2009 . jerome busemeyer , zheng wang , and james townsend . quantum dynamics of human decision making . , 50:220241 , 2006 . taksu cheon and taiki tahahashi . interference and inequality in quantum decision theory . , 375:100104 , 2010 . elio conte , andrei khrennikov , orlando todarello , antonio federici , leonardo mendolicchio , and joseph zbilut . mental states follow quantum mechanics during perception and cognition of ambiguous figures . , 16:117 , 2009 . elio conte , orlando todarello , antonio federici , francesco vitiello , michele lopane , andrei khrennikov , and joseph zbilut . some remarks on an experiment suggesting quantum like behavior of cognitive entities and formulation of an abstract quantum mechanical formalism to describe cognitive entity and its dynamics . , 31:10761088 , 2007 . ratchel croson . the disjunction effect and reason - based choice in games . , 80:118133 , 1999 . vladimir danilov and ariane lambert - mogiliansky . measurable systems and behavioral sciences . , 55:315340 , 2008 . vladimir danilov and ariane lambert - mogiliansky . expected utility theory under non - classical uncertainty . , 68:2547 , 2010 . jos acacio de barros . quantum - like model of behavioral response computation using neural oscillators . , 110:171182 , 2012 . jose acacio de barros and patrick suppes . quantum mechanics , interference , and the brain . , 53:306313 , 2009 . morris degroot and mark schervish . pearson education ( 4th edition ) , 2011 . daniel ellsberg . risk , ambiguity and the savage axioms . , 75:643669 , 1961 . richard feynman , robert leighton , and mathew sands . . addison - wesley , 1965 . riccardo franco . the conjunction fallacy and interference effects . , 53:415422 , 2009 . liane gabora and diederik aerts . contextualizing concepts using a mathematical generalization of the quantum formalism . , 14:327358 , 2002 . gerd gigerenzer and daniel g. goldstein . reasoning the fast and frugal way : models of bounded rationality . , 103:650669 , 1996 . gerd gigerenzer and ulrich hoffrage . how to improve bayesian reasoning without instruction : frequency formats . , 102:684704 , 1995 . stuart hameroff . how quantum brain biology can rescue conscious free will . 6:117 , 2012 . emmanuel haven and andrei khrennikov . . cambridge university press , 2013 . mika hirvensalo . . springer , 2003 . evgenia hristova and maurice grinberg . disjunction effect in prisonner s dilemma : evidences from an eye - tracking study . in _ proceedings of the 30th annual conference of the cognitive science society _ , 2008 . andrei khrennikov . classical and quantum mechanics on information spaces with applications to cognitive psychological , social and anomalous phenomena . , 29:10651098 , 1999 . andrei khrennikov . interference of probabilities in the classical probabilistic framework . , 155:417 , 2005 . andrei khrennikov . quantum - like brain : interference of minds . , 84:225241 , 2006 . andrei khrennikov . can quantum information be processed by macroscopic systems ? , 6:401429 , 2007 . andrei khrennikov . description of composite quantum systems by means of classical random fields . , 40:10511064 , 2009 . andrei khrennikov . quantum - like model of cognitive decision making and information processing . , 95:179187 , 2009 . andrei khrennikov and emmanuel haven . quantum mechanics and violations of the sure - thing principle : the use of probability interference and other concepts . , 53:378388 , 2009 . daphne koller and nir friedman . . the mit press , 2009 . anton kuhberger , dagmara komunska , and perner josef . the disjunction effect : does it exist for two - step gambles ? , 85:250264 , 2001 . charles lambdin and charles burdsal . the disjunction effect reexamined : relevant methodological issues and the fallacy of unspecified percentage comparisons . , 103:268276 , 2007 . ariane lambert - mogiliansky , shmuel zamir , and herve zwirn . type indeterminacy : a model for the kt(kahneman - tversky)-man . , 53:349361 , 2009 . mathew leifer and david poulin . quantum graphical models and belief propagation . , 323:18991946 , 2008 . eprint arxiv - ph/0708.1337 [ quant - ph ] . shu li and john taplin . examining whether there is a disjunction effect in prisoner s dilemma game . , 44:2546 , 2002 . donald metzler and w. bruce croft . linear feature - based models for information retrieval . , 16(28):123 , 2007 . pierfrancesco la mura . projective expected utility . , 53:408414 , 2009 . pierfrancesco la mura and lukasz swiatczak . markov entangled networks . technical report , aaai spring symposium , 2007 . judea pearl . fusion , propagation , and structuring in belief networks . , 29:241288 , 1986 . judea pearl . . morgan kaufmann publishers , 1988 . edward piotrowski . quantum - like approach to financial risk : quantum anthropic principle . , 32:38733879 , 2001 . emmanuel pothos and jerome busemeyer . a quantum probability explanation for violations of rational decision theory . , 276:21712178 , 2009 . emmanuel pothos and jerome busemeyer . can quantum probability provide a new direction for cognitive modeling ? , 36:255327 , 2013 . sturat russel and peter norvig . . pearson education ( 3rd edition ) , 2010 . leonard savage . . john wiley , 1954 . eldar shafir and amos tversky . thinking through uncertainty : nonconsequential reasoning and choice . , 24:449474 , 1992 . terrence stewart and chris eliasmith . realistic neurons can compute the operations needed by quantum probability theory and other vector symbolic architectures . , 36:307308 , 2013 . katya tentori , vincenzo crupi , and selena russo . on the determinants of the conjunction fallacy : confirmation versus probability . , 142:23555 , 2013 . robert tucci . quantum bayesian nets . , 9:295337 , 1995 . amos tversky and daniel kahnenman . judgment under uncertainty : heuristics and biases . , 185:11241131 , 1974 . amos tversky and derek koehler . support theory : a nonextensional representation of subjective probability . , 10:547567 , 1994 . amos tversky and eldar shafir . the disjunction effect in choice under uncertainty . , 3:305309 , 1992 . vyacheslav yukalov and didier sornette . decision theory with prospect interference and entanglement . , 70:283328 , 2011 . the most naive approach of parameter search is the grid search method . in this approach , it is placed a grid over the parameter space and the data is evaluated at every grid intersection , returning the parameters , which lead to the maximum performance of an algorithm @xcite . however , grid search has the problem of being unbounded , since an infinite set of parameters are available to be tested . in the scope of this work , this is not a problem , because we always have a boundary . the values of a cosine function can all be specified for angles between the range @xmath143 $ ] . the grid search approach is a very naive method for finding parameters in the parameter space . in fact , there are several advanced parameter search algorithms in the literature , which do not have a heavy computational cost . the motivation behind the usage of the grid search approach in this work is that the computational time required by any of the other advanced methods is almost the same as using grid search . in addition , many of the advanced search parameter methods in the literature perform approximations , which can be avoided by the direct parameter search of the grid search approach @xcite . classical probability when it is true ( pr(a = t ) ) , @xmath144 , + interference term correspondent to @xmath144 , @xmath145 , + classical probability when it is true ( pr(a = f ) ) , @xmath146 , + interference term correspondent to @xmath146,@xmath147 , + + maximum quantum probabilities , + list of @xmath104 parameters that maximise the quantum probability + + @xmath148 // initialise variable that will store the maximum probability found

an exotic and distinctive feature of interacting one - dimensional systems is the phenomenon of spin - charge separation where the elementary excitations of the fermionic system are decoupled charge and spin bosonic modes that propagate at different velocities.@xcite at low energies , luttinger liquid theory@xcite has been quite successful at describing the properties of one - dimensional systems such as quantum wires@xcite and nanotubes.@xcite one of the elegant features of the theory is the ease with which one can obtain exact results , thanks to the quadratic form of the hamiltonian describing the low energy properties , @xmath0 , \label{1dhamltn}\ ] ] where @xmath1 stands for the charge and spin sectors and @xmath2 and @xmath3 are bosonic fields satisfying @xmath4=-i\frac{\pi}{2 } \delta_{ij}{\rm sgn}(x - x')$ ] , and representing charge / spin density fluctuations and charge / spin current density fluctuations , respectively , and @xmath5 are the collective mode velocities . the parameters @xmath6 describe the interactions in the 1-d system . for repulsive interactions @xmath7 , and for su(2 ) symmetry in the spin sector @xmath8 . generally , the theory is applied under the assumption that both the spin and charge modes operate at low energy . but at low densities , @xmath9 , where @xmath10 ( with @xmath11 the bohr radius for the material ) , strong interactions open up a wide window between the characteristic energies of the charge and the spin sectors by suppressing the spin exchange energy @xmath12 while enhancing the charge energy @xmath13 . when the temperature @xmath14 is raised into this window of energy , @xmath15 ( where @xmath16 boltzmann s constant ) , the spin sector consists of thermally randomized spins that are no longer described by ( [ 1dhamltn ] ) while the charge sector effectively becomes a spinless luttinger liquid ( ll ) . a one - dimensional system in this regime is known as a spin - incoherent luttinger liquid ( sill).@xcite several properties of the sill have already been established theoretically,@xcite but the experimental effort has been slowed due to the difficulty of reaching the aforementioned window of energy . nevertheless , there are experimental indications of this regime in high quality quantum wires grown with the cleaved - edge technique,@xcite in recent split - gate devices,@xcite and possibly even in low - density carbon nanotubes.@xcite given the recent progress in manufacturing hybrid devices consisting of superconductors ( sc ) and ferromagnets ( fm ) , and one - dimensional systems such as nanotubes , we propose here that the properties of the sill be experimentally investigated by studying junctions of superconductors and ferromagnets with a sill . as we show below , these sill hybrid systems lead to distinctive behaviors ( relative to a ll or non - interacting 1-d system ) that will allow the sill to be identified in various types of transport measurements , and in the dynamics of a coupled ferromagnet . in early work on transport in the spin - incoherent regime , matveev found that a sill adiabatically connected to non - interacting leads resulted in a universal reduction of the conductance of a single - channel quantum wire from @xmath17 to @xmath18.@xcite this effect has been ascribed to the deactivation of the spin channel , whereby the spin excitations are reflected while the charge ones propagate freely . in a recent related work , the authors studied the problem of a sill adiabatically connected to two superconductors . it was found that the critical josephson current through such a sc - sill - sc systems suffers the same fate : the critical value of the josephson current is halved relative to the ll case.@xcite these universal results follow from the assumption of adiabaticity of the contacts between the sill and the leads / sc , which may not be satisfied in some experimental situations . in this work we study the transport properties of a sill in the opposite limit where it is contacted via tunnel junctions to a superconductor and/or a ferromagnet . we will discuss several scenarios that may be implemented experimentally with existing technologies . the central aim of this work is to contribute to our fundamental understanding of the sill by addressing its properties in hybrid structures , and to add to the arsenal of experiments that can be carried out to probe it and infer its existence . our main results are the following . for the case of a fm tunnel coupled to a sill , we compute the spin current pumped into the sill due to a time - dependent magnetization in the ferromagnet . the algebraic form of the expression for the pumped spin current is the same as that for non - interacting fermions , or a luttinger liquid , but the coefficients appearing have a temperature dependence characteristic of the sill . we argue the spin transport is diffusive and under this assumption compute the renormalization ( due to the pumped spin current and its back - flow ) of the gilbert damping @xmath19 of the magnetization motion in the fm . this also has a characteristic temperature dependence in the spin - incoherent regime . for the case of a sc tunnel coupled to a sill , we evaluate the ac and dc josephson current in a wire that can be either side - coupled ( bulk - coupled ) or end - coupled to the sc . we find the dc josephson current suffers an exponential decay in real space due to the incoherence of the spin sector ( in contrast to the inverse length suppression that was found for adiabatically contacted scs@xcite ) . finally , we show that the ac josephson effect also contains important information about spin - incoherent effects in its voltage and temperature dependence . this paper is organized as follows . in sec . [ sec : bosonization ] we review bosonization for a strongly interacting electron system for energy scales small compared to the charge energy , but arbitrary compare to the spin energy . in sec . [ sec : fm ] we study spin pumping and magnetization dynamics in the fm - sill system . in sec . [ sec : sc - sill ] we study ac and dc josephson effects in the sc - sill - sc system . finally in sec . [ sec : discussion ] we discuss prospects for experimental realizations of the physics we discussed and directions for future work . some technical details and results appear in various appendicies . in this paper we will be studying a strongly interacting one dimensional electron system coupled to either a ferromagnet or a superconductor . strong interactions imply@xcite that @xmath20 and therefore it is possible to be in a regime of temperatures that may not be small compared to the spin energy @xmath12 , but are still small compared to the charge energy @xmath13 . this means that we are no longer free to use standard bosonization procedures@xcite for the electron operator because these formulas rely on the assumption that the relevant energy scales on which the system is probed are small compared to both spin and charge energies . for a strongly interacting system we require a formalism that applies to all energy scales relative to the spin energy ( and in particular to @xmath21 ) , but only low energy scales to relative to the charge energy . such a formalism was recently developed by matveev , furusaki , and glazman,@xcite and applied to the evaluation of the spectral function and tunneling into a strongly interacting electron system at arbitrary temperatures relative to the spin energy.@xcite in this section we review the essential elements of their work and show how to represent the electron operator in a strongly interacting one dimensional electron system . the general expressions for electron correlation functions obtained in this way provide a launching point for numerical studies in the intermediate ( with respect to spin ) temperature regime @xmath22 , which is not well handled by existing analytical methods . as is typical of interacting one dimensional electron systems , we assume the hamiltonian in the strongly interacting case is spin - charge separated,@xcite @xmath23 . here , @xmath24 is identical to that of the ll , and is given by eq . . on the other hand , the spin hamiltonian at arbitrary temperatures is to a very good approximation given by a nearest neighbor antiferromagnetic heisenberg spin chain,@xcite @xmath25 where evidently the spin energy is set by @xmath26 : @xmath27 . the basic idea is to represent the electron operator as a product of operators that describe the holons @xmath28 ( spinless fermions that naturally arise in the context of strongly interacting fermions and the spin - incoherent regime@xcite ) and the spin degrees of freedom @xmath29 . the holon operators ( denoted by @xmath30 ) by construction satisfy the equation @xmath31 where @xmath32 is the electron annihilation operator for electrons of spin projection @xmath33 , and @xmath34 the corresponding electron creation operator . the issue of how to bosonize the electron operator for a strongly interacting system earlier arose in the context of the large @xmath35 limit of the one dimensional hubbard model . et al._@xcite wrote the electron creation operator as @xmath36 where @xmath37 creates a site on the spin chain with spin projection @xmath33 . the expression can be physically motivated as follows . from it is clear that the creation of an electron is also accompanied by the creation of a holon . however , electrons also carry spin so there must be a component of the electron operator that also creates spin . this is accomplished by @xmath37 . in general , one has @xmath38 as the object that adds a new site to the spin chain between @xmath39 and @xmath40 . while this appears physically intuitive , the expression suffers from the drawback that it does not naturally account for the variation of electron density with position in a real electron gas.@xcite matveev , furusaki , and glazman showed@xcite that a remedy for this issue is to define the position at which the spin site is added to the chain in terms of the number of holons to the left of the site , @xmath41 in terms of the electron creation and annihilation operators are defined as @xmath42 the operators given above explicitly account for the fact that the spins are attached to electrons , and the formulas are valid at all energy scales . it is perhaps worth noting that even though the hamiltonian is spin - charge separated , the electron operators are not written as a product of a spin piece and a charge piece because the spin " pieces @xmath43 also depend on the electron density via . in writing eqs . , no assumptions have been made about the energy scale relative to the spin and charge energies . we now restrict our considerations to energies small compared to @xmath13 , but arbitrary with respect to @xmath12 . in this case , we are free to bosonize the holon sector . the hamiltonian for the holons must then necessarily take the low energy form@xcite @xmath44 , \label{eq : h_holon}\ ] ] where the interaction parameter @xmath45 of the holon is related to the interaction parameter of the charge sector as@xcite @xmath46 , and the spinless fields @xmath47 and @xmath48 can be related to the holon density as@xcite @xmath49,\ ] ] where the holon fermi wave vector is twice the electron fermi wavevector@xcite @xmath50 . the bosonic fields satisfy the commutation relations @xmath51=i\pi\delta(x - y)$ ] . since we are interested in low energies with respect to the charge energy , the electron operator may be expanded about the two holon fermi points at @xmath52 , @xmath53 where @xmath54 destroys an holon near the right fermi point and @xmath55 destroys an electron near the left fermi point . the left and right holon operaters are bosonized as @xmath56},\ ] ] where @xmath57 is a short distance cut - off of order the interparticle spacing @xmath58 . combining the results of , , , and one obtains the bosonized form of the electron annihilation operator for spin @xmath33 @xmath59}+e^{-i[k_f^hx+\theta(x)]}\right)\nonumber \\ \times z_{l , s}\bigg |_{l=\frac{1}{\pi}[k_f^hx+\theta(x)]},\end{aligned}\ ] ] and an analogous expression for the electron creation operator @xmath60 . expression , however , it not quite complete as it does not account for the discreteness of the charge of the electron . this can be accomplished by interpreting @xmath61 } \to \sum_l z_{l , s}\delta\left({1 \over \pi}[k_f^hx+\theta(x)]-l\right),\ ] ] after which the full electron annihilation operator ( include both left and right moving parts ) becomes@xcite @xmath62},\ ] ] where @xmath63 the expression for the electron annihilation operator is the key result of the this section , obtained earlier by matveev , furusaki , and glazman , who also showed that in the limit of small energies compared to @xmath12 the expression correctly recovers the standard ll formulas for the electron annihilation operator.@xcite with correlation functions involving electron operators can be expressed in terms of the correlation functions of the holon and spin sectors at arbitrary temperatures with respect to @xmath12 , but small energies compared to @xmath13 . see appendix [ app : spin_current ] and appendix [ app : andreev_tdos ] for examples of the evaluation of the single particle green s function near different types of boundaries using the bosonization formula and the boundary conditions to be discussed next . in this work we will be primarily interested in two types of boundary conditions on the electron operators : ( 1 ) open or hard wall " boundary conditions appropriate for tunnel junctions and ( 2 ) andreev boundary conditions appropriate for adiabatically contacted ( no electron backscattering ) superconductors . other , more general , intermediate " boundary conditions are possible though less generic.@xcite to help orient the reader , in each of the two cases above we will briefly summarize the result appropriate for the ll regime for the purpose of drawing contrast with the strongly interacting , finite temperature sill regime . our discussion of the boundary conditions extends the results of ref . [ ] . open boundary conditions , or hard wall " boundary conditions result in electron waves that are perfectly reflected at the boundary . this implies there is no charge current or spin current through the boundary . for concreteness , let us assume that our boundary is located at @xmath64 with the interacting one dimensional system living on @xmath65 . then an electron traveling to the left with spin @xmath33 will be reflected to a right moving electron with the same spin @xmath33 @xmath66 with a phase shift @xmath67 that depends on details of the boundary scattering potential.@xcite from the standard ll bozonization formulas ( in our convention ) @xmath68},\ ] ] we see that implies that @xmath69 , and therefore also that @xmath70 and @xmath71 . in computing correlation functions at the boundary , this means that we must take @xmath72 and @xmath73 to be non - fluctuating quantities , allowing only @xmath74 and @xmath75 to fluctuate . in the strongly interacting case , the reflection condition implies that @xmath76 , where we have used . since the spin site @xmath40 is related to the @xmath47 field via and we have @xmath77 at @xmath64 . this implies any correlation function involving electron operators evaluated at the boundary will depend only on @xmath78 or its hermitian conjugate , and the fluctuations of the field @xmath48 at @xmath64 . we will apply these boundary conditions in a calculation of the single particle green s function in appendix [ app : spin_current ] . for completeness , below we give the expansions of the bosonic fields for a system of finite length @xmath79 with open boundary conditions at @xmath64 and @xmath80:@xcite @xmath81 where @xmath82 , @xmath83=\delta_{nm}$ ] , and @xmath84=i$ ] . for a semi - infinite system , we take @xmath85 and the discrete sums over @xmath86 become integrals over momentum @xmath87 . in a certain sense the andreev limit is the opposite limit of hard wall " scattering from a sc - m , or sc - ll , or sc - sill interface . whereas open boundary conditions imply an incident electron is perfectly reflected , andreev boundary conditions imply that an electron is perfectly absorbed by the sc ( with a concomitant reflected hole of the opposite spin).@xcite however , compared to the open boundary conditions the andreev boundary conditions are more subtle as they involve an energy scale , the superconducting gap @xmath88 , that under many circumstances can not be taken to be infinitely large relative to the energy scales of interest ( such as @xmath89 , or an applied voltage ) in the interacting one - dimensional system . one of the most important consequences of a finite @xmath88 is an energy - dependent reflection coefficient , which ultimately leads to the proximity effect in the normal material.@xcite in the context of interacting one dimensional systems , andreev boundary conditions have been discussed by a number of authors.@xcite the conclusion of these works is that when a left moving electron with spin @xmath33 is reflected as a right moving hole with the opposite spin there is an energy dependent phase shift @xmath90 ( proportional to the momentum difference @xmath91 with respect to the fermi point)@xcite that multiplies a factor@xcite @xmath92 that encodes the phase @xmath93 of the superconducting order parameter ( assumed non - zero for @xmath94),@xcite @xmath95 where @xmath96 is the superconducting coherence length . the function @xmath97 for @xmath98 and @xmath99 for @xmath100 . the boundary conditions are valid only at energy scales much smaller than @xmath88 , which we will assume throughout this work . applying the boundary conditions to the ll case ( with bosonized electron operator below ) gives @xmath101 and @xmath102 . therefore , we find that much like the situation of perfect reflection there is no spin - current through the interface , but there is is a net charge current . moreover , analysis of singlet superconductivity and spin density wave correlation functions using in the ll regime shows that there are suppressed spin fluctuations near ( distances less than @xmath103 ) the interface.@xcite it is perhaps worth noting that if the sc - ll interface has a very weak electron backscattering , the interactions in the ll tend to renormalize the interface scattering.@xcite for strongly interacting electrons the boundary condition implies @xmath104 and @xmath105 at the interface , where we have again used and also . we show in appendix [ app : andreev_tdos ] that we recover our earlier results@xcite for sc - sill junctions in the andreev limit using this formalism . finally , for completeness and for unification of our notation , below we give the expansions of the bosonic fields@xcite for andreev boundary conditions on a system of length @xmath79 ( see fig . [ fig : sc_tunnel ] ) with superconducting order parameter phase difference @xmath106 : @xmath107(b_m + b^\dagger_m)+\theta^0 , \nonumber \\ \phi(x)&=&i\sum_{m=1}^\infty \sqrt{\frac{1}{2k_\rho m}}\sin\left[\frac{m \pi } { l}\left(x+{\xi \over 2}\right)\right](b_m -b^\dagger_m)\nonumber \\ & & \hspace{5cm}+\phi(x),\end{aligned}\ ] ] where @xmath108 , @xmath83=\delta_{nm}$ ] , and @xmath109=i$ ] . the topological number @xmath110 and the total spin of the system @xmath111 must satisfy the constraint @xmath112even.@xcite we have also included the proximity effects@xcite via the length @xmath103 in the field expansions that appeared earlier in the general expression for the boundary conditions appropriate to andreev reflection.@xcite note that compared to the field expansions for open boundary conditions , the expansions for andreev boundary conditions have @xmath47 and @xmath48 switched " . this can be understood in simple physical terms : @xmath47 and @xmath48 are conjugate fields so if one is a constant , the other is strongly fluctuating . thus , the switching " of the fields comes from opposite " nature of the two boundary conditions in the charge sector . for open boundary conditions there is no charge current through the boundary , while for andreev boundary conditions the charge current on either side of the boundary is unchanged by its presence because there is no electron backscattering there . had we been concerned with the bosonization of the spin sector , we would have found that due to the absence of spin current through the boundary that sector would have expansions appropriate for open boundary conditions similar to .@xcite having spent the time to develop the formalism used in our calculations , we now turn our attention squarely to the physics of hybrid junctions involving spin - incoherent luttinger liquids . we first discuss a junction consisting of a ferromagnet tunnel coupled to a spin - incoherent luttinger liquid . we consider the set - up shown schematically in fig.[fig : fm ] whereby a ferromagnet is coupled via a tunnel juntion to a spin - incoherent luttinger liquid . a similar situation was considered by bena and balents in the context of a fm - ll junction.@xcite they found that when the magnetization @xmath113 of the fm acquires a time dependence ( perhaps by the application of external fields ) , spin current is pumped into the adjoined ll at the rate @xmath114 where @xmath115 and @xmath116 are temperature dependent parameters depending on the interface tunneling and interactions in the ll . below we show that an identical expression is obtained for a sill and we derive explicit expressions for @xmath115 and @xmath116 in this case . we schematically indicate how the temperature dependence crosses over from the ll to the sill case . we note that was earlier derived for a non - interacting metal attached to a fm . in this case , @xmath115 and @xmath116 are temperature independent to lowest order.@xcite thus , the temperature dependence of @xmath115 and @xmath116 encodes information about the interactions and may also indicate whether the attached 1-d system is in the ll or sill regime . we now turn to a derivation of these results . we assume that the ferromagnet is itinerant , such as fe or co , and can be described by a stoner - type model in which there is a different density of states for spin - up and spin - down electrons . the different density of states can be absorbed into distinct effective tunneling matrix elements , @xmath117 , for spin - up and spin - down electrons.@xcite this also implies that the local action@xcite describing the spin up and spin down electrons are identical to that of non - interacting electrons @xmath118 where @xmath119 is the inverse temperature and @xmath120 are the bosonic fields describing local fluctuations at the tunneling point , @xmath64 , of spin up and spin down electrons in the fm . for the sill we have the following spin - charge separated form of the local action @xmath121 where @xmath122 is the luttinger parameter of the charge sector of the sill , @xmath123 is the bosonic charge field , and @xmath124 is the action for the spin sector which will actually play no role in the evaluation of correlation functions in the spin - incoherent regime as it effectively drops out leading to super universal spin physics".@xcite for an arbitrary spin quantization axis relative to the magnetization @xmath125 it is useful to define @xmath126 which projects the spin quantization axis onto the magnetization direction @xmath127 . here @xmath128 is a @xmath129 matrix and @xmath130 are the pauli spin matricies . the tunneling hamiltonian then takes the form@xcite @xmath131 where @xmath132 ( @xmath133 ) annihilates / creates an electron in the fm and @xmath134 ( @xmath135 annihilates / creates an electron in the sill . the tunneling is assumed to occur at @xmath64 as shown in fig.[fig : fm ] . the @xmath136 tunneling matrix @xmath137 is then @xmath138 the tunneling hamiltonian can be expressed more explicitly as @xmath139 where @xmath140 is the time - dependent magnetization of the ferromagnet , and @xmath141 and @xmath142 . ( from here onwards we will assume @xmath143 . ) the spin current operator is obtained from the relation @xmath144 $ ] , where @xmath145 is the spin density in the sill at the boundary and summation over repeated spin indices is understood . it follows then that the spin current operator is given by @xmath146 the spin current , @xmath147 \rangle$ ] , is obtained@xcite from second - order time dependent perturbation theory which gives @xmath148-\int \frac{d \omega}{h } c(\omega)e^{-i\omega t}\times\nonumber\\\left(\text{im } ( u_2^*u_1)\vec{m}(\omega)+\frac{|u_2|^2}{2}\vec{m}(\omega)\times\vec{m}(t)\right),\end{aligned}\ ] ] where @xmath149 is the fourier transform of the retarded greens function @xmath150\rangle$ ] . if we assume the typical frequencies of the magnetization precession are in the ghz range@xcite ( which corresponds to energies of roughly 100 mk ) or smaller , then this is a small energy scale in the problem and we may safely expand @xmath149 for small @xmath151 . it can be easily checked that the zero frequency terms cancel exactly , leaving only the contributions linear in @xmath151 , provided we drop terms proportional to @xmath152 and all higher powers . upon integration over frequency , the linear frequency terms are converted to time derivatives yielding expression where the coefficients @xmath115 and @xmath116 are given by @xmath153 since only the imaginary part of @xmath149 is odd with respect to frequency , only this piece will contribute to @xmath154 and those parameters will therefore be real quantities . the final step is to compute the temperature dependence of @xmath155 in the spin - incoherent regime . since we are working within linear response , the commutator in @xmath156\rangle$ ] is evaluated in the state where there is no tunneling between the ferromagnet and the sill , and a hard - wall at @xmath64 . by writing @xmath157 $ ] , and noting that the long time behavior of both terms has the same functional dependence , we can easily extract the long time behavior@xcite of @xmath158 using and . ( a more careful calculation that yields the same result is given in appendix [ app : spin_current ] . ) at finite temperatures ( small with respect to the fermi energy of the ferromagnet and charge energy of the sill , but large compared to @xmath159 ) , we have @xmath160 and @xmath161^{1/(2k_\rho)}$ ] . the crucial difference with the luttinger liquid is that the exponent for the correlations @xmath162 have changed : in the case of a ll @xmath163 is replaced by @xmath164 . the remainder of the computations carry through exactly as they would for a ll and we find for @xmath165 @xmath166 \propto \hbar \omega \left(k_b t\right)^{\delta^{sill}},\ ] ] where @xmath167 . this implies @xmath168 quite generally so that @xmath169 where @xmath170 interpolates between the ll and sill regimes as the temperature is swept through the spin energy @xmath171 . the temperature dependence of the exponent @xmath170 is shown schematically in fig . [ fig : delta ] . it is interesting to note that @xmath172 for @xmath122 in the luttinger liquid regime implying that less spin current is pumped as the temperature decreases . on the other hand , so long as the system remains in the spin - incoherent regime the opposite behavior may be obtained if @xmath173 : since @xmath174 _ more _ spin current is pumped as the temperature is lowered ( for @xmath175 ) . this is related to the diverging density of states at the boundary@xcite when @xmath173 . in gated cleaved - edge overgrowth quantum wires it appears possible to lower @xmath176 down to values of order @xmath177 and so it may be possible in the spin - incoherent regime to tune between the positive and negative exponent regimes.@xcite this qualitative difference should be easily seen in experiment . having computed the spin current pumped into the sill by a time - dependent magnetization vector , it is important to ask how the spin accumulation in the sill in turn affects the magnetization motion . we address this question by computing the renormalization of the gilbert damping constant due to the spin flow into the sill . we again closely follow the notation of bena and balents@xcite to clearly establish a connection to the ll case . the landau - lifshitz - gilbert equation for magnetization @xmath178 precessing around effective magnetic field @xmath179 is @xmath180 where the spin current @xmath181 flows ( @xmath182 is given by for @xmath183 and is non - zero only for time dependent @xmath178 ) into the sill and @xmath184 is the saturation magnetization of the ferromagnet . here @xmath185 is the gyromagnetic ratio which is typically equal to its free electron value,@xcite @xmath186 , in transition metal ferromagnets , and @xmath187 is the dimensionless gilbert damping parameter in the absence of the spin current . its value is typically of order @xmath188 . the spin backflow due to spin accumulation in the interacting 1-d system can be described by boundary conditions on left and right moving spin currents and is expressed as@xcite @xmath189 where @xmath190 is the spin chemical potential in the wire related to the magnetization by @xmath191 , with @xmath192 the ( generally temperature dependent ) spin susceptibility , and @xmath193 describes the effective exchange coupling between electrons in the sill and the fm.@xcite the current @xmath194 arises from the electron tunneling contribution to the spin current.@xcite note that the back - scattered spin current contains two terms : ( i ) a term arising from electron transfer from the fm to the interacting 1-d system and ( ii ) a second term arising from exchange between the local spin density and the precessing magnetization vector @xmath178 . since @xmath181 this implies the spin current itself has a contribution due purely to exchange effects , even in the absence of electron transfer.@xcite our goal in this section is to express as @xmath195 and determine the renormalized parameters @xmath196 and @xmath19 . from it is evident that we must find @xmath197 . we have already computed @xmath182 , eq . , in the previous subsection . now we must determine @xmath198 , , which depends on @xmath190 and @xmath194 . we start with @xmath190 which is a function of the spin dynamics in the interacting 1-d system . it has earlier been shown that su(2 ) invariant electron backscattering leads to diffusive spin behavior in the weakly interacting regime.@xcite in the strongly interacting regime where the spin sector may effectively be described by a heisenberg spin chain coupled to phonon distortions@xcite the spin dynamics has also been shown to be diffusive.@xcite in both limits , the mean free path @xmath199 . following the arguments@xcite that the spin - incoherent regime can be approached from below " ( @xmath200 ) , we expect the diffusion length @xmath40 to saturate at the interparticle spacing , @xmath58 , for @xmath175 . since the diffusion constant @xmath201 , this implies the diffusion constant in the sill becomes independent of temperature and takes the value @xmath202 . note that in the limit @xmath203 , there is no spin diffusion , as the spin excitations can not propagate in the system . spin will simply pile up at the boundary of the fm without moving further into the interacting 1-d system . unlike our earlier results for the temperature dependence of pumped spin into the sill , eq . with @xmath204 , which were valid for vanishing spin velocity , here our results are qualitatively dependent on keeping @xmath205 finite , although still small enough to be in the spin - incoherent regime . we also assume the finite length sill is characterized by a spin - flip time @xmath206 ( due to impurities , spin - orbit effects , etc ) . the diffusion equation for spin in the sill is then @xmath207 with the boundary conditions @xmath208 at @xmath64 and vanishing spin current @xmath209 at @xmath80 . see fig . [ fig : fm ] for the set - up . the solution to this equation is simple to obtain and is given by @xmath210}{\kappa \sinh[\kappa l]}\left(\frac{1}{d_s \chi(t)}\right ) \vec i(x=0),\ ] ] where @xmath211 . if the precession frequency @xmath151 is small compared to the inverse spin relaxation time , then to a good approximation @xmath212 . at the boundary @xmath64 we have @xmath213 where @xmath214 \frac{1}{\kappa d_s \chi(t)}$ ] . note that @xmath215 so the spin current decays exponentially with distance into the sill on a length scale set by the product of the diffusion constant and the spin - relaxation time . for a long spin - relaxation time , this length scale can be large compared to the inter - particle spacing . we have now determined all parts of the spin back flow , except the tunneling current @xmath194 , which we now do . the tunneling current is proportional to the imaginary part of the fourier transformed correlation function@xcite @xmath149 defined below eq . , @xmath216\nonumber \\ \propto |u_1|^2(k_b t)^{\delta(t)+1 } \sinh\left(\frac{\mu_s}{4 k_b t}\right ) \nonumber \\ \times \left|\gamma\left(1+\frac{\delta(t)}{2}+i\frac{\mu_s}{4\pi k_b t}\right)\right|^2 \nonumber \\ \propto \mu_s |u_1|^2(k_b t)^{\delta(t)}\left|\gamma\left(1+\frac{\delta(t)}{2}\right)\right|^2,\end{aligned}\ ] ] where we have used the results of appendix [ app : spin_current ] and taken the limit @xmath217 . evidently , the main effect of the spin - incoherent physics is to change the exponent @xmath170 to the spin - incoherent value , @xmath204 , so that we have @xmath218 in the spin - incoherent regime . let us define @xmath219 which is temperature dependent and @xmath220 independent . at this point the determination of @xmath196 and @xmath19 is identical to the ll case.@xcite we find @xmath221 and @xmath222 where @xmath223 $ ] , @xmath224 $ ] , and @xmath154 are given in eqs.([eq : a_gen],[eq : a_temp ] ) . the temperature dependence in @xmath225 is entirely contained in @xmath226 and @xmath103 . see table [ tb : ll_sill ] below for a comparison of the ll and sill regimes . [ h ] .temperature dependence of key quantities in ll - fm and sill - fm hybrid systems . here @xmath227 is a constant that depends on @xmath228 , and @xmath229 . [ cols="^,^,^,^,^,^",options="header " , ] as with the case of a ll , for a sill we expect there to be little renormalization of @xmath196 relative to @xmath185 , but the smallness of the gilbert damping @xmath187 means this may obtain a significant temperature dependent correction depending on @xmath230 . one could easily imagine other scenarios such as a fm - sill - fm junction , a fm - sill - m junction , or even a fm - sill - sc junction . however , because the spin transport is generally diffusive , for junctions whose length @xmath79 is long compared to @xmath231 , the two contacts to the sill will more or less behave independently from the point of view of spin transport . to the extent that the two leads are coupled , it is evident from the discussion of the previous section that most of the physics ( and corresponding general formulas ) present for a ll system also apply to the sill system only with some modifications in the temperature dependence of parameters appearing in the theory . while this may seem like a somewhat trivial result , it is not . the temperature dependence serves as a means to determine whether spin - incoherent physics is present in the system . in particular , we note quite generally that if @xmath232 the temperature dependence of many quantities ( e.g. , @xmath233 and those derived from them ) change _ qualitatively _ relative to the expectations for a ll . such qualitative differences should be observable in experiment . it is worthwhile to step back and emphasize some of the essential differences between the sill and ll cases . first , we note that the different temperature dependence can be traced to three quantities : ( 1 ) the correlation function @xmath158 defined below eq . . ( 2 ) the temperature dependence of the spin susceptibility @xmath93 . ( 3 ) the temperature dependence of the spin diffusion constant @xmath229 . so long as the sill is tunnel contacted to the fm , the correlation function @xmath158 will always appear at lowest order in perturbation theory and carry along with it the characteristic temperature dependence of the sill . this quantity appears in both the description of the pumped spin current and the magnetization dynamics . on the other hand , when we are specifically interested in how the spin propagates in the sill [ as we saw for the backscattered spin current ] the diffusion constant @xmath229 enters , and also the spin susceptibility via the einstein relation for the spin conductivity , @xmath234 . we remark that the spin transport is generically diffusive in the spin - incoherent regime , while it may be either ballistic or diffusive in the ll regime.@xcite having now flushed out what we feel are the central considerations and results for fm - sill hybrid structures , we now turn our attention to sc - sill hybrid structures . throughout this work , we assume that all energy scales are small compared to the superconducting gap , @xmath88 , unless explicitly stated otherwise , such as in the limit of a short junction ( defined below ) . we are primarily interested in the case where the superconductor is contacted via a tunnel junction with a spin - incoherent luttinger liquid . let us begin our discussion of sc - sill hybrid systems by considering the simplest case : where the fm in fig . [ fig : fm ] is replaced by a sc . earlier we studied the properties of such a junction in the andreev limit.@xcite we found a number of remarkable properties , including a completely universal ( independent of the charge and spin hamiltonians ) tunneling density of states in the spin - incoherent regime.@xcite in the opposite limit of a tunnel contact between the sc and sill , no proximity effects of the sc are felt other than those perturbative in the tunneling.@xcite ( in contrast to the andreev limit where the behavior of the pair correlations and tunneling density of states follows directly from the form of the field expansions imposed by the boundary conditions.@xcite ) as the proximity effects are already weak in the andreev limit ( they decay exponentially fast with distance , unlike the power law decay in the ll@xcite ) we do not pursue the very weak proximity effects in the spin - incoherent regime in the limit of a weak tunneling sc - sill junction . on the other hand , it is worthwhile to briefly discuss the behavior expected in opposite limit of nearly perfect andreev reflection with weak ( perturbative ) electron backscattering at the sc - sill interface . in the ll context , this was addressed by vishveshwara _ et al._@xcite in a sc - ll - m system who found dips in the conductance of the junction as a function of gate voltage . these dips correspond to multiple electron reflections off the sc interface and occur at voltages corresponding to integer values , @xmath235 , of the inverse electron traversal time , @xmath236 , where @xmath9 is an integer counting the number of traversals in the ll.@xcite for a sc - sill - m system we would expect similar qualitative effects , although with an overall suppression in the conductance.@xcite finally , we briefly remark that the conductance per channel of an adiabatic sc - sill junction should drop to 1/2 the value expected for a sc - ll junction:@xcite @xmath237 . recall that the conductance of the adiabatic sc - ll interface @xmath238 generalizes the result for the adiabatic sc - m interface@xcite @xmath239 realized for @xmath240 . we now turn our attention to the final topic of this work , sc - sill - sc junctions of the types shown in fig . [ fig : sc_tunnel ] . our primary focus will be on the behavior of the ac and dc josephson effects in such a hybrid structure . in the context of sc - ll - sc structures there have already been a number of theoretical works@xcite and some experimental results.@xcite earlier we studied the dc josephson effect in a junction like that shown in fig . [ fig : sc_tunnel](a ) only with andreev boundary conditions at the interfaces of the sill with the scs.@xcite we found that in spite of the fact that the pair correlations decay exponentially with distance from the boundary of the sc , the josephson critical current scaled inversely with the length of the sill and was reduced by a factor of 2 relative to either the sc - ll - sc or sc - m - sc case ( which have identical critical currents that also scale inversely with the length).@xcite we remarked that this robustness " is essentially related to the fact that the superconductor phase difference only couples to the charge degrees of freedom in the sill , which remain completely coherent . this effect can also be taken to support the notion@xcite that in some sense superconductors are spin - charge separated.@xcite yet another perspective on this result can be obtained by noting that the andreev boundary conditions on a system of finite length can be mapped onto a system with _ periodic _ boundary conditions of twice the length.@xcite in this way , the josephson response becomes equivalent to the problem of persistent currents for _ spinless _ electrons in a ring threaded by magnetic flux , with the superconductor phase difference being simply related to the flux . therefore , the dc josephson current for andreev boundary conditions is essentially given by the physics of _ spinless _ electrons which is why spin - incoherent physics does not manifest in a dramatic way . the factor of 2 reduction in the critical current and josephson period are a direct consequence of the mapping of the sill onto spinless electrons . we now turn our attention to the case of open boundary conditions at the interfaces between the scs and sill . in this case phase information must be carried between the two superconductors by cooper pair tunneling between them . because the _ pairs _ must hop from one sc onto the sill , then onto the other sc , the josephson current will occur at 4th order in this tunneling process . moreover , because the cooper pair tunneling process creates a pair of electrons locally in the sill near the barrier , many fourier modes of the electron operator will come into play . this is in contrast to the case of andreev boundary conditions where only the zero mode enters in the evaluation of josephson current.@xcite this crucial difference will turn out to dramatically suppress the josephson response relative to what one would have for a ll when the sill is tunnel coupled to scs . in short , we find the critical current decays as an exponential of the junction length ( compared to a power of the inverse junction length for a ll at zero temperature ) with the characteristic decay length the inter - particle spacing . the sc - sill - sc tunnel junctions shown in fig . [ fig : sc_tunnel ] are modeled by the following hamiltonian , @xmath241 where @xmath242 are @xmath33-wave bcs hamiltonians for the superconductors and @xmath243 is the sum of eqs . and . the tunneling hamiltonian @xmath244 is given by @xmath245 we assume the scs support a phase difference of @xmath246 and have the same superconducting gap @xmath88 . the josephson current is obtained from the relation @xmath247 where @xmath248 is the electronic charge and @xmath249 is the partition function . we employ imaginary time perturbation theory , where terms dependent on @xmath93 appear in fourth order,@xcite @xmath250 where @xmath251 is a two - particle ( cooperon ) propagator in the one - dimensional interacting system . the similar terms " account for time ordered permutations and spin projections ( there are 23 of them , @xmath252 all together ) . the propagation of cooper pairs in the superconductors is described by the time - ordered anomalous green function @xmath253 , @xmath254 where the expectation value is taken with respect to @xmath255 and the electron operators in are evaluted at @xmath64 for @xmath256 and @xmath80 for @xmath257 . of the 24 terms ( plus their hermitian conjugates ) that appear in the partition function , eq . , physical considerations aid us in choosing the most relevant ones depending on the length of the sc separation distance @xmath79 . there are two important length scales in the junctions we study that give rise to corresponding time scales : the superconducting coherence length @xmath103 and the junction length @xmath79.@xcite we define long " to mean @xmath258 and short " to mean @xmath259 . we now turn to a discussion of the general expression for @xmath251 in the limit of long and short junctions . for long junctions the tunneling into the sill is fast " and the propagation of the cooper pairs is slow " . the relevant two - body propagator is @xmath260 where @xmath32 is the bosonized electron annihilation operator given by eq . and the averaging is taken with respect to @xmath243 . we have @xmath261 , @xmath262 , and @xmath263 . in the opposite limit of a short junction @xmath259 , the propagation through the sill is fast " compared to the slow " tunneling of the cooper pairs , so the two - particle propagator approximately separates into a product of two single - particle green s functions@xcite @xmath264 where we have @xmath261 , @xmath262 , and @xmath265 ( note the re - ordering of electron operators relative to ) . _ end contacted sill when @xmath266 _ : we first consider the case of an end - contacted " sill , shown in fig.[fig : sc_tunnel]a . for this geometry the expansions of the holon field operators are the appropriate ones . our assumption that all energy scales are small compared to @xmath88 ( specifically @xmath267 , and @xmath268 ) enables us to approximate the anomalous green s function with a delta function in time , @xmath269 where @xmath270 is the density of states of the sc @xmath271 at the fermi energy in its normal state . this approximation forces the times the electrons tunnel into or out of the sill to coincide , @xmath272 and @xmath273 . without loss of generality , we let @xmath274 and define @xmath275 . using the expression for the electron operator , eq . , the propagator reads @xmath276 the calculation of the cooper pair propagator is carried out in the same manner as that of the green s function presented in appendix [ app : spin_current ] . we first evaluate the expectation value of the spin chain site creation / annihilation operators , @xmath277 in the spin - incoherent regime . to this end it is convenient to fully exploit the symmetries and boundary conditions of the problem . we first note that the open boundary conditions at @xmath278 force @xmath279const at both these points . as before , we take this constant to be zero so that all @xmath47 fields effectively drop out of eq .. the momentum integrals in can then be done trivially to give delta functions on the sites @xmath280 which sets @xmath281 and @xmath282 . therefore , we have @xmath283 which in the spin - incoherent regime can be evaluated quite simply by appealing to the physics of the spins . here @xmath284 measures the amplitude for two spins of opposite orientation introduced at @xmath80 to arrive at @xmath64 . deep in the sill regime where all the spins are randomized and spin exchange is highly suppressed , this is just the probability of finding neighboring sites of the spin chain housing oppositely aligned spins , @xmath285 from which it follows that @xmath286 the remaining correlations over the charge degrees of freedom can be computed using the expansions and the identity @xmath287/2}$ ] valid when @xmath288 $ ] commutes with both @xmath289 and @xmath290 . the final result is @xmath291 where @xmath292 is a short distance cut - off that ensures ultra violet convergence of the integrals over the charge fluctuations in the wire . up to unimportant phase factors and other overall multiplicative constants , the josephson current in the wire is then @xmath293 where the integral has been cut off at times @xmath294 , the time it takes for charge to propagate between the two ends of the sill . compared to the ll case , the most important difference is that the critical current scales as an exponential of the length , @xmath295 rather than a power law @xmath296 in the ll regime.@xcite a measurement of the length dependence of the critical current then serves as clear signature of spin - incoherent physics ( or lack thereof ) , provided one has some knowledge of the density to infer @xmath297 . we also note that the superconducting phase difference @xmath93 appears as an argument to the @xmath298 function , rather than the saw - tooth form we found for andreev contacts.@xcite the @xmath298 form follows from the assumption of tunneling contacts.@xcite _ bulk contacted sill when @xmath258 _ : we now consider the case of a bulk - contacted " sill , shown in fig.[fig : sc_tunnel]b . for this case , we assume that the perturbations induced by the contacts are irrelevant , which is so for @xmath299.@xcite otherwise , if the perturbations induced by the contacts are relevant this case reduces to an end - contacted " sill discussed above . starting again with we note that translational symmetry implies that the spin correlations satisfy @xmath300 which allows us to shift the sums : @xmath301 in . the summation over @xmath302 leads to a delta function setting @xmath303 and the summation over @xmath304 leads to a delta function setting @xmath305 . making the change of variables @xmath306 and @xmath307 one can do the integration over @xmath308 which sets @xmath309 . the result is @xmath310-[\phi(0,\tau)-\phi(l,0)]\right\ } } \rangle . \label{eq : pi_bulk_gen}\end{aligned}\ ] ] the expression is valid quite generally under the assumption that the contacts are irrelevant perturbations and do not lead to the end contacted " result described earlier . if we now specialize to the spin - incoherent case we have @xmath311 substituting into and integrating over momentum , @xmath312\nonumber \\ \times e^{i2\left\{\left [ \theta(0,\tau)-\theta(l,0)\right]-[\phi(0,\tau)-\phi(l,0)]\right\}}\rangle . \end{aligned}\ ] ] finally recalling that for @xmath313 we can replace the sum over @xmath40 by an integral,@xcite we find @xmath314 \ln2 } e^{i2\left\{\left [ \theta(0,\tau)-\theta(l,0)\right]-[\phi(0,\tau)-\phi(l,0)]\right\ } } \rangle , \nonumber \\\end{aligned}\ ] ] which can be evaluated ( up to phase factors ) to give @xmath315 where @xmath316.$ ] the josephson current in the wire is then @xmath317 where we have again cut off the integral at times @xmath318 . evidently , the critical current again scales as an exponential of the length , @xmath319 rather than a power law@xcite @xmath296 in the ll regime . this luttinger liquid result corrects the result originally obtained by fazio _ et al._@xcite by taking into account the proximity effect . comparing the critical currents of the end contacted case with the bulk contacted case , one sees that the length dependence is not materially different : they only differ in the inconsequential power law that multiplies the exponential . _ end contacted sill when @xmath259 _ : to compute the josephson current we must evaluate the single - particle green s functions that appear in . since @xmath103 is a property of the sc which we assume is tunnel contacted to the sill , it is independent of the properties of the sill . we will further assume @xmath320 . the opposite limit of the wire length being shorter than the inter - particle spacing is not well motivated physically . the single - particle green s functions appearing in can readily be evaluated following the method of appendix [ app : spin_current ] . this gives @xmath321}\rangle , \nonumber \\ & = & \frac{1}{2\pi \alpha_c } e^{-\frac{k_f^h l}{\pi}\ln 2 } \left(\frac{\eta}{1+e^{-\eta}e^{-v_\rho \pi \tau / l}}\right)^{\frac{1}{4k_\rho}},\end{aligned}\ ] ] which immediately leads to the josephson current @xmath322 ^ 2 , \label{eq : endjoscurntshort}\end{aligned}\ ] ] with a critical current @xmath323 . _ bulk contacted sill when @xmath259 _ : for a bulk contacted sc in the short wire limit we again apply the formula where the green s functions are those appropriate for an infinite system , computed earlier in the literature in ref . [ ] . the result is @xmath324 ^ 2 , \label{eq : bulkjoscurntshort}\end{aligned}\ ] ] which evidently has the critical current @xmath325 , where @xmath326 $ ] . comparing the results for bulk and end contacted , as well as short verses long wires , we find the most important result is how the josephson critical current scales with the length of the wire . suppressing the relatively unimportant power law decay that multiplies the dominant exponential decay we find @xmath327 so that the decay of critical current with the length of the sill is twice as fast as for a short wire . this originates in the fact that for a short wire electrons propagate independently so the spin incoherence affects each electron independently , rather than a single coherent pair as in the case of the long wire . it is also worth emphasizing that the assumption of tunneling contacts always results in a josephson current that is proportional to the product of the bare conductances of the two contacts with a sinusoidal dependence on the phase difference of the superconducting order parameter , @xmath328 . therefore it is the length dependence of the critical current that provides a smoking gun signature of spin - incoherence in the dc josephson effect with tunnel contacts . for andreev ( adiabatic ) contacts , the length dependence is identical to those of a ll or non - interacting one dimensional system.@xcite instead , the sill physics is manifest clearly in the flux dependence of the josephson current which takes on a saw - tooth form of half the usual period.@xcite the ac josephson effect occurs when there is a finite voltage @xmath329 across the sc - sill - sc system . the josephson phase acquires a time dependence @xmath330 leading to a josephson current that oscillates in time . a sub - gap dissipative current is also induced , but at small voltages this can be estimated to be small.@xcite as in the case of the dc josephson effect , the qualitative features of the ac josephson effect also depend on whether the wire is short or long , as defined in the dc case . _ long wire case _ : we have seen that the qualitative features of the dc josephson current in the spin - incoherent regime for bulk and end contacts are not very different . that is also true of the ac josephson effect . the ac josephson current can be computed from@xcite @xmath331,\end{aligned}\ ] ] where @xmath332 is the cooper pair propagator , eq . , evaluated at real times . the ac josephson current can be decomposed into sinusoidal and cosinusoidal components , @xmath333,\ ] ] where @xmath334,\;\;\;\;\\ \label{eq : j_c } j_{c}&=&{\rm re}\left[\int_0^\infty dt ' \left(e^{-i evt ' } \pi(t ' ) - e^{+i evt ' } \pi(t')\right)\right].\end{aligned}\ ] ] in the luttinger liquid regime @xmath335 and @xmath336 oscillate with voltage across the junction , as does the amplitude @xmath337.@xcite the frequency of the oscillations of @xmath338 with voltage depends on the relative spin and charge velocities . one period occurs when the spin and charge parts differ by @xmath339 , that is when @xmath340.@xcite in the spin - incoherent regime , we have @xmath203 , implying vanishing voltages will lead to oscillations and they may cease to be observed . the lack of amplitude oscillations will persist into the spin - incoherent regime , which has effectively only one velocity , the charge velocity . on the other hand , if @xmath205 is not too different from @xmath341 ( say , @xmath342 ) , then one can expect to find a temperature dependence of the voltage oscillations that reveals spin - incoherent physics in a way analogous to coulomb drag@xcite or charge fluctuation noise:@xcite at temperatures below the spin energy , there will be amplitude oscillations in the josephson current as a function of voltage , while for temperatures above the spin energy there will be no such oscillations because the spin mode effectively does not propagate . this washing out " of the high frequency ( because of the ratio @xmath343 ) oscillations with temperature is the signature of spin - incoherent physics in the ac josephson effect . _ short wire case _ : for a short wire , the voltage dependence is independent of the properties of the 1-d system,@xcite be it a ll or a sill . in this case the ac josephson current is @xmath344 where @xmath345 is the zero voltage critical current for the short wire given in sec . [ sec : dc - josephson ] , and @xmath346 is an elliptical integral . recall that for a short wire , @xmath347 . thus , the ac josephson effect is only effective at revealing spin - charge separation in the long wire limit , and there are no new spin - incoherent features that appear relative to the ll aside from the length dependence of the dc critical current that enters in . in this paper we have touched on what we believe are some of the most easily observed consequences of spin - incoherent behavior in ferromagnet / spin - incoherent luttinger liquid and superconductor / spin - incoherent luttinger liquid junctions . for the case of fm - sill junctions , we computed the spin current pumped into the sill as a result of magnetization dynamics and the effect of spin accumulation in the sill on the parameters governing the magnetization dynamics . we found that for interactions in the sill with @xmath348 the temperature dependence of the spin current and magnetization dynamics is _ qualitatively different _ from the case of a ll and should thus be observable experimentally . some of the key differences between fm - sill and fm - ll systems are summarized in table [ tb : ll_sill ] . the crossover from fm - ll to fm - sill in the exponent @xmath349 governing the temperature dependence of several key quantities is shown in fig . [ fig : delta ] . in the case of sc - sill junctions our results greatly extend those obtained earlier by us.@xcite in that earlier work we were concerned only with the case of adiabatic ( andreev ) contact of the spin - incoherent luttinger liquid to the superconductor . here we have developed those results further and also discussed the opposite limit of tunnel contacts to the sc . in the tunneling limit we have computed the ac and dc josephson response in the geometries shown in fig [ fig : sc_tunnel ] . we find that in contrast to the case of adiabatic contacts , the tunnel contacts lead to a josephson critical current that is exponentially suppressed with the length of the sill region . this difference arises from the fact that in the andreev limit the dc josephson effect is determined solely by the zero modes of the hamiltonian , while for tunnel contacts the non - zero modes generally dominate the response . these non - zero modes enter because in the tunneling process an electron is created locally near the tunnel barrier and thus requires many wavevectors to build its wavepacket . with the aim of providing a general discussion of junctions of ferromagnets or superconductors with a strongly interacting one dimensional system , we have couched many of our calculations in the recently developed scheme for bosonizing strongly correlated electron systems.@xcite this formalism is valid for arbitrary temperatures with respect to the spin energy @xmath12 , but requires that the temperatures remain small compared to the charge energy @xmath13 . in order to adapt that formalism to the systems we discussed here , we extended those results to express open and andreev boundary conditions in a language valid for arbitrary temperatures with respect to the spin sector . through several examples , we showed how various correlations functions could be evaluated and verified that in the spin - incoherent regime the results properly reduce to the results obtained using the world - line picture.@xcite one of our primary motivations for using the language of ref . [ ] is to help provide a clear starting point for numerical studies of the spin - incoherent regime and the many interesting ( and probably experimentally relevant ) crossovers that occur between it and the luttinger liquid regime . within this formalism the charge physics can be computed analytically via a standard bosonization scheme , but the spin sector must be addressed numerically for temperatures of order the spin energy @xmath12 . the types of correlation functions that must be computed numerically are those that add and remove a site ( or multiple sites ) from a spin chain , such as @xmath350 , that appear in the evaluation of a single - particle ( or multi - particle ) greens function . ( see eq . , for example . ) we would like to emphasize that numerical studies of strongly interacting one - dimensional electron systems with appreciable temperature compared to the spin energy is an entirely untouched area and is now ripe for investigation . on the experimental side , the lack of a clear experimental smoking gun " observation of the sill remains a key issue to be addressed . however , there are mounting experimental indications we are not far away.@xcite our best numerical estimates@xcite suggest that many low density quantum wires sit right on the edge of the spin - incoherent regime and perhaps all that is needed is a focused experimental effort to search for its signatures , rather than any key technical breakthough . one of our aims in this work it to highlight certain classes of hybrid structures where spin - incoherent physics should be observable . finally , we would like to close with what we regard as some of the outstanding theoretical issues surrounding the spin - incoherent luttinger liquid . perhaps the main one is the behavior on temperature scales @xmath22 that we already alluded to above . related to this is a better understanding of the crossover between the luttinger liquid and the spin - incoherent luttinger liquid regimes . both of these will likely require a numerical attack as there are no obvious analytical methods available to address them . there is also the issue of spin - orbit coupling that has so far received no attention . for very strong spin - orbit coupling is there novel behavior in the regime @xmath351 ? the subject of noise in hybrid structures involving a sill will be discussed in a forthcoming work.@xcite we thank leon balents for discussions on spin transport in one dimensional systems and oleg starykh for discussions on pair correlations in superconductors . this work was supported by nsf grants phy05 - 51164 , dmr-0606489 , the lee a. dubridge foundation , and the welch foundation . in this appendix we compute the correlator @xmath158 and its fourier transform @xmath149 using the general formalism for bosonization of strongly correlated electrons in one dimension developed by matveev , furusaki , and glazman who also applied it to the evaluation of the single particle green s function for an infinite system and its fourier transform.@xcite we summarized the main results of the bosonization scheme in sec . [ sec : bosonization ] . the calculation here is for a semi - infinite , or finite system with open boundary conditions , so is different in detail from what has been discussed in refs . [ ] , but the basic elements of the bosonization are the same . this appendix is meant to illustrate clearly in a specific example which parts of correlation functions can be computed analytically at finite temperatures and which pieces in general require methods yet to be developed , or a numerical attack . as was already discussed in sec . [ sec : bosonization ] the chief difficulty lies in computing the correlations at arbitrary temperatures in the spin sector . it is hoped that the details given here will provide a good starting point for those skilled in numerics to enter the study of strongly interacting one dimensional systems where there is currently _ no _ quantitative understanding of the regime @xmath352 . given the typical values of @xmath353 present@xcite in quantum wires , this intermediate " temperature regime may turn out to be the most relevant experimentally . in the evaluation of the spin current ( pumped from a ferromagnet into an adjoined one dimensional system coupled via a tunnel junction ) , the correlation function @xmath354\rangle\ ] ] arises in lowest ( second ) order perturbation theory . it contains the difference of the product of single particle greens functions for both the fm and the interacting one dimensional system @xmath355 , \end{aligned}\ ] ] where the brackets @xmath356 denote a thermal average computed with the open boundary conditions described in sec . [ subsec : open ] . the correlators @xmath357 and @xmath358 can be computed by standard bosonization methods.@xcite the result is @xmath359 where @xmath360 is a high energy cut off . our real objects of interest here are the boundary greens functions @xmath361 and @xmath362 evaluated at arbitrary temperature with respect to @xmath12 , but small compared to @xmath13 . to evaluate these we make use of expression for the electron operator . straight forward substitution gives @xmath363 } z_{l_1,s } z^\dagger_{l_2,s } e^{-i[(1+\frac{q_2}{\pi})\theta(0)-\phi(0)]}\rangle.\end{aligned}\ ] ] open boundary conditions at @xmath64 prevent fluctuations of the @xmath47 field which sets it to a constant , which we take to be 0 . with @xmath47 set to a constant , the integrals over momentum can be done trivially resulting in @xmath349-functions for @xmath364 which kill those sums and selects @xmath281 . the final result is remarkably simple @xmath365}\rangle.\ ] ] similar manipulations yield @xmath366}\rangle,\ ] ] where we have made explicit the time dependence in the spin correlators involving @xmath37 and @xmath367 . the general theoretical challenge is then to compute those correlation functions which involve adding and removing a site from the spin chain , which is an interacting many - body problem analogous to fermi edge physics@xcite only it involves knowledge of the hamiltonian on potentially many energy scales . however , for @xmath20 the time evolution of the spin degrees of freedom are very slow compared to the charge degrees of freedom , regardless of temperature , and can be neglected for times @xmath368 . note that in the limit @xmath369 , the spin dynamics can always be neglected . furthermore , if @xmath370 , then the spins become essentially non - interacting , and therefore non - dynamical . then the correlation functions simplify considerably . in the spin - incoherent regime one has@xcite @xmath371 and @xmath372 independent of time which can be straightforwardly generalized to include an applied external magnetic field.@xcite the result is @xmath373 where @xmath374 is the probability of having spin projection @xmath33 . it takes the values @xmath375 , where @xmath376 is the zeeman energy . note that the correlation functions @xmath377 and @xmath378 in the spin - incoherent regime are identical to their values in the infinite system.@xcite this is because in the spin - incoherent regime the spins are non - dynamical , so the open boundary conditions which tend to suppress fluctuations have essentially no effect on the spins which are rendered non - dynamical by @xmath370 . finally , we are left to evaluate @xmath379}\rangle$ ] , which is identical to @xmath357 and @xmath358 only with the exponent changed @xmath380}\rangle = \frac{(-i)^{1/2k_\rho}}{2\pi \alpha_c}\left[\frac{\left(\pi k_b t/\epsilon'_c\right)}{\sinh\left ( \frac{\pi k_b t t}{\hbar}\right)}\right]^{1/2k_\rho},\ ] ] where @xmath381 is a high energy cut off for the charge sector in the sill . combining the results , , , and , we find @xmath382^{\delta^{sill}+2},\ ] ] where @xmath383 is given in table [ tb : ll_sill ] , and @xmath384 is an effective high energy cut off given by @xmath385^{2k_\rho/(2k_\rho+1)}$ ] . for making the following formulas more compact , we define @xmath386 . we are now ready to compute the fourier transform of , @xmath387 . making the change of variables @xmath388 , we have @xmath389^{\tilde \delta}}.\ ] ] the integral is standard @xmath390^{\tilde \delta } } = 2^{\tilde \delta-1}\gamma(1-\tilde \delta)\frac{\gamma\left ( { \tilde \delta \over 2}-i\frac{\hbar \omega}{2\pi k_b t}\right)}{\gamma\left(1-{\tilde \delta \over 2}-i\frac{\hbar \omega}{2\pi k_b t}\right)}.\ ] ] the result of the integral can be transformed to a more convenient form using the following identities for complex number @xmath391 : @xmath392 and @xmath393 . this gives @xmath394 { 1 \over \pi},\end{aligned}\ ] ] where the sin can be expanded to pick out the real and imaginary parts . selecting the imaginary part of @xmath149 we find @xmath395\propto \frac{1}{(2\pi \alpha_c)^2}\left(\frac{\hbar}{\pi k_b t}\right ) \left(\frac{\pi k_b t}{\tilde \epsilon_c}\right)^{\tilde\delta}\nonumber \\ \times \sinh\left(\frac{\hbar \omega } { 2\pi k_b t}\right)\bigg|\gamma\left({\tilde \delta \over 2}+i\frac{\hbar \omega}{2\pi k_b t}\right)\bigg|^2,\end{aligned}\ ] ] which for @xmath165 gives @xmath396\propto \hbar \omega ( k_b t)^{\delta^{sill}},\ ] ] in agreement with our , and the general analytical form earlier obtained by bena and balents@xcite up to the exponent describing the temperature dependence which is changed in the spin - incoherent regime , as illustrated in fig [ fig : delta ] . the calculation above is valid provided the temperature is larger than the level spacing of the finite length wire : @xmath397 . for @xmath398 the finite length results of ref . [ ] can readily be generalized to finite temperatures . no simple power laws emerge , but rather a more complicated dependence involving hyperbolic trig functions . in this appendix we show that the bosonization scheme introduced in sec . [ sec : bosonization ] recovers our earlier results for the single - particle greens function , the tunneling density of states , and the pair correlations.@xcite we are interested in computing the fourier transform of the single - particle green s function @xmath399 where here @xmath400 and @xmath401 measure the distance from the boundary with the sc . taking @xmath402 and using the general formula we have @xmath403-\phi(x , t)]}\nonumber \\ \times e^{-i[(1+\frac{q_2}{\pi})[k_f^hx+\theta(x,0)]-\phi(x,0)]}\rangle,\;\;\;\end{aligned}\ ] ] where the @xmath47 and @xmath48 fields have the expansion given in . we take the limit @xmath85 as we did in ref . [ ] so that the zero modes play no role . since the holon sector is described the the gaussian theory we can make use of the identity @xmath404 for operator @xmath289 and evaluate the correlation functions at zero temperature with respect to the charge energy but infinite temperature with respect to the spin energy : @xmath405 , that is we take both the spin energy and the temperature to zero but alway maintain @xmath406.@xcite in the spin - incoherent regime , the spin correlations are translationally invariant so we have @xmath407 . making a change of variables @xmath408 , the summation over the sites of the spin chain results in a delta function , @xmath409 , which immediately kills one of the momentum integrals and sets @xmath410 . we re - label the remaining momentum variable @xmath91 , @xmath411 } e^{-i[(1+\frac{q}{\pi})\theta(x,0)-\phi(x,0)]}\rangle.\;\;\;\end{aligned}\ ] ] next , we define a variable @xmath412 and re - express the integration in terms of this variable , @xmath413 } e^{-i[y\theta(x,0)-\phi(x,0)]}\rangle.\end{aligned}\ ] ] finally , noting that @xmath414)=\int _ { -\infty}^{\infty } \frac{dy}{2\pi } e^{-i[\pi l - ( \theta(x , t)-\theta(x,0))]y}$ ] we see that eq . correctly reproduces eq.(4 ) of ref . [ ] after we recall the relations @xmath415 and @xmath416.@xcite the identical result therefore follows for the greens function and the tunneling density of states derived from it . here we compute the pair correlation , @xmath417 , a distance @xmath400 from the superconductor boundary . by direct substitution of and we have @xmath418-\phi(x)]}\nonumber \\ \times e^{-i\phi(x ) } z_{l_2,\downarrow } e^{i[(1+\frac{q_2}{\pi})[k_f^hx+\theta(x)]-\phi(x)]}\rangle.\end{aligned}\ ] ] at the boundary , we showed in sec . [ subsec : andreev ] that andreev reflection implies that @xmath419 . since @xmath420 at the boundary we may fluctuate , we expect also that the result @xmath419 approximately holds near ( within @xmath103 ) of the boundary . hence , @xmath421}\nonumber \\ \times e^{-i\phi(x ) } z^\dagger_{l_2,\uparrow } e^{i[(1+\frac{q_2}{\pi})[k_f^hx+\theta(x)]}\rangle.\end{aligned}\ ] ] we now assume that we are in the spin - incoherent regime where the spin correlations are translationally invariant : @xmath422 . then repeating the steps used above to compute the single particle greens function ( changing the summation variable @xmath408 , killing the momentum integral with @xmath349-function ) , we find @xmath423 } e^{-i2\phi(x ) } \rangle.\end{aligned}\ ] ] we then make use of the integral representation of the @xmath349-function , @xmath424\right ) } = \delta \left(l-\frac{2}{\pi}[k_f^hx+\theta(x)]\right)$ ] to obtain @xmath425\right)e^{-i2\phi(x ) } \rangle,\ ] ] which for @xmath426 can be evaluated by taking the discrete sum to an integral : @xmath427 . doing so we recover the result ( 9 ) of ref . [ ] after again recalling the relations @xmath415 and @xmath428.@xcite

online social networks ( osn ) like facebook and twitter have become an essential part of our daily life , e.g. facebook has over 1 billion monthly active users 1/7 of the world population . despite their overwhelming success , the centralized control of these services have led to serious concerns about user privacy , censorship @xcite and operational robustness . they are now obvious targets or even vehicles of many totalitarian regimes which constantly seek to monitor and control information dissemination among their people . some even argues that the centralized and absolute control power associated with osn services have resulted in many non - user - friendly management styles or policies , e.g. the real world cases reported by indiewebcamp participants . the aforementioned concerns have motivated the active development of decentralized social networks ( dsn ) in the recent years with the goal to let users regain better control over their personal data . although numerous dsn projects have been launched in the past few years , to date , very few of them have managed to proceed beyond the prototyping or paper - publishing stage . even the most successful , often cited example of dsn , namely , diaspora @xcite , has only about 0.4 million users , let alone active ones . while this may not be surprising given the wide range of challenges in implementing , bootstrapping and operating a dsn @xcite , we believe that the biggest hurdle for the widespread adoption of dsn services is the lack of a gradual transition path for a user to migrate to a dsn without leaving behind a large portion of his / her friends who are likely to stick with only existing osns for convenience . as such , even for the minority early adopters who are determined to move to a dsn service for its enhanced privacy protection and possibly richer functionalities , the ability to perform cross - platform socialization will be critical . towards this end , we propose snsapi , a middleware which enables an end - user to aggregate and stitch together all of his / her online social activities and become a node in a `` meta social network '' @xcite as illustrated in fig . [ fig : instantiated - metasn ] . in fact , many existing sns users already consciously or subconsciously `` stitch '' the heterogeneous platforms together by selectively relaying messages across different platforms after some manual filtering and editing . for example , after reading a `` juicy '' gossip from a blog , you may forward it manually to your personal friends on facebook . such manual cross - platform forwarding operations can be viewed as the formation of a meta social network which overlays on top of the existing sns . our objective is to provide the tools and systems that can help users to better perform cross - platform socialization . we expect those tools and systems can gradually detach users from existing centralized osns and allows a smooth transition to the decentralized ones . in summary , this paper has made the following technical contributions : @xmath0 after reviewing related work in section [ sec : related - works ] , we analyze the challenges of building a dsn and propose a meta social networking approach in section [ sec : a - meta - social - networking - approach - for - decentralization ] . @xmath0 as described in section [ sec : current - design - and - implementation ] , we have designed , implemented and released multiple iterations of an open - source middleware called snsapi to support cross - platform socialization over existing sns . sample applications built using snsapi are also presented to demonstrate its flexibility and extensibility . @xmath0 design choices , observations and maintenance experience of snsapi are discussed in section [ sec : design - maintenance - and - refactor - experience ] . @xmath0 we have conducted an experimental deployment of a 6000-node dsn over planetlab based on the snsapi middleware and measured its empirical performance . as a comparison , we also provide an analytical model to characterize the performance of the system with results presented in section [ sec : a - medium - scale - dsn - in - the - wild ] . @xmath0 we conclude our findings and propose follow - up work for the future in section [ sec : conclusions ] . to overcome the drawbacks of centralized sns , decentralized social networks ( dsn ) such as diaspora @xcite , musubi @xcite and onesocialweb @xcite have recently been proposed and implemented . according to our classification in section [ sec : a - meta - social - networking - approach - for - decentralization ] , most works are distributed social network ( dissn ) . two examples of federation protocols are onesocialweb @xcite and ostatus @xcite , which are not bound to any particular software implementation . for the dissns , there are several different system architectures . prpl @xcite and musubi @xcite are two examples of fully decentralized systems . each user corresponds to a dsn node and they only have the view of their direct connections . diaspora @xcite is a super - node based system . users can setup their own `` pod '' ( server ) or register on other `` pods '' . peoplenet @xcite and rflex @xcite are systems combining heterogeneous communication technologies . peoplenet @xcite first routes messages to the neighbourhood of the target using cellular infrastructures and then leverages opportunistic forwarding to complete the last hop(s ) of the delivery . rflex @xcite targets spontaneous group messaging and transparently switches between the cloud backend and nfc/ d2d connections . epost @xcite and peerson @xcite are two dht - based solutions . while epost @xcite follows dht loop for message delivery , peerson @xcite only uses dht to locate users before establishing direct connections for data transfer . these systems and protocols adopt a clean - slate design and have few or none provisions to support migration . although some systems have import functions from existing service providers , they still have the lock - in effect because they do not give a common data structure abstraction to support convenient export and inter - operation . in terms of the ability to perform cross - platform operation , there are many existing aggregation services on the internet . for example , ifttt @xcite abstracts internet - based information services as `` channels '' and allows users to define `` if - this - happens - then - do - that '' ( ifttt ) forwarding recipes . yahoo pipes @xcite supports fewer platforms but allows more sophisticated processing by designing a story - board of logical operations ( e.g. condition , loop , etc ) . yoono @xcite , socialoomph @xcite , and hootsuite @xcite are other example services which provide partial interoperability between heterogeneous platforms with some basic personalization functions . while these services are useful for novice users , they have at least one of the following problems : 1 ) only configurable but not programmable , which severely limits their functions ; 2 ) non open - source and thus less reusable for automation ; 3 ) not readily extensible to other platforms . the open - source projects , opensocial @xcite provides an abstraction for different osns but it requires a steep learning for simple tasks . thinkup @xcite , an open - source web service , reads messages from several osns and stores them in a local database to facilitate the mining of important information . however , it does not provide a comprehensive abstraction of heterogeneous sns like snsapi and its web service nature also requires more infrastructure support ( e.g. lamp environment ) , making it hard to run under resource critical environment , e.g. mobile devices . before delving into the details of our approach for the decentralization of sns , we briefly discuss three alternative paths towards this goal . according to the degree of decentralization , there are three types of dsn as illustrated in fig . [ fig : degree - of - decentralization ] : @xmath0 distributed social network ( dissn ) . in a distributed social network , nodes are homogeneous and runs the same software package the solid circles . they use the same protocol the solid lines . most dsn proposals and implementations are actually dissn , e.g. diaspora @xcite , musubi @xcite etc . in this sense , the dissn approach is * software - package oriented*. @xmath0 federated social network ( fedsn ) . nodes in a fedsn can adopt different software packages the `` solid '' and `` dashed '' circles in fig . [ fig : fedsn ] as long as they support a common protocol , a.k.a . the federation protocol represented by solid lines in fig . [ fig : fedsn ] . email and ostatus @xcite are two examples of this approach . in short , fedsn is * protocol oriented*. @xmath0 meta social network ( metasn ) @xcite . metasn is the network of different social networks : in this case , 1 ) there are many different software packages ( server/ client implementations ) and 2 ) nodes may use different protocols the solid or dashed lines in fig . [ fig : metasn ] . information diffusion on metasn does not rely on a single software package or a single protocol . instead , it is done via multi - hop bilateral communications . there is a common object to hold information and it can have different representations on different links . in other words , metasn is * object oriented*. one can see that metasn is highly reminiscent of our real life social network in which different people may talk different languages . socialization is not dependent on a single language but on multi - hop bilateral communications . we have designed a cross - platform middleware , snsapi , to realize the metasn approach described above . [ fig : instantiated - metasn ] illustrates a metasn formed by snsapi and existing osns . users running an snsapi - based app can socialize with users of mainstream osns , e.g. twitter and facebook , using the corresponding protocol . they can also conduct the same types of online social activities with each other in a transparent manner via other communication channels supported by snsapi , e.g. ( e)mail and ( r)ss . in contrast , the twitter user and the facebook user at bottom of the figure can not communicate with each other due to the lack of a common protocol . snsapi users also have the ability to aggregate , filter and/or relay messages in this metasn . as such , the snsapi - based nodes can enable multi - hop social interactions among users who are originally locked into their own silo of existing osns . in short , the snsapi can help to jumpstart and then grow a dsn in an incremental , gradual manner . our cross - platform middleware , snsapi , is designed based on the following principles : 1 ) focusing on solving the 80% problems ; 2 ) staying open to support future service evolution ; 3 ) trading execution performance for script development efficiency . in this section , we first introduce the overall architecture and the core abstractions . after that , we present some sample applications to show the flexibility and programmability of snsapi . figure [ fig : the - snsapi - architecture ] depicts the architecture of snsapi , which consists of the following three layers : @xmath0 interface layer ( il ) . is the base class for all kinds of sns and can be derived to implement real logic that interfaces with those platforms . in snsapi terminology , the module containing derived classes is called `` plugin '' ; the derived class from is called `` platform '' ; the instance of the class is called `` channel '' . message and message list classes are also defined in this layer to provide the core abstraction to support metasn . @xmath0 physical layer ( pl ) . there are many common operations when interfacing with different sns , e.g. http , oauth , error definition , etc . we implement them in the pl so that plugins can reuse them . to enhance flexibility , we provide a wrapper class for most 3rd party modules so that users can substitute ( all or part of ) them with others without modifying the core of snsapi . @xmath0 application layer ( al ) . to reduce repeated works in batch operations , we developed a `` pocket '' class in al . it is a container to hold multiple channels . for most applications , pocket should be the service access point ( sap ) to snsapi . in this way , end users can enable new channels by simple configurations and no intervention from app developer is needed . , title="fig : " ] we abstract the interfaces of different sns platforms to support a set of common primitives . since not all platforms have direct support for those primitives , we often have to translate the functions in order to provide a unified view to the upper layers . the three fundamental primitives and some examples are as follows : @xmath0 this name is short for either `` authorization '' or `` authentication '' . for example , in order to access osns like facebook , we need to go through the oauth flow . in order to access email platform , the usual approach is to authenticate via username/ password . @xmath0 this function gets the latest messages targeted to the user on a certain channel . for an osn , it retrieves the home timeline . for email platform , this function gets latest messages from inbox , which resembles the home timeline of osns . is the basic `` read '' function for a platform . @xmath0 it provides the basic `` write '' function for a platform . on an osn , this can be a status update , blog update , or others , depending on the platform . on rss platform , this function writes a new entry to the feed . although and can be named as `` read '' and `` write '' from a modeling viewpoint we stick to the osn convention . this philosophy is adopted throughout snsapi : instead of providing a unification scheme from scratch and requiring other existing/ new osn platforms to follow ( as in the case of opensocial @xcite ) , we adapt snsapi to existing platforms and work out a common divisor . from the basic primitives , two more `` write '' functions can be derived : reply and forward . the definition is different across osns due to differences in the positioning of each platform . the model of `` write '' operations observed on existing osns are further discussed in section [ sec : the - model - of - write - in - osn - platforms ] . following are the definitions within the context of snsapi : @xmath0 by replying a message , the user is able to add comments to the original message . if bob replies a message posted by alice , she will get a notification in some way . where to post the comment is not regularized and can be different across platforms . @xmath0 by forwarding a message , the user is also able to add comments to the original message . if bob forwards a message posted by alice , she may or may not get a notification , depending on the platform . however , the forward message together with the original message will appear in the message update list of bob . note that can be readily implemented via and we have already included this cross - platform forwarding function in the base class . plugins can implement platform - specific forwarding functions to enrich the functionalities . as is discussed in section [ sec : a - meta - social - networking - approach - for - decentralization ] , we need to design a common object to allow the formation of metasn . message is the most important data structure sns but comes in different forms , e.g. json object returned by the api of one osn , rfc822 formatted texts of an email , or an xml entry of a rss feed . even if we only consider conventional osns , the json objects returned can still be quite different . in order to facilitate cross - platform operations , we abstract a object . it has the following four components : @xmath0 it contains sufficient information for one to identify a across platforms . it is only designed to be used by snsapi plugins . users are not supposed to tap into the fields of . @xmath0 mandatory fields this includes , , , , and . those fields are actually the `` greatest common divisor '' of different sns platforms according to our year - long refactoring experience . the field is used to convey non - text - based information like url , image , and video files and can be an empty list . @xmath0 optional fields although mandatory fields provide a common base to perform cross - platform socialization , we may sometimes expect a richer set of functionalities . examples are like the `` like count '' and `` share count '' on some osns . those fields are also unified via re - formatting and re - assembling of the raw response from osn providers . app developers should test the existence before using it since they are optional . @xmath0 raw data this is the original data obtained from an osn service provider . it gives the largest amount of information . however , app developer should refer to the manual of specific platforms in order to use it . plugin implements the logic to enable transaction with any specific sns platform . a standard plugin implements the 5 interfaces defined in section [ sec : abstraction - of - interfaces ] . it also performs essential data conversions from raw response to the data structures regularized in section [ sec : abstraction - of - data - structures ] . following is the list of plugins we have already developed : @xmath0 renren , sina weibo , and tencent weibo we started with these platforms , which are the three largest osn services in mainland china . the plugins use restful apis of the corresponding osn . based on the experience of developing these plugins , we have derived some common building blocks , e.g. http , oauth , and put them in the physical layer ( pl ) . @xmath0 rss , email and sqlite these three platforms are not regarded as social networking services in the traditional sense . however , they are fundamentally the same as the osns mentioned above , i.e. a means to read/ write messages , so it is easy to make them snsapi - compliant . @xmath0 twitter and facebook these two platforms are developed based on existing 3rd - party wrappers . their existence shows that one can easily adapt existing wrappers to be snsapi - compliant , so as to enable seamless cross - platform operation on new platforms . pros/ cons of using 3rd - party wrappers are further discussed in section [ sec : user-3rd - party - wrappers - for - plugins ] . the so - called snscli , a command - line interface in form of a python shell , has been developed as an application ( app ) of snsapi . some typical usages include : 1 ) use snscli to interface with programmes / scripts written in other languages via stdin/ stdout ; 2 ) support interactive debugging for a platform . researchers have also developed crawlers in snscli using the vertical interfaces ( section [ sec : vertical - interface - and - horizontal - interface ] ) to facilitate the collection of data over large - scale osns . after our third round of refactoring at the beginning of 2013 , an open - source contributor from github , tommy alex , built a graphical user interface ( gui ) called snsgui ( fig . [ fig : screenshots - of - gui - on - pc - and - android ] ( a ) ) which runs on major computing platforms including linux , windows , and os x. later tommy also ported snsapi to android with a new ui ( snsdroid in fig . [ fig : screenshots - of - gui - on - pc - and - android ] ( b ) ) . it is noteworthy that tommy managed to develop snsgui and snsdroid in 10 days and 7 days , respectively . before developing snsgui , he has no prior knowledge of snsapi . this demonstrates the flexibility and portability of snsapi . we have also developed another app called snsrouter @xcite which provides a web ui and a ranking framework for cross - platform socialization . based on the unified abstraction of snsapi , the ui was developed in one week . user efficiency is improved due to the convenient cross - platform operations plus prioritization of incoming messages . among all kinds of benefits , it is worth to note that snsrouter can readily become a node in an ad hoc dsn by properly configuring some channels supported by snsapi . due to space limit , we can not present other existing apps that have been developed for snsapi , e.g. automatic forwarder , automatic replier , automatic backup script . they can be found in our open - source repository . they all share one distinguishing feature when comparing to their traditional counterparts : these apps all work in a platform - independent fashion and are ready to operate on new platforms . since the initiation of snsapi project , we have gone through three major rounds of refactoring . the project is still under active development and there will be more forthcoming features . in this section , we provide a comprehensive discussions of our experience in designing , maintaining and refactoring snsapi . the philosophical shifts and key observations extracted from our experience may shed light on future design of similar systems . we categorize the functions provided by different social platforms into three groups . as is described in section [ sec : abstraction - of - interfaces ] , base functions include , and ; derived functions include and . any other osn - specific functions are regarded as extra functions , e.g. album list query and like . extra functions are not very common across platforms , so we do not try to unify them . [ fig : interface - horizontal - vertical.eps ] illustrates the three types of functions . based on this categorization , we can deal the aforementioned functions in a horizontal manner . while providing horizontal interfaces is the original goal of snsapi , we found that some users expected snsapi to be a full - fledged wrapper for a given osn platform . such mis - matched perception roots from the fact that there exists many ad hoc wrappers in all kinds of languages for a wide range of osn platforms . we realized that the support of the horizontal interfaces alone will not be adequate : in order to provide a unified set of horizontal interfaces , we are forced to trim down existing operations in some platforms which results in a loss of functionalities . to overcome this dilemma , we have to expose an additional set of vertical interfaces which support a richer set of functions but are allowed to be drastically different across various platforms . the horizontal interfaces can be used to systematically stitch different social platforms together and they require a very mild learning curve . the vertical interfaces must be used with the assistance of documents from service providers . they are not the recommended way of using snsapi because apps have to deal with upcoming changes from the osn providers . nevertheless , developers may find it useful when they are dealing with a specific platform , e.g. large - scale crawling as is used in section [ sec : a - medium - scale - dsn - in - the - wild ] . although snsapi tried to conform to existing services from the very beginning , the models provided by existing platforms may not be ultimately ideal . we observed different models of `` write '' operations , i.e. , and , from existing osns . a summary of the write models is presented in table [ tab : write - models - of - different - osn - platforms ] and we discuss the advantages/ disadvantages as follows . @xmath0 twitter : there is only one operation twitter has a `` retweet '' built - in ( official retweet , ort ) but users find it not very useful because they can not put comments while retweeting . towards this end , many twitter users are still using the old user - invented conventions ( user - invented retweet , urt ) by updating a status in the form : . this user invented `` rt '' notation is closer to the `` forward '' operation defined in section [ sec : abstraction - of - interfaces ] . if the `` '' appears at the beginning of a status update , it is regarded as a `` reply '' . in other words , the reply operation on twitter is still an update operation in essence . @xmath0 facebook : there are three operations regarding a status update , share , and reply . facebook s reply is the same as our definition in section [ sec : abstraction - of - interfaces ] . we will term this type of reply as facebook - style reply ( fbsr ) , in contrast to the above mentioned twitter - style reply ( tsr ) . fbsr is more privacy preserving but snsapi will lose track of those reply messages because they do not enter the timeline of the replier . when `` share '' one object on facebook , users can add comments to it . at first glance , it appears as a forward operation in our context . however , all share operations go to the original status and are threaded there . compared with our notion of `` forward '' , this `` share '' can not track the forwarding traces . in this case , it is better to textually construct a `` forward'' message by the method directly ( i.e. a urt ) on facebook . in fact , the general method in is designed to perform a urt if the platform does not have a built - in `` forward '' function in our sense . @xmath0 renren , sina weibo , tencent weibo : there are three `` write '' operations update , forward , and reply . in terms of product positioning , renren is a facebook - like service in china ; sina weibo and tencent weibo are two twitter - like ones . interestingly , their write models are the same and partly differ from both twitter and facebook as presented in table [ tab : write - models - of - different - osn - platforms ] . .write models of different osn platforms[tab : write - models - of - different - osn - platforms ] [ cols="<,<,<",options="header " , ] + ( the more * s the better ) friend - list management for osns is so far only available via platform - specific vertical interfaces ( section [ sec : vertical - interface - and - horizontal - interface ] ) . as the designers of snsapi , we have two strong reasons not to horizontally abstract friend - list management functions . firstly , friend - list management is not a common and frequent operation ( so - called 80% problems ) , so out - of - band ( oob ) management is acceptable . secondly , the channel - list in snsapi is actually an implementation of friend - list under the fully decentralized scenario . consider a social network formed using the rss platform ( section [ sec : a - medium - scale - dsn - in - the - wild ] ) . one user can update status to a rss feed and publish it to some common location known by his/ her friends . the other users can instantiate a rss channel with the url pointing to this location . in this way , the snsapi users form a dsn and the channel - list of snsapi is actually the friend - list . instead of unifying friend - list management functions on different osns , a better way is to build a channel management application on snsapi to provide a more elegant abstraction . although our primary goal is to evolve towards a decentralized social networks structure , centralized osn platforms are ( and will remain to be ) very important components under the bigger meta social network paradigm . as discussed before , one reason is that those centralized osn services are already established and we must avoid link loss during the migration process . another reason is that centralized osn platforms can actually serve as good message aggregators . consider a special platform we built in snsapi . with this platform , one can directly configure the list of users he / she wants to follow . without explicitly adding the target parties as friends via the request/ approval process , the snsapi user can readily get their public status updates . this approach gives us a clean model for friend - list management as is discussed in section [ sec : friend - list - management ] . however , more resource is used for querying those `` friends '' one by one . the alternative approach , adding those friends on an osn first and configure only one normal channel in snsapi , is more resource saving because a single query is enough to retrieve an aggregated timeline . this section describes our experimental deployment of a large - scale dsn on planetlab using snsapi . after instantiating a network of 6000 snsapi - based nodes on planetlab , we replay real - world update/ forward activity traces collected from sina weibo , one of the top osns in china . to our best knowledge , this is not only the largest experimental validation and systematic evaluation of any dsn solutions in the literature but also the first one which involves the replaying of real osn traces in this scale . our experiment platform has three major components : @xmath0 snsapi : the cross - platform middleware described in section [ sec : current - design - and - implementation ] the basic building block for this dsn . @xmath0 bot : when the experiment starts , a bot running on each instance of a snsapi - based node will continuously perform four actions : 1 ) periodically query timeline from their friends ( ) ; 2 ) update sns status according to a schedule ( ) ; 3 ) forward status according to a data - driven user behavior model to be described in section [ sec : data - driven - user - action - model ] ( ) ; 4 ) monitor nodal resource usage . we define an important system parameter called query gap ( qg ) which is the interval between two rounds of the timeline query action . in our experiment , we use a system - wide constant value for qg and show that this simple strategy suffices for the initial formation of a medium sized dsn ( e.g. 6000 nodes ) . to avoid synchronization problems , every bot sleeps for a uniformly random duration between 0 and qg before the issuing its first timeline - query to its neighbours . @xmath0 planetlab manager ( plm ) : planetlab is a global research infrastructure widely used to evaluate new networking systems . planetlab nodes provide no guarantee on their service availability and can go online / offline at any time . this is ideal to evaluate the robustness and elasticity of our proposed dsn architecture . we have also developed a toolbox to manage our experiment on pl which includes modules for 1 ) base environment deployment such as the nodal installation of python2.7 and virtualenv ; 2 ) node - screening which help us to avoid machines with large time - drift and those which are behind incompatible firewall ; 3 ) experiment workflow management like workspace cleaning , bot distribution , initialization/ starting/ terminating each experiment , and nodal log collection . in principle , we can connect our bots through all kinds of platforms supported by snsapi . in reality , most osn platforms require user s intervention before snsapi can access it : app - key application , configuration on the osn , and completion of an authorization flow . to allow automatic deployment of thousands of bots , we choose to use rss in this experiment . for practical deployment in the wild , each pair of snsapi users can choose their preferred communication channel as our metasn model only requires bilateral communications . we collected real data from sina weibo , the largest microblogging service in china , to evaluate the proposed ad hoc dsn based on snsapi . we leverage the vertical interface ( section [ sec : vertical - interface - and - horizontal - interface ] ) of snsapi to crawl the profile , status update , forwarding history , as well as the followee list of each user . in order to get a meaningful trace to evaluate our system , we need a tightly knitted group of active users with content frequently originated from and forwarded within the group . the following heuristic to used to identify such group of candidate users : 1 ) we use 60 + users categorized as `` programmers '' by sina weibo as our seed ; 2 ) from the forwarding history of these seed users , we extract 6000 + more users . the overall development plus execution time of this snsapi - based sina weibo crawler is less than two days . this shows the remarkable development and execution efficiency provided by snsapi . in total , we collected 33 gb of raw json data , which includes 12 million status update / forward actions , involving 480,000 users in the forwarding chain . one of the primary reasons for users to use osn is to identify and disseminate interesting messages . using only passive measurements , it is impossible to determine what messages are seen and what messages are interesting to each user . as a surrogate , we consider the forwarding behaviors , which are strong indicators of user interest and message relevance . a key objective is to determine the additional forwarding latency caused by our snsapi - based dsn when comparing to that of a centralized osn . consider the process illustrated in fig . [ fig : user - model.eps ] : user @xmath1 post the original message @xmath2 at @xmath3 ; user @xmath4 forward this message by posting @xmath5 at @xmath6 . the following factors will affect the end - to - end delay from @xmath3 to @xmath6 : @xmath0 user online / offline pattern : from @xmath7 to @xmath8 the user is offline . the original message @xmath2 is posted during this period but can not be forwarded by @xmath4 . @xmath0 system delay : this is the time required by @xmath2 to go from @xmath1 s database to @xmath4 s database . for a centralized osn , there is conceptually only one database so this delay component should be negligible . that is , as soon as @xmath1 post @xmath2 , @xmath4 is able to get @xmath2 in his / her database . for dsn , it can take considerable time for @xmath2 to go to @xmath4 s database , which may cause a major increase in latency . @xmath0 application delay : users rely on certain application to retrieve their home timeline , e.g. the web ui or the mobile app of facebook . additional delay is added due to various reasons like refreshing frequency or message re - ranking strategy . users rely on many different applications to access social networks , so this part of delay can be highly variant . @xmath0 user delay : after the above stages , @xmath4 already has @xmath2 in the home timeline . in order to identify the interesting messages , @xmath4 will spend time on reading the messages , watching attached videos , or even following the embedded links . additional delay is added until @xmath5 is finally posted at time @xmath6 . our analysis of the user action cycle shows that : 1 ) system delay causes the major difference between centralized and decentralized social networks ; 2 ) factors like application delay and user delay are too complex to be characterised by simple models ; 3 ) precise user online / offline pattern is only available to the sns providers ; 4 ) the end - to - end delay , i.e. @xmath9 , is observable via passive measurement . towards this end , we adopt an overall data - driven model , established based on two assumptions : 1 ) a bot is always online , trying to identify and disseminate interesting messages ; 2 ) there is an intrinsic delay for each message to be forwarded . we use @xmath10 to model the intrinsic delay for each message . suppose one bot in our dsn see the message @xmath2 at time @xmath11 , then the forward time @xmath12 is determined by the two rules : @xmath0 1 ) if @xmath13 , let @xmath14 @xmath0 2 ) if @xmath15 , let @xmath16 the first rule is to respect the delay observed from real traces . if our dsn delivers messages quickly enough , then the bot waits until the intrinsic delay is passed . the second rule comes from the bot assumption . since the bot is always online and the message is seen after the intrinsic delay period , it forwards the message immediately . [ fig : user - model - bot.eps ] depicts this data - driven user behavior model . based on this model , our core performance metric is the extra forwarding delay ( efd ) , defined as @xmath17 . to launch our experiment , we first distribute 6733 bots to 450 planetlab nodes randomly . the experiment lasts for one day wall - clock time . we filter out update and forward activities of those users from aug 1 , 2013 0:00:00 to aug 1 , 2013 23:59:59 ( utc ) . during this period , there are 1168 messages originated by the 6733 sina weibo users and are further forwarded within this group . the query gap ( qg ) is empirically set to 5 minutes . in real life , users can manually trigger the function if one wants to get the updates of his friends immediately . after a replaying a 24-hour real trace , we collect 57 gb of logs from the planetlab nodes . we then compute the average resources consumed by each bot . the histograms and summary statistics are presented from fig . [ fig : avg_cpu_hist.eps ] to fig . [ fig : sqlite_db_size_hist.eps ] . notice that the nodal memory usage includes both the bot and snsapi . without the extra bot logic , the typical value is 10 mb , i.e. for running the snscli . there are two major disk - space consumption in this experiment . first is the rss feed file : each bot writes its feed to a directory which is then served by a lightweight http server . second is a sqlite database to support asynchronous operations via the buffering of incoming and outgoing messages . from fig . [ fig : atom_size_hist.eps ] and fig . [ fig : sqlite_db_size_hist.eps ] , we see an exponential decay and the typical values are smaller enough to hold in the ram so that disk storage and access can be avoided in real applications . the c.d.f . and some percentiles of extra forwarding delay ( efd ) are shown fig . [ efd_cdf.eps ] . we see that 68% , 73% , 88% and 99% messages are forwarded with no efd , @xmath18 1min efd , @xmath18 10min efd and @xmath18 1hour efd , respectively . since social networking is not a real - time service and there are already considerable delay other than system delay , we remark that this efd is acceptable . the mild resource consumption and small efd shows that snsapi is a viable solution to enable rapid and incremental deployment of an ad hoc dsn . furthermore , users can run snsapi not only on all major desktop operating systems ( as demonstrated in section [ sec : applications ] ) but also resource - constrained devices like smartphones . although snsapi supports both pull and push channels ( section [ sec : pull - channels - vs - push - channels ] ) , we only used pull channels , namely rss , in our experiment . the major parameter for this pull - based dsn is the timeline query gap ( qg ) and we denote it by @xmath19 . we want to characterise the relation between @xmath19 and the following random variables ( r.v . ) : @xmath0 @xmath20 : an abstract notation for certain resource . it is an r.v . parameterized by @xmath19 . the expectation of @xmath20 is inversely proportional to @xmath19 . without loss of generality , we let @xmath21 = \frac{1}{h}$ ] . @xmath0 @xmath22 : cpu usage . @xmath24 is the cpu usage of the bot logic , i.e. checking and forwarding messages every second . it is constant w.r.t . the poll parameter @xmath19 . @xmath25 is the resource consumption of snsapi . @xmath26 is the coefficient to be fit later . taking expectation , we have the relationship : @xmath27 = { \mathrm{e}}[c_b ] + \beta_c { \mathrm{e}}[r ] = { \mathrm{e}}[c_b ] + \frac{1}{h } \beta_c$ ] . however , we already know from the system design that bot logic will take up the major portion of cpu usage . we expect @xmath26 to be a near - zero coefficient . @xmath0 @xmath28 : memory usage . as @xmath19 varies , only query frequency is changed . memory consumption is not affected by it . it is a r.v . independent of @xmath20 and @xmath19 . @xmath0 @xmath29 : disk i / o . there are two parts of major disk i / o as discussed in section [ sec : a - case - study - of-6k - node - dsn - based - on - snsapi ] . the maximum size of rss feed and sqlite db are 20 kb and 8 mb , respectively . in real applications , they can be fully held in ram , thus eliminating all disk i / o . since this only adds a constant to @xmath28 , we do not build an individual model for disk i / o . @xmath0 @xmath31 : http usage . in our experiment , the major network i / o is caused by http . others like domain name lookup only create very little network load . we measure http usage from the followee s point of view . @xmath32 is the number of http query one bot receives and @xmath33 is the size of http query it serves . since the traces are from a micro - blogging service , the size of messages are upper bounded . thus , we can assume @xmath34 and formulate the following relationships : @xmath35 = \beta_{h_s } { \mathrm{e}}[r ] = \frac{1}{h } \beta_{h_s}$ ] and @xmath36 = \beta_{h_q } { \mathrm{e}}[r ] = \frac{1}{h } \beta_{h_q}$ ] . @xmath0 @xmath37 : the delay from the original message to the forward message in our dsn system . this depends on @xmath38 , the intrinsic delay . most of the relationships are clear and ready for model fitting . the end - to - end forwarding delay @xmath37 can be characterized based on the following additional assumptions : @xmath0 a1 : @xmath19 is the dominant delay component in this system so that other small delays can be neglected . @xmath0 a2 : we assume an ideal internet connection . given current typical access rate and the size of rss feeds in fig . [ fig : atom_size_hist.eps ] , one bot can finish the query to all its neighbours within seconds . this is also very small compared to typical values of @xmath19 so that we can further assume one round of query can be completed immediately upon its issuing . in reality , the forwarding traces form a tree . since our target is @xmath39 $ ] , we can invoke the linearity of expectation , namely @xmath39 $ ] is the summation of per - segment delay . in this way , we essentially decompose the tree into individual forwarding chains . denote the length of forwarding sequence by @xmath40 , which will be fit later using data from real traces . then the expectation of end - to - end delay can be expressed as : ( total expectation ) @xmath41 = \int { \mathrm{e}}[\delta |l = l ] \pr \ { l = l \ } { \mathrm{d}}l\ ] ] next , we focus on the analysis of a chain of length @xmath42 . we denote the original message as @xmath43 and the forward messages as @xmath44 . we denote the absolute time when message @xmath45 is posted as @xmath46 . let @xmath47 be the per - segment delay so that @xmath48 according to definitions . the per - segment intrinsic delay is denoted as @xmath49 . denote the absolute time that a bot starts to pull messages by @xmath31 . note that the bots sleep a uniformly random duration before the first pull . then we have @xmath50 $ ] for any fixed time @xmath51 . we use the deferred decision principle to study the forwarding behaviour of @xmath45 . that is , we first fix @xmath52 , and defer the realization of r.v . @xmath31 . @xmath53 is the duration from first pull to the post of @xmath54 . by modular @xmath19 , we get the time elapse from last pull to the post of @xmath54 . then @xmath55 is the waiting time until next pull and we have a compact expression for @xmath47 according to the data - driven user behavior model described in section [ sec : data - driven - user - action - model ] : @xmath56 if we fix @xmath57 , @xmath58 $ ] and further @xmath59 $ ] . then the overall delay conditioned on the length being @xmath42 is : @xmath60 we now calculate the extra forwarding delay ( efd ) , denoted by @xmath61 . this is to use end - to - end delay in our system minus intrinsic delay ( the delay expected in a centralized system ) : @xmath62\ ] ] taking expectation , we have : @xmath63 = \sum_{k = 1}^l { \mathrm{e}}[\delta_k - i_k ] = \sum_{k = 1}^l { \mathrm{e}}[\max \ { i_k , w \ } - i_k]\ ] ] where @xmath64 is a uniform random variable in @xmath65 $ ] . the overall efd is : @xmath66 = \int { \mathrm{e}}[\delta_{{\ensuremath{\operatorname{efd } } } } |l = l ] \pr \ { l = l \ } { \mathrm{d}}l\ ] ] we assume @xmath49 to be i.i.d . and use @xmath38 for a shorthand . then the expression is simplified to : @xmath66 = { \mathrm{e}}[l ] { \mathrm{e}}[\max \ { i_k , w \ } - i_k]\ ] ] we can fit @xmath38 and @xmath40 from real traces and give the final expression for efd . in this section , we run 20 experiments parameterized by @xmath19 and compare the corresponding empirical measurement results with those predicted by the analytical model . we replay 5 days worth of traces from aug 1 , 2013 0:00:00 to aug 5 , 2013 23:59:59 ( utc ) to drive the experiment in order to reduce diurnal variation . to make the experiment tractable and also reduce planetlab machine variance , we accelerate the experiment time by 12 times . that is , the 5-day traces are replayed in 10 real - life hours . we also trim the dsn topology to 4881 nodes by excluding those nodes having no update or forward activities during this period . the empirical distribution of @xmath38 and @xmath40 are plotted in figs . [ fig : distri_intrinsic_delay.eps]-[fig : distri_sequence_length.eps ] . observe that the intrinsic delay of messages follows a power - law decay while exponential decay is exhibited by forwarding sequence length with their distributions given by : @xmath67 where @xmath68 , @xmath69 , @xmath70 and @xmath71 based on fitting against experimental measurements . the normalization factors can be calculated as : @xmath72 \nonumber\end{aligned}\ ] ] and @xmath73 where @xmath74 and @xmath75 are the maximum and minimum intrinsic delay from observed from the trace data . the expression for efd , @xmath76 = { \mathrm{e}}[l ] { \mathrm{e}}[\max \ { i , w \ } - i]$ ] , can be derived in two steps . instead of going through the fitted parameters , we estimate @xmath77 $ ] directly by averaging over the forwarding sequence lengths and get @xmath77 = 1.14 $ ] . this is to reduce possible errors in parameter fitting . the second part can be solved by double integral : @xmath78\\ & = & \int_1^{i_{(\max ) } } \int_0^h p ( i , w ) [ \max \ { i , w \ } - i ] { \mathrm{d}}w { \mathrm{d}}i\\ & = & \frac{1}{h } \frac{1}{z_i}10^b [ \frac{1}{(a + 1 ) ( a + 2 ) ( a + 3 ) } h^{a + 3 } - \frac{1}{2 } h^2 \frac{1}{a + 1 } \\ & & + \frac{h}{a + 2 } - \frac{2-a(a+3)}{(a + 1 ) ( a + 2 ) ( a + 3 ) } ] \end{aligned}\ ] ] fig . [ fig : h_vs_efd.eps ] depicts the scattered plots of the experiment results . the analytical prediction is plotted as a solid line . one can see that the analytical result matches experimental result when @xmath19 is relatively large ( @xmath79 10 mins ) . for smaller @xmath19 , the analytical result under estimates the efd which can be explained by the following reasons : firstly , we assume @xmath19 is the only delay but there are other delays . those small delays become dominating when @xmath19 is small . secondly , we assume the statuses are obtained immediately if one bot pulls its neighbour . however , given the fully decentralized construction of our dsn , one needs to pull rss feeds from multiple friends . since the network performance between planetlab nodes varies widely , some bots may take much longer time than qg to finish one round of pull . this also makes the experiment results deviate from analytical ones with small @xmath19 . one extreme case is when @xmath80 , the analytical model , which assumes @xmath19 to be the only delay factor , will predict efd=0 . this is obviously not the case and we are working on the refinement of the model for efd . various resource consumption is depicted from fig . [ fig : h_vs_cpu.eps ] to fig . [ fig : h_vs_http_size.eps ] . the results align well with our analysis in section [ sec : analytical - model - for - the - pull - based - backbone ] , namely 1 ) cpu and memory usage are just a constant plus experimental variance ; 2 ) http query and size are reciprocal of query gap , @xmath19 . in this paper , we analyze the challenge towards a decentralized paradigm of social networking services and proposed a meta social networking approach to solve the migration problem . we have built a cross - platform middleware , which is lightweight , programmable , and flexible , to support the transition from centralized osns to decentralized ones . we have presented the design and implementation of snsapi and discussed design choices and observations based on our year - long experience in maintaining and refactoring the corresponding open - source project . we deployed a 6000-node dsn on planetlab based on snsapi to demonstrate that it is a viable approach to form a dsn . using real traces collected from a mainstream osn , we show that with only mild resource consumption , we can achieve acceptable forwarding latency comparable to that of a centralized osn . we have also developed an analytical model to characterize the tradeoffs between resource consumption and message forwarding delay in such a dsn . via 20 parameterized experiments on planetlab , we have found that the empirical measurement results match reasonably well with performance predicted by our analytical model . future works include using non - uniform and even adaptive query gap , improving the analytical model , formation of a dsn using push channels ( e.g. email platform also supported by snsapi ) , developing a distributed back - off protocol to alleviate the congestion in the dsn . during the deployment of the 6000-node dsn and the series of parameterized experiments , we have also built a full set of tools for conducting and managing of similar large - scale dsn experiments . like the snsapi and its apps , this toolbox as well as the osn activity traces we collected from sina weibo will be open - sourced in the near future . we hope they will provide a foundation to support large - scale experimental deployment and performance evaluation for dsn and the associated algorithms .

the cosmic microwave background ( cmb ) is a tremendous tool for studying cosmology . the anisotropies in the cmb provide a wealth of cosmological information ( for a recent review see hu & dodelson 2002 ) and upcoming experiments will provide high sensitivity ( approaching 1 @xmath6 ) and high angular resolution ( approaching 1 ) . such precise measurements will provide interesting constraints on secondary anisotropies in the cmb , imprinted by material along the line of sight at redshifts @xmath7 . the largest secondary anisotropy is expected to be caused by galaxy clusters , through the sunyaev - zeldovich ( sz ) effect @xcite . compton scattering of cmb photons by intra - cluster electrons leaves a spatial and spectral imprint in the cmb . recent measurements of the sz effect have provided detailed maps of the electron distribution in galaxy clusters ( see birkinshaw 1999 and carlstrom 2002 for recent reviews ) but new experiments offer the promise of an order of magnitude increase in sensitivity . with @xmath6 sensitivity it should be possible to probe other cluster properties , such as bulk velocities and electron temperatures . in this paper we investigate the feasibility of measuring peculiar velocities and cluster temperatures at cm and mm wavelengths in the presence of contaminating sources , focusing on the effects of dusty high - redshift galaxies . these issues have been partly addressed previously by @xcite and @xcite but both our understanding of dusty distant galaxies and cmb experimental capabilities have advanced tremendously in recent years . we forecast errors by calculating the sz - parameter fisher matrix for multi - frequency maps that contain signals from the clusters as well as noise from cmb anisotropy , dusty galaxies and the measurement process . our work extends previous studies of sz parameter reconstruction @xcite by considering point source and cmb anisotropy contamination and the influence of the atmosphere on the relative sensitivity of different frequency bands . while @xcite did include the effect of cmb anisotropy contamination , explicitly subtracting it with an optimal filter , point source contamination was ignored . and while @xcite included both cmb and point source contamination , forecasts were for one specific experiment : planck . here we vary experimental parameters in order to guide experimental design . in 2 we outline the sz effect and how it relates to cluster properties . 3 summarizes the relevant known properties of dusty galaxies observed in the direction of galaxy clusters , while 4 outlines our methods for estimating the effects of contaminating sources on cluster measurements . we discuss complications that will arise in analyses of real data in 5 and describe the experiments we consider in 6 . our primary results are in 7 , where we forecast uncertainties for a four - channel reference experiment as a function of cluster gas temperature and core radius . we explore the impact of varying experimental parameters such as angular resolution , observing time and number and placement of frequency channels . we follow that with a discussion of possible radio point source contamination and close with a discussion of our results and implications for future instrumentation . the sz effect is the change in energy of cmb photons from compton scattering with electrons , primarily in the intra - cluster medium . for recent reviews see @xcite and @xcite . the high temperature ( several kev ) of the electrons relative to the cmb photons leads to a net increase of energy of the scattered photons , while a bulk motion of the cluster electrons along the line of sight leads to either a net redshift ( moving away from observer ) or blueshift ( moving toward observer ) of the scattered cmb photons . the first effect is the thermal sz effect and it has a distinctly non - thermal spectrum ; the scattering preserves photon number and mainly just shifts the energy of each photon , distorting the original blackbody spectrum . the effect of the bulk motion is the kinetic sz effect , and the emergent spectrum is that of a blackbody with a slightly different temperature . for the thermal sz effect , the main physical parameters are the optical depth to thomson scattering , @xmath3 and the fractional energy gain per scattering @xmath8 . the combination @xmath9 sets the amplitude of the spectral distortion . for the kinetic sz effect , the redshift or blueshift is set by @xmath10 , so the relevant combination is @xmath11 . typical electron temperatures are on the order of 0.01 @xmath12 , making relativistic effects modestly important @xcite . the relativistic corrections are sensitive to the electron temperature through the relativistic correction to the thomson cross - section and through the relativistic form of the thermal maxwellian velocity distribution . these corrections can be on the order of @xmath13 or higher at many observing frequencies . thus the temperature difference due to the kinetic sz effect is given by @xmath14 while the temperature difference due to the thermal sz effect as a function of frequency @xmath15 ( expressed in dimensionless units @xmath16 ) is given by @xmath17 \quad , \ ] ] where the relativistic corrections have been represented by @xmath18 . typical values for massive clusters of @xmath19 , @xmath20 and @xmath21 lead to typical amplitudes ( @xmath22 ) for the thermal effect , kinetic effect and relativistic corrections of @xmath23 , @xmath24 and @xmath24 respectively . in this work , we take our fiducial galaxy cluster to have @xmath25 , @xmath26 kev and @xmath27 . we vary the thermal sz properties using the scaling relations of @xcite . the central @xmath0 parameter was taken to scale as @xmath28 , which leads to @xmath29 . self - similar evolution would result in @xmath30 ; we use the slightly steeper relation to be consistent with evidence for excess entropy . we ignore evolution of intrinsic cluster properties with redshift , such as the expectation that clusters ( of a fixed mass ) become more compact ( and therefore hotter ) at higher redshift . it is expected that the effect of this evolution is to enhance the thermal sz effect , but the details of this redshift evolution are sensitive to poorly understood gas processes @xcite . star - forming galaxies can be very luminous in the submillimeter wavelength range , as recent measurements by the sub - millimeter common user bolometer array ( scuba ) and other experiments have discovered ( see blain 2002 for a recent review ) . in particular , galaxy clusters are often found to coincide with relatively bright submillimeter emitters , probably because of gravitational lensing effects @xcite . the emission is mainly from dust at a temperature of several tens of kelvin , putting the peak of the radiation spectrum at submillimeter wavelengths , but there can still be significant emission at millimeter wavelengths . many galaxy clusters have been observed to coincide with submillimeter sources with fluxes of several mjy at @xmath15=350 ghz . assuming a typical spectrum for these sources would lead to expectations of fluxes at 150 ghz on the order of 1 mjy . in a beam of 1x1 this would correspond to nearly 30 @xmath6 of contamination if not correctly taken into account . this makes dusty galaxies a non - negligible contaminant for studies of galaxy clusters at mm wavelengths @xcite . throughout we often characterize a given intensity as the equivalent departure from the mean cmb temperature , @xmath31 . the conversion factor , for @xmath32 , is ^-1 & = & ( ) ^2 + & = & ( ) the experiments we consider are not sensitive to the absolute flux , but to spatial variations in the flux . these variations contribute to the variance of temperature fluctuations on the sky , observed with a gaussian beam profile with full width at half maximum of @xmath33 , ( t)^2 = _ l 2l+1 4 c_l where c_l = [ b_t]^-2 s^3 ds . this variance decreases with beam size as @xmath34 . for the convenience of using a quantity that is a property of the sky only , and not the angular resolution of the telescope , we define @xmath35 where @xmath36 is the solid angle of the beam . the shot noise at 350 ghz can be inferred from scuba observations . @xcite used these observations to fit a double power - law form for @xmath37 : dn(>s)/ds = ^-1 with @xmath38 mjy , @xmath39 , @xmath40 and @xmath41 the result is shot noise with an rms of @xmath42k - arcmin where the subscript `` 350 '' is used to denote the frequency . if the spectral behavior of these sources is known , then it would be straightforward to measure the flux at a higher frequency , where the emission is stronger and the sz effects are smaller , and subtract the appropriate levels from the measurements at other frequencies . however , the spectral behavior is not perfectly known , so there will be uncertainty due to imperfect subtraction , as well as uncertainty due to noise in the measurement at higher frequencies . to investigate the homogeneity of the spectral behavior , we used submillimeter observations of local galaxies @xcite . local galaxies were modeled as having a spectrum of the form @xmath43 , where @xmath44 is the blackbody function and @xmath45 here refers to the emissivity spectral index ( not the bulk velocity of the cluster ! ) . dunne provide best - fit values of @xmath45 and dust temperatures . we used these best - fit spectra for the local galaxies , placed them at a range of redshifts and fit them over the range 150 to 350 ghz assuming a power - law @xmath46 . as a function of redshift the mean spectral index was well - fit by @xmath47 , with a fairly constant scatter at each redshift of @xmath48 . the moderate flattening with increasing redshift arises because the radiation that is observed at 350 ghz originates at a higher frequency , closer to the peak of the dust emission . without source redshifts , it is not clear what to take as the mean spectral index . assuming a uniform distribution in redshift between 0 and 5 , the mean spectral index is 2.6 and the @xmath49 scatter around this mean is 0.4 . the sources are unlikely to be at very low redshift , simply due to volume considerations , but the redshift distribution of the observed submillimeter - emitting galaxies is very uncertain . we set our mean value for @xmath50 as @xmath51 and set @xmath52 , unless specified otherwise . we set @xmath53 since the redshift distribution will be more peaked than the uniform one that leads to 0.4 . the quantity @xmath54 is the spectral index rms for _ individual _ galaxies . for @xmath55 galaxies of similar brightness the composite intensity will have a scatter in spectral index that is smaller by @xmath56 . we define an effective number density of galaxies by weighting them according to their contribution to shot - noise variance . thus , n_eff & = & s^3 d / s^3 d + & = & 0.64/(sq . arcmin ) where the last equality follows from the @xcite @xmath37 . for simplicity we set @xmath57(sq . finally we define @xmath58 and take our fiducial value to be @xmath59-arcmin . the @xcite @xmath37 is for an unlensed population of sources . lensing changes both the brightnesses of galaxies and their number densities . indeed , the magnification provided by galaxy clusters has been exploited to study the population to fainter limiting magnitudes @xcite . the magnitude , and even the sign , of the effect on shot noise depends on @xmath37 . magnification increases @xmath60 which increases shot noise , but also decreases @xmath37 ( at fixed unlensed @xmath60 ) which decreases shot noise . for a power - law , @xmath61 , the net effect is an increase in shot noise if @xmath62 . for the relatively bright population at 350 ghz of interest for this work , the shape of @xmath63 is such that lensing leads to an enhancement of shot noise . @xcite pointed out that lensing would significantly enhance the level of contamination by ir luminous galaxies of sz observations . the key quantity for lensing is the einstein radius , @xmath64 . a point source with source plane location coincident with the cluster center ( @xmath65 ) will form a ring of infinite magnification appearing in the image plane at distance @xmath64 . in the image plane , magnification rises from the center towards @xmath64 and then drops beyond @xmath64 . thus , as @xcite showed , the enhancement of confusion noise for an observation towards the cluster center increases with increasing beam size until peaking at @xmath66 as more weight is placed on galaxies near @xmath64 , and then drops with further increase in @xmath33 as the effect is diluted by adding in more galaxies from beyond the einstein radius . for a rough model of the effect of lensing we introduce the confusion noise enhancement factor , @xmath67 which is the ratio of lensed confusion noise to unlensed confusion noise at distance @xmath68 from the cluster center and for observations with angular resolution @xmath33 . for the dependence of @xmath69 on @xmath33 we use the results of @xcite for his galaxy evolution model i1 for a cluster with velocity dispersion of @xmath70 km / s and @xmath71 when it is placed at @xmath72 . for this case @xmath73 1.1 , 1.5 , 2.5 , 2.1 , 1.8 and 1.5 for @xmath74 10 , 20 , 40 , 60 , 100 and 200 arcsec respectively . we ignore the redshift - dependence of @xmath69 since for @xmath75 @xcite shows that @xmath69 is highly independent of the cluster redshift . we also ignore the @xmath76 dependence of @xmath69 since @xcite shows this dependence to be quite slow , covering a range of 30% as @xmath76 varies from 800 km / s to 2000 km / s . for @xmath77 km / sec one expects a gas temperature of about @xmath78 kev @xcite . for dependence on @xmath68 we simply set @xmath79 for @xmath80 and @xmath81 otherwise . for all cases that we consider here the beam size will be larger than a typical einstein radius for a cluster . we model the data at sky location @xmath82 and frequency @xmath83 as having contributions from both signal and noise : ( _ a,_i ) = s(_a,_i ) + n(_a,_i ) . we consider each component in detail below . the signal is due to three components , s(_a , x_i ) = _ m=1 ^ 3 p_m h(x_i)f_m(_a ) , [ eqn : signals ] where p_1 & & y = dl _ t n_e = | _ + p_2 & = & dl _ t n_e v / c = | _ and + p_3 & = & dl_t n_e ^2 = |_^2 = y | _ y. here @xmath84 is the electron number density and @xmath85 is the dimensionless gas temperature . the subscripts on the average quantities indicate the weighting along the line of sight . for @xmath3 the weighting is with the electron - number density and for @xmath0 the weighting is with the pressure , @xmath86 . the optical depth , @xmath3 is given by @xmath87 . the integrals defining @xmath88 and @xmath89 are along the line of sight through the cluster center . we take the angular profile of the cluster to be h ( ) = ( 1+(||/_c)^2)^-1/2 the beam - convolved cluster profile is denoted by @xmath90 , and we assume gaussian beams in all that follows . note that the assumption of the same cluster profile for all three signal components in equation [ eqn : signals ] is only valid for an isothermal intracluster medium . the frequency dependences are given by @xcite f_1 ( ) & & x -4 + f_2 ( ) & & 1 + f_3 ( ) & & -10 + x - x^2 + x^3 + ( -+ x)s^2 where x & & x coth(x/2 ) + s & & x / sinh(x/2 ) . the first and third components are due to the scattering of cmb photons off of the hot electrons in the cluster and represent the thermal sz effect , while the second component is the kinetic sz effect . we have neglected contributions to the sz effects that are @xmath3 times higher order products of @xmath91 ( @xmath92 ) and @xmath68 ( @xmath93 ) . for clusters with temperatures of 10 kev or lower the higher order corrections are less than five percent @xcite at all frequencies except near the null of the thermal sz spectrum ( where the thermal sz signal is sub - dominant ) . because they are the amplitudes of the different frequency shapes , @xmath94 , @xmath95 and @xmath89 are the theoretical parameters most directly related to the data . from them one can estimate the physically interesting quantities : _ y & = & p_3/p_1 = & = & p_1 ^ 2/p_3 = = + & = & p_2/ = | thus the temperature determinations from sz measurements are measurements of the pressure - weighted average gas temperature . velocity measurements are of the mass - weighted average velocity ( assuming electron density traces mass ) times a correction factor dependent on differences in different temperature averages . the optical depth determinations are of the optical depth times another ratio of different average temperatures . throughout , we assume isothermality in order to avoid these complications ; the distinctions must be kept in mind when analyzing real data , particularly for comparison of x - ray and sz gas temperatures . in a recent preprint @xcite it was noted that peculiar velocity measurements of non - isothermal clusters could be biased . the source of this bias can be clearly seen in the above expression for @xmath96 , where the true @xmath97 is multiplied by the ratio of two different weightings of the electron temperature . in general the optical depth weighted temperature will not be the same as the pressure weighted temperature , resulting in a biased velocity estimate . this is a direct consequence of the biased estimate of the optical depth . the noise has a contribution from measurement error , @xmath98 , from primary cmb fluctuations , @xmath99 and from emission from galaxies in the direction of the cluster @xmath100 . we assume the composite dusty galaxy spectrum has a power - law form in intensity , @xmath101 where @xmath50 is spatially varying with @xmath102 and @xmath103 as discussed earlier . linearizing the dependence of @xmath104 on @xmath105 we can write @xmath106 as n_gal(_a,_i ) = ( _ a,_i ) e(_i,_b ) where @xmath107 is the shot noise enhancement factor discussed earlier . the frequency dependences are given by f_4 ( ) & = & ^|-2 x^-2 e^-x(e^x-1)^2/f_4(350 ghz ) + f_5 ( ) & = & f_4 ( ) the cmb noise contribution is independent of frequency , and the instrument noise contributions as a function of frequency depend on the experiment being considered . the total noise in the map is then @xmath108 . the noise in the map therefore has the two - point function , considering points in the map @xmath82 at frequency @xmath83 and @xmath109 at frequency @xmath110 : @xmath111 ^ 2 f_5(\nu_a)f_5(\nu_b ) \\ \nonumber & & + \sigma^2(\nu_a)/\omega_{\rm pix}\delta_{ab}\delta_{i j}\end{aligned}\ ] ] in the above , @xmath112 is a shot - noise covariance matrix ( with off - diagonal correlations only due to beam - smoothing ) normalized to unit variance , and the last term gives the contribution of instrument noise , which will be considered in detail in 6 . @xmath113 is the covariance matrix of the cmb fluctuations , given by c^cmb_i j = _ l c_l p_l(_i j ) where @xmath114 is the angular separation between pixels @xmath115 and @xmath116 . this analysis implicitly assumes that we know the statistical properties of all sources of noise , including the dusty galaxies . this is an excellent approximation for the cmb fluctuations , but the dusty galaxy noise power spectrum will be affected by both clustering and gravitational lensing . on arcminute scales the clustering effects should be relatively small , but the effects of gravitational lensing will modify the noise properties . we forecast how well the three parameters of our model can be measured by calculating a fisher matrix . combining the spatial and spectral indicators @xmath115 and @xmath117 into a combined index @xmath4 ( and @xmath116 and @xmath118 into @xmath15 ) we can write the fisher matrix as f_pp = ( c^noise)^-1 _ to this fisher matrix we usually add a prior fisher matrix which carries the information we have that is not contained in our measurements in the 30 to 350 ghz range . the error covariance matrix for our three parameters is then given by c_pp = _ pp. note that since our data depend linearly on our parameters @xmath119 , this fisher matrix calculation of the expected error covariance matrix is exact , given our model of the data . we include prior information due to x - ray measurements of the gas temperature with error @xmath120 so that f_prior(1,1 ) & = & ( _ ) ^-2 + f_prior(3,3 ) & = & ( _ y)^-2 and + f_prior(1,3 ) = f_prior(3,1 ) & = & - . including the prior information in this manner is approximate since it results from taylor expanding the dependence of @xmath68 on @xmath94 and @xmath89 . to calculate errors on a new set of parameters , @xmath121 , that are functions of our original parameters @xmath119 , we taylor expand the dependence of @xmath121 on @xmath119 to first order about the fiducial value . we then calculate the fisher matrix in these new coordinates by : f_ii = _ pp r_ipf_ppr^t_pi where the transformation matrix @xmath122 . for example , such a variable transformation is necessary to get errors on @xmath123 . dropping the higher - order terms in the taylor expansion makes this procedure approximate as well . calculating these fisher matrices can be computationally expensive as matrices of size @xmath124 have to be inverted and multiplied where @xmath125 is the number of map pixels and @xmath126 is the number of frequencies . we therefore wish to pixelize no more finely than necessary , and keep the physical area of the map as small as possible . if the map is too small compared to the beam - convolved cluster profile we will lose the large - scale information necessary for subtracting off the cmb contamination . if the pixel size is too large we will lose small - scale information . we have found the following prescriptions to ensure we are neither losing information , nor using much larger @xmath125 than necessary : _ m & = & 4(2_c + _ b ) + _ pix & = & ( _ c+_b)/2.5 where the map and pixels are squares of length @xmath127 and @xmath128 respectively . our model is fairly detailed and sufficient in many ways , but there are several respects in which real clusters could present some challenges , such as temperature structure in the intra - cluster medium , incomplete knowledge of the lensing properties of galaxy clusters , and internal bulk flows in the cluster . temperature gradients in the intra - cluster medium would lead to a mis - match between the signal maps of equation [ eqn : signals ] . observed clusters show significant departures from isothermality @xcite at relatively large radii and at small radii , but the largest scales are already obscured by the primary cmb anisotropies and the small scales contribute relatively little to the sz effect . it is therefore not expected to be a significant effect but will complicate data analysis . similarly , we have assumed that we know the cluster template . in the cases where we assume complementary x - ray spectroscopy it is quite reasonable to assume that the x - ray image gives an estimate of the electron spatial distribution . without x - ray information the best estimate of the spatial template will come from the thermal sz map itself , where a large mismatch between the assumed spatial template and the observed emission should be evident . a parametrized model for the cluster could be introduced into the fit without significantly affecting the constraints on peculiar velocities . we have assumed that the magnification due to lensing is a simple step function in radius . the main effect of lensing is to amplify the poisson noise , and the amount of amplification depends on the details of the mass profile . incomplete knowledge of the mass distribution will therefore lead to less efficient component separation . deep optical images will be useful for the purpose of studying the strong lensing properties of galaxy clusters . on a related note , we have assumed that we know the statistical properties of the galaxy contamination . lensing modifies the noise properties of the background galaxies , and there is uncertainty in the statistics of dusty galaxies . multi - frequency observations of many fields , both with and without clusters , will lead to a good understanding of the statistics of dusty galaxies . we have also assumed a single peculiar velocity for the cluster whereas real clusters show evidence for internal flows , sometimes as large as 3 - 4000 km s@xmath129 @xcite . some striking visualizations of the kinetic sz fluctuations induced by these internal flows are given in @xcite . it has been shown @xcite that the average peculiar velocity provides an unbiased estimate of the true bulk velocity but with an added dispersion of roughly 100 kms@xmath129 . for our purposes this can be considered as an extra source of noise which is small compared to the uncertainties due to astrophysical confusion . we have neglected calibration uncertainty . since the total signal , anywhere other than near the thermal sz null , is about ten times as large as the kinetic sz signal , accurate calibration of one band relative to another is important . calibration errors will have to be controlled to better than about 10% for peculiar velocity measurements to be better than about one @xmath130 . the same argument applies to gas temperatures for clusters with temperatures near 6 kev , since the relativistic correction is also about 10% of the total signal . but gas temperatures , since they do not suffer from cmb anisotropy confusion , can be measured to much better than 1 @xmath130 . achieving @xmath131 kev , as can be done if we neglect calibration uncertainty , may require calibration uncertainties as small as 2% . measurements closer to the thermal sz null will reduce these sensitivities to calibration error . as our reference experiment we have modeled the sensitivity achievable for a generic ground - based bolometric experiment with three frequency bands selected to lie in the atmospheric windows available at a good millimeter site . we have also assumed that a 30 ghz measurement is available , with sensitivity similar to that expected from the sza . we assume that the sensitivity in each channel of the bolometric experiment is limited only by fluctuations in the photon background itself and that other potential sources of noise , such as phonon fluctuations in the bolometers or electronic readout noise , are negligible ( this can be achieved with current detector technology , see lange 2002 ) . we assume that @xmath132 noise introduced by fluctuations in atmospheric water vapor emission can be adequately subtracted from all of the data . this can be accomplished by some kind of spatial chopping or removal of common - mode signals from array data , or alternatively the spectral properties of the atmospheric noise can be used . the latter technique has been demonstrated by the sunyaev - zeldovich infrared experiment ( suzie ) which operates at 150 , 220 and 350 ghz and which has been used to set limits to the peculiar velocities of 11 galaxy clusters @xcite . for the purposes of this paper we do not include in our calculations any of the overhead ( increase in real observation time ) that this process will produce , since this will depend on the individual experiments . for example , since the atmosphere is in the near field of most large telescopes , many detectors view the same column of atmosphere , yet observe different parts of the cluster , or none of the cluster at all . by using detectors that are a large distance from the cluster to subtract common - mode fluctuations , it may be possible to subtract atmospheric noise with little or no degradation in sensitivity . this is a commonly - made assumption that we also include here . with the above assumptions the performance at each frequency is characterized by the noise equivalent power ( nep ; quoted in whz@xmath133 ) , given by : @xmath134 where @xmath135 is the total background loading on the detector in watts , @xmath136 is the detection bandwidth in ghz and @xmath137 is the number of waveguide modes that are detected ( @xmath138 for a diffraction - limited system ) . the first term in equation [ model : e1 ] is caused by shot noise due to poisson statistics of the incoming photons while the second term accounts for the effect of photon ( boson ) correlation . for more detail , see @xcite . sources of background loading include the atmosphere , warm emission from the telescope and surroundings , and of course the cmb itself . the total background loading is then : @xmath139 where the sum is over all sources of power on the detector . for each source of loading , the power at the detector can be calculated from : @xmath140 where @xmath141 is the equivalent rayleigh - jeans ( rj ) temperature of the @xmath115th background source , @xmath142 is the fraction of photons from the background source that reach the detector , @xmath143 is the effective telescope area and @xmath144 is the solid angle response of each detector on the sky . the integral is over the spectral band to which the detector responds . we assume for simplicity , a square band of width @xmath136 . we now consider each source of background power in turn . for the telescope the equivalent rj temperature @xmath145 is : @xmath146 where @xmath147 is the physical temperature of the telescope , @xmath148 is a frequency - dependent emissivity and @xmath149 . for the atmosphere : @xmath150\right\ } \label{model : e5}\ ] ] where @xmath151 is the physical temperature of the telescope , @xmath152 is the frequency - dependent zenith optical depth and @xmath153 is the zenith angle of the observation . for the cmb : @xmath154 where @xmath155k and @xmath156 . the optical efficiency is assumed to be : @xmath157 & \mbox{for $ i=3 $ ( cmb ) } \end{array } \right . \label{model : e7}\ ] ] which are values typical of those measured for bolometric systems @xcite . once the nep is determined , the sensitivity to cmb fluctuations , including the sz effect , is obtained by calculating the noise equivalent temperature ( net ) as follows : @xmath158 where @xmath159 is given above . we further assume that the focal plane is fully sampled with detectors spaced at half the beam size . each detector will have an angular response that is diffraction - limited at each frequency , in which case the solid angle response of each detector is just @xmath160 . with a suitable scanning strategy , an oversampled map can always be produced with a pixel solid angle @xmath161 . the rms sensitivity per map pixel is then : @xmath162 where @xmath163 is the integration time per detector and the net is per focal plane element . the factor @xmath164 accounts for the increase in sensitivity that a fully sampled focal plane can achieve over a focal plane with feed horns that are exactly diffraction - limited . we assume @xmath165 which is the ideal limit . the real value of @xmath164 depends upon factors such as detector noise and instrument photon background @xcite . we include a map pixel size here to allow us to compare experiments with different angular resolutions . if the array is large enough the scanning can be arranged so that there are always focal plane elements viewing the cluster . no integration time is lost in this case . other chopping schemes could degrade the sensitivity by a factor that depends on the details of the particular scheme being used . thus our final numbers may be taken as best case values . in order to use equations [ model : e1 ] through [ model : e9 ] , we now define the details of the experiments . for our reference experiment we assume a ground - based off - axis 8-m telescope located at a site with an altitude of 17,000@xmath166 and precipitable water vapor of 0.5 mm . the telescope and instrument are assumed to have an equivalent rayleigh - jeans temperature of @xmath167k that is frequency independent . figure [ model : f1 ] shows the dependence of @xmath168 on frequency , for a map pixel size of 1 sq arcmin . two models are shown , one with a bandwidth of 20% at each frequency and one with a bandwidth of 10% . there are broad atmospheric windows visible at 135 - 165ghz and at 215 - 290ghz . the window at 345ghz is sufficiently narrow that only a 10% wide frequency band can be utilized but it is a very useful atmosphere and/or point source monitor . based on this figure we have placed three of the four channels of the reference experiment at 150 , 220 and 280ghz . the sensitivities are shown in table [ tab : experiments ] . .assumed experimental specifications and sensitivities per square arcminute pixel . [ cols="^,^,^,^ " , ] we begin with our fiducial cluster , studying the error forecasts as a function of confusion noise and spectral index rms in fig . [ fig : dusty ] . for the lower @xmath169 values one can infer from the flattening of the error contours with increasing shot noise amplitude that the data themselves are capable of fitting for the dusty galaxy component with residuals at about the 100 to 200 @xmath4 k level . if instrument noise were to increase , this transition from dependence on @xmath170 to independence would occur at higher values . the trend of error level with @xmath169 is as expected : flat at low @xmath171 and steeper at higher values . peculiar velocities are more affected , particularly at non - zero @xmath169 , than the other parameters . we now fix the point source contamination parameters and study our results as a function of galaxy cluster core radius , @xmath172 , and gas temperature , @xmath173 . note that as we vary @xmath172 the map size we assume increases in order to assure complete coverage of the cluster and good cmb subtraction . for large - format bolometer arrays , expected to have more than 1000 elements , the larger map size does not require longer integration times because the field of view will be much larger than the galaxy cluster . therefore , despite the added coverage , one hour of observation is still sufficient to obtain the reference experiment sensitivities in table 1 . on the left of fig . [ fig : reference ] we set the point source contamination parameters to zero and on the right to their fiducial values @xmath174k - arcmin and @xmath175k - arcmin . because of the linear response of the signal to @xmath0 , @xmath176 is independent of @xmath0 and therefore @xmath173 as seen in the left side of fig . [ fig : reference ] . in contrast , at fixed @xmath0 and @xmath177 the signal is proportional to @xmath178 so one expects @xmath179 . confusion with @xmath0 , more important at high @xmath173 , leads to the error decreasing more slowly than @xmath180 . the @xmath181 has a similar dependence on @xmath173 since the negligible error on @xmath0 means @xmath182 and our scaling assumes @xmath29 . for fixed three - dimensional distribution of gas pressure , central @xmath0 values do not vary as a function of redshift . core radii do as @xmath183 where @xmath184 is the angular - diameter distance to redshift @xmath185 . thus to study dependence on @xmath185 we vary @xmath172 with fixed @xmath0 . note that @xmath186 0.5 , 0.7 , 1.3 , 1.4 for @xmath187 , 0.3 , 1 , 2 respectively , assuming @xmath188 and @xmath189 . note that the number of pixels with sz signal in any given range scales as @xmath190 , so the errors on @xmath3 , @xmath0 and @xmath173 decrease as @xmath191 . the peculiar velocities on the other hand suffer from contamination from the cmb . as @xmath172 increases it becomes more difficult to distinguish the signal proportional to @xmath177 from the very red cmb power spectrum , hence @xmath192 goes up . at large @xmath172 , where the error in @xmath97 is dominated by the cmb - induced error in @xmath177 , the error in @xmath3 is negligible . in this case , @xmath193 so @xmath194 , which is the scaling we see . at smaller @xmath172 uncertainty in @xmath3 is no longer negligible and the scaling with @xmath173 is more complicated . including the effect of the fiducial point source contamination ( right side of fig . [ fig : reference ] ) does not alter the error forecasts dramatically . for @xmath1 the largest impact is at the smaller @xmath172 which is unfortunate since this is where @xmath1 is measured best . at small @xmath172 the cmb noise is smaller and ( less importantly ) the lensing enhancement of confusion noise is largest . at @xmath195 the velocity errors are almost doubled . figure [ fig : reference_time ] shows how the error forecasts depend on observing time . on the left side one can see @xmath196 indicating that for the reference experiment the dominant source of uncertainty on @xmath0 is instrument noise , not dusty galaxies . similar scalings are seen for @xmath181 and @xmath197 . the @xmath198 in contrast , quickly saturate with increased observing time producing very little improvement since the dominant source of error is the cmb contamination . the saturation occurs even sooner for larger @xmath172 . as can be seen in fig . [ fig : reference_fwhm ] parameters other than @xmath1 are highly independent of the angular resolution @xmath33 . this is because none of the signal is lost with increasing @xmath33 : the signal is simply spread out over greater area . errors would go up with increasing @xmath33 for single - pixel observations as opposed to the map - making observations we consider . the peculiar velocity dependence on @xmath33 is steep for @xmath199 , since spreading out the signal degrades the ability to subtract off the cmb contamination . note that to explore the effect of angular resolution independent from sensitivity , we have not scaled the sensitivity with beam size as one would expect from eq . [ model : e9 ] , but rather fix the weight per solid angle . the apparent decrease in @xmath198 with increasing @xmath33 at fixed @xmath200 is an artifact of our sparse sampling of the @xmath172 , @xmath33 space . we now consider the importance of each of the channels of our reference experiment . on the left side of fig . [ fig : reference_minus ] we show error forecasts for the reference experiment with no 30 ghz channel , ` reference-30 ' . without this channel it is harder to distinguish a change in @xmath0 from a change in @xmath173 , thus the errors in both of these increase considerably . their ratio , @xmath201 , suffers similarly . the impact on peculiar velocities is less dramatic , especially at high @xmath172 where the cmb contamination remains dominant . changes are significant at smaller @xmath172 where , e.g. , @xmath198 goes from 250 to 400 km / sec at @xmath202 kev . the strong degeneracy between @xmath0 and @xmath173 for reference-30 can be seen on the right side of fig . [ fig : reference_minus ] . the correlation coefficient , @xmath203 increases from 0.93 for reference to 0.998 for reference-30 . this near - maximal correlation coefficient means that the error in @xmath173 ( @xmath0 ) is increased by a factor of 14 compared to the error in @xmath173 ( @xmath0 ) if @xmath0 ( @xmath173 ) were held fixed . including a reasonable prior on the velocity of 300 km / sec decreases the @xmath173 and @xmath0 errors by about 20% . the importance of a low frequency measurement for determining @xmath173 has been emphasized by @xcite and @xcite . after the 30 ghz channel , the most critical channels in decreasing order are 150 ghz , 220 and 280 . removal of the 150 and 220 ghz channels can increase @xmath197 by up to a factor of 5 , although only about a factor of 2 at our small fiducial core radius of @xmath195 assumed in the right side of fig . [ fig : reference_minus ] . these increases , at the assumed level of instrument noise , are not sufficient to degrade the peculiar velocity errors . the dashed line on the right side of fig . [ fig : reference_minus ] is @xmath28 , through our fiducial values of @xmath0 and @xmath173 . we see that the degeneracy direction is such that assuming the scaling relation and a normalization would lead to much more precise determinations of both @xmath0 and @xmath173 . on the other hand , the orientation of the degeneracy makes testing the scaling relation more difficult than if it lay along the line . = .67 we saw that for the reference experiment the loss of the 30 ghz channel increased @xmath197 , as well as @xmath198 for small @xmath172 . the lack of a low frequency channel can therefore be compensated by an x - ray determination of the temperature . thus we show forecasts for the act survey and suzie - iii observations both complemented by x - ray temperature determinations of @xmath131 kev . the results for the two different experiments are very similar . we assume one hour of integration with suzie - iii for every 48 sq . arcmin covered . recall that in order to assure good subtraction of the cmb , we assume a map of linear extent @xmath204 . we have made no attempt to optimize this observing strategy . the right - most figure in figure [ fig : planned ] shows that adding a 30ghz measurement to the millimeter data can break this degeneracy , as we expect from our previous forecasts for the reference experiment in fig . [ fig : reference ] . in 1 hour of observation with suzie iii ( per 48 sq . arcmin of map ) and a 12-hour sza observation , the peculiar velocity of a cluster can be measured to 250 km s@xmath129 _ even without an x - ray measurement of temperature_. a low frequency channel will thus be crucial for obtaining peculiar velocity and gas temperatures at high redshift where @xmath205-ray brightness is very low . we also see in figure [ fig : planned ] a temperature error for our fiducial cluster of @xmath206 kev ( and even smaller for larger @xmath172 ) . at this accuracy level , one expects to begin to see discrepancies between x - ray temperature measurements which are more sensitive to the hotter , inner regions and sz temperature measurements which are more sensitive to the cooler , outer regions @xcite . deeper integrations can slowly improve these measurements as seen from the study of dependence of forecasts for the reference experiment on observation time . we have assumed that ( i ) the frequency spectrum of cluster radio sources will be sufficiently steep at high frequencies that measurements at 150 ghz and higher will not be contaminated and ( ii ) that measurements at 30 ghz will be sufficiently high resolution that point sources can be subtracted at that frequency ( this is possible with interferometric experiments such as the sza ) . consequently we have concentrated only on the contamination introduced by dusty galaxies known to be strong submillimeter - wavelength emitters and have excluded radio sources from our analysis . we will now justify this omission . clusters often contain radio point sources at mjy levels , as seen from 30 ghz sz measurements @xcite . it is important to note that radio point sources and submillimeter point sources are _ not _ the same objects . the excess of radio sources toward galaxy clusters is most likely due to emission from cluster galaxies themselves @xcite , whereas the dusty galaxies are typically background sources that are not associated with cluster members . very little is known about the emission from radio point sources at millimeter wavelengths , but their spectra are expected to steepen due to the energy losses of the most energetic electrons to synchotron emission . this is confirmed to some extent by spectral indices measured between 10 - 90 ghz for bright sources @xcite , although there is substantial scatter in the measured index . @xcite find a mean index of @xmath2070.5 between 10 and 90 ghz , while @xcite find a mean index at 40 ghz of @xmath208 . measurements by @xcite of the spectral indices of the 208 point sources detected by wmap at 40 ghz @xcite , suggest a mean of @xmath209 at 20ghz with large scatter , but also show that the spectra of radio sources are often not well - approximated by a simple power law . sources of only a few mjy at 30 ghz could easily still have mjy fluxes at 150 ghz if the spectrum does not steepen significantly above 90 ghz . we have used the point source counts from wmap , dasi , vsa and cbi summarized in @xcite to determine a best fit model of : @xmath210 where @xmath211 per sq . deg . and @xmath212 mjy . our expression gives number counts for mjy sources that are slightly lower than @xcite , but that are consistent with those determined by @xcite based on sz measurements of clusters at 30ghz . consequently we expect that the number of sources above 0.1 mjy at 30 ghz is less than one per 4 sq arcmin ( roughly the size of a moderate cluster ) . repeating our fisher analysis with radio sources at this level , we find that for all the experimental situations we have considered , the peculiar velocity uncertainty for our canonical source is unaffected , even if a spectral index of @xmath213 is assumed , and the number counts are 10 times higher than we have assumed . the spectra of mjy radio sources will soon be much better understood when the sza and the array for microwave background anisotropy ( amiba ) start operation at 90 ghz . however , a complete understanding of faint radio sources at mm wavelengths will likely require the sensitivity and resolution of the atacama large millimeter array ( alma ) . at roughly the level of 1 kev the distinction between x - ray emission - weighted temperatures and electron - weighted temperatures becomes important @xcite , so a comparison of the two `` gas temperatures '' could be valuable for studies of the physics of the intra - cluster medium . furthermore , the pressure - weighted temperature should be an excellent indicator of cluster mass . this could allow independent calibration of the mass scale for galaxy cluster surveys that use x - ray , optical or sz selection criteria . the measurement of peculiar velocities using the kinetic sz effect has been a goal of sz measurements for many years . peculiar velocities at large distances will allow measurements of cosmological parameters and reconstruction of the large scale gravitational potential . the two - point function of radial peculiar velocities is sensitive mainly to @xmath214 and largely insensitive to the dark energy equation - of - state parameter , @xmath215 or other cosmological parameters . a linear theory calculation of how well the two - point function can constrain @xmath214 gives @xcite [ eqn : peelknox ] _ m = 0.04 . the suzie iii and act collaborations plan to measure optical redshifts for 400 clusters with masses greater than @xmath216 . since typical errors for peculiar velocities on these may be 300 km / sec and the velocity rms is about 400 km / sec , these data may be able to achieve @xmath217 . however , equation [ eqn : peelknox ] assumes that the variance in our peculiar velocity measurement errors is perfectly well known . in reality , the statistical properties of the errors will be difficult to know well because of their dependence on the subtraction of the contamination of dusty galaxies . if the assumed error variance is different from the actual error variance by @xmath218 then there will be a systematic error in @xmath214 of _ m = f()^-1 0.02 ( f/0.1 ) independent of the number of clusters measured . to control this systematic error , we will need to learn more about the spectral properties and luminosity functions of sub - mm galaxies . to avoid this source of systematic error one could use a linear statistic , the average difference between the radial velocities of two clusters separated by distance @xmath219 @xcite . how this statistic depends on cosmological parameters is under investigation . another application of peculiar velocities is gravitational potential reconstruction @xcite . as @xcite point out , due to the low number density of clusters this is only possible on very large scales . comparison with reconstruction from galaxy number counts can determine galaxy biasing properties . a good understanding of the large scale potential would be valuable for investigating environmental effects on galaxy formation and would be a starting point for constrained realizations of simulations of the large scale structure of our universe . contamination by emission from dusty galaxies will make accurate measurements of peculiar velocities difficult . with four observing frequencies , few @xmath6 sensitivity , and arcminute resolution it will be possible to measure peculiar velocities at the level of roughly 200 km / s for massive galaxy clusters . for comparison , previous estimates of uncertainties in peculiar velocities by holder ( 2003 ) assumed @xmath220 sensitivity and perfect point source removal , which on figure [ fig : reference ] roughly correspond to @xmath221 and an observing time of several tens of hours . in that case it was found that uncertainties well below 100 km / s were possible , as could be extrapolated from figure [ fig : reference ] . such exposure times may be prohibitively long and removing point source contributions to this level will likely require alma . alma will provide both a better understanding of the physical nature of dusty galaxies in general as well as measurements of the contaminating galaxy fluxes at exactly the frequencies of interest . the reference experiment , as well as combinations of planned observations , can provide measurements of peculiar velocities at the level of 150 - 200 km / s , even in the presence of contamination by dusty galaxies . this is sufficient accuracy to constrain @xmath214 to better than 10% . at the same time galaxy cluster temperature measurements will be possible at the kev level . this will allow interesting comparisons with x - ray spectroscopic measures of the x - ray emission - weighted temperature . peculiar velocities measured using the kinetic sz effect , a long - standing challenge to cmb experimentalists , may soon be a reality . the final limiting obstacle appears to be contamination from dusty galaxies , but sensitive multi - frequency measurements should clear this hurdle . + , c. l. , hill , r. s. , hinshaw , g. , nolta , m. r. , odegard , n. , page , l. , spergel , d. n. , weiland , j. l. , wright , e. l. , halpern , m. , jarosik , n. , kogut , a. , limon , m. , meyer , s. s. , tucker , g. s. , & wollack , e. 2003 , , 148 , 97 , m. c. , ade , p. a. r. , bhatia , r. s. , bock , j. j. , daub , m. d. , goldstein , j. h. , haynes , c. v. , holzapfel , w. l. , kuo , c. l. , lange , a. e. , leong , j. , lueker , m. , newcomb , m. , peterson , j. b. , ruhl , j. , sirbi , g. , torbet , e. , tucker , c. , turner , a. d. , & woolsey , d. 2003 , , submitted , astroph/0303515

the copernican principle implies that the earth , and , by extension , the solar system , do not hold a central or specifically favored position . this viewpoint is related to the so - called _ mediocrity principle _ @xcite , which notes that an item drawn at random is more likely to come from a heavily populated category than one which is sparsely populated . these principles , however , have not had particularly apparent success when applied in the context of extrasolar planets . @xcite used their high precision doppler survey data to deduce that of order 50% ( or more ) of the chromospherically quiet main - sequence dwarf stars in the solar neighborhood are accompanied by a planet ( and in many cases , by multiple planets ) with @xmath3 , and orbital periods of @xmath4 . taken strictly at face value , this result implies that our own solar system , which contains nothing interior to mercury s @xmath5 orbit , did not participate in the galaxy s dominant mode of planet formation . yet the eight planets of the solar system have provided , and continue to provide , the _ de - facto _ template for most discussions of planet formation . indeed , where extrasolar planets are concerned , m - dwarfs and mediocrity appear to be effectively synonymous . recent observational results suggest that low - mass planets orbiting low - mass primaries are by no means rare . numerous examples of planets with @xmath6 and m - dwarf primaries have been reported by the doppler surveys ( e.g. @xcite , and many others ) , and the _ kepler _ mission has indicated that small planets are frequent companions to low mass stars . for example , @xcite report that among dwarf stars with @xmath7 , the occurrence rate of @xmath8 planets with @xmath9 is @xmath10 planets per star . improved statistics , however , are required for a definitive statement that is couched in planetary masses as well as in planetary radii . figure [ fig : mdwarfcensus ] shows the current distribution of reported planets and planetary candidates orbiting primaries with @xmath11 , which we adopt as the functional border between `` m - type '' stars and `` k - type stars '' . the census of low - mass planets orbiting low - mass primaries can be accessed using a variety of techniques . for objects near the bottom of the main sequence , it appears that transit photometry from either ground @xcite or space @xcite offer the best prospects for planetary discovery and characterization . for early to mid m - type dwarfs , there is a large enough population of sufficiently bright primaries that precise doppler detection ( see , e.g. @xcite ) can play a lead role . for the past decade , we have had a sample of @xmath12160 nearby , photometrically quiet m - type stars under precision radial velocity surveillance with the keck telescope and its hires spectrometer . in recent months , this survey has been supplemented by data from the automated planet finder telescope @xcite . here , we present 16.6 years of doppler velocity measurements for the nearby m3 dwarf gj 687 ( including 122 velocity measurements from keck , 20 velocity measurements from the apf , and 5 velocity measurements made with the hobby - eberly telescope ) and we report the detection of the exoplanet that they imply . we use this discovery of what is a highly archetypal representative of a planet in the milky way in terms of its parent star , its planetary mass , and its orbital period to motivate a larger discussion of the frequency of occurrence , physical properties , and detectability of low - mass planets orbiting m - type stars . the plan for this paper is as follows . in 2 , we describe the physical and spectroscopic properties of the red dwarf host star gliese 687 . in 3 , we describe our radial velocity observations of this star . in 4 we describe our keplerian model for these observations , along with an analysis that assesses our confidence in the detection . in 5 , we describe our photometric time series data for the star , which aids in the validation of the planet by ruling out spot - modulated interpretations of the doppler variations . in 6 , we discuss the ongoing refinement of the planet - metallicity correlation for low - mass primaries , in 7 , we discuss the overall statistics that have emerged from more than 15 years of precision doppler observations of m - dwarf stars with the keck telescope , and in 8 we conclude with an overview that evaluates the important future role of the apf telescope in precision velocimetry of nearby , low - mass stars . gliese 687 ( lhs 450 , bd+68@xmath13946 ) lies at a distance , @xmath14 pc , is the 39th - nearest known stellar system , and is the closest star north of @xmath15 declination . figure [ fig : hrdiag ] indicates gj 687 s position in the color - magnitude diagram for stars in the lick - carnegie survey s database of keck observations . due to its proximity and its brightness ( v=9.15 , @xmath16=4.548 ) , gliese 687 has been heavily studied , and in particular , the chara array has recently been used to obtain direct interferometric angular diameter measurements for the star . @xcite find @xmath17 , and derive @xmath18 , @xmath19 , and use the mass - radius relation of @xcite to obtain @xmath20 . as illustrated in figure [ fig : sindex ] , gliese 687 s mean mt . wilson s value and the dispersion of its s - index measurements from our spectra indicate that it has a moderate degree of chromospheric activity . this conclusion is in concordance with our long - term photometric monitoring program , which also indicates that the star is somewhat active . ccc spectral type & m3 v & @xcite + mass ( @xmath21 ) & @xmath22 & @xcite + radius ( @xmath23 ) & @xmath24 & @xcite + luminosity ( @xmath25 ) & @xmath26 & @xcite + distance ( pc ) & @xmath27 & @xcite + @xmath28 & 1.5 & simbad + @xmath29 mag . & 9.15 & @xcite + @xmath30 mag . & 5.335 & @xcite + @xmath31 mag . & 4.77 & @xcite + @xmath32 mag . & 4.548 & @xcite + avg . s - index & 0.811 & this work + @xmath33 & 0.096 & this work + @xmath34 ( days ) & @xmath35 & this work + @xmath36 ( k ) & @xmath37 & @xcite + doppler shifts from both the keck ( 122 observations ) and apf ( 20 observations ) platforms were measured , in each case , by placing an iodine absorption cell just ahead of the spectrometer slit in the converging beam of stellar light from the telescope @xcite . the forest of iodine lines superimposed on the stellar spectra generates a wavelength calibration and enables measurement of each spectrometer s point spread function . the radial velocities from keck were obtained by operating hires at a spectral resolving power r@xmath1270,000 over the wavelength range of 3700 - 8000 @xmath38 , though only the region 5000 - 6200 @xmath38 containing a significant density of iodine lines was used in the present doppler analysis @xcite . the apf measurements were obtained over a similar spectral range , but at a higher spectral resolving power , r@xmath12108,000 . for each spectrum that was obtained , the region containing the iodine lines was divided into @xmath12700 chunks , each of @xmath39 width . each chunk produces an independent measure of the wavelength , psf , and doppler shift . the final measured velocity is the weighted mean of the velocities of the individual chunks . all radial velocities ( rvs ) have been corrected to the solar system barycenter , but are not tied to any absolute velocity system . as such , they are `` relative '' velocities , with a zero point that can float as a free parameter within an overall system model . the internal uncertainties quoted for all the radial velocity measurements in this paper reflect only one term in the overall error budget , and result from a host of errors that stem from the characterization and determination of the point spread function , detector imperfections , optical aberrations , consequences of undersampling the iodine lines , and other effects . two additional major sources of error are photon statistics and stellar `` jitter '' . the latter varies widely from star to star , and can be mitigated to some degree by selecting magnetically inactive older stars and by time - averaging over the star s unresolved low - degree surface @xmath40-modes . all observations in this paper have been binned on 2-hour timescales . in addition to the radial velocities that we have obtained at keck and apf , we also use five doppler measurements obtained by @xcite at the hobby - eberly telescope located at mcdonald observatory . these radial velocity observations are presented in the appendix . the combined radial velocity data sets show a root - mean - square ( rms ) scatter of 7.58 @xmath41 about the mean velocity . this scatter is measured after we have applied best - fit telescope offsets of 0.64 @xmath42 for keck , -1.71 @xmath42 for apf , and 1.27 @xmath42 for het . a lomb - scargle periodogram of the 149 velocity measurements of gj 687 is shown in figure [ fig : periodogram ] . false alarm probabilities are calculated with the bootstrap method , as described in @xcite , iterating 100,000 times for a minimum probability of @xmath43 as easily met by the tallest @xmath44=38.14 day peak in figure [ fig : periodogram ] . this signal in the data is modeled as a @xmath45 planet with an orbital eccentricity , @xmath46 . lllr period ( d ) & 38.14&(0.015 ) + mass ( @xmath47 ) & 0.058&(0.007 ) + mass ( @xmath48 ) & 18.394&(2.167 ) + mean anomaly ( deg ) & 234.62&(87.962 ) + eccentricity & 0.04&(0.076 ) + longitude of periastron ( deg ) & 359.43&(120.543 ) + semi - major axis ( au ) & 0.16353&(0.000043 ) + & & & + time of periastron ( jd ) & 2450579.11&(9.32 ) + & & & + rv half amplitude ( @xmath42 ) & 6.43&(0.769 ) + & & + first observation epoch ( jd ) & 2450603.97 & + velocity offsets & & + keck / hires & 0.64 @xmath49 & ( 0.63 ) + apf / levy & -1.71 @xmath49 & ( 1.68 ) + het & 1.27 @xmath49 & ( 0.98 ) + @xmath50 & 18.55 & + rms & & + keck / hires & 6.62 @xmath49 & + apf / levy & 3.95 @xmath49 & + het & 2.44 @xmath49 & + jitter & 5.93 @xmath49 & + using levenberg - marquardt optimization , we obtained a best - fit keplerian model for the system . this fit , which assumes @xmath51 and @xmath52 for the planet , is listed in table [ tab : fit ] . the phased rv curve for the planet in table [ tab : fit ] is shown in figure [ fig : phasedcurves ] . a power spectrum of the residuals to our one - planet fit is shown in figure [ fig : residperiodogram ] and indicates no significant periodicities . also shown in this figure is a periodogram of our mt . wilson s - index measurements from the spectra , which are a proxy for the degree of spot activity on the star at a given moment . none of the peaks in the periodogram of s - index values coincide with the peak that we suspect to be a planet . the reduced chi - squared statistic for our fit is @xmath53 18.55 and results in a fit with a combined rms of 6.16 @xmath42 and estimated excess variance of @xmath54 = 5.93 @xmath42 ( the estimate of the jitter that is required to bring the reduced chi - squared statistic of the fit down to unity ) . this value accounts for variance in both the stellar signal and from the telescope itself , though for a moderately active star such as gj 687 a stellar jitter of order 6.0 @xmath42 is reasonable , and could account for the majority of the observed variance . in order to compute parameter uncertainties for our orbital fit , we implement a markov chain monte carlo algorithm @xcite . the mcmc algorithm returns a chain of state vectors , @xmath55 ( a set of coupled orbital elements , e.g. period , mass , etc . and the three velocity offset parameters ) . the goal of the markov chain calculation is to generate an equilibrium distribution proportional to @xmath56 $ ] . we adopt non - informative priors on all parameters ( and uniform in the log for masses and periods ) . the resulting error correlations are shown in figure [ fig : errorcorr ] , and a set of 100 states drawn randomly from the converged chain are shown in the orbital diagram in figure [ fig : orbit ] . the error correlation diagram indicates that all parameters are well determined , save the usual degeneracies between mean anomaly and @xmath57 for the low - eccentricity orbit . the distribution of the residuals relative to the best - fit model shows no evident pathologies . indeed , a quantile - quantile plot ( shown in figure [ fig : fitplot ] ) indicates that the distribution of residuals is well described by a normal distribution . we note that the smaller scatter of points obtained with the apf telescope could be a consequence of the fact that they were all taken within a @xmath58 = 140d period , and thus sample only one segment of the stellar activity cycle . a potentially significant challenge to correctly identifying the orbital period of a proposed exoplanet arises from the discrete and uneven sampling inherent in radial velocity surveys . the spacing of observations leads to increased noise and the presence of aliases within the star s periodogram which can be mistaken for a true orbital signature . for a real signal occurring at a frequency @xmath59 we expect alias signatures at @xmath60 where n is an integer . in order to aid confirmation that the periodic signal we observe is actually a planetary signature , we must be able to calculate where aliases due to our observing cadence will occur , and then verify that they are not the source of the signal . the aliases are determined using a spectral window function as defined by @xcite @xmath61 where n is the total number of observations and t is the date on which they were taken . plotting this function will result in peaks that are due solely to the sampling cadence of the data . because our observations are constrained by when the star is visible in the night sky , and because keck telescope time is allocated to doppler surveys primarily when the moon is up , we expect aliases at periods of 1 solar day , 1 sidereal day , 1 synodic month and 1 sidereal year . examining the window function in figure [ fig : windowfunc ] we do see peaks resulting at these periods , but careful analysis of the periodogram for our radial velocity observations shows no evidence of strong signals occurring at the locations necessary for our @xmath1 day signal to be a potential alias instead of a true keplerian signature . with an apparent @xmath16-band magnitude of 4.54 , gliese 687 is brighter ( in the near infrared ) than _ all _ known hosts of transiting extrasolar planets other than 55 cancri . as a consequence , transits by gliese 687 s planetary companion ( which has an equilibrium temperature , @xmath62 ) , were they to occur , would be of substantial scientific value . in particular , transmission spectroscopy with jwst would give insights into what is likely a dynamic and chemically rich planetary atmosphere . the _ a - priori _ geometric transit probability for gliese 687 b , however , is a scant @xmath63% , and as we describe below , there is no evidence that transits occur . with @xmath64 , the currently observed mass - radius range for exoplanets indicates that the planetary radius , @xmath65 could credibly range from @xmath66 to @xmath67 , implying potential transit depths in the @xmath68% to @xmath69% range . during the 20092013 observing seasons , we acquired a total of 866 photometric observations of gj 687 on 519 nights with the tennessee state university ( tsu ) t12 0.80 m automatic photoelectric telescope ( apt ) at fairborn observatory in arizona . the t12 apt is one of several tsu automatic telescopes operated at fairborn @xcite . it is equipped with a two - channel precision photometer that employs a dichroic filter and standard strmgren @xmath70 and @xmath71 filters to separate the two passbands and two emi 9124qb bi - alkali photomultiplier tubes to measure the @xmath70 and @xmath71 count rates simultaneously . we observed gj 687 , designated our program star ( p ) , differentially with respect to three neighboring comparison stars : c1 ( hd 156295 , @xmath72 , @xmath73 , f0 iv ) , c2 ( hd 160198 , @xmath74 , @xmath75 , f2 v ) , and c3 ( hd 161538 , @xmath76 , @xmath77 , f2 v ) . a detailed description of the observing sequence and the data reduction and calibration procedures are given in @xcite . we computed all pairwise differential magnitudes @xmath78 , @xmath79 , @xmath80 , @xmath81 , @xmath82 and @xmath83 in both the @xmath70 and @xmath71 passbands , corrected them for atmospheric extinction , and transformed them to the standard strmgren photometric system . observations with internal standard deviations greater than 0.01 mag were discarded to remove data taken in non - photometric conditions . intercomparison of the six sets of differential magnitudes demonstrated that hd 156295 ( c1 ) is a low - amplitude variable while both hd 160198 ( c2 ) and hd 161538 ( c3 ) are constant to the expected measurement precision . to improve our precision , we combined the separate differential @xmath70 and @xmath71 observations into a single @xmath84 passband . " we also computed the differential magnitudes of gj 687 with respect to the mean brightness of the two good comparison stars : @xmath85 . the standard deviation of the @xmath81 comparison star differential magnitudes is 0.0020 mag , which we take to be the precision of a single measurement . a total of 606 nightly measurements in the five observing seasons survived the cloud - filtering process . these data are plotted as @xmath85 differential magnitudes in the top panel of figure [ fig : phot1 ] . the five individual observing seasons are plotted in the remaining panels . the standard deviations for the yearly light curves are given in each panel . these range from 0.0049 to 0.0092 mag , compared to the measurement precision of 0.0020 mag . gaps of 1012 weeks in the yearly light curves for 2009 through 2012 are due to southern arizona s july september rainy season when good photometry is not possible . low - amplitude variability is seen in gj 687 during each observing season , resembling light curves typical of modestly active stars with spot filling factors of a few percent ( see , e.g. , * ? ? ? the 2010 light curve has the largest amplitude variability ( @xmath86 mag ) and reveals cyclic variation with a time scale of @xmath87 days . the other light curves have lower amplitudes and include cyclic variations of @xmath87 and also @xmath88 days . these year - to - year and cycle - to - cycle variations are also typical of modestly active stars . we interpret the 60-day variability as the signature of the star s rotation period and the 30-day variability as a sign of spot activity on opposite hemispheres of the star . frequency spectra of the complete 2009 - 2013 data set and of the 2010 data alone are shown in the top and bottom panels of figure [ fig : phot2 ] respectively . the rotational modulation signal is seen most clearly in the 2010 data , which matches up with the most coherent light curve in figure [ fig : phot1 ] . therefore , we take the @xmath89d signal as our best measurement of the star s rotation period . inspection of the 2010 photometric segment of figure [ fig : phot1 ] clearly shows the overall 60-day modulation that generates the periodogram peak . departures from perfect periodicity are presumably caused by the evolution of the spot activity on the surface of the star . finally , we search for transits of gj 687 b by first removing the spot variability from each of the yearly light curves . we do this by successively subtracting multiple frequencies from each yearly light curve using the method described in @xcite . we removed three to six frequencies from each light curve until each set of residuals approached the precision of a single observation . the residuals from all five observing seasons are plotted in the top panel of figure [ fig : phot3 ] , phased with the 38.14-day best - fit planetary orbital period and a time of mid transit computed from the orbital parameters . the vertical bar represents the 0.0022 mag standard deviation of the residuals from their mean , very close to the measurement precision given above . a sine fit to the phased data gives a formal semi - amplitude of just @xmath90 mag . since none of the frequencies removed from the yearly light curves were similar to the orbital frequency or its harmonics , this result limits any periodic brightness variability of the star on the observed radial velocity period to a very small fraction of one milli - magnitude ( mmag ) . this rules out the possibility that the 38.14-day radial velocity variations in gj 687 are induced by stellar activity , as has been documented in somewhat more active stars , for instance , by @xcite , @xcite , and @xcite . instead , this lack of photometric variability confirms that the radial velocity variations in gj 687 result from true planetary reflex motion . the photometric observations within @xmath91p of mid - transit are replotted with an expanded scale in the bottom panel of figure [ fig : phot3 ] . the solid curve shows the predicted phase , depth ( assuming neptune - like density ) , and duration of a central transit , computed from the stellar radius in table [ tab : stellarparams ] and the orbital elements in table [ tab : fit ] . the horizontal error bar under the predicted transit time gives the @xmath92 uncertainty in the timing of the transit . the photometric observations when filtered using the @xcite procedure described above , and when folded at the @xmath1 day best - fit period for the planet , give no indication that transits occur . we note , however , that the markov chain models generate a five - day window for possible transits , and so a more conservative approach is also warranted . in figure [ fig : transit ] , we plot the unfiltered photometric data , indicating the range of photometric points that potentially could have been affected by transits were they to occur . because of uncertainties in the orbit , the potential transit duration , the potential size of the planet , and the error in the photometric filtering , we recommend that continued photometric monitoring be carried out to confirm that transits do not occur . gliese 687 appears to have a slightly sub - solar metallicity . @xcite use na i , ca i , and h@xmath93o - k@xmath94 calibrations to estimate [ fe / h]=-0.09 for gliese 687 , whereas the m - dwarf metallicity calibration of schlaufman & laughlin 2010 yields a value [ fe / h]=-0.02 . the connection between the detectable presence of a giant extrasolar planet and the metallicity of the host star was noticed soon after the first extrasolar planets were detected @xcite , and has been studied in many previous works , see , e.g. @xcite . for m dwarfs , recent work , such as that by @xcite , suggests that the giant planet stellar metallicity correlation holds robustly for m - dwarf primaries , but that for planets with mass , @xmath95 , no correlation is found with host star metallicity , and indeed , @xcite report a hint of _ anti_-correlation between the presence of a low - mass planet and host star @xmath96 $ ] . our detection of a neptune - mass companion to gliese 687 , and our lick - carnegie database of doppler measurements of m dwarf stars provides an opportunity to revisit this topic . our database of radial velocity observations taken at the keck telescope contains 142 m - type stars with the necessary spectral information to assess metallicity , 17 of which are known to host planets published in the peer - reviewed literature . we break the planet - hosting stars into two subgroups based on their masses - stars with @xmath97 planets less than 30 @xmath98 are described as neptune hosting while stars with @xmath97 planets greater than 30 @xmath98 are listed as jupiter hosting . we replicate the procedure of @xcite and examine how horizontal distance from a field m dwarf main sequence in a @xmath99 vs. @xmath100 color - magnitude diagram ( cmd ) correlates with metallicity , as noted e.g. , by @xcite . the top panel of figure [ fig : metallicity ] displays all of the lick - carnegie survey m dwarf stars plotted in @xmath99 vs. @xmath100 , with grey dots denoting survey stars without known planets , red dots denoting survey stars that host `` neptune - mass '' planets and blue dots representing the survey stars that host `` jupiter - mass '' planets . it can be seen that most planet hosting stars fall to the right of the field m dwarf main sequence presented in @xcite ( black line ) , which is taken to be a [ fe / h]=0.017 isometallicity contour in this cmd . in order to quantify the likelihood that a star s horizontal distance from the isometallicity contour is related to its propensity to host planets , we compare the distances for our actual planet - hosting stars with randomly drawn samples from the collection of m dwarfs in the survey . we characterize the position of each m dwarf by obtaining v - band and @xmath101 photometry and then using them to calculate the distance statistic @xmath102 : @xmath103 to determine if the @xmath102 of our known planet hosting subgroups is significant or , alternatively , if it could be produced by chance , we make use of a monte carlo simulation that calculates the cumulative sample distance of survey m dwarfs from the field m dwarf main sequence presented by johnson and apps 2009 . for the simulation , we randomly select a subset of m dwarfs from the lick - carnegie field star list , setting the sample size equal to the number of m dwarfs known to host either jovian or neptune mass planets . then we compute the cumulative horizontal distance of those stars from the field m dwarf ms , where stars to the right of the ms add their distance to the sum and stars to the left of the ms subtract their distance . we repeat this process 10,000 times to determine the distribution of cumulative horizontal distances from the ms given no correlation between whether the star hosts an exoplanet and its location in the @xmath104 cmd . our results show that for hosts of jupiter mass planets , @xmath105 , which corresponds to a probability of @xmath106 that the stars cumulative distance from the isometallicity contour occurred by chance . for the neptune hosts , we find that @xmath107 leading to @xmath108 , and for the combination of all planet hosts , we obtain @xmath109 or @xmath110 . the distributions resulting from the monte carlo simulation and the locations of the actual planet hosting stars @xmath102 values are displayed in the bottom panel of figure [ fig : metallicity ] . the points plotted on top of each curve in figure [ fig : metallicity ] represent the actual cumulative distance of our planet hosting star samples from the field m dwarf ms . our results thus indicate that the planet - metallicity correlation is robust for m - dwarf hosts of planets with @xmath111 , but that at smaller masses there is , at present , no evidence a correlation exists . the lick - carnegie exoplanet survey and its predecessors have carried out a long - term monitoring program of the brightest m - dwarf stars in the sky . our database of observations contains 159 stars that have more than 10 observations apiece , and which , additionally , have median internal uncertainty @xmath112 . within this group , there is a subset with extensive data sets . for example , 11 stars have @xmath113 observations and median internal uncertainties @xmath114 . a question of substantial interest , therefore , is the degree to which the observations taken to date have probed the true aggregate of planetary companions to the m - dwarf stars in our survey . the effort required to obtain the existing data has been substantial . among the m - type stars alone , our database contains a total of 5,468 velocity measurements from keck i , totaling 2,579,862 seconds ( 29.86 days ) of on - sky integration . overheads , including the acquisition of high s / n spectra , ccd readout time , and weather losses , add materially to this time investment . furthermore , the distribution of total observing time allotted to the stars on the list has been highly uneven . targets such as gliese 436 and gliese 876 , which harbor planetary systems of particular interest , have received much more attention than the typical red dwarf in the survey . for example , gliese 436 has 148 observations and gliese 876 has 204 observations obtained with the keck telescope . the stars themselves also exhibit a range of chromospheric activity levels . the resulting star - to - star dispersion in stellar jitter " ( tantamount to a measurement uncertainty , @xmath115 ) complicates the evaluation of threshold levels for @xmath97 as a function of orbital period to which planetary companions can be excluded . there are a variety of approaches to the measurement of false alarm probabilities ( fap ) in the context of spectral analysis of unevenly sampled data . see , e.g. @xcite for a recent discussion . a very simple approach is described by @xcite . for a gaussian random variable . ] , the probability distribution for obtaining a peak at frequency @xmath57 of lomb - normalized power @xcite , @xmath116 , is exponential with unit mean . if a data set drawn from measurements of a white noise ( gaussian ) distribution supports measurement of @xmath117 independent frequencies , the probability that no peak exceeds power @xmath118 ( the fap ) is @xmath119 . we adopt a fap of @xmath120 , calculated with the above method ( and using monte - carlo simulations to determine @xmath117 ) as the generic threshold for attributing a given planetary signal to a given dataset . with this detectability threshold , we use the systemic console 2.0 software package @xcite to determine the number of readily detectable planets in our m - dwarf data set . a `` readily detectable '' planet generates a signal that can be isolated algorithmically ( and automatically ) by straightforward periodogram analysis and levenberg - marquardt minimization . the results of this exercise are shown in figure [ fig : mdwarftable ] , which locates signals corresponding to 19 previously published planets orbiting 14 separate m - dwarf primaries . other than gliese 667c , there are no stars on our 159-star list for which a planet has been published by another group , and for which the automated algorithm finds no planets . regarding gj 667c , 40 observations have been made at keck , and these were used in a characterization of the gj 667c system @xcite , however the peak planetary signal for this set fell below our fap threshold when utilizing only the keck data . the bright planet - hosting red dwarfs gliese 832 , 3634 , and 3470 all have declinations that are too far south to be observed from mauna kea , and hip 79431 ( ra 16h 12 m 41.77s dec -18@xmath13 52@xmath121 31.8@xmath122 ) is not on the list of m - dwarfs being monitored at keck . the kepler mission s photometric data have been used to infer that small planets orbiting m - dwarfs are very common . for example , @xcite find an occurrence rate of 0.9 planets per star in the range @xmath8 with @xmath123 days . given the existence of this large number of small - radius planets , it is of interest to make a quantitative analysis of how deep into the expected population of super - earth type planets suggested by the kepler mission the keck radial velocity survey has probed . to answer this question , we have created synthetic radial velocity data sets that contain test planets , and which conform with the timestamps , the internal measurement uncertainties , and the stellar properties ( namely mass ) for all 104 m - dwarfs under surveillance at keck with at least 20 radial velocity observations . to address the error source arising from stellar jitter , @xmath124 , we use the median value provided in @xcite of 3.9 @xmath42 as the expected level of @xmath124 for our m - dwarf stars . this value is then added in quadrature with the internal uncertainty and applied to the synthetic data set to create a more accurate representation of the system . for each of the 104 stars in the keck survey , we have created 400 synthetic data sets . each set contains a single planet . the planets are evenly spaced in log period from 2 to 100 days , and evenly spaced in log mass from 1 @xmath48 to 1 @xmath125 . we assign a circular orbit to these test planets and assume @xmath126 and @xmath52 in each case . we then calculate the radial velocity each of these planets would induce on a parent star . a gaussian distribution with @xmath127 is used to perturb the predicted radial velocity value . each of the @xmath128 synthetic systems is passed to the planet search algorithm . figure [ fig : massperpanel ] shows examples of the returned planets for 4 of the 104 stars in this experiment including the star of main interest here , gj 687 . in figure [ fig : massperpanel ] the black lines from bottom to top represent constant @xmath129 values of 1 , 10 , and 100 respectively . as expected , the detectablity thresholds lie roughly along lines of constant @xmath32 . to determine the smallest @xmath32 value we could reliably detect for each star , we find the smallest value of @xmath32 for which a planet was found for at least 50% of the chosen periods . this median value generates the green lines seen in the figure . the top panel of figure [ fig : mink ] shows these minimum @xmath32 values for each of the 104 stars that we tested . for clarity , the stars in this figure have been ordered by increasing minimum @xmath32 , and are colored by the number of observations we have for each . if a test planet lies in its star s habitable zone , ( defined as the semi - major axis at which the flux received by the planet is the solar constant received at earth ) we can ask how large the planet needs to be to be detectable by our radial velocity survey . figure [ fig : mink ] shows these threshold masses for each of the 104 m - dwarf stars which we analyzed . these stars maintain the ordering from the top panel , but now have been colored by the mass of the parent star . we see that while the keck survey has probed substantially into the regime occupied by neptune - mass planets , it has not made significant inroads into the super - earth regime for periods that are of astrobiological interest . gj 687 is the second planetary system to be detected using data from the apf telescope , with the first being hd 141399 b , c , d , and e @xcite . apf has successfully navigated its commissioning stage , and , since q2 2013 , it has routinely acquired science - quality data that presents sub - m / s precision on known radial velocity standard stars @xcite . in recent months , the degree of automation for apf has increased substantially . the facility currently works autonomously through an entire night s operations , calibration , and observing program . the apf and its accompanying high - resolution levy spectrograph together form a dedicated , cost - effective , ground - based precision radial velocity facility that is capable of detecting terrestrial - mass planets at distances from their parent stars at which surface liquid water could potentially be present . unlike other highly successful rv facilities , the apf uses neither image scrambling nor image slicing . with a peak efficiency of 15% and typical spectral resolutions of @xmath130 , the apf represents a critical new resource in the global quest to detect extrasolar planets . initial speed comparisons indicate that in order to match the signal - to - noise acquired using the keck telescope / hires spectrograph combination , the apf needs only a factor of 6 increase in observing time . since the amortized cost of a night on keck is @xmath131 more expensive than a night on the apf , and because 80% of the apf s nights are reserved for exoplanetary work , the apf ( with its sub - m / s precision and dedicated nightly cadence abilities ) will likely provide key contributions to exoplanet detection and characterization in the coming years . gliese 687 b s radial velocity half - amplitude , @xmath132 , is substantially greater than the current state - of - the - art detection threshold for low mass planets . the lowest measured value for @xmath32 in the catalog of doppler - detected extrasolar planets stands at @xmath133 @xcite . on the other hand , gliese 687 s status as one of the nearest stars to the sun imbues it with a great deal of intrinsic interest . in our view , the relatively recent date for gliese 687 b s detection can be attributed both to the substantial amount of stellar - generated radial velocity noise ( as evidenced by figures [ fig : sindex ] and [ fig : phasedcurves ] ) , but also to its location in draco , high in the northern sky , where apf , along with harps north , are the only facilities that can routinely observe at sub-@xmath134 precision . ( as evidenced by the data in this paper , keck can observe at these high declinations , but at significantly higher expense in comparison to stars lying closer to the celestial equator . ) indeed , gliese 687 s stellar coordinates ( r.a . 17h , 36 m , dec @xmath135 ) place it very close to the north ecliptic pole , located at @xmath136h , dec@xmath137 . this location flags it as a star of potentially great importance for the forthcoming nasa tess mission . as currently envisioned ( and as currently funded ) , tess is a two - year , all - sky photometric survey to be carried out by a spacecraft in a 27-day p/2 lunary resonant orbit . tess will photometrically monitor @xmath138 bright stars with a @xmath139 1-hour systematic error floor . ( for reference , a central transit of the sun by the earth produces an 86 ppm transit depth . ) the northern ecliptic hemisphere will be mapped during the first year of the mission via a sequence of 13 sectors with 27 days of continuous observation per sector . these sectors overlap at the north ecliptic pole , and create an area of @xmath121,000 square degrees ( 1/50th of the sky ) for which photometric baselines will approach 365 days . gliese 687 lies at the center of this tess `` overlap zone '' ( which also coincides with jwst s continuous viewing zone ) . because much longer time series are produced in the overlap zone , the highest - value transiting planets found by the mission will emerge from this part of the sky ( along with the sister segment covering the south ecliptic pole ) . as mentioned above , however , the tess overlap zone has received relatively little attention from the highest - precision doppler surveys . about 10,000 target stars from the tess dwarf star catalog , all with v@xmath14012 , are present in the overlap zone . this , of course , is far too many stars to survey with doppler rv , but there appears to be substantial value inherent in monitoring the _ brightest _ , _ nearest _ , and _ quietest _ members of the cohort of tess overlap stars . the latest estimates @xcite suggest that @xmath1250% of main sequence stars in the solar vicinity harbor @xmath141 planets with @xmath4 . assuming a uniform distribution in period between 5 and 100 days , the average transit probability for these planets is @xmath142 , suggesting that of order @xmath143 low mass transiting planets ( and systems of transiting planets ) will be detected by tess within the overlap zone . of these , a small handful , of order 5 systems total ( and perhaps , with probability @xmath144% , including gliese 687 b ) will garner by far the most attention from follow - up platforms such as jwst , due to their having optimally bright parent stars . our detection of gliese 687 b suggests that by starting now , with a systematic program of doppler observations of a target list of @xmath12200 carefully vetted g , k & m dwarf stars with v@xmath127.5 to v@xmath1210.5 in the 1000-square degree tess overlap zone , apf can ensure that a precise multi - year doppler velocity time series will exist for the most important tess planet host stars at the moment their transiting planets are discovered . gl acknowledges support from the nasa astrobiology institute through a cooperative agreement between nasa ames research center and the university of california at santa cruz . ssv gratefully acknowledges support from nsf grants ast-0307493 and ast-0908870 . rpb gratefully acknowledges support from nasa oss grant nnx07ar40 g , the nasa keck pi program , and from the carnegie institution of washington . s.m . acknowledges support from the w. j. mcdonald postdoctoral fellowship . gwh acknowledges support from nasa , nsf , tennessee state university , and the state of tennessee through its centers of excellence program . the work herein is based on observations obtained at the w. m. keck observatory , which is operated jointly by the university of california and the california institute of technology , and we thank the uc - keck , uh , and nasa - keck time assignment committees for their support . this research has made use of the keck observatory archive ( koa ) , which is operated by the w. m. keck observatory and the nasa exoplanet science institute ( nexsci ) , under contract with the national aeronautics and space administration . we thank those who collected the data in the koa including : d. fischer , g. marcy , j. winn , w. boruki , g. bakos , a. howard , and t. bida . we also wish to extend our special thanks to those of hawaiian ancestry on whose sacred mountain of mauna kea we are privileged to be guests . without their generous hospitality , the keck observations presented herein would not have been possible . this research has made use of the simbad database , operated at cds , strasbourg , france . this paper was produced using @xmath145 . , , . ccc 2450603.965 & -15.930 & 1.770 + 2450956.065 & -4.740 & 2.040 + 2450982.977 & -12.880 & 1.560 + 2451013.879 & -15.550 & 1.610 + 2451312.036 & -6.870 & 1.910 + 2451368.798 & -17.330 & 1.790 + 2451439.745 & -25.500 & 2.170 + 2451704.957 & -14.920 & 1.810 + 2452007.029 & -3.280 & 1.970 + 2452009.077 & -7.440 & 2.020 + 2452061.913 & 7.450 & 2.040 + 2452062.935 & 4.490 & 1.940 + 2452094.853 & 2.450 & 1.550 + 2452096.894 & 4.560 & 1.850 + 2452097.982 & 5.180 & 1.610 + 2452127.919 & 1.940 & 2.200 + 2452133.737 & 2.830 & 2.010 + 2452160.872 & -6.120 & 2.170 + 2452161.821 & -6.840 & 2.370 + 2452162.787 & -8.300 & 2.370 + 2452445.988 & 17.880 & 2.100 + 2452537.743 & -4.900 & 2.340 + 2452713.113 & 8.860 & 1.590 + 2452806.027 & -4.580 & 1.990 + 2452850.901 & 6.090 & 2.040 + 2453179.985 & 12.360 & 1.730 + 2453479.066 & 8.000 & 1.210 + 2453549.857 & -2.830 & 0.880 + 2453604.881 & -2.530 & 1.220 + 2453838.109 & 3.550 & 1.220 + 2453932.905 & 3.890 & 1.380 + 2453960.873 & 3.880 & 1.630 + 2453961.820 & -0.510 & 1.800 + 2453981.789 & -14.390 & 1.190 + 2453982.904 & -1.880 & 1.900 + 2453983.831 & -8.010 & 1.220 + 2453984.887 & -0.360 & 1.290 + 2454248.023 & 7.800 & 2.090 + 2454248.990 & 8.410 & 2.030 + 2454249.945 & -3.960 & 1.730 + 2454252.032 & -12.260 & 1.870 + 2454255.924 & -12.790 & 1.280 + 2454277.851 & 8.200 & 1.500 + 2454278.896 & 11.300 & 1.700 + 2454279.933 & 11.850 & 1.580 + 2454294.903 & -10.850 & 1.260 + 2454304.888 & -4.110 & 1.570 + 2454306.037 & 2.450 & 1.600 + 2454307.015 & -2.050 & 1.500 + 2454308.072 & 5.540 & 1.600 + 2454309.050 & 7.370 & 1.810 + 2454310.042 & 8.730 & 1.720 + 2454311.024 & 1.600 & 1.480 + 2454312.017 & 0.860 & 1.590 + 2454312.875 & 2.350 & 1.450 + 2454313.875 & -1.170 & 1.320 + 2454314.908 & 6.220 & 1.630 + 2454318.921 & 5.360 & 1.560 + 2454335.826 & -13.560 & 1.940 + 2454338.888 & -16.490 & 1.790 + 2454339.883 & -9.590 & 1.780 + 2454343.791 & -8.170 & 1.950 + 2454396.717 & 3.080 & 2.390 + 2454397.729 & -4.380 & 2.130 + 2454548.052 & 12.600 & 1.330 + 2454549.092 & 3.010 & 1.280 + 2454633.963 & -6.780 & 1.570 + 2454634.903 & -5.330 & 1.650 + 2454635.946 & -2.500 & 1.790 + 2454636.901 & 2.170 & 1.650 + 2454637.939 & -5.830 & 1.610 + 2454638.838 & -5.010 & 1.450 + 2454639.998 & 3.970 & 1.850 + 2454641.954 & 2.760 & 1.540 + 2454674.860 & -0.820 & 1.440 + 2454688.892 & -4.450 & 1.710 + 2454689.917 & 6.340 & 1.890 + 2454717.853 & 4.220 & 2.000 + 2454718.914 & -2.870 & 2.070 + 2454719.866 & 1.210 & 2.040 + 2454720.871 & 0.570 & 2.140 + 2454721.879 & 4.620 & 2.040 + 2454722.797 & 5.810 & 2.210 + 2454723.814 & 2.610 & 2.120 + 2454724.848 & 1.350 & 1.990 + 2454968.017 & 5.210 & 0.970 + 2455016.035 & 0.730 & 1.370 + 2455022.084 & 1.680 & 1.670 + 2455023.044 & 6.360 & 1.070 + 2455024.823 & -11.730 & 0.920 + 2455049.825 & 0.220 & 1.090 + 2455050.858 & 3.990 & 1.650 + 2455051.772 & 6.360 & 1.790 + 2455052.780 & 8.540 & 1.930 + 2455053.925 & -12.540 & 1.510 + 2455143.692 & -3.080 & 2.620 + 2455167.704 & 3.210 & 1.710 + 2455259.119 & 3.760 & 1.610 + 2455260.144 & 0.000 & 1.470 + 2455371.991 & -5.030 & 1.070 + 2455407.955 & -1.480 & 1.060 + 2455463.718 & -8.590 & 1.520 + 2455516.696 & -1.120 & 1.740 + 2455518.722 & -3.780 & 2.320 + 2455609.135 & 11.250 & 1.310 + 2455638.121 & -3.930 & 0.910 + 2455639.143 & -2.250 & 0.910 + 2455665.046 & -5.130 & 1.190 + 2455670.082 & -8.090 & 1.290 + 2455721.057 & 6.840 & 1.850 + 2455825.832 & 3.030 & 2.280 + 2455840.748 & 3.820 & 1.390 + 2456027.098 & 11.390 & 1.940 + 2456116.938 & 16.010 & 1.130 + 2456117.906 & 14.270 & 1.050 + 2456329.167 & 5.420 & 1.970 + 2456432.892 & -0.700 & 1.300 + 2456433.935 & 1.610 & 1.340 + 2456548.855 & -2.340 & 1.490 + 2456549.793 & -7.780 & 1.430 + 2456550.849 & -3.840 & 2.150 + 2456551.724 & -2.520 & 1.570 + ccc 2456484.738 & 11.650 & 0.910 + 2456486.839 & 9.500 & 0.950 + 2456488.845 & 11.640 & 1.120 + 2456490.726 & 8.300 & 1.260 + 2456492.816 & 7.020 & 0.900 + 2456494.776 & 5.260 & 0.750 + 2456508.764 & -8.650 & 0.780 + 2456510.800 & -9.020 & 0.800 + 2456513.772 & -8.920 & 0.860 + 2456521.823 & -2.120 & 0.730 + 2456522.790 & -2.920 & 0.930 + 2456533.797 & 2.910 & 0.820 + 2456551.743 & -8.410 & 0.810 + 2456552.717 & -6.660 & 0.970 + 2456553.727 & -7.610 & 0.850 + 2456554.720 & -4.640 & 0.890 + 2456564.762 & 1.530 & 0.940 + 2456568.692 & 0.460 & 1.140 + 2456571.657 & 0.580 & 0.940 + 2456622.032 & -9.750 & 1.280 + 2456677.921 & 2.390 & 0.980 + 2456681.011 & 4.660 & 1.360 + , s. s. , allen , s. l. , bigelow , b. c. , et al . 1994 , in society of photo - optical instrumentation engineers ( spie ) conference series , vol . 2198 , society of photo - optical instrumentation engineers ( spie ) conference series , ed . d. l. crawford & e. r. craine , 362+

the collapse of a massive star down to a neutron star releases over @xmath1erg of potential energy . just one percent of this energy is required to power the observed type ib / c and type ii supernovae . most of the energy is emitted in the form of neutrinos . how this small fraction of the energy is converted into explosion energy is still a matter of debate , but it is believed by most that neutrinos play a role in depositing energy above the collapsed core to drive this explosion ( see fryer 2003 for a review ) . whether or not neutrinos are important for the explosion , they do provide astronomers a window into the explosion mechanism behind supernovae . neutrinos also provide a window into the material conditions at the core of a supernova explosion . with temperatures above 10mev and densities above nuclear densities , core - collapse supernovae make ideal laboratories for nuclear physics . to extract information about nuclear physics from such laboratories , we must understand the systematics in our supernova experiment . when the explosion engine is active , convective motions in the engine ( herant et al . 1994 ) make it very difficult to interpret the neutrino signal . an observed elevated neutrino luminosity could be a change in the neutrino opacity or it could be a difference in the convective engine . one approach to eliminate issues with convection would be to wait until all convective activity ceases . a number of studies have now been presented following the evolution of a collapsing star over 0.5 - 1.5s after bounce ( fryer & heger 2000 ; burrows et al . 2006 ; scheck et al . 2006 ; and , in 3-dimensions , fryer & young 2007 ) . in all these calculations , hydrodynamic motions near the proto - neutron star continue to dramatically affect the neutrino emission through the end of the simulations . to achieve a clean neutrino signal , we must wait until after the launch of the explosion . we must also wait a few seconds after the launch of the explosion because the proto - neutron star may experience deep convection ( keil et al . 1996 ) . after this time ( roughly a few seconds to 10 - 20s ) , the supernova finally seems to have achieved an ideal condition as a physics laboratory ( reddy et al . after 10 - 20s , the neutrino signal will be too weak to do much experimental science , so we truly are limited to this narrow time window . unfortunately , even at these late times , nature does not allow completely pristine conditions . material falling back from the supernova explosion ( `` fallback '' ) may well produce a new round of convection and confusion to our neutrino signal . this fallback has been studied both in its important role in calculating the initial mass of the neutron star formed in a supernova explosion ( e.g. fryer & kalogera 2001 ) and in estimating the r - rpocess yields in supernovae fryer et al . ( 2006 ) . in this paper , we study its role in determining the neutrino luminosity arising after the first few seconds of a supernova . in 2 , we review the history of supernova fallback and present new calculations of supernova explosions estimating the fallback for a range of explosion energies and stellar masses . 3 describes the 2-dimensional code used to model the neutrino emission from the this fallback and shows results for the suite of simulations run for this paper . different than the multi - dimensional simulations modeling stellar collapse , these simulations do not start at the onset of collapse ( or bounce ) and end at the launch of the explosion . instead , they start a 2 - 10s ( depending on the explosion energy , etc . from 2 ) after the launch of the explosion when material begins to fall back onto the neutron star . we conclude with a brief discussion on how neutrino signals can be used to help better understand the supernova explosion , neutron star birth masses and neutrino cross - sections . the idea of fallback was first discussed by colgate ( 1971 ) to overcome nucleosynthesis issues arising from the supernova ejection of neutron rich material produced in stellar cores ( arnett 1971 , young et al . colgate argued that the inner layers of the ejected material would deposit its energy to the stellar material above it , ultimately reducing its energy below that needed to escape the neutron star , and it would fall back onto the neutron star . in such a scenario , one would expect the inner material to fall back quickly ( within the first few to ten seconds ) . it was argued that this material ( the neutron rich material from the initial explosion ) would accrete onto the neutron star , alleviating any nucleosynthesis issues . since this work by colgate , supernova explosion calculations have confirmed that fallback does occur ( bisnovatyi - kogan & lamzin 1984 ; woosley 1989 ; fryer et al . 1999 ; macfadyen et al . 2001 ) and new arguments for the cause of this fallback were suggested . for example , woosley ( 1989 ) argued that when the supernova shock decelerates in the hydrogen layers of the star , it sends a reverse shock that drives fallback . such a model argues that the fallback will happen at late times , long after it can affect the neutrino luminosity from the cooling proto - neutron star . it also suggested that close binary systems ( where a star s hydrogen envelope was removed prior its collapse ) might experience a very different amount of fallback than the amount of fallback in the collapse of a single star . which is the true cause of the fallback ? and more to the point for neutrino emission , when does fallback occur ? this confusion has mostly exists because , without a quantitatively reliable explosion mechanism , scientists have artificially driven supernova explosions to be able to study the results of these explosions ( e.g. supernova light curves , nucleosynthetic yields , and fallback ) . much of the past work ( e.g. woosley 1989 ; fryer et al . 1999 ; macfadyen et al . 2001 ) used piston driven explosions . by moving the piston out , scientists artificially lowered the amount of fallback and delayed this fallback considerably . these calculations all predicted that fallback would occur more than 100s after the launch of the explosion . but the piston - driven explosion mechanism may not accurately model the nature of the fallback . young & fryer ( 2007 ) found that piston - driven explosions produced both different fallback rates and different nucleosynthetic yields than energy - driven explosions of the same final energy . the errors in the yields or light - curves are on par with slight changes in the explosion energy . such small errors seemed unimportant in matching the observations . for fallback , the differences are much more dramatic . not only does the amount of fallback change , but the timescale at which the fallback occurs can change by more than an order of magnitude , moving a falback time of a few hundred seconds down to just 3 - 15s . this changes the fallback accretion rate by over an order of magnitude and ultimately determines whether or not fallback is important in estimating the observed neutrino signal . it also means that the the amount of fallback will not change between binary and single stars ( unless binary interactions change the internal structure of the star ) . in this paper , we focus on fallback calculations from energy - injected explosions . energy - injection is much better at mimicking the currently - favored supernova mechnanisms . let s take the convection - enhanced , neutrino - driven supernova engine as an example ( see fryer 2003 for a review ) . in this mechanism , the basic energy source is neutrinos ( either diffusing out of the proto - neutron star core or from newly accreting material ) that heat the atmosphere above the neutron star . this atmosphere is topped by the accretion shock of the infalling star . when enough energy is deposited in this region , the accretion shock will be pushed outward and an explosion occurs . convection aids this mechanism by both allowing heated material to rise ( cooling by adiabatic expansion instead of forcing it to continue to heat until it can cool by neutrino emission ) and allowing shocked material at the top of the convective region to flow down to the neutron star surface to accrete onto the neutron star ( emitting neutrinos ) or be heating to be part of the rising bubble . an energy injection method , although not mimicking the effects of convection directly , can mimick the basic tenets of this model - heating just above the neutron star surface to blow off the accretion shock . ideally , once the region above the neutron star becomes more rarefied , the energy injection will essentially halt ( aside from a weak , by supernova standards , neutrino - driven wind ) . some groups have gone so far as to only inject this energy through a neutrino flux ( frlich et al . 2006 ) . in this manner , once the region becomes rarefied , the energy injection drops naturally . a piston - driven explosion can not mimic this effect well unless very specific attention is paid to the input of the piston . the very different fallback results from piston explosions demonstrate just how far off such explosions are from the energy drive of the standard neutrino model . this does not mean that by using energy - injection , we can produce a definitive fallback estimate for a supernova explosion . the explosion energy is one of our primary uncertainties in estimating the fallback rate . supernova scientists have yet to agree on the exact mechanism behind the supernovae and we are far from achieving quantitative predictions of these explosions . with accurate light - curve and spectra calculations , we may be able to estimate the explosion energy for a particular supernova based on its observations . by studying a range of explosion energies for a given progenitor , we can provide a template for the neutrino flux ( from fallback ) arising from these systems , allowing us to possibly extract this effect and once more focus a study on the physics of dense , hot matter . our fallback calculations will all be based on fallback estimates from energy - driven explosions . we use the same multi - step technique employed in young et al . ( 2006 ) and young & fryer ( 2007 ) . we start with progenitor stars modeled to collapse ( either from heger et al . 2000 or young et al . these progenitors are mapped into the 1-dimensional core - collapse code from herant et al . this code includes equations of state valid from stellar densities up to nuclear densities , a 3-flavor flux - limited diffusion neutrino transport scheme , and a simple nuclear network . with this code , we follow the collapse and formation of the proto - neutron star and the propogation of the shock produced when the collapse halts due to nuclear forces . after this shock stalls ( as it loses its energy via neutrino losses ) , we have a structure defined by a shocked `` atmosphere '' produced by the now - stalled bounce shock above a dense proto - neutron star . the proto - neutron star typically has a baryonic mass between 1.1 and 1.3m@xmath2 . the edge of the proto - neutron star is determined by the mass where the density drops below @xmath3 . at these densities , there is typically a well - defined edge where the density drops from @xmath4 to @xmath5 over a very narrow mass cut . the exact location of this edge depends upon the progenitor mass and probably the exact code used to model this collapse phase ( e.g. 1-dimensional versus multi - dimensional results ) . in general , 1-dimensional calculations have not produced supernova explosions . with the proto - neutron star removed , we have also removed the energy source for any explosion . to induce an explosion in 1-dimension , we must source in energy . the simulations here source the energy directly into the inner 15 cells ( roughly 0.1m@xmath2 ) of the star . we keep this energy source on for a limited time ( between 50 and 300ms ) , varying the energy injection rate and time to produce a range of explosion energies . young & fryer ( 2007 ) found that such energy sourcing was more flexible than a simple neutrino enhancement to modeling the full range of proposed explosion mechanisms . in our calculations , we use the shorter ( 50ms ) injection duration for the lowest energy explosions and the lowest mass stars and the longer ( 300ms ) injection duration for the more energetic explosions . we follow this explosion for 400s . this allows us to follow the shock as it moves well into the star ( and , for binary systems , out of the star surface ) . our simulation space includes enough matter to ensure that for this 400s duration , the shock is well within the simulation space ( there are at least 100 zones between the final shock postion and the outer zone of our star ) . in the case of the binary systems , we have included a mass loss estimated from the stellar models . typically , we model this star with roughly 2000 zones . young et al . ( 2008 ) have done a resolution study , comparing 1000,2000 and 4000 zones . they do not find an appreciable difference in the ejected ( and hence fallback ) mass based on this resolution . it will make a large difference on how the fallback mass tapers off . however , note that with the lagrangian code in this paper , we can not accurately calculate low fallback rates . but , as we shall see in our 2-dimensional models , it may well be that fallback powered explosions might cut off this low - rate fallback . we use the same 1-dimensional code to model the fallback . as material falls back onto the edge of our proto - neutron star and its density rises above @xmath6 , we remove the particle and add its mass to our proto - neutron star . although we could follow the accretion and neutrino cooling of this material to higher densities with our code , as the density rises , the sound speed increases and the cell size decreases , both of which cause the time step to decrease . to make this problem tractable , we choose a low enough density that the material accretes before dropping the timestep below 1 microsecond . even so , these simulations typically take between 1 million to 10 million timesteps . as we shall find in 3 , the modeling the true behavior of this matter requires modeling the accretion in multi - dimensions . since these simulations are focused on calculating the infall rate ( not true accretion rate ) , our assumptions do not introduce large errors in our analysis . in this paper , we present the results from a suite of 1-dimensional explosion models models using 3 different progenitor masses and explosion energies ( for a summary , see table [ tab : fallback ] ) . table [ tab : fallback ] also shows the peak accretion rates and total mass accreted for these models . note that we predict a range of neutron star masses based on both progenitor mass and explosion energy , in agreement with fryer & kalogera ( 2001 ) . but remember that the neutron star masses assume that all of the fallback remains on the neutron star . as we shall see in 3 , some of this matter is re - ejected . for normal explosion energies , it is likely that 12 - 15m@xmath2 stars produce neutron stars with gravitational masses in the 1.3 - 1.5m@xmath2 range . if the explosions are stronger , the gravitational masses may well be as low as 1.2m@xmath2 . for weak explosions , or more - massive stars , the remnant masses may be so large that the neutron star collapses to a black hole . figure [ fig : macc ] shows the mass accreted in the first 15s for our models . note that in all casses , fallback occurs almost immediately ( with some delays of 2 - 7 seconds ) . unless large amounts of fallback occur ( above 1m@xmath2 ) , the fallback is likely to be mostly over after 10 - 15s . with over 0.1m@xmath2 falling back in 10 - 15s , accretion rates above 0.01m@xmath2s@xmath7 are expected , corresponding to neutrino luminosities in excess of @xmath8 . the bottom line is that , for normal supernovae , fallback occurs , at least for stars with initial masses at or above 12m@xmath2 . this fallback occurs early , so it will definitely play some role in the neutrino emission at the 3 - 15s timescale . an additional feature of the progenitor that affects the fallback and the neutrino luminosity is the angular momentum in the progenitor . figure [ fig:1drot ] shows the angular momentum of the inner 4m@xmath2 of a star for a variety of stars with and without magnetic braking ( heger et al . 2000,2005 ) . for neutron star accretion , the relevant angular momentum is that within the inner 1.4 - 2.0m@xmath2 . such mass zones have low - angular momenta : a few@xmath9 for stars with magnetic braking , a few@xmath10 for stars without . our typical simulations use a value of @xmath11 . these low angular momenta had little effect on our results , and we ran some simulations with twice that amount . for our black hole systems , we ran even higher angular momenta . for neutron stars , the angular momentum is not enough for this material to form a true disk in the star , but it can alter the downflow and it is this effect that we would like to study in this paper . note that although there is quite a bit of structure in the angular momenta ( caused by incomplete angular momentum transport across elemental boundaries ) , in general , the angular momenta of the stars increases as one moves to higher and higher mass shell . this is one reason why black - hole forming systems are more likely to produce asymmetric explosions than typical neutron - star forming systems . the increased angular momentum means that angular momentum plays a bigger role in shaping the explosion , possibly producing larger asymmetries . our 1-dimensional calculations provide us with a fallback rate . if we assume that any fallback material emits all of the potential energy released from its downfall at the moment it hits the proto - neutron star , we can estimate the neutrino luminosity from the fallback . the accretion of 0.1m@xmath2 masses onto a 10 km , 1.4m@xmath2 neutron star over 10s would correspond to a neutrino luminosity of @xmath12 . for our more massive stars , this fallback can be ten times higher , corresponding to a neutrino luminosity of @xmath13 for over 10s after the launch of the explosion . assuming pair - annihilation dominates the neutrino emission , this neutrino luminosity would be nearly 50% electron and 50% anti - electron neutrinos . typical neutron star luminosities after 1s are below @xmath14 and can be as low as a few @xmath15 ( e.g. bruenn 1987 , keil & janka 1996 ) . even at 1s , fallback can dominate the neutrino emission if the fallback is heavy . after @xmath165s our fallback estimates argue that fallback neutrinos dominate the neutrino emission . but we have made several assumptions in this estimate for the neutrino luminosity . first , we have assumed that all of the potential energy released is immediately emitted in neutrinos . this energy can go into heating the proto - neutron star which will then cool on a longer timescale . this energy also may go into ejecting other infalling material . as the material falls onto the proto - neutron star , it is shock heated and can rise . fryer et al . ( 1996,2006 ) found that these rising shocked bubbles can accelerate above the escape velocity and actually be re - ejected . so it is quite possible that a fraction of the `` fallback '' material does not end up on the neutron star if we account for these multi - dimensional effects . this re - ejected material both does not contribute to the total energy available to produce neutrinos , but it takes some fraction of the potential energy released to power its explosion . finally , @xmath17 and @xmath18 neutrinos may also be released and we must include these neutrinos in our energy budget . in this section , we study the effects of the above assumptions to get a more accurate neutrino luminosity from fallback . our primary concern with our simple estimate of fallback neutrino emission is the fact that we ignore multi - dimensional effects . we have known for some time that if a compact object is accreting mass with considerable angular momentum ( and inefficient cooling ) , outflows occur ( e.g. blandford & begelman 1999 ) . but even if the angular momentum is minimal , if the compact object is a neutron star , a sizable fraction of the infalling material can be re - ejected ( fryer et al . 2006 ) . to study the fate of supernova fallback , we want to understand a number of effects . for example , how do the results vary with accretion rate ? but we would like to also understand the role of angular momentum and the differences between hot or cold neutron stars . finally , numerical effects , such as boundary conditions , are bound to play a role . before we discuss the results of our simulations , let s discuss the code used for these calculations and our tests of the physical and numerical effects . our code must model the physics of downflows and allow us to answer the questions in the preceding paragraph . first and foremost , we must follow the evolution in a multi - dimensional manner . as a first step , we use the two - dimensional smooth particle hydrodynamics code described in fryer et al . ( 1996,2006 ) . we model the region from 10,000 km above the proto - neutron star down to the proto - neutron star surface . typical runs range from an initial set of @xmath19 particles moving up to 50 - 70,000 particles by the end of the simulation . the code includes an equation of state valid from densities below 1g@xmath20 up to nuclear densities ( including an estimate of nuclear statistical equilibrium ) . neutrino transport is followed using flux - limited diffusion neutrino scheme for three neutrino species ( herant et al . the neutrino emission and cross - sections are also outlined in herant et al . ( 1994 ) . our simulations are set to mimic the conditions 2 - 10s after the launch of the explosion when material begins falling onto the proto - neutron star . the initial condition begins with a set of particles ranging from 1000 km up to 10,000 km . the material is given velocities set to the free - fall velocity ( in our 1-dimensional simulations , the infall velocity is within 1 - 10velocity ) at their radial position with densities set to give the desired accretion rate ( recall , @xmath21 . with @xmath22 within 1 - 10% of the free - fall velocity , we can accurately set up our initial conditions : our 3 separate rates correspond to 0.001,0.01,0.1m@xmath2s@xmath7 ( table [ tab : neut ] ) . for these simulations , we model only constant fallback rates . note that in nature , the fallback rate varies quite a bit . but we are focusing our study on the neutrino emission at peak fallback rates , and this suite of simulations will definitely bracket the range of results . the entropy of the infalling material is generally set to a few . the results are fairly insensitive to this initial entropy as the shock resets the entropy . boundary conditions are probably the biggest uncertainty in our calculations . particles are fed in through the outer boundary and are allowed to accrete through the inner boundary and be ejected out of the outer boundary . the infall through the outer boundary is determined by a fixed infall rate . this infall is only altered if material is flowing out of this outer boundary . at any point where an outflow occurs , the inflow is temporarily halted . this models the effect of outflows choking off the accretion . the inner boundary is more difficult . ideally , we would model the matter until its density reaches neutron star densities ( @xmath23 ) and it is mostly deleptonized . at such time , the matter has lost most of its energy and we can be sure that we have accounted for the total neutrino luminosity . however , the sound speeds at such densities and the size of our lagrangian particles near the proto - neutron star would decrease the timestep to fractions of a microsecond . such small timesteps prohibit us from following the evolution of the fallback for more than a fraction of a second . instead we opt to use slightly less demanding criteria for the removal of particles on the inner boundary . our standard set of models accretes particles whose density rises above @xmath24 with electron fractions below 0.3 . this means that we are assuming any further energy released by the matter as it accretes goes into heating the neutron star which will cool on longer timescales . in our suite of models , we include a test of this boundary condition and find that the total neutrino luminosity is not too sensitive to our assumption for accretion . we assume the axis of rotation lies along our axis of symmetry in our 2-dimensional caclulation . each particle is given an angular momentum and it retains that angular momentum for the duration of the calculation . the angular momentum for each particle is determined to fit the angular velocities of our progenitor stars ( fig . 2 ) , so material along the axis has a low angular momentum whereas material in the equator has the highest angular momentum . the specific angular momentum @xmath25 is given by @xmath26 where @xmath27 is the angular velocity taken from the stellar models . the inner material starts below the angular momentum given by @xmath27 , but by the end of our simulations , most of the material near the neutron star surface has the full angular momentum set by this @xmath27 value . there is no angular momentum transport and hence , technically , no heating from this transport . however , the angular momentum will slow the materials inflow , breaking the symmetry in the downflow . fryer & heger ( 2000 ) also found that the angular momentum alters the instability criterion , preventing convection in the equatorial region where the angular momentum gradient is highest . to study the effect of a hot neutron star , we introduced a non - zero neutrino flux arising from our inner , neutron - star , boundary . in particular , we are interested in how the neutrinos from a hot neutron star affect the hydrodynamics , so we have chosen rather large luminosities ( @xmath28 electron neutrino luminosities with energies of 10mev with corresponding @xmath29 anti - electron neutrino luminosities with energies of 15mev ) . these neutrinos are a boundary source for our flux - limited diffusion transport scheme and transported out of the system , heating the inflowing material . however , in most of our simulations , the neutrino optical depth is fairly low , and the total energy deposited is also minimal ( see 3.2 ) . finally , there is always concern that the artificial viscosity used in sph is introducing spurious effects into our calculations . for the most part , our studies have found this numerical artifact to play a small role in results studying core - collapse supernovae , and we do not expect it to play a large role in these calcualtions . nonetheless , we have included a simulation where we have increased both the bulk and von neumann - richtmyer viscosities by a factor of 2 ( 3.0,6.0 respectively versus our standard values of 1.5,3.0 ) . with both numerical and physical effects to study , we have run a suite of simulations to test the dependence of the neutrino luminosity on the initial conditions and on the numerics . a summary of this suite of models , along with their basic results , is summarized in table [ tab : neut ] . our base model has an accretion rate of 0.01m@xmath2 s@xmath7 with an angular momentum equal to that shown in the circle in figure [ fig:1drot ] . we also include higher rotating runs with angular momenta set by equating the angular momenta to the value denoted by the square in figure [ fig:1drot ] . in both cases , the angular momentum is low , so we do nt expect the formation of a full accretion disk . we study the inner boundary in a number of ways . we include a simulation where the criteria for accretion is more strict : @xmath30 with electron fractions below 0.1 . we also include a set of simulations with an absorbing boundary at 100 km ( a black hole boundary condition ) . we study accretion rates 10 times higher and ten times lower than our canonical rate . finally , since the accretion occurs at early times , we have also included a few simulations where the neutron star itself is still emitting neutrinos . in [ sec : dynamics ] , we compare the results on the dynamics for this suite of calculations . in [ sec : neutrinos ] we compare the resulting neutrino luminosities . before we discuss the neutrino fluxes from this suite of calculations , let s compare the dynamics in the accretion . we expect material to shock as it hits the proto - neutron star , and some of this shocked material will begin to rise , driving convection . these simulations are 2-dimensional . even without angular momentum , perturbations along our symmetry axis would allow the instability to develop stronger along this axis . but we have studied this in some detail in core - collapse calculations and have minimized this effect ( see , for example , fryer & heger 2000 ) . however , the fact that the angular momentum axis also lies along the symmetry axis drives convection along our axis of symmetry . the growth of convective instabilities is stabilized by angular momentum gradients and the deceleration of material ( along with the fact that our initial condition has an angular momentum gradient which is strongest along the equator ) , the convection initially grows strongest in the poles . figure [ fig : comp - nsvbh ] shows the results of our standard calculation showing this outflow . this figure shows the evolution at 0.15s . the right panel shows the evolution of the corresponding `` black hole '' simulation with the absorptive boundary . the true innermost stable circular orbit for a slowly rotatign 3m@xmath2 black hole is closer to 20 km , so our 100 km absorption radius understimates the activity from a real black hole . however , we can see that although the angular momentum alters the inflow , it is insufficient to stop it , and the material continues to accrete directly onto the black hole . as our neutron star model evolves the outflow expands , constraining the accretion to a funnel roughly @xmath31 from the plane . figure [ fig : comp - late ] shows the evolution of our standard model to 0.3 and 0.45s . over half of the inflowing material is ultimately ejected . not only does the ejected material not contribute to the energy available for neutrino emission , some of the energy of that accreted material goes toward accelerating this ejecta . such outflows have been studied for over a decade in supermassive black hole systems such as active galactic nuclei ( see blandford & begelman 1999 ) for a review . the amount of potential energy released in the accretion of material onto a compact object is enormous . if this energy is tapped ( either by viscous forces or neutrino emission ) , it can drive an explosion . rockefeller et al . ( 2007 ) and fryer et al . ( 2006 ) have recently applied this physics to stellar - massed compact objects to better understand their simulations of accreting systems . in rotating black hole systems , material hangs up in a disk . viscous forces that transport out angular momentum also transport energy , driving outflows . what prevents this ejection is cooling ( either via photon radiation in the case of supermassive black holes or neutrinos in the case of most collapsing systems ) . of course , for black hole systems that do not have a hard surface preventing the inflow of material , high angular momentum is required to prevent the energy from all flowing directly into the black hole . in neutron star systems , outflows ( or at least vigorous convection ) have been expected for more than a decade ( chevalier 1989 ; fryer et al . 1996 ) . if the angular momentum were high enough , the infalling material would hang up in a disk . this is most important in the black hole systems . we ran a series of models increasing the specific angular momentum ( @xmath25 ) a factor of 3 and , for our black hole models , a factor of 10 . macfadyen & woosley ( 1999 ) found that the specific angular momentum must be at least @xmath32 ( a factor of 1000 higher than our standard @xmath33 value ) . as we can see from figure [ fig : comp - rot ] , our factor of 10 increase in @xmath25 is not enough to change the fate of material falling back on our `` black hole '' simulation with its large absorptive boundary . in a true black hole system , the angular momentum available would increase as we move beyond 3m@xmath2 ( well above the angular momenta shown in fig . [ fig:1drot ] ) . this is one reason why the collapsar grb model argues for systems where the black hole mass exceeds 3m@xmath2 when the angular momentum in the star is sufficient to produce an accretion disk . neutrinos from these systems have been considered in detail elsewhere ( macfadyen & woosley 1999 , hungerford et al . 2006 , rockefeller et al . 2007 ) and we do not study them in this paper . if the neutron star is still hot and emitting neutrinos , it can also alter the inflow of fallback . figure [ fig : comp - hot ] shows two simulations with varying amounts of neutrino energy ( 18 and 36@xmath34 ) arising off the neutron star surface . the effect on the dynamics is minimal , although a slight increase in the velocity ( and hence position at a given time ) can be seen . as we shall see below , the neutrinos emanating from the proto - neutron star surface will dominate the total neutrino flux in such cases . but we do nt expect these high luminosities at 10s and the accretion luminosity will dominate at these later times . but at early times , the modification of the downflow on the neutrino opacities is the dominant effect . finally , we have varied the infall rate from 0.001 to 0.1m@xmath2s@xmath7 . the effect this has on the dynamics is shown in figure [ fig : comp - mdot ] . as with many of our models , the nature of the dynamics is not altered significantly by these changes . but we shall see that the neutrino flux very much depends on the accretion rate . we found in the last section that the dynamical behavior of the fallback was relatively insensitive to both uncertainties in the numerics as well as uncertainties initial conditions : e.g. fallback rate and rotation . but what about the neutrino luminosity ? first let s study the uncertainties in the numerics . fig [ fig : nunum ] shows the neutrino luminosity and mean energy for 3 different models studying the numerical effects ( both the artificial viscosity and the treatment of the inner boundary ) on the calculation . the variations in the viscosity and the inner boundary lead to differences that are less than a factor of 2 in the neutrino luminosity and 5% variations in the neutrino mean energy . many of these errors could be dominated by the explicit transport scheme used in these calculations and it is possible that these errors can be significantly diminished with implicit schemes that are more stable . figure [ fig : nunum ] also shows the results for a fast rotating ns model ( ns - rot2 ) . with the low angular momenta in our calculations based on the rotation velocities in the inner core material of massive stars ( fig . [ fig:1drot ] ) , rotation does not alter the neutrino luminosity noticeably . black hole systems depend more sensitively on the rotation because it is the angular momentum that prevents the material from accreting directly into our black hole . but as we expected from the fact that our angular momentum is too low to produce accretion disks , the neutrino emission from our black hole systems is negligible ( fig . [ fig : nubh ] ) . in such low - angular momentum systems , we expect essentially no emission after the collapse of the neutron star down to a black hole . contrast this to typical collapsar conditions , where the material falling back onto the black hole has enough angular momentum to hang up in a disk and its falback accretion rate is high enough to produce high - density , high - temperature structures . in this case , the neutrino emission from a black hole can be quite large - with luminosities on par with the neutrino burst of the original collapse ( macfadyen & woosley 1999 ; hungerford et al . 2006 ; rockefeller et al . 2007 ) . in our models , the @xmath17 and @xmath18 neutrinos tend to be an order of magnitude lower than the corresponding electron neutrino fluxes . so our assumption that most of the neutrinos are emitted as electron or anti - electron neutrinos is reasonably valid . hungerford et al . ( 2006 ) and rockefeller et al . ( 2007 ) found that for collapsar models , the @xmath17 and @xmath18 neutrinos make up a sizable fraction of the total emission . at higher accretion rates ( and higher angular momentum in the infalling material ) , the fraction of the luminosity emitted in @xmath17 and @xmath18 neutrinos will likely increase . the strongest dependency on our initial conditions in the mass accretion rate . figure [ fig : numdot ] shows the neutrino emission from 3 different accretion rates onto our neutron star surface . our basic potential energy released estimate would argue that the neutrino luminosity scales with the accretion rate . and this trend is basically true for our simulations . however , note that the electron and anti - electron neutrinos for our highest accretion rate ( 0.1m@xmath2s@xmath7 ) case are not quite an order of magnitude higher than our standard accretion rate ( 0.01m@xmath2s@xmath7 ) . this is because the electron neutrinos become trapped in this highest accretion rate case , lowering the escaping luminosity . the @xmath17 and @xmath18 remain nearly an order of magnitude higher as they are not trapped in any of our simulations . figure [ fig : nuhot ] shows the neutrino emission for 3 different `` hot '' neutron star models . fallback contributes an additional 10 - 20% of the luminosity on average for the ns2-hot2 model , 20 - 40% to the ns2-hot1 model , and over 50% to the ns1-hot model . the neutrino energy is also altered by an amount comparable to the change in the luminosity . although a hot neutron star may dominate the neutrino luminosity , fallback clearly can contribute a sizable fraction of the observed neutrino flux . finally , note that in general , our predicted fallback neutrino energies are high . the most notable exception is that the mean neutrino energy is 5 - 10mev cooler for our low accretion - rate simulations . as the accretion rate decreases , so too will the mean energy of the emitted neutrinos . fallback ( of at least 0.1m@xmath2 ) is likely to occur in normal ( @xmath35erg ) supernova explosions from stars with initial masses of 12m@xmath2 or greater . with energy - injected ( more realistic than piston - driven ) explosion calculations , this fallback occurs quickly , generally in the first 15s and peak accretion rates in the 0.01 - 0.1m@xmath2s@xmath7 range should be expected . if the engine stays active long after the launch of the supernova shock , the total fallback will be lower . but neutrino engines weaken significantly quickly after the shock is launched and the material above the neutrinosphere ( absorbing the energy ) is ejected . magnetic - driven explosions could be very different . the mechanism for this fallback is essentially that proposed by colgate ( 1971 ) whereby the outflowing material decelerates as it pushes against the material above it , ultimately decreasing its velocity below the escape velocity and causing it to fall back onto the proto - neutron star . this means that the bulk of the fallback is fairly insensitive to the structure on the outer envelope of the star ( so fallback is roughly the same whether or not the star is in a binary ) . our multi - dimensional simulations of this fallback suggest that not all of this fallback material is actually incorporated into the proto - neutron star . some of it flows out of the system . the corresponding neutrino luminosity is also lower , both because less material is accreted and some of the energy released goes toward driving outflows . but the neutrinos from fallback could still dominate the neutrino emission from a proto - neutron star 10s into the explosion . at this time , observations of neutrinos from supernovae are limited to those of supernova 1987a ( hirata et al . 1987 , bionta et al . 1987 ) . with only 20 neutrinos over @xmath1615s , it is difficult to place many constraints on the fallback . some authors have claimed that claimed that the late - time neutrinos could not be easily explained by a cooling neutron star ( e.g. suzuki & sato 1987 ) , but others have argued that the observed neutrino signal is consistent with neutrinos diffusing out of a hot proto - neutron star ( burrows & lattimer 1987 ; bruenn 1987 ) . in our neutron - star forming models , fallback ends within the first 10 - 15s , comparable to the duration of the observed neutrino burst and the flux is consistent with our more standard accretion rates . the observations are consistent with fallback , but could easily be explained by a neutron star without any fallback . given that the progenitor star is believed to be greater than 15m@xmath2 , fallback is likely to have occured . but with the current errors in the time - dependent luminosity , it is difficult to determine whether or not accretion is taking place . if we assume fallback accretion did occur , the fact that the mean neutrino energy appears to be dropping with time in the observations suggests that the accretion rate is dropping at 10s . but neutrinos are still observed at 10s . unless the fallback material has considerable angular momentum , the compact remnant is not a black hole at 10s . in addition , because the accretion rate is dropping dramatically at this time , it is unlikely that the remnant will accrete much additional mass . it will remain a neutron star . in our fallback calculations , systems that accrete enough material to form a black hole are either already a black hole at 10s or still accreting rapidly at 10s , neither of which is supported by the observations . but we really need a better neutrino signal to say anything definitive . with such high accretion rates , fallback can easily dominate the neutrino luminosity after a few seconds , more than doubling the emission from the neutron star at thearly times . if we wish to study neutrino opacities in the supernova explosion , we will have to be able to calculate this fallback . it may be possible to estimate the fallback rate from detailed study of supernova light - curves . fortunately , any system that is detected in neutrinos will have a wealth of data in photons of all wavelengths . if the fallback rate is high , the infalling material will alter the position of the neutrinosphere , and the problem becomes nearly as complex as the supernova explosion modeling itself . we have only touched the surface on the difficulties in modeling such systems . in this paper , we have not addressed the production or feedback of magnetic fields , which may definitely alter the fate of the fallback . we have also not discussed the nuclear yields from the ejecta ( a first paper on this is by fryer et al . 2007 and we plan future projects studying this nucleosynthesis ) . however , note that the old `` wind '' r - process picture will not work in any system with fallback ( stars more massive than @xmath36m@xmath2 ) . the matter trajectories just ca nt be explained by the wind solution in the cases where fallback dominates the motion near the neutron star . we also have not studied the actual accretion onto the neutron star in detail . note that the final neutron star masses in table [ tab : fallback ] assumed no mass ejecta . but as we have found in this study , over half of the fallback material may be ejected , leading to smaller neutron star masses and a narrower mass range for these neutron stars . it is a pleasure to thank patrick young , frank timmes and aimee hungerford for useful conversations on this project . this project was funded in part under the auspices of the u.s . dept . of energy , and supported by its contract w-7405-eng-36 to los alamos national laboratory , and by a nasa grant swif03 - 0047 . 12 & & & & + & 0.55 & 0.33 & 1.64 & @xmath370.02 & 120 + & 0.92 & 0.23 & 1.54 & @xmath370.01 & 86 + & 2.4 & 0.030 & 1.34 & @xmath370.003 & 11 + & 3.6 & 0.015 & 1.33 & @xmath370.001 & 5.6 + 15 & & & & + & 0.9 & 0.34 & 1.75 & @xmath370.05 & 130 + & 1.65 & 0.25 & 1.66 & @xmath370.05 & 93 + & 7.0 & 0.11 & 1.52 & @xmath370.02 & 41 + & 17 & 0.08 & 1.49 & @xmath370.01 & 30 + 23 & & & & + & 1.25 & 2.2 & 3.9 & @xmath370.15 & 820 + & 2.0 & 1.75 & 3.45 & @xmath370.1 & 650 + & 2.5 & 0.86 & 2.56 & @xmath370.1 & 320 + & 6.6 & 0.02 & 1.72 & @xmath370.01 & 7.5 + ns3 & @xmath38 & @xmath39 & @xmath40 & 0 . & + bh3 & @xmath38 & @xmath41 km & @xmath40 & 0 . & + ns2 & @xmath42 & @xmath39 & @xmath40 & 0 . & + ns2-rot2 & @xmath42 & @xmath39 & @xmath43 & 0 . & + ns2-hot1 & @xmath42 & @xmath39 & @xmath40 & 18 . & + ns2-hot2 & @xmath42 & @xmath39 & @xmath40 & 36 . & + ns2alpha & @xmath42 & @xmath39 & @xmath40 & 0 . & @xmath44 + ns2bound & @xmath42 & @xmath45 & @xmath40 & 0 . & @xmath44 + bh2 & @xmath42 & @xmath46 & @xmath40 & 0 . & + bh2-rot0 & @xmath42 & @xmath46 & 0 . & 0 . & + bh2-rot2 & @xmath42 & @xmath46 & @xmath43 & 0 . & + bh2-rot10 & @xmath42 & @xmath46 & @xmath47 & 0 . & + ns1 & @xmath48 & @xmath39 & @xmath40 & 0 . & + ns1hot & @xmath48 & @xmath39 & @xmath40 & 36 . & + bh1 & @xmath48 & @xmath46 & @xmath40 & 0 . &

this paper is based on a talk at the jenam2003 meeting in budapest , hungary in august 2003 , and will cover several topics . in 2 , reference systems are defined and a brief historical sketch is given . in 3 , the fundamental synthesis formula used for the determination of radio sources positions , and two basic observed quantities , the phase and the group delay , are discussed . the international celestial reference frame ( icrf ) is described in some detail in 4 . the measurement of relative positions to obtain the motion of radio sources , and calibrator choice concerns are then given in 5 , and in 6 , the description of four experiments highlight current techniques . a brief conclusion is given in 7 , and an appendix concerning the controversy of the speed of gravity ends the paper . in order to study the position and motion of objects , a reference frame is needed . in a normal three - dimensional space ( a flat minkowski space ) , three coordinate numbers specify the location of an object . these coordinate numbers are defined on a reference frame which is given by an origin ( zero point ) and the direction of two of three orthogonal axes . in principle , one can choose an arbitrary origin and axes directions , but some reference frames are _ better _ than others . if a rotating or accelerating reference frame is used , then the objects in this frame will have peculiar motions which can not be understood by application of dynamical laws to their motions . for example , within the earth - centered reference frame ( right ascension and declination ) , the stars show an annual parallactic motion which depends on their distance and direction in the sky . these apparent motions are , of course , caused by the earth orbital motion ; hence our adopted reference frame was rotating in space . in fact , the best determination of the quality of a reference frame is to look for un - dynamic and strange correlated motions of objects in the frame , and try to ascertain what frame motion is causing the aberrant behavior of the stars . non - rotating and non - accelerating reference frames are called _ inertial _ frames , and all of the laws of mechanics are frame - independent as long as an inertial reference frame is used . but , how can these frame properties be determined without some knowledge of an _ absolute space _ with no motion and rotation a preferred reference frame ? it is generally agreed that such a frame does not exist according to mach s principle and einstein s theory of general relativity ( gr ) ; nevertheless , we can choose a reference frame which is quasi - inertial in that its origin has negligible acceleration and the reference axes have little rotation at least to the accuracy of the measurements . for the studies of celestial bodies , the most convenient celestial reference frame has its origin at the barycenter position of the solar system . the motion of the sun around the center of the galaxy , the gravitational bending from the bulge of our galaxy , and the motion of the galaxy in space produce both large static corrections ( which can be removed from the source positions ) and small differential corrections over the sky at the 0.01 mas level ( sovers , fanslow & jacobs @xcite ) . the precise position of the barycenter is in error by several km because of the uncertainty in the masses of the major solar system members , predominantly the sun and jupiter . the two axes of the barycenter frame ( usually called the pole and principle plane directions ) of the celestial reference frame are generally defined by the orientation of the earth in space at some specified time . since most observations are made from a slowly moving crust on the surface of the earth which also orbits the sun and rotates somewhat non - uniformly around a wobbling axis , the difference of the earth orientation at the observation time to that at the fiducial time is a major challenge in obtaining the accurate position of celestial objects on the adopted quasi - inertial reference frame . after defining a suitable reference frame , a method must be developed by which observations of celestial objects can be located on this frame . this is commonly done in two ways : a _ kinematic frame _ is one in which the positions of a suitable set of celestial objects are known at any time ( some of the distant objects can be considered fixed in the sky ) . thus , by comparing observations of target object with these fiducial objects , the target coordinates can be placed on the adopted reference frame . in principle , only two fiducial objects are needed to specify the frame completely , but observations of both objects and the target will often be impossible ; hence , the number of fiducial objects should be much larger so that several can be accessible near the time and position of the target observation . a _ dynamic frame _ is defined , not by a fiducial set of objects in the sky , but by an accurate ephemeris of the solar system bodies and the earth rotation and orientation . this information can be obtained , for example , by the motion of the stars which reflect the earth motion . in the 1970 s and 1980 s , the use of spacecraft tracking , laser - ranging to non - earth surfaces and planetary radar observations added significant accuracy in the determination of the orbital , rotational and orientation motions of the earth . of course , the realization of a reference frame can consist of a combination of both kinematic and dynamic information . newcomb s studies of the relative motion of stars in the @xmath0 century produced accurate values of the precessional motion of the earth s pole , the length of a day and year , by measuring the transit time of about 1000 bright cataloged stars . with these constants , the fk3 catalog ( fundamental katalog 3 ; kopff @xcite ) of the star positions and the assumed precession and nutation constants defined a celestial reference system which could be used to determine the coordinates of any star or solar system object to an accuracy of @xmath1 . the extension of catalogs to more than 10000 bright stars and more accurate proper motion determinations led to the compilation of the fk4 catalog ( fricke @xcite ) along with better precessional constants . the accuracy of this improved reference frame increase to about @xmath2 , with some degradation in the southern sky . both frames are dynamical because they relied on the modeling of the earth s orbital and spin motion using the residual motion of many stars . the fk4 system was not keeping up with the accuracy of the observations , and created pressure for continuing revision of the fundamental constants of the system . thus , the fk5 system ( fricke @xcite ) was developed around 1980 . the procedures which improved the fk5 frame were better measurements of the proper motion of thousands of stars , and the specific assumption that distant galaxies were fixed in the sky . this added a kinematical frame component to fk5 reference frame definition and improved the accuracy to about @xmath3 by 1990 . other astronomical observations the motion of near - earth asteroids and occultations of stars by planets also increased the accuracy of this frame . but , the seeds were planted for a kinematical approach to future reference frames . beginning around 1980 , the jet propulsion laboratory ( jpl ) determined a celestial reference using many types of observations , mostly of the dynamic type . the data were gathered from : optical observations of the planets , viking spacecraft range observations , radar observations of mercury , venus and mars , lunar - laser ranging , asteroid perturbations , and precision lunar modeling . this de200 ephemeris was tied to the fk5 system by proper overlap of planetary objects ( standish @xcite ) . the technique of very long baseline interferometry ( vlbi ) demonstrated in the 1970 s that many radio sources have significant emission within a component ( radio core ) less than 1 mas . these sources are identified with quasars ( galaxies with intense point - like stellar nucleii ) at distance on the order of 1 gpc . thus , it is an excellent assumption that these sources are virtually fixed in the sky , and they could define an inertial reference frame to much higher accuracy than the optical - based fk5 frame or the planetary ephemerides . over the next 20 years , improvements in the stability and observing procedures of vlbi , with more precise modeling of the earth rotation , nutation , orientation and planetary motions , led to quasar position accuracy approaching 1 mas over the sky . in the early 1990 s the astronomical community formed working groups in order to define an international celestial reference system ( icrs ) ( ferraz - mello @xcite ; feissel & mignard @xcite ) which could best utilize the accurate astrometric theory and results . the icrs was a set of rules and conventions , with the modeling required , to define at any time the orientation of the three coordinate axes ( only two are independent ) , located at the barycenter of the solar system . the axis directions were fixed relative to a suitable number of distant extragalactic sources . for reasons of continuity with the fk5 system , the icrs - defined pole direction at epoch j2000.0 was set to the fk5 pole at that epoch , and the origin of right ascension was defined by a small radio component in the source 3c273 with an accurate position measurement determined by a lunar occultation ( hazard @xcite ) . the catalog of positions of the fiducial objects needed to realize the icrs is called the international celestial reference frame ( icrf ) ( ma @xcite ) . since many telescopes around the globe participate in vlbi observations , a complementary terrestrial reference frame was needed . thus , the international terrestrial reference system ( itrs ) was formulated to describe the rules for determining specific locations on or near the earth . in order to realize the itrs , the international terrestrial reference frame ( itrf ) consists of a catalog of about 50 fiducial locations of the earth . global position service ( gps ) observations over the last 10 years have also provided high accuracy in the determination of the itrf ( see http://www.iers.org/iers/products/itrf ) . this paper will describe two aspects of astrometry : all - sky _ absolute _ astrometric techniques to determine the accurate positions used in defining the icrf , and _ relative _ astrometric techniques used to determine the motion of individual objects . however , a brief introduction to radio interferometry is needed in order to understand the calibration methods and the different observing strategies used for absolute and relative astrometry . a radio source signal , intercepted by two telescopes pointed in the same direction of sky , is essentially identical except for the different travel time ( delay ) of the radio wave from the quasar to each telescope . the correlation ( multiplying and averaging ) of the two signals produces a response called the spatial coherence function . its amplitude is related to the strength of the source and its angular size ; its phase is related to the delay of the signal between the telescope , with a minor contribution from the source structure . for an array of many telescopes , each telescope - pair ( baseline ) gives an independent measure of the spatial coherence function . using a priori calibrations of the telescope amplification properties , the coherence amplitude can be converted into true energy units ( flux density ) , and is denoted as the visibility amplitude . the coherence phase is a rapidly changing function of time because the earth rotation changes the delay between each baseline . however , an accurate model of the parameters describing the observations ( the location of the telescopes on the earth surface , the orientation and rotation of the earth in space , the position of the radio source , propagation delays in the troposphere and ionosphere , etc ) can be calculated at any time for any baseline . when this model coherence phase is subtracted from the observed spatial coherence phase , a slowly changing residual phase , the observed visibility phase , is obtained . the basic components of the observed visibility phase , @xmath4 observed at frequency @xmath5 for source @xmath6 at time @xmath7 , between telescopes @xmath8 and @xmath9 are where the structure phase of the @xmath11 source is given by @xmath12 , and is zero for a point source . @xmath13 is the model location of telescope @xmath8 , and @xmath14 is the unknown location offset ; @xmath15 is the model position of the @xmath6th source , and @xmath16 is the unknown position offset , @xmath17 is the residual instrumental and clock delay error for telescope @xmath8 , and @xmath18 is the residual tropospheric propagation delay in the direction to the @xmath11 source for telescope @xmath8 . the residual ionospheric refraction for telescope @xmath8 is @xmath19 . since this delay is produced by plasma refraction , the phase varies as @xmath20 . those quantities in bold face are vectors on the ground or directions in the plane of the sky . however , _ the measured visibility phase is only defined between @xmath21 and @xmath22 turn . _ the first line in eq . ( [ eq1 ] ) gives the terms which are associated with the source properties : its structure and position . when combined with the visibility amplitude , the ensemble of visibility amplitudes and phases for all baselines and times is the two - dimensional fourier - pair of the source brightness distribution in the sky ( thompson , moran & swenson @xcite ) . hence , the fourier transform of the visibility amplitude and phase will produce an image from which positional information can be obtained . the second line in eq . ( [ eq1 ] ) shows the major error terms which have significantly different spatial and temporal properties . for example , @xmath23 represents the slowly changing location of the telescope ( with respect to the a priori model ) which are caused by continental drift , or earth orientation , rotation and nutation uncertainties . this error term ( by virtue of the dot product with the source position ) has a period of 24 hours . the @xmath24-term represents all temporal delay errors associated with a telescope , including that from the independent maser clocks , with a typical one hour time variation time scale . the propagation delay through the troposphere @xmath25 ( or the ionosphere @xmath26 at low observing frequencies ) is the most intractable of the error sources . even with detailed ground meteorological measurements and tropospheric / ionospheric path - length monitoring from gps satellites , the a priori delay model can be significantly different than the actual path delay above each telescope . see treuhaft and lanyi ( @xcite ) , niell ( @xcite ) and macmillan & ma ( @xcite ) for more information on modeling the troposphere . in order to obtain _ absolute _ positions , all of the error contributions must be determined and parameterized as accurately as possible , and the necessary observing strategies are described in 4 . on the other hand _ relative _ positions are generally obtained by alternating observations of the target with a nearby fixed radio source ( calibrator ) with known properties , and assuming that the errors affecting the observations of the calibrator are nearly the same as that for the target source . this technique is described in 5 . for small arrays ( @xmath27 km ) , the experimental a priori model is sufficiently accurate so that cycle ambiguities of the phase are not a problem , and eq . ( [ eq1 ] ) can be used directly to obtain accurate positions ( wade @xcite ) . however , for most vlbi experiments with baselines well in excess of 1000 km , even the most accurate a priori model now available produce residual delays at telescope baselines of 5000 km which are @xmath28 psec , equivalent to a 4 cm path - length , more than one cycle of phase at @xmath29ghz . at frequencies above 2 ghz , the dominant error is caused by the troposphere ; at lower frequencies the ionospheric delay becomes dominant , and delay changes over short times scales and in small patches of sky can exceed 5 cm . thus , the measured visibility phases have cycle ambiguities and are difficult to use directly for astrometric analysis . the phase ambiguity problem can be overcome by the measurement of the derivative of the visibility phase with frequency , which is called the group delay , @xmath30 . it can be determined by observing at many frequencies simultaneously , or nearly simultaneously , to derive the phase slope . as long as some selected observing frequencies are not too widely separated , the group delay is not affected by phase cycle ambiguities . the use of the group delay for astrometric work and the need for special calibrations and frequency sampling were first described in detail by rogers ( @xcite ) , and the technique is often called bandwidth - synthesis . the derivative of eq . ( [ eq1 ] ) with respect to @xmath5 is trivial and simply removes the @xmath5 in front of the large bracket . two additional terms are produced by source structure ( sovers @xcite ) , which causes errors at the level of 5 psec , and the ionospheric refraction which can be removed by observing with a wide frequency range . hence , the solution form using the phase or the group delay is identical upon differentiation with time is more complicated . the phase rate is less accurate than the group delay because of the short - term contamination by the tropospheric phase changes , although at low elevations the rate term is a important contribution to the data . the rate is also useful for narrow - band line emission which have poorly defined group delays . ] . however , the group delay can only be adequately measured for relatively strong sources for which an accurate visibility phase can be determined in most of the simultaneously - observed channels . astrometry using weak sources is not possible ( see 5 for more details ) . the majority of the vlbi observations used for the icrf comes from observations by the nasa crustal dynamics project ( cdp ) , the united states naval observatory ( usno ) , the jet propulsion laboratory ( jpl ) , and other groups . although astrometric observations began as early as 1971 at jpl , the wide - band , dual - frequency observations which are the basis of the icrf solutions started in 1979 . this set of data will be collectively designated in this paper as cdp observations . the observations were made simultaneously at 2.3 ghz and 8.4 ghz with at least four frequency channels ( each with 16 subchannels or 16 delay lags ) at both frequencies . the spanned bandwidth at 2.3 ghz was about 0.1 ghz , and at 8.4 ghz about 0.4 ghz , from which accurate group delays were determined . by a suitable combination of the two - frequency data , the ionospheric refraction was determined and removed , to produce 8.4 ghz data which is then ionospheric - free ( sovers , fanslow & jacobs @xcite ) . the observing strategy was carefully planned in order to obtain robust solutions for all of the unknown parameters ( ma @xcite ) . each observing period ( denoted as a session ) lasted 24-hours in order to fully sample the @xmath31 terms in eq . ( [ eq1 ] ) . by about 1990 each session had evolved into a schedule which consisted of about 150 observations in which about 50 good quality sources were observed for one to ten minutes with no set number of scans per source . sources were scheduled over the entire sky in a relatively short time in order to determine the variable tropospheric refraction which produces the largest source of error . determination of each telescope clock delay and mean atmosphere zenith - path delay and gradient were determined at hourly intervals . many different arrays and telescopes have been used over the years , but the observational plan has remained essentially unchanged . the group delay and rates are analyzed by several available software packages ( modest sovers , fanslow & jacobs @xcite , calc / solve ryan @xcite ) in order to obtain all of the parameters associated with the experiment . some parameters are global , unchanged over the entire period covered by the observations ( source positions , some telescope properties ) , some parameters are slowly variable but considered as constant over each 24-hour session ( precession / nutation , ut1 offset and rate , telescope locations ) , and some parameters are extremely variable with parameter estimates made every hour ( zenith path delays and gradients , telescope clock drifts ) . the typical residual delay error for each 24-hour session after obtaining the best solution is about 25 psec , and is dominated by residual tropospheric delay residuals . this corresponds to a phase error at 8 ghz frequency of @xmath32 . cos@xmath33 and @xmath33 ( from ma 1998 , fig . 6a and 6b ) . * ( right ) * the measured nutation terms , with respect to the iau 1980 model , observed for each session . the plotted curve is fit to the data using known nutation periods with 18 yr , 9 yr , 1 yr , 0.5 yr , and 14 days period ( from ma 1998 , fig . 1a and 1b).,title="fig:",width=302 ] cos@xmath33 and @xmath33 ( from ma 1998 , fig . 6a and 6b ) . * ( right ) * the measured nutation terms , with respect to the iau 1980 model , observed for each session . the plotted curve is fit to the data using known nutation periods with 18 yr , 9 yr , 1 yr , 0.5 yr , and 14 days period ( from ma 1998 , fig . 1a and 1b).,title="fig:",width=151 ] in the fall of 1995 , a combined solution of all previous cdp observations determined thousands of parameters . the global parameters of the source positions and some telescope properties were held fixed over the whole set of data . the details concerning this complex least - squares fit to determine the most consistent set of source positions are given elsewhere ( ma @xcite ) . the resulting catalog contains three sets of sources . the ` best ' 212 sources have frequent observations , relatively small position errors , and are dominated by a small - diameter radio component . they were used to define the orientation of the axes of the icrf . in addition , there are about 300 sources with a smaller number of observations , less accurate positions , or obviously unstable positions ( due to large source structure changes ) . about 100 sources with bright optical counterparts were included for potential frame - tie connections . the sky distribution of the defining 212 defining sources in the icrf is shown in fig . 1 ( left ) . the distribution is relatively evenly spread north of the equator . the distribution for the entire set of icrf sources is given in fig . 1 ( right ) . observations are continuing and an additional 59 good quality sources have been added to the icrf - extension1 catalog . the distribution of the root - mean - square ( rms ) radio source position error is given in fig . 2 ( left ) . in order to obtain realistic , but conservative error estimates , the internal errors , generated from the scatter among the sessions , were multiplied by 1.5 and then added in quadrature to a 0.25 mas error . these errors should then include structure effects over time and frequency . the accuracy of the icrf realization of the icrs axes is about 0.02 mas . an indication of the accuracy of the cdp observations can be seen in fig . 2 ( right ) which shows the derived correction to the pole orientation ( nutation ) obtained for each 24-hour cdp session . it was fit to a realistic model known from geophysical and gravitational considerations with an rms of the fit of about 0.3 mas . this is another indication of the accuracy of the frame realization for a 24-hour long session . since the icrf frame is accurate at the level of a few 0.02 mas and is stable over time scales of decades , it is advantageous to link other frames to it . the link to the fk5 frame was discussed in connection with continuity of the direction of the pole and origin of right ascension , with the icrf adopting the fk5 pole at a specific epoch of time . the hipparcos catalog ( perryman @xcite ) contains observations between 1989.85 and 1993.21 of over 100,000 stars brighter than 12-mag . these observations were placed on the hipparcos reference frame ( hg , e. @xcite ) to an accuracy of 1.0 mas . the link between the hipparcos optical frame and the icrf was accomplished in several ways . first , the few bright stars with significant radio emission were observed with the vlbi , the vla and with merlin in order to determine the radio positions with respect to the icrf ( lestrade @xcite ) . accurate parallaxes and proper motions were also obtained for these radio emitting stars ( boboltz @xcite ) . secondly , comparison of the hipparcos optical with hst observations linked these stars to extragalactic objects , and then to nearby icrf objects which could be detected by hst . next , additional optical observations using ground based telescopes also helped link hipparcos positions to nearby galaxies . finally , the earth orientation parameters were obtained from the hipparcos observations and then compared with that obtained from that from nearly concurrent vlbi observations ( lindegren & kovalevsky @xcite ) . with these comparisons , the hipparcos optical catalog is now consistent with the icrf to about 0.6 mas in position and 0.25 mas / yr for axis rotation . however , the position of an individual star and the reference frame tie decreases in accuracy with time because of the proper motion uncertainties of the stars of 1 mas per year . the link of the icrf to the planetary dynamic ephemeris was accomplished using several techniques . first , the time of arrival of a pulsar signal is sensitive to the orbital motion of the earth . by comparing the pulsar position derived from timing analysis with that obtained with vlbi observations , improvements in the dynamics of the earth motion could be obtained ( dewey @xcite ) . secondly , occultations of radio sources with solar system objects also tie the icrf to that of the planetary ephemeris . finally , radio interferometry of spacecraft with respect to nearby icrf source also improved the frame tie between the two systems ; for example , the observations the magellan and pioneer spacecraft with respect to nearby quasars ( folkner @xcite ) . comparison of the lunar ranging with vlbi earth orientation results were also important in tying the two frames together ( folkner @xcite ) . the resulting jpl de405 ephemeris is connected to the icrf to about 1.0 mas in position and about 0.05 mas / yr in rotation ( standish @xcite ) , and the tie has recently been improved by a factor of two or three ( border 2003 , private communication ) the icrf will be the best realization of a quasi - inertial system at least until 2015 when the space interferometry mission ( sim ) has been in operation for several years . hence , improvements in the icrs and the accuracy of the icrf positions are needed , and may be obtained in several ways . first , many more observations of the icrf sources have been made since the original solution was made in 1997 , and improved positions are designated as the icrf - ext1 catalog . a further update , icrf - ext2 , is expected ( fey 2004 , in preparation ) . secondly , vlba observations , in its search for suitable calibrators needed for phase referencing ( beasley @xcite ; fomalont @xcite ) , have produced a catalog of over 2000 radio sources , some of which are candidates for additions to the icrf catalog . also , the number of experiments for sources in the southern hemisphere ( roopesh @xcite ) has enlarged over the last ten years . a new defining list of icrf sources may be generated in the next five years , with an expected improvement in the grid accuracy of about a factor of two . however , discontinuities between the old and new frame definitions must be avoided . a major source of error in the parameters associated with the icrf is caused by the variable tropospheric delay ; it is the major contributor to the current limit of 0.25 mas accuracy for the good quality sources . although the troposphere zenith path - delay and gradient are modeled every hour during most 24-hour sessions , the structure and kinematics of the troposphere are extremely complicated , and the typical rms delay scatter , after obtaining the best solution , is about 25 psec at a 5000-km baseline , which corresponds to a positional error of about 0.2 mas at 8 ghz . more sophisticated models , better monitoring of the tropospheric delay using gps satellites , and ground measurements of water vapor emission in the line of sight to a source could decrease this uncertainty to one - half of the present level , but progress has been slow . an example of the gain in precision by an improvement in the tropospheric modeling is shown in fig . 3 ( left ) . the plot shows the difference in the derived declination of sources between two reductions of all of the 24-hour sessions in 1995 : one reduction used an azimuthally - symmetric tropospheric delay model over each telescope and the other included a tropospheric gradient term ( which is now routinely used ) . the differences as a function of source declination are up to 0.4 mas and show that systematic errors would have been present in the icrf frame if the simpler tropospheric model had been used . any remaining systematic error could be as large as 0.1 mas . another source of error is associated with the evolution of quasars their changing structure . while these distant objects are fixed in the sky , the radio emission does not emanate from the galactic nucleus , but from the inner jet region which may be located 0.1 mas from the core at 8 ghz , and could be somewhat variable in position . another source of error is caused by radio - emitting clouds of magnetic plasma which propagate down the jet and are often much brighter than the radio core nearer the galaxy nucleus . 3 ( right ) shows the apparent position versus time for several quasars which have undergone such evolution . the typical time - scale of changes are several months to years , and apparent position movements of 0.1 mas are common , and occasional changes as large as 1 mas can occur . with our knowledge of the nature of quasar radio emission , and with frequent monitoring of the radio source emission structure , it should be possible to model these evolutionary changes to remove the effects of these apparent positional changes ( fey & charlot @xcite ; sovers @xcite ) ) . finally , the icrf positions are based on radio source properties at 8.4 ghz , but observations and studies are now underway to define a complementary icrf at a frequency of 32 ghz ( lanyi @xcite ) . at this higher frequency , the ionospheric refraction is smaller and the radio source structures are more compact . however , radio variability and structure changes will still be a problem . nasa - jpl plans to use 32 ghz for spacecraft telemetry , beginning in 2005 . with an icrf defined at this frequency , the tie between the planetary ephemerides and the quasar reference reference could be strengthened . however , such a tie is now perturbed by the limited modeling of the asteroids at the level of 0.2 mas per year which may be the dominant error ( standish & fienga @xcite ) . the advances in absolute astrometric measurements over the next 20 years ( including space interferometry ) may reach accuracies at the level of 0.01 mas , when a true space - time reference frame must be considered . this means that the astrometric system must become four - dimensionally based rather than the current system which is three dimensional , but with correction factors associated with the variable gravitational field of the solar system . also , there is some discussion that the origin of the icrf should be moved to the barycenter of the earth - moon system , rather than that of the solar system , because of the lower gravitation potential and diminished gravitational space distortion near the earth compared with that near the sun ( soffel @xcite ) . many astrophysical phenomena are associated with the accurate space motion of celestial objects , as well as their accurate position . for a source in the milky way , the parallax is the only direct method of determining the distance , and its proper motion is often related to the evolution and formation of the object . from the orbital motion of a source , many properties of a binary or multiple system , including the detection of planet - like objects , can be made . for solar system objects the accuracy of radio interferometry is similar to that of lunar - laser ranging and range / doppler measurements , and is complementary since interferometry determines positions on the celestial sphere and the other techniques determine the distance . finally , the effect of the gravitational field in the solar system on the propagation of radio waves can be measured and compared with that predicted by gr or other theories of gravity . the determination of the motion of a radio source requires less complicated observations and reductions than that associated with all - sky icrf observations . the reason is that the motion of a radio source can be obtained by measuring its position with respect to any detectable radio source , called a calibrator , that is _ fixed _ in the sky . if the calibrator - target separation in the sky is small and observations are switched rapidly between the calibrator and target , then the major error terms , shown in eq . ( [ eq1 ] ) , are similar for the two sources and nearly cancel when their observed visibility phases are differed . this observation scheme is called _ phase referencing _ ( beasley & conway @xcite ) . from eq . ( [ eq1 ] ) the phase difference , @xmath34 at any time ( t ) between the target source @xmath35 and the calibrator @xmath36 , for the baseline between telescopes @xmath8 and @xmath9 is @xmath37 + \frac{1}{\nu c}\delta i^{s - c}_{l , m}(t)\end{aligned}\ ] ] the first line shows the phases which dependent on the target and calibrator source structures and positions . the next two lines contains the differential delay errors . the small separation of the two sources , @xmath38 , significantly decreases the effect of the telescope - location errors ( in the most general sense ) , for example , by a factor of 50 for a one degree target - calibrator separation . similarly , @xmath39 is the difference of the tropospheric delay error in the direction of the two sources , and @xmath40 is the ionospheric delay . the quasi - random short - term delays from small clouds will not cancel particularly well , even for close calibrators , but larger angular - scale deviations from the a priori model will cancel . this term will be discussed in more detail using multi - calibrator observations . finally , the purely temporal ( mostly clock ) delay variations would cancel precisely if the calibrator and target were observed simultaneously , simultaneously ( honma @xcite ) . ] . otherwise , time - interpolation between calibrator observations is needed , and small second order clock errors will remain in the difference . the switching time between observations of the calibrator and target depend on temporal characteristics of the @xmath24 and @xmath25 terms in eq . ( [ eq2 ] ) . the longest time for which interpolation between two observations will produce an accurate phase is called the _ coherence _ time , and varies from 30 sec at 23 ghz to 5 min at 1.4 ghz . during periods of inclement whether or strong ionospheric activity , the coherence time can decrease to 10 sec , making phase referencing impossible unless the calibrator and target are observed simultaneously . the differential delay error between the calibrator and target are often less than 5 psec . even at a high frequency of 23 ghz , the corresponding phase difference in @xmath41 is less than one cycle and ambiguities of the phase are unlikely to occur . ( changes more than one cycle can be used as long as the phase between subsequent calibrator observations are connected properly . ) the group delays can still be obtained by observing at several frequencies ; however , this is not recommended , unless necessary , for several reasons . first , the precision of the visibility phase is greater than that of the group delay by the factor @xmath42 where @xmath5 is the observing frequency and @xmath43 is the spanned frequency range used to determine the group delay . for a reasonably strong sources , the typical delay error using the measured phase is @xmath44 psec , while that for the group delay is about 10 psec . since the all - sky observations , described for the cdp observations have typical residuals of 25 psec , which are dominated by the troposphere errors , the use of the much more accurate phase data ( if it were possible to sort out cycle ambiguities ) would not provide more accurate solutions than use of the group delay . on the other hand , the delay errors from a phase - referencing experiment are often @xmath45 psec , which is less than the inherent group delay accuracy , but not the phase accuracy . another reason for using the phase , rather than the group delay , is to preserve the imaging capability afforded by the visibility phase . if the error terms in the second line of eq . ( [ eq2 ] ) can be determined ( or assumed to be negligible ) , then fourier imaging of the residual phase will produce an image of the source with its position relative to that of the calibrator . even if the target source is extremely weak and can not be detected during a single observation , such fourier imaging of the entire target data set will produce an image from which accurate astrometric information can be obtained . for these weak sources , the group delay can not be measured . the two most important properties of a calibrator source are its proximity in the sky to the target , and the strength of a compact radio component which must be detectable within a coherence time . in order for the calibrator phase to be accurately determined in a few minutes , the radio core must contain no less than 20 mjy ( using the vlba at 8 ghz at 64 mhz bandwidth which gives an rms noise of 4 mjy ) . the use of a much stronger calibrator source which is more distant from the target than a fainter , but otherwise suitable calibrator , is not recommended . the decrease in the measurement phase error from a stronger source is more than balanced by the increase in differential errors between the calibrator and target . many calibrators contain very extended emission and even most compact radio cores are not true point sources . the phase effect of this structure is contained in the @xmath46 term in eq . ( [ eq2 ] ) . using the self - calibration techniques for which images of strong sources can be made ( cornwell & fomalont @xcite ) , the structure phase can be obtained and removed . thus , imaging of all calibrators should be done routinely . a more serious problem is the variable position offset of the calibrator @xmath47 between observing sessions , since this change is translated directly into an apparent position change of the target . this problem is associated with extended sources where the maximum radio brightness or centroid brightness moves along the jet and with respect to the ( fixed ) galaxy nucleus ( see fig . 3 ( right ) for some examples ) . even for compact sources , it is possible that its position shifts with time since the radio emission probably emanates from the inner part of the radio jet which may not be a stationary position with respect to the galaxy nucleus . when trying to reach astrometric levels @xmath48 mas in a target over a period of time longer than a few months , it is crucial to image the calibrator in order to determine the probable angular change from the calibrator evolution . only a limited number of suitable calibrators are contained in the icrf catalog . over 2000 additional calibrators can be found in the vlba surveys of calibrations ( beasley @xcite ; fomalont @xcite ) , and most have derived positions which are within 1 mas of the icrf grid because they have been observed in one or more of the cdp 24-hour sessions . independent searches for calibrator sources , which are near a desired target source , also can be undertaken using preliminary observations of km - sized arrays ( vla , wsrt , atca ) to search for faint , compact flat - spectrum , candidate sources . at 1.4 ghz , the field of view of many vlbi arrays are sufficiently large so that a calibrator can usually be found sufficiently close to the desired target , and both can then be observed simultaneously . away . the fit is shown by the curve . * ( right ) * the motion of pulsar b0950 + 08 using a calibrator that is @xmath49 away , but using a multi - frequency technique for removing the differential ionospheric refraction . the best fit proper motion and parallax are shown by the curve.,title="fig:",width=226 ] away . the fit is shown by the curve . * ( right ) * the motion of pulsar b0950 + 08 using a calibrator that is @xmath49 away , but using a multi - frequency technique for removing the differential ionospheric refraction . the best fit proper motion and parallax are shown by the curve.,title="fig:",width=113 ] in this section several phase - referencing astrometric experiments illustrate the techniques which are currently being used . the first example for b0656 + 14 illustrates a simple phase - referencing ( in - beam ) experiment . the second example for b0950 + 08 describes a scheme for removing ionospheric refraction using multi - frequency observations . the third example for hd8703 illustrates observations over many years to determine the orbital motion of a resolved radio source , and the problems associating with calibrator structure changes over time . all three of these examples assume that the phase error of the calibrator is a perfect measure of the phase errors of the target . the last example , the measurement of the deflection of a quasar by the jovian gravitational field , shows that by using more than one calibrator , relative positional accuracies less than 0.01 mas can be obtained from a six - hour observation . the pulsar psr b0656 + 14 was observed with the vlba at 1.67 ghz for five sessions , separated by about six months each , between epochs 2000.9 to 2002.5 ( briskin @xcite ) . a calibrator , j0658 + 1410 with flux density 35 mjy , was found by preliminary vla observations near the pulsar , only @xmath50 away , within the reception area of the vlba antennas , so that both ` in - beam ' calibrator and target were observed simultaneously . because of the proximity of the calibrator and target , the differential delay errors were assumed to be zero . the calibrator image was sufficiently point - like ; hence , its structure phase was assumed to be zero and its apparent position was assumed fixed in the sky since it is almost certainly an extra - galactic radio source . images of the pulsar were then made from @xmath51 using the standard fourier - techniques . the pulsar position was determined from the location of the peak of the pulsar image , with error estimates . its peak flux density was 3.6 mjy and was far too weak to be detected in a single observation and to derive a group delay . using the images of the pulsar made from each session , the proper motion and parallax were accurately determined . the motion associated with the parallax motion ( after removal of the proper motion ) is shown in fig . 4 ( left ) . the parallax @xmath52 mas , corresponding to a distance of @xmath53 pc . the rms of the fit for each epoch was about 0.6 mas which is impressive for a relatively weak radio source . another experiment , also a measurement of the motion of a pulsar , did not have the luxury of an available in - beam calibrator . using the vlba , alternating observations of the pulsar b0950 + 08 and calibrator j0946 + 1017 , about @xmath49 away , were made every minute over a period of 7 hours , for 4 sessions between 1998.33 and 1999.85 ( briskin @xcite ) . with the relatively large calibrator - target separation , the ionospheric refraction at 1.6 ghz was expected to be the dominant source of error . since the pulsar is strong , about 64 mjy , the observations of the pulsar and calibrator were made at eight frequencies , simultaneously , between 1.31 ghz and 1.71 ghz . the differential phase error terms in eq . ( [ eq2 ] ) can then be written more generally as @xmath54 where @xmath55 contains all of the non - dispersive delay terms ( independent of frequency ) , and @xmath56 is the expected ionospheric delay component between the two sources . by fitting the differential phase observed over the eight frequencies to a function of the form @xmath57 , both delay terms can be determine , and the ionospheric contribution removed from the observed differential phase . for telescopes more than 1000 km apart , the ionospheric - induced position shift can be as large as 10 mas per degree of source - calibrator separation , changing significantly within a time - scale of a few minutes ; however a shift of 1 mas per degree over five to ten minutes of time is more typical for the source position jitter . pulsar images , made from the phase before the ionosphere correction , were severely distorted . the pulsar position was then obtained from the ionospheric - free images , and the position of b0950 + 08 determined in the same manner as for b0656 + 14 . the structure phase of the calibrator j0946 + 1017 was determined at each observation epoch and removed from the differential phase . the calibrator appearance did not change significantly , hence it was assumed to be stationary in the sky . the derived motion of the pulsar is shown in fig . 4 ( right ) . the proper motion is @xmath58 mas yr@xmath59 , @xmath60 mas yr@xmath59 , and @xmath61 mas . the gravity probe b ( gp - b ) mission , developed by nasa and stanford university , will fly four precision gyroscopes in earth orbit in order to measure two general relativistic precessions , a geodetic effect and a frame - dragging effect ( turneaure @xcite ) . these precessions will be measured with respect to a suitably bright star whose position must be known to an accuracy of @xmath62 mas . thus , the goal of the associated radio observations was to find a suitable bright star with sufficient radio emission , and to determine its position and motion with the necessary accuracy . after exploratory observations , the rs can ven binary - star system , hr8703 , was chosen . the radio emission comes from a main - sequence f - g star and the companions is a fainter k star with a known period of 24.65 days . this project illustrates some of the problems associated with both the target and the calibrator with variable source structure . the radio position of hr8703 has now been measured from more than 20 observation sessions since early 1997 , and is still continuing . the vlba , the vla and the three nasa deep - space network telescopes , at 8.4 ghz , had sufficient sensitivity that accurate group delays could be obtained for the star when it was stronger than 5 mjy , which was most of the time . the primary calibrator used was 3c454.3 ( 2251 + 158 ) , about @xmath63 from the star . a second calibrator source , b2250 + 194 , about @xmath49 away , was also used , as a control object to determine the accuracy of the observations . there are two groups performing the data reduction . one group is using the differential phase between hr8703 and 3c454.3 , and the preliminary result reported here is from this group ( ransom ( @xcite ) . another group is analyzing the experiment using the group delay . using phase - referencing of hd8703 with 3c454.3 as the calibrator , good quality images of the binary system were made for each session . for some of the sessions , the radio emission showed two radio components separated by 1 mas , in some cases only one component was observed determined . the best fit to the parallax and proper motion ( fig . 5 ( left ) ) , was obtained by using the peak position of the radio emission when one component was detected , and using the geometric center when two components were detected . ( this interpretation is consistent with the radio emission mechanism of fron rs can ven stars . ) also , the residual motion of the star , after removing the parallax and proper motion fit , showed a repeating pattern of 24.65 day period and , hence , this motion was associated with the binary orbit of the radio emission ( fig . 5 ( right ) . the estimated error for the proper motion , parallax and orbital elements is about 0.1 mas . the positions of hd8703 are with respect to that of 3c454.3 . however comparison of the relative positions of 3c454.3 ( see fig . 4 ( right , bottom ) ) and b2250 + 194 over the experiment period show relative motion between them in the order of 0.3 mas with a time scale of one year . since both calibrators have extended , evolving structure , it is difficult to determine which source contributes to the apparent motion , and some of this shift may appear in the positions derived for hd8703 . taken over the five years of observations , these position errors lead to an proper motion uncertainty of @xmath64 mas yr@xmath59 , which is still within the limited needed for gp - b mission . however , a more detailed analysis of the calibrator changes and other phase error terms in eq . ( [ eq2 ] ) is currently underway . on september 8 , 2002 when jupiter passed within @xmath65 of a background quasar j0842 + 1835 , its gravitational field produced an apparent change ( deflection ) in the position of the quasar . at closest approach , according to gr , the deflection has two major components : a radial deflection of 1.19 mas , and a smaller deflection in the direction of motion of jupiter of 0.051 mas . this smaller deflection component is caused by the gravitational aberrational produced by the relative motion of jupiter and the earth , and is related to the speed of propagation of gravity ( kopeikin @xcite ; frittelli @xcite ) . thus , by measuring this aberrational component to an accuracy less than 0.01 mas , an estimate of the speed of gravity could be obtained . however , this precision was about a factor of three to five better than had been previously obtained ( see previous experiments as good examples ) . such a gain in positional accuracy could only be obtained by determining the phase error terms associated with eq . ( [ eq2 ] ) more accurately . although phase referencing between a calibrator and target removes much of the temporal dependence of the phase error , any coherent angular dependence in the sky will produce systematic phase errors between the target and calibrator , and produce distorted images and systematic errors in the position of the peak of the emission . this error can be diminished by choosing a calibrator closer to the target ( if you can find one ) , or by using _ more than one calibrator _ to determine the angular dependence of the phase error . preliminary vlba test observations in 2001 suggested that such a multi - calibrator scheme was effective . additional details on multi - source calibrations are given by fomalont ( @xcite ) . the design of the jupiter deflection experiment is shown in fig . 6 ( left ) . the observations were made at 8.4 ghz with the vlba and effelsberg , germany telescope . in order to determine the phase errors toward the target source j0842 , two calibrators , j0839 and j0854 on opposite sides of the target , were observed . each source was observed for 1.5 min in turn , and the trio of scans were repeated every 4.5 min for ten hours over the day . although the deflection of j0842 was strongest on 2002 september 8 , identical observations on five days ( september 4 , 7 , 8 , 9 , 12 ) were made in order to obtain sufficient redundancy to estimate realistic positional errors of the deflection . more details concerning the design and analysis of this experiment are given by fomalont & kopeikin ( @xcite ) . 6 ( right , top ) shows the measured phase for each of the three sources for a typical observing day and baseline . the overall temporal behavior of the phases , dominated by the clock and gross troposphere delay errors , are similar for the three sources . however , after ut=17h the phases for the three sources show displacements which are consistent with a simple phase gradient error covering the three sources : the displacement for j0839 from j0842 is in the opposite sense and about 25% of that between j0854 and j0842 , as expected from their relative positions in the sky . thus , the interpolation of the phase of j0839 and j0854 ( weighted by their inverse distance from j0842 ) gives a very good estimate of the phase associated with j0842 . the corrected phase of j0842 , @xmath66 , after the two source calibration , is shown in fig . 6 ( right , bottom ) . the phase scatter is small with a rms less than 0.05 cycle , or about 0.040 mas . the slight offset from zero phase is associated with the sum of the position offsets from the a priori values , as seen in the expression for @xmath66 , @xmath67 \nonumber \\ + \hbox{terms of 2$^{nd}$ order in time and angle } \nonumber\end{aligned}\ ] ] where the 0.80 and 0.20 are the calibrator weighting factors for this particular configuration . the position of the j0842 can be obtained by fourier inversion of the corrected phase and measured visibility amplitude , or by a least - square fit of the phases if the source is sufficiently strong . the resultant displacement of the radio from the image center is a measure of the linear combination of offset positions ( from the a priori positions ) , shown in eq . ( [ eq4 ] ) . assuming that the calibrator positions have not varied , any changes in the target position with time reflect a change in the target position . the goal of the september 2002 observation was to determine the speed of propagation of gravity , @xmath68 , from the determination of the aberrational part of the quasar deflection , as outlined in fig . 7 ( left ) . for @xmath69 , the expected deflection in the direction of jupiter s motion is 0.051 mas at closest approach . more generally , the deflection is proportional to @xmath70 . a graphical display of the experimental determination of this deflection term is given in fig . 7 ( right ) which shows the measured position of j0842 every hour on september 8 , relative to the other four observing days . these positions were determined from images of the residual phases and the observed visibility amplitudes . the measured aberrational deflection , averaged over september 8 , was @xmath71 mas , which gives a deflection that is @xmath72 times that predicted by gr . the error estimates are consistent with the position of the target on the four off - jupiter days when no change of position is expected . when converted into the speed of propagation of gravity , the result is @xmath73c . a similar analysis on the data using only one calibrator , j0839 , gave a positional error of 0.027 mas compared with 0.009 mas achieved using a second calibrator . after a discussion of inertial frames and reference systems , the methods for determining the position of radio sources were discussed . first , all - sky observations used for the icrf were used to determine absolute positions of radio sources which formed the definition of the celestial inertial reference system now in use . secondly , phase referencing was described using four different experiments to illustrate several aspects of measuring the motion of radio sources to accuracies approaching 0.010 mas per session . although the experimental result is not in doubt , there is disagreement over the correct role of gravity in this experiment . kopeikin ( @xcite ) , using the einstein equations directly , interprets the retarded deflection component as a measure of the speed of propagation of gravity ( see also kopeikin & fomalont ( @xcite ) ) . will ( @xcite ) concludes , however , that the retarded deflection component is associated with the aberration of light , not of gravity . other interpretations are given by asada ( @xcite ) and samuel ( @xcite ) . at the jenam2003 meeting in budapest , a lively give - and - take session took place during the late afternoon of august 28 . there were about 20 - 30 participants . the main points of view were presented by g. schfer ( institute for theoretical physics , friedrich schiller university , jena , germany ) , sergei kopeikin ( university of missouri - columbia , mo , usa ) , and gabor lanyi ( jet propulsion laboratories , pasadena , ca , usa ) schfer supported the point of view of asada and will that the experiment measured the speed of radio waves used in observations . the main arguments were : 1 . gravity and light move along null geodesics and their speed can not be disentangle in the experiment ; 2 . the speed of gravity is associated with accelerated motion of light - ray deflecting bodies and thus , only with gravity waves ; 3 . the propagation of gravity effect shows up only in terms of order of @xmath74 in the metric tensor but the experiment is sensitive to @xmath75 terms only . kopeikin interpretation is that that the experiment measured the speed of propagation of gravity in the near zone of the solar system . in answer to schfer s points , kopeikin pointed out that : 1 . the relativistic deflection of light by jupiter depends on the euclidean product of two null vectors at the point of observation the vector of light propagation from the quasar and that of gravity propagation from jupiter . because the quasar and jupiter are located at different positions on the sky , the gravity propagation direction can be separated from the light propagation direction and , thus , the speed of gravity could be measured . 2 . there are different types of gravitational fields according to algebraic classification b. y. petrov . only gravitational fields of type n ( plane gravitational waves ) are generated by the accelerated motion of bodies , and they exist only in the radiative zone of the solar system . other types of gravitational fields are more general and can be associated with the velocity of the bodies . they dominate in the near zone of the solar system where the jovian deflection experiment has been done . nevertheless , gravitational fields of the other petrov s types also propagate with finite speed which can be measured . it is true that propagation of gravity effect shows up only in terms of order of @xmath74 in the metric tensor . but the metric tensor is not directly observable quantity in the jovian deflection experiment because it depends on specific choice of coordinates . therefore , any reference to the metric tensor in connection with measured gravitational effects is irrelevant . what is observed is the phase of radio waves coming from the quasar which is perturbed by gravitational field of jupiter from its retarded position , in accordance with einstein s equation predicting that gravity travels with finite speed equal numerically to the speed of light in vacuum . this relativistic effect of the retardation of gravity in the electromagnetic phase is gauge - independent and shows up already in terms of order of @xmath75 which we have observed . the sections on reference frames and the icrf were significantly improved from the suggestions and corrections by chris jacobs . i also thank chopo ma for his contributions , and ryan ranson for the preliminary hr8703 results . ryan , j.w . , clark , t.a . , ma , c. , gordon , d. , capretee , d. & himwich , w.e . 1993 , in contributions of space geodesy to geodynamics : earth dynamics , ed d.e . smith & d. l. turcotte ( washington : am.geophys . unions ) , 37

the ability to coherently transfer atoms out of a bose - einstein condensate ( bec ) in a trapped atomic gas is of fundamental importance in the creation of coherent matter wave generators , often called `` atom lasers '' . theoretical interest has developed toward characterizing the out - coupling stage of such an atom laser , with the hope of developing efficiency criteria for this process . all this work has been done in the zero temperature limit , using either a non - interacting gas coupled to a laser mode @xcite or the modified gross - pitaevskii dynamics of the condensate order parameter @xcite . a complete understanding of the out - coupling process necessarily must take into account the presence of the thermal cloud composed of noncondensate atoms . recently , the role that the noncondensate atoms play in the out - coupling process has been discussed by japha , choi , burnett and band @xcite . in their work , a `` pair - breaking '' contribution was obtained , in which a correlated pair of atoms take part in the output process , with one output atom appearing simultaneously with the creation of an excitation in the trapped bose gas . out - coupling experiments are only a small class of experiments that can be performed in which a bec is coupled to another gas of atoms ( that may or may not be bose - condensed ) . the mit group @xcite has performed work in which a bec is scattered from an optical bragg grating . they show that the excitation spectrum of the scattered atoms gives a direct experimental measure of the dynamic structure factor of the initial bose - condensate . since the detailed form of the dynamic structure factor @xmath2 is dependent upon the many - body correlations of the bose - condensate , this kind of experiment provides a means of probing the unique correlations induced by a bose - condensate @xcite . despite the large amount of work that has been performed with regard to the coupling of bose - condensed gases with another ( possibly bose - condensed ) gas , there still does not exist in the literature a clean formulation of such a coupling . even though refs.@xmath3@xcite have sparked interest in this issue , the formalism used in these papers does not show in a transparent way which correlations present in a bose - condensed gas can be probed by such experiments . moreover , the case in which both trapped gases are bose - condensed was not considered . in this paper , we set up a general linear response theory approach for treating the coupling of two bose gases at finite temperatures , which gives a formally exact expression for the tunneling current . the general expressions we derive within linear response theory are independent of the specific approximation used for the many - body states of the two uncoupled , trapped bose gases . in this sense , our results are the analogue of showing how the inelastic neutron scattering cross - section is related to the dynamic structure factor @xcite and the general theory of tunneling in superconductors @xcite . the work closest to ours is by meier and zwerger @xcite , who applied a similar formalism in discussing coherent josephson tunneling between two uniform bose - condensed gases . in sections [ section : linear response theory - general ] and [ section : tunneling currents in terms of correlations functions ] , we give a detailed account of the many - body theory needed to find formal expressions for the tunneling currents in terms of single - atom correlation functions . we apply our theoretical formalism to a class of experiments involving raman scattering of atoms between trapped bose gases . we calculate the tunneling current expression at finite temperatures using the bogoliubov - popov approximation for the single - particle excitations in the trapped bose gases @xcite . we discuss coherent tunneling of atoms between two bose - condensed gases , as well as the out - coupling of atoms from a bec in a trap to the vacuum ( i.e. , to free particle states ) . in this latter case , as expected , we show that our more general formalism reproduces the results of refs.@xmath3@xcite in the appropriate limit . in section [ sec : lda ] , we give an explicit calculation of the out - coupling current based on treating the noncondensed atoms within a simple local density approximation ( lda ) , such as used in refs . @xcite . in essence , we show that besides the dominant contribution from tunneling of atoms out of the condensate , one finds a much smaller contribution which arises from the atoms tunneling out of the noncondensate . this latter contribution , however , is of special interest since it is directly related to the single - particle correlation function of the noncondensate atoms . one component has weight proportional to the bogoliubov amplitude @xmath4 , where @xmath5 is the bose - distribution function for an excitation of energy @xmath6 , and increases with temperature . the other , more interesting , component has weight proportional to the bogoliubov amplitude @xmath7 . this latter process is formally related to the `` negative energy pole '' of the bogoliubov spectral density . this so - called `` pair - breaking '' term @xcite is of interest as a unique signature of the unusual correlations induced by a bose condensate . it corresponds to the out - coupling of an atom and the simultaneous creation of a bogoliubov excitation in the trapped gas . this contribution already is present at zero temperature and , in contrast to the tunneling current from the condensate , contributes to an output tunneling current of atoms with a wide spectrum of energies . as the temperature increases , the weight of the @xmath8 term decreases while the @xmath9 term starts to build up ( we recall that as the condensate density goes to zero , @xmath10 and @xmath11 ) . the analogue of this pair - breaking process was discussed earlier in the context of a bose - condensed gas of excitons @xcite . in contrast to our work , a proposal has recently been put forward @xcite to measure these bogoliubov amplitudes using an optical bragg scattering technique . in this approach , a two - photon bragg scattering beam is used to excite quasiparticles out of the condensate and then a second bragg beam is used to probe the altered momentum distribution of the sample . recent work at mit has beautifully verified these predictions @xcite . unlike this situation , in the raman scattering experiment the energy absorbed by the atoms imparts a momentum kick in addition to causing a change of hyperfine state , allowing the scattered atom to no longer be trapped . our results give another fairly direct method of measuring the bogoliubov quasiparticle amplitudes . furthermore , since the weight of the out - coupling current is proportional to the bogoliubov amplitude @xmath1 , it is significantly amplified for larger values of the s - wave scattering length @xmath12 such as can be achieved by working close to a feshbach resonance . we hope our work further stimulates experimental interest in this kind of out - coupling experiment , even though the current from the noncondensate atoms is quite small compared to that from the condensate . our goal is to calculate the current of atoms tunneling between two atomic gases coupled through a weak external field , which can be treated as a time - dependent perturbation to a time - independent many - body hamiltonian using linear response theory ( see , for example , refs.@xmath3@xcite ) . the total hamiltonian in the schrdinger representation is@xmath13 where @xmath14 is the hamiltonian for the many - body systems , @xmath15 and @xmath16 are the atomic hyperfine level energies of an atom in gas 1 and gas 2 respectively , @xmath17 is the atomic number operator for gas 1(2 ) , given by@xmath18 the external probe field @xmath19 is explicitly time - dependent . we assume that the two many - body systems do not interact with one another in the absence of @xmath20 , so that @xmath21 . this amounts to assuming that the two gases do not overlap , _ _ _ the coupling term of the kind @xmath22 can be neglected . in the absence of @xmath19 , the number of atoms in the two systems is fixed by their chemical potentials @xmath23 and @xmath24 . the tunneling perturbation @xmath19 couples the two many - body systems together by introducing a mechanism by which an atom can tunnel between the two systems . this perturbation has terms which create an atom in one system while destroying an atom in the other system , and vice - versa . an example of the kind of experiment described by the tunneling perturbation @xmath19 is the out - coupling phenomenon considered in refs . @xcite . in this experiment , an external electromagnetic ( em ) field causes transitions between two atomic hyperfine levels , one of which is not trapped . we label the quantum states corresponding to these hyperfine levels by @xmath25 and @xmath26 . assuming that hyperfine level @xmath25 is lower in energy than hyperfine level @xmath26 , energy will be absorbed from ( emitted to ) the external field facilitating the transition @xmath27 ( @xmath28 ) . in more general experiments , however , the transition can take place between two atomic hyperfine levels that are both trapped . an example of such a process is a stimulated raman transition between the magnetic hyperfine levels @xmath29 and @xmath30 ( for a detailed discussion , see ref . we will return to the specifics of a bec coupling experiment in section [ section : coupling between two trapped gases ] when we calculate the tunneling current between two trapped bose - condensed gases . for the moment it suffices to write down the semiclassical form of the tunneling perturbation @xmath20 arising from the electric - dipole and rotating wave approximations @xcite , @xmath31 here @xmath32 is the effective energy transfer from the laser fields , @xmath33 is the momentum transfer from the laser fields , and @xmath34 ( taken to be real ) is the interaction strength that is associated with the rabi frequency of the transition between atomic states . we have also introduced a factor @xmath35 that models the adiabatic switching on of the interaction at @xmath36 . in our final results , we take the limit @xmath37 . we use units in which @xmath38 and @xmath39 , unless otherwise specified . we next outline the formalism for calculating the tunneling current between two bose gases ( resulting from the action of @xmath20 ) . the state of the combined system evolves in time according to @xmath40 , where the evolution operator @xmath41 is unitary and satisfies the equation of motion@xmath42 this equation of motion can be formally solved to give @xmath43 , where @xmath44 denotes the time - ordering operator . in the heisenberg picture , operators are written in terms of the corresponding schrdinger picture operators as @xmath45 , while the state of the system is given by the initial state of the system at @xmath36 , namely @xmath46 . we use the subscript @xmath47 to denote an operator or state in the heisenberg picture with respect to the full hamiltonian @xmath48 . the heisenberg equation of motion for any operator @xmath49 is then given as@xmath50 .\ ] ] we define the atomic current _ from _ gas 1 _ to _ gas 2 in terms of the thermal equilibrium ensemble average of @xmath51@xmath52 we have introduced a negative sign in this expression for the atomic current so that the current flow from gas 1 to gas 2 is positive . the ensemble average in ( [ current definition ] ) is given by @xmath53 \hat{u}(t)\right\ } , \label{ensemble average for n1dot 1 } \end{aligned}\ ] ] where the initial equilibrium density operator @xmath54 describes the state of the system at @xmath36 before the coupling @xmath20 is turned on . since @xmath55 = \left [ \hat{v}(t),\hat{n}_{1}\right ] $ ] , the average in ( [ ensemble average for n1dot 1 ] ) takes the form@xmath56 \hat{u}(t)\right\ } .\ ] ] we now turn to the task of evaluating this ensemble average within linear response theory @xcite . we first write the average ( [ ensemble average for n1dot 2 ] ) in an interaction picture , in which @xmath20 is taken as a perturbation to the many - body hamiltonian @xmath57 of the uncoupled gases . later we will introduce an interaction picture involving @xmath58 . with respect to @xmath59 , we define the interaction representation as @xmath60 in this interaction picture , we also define the time - dependent density operator as @xmath61 using these definitions , we now can write ( [ ensemble average for n1dot 2 ] ) as @xmath62 \mathrm{e}^{-i\hat{h}_{0}^{\prime } t}\right\ } .\ ] ] evaluating the commutator @xmath63 $ ] in ( [ ensemble average for n1dot 3 ] ) , we obtain@xmath64.\ ] ] the density operator in the interaction picture defined in ( [ density matrix in interaction picture ] ) satisfies the liouville equation @xmath65 , \ ] ] which has a formal solution@xmath66 .\ ] ] iterating ( [ liouville equation in integral form ] ) to first order in @xmath20 , we obtain @xmath67 here@xmath68\ ] ] is the perturbation to the initial equilibrium density operator @xmath54 to first order in @xmath20 , the linear response approximation to the density matrix . within this first - order approximation for @xmath69 the ensemble average of an operator @xmath49 is given by @xmath70 where @xmath71 the atomic current given by ( [ current definition ] ) can thus be written in linear response theory as@xmath72 where @xmath73 is the equilibrium current and @xmath74 is the linear response current . substituting our approximation for @xmath69 of ( [ linear response theory approximation to density matrix in interaction picture ] ) together with ( [ v ] ) for @xmath20 into ( [ ensemble average for n1dot 4 ] ) , we obtain , after some algebra , explicit expressions for the equilibrium current @xmath75 and the linear response current @xmath76}\left\langle \left [ \hat{\psi } ^{\dagger } _ { 1h_{0}^{\prime } } ( \mathbf{r}^{\prime } , t^{\prime } ) \hat{\psi } _ { 2h_{0}^{\prime } } ( \mathbf{r}^{\prime } , t^{\prime } ) , \hat{\psi } _ { 1h_{0}^{\prime } } ( \mathbf{r},t)\hat{\psi } ^{\dagger } _ { 2h_{0}^{\prime } } ( \mathbf{r},t)\right ] \right\rangle _ { eq}\nonumber \\ & & + \gamma ^{2}\mathrm{e}^{i[\mathbf{q}\cdot ( \mathbf{r}+\mathbf{r}^{\prime } ) -\omega ( t+t^{'})]}\left\langle \left [ \hat{\psi } _ { 1h_{0}^{\prime } } ( \mathbf{r}^{\prime } , t^{\prime } ) \hat{\psi } ^{\dagger } _ { 2h_{0}^{\prime } } ( \mathbf{r}^{\prime } , t^{\prime } ) , \hat{\psi } _ { 1h_{0}^{\prime } } ( \mathbf{r},t)\hat{\psi } ^{\dagger } _ { 2h_{0}^{\prime } } ( \mathbf{r},t)\right ] \right\rangle _ { eq}\biggr \}.\label{linear response current 1 } \end{aligned}\ ] ] the expressions in ( [ equilibrium current 1 ] ) and ( [ linear response current 1 ] ) are the basis of the rest of this paper . one sees that there are two different contributions to the linear response current in ( [ linear response current 1 ] ) , corresponding to phases depending on the times through @xmath77 or @xmath78 . the former will be seen shortly to correspond to a _ normal _ current of single atoms between the two gases , while the latter corresponds to an _ anomalous _ current of atoms between the two gases . we next want to express the ensemble averages in ( [ equilibrium current 1 ] ) and ( [ linear response current 1 ] ) in terms of thermal green s functions given by standard techniques in finite temperature many - body theory @xcite . to this end , we first express the time - dependent field operators in ( [ equilibrium current 1 ] ) and ( [ linear response current 1 ] ) in the usual `` heisenberg picture '' used in many - body theory , namely the interaction picture in which the free hamiltonian is taken to be the many - body `` grand canonical hamiltonian '' @xmath79 using the identity @xmath80 , it is easily verified that @xmath81 where we define @xmath82 . using ( [ relation between destruction operators in h0prime and k0 interaction pictures ] ) , it is easy to rewrite ( [ linear response current 1 ] ) in terms of the @xmath83 operators . we also note that the grand canonical ensemble average of a product of operators of the two gases factorizes @xmath84 where @xmath85 . below the transition temperature of a bose gas , the new phase of matter is described by a symmetry - broken state in which the equilibrium density operator does _ not _ commute with the number operator : @xmath86 \neq 0 $ ] ( the state breaks the system s @xmath87 symmetry ) . thus , the equilibrium ensemble averages are truly over the @xmath88-ensemble introduced by bogoliubov @xcite , which is symbolically defined as @xmath89 where @xmath90 is the symmetry - breaking grand canonical equilibrium density operator . the symmetry - breaking hamiltonian is defined as @xcite @xmath91 @xmath92 is the broken symmetry grand canonical ensemble partition function . we will use the notation @xmath93 to denote equilibrium ensemble averages , with the understanding that they are over a broken symmetry state . an equivalent way of stating this is that atomic field operators can be decomposed into two parts : @xmath94 where @xmath95 is the symmetry - breaking order parameter field and @xmath96 is the fluctuation field operator associated with the noncondensate atoms . in ( [ order - parameter ] ) , @xmath97 is the local condensate density while @xmath98 is the condensate phase . in our formalism based on linear response from thermal equilibrium , the bose order parameter in ( [ order+fluctuation - field ] ) is that appropriate to thermal equilibrium . a phase @xmath99 would arise if we were dealing , for example , with a condensate with a vortex . we also note that the usual time - dependence of the bose - order parameter ( @xmath100 ) is absent because we are working in the @xmath101 interaction representation . we do not need to assume that the inverse temperature is the same for both gases . indeed , because the two gases are assumed to be uncoupled in the absence of the perturbation @xmath20 , no mechanism exists to force both gases to reach a common thermodynamic equilibrium . therefore , the equilibrium ensemble averages in ( [ equilibrium current 1 ] ) , ( [ linear response current 1 ] ) and ( [ ensemble average decomposition ] ) may be viewed as being over two independent gases . the equilibrium tunneling current in ( [ equilibrium current 1 ] ) is first order in the perturbation @xmath20 , whereas the linear response current in ( [ linear response current 1 ] ) is second order in @xmath20 . using ( [ relation between destruction operators in h0prime and k0 interaction pictures ] ) , ( [ equilibrium current 1 ] ) can be written as @xmath102 where @xmath103 , @xmath104 . introducing ( [ order - parameter ] ) into ( [ equilibrium current 2 ] ) , the equilibrium current is reduced to@xmath105.\ ] ] the magnitude of this equilibrium current is seen to be proportional to the overlap of the two condensate density profiles and thus serves as a probe of the condensate surface structure . in a non - bose condensed gas , the ensemble averages @xmath106 and @xmath107 in ( [ equilibrium current 2 ] ) vanish . as noted in ( [ order - parameter ] ) and the discussion preceding it , these ensemble averages do _ not _ vanish in a bose - condensed system . thus , the equilibrium current ( [ equilibrium current 3 ] ) is a direct manifestation of bose broken - symmetry . this expression for the equilibrium current is a more general form of the _ linearized _ `` internal '' josephson current @xcite for bose - condensed gases . our expression ( [ equilibrium current 3 ] ) differs from previous results in several ways . firstly , there is the phase associated with the momentum @xmath33 transferred by the two laser beams . this will introduce momentum kicks ( bragg scattering ) to the atoms in the two condensates @xcite . secondly , we have allowed for the case of a spatially varying condensate phase difference , @xmath108 , such as would arise if one of the condensates had a vortex . the quantity @xmath109 in ( [ equilibrium current 3 ] ) is analogous to the voltage in the standard discussion of the josephson current in superconductors @xcite . we will show later ( see ( [ deltaiaa ] ) and ( [ deltaiab ] ) ) that this quantity also enters into the phase of the anomalous component of the linear response current in ( [ linear response current 1 ] ) as a second harmonic . we may interpret the atomic energy level splitting of the atomic states @xmath110 in addition to the chemical potential difference @xmath111 of the two bose gases as a negative voltage impeding the atomic current flow _ from _ gas 1 _ to _ gas 2 . the energy difference between the laser photons acts as a forward biased voltage serving to negate this reverse bias . in the regime where @xmath112 , we recover what is termed the d.c . josephson effect in superconducting systems @xcite . in this regime , as with superconductors , a non - zero difference between the phases of the two condensate order parameters drives a d.c . atomic current . we next turn to the linear response current in ( [ linear response current 1 ] ) , which is second order in the tunneling perturbation @xmath20 . we first want to write it in terms of thermal green s functions by introducing correlation functions of the atomic field operators . as in the previous section , we rewrite our time - dependent interaction picture field operators in terms of a heisenberg picture defined with respect to @xmath101 , rather than @xmath59 ( see ( [ relation between destruction operators in h0prime and k0 interaction pictures ] ) ) . we consider the first term in the expression of the linear response current in ( [ linear response current 1 ] ) . we refer to this term as the normal current , for reasons which shall become clear . the normal current is then given by ( @xmath113 ) @xmath114.\label{normal current definition } \end{aligned}\ ] ] we have evaluated the commutators in ( [ linear response current 1 ] ) making use of ( [ ensemble average decomposition ] ) to write the ensemble averages as a product of two - point correlation functions such as @xmath115 in subsequent discussion , we will refer to @xmath116 as @xmath117 for notational simplicity . the normal current given in ( [ normal current definition ] ) is independent of the time @xmath118 . we make this more explicit by rewriting the integral and noting that factors of @xmath35 factor can be neglected in the limit @xmath37 . introducing the ( heaviside ) step function @xmath119 , the normal current may be written in the explicitly time - independent form @xmath120.\label{normal current 1 } \end{aligned}\ ] ] we observe that , as expected for a current , the second term in ( [ normal current 1 ] ) breaks time - reversal symmetry ; the current in the negative time direction is the negative of that in the positive time direction . the correlation functions appearing in ( [ normal current 1 ] ) may now be written in terms of their fourier transforms @xmath121 and hence ( [ normal current 1 ] ) becomes @xmath122.\label{normal current 2 } \end{aligned}\ ] ] the integral over the time @xmath123 in ( [ normal current 2 ] ) is trivially given by @xmath124 in the physical limit @xmath37 , we may use the identity @xmath125 where @xmath126 denotes the principal value of an integral . therefore , the normal current ( [ normal current 2 ] ) may be written in the form @xmath127 where we have introduced the two contributions @xmath128\label{deltaina } \end{aligned}\ ] ] and @xmath129}{\omega ^{\prime \prime } -\omega ^{\prime } -\omega + \delta \varepsilon + \delta \mu } .\label{short time normal current } \end{aligned}\ ] ] we next show how the correlation functions in ( [ deltaina ] ) and ( [ short time normal current ] ) may be converted to spectral densities , which in turn can be directly calculated from time - ordered single - particle green s functions @xcite . thus , the problem is reduced to the application of well - known techniques from many - body theory to calculate these green s functions @xcite . we first rewrite the correlation functions in terms of the order parameter fields and the single - particle correlation functions of the noncondensate field operators . the diagonal correlation functions above become @xmath130 and@xmath131 using the detailed balance identity @xcite@xmath132 in addition to the bose identity @xmath133 one may easily write @xmath134 in terms of a spectral density ( see p.@xmath37 of ref.@xmath3@xcite ) @xmath135 here @xmath136 is the bose distribution function @xmath137 and the single - particle spectral density is @xmath138 \right\rangle _ { eq}.\ ] ] note that for an inhomogeneous system , this spectral density is in general not real . in turn , @xmath139 is given by the difference of the retarded and advanced green s functions , @xmath140 and @xmath141 respectively , for the noncondensate quantum field operators of a bose gas ( see p.@xmath3150 of ref.@xmath3@xcite ) : @xmath142 where @xmath143 is the fourier transform of @xmath144 \right\rangle _ { eq}\ ] ] and @xmath145 is the fourier transform of @xmath146 \right\rangle _ { eq}.\ ] ] the point of all this formalism is that standard finite - temperature many - body theory @xcite most naturally determines thermal green s functions at imaginary bose matsubara frequencies @xmath147 . fortunately , one can determine @xmath148 by a unique analytic continuation @xmath149 to the real frequency axis . this procedure thus gives the spectral density @xmath150 . we may express ( [ deltaina ] ) and ( [ short time normal current ] ) in terms of the condensate order parameters and single - particle noncondensate spectral densities . we find@xmath151 and@xmath152 in these expressions we have introduced the normal current densities @xmath153 which corresponds to atomic tunneling of single particles between the condensate and noncondensate components of the two gases . in contrast , @xmath154 is associated with atomic tunneling of single atoms between the noncondensates of the two gases . we observe that in both @xmath155 and @xmath156 , atoms make only two types of transitions between the states of the two gases . the first transition process involves tunneling of an atom between the condensate of one gas and the noncondensate of the other . it is represented in @xmath155 and @xmath156 by terms of the form @xmath157 . this contributes at all temperatures including @xmath158 . the second transition process involves the tunneling of an atom between the _ noncondensates _ of the two gases . it appears in @xmath155 and @xmath156 as a term of the form @xmath159 . this kind of transition can occur at zero temperature only when one of the gases is bose - condensed . finally we consider the `` anomalous current '' given by the second term in ( [ linear response current 1 ] ) in terms of the condensate order parameters and noncondensate spectral density functions . we follow the identical procedure used to find ( [ deltaina 1 ] ) and ( [ deltainb 1 ] ) for the contributions @xmath155 and @xmath156 to the normal current . we obtain finally@xmath160.\label{anomalous current definition } \end{aligned}\ ] ] clearly this current may be written entirely in terms of the so - called anomalous correlation functions involving creation or destruction of _ pairs _ of atoms , hence the name anomalous current . using @xmath161 , we first rewrite ( [ anomalous current definition ] ) as @xmath162\bigg \}.\label{anomalous current 1 } \end{aligned}\ ] ] this shows that the anomalous current is explicitly _ time - dependent _ , in contrast to the normal quasiparticle current given by ( [ normal current 2 ] ) . it clearly involves a second harmonic of the first - order josephson current in ( [ equilibrium current 3 ] ) . following the same procedure we used in writing the normal current in terms of the condensate order parameters and noncondensate spectral density functions of the of the two gases , the anomalous current may be rewritten as @xmath163 where we have introduced the quantities @xmath164\label{deltaiaa } \end{aligned}\ ] ] and @xmath165.\label{deltaiab } \end{aligned}\ ] ] the quantities @xmath166 and @xmath167 in the above expressions for @xmath168 and @xmath169 are given by @xmath170 and @xmath171 where we have introduced the total anomalous current density @xmath172 , which is written as @xmath173 here @xmath174 is the anomalous noncondensate single - particle spectral density function , analogous to the noncondensate single - particle spectral density function @xmath139 defined in ( [ spectral - density - definition ] ) . the first and second terms in ( [ total anomalous tunneling current ] ) correspond to a tunneling current between the condensate of one gas and the noncondensate of the other gas . the third term in ( [ total anomalous tunneling current ] ) represents a tunneling current between the noncondensates of the two gases . one can rewrite @xmath166 and @xmath167 such that the contributions from the condensate - noncondensate and noncondensate - noncondensate tunneling processes are made more explicit . thus , we have ( see also ref . @xcite ) @xmath175 and @xmath176 where we have introduced the condensate - noncondensate current density ( the analogue of ( [ condensate - noncondensate current density ] ) ) @xmath177 and the noncondensate - noncondensate current density ( the analogue of ( [ noncondensate - noncondensate current density ] ) ) @xmath178 we note that @xmath169 in ( [ deltaiab ] ) is the current most directly associated with the _ coherent _ tunneling responsible for the josephson effect in superconductors . this current is in - phase with the josephson current in ( [ equilibrium current 3 ] ) but involves a second harmonic with a frequency @xmath179 . it clarifies the role of the noncondensate in the bec josephson effect . on the other hand , the @xmath168 current in ( [ deltaiaa ] ) is directly related to a _ dissipative _ interference current , just as one finds for the josephson effect in superconductors @xcite . in this section , we use the results of section [ section : tunneling currents in terms of correlations functions ] to calculate the second order tunneling current involved in a raman scattering experiment . the spectral density functions are calculated in the finite temperature bogoliubov - popov approximation @xcite . the josephson tunneling current has already been discussed and is given by ( [ equilibrium current 3 ] ) . a two - photon bragg scattering experiment involves two laser beams at different frequencies and with different propagation vectors @xcite . however , we are interested in a different type of two - photon process in which an atom changes its hyperfine state . this is a raman process and is depicted in fig.@xmath3[fig : raman - figure ] . we now discuss how the effective frequency @xmath32 and wavevector @xmath33 in ( [ v ] ) are defined in such a raman scattering experiment @xcite . assuming that each laser beam has equal intensity @xmath180 , the combined intensity of the two beams at space - time point @xmath181 is given by @xmath182 where @xmath32 is the frequency difference between the two laser beams@xmath183 the wavevector difference between the two laser beams is @xmath184 where @xmath185 is the wavevector of laser beam @xmath186 . we may take @xmath187 such that the standard bragg scattering relation ( see , for example , p.@xmath3103 of ref.@xmath3@xcite ) holds and @xmath188 is given by @xmath189 where @xmath119 is the angle between the two counter - propagating beams . finally the coupling parameter @xmath34 in ( [ v ] ) is related to the single - photon rabi frequency @xmath190 ( assumed to be the same for both hyperfine level transitions ) and the laser beam detuning @xmath191 from the intermediate level @xmath192 ( see fig . [ fig : raman - figure ] ) according to @xcite @xmath193 where @xmath194 is the so - called two - photon rabi frequency " . the many - body hamiltonian ( in the absence of a tunneling perturbation ) is @xmath21 , where @xmath195 \hat{\psi } _ { 1}(\mathbf{r})+\frac{g_{1}}{2}\hat{\psi } _ { 1}^{\dagger } ( \mathbf{r})\hat{\psi } _ { 1}^{\dagger } ( \mathbf{r})\hat{\psi } _ { 1}(\mathbf{r})\hat{\psi } _ { 1}(\mathbf{r})\right\ } .\ ] ] in this expression for @xmath196 , @xmath197 is the trapping potential and @xmath198 is the s - wave scattering pseudo - potential with @xmath199 being the s - wave scattering length for atoms in gas @xmath200 . the expression for @xmath201 has an analogous form . we recall that the spectral densities in section [ section : tunneling currents in terms of correlations functions ] are defined in terms of the grand canonical hamiltonian @xmath202 . thus the excitation energies of gas @xmath203 are measured relative to the chemical potential @xmath204 . in the standard static bogoliubov - popov approximation for excitations in a trapped bose gas @xcite , the noncondensate field operator @xmath96 can be written as@xmath205 where @xmath206 and @xmath207 are the bogoliubov quasiparticle amplitudes , @xmath208 is the energy of bogoliubov state @xmath186 and @xmath209 ( @xmath210 ) annihilate ( create ) a quasiparticle in state @xmath186 , respectively . these annihilation and creation operators are defined to satisfy the boson commutation relations@xmath211 = \left [ \hat{\alpha } _ { m}^{\dagger } , \hat{\alpha } _ { n}^{\dagger } \right ] = 0;\qquad \left [ \hat{\alpha } _ { m},\hat{\alpha } _ { n}^{\dagger } \right ] = \delta _ { mn}.\ ] ] the bogoliubov quasiparticle amplitudes must satisfy the finite temperature coupled bogoliubov - popov equations of motion @xcite@xmath212 in order for @xmath101 to be reduced to the diagonal form @xmath213 in ( [ bogoliubov - equations - of - motion ] ) , we have introduced the differential operator@xmath214 where @xmath215 is the total static density . using the commutation relations ( [ bogoliubov - commutation - relations ] ) , one finds that the single - particle spectral density function can be expressed in terms of bogoliubov quasiparticles as@xmath216,\ ] ] which has the fourier transform@xmath217.\ ] ] we note the appearance in ( [ bogoliubov - spectral - density - frequency ] ) of negative energy poles ( @xmath218 ) in addition to the usual positive energy poles ( @xmath219 ) . these are a result of unique processes induced by the bose condensate . the two poles follow immediately from ( [ bogoliubov - quasi - particle - quantization ] ) , which shows how the destruction of an atom involves a linear superposition of the destruction ( amplitude @xmath220 ) and creation ( amplitude @xmath221 ) of bogoliubov excitations . similarly , the off - diagonal or anomalous single - particle spectral density function of relevance to the anomalous current densities @xmath222 in ( [ condensate - noncondensate anomalous current density ] ) and @xmath223 in ( [ noncondensate - noncondensate anomalous current density ] ) is given by @xmath224.\ ] ] the spectral density functions can be shown to satisfy @xmath225 and @xmath226 . using the results in ( [ bogoliubov - spectral - density - frequency ] ) and ( [ bogoliubov spectral density for anomalous correlations ] ) , we can now write the normal and anomalous linear response currents in terms of bogoliubov excitations . in section [ section : tunneling currents in terms of correlations functions ] , we found that two distinct physical processes contribute to the normal current . the current density of atoms tunneling between the condensates and noncondensates of the two gases is given by ( [ condensate - noncondensate current density ] ) , or @xmath227.\label{condensate - noncondensate normal current density in bogoliubov approximation } \end{aligned}\ ] ] this type of process involves both condensate atoms and quasiparticles . the current density associated with the tunneling of atoms between the noncondensates of the two gases has a more complicated form given by ( [ noncondensate - noncondensate current density ] ) , or @xmath228.\label{noncondensate - noncondensate normal current density in bogoliubov approximation } \end{aligned}\ ] ] as in the case of the normal current , two different physical processes contribute to the anomalous current in ( [ anomalous current definition ] ) . the first process involves pairs of condensate atoms tunneling to noncondensate states of the other bose gas . this anomalous current density is given by ( [ condensate - noncondensate anomalous current density ] ) , namely @xmath229.\label{condensate - noncondensate anomalous current density in bogoliubov approximation } \end{aligned}\ ] ] the expression for the anomalous current density between condensate 1 and noncondensate 2 has a similar form . the final process involves quasiparticles of both gases and is given by ( [ noncondensate - noncondensate anomalous current density ] ) , namely@xmath230.\label{noncondensate - noncondensate anomalous current density in bogoliubov approximation } \end{aligned}\ ] ] as with the josephson current given in ( [ equilibrium current 3 ] ) , one sees from ( [ condensate - noncondensate normal current density in bogoliubov approximation])-([noncondensate - noncondensate anomalous current density in bogoliubov approximation ] ) that the magnitude of the normal and anomalous quasiparticle currents is determined by the spatial overlap of the bogoliubov wavefunctions described by @xmath231 and @xmath232 for the two systems . when one uses ( [ condensate - noncondensate normal current density in bogoliubov approximation])-([noncondensate - noncondensate anomalous current density in bogoliubov approximation ] ) in the expressions for the normal ( see ( [ deltaina 1 ] ) and ( [ deltainb 1 ] ) ) and anomalous ( see ( [ deltaiaa ] ) and ( [ deltaiab ] ) ) currents , our results are essentially equivalent to expressions discussed in ref . the main differences are that meier and zwerger approximated the two bose gases as being uniform ( or homogeneous ) and used the single - particle spectral densities at @xmath158 . in this section , we consider the special case of atoms out - coupling to the vacuum as discussed in refs . @xcite . more precisely , we first deal with the case when gas 2 is a uniform non - interacting gas . the corresponding spectral density function then has the well - known form @xcite@xmath233 where@xmath234 is the the excitation energy of an atom in gas @xmath235 . because the gas is assumed to be in the classical limit , we have @xmath236 . the only terms in ( [ condensate - noncondensate normal current density in bogoliubov approximation ] ) which are finite reduce to @xmath237 and in ( [ noncondensate - noncondensate normal current density in bogoliubov approximation])@xmath238v_{1j}(\mathbf{r}^{\prime } ) v^{*}_{1j}(\mathbf{r})\delta \left ( \omega + e_{1j}\right ) \biggr \}\delta ( \tilde{\varepsilon } _ { \mathbf{k}2}-\omega ^{\prime } ) .\label{outcoupling current density noncondensate - noncondensate } \end{aligned}\ ] ] we have used the bose identity @xmath239 $ ] in writing the last term in ( [ outcoupling current density noncondensate - noncondensate ] ) . the condition that @xmath240 results in negligible tunneling from gas 2 back to gas 1 . the current density @xmath241 in ( [ outcoupling current density condensate - noncondensate ] ) corresponds to the _ destruction _ of a condensate atom in gas @xmath200 and the _ creation _ of an excitation in gas @xmath235 with energy @xmath242 . the first term in the density @xmath243 of ( [ outcoupling current density noncondensate - noncondensate ] ) represents what might be called the `` quantum evaporation '' of an atom , _ destroying _ a bogoliubov quasiparticle with energy @xmath244 and _ creating _ an atom in gas @xmath235 with @xmath245 . the second term in ( [ outcoupling current density noncondensate - noncondensate ] ) corresponds to _ creating _ an atom in gas @xmath235 with energy @xmath245 and simultaneously _ creating _ a quasiparticle excitation in gas @xmath200 with energy @xmath246 $ ] . this second out - coupling channel in ( [ outcoupling current density noncondensate - noncondensate ] ) is a direct consequence of the unusual condensate - induced correlations in a bose gas . it corresponds to the `` pair - breaking '' process discussed by the oxford group @xcite . the analogue of this process appears in the predicted lineshape of the recombination photons emitted in the decay of an exciton in a bose - condensed gas of optically - excited excitons @xcite . all three out - coupling contributions will be calculated in section [ sec : lda ] . using ( [ outcoupling current density condensate - noncondensate ] ) and ( [ outcoupling current density noncondensate - noncondensate ] ) , the out - coupling current @xmath247 is@xmath248v_{1j}(\mathbf{r}^{\prime } ) v^{*}_{1j}(\mathbf{r})\delta \left ( k^{2}/2m-(\omega -\delta \varepsilon + \mu _ { 1})+e_{1j}\right ) \biggr ] \biggr \}.\label{bogoliubov outcoupling current } \end{aligned}\ ] ] we refer to the second term as the `` @xmath0 contribution '' and the third term as the `` @xmath1 contribution '' . it is useful to describe the main features of the out - coupling current spectrum arising from the three terms in ( [ bogoliubov outcoupling current ] ) , as a function of the energy @xmath245 of the out - coupled atoms , for fixed @xmath32 and * @xmath33*. the energy conserving delta functions in ( [ bogoliubov outcoupling current ] ) determine the energies of the out - coupling current peaks for the different processes . there is a delta function peak at @xmath249 from atoms coming directly from the condensate , where @xmath23 is the condensate energy of a trapped atom in the hyperfine state @xmath25 and @xmath250 is the hyperfine state energy . relative to this condensate pole , the peak associated with the quantum evaporation term is shifted by a positive amount @xmath251 , while the peak of the `` pair - breaking '' channel is shifted by a negative amount @xmath252 . these shifts follow from the fact that the quantum evaporation channel arises from the positive energy pole ( related to @xmath0 ) and the pair - breaking channel corresponds to the negative energy pole ( related to @xmath1 ) in the single - particle spectral density . we next consider the behavior of the three out - coupling current contributions in ( [ bogoliubov outcoupling current ] ) as a function of temperature . for @xmath253 , since @xmath254 , ( i.e. no thermally excited quasiparticles ) the second channel ( @xmath0 ) has zero weight . thus , at @xmath253 , only the first and third processes occur . when gas 1 is at a finite temperature ( @xmath255 ) , the second channel ( @xmath0 ) begins to contribute to the out - coupling current , taking weight away from the pair - breaking channel ( @xmath1 ) . all these features are shown explicitly in the numerical results given in section [ sec : lda ] ( see figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5nm]-[fig : total out - coupling current for q=3d3d10/aho a=3d3d5 nm t=3d3d0.2tbec ] ) . the out - coupling current derived in ref . @xcite is given as the rate of atoms excited out of gas 1 with energy @xmath242 and may be written as ( where we use a notation more consistent with our own)@xmath256v_{1j}(\mathbf{r}^{\prime } ) v^{*}_{1j}(\mathbf{r})\delta \left ( k^{2}/2m-(\omega -\delta \varepsilon + \mu _ { 1})+e_{1j}\right ) \biggr ] \biggr \},\label{oxford - result } \end{aligned}\ ] ] where @xmath257 is the wavefunction for the out - coupled excitation with energy @xmath242 . we can make a direct comparison between the integrands of ( [ bogoliubov outcoupling current ] ) and ( [ oxford - result ] ) . indeed taking @xmath258 ( a free atom with momentum @xmath259 ) reduces ( [ oxford - result ] ) to the integrand of ( [ bogoliubov outcoupling current ] ) . thus our result for the out - coupling current reduces to that derived in refs.@xmath3@xcite , where the out - coupled atoms were in free - particle states . in this section , we evaluate the out - coupling current in ( [ bogoliubov outcoupling current ] ) and estimate the magnitude and temperature - dependence of the quantum evaporation and pair - breaking effects discussed in section [ sec : case of out - coupling to a non - bose - condensed gas ] . in previous work @xcite , these noncondensate processes were found to be a few percent , relative to the dominant contribution of atoms tunneling out of the high density condensate . it has also been found @xcite that in current bose gas experiments , the contribution of the noncondensate atoms to the dynamic structure factor @xmath2 at @xmath260 is about @xmath261 of the contribution from the condensate fluctuations ( the latter has been studied in refs . @xcite at @xmath158 ) . this small amplitude is , of course , a direct reflection of the strongly peaked condensate at the centre of the trap compared to the broad , low density profile of the noncondensate atoms . however , while small , the tunneling current from the noncondensate atoms gives direct information about the bose - induced correlations and thus is a worthwhile goal for future bragg scattering and raman experiments . the condensate current in ( [ bogoliubov outcoupling current ] ) can be rewritten in the form@xmath262 where @xmath263 is the fourier transform of the bose order parameter @xmath264 and we have introduced the abbreviation @xmath265 . it is also useful to introduce a `` detuned frequency '' @xmath266 the noncondensate contribution in ( [ bogoliubov outcoupling current ] ) is more conveniently written in terms of the trapped gas bogoliubov spectral density in ( [ bogoliubov - spectral - density - frequency ] ) , @xmath267 in the rest of this section , for notational simplicity , we shall drop the index `` 1 '' on the quantities associated with the trapped gas . equations ( [ condensate outcoupling current ] ) and ( [ noncondensate outcoupling current ] ) summarize our main results for the tunneling currents into the vacuum , showing very clearly which properties of the trapped bose gas they depend on . in the standard thomas - fermi approximation , the condensate density profile is given by @xcite @xmath268 where the chemical potential is@xmath269 and @xmath270 the thomas - fermi radius @xmath271 defines the size of the condensate , with @xmath272 the total number of particles in the bose - condensate at temperature @xmath273 and @xmath274 is the oscillator length . the @xmath158 thomas - fermi results have been extended to finite temperature by adjusting @xmath275 according to the noninteracting trapped gas result @xcite @xmath276 where @xmath277 is the noninteracting trapped gas bose - condensation temperature . using ( [ thomas - fermi condensate density ] ) for the order parameter @xmath278 , the fourier transform in ( [ order parameter fourier transform definition ] ) can be calculated analytically @xmath279 where @xmath280 is the ordinary bessel function of order 2 . this result is the same as previously obtained @xcite , taking into account that we use harmonic oscillator units ( length is measured in units of @xmath274 and energy is measured in terms of @xmath281 ) . writing the integral in ( [ condensate outcoupling current ] ) in terms of spherical coordinates , we obtain using ( [ order parameter fourier transform ] ) @xmath282 where @xmath283 . the result ( [ lda condensate outcoupling contribution in harmonic oscillator units ] ) may be easily generalized to asymmetric harmonic traps , having trapping potential @xmath284 by introducing the coordinate transformation@xmath285 where @xmath286 . the same procedure as used above follows through with the same final expressions , apart from @xmath287 being replaced by @xmath288 . in addition , any wavevector @xmath259 in our tunneling expressions such as ( [ condensate outcoupling current ] ) needs to be interpreted as @xmath289 in an anisotropic trap @xcite . the total out - coupling current from the condensate given by ( [ condensate outcoupling current ] ) , as a function of the detuning parameter @xmath290 in ( [ definition of detuning parameter ] ) , is shown by the solid lines in figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5 nm ] and [ fig : total out - coupling current for q=3d3d2/aho a=3d3d5 nm t=3d3d0.2tbec ] . in all our numerical calculations , for the two - photon rabi coupling we have used @xmath291 . note the logarithmic scale for the out - coupled atom current . the current magnitude only decreases as a function of increasing temperature by about @xmath292 . from fig . [ fig : condensate - noncondensate out - coupling for q=3d3d10/aho and a=3d3d5 nm ] , it can be seen that the maximum condensate tunneling current amplitude is sensitive to the magnitude of the momentum transfer @xmath33 . the decrease in the maximum tunneling current with increasing values of @xmath188 is a result of the fact that the momentum distribution of atoms in the condensate is peaked at zero momentum , decreasing for increasing momentum . for an asymmetric trap , this means that the directions of strongest confinement ( largest trap frequency ) will out - couple most efficiently . we observe also that the peak of the condensate tunneling current shifts toward positive @xmath290 as @xmath188 increases . the out - coupling current ( [ noncondensate outcoupling current ] ) from the noncondensate atoms will be estimated using a simple local density approximation ( lda ) . the lda has also been recently used to calculate @xmath2 in trapped gases @xcite . we first introduce a new coordinate system defined in terms of the relative coordinate@xmath293 and the center - of - mass coordinate@xmath294 writing the single - particle spectral density function @xmath295 in ( [ noncondensate outcoupling current ] ) in terms of these new coordinates , we note that for large condensates and quasiparticle energies much larger than @xmath281 , the spectral density will vary slowly as a function of @xmath296 . it is thus plausible to approximate @xmath139 by its homogeneous form , but adjusted so that at each point @xmath296 the appropriate condensate density @xmath297 is used . thus , in the lda , the single - particle spectral density function in ( [ noncondensate outcoupling current ] ) is approximated by@xmath298 where @xmath299 denotes the _ uniform _ gas expression for the spectral density function for a condensate density @xmath97 . in the bogoliubov - popov approximation , the lda form ( [ lda approximation for spectral density ] ) for the spectral density in energy - momentum space is @xcite@xmath300.\label{lda bogoliubov spectral density in momentum space } \end{aligned}\ ] ] thus , in the lda the quasiparticle energies become @xmath301 while the bogoliubov quasiparticle mode amplitudes are given by@xmath302 we have introduced the free atom energy @xmath303 and@xmath304 plays the role of an effective " local chemical potential , modifying the chemical potential @xmath305 as a result of the inhomogeneity of the trap . within the lda , the noncondensate contribution ( [ noncondensate outcoupling current ] ) to the out - coupling current is given by@xmath306 where as before we have introduced the variable @xmath307 for notational simplicity . using the identity@xmath308 where @xmath309 , we may evaluate the integral over @xmath296 in ( [ lda noncondensate outcoupling current version 1 ] ) assuming a spherically symmetric trap . after some algebra , one obtains the following expression for the noncondensate contribution to the out - coupling current@xmath310 where we have introduced the lda spatially averaged spectral density function in energy - momentum space@xmath311 where@xmath312 and harmonic oscillator units are used . note that the noncondensate out - coupling current in ( [ lda noncondensate outcoupling current in harmonic oscillator units ] ) may be generalized to an asymmetric trap in the same manner as discussed earlier for the condensate out - coupling current ( see remarks below ( [ coordinate transform for asymmetric trap to symmetric trap ] ) ) . the formula in ( [ lda noncondensate outcoupling current in harmonic oscillator units ] ) has a simple physical interpretation . the integrand gives the number of emitted atoms with momentum @xmath259 per second . the first term in ( [ lda energy - momentum space spectral density function ] ) is associated with the positive energy pole in ( [ bogoliubov - spectral - density - frequency ] ) , with amplitude @xmath0 , and only occurs for positive frequency @xmath313 . the second term in ( [ lda energy - momentum space spectral density function ] ) arises from the negative energy pole in ( [ bogoliubov - spectral - density - frequency ] ) , with amplitude @xmath1 , and only occurs for negative frequency @xmath314 . these first two terms are associated with quasiparticles in the region of the ( thomas - fermi ) condensate , i.e. for @xmath315 . the third term , in contrast , is associated with quasiparticles lying outside of the condensate region . these quasiparticles comprise a hartree - fock gas in our simple model calculation . the dominant weight in ( [ lda energy - momentum space spectral density function ] ) comes from small @xmath316 and @xmath32 . the quasiparticles with @xmath0 and @xmath1 weights in the region where there is a condensate are the dominant contribution . for @xmath317 , one also finds a contribution from the normal hartree - fock gas of atoms existing outside of the condensate . as the temperature increases toward @xmath318 , this contribution from the normal hartree - fock gas becomes more significant relative to the bogoliubov quasiparticles . of course , for @xmath319 , it is the sole contribution . we have evaluated the integral appearing in ( [ lda noncondensate outcoupling current in harmonic oscillator units ] ) numerically . some results , as a function of the detuning parameter @xmath290 for various temperatures , are shown in figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5 nm ] and [ fig : total out - coupling current for q=3d3d2/aho a=3d3d5 nm t=3d3d0.2tbec ] . note the logarithmic scale for the out - coupled atom current . as a function of detuning , the out - coupling current separates into two regions . from ( [ lda noncondensate outcoupling current in harmonic oscillator units ] ) and ( [ lda energy - momentum space spectral density function ] ) , we observe that for @xmath320 ( or @xmath321 ) , we have @xmath322 for all values of the wavevector @xmath259 . in this region , we find a contribution from the @xmath0 bogoliubov modes , in addition to a smaller contribution from the atoms coming from outside the condensate region , at finite temperatures . this contribution to the out - coupling current @xmath323 represents the `` quantum evaporation '' process discussed at the end of section [ sec : case of out - coupling to a non - bose - condensed gas ] , being associated with positive energy poles of the single - particle noncondensate green s function . it is only present at finite temperatures . in fig . [ fig : total out - coupling current for q=3d3d2/aho a=3d3d5 nm t=3d3d0.2tbec ] , we have plotted the _ total _ out - coupling current as a function of the detuning @xmath290 at an intermediate temperature of @xmath324 . however , for @xmath325 ( or @xmath326 ) , @xmath327 can be less than zero for a restricted set of wavevectors @xmath259 . in this case , there are two contributions to the out - coupling current @xmath323 . one arises from the @xmath1 bogoliubov modes and hence represents the so - called pair - breaking process associated with the negative energy poles of the single - particle noncondensate spectral density function ( i.e. , a tunneling atom creates a bogoliubov excitation in the trapped gas ) . this is the only process which contributes to the noncondensate out - coupling current at zero temperature . at finite temperature , a second process arises which is associated with quantum evaporation from high momentum states . this involves the positive energy poles of the single - particle noncondensate spectral density function . this second contribution decays more quickly as the detuning is increased , compared with the @xmath1 contribution . we note that these two regions will be doppler shifted toward positive @xmath290 by the nonzero momentum kick @xmath33 from the laser fields . the overall behavior of our lda results in figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5nm]-[fig : total out - coupling current for q=3d3d10/aho a=3d3d5 nm t=3d3d0.2tbec ] is similar to that reported in refs . @xcite , which were based on solving a one - dimensional gross - pitaevskii equation for an isotropic trap . we have also evaluated the noncondensate contribution to the out - coupling current for @xmath328 , with the results shown in figs . [ fig : condensate - noncondensate out - coupling for q=3d3d10/aho and a=3d3d5 nm ] and [ fig : total out - coupling current for q=3d3d10/aho a=3d3d5 nm t=3d3d0.2tbec ] . we observe that the point at which there is no longer any contribution from the @xmath1 bogoliubov modes shifts increasingly toward positive @xmath290 as @xmath188 is increased . however , unlike for the case of @xmath329 , the contribution at finite temperatures from the @xmath0 bogoliubov modes dominates that from the @xmath1 bogoliubov modes for the range of @xmath290 considered . this results from the fact that the higher momentum quasiparticles have greater spectral weight . we have plotted the _ total _ out - coupling current at a temperature of @xmath324 in fig . [ fig : total out - coupling current for q=3d3d10/aho a=3d3d5 nm t=3d3d0.2tbec ] . because of the strength of the condensate contribution , some comment is required about the observability of the noncondensate contribution in the total out - coupling current . since the contribution from the condensate is smallest at high @xmath188 , we compare the condensate and noncondensate contributions at a momentum kick of @xmath328 ( see fig . [ fig : condensate - noncondensate out - coupling for q=3d3d10/aho and a=3d3d5 nm ] ) . since the out - coupling current as a function of @xmath290 is much broader for the noncondensate contribution than the condensate contribution , there are regions in which the strong contribution from the condensate will not dwarf that from the noncondensate . at @xmath260 , the maximum contribution from the quantum evaporation process is about @xmath330 of the maximum contribution from the condensate ( see fig . [ fig : condensate - noncondensate out - coupling for q=3d3d10/aho and a=3d3d5 nm ] ) . the small but interesting @xmath1 bogoliubov contribution is perhaps more visible as a `` distinct contribution '' in the limit of very low temperatures ( see the @xmath158 results in figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5 nm ] and [ fig : condensate - noncondensate out - coupling for q=3d3d10/aho and a=3d3d5 nm ] ) . the unique features associated with the out - coupled atoms coming from the noncondensate can also be enhanced by working with larger values of the s - wave scattering length @xmath12 . in particular , one can `` tune '' the scattering length @xmath12 significantly by the use of feshbach resonances @xcite . the amplification of the out - coupling current of atoms associated with the negative energy poles of the single - particle noncondensate green s function could be looked for in trapped gases near a feshbach resonance . out - coupling experiments may thus provide important information about the bose - condensed gas excitation spectrum if we can increase the s - wave scattering length substantially . in this case , of course , we expect observable changes from the simple bogoliubov - popov approximation we introduced in section [ section : coupling between two trapped gases ] for the single - particle spectral density . such many - body effects can thus be probed `` directly '' using the out - coupling kind of experiment we have been discussing . this paper has studied the weak coupling of two trapped dilute bose gases using a tunneling hamiltonian approach . when considering tunneling between two trapped condensates , we obtained a coherent josephson - type time - dependent contribution given in ( [ equilibrium current 3 ] ) , with a frequency @xmath331 ( see also ref . the expression in ( [ equilibrium current 3 ] ) , which corresponds to coherent transfer of atoms between two condensates , is first order in the coupling @xmath34 . in addition , we found that two types of quasiparticle currents emerge from treating a perturbation of the kind shown in ( [ v ] ) to second order : a static , time - independent current given by ( [ deltaina ] ) and ( [ short time normal current ] ) involving only regular correlation functions of the form @xmath332 , and a time - dependent current given by ( [ deltaiaa ] ) and ( [ deltaiab ] ) involving a second harmonic with frequency @xmath333 . the latter is associated with anomalous correlation functions of the kind @xmath334 , and is reminiscent of the josephson current carried by cooper pairs in superconductors @xcite . these tunneling currents were written in terms of single - particle spectral density functions that contain all the details of the microscopic correlations of the bose - condensed gases . in this manner , the tunneling dynamics is cleanly _ separated _ from the specific approximation used for the many - body description of the bose - condensed gas . this separation was one of the main goals of our work . our paper further serves to unify and extend the work of ref . @xcite on josephson - type and other kinds of tunneling phenomena in trapped bose gases , as well as recent work @xcite dealing with out - coupling from a single bose - condensed gas . it provides a transparent way of expressing the various contributions to the tunneling processes involving an inhomogeneous trapped bose gas , before any specific approximation is introduced . the present paper thus provides a foundation upon which one can study the details of different tunneling processes within various many - body approximations for the dynamics of a bose gas . our work re - emphasizes the potential usefulness of the kind of out - coupling experiments discussed in refs.@xmath3@xcite as a probe of the many - body physics of a bose - condensed gas , including systems of reduced dimensionality . as we have noted in the text , our present study contrasts with the work reported in refs . the latter deal with the density fluctuation spectrum while our present paper ( see also ref . @xcite ) probes the single - particle green s function of a trapped bose gas , which is perhaps a more basic measure of the condensate - induced correlations between atoms . for illustration , we have used our formalism in sections [ section : coupling between two trapped gases ] and [ sec : case of out - coupling to a non - bose - condensed gas ] to find a general expression for the out - coupling atom current using the bogoliubov - popov quasiparticle approximation . our results exhibit the same three processes obtained by the oxford group @xcite . the first process is the tunneling of an atom out of the condensate of the trapped gas . the second process is the `` quantum evaporation '' of an atom from the trapped gas , _ destroying _ a quasiparticle excitation in the trapped gas and _ creating _ an atom in the out - coupled gas . the third process , referred to as a `` pair - breaking '' process in ref . @xcite , involves an atom tunneling out of the trapped gas concurrently with the _ creation _ of a quasiparticle excitation in the trapped gas ( see also ref . @xcite ) . in section [ sec : lda ] , we have numerically evaluated the relative contributions of these three processes as a function of the laser detuning parameter @xmath335 , for various values of the wavevector difference between the two laser beams @xmath336 and the temperature @xmath273 . this kind of experiment would be especially promising as a probe of the correlations in a bose gas in the case of a large scattering length @xmath12 made possible by working near a feshbach resonance . this enhances the tunneling contribution from the noncondensate atoms and moreover requires the use of an improved single - particle correlation function which includes quantum depletion of the condensate ( such as the second order beliaev approximation ) . the tunneling currents from the noncondensate atoms are admittedly small and will require high precision experiments . we argue that the unique information that can be obtained justifies the effort . in future work , we plan on using the formalism set up in this paper to discuss the case when the trapped bose condensate is in a vortex state . in addition , information about the single - particle correlation functions in a one dimensional bose gas @xcite is already present in the out - coupled tunneling current . we would like to thank j.e . williams for discussions of josephson tunneling in trapped bose gases , as well as m. imamovic - tomasovic and t. nikuni for discussions concerning the lda . the research was supported by nserc of canada . out - coupling atom current for @xmath337 and @xmath338 . the detuned frequency @xmath339 is defined in ( [ definition of detuning parameter ] ) . the plot on the left is at @xmath340 . the plot on the right is at @xmath341 . in figs . [ fig : condensate - noncondensate out - coupling current for q=3d3d2/aho and a=3d3d5nm]-[fig : total out - coupling current for q=3d3d10/aho a=3d3d5 nm t=3d3d0.2tbec ] , we take @xmath342 atoms , @xmath343 and @xmath344 . ] out - coupling atom current for @xmath337 and @xmath338 . the plot on the left shows each component contributing to the out - coupling atom current . the plot on the right shows the total out - coupling atom current . in both plots @xmath345 . ] out - coupling atom current for @xmath337 and @xmath346 . the plot on the left shows each component contributing to the out - coupling atom current . the plot on the right shows the total out - coupling atom current . in both plots @xmath345 . ]

spectrum scarcity and ever - increasing data rate requirements are the two main motivating factors behind introducing carrier aggregation ( ca ) in modern wireless communication systems , such as 3gpp lte - advanced @xcite . in ca transmission , multiple component carriers ( ccs ) at different rf frequencies are adopted simultaneously , either within the same rf band ( intraband ca ) , or at different rf bands ( interband ca ) @xcite . while interband ca naturally leads to a _ non - contiguous transmit spectrum _ , this can also happen in the intraband ca case when the adopted component carriers are located in a non - contiguous manner . adopting non - contiguous ca provides lots of flexibility in rf spectrum use , but also poses substantial challenges in the transmitter design , especially in lower - cost devices such as lte - advanced user equipment ( ue ) . moreover , multicarrier type modulations with very large peak - to - average power ratio ( papr ) are used in most state - of - the - art wireless communication systems . when combined with noncontiguous transmission schemes and high power - efficiency requirements , controlling the transmitter unwanted emissions due to power amplifier ( pa ) nonlinearity becomes a true challenge @xcite . in @xcite , a comparison of different transmitter architectures for ca based transmission schemes was presented , showing that it is more power efficient to combine the ccs before the pa , and thus use only a single pa , instead of having a separate pa for each cc . this is also technically feasible , especially in the intraband ca case . the power savings can be quite significant since the overall transmitter efficiency is dominated by the pa @xcite , and is particularly critical in mobile terminals and other lower - cost devices , such as small - cell base - stations , with limited computing and cooling capabilities @xcite . however , the power efficiency gained from using a single pa for all ccs comes at the expense of more severe unwanted emissions , stemming from the pa nonlinear characteristics , that can occur at adjacent channels or bands , or even at own receiver band in fdd transceiver case @xcite . the levels of different unwanted emissions are , in general , regulated by the standardization bodies and itu - r @xcite . it has been recently demonstrated that in the context of lte - advanced mobile transmitter with noncontiguous ca or multicluster type transmission , the pa nonlinearities lead to spurious emissions that can seriously violate the given spectrum and spurious emission limits if not properly controlled @xcite . furthermore , in fdd devices , in addition to violating the general spurious emission limits , the generated spurious components can also overlap with the device s own receive band , causing own receiver desensitization @xcite . one obvious solution to decrease the levels of unwanted emissions is to back off the transmit power from its saturation region . in 3gpp ue terminology , this is known as maximum power reduction ( mpr ) , and mpr values up to 16 db are allowed in some use cases for mobile terminals @xcite . however , this approach will end up yielding a significantly lower pa efficiency as well as a substantial reduction in the network coverage . there is thus a clear need for alternative pa linearization solutions which do not have such drastic adverse effects . digital predistortion ( dpd ) is , in general , one of the most effective solutions for mitigating transmitter nonlinear distortion . there have been substantial research efforts in the past 20 years in developing ever more efficient and elaborate dpd techniques , mostly for high - end macro base - station type of devices . the conventional dpd approaches seek to linearize the full composite transmit signal , and we thus refer to such solutions as full - band dpd in this article . there have also been a handful of recent works on efficient concurrent linearization techniques for multi - band transmitters that employ only a single pa @xcite . these works assume that the component carriers are separated by a large distance such that the spurious emissions are filtered out by the transmit rf filter , and hence linearization of only the main carriers is pursued . first , the emission limits in the spurious region are generally stricter than in the spectral regrowth region around the component carriers , and are thus more easily violated . this has been recognized also in 3gpp recently , in the context of intraband noncontiguous carrier aggregation @xcite . moreover , even in interband carrier aggregation scenarios in fdd devices , some of the spurious components can be hitting the own rx band causing own receiver desensitization @xcite . second , by concentrating the linearization efforts to the most critical spurious emissions only , the processing and instrumentation requirements can be significantly relaxed , thus potentially facilitating the dpd processing also in mobile terminals and other lower - cost devices . this applies not only to the dpd main path processing complexity but also to the feedback receiver instrumentation complexity which can also be substantially reduced . there have been some recent studies in the literature that consider the mitigation of the spurious emissions explicitly . in @xcite , such processing was added to complement a concurrent linearization system while assuming a frequency - flat pa response within the spurious sub - band . in @xcite , the dpd parameter estimation was done offline , based on extracting the quasi - memoryless pa parameters using large - signal network analyzer ( lsna ) measurements , and covering up to fifth - order processing . in @xcite , a memoryless least - squares fit between the observed intermodulation distortion ( imd ) at the considered sub - band and certain basis functions was performed . the basis functions were generated using a wideband composite carrier baseband equivalent signal , followed by sub - band filtering , implying a very high sample rate and processing complexity , especially for widely spaced carriers . the estimated and regenerated imd was then applied at the pa input , oppositely phased , such that it is canceled at the pa output . in this article , we extend the elementary third - order im3 sub - band dpd solution , proposed by the authors in @xcite , in two ways . first , the third - order im3 sub - band dpd is extended to incorporate higher - order processing based on explicit modeling of the higher - order spurious components at the im3 sub - band . this will enhance the im3 spurious emission suppression considerably . furthermore , we also extend the sub - band dpd solution to include higher - order sub - bands , i.e. , im5 , im7 , etc . , thus offering more flexibility and linearization capabilities beyond the basic im3 sub - band . all the proposed solutions are derived for wideband nonlinear pas with memory . furthermore , we also formulate novel decorrelation - based parameter estimation methods , covering both sample - adaptive and block - adaptive learning rules , to efficiently identify the needed dpd parameters with low complexity . the proposed learning solutions are also shown to offer better performance than the earlier proposed third - order or fifth - order inverse based methods . we also provide comprehensive simulation and rf measurement results , using a commercial lte - advanced mobile pa , to evaluate and validate the effectiveness of the proposed solutions in real world scenarios . the rest of this article is organized as follows : section [ sec : dpd_modeling ] presents the mathematical modeling of the different considered spurious components at different sub - bands , produced by a nonlinear pa with memory . stemming from this modeling , also the corresponding core processing principles of the proposed dpd solutions at different sub - bands are formulated . then , section [ sec : dpd_learning ] presents the proposed dpd parameter learning solutions , covering both sample - adaptive and block - adaptive decorrelation solutions at different sub - bands . section [ sec : complexity ] addresses then different implementation alternatives of the proposed sub - band dpd concept , and also analyzes the computing complexity of the sub - band dpd in terms of the number of floating point operations together with some hardware complexity aspects . finally , sections [ sec : simulations and measurements ] and [ sec : rf measurement results ] report comprehensive simulation and rf measurement results , evidencing excellent spurious emission suppression in different realistic scenarios . in this manuscript , we assume a practical case of non - contiguous carrier aggregation with two component carriers . when such noncontiguous dual - carrier signal is applied at the pa input , the pa nonlinearity leads to intermodulation distortion at different sub - bands , as shown in fig . [ fig : psd ] . assuming a cc separation of @xmath0 , in addition to the spectral regrowth around the main carriers , intermodulation between the two ccs yields strong imd at integer multiples of @xmath0 from the main ccs . in this article , we refer to the intermodulation ( i m ) sub - bands located at @xmath1 from the main ccs as the _ im3 sub - bands_. similarly , the _ im5 sub - bands _ are located at @xmath2 from the main ccs , and so on . in general , each i m sub - band includes nonlinear distortion components of different orders as shown in fig . [ fig : psd ] . for example , in the case of a seventh - order pa nonlinearity , the im3 sub - band contains third , fifth and seventh - order nonlinearities , while the im5 sub - band contains fifth and seventh - order components . in this section , we start with the fundamental modeling of the nonlinear distortion at the im3 sub - band as a concrete example . stemming from that modeling , we then formulate the basic processing of the proposed im3 sub - band dpd . after that , the modeling is extended to cover the nonlinear distortion components at the higher - order i m sub - bands , followed by the corresponding higher - order sub - band dpd processing . the actual parameter estimation and learning algorithms for the proposed sub - band dpd structures are then presented in details in section [ sec : dpd_learning ] . in all the modeling and developments , we adopt the widely - used wideband parallel hammerstein ( ph ) pa model @xcite to describe the fundamental nonlinear behavior of the pa , as it has been shown to model accurately the measured nonlinear behavior of different classes of true pas . the modeling is carried out at composite baseband equivalent level , where the two component carriers are assumed to be separated by @xmath3 . the composite baseband equivalent input and output signals of the @xmath4 order parallel hammerstein pa model , with monomial nonlinearities and fir branch filters , respectively , read @xmath5 where @xmath6 and @xmath7 are the baseband component carrier signals , @xmath8 denotes the ph branch filter impulse responses of order @xmath9 , and @xmath10 is the convolution operator . intermodulation between the two component carriers leads to the appearance of imd components at @xmath11 , @xmath12 , etc . , as shown in fig . [ fig : psd ] for the corresponding rf spectrum . @xmath13 here , @xmath14 denote the baseband equivalent impulse responses corresponding to the wideband ph pa model filters @xmath8 , evaluated at the im3@xmath15 sub - bands , formally defined as @xmath16 with @xmath17 denoting an ideal low pass filtering operation with a passband width @xmath18 times the bandwidth of the wider cc . furthermore , @xmath19 and @xmath20 in ( 3 ) are the corresponding @xmath21 order static nonlinear ( snl ) basis functions , related to the nonlinear distortion at the im3 + and im3- sub - bands , respectively . assuming then an @xmath22 order pa model ( i.e. , @xmath23 ) , as a concrete high - order example , the im3 + sub - band basis functions read @xmath24 the corresponding basis functions for the im3- sub - band , i.e. , @xmath20 , can be obtained by simply interchanging @xmath6 and @xmath7 in the above expressions . next , these behavioral modeling results are utilized to formulate the proposed im3 sub - band dpd concept , specifically tailored to suppress the distortion at the im3 sub - band . for presentation purposes , we focus here on suppressing the spurious emissions at the im3 + sub - band , while the corresponding emissions at the im3- sub - band can also be easily mitigated using a very similar structure with minor changes as will be elaborated later in section [ sec : higher - order dpd principle ] . the key idea of the sub - band dpd concept is to inject a proper additional low - power cancellation signal , with structural similarity to ( [ eq : im3_output_mem ] ) and located at @xmath25 , into the pa input , such that the level of the im3 + term at the pa output is reduced . now , stemming from the imd structure in ( [ eq : im3_output_mem ] ) , an appropriate digital injection signal can be obtained by adopting the im3 + basis functions @xmath19 in ( [ eq : im3_basisfunctions1])-([eq : im3_basisfunctions5 ] ) , combined with proper filtering using a bank of sub - band dpd filters @xmath26 . incorporating such sub - band dpd processing with polynomial order @xmath27 , the composite baseband equivalent pa input signal reads @xmath28e^{j 2\pi \frac{3 f_{if}}{f_s } n } \label{eq : pa_in_with_im3_dpd}\end{aligned}\ ] ] here , and in the continuation , we use @xmath29 variables to indicate sub - band dpd processing and the corresponding predistorted signals . this dpd processing principle is illustrated in fig . [ fig : sub - band dpd ] at a conceptual level , where the tx / rx duplexer filters are omitted for simplicity of the presentation since they do not directly impact the sub - band dpd processing or learning . similar convention is followed also in other figures . different implementation alternatives as well as the parameter learning and feedback receiver aspects are addressed in more details in sections iii and iv . the direct substitution of ( [ eq : pa_in ] ) in ( [ eq : pa_out_mem ] ) leads to the appearance of spurious intermodulation terms also at higher - order i m sub - bands , in addition to the previously considered im3 sub - bands , as illustrated already in fig . [ fig : psd ] . these higher - order sub - band emissions can also be harmful , since they can violate the emission limits , cause own receiver desensitization , or both . thus , developing a sub - band dpd solution that can tackle also the distortion at these higher - order i m sub - bands with feasible complexity is important . similar to the above im3 + sub - band developments , we next extract the imd terms at the higher - order i m sub - bands . the baseband equivalent imd terms at the im5 + , im7 + , im9 + , and im11 + sub - bands , as concrete examples , can be extracted using ( [ eq : pa_in ] ) and ( [ eq : pa_out_mem ] ) , interpreted at proper sub - bands , yielding @xmath30 in the above models , @xmath31 , @xmath32 , @xmath33 , and @xmath34 denote the baseband equivalent impulse responses corresponding to the wideband pa model filters @xmath8 in ( 2 ) , evaluated at the im5 + , im7 + , im9 + , and im11 + sub - bands , respectively . these are formally obtained similar to ( 4 ) but replacing @xmath35 with either @xmath36 , @xmath37 , @xmath38 , or @xmath39 , respectively . the @xmath21 order basis functions at the these higher - order i m sub - bands , denoted by @xmath40 , @xmath41 , @xmath42 , and @xmath43 , and assuming again an @xmath22 order ph pa model , read @xmath44 the corresponding baseband equivalent imd terms at the negative i m sub - bands can be obtained by simply interchanging @xmath6 and @xmath7 in ( [ eq : imd_basisfunctions1])-([eq : imd_basisfunctions2 ] ) . then , stemming from the distortion structure in ( [ eq : im5_output])-([eq : im11_output ] ) , and adopting similar ideology as in the previous im3 sub - band dpd case , a natural injection signal for suppressing the spur components at higher - order i m sub - bands can be obtained by properly filtering and combining the above higher - order sub - band basis functions . the sub - band specific filters for the @xmath21 order basis functions are denoted with @xmath45 , @xmath46 , @xmath47 , and @xmath48 . incorporating such dpd processing with dpd polynomial order @xmath27 , and aggregating at the same time parallel sub - band dpds simultaneously for the im3 + , im5 + , and im7 + sub - bands , as a concrete example , the composite baseband equivalent pa input signal now reads @xmath49e^{j 2\pi \frac{3 f_{if}}{f_s } n } \nonumber\\ & + \left[\sum_{\substack{p=5 \\ p \text { odd}}}^{q } \alpha_{5+,p , n } \star u_{5+,p}(n)\right]e^{j 2\pi \frac{5 f_{if}}{f_s } n } \nonumber\\ & + \left[\sum_{\substack{p=7 \\ p \text { odd}}}^{q } \alpha_{7+,p , n } \star u_{7+,p}(n)\right]e^{j 2\pi \frac{7 f_{if}}{f_s } n } \label{eq : pa_in_dpd_imds}\end{aligned}\ ] ] in general , the achievable suppression of the distortion at different i m sub - bands depends directly on the selection and optimization of the different filters @xmath26 , @xmath45 , @xmath46 , and so on . this is addressed in detail in the next section . in this section , based on the previous spurious component modeling and the proposed sub - band dpd principle , we formulate computing feasible and highly efficient estimation algorithms for learning and optimizing the sub - band dpd filter coefficients such that the spurious emissions are minimized at the considered i m sub - bands . we start by deriving some analytical reference solutions , taking the third - order im3 sub - band dpd as a simple and tractable example , in order to demonstrate that minimizing the im3 power is essentially identical to decorrelating the im3 observation against the corresponding sub - band dpd basis functions . then , both sample - adaptive and block - adaptive decorrelation based learning rules are devised , covering the general cases of higher - order processing with memory at im3 and higher - order i m sub - bands . in this subsection , we derive three analytical reference solutions for calculating the im3 sub - band dpd coefficients . the considered approaches are the third - order inverse solution , also adopted in @xcite , the im3 power minimization solution , and the analytical im3 decorrelation - based solution . to keep the analytical developments simple and tractable , we consider a simplified case with a third - order memoryless im3 + sub - band dpd and a third - order memoryless pa . the actual sample - adaptive and block - adaptive decorrelation based learning solutions , devised later in this section , are formulated for the general cases of higher - order processing with memory . now , starting with the dual carrier signal in ( [ eq : pa_in ] ) , and limiting the study to the simplified case of a third - order memoryless im3 + sub - band dpd and a third - order memoryless pa , the basic signal models are given by ( [ eq : thirdorderanalysis_1])-([eq : thirdorderanalysis ] ) . @xmath50 here , @xmath51 and @xmath52 are the memoryless pa model parameters , @xmath53 denotes the memoryless sub - band dpd parameter to be optimized , while @xmath54 and @xmath55 refer to the baseband equivalent pa output at the positive im3 sub - band without and with sub - band dpd , respectively . now , as ( [ eq : thirdorderanalysis ] ) clearly shows , the im3 sub - band distortion at the pa output , with dpd adopted , depends directly on , and can thus be controlled by , the dpd coefficient @xmath53 . in the well - known _ third - order inverse solution _ , the dpd parameter @xmath53 is selected such that the third - order term in ( [ eq : thirdorderanalysis ] ) is canceled , i.e. , @xmath56 . the corresponding solution , denoted by @xmath57 , thus reads @xmath58 however , this will not remove all the distortion , because higher - order terms will be created due to the predistortion , as shown in ( [ eq : thirdorderanalysis ] ) . a more elaborate method for selecting the dpd parameter @xmath53 is the one that minimizes the power of the total im3 sub - band signal , @xmath55 , referred to as the _ minimum im3 power _ or _ minimum mean - squared error ( mmse ) _ solution in the continuation . for notational convenience , we define the so - called error signal as @xmath59 , as with ideal predistortion the im3 sub - band signal would be zero , and thus the optimization means minimizing the power of this error signal . the detailed derivation of the optimum dpd parameter which minimizes the mean squared error @xmath60 $ ] is given in appendix a , yielding @xmath61^*}{\left[|f_1|^2 { { \rm i\kern-.3em e}}_{42 } + 4 \mathbb{r}(f_1 f_3^*)({{\rm i\kern-.3em e}}_{62 } + { { \rm i\kern-.3em e}}_{44 } ) + 4 |f_3|^2 ( { { \rm i\kern-.3em e}}_{46 } + 2 { { \rm i\kern-.3em e}}_{64 } + { { \rm i\kern-.3em e}}_{82})\right]^ * } , \label{eq : mmse}\end{aligned}\ ] ] where @xmath62 refers to the products of the cc signals higher - order moments of the form @xmath63 { { \rm i\kern-.3em e}}[|x_2|^j]$ ] , while @xmath64 $ ] denotes the statistical expectation operator . as will be shown with concrete examples in section v , this mmse solution provides better linearization performance compared to the third - order inverse solution . however , as shown in ( [ eq : mmse ] ) , the analytical mmse solution requires the knowledge of various higher - order moments of the cc signals . furthermore , the above solution is valid only in the case of a memoryless third - order nonlinear system , beyond which obtaining an analytical expression for the dpd coefficients becomes overly tedious . . the analytical reference solution for the decorrelation - based sub - band dpd structure is obtained by setting the correlation between the error signal , @xmath65 , and the third - order basis function , @xmath66 , to zero , i.e. , @xmath67 = 0 $ ] . then , through straight - forward algebraic manipulations , the decorrelation - based dpd coefficient , denoted by @xmath68 , can be shown to read @xmath69 assuming then , for simplicity , that the cc baseband equivalents are complex gaussians , ( [ eq : alpha_0 ] ) simplifies to @xmath70 where @xmath71 $ ] and @xmath72 ^ 2 $ ] are the variances of the two ccs @xcite . a clear advantage of the decorrelation - based approach compared to the earlier mmse solution lies in the simple and straightforward adaptive filtering based practical computing solutions , sketched initially in @xcite for simple third - order processing at im3 , which do not require any prior knowledge about signal moments or any other parameters . furthermore , opposed to the mmse solution , the decorrelation based approach can be easily extended to include higher - order nonlinearities and memory effects , at both the im3 sub - band as well as the other higher - order i m sub - bands , as will be described in details in the next subsection . in this subsection , we provide the actual sample - adaptive learning rules for both im3 and higher - order i m sub - band dpd structures with memory , whose basic operating principles were described in section ii . to facilitate the learning , we assume a feedback or observation receiver measuring the particular i m sub - band whose sub - band dpd filter coefficients are currently under learning . notice that opposed to classical wideband dpd principles , where a wideband observation receiver is typically needed , a more narrowband receiver is sufficient here since only the particular i m sub - band is observed . as both the im3 and the higher - order i m sub - band dpds are based on multiple strongly correlated basis functions , given in ( [ eq : im3_basisfunctions1])-([eq : im3_basisfunctions5 ] ) and ( [ eq : imd_basisfunctions1])-([eq : imd_basisfunctions2 ] ) , respectively , we start by introducing a basis function orthogonalization procedure , after which the actual proposed adaptive decorrelation algorithms are described . when a nonlinearity order higher than the i m sub - band order is considered , multiple basis functions are adopted in the sub - band dpd processing , as described in section ii . taking as an example case the im3 + sub - band , the snl basis functions @xmath19 , @xmath73 , given in ( [ eq : im3_basisfunctions1])-([eq : im3_basisfunctions5 ] ) , are highly correlated . this can negatively impact the convergence and stability of the adaptive decorrelation - based learning . therefore , the snl basis functions are first orthogonalized @xcite , which yields a new set of dpd basis functions @xmath74 , @xmath73 , written formally at sample level as @xmath75 where @xmath76^t , \\ \textbf{s}_{3+}(n ) & = [ s_{3+,3}(n)\:\:s_{3+,5}(n ) \:\ : ... \:\ : s_{3+,q}(n)]^t , \\ \textbf{w } & = \begin{bmatrix } 1 & 0 & 0 & 0 & 0 \\ w_{5,3 } & w_{5,5 } & 0 & 0 & 0 \\ w_{7,3 } & w_{7,5 } & w_{7,7 } & \dots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ w_{q,3 } & w_{q,5 } & w_{q,7 } & \dots & w_{q , q } \end{bmatrix}.\end{aligned}\ ] ] the lower triangular matrix @xmath77 can be obtained , e.g. , through gram - schmidt orthogonalization , qr or singular value decomposition , or using a lower complexity iterative orthogonalization algorithm @xcite . similarly , the snl basis functions in ( [ eq : imd_basisfunctions1])-([eq : imd_basisfunctions2 ] ) corresponding to the higher - order i m sub - bands are also orthogonalized , to obtain new sets of orthogonal basis functions @xmath78 , @xmath79 , etc . we now present the actual sample - adaptive decorrelation based learning algorithm for the im3 + sub - band dpd coefficients . for notational convenience , we introduce the following vectors @xmath80^t , \\ \bar{\boldsymbol{\alpha}}_{3+}(n ) & = [ \boldsymbol{\alpha}_{3+,0}(n)^t \:\ : \boldsymbol{\alpha}_{3+,1}(n)^t \:\ : ... \:\ : \boldsymbol{\alpha}_{3+,n}(n)^t]^t , \\ \bar{\textbf{s}}_{3+}(n ) & = [ \textbf{s}_{3+}(n)^t \:\ : \textbf{s}_{3+}(n-1)^t \:\ : ... \:\ : \textbf{s}_{3+}(n - n)^t]^t,\end{aligned}\ ] ] where now @xmath81 denotes the @xmath82 adaptive filter coefficient of the @xmath21 order orthogonalized im3 + basis function @xmath74 at time index @xmath83 , and @xmath84 denotes the adaptive filter memory depth . furthermore , the vectors @xmath85 and @xmath86 incorporate all the coefficients and basis function samples up to polynomial order @xmath27 . adopting this notation , the instantaneous sample of the composite baseband equivalent pa input signal @xmath87 now reads @xmath88 where the instantaneous sample of the baseband equivalent im3 + injection signal @xmath89 reads @xmath90 then , in order to adaptively update the filter coefficients @xmath85 , the im3 + sub - band is observed with the feedback receiver , and the coefficient is updated as @xmath91 where @xmath92 denotes the baseband equivalent observation of the pa output at the im3 + sub - band with the current dpd coefficients . the scaling factor @xmath93 normalizing the learning step - size @xmath94 in ( [ eq : dpd_filterupdate_mem ] ) is philosophically similar to that of the normalized least mean square ( nlms ) algorithm , in effect making the learning characteristics more robust against the input data dynamics . the proposed coefficient update in ( [ eq : dpd_filterupdate_mem ] ) is seeking to decorrelate the im3 + observation and the adopted orthogonalized basis functions . this type of learning algorithm can also be interpreted as a stochastic newton root search in the objective function @xmath95 $ ] , where the target is to search for the dpd coefficients @xmath96 that minimize the ensemble correlation between @xmath97 and the baseband equivalent of the im3 + injection signal @xmath89 in ( [ eq : injected_im3 ] ) . the overall processing flow is graphically illustrated in fig . [ fig : decorr_dpd ] . we now extend the decorrelation based learning to higher - order i m sub - bands . first , adopting vector - based notations again , we introduce the following vectors for notational convenience @xmath98^t , \\ \boldsymbol{\alpha}_{7+,l}(n ) & = [ \alpha_{7+,7,l}(n)\:\:\alpha_{7+,9,l}(n)\:\ : ... \:\ : \alpha_{7+,q , l}(n)]^t , \\ \bar{\boldsymbol{\alpha}}_{5+}(n ) & = [ \boldsymbol{\alpha}_{5+,0}(n)^t \:\ : \boldsymbol{\alpha}_{5+,1}(n)^t \:\ : ... \:\ : \boldsymbol{\alpha}_{5+,n}(n)^t]^t , \\ \bar{\boldsymbol{\alpha}}_{7+}(n ) & = [ \boldsymbol{\alpha}_{7+,0}(n)^t \:\ : \boldsymbol{\alpha}_{7+,1}(n)^t \:\ : ... \:\ : \boldsymbol{\alpha}_{7+,n}(n)^t]^t , \\ \textbf{s}_{5+}(n ) & = [ s_{5+,5}(n)\:\:s_{5+,7}(n ) \:\ : ... \:\ : s_{5+,q}(n)]^t , \\ \textbf{s}_{7+}(n ) & = [ s_{7+,7}(n)\:\:s_{7+,9}(n ) \:\ : ... \:\ : s_{7+,q}(n)]^t , \\ \bar{\textbf{s}}_{5+}(n ) & = [ \textbf{s}_{5+}(n)^t \:\ : \textbf{s}_{5+}(n-1)^t \:\ : ... \:\ : \textbf{s}_{5+}(n - n)^t]^t , \\ \bar{\textbf{s}}_{7+}(n ) & = [ \textbf{s}_{7+}(n)^t \:\ : \textbf{s}_{7+}(n-1)^t \:\ : ... \:\ : \textbf{s}_{7+}(n - n)^t]^t.\end{aligned}\ ] ] then , similar to @xmath89 in ( [ eq : injected_im3 ] ) , the baseband equivalents of the im5 + and im7 + sub - band dpd injection signals , denoted by @xmath99 and @xmath100 , respectively , read @xmath101 thus , the composite baseband equivalent pa input signal when the im3 + , im5 + , and im7 + sub - band dpds are all included , reads @xmath102 then , similar to the im3 + coefficient update in ( [ eq : dpd_filterupdate_mem ] ) , the higher - order im5 + and im7 + coefficient updates read @xmath103 where @xmath104 and @xmath105 denote the baseband equivalents of the pa output at the im5 + and im7 + sub - bands , with the corresponding sub - band dpds included adopting the current coefficients , @xmath106 and @xmath107 , respectively . from the learning perspective , observing a single i m sub - band at a time is the most obvious alternative , which means that the learning for multiple sub - band dpds happens one at a time . furthermore , extending the decorrelation - based learning for the negative i m sub - bands is straight - forward . this can be obtained by interchanging @xmath6 and @xmath7 in the snl basis functions expressions , and observing the pa output at the corresponding negative i m sub - bands . as fig . [ fig : decorr_dpd ] illustrates , the previous sample - adaptive decorrelation - based learning concept is in principle a closed - loop feedback system with nonlinear adaptive processing inside the loop . during the dpd learning phase , and under the potential hardware processing and latency constraints , the dpd parameter convergence and consequently the dpd linearization performance can be affected , especially if the learning loop delay becomes large . stemming from this , an alternative and new block - adaptive decorrelation - based learning solution is developed next . the proposed block - adaptive learning rule implies defining two distinct blocks in the processing , as illustrated in fig . [ fig : blockdpd_learning ] . a single update cycle of the learning algorithm will utilize @xmath108 samples whereas the dpd parameter update interval is @xmath109 samples , with @xmath110 . thus , by proper choice of @xmath108 and @xmath109 , arbitrarily long loop delays can in principle be tolerated , facilitating stable operation under various hardware and software processing latency constraints . notice that proper timing synchronization between the observation receiver output and the basis functions is , in general , needed , which can be accomplished prior to executing the actual coefficient learning procedure . are applied in the sub - band dpd main path processing during the next block . ] [ tab : flops ] now , assuming an estimation block size of @xmath108 samples and dpd filter memory depth of @xmath84 per each of the im3 + orthogonalized basis functions , the following vectors and matrices , which stack all the samples and the corresponding dpd filter coefficients within block @xmath111 , can be defined : @xmath112^t , \\ \bar{\textbf{s}}_{3+}(n_m ) & = [ \textbf{s}_{3+}(n_m)^t \:\ : ... \:\ : \textbf{s}_{3+}(n_m - n)^t]^t , \\ \textbf{s}_{3+}(m ) & = [ \bar{\textbf{s}}_{3+}(n_m)\:\ : ... \:\ : \bar{\textbf{s}}_{3+}(n_m+m-1 ) ] , \\ \boldsymbol{\alpha}_{3+,l}(m ) & = [ \alpha_{3+,3,l}(m)\:\:\alpha_{3+,5,l}(m)\:\ : ... \:\ : \alpha_{3+,q , l}(m)]^t , \\ \bar{\boldsymbol{\alpha}}_{3+}(m ) & = [ \boldsymbol{\alpha}_{3+,0}(m)^t \:\ : \boldsymbol{\alpha}_{3+,1}(m)^t \:\ : ... \:\ : \boldsymbol{\alpha}_{3+,n}(m)^t]^t.\end{aligned}\ ] ] here , the index of the first sample of block @xmath111 is denoted by @xmath113 . the block - adaptive im3 + sub - band dpd coefficient update then reads @xmath114^t , \\ \bar{\boldsymbol{\alpha}}_{3+}(m+1 ) & = \bar{\boldsymbol{\alpha}}_{3+}(m ) - \frac{\mu}{||\textbf{s}_{3+}(m)||^2 + c } \ : \textbf{s}_{3+}(m)\textbf{e}_{3+}^*(m ) , \label{eq : blockadaptive}\end{aligned}\ ] ] where @xmath115 refers to the element - wise conjugated error signal vector , while @xmath116 denotes the filter input data matrix , all within the processing block @xmath111 . the obtained new dpd coefficients @xmath117 are then applied on the next block of @xmath109 samples as illustrated in fig . [ fig : blockdpd_learning ] . while the above presentation describes the block - adaptive learning only at the im3 + sub - band , extending the principle to higher - order i m sub - bands is straightforward , and thus not explicitly shown . one of the main advantages of the proposed decorrelation - based sub - band dpd technique is its reduced complexity compared to the classical full - band dpd processing , especially in scenarios where the ccs are widely spaced and thus very high speed adcs and dacs are required in the classical solutions . in this section , we first address some implementation aspects of the proposed sub - band dpd concept . we then provide a thorough comparison of the computing and hardware complexity perspectives between the full - band and the proposed sub - band dpds . finally , some system power efficiency considerations are also presented , with particular focus on low - cost mobile devices . fig . [ fig : implementationalternatives ] shows two alternative architectures of the overall sub - band dpd processing with multiple i m sub - bands . the first architecture shown in fig . [ fig : decorr_dpd_im5_im7_im9 ] adds an upsampling and digital if upconversion block after each sub - band dpd stage in order to digitally place the generated injection signals at proper intermediate frequencies . after adding the sub - band dpd outputs in the digital domain , a single wideband dac per i and q branch is used after which the signal is upconverted to rf and amplified by the pa . the second architecture in fig . [ fig : decorr_dpd_im5_im7_im9_2 ] adds the outputs of the sub - band dpds in the analog domain , implying that each sub - band dpd is followed by a narrow - band dac per i and q branch , together with an analog complex if upconversion . a common i / q rf modulator is then used for all sub - bands prior to the pa module . both of these two architectures have their own advantages and disadvantages . in particular , if the carrier spacing between the component carriers is not very large , architecture i is likely to be more suitable . on the other hand , when the carrier spacing starts to increase , using a single wideband dac may not be efficient from cost and power consumption perspectives , so in this case architecture ii is likely to be more attractive . however , some extra processing may be required in this architecture in order to achieve proper synchronization between the dacs . this forms an interesting topic for future research . a common advantage of both architectures is that each sub - band dpd block can be switched on or off according to the prevailing emission levels and limits at the considered sub - bands . such flexibility is not available in the classical full - band dpd solutions since , by design , a full - band predistorter always tries to linearize the full composite transmit band . with the sub - band dpd concept , the linearization can be flexibly tailored and optimized to those frequencies which are the most critical from the emission limits perspective . in general , the computational complexity of any dpd can be classified into three main parts @xcite : identification complexity , adaptation complexity and running complexity . the identification part is basically the estimation complexity of the dpd parameters , while the adaptation complexity includes the required processing in order to adapt to new operating conditions or device aging . finally , the running complexity , which is the most critical especially for mobile - type devices , involves the number of computations done per second while the dpd is operating . in this subsection , we will focus on the dpd running complexity in details , while the dpd parameter identification and feedback receiver complexity perspectives are shortly discussed in section [ sec : fb_complexity ] . for a quantitative comparison of the running complexities , we shall use the number of floating point operations ( flops ) per sample , the number of dpd coefficients , and the required sample rate in the predistortion path as the main quantitative metrics . in general , the running complexity is divided into two main parts : the first is the basis function generation , and the second is the actual predistortion filtering using the basis functions @xcite . the number of flops required to perform these two operations is shown in table [ tab : flops ] for ninth - order sub - band and full - band dpds , respectively , where @xmath118 and @xmath119 are the corresponding memory depths per adopted basis function . the full - band dpd architecture that we use also in our comparative performance simulations in section v , and which is also widely applied otherwise , is based on the ph architecture with ninth order nonlinearity , while the sub - band dpd is based on architecture ii shown in fig . [ fig : decorr_dpd_im5_im7_im9_2 ] . frequency selectivity of the nonlinear pa is another important factor to be considered when comparing the two dpd architectures . this implies that when memory effects of the pa are considered , substantially longer filters are needed for the full - band dpd compared to the sub - band dpd for a certain performance requirement . in the complexity analysis , in the sub - band dpd case , we thus assume a memory depth of @xmath120 per basis function , while a memory depth of @xmath121 is assumed for the full - band ph dpd per basis function . a substantial reduction in the needed sample rates can be achieved when adopting the sub - band dpd . as a concrete example , if we consider a challenging lte - a ul scenario of two 1 mhz ccs with 20 mhz carrier spacing , the required sample rate by the full - band dpd is @xmath122 , while for the sub - band dpd it becomes @xmath123 . . furthermore , the processing complexity of the sub - band dpd solution is clearly feasible for modern mobile device processing platforms , in terms of the gflops count , while that of the classical fullband dpd is clearly infeasible . however , in some scenarios , when the carrier spacing between the ccs is decreased and/or the cc bandwidth increases , the benefit from using the sub - band dpd approach is reduced since the sample rates of the sub - band and full - band dpd will become more comparable . notice also that , e.g. , in interband ca cases , one can adopt a concurrent 2d - dpd @xcite to linearize the main carriers , which can then be easily complemented with the sub - band dpd processing to protect the own receiver if tx filtering does not offer sufficient isolation . such a scenario could take place , e.g. , in uplink band 1 ( 1920 - 1980 mhz ) + band 3 ( 1710 - 1785 mhz ) interband ca where it is technologically feasible to adopt a single multiband / multimode pa module for amplification . in such scenario , the im3 sub - band is then directly at the own rx frequencies of band 1 ( 2110 - 2170 mhz ) . in addition to the complexity reduction in the dpd main processing path , the complexity of the feedback or observation receiver , used for dpd parameter estimation and adaptation , is also greatly reduced . in order to estimate , e.g. , the parameters of the im3 + sub - band dpd , we only need to observe the im3 + sub - band at the pa output , instead of observing the full - band ( including all the i m sub - bands ) which is the case in the full - band dpd . this reduces the cost , complexity and power consumption in the feedback path allowing use of simpler instrumentation , in particular the adc . moreover , only a single observation receiver is required , even if linearizing multiple sub - bands , since the parameter learning and corresponding observation of the pa output can be done sequentially , in a per sub - band manner . finally , in terms of the parameter estimation algorithmic complexity , the proposed decorrelation based solutions are extremely simple , when compared to any classical full - band dpd related method , such as least - squares based parameter fitting and indirect learning architecture @xcite that are commonly adopted . when a dpd is adopted , a less linear but more efficient pa , that can operate near its saturation region , can generally be used . however , the overall power efficiency of the device is only improved if the extra power consumed by the dpd stage is less than the power savings due to increased pa efficiency . here , we address this aspect from a mobile device perspective . we consider a practical scenario where the transmit power at the output of the mobile pa is + 26 dbm ( i.e. , 400 mw ) , stemming from 3gpp lte - advanced requirements @xcite , and assuming that the tx duplexer filter and connector insertion losses are 3 db . then , good examples of practical pa power efficiency figures when operating in highly nonlinear or linear modes are around 35@xmath124 and 20@xmath124 ( or even less ) , respectively . this means that the power consumed by the highly nonlinear pa is roughly 1150 mwatt , while the corresponding linear pa consumes roughly 2000 mwatt . in other words , adopting a highly power - efficient nonlinear pa saves 850 mwatt of power in this particular example . then , in order to suppress the spurious emissions at the pa output , in the nonlinear pa case , we adopt the sub - band dpd solution . in @xcite , a state of the art 28 nm implementation of the qualcomm hexagon dsp capable of supporting up to 4.8 gflops at 1.2 ghz is reported , which is more than enough for carrying out the needed sub - band dpd processing for the im3 and im5 sub - bands in our example . in the full - band dpd case , on the other hand , which requires 31.9 gflops as shown in table [ tab : flops ] , this is clearly insufficient . the power consumption of the dsp platform in @xcite is shown to be approximately 100 mwatt , when running at 840 mhz ( i.e. , 3.36 gflops assuming 4 flops per cycle ) , which is again sufficient for linearizing the im3 and im5 sub - bands in our example . thus , adopting the nonlinear pa already saves 850 milliwatt , when it comes to the pa interface , as explained above , while when complemented with the sub - band dpd processing for enhanced linearity , only 100 milliwatt of additional power is consumed . thus , the overall power budget and power - efficiency are clearly in favor of using a highly non - linear pa , complemented with the sub - band dpd structure , even in a mobile device with non - contiguous uplink carrier aggregation . furthermore , if the sub - band dpd processing is implemented using a dedicated hardware solution ( e.g. , a digital asic ) , an even more power - efficient dpd stage can most likely be realized . in this section , a quantitative performance analysis of the proposed sub - band dpd solution is presented using matlab simulations with practical models for mobile - like pas designed for low - cost devices . in general , we quantify the suppression of intermodulation power at the im3 and im5 sub - bands through the power ratios relative to the component carrier wanted signal power as shown in fig . [ fig : psd ] , which are defined as @xmath125 the inband transmit waveform purity is measured through the error vector magnitude ( evm ) , which is defined as @xmath126 here , @xmath127 is the power of the error signal , while @xmath128 is the reference power of the ideal symbol constellation . the error signal is defined as the difference between the ideal symbol values and the corresponding synchronized and equalized samples at the pa output , both normalized to have identical linear gains . in section [ sec : dpd_analytic_solutions ] , analytical reference expressions for the third - order inverse , minimum mse , and decorrelation - based sub - band dpd solutions were presented . in this subsection , we shortly evaluate and compare the performance of these analytical solutions , assuming a third - order memoryless pa and known parameters , and focus only on the positive im3 sub - band for simplicity . the memoryless pa model has been identified using a true mobile pa , with transmit power of + 23dbm . the main objective is to compare the performance of the three analytical solutions in terms of linearization performance , and thereon to verify that the decorrelation based solution is essentially identifcal to the minimum mse solution . the signal used in this performance evaluation is composed of two 1mhz lte - a ul sc - fdma ccs with qpsk data modulation , and the cc spacing is 10mhz . the obtained results shown in table [ tab : im3r_evm_analytical ] demonstrate that the decorrelation - based solution is giving almost the same performance as the minimum mse solution , and also that both of them are substantially better than the classical third - order inverse solution @xcite in terms of the imd suppression at the considered sub - band . it can also be seen from table [ tab : im3r_evm_analytical ] that the evm is essentially not affected by the sub - band dpd processing . [ tab : im3r_evm_analytical ] next we evaluate and compare the performance of 5th - order inverse and decorrelation - based im3 sub - band dpd solutions . the 5th - order inverse reference solution for im3 + sub - band is derived in appendix b. in these simulations , the transmit waveform is again composed of two 1 mhz lte - a ul sc - fdma component carriers with qpsk data modulation , and the cc spacing is 10 mhz . the pa model , in turn , is a memoryless 5th - order model whose parameters have been identified using a true mobile pa transmitting at + 23 dbm . the 5th - order inverse dpd is using known parameters , while the decorrelation - based one is adopting the proposed sample - adaptive learning described in detail in section iv . the pa output spectra with different solutions are illustrated in fig . [ fig : psd_pth_decorr ] . it can clearly be seen that the decorrelation - based sub - band dpd substantially outperforms the @xmath4-order inverse based sub - band dpd , in both third and fifth - order cases , despite the fact that the @xmath4-order inverse solutions are using known parameters . the reason is that the inverse solutions cancel only the third and fifth order terms , as described in appendix b , but do not suppress the other induced higher - order terms , which have structural similarity and correlation with the third and fifth - order basis functions . on the other hand , the proposed decorrelation - based solution takes this explicitly into account , and thus achieves clearly better spurious emission suppression . the figure also illustrates that the performance of the fifth - order decorrelation - based sub - band dpd is clearly better than that of the third - order one . in the remaining parts of this section , we shall focus on more detailed performance evaluations of the proposed decorrelation - based sub - band dpd solution , incorporating memory both in the pa and in the predistortion processing . from now on , we refer to the decorrelation - based sub - band dpd simply as sub - band dpd , to simplify the presentation . -order inverse and decorrelation - based im3 + sub - band dpd solutions are compared , with both third - order and fifth - order nonlinear processing.,scaledwidth=48.0% ] in this subsection , we evaluate the performance of the proposed sub - band dpd in the more realistic case of having memory in the pa . the pa model is a 9th - order parallel hammerstein model , with four memory taps per branch , and the parameters have been identified using measurements with a true mobile pa transmitting at + 24 dbm . the sub - band dpd structure contains now also memory , with two taps ( @xmath129 ) per basis function . block - adaptive learning principle is adopted , with 200 blocks each containing 1000 samples . the transmit waveform is otherwise identical to earlier cases , but the cc separation is now 12 mhz . [ tab : im3_im5_evm ] fig . [ fig : psd_im3 ] shows the effectiveness of the proposed higher - order sub - band dpds , here processing the im3 + sub - band , compared to the basic third - order solution presented earlier in @xcite . up to 40 db suppression for the im3 spurious emissions is shown when adopting ninth - order processing , something that has not been reported before in any prior works using a practical pa model with memory effects . [ fig : im3_pwr ] shows then how the im3 spurious emission level changes with varying the tx power and using different sub - band dpd orders . the im3 spurious emissions are clearly below the general spurious emission limit ( -30 dbm / mhz ) even at very high tx powers up to + 23 dbm , when using a seventh - order , or higher , sub - band dpd . another main contribution in this paper is the extension of the im3 sub - band dpd solution to include also higher - order i m sub - bands . [ fig : psd_im3_5_7_9 ] shows the performance of the negative im5 and im7 sub - band dpds , in addition to the negative im3 sub - band dpd , when ninth - order nonlinear processing is adopted in the sub - band dpds . up to 40 db , 35 db , and 18 db of suppression is achieved at the negative im3 , im5 , and im7 sub - bands , respectively . this shows the effectiveness of the proposed solutions in processing and suppressing also the spurious emissions at higher i m sub - bands . in this subsection , we compare the performance of our proposed sub - band dpd technique against that of a classical full - band dpd adopting parallel hammerstein based wideband linearization and indirect learning architecture ( ila ) . the evaluation considers the linearization performance , complexity , and transmitter efficiency utilizing our earlier complexity analysis results reported in section [ sec : complexity ] . the ninth - order full - band ph ila dpd uses 100k samples for parameter learning , per ila iteration , with a total of 2 ila iterations , and the number of memory taps per ph branch is 4 . the ninth - order block - adpative sub - band dpd , in turn , uses also a total of 200k samples with block size @xmath130 , and adopts two memory taps per dpd snl basis function . the results are collected in table [ tab : im3_im5_evm ] , which shows that in addition to the significantly lower complexity measured by the number of gflops , the sub - band dpd achieves better linearization performance in terms of spurious imd suppression at the considered im3 and im5 sub - bands . on the other hand , the full - band dpd outperforms the sub - band dpd in terms of the inband distortion mitigation ( i.e. , evm ) . this is expected , since the full - band dpd linearizes the whole transmit band , including the main ccs and the imd spurious emissions . however , the evm with the sub - band dpd is only around @xmath131 which is by far sufficient for modulations at least up to @xmath132-qam . . thus the transmitter becomes more power efficient when using the sub - band dpd as shown in table [ tab : im3_im5_evm ] . for fairness , it is to be acknowledged that a full - band dpd can typically enhance the evm and aclr , while the sub - band dpd concept is specifically targeting only the spurious emissions . in order to further demonstrate the operation of the proposed sub - band dpd solution , we next report results of comprehensive rf measurements using a commercial lte - advanced mobile terminal pa together with a vector signal transceiver ( vst ) , which is implementing the rf modulation and demodulation . the actual sub - band dpd processing and parameter learning algorithms are running on a host processor . the acpm-5002-tr1 mobile power amplifier used in our measurements is designed for lte - a ul band 25 ( 1850 - 1915 mhz ) , with 29 db gain . the national instruments ( ni ) pxie-5645r vst includes both a vector signal generator ( vsg ) , and a vector signal analyzer ( vsa ) with 80 mhz instantaneous bandwidth . in these experiments , the digital baseband waveform is divided into 50 blocks of size @xmath133 samples which are first generated locally on the host processor , and then transferred to the vsg to perform rf i / q modulation at the desired power level at the pa input . the vst rf output is then connected to the input port of the external power amplifier , whose output port is connected to the vst rf input through a 40 db attenuator , implementing the observation receiver as illustrated also in fig . [ fig : measphoto ] . the vsa performs rf i / q demodulation to bring the signal back to baseband . the baseband i / q observation block is then filtered to select the im3 sub - band which is used for block - based sub - band dpd learning , after proper alignment with the locally generated basis functions , as explained in section [ sec : block_based_dpd ] . the sub - band dpd block size @xmath108 used in these experiments is @xmath134 , and the dpd memory depth @xmath84 is @xmath120 . in general , two different intra - band lte - a ca rf measurement examples are demonstrated in this subsection . the first experiment demonstrates the violation of the spurious emission limit due to the inband emission of the im3 + spur , thus not being attenuated by the tx filter . the second experiment demonstrates an own rx desensitization example , in fdd transceiver context , where the im3 + spur is located at the own rx band ( 1930 - 1995 mhz ) and is not sufficiently attenuated by the duplexer tx filter . notice that in principle , when tackling specifically the own rx desensitization problem with the proposed sub - band dpd solution , the main receiver of the fdd device could potentially be used as the observation receiver ( i.e. , learning the sub - band dpd coefficients without an extra auxiliary / observation receiver ) . however , such an approach would indeed be applicable only in the own rx desensitization case , while when mitigating other harmful emissions , an auxiliary observation receiver would anyway have to be adopted . thus , in these rf measurements , we adopt the auxiliary receiver based approach , while consider further developments of the main receiver based parameter learning as an important topic for our future work . [ fig : psd_lab_im3 ] shows the measured power spectral density at the pa output using the im3 sub - band dpd with different orders . the adopted waveform is an intra - band ca lte - a ul signal with 3 mhz per cc and 20 mhz carrier spacing . the power level at the pa output in this example is + 25 dbm . the intermodulation distortion at the im3 + sub - band is emitted inband , since the total tx band covers 1850 - 1915 mhz , and is clearly violating the spurious emission limit ( -30 dbm / mhz ) when no dpd is used . when using the sub - band dpd , the spurious emission level is well below the emission limit , given that at least 5th - order processing is deployed . . notice that there is no need for predistorting the im3- sub - band in this example , since those emissions will be filtered out by the tx filter . [ fig : psd_rxsens_lab ] illustrates another lte - a intra - band ca ul band 25 example , now with 5 mhz cc bandwidths and 40 mhz carrier spacing . the power level at the pa output in this example is + 25 dbm . then , by adopting a ninth - order im3 + sub - band dpd , more than 20 db of spurious emission suppression is achieved , as can be seen in fig . [ fig : psd_rxsens_lab ] . again , two taps ( @xmath129 ) per basis function are used , and block - adaptive parameter learning is deployed . [ fig : dpd_coeff_9th ] shows the real - time convergence of the sub - band dpd coefficients for the third , fifth , seventh , and ninth - order basis functions , respectively . it can be seen that the coefficients converge in a stable manner in a real rf environment due to the orthogonalization of the snl basis functions , as explained in section [ sec : dpd_adaptive_filter ] . per basis function , when the positive im3 sub - band is considered and block - adaptive learning is deployed . an rf measurement example with two 5 mhz lte - a ul ccs and 40 mhz carrier spacing is used with a real commercial mobile pa operating at + 25 dbm.,scaledwidth=48.0% ] . this is 25 db above the effective rx noise floor when assuming 9 db ue rx noise figure ( nf ) @xcite . this would thus cause significant own receiver desensitization that could lead to a complete blocking of the desired rx signal . on the other hand , when the proposed im3 + sub - band dpd is deployed , the integrated power of the im3 + spur at the rx band will be approximately -95 dbm/5mhz , which is only 3 db above the effective rx noise floor , as shown also in fig . [ fig : im3_pwr_rxband ] . though the residual spur is still slightly above the effective rx noise floor , the sensitivity degradation is substantially relaxed , despite operating at a maximum pa output power of + 25 dbm . we elaborate on this further in fig . [ fig : im3_pwr_rxband ] , showing the integrated power of the im3 + spur at the rx lna input while changing the pa output power level . with a ninth order sub - band dpd , we can transmit up to + 18 dbm pa output power with an effectively perfectly linear tx , in terms of im3 + spur level , compared to + 10 dbm pa output power without dpd . additionally , with dpd , the integrated im3 + power is less than the effective rx noise floor up to + 23 dbm pa output power , while without dpd it is already 20 db above the noise floor at the same power level . in this article , a novel low - complexity sub - band digital predistortion ( dpd ) solution was proposed for suppressing unwanted spurious emissions in non - contiguous spectrum access . novel decorrelation based adaptive parameter learning methods were also formulated , allowing efficient estimation and tracking with low computational complexity . all algorithm derivations and modeling were carried out in the general case of having memory in the pa as well as in the sub - band dpd processing . different nonlinear distortion and processing orders , beyond classical third - order cases , were also reported . the proposed technique can find application in suppressing inband spurs which would violate the spurious emission limit , suppressing out - of - band spurs falling , e.g. , on the own receiver band , or in protecting primary user transmissions in cognitive radio systems . a quantitative complexity analysis was presented , comparing the proposed solution to conventional full - band dpd solutions available in the literature . the performance was evaluated in a comprehensive manner , showing excellent linearization performance despite the considerably reduced complexity compared to classical full - band solutions . finally , extensive rf measurement results using a commercial lte - advanced mobile power amplifier were reported , evidencing up to 22 db suppression of the most problematic third - order spurious emissions . here , we derive the analytical minimum mean - square error solution shown in ( 32 ) , which is used as a reference solution in simulations in section [ sec : analytical_sim ] . we first define the so - called error signal as @xmath59 , since with ideal predistortion the im3 + sub - band signal would be zero , and thus the optimization means minimizing the power of this error signal . from ( [ eq : thirdorderanalysis ] ) , the error signal reads @xmath135 where @xmath51 and @xmath52 are the third - order memoryless pa model parameters , @xmath53 denotes the dpd coefficient to be optimized , and @xmath6 and @xmath7 are the baseband equivalents of the two component carriers . the statistical expectation @xmath60={{\rm i\kern-.3em e}}[e(n)e^*(n)]$ ] , assuming that the component carrier signals @xmath6 and @xmath7 are statistically independent , and ignoring some high - order vanishingly small terms , then reads @xmath136 = { { \rm i\kern-.3em e}}_{42 } \left[|f_3|^2 + |f_1|^2|\alpha|^2 + \alpha f_1 f_3^ * + \alpha^ * f_1^ * f_3 \right ] \nonumber\\ & + ( 2 { { \rm i\kern-.3em e}}_{62 } + 2 { { \rm i\kern-.3em e}}_{44 } ) \left[|f_3|^2 ( \alpha + \alpha^ * ) + |\alpha|^2(f_1 f_3^ * + f_1^ * f_3 ) \right ] \nonumber\\ & + ( 4 { { \rm i\kern-.3em e}}_{46 } + 8 { { \rm i\kern-.3em e}}_{64 } + 4 { { \rm i\kern-.3em e}}_{82 } ) |f_3|^2 |\alpha|^2 , \label{eq : meansquarederror}\end{aligned}\ ] ] where @xmath62 is used as a shorthand notation for @xmath63 { { \rm i\kern-.3em e}}[|x_2|^j]$ ] . now , differentiating ( [ eq : meansquarederror ] ) with respect to @xmath53 , yields @xmath137 & = { { \rm i\kern-.3em e}}_{42}\left[\alpha^ * |f_1|^2 + f_1 f_3^ * \right ] \nonumber\\ & + ( 2 { { \rm i\kern-.3em e}}_{62 } + 2 { { \rm i\kern-.3em e}}_{44 } ) \left[|f_3|^2 + \alpha^ * ( f_1 f_3^ * + f_1^ * f_3 ) \right ] \nonumber\\ & + ( 4 { { \rm i\kern-.3em e}}_{46 } + 8 { { \rm i\kern-.3em e}}_{64 } + 4 { { \rm i\kern-.3em e}}_{82 } ) |f_3|^2 \alpha^*. \label{eq : meansquarederrordiff}\end{aligned}\ ] ] then , setting ( [ eq : meansquarederrordiff ] ) to zero and solving for @xmath53 yields the optimal mmse dpd parameter , given by @xmath61^*}{\left[|f_1|^2 { { \rm i\kern-.3em e}}_{42 } + 4 \mathbb{r}(f_1 f_3^*)({{\rm i\kern-.3em e}}_{62 } + { { \rm i\kern-.3em e}}_{44 } ) + 4 |f_3|^2 ( { { \rm i\kern-.3em e}}_{46 } + 2 { { \rm i\kern-.3em e}}_{64 } + { { \rm i\kern-.3em e}}_{82})\right]^*},\end{aligned}\ ] ] where @xmath138 denotes the real - part operator . this concludes the proof . here , we derive the @xmath139 order inverse solution for the im3 + sub - band dpd , to be used as a reference solution in the simulations in section [ sec : inv_vs_decorr_sim ] . the output of a memoryless @xmath139 order polynomial pa model is @xmath140 where @xmath51 , @xmath52 and @xmath141 are the polynomial coefficients , and @xmath142 is the composite baseband equivalent input signal , as given in ( [ eq : pa_in ] ) . through direct substitution of ( [ eq : pa_in ] ) in ( [ eq : pa_5th_nomem ] ) , the baseband equivalent im3 + distortion term , located at three times the if frequency , can be extracted and reads @xmath143 stemming from the signal structure in ( [ eq : im3_out_5th_nomem ] ) , the sub - band dpd injection signal is composed of three basis functions of the form @xmath144 , @xmath145 and @xmath146 . in case of fifth - order inverse based sub - band dpd , these basis functions are multiplied by proper coefficients such that all the distortion terms at the im3 + sub - band at the pa output , up to order five , are cancelled . thus , incorporating such dpd processing , yet with arbitrary coefficients , the composite baseband equivalent pa input signal reads @xmath147 where the subscript @xmath148 in the coefficients is emphasizing the @xmath4-order inverse based solution . substituting ( [ eq : pa_in_fifthdpd ] ) in ( [ eq : pa_5th_nomem ] ) , and extracting the third and fifth order im3 + terms , yields r4 - 124353 , non - contiguous intraband unwanted emission . , 3gpp tech . , feb . 2013 . h. gandhi , d. greenstreet , and j. quintal , `` digital radio front - end strategies provide game - changing benefits for small cell base stations , '' tech . texas instruments , may 2013 . l. guan and a. zhu , `` green communications : digital predistortion for wideband rf power amplifiers , '' , vol . 15 , no . 7 , pp . 8489 , dec . r4 - 123797 , `` ue reference sensitivity requirements with two ul carriers . , '' 3gpp tech . , feb . 2013 . . chao yu , wenhui cao , yan guo , and a. zhu , `` digital compensation for transmitter leakage in non - contiguous carrier aggregation applications with fpga implementation , '' , vol . 43064318 , dec 2015 . a. kiayani , m. abdelaziz , l. anttila , v. lehtinen , and m. valkama , `` digital mitigation of transmitter - induced receiver desensitization in carrier aggregation fdd transceivers , '' , vol . 11 , pp . 36083623 , nov 2015 . bassam , f.m . ghannouchi , and m. helaoui , `` 2-d digital predistortion ( 2-d - dpd ) architecture for concurrent dual - band transmitters , '' , vol . 10 , pp . 25472553 , oct . p. roblin , c. quindroit , n. naraharisetti , s. gheitanchi , and m. fitton , `` concurrent linearization , '' , vol . 14 , no . 7 , pp . 7591 , nov . y. liu , j. yan , and p. asbeck , `` concurrent dual - band digital predistortion with a single feedback loop , '' , vol . 5 , pp . 15561568 , may 2015 . p. roblin , s. k. myoung , d. chaillot , y. g. kim , a. fathimulla , j. strahler , and s. bibyk , `` frequency - selective predistortion linearization of rf power amplifiers , '' , vol . 1 , pp . 6576 , jan . 2008 . j. kim , p. roblin , d. chaillot , and z. xie , `` a generalized architecture for the frequency - selective digital predistortion linearization technique , '' , vol . 1 , pp . 596605 , jan . bassam , m. helaoui , and f.m . ghannouchi , `` channel - selective multi - cell digital predistorter for multi - carrier transmitters , '' , vol . 60 , no . 8 , pp . 23442352 , m. abdelaziz , l. anttila , a. mohammadi , f. ghannouchi , and m. valkama , `` reduced - complexity power amplifier linearization for carrier aggregation mobile transceivers , '' in _ ieee international conference on acoustics , speech , and signal processing _ , pp . 39083912 , may 2014 . m. abdelaziz , l. anttila , j.r . cavallaro , s. bhattacharyya , a. mohammadi , f.m . ghannouchi , m. juntti , and m. valkama , `` low - complexity digital predistortion for reducing power amplifier spurious emissions in spectrally - agile flexible radio , '' in _ 9th international conference on cognitive radio oriented wireless networks _ , pp . 323328 , june 2014 . tehrani , h. cao , s. afsardoost , t. eriksson , m. isaksson , and c. fager , `` a comparative analysis of the complexity / accuracy tradeoff in power amplifier behavioral models , '' , vol . 58 , no . 6 , pp . 15101520 , june 2010 . mahmoud abdelaziz ( s13 ) received the b.sc . degree ( with honors ) and m.sc . degree in electronics and electrical communications engineering from cairo university , egypt , in 2006 and 2011 . he is currently pursuing the doctoral degree at tampere university of technology , finland where he works as a researcher with the department of electronics and communications . from 2007 to 2012 he has been working as a communication systems and signal processing engineer at newport media , as well as other companies in the wireless industry . his research interests include statistical and adaptive signal processing in flexible radio transceivers , wideband digital pre - distortion , and cognitive radio systems . lauri anttila ( s06 , m11 ) received the m.sc . degree and d.sc . ( tech ) degree ( with honors ) in electrical engineering from tampere university of technology ( tut ) , tampere , finland , in 2004 and 2011 . currently , he is a senior research fellow at the department of electronics and communications engineering at tut . his research interests are in signal processing for wireless communications , in particular radio implementation challenges in 5 g cellular radio and full - duplex radio , flexible duplexing techniques , and transmitter and receiver linearization . he has co - authored over 60 peer reviewed articles in these areas , as well as two book chapters . chance tarver received a b.s . degree in electrical engineering from louisiana tech university , ruston , la , in 2014 , and his m.s . degree in electrical and computer engineering in 2016 from rice university , houston , tx . he is currently a ph.d . student in the department of electrical and computer engineering at rice university . his research interests include software defined radio and signal processing for wireless communications . kaipeng li ( s14 ) received his b.s . degree in physics from nanjing university , nanjing , china , in 2013 , and his m.s . degree in electrical and computer engineering from rice university in 2015 . he is currently a ph.d . candidate in the department of electrical and computer engineering at rice university , houston , texas . his research interests include digital signal processing , parallel computing on gpgpu and multicore cpu , software - defined radios and massive mimo systems . joseph r. cavallaro ( s78 , m82 , sm05 , f15 ) received the b.s . degree from the university of pennsylvania , philadelphia , pa , in 1981 , the m.s . degree from princeton university , princeton , nj , in 1982 , and the ph.d . degree from cornell university , ithaca , ny , in 1988 , all in electrical engineering . from 1981 to 1983 , he was with at@xmath151 t bell laboratories , holmdel , nj . in 1988 , he joined the faculty of rice university , houston , tx , where he is currently a professor of electrical and computer engineering . his research interests include computer arithmetic , and dsp , gpu , fpga , and vlsi architectures for applications in wireless communications . during the 1996 - 1997 academic year , he served at the national science foundation as director of the prototyping tools and methodology program . he was a nokia foundation fellow and a visiting professor at the university of oulu , finland in 2005 and continues his affiliation there as an adjunct professor . he is currently the director of the center for multimedia communication at rice university . he is a fellow of the ieee and a member of the ieee sps tc on design and implementation of signal processing systems and the chair - elect of the ieee cas tc on circuits and systems for communications . he is currently an associate editor of the ieee transactions on signal processing , the ieee signal processing letters , and the journal of signal processing systems . he was co - chair of the 2004 signal processing for communications symposium at the ieee global communications conference and general / program co - chair of the 2003 , 2004 , and 2011 ieee international conference on application - specific systems , architectures and processors ( asap ) , general / program co - chair for the 2012 , 2014 acm / ieee glsvlsi , finance chair for the 2013 ieee globalsip conference , and tpc co - chair of the 2016 ieee sips workshop . he was a member of the ieee cas society board of governors during 2014 . mikko valkama ( s00 , m01 , sm15 ) was born in pirkkala , finland , on november 27 , 1975 . he received the m.sc . and ph.d . degrees ( both with honors ) in electrical engineering ( ee ) from tampere university of technology ( tut ) , finland , in 2000 and 2001 , respectively . in 2002 , he received the best ph.d . thesis -award by the finnish academy of science and letters for his dissertation entitled `` advanced i / q signal processing for wideband receivers : models and algorithms '' . in 2003 , he was working as a visiting researcher with the communications systems and signal processing institute at sdsu , san diego , ca . currently , he is a full professor and department vice - head at the department of electronics and communications engineering at tut , finland . his general research interests include communications signal processing , estimation and detection techniques , signal processing algorithms for software defined flexible radios , cognitive radio , full - duplex radio , radio localization , 5 g mobile cellular radio , digital transmission techniques such as different variants of multicarrier modulation methods and ofdm , and radio resource management for ad - hoc and mobile networks .

in the last decades point contact spectroscopy has been proved to be a very powerful method to study electron - phonon interaction in metals ( for reviews see , and ) . in the recent years it has been realized that in mesoscopic metallic systems the electron scattering by some dynamical defects may play a major role in understanding the energy relaxation and dephasing processes of the electrons at low temperature ( for reviews see and ) . such defects can also be studied by point contact spectroscopy provided that they are situated in the contact region . these defects can either be magnetic impurities or structural defects where an atom or a group of atoms have a metastable , almost double degenerate groundstate forming a two - level system ( tls ) . the tls spectroscopy is very similar to the phonon spectroscopy , but as the tlss can not propagate like phonons , they are not in thermal equilibrium with the bath . the tls can be formed as defects in the bulk region of crystalline materials , or at dislocations , at surfaces due to impurities and also in amorphous materials . in most of the cases the internal structure of the tlss are unknown and they can behave very differently . the dynamics of magnetic impurities shows the kondo effect@xcite where by lowering the temperature a kondo resonance is developed at the range of the characteristic kondo temperature , @xmath0 and approaching zero temperature a magnetic singlet is formed by the impurity spin and the electrons . in that case near zero temperature the magnetic singlet acts just like a static impurity and the dynamical processes are frozen out . the subject of slow defects can be described by the conventional theory of tlss and its application to the point contact is the present review . there are , however , many experimental facts which can not be described in that way , and show resemblance to the magnetic kondo anomalies even if their magnetic origin is very unlikely . we make an effort to point at those experiments by comparing the observed behavior with the predictions of the slow tls model . several of these are believed to be due to dynamical defects , different from slow tls . it must be emphasized that there is no theory which is generally accepted and can be applied . earlier it has been proposed that they are fast tlss , which can be described by the so called two - channel kondo model.@xcite that proposal has recently been criticized and debated.@xcite these scatterers now appear rather as `` phantoms '' and the present review does not cover that issue , we , however , present the earlier interpretation just to help in reading the earlier experimental publications . * . in that case an atom or a group of atoms can have a transition between the two states of lowest energies by direct tunneling or at higher temperature by thermal excitation over the barrier . in the former case those systems are also called as tunneling centers ( for review see @xcite ) . the transition rate can be on the scale of seconds and even on longer scales . in that case e.g. the specific heat may depend on the speed of the measurement as to reach thermal equilibrium takes longer time . the resistivity of such mesoscopic system may jump in time between two or several values on the above time scale , as in small mesoscopic systems the resistivity depends on the position of a single atom even if the two possible atomic positions are much smaller than an atomic distance . the study of point contacts containing tlss in the contact region turned out to be very powerful as the tls may result in zero bias anomalies shown by the @xmath1 characteristics . the electron passing through the point contact can interact dynamically with the tls or it can be elastically scattered , which scattering depends on the actual position of the atom forming the tls . the inelastic scattering may result in a back - scattering of the electron in which case the electron does not proceed from one electrode to the other when the applied voltage is larger than the splitting of the tls and the process shows up as a resistivity minimum at zero bias . in the case of elastic scattering the transmission rate may depend on the position of a single atom and the anomalies can have either sign . the characteristic relaxation times from the tls causing slow telegraph fluctuation of resistance to those being responsible for zero bias anomalies in the @xmath1 curve varies over several orders of magnitude , but for all these so - called slow tlss the average transition time is much longer than the electron - tls interaction time . * the fast tlss are strongly debated whether they can show kondo anomaly.@xcite here we present the previous ideas just to make contact with the extended literature , and the present status of the problem will be discussed at the end of the paper . if a conduction electron experiences several switching occasions one after the other we speak about fast tls . these fast tlss may show kondo anomaly as the tls s two states can be described by a quasi - spin replacing the real impurity spin of the magnetic kondo problem and the scattered electron can have different spherical momenta e.g. @xmath2- and @xmath3-states playing similar role to the real spin of the conduction electron . the spin - flip process of the magnetic kondo problem is replaced in the latter case by electron assisted transition between the two states of the tls . that trasnition was is originally believed to be electron assisted tunneling , which is unlikely according to the present theories,@xcite however an electron assisted transition without tunneling is still possible.@xcite in point contacts containing magnetic impurities the kondo resonance scattering contributes to the back - scattering rate , thus the kondo impurities result in zero bias peak in the dynamical resistivity , which has a width characterized by the kondo temperature , @xmath0 . the sign of the peak is just opposite to those caused by slow tlss . i.k . yanson and his co - workers@xcite and also other groups@xcite studied the zero bias anomalies due to intentionally placed magnetic impurities . considering the kondo effect there is a significant difference between the magnetic case and the tls . in the latter case the conduction electrons , additionally to their angular momenta occurring in the hamiltonian and the coupling , have a further degree of freedom , the real spin , which makes the electronic sea double degenerate although it does not appear in the coupling constants . referring to that extra degeneracy the phenomenon is called two - channel kondo problem ( 2-ck ) in contrast to the magnetic one - channel case ( 1-ck ) ( for earlier review see ) . that extra degeneracy prevents the system to have a singlet ground state at low temperature , it has finite entropy and , therefore , the fermi liquid behavior does not show up as far as a low energy cutoff ( e.g. the energy difference between the two lowest states of the tls , known as splitting ) is reached as the temperature is lowered . + the condition for the formation of a 2-ck ground state is that the energy splitting is negligible compared to the kondo temperature , which is generally not expected . the zero bias anomalies have been believed to be powerful tools to make difference between 1-ck and 2-ck problems ( for a review see ) . for example annealing or electromigration could result in disappearance or modification of the zero bias anomalies in the case of structural defects while such effects are not expected for magnetic impurities.@xcite similar issues also appear in the electron dephasing time which has been intensively debated recently@xcite . another version of the two - channel kondo effect was proposed to explain the zero bias anomalies in uru@xmath4si@xmath4 point contacts@xcite where two degenerate localized electron orbitals of the u atom play the role of the tls which interact with the conduction electrons.@xcite the issue of the 2-ck problem was responsible for the great interest in the zero bias anomalies due to tls and , therefore , it has a great importance to distinguish between anomalies due to slow tlss and fast scatterers . even if a considerable amount of the discussions are carried out considering fast tlss there is no doubt that in many cases the slow ones are dominant e.g. where the zero bias anomaly in the point contact spectrum has a positive sign.@xcite working on the issues listed above we learned that the literature on the theory of slow tlss in point contacts is very much scattered over many papers and journals , the predictions should be collected from a large number of papers , especially , published in russian and ukrainian physical journals some of them not available at smaller libraries . that experience encouraged us to collect and at several occasions to complete those results . the present review is written in a self - contained manner , thus no further reading is necessary to follow the theory . we want to emphasize that the authors did not consider as a task to give the historically proper complete list of references but they picked up those which are the most appropriate ones to get further information . the model of two level tunneling systems was at first introduced to explain the unexpected low - temperature behavior of specific heat and thermal conductivity in disordered solids which can not be described by phonon excitations.@xcite the model based on two level systems , which was able to explain the linear temperature dependence of specific heat of amorphous dielectrics at low temperatures was suggested by anderson , halperin and varma@xcite and phillips.@xcite this so - called tunneling model assumes that in an amorphous state some atoms or groups of atoms may switch between two , energetically nearly equivalent configurations . this situation is modeled as a double well potential with two stable states differing by energy @xmath5 , and a potential barrier between them with a tunneling rate @xmath6 ( see fig . [ tls ] ) . between the two positions , and a tunneling probability @xmath6 for crossing the barrier between the two metastable states . @xmath7 and @xmath8 denote the width of the barrier and the distance between the minima , respectively . _ _ ] the tunneling probability can be approximated in terms of the gamow parameter ( @xmath9 ) : @xmath10 , where @xmath9 is defined by the characteristic height , @xmath11 and width , @xmath7 of the potential barrier , and the mass , @xmath12 of the tunneling particle : @xmath13 furthermore it is assumed that the tlss in a solid are uniformly distributed in terms of the parameters @xmath5 and @xmath9 , that is the density of states @xmath14 for the first excited states of the tlss is constant : @xmath15 experimental studies do indeed support this assumption.@xcite the two positions of the system are described by a quasispin , and the hamiltonian of the tls is expressed in terms of the quasispin s pauli - operators : h_tls=12 _ , ( b^+_^z_b_+ _ 0b^+_^x_b _ ) , [ h_tls ] where the indexes @xmath16 and @xmath17 correspond to the two states of the tls ( @xmath18 and @xmath19 ) , and @xmath20 and @xmath21 are the creation and annihilation operators for the state @xmath16 . this hamiltonian can be diagonalized by writing the eigenstates as:|e_+&=&|1+|2 + between the two eigenstates is : e=.the coefficients @xmath22 and @xmath23 can be expressed as : = + = .[munu]for a highly asymmetric tls , where @xmath24 the energy eigenstates are the two positions of the tls , that is @xmath25 and @xmath26 . in the opposite case , where the tunneling energy is much larger than the splitting ( @xmath27 ) the energy eigenstates are delocalized between the two positions , and @xmath28 . in a disordered material two level systems are not isolated objects , but they interact with their neighborhood . for instance a strain induced by an elastic wave may change the parameters of the two level systems . it is assumed that only the asymmetry parameter @xmath5 is influenced by an elastic wave , thus the interaction between two level systems and phonons can be treated as a perturbation in @xmath5 : @xmath29 where @xmath30 is the coupling constant , and @xmath31 is the appropriate local elastic strain tensor . in metallic samples the coupling to conduction electrons has to be taken into account as well . the general form of the hamiltonian describing the interaction between the electrons and the tls is:@xcite h_el - tls=__1,_2 , + , , b^+_a^+__2v^__1_2a__1b _ , [ gen_int_tls ] where @xmath32 and @xmath33 are the electron creation and annihilation operators , @xmath34 is the spin index of the electron and @xmath11 is the interaction matrix element , which can be decomposed in terms of pauli operators:@xcite v^__1_2=_i = x , y , z v^i__1_2^i_+v^0__1_2_. [ v_pauli ] the last term stands for potential scattering on the average positioned tls , the term with @xmath35 describes the difference between scattering amplitudes for the two positions of the tls , while @xmath36 and @xmath37 correspond to the electron assisted transitions between the two states of the tls . the interaction hamiltonian is further simplified if the electron - tls interaction time is much shorter than the transition time of the tls , thus the tls stays in a given position during the electron scattering . for these so - called slow two level systems the electron assisted transitions described by @xmath38 and @xmath39 can be neglected , and the hamiltonian is : h_el - tls=__1,_2 , v^z__1_2a^+__2a__1 _ , ^z_b^+_b_+__1,_2,v^0__1_2a^+__2a__1 _ , _ b^+_b _ , [ int_slowtls]the interaction matrix elements between the delocalized energy eigenstates are : e_-|h_el - tls|e_+=__1,_2,a__2^a__1,[matrix_elements_inel]which gives the amplitude of the electron s inelastic scattering on the tls , which is associated with an energy change , @xmath40 . the matrix elements corresponding to the elastic processes are written as : e_+|h_el - tls|e_+&=&__1,_2,a__2^a__1 + e_-|h_el - tls|e_-&=&__1,_2,a__2^a__1.[matrix_elements_elastic]in both cases the energy of the scattered electron is conserved , however the scattering cross section is different if the tls is in the @xmath41 or @xmath42 state . by simple argumentation the interaction strength , @xmath43 can be estimated as : @xmath44,@xcite as the difference between the electron scattering amplitudes in the two positions must be proportional to the distance of those positions on the scale of the fermi wavelength , and the strength of the scattering , @xmath11 . @xmath45 is very model dependent , e.g. the tunneling atom can be different from the most of the atoms . the electron assisted transition processes are only important in the case of fastly relaxing two level fluctuators . in this case the hamiltonian given by eq . ( [ gen_int_tls ] ) can be scaled to the two channel kondo problem ( 2ck ) with the hamilton operator : h_int=_tlsv_p_1,p_2 , , , a^+_p_1_a_p_2 , [ 2ck ] where the the momentum dependence of @xmath46 is decomposed into appropriately choosen spherical waves with indices @xmath47 , and @xmath48 ( for details see ) . @xmath49 is a pseudo - spin corresponding to the spherical indices of the electrons , whereas @xmath50 is a pseudo - spin corresponding to the states of the tls , and @xmath11 is the interaction strength . in this model the electron spin is an extra degree of freedom compared to the single channel magnetic impurity kondo problem , and this quantum number is conserved during the interaction and the coupling is independent of that . in contrast to the single - channel kondo problem , which has a non - degenerated , @xmath51 ground state because of the spin compensation cloud formed by the conduction electrons , the ground state of the two channel kondo problem has a non - zero entropy . in the two channel kondo model the conduction electrons overscreen the impurity spin by forming a non - trivial spin state . the low temperature regime of the two channel kondo model was found to have non - fermi - liquid behavior . this paper is devoted to review the literature of the slow tlss and point contacts . the kondo regime of the tlss is out of present consideration , but one can find an up - to - date review of this field in the work of cox and zawadowski.@xcite in the previous part of this section the scattering rates due to tlss are calculated in born approximation . considering only the screening interaction @xmath43 and ignoring the assisted tunneling rates ( @xmath52 ) kondo@xcite has found power law correction to the tunneling rate which is due to the building up of a screening cloud or friedel oscillation around the changed position of the tunneling atom . that is closely related to the anderson s orthogonality catastrophe@xcite and the x - ray absorption problem ( see e.g. ) . that phenomenon is associated with creating a large number of electron - hole pairs as the position of the tunneling atom is changed . the renormalized dimensionless coupling constant ( usually denoted by @xmath53 ) can be expressed in terms of the phase shift @xmath54 describing the scattering of the electrons by the atom in the s - wave channel ( see e.g. ) and the separation distance , @xmath8 between the two positions of the tls @xmath55 ^ 2 , \label{k}\ ] ] where @xmath56 with the spherical bessel function , @xmath57 and @xmath58 for only @xmath2-wave scattering . in the weak coupling limit @xmath59 and @xmath60 that expression is proportional to @xmath61 . in many other publications the notation @xmath16 is used for @xmath53 ( see e.g. ) . the behavior of a tls coupled to electronic heat bath ( ohmic region ) by the coupling @xmath53 has been studied in great detail ( see e.g. and ) . in general there are two different regions regarding the coupling strength @xmath53 . here the asymmetry parameter @xmath5 is disregarded and @xmath62 is taken . * . at @xmath63 the atom has a periodic motion between the two positions . as @xmath53 is increased the oscillation is more damped and the renormalized tunneling frequency @xmath64 is given by @xmath65 where @xmath66 is the high frequency cutoff , which has first believed to be the electronic bandwidth ( see e.g. , but according to the adiabatic renormalization it is rather the energy scale of the lower excitations in the potential well where several excited levels are considered.@xcite * . the motion is incoherent and the atom is spending longer times in a position before tunneling as @xmath53 is increased . many experimental facts indicate ( see next subsection ) that the tlss are in the first region ( @xmath58 ) and even can be in the weak coupling limit ( @xmath67 ) . the main issue is how the quasielastic and inelastic scattering rates are modified by the coupling @xmath53 . that problem was investigated by h. grabert , s. linkwitz , s. dattagupta , u. weiss@xcite considering the neutron scattering by tls . those results are very instructive for the present case , there are , however , essential differences : * korringa type of relaxations are very important in the relaxation of the tls coupled to the ohmic electronic heat bath . if energy @xmath68 is transferred to the tls the spectrum of the tls is broadened by a relaxation rate proportional to @xmath68 . in the case of point contacts that energy may arise from the energy of the electron passing through the point contact . at finite bias , @xmath11 the system is , however , out of equilibrium and in the stationary state of the tls the averaged energy of the tls is proportional to @xmath11 also . thus , the nonequlibrium situation should be treated by using e.g. the keldysh technique.@xcite * the energy transfer is not determined by the applied voltage , in contrast to the neutron scattering experiment where the energy of the incoming and outgoing neutrons are simultaneously measured . in the present case by small change of the voltage the energy of the extra electrons passing through the point contact are given ( @xmath69 ) but the energy of the scattered electrons is not determined , thus the energy of the outgoing electron shows some distribution . that results in an additional smearing in the spectrum . in the following we give a brief summary of the neutron scattering case.@xcite very accurate results are obtained for low temperature and energies except the range @xmath70 . that is certainly the interesting range considering especially for the long range tails of the characteristics which are crucial for the interpretations of the experiments.@xcite the calculation was carried out by expanding in the tunneling events and summing up . more comprehensive studies were performed by using numerical renormalization group@xcite for the entire energy range of energy and even including the asymmetry , @xmath5 and later even for nonequilibrium.@xcite . the previous results@xcite have a very good fit , by the formula @xmath71 where @xmath33 , @xmath72 , and @xmath73 are fitted parameters . that clearly shows a double peak structure with energies @xmath74 . the lines are broadened by a korringa type energy relaxation rate proportional to @xmath68 . in the nearly lorentzian lineshape no direct role of coupling @xmath53 was discovered . those results are strongly indicating that in weak coupling region that would be very hard to discover any sign of anomalous relaxation rate in the tunneling characteristics . that situation is further complicated by the averaging over the distribution of the energy of the scattered electrons and the out of equilibrium situation . we find , that the experimental studies of the tunneling characteristics are not appropriate tool to look for the interaction effects , except that @xmath6 is renormalized . that situation is very different from the cases of one and two channel kondo impurities , where the characteristics are determined by the kondo resonances . there are several experimental methods to investigate the kinetics of tlss . the first two methods to be discussed are studying the tlss by their interaction with phonons . in heat release measurements the time dependence of heat release or specific heat is measured : after a sudden cool down of the sample due to the energy relaxation of the tlss it takes a long period of time to reach thermal equilibrium . the second group of measurements investigates the propagation of sound in amorphous materials measuring sound attenuation and sound velocity.@xcite heat release measurements basically focus on amorphous dielectrics and superconductors , as in metals the contribution of conduction electrons to the heat capacity usually dominates that of tlss , thus the estimation of @xmath75 is more difficult . according to the heat release measurements of kol et al.@xcite in amorphous metal fe@xmath76b@xmath77si@xmath78 the density of the tlss is about @xmath79 . though in acoustic measurements basically the phonons are investigated , if the electron - tls coupling is strong , the results may significantly differ from that on amorphous dielectrics.@xcite a simple description for ultrasound absorption by two level systems is as follows . the occupation numbers for the upper and the lower states of the tls are @xmath80 and @xmath81 , respectively . the probability for ultrasound absorption is @xmath82 , where @xmath14 is the ultrasound intensity , and @xmath16 is the absorption coefficient . the probability of relaxation is @xmath83 , where @xmath84 is the relaxation time of the tls . in balance these two quantities are equal , thus one obtains : @xmath85 the absorbed energy is : @xmath86 it is easy to see , that at large enough intensities ( @xmath87 ) , the energy absorption saturates to a value inversely proportional to the relaxation time ( @xmath88 ) . in insulators and semiconductors the available maximal ultrasound power is enough to drive the absorption into saturation at @xmath89k , but in metals such saturation is only obtained at temperatures as low as @xmath90mk . it implies that the relaxation times in metallic samples are much shorter than in disordered insulators . the interpretation for this experimental observation is , that the electron - tls scattering processes are dominating in the relaxation of tlss . assuming a standard korringa - like electron - tls relaxation , where an electron - hole pair is created , at low enough temperatures one obtains : @xmath91 that is , the ultrasound absorption measurements give the opportunity to make an estimation for the electron - tls coupling parameter ( @xmath53 ) , and the density of the tlss ( @xmath75 ) . in fact the acoustic properties of disordered metals do not show universal behavior as in the case of insulating glasses . there are several theoretical approaches ( the mentioned standard korringa - like relaxation,@xcite the strong coupling theory of vladr and zawadowski,@xcite the electron - polaron effect considered by kagan and prokofev@xcite ) that may explain the aspects of the relaxation of tlss due to conduction electrons.@xcite still , taking into account the simple electron - tls coupling given by eq . ( [ gen_int_tls ] ) , the acoustic measurements can give estimations for the density of two level systems ( @xmath75 ) or the electron - tls coupling constant ( @xmath53 ) . in case of superconductors the korringa relaxation is suppressed by the presence of the superconducting gap , but the gap disappears if a high enough magnetic field is applied on the sample , switching on the tls - electron interaction.@xcite according to the measurements of esquinazi et al.@xcite in the normal and superconducting state of @xmath92 compared with the theory of kagan and prokofev@xcite the coupling parameter is approximately @xmath93 . the acoustic measurements of coppersmith and golding@xcite on the normal conducting amorphous metal pd@xmath94si@xmath95cu@xmath96 estimated the coupling constant as @xmath97 . for the density of the two level systems a lot of results are available in the literature , and it can be stated generally , that it is the same within a factor of less than 10 over a large variety of disordered material , being metallic glasses or dielectric amorphous material . ( for tlss with energy splitting less than 1 kelvin it is approximately 1 - 10ppm . ) a detailed review of heat release and acoustic measurements in disordered material can be found in , and for earlier data see . electron - tls interaction can be studied directly by measuring electric transport in disordered metals . point contact spectroscopy offers the possibility to investigate the properties of even a single two level system centered in the vicinity of the contact . the most spectacular sign of two level systems in point contacts is the so - called telegraph noise : the resistance of the contact is fluctuating between two ( or more ) discrete values on the timescale of seconds . one can estimate the average lifetimes in the excited ( @xmath98 ) and ground state ( @xmath99 ) by recording hundreds of transitions , and fitting the resulting histograms of lifetimes ( @xmath100 and @xmath101 ) with exponential decay functions ( @xmath102 and @xmath103 , respectively ) . the two life times are related to each other by the detailed balance , @xmath104 , where @xmath5 is the asymmetry parameter . if the tls jumps to the other state , the electron screening cloud also needs to rearrange , which makes the jumps of the tls slower . ( the building up of the electronic screening cloud is related to a process which is similar to the x - ray absorption in metallic systems@xcite , and which is also called as electron - polaron effect.@xcite ) from the theoretical point of view this slow - down of the tls motion is treated as a renormalization of the tunneling amplitude:@xcite @xmath105 where @xmath66 is the bath cutoff frequency for which originally the electron bandwidth had been taken,@xcite but later it was proposed that it is replaced by a typical energy of the next higher excitation in the potential well.@xcite golding et al.@xcite studied the two level fluctuation in mesoscopic disordered bi samples . they argued that their samples were in the strong coupling limit , where the coupling of the tls to conduction electrons is very strong compared to the tunneling matrix element , thus the latter can be treated as a small perturbation . in this limit the tunneling of the tls is incoherent , because of the dephasing due to the fast oscillations of the electron bath . here a theoretical scaling function can be set up : @xmath106 , which does indeed agree with the experimental result . ( the energy splitting of the the tls can be tuned by changing the electron density , which can be reached by applying weak magnetic field.@xcite ) in the limit @xmath107 the scaling function is constant , i.e. @xmath108 , that is the coupling parameter can be determined from a single fit , giving @xmath109 for the [ couplingconst ] particular tls measured . the effect of fastly relaxing tls in point contacts can not be resolved as a telegraph fluctuation of the resistance , it causes an anomalous behavior in the voltage dependence of the differential resistance around zero bias , a so - called zero bias anomaly ( zba ) , which will be a subject of detailed discussion in this review . metallic systems , where two macroscopic electrodes are connected via a contact with small cross section are called `` point contacts '' ( pc ) in general , regardless of the actual size of the contact area . it can be generally stated , that the resistance of a point contact is mostly determined by the narrow neighborhood of the junction ; therefore , a pc acts as a `` microscope '' magnifying all kinds of phenomena occurring in the small contact region . + ( a ) ( b ) ( c ) ( d ) several methods have been worked out to produce extremely small contacts between two conducting leads . figure [ 4methods ] presents the four most important experimental techniques . the first one ( fig . [ 4methods]a ; ) is referred to as mechanically controllable break junction technique ( mcbj ) . here the sample ( practically a piece of metallic wire ) is fixed on the top of a flexible beam , and a small notch is established between the anchoring points . the contact is created in situ at low temperature by breaking the sample with bending the beam , thus one obtains clear and adjustable junction on atomic length scale . the second method ( fig . [ 4methods]b ; ) uses nanolithography to establish a small hole in a silicon nitride insulating membrane by etching . if the etching is stopped just when the hole breaks through , the diameter at the bottom edge remains extremely small ( @xmath110 3 nm ) , well below the usual resolution of lithography ( @xmath110 40 nm ) . in the next step metal is evaporated on both sides , creating a high - quality point contact device . this method can provide extremely stable and clear contacts on atomic length scale , however the contact diameter can not be varied during the measurement . the third drawing ( fig . [ 4methods]c ) shows a similar arrangement to a scanning - tunneling microscope , a so - called spear - anvil geometry : a vertically movable , sharply tapered needle is pressed onto a flat surface . finally , fig . [ 4methods]d shows a simple technique , where the edges of two wires are brought into contact . in arrangements ( a ) , ( c ) and ( d ) usually a differential screw mechanism is used to adjust the contact supplemented with a piezo - crystal for fine tuning . the first application of point contacts was carried out by igor yanson@xcite to investigate electron - phonon scattering in nanojunctions . he found , that the point contact spectrum , obtained as the second derivative of current with respect to the bias voltage ( @xmath111 ) contains structure due to the electron - phonon interaction described by the eliashberg function @xmath112.@xcite this simple method for measuring electron - phonon interaction spectrum became a popular application of pc spectroscopy , however , it can be used to probe other electron scattering processes as well , like electron - tls interaction , which is the central topic of this paper . theoretically , point - contacts are considered as two bulk electrodes connected through a narrow constriction . the simplest , and most commonly used pc model is presented in fig . [ pcmodels.fig]a . this so - called opening - like point contact is an orifice with diameter @xmath8 in an infinite isolating plane between the two electrodes . another extreme limit is the channel - like pc : a long , narrow neck between the bulk regions with the length being much larger than the diameter , @xmath113 ( fig . [ pcmodels.fig]b ) . the crossover between the both cases can be obtained by considering the point - contacts as rotational hyperboloids with different opening angles ( fig . [ pcmodels.fig]c ) . in most of the cases the shape of the pc does not influence the character of physical processes in the constriction radically , and the main parameter is the ratio of the contact diameter ( @xmath8 ) and other characteristic length scales in the system . three fundamental length scales are the mean free paths connected to different scattering processes ( @xmath114 ) ; the fermi wavelength of electrons ( @xmath115 ) ; and the atomic diameter ( @xmath116 ) . '' in an insulating screen between two metallic half - spaces . ( b ) two bulk regions are connected with a long , narrow conducting neck . ( c ) rotational hyperboloid.__,title="fig:",height=170 ] '' in an insulating screen between two metallic half - spaces . ( b ) two bulk regions are connected with a long , narrow conducting neck . ( c ) rotational hyperboloid.__,title="fig:",height=170 ] '' in an insulating screen between two metallic half - spaces . ( b ) two bulk regions are connected with a long , narrow conducting neck . ( c ) rotational hyperboloid.__,title="fig:",height=170 ] + ( a ) ( b ) ( c ) if the fermi wavelength or the atomic diameter becomes comparable to the contact size , we speak about _ quantum point contacts _ and _ atomic - sized point contacts _ , respectively . these systems are reviewed in detail in . in this paper we consider contacts , that are neither _ atomic _ nor _ quantum _ , but that are small enough compared to certain mean free paths , and the calculations are basically performed for an opening type geometry . concerning the mean free paths , we have to make difference between the elastic and inelastic scatterings . under usual experimental conditions the elastic mean free path ( @xmath117 ) is smaller than the inelastic one ( @xmath118 ) . here the inelastic mean free path is the _ length of the path _ an electron travels between two inelastic scatterings ( @xmath119 ) . mostly the important parameter is not @xmath118 , but the inelastic diffusive length , that is the average _ distance _ an electron can diffuse between two inelastic scatterings : @xmath120 . if the contact diameter is much smaller than any of the mean free paths @xmath121 , we speak about a _ ballistic _ contact . in this case the electron travels through the constriction without any scattering ( except for the reflection on the walls ) . on the other hand , if @xmath122 , the electron makes a diffusive motion in the contact , and accordingly we speak about the _ diffusive _ regime . at contact diameters exceeding the inelastic diffusive length ( @xmath123 ) the excess energy of the electrons is dissipated inside the constriction , which causes a considerable joule heating in the contact . this limit is called _ thermal _ regime . in the following subsections these different limits are discussed . in many cases the system is characterized by the knudsen number , @xmath124 , which was first introduced for the problem of the gas outflowing from a tank through a hole,@xcite but in the recent decades it has been used to characterize point contacts as well . first the diffusive contacts are treated , for which the electric potential ( @xmath125 ) can be determined by classical equations . ( for a general discussion see ) if the mean free path of the electrons is much shorter than the dimension of the contact ( @xmath126 ) then the current density , @xmath127 is given by ohm s law in terms of the electric field , @xmath128 or the electric potential , @xmath125 : @xmath129 where @xmath34 is the conductivity of the metal . furthermore due to the charge neutrality in metals the continuity equation holds : @xmath130 if the conductivity is considered to be constant , these two equations result in the laplace equation for the electric potential : @xmath131 in this phenomenological approach the scattering processes are included in the conductance , which is inversely proportional to the mean free path , @xmath132 . this treatment does not distinguish between elastic and inelastic electron scattering . in the further treatment of the structures in the dynamical conductance ( sec . [ scatteringinpc ] ) the inelastic scatterings play crucial role , therefore , instead of the phenomenological theory the kinetic equations must be used . here the solution of the laplace equation is presented in order to determine the linear resistance of a diffusive contact , and to determine the position dependence of the electric potential . furthermore , we note that the laplace equation is only applicable if @xmath34 is constant is space . this assumption is certainly not valid if the temperature of the contact neighborhood is inhomogeneous , which happens in the thermal regime . this situation is treated in the next subsection . here we assume that the inelastic diffusive length is much larger than the dimension of the contact , thus the temperature of the contact region is constant . the laplace equation should be solved with the boundary conditions : ( z ) . [ bound ] this problem was first solved by maxwell.@xcite due to the geometry of the problem the laplace equation is most easily solved in an elliptical coordinate system demonstrated in fig . [ elliptic_system ] ( see e.g. , ) . = 8.0 cm 0.7truecm the transformation between a traditional cylindrical coordinate system @xmath133 and the elliptical coordinates @xmath134 is determined by the following equations of the ellipses and hyperbolas : + z^2&=&1 + r^2-z^2&=&1 , [ ell_hyp ] where @xmath135 is the radius of the contact ; @xmath136 ; and @xmath137 . using eq . ( [ ell_hyp ] ) the old coordinates can be expressed in terms of the new ones : r = a , z = a . in this coordinate system the orifice is given by the @xmath138 surface , while the isolating layer is defined by @xmath139 . the next task is to determine the components of the metric tensor in the new coordinates : s^2&=&dr^2+dz^2+r^2d^2 + or , using @xmath9 and @xmath22 , s^2&=&a^2 ^ 2+^2d^2 + a^2 ^ 2+^2d^2 + a^2(1+^2)(1-^2)d^2 . the prefactors are the components of the metric tensor in the new ( also orthogonal ) coordinate system : g_=a^2 ^ 2+^2 , g_=a^2 ^ 2+^2 , g_=a^2(1+^2)(1-^2 ) . substituting these into the form of the laplace operator expressed by the general orthogonal coordinates : = 1 , one can get easily = 1 . taking the geometry of the problem into account it is obvious that = 0 . the boundary condition eq . ( [ bound ] ) can only be satisfied if at @xmath140 the potential is independent of @xmath22 , that is @xmath141 . as an ansatz we extend this condition to arbitrary @xmath9 values which means that the equipotential surfaces are the @xmath142 surfaces . as these surfaces are orthogonal to the @xmath143 plane it is obvious that electric field has no normal component at the insulating layer . after this consideration the laplace equation takes the following simple form = 0 . [ laplace_apert ] the solution of this equation with the boundary conditions is : ( ) = -v ( ) . [ phi_lambda ] along the @xmath144-axis ( @xmath145 ) , the potential is changing as ( z)=-v , [ eq31 ] which is plotted in fig . [ potent_fig ] . = 0.6 0.7truecm inside the opening ( @xmath146 ) the electric field has only @xmath144 component : e_z=-=-1|_=0=v1.[ez]it should be noted that the field is the largest at the edge of the opening ( @xmath147 ) and most of the current flows through that region . the total current flowing through the contact is : i=_0^a 2rdre_z(r)=2av , and thus the resistance of a diffusive point contact is : r=1 . this calculation was performed for an opening type contact . in case of a channel type geometry with @xmath113 the potential drops in the channel , and the resistance is easily written using ohm s law : @xmath148 if the dimension of the contact is larger than the inelastic mean free path , then the electron s excess energy is dissipated in the contact region , and joule heating takes place . in this case the contact center can be considerably overheated compared to the bath temperature , and thus the conductivity is also position dependent : @xmath149 . this phenomenon can be treated by considering the equations both for the electrical and the heat conduction . the electrical and thermal current densities are : @xmath150 where @xmath151 is the heat conductivity and @xmath152 is the position dependent temperature . the continuity equations for the electrical and thermal current are determined by the charge neutrality and the joule heating respectively : @xmath153 furthermore , we assume the validity of the the wiedemann - franz law : = , [ wiedemann ] where @xmath154 is the lorentz number . combining these equations we get : @xmath155 from these equations it can be easily seen that the temperature is generally related to the electric potential as : @xmath156 the constant term is determined by the boundary condition : @xmath157 thus : @xmath158 in an opening type contact the potential is zero at the contact surface , so the local temperature is written as : @xmath159 according to this relation a bias voltage of @xmath160mv already heats up a contact from cryogenic temperatures ( @xmath161k ) to room temperature . this feature has crucial importance in experimental studies . if the measurements are performed on a large contact or at high bias voltages , it can easily happen that the @xmath1 curve presents the temperature dependence of the conductance instead of the spectroscopic features,@xcite which will be discussed in sec . [ scatteringinpc ] . if the linear dimension of the contact is much smaller than the minimum of the mean free paths the electrons are travelling along straight trajectories , and in the first approximation no scatterings are taken into consideration . this system is described by the equations of semiclassical dynamics , so the space and momentum dependent distribution function , @xmath162 is to be determined beside the electric potential , @xmath163 . in this limit the electrons `` remember '' which side of the contact they are coming from , thus the distribution function at position @xmath164 can be expressed as a sum of the terms corresponding to electrons coming from the left or the right side of the contact.@xcite the distribution function can be determined by solving the homogeneous boltzmann equation which contains no collision integral:@xcite @xmath165 where @xmath128 is the electronic field , @xmath166 , and @xmath167 is the velocity of the electron with momentum @xmath168 . note that the electronic charge @xmath169 is _ positive _ , so the charge of an electron is @xmath170 . far from the constriction the boundary condition for the distribution function is that it has to tend to the equilibrium distribution function . @xmath171 where @xmath172 . if we apply @xmath11 voltage on the sample , the boundary condition for the potential @xmath125 takes the form @xmath173 the solution of the boltzmann equation can be found by using the trajectory method.@xcite ( the trajectories of the electrons are considered to be straight because we are interested in the limit @xmath174 and the current we are searching for , depends linearly on the electrical field . it can be shown that the bending of trajectory only contributes to the current in the higher order of the field . ) in this way @xmath175 holds combined with the charge neutrality condition @xmath176=0 , \label{charge}\ ] ] where @xmath177 denotes the integration over the trajectory of the electron coming from the distant reservoir to the contact region at point @xmath164 . the distribution function of an electron satisfying the conditions ( [ charge ] ) , ( [ trajectory ] ) and ( [ limit1 ] ) takes the form in the linear order of the field @xmath178 which is constant along the trajectory as the energy @xmath179 is conserved . the expression @xmath180 describes the different value of the chemical potential in the right and left hand sides of the constriction ( @xmath181 ) : @xmath182 ( for an illustration see fig . [ eta.fig ] . ) that can be also expressed in terms of the solid angle , @xmath183 at which the opening of the contact is seen from point @xmath164 : @xmath184 a simple visualization of the distribution function is presented in fig . [ distr_ball ] . = 0.25 0.7truecm = 0.7 0.7truecm the potential can be derived by substituting the form ( [ distrfunc ] ) of the distribution function into the neutrality condition ( [ charge]):@xcite 2e^3=0 , where the integral for equienergetic surfaces is easily expressed by the solid angle @xmath185 , and just an energy integral remains : = 0 , [ neutr2 ] where the upper and lower signs correspond to the cases @xmath186 and @xmath187 , respectively . the equation can be solved easily by using the identity ( f_0(+a)-f_0())d =- a[identity]for an arbitrary energy shift @xmath33 . we can get the potential in the whole space as:@xcite @xmath188 \mathrm{sign}(z ) . \label{potent_op}\ ] ] here @xmath185 is the solid angle at which the contact is seen from the position @xmath164 . the potential depending on the distance measured from the contact along the @xmath144 axis is shown in fig . [ potent_fig ] ( solid curve ) . the current density , @xmath127 is determined by the distribution function as follows : ( ) = -2e . [ currdens ] the current through the contact is calculated by integrating the @xmath144-component of the current density over the area of the contact , @xmath189 : @xmath190 at the contact surface half of the electrons go to the left with a distribution function @xmath191 and half of the electrons goes to the right with a distribution function @xmath192 , thus the current is written as i =- e_a da__=(f_0(+)-f_0(- ) ) , where the integral @xmath193 is taken over an equienergetic surface in the @xmath168-space . at low enough temperature and voltage ( @xmath194 and @xmath195 ) the expression @xmath196 equals @xmath197 according to the identity ( [ identity ] ) , so the current is written as : i=()_fsv,[sharvin]where @xmath189 is the area of the contact , @xmath198 is the area of the fermi surface , @xmath199 and @xmath200 denotes the average taken over the fermi surface . the conductance of a ballistic point - contact given by the formula ( [ sharvin ] ) is known as the sharvin conductance.@xcite for free electron gas the sharvin conductance is simplified as : g_s= ( ) ^2 . [ sharvinconductance ] in the intermediate region between the diffusive and ballistic regime an interpolating formula can be set up by numerically solving the boltzmann equation for arbitrary ratio of the contact diameter and mean free path , @xmath114:@xcite @xmath201 where @xmath202 is a numerically determinable monotonous function , @xmath203 ; @xmath204 . note that the first term is exactly the sharvin resistance by putting the drude resistivity into the formula , @xmath205 , thus it is actually independent of @xmath114 . this formula provides the possibility to estimate the contact diameter from the contact resistance . the voltage applied on a point contact results in a nonequilibrium distribution of the conduction electrons in the contact region . an electron coming from the right reservoir has an energy larger by @xmath197 than those coming from the left reservoir . this energy can be released through inelastic scattering processes , which can happen in such a way that an electron that has already crossed the contact is scattered back through the opening . this so - called backscattering reduces the current , thus the energy dependence of the scattering probability can be traced by measuring the nonlinearity of the current as the bias voltage is varied . this phenomenon was first used by igor yanson@xcite ( for a review see ) to study the phonon spectra and the electron - phonon interaction and since that it is widely applied . in this section we discuss the nonlinearities in the current voltage characteristics due to the scattering on slow tlss , which show strong similarities to localized phonons . there is , however , an essential difference between the two cases , while the phonons contribute to the voltage region up to @xmath206mev , the tlss manifest themselves below or even well below @xmath207mev due to their characteristic energies . the main contribution to the @xmath1 characteristics comes from the close region of the contact , therefore a microscopic process like a transition of one atom from one position to another in the contact region occurs as a significant measurable change in the current . considering the tlss the main advantage of the point contact spectroscopy is that in case of small enough contacts one can investigate just one or few scattering centers . the current through the contact can be derived by solving the classical stationary boltzmann equation @xmath208 where @xmath209 is the collision integral for the electron with momentum @xmath168 at position @xmath164 , which is assumed to be local in real space . in the case of bulk phonons the collision integral is @xmath210 \delta(\varepsilon_{{\z{p}}+{\z{q}}}-\varepsilon_{\z{p}}-\omega_{\z{q}}^\alpha)+ \nonumber\\ & { } & \left[f_{{\z{p}}-{\z{q}}}(1-f_{\z{p}})n_{\z{q}}^\alpha - f_{\z{p}}(1-f_{{\z{p}}-{\z{q}}})(n_{\z{q}}^\alpha+1)\right ] \delta(\varepsilon_{{\z{p}}-{\z{q}}}-\varepsilon_{\z{p}}-\omega_{\z{q}}^\alpha)\bigg\ } , \label{eq70}\end{aligned}\ ] ] where @xmath211 is proportional to the squared matrix element of the electron - phonon coupling , @xmath212 is the phonon occupation number for momentum @xmath213 and energy @xmath214 in phonon branch @xmath16 . there is another term of the collision integral which is due to the elastic ( impurity ) scattering but that is not given here . the collision integral vanishes far from the point contact because well inside the electrodes thermal equilibrium exists . the situation is further simplified when the following assumptions are valid : ( i ) the electron - phonon interaction is homogeneous in the real space ; ( ii ) the phonon distribution corresponds to thermal equilibrium . in that case the information on the contact geometry and also the elastic scattering due to impurities in the dirty limit can be expressed by a geometrical factor @xmath215 in the expression of the current which depends only on the momenta @xmath168 and @xmath216 of the incoming and outgoing electrons . that factor is frequently called as the @xmath53-factor ( see e.g. ) . in the case of phonons the @xmath53-factor for a ballistic opening like contact is written as : @xmath217 where @xmath218 @xmath219 are the unit vector parallel with @xmath168 and @xmath216 , while @xmath220 and @xmath221 are the @xmath144 components of these unit vectors , where the @xmath144 direction is the axis of the contact . the @xmath53-factor can be also calculated for a diffusive contact : @xmath222 where @xmath117 is the elastic mean free path , and @xmath8 is the contact diameter . at zero temperature the logarithmic derivative of the resistance can be expressed by a function @xmath223 , where @xmath224 is the density of states of phonons and @xmath225 is the modified electron - phonon coupling which contains the @xmath53-factor . without that factor it is just the electron - phonon coupling strength and in that case @xmath226 is the eliashberg function known in the theory of superconductivity . the main result is @xmath227 in case of single crystals @xmath226 depends also on the orientation of the crystal , which makes it possible to study the anisotropy of the phonon spectrum.@xcite measuring @xmath228 the phonon spectra were determined for many different cases and the structures in them were identified as the details of the spectra.@xcite those structures , however , are superimposed on a continuous background . until now , it is assumed that the phonons are in thermal equilibrium , which is the correct assumption as far as the phonons at the contact region arrive from a large distance without any collision where the thermal bulk distribution is realized . that is , however , not the case when in the contact region the phonon mean free path is comparable with the contact size . then the phonons generated by the non - equilibrium electrons in the contact region can not escape thus the phonons are also out of equilibrium . the measure of non - equilibrium is limited by the energy relaxation time of the phonons @xmath229 . the non - equilibrium localized phonons contribute to the background resistance.@xcite that can be studied by measuring the amplitude of the background as a function of the frequency of the applied voltage@xcite and at high frequency ( @xmath230 ) that is decreasing . these phenomena will be studied in the context of scattering on tlss as well . in the following the boltzmann equation will be solved considering only the electron - tls interaction . it will be assumed that the density of the tlss is low thus an electron scatters only once in the orifice region and the contribution of the different tlss are additive . first the ballistic limit is treated , where both the elastic and inelastic mean free paths are larger than the size of the contact . the collision term related to a single tls is very similar to the case of phonons , the momentum is , however , not conserved . furthermore , the tls has only two states with population @xmath231 and @xmath232 ( @xmath233 ) corresponding to the tls energy levels @xmath234 and @xmath235 ( see sec . [ tlsinsolids ] ) . the situation is different from those cases of phonons , where the phonon mean free path is large compared to the size of orifice , @xmath236 and the phonon distribution is the bulk one which is in thermal equilibrium with the contact regions . in the case of tlss the interaction is localized in space , thus the tlss in the contact region are decoupled from the bulk region and they can be considerably out of thermal equilibrium due to the voltage drop at the contact . therefore , the situation is similar to the special case of localized phonons with @xmath237 shortly discussed in the introductory remarks of this section . the collision term similar to eq . ( [ eq70 ] ) can be written as the sum of inelastic and elastic terms as @xmath238 . in the inelastic case the tls jumps to its other state due to the scattering , and thus the energy of the electron changes by @xmath239 . the strength of the interaction is characterized by @xmath240 , which is obtained from the interaction matrix elements ( eq . [ matrix_elements_inel ] ) by fermi s golden rule as @xmath241 the interaction strength is symmetric in the incoming and outgoing momentum , and in the neighborhood of the fermi surface it depends only on the unit vectors @xmath218 and @xmath219 , i.e.@xmath242 . the collision integral for the inelastic scattering is written as : @xmath243 where @xmath244 is the place of the tls . the current correction due to the elastic scattering processes can be written similarly with two remarks : ( i ) in an elastic process the energy of the tls does not change , thus @xmath245 must be inserted ; ( ii ) the scattering cross section is different for the tls being in the `` @xmath246 '' and the `` @xmath247 '' state . according to the matrix elements in eq . ( [ matrix_elements_elastic ] ) , the two scattering strengths are given as : @xmath248 after introducing these notations the correction to the collision integral for the elastic scattering is : @xmath249\delta(\varepsilon_{{\z{p}}}-\varepsilon_{{\z{p}}'})\label{irtlsel}.\end{aligned}\ ] ] the electron distribution function @xmath250 and the electric potential is expanded in terms of @xmath251 . at a large distance measured from the orifice the potential is constant and @xmath252 . thus , at @xmath253 the electrons are in thermal equilibrium . the distribution function @xmath250 and the potential @xmath254 can be expanded as @xmath255 and similarly the electric field @xmath256 where the upper indices @xmath257 , @xmath258 label the order in the strength of the electron - tls coupling . the zero order terms have been previously calculated and the following treatment is restricted to the first order terms . the previous results are @xmath259 and according to eq . ( [ potent_op ] ) @xmath260 for a round orifice perpendicular to the @xmath144-axis , where @xmath261 is the fermi function . the first order terms in the boltzmann equation given by eq . ( [ boltzmann_new ] ) can be written as @xmath262 where the label @xmath257 of the collision term indicates that the collisions are calculated with the distribution functions of zeroth order , @xmath263 . the change in the potential is determined again by the neutrality condition given by eq . ( [ charge ] ) for @xmath263 , which is @xmath264 the boundary conditions are @xmath265 and @xmath266 in order to determine @xmath267 and @xmath268 the trajectory method is used ( see e.g. ) dealing with phonons . electron trajectories are considered where the electron comes from the left or right and arrives to the plane of the orifice at time @xmath269 at point @xmath164 with momentum @xmath168 to calculate the current flowing through the orifice . these trajectories correspond to zeroth order . moving along the trajectory @xmath270 the distribution function varies as @xmath271 where @xmath272 , and @xmath273 . this equation can be integrated and using the boltzmann equation one finds for @xmath267 at @xmath269 @xmath274 here the collision term describes the adding or taking off electrons to the trajectory arriving at the contact at @xmath269 . in the expression of @xmath267 only the terms linear in the voltage are kept , thus @xmath275 in the above equation can be approximated by the zeroth order term @xmath276 . taking into account that @xmath277 the first term of the integral can be performed , thus @xmath278 the equation of electrical neutrality taken at @xmath269 with @xmath279 and @xmath280 combined with the equation above determines @xmath281 @xmath282 the general expression of the electrical current flowing through the surface of the contact is @xmath283 where @xmath284 is the surface element of the contact and the integral is taken over the orifice , the factor @xmath285 is due to the electron spin . the change @xmath286 in the total current due to the presence of the tls can be separated as @xmath287 and @xmath288 can be further split , whether the electron scattering is elastic or inelastic @xmath289 making use of the expression ( [ eq : f1.2 ] ) for @xmath267 only the collision term contributes to the current because the other term is even in the momentum . the integral according to the time can be transformed to the one along the path using @xmath290 where @xmath291 is the element of the path . the expression for the change of the current due to the presence of collision on tls is @xmath292 the path of the integral is changed due to the collisions but it may contain reflections by the insulating surface of the contact . the next step is to introduce the collision term due to a single tls at position @xmath244 , given by eq . ( [ irtlsin ] ) . as the paths arriving at the surface element of the orifice @xmath293 are straight lines , the volume element for the scattering event is @xmath294 where @xmath295 is the angle between the path and @xmath144 direction . now the integration over different paths and position of the collision gives @xmath296 where @xmath297 is defined by @xmath298 and the momentum integral is restricted to the solid angle @xmath299 at which the contact can be seen from the tls at position @xmath244 . furthermore , @xmath300 as the electron passes the contact from the tls ( for an illustration see fig . [ pctls1 ] ) . the final expression is obtained using @xmath301 where @xmath302 is the solid angle element in the momentum space and @xmath303 is the conduction electron density of states for one spin direction . assuming a spherical fermi surface @xmath304 and the angle @xmath295 is also indicated . _ _ ] using the expression of the collisison term @xmath305 ( [ irtlsin ] ) the change in the current due to the inelastic scattering can be written as @xmath306 \delta(\varepsilon+e-\varepsilon ' ) + \left[f_{\z{p}'}^0(1-f_{\z{p}}^0)n_+-f_{\z{p}}^0(1-f_{\z{p}'}^0)n_-\right ] \delta(\varepsilon - e-\varepsilon')\big\},\end{aligned}\ ] ] where the distribution functions @xmath307 are given by eq . ( [ eq : f0 ] ) and taken at @xmath308 , and the dimensionless notation @xmath309 is introduced . the potential @xmath310 appearing in these expressions can be dropped as the energy integral variables can be shifted . the essential contribution is given by @xmath311 , which has a positive sign if the electron at @xmath308 arrives from the left contact and a negative sign if it is arriving from the right contact . the final expression is @xmath312 \delta(\varepsilon+e-\varepsilon')+\nonumber\\ & & \big[f_0(\varepsilon'+{ev\over2}\eta(\z{p}'))(1-f_0(\varepsilon+{ev\over2}\eta(\z{p})))n_+- f_0(\varepsilon+{ev\over2}\eta(\z{p}))(1-f_0(\varepsilon'+{ev\over2}\eta(\z{p}')))n_- \big ] \delta(\varepsilon - e-\varepsilon')\bigg\}.\nonumber\\\end{aligned}\ ] ] this expression can be divided into two parts . if the electron with momentum @xmath216 arrives from the same side of the contact ( @xmath313 ) as the unscattered electron with momentum @xmath168 , thus @xmath314 , then the electron is scattered forward , while for @xmath315 it is scattered backward ( see fig . [ fig : paths ] ) . the total current can be divided according to that , thus @xmath316 signs are also indicated.,scaledwidth=70.0% ] the forward scattering cancels out . that can be seen by looking e.g. the first and the last term in the previous expression and changing @xmath317 , @xmath318 in the calculation of the backscattering the case @xmath319 is taken first . then the electron contributing to the current originally comes from the left ( @xmath320 ) and after the scattering goes from right to the left ( @xmath321 ) . @xmath322 the factor due to the angular integrals @xmath323 plays the role of the geometrical factor for the bulk phonons . only the backscattering contributes like in the case of phonons as it can be seen from eq . ( [ kfactorball ] ) for ballistic regions where the @xmath53-factor contains the factor @xmath324 . in the following the geometrical factor is introduced which depends on the position of the tls , @xmath244 , and it is independent of the strength of the interaction @xmath325 where @xmath326 if the momentum dependence of the interaction is ignored then @xmath327 and @xmath328 is a pure geometrical factor . further on this simplification is assumed . if the tls is positioned on the geometrical axis of a circular contact then the above integral can be evaluated as @xmath329 in the second part of the integral in eq . ( [ diin.right ] ) the variables are changed as @xmath317 and the integral with respect @xmath330 is performed , furthermore , in some of the integrals the variable is shifted as @xmath331 . the result is @xmath332n_-\nonumber\\ & + & \left[f_0(\varepsilon+e+ev)-f_0(\varepsilon+e - ev)\right]n_+ \big\}.\label{eq:112}\end{aligned}\ ] ] now the tls on the left is considered ( sign@xmath333 ) . then for backscattering the signs of @xmath334 and @xmath335 are reversed ( see fig . [ fig : paths]b ) that is equivalent of changing the sign of @xmath11 ( @xmath336 ) in eq . ( [ diin.right ] ) and then the expression in the curly bracket also changes its sign as it can be seen by inserting @xmath317 . that negative sign is cancelled by sign@xmath333 , thus the total current is unchanged . therefore , the current does not depend on whether the tls is on the right or left as it should be . after inserting the fermi functions and performing the energy integral , the following result is obtained for the inelastic current correction : @xmath337 where @xmath338- { e - ev\over { 2}}\bigg[\coth({e - ev\over { 2k_bt}})-1\bigg]\right\},\nonumber\\ d=\frac{2e}{\hbar}wk_{\z{r}}\left\ { { ev - e\over { 2}}\bigg[\coth({ev - e\over { 2k_bt}})-1\bigg]- { -e - ev\over { 2}}\bigg[\coth({-e - ev\over { 2k_bt}})-1\bigg]\right\}.\end{aligned}\ ] ] these expressions are simplified at @xmath62 as @xmath339 if the tls was in thermal equilibrium with the bath , at zero temperature @xmath340 and @xmath341 would hold . in that case the voltage dependence of the conductance is easily evaluated , it contains a jump - like decrease at the excitation energy of the tls : @xmath342 where @xmath343 . that result means if the tls is in the ground state @xmath341 then the tls can be excited only by a voltage @xmath344 , where the backscattering reduces the current . the second derivative has a dirac delta peak at the excitation energy @xmath345 this result is similar to the phonon result , that is the @xmath346 shows the structure of the excitation spectrum of a scatterer in the contact region . however , for a tls situated in the contact region the occupation of the upper level is not zero , as the system is driven out of equilibrium . due to this feature a nontrivial contribution occurs in the background resistance beside the spectroscopic feature at the excitation energy . this can be evaluated after calculating the voltage dependence of the occupation number , @xmath347 , which is performed later in this section . a similar calculation can be performed for the current correction due to the elastic scattering on tlss by using the collision integral @xmath348 ( [ irtlsel ] ) instead of @xmath349 ( [ irtlsin ] ) . the formulas for the elastic scattering can be derived easily by inserting @xmath245 and modifying the scattering strength in the results for the inelastic case . for elastic scattering the strength of the interaction can be different for the tls being in the `` @xmath246 '' and `` @xmath247 '' state , thus in eq . ( [ cd ] ) @xmath350 must be replaced by @xmath351 ^ -$ ] and @xmath351^+$ ] in the expression for @xmath352 and @xmath353 , respectively . after these considerations the result is : @xmath354 ^ -+n_+\cdot[wk_\z{r}]^+\right).\ ] ] note , that in the elastic case the voltage dependent `` @xmath355 '' terms cancel out , thus the only voltage dependence comes from the voltage dependence of the occupation number . if the tls is in thermal equilibrium with the bath , then @xmath356 holds , thus the elastic current correction is a voltage independent constant term . similarly , if the scattering cross sections for the two states of the tls are equal , then the voltage dependence cancels out due to the condition @xmath233 . in the following we use a simplified notation for the elastic current correction : @xmath357 where @xmath358 and @xmath359 are considered as the reduction of the conductance due to the tls being in the `` @xmath246 '' and the `` @xmath247 '' state , respectively.@xcite the average change in the conductance due to the elastic scattering is characterized by @xmath360 , which is written as follows according to formulas ( [ diel ] , [ kr ] , [ kisw ] , [ wel ] ) : @xmath361 ^ 2+\left[(\mu^2-\nu^2)v^z\right]^2\right),\ ] ] where @xmath45 and @xmath43 are the the interaction matrix elements from eq . ( [ int_slowtls ] ) assuming isotropic scattering . the difference between the conductances corresponding to the two states of the tls are written as : @xmath362 again , the voltage dependence of the current correction can only be determined after calculating the occupation number as a function of bias voltage , which is done in the following part . it is important to note that in case @xmath363 the scattering strength is symmetric for the two states , thus @xmath364 . in the following @xmath231 will be calculated in the nonequilibrium case , where it is determined by the interaction with the nonequilibrium electrons . the transition probability from the `` @xmath247 '' to the `` @xmath246 '' state due to an electron scattered from momentum state @xmath168 to @xmath216 has been considered in the collision term ( [ irtlsin ] ) and it is : @xmath365 and similarly : @xmath366 the transition probabilities for the tls are obtained after integrating for the electron momenta : @xmath367 where the expression ( [ eq : f0 ] ) for @xmath368 is used . the kinetic equation for @xmath231 is : @xmath369 in the fermi functions the electron momentum is only involved in @xmath334 , which tells which reservoir the electron at position @xmath244 with momentum @xmath168 is coming from . therefore the integral can be performed by separating the regions of integration into four cases with respect to the possible values of @xmath334 and @xmath335 . in all of these four regions the factor containing the fermi functions is independent of @xmath370 and @xmath371 . there is an essential simplification if @xmath372 is replaced by a momentum independent averaged one , @xmath373 . take e.g. the case where the tls is on the right hand side of the contact ( @xmath374 ) and introduce the solid angle @xmath299 at which the opening can be seen from point @xmath244 . in this case the integrals for the different regions are written as : @xmath375 this simplification is justified only after taking an average over large number of tlss but that simplifies the calculation and makes the result more transparent . in this simplified case the kinetic equation for the tls is obtained after calculating the energy integrals for the different cases . ( as before , @xmath376 is eliminated by shifting the integral variables . ) the final result is independent whether the tls is on the right or left , and it can be written in the form : @xmath377 where the coefficients @xmath189 and @xmath72 are : @xmath378+\nonumber\\ + \left ( \left(1-\frac{\omega_\z{r}}{4\pi}\right)^2 + \left ( \frac{\omega_\z{r}}{4\pi}\right)^2\right ) \frac{e}{2 } \left[\coth(\frac{e}{2k_bt})-1\right ] \bigg\ } , \nonumber\\ b={w\over\hbar}\bigg\{\left({1-{\omega_\z{r}\over { 4\pi}}}\right){\omega_\z{r}\over { 4\pi}}\left[{{ev - e\over{2}}\left({\coth ( { ev - e\over{2k_bt}})-1}\right)+ { -ev - e\over{2}}\left({\coth ( { -ev - e\over{2k_bt}})-1}\right)}\right]+\nonumber\\ + \left({\left({1-{\omega_\z{r}\over { 4\pi}}}\right)^2 + \left({{\omega_\z{r}\over{4\pi}}}\right)^2}\right ) { -e\over 2}\left[{\coth ( { -e\over{2k_bt}})-1}\right]\bigg\},\end{aligned}\ ] ] respectively . the relaxations are of korringa types , which depend on @xmath11 only if the two electrons involved have different chareacters in terms of @xmath334 . the stationary value of @xmath231 is obtained as @xmath379 which is an even function of @xmath11 . at @xmath62 this expression simplifies essentially : @xmath380 where @xmath381 . @xmath231 is a continuous function of @xmath11 . far from the opening @xmath382 thus @xmath383 at voltages for which @xmath384 as the electron gas is in thermal equilibrium . for @xmath385 and @xmath386 @xmath387 as for large enough voltage the levels are equally occupied @xmath388 . if the tls is in the middle of the contact , then @xmath389 . if the tls is far away from the contact then it is in thermal equilibrium , which can be obtained by taking @xmath390 and the deviation from equilibrium is large when the tls is in the middle of the contact region @xmath391 . here other relaxation mechanism different from the scattering of electrons is not taken into account as the generation of bulk phonons is very weak as the relevant phase space is very small at low energies . at large concentration of tlss the tlss are interacting and the collective effects may modify the stationary values of the occupation numbers . now the inelastic and elastic contributions to the current and the conductance are calculated using the stationary occupation numbers obtained . the expressions will be given when the tls is just in the middle of the round opening , thus @xmath391 ( @xmath392 ) , but the calculation for arbitrary @xmath393 can be easily performed as well . at zero temperature the inelastic current correction is obtained by inserting the stationary value of @xmath231 given by eq . ( [ nt0 ] ) into ( [ iinel2 ] ) : @xmath394 which is a continuous function of @xmath11 . after differentiation with respect to the voltage the correction to the conductance is obtained as : @xmath395 and the second derivative of the current is : @xmath396 at positive voltages the second derivative of the current shows a negative dirac delta peak at @xmath397 , which reflects the energy spectrum of a single tls . above the excitation energy a positive background is seen due to the nonequilibrium distribution of the tls occupation number . as a next step the elastic contribution is determined at @xmath62 for a single tls positioned in the middle of the contact . the change in the conductance due to the elastic scattering can be calculated by differentiating eq . ( [ iel ] ) . the elastic current correction contains a linear term , which causes a constant , voltage independent reduction of the conductance . experimentally it is hard to separate this constant contribution , thus we calculate only the voltage dependent part by subtracting the zero bias conductance : @xmath398 the second derivative of the @xmath1 curve is @xmath399 again , a dirac delta peak reflects the spectrum of the tls , and a continuous background arises at @xmath344 due to the nonequilibrium distribution . contrary to the inelastic case , the dirac delta peak and the background have the same sign ; furthermore , in the elastic case the sign of the peak can either be positive or negative depending on the sign of @xmath400 . the amplitude @xmath400 is related to the universal conductance fluctuation @xmath401 ( see e.g. ) . for arbitrary position of the tls the occupation number @xmath231 is zero at @xmath402 and @xmath403 at @xmath404 . therefore , the total amplitudes for the change in the conductance in the elastic and inelastic case can be generally written as : @xmath405 where isotropic scattering was assumed , and the formulas ( [ scatteringstrength ] , [ kisw ] , [ kr ] , [ dgamma ] ) were used . the results are given in the unit of the universal conductance quantum , @xmath401 . according to eq . ( [ munu ] ) for a highly asymmetric tls , where the energy splitting @xmath5 is much larger than the transition term @xmath6 , the equations @xmath406 and @xmath407 hold . in this case the inelastic term is suppressed . in the opposite case , where @xmath408 the elastic term is suppressed as @xmath409 . it must be also noted that the inelastic term depends only on the matrix element @xmath43 , whereas the elastic term is influenced both by @xmath43 and @xmath45 . the inelastic term can be roughly estimated as @xmath410 using @xmath411 , @xmath412 , @xmath413 ( see ) . the amplitude sharply drops by moving the tls further from the orifice than its diameter because of the geometrical factor . the estimation of the elastic term is more difficult as @xmath45 is strongly model dependent . in the following the effect of a tls is considered at finite temperature . the formulas for the conductance and the second derivative of the @xmath1 curve can be explicitly calculated using equations ( [ iinel ] , [ cd ] , [ diel ] , [ ab ] , [ nplus ] ) ; however , these equations are very complicated , thus the results are demonstrated by figures . figure [ ball_cent ] shows the voltage dependence of the conductance for the inelastic and elastic case respectively . in fig . [ secder_ball ] the second derivative of the @xmath1 curve , the so - called point contact spectrum is presented at various temperatures both for the elastic and inelastic case . in fig . [ secder_ballq ] both the inelastic and elastic contributions are compared for a tls positioned in the contact center ( @xmath391 ) and for a tls being farther away ( @xmath414 ) . and @xmath415 , respectively ) . the temperature is @xmath416 . the sign of the elastic contribution can be either positive or negative depending on the sign of ( @xmath400 ) . on the other hand , the inelastic scattering always causes a decrease in the conductance.__,title="fig:",scaledwidth=50.0% ] and @xmath415 , respectively ) . the temperature is @xmath416 . the sign of the elastic contribution can be either positive or negative depending on the sign of ( @xmath400 ) . on the other hand , the inelastic scattering always causes a decrease in the conductance.__,title="fig:",scaledwidth=50.0% ] curve for inelastic and elastic scattering at various temperatures for a ballistic point contact ( @xmath417 , and @xmath418 , respectively ) . the inelastic contribution shows a negative peak at the excitation energy of the tls and a positive background tail at higher voltages . the elastic contributions similarly shows a peak at @xmath397 and a background tail at higher voltages , but here the sign of the peak and the tail is the same . in the elastic case the sign of the signal can either be positive or negative depending on the sign of ( @xmath400).__,title="fig:",scaledwidth=50.0% ] curve for inelastic and elastic scattering at various temperatures for a ballistic point contact ( @xmath417 , and @xmath418 , respectively ) . the inelastic contribution shows a negative peak at the excitation energy of the tls and a positive background tail at higher voltages . the elastic contributions similarly shows a peak at @xmath397 and a background tail at higher voltages , but here the sign of the peak and the tail is the same . in the elastic case the sign of the signal can either be positive or negative depending on the sign of ( @xmath400).__,title="fig:",scaledwidth=50.0% ] curve for inelastic and elastic scattering for two different positions of the tls : in one case the tls is in the contact center , in the other case it is farther away on the contact axis , so that the opening is seen in a solid angle @xmath419 . the tls being farther away from the contact is closer to thermal equilibrium , which is reflected by the reduced amplitude of the background tail with respect to the spectroscopic peak in the inelastic contribution.__,title="fig:",scaledwidth=50.0% ] curve for inelastic and elastic scattering for two different positions of the tls : in one case the tls is in the contact center , in the other case it is farther away on the contact axis , so that the opening is seen in a solid angle @xmath419 . the tls being farther away from the contact is closer to thermal equilibrium , which is reflected by the reduced amplitude of the background tail with respect to the spectroscopic peak in the inelastic contribution.__,title="fig:",scaledwidth=50.0% ] as it has been discussed in sec . [ maxwell ] the phenomenological treatment must be replaced by a theory based on kinetic equation , where effect of the inelastic scatterings are taken into account in the distribution function . in the following the elastic and inelastic scatterings are treated in equal footing . as a first step the distribution function is determined in the presence of elastic scattering and next the contributions of the tls are treated as a perturbation . consider first the elastic scattering in the limit @xmath420 , where the resistance is very large . because of the very strong elastic scattering , the electrons are immediately redistributed concerning the direction of their momenta . therefore , in the limit @xmath421 the distribution function depends only on the energy of the electrons , @xmath422 . as the electron arriving at the contact is either coming from the left or the right reservoirs and the electron energies are changed due to the external potential @xmath254 the distribution function must be the superposition of the distribution of the electrons coming from the left or right with amplitudes @xmath423 and @xmath424 , respectively : f_()=_0()f_0(_-e ( ) + ev2)+(1-_0())f_0(_- e()-ev2 ) . [ diffdistr ] using the charge neutrality condition ( @xmath425 @xmath423 can be determined by a similar treatment used earlier in case of eq . ( [ potent_op ] ) : _ 0()=12+()v . [ alpha ] this formula is valid in case of arbitrary geometry because only the shape of the potential function contains information about the details of the geometry . for an opening type circular contact the potential was determined in sec . [ maxwell ] by solving the laplace equation using a hyperbolic coordinate system . = 0.6 0.7truecm the distribution function contains two sharp steps due to the two different fermi energies in the two reservoirs ( see fig . [ distrmax ] ) . such steps are measured by the saclay group studying short metallic wires of type fig . [ pcmodels.fig]b by attaching an extra tunneling diode on the side of the wire . ( see ) . in the ballistic limit the distribution function is very similar only the factors @xmath423 and @xmath424 are replaced by the geometrical factors @xmath426 and @xmath427 determining the solid angles in which the ballistic electron arrives from the left or right reservoirs to the point @xmath164 . the electron distribution function ( [ diffdistr ] ) obtained for @xmath421 does not result in any current in agreement that the resistivity is infinite in this limit . for finite @xmath428 the factor @xmath423 must be replaced by a momentum dependent one @xmath429 , and the current is due to @xmath430 , which has a strong asymmetric momentum dependence and @xmath431 as @xmath421 . the @xmath432 will be determined by the very elegant theory of kulik , shekter and shkorbatov,@xcite who have extended the treatment of the ballistic limit to the diffusive one treating the scattering on phonons and defects which are large in space . the strong elastic scattering on defects is combined with a weak inelastic scattering . the limit considered is specified as @xmath433 where @xmath8 is the size of the contact and @xmath117 and @xmath118 are the elastic and inelastic mean free paths of the electrons , respectively . introducing the corresponding relaxation times , @xmath434 and @xmath435 the inequality can be rewritten as @xmath436 . that means that an electron can diffuse out from the contact region of size @xmath8 with small probability of inelastic scattering , thus double inelastic scatterings can be neglected . in the following @xmath435 is related to the relaxation time due to tlss . also in the diffusive limit the current correction due to the tlss is derived by solving the boltzman equation ( [ boltzmann_new ] ) with an appropriate collision term . in this case the collision integral @xmath209 has two contributions as the elastic impurity part @xmath437 and the contribution of individual tlss at position @xmath244 , @xmath438 thus @xmath439 where the sum is due to different tlss , but in the following it is assumed that the concentration of the tlss is so low that first a single tls at position @xmath244 is considered and in the final result the contributions of the different tlss are additive . the kinetic equation can be arranged as @xmath440 first the left hand side is treated in the diffusive limit and the right hand side is considered as a weak perturbation . the impurity part of the collision integral is @xmath441 where @xmath442 is the elastic transition probability , @xmath443 is the velocity at momentum @xmath216 perpendicular to the equienergetic surface for which the integral is performed . the electric field can be expressed by the electric potential @xmath254 as @xmath444 , where @xmath125 is determined by the neutrality condition as in the ballistic case . the distribution function @xmath445 satisfies the following boundary condition : at the boundary , @xmath446 between the insulator and the metal we assume mirror reflection @xmath447 where an incoming electron with momentum @xmath168 is reflected with the momentum @xmath448 . furthermore , very far from the contact @xmath313 the equilibrium distribution is recovered with chemical potential @xmath22 , thus @xmath449 and @xmath22 satisfies the charge neutrality . the distribution function in the absence of the collision term due to tls is denoted by @xmath450 and the first order correction due to the tls is @xmath451 , thus @xmath452 where the higher order corrections are neglected . the kinetic equation for @xmath450 contains the static impurity contributions , thus @xmath453 and @xmath451 satisfies the equation @xmath454 where the electric field is also expanded as @xmath455 . the term @xmath456 is linearized in the collision , thus @xmath445 is replaced by @xmath450 . similarly , the current is @xmath457 , where @xmath458 and the integral with respect @xmath459 is taken on a dividing surface representing the point contact . according to the introductory remarks of this section , at finite elastic mean free path the distribution function can be expressed with a momentum dependent parameter , @xmath429 : @xmath460 where the electrons arriving from far left ( right ) at the contact have distribution function @xmath461 ( @xmath462 ) , thus @xmath463 as the collision term ( [ eq : i_i ] ) is linear in the distribution function the equation for @xmath450 is satisfied with @xmath429 as well : @xmath464 in the @xmath421 limit the solution of this equation is @xmath423 , given by ( [ alpha ] ) . at finite elastic mean free path @xmath429 can be expanded as : @xmath465 where @xmath432 is the first order term in the small parameter , @xmath466 . the momentum dependent correction , @xmath467 can be determined by a simple argumentation . at non - zero relaxation time the probability @xmath468 can be considered as the momentum independent probability @xmath469 taken at the position of the last elastic collision , that is : @xmath470 the expansion of this formula in @xmath434 gives : @xmath471 the value of @xmath469 is determined by the potential , which drops in the contact region , thus @xmath472 , and @xmath473 . the value of @xmath467 can also be obtained by inserting ( [ deltaalpha ] ) into the boltzmann equation for @xmath468 ( [ eq : alpha_kineq ] ) . after neglecting the higher order terms in the small parameters @xmath474 and @xmath475 the following simple formula is achieved : @xmath476 for isotropic scattering @xmath477 holds , the relaxation time approximation is appropriate , thus the collision integral is expressed as : @xmath478 which provides the relaxation time , @xmath434 due to the elastic scattering by the impurities . inserting this result into ( [ boltzalpha ] ) the formula ( [ deltaalpha2 ] ) for @xmath467 is regained . the next step is to determine @xmath479 using the equation ( [ boltzalpha ] ) for @xmath468 . the boundary condition for @xmath480 as very left the electron comes from the left contact where the electrons are in equilibrium thus @xmath481 furthermore , the collision term contributes to the distribution function only if the system is out of equilibrium , thus @xmath479 is proportional to the electric field . therefore , the second term @xmath482 in equation ( [ eq : kineq_sep ] ) can be dropped as it is proportional to @xmath483 . the correction to the kinetic equation is determined by eq . ( [ eq : kineq_sep ] ) where after linearizing each term in the field the ansatz with the function @xmath484 @xmath485 can be used . the terms proportional to @xmath486 cancel out in eq . ( [ eq : kineq_sep ] ) . thus @xmath487 with the boundary condition @xmath488 and @xmath489 . as it will be shown below the change in the distribution due to the collision term containing the interaction with the tls can be expressed as linear response with the aid of a green s function @xmath490 : @xmath491 where @xmath492 and @xmath493 with the boundary condition @xmath494 and @xmath495 . as @xmath490 is closely related to the correlation function @xmath496 the time reversal symmetry can be applied as @xmath497 the function @xmath498 is defined as @xmath499 where the integral with respect to @xmath500 is a surface integral of variable @xmath501 taken on the orifice . this function satisfies the equation @xmath502 which follows from the equation ( [ chipr2 ] ) defining @xmath490 . the boundary conditions are the usual ones @xmath503 and @xmath504 . this equation for @xmath498 is closely related to eq . ( [ boltzalpha ] ) for @xmath429 introduced earlier , but without the inhomogeneous term . ( the sign in front of the collision integral can be changed by replacing @xmath505 and taking the relation @xmath506 into account . ) by comparing the two equations the expression @xmath507 is obtained , which automatically satisfies the boundary condition for @xmath498 . using the expression ( [ eq : galpha ] ) for @xmath498 the current can be calculated and only the @xmath168-dependent part is contributing , as the total number of electrons at position @xmath164 is not changed by the collision . therefore : @xmath508 the kinetic equation ( [ eq : chi_kineq ] ) for @xmath509 can be checked using equations ( [ chipr1 ] ) , ( [ chipr2 ] ) , ( [ chipr3 ] ) and inserting twice the identity ( [ gpp ] ) . the next task is to calculate the correction to the current due to the inelastic scattering on the tls . using the expression ( [ chi ] ) for @xmath479 ( dropping the first term which is even in momentum @xmath168 in leading order ) : @xmath510 where the variable @xmath164 is taken over the contact area , @xmath511 . inserting the expression of @xmath512 the result can be rewritten as @xmath513 that result can be further rewritten in terms of @xmath498 as @xmath514 using eq . ( [ eq : galpha ] ) for @xmath498 the momentum independent terms do not contribute due to eq . ( [ qwerty ] ) . the remaining term contains @xmath467 which was determined earlier and @xmath515 holds . furthermore , @xmath516 and in this way the final expression for the current is @xmath517 where @xmath467 is taken at the position of the tls , @xmath244 . because of the odd parity of @xmath467 only the odd part of the collision integral contributes . using the equation ( [ falpha ] ) for @xmath450 the collision integral for the inelastic scattering can be written similarly to the ballistic case ( eq . [ irtlsin ] ) : @xmath518 \left[\alpha_{\hat{\z{p}}}(1-f_0(\varepsilon+{ev\over2 } ) ) + ( 1-\alpha_{\hat{\z{p}}})(1-f_0(\varepsilon-{ev\over2}))\right ] \right.\cdot\nonumber\\ & { } & \;\;\;\bigg[n_-\delta(\varepsilon'-\varepsilon - e)+ n_+\delta(\varepsilon'-\varepsilon+e)\bigg]\nonumber\\ & { } & \;-\bigg[\quad\hat{\z{p}}'\leftrightarrow\hat{\z{p}}\quad \varepsilon'\leftrightarrow\varepsilon\quad\bigg ] \bigg[\quad\hat{\z{p}}'\leftrightarrow\hat{\z{p}}\quad \varepsilon'\leftrightarrow\varepsilon\quad\bigg ] \bigg[\quad \varepsilon'\leftrightarrow\varepsilon\quad\bigg ] \bigg\}\end{aligned}\ ] ] is obtained , where in the second set of brackets the variables are changed as @xmath317 , @xmath519 and all the energy variables @xmath520 and @xmath330 are shifted by @xmath521 thus @xmath254 drops out . now using eq . ( [ deltaalpha ] ) @xmath522 and keeping only the odd part of the collision integral , after changing the variables the current is @xmath523 \left[\alpha_0(1-f_0(\varepsilon+{ev\over2}))+ ( 1-\alpha_0)(1-f_0(\varepsilon-{ev\over2}))\right ] \right.\right.\nonumber\\ & { } & \;\;\;-\delta\alpha_{\hat{\z{p } } } \left[\alpha_0f_0(\varepsilon'+{ev\over2}))+ ( 1-\alpha_0)f_0(\varepsilon'-{ev\over2})\right ] \left[f_0(\varepsilon+{ev\over2})-f_0(\varepsilon-{ev\over2})\right ] \bigg]\times\nonumber\\ & { } & \;\;\;\left[n_-\delta(\varepsilon'-\varepsilon - e)+ n_+\delta(\varepsilon'-\varepsilon+e)\right]\nonumber\\ & { } & \;\;-\bigg[\quad\hat{\z{p}}'\leftrightarrow\hat{\z{p}}\quad \varepsilon'\leftrightarrow\varepsilon\quad\bigg]\bigg\},\end{aligned}\ ] ] where @xmath469 and @xmath467 are taken at the site of the tls ( @xmath308 ) . in this expression two squared averaged matrix elements occur @xmath524 where the expression ( [ deltaalpha2 ] ) for @xmath467 was used . furthermore @xmath525 holds as @xmath526 and @xmath527 integrating over the variable @xmath330 the following expression is obtained @xmath528\right\},\nonumber\end{aligned}\ ] ] where the geometrical factor @xmath529 is introduced for the diffusive limit as @xmath530 the expression ( [ idiff ] ) can be brought to a form which is comparable to the result ( [ eq:112 ] ) in the ballistic limit . the key of the algebra applied is in the integral @xmath531 the variable can be shifted by an arbitrary energy @xmath532 as the contribution comes from the neighborhood of the fermi energy and the bandwidth cutoff does not play any role . the result of a lengthy algebra is i^(1)=-wk_^diff f_0()\{(f_0(-e - ev)- f_0(-e+ev))n_-+(f_0(+e - ev)- f_0(+e+ev))n_+}.this expression coincides with the one obtained in the ballistic limit ( [ eq:112 ] ) after replacing the geometrical factor @xmath533 . therefore , the final results in the ballistic limit ( [ iinel ] , [ cd ] , [ iinel2 ] ) are also valid in the diffusive limit , just the geometrical factor , @xmath534 is different . the elastic contribution can be obtained by inserting @xmath245 and considering different scattering strengths for the two states of the tls in the diffusive limit as well . therefore , the ballistic result for the elastic case ( [ diel ] ) is also valid in the diffusive regime with the replaced geometrical factor . the next step is to calculate @xmath231 . in the present case whether the electrons come from the left or right reservoir is determined by @xmath469 or @xmath535 , thus the previous results in the ballistic regime ( [ nplus ] , [ ab ] , [ nt0 ] ) can be adopted after changing @xmath536 the momentum dependent correction @xmath467 to @xmath469 contributes only if the momentum dependence of @xmath537 is kept but that is small in case @xmath538 . concluding , in the diffusive limit the shape of the current correction due to the scattering on a tls is the same as in the ballistic case , the only essential difference is in the geometrical factors , which are independent of the strength of the interaction between the electrons and the tlss . in the ballistic case @xmath534 is in the range of unity if the tls is situated in the contact region . however , @xmath529 contains other quantities proportional to @xmath539 which indicates how much @xmath16 is changed in space on the scale of the impurity scattering mean free path . we have seen that @xmath540 scales with the inverse of the contact diameter , @xmath8 , thus @xmath541 in case where the tls is nearby the orifice . far from the orifice ( @xmath313 ) @xmath542 . the amplitudes of the change in the conductance for the important tls ( being close to the contact ) are scaled down by @xmath543 which can be a very small factor in the dirty limit . the largest contribution arises from the tlss for which @xmath544 is the largest , thus for the tls just nearby the edge of the orifice . ( see eq . [ ez ] ) the change of the geometrical factor due to the impurity scattering length is similar to the case of electron - phonon interaction , where integration is taken according to location of the interaction.@xcite the @xmath53-factors for electron - phonon scattering are given in ( [ kfactorball ] , [ kfactordiff ] ) for the ballistic and diffusive case , respectively . in the phonon case the amplitude of the signal in the diffusive limit is reduced only by @xmath545 . up to this point the scattering on tlss in point contacts was investigated as a function of bias voltage , which provides information about the excitation spectrum of the tls , and also about the nonequilibrium distribution of the tls states . a further possibility to study the dynamical properties of the tlss is the study of the response to a harmonic excitation . let us consider that a voltage of @xmath546 is applied on the point contact , where the dc bias voltage , @xmath547 is varied on the scale of the excitation energy of the tls to cover the spectroscopic peak at @xmath548 , and also the background tail at @xmath344 . the amplitude of the harmonic excitation is considered to be much smaller than the excitation energy , @xmath549 . furthermore , it is assumed that the energy @xmath550 is much smaller than the excitation energy of the tls , thus the quantum nature of the radiation should not be taken into account . in this case the current flowing through the contact can be expanded in second order in terms of the time dependent voltage contribution : @xmath551 where the equation @xmath552 is used . the voltage dependence of the harmonic response , @xmath553 provides the differential resistance curve , @xmath554 . the second harmonic response , @xmath555 directly provides the point contact spectrum , i.e. the second derivative of the @xmath1 curve . the dc component , which can be measured as a time average of the response signal : @xmath556 also shows the second derivative of the @xmath1 curve after subtracting the current without harmonic excitation , @xmath557 . further on this term is referred to as the dc shift signal , denoted by @xmath558 . the above considerations are , however , only valid if the system is fast enough to follow the time dependent voltage . in the case of two level systems , the dynamical behavior is determined by the kinetic equation ( [ dndt ] ) : @xmath559 where the coefficients @xmath189 and @xmath72 are given by eq . ( [ ab ] ) . both of the coefficients depend on the actual voltage @xmath560 at given @xmath561 . for simplicity the notation @xmath562 is introduced , which is the inverse of the relaxation rate of the tls ( @xmath563 ) . for slow two level systems the value of the relaxation time can be estimated using eq . ( [ dndt],[kisw],[scatteringstrength ] ) as follows : |b ( 2)^2(_0 v^z)^2[relaxtime]inserting @xmath412 , @xmath564mev , and @xmath413@xcite into this equation a relaxation time of @xmath565s is obtained . if the alternating voltage is faster than the relaxation of the tls ( @xmath566 ) then a more complicated procedure is necessary to determine the time dependent response of the system . this phenomenon , however can be used to trace the typical relaxation time of the tls by studying the response signal as a function of the frequency @xmath68 . the first calculation with time dependent applied voltage was performed by kozub and kulik.@xcite in the next part a general solution is given for the time dependent occupation number @xmath567 at arbitrary frequency and amplitude of the harmonic excitation signal . after that various limits are treated , where the formulas are essentially simplified . the kinetic equation ( [ mw1 ] ) is a linear inhomogeneous differential equation which has a general solution @xmath568 where the coefficient @xmath569 is determined by the initial condition at time @xmath570 . according to the definitions given by eq . ( [ ab ] ) @xmath571 and @xmath572 hold , thus the exponent in the last term of eq . ( [ mw2 ] ) monotonically increases with time . this means that @xmath569 is essential only for the transient behavior and for long time it can be dropped . for periodic voltage @xmath189 and @xmath573 are also a periodic function of time : @xmath574 , @xmath575 . the time integrals can be divided according to the periods as : @xmath576 where @xmath577 , @xmath578 is an integer and @xmath579 . in this way : @xmath580 where @xmath581 and @xmath582 using these identities , one gets : @xmath583 the final result is : @xmath584 which gives the following stationary solution in the long time @xmath585 limit : @xmath586 turning to dimensionless variables like @xmath587 , and @xmath588 and also new notations , @xmath589 , @xmath590 , @xmath591 and @xmath592 one gets : @xmath593 it is important to note that @xmath594 is proportional to the relaxation rate @xmath595 and @xmath596 to the excitation rate of the upper level , which are independent of the frequency @xmath68 . in the following two simple limits are treated : @xmath597 ( @xmath598 ) ; and @xmath599 ( @xmath600 ) . [ [ low - frequency - limit - omegataull-1 ] ] low frequency limit ( @xmath601 ) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + assuming that @xmath602 ( @xmath603 ) the first part of the right hand side of eq . ( [ mw10 ] ) disappears as @xmath604 . ( note that @xmath594 is a monotonous function . ) the contribution to the second part also disappears as far as @xmath605 is finite . in the vicinity of the upper limit of the integral ( @xmath606 ) , however , @xmath607 can be expanded as @xmath608 , where @xmath609 , thus @xmath610 where the term arising from the lower limit is dropped as it is exponentionally decreasing with @xmath611 . the final result for the limit @xmath611 is : @xmath612 that result also holds for @xmath613 , where the condition @xmath614 is not satisfied any more . without making a rigorous proof , it can be argued , that @xmath613 is at the beginning of the time period @xmath615 . in the long time limit ( @xmath616 ) the time can be shifted , therefore , that time is not special , thus the result ( [ mw11b ] ) must be generally valid . in this way the static result is recovered with @xmath189 and @xmath617 taken with the actual voltage at time @xmath561 . therefore , at arbitrary time the system is in the stationary state , and the the expansion of the current in ( [ iexpansion ] ) is valid . [ [ high - frequency - limit - omegataugg-1 ] ] high frequency limit ( @xmath566 ) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + taking @xmath618 only the first part of the right hand side of eq . ( [ mw10 ] ) contributes . the denominator can be expanded as @xmath619 and the result is @xmath620 where @xmath621 stands for an average for a time period and the definition of @xmath622 is used . as the tls reacts slowly on that time scale , thus it reacts on the average value of the parameters @xmath189 and @xmath617 . note that @xmath623 . ( a)(b)(c ) in order to calculate the time dependent current first the occupation number @xmath567 must be calculated , which is not an easy task in the general case as the formula given by eq . ( [ mw10 ] ) contains double integrals . the next step is to insert the obtained @xmath567 into the equation for the current ( [ iinel ] ) . if one is interested in the average current ( dc ) and the harmonics of the current , it is not enough just to determine the the averaged @xmath231 and its harmonics . that can be demonstrated easily at @xmath62 by eq . ( [ iinel2 ] ) . at @xmath624 the expression contains a term , which is independent of the occupation number . this causes a nonlinearity in the @xmath1 curve in every case , regardless of the voltage dependence of the occupation number . for instance , if the tls was in thermal equilibrium with the bath then @xmath340 would hold at any bias voltage . in this case the @xmath1 curve is linear both at @xmath625 and at @xmath344 , but it has a breaking point at @xmath548 . if the alternating voltage is smaller than @xmath40 in the whole time period ( @xmath626 ) , then the response to the excitation is completely linear , and no higher harmonics are generated . this situation is demonstrated in fig . [ rectification1.fig]a . similarly , the response is linear if @xmath627 at any time . however , if the condition @xmath626 is only valid in a part of the time period , the system gives a rectified response to the harmonic excitation , thus higher harmonics are also generated . this case is shown in fig . [ rectification1.fig]b . the tls can be excited only in the part of the time that is demonstrated by the shadowed areas , where the conductance is reduced due to the possibility of backscattering . in fig . [ rectification1.fig]c a third case is presented , where the amplitude of the alternating voltage is so large , that rectification occurs both at @xmath397 and at @xmath628 . in this case the contributions from the both polarities work against each other , and the higher harmonic generation is reduced . in a general case both the nonlinearity of the occupation number @xmath347 and the above discussed intrinsic nonlinearity of the @xmath1 curve cause higher harmonic generation . however , if the period time of the excitation is much shorter than the characteristic relaxation time of the tls then the occupation number can not follow the alternating voltage , thus it sets at a constant steady state value . in this situation the higher harmonic generation is still present due to the intrinsic nonlinearity of the @xmath1 curve . in the following section these features are discussed in two special cases . in sec . [ rectification.sec ] the @xmath62 the limit is treated in the high frequency limit ( @xmath629 ) , where the rectified current and the harmonics can be easily calculated at arbitrary value of the amplitude of the excitation , @xmath630 . these results are also valid at finite temperature if @xmath631 . in the second case ( sec . [ expansion.sec ] ) an expansion with respect to @xmath630 is applied . this approach is valid if the amplitude of the excitation is small ( @xmath632 ) , however in this limit the frequency dependence can be investigated in a wide range of frequencies . the rectification due to a tls in the contact can be nicely demonstrated at @xmath62 in the @xmath629 limit . we have seen that at @xmath629 the occupation number @xmath231 becomes time independent which makes it possible to perform the calculation analytically . according to eq . [ mw11c ] @xmath633 , where @xmath634 and @xmath635 are the time average of these coefficients . inserting zero temperature into the formula ( [ ab ] ) the voltage dependence of @xmath189 and @xmath573 can be written as : @xmath636 therefore , in case of amplitudes for which @xmath637 holds for the total period of time the occupation number is zero . if that is not satisfied , then the actual voltage at a certain part of the time can excite the tls out of the groundstate . two cases must be distinguished : ( i ) where @xmath638 but @xmath639 ; ( ii ) where @xmath638 and @xmath640 . case ( i)case ( ii ) figure [ rectification2.fig ] show the actual voltage as a function of time using @xmath641 in these two cases . the shadowed areas indicate the regions where @xmath642 . the phases , where the shadowed areas are ending are denoted by @xmath643 ( in case ( i ) ) and @xmath644 ( in case ( ii ) ) . all of the following results for case ( i ) can be applied also for case ( ii ) by inserting @xmath645 . these phases can be expressed by @xmath547 , @xmath630 and @xmath40 as follows : @xmath646 the time average of @xmath189 is calculated as : @xmath647{\rm d}\varphi+ \int\limits_{\bar\varphi_0}^{\pi}\left[-e(v_0+v_1\cos\varphi)-e\right]{\rm d}\varphi\right\}=\\ = \frac{\kappa}{\pi}e\left [ v_0(\varphi_0-\varphi_0^{\star})+v_1(\sin\varphi_0+\sin\varphi_0^{\star})\right]-\frac{\kappa}{\pi}e(\varphi_0+\varphi_0^{\star}),\end{aligned}\ ] ] and the occupation number is written as @xmath648 after having the result for @xmath231 the inelastic current correction can be calculated using eq . [ iinel2 ] . at the points , where @xmath649 the time derivative of the current is discontinuous , which results in higher harmonic generation . the second harmonic generation with @xmath650 and @xmath651 and also the shift of the dc signal can be easily calculated : @xmath652 the results for the @xmath650 are plotted for different values of the excitation amplitude , @xmath630 . in this case the width of the peak at @xmath653 is determined by the amplitude @xmath630 , and these results are also valid at finite temperature if @xmath654 . the curves are plotted for different amplitudes of the excitation signal . _ _ ] at finite frequencies the occupation number is already time dependent ( @xmath567 ) , thus the above considerations can not be applied . a calculation for intermediate frequencies can be performed in the @xmath655 limit . in this case the integrals in the general solution for @xmath231 ( eq . [ mw2 ] ) are calculated after expanding the coefficients @xmath656 and @xmath657 with respect to @xmath630 . the expansion for @xmath189 is : @xmath658 and similarly for @xmath617 . the derivatives are taken at @xmath659 . in the general solution for @xmath231 first the quantity @xmath660 must be calculated . the non - oscillating part of the exponent is linearly increasing with time , thus the expansion can be carried out only for the oscillating part assuming that @xmath661 and @xmath662 : @xmath663 where @xmath664 the shift of the relaxation rate by a term proportional to @xmath665 is due to the nonlinear behavior . the next step is to calculate the integrals @xmath666 . all the terms obtained are proportional to the exponential , @xmath667 , which is cancelled out by the factor @xmath668 in the expression ( [ mw2 ] ) . in the remaining part the higher order terms @xmath669 are dropped . the next tedious task is to collect all the terms in the coefficients of the expansion . the occupation number can be written in the following form : @xmath670 where @xmath671 is the stationary occupation number at @xmath672 ( @xmath673 ) , @xmath674 and @xmath675 define the amplitudes of the harmonic response to the excitation , @xmath676 is the shift of the stationary value due to the nonlinear behavior , while @xmath677 and @xmath678 define the amplitudes of the second harmonic generation . the values of these coefficients are given by the following equations : @xmath679 the conditions for the expansion have excluded the narrow frequency region around @xmath680 . accordingly , the above results contain singular terms in @xmath68 , which diverges as @xmath681 . on the other hand , it has been seen that the @xmath681 limit is well defined for @xmath231 , and indeed , it can be shown that the singular terms cancel out in the low frequency limit . ) and the second harmonic terms ( @xmath682 , @xmath683 ) as a function of bias voltage at different excitation frequencies in the inelastic case . the right panels presents the frequency dependence of the corresponding terms at a fixed voltage of @xmath397 . the temperature is @xmath416.__,scaledwidth=60.0% ] the expression for the current correction can be expressed similarly to ( [ mw20 ] ) : @xmath684 in the inelastic case the current correction given by eq . ( [ iinel ] ) can be written as @xmath685 where @xmath686 . these must be expanded similarly to equation ( [ mw17 ] ) given for @xmath189 . after this step the coefficients in ( [ mw23 ] ) can be given as follows : @xmath687 these results can be analytically calculated , and for demonstration the experimentally important quantities ( @xmath688 , @xmath689 , and @xmath690 ) are plotted in fig . [ mwin.fig ] . the left panels show the voltage dependence at different @xmath68 values , while the right panels demonstrate the @xmath68 dependence at the fixed bias voltage value of @xmath397 . at low frequencies , the tls occupation number can follow the alternating voltage , thus both @xmath690 and @xmath558 are proportional to the second derivative of the @xmath1 curve ( see eq . [ iexpansion ] ) . in this limit there is no phase shift , and thus @xmath691 . in accordance with the static results , the curves show a negative peak at @xmath397 , and a positive tail at higher voltages . in the high frequency limit the tls can not follow the excitation , thus the occupation number sets at a constant value . due to the intrinsic nonlinearity of the @xmath1 curve , even in the high frequency limit a pronounced negative peak is observed at @xmath397 , but the positive background disappears . ( at zero temperature the amplitude of the negative dirac delta peak at @xmath397 in @xmath690 is exactly @xmath692 times smaller in the @xmath693 limit than in the @xmath681 limit . this factor of @xmath692 is well demonstrated by the @xmath68 dependence of @xmath694 in fig . [ mwin.fig ] as well . ) around the characteristic frequency of the tls ( @xmath695 ) a significant phase shift is observed , as demonstrated by the curves for @xmath689 . ) and the second harmonic terms ( @xmath696 , @xmath697 ) as a function of bias voltage at different excitation frequencies in the elastic case . the right panels presents the frequency dependence of the corresponding terms at a fixed voltage of @xmath397 . the temperature is @xmath416.__,scaledwidth=60.0% ] in the elastic case the different current components can be expressed by using eq . ( [ iel ] ) : @xmath698 in fig . [ mwel.fig ] the second order terms are plotted for the elastic case . the nonlinearity of the @xmath1 curve of the elastic contribution is basically determined by the nonlinearity of @xmath347 . above the characteristic frequency of the tls the occupation number can not follow the excitation , thus both @xmath699 and @xmath699 are suppressed . the shift of the dc component ( @xmath700 ) shows the spectroscopic peak at @xmath397 even at @xmath701 , just the amplitude is reduced by @xmath702 . around the characteristic frequency of the tls a significant @xmath699 contribution is observed . the dynamical properties of the system are also demonstrated by the phaseshift in the second harmonic contributions : @xmath703 . in fig . [ phaseshift.fig ] the phaseshift @xmath704 is plotted at @xmath397 both for the inelastic and elastic case . in the elastic contribution the phaseshift grows from @xmath705 to @xmath706 between the @xmath707 and the @xmath701 limit . contrary , in the inelastic case the phaseshift is zero in both limits , and even at @xmath708 a only minor phaseshift of @xmath709 is observed , as the nonlinearity is basically related to the intrinsic nonlinearity of @xmath1 curve and the nonlinearity of @xmath347 plays a less important role . for the elastic and the inelastic contribution , respectively . the temperature is @xmath416.__,scaledwidth=35.0% ] in this section the theory of the scattering on slow tlss in point contacts was reviewed . in sec . [ ballistic ] it was seen that the electron distribution functions and the electric potential around a ballistic contact can be determined by solving the homogeneous boltzmann equation for a point contact geometry . the solution yields a voltage independent conductance , which is known as the `` sharvin conductance '' ( eq . [ sharvinconductance ] ) . the effect of the scattering on a slow tls near the contact can be considered by perturbation in the collision term of the homogeneous boltzmann equation . the electron - tls interaction may result in backscattering processes , where an electron that has already crossed the contact is scattered back through it . these processes give a strongly voltage dependent correction to the current , for instance at zero temperature backscattering can only occur if the voltage bias exceeds the excitation energy of the tls . the nonlinear current correction due to the inelastic scattering of the electrons on a single tls can be explicitly given , if the occupation number of the upper level @xmath231 , the electron - tls coupling strength @xmath373 , and the geometrical factor @xmath710 are known . it has crucial importance that the interaction of a tls with the neighborhood is dominated by the electron - tls scattering , and the phonon contribution is significantly smaller because of the relatively small phonon phase space at energies much smaller than the debye energy . therefore , the excited tls mainly relaxes its energy through the electrons , which have a nonequilibrium distribution near the contact . as a consequence , the tls occupation number is also not the equilibrium one , and it is strongly voltage dependent . the value of @xmath347 can be determined by a simple dynamical equation using fermi s golden rule . due to voltage dependence of @xmath231 a nonlinear current correction arises from the elastic scattering of the electrons on the tls as well , provided that the scattering cross sections are different for the two states of the tls . both for the elastic and inelastic scattering the point contact spectrum ( i.e. the second derivative of the @xmath711 curve ) shows a sharp peak at the excitation energy of the tls , and a background tail at higher voltages due to the nonequilibrium occupations . the current correction due to the scattering on a single tls can also be determined for a diffusive point contact . in this case the equations are more difficult , as the collision integral due to the elastic scattering on defects should also be included in the boltzmann equation . surprisingly , the lengthy calculations give exactly the same results as in ballistic contacts , just the geometrical prefactors are different . the response to a time dependent bias voltage was also reviewed . due to the nonlinearity of the @xmath711 curve the application of a small ac modulation beside the dc bias causes two important features in the response signal : ( i ) a second harmonic signal is generated , ( ii ) the average value of the current is shifted , which is referred to as the dc shift signal . if the frequency of the ac modulation is much smaller than the relaxation frequency of the tls and its amplitude is small compared to the energy splitting , then both of these signals are proportional to the second derivative of the @xmath711 curve . on the other hand , if the modulating signal is fast compared to the relaxation of the tls then the occupation number @xmath231 can not follow the variation of the bias voltage , which has important consequences . previously it was believed that in the high frequency limit both the elastic and inelastic contributions from the scattering on the tls are completely suppressed in the response signal . our more detailed calculations have shown that the dc shift signal contains a frequency independent contribution due to the rectification at voltages in the range of the tls energies thus it is not decisive for the relaxation dynamics of the tls . we have demonstrated that both the elastic and inelastic contributions are almost independent of the irradiation frequency in the dc shift signal . a more appropriate quantity for these studies is the second harmonic generation signal . in this case the elastic contribution is completely suppressed and the inelastic contribution shrinks by a factor of @xmath692 if the irradiation frequency exceeds the relaxation frequency of the tls . we have also investigated the phase shift in the second harmonic signal which might be examined in subsequent experimental studies . there is another suggestion by kozub and rudin@xcite to explain zero bias anomalies based on the adiabatic renormalization of the splitting of the tls by electrons . that gives , however , a weak temperature dependence because the main contribution comes from the electrons far from the fermi energy , thus that is not related to infrared singularities . the interaction between the tls due to the friedel oscillation was also considered and magnetic field dependence is suggested . however , it is hard to compare the theory with experiment in its present form . it is also interesting to note that tlss and magnetic impurities interacting with metallic electrons show several similar features . both could lead to logarithmic corrections known as the kondo effect . the main similarity is that the spin @xmath712 has also two states . without external magnetic field or internal magnetic field characteristic to the spin glass systems the two states @xmath713 are degenerate , just like a tls without splitting ( @xmath714 ) . in that case the occupations of the two spin states @xmath715 and that can not be affected by the applied voltage on the point contact because of spin degeneracy . in born approximation that gives a single peak at zero bias in the point contact spectra @xmath716 , and no background occurs . the logarithmic kondo corrections , however , broaden the peak which looks very similar to background in case of phonons and tls , but that is not due to a nonequilibrium distribution.@xcite the situation is different , where the spin degeneracy is lifted by the magnetic field . in that case @xmath717 , thus the occupation numbers depend on the applied voltage and background occurs even without kondo corrections . in the following section the former experimental results are discussed and confronted with the model of slow two level systems . this review is aimed at collecting and summarizing the main results concerning the theory of the scattering on slow two - level systems in point contacts , and to compare those with the experiments . as a point contact probes the restricted area around the contact region , it can be used to investigate a few or even a single tls located in the contact area . therefore , some measurements can be directly used to study the nature of tlss in the contact region , and to check the relevance of the model presented . in disordered metallic point contacts tlss with a wide variety of relaxation times can be present . a spectacular sign of the presence of tlss in a point contact is the telegraph fluctuation of resistance , caused by defects fluctuating between metastable configurations on an experimentally resolvable timescale.@xcite examples for such so - called two level fluctuators are shown in fig . [ buhrmantls ] . panel ( a ) shows the simplest telegraph fluctuation : the resistance is switching between two values , as discussed by considering the elastic processes ( see sec . [ elasticscattering ] ) , while the switching itself is an inelastic process . in panel ( b ) two independent telegraph fluctuations are superimposed upon each other , while on panel ( c ) and ( d ) one two - level fluctuator modulates the amplitude or the average switching time of the other . the screening of these two - level systems by electrons causes a slow - down of the tls motion which can be described by the renormalization of the tunneling amplitude.@xcite this phenomenon can be used to estimate the electron - tls coupling constant , as it is discussed in and . this review concentrates on slow tlss where the relaxation time is much larger than the electron - tls interaction time , but still the tls is fast enough not to be able to resolve it as a telegraph fluctuation of the resistance . it was seen , that such `` slow '' tlss can be studied by point - contact spectroscopy , a singularity in the pc spectrum ( @xmath718 ) is expected if the voltage bias coincides with the energy splitting of the tls close to the contact . figures [ ball_cent ] and [ secder_ball ] demonstrated that the contributions of both the elastic and inelastic scattering exhibit this anomaly at @xmath397 . in the inelastic case this anomaly always corresponds to a maximum of the differential conductance , @xmath719 , which is reflected by a negative peak in the point contact spectrum @xmath720 at @xmath397 . in the elastic case a similar peak is observed , however its sign can either be negative or positive , depending on the sign of @xmath400 ( see eq . [ iel ] ) . a similar peak is expected in the pc spectrum for fast kondo - like tlss , for which the electron assisted transition processes are not negligible ( see sec . [ fasttls ] ) . for the fast tlss the anomaly appears as a minimum in the differential conductance around zero bias , and the width of the zero bias minimum is determined by the kondo temperature @xmath0 . this feature is shown as a positive peak in the @xmath720 curve at @xmath721.@xcite that peak shows very strong similarity to the kondo peak in the presence of magnetic impurities,@xcite but no magnetic field dependence is observed , and the peak is changing by annealing . the sign of the anomaly has crucial importance in distinguishing the slow tlss from fast tlss . along this discussion the anomaly exhibiting a conductance maximum at zero bias is called `` positive '' zero bias anomaly ( zba ) , while the conductance minimum is called `` negative '' zba . we note that the labelling of the sign of the anomaly can be confusing , as in some cases the voltage dependence of either the differential conductance or the differential resistance is plotted , in other studies the point contact spectrum is presented . furthermore , concerning the pc spectrum , in theory the @xmath720 curves have direct physical meaning ( showing the excitation spectrum ) , whereas in experimental studies usually the second derivative of the inverse curve , @xmath720 is plotted . regardless of the way how the pc spectrum is plotted , the sign of the anomaly can be determined by comparing it with the sign of the contribution of the phonons . for the inelastic scattering on tlss the increase of the bias voltages causes an enhancement of the back - scattering , thus this anomaly has the same sign as the phonon part of the spectrum . in contrast , for kondo - like anomalies a resonant scattering is observed , which is restricted to the energy range of the kondo temperature . in this case the back - scattering is enhanced at zero bias , and it decreases as the bias is elevated above @xmath722 , thus the anomaly has negative sign compared to the phonon contribution . in terms of this labelling , the fast tlss should always show a negative anomaly , the inelastic contribution from slow tlss shows up as a positive anomaly , while the elastic contribution has an indefinite sign . in the experiments different zero bias anomalies were observed . a basic distinction should be made between the experiments performed on pure metals , and those , performed on disordered systems , like metallic glasses . in the first case the temperature dependence of the bulk resistance does not show any anomalous behavior , but due to the magnifying effect of point contacts , the scattering on a few defects near the contact might be seen in the point contact spectrum . in this case the appearance of the anomaly depends on the way the contact is created , and usually the anomaly can be destroyed by annealing . in metallic glasses , however , a high degree of disorder is initially present , and the anomalous feature seen in the pc spectrum is also reflected by the anomalous low - temperature behavior of the bulk resistivity@xcite ( see also ) . several metallic contacts show anomalies with different signs depending on the way the contact is prepared . in the following we concentrate on a few typical and most interesting cases . the zero bias anomalies were widely studied in copper nanocontacts created with different methods . the first experiments by touching two cu electrodes@xcite reported a pronounced positive anomaly at @xmath723mev [ fig . [ akimenko.fig]a ] . the zba has a highly asymmetric shape , which can be well fitted with the theoretical prediction for the elastic scattering on a single slow tls . the inelastic scattering could be excluded , as in this case the anomaly should show a basically symmetric peak with a background of opposite sign [ see fig . [ secder_ballq ] ] . furthermore , as the zba has consequently shown a positive sign , a kondo - like explanation could also be excluded . in some occasions further low - energy peaks were also observed [ fig . [ akimenko.fig]b ] , which were argued to be due to the higher excitations of the tls . , title="fig : " ] , title="fig : " ] similar zbas were seen in the experiments of keijsers et al.@xcite on copper thin film mechanically controllable break junctions [ fig . [ posneganomaly]a ] . these samples were created by vapor deposition , and judging from the residual resistance ratio and the intensity of the phonon peaks in the pc spectrum they were of low crystalline quality implying that the contacts were in the diffusive regime . however , the authors claim that the zba is not due to the intrinsic disorder of the sample , but it correlates with the amount of bending and stretching of the film . in these thin film samples the zba has consequently shown a positive sign , and it could be well fitted with the elastic scattering on a slow tls , similarly to the above discussed measurements of akimentko et al . the same work of keijsers et al.@xcite has also presented measurements on single crystal cu break junctions [ fig . [ posneganomaly]b ] . in this case the well resolved electron - phonon interaction spectrum implies a large elastic mean free path . in these measurements the zba has consequently shown a negative sign . here , the elastic scattering on a slow tls is still a reasonable explanation , however , the kondo - like behavior also has to be considered . ralph et al.@xcite have investigated nanofabricated cu contacts . in this case the size of the contact naturally could not be varied during the measurements , but as a great advantage , the contact was `` fixed '' , and thus a detailed study of the temperature and magnetic field dependence could be performed . in these experiments a well defined negative anomaly was observed if the sample was quenched [ fig . [ ralphzba.fig ] ] , and the anomaly disappeared after annealing . the detailed theoretical analysis of the curves ( including scaling arguments)@xcite has shown very good agreement with the theory of fast kondo - like tlss , even if their existence is still an open question.@xcite the relatively high kondo temperature ( @xmath724k ) can be obtained in the framework of recent theory@xcite only if the center is light ( e.g. h or he ) as the energy scale is the kinetic energy.@xcite mv is attributed to the scattering on fast two level systems , while the structures at higher voltages originate from the scattering on phonons . [ taken from with kind permission of the american institute of physics . ] ( b ) differential conductance of a cu nanoconstriction at @xmath725mk . the curve measured in @xmath726 t magnetic field is shifted down by @xmath727 for clarity , and only a slight dependence on the field is observed , which supports the nonmagnetic origin of the anomaly . for comparison the inset shows the conductance of cu constriction with @xmath728ppm mn at @xmath160mk , showing a pronounced zeeman splitting in magnetic field . [ taken from with kind permission of the american institute of physics . ] similar zero bias anomaly was also observed in ti nanojunctions.@xcite , title="fig : " ] mv is attributed to the scattering on fast two level systems , while the structures at higher voltages originate from the scattering on phonons . [ taken from with kind permission of the american institute of physics . ] ( b ) differential conductance of a cu nanoconstriction at @xmath725mk . the curve measured in @xmath726 t magnetic field is shifted down by @xmath727 for clarity , and only a slight dependence on the field is observed , which supports the nonmagnetic origin of the anomaly . for comparison the inset shows the conductance of cu constriction with @xmath728ppm mn at @xmath160mk , showing a pronounced zeeman splitting in magnetic field . [ taken from with kind permission of the american institute of physics . ] similar zero bias anomaly was also observed in ti nanojunctions.@xcite , title="fig : " ] in all of the above experiments the zbas were attributed to a single or a few tlss . in the measurements , where consequently a positive zba was observed both the inelastic scattering on a slow tls and the kondo - like behavior can be ruled out due to shape and the sign of the anomaly , respectively . therefore , the elastic scattering on a slow tls remains as an explanation , which indeed gives good fit for the curves . however , it should be noted that for elastic scattering the sign of the anomaly is expected to be contact dependent . a consequent positive sign might be explained by the following argument : the atom in the higher level of the tls can be considered to be more far away from the ordered arrangement of the atoms , thus its scattering cross section is larger than that of the low energy state , which results in an enhanced backscattering as the bias is increased . it is interesting to note that the positive zbas in exhibit a pronounced contact size dependence : as the diameter of the junction is decreased the zba shifts to higher energy values [ fig . [ posneganomaly]a ] . the zbas showing kondo effect in magnetically doped contacts have exhibited a similar size effect.@xcite the latter phenomenon was explained by zarnd and udvardi as a result of density of states oscillations nearby the surface.@xcite similar size effect is expected for the kondo - like scattering on fast tlss , but we stress that this explanation does not apply to the zbas in fig . [ posneganomaly]a due to the positive sign of the anomaly . the zero bias anomalies are also characteristic for metallic glass samples , which were always found to show negative anomaly@xcite [ for examples see fig . [ telegraphzba ] ] in accordance with the anomalous low - temperature increase of the bulk resistance.@xcite in metallic glasses tlss with a wide distribution of parameters can be present , thus the interpretation of the results is even more difficult . usually the zba is attributed to the scattering on tlss , however the magnetic origin also can not be excluded . the anomalies can not be suppressed by magnatic field in contrast to the magnetic kondo effect , however the anomalous part of the temperature dependence of the resistance shows a small but well defined magnetoresitance , which is not present at higher temperatures.@xcite another interesting feature is the telegraph fluctuation of the zba demonstrated by the upper curves in fig . [ telegraphzba ] . this telegraph noise is only present in the narrow voltage region of the zba , that is the amplitude of the fluctuation vanishes as the voltage is increased ( fig . [ diffzba ] ) . this extraordinary behavior may be understood by assuming that a slow two level fluctuator is not directly changing the resistance ( like in fig . [ buhrmantls ] ) , but it is modulating the parameters of a fast tls in the neighborhood . therefore , the amplitude of the fluctuation only depends on the parameters of a _ single _ or a few tlss with the slowly moving defect being in one or the other metastable position , and the contributions of electron - phonon interaction , or other tlss being more far away from this particular defect are less significant . comparing the voltage dependence of the fluctuation with theory , it turns out that the model of fast kondo - like tlss is in good agreement with the measurement , while the theory of slow tlss gives a poor fit because of the long tail of the pc spectrum.@xcite the common feature in the yet reported zbas is that the anomaly appears in the voltage region @xmath729mv . usually the anomaly shows a negative sign , however even in these cases the elastic contribution from slow tlss should be considered . a direct distinction between the fast and slow tlss could be made by studying the relaxation dynamics of the system . in terms of the two models the position of the singularity reflects the kondo temperature ( @xmath0 ) or the energy splitting ( @xmath40 ) , respectively . inserting the typical voltage range of the anomaly into the models , the standard korringa - like relaxation rate of slow tlss due to the creation of an electron and hole pair is @xmath730s ( see eq . [ relaxtime ] ) , while in the two channel kondo model the relaxation time is in the range of @xmath731s.@xcite therefore the boundary between the two cases can be investigated with microwave irradiation measurements in the ghz frequency range . experimentally the high frequency measurements were usually performed by irradiating the contact with a chopped microwave signal , and measuring the change of the dc current as the microwave is on and off . according to eq . [ iexpansion ] the voltage dependence of the dc shift signal is proportional to the second derivative of the @xmath711 curve if the frequency of the irradiation signal is slow compared to the relaxation of the tls , and the perturbation due to the alternating voltage is weak . previously , it was believed that above the characteristic frequency of the tls the dc shift signal should vanish , however the calculations were only performed for the second generation signal in the case of elastic scattering.@xcite measurements on various systems have shown that the peak in the dc shift signal can not be suppressed in the frequency range of the measurements ( @xmath732ghz).@xcite therefore , these studies concluded that the anomaly comes from fast tlss with a characteristic frequency @xmath733s@xmath734 . however , according to our results in sec . [ expansion.sec ] these measurements are not decisive as the dc shift signal _ should not vanish for slow tlss either _ , it should be only slightly reduced . ( see upper panels in fig . [ mwin.fig ] and fig . [ mwel.fig ] ) . the arbitrariness of the substraction of the background signal makes it even more difficult to draw a clear conclusion concerning the frequency dependence of the peak . more conclusive results can be obtained by measuring the second harmonic response to the high frequency irradiation , which is , however , experimentally more difficult . according to our calculations , in the second harmonic signal the peak due to the elastic scattering should disappear at high frequency , while the peak due to inelastic scattering only decreases by a factor of 2/3 ( bottom panels in fig . [ mwel.fig ] and [ mwin.fig ] , respectively ) . recent second harmonic generation measurements on ni@xmath735nb@xmath736 metallic glasses have shown,@xcite that the zba can not be suppressed with frequencies up to @xmath737ghz . this measurement excludes the elastic contribution of slow tlss due to the lack of frequency dependence , and it also excludes the inelastic contribution due to slow tlss as the sign of the peak is negative . however , the authors propose that the frequency dispersion of the smooth background signal might come from the scattering on slow tlss . an interesting result has been presented by balkashin et al . on annealed ni@xmath735nb@xmath736 metallic glass samples.@xcite the dc shift response signal was measured at various microwave irradiation frequencies both before and after annealing . the unannealed samples have shown insignificant frequency dispersion up the maximum frequency of the measurements ( @xmath737ghz ) , in agreement with the previous studies . in the annealed samples , however , the zba was completely suppressed at high microwave frequencies [ fig . [ annealing.fig ] ] . as we have discussed , the zba in the dc shift response signal should not dissappear for slow tlss either , thus the experimental observations on the annealed samples can not be interpreted in terms of the presented theory for slow tlss . this observation questions whether the zba originates from the scattering on tlss in the studied ni@xmath738nb@xmath739 metallic glasses . up to now , the high frequency measurements were only performed on metallic glass samples . unfortunately in metallic glasses various interactions coexist , which makes the interpretation of the results difficult . especially serious problem is the normalization of the measured response signals . in pure samples the curves can be normalized to the phonon peaks , as the electron - phonon relaxation time is very small ( @xmath740s ) . in metallic glasses , however the phonon peaks can not be identified , and a smooth background is observed , which can contain various contributions . in disordered systems the disorder induced electron - electron interaction ( see e.g. ) and the scattering on nonequilibrium phonons@xcite are also important , furthermore the scattering on a variety of tlss with different excitation energies can also contribute to the background . additionally , the materials under study contain magnetic ions , thus the interactions of magnetic origin also can not be excluded . due to the complexity of the system the background signal can also have frequency dispersion in the frequency range of the measurements , which hampers the comparison of the response signals measured at different frequencies . a more clear picture could be obtained by investigating the high frequency behavior of the zbas observed in clean contacts of pure metals . in this case the zba is attributed to the scattering on a single or a few tlss , and due to the ballistic nature of the contact clear phonon peaks are observed at higher energies . in this case the microwave response signals are easily compared by normalizing the curves to the phonon peaks , and thus the measurement of the second harmonic response signal provides explicit information about the relaxation dynamics of the system . especially interesting would be the study of positive zbas in copper nanojunctions ( or in other metals like ag , au and pt ) , which are explained by the elastic scattering on slow tlss . nb@xmath736 with ag in measurements at different frequencies : a sonic frequency ( 1 ) and microwave frequencies @xmath741 0.48ghz ( 2 ) , 2.036ghz ( 3 ) , and 4.72ghz ( 4 ) . the anomaly is strongly suppressed with the microwave irradiation , whereas the zba in unannealed samples is unaffected by the same irradiation . [ taken from with kind permission of the american institute of physics . ] ] a large variety of point contact spectroscopy measurements show zbas , which may be attributed to the scattering of electrons on tlss . certain measurements on cu nanocontacts show positive zbas , which can be well explained by the elastic scattering on slow tls . in this case , however , the reason for the consequently positive sign of the anomaly is not well understood . another group of measurements show negative zbas , and among these the experiments on cu nanocinstrictions@xcite are well interpreted in the framework of the two channel kondo model . in these studies the zbas are attributed to a few or even a single tls , which allows a direct comparison of the experimental observations with the theoretical predictions . zbas are frequently observed in metallic glass samples , however , in this case the complexity of the system makes the interpretation of the results more difficult . interesting attempts were made by investigating the frequency dispersion of the zbas in metallic glasses , however some of the results were misinterpreted due to the lack of the theoretical calculation for the frequency dependence of the dc shift signal . novel second harmonic generation measurements on ni@xmath738nb@xmath739 metallic glasses has proved that in this system the zba can not be explained by the scattering on slow tlss , however slow tlss can give contribution to the background . the study of the high frequency behavior of a few tlss near clean metallic contacts might provide a more clear information about the relaxation dynamics of tlss . in the cases cited above the origins of the tlss are not well known . tlss can occur , however , in crystalline material with deviation from the stochiometric composition or disorder as well . the tls is formed in crystalline neighborhood where the size or the shape of the cavity in which the atom is sitting can be modified by the surroundings . in such cases clear indication of orbital kondo effect was found in pbte if some pb atoms are replaced by smaller ge atoms.@xcite however , this material is not so easy to produce in a controlled way . recently a new family of material was discovered , which are metallic and they are composed of u , th , as , se and s.@xcite the systematic study of these well - controlled materials may also provide the possibility to understand the tls - electron interaction in more detail . the point contact spectroscopy is a very powerful method to investigate the dynamics of slow tlss and magnetic kondo impurities and resonances of unknown , nonmagnetic origin , presumably due to some structured defects at zero bias . that spectroscopy is complementary of tunneling spectroscopy with oxide barrier@xcite or break junction in the tunneling limit,@xcite but there are similarities and differences as well . the local excitations in the barrier as molecular vibrations or magnetic spin flip processes open new tunneling channels resulting in extra current at voltages above the excitation energy . that gives just the opposite sign compared to the point contact case , as those anomalies are the negative ones . the magnetic impurities in the electrodes nearby the barrier can also depress the local tunneling density of states around the fermi level in the energy range of the kondo temperature.@xcite that gives a negative zero bias anomaly just like in the point contact case . in the future it would be very important to make systematic studies concerning the tls , starting with the slow tls with contributions of arbitrary signs due to dominating elastic processes . the detailed studies of the frequency dependence based on the theoretical part of the present review would be especially useful and that would clarify some of the contradictions of the present time . it is very important to study further the resonances of unknown origin which have strong resemblance to nonmagnetic kondo behavior@xcite and to make attempts to determine their origin in at least a few cases . the origin of those resonances has been the subject of theoretical debates in the recent years . originally it was proposed that kondo effect could arise from tunneling tlss with non - commutative couplings e.g. due to the screening by electrons and electron assisted tunneling ( see e.g. ) . it was shown , however , that the adiabatic screening of the moving atom by electrons depresses the tunneling rate in such a strong manner,@xcite that the actual kondo temperature is negligible . recently , it has been suggested that the situation can be changed if the atom in the double potential well has the ground - state below the potential barrier , but the first excited state is above it.@xcite then this objection could be avoided , but the value of the splitting raises many questions . that theory for an intermediate heavy atom gives too small @xmath0 , not more than @xmath742k and that can not explain the resonances of typical width of @xmath743k . there is , however , the possibility of light atoms like hydrogen,@xcite as the dominating energy scale is the kinetic energy of the atom in the potential well . therefore , it would be very important in the future to study hydrogen defect either in atomic states or water contamination , where the hydrogen could still change its position . summarizing , the origin of slow tls and the dubious resonances deserve extended further studies . we acknowledge to all the authors laying down the principles of slow tlss in point contacts . we are especially thankful to i.k . yanson and sz . csonka for the careful reading of the manuscript , and also to k.e . nagaev , j. von delft , j.m . van ruitenbeek , o.p . balkashin , a.n . omelyanchouk , i.o . kulik , g. zarnd , d. c. ralph , g. mihly , r. a. buhrman and h. van kempen for their useful remarks . part of this work was supported by the hungarian research funds otka t026327 , ts040878 , t037451 , t038162 , t034243 .

coherent scattering of neutrinos on complex nuclei was proposed long ago @xcite as a prominent probe to study neutral - current ( nc ) @xmath0-nucleus processes , but up to now no events have been experimentally measured . neutrino detection , constitutes an excellent probe to search for a plethora of conventional neutrino physics applications and new - physics open issues @xcite . in principle , low - energy astrophysical and laboratory neutrino searches provide crucial information towards understanding the underling physics of the fundamental electroweak interactions within and beyond the sm @xcite . well - known neutrino sources include ( i ) supernova neutrinos ( with energies up to 60 - 100 mev ) and ( ii ) laboratory neutrinos ( with energies up to 52.8 mev ) emerging from stopped - pion and muon decays at muon factories ( fermilab , psi , jparc , etc . ) and at the spallation neutron source ( sns ) at oak ridge national lab @xcite . recently , it became feasible @xcite to detect neutrinos by exploiting the neutral current interactions and measuring the nuclear recoil signals through the use of very low threshold - energy detectors @xcite . to this purpose , great experimental effort has been put and new experiments have been proposed to be performed at facilities with stopped - pion neutrino beams , based on promising nuclear detectors like those of the coherent experiment @xcite and others @xcite at the sns , or alternative setups at the booster neutrino beam ( bnb ) at fermilab @xcite . the nuclear @xmath0-detectors adopted by the relevant experiments include liquid noble gases , such as @xmath2ne , @xmath3ar , @xmath4xe as well as , @xmath5ge and csi[na ] detection materials @xcite . on the theoretical side , the @xmath0-signals of low - energy neutrinos , expected to be recorded in sensitive nuclear detectors @xcite , could be simulated through nuclear calculations of @xmath0-nucleus scattering cross sections . such results may provide useful information relevant for the evolution of distant stars , the core collapse supernovae , explosive nucleosynthesis and other phenomena @xcite . in fact , coherent neutral current @xmath0-nucleus scattering events are expected to be observed by using the high intensity stopped - pion neutrino beams @xcite and nuclear targets for which recoil energies are of the order of a few to tens of kev , and therefore appropriate for detection of wimps @xcite , candidates of cold dark - matter @xcite . such detectors are e.g. the supercdms @xcite , gerda @xcite and other multi - purpose detectors @xcite . for low - energies , the dominant vector components of nc interactions lead to a coherent contribution of all nucleons ( actually all neutrons ) in the target nucleus @xcite . it is worth mentioning that , after the discovery @xcite of neutrino oscillations in propagation , the challenge of neutral and charged lepton flavour violation ( lfv ) is further investigated by extremely sensitive experiments @xcite searching for physics beyond the current standard model ( sm ) @xcite . to this end , neutrino - nucleus coherent scattering experiments may probe new physics beyond the sm involved in exotic neutrino - nucleus interactions @xcite , an undoubtable signature of non - standard physics . therefore , new data and insights will be provided to the physics of flavour changing neutral current ( fcnc ) processes , in the leptonic sector , in non - standard neutrino oscillation effects @xcite , in neutrino transition magnetic moments @xcite , in sterile neutrino search @xcite and others @xcite . furthermore , such experimental sensitivity may also inspire advantageous probes to shed light on various open issues in nuclear astrophysics @xcite . in recent works @xcite , neutral - current ( nc ) non - standard interactions ( nsi ) involving ( anti)neutrino scattering processes on leptons , nucleons and nuclei have been investigated . the reactions of this type that take place in nuclei are represented by @xmath6 ( @xmath7 with @xmath8 ) . it has been suggested @xcite that , theoretically the latter processes , can be studied with the same nuclear methods as the exotic clfv process of @xmath9 conversion in nuclei @xcite . the corresponding lagrangians may be derived within the context of various extensions of the sm @xcite , like the four fermion contact interaction , seesaw model @xcite , left - right symmetric models @xcite , gluonic operator model @xcite , etc . it is well - known that neutrino nsi may have rather significant impact in many areas of modern physics research and thus , motivate a great number of similar studies @xcite . especially in astrophysical applications , constraints coming out of some supernova explosion scenarios @xcite , may be affected and eventually lead to the necessity of further investigation of nsi in both lfv and clfv processes that may occur in solar and supernova environment @xcite . such open issues motivated our present work too . one of our main purposes in this paper , which is an extension of our previous study @xcite , is to comprehensively study the above issues by performing nuclear structure calculations for a set of experimentally interesting nuclei . we estimate reliably the nuclear matrix elements describing both interaction channels , the exotic and the standard model ones , but we mainly focus on the sm component of the neutrino - nucleus processes , i.e. we consider @xmath10 in the reactions of eq . ( [ neutrin - nsi ] ) . exotic neutrino - nucleus events are also computed . by exploiting our accurate original cross sections , we obtain simulated @xmath0-signals and flux averaged cross sections which are experimentally interesting quantities for both supernova and sns neutrinos . the total number of events expected to be recorded over the energy - threshold for the studied nuclear targets are also presented for both cases . we stress that , we have devoted special effort to obtain results of high accuracy by constructing the nuclear ground state within the context of the quasi - particle random phase approximation ( qrpa ) , i.e. by solving iteratively the bcs equations for realistic pairing interactions ( the bonn c - d potential ) @xcite , and achieving high reproducibility of the available experimental data @xcite . in addition , we made comparisons with the results of other methods evaluating the nuclear form factors that enter the coherent rate @xcite as the one which employs fractional occupation probabilities ( fop ) of the states ( on the basis of analytic expressions ) @xcite , and other well - known methods @xcite . in this section we present briefly the necessary formalism for describing all channels of the nsi processes of the reactions ( [ neutrin - nsi ] ) , derived by starting from the corresponding nuclear - level feynman diagrams . -nucleus reactions , ( _ b _ ) non - standard z - exchange @xmath0-nucleus reactions , and ( _ c _ ) z - exchange and photon - exchange @xmath11 in the presence of a nucleus ( muon - to - electron conversion ) . the non - standard ( clfv or lfv ) physics enters in the complicated vertex denoted by the bullet @xmath12 @xcite.,scaledwidth=65.0% ] in fig . [ fig.1 ] the exchange of a @xmath13-boson between a lepton and a nucleon is represented , for the sm @xmath0-nucleus scattering , fig . [ fig.1](a ) , and for the exotic @xmath0-nucleus scattering , fig . [ fig.1](b ) . as already mentioned in the introduction , the non - standard @xmath0-nucleus processes @xcite and the exotic clfv @xmath9 conversion in nuclei @xcite , can be predicted within the context of the same new - physics models @xcite . for this reason , in fig . [ fig.1](c ) we also show the exchange of a @xmath13-boson or a virtual @xmath14-photon leading to the nuclear @xmath9 conversion , @xcite . thus , the leptonic vertex in the cases of fig . [ fig.1](b),(c ) is a complicated one . a general effective lagrangian that involves sm interactions ( @xmath15 ) and nsi ( @xmath16 ) with a non - standard flavour preserving ( fp ) term , a non - universal ( nu ) term and a flavour changing ( fc ) term reads @xmath17 each of the components @xmath15 and @xmath16 , the individual terms @xmath18 and @xmath19 as well as the nuclear matrix elements that arise from each part , are discussed below . the quark - level lagrangian for neutral current non - standard neutrino interactions @xmath16 , at the four fermion approximation , ( energies @xmath20 ) is parametrized as @xcite @xmath21\left[\bar{f}\gamma^\rho p f\right]\ , , \label{nsi_lagr}\ ] ] where @xmath22 denotes a first generation sm quark , @xmath23 are three light neutrinos with majorana masses and @xmath24 are the chiral projectors . in the latter lagrangian ( [ nsi_lagr ] ) , two classes of non - standard terms are considered ( i ) flavour preserving non - sm terms that are proportional to @xmath25 ( known as non - universal , nu interactions ) and ( ii ) flavour - changing ( fc ) terms proportional to @xmath26 , @xmath27 . these couplings are defined with respect to the strength of the fermi coupling constant @xmath28 @xcite . in the present work we examine spin - zero nuclei thus , the polar - vector couplings defined as @xmath29 are mainly of interest . for the axial - vector couplings it holds @xmath30 . following ref . @xcite , the nuclear physics aspects of the neutrino - matter nsi can be explored by transforming the quark - level lagrangian ( [ nsi_lagr ] ) , eventually to the nuclear level where the hadronic current is written in terms of nc nucleon form factors that are functions of the four momentum transfer . generally , for inelastic @xmath0-nucleus scattering , the magnitude of the three momentum transfer , @xmath31 , is a function of the scattering angle of the outgoing neutrino @xmath32 ( in laboratory frame ) , the initial , @xmath33 , and final , @xmath34 , nuclear energies , as well as the excitation energy of the target nucleus , @xmath35 , and takes the form @xmath36 @xcite . our analysis in the present paper , concentrates on the dominant coherent ( elastic ) channel where only @xmath37 transitions occur ( @xmath38 , @xmath39 ) and the momentum transfer in terms of the incoming neutrino energy , @xmath40 , becomes @xmath41 or equivalently @xmath42 . the nsi coherent differential cross section of neutrinos scattering off a spin - zero nucleus , with respect to the scattering angle @xmath32 reads @xcite @xmath43 @xmath44 , denotes the flavour of incident neutrinos and @xmath45 represents the nuclear ground state ( for even - even nuclei assumed here , @xmath46 ) . the nuclear matrix element , that enters the cross section of eq . ( [ nsi_dcostheta ] ) , is written as @xcite @xmath47 ^ 2 \\ & \ , \ , \ , \ , + \sum_{\beta \neq \alpha } \left[\left ( 2 \epsilon_{\alpha \beta}^{uv}+ \epsilon_{\alpha \beta}^{dv } \right ) z f_z ( q^2)+ \left(\epsilon_{\alpha \beta}^{uv}+ 2 \epsilon_{\alpha\beta}^{dv } \right)n f_n ( q^2)\right]^2 , \end{aligned } \label{gv}\ ] ] ( @xmath48 ) where @xmath49 denote the nuclear ( electromagnetic ) form factors for protons ( neutrons ) . we stress that , in the adopted nsi model , the coherent nc @xmath0-nucleus cross section is not flavour blind as in the sm case . obviously , by incorporating the nuclear structure details , in eqs . ( [ nsi_dcostheta ] ) and ( [ gv ] ) , the cross sections become more realistic and accurate @xcite . the structure of the lagrangian ( [ tot_lagr ] ) , implies that in the r.h.s . ( [ gv ] ) the first term is the nu matrix element , @xmath50 , and the summation is the fc matrix element , @xmath50 , hence we write @xmath51 from experimental physics perspectives , it is rather crucial to express the differential cross section with respect to the recoil energy of the nuclear target , @xmath52 . in recent years it became feasible for terrestrial neutrino detectors to detect neutrino events by measuring nuclear recoil @xcite . therefore , it is important to compute also the differential cross sections @xmath53 . in the coherent process , the nucleus recoils ( intrinsically it remains unchanged ) with energy which , in the approximation @xmath54 takes the maximum value @xmath55 , with @xmath56 denoting the nuclear mass @xcite . then , to a good approximation , the square of the three momentum transfer , is equal to @xmath57 , and the coherent nsi differential cross section with respect to @xmath58 can be cast in the form @xmath59 we note that , compared to previous studies @xcite , we have also taken into consideration the interaction @xmath60 quark [ see eq . ( [ gv ] ) ] , in addition to the momentum dependence of the nuclear form factors @xcite . both eqs . ( [ nsi_dcostheta ] ) and ( [ nsi_dt ] ) are useful for studying the nuclear physics of nsi of neutrinos with matter . furthermore , by performing numerical integrations to eq . ( [ nsi_dcostheta ] ) over the scattering angle @xmath32 or to eq . ( [ nsi_dt ] ) over the recoil energy @xmath52 , one can obtain integrated ( total ) coherent nsi cross sections , @xmath61 . following eq . ( [ nu - fc - terms ] ) , the individual cross sections @xmath62 and @xmath63 may be evaluated accordingly . the effective ( quark - level ) sm @xmath0-nucleus interaction lagrangian , @xmath15 at low and intermediate neutrino energies , is written as @xmath64 \left [ \bar{f } \gamma^{\rho } p f \right ] \ , , \label{sm_lagr}\ ] ] where @xmath65 and @xmath66 are the left- and right - handed couplings of the @xmath67-quark to the @xmath13-boson and @xmath68 and @xmath69 are the corresponding couplings of the @xmath70-quark ( @xmath71 is the weinberg mixing angle ) . for coherent @xmath0-nucleus scattering , the sm angle - differential cross section reads @xmath72 the operator @xmath73 in the nuclear matrix element of the latter equation is the coulomb operator which is equal to the product of the zero - order spherical bessel function times the zero - order spherical harmonic @xcite . this matrix element can be cast in the form @xcite @xmath74 ^ 2 \ , , \label{sm - me}\ ] ] where , the polar - vector couplings of protons @xmath75 and neutrons @xmath76 with the @xmath13 boson ( see fig . [ fig.1](a ) ) , are written as @xmath77 and @xmath78 , respectively . as can be easily seen , the vector contribution of all protons is very small ( @xmath79 ) , hence the coherence in eq . ( [ sm - me ] ) essentially refers to all neutrons only of the studied nucleus . after some straightforward elaboration the differential cross section with respect to the nuclear recoil energy , @xmath52 , takes the form @xmath80 the lagrangian @xmath81 of eq . ( [ tot_lagr ] ) , contains the flavour preserving ( fp ) part , equal to @xmath82 , which can be evaluated through the coulomb matrix element @xmath83 subsequently , the total coherent cross section may be computed on the basis of the matrix element @xmath84 in a previous work @xcite we evaluated original differential cross sections @xmath85 and @xmath86 , as well as individual angle - integrated cross sections of the form @xmath87 , with @xmath44 , and @xmath88 ( fc stands for the six flavour changing processes @xmath89 ) . in this work , we perform standard model cross sections calculations ( for convenience from now on we drop the index @xmath90 and always consider @xmath91 ) for a set of nuclei throughout the periodic table up to @xmath92pb . we adopt various nuclear models ( see section [ nucl - methods ] ) to compute the nuclear form factors . then , for a great part of the cross section results ( except differential cross sections ) we evaluate folded cross sections , and event rates . at first , we study the nuclear structure details of the matrix elements entering eq . ( [ sm - me ] ) , such results reflect the dependence of the coherent cross section on the incident - neutrino energy @xmath93 and the scattering angle @xmath32 ( or the recoil energy @xmath58 ) . we mention that for the even - even nuclei this study involves realistic qrpa calculations for the differential cross sections @xmath94 and @xmath95 , performed after constructing the nuclear ground state @xmath96 by solving iteratively the bardeen cooper schriffer ( bcs ) equations . the solution of these equations provides the probability amplitudes @xmath97 and @xmath98 of the @xmath99-th single nucleon level to be occupied or unoccupied , respectively . moreover , the latter equations provide the single quasi - particle energies , based on the single particle energies of the nuclear field ( a coulomb corrected woods - saxon potential in our case ) as well as the pairing part of the residual two - body interaction ( bonn c - d potential in our case ) . then , the nuclear form factors for protons ( neutrons ) are obtained as @xcite @xmath100\ , \langle j\vert j_0(qr)\vert j\rangle\left(\upsilon^j_{n_n}\right)^2 \ , , \ ] ] with @xmath101=\sqrt{2 j + 1}$ ] , @xmath102 . for each nuclear system studied , the chosen active model space , the harmonic oscillator ( h.o . ) parameter @xmath103 and the values of the two parameters @xmath104 for proton ( neutron ) pairs that renormalise the monopole ( pairing ) residual interaction ( obtained from the bonn c - d two - body potential describing the strong two - nucleon forces ) , are presented in table [ table1 ] . the adjustment of @xmath104 is achieved through the reproducibility of the pairing gaps @xmath105 ( see e.g. @xcite ) . c|cccccc nucleus & model - space & @xmath103 & @xmath106 & @xmath107 & @xmath108 & @xmath109 + @xmath110 & 8 ( no core ) & 1.522 & 4.68536 & 4.84431 & 1.12890 & 1.19648 + @xmath111 & 8 ( no core ) & 1.675 & 3.36181 & 3.49040 & 1.06981 & 1.13636 + @xmath112 & 10 ( no core ) & 1.727 & 3.81516 & 3.83313 & 1.15397 & 1.27600 + @xmath113 & 10 ( no core ) & 1.809 & 3.03777 & 3.14277 & 1.15568 & 1.23135 + @xmath114 & 15 ( no core ) & 1.902 & 1.75518 & 1.76002 & 0.94388 & 1.01348 + @xmath115 & 15 ( no core ) & 1.952 & 1.91109 & 1.55733 & 1.05640 & 0.99890 + @xmath116 & 15 ( no core ) & 2.086 & 1.52130 & 1.56935 & 0.95166 & 1.17774 + @xmath117 & 15 ( core @xmath118 ) & 2.262 & 1.19766 & 1.20823 & 0.98207 & 1.13370 + the nuclear form factor , which is the fourier transform of the nuclear charge density distribution @xmath119 , is defined as @xmath120 with @xmath121 being the zero order spherical bessel function . due to the significance of the nuclear form factors in our calculations and for the benefit of the reader we devote a separate discussion to summarise some useful possibilities of obtaining these observables . \i ) _ use of available experimental data _ + for many nuclei and especially for odd - a isotopes , the proton nuclear form factors @xmath122 , are computed by means of a model independent analysis ( using a fourier - bessel expansion model or others ) of the electron scattering data for the proton charge density @xmath119 @xcite wherever , possible . the absence of similar data for neutron densities , restricts us to assume that @xmath123 . in this work , we consider this method only for the case of the very heavy doubly closed @xmath124 nucleus . ar and @xmath125ti nuclei . the introduction of fractional occupation probabilities ( fop ) of the states provides higher reproducibility of the experimental data , compared to the simple shell - model and that of eq . ( [ f - bessel ] ) . the bcs nuclear neutron form factor @xmath126 is also presented and compared . ] \ii ) _ fractional occupation probabilities in a simple shell - model _ @xcite the form factor @xmath127 , for h.o . wavefunctions has been written as @xcite @xmath128 the radial nuclear charge density distribution @xmath119 , entering the definition of eq . ( [ definition - ff ] ) , is written in the following compact form @xcite @xmath129 where @xmath130 , with @xmath103 denoting the h.o . size parameter . @xmath131 stands for the number of quanta of the highest occupied proton ( neutron ) level . the coefficients @xmath132 are expressed as @xmath133 where @xmath134 is the gamma function . for the coefficients @xmath135 , @xmath136 and further information see ref . @xcite . up to this point , that the proton occupation probabilities entering eq . ( [ definition - ff ] ) eq . ( [ forfac ] ) have been considered equal to unity for the states below the fermi surface and zero for those above the fermi surface . in ref . @xcite , the authors introduced depletion and occupation numbers , to parametrise the partially occupied levels of the states . these parameters satisfy the relation @xmath137 within this context , the `` active '' surface nucleons ( above or below the fermi level ) have non - zero occupation probability @xmath138 , smaller than unity , while the `` core '' levels have occupation probability @xmath139 . in this paper we extend the work of ref . @xcite where three parameters @xmath140 are used to describe the partial occupation probabilities of the surface orbits . we improve the formalism by introducing more parameters , increasing this way the number of `` active '' nucleons in the studied nuclear system and come out with higher reproducibility of the experimental data @xcite . to this aim , we introduce four parameters @xmath141 in eq . ( [ depletion ] ) . then , the assumed `` active '' single - particle levels are five and eq . ( 16 ) of ref . @xcite becomes @xmath142 \\ & + \pi(\chi , z_{c } ) \left[\frac{z^{\prime}-z}{z^{\prime}-z_{c}}-\frac{\alpha_{2}}{z_{c}-z_{1}}-\frac{\alpha_{3}}{z^{\prime}-z_{c } } \right ] \\ & + \pi(\chi , z^{\prime } ) \left[\frac{z - z_{c}}{z^{\prime}-z_{c } } + \frac{\alpha_{3}}{z^{\prime}-z_{c}}- \frac{\alpha_{4}}{z^{\prime \prime}-z^{\prime } } \right ] \\ & + \pi(\chi , z^{\prime \prime } ) \left[\frac{\alpha_{4}}{z^{\prime \prime}-z^{\prime } } - \frac{\lambda}{z^{\prime \prime \prime}-z^{\prime \prime } } \right ] + \pi(\chi , z^{\prime \prime \prime } ) \frac{\lambda}{z^{\prime \prime \prime}-z^{\prime \prime}}\ , , \end{aligned } \label{fractional}\ ] ] with @xmath143 . by substituting the polynomial @xmath144 of eq . ( [ rho(r ) ] ) with that of the latter expression and using the experimental data @xcite , we we fit the parameters @xmath145 ( and similarly for the form factor of eq . ( [ forfac ] ) ) . as an example , for the @xmath3ar isotope we have , @xmath146 . the resulting fractional occupation probabilities that fit the experimental charge density distribution are @xmath147 , @xmath148 , @xmath149 , @xmath150 . similarly for the @xmath125ti nucleus , we have @xmath151 and the fitting parameters are @xmath152 , @xmath153 , @xmath154 , @xmath155 . in fig . [ fig.2 ] the prediction of the method is compared with that of the simple shell - model and the experimental data . we note that in the momentum transfer range of our interest ( i.e. @xmath156 ) the form factor has excellent behaviour . we however mention that even though the fop method presents very high reproducibility of the experimental data , it is not always applicable , e.g. for deformed nuclei ( where bcs appears to be still successful ) . \iii ) _ use of effective expressions for the nuclear form factors _ + we finally discuss one of the most accurate effective methods for calculating the nuclear form factor by ref . @xcite @xmath157\ , , \label{f - bessel}\ ] ] where @xmath158 is the known first - order spherical - bessel function and @xmath159 , with @xmath160 and @xmath161 being the radius and surface thickness parameters of the nucleus respectively . the radius parameter is usually given from the semi - empirical form @xmath162 while @xmath161 is of the order of @xmath163 ( see ref . @xcite ) . it is worth noting that , by inserting the form factors @xmath49 obtained as described above in eq . ( [ sm - me ] ) , the resulting cross sections have a rather high confidence level . in the next part of the paper the results show that the momentum dependence of the nuclear form factors becomes crucial , especially for intermediate and high energies . in some cases , differences of even an order of magnitude may occur as compared to the calculations neglecting the momentum dependence of the nuclear form factors . the next phase of our calculational procedure is related to the total coherent @xmath0-nucleus cross sections , obtained through numerical integration of eq . ( [ sm_dcostheta ] ) over angles [ or eq . ( [ sm_dt ] ) over @xmath52 ] as @xmath165 the results for the standard model cross sections , for a set of different promising targets throughout the periodic table , are presented in fig . [ fig.3 ] . as can be seen , the present nuclear structure calculations indicate that between light and heavy nuclear systems , the cross sections may differ by even two orders of magnitude ( or more ) as a consequence of the dependence on the nuclear parameters ( i.e. mass , form factors , etc . ) . we also see that for heavier nuclei the cross sections flatten more quickly ( at lower neutrino energies ) compared to that of lighter nuclear isotopes . the latter conclusion originates mainly from the fact that , for heavy nuclei the suppression of the cross sections due to the nuclear form factors becomes more significant . thus , for heavy material the nuclear effects become important even at low energies . such original cross section results are helpful for the simulations of the standard and non - standard model signals of @xmath0-detection experiments ( see below ) . in units @xmath166 for a set of nuclei as a function of the incoming neutrino energy @xmath40 , for the sm neutrino processes @xmath167.,scaledwidth=50.0% ] as discussed previously , our present calculations may also be useful for ongoing and future neutrino experiments related to supernova ( sn ) neutrino detection , since as it is known , the neutrinos emitted in sn explosions transfer the maximum part of the total the energy released . then , the total neutrino flux , @xmath168 , arriving at a terrestrial detector as a function of the sn neutrino energy @xmath40 , the number of emitted ( anti)neutrinos @xmath169 at a distance @xmath70 from the source ( here we consider @xmath170 ) , reads @xcite @xmath171 ( @xmath172 ) where @xmath173 denotes the energy distribution of the ( anti)neutrino flavour @xmath174 . the emitted sn - neutrino energy spectra @xmath175 may be parametrised by maxwell - boltzmann distributions that depend only on the temperature @xmath176 of the ( anti)neutrino flavour @xmath177 or @xmath178 ( the chemical potential is ignored ) we have @xmath179 ( @xmath180 , @xmath181 , @xmath182 , @xmath183 @xcite ) . for each flavour , the total number of emitted neutrinos @xmath169 is obtained from the mean neutrino energy @xcite @xmath184 and the total energy released from a sn explosion , @xmath185 @xcite . the spallation neutron source ( sns ) at oak ridge national lab @xcite produces neutrons by firing a pulsed proton beam at a liquid mercury target @xcite . the main aim of the coherent proposal @xcite ( or of other similar conceptual @xcite ) concerns with possible detection of neutrino - nucleus coherent scattering events at the sns . our simulations here , are mainly motivated by previous studies @xcite and the hope to provide our accurate nuclear structure calculations . in stopped pion - muon sources , neutrinos are produced by the pion decay chain . pion decay at rest @xmath186 ( @xmath187 ) produces monochromatic muon neutrinos @xmath188 at 29.9 mev , followed by electron neutrinos @xmath189 and muon antineutrinos @xmath190 that are produced by the muon - decay @xmath191 ( @xmath192 ) @xcite . for pulsed beams in time - scales narrower than @xmath193 , @xmath194 s and @xmath190 s will be delayed with the beam while @xmath188 s will be prompt with the beam @xcite . the emitted @xmath194 and @xmath195 neutrino spectra are described by the high precision normalized distributions , known as the michel spectrum @xcite @xmath196 ( @xmath197 is the muon rest mass ) . the maximum neutrino energy in the latter distributions is @xmath198 ( see e.g. @xcite ) . the spallation neutron source ( sns ) at oak ridge national lab is currently the most powerful facility to detect for a first time neutrino - nucleus coherent scattering events , since it provides exceptionally intense fluxes @xmath199 at 20 m and @xmath200 at 40 m from the source @xcite . the simulated laboratory neutrino signals @xmath201 coming out of our calculations for the adopted nuclear targets are discussed below . by weighting the integrated cross section @xmath202 with the neutrino distributions of eq . ( [ eta - sn ] ) , for sn neutrinos , or eq . ( [ labor - nu ] ) , for laboratory neutrinos , the total signal produced on a terrestrial detector is described by @xcite @xmath203 the resulting signals , @xmath204 , obtained by inserting in eq . ( [ eq.signal ] ) the cross sections @xmath205 of fig . [ fig.3 ] , are plotted in fig . [ fig.4 ] . -detector , ( _ left _ ) for supernova neutrinos ( @xmath206 ) , evaluated with maxwell - boltzmann distributions at @xmath170 , and ( _ right _ ) for sns neutrinos ( @xmath207 ) , at 20 m from the source . for the case of sns neutrinos the figure takes into account only the delayed beam , evaluated with the generic flux of @xmath208 . different nuclear detectors have been studied . ] in our previous work @xcite it was shown that the simulated cross sections reflect the characteristics of the incident neutrino spectrum of the specific neutrino flavour @xmath174 and therefore , such a simulated signal is characterised by its own position of the maximum peak and width of the distribution @xmath209 . we , however , remind that within the framework of the sm , coherent neutrino scattering is a flavour blind and a particle - antiparticle blind process . for this particular case our results are shown in fig . [ fig.4 ] for supernova and laboratory ( sns ) neutrinos . in neutrino simulations , another useful quantity is the flux averaged cross section @xcite which in our notation is written as @xmath210 c|cccccccccccc nucleus & @xmath110 & @xmath111 & @xmath112 & @xmath113 & @xmath114 & @xmath115 & @xmath211 & @xmath116 & @xmath117 & @xmath124 + @xmath212 & 1.46 & 2.51 & 3.91 & 7.52 & 18.59 & 25.43 & 33.29 & 70.63 & 207.56 & 514.93 + @xmath213 & 3.07 & 5.33 & 8.13 & 15.52 & 37.91 & 51.50 & 67.02 & 139.83 & 395.59 & 949.50 + [ table2 ] the results for @xmath214 , obtained by using the angle - integrated cross sections of fig . [ fig.3 ] are listed in table [ table2 ] for both neutrino sources . from experimental physics perspectives , predictions for the differential event rate , @xmath215 , of a @xmath0-detector are crucial @xcite . the usual expression for computing the yield in events is based on the neutrino flux , and is defined as @xcite @xmath216 where @xmath217 accounts for the total number of nuclei ( atoms ) in the detector material @xmath218 times the total time of exposure @xmath219 . using the latter equation , one concludes that , the lower the energy recoil , the larger the potentially detected number of events ( see fig . [ fig.5 ] and fig . [ fig.6 ] ) . in principle , in order to maximize the potential detection of a rare event process like the @xmath0-nucleus scattering , detector materials with very low energy - recoil threshold and low - background are required . ( _ right _ ) , for supernova neutrinos at @xmath170 . here , 1 ton of perfectly efficient @xmath2ne and @xmath3ar detectors have been considered and also possible neutrino oscillation in propagation effects are neglected . for heavier nuclear targets the differences become rather significant . in this figure , @xmath220 stands for eq . ( [ f - bessel ] ) and fop for the method of fractional occupation probabilities of the states . for more details see the text . ] but for @xmath5ge and @xmath4xe . ] in the last stage of our study we make predictions for the total number of coherent scattering events , the most important quantity , both from theoretical and experimental perspectives . to this purpose , we evaluate the number of expected counts , for the studied detector materials , by performing numerical integration of eq . ( [ diff.rate ] ) over the nuclear recoil threshold @xmath221 ( see table [ table3 ] ) . as has been discussed previously @xcite , sn neutrino detection might become possible by the massive dark matter detectors @xcite which have very good energy resolution and low threshold capabilities @xcite . these experiments are designed ( or planned ) to search for wimps @xcite and/or other rare events such as the neutrinoless double beta decay . the latter , use heavy nuclei as nuclear detectors , e.g. ge ( gerda @xcite and supercdms @xcite experiments ) . in addition we report that sn neutrino events can be potentially detected by experiments using noble gases like ne ( clean detector @xcite ) , ar ( warp programme @xcite ) and xe ( xenon 100 collaboration @xcite ) . c|cccccc nucleus & @xmath58 & @xmath222 kev & @xmath223 kev & @xmath224 kev & @xmath225 kev + @xmath110 & 2.52 & 2.25 & 2.05 & 1.60 & 1.14 + @xmath111 & 3.29 & 2.84 & 2.51 & 1.83 & 1.19 + @xmath112 & 4.03 & 3.35 & 2.87 & 1.96 & 1.16 + @xmath114 & 9.46 & 6.63 & 5.01 & 2.53 & 1.00 + @xmath115 & 10.73 & 7.04 & 5.06 & 2.27 & 0.76 + @xmath211 & 12.00 & 7.36 & 5.04 & 2.01 & 0.57 + @xmath116 & 18.58 & 9.61 & 5.82 & 1.70 & 0.30 + @xmath117 & 30.68 & 9.84 & 4.16 & 0.46 & 0.01 + @xmath124 & 46.93 & 7.86 & 1.95 & 0.03 & @xmath226 + [ table3 ] as mentioned in section [ nucl - methods ] , in order to test our nuclear calculations we have also employed other nuclear methods . to this purpose , we have compared our original results evaluated with the bcs method with those obtained as discussed in subsection [ other - methods ] and concluded that for the case of the coherent channel all available nuclear methods are in good agreement , but their results differ significantly from those obtained assuming @xmath227 ( see fig . [ fig.5 ] and [ fig.6 ] ) . we stress however , that , since the cross section is mostly sensitive to the neutron distribution of the target nucleus , the most accurate method ( at low and intermediate energies ) is the bcs method which provides realistic proton as well as neutron form factors . all other methods employed here consider only the proton distribution and assume @xmath228 , which especially for heavy nuclei , is a rather crude approximation . we remark , however , that the aforementioned nuclear methods offer reliable results on the differential and total event rates for low energies ( see fig . [ fig.5 ] and fig . [ fig.6 ] ) , but in order to correctly estimate the neutron form factor , methods like the bsc are probably more appropriate . our present nuclear structure calculations for laboratory ( sns ) neutrinos @xcite ( see fig . [ fig.7 ] ) , are in good agreement with previous results @xcite . they imply that a comparably large number of coherent neutrino scattering events is expected to be measured by using lne , lar , lxe , ge and csi[na ] materials adopted by the coherent collaboration @xcite . the predictions of the bcs method for these nuclei are illustrated in fig . [ fig.7 ] and compared with those of other promising nuclear targets . because the neutrino flux produced at the sns is very high , ( of the order of @xmath229 per flavour at 20 m from the source @xcite ) , even kg - scale experiments expect to measure neutrino - nucleus coherent scattering events at significantly higher rates than those of supernova neutrinos . it is worth noting that , the choice of the target nucleus plays also a crucial role , since a light nuclear target may yield almost constant number of events throughout the energy range , but small number of counts . on the other hand , a heavy nuclear target provides more counts , but yields low - energy recoil making the detection more difficult . this leads to the conclusion that the most appropriate choice for a nuclear detector might be a combination of light and heavy nuclear isotopes , like the scintillation detectors discussed in ref . @xcite . ) . the upper ( lower ) panel assumes the delayed ( prompt ) flux of laboratory stopped - pion neutrino sources . this figure assumes a perfectly efficient detector and negligible neutrino oscillation effects.,scaledwidth=50.0% ] the multi - target approach of the coherent experiment @xcite aiming on neutrino detection can also explore non - standard physics issues such as nsi @xcite , neutrino magnetic moment @xcite and sterile neutrino @xcite . in this subsection we find it interesting to evaluate the non - standard neutrino - nucleus events that could be potentially detected by this experiment in each of the proposed nuclear targets . the high intensity sns neutrino beams @xcite and the two promising @xmath0-detectors , liquid @xmath230 ( 391 kg ) and liquid @xmath114 ( 456 kg ) @xcite , firstly proposed by the clear @xcite and clean @xcite designs ( located at distance 20 m from the source ) , constitute excellent probes to search for the exotic @xmath0-reactions . other possibilities @xcite include medium and heavy weight targets like @xmath231 ( 100 kg ) inspired by the dark matter supercdms @xcite detector ( located at 20 m ) and @xmath232 ( 100 kg located at 40 m ) . and ( _ right _ ) 456 kg of liquid @xmath114 , both located at a distance of 20 m ( @xmath233 ) from the source . a perfectly efficient detector and negligible neutrino oscillation effects are assumed.,scaledwidth=100.0% ] , but for 100 kg of @xmath5ge at 20 m ( @xmath234 ) and 100 kg of liquid @xmath4xe at 40 m ( @xmath235 ) from the source.,scaledwidth=100.0% ] in fig . [ fig.8 ] and fig . [ fig.9 ] the resulting number of exotic events are illustrated and compared with the sm predictions . we note , however , that especially for the case of the flavour changing ( fc ) channel @xmath236 , by using the extremely high sensitivity of the ongoing @xmath1 conversion experiments ( comet @xcite and mu2e @xcite ) , very robust bounds have been set on the vector parameters @xmath237 @xcite . to this end , we conclude that , if the mu2e and comet experiments will not detect muon to electron conversion events , then the new @xmath237 parameters extracted in @xcite will lead to undetectable coherent rates at the sns facility for this channel . for our present calculations we use the current bounds @xcite set by the sensitivity of the psi experiment @xcite and found countable number of events for the near detectors in the case of the corresponding @xmath236 reaction . the other exotic parameters , i.e. @xmath238 with @xmath239 and @xmath240 have been taken from ref . @xcite . as discussed in ref . @xcite , we do not take into account the @xmath241 contribution , since the corresponding limits are poorly constrained and eventually predict unacceptably high rates . before closing , it is worth noting that , the present calculations indicate significant possibility of detecting exotic neutrino - nucleus events through coherent scattering in the aforementioned experiments . since neutrino - physics enters a precision era @xcite , a difference from the standard model predictions leads to an undoubtable evidence of non - standard neutrino - nucleus interactions ( nsi ) . we recall that , in order to experimentally constrain simultaneously all the exotic parameters at high precision , the detector material should consist of maximally different ratio @xmath242 @xcite . our future plans include estimation of the incoherent channel which may provide a significant part of the total cross section , especially for energies higher than @xmath243 mev ( depending on the nuclear target @xcite and the particle model predicting the exotic process ) . initially , in this paper the evaluation of all required nuclear matrix elements , related to standard model and exotic neutral - current @xmath0-nucleus processes is formulated , and realistic nuclear structure calculations of @xmath0-nucleus cross sections for a set of interesting nuclear targets are performed . the first stage involves cross sections calculations for the dominant coherent channel in the range of incoming neutrino - energies @xmath244 mev ( it includes @xmath0-energies of stopped pion - muon neutrino decay sources , supernova neutrinos , etc ) . additionally , new results for the total number of events expected to be observed in one ton of various @xmath0-detector materials are provided and the potentiality of detecting supernova as well as laboratory neutrino - nucleus events is in detail explored . the calculations are concentrated on interesting nuclei , like @xmath112 and @xmath114 , @xmath5ge and @xmath4xe which are important detector materials for several rare event experiments , like the coherent at oak ridge national laboratory , and also experiments searching for dark matter events as the gerda , supercdms , xenon 100 , clean , etc . by comparing our results with those of other methods , we see that the nuclear physics aspects ( reflecting the accuracy of the required @xmath0-nucleus cross sections ) , appreciably affect the coherent @xmath37 transition rate , a result especially useful for supernova @xmath0-detection probes . in the present work , the qrpa method that considers realistic nuclear forces has been adopted in evaluating the nuclear form factors , for both categories of @xmath0-nucleus processes , the conventional and the exotic ones . also , a comparison with other simpler methods as ( i ) effective methods and ( ii ) the method of fractional occupation probabilities , which improves over the simple shell - model and gives higher reproducibility of the available experimental data , is presented and discussed . we conclude that among all the adopted methods the agreement is quite good , especially for light and medium nuclear isotopes . however , since coherent neutrino - nucleus scattering can probe the neutron nuclear form factors , methods like the bcs provide more reliable results . in view of the operation of extremely intensive neutrino fluxes ( at the sns , psi , j - parc , fermilab , etc . ) , the sensitivity to search for new physics will be largely increased , and therefore , through coherent neutrino - nucleus scattering cross section measurements , several open questions ( involving non - standard neutrino interactions , neutrino magnetic moment , sterile neutrino searches and others ) may be answered . towards this purpose , we have comprehensively studied the non - standard neutrino - nucleus processes and provided results for interesting nuclear detectors . our predictions for the total number of events indicate that , within the current limits of the respective flavour violating parameters , the coherent experiment may come out with promising results on nsi . moreover , this experiment in conjunction with the designed sensitive muon - to - electron conversion experiments ( mu2e , comet ) , may offer significant contribution for understanding the fundamental nature of electroweak interactions in the leptonic sector and for constraining the parameters of beyond the sm lagrangians . one of us , ( dkp ) wishes to thank dr . kosmas for technical assistance . 00 d.z . freedman , phys . d 9 ( 1974 ) 1389 . a. drukier and l. stodolsky , phys . rev . d 30 ( 1984 ) 2295 . k. langanke , acta phys . pol . b 39 ( 2008 ) 265 . h. ejiri , phys . 338 ( 2000 ) 265 . kosmas and e. oset , phys . c 53 ( 1996 ) 1409 . j. schechter and j.w.f . valle , phys . d 22 ( 1980 ) 2227 ; + j. schechter and j.w.f . valle , phys . d 25 ( 1982 ) 774 . k. scholberg , phys . d 73 ( 2006 ) 033005 . a.a . aguilar - arevalo et al . , phys . rev . d 79 ( 2009 ) 072002 ; arxiv0806.1449 . louis , prog . 63 ( 2009 ) 51 - 73 . a. bolozdynya et al . , arxiv:1211.5199 . d. akimov et al . ( csi collaboration ) , arxiv:1310.0125 . k. scholberg et al . , arxiv:0910.1989 . j. measuring coherent - nc@xmath0as at fermilab . " ( 2011 ) . url : http://if-neutrino.fnal.gov/neutrino1-pagers.pdf . s.j . brice et . al . , phys . d 89 ( 2014 ) 072004 . collar et al.,arxiv:1407.7524 , submitted to nucl . k. hirata , et al . 58 ( 1987 ) 1490 ; + r.m . bionta , et al . 58 ( 1987 ) 1494 . keil , g.g . raffelt and h .- janka , astrophys . j. 590 ( 2003 ) 971 . haxton , phys . lett . 60 ( 1988 ) 768 . giannaka , t.s . kosmas , j. phys . ( 2013 ) 012124 . avignone and y.v . efremenko , j. phys . g29 , ( 2003 ) 2615 . y. efremenko and w. hix , j. phys . ( 2009 ) 012006 . horowitz , k.j . coakley , and d.n . mckinsey , phys . d 68 ( 2003 ) 023005 . a.j . anderson , j.m . conrad , e. figueroa - feliciano , k. scholberg and j. spitz , phys . rev . d 84 ( 2011 ) 013008 . kosmas and j.d . vergados , phys . d 55 ( 1997 ) 1752 . vergados , t. s. kosmas , phys . atom . 61 ( 1998 ) 1066 . p. brink et al . , arxiv : astro - ph/0503583v1 . gerda collaboration , acta phys.polon , b41 ( 2010 ) 1469 . mckinsey , k.j . coakley astropart . ( 2005 ) 355 . r. brunetti et al . , new astron . rev.49 ( 2005 ) 265 . xenon 100 collaboration , arxiv:1107.2155v1 . y. giomataris , j.d . vergados , phys . b 634 ( 2006 ) 23 . j. monroe and p. fisher , phys . d 76 ( 2007 ) 033007 . m. biassoni and c. martinez , astropart . 36 ( 2012 ) 151 . super - kamiokande collaboration , y. fukuda , et al . , 86 ( 2001 ) 5651 . sno collaboration , q.r . ahmad , et al . ( 2002 ) 011302 . kamland collaboration , k. eguchi , et al . ( 2003 ) 021802 , hep - ph/0212021 . forero , m. tortola , j.w.f . valle phys . d 86 ( 2012 ) 073012 ; arxiv:1405.7540 [ hep - ph ] . j.w.f . valle , nucl . 229 - 232 ( 2012 ) 23 - 29 and references therein . cui et al . ( comet collaboration ) , kek report 2009 - 10 ; + comet collaboration , the comet proposal to jparc , jparc proposal , 2007 ; + comet collaboration , a. kurup , nucl . suppl . 218 ( 2011 ) 38 . abrams et al . ( mu2e collaboration and project ) , mu2e conceptual design report , arxiv:1211.7019 ; + mu2e collaboration , f. cervelli , j. phys . ( 2011 ) 012073 . bernstein , p.s . cooper , phys . rept . 532 ( 2013 ) 27 . barlow , nucl . suppl . 218 ( 2011 ) 44 . y. kuno , nucl . 225 - 227 ( 2012 ) 228 - 231 . y. kuno and y. okada , rev . 73 ( 2001 ) 151 . s. davidson , c. pena - garay , n. rius , and a. santamaria , jhep 03 ( 2003 ) 011 . j. barranco , o.g . miranda and t.i . rashba , jhep 0512 ( 2005 ) 021 . papoulias and t.s . kosmas , phys . b 728 ( 2014 ) 482 . a. friedland , c. lunardini , and c. pena - garay , phys . b 594 ( 2004 ) 347 . a. friedland , c. lunardini and m. maltoni , phys . d 70 ( 2004 ) 111301 . a. friedland , c. lunardini , phys.rev . d 72 ( 2005 ) 053009 . k.j . healey , a.a . petrov , d. zhuridov , phys.rev . d87 ( 2013 ) 11 , 117301 . anderson , j.m . conrad , e. figueroa - feliciano , c. ignarra , g. karagiorgi , k. scholberg , m.h . shaevitz , j. spitz , phys . d 86 ( 2012 ) 013004 . p.s . amanik and g.c . mclaughlin , j.phys . g 36 ( 2009 ) 015105 . amanik , g.m . fuller and b. grinstein , astropart . ( 2005 ) 160 . a. esteban - pretel , r. tomas , and j.w.f . valle , phys . d 81 ( 2010 ) 063003 . d.k . papoulias and t.s . kosmas , j. phys . 410 ( 2013 ) 012123 . chiang , e. oset , t.s . kosmas , a. faessler and j.d . vergados , nucl . a 559 ( 1993 ) 526 . kosmas and j.d . vergados , phys . ( 1996 ) 251 . kosmas , nucl . phys . a 683 ( 2001 ) 443 ; + t.s . kosmas , prog . part . 48 ( 2002 ) 307 . kosmas , g.k . leontaris and j.d . vergados , prog . part . nucl . phys . 33 ( 1994 ) 397 . f. deppisch , t.s . kosmas and j.w.f . valle , nucl . b 752 ( 2006 ) 80 . forero , s. morisi , m. tortola , j.w.f . valle , jhep 1109 ( 2011 ) 142 . s.p das , f.f . deppisch , o. kittel , j.w.f . valle , phys . d 86 ( 2012 ) 055006 . petrov , d.v . zhuridov , phys . d 89 ( 2014 ) 033005 . amanik and g.m . fuller , phys . d 75 ( 2007 ) 083008 . j. barranco , o.g . miranda , c.a . moura and j.w.f . valle , phys . rev . d 73 ( 2006 ) 113001 . o.g . miranda , m.a . tortola and j.w.f . valle , jhep 0610 ( 2006 ) 008 . j. barranco , o.g . miranda and t. i. rashba , phys . d 76 ( 2007 ) 073008 . t.s . kosmas and j.d . vergados , phys . b 215 ( 1988 ) 460 ; + t.s . kosmas and j.d . vergados , nucl . phys . a 510 ( 1990 ) 641 . kosmas , j.d . vergados , o. civitarese and a. faessler , nucl . phys . a 570 ( 1994 ) 637 . kosmas , s. kovalenko and i. schmidt , phys . b 511 ( 2001 ) 203 ; + t.s . kosmas , s. kovalenko and i. schmidt , phys . b 519 ( 2001 ) 78 . chasioti and t.s . kosmas , nucl . a 829 ( 2009 ) 234 . v. tsakstara and t.s . kosmas , phys . c 83 ( 2011 ) 054612 ; + v. tsakstara and t.s . kosmas , phys . c 84 ( 2011 ) 064620 . h. de vries , c.w . de jager and c. de vries , at . data and nucl . data tables 36 ( 1987 ) 495536 . donnelly and j.d.walecka , nucl . phys . a 274 ( 1976 ) 368 . alberico , et al . c 79 ( 2009 ) 065204 . kosmas and j.d . vergados , nucl . phys . a 536 ( 1992 ) 72 . j. engel , phys . b 264 ( 1991 ) 114 . c. dohmen et al . ( sindrum - ii collaboration ) , phys . b 317 ( 1993 ) 631 . p. wintz , in proceedings of the first international symposium on lepton and baryon number violation ( 1998 ) , edited by h.v . klapdor - kleingrothaus and i.v . krivosheina ( institute of physics publishing , bristol and philadelphia ) , p.534 . unpublished .

hidden variable resultant methods are a popular class of algorithms for global multidimensional rootfinding @xcite . they compute all the solutions to zero - dimensional polynomial systems of the form : @xmath0 where @xmath1 and @xmath2 are polynomials in @xmath3 with complex coefficients . mathematically , they are based on an elegant idea that converts the multidimensional rootfinding problem in into one or more eigenvalue problems @xcite . at first these methods appear to be a practitioner s dream as a difficult rootfinding problem is solved by the robust qr or qz algorithm . desirably , these methods have received considerable research attention from the scientific computing community @xcite . despite this significant interest , hidden variable resultant methods are notoriously difficult , if not impossible , to make numerically robust . most naive implementations will introduce unwanted spurious solutions , compute roots inaccurately , and unpredictably miss zeros @xcite . spurious solutions can be removed by manually checking that all the solutions satisfy , inaccurate roots can usually be polished by newton s method , but entirely missing a zero is detrimental to a global rootfinding algorithm . the higher the polynomial degree @xmath4 and the dimension @xmath5 , the more pronounced the numerical difficulties become . though our conditioning bounds do hold for small @xmath4 and @xmath5 , this paper deals with a worst - case analysis . hence , our conclusions are not inconsistent with the observation that ( at least when @xmath4 and @xmath5 are small ) resultant methods can work very well in practice for some problems . when @xmath6 and real finite solutions are of interest , a careful combination of domain subdivision , regularization , and local refinement has been successfully used together with the cayley resultant ( also known as the dixon or bzout resultant ) for large @xmath4 @xcite . this is the algorithm employed by chebfun for bivariate global rootfinding @xcite . moreover , for @xmath7 , randomization techniques and the qz algorithm have been combined fruitfully with the macaulay resultant @xcite . there are also many other ideas @xcite . however , these techniques seem to be less successful in higher dimensions . in this paper , we show that any plain vanilla hidden variable resultant method based on the cayley or sylvester matrix is a numerically unstable algorithm for solving a polynomial system . in particular , we show that the hidden variable resultant method based on the cayley resultant matrix is numerically unstable for multidimensional rootfinding with a factor that grows exponentially with @xmath5 . we show that for @xmath7 the sylvester matrix leads to a hidden variable resultant method that can also square the conditioning of a root . we believe that this numerical instability has not been analyzed before because there are at least two other sources of numerical issues : ( 1 ) the hidden variable resultant method is usually employed with the monomial polynomial basis , which can be devastating in practice when @xmath4 is large , and ( 2 ) some rootfinding problems have inherently ill - conditioned zeros and hence , one does not always expect accurate solutions . practitioners can sometimes overcome ( 1 ) by representing the polynomials @xmath2 in another degree - graded polynomial basis for @xmath8 $ ] is degree - graded if the degree of @xmath9 is exactly @xmath10 for @xmath11 . ] however , the numerically instability that we identify can be observed even when the roots are well - conditioned and for degree - graded polynomial basis ( which includes the monomial , chebyshev , and legendre bases ) . we focus on the purely numerical , as opposed to symbolic , algorithm . we take the view that every arithmetic operation is performed in finite precision . there are many other rootfinders that either employ only symbolic manipulations @xcite or some kind of symbolic - numerical hybrid @xcite . similar careful symbolic manipulations may be useful in overcoming the numerical instability that we identify . for example , it may be possible to somehow transform the polynomial system into one that the resultant method treats in a numerical stable manner . this paper may be considered as a bearer of bad news . yet , we take the opposite and more optimistic view . we are intrigued by the potential positive impact this paper could have on rootfinders based on resultants since once a numerical instability has been identified the community is much better placed to circumvent the issue . we use the following notation . the space of univariate polynomials with complex coefficients of degree at most @xmath4 is denoted by @xmath8 $ ] , the space of @xmath5-variate polynomials of maximal degree @xmath4 in the variables @xmath3 is denoted by @xmath12 $ ] , and if @xmath13 is a vector space then the cartesian product space @xmath14 ( d - times ) is denoted by @xmath15 . finally , we use @xmath16 to be the vectorization of the matrix or tensor @xmath17 to a column vector ( this is equivalent to ` v ( : ) ` in matlab ) . our setup is as follows . first , we suppose that a degree - graded polynomial basis for @xmath8 $ ] , denoted by @xmath18 , has been selected . all polynomials will be represented using this basis . second , a region of interest @xmath19 is chosen such that @xmath20 , where @xmath20 is the tensor - product domain @xmath21 ( @xmath5 times ) , contains all the roots that would like to be computed accurately . the domain @xmath22 can be a real interval or a bounded region in the complex plane . throughout , we suppose that @xmath23 for @xmath11 , which is a very natural normalization . our two main results are in theorem [ thm : condtheorem ] and theorem [ thm : condsylvester ] . together they show that there exist @xmath2 in such that @xmath24 where @xmath25 is either the cayley ( for any @xmath26 ) or sylvester ( for @xmath6 ) resultant matrix . such a result shows that in the absolute sense the eigenvalue problem employed by these two resultant - based methods can be significantly more sensitive to perturbations than the corresponding root . together with results about relative conditioning , we conclude that these rootfinders are numerically unstable ( see section [ sec : absrel ] ) . in the next section we first introduce multidimensional resultants and describe hidden variable resultant methods for rootfinding . in section [ sec : cayley ] we show that the hidden variable resultant method based on the cayley resultant suffers from numerical instability and in section [ sec : sylvester ] we show that the sylvester matrix has a similar instability for @xmath7 . in section [ sec : absrel ] we explain why our absolute conditioning analysis leads to an additional twist when considering relative conditioning . finally , in section [ sec : futurework ] we present a brief outlook on future directions . this paper requires some knowledge of multidimensional rootfinding , hidden variable resultant methods , matrix polynomials , and conditioning analysis . in this section we briefly review this material . global rootfinding in high dimensions can be a difficult and computationally expensive task . here , we are concerned with the easiest situation where has only simple finite roots . let @xmath27 be a solution to the zero - dimensional polynomial system . then , we say that @xmath28 is a simple root of if the jacobian matrix @xmath29 is invertible , where @xmath30 \vdots & \ddots & \vdots \\[5pt ] \frac{\partial p_d}{\partial x_1}(\underline{x}^{\ast } ) & \ldots & \frac{\partial p_d}{\partial x_d}(\underline{x}^{\ast})\end{bmatrix}\in\mathbb{c}^{d\times d}. \label{eq : jacobian}\ ] ] if @xmath29 is not invertible then the problem is ill - conditioned , and a numerically stable algorithm working in finite precision arithmetic may introduce a spurious solution or may miss a non - simple root entirely . we will consider the roots of that are well - conditioned ( see proposition [ def : conditioning ] ) , finite , and simple . our focus is on the accuracy of hidden variable resultant methods , not computational speed . in general , one can not expect to have a `` fast '' algorithm for global multidimensional rootfinding . this is because the zero - dimensional polynomial system in can potentially have a large number of solutions . to say exactly how many solutions there can be , we first must be more precise about what we mean by the degree of a polynomial in the multidimensional setting @xcite . a @xmath5-variate polynomial @xmath31 has total degree @xmath32 if @xmath33 for some tensor @xmath34 . it is of total degree @xmath4 if one of the terms @xmath35 with @xmath36 is nonzero . moreover , @xmath31 has maximal degree @xmath32 if @xmath37 for some tensor @xmath34 indexed by @xmath38 . it is of maximal degree @xmath4 if one of the terms @xmath35 with @xmath39 is nonzero . bzout s lemma says that if involves polynomials of total degree @xmath4 , then there are at most @xmath40 solutions ( * ? ? ? * chap . 3 ) . for polynomials of maximal degree we have the following analogous bound ( see also ( * ? ? ? * thm . 5.1 ) ) . the zero - dimensional polynomial system in , where @xmath2 are of maximal degree @xmath4 , can have at most @xmath41 solutions . [ lem : maxsolutions ] this is the multihomogeneous bzout bound , see ( * ? ? ? 8.5.2 ) . for polynomials of maximal degree @xmath4 the bound is simply @xmath42 , where @xmath43 is the @xmath44 identity matrix and @xmath45 is the permanent of @xmath34 . we have selected maximal degree , rather than total degree , because maximal degree polynomials are more closely linked to tensor - product constructions and make later analysis in the multidimensional setting easier . we do not know how to repeat the same analysis when the polynomials are represented in a sparse basis set . suppose that the polynomial system contains polynomials of maximal degree @xmath4 . then , to verify that @xmath46 candidate points are solutions the polynomials @xmath2 must be evaluated , costing @xmath47 operations . thus , the optimal worst - case complexity is @xmath47 . for many applications global rootfinding is computationally unfeasible and instead local methods such as newton s method and homotopy continuation methods @xcite can be employed to compute a subset of the solutions . despite the fact that global multidimensional rootfinding is a computationally intensive task , we still desire a numerically stable algorithm . a survey of numerical rootfinders is given in ( * ? ? ? * chap . 5 ) . when @xmath48 , global numerical rootfinding can be done satisfactorily even with polynomial degrees in the thousands . excellent numerical and stable rootfinders can be built using domain subdivision @xcite , eigenproblems with colleague or comrade matrices @xcite , and a careful treatment of dynamic range issues @xcite . the first step of a hidden variable resultant method is to select a variable , say @xmath49 , and regard the @xmath5-variate polynomials @xmath2 in as polynomials in @xmath50 with complex coefficients that depend on @xmath49 . that is , we `` hide '' @xmath49 by rewriting @xmath51 for @xmath52 as @xmath53(x_1,\ldots , x_{d-1 } ) = \sum_{i_1,\ldots , i_{d-1}=0}^{n } c_{i_1,\ldots , i_{d-1}}(x_d)\prod_{s=1}^{d-1}\phi_{i_s}(x_s),\ ] ] where @xmath54 is a degree - graded polynomial basis for @xmath8 $ ] . this new point of view rewrites as a system of @xmath5 polynomials in @xmath55 variables . we now seek all the @xmath56 such that @xmath57,\ldots , p_d[x_d^{\ast}]$ ] have a common root in @xmath58 . algebraically , this can be achieved by using a multidimensional resultant ( * ? ? ? let @xmath1 and @xmath59 . a functional @xmath60)^{d}\rightarrow\mathbb{c}$ ] is a multidimensional resultant if , for any set of @xmath5 polynomials @xmath61 $ ] , @xmath62 is a polynomial in the coefficients of @xmath63 and @xmath64 if and only if there exists an @xmath65 such that @xmath66 for @xmath52 , where @xmath67 denotes the extended complex plane . [ def : resultantmulti ] definition [ def : resultantmulti ] defines @xmath68 up to a nonzero multiplicative constant ( * ? ? ? 1.6.1 ) . in the monomial basis it is standard to normalize @xmath68 so that @xmath69 ( * ? ? ? 1.6.1(ii ) ) . for nonmonomial bases , we are not aware of any standard normalization . assuming only has finite solutions , if @xmath68 is a multidimensional resultant then for any @xmath56 we have @xmath70,\ldots , p_d[x_d^{\ast}])\!=\!0 \!\quad \!\longleftrightarrow\!\quad\ ! \exists ( x_1^{\ast},\ldots , x_{d-1}^{\ast})\in\mathbb{c}^{d-1}\text { s.t . } p_1(\underline{x}^{\ast})\!=\!\cdots\!=\!p_d(\underline{x}^{\ast})\!=\!0,\ ] ] where @xmath71 . thus , we can calculate the @xmath5th component of all the solutions of interest by computing the roots of @xmath72,\ldots , p_d[x_d])$ ] and discarding those outside of @xmath73 . in principle , since @xmath72,\ldots , p_d[x_d])$ ] is a univariate polynomial in @xmath49 it is an easy task . however , numerically , @xmath68 is typically near - zero in large regions of @xmath74 , and spurious solutions as well as missed zeros plague this approach in finite precision arithmetic ( see figure [ fig : resultanttrouble ] ) . thus , directly computing the roots of @xmath68 is spectacularly numerically unstable for almost all @xmath4 and @xmath5 . this approach is rarely advocated in practice . typicalresultants ( 15,52 ) missed zeros ( 30,10 ) spurious solution ( 50,25 ) inaccurate root ( 30,50 ) ( -1,-1)7 ( 42,15 ) ( -1,2)7 ( 70,29 ) ( 2,1)10 ( 50,-2.5 ) @xmath75 instead , one often considers an associated multidimensional resultant matrix whose determinant is equal to @xmath68 . working with matrices rather than determinants is beneficial for practical computations , especially when @xmath6 @xcite . occasionally , this variation on hidden variable resultant methods is called _ numerically confirmed eliminants _ to highlight its improved numerical behavior ( * ? ? ? however , we will show that even after this significant improvement the hidden variable resultant methods based on the cayley and sylvester resultant matrices remain numerically unstable . let @xmath1 , @xmath59 , @xmath76 , and @xmath68 be a multidimensional resultant ( see defintion [ def : resultantmulti ] ) . a matrix - valued function @xmath77)^d\rightarrow \mathbb{c}^{n\times n}$ ] is a multidimensional resultant matrix associated with @xmath68 if for any set of @xmath5 polynomials @xmath61 $ ] we have @xmath78 there are many types of resultant matrices including cayley ( see section [ sec : cayley ] ) , sylvester ( see section [ sec : sylvester ] ) , macaulay @xcite , and others @xcite . in this paper we only consider two of the most popular choices : cayley and sylvester resultant matrices . theoretically , we can calculate the @xmath5th component of the solutions by finding all the @xmath56 such that @xmath79,\ldots , p_d[x_d^{\ast}]))=0 $ ] . in practice , our analysis will show that this @xmath5th component can not always be accurately computed . each entry of the matrix @xmath80,\ldots , p_d[x_d])$ ] is a polynomial in @xmath49 of finite degree . in linear algebra such objects are called matrix polynomials ( or polynomial matrices ) and finding the solutions of @xmath81,\ldots , p_d[x_d]))=0 $ ] is related to a polynomial eigenproblem @xcite . since multidimensional resultant matrices are matrices with univariate polynomial entries , matrix polynomials play an important role in the hidden variable resultant method . a classical reference on matrix polynomials is the book by gohberg , lancaster , and rodman @xcite . let @xmath76 and @xmath82 . we say that @xmath83 is a ( square ) matrix polynomial of size @xmath84 and degree @xmath85 if @xmath83 is an @xmath86 matrix whose entries are univariate polynomials in @xmath87 of degree @xmath88 , where at least one entry is of degree exactly @xmath85 . in fact , since is a zero - dimensional polynomial system it can only have a finite number of isolated solutions and hence , the matrix polynomials we consider are regular @xcite . we say that a square matrix polynomial @xmath83 is regular if @xmath89 for some @xmath90 . a matrix polynomial @xmath83 of size @xmath84 and degree @xmath85 can be expressed in a degree - graded polynomial basis as @xmath91 when the leading coefficient matrix @xmath92 in is invertible the eigenvalues of @xmath83 are all finite , and they satisfy @xmath93 . let @xmath83 be a regular matrix polynomial of size @xmath84 and degree @xmath85 . if @xmath90 is finite and there exists a non - zero vector @xmath94 such that @xmath95 ( resp . @xmath96 ) , then we say that @xmath97 is a _ right ( resp . left ) eigenvector _ of @xmath83 corresponding to the eigenvalue @xmath87 . [ def : matpolyeigenvector ] for a regular matrix polynomial @xmath83 we have the following relationship between its eigenvectors and determinant @xcite : for any finite @xmath90 , @xmath98 in multidimensional rootfinding , one sets @xmath99,\ldots , p_d[\lambda])$ ] and solves @xmath100 via the polynomial eigenvalue problem @xmath95 . there are various algorithms for solving @xmath95 including linearization @xcite , the ehrlich aberth method @xcite , and contour integration @xcite . however , regardless of how the polynomial eigenvalue problem is solved in finite precision , the hidden variable resultant method based on the cayley or the sylvester matrix is numerically unstable . for the popular resultant matrices , such as cayley and sylvester , the first @xmath55 components of the solutions can be determined from the left or right eigenvectors of @xmath101,\ldots , p_d[x_d^{\ast}])$ ] . for instance , if linearization is employed , the multidimensional rootfinding problem is converted into one ( typically very large ) eigenproblem , which can be solved by the qr or qz algorithm . practitioners often find that the computed eigenvectors are not accurate enough to adequately determine the @xmath55 components . however , the blame for the observed numerical instability is not only on the eigenvectors , but also the eigenvalues . our analysis will show that the @xmath5th component may not be computed accurately either . not even a numerically stable algorithm can be expected to accurately compute a simple root of if that root is itself sensitive to small perturbations . finite precision arithmetic almost always introduces roundoff errors and if these can cause large perturbations in a root then that solution is ill - conditioned . the absolute condition number of a simple root measures how sensitive the location of the root is to small perturbations in @xmath2 . let @xmath27 be a simple root of . the absolute condition number of @xmath28 associated with rootfinding is @xmath102 , i.e. , the matrix @xmath103-norm of the inverse of the jacobian . [ def : conditioning ] see @xcite . as a rule of thumb , a numerically stable rootfinder should be able to compute a simple root @xmath104 to an accuracy of @xmath105 , where @xmath106 is the unit machine roundoff . in contrast , regardless of the condition number of @xmath28 , a numerically unstable rootfinder may not compute it accurately . worse still , it may miss solutions with detrimental consequences . a hidden variable resultant method computes the @xmath5th component of the solutions by solving the polynomial eigenvalue problem @xmath80,\ldots , p_d[x_d])v = 0 $ ] . the following condition number tells us how sensitive an eigenvalue is to small perturbations in @xmath25 ( * ? ? ? ( 12 ) ) ( also see @xcite ) : let @xmath56 be a finite eigenvalue of @xmath107,\ldots , p_d[x_d])$ ] . the condition number of @xmath108 associated with the eigenvalue problem @xmath109 is @xmath110 where the supremum is taken over the set of matrix polynomials @xmath111 such that @xmath112 . [ def : condeigenvalue ] a numerical polynomial eigensolver can only be expected to compute the eigenvalue @xmath108 satisfying @xmath113 to an accuracy of @xmath114 , where @xmath106 is unit machine roundoff . we will be interested in how @xmath115 relates to the condition number , @xmath102 , of the corresponding root . it can be quite difficult to calculate @xmath115 directly from , and is usually more convenient to use the formula below . ( related formulas can be found in ( * ? ? ? 1 ) for symmetric matrix polynomials and in ( * ? ? ? 5 ) for general matrix polynomials . ) let @xmath116 be a regular matrix polynomial with finite simple eigenvalues . let @xmath56 be an eigenvalue of @xmath116 with corresponding right and left eigenvectors @xmath117 . then , we have @xmath118 where @xmath119 denotes the derivative of @xmath25 with respect to @xmath49 . [ lem : conditionexpression ] the first part of the proof follows the analysis in @xcite . let @xmath116 be a regular matrix polynomial with a simple eigenvalue @xmath56 and corresponding right and left eigenvectors @xmath117 . a perturbed matrix polynomial @xmath120 will have a perturbed eigenvalue @xmath121 and a perturbed eigenvector @xmath122 such that @xmath123 , where @xmath124 . expanding , keeping only the first order terms , and using @xmath125 we obtain @xmath126 multiplying by @xmath127 on the left , rearranging , and keeping the first order terms , we obtain @xmath128 where the derivative in @xmath129 is taken with respect to @xmath49 . thus , from we see that @xmath130 we now show that the upper bound in can be attained . take @xmath131 . then , @xmath132 and @xmath133 the result follows by definition [ def : condeigenvalue ] . for the cayley resultant matrix ( see section [ sec : cayley ] ) , we will show that @xmath134 can be as large as @xmath135 ( see theorem [ thm : condtheorem ] ) . thus , there can be an exponential increase in the conditioning that seems inherent to the methodology of the hidden variable resultant method based on the cayley resultant matrix . in particular , once the polynomial eigenvalue problem has been constructed , a backward stable numerical eigensolver may not compute accurate solutions to . we now must tackle the significant challenge of showing that the cayley and sylvester resultant matrices do lead to numerical unstable hidden variable resultant methods , i.e. , for certain solutions @xmath28 the quantity @xmath134 can be much larger than @xmath102 . the hidden variable resultant method when based on the cayley resultant @xcite finds the solutions to by solving the polynomial eigenvalue problem given by @xmath136 , where @xmath137 is a certain matrix polynomial . to define it we follow the exposition in @xcite and first introduce a related cayley function @xmath138 . the cayley function associated with the polynomials @xmath139 $ ] is a multivariate polynomial in @xmath140 variables , denoted by @xmath141 , and is given by @xmath142 q_1(t_1,s_2,\ldots , s_{d-1 } ) & \ldots & q_d(t_1,s_2,\ldots , s_{d-1})\\[3pt ] \vdots & \ddots & \vdots\\[3pt ] q_1(t_1,t_2,\ldots , t_{d-1 } ) & \ldots & q_d(t_1,t_2,\ldots , t_{d-1})\\[3pt ] \end{pmatrix}\bigg/\prod_{i=1}^{d-1 } ( s_i - t_i ) . \label{eq : dixonresultant}\ ] ] in two dimensions the cayley function ( also known as the bzoutian function @xcite ) takes the more familiar form of @xmath143 q_1(t_1 ) & q_2(t_1)\end{pmatrix } = \frac{q_1(s_1)q_2(t_1)-q_2(s_1)q_1(t_1)}{s_1-t_1},\ ] ] which is of degree at most @xmath144 in @xmath145 and @xmath146 . by carefully applying laplace s formula for the matrix determinant in , one can see that @xmath138 is a polynomial of degree @xmath147 in @xmath148 and @xmath149 for @xmath150 . note that @xmath138 is not the multidimensional resultant ( except when @xmath151 for all @xmath10 ) . instead , @xmath138 is a function that is a convenient way to define the cayley resultant matrix . let @xmath152 be the selected degree - graded polynomial basis . the cayley resultant matrix depends on the polynomial basis and is related to the expansion coefficients of @xmath138 in a tensor - product basis of @xmath152 . that is , let @xmath153 be the tensor - product expansion of the polynomial @xmath138 , where @xmath34 is a tensor of expansion coefficients of size @xmath154 . the _ cayley resultant matrix _ is the following unfolding ( or matricization ) of @xmath34 ( * ? ? ? 2.3 ) : the cayley resultant matrix associated with @xmath155 $ ] with respect to the basis @xmath152 is denoted by @xmath156 and is the @xmath157 matrix formed by the unfolding of the tensor @xmath34 in . this unfolding is often denoted by @xmath158 , where @xmath159 and @xmath160 _ ( * ? ? ? 2.3)_. for example , when @xmath161 for @xmath150 we have for @xmath162 @xmath163 this is equivalent to n = factorial(d-1)*n^(d-1 ) ; r = reshape(a , n , n ) ; in matlab , except here the indexing of the matrix @xmath156 starts at @xmath164 . for rootfinding , we set @xmath165,\ldots , q_d = p_d[x_d]$ ] ( thinking of @xmath49 as the `` hidden '' variable ) . then , @xmath166 is a square matrix polynomial ( see section [ sec : matrixpoly ] ) . if all the polynomials are of maximal degree @xmath4 , then @xmath156 is of size @xmath167 and of degree at most @xmath168 . the fact that @xmath169 is the maximum number of possible solutions that can possess ( see lemma [ lem : maxsolutions ] ) is a consequence of @xmath156 being a resultant matrix . in particular , the eigenvalues of @xmath137 are the @xmath5th components of the solutions to and the remaining @xmath55 components of the solutions can in principle be obtained from the eigenvectors . it turns out that evaluating @xmath138 at @xmath170 is equivalent to a matrix - vector product with @xmath156 . this relationship between @xmath156 and @xmath138 will be essential in section [ subsec : cayleyeigenvectorstructure ] for understanding the eigenvectors of @xmath156 . let @xmath1 , @xmath171 , and @xmath138 and @xmath156 be the cayley function and matrix associated with @xmath139 $ ] , respectively . if @xmath17 is the tensor satisfying @xmath172 for @xmath173 , then we have @xmath174 where @xmath175 is the tensor that satisfies @xmath176 [ lem : rightcayleyvectors ] the matrix - vector product @xmath177 is equivalent to the following sums : @xmath178 for some tensor @xmath175 . the result follows from . in this section we show that for systems of linear polynomials , i.e. , of total degree @xmath179 , the cayley resultant is precisely cramer s rule . we believe this connection is folklore , but we have been unable to find an existing reference that provides a rigorous justification . it gives a first hint that the hidden variable resultant method in full generality may be numerically unstable . let @xmath34 be a matrix of size @xmath44 , @xmath180 , and @xmath181 a vector of size @xmath182 . then , solving the linear polynomial system @xmath183 by the hidden variable resultant method based on the cayley resultant is equivalent to cramer s rule for calculating @xmath49 . [ thm : cayleycramer ] let @xmath184 be the last column of @xmath34 and @xmath185 , where @xmath186 is the @xmath5th canonical vector . recall that cramer s rule computes the entry @xmath49 in @xmath187 via the formula @xmath188 . we will show that for the linear polynomial system @xmath183 we have @xmath189 . observe that this , in particular , implies that ( since @xmath138 has degree @xmath164 in @xmath190 for all @xmath191 ) @xmath192 . hence , the equivalence between cramer s rule and rootfinding based on the cayley resultant . first , using , we write @xmath193 where the matrices @xmath194 and @xmath17 are @xmath195 s_2 & s_2 & t_2 & \dots & t_2\\[3pt ] \vdots & \vdots & \vdots & \ddots & \vdots\\[3pt ] s_{d-1 } & s_{d-1 } & s_{d-1 } & \dots & t_{d-1 } \\[3pt ] 1 & 1 & 1 & \dots & 1 \end{bmatrix } , \qquad m = b v + x_d a_d e^t,\ ] ] where @xmath196 is the @xmath182 vector of all ones . ( it can be shown by induction on @xmath5 that @xmath197 , as required . ) using the matrix determinant lemma , we have @xmath198 where @xmath199 is the algebraic adjugate matrix of @xmath200 . now , recall that @xmath201 and observe that @xmath202 . hence , we obtain @xmath203 using @xmath204 and the matrix determinant lemma one more time , we conclude that @xmath205 thus , @xmath189 and the resultant method calculates @xmath49 via cramer s formula . it is well - known in the literature that cramer s rule is a numerically unstable algorithm for solving @xmath206 ( * ? ? ? thus , theorem [ thm : cayleycramer ] casts significant suspicion on the numerical properties of the hidden variable resultant method based on the cayley resultant . ultimately , we wish to use lemma [ lem : conditionexpression ] to estimate the condition number of the eigenvalues of the cayley resultant matrix . to do this we need to know the left and right eigenvectors of @xmath156 . the following lemma shows that the eigenvectors of @xmath156 are in vandermonde form vector @xmath97 is in vandermonde form if there is an @xmath207 such that @xmath208 for @xmath209 . in higher dimensions , the vector @xmath210 is in vandermonde form if @xmath211 for some @xmath212 . ] . to show this we exploit the convenient relationship between evaluation of @xmath138 and matrix - vector products with @xmath156 . suppose that @xmath213 is a simple root of . let @xmath17 and @xmath214 be tensors of size @xmath215 and @xmath216 , respectively , defined by @xmath217 and @xmath218 then , the vectors @xmath219 and @xmath220 are the right and left eigenvectors of the matrix @xmath221,\ldots , p_d[x_d^{\ast}])$ ] that correspond to the eigenvalue @xmath108 . [ lem : cayleyeigenvectorstructure ] let @xmath222,\ldots , p_d[x_d^{\ast}])$ ] be the cayley function associated with @xmath57,\ldots , p_d[x_d^{\ast}]$ ] . from we find that @xmath223 because the determinant of a matrix with a vanishing last row is zero . moreover , by lemma [ lem : rightcayleyvectors ] we have @xmath224 since @xmath152 is a polynomial basis we must conclude that @xmath225 , and hence , @xmath226 with @xmath227 . in other words , @xmath97 is a right eigenvector of @xmath156 corresponding to the eigenvalue @xmath108 ( see definition [ def : matpolyeigenvector ] ) . an analogous derivation shows that @xmath220 is a left eigenvector of @xmath156 . to bound @xmath228 we need to bound the absolute value of the generalized rayleigh quotient of @xmath229 ( see lemma [ lem : conditionexpression ] ) , whenever @xmath104 is such that @xmath108 is a simple eigenvalue of @xmath137 , i.e. , there are no other solutions to with the same @xmath5th component . in a similar style to the proof of lemma [ lem : cayleyeigenvectorstructure ] we show this by exploiting the relation between evaluating the derivative of @xmath138 and matrix - vector products with @xmath229 . let @xmath230 be the polynomials in , @xmath231 a solution of , and @xmath232 the cayley function associated with @xmath233,\dots , q_d = p_d[x_d]$ ] . we have @xmath234 where @xmath29 is the jacobian matrix in . that is , @xmath235 evaluated at @xmath236 for @xmath150 is equal to the determinant of the jacobian . [ thm : jacobianexpression ] recall from that @xmath232 is a polynomial in @xmath237 and @xmath238 written in terms of a matrix determinant , and set @xmath233,\ldots , q_d = p_d[x_d]$ ] . the determinant in for @xmath232 can be expanded to obtain @xmath239(t_1,\dots , t_{i-1},s_i,\dots , s_{d-1}),\ ] ] where @xmath240 is the symmetric group of @xmath241 and @xmath242 is the signature of the permutation @xmath243 . when we evaluate @xmath232 at @xmath236 for @xmath150 the denominator vanishes , and hence , so does the numerator because @xmath232 is a polynomial . thus , by lhospital s rule , @xmath235 evaluated @xmath236 for @xmath150 is equal to @xmath244(t_1,\dots , t_{i-1},s_i,\dots , s_{d-1 } ) \label{eq : diffdet}\ ] ] evaluated at @xmath245 , @xmath246 , and @xmath247 . in principle , one could now apply the product rule and evaluate the combinatorially many terms in . instead , we note that after applying the product rule a term is zero if it contains @xmath248 for any @xmath249 and @xmath250 ( since @xmath28 is a solution to ) . there are precisely @xmath5 partial derivatives and @xmath5 terms in each product so that any nonzero term when expanding [ eq : diffdet ] has each @xmath251 differentiated precisely once . finally , note that for each @xmath150 only the @xmath252 terms in the product depend on @xmath148 . hence , from we obtain @xmath253 the result follows because the last expression is the determinant of the jacobian matrix evaluated at @xmath28 . as a consequence of theorem [ thm : jacobianexpression ] we have the following unavoidable conclusion that mathematically explains the numerical difficulties that practitioners have been experiencing with hidden variable resultant methods based on the cayley resultant . let @xmath1 . then , there exist @xmath230 in with a simple root @xmath231 such that @xmath254 and @xmath255 . thus , an eigenvalue of @xmath137 can be more sensitive to perturbations than the corresponding root by a factor that grows exponentially with @xmath5 . [ thm : condtheorem ] using lemma [ lem : rightcayleyvectors ] , theorem [ thm : jacobianexpression ] has the following equivalent matrix form : @xmath256 where @xmath227 , @xmath257 , and @xmath17 and @xmath214 are given in lemma [ lem : cayleyeigenvectorstructure ] . since @xmath258 , we know that @xmath259 and @xmath260 . hence , by lemma [ lem : conditionexpression ] @xmath261 denoting the singular values ( * ? ? ? 7.3 ) of the matrix @xmath29 by @xmath262 , select @xmath230 and @xmath231 such that @xmath263 . such polynomial systems do exist , for example , linear polynomial systems where @xmath264 and @xmath194 is a matrix with singular values @xmath265 . to ensure that @xmath255 we also require @xmath266 . then , we have @xmath267 the result follows . [ exc ] let @xmath268 be a @xmath269 orthogonal matrix , @xmath270 , having elements @xmath271 for @xmath272 , and let @xmath273 . consider the system of polynomial equations @xmath274 the origin , @xmath275 , is a simple root of this system of equations . the jacobian of the system at @xmath164 is @xmath276 , and hence , the absolute conditioning of the problem is @xmath277 . constructing the cayley resultant matrix polynomial in the monomial basis , one readily sees that for this example the right and left eigenvectors for the eigenvalue @xmath278 satisfy @xmath279 . as a consequence , @xmath280 . we emphasize that this numerical instability is truly spectacular , affects the accuracy of @xmath108 , and can grow exponentially with the dimension @xmath5 . moreover , theorem [ thm : condtheorem ] holds for any degree - graded polynomial basis selected to represent @xmath2 as long as @xmath258 . in particular , the associated numerical instability can not be resolved in general by a special choice of polynomial basis . theorem [ thm : condtheorem ] is pessimistic and importantly does not imply that the resultant method always loses accuracy , just that it might . in general , one must know the solutions to and the singular values of the jacobian matrix to be able to predict if and when the resultant method will be accurate . one should note that theorem [ thm : condtheorem ] concerns absolute conditioning and one may may wonder if a similar phenomenon also occurs in the relative sense . in section [ sec : absrel ] we show that the relative conditioning can also be increased by an exponential factor with @xmath5 . a popular alternative in two dimensions to the cayley resultant matrix is the sylvester matrix ( * ? ? ? 3 ) , denoted here by @xmath281 . we now set out to show that the hidden variable resultant based on @xmath281 is also numerically unstable . however , since @xmath7 the instability has only a moderate impact in practice as the conditioning can only be at most squared . with care , practical bivariate rootfinders can be based on the sylvester resultant @xcite though there is the possibility that a handful digits are lost . a neat way to define the sylvester matrix that accommodates nonmonomial polynomial bases is to define the matrix one row at a time . let @xmath282 and @xmath283 be two univariate polynomials in @xmath284 $ ] of degree exactly @xmath285 and @xmath286 , respectively . then , the sylvester matrix @xmath287 associated with @xmath282 and @xmath283 is defined row - by - row as @xmath288 where @xmath289 is the row vector of coefficients such that @xmath290 and @xmath291 where @xmath292 is the row vector of coefficients such that @xmath293 . [ def : sylvestermatrix ] in the monomial basis , i.e. , @xmath294 , definition [ def : sylvestermatrix ] gives the sylvester matrix of size @xmath295 as ( * ? ? ? 3 ) : @xmath296 & \ddots & \ddots & \ddots & \ddots \\[3pt ] & & a_{0 } & a_{1 } & \ldots & a_{\tau_1 } \\[3pt ] b_{0 } & b_{1 } & \ldots & b_{\tau_2 } & & \\[3pt ] & \ddots & \ddots & \ddots & \ddots \\[3pt ] & & b_{0 } & b_{1 } & \ldots & b_{\tau_2 } \\[3pt ] \end{pmatrix}\begin{matrix } { \left.\vphantom{\begin{matrix } x\\[3pt]y\vphantom{\ddots}\\[3pt]y\\[3pt ] \end{matrix}}\right\}\\ { \left.\vphantom{\begin{matrix } x\\[3pt]y\vphantom{\ddots}\\[3pt]y\\[3pt ] \end{matrix}}\right\}\\ \end{matrix}\label{eq : monomial}\ ] ] where @xmath297 and @xmath298 . our goal is to use lemma [ lem : conditionexpression ] to bound the condition number of the eigenvalues of the sylvester matrix . it turns out the right eigenvectors of @xmath281 are in vandermonde form . however , the left eigenvectors have a more peculiar structure and are related to the byproducts of a generalized clenshaw s algorithm for degree - graded polynomial bases ( see lemma [ lem : sylvestereigenvectorstructure ] ) . we develop a clenshaw s algorithm for degree - graded bases in this section with derivations of its properties in appendix [ sec : appendix ] . the selected polynomial basis @xmath299 is degree - graded and hence , satisfies a recurrence relation of the form @xmath300 where @xmath301 and @xmath302 . if @xmath299 is an orthogonal polynomial basis , then is a three - term recurrence and it is standard to employ clenshaw s algorithm @xcite to evaluate polynomials expressed as @xmath303 . this procedure can be extended to any degree - graded polynomial basis . let @xmath304 be expressed as @xmath303 , where @xmath305 is a degree - graded polynomial basis . one can evaluate @xmath304 via the following procedure : let @xmath306(x ) = 0 $ ] , and calculate @xmath307(x),\ldots , b_1[p](x)$ ] from the following recurrence relation : @xmath308(x ) = a_k + ( \alpha_kx + \beta_k)b_{k+1}[p](x ) + \sum_{j = k+1}^{n-1 } \gamma_{j , k+1}b_{j+1}[p](x ) , \qquad 1\leq k\leq n. \label{eq : clenshawlikealg}\ ] ] we refer to the quantities @xmath309(x),\ldots , b_{n+1}[p](x)$ ] as _ clenshaw shifts _ ( in the monomial case they are called horner shifts @xcite ) . the value @xmath304 can be written in terms of the clenshaw shifts is stated in a general form and holds for _ any _ degree - graded basis , in this paper we fix the normalization @xmath310 , that implies in particular @xmath311 simplifying . ] . let @xmath4 be a positive integer , @xmath207 , @xmath305 a degree - graded basis satisfying , @xmath303 , and @xmath306(x),\ldots , b_1[p](x)$ ] the clenshaw shifts satisfying . then , @xmath312(x ) + \sum_{i=1}^{n-1 } \gamma_{i,1}b_{i+1}[p](x ) . \label{eq : clenshawdegreegraded}\ ] ] [ lem : clenshawdegreegraded ] see appendix [ sec : appendix ] . clenshaw s algorithm for degree - graded polynomial bases is summarized in figure [ fig : clenshawalgorithm ] . we note that because of the full recurrence in the algorithm requires @xmath313 operations to evaluate @xmath304 . though this algorithm may not be of significant practical importance , it is of theoretical interest for the conditioning analysis of some linearizations from the so - called @xmath314- or @xmath315-spaces @xcite when degree - graded bases are employed @xcite . there is a remarkable and interesting connection between clenshaw shifts and the quotient @xmath316 , which will be useful when deriving the left eigenvectors of @xmath281 . with the same set up as lemma [ lem : clenshawdegreegraded ] we have @xmath317(y)\phi_i(x ) , \quad x\neq y \label{eq : diffdegreegraded}\ ] ] and @xmath318(x)\phi_i(x ) . \label{eq : diffclenshaw}\ ] ] [ thm : diffdegreegraded ] see appendix [ sec : appendix ] . the relation between the derivative and clenshaw shifts in has been noted by skrzipek for orthogonal polynomial bases in @xcite , where it was used to construct a so - called _ extended clenshaw s algorithm _ for evaluating polynomial derivatives . using theorem [ thm : diffdegreegraded ] and @xcite an extended clenshaw s algorithm for polynomials expressed in a degree - graded basis is immediate . we now set @xmath319 $ ] and @xmath320 $ ] ( considering @xmath75 as the hidden variable ) , and we are interested in the eigenvectors of the matrix polynomial @xmath321 , when @xmath322 is a solution to when @xmath6 . it turns out that the right eigenvectors of @xmath321 are in vandermonde form , while the left eigenvectors are related to the clenshaw shifts ( see section [ subsec : generalizedclenshaw ] ) . suppose that @xmath323 is a simple root of and that @xmath324 $ ] and @xmath325 $ ] are of degree @xmath285 and @xmath286 , respectively , in @xmath326 . the right eigenvector of @xmath321 corresponding to the eigenvalue @xmath327 is @xmath328 and the left eigenvector is defined as @xmath329(x_1^{\ast } ) , & 0\leq i\leq \tau_2 - 1,\\ \alpha_{i-\tau_2 } b_{i-\tau_2 + 1}[q_{1}](x_1^{\ast } ) , & \tau_2\leq i\leq \tau_1+\tau_2 - 1 , \end{cases}\ ] ] where @xmath330 $ ] and @xmath331(x_1^{\ast})$ ] are the clenshaw shifts with respect to @xmath152 , while the coefficients @xmath332 are defined as in . [ lem : sylvestereigenvectorstructure ] by construction we have , for @xmath333 , @xmath334 and , for @xmath335 , @xmath336 thus , @xmath97 is a right eigenvector of @xmath321 corresponding to the eigenvalue @xmath327 . for the left eigenvector , first note that for any vector @xmath337 of the form @xmath338 for @xmath339 we have by theorem [ thm : diffdegreegraded ] @xmath340(x_1^{\ast})\phi_{i}(x)q_1(x ) + \sum_{i=0}^{\tau_1 - 1 } \alpha_{i}b_{i+1}[q_1](x_1^{\ast})\phi_{i}(x)q_2(x)\\ & = -\frac{q_2(x ) - q_2(x_1^{\ast})}{x - x_1^{\ast}}q_1(x ) + \frac{q_1(x ) - q_1(x_1^{\ast})}{x - x_1^{\ast}}q_2(x)\\ & = -\frac{q_2(x)}{x - x_1^{\ast}}q_1(x ) + \frac{q_1(x)}{x - x_1^{\ast}}q_2(x ) = 0 , \end{aligned}\ ] ] where the second from last equality follows because @xmath341 . since holds for any @xmath342 and @xmath343 is a basis of @xmath344 $ ] , we deduce that @xmath345 , and hence , @xmath346 is a left eigenvector of @xmath281 corresponding to the eigenvalue @xmath327 . to bound @xmath347 we look at the absolute value of the generalized rayleigh quotient of @xmath348 , whenever @xmath28 is such that @xmath327 is a simple eigenvalue of @xmath349 . lemma [ lem : sylvestereigenvectorstructure ] allows us to show how the generalized rayleigh quotient of @xmath350 relates to the determinant of the jacobian . with the same assumptions as in lemma [ lem : sylvestereigenvectorstructure ] , we have @xmath351 where @xmath346 and @xmath97 are the left and right eigenvectors of @xmath281 , respectively , and @xmath29 is the jacobian matrix in . [ lem : sylvesterjacobian ] by lemma [ lem : sylvestereigenvectorstructure ] we know the structure of @xmath97 and @xmath346 . hence , we have @xmath352(x_1^{\ast})\phi_{i}(x_1^{\ast})\frac{\partial q_1}{\partial x_2}(x_1^{\ast } ) + \!\!\sum_{i=0}^{\tau_1 - 1}\!\ ! \alpha_{i}b_{i+1}[q_1](x_1^{\ast})\phi_{i}(x_1^{\ast})\frac{\partial q_2}{\partial x_2}(x_1^{\ast})\\ & = -\frac{\partial q_1}{\partial x_2}(x_1^{\ast})\frac{\partial q_2}{\partial x_1}(x_1^{\ast } ) + \frac{\partial q_1}{\partial x_1}(x_1^{\ast})\frac{\partial q_2}{\partial x_2}(x_1^{\ast } ) , \end{aligned}\ ] ] where the last equality used the relation in . the result now follows since this final expression equals @xmath353 and since @xmath258 we have @xmath354 . there exist @xmath355 and @xmath356 in with a simple root @xmath357 such that @xmath358 and @xmath359 . thus , an eigenvalue of @xmath349 can be squared more sensitive to perturbations than the corresponding root in the absolute sense . [ thm : condsylvester ] we give an example for which @xmath360 in lemma [ lem : sylvesterjacobian ] . for some positive parameter @xmath106 and for some @xmath361 consider the polynomials @xmath362 one can verify that @xmath363 is a common root . since @xmath364(0 ) | = \alpha_{n-1 } \alpha_{n-1}^{-1 } = 1 $ ] we have @xmath365 . the result then follows from @xmath366 and lemma [ lem : sylvesterjacobian ] . [ exs ] let us specialize example [ exc ] to @xmath7 , i.e. , for some @xmath273 and @xmath367 let us consider the system @xmath368 again , for the solution @xmath369 we have @xmath370 . building the sylvester matrix in the monomial basis , we obtain @xmath371 as predicted by the theory , @xmath372 is an eigenvalue with corresponding right and left eigenvectors , respectively , @xmath373 and @xmath374 . moreover , it is readily checked that , as expected , @xmath375 . therefore , @xmath376 theorem [ thm : condsylvester ] mathematically explains the numerical difficulties that practitioners have been experiencing with hidden variable resultant methods based on the sylvester resultant . there are successful bivariate rootfinders based on this methodology @xcite for low degree polynomial systems and it is a testimony to those authors that they have developed algorithmic remedies ( not cures ) for the inherent numerical instability . we emphasize that theorem [ thm : condsylvester ] holds for any normalized degree - graded polynomial basis . thus , the mild numerical instability can not , in general , be overcome by working in a different degree - graded polynomial basis . the example in the proof of theorem [ thm : condsylvester ] is quite alarming for a practitioner since if @xmath106 is the unit machine roundoff , then we have @xmath377 and @xmath378 . thus , a numerical rootfinder based on the sylvester matrix may entirely miss a solution that has a condition number larger than @xmath379 . a stable rootfinder should not miss such a solution . when @xmath6 , we can use theorem [ thm : condtheorem ] and lemma [ lem : sylvesterjacobian ] to conclude that the ratio between the conditioning of the cayley and sylvester resultant matrices for the same eigenvalue @xmath380 is equal to @xmath381 , where @xmath97 and @xmath346 are the right and left eigenvector of @xmath321 associated with the eigenvalue @xmath380 . this provides theoretical support for the numerical observations in @xcite . however , it seems difficult to predict _ a priori _ if the cayley or sylvester matrix will behave better numerically . for real polynomials and @xmath7 , the cayley resultant matrix is symmetric and this structure can be exploited @xcite . in the monomial basis , the sylvester matrix is two stacked toeplitz matrices ( see ) . it may be that structural differences like these are more important than their relatively similar numerical properties when @xmath6 . let @xmath382 be the solution of a mathematical problem depending on data @xmath383 . in general , with the very mild assumption that @xmath383 and @xmath384 lie in banach spaces , it is possible to define the absolute condition number of the problem by perturbing the data to @xmath385 and studying the behaviour of the perturbed solution @xmath386 : @xmath387 similarly , a relative condition number can be defined by looking at the limit ratios of _ relative _ changes . @xmath388 in this paper , we have compared two absolute condition numbers . one is given by proposition [ def : conditioning ] : there , @xmath389 is a solution of while @xmath390 is the set of polynomials in . the other is given by lemma [ lem : conditionexpression ] , where @xmath383 is a matrix polynomial and @xmath391 is the @xmath5th component of @xmath392 . to quote n. j. higham @xcite : `` usually , it is the relative condition number that is of interest , but it is more convenient to state results for the absolute condition number '' . this remark applies to our analysis as well . we have found it convenient to study the absolute condition number , but when attempting to solve the rootfinding problem in floating point arithmetic it is natural to allow for relatively small perturbations , and thus to study the relative condition number . hence , a natural question is whether the exponential increase of the absolute condition number in theorem [ thm : condtheorem ] and the squaring in theorem [ thm : condsylvester ] causes a similar effect in the relative condition number . it is not immediate that the exponential increase of the absolute condition number leads to the same effect in the relative sense . we have found examples where the exponential increase of the absolute condition number is perfectly counterbalanced by an exponentially small cayley resultant matrix . for instance , linear polynomial systems , when the cayley resultant method is equivalent to cramer s rule , fall into this category . in the relative sense , it may be possible to show that the hidden variable resultant method based on cayley or sylvester is either numerically unstable during the construction of the resultant matrix or the resultant matrix has an eigenvalue that is more sensitive to small relative perturbations than hoped . we do not know yet how to make such a statement precise . instead , we provide an example that shows that the hidden variable resultant method remains numerically unstable in the relative sense . let @xmath106 be a sufficiently small real positive parameter and @xmath1 . consider the following polynomial system : @xmath393 where if @xmath5 is odd then take @xmath394 . selecting @xmath395^d$ ] , we have that @xmath396 for @xmath250 , except possibly @xmath397 if @xmath5 is odd . it can be shown that the origin^d$ ] . ] , @xmath28 , is a simple root , @xmath398 , @xmath399 , and that @xmath400 thus , neither the polynomials @xmath401 or the resultant matrix @xmath137 are small . in such an example , the relative condition number will exhibit the same behavior as the absolute condition number . in particular , the relative condition number of an eigenvalue of @xmath137 may be larger than the relative condition number of the corresponding solution by a factor that grows exponentially with @xmath5 . the same example ( for @xmath7 ) , and a similar argument , applies to the sylvester matrix showing the conditioning can be squared in the relative sense too . in this paper we have shown that two popular hidden variable resultant methods based on the sylvester and cayley matrices are numerically unstable . our analysis is for degree - graded polynomial bases and does not include the lagrange basis or certain sparse bases . we believe that the analysis of the cayley matrix in section [ sec : cayley ] could be extended to include general polynomial bases , though the analysis in section [ sec : sylvester ] for the sylvester matrix is more intimately connected to degree - graded bases . we hesitantly suggest that hidden variable resultant methods are inherently plagued by numerial instabilities , and that neither other polynomial bases nor other resultants can avoid a worst - case scenario that we have identified in this paper . we do not know exactly how to formulate such a general statement , but we note that practitioners are widely experiencing problems with hidden variable resultant methods . in particular , we do not know of a numerical multidimensional rootfinder based on resultants that is robust for polynomial systems of large degree @xmath4 and high @xmath5 . however , at the moment the analysis that we offer here is limited to the cayley and sylvester matrices . despite our doubts that it exists , we would celebrate the discovery of a resultant matrix that can be constructed numerically and that provably does not lead to a numerically unstable hidden variable resultant method . this would be a breakthrough in global rootfinding with significant practical applications as it might allow to be converted into a large eigenproblem without confronting conditioning issues . solving high - dimensional and large degree polynomial systems would then be restricted by computational cost rather than numerical accuracy . finally , we express again our hope that this paper , while appearing rather negative , will have a positive long - term impact on future research into numerical rootfinders . we thank yuji nakatsukasa , one of our closest colleagues , for his insightful discussions during the writing of @xcite that ultimately lead us to consider conditioning issues more closely . we also thank anthony austin and martin lotz for carefully reading a draft and providing us with excellent comments . while this manuscript was in a much earlier form martin lotz personally sent it to gregorio malajovich for his comments . gregorio s comprehensive and enthusiastic reply encouraged us to proceed with renewed vigor . this appendix contains the tedious , though necessary , proofs required in section [ subsec : generalizedclenshaw ] for clenshaw s algorithm for evaluating polynomials expressed in a degree - graded basis . by rearranging we have @xmath402(x ) - ( \alpha_kx + \beta_k)b_{k+1}[p](x ) - \sum_{j = k+1}^{n-1 } \gamma_{j , k+1}b_{j+1}[p](x)$ ] . thus , @xmath403(x ) - ( \alpha_kx+\beta_k)b_{k+1}[p](x ) -\!\ ! \sum_{j = k+1}^{n-1 } \gamma_{j , k+1}b_{j+1}[p](x)\right]\!\phi_k(x).\ ] ] now , by interchanging the summations and collecting terms we have @xmath404(x ) - \sum_{k=2}^n \left(\alpha_{k-1}x+\beta_{k-1}\right)\phi_{k-1}(x)b_k[p](x)\\ & \qquad\quad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad -\sum_{j=2}^{n-1}\left[\sum_{k=1}^{j-1}\gamma_{j , k+1}\phi_{k}(x)\right]b_{j+1}[p](x)\\ & = a_0\phi_0(x ) + \phi_1(x)b_1[p](x)\\ & \qquad\qquad\qquad+\sum_{j=1}^{n-1}\left[\phi_{j+1}(x ) -(\alpha_{j}x+\beta_{j})\phi_{j}(x ) - \sum_{k=1}^{j-1 } \gamma_{j , k+1}\phi_{k}(x)\right]b_{j+1}[p](x)\\ \end{aligned}\ ] ] finally , using we obtain @xmath405(x ) + \sum_{j=1}^{n-1 } \gamma_{j,1}\phi_0(x)b_{j+1}[p](x),\ ] ] as required . section [ subsec : generalizedclenshaw ] also shows that clenshaw s algorithm connects to the quotient @xmath316 . to achieve this we need an immediate result that proves a different recurrence relation on the clenshaw shifts to . the proof involves tedious algebraic manipulations and mathematical strong induction . let @xmath4 be an integer , @xmath305 a degree - graded basis satisfying , and @xmath306,\ldots , b_1[p]$ ] the clenshaw shifts satisfying . then , for @xmath406 , @xmath407(x ) = ( \alpha_nx+\beta_n)b_{j}[\phi_{n}](x ) + \sum_{s = j+1}^n\gamma_{n , s}b_{j}[\phi_{s-1}](x).\ ] ] [ lem : newrecurrence ] we proceed by induction on @xmath408 . let @xmath409 . we have , by , @xmath410(x ) = ( \alpha_nx + \beta_n)b_{n+1}[\phi_{n+1}](x ) = ( \alpha_nx + \beta_n)b_{n}[\phi_{n}](x),\ ] ] where the last equality follows because @xmath411(x ) = b_n[\phi_n](x ) = 1 $ ] . now , suppose the result holds for @xmath412 . we have , by and the inductive hypothesis , @xmath413(x ) & = ( \alpha_kx + \beta_k)b_{k+1}[\phi_{n+1}](x ) + \sum_{j = k+1}^{n } \gamma_{j , k+1}b_{j+1}[\phi_{n+1}](x)\\ & = ( \alpha_kx + \beta_k)\left[(\alpha_nx+\beta_n)b_{k+1}[\phi_{n}](x ) + \sum_{s = k+2}^n \gamma_{n , s}b_{k+1}[\phi_{s-1}](x)\right ] \\ & \hspace{1cm}+ \sum_{j = k+1}^{n-1 } \gamma_{j , k+1}\left[(\alpha_nx+\beta_n)b_{j+1}[\phi_{n}](x ) + \sum_{s = j+2}^n \gamma_{n , s}b_{j+1}[\phi_{s-1}](x)\right ] \\ & \hspace{2cm}+ \gamma_{n , k+1}b_{n+1}[\phi_{n+1}](x).\\ \end{aligned}\ ] ] by interchanging the summations and collecting terms we have @xmath414(x ) & = ( \alpha_nx+\beta_n)\left[(\alpha_kx + \beta_k)b_{k+1}[\phi_n](x ) + \sum_{j = k+1}^{n-1}\gamma_{j , k+1}b_{j+1}[\phi_n](x)\right ] \\ & \hspace{.6 cm } + \sum_{s = k+3}^n \gamma_{n , s}\left[(\alpha_kx + \beta_k)b_{k+1}[\phi_{s-1}](x ) + \sum_{j = k+1}^{s-2}\gamma_{j , k+1}b_{j+1}[\phi_{s-1}](x)\right]\\ & \hspace{2cm}+ \gamma_{n , k+2}(\alpha_kx+\beta_k)b_{k+1}[\phi_{k+1}](x ) + \gamma_{n , k+1}b_{n+1}[\phi_{n+1}](x)\\ & = ( \alpha_nx+\beta_n)b_k[\phi_n ] + \sum_{s = k+1}^n \gamma_{n , s } b_k[\phi_{s-1 } ] , \end{aligned}\ ] ] where in the last equality we used , @xmath415(x ) = b_k[\phi_{k+1}](x)$ ] , and @xmath411(x ) = b_k[\phi_k](x ) = 1 $ ] . * case 1 : @xmath416 . * since for a fixed @xmath417 the clenshaw shifts are linear , i.e. , @xmath418(y ) = c_1b_j[\phi_i](y ) + c_2 b_j[\phi_k](y)$ ] for constants @xmath419 and @xmath420 , it is sufficient to prove the theorem for @xmath421 for @xmath422 . assume that the result holds for @xmath425 . from the inductive hypothesis , we have @xmath426(y)\phi_j(x ) \\ & \qquad\qquad\qquad\qquad\qquad\qquad+ \sum_{j=1}^k\gamma_{k , j}\sum_{s=0}^{j-2}\alpha_sb_{s+1}[\phi_{j-1}](y)\phi_s(x ) . \end{aligned}\ ] ] moreover , by interchanging the summations and collecting terms we have @xmath427(y)\phi_{k-1}(x ) \\ & \!\!\!+ \sum_{j=0}^{k-2}\alpha_j\left[(\alpha_kx+\beta_k)b_{j+1}[\phi_k](y ) + \sum_{s = j+2}^k\gamma_{k , s}b_{j+1}[\phi_{s-1}](y)\right]\phi_j(x ) . \end{aligned}\ ] ] finally , since @xmath428(y ) = 1 $ ] , @xmath429(y ) = ( \alpha_kx+\beta_k)b_k[\phi_k](y)$ ] , and by , we have @xmath430(y)\phi_k(x ) + \alpha_{k-1}b_{k}[\phi_{k+1}](y)\phi_{k-1}(x ) \\ & \qquad\qquad\qquad + \sum_{j=0}^{k-2 } \alpha_jb_{j+1}[\phi_{k+1}](y)\phi_j(x ) \end{aligned}\ ] ] and the result follows by induction . 3 , _ the method of resultants for computing real solutions of polynomial systems _ , siam j. numer . , 29 ( 1992 ) , pp . 831844 . , _ a numerical method for polynomial eigenvalue problems using contour integral _ , japan j. indust . appl . math . , 27 ( 2010 ) , pp . , _ numerically solving polynomial systems with bertini _ , siam , 2013 . , _ computing curve intersection by means of simultaneous iterations _ , numer . algor . , 43 ( 2006 ) , pp . 151175 . , _ solving polynomial eigenvalue problems by means of the ehrlich - aberth method _ , linear algebra appl . , 439 ( 2013 ) , pp . 11301149 . , _ solution of a polynomial system of equations via the eigenvector computation _ , linear algebra appl . , 319 ( 2000 ) , pp . 193209 . , _ computing zeros on a real interval through chebyshev expansion and polynomial rootfinding _ , siam j. numer . , 40 ( 2002 ) , pp . 16661682 . , _ solving transcendental equations : the chebyshev polynomial proxy and other numerical rootfinders _ , siam , 2014 . , _ an algorithm for finding the basis elements of the residue class ring of a zero dimensional polynomial ideal _ , j. symbolic comput . , 41 ( 2006 ) , pp 475511 . , _ resultant - based methods for plane curves intersection problems _ , computer algebra in scientific computing , springer berlin heidelberg , 3718 ( 2005 ) , pp . 7592 . , _ introduction to residues and resultants _ , in , a. dickenstein and i. z. emiris , editors , _ solving polynomial equations . foundations , algorithms , and applications _ , springer , 2005 . , _ on the theory of elimination _ , cambridge and dublin math . j. iii , ( 1848 ) , pp . , _ fast computation of the bzout and dixon resultant matrices _ , j. symb . comput . , 33 ( 2002 ) , 1329 . , _ a note on the summation of chebyshev series _ , math . comput . , 9 ( 1955 ) , 118120 . , _ using algebraic geometry _ , springer , 2013 . , _ fiedler companion linearizations and the recovery of minimal indices _ , siam j. mat . appl . , 31 ( 2010 ) , pp . 21812204 . , _ back to the roots : polynomial system solving , linear algebra , systems theory _ , proc . 16th ifac symposium on system identification ( sysid ) , 2012 , pp . 12031208 . , _ sparse elimination and applications in kinematics _ , dissertation , university of california , berkeley , 1994 . , _ symbolic and numeric methods for exploiting structure in constructing resultant matrices _ , j. symbolic comput . , 33 ( 2002 ) , pp . 393413 . , _ discriminants , resultants , and multidimensional determinants _ , springer , birkhuser , boston , 2008 . , _ the ehrlich aberth method for palindromic matrix polynomials represented in the dickson basis _ , linear algebra appl . , 438 ( 2013 ) , pp . , _ matrix polynomials _ , siam , philadelphia , usa , 2009 , ( unabridged republication of book first published by academic press in 1982 ) . , _ the colleague matrix , a chebyshev analogue of the companion matrix _ , the quarterly journal of mathematics , 12 ( 1961 ) , pp . , _ accuracy and stability of numerical algorithms _ , siam , 2nd edition , 2002 . , _ function of matrices : theory and computation _ , siam , 2008 . , _ matrix analysis _ , cambridge university press , 2nd edition , new york , 2013 . , _ accurate solution of polynomial equations using macaulay resultant matrices _ , math . , 74 ( 2005 ) , pp . 221262 . , _ comparison of various multivariate resultant formulations _ , acm proceedings of the 1995 international symposium on symbolic and algebraic computation , 1995 . , _ complex algebraic curves _ , cambridge university press , cambridge , 1992 . , _ a simple solution to the six - point two - view focal - length problem _ , computer vision eccv , springer berlin heidelberg , ( 2006 ) , pp . 200213 . , _ vector spaces of linearizations for matrix polynomials _ , siam j. matrix anal . appl . , 28 ( 2006 ) , pp . , _ multipolynomial resultants and linear algebra _ , papers from the international symposium on symbolic and algebraic computation . acm , 1992 . , _ algorithms for intersecting parametric and algebraic curves i : simple intersections _ , acm trans . graphics , 13 ( 1994 ) , pp . _ vector spaces of linearizations for matrix polynomials : a bivariate polynomial approach _ , submitted . , _ computing the common zeros of two bivariate functions via bzout resultants _ , numer . , 129 ( 2015 ) , pp . , _ block tensor unfoldings _ , siam j. mat . anal . appl . , 33 ( 2012 ) , pp . 149169 . , _ polynomial evaluation and associated polynomials _ , numer . , 79 ( 1998 ) , pp . 601613 . , _ the numerical solution of systems of polynomials arising in engineering and science _ , world scientific , 2005 . , _ numerical solution of bivariate and polyanalytic polynomial systems _ , siam j. numer . , 52 ( 2014 ) , pp . 15511572 . , _ finding all real zeros of polynomial systems using multi - resultant _ , j. comput . , 167 ( 2004 ) , pp . 417428 . , _ an algorithm for quadratic eigenproblems with low rank damping _ , siam j. matrix anal . appl . , 36 ( 2015 ) , pp . , _ backward error and condition of polynomial eigenvalue problems _ , linear algebra appl . , 309 ( 2000 ) , pp . 339361 . , _ the quadratic eigenvalue problem _ , siam review , 43 ( 2001 ) , pp . 235286 . , _ computing with functions of two variables _ , dphil thesis , the university of oxford , 2014 . , _ an extension of chebfun to two dimensions _ , siam j. sci . comput . , 35 ( 2013 ) , c495c518 . , _ resultant methods for the inverse kinematics problem _ , computational kinematics , springer netherlands , 28 ( 1993 ) , pp .

interactions among individuals in a population can be described by networks of who - contacts - whom . studies of contact networks in sexually transmitted diseases have long revealed a high variability in the number of contacts per individual and highlighted the importance of those individuals described as `` super - spreaders '' for the onset of an epidemic @xcite . similar conclusions about the importance of super - spread events were drawn from contact tracing data collected from recent epidemic outbreaks of airbone - transmitted diseases like those of the severe acute respiratory syndrome ( sars ) in 2002 and 2003 @xcite . on the other hand , the risk perception among people during an epidemic outbreak triggers behavioural responses to lower the risk of contagion @xcite , the avoidance of contacts with infected individuals being an example of such responses @xcite . this sort of social distancing led to the idea of disease - avoiding link rewiring and is one of the basis of the so - called adaptive or dynamic networks . such a preventive rewiring assumes transmission of information which allows people to gather knowledge about the disease status of their neighbours . therefore , in such networks the contact pattern is no longer static but evolves with the spread of an infectious disease according to the rules defining the rewiring process @xcite . pairwise models have been the main approach adopted for the analysis of epidemic dynamics on adaptive networks @xcite . this class of models was initially developed to deal with processes defined on regular ( random ) networks and offers a good description of their dynamics . in their classic formulation and over heterogeneous networks , however , their accuracy is far from being satisfactory , especially for its prediction of the epidemic threshold . the so - called effective degree models are extensions of them with a higher accuracy in their predictions ( but also with a higher complexity ) . in these models , in addition to the disease status of nodes , the number of neighbours for each status is also considered @xcite . at the individual level , pair - based epidemic models have been developed in terms of master equations for the probabilities of the individual pairs @xcite . as with effective degree models , they show a higher accuracy than pairwise models formulated at the population level , but at the price of a higher computational complexity @xcite . to analyze epidemic outbreaks on dynamic networks , stochastic models have also been used . if the contact network has a large size and no cycles , it can be locally described as a tree and the initial phase of an epidemic can be approximated by a branching process @xcite . an example of a stochastic model defined on a dynamic network is the one developed in @xcite . the model assumes that , at a given rate , the identities of neighbours change stochastically by means of an instantaneous edge swap between a randomly selected pair of links . so , this neighbour - exchange mechanism is independent of the epidemic dynamics because it does not depend on the disease status of the involved nodes . in other words , it is not an example of behavioural response against the presence of the disease . other models defined on dynamic networks whose architecture evolves by random edge swapping can be found in @xcite . this paper aims mainly at comparing the predictions from both modelling methodologies ( pairwise / stochastic ) for the initial phase of susceptible - infectious - recovered ( sir ) and susceptible - exposed - infectious - recovered ( seir ) epidemics with preventive rewiring among individuals ( so , with an interplay between the spread of the disease and the rewiring process , that is , between disease s dynamics and network dynamics ) . in particular , in the sir model we will assume that susceptible individuals break off connections with infectious neighbors at a given rate @xmath0 and , in place of them , new connections to susceptible and recovered individuals are created with probability @xmath1 . as for the seir model , we consider two alternative scenarios for the dynamics of exposeds ( i.e. , infected but not infectious individuals ) . in the first one , exposed individuals break off with their infectious neighbors at a rate @xmath6 and , with probability @xmath1 , they reconnect to any non - infectious individual in the population . in turn , susceptibles can also reconnect , with the same probability @xmath1 , to exposeds ( in addition to other susceptibles and removed individuals ) when breaking off with infectious neighbors . in the second scenario , exposed individuals do not rewire at all ( @xmath7 ) , and susceptibles rewire away from both exposeds and infectives , and create new connections with probability @xmath1 . from a modeling viewpoint , the rewiring scenario can depend on whether exposed individuals realize they have been infected ( for instance , because they show symptoms ) or not ( they are asymptomatic ) . in both scenarios , the degree distribution changes over time and its mean degree is preserved only when @xmath8 . the introduction of a reconnection probability @xmath1 allows us to consider different degrees of rewiring , ranging from a situation where each deleted link is replaced by a newly created one ( @xmath8 ) to the limit case where no new connection is made and edges are simply deleted ( @xmath2 ) . in other words , @xmath1 can be though of a measure of the intensity of social distancing of rewiring individuals . as it is claimed in @xcite , people value person - to - person contacts and are willing to accept some disease risk to gain contact - related benefits . so , different values of @xmath1 could be considered according to the type of social relationship modeled by the network . the basic reproduction number @xmath3 , namely , the average number of infections produced by a _ typical _ infectious individual when the _ fraction _ infected is still negligible , is one of the compared quantities . its predicted value will be checked against stochastic simulations carried out on contact networks with degree distributions that follow a poisson distribution and a power law , respectively . it is worth noting that , while the meaning of @xmath3 in randomly mixing homogeneous populations is straightforward because any infectious individual is as likely as any other to infect a susceptible one , in heterogeneous networks it requires that one specifies the meaning of typical individual @xcite . in most network models ( in particular , for those without multiple levels of mixing ) , this definition implies that one has to compute @xmath3 from the average number of infections per infective once the early correlations of disease status around infectious individuals have been formed , which takes a couple of generations after the occurrence of the primary cases . interestingly , this computation / redefinition of @xmath3 has been obtained under both previous modelling approaches @xcite . in fact , it is well known that both pairwise models and branching process approximations lead to the same epidemic ( or invasion ) threshold in networks without rewiring @xcite . in section [ sec - model ] we present the sir and seir epidemic models with rewiring , here denoted by sir-@xmath0 and seir-@xmath0 respectively . in section [ sir - branching ] we use the branching process approximation to analyse the early phase of an sir-@xmath0 epidemic . in particular , we compute @xmath3 and the expected degree of infectives during this initial phase as a function of the rewiring rate @xmath0 . in section [ seir - branching ] the same approximation is applied to the study of the early stage of an seir-@xmath0 epidemic under different types of rewiring processes . section [ pairwise ] contains the results of the initial phase obtained for both epidemic models using the pair approximation with the triple closure introduced in @xcite for heterogeneous networks . in particular , the seir-@xmath0 pairwise model extends the one considered in @xcite to account for the rewiring of exposed individuals and the possibility that susceptibles reconnect to exposed ones after breaking off an infectious link . in section [ simus ] numerical estimates of @xmath3 and the mean degree of infectives during the initial phase are obtained from continuous - time stochastic simulations on heterogeneous networks . finally , in section [ discuss ] we discuss the analytical results obtained from both approximations and compare them to the output of the stochastic simulations . moreover , we comment about the new insight into the role of the rewiring process in the seir-@xmath0 epidemic model . let us define our stochastic network epidemic model . the population consists of a fixed number @xmath9 of individuals and the stochastic network model is given by the configuration model with degree distribution @xmath10 having finite mean @xmath11 and finite variance @xmath12 ( e.g. @xcite ) . this model is defined by all individuals having i.i.d . degrees @xmath13 and edge - stubs being pairwise connected completely at random with any loop or multiple edge being removed making the graph simple . we are primarily interested in the situation where @xmath9 is large , and the approximations will rely on this . on this network we now define an epidemic model where susceptible individuals may rewire if they are neighbors of infectious individuals . we start by defining an sir epidemic where infected people immediately become infectious and later recover , and then extend the model to an seir model in which infected people are at first exposed ( latent ) , then they become infectious , and eventually they recover . the latter model is a bit more complicated in that now the rate of rewiring could differ depending on whether the person rewiring is susceptible or exposed , and depending on whether the person he / she rewires away from is exposed or infectious . note that such a rewiring process causes the degree distribution to change over time . however , since we will focus our analysis on the early stage of an epidemic outbreak where only a tiny fraction of individuals is initially infected , we will assume that the degree distribution @xmath14 does not change significantly during the initial phase of the epidemic . in this model individuals are at first susceptible . if an individual gets infected he / she immediately becomes infectious , and after some random time he / she recovers and becomes immune for the rest of the outbreak . sir hence stands for susceptible - infectious - recovered ( e.g. @xcite for more on sir and seir epidemic models ) . the sir-@xmath0 model is defined on the network ( described above ) as follows . initially one randomly selected individual is infectious and the rest are susceptible . an infectious individual transmits the disease to each of its susceptible neighbors at a rate @xmath15 , and the infectious periods are i.i.d . following an exponential distribution with rate parameter @xmath16 ( so infectious individuals recover at a rate @xmath16 ) further , susceptible individuals that are neighbors with infectious ones break off with such neighbors independently at rate @xmath0 and , with probability @xmath1 , replace each lost connection by reconnecting to a randomly selected non - infectious ( i.e. , susceptible or recovered ) individual in the community . therefore @xmath17 is the effective rewiring rate , i.e. , the rate at which new links are created by susceptibles in substitution for those previously deleted . the sir-@xmath0 network epidemic has the following parameters : @xmath15 ( infection rate ) , @xmath16 ( recovery rate ) , @xmath0 ( rewiring rate ) , @xmath1 ( reconnection probability ) , and the degree distribution @xmath14 with mean @xmath11 and variance @xmath12 . we will focus on what happens early on in the epidemic , before a substantial fraction of the community has been infected . in the seir-@xmath0 model an infectious individual transmits the disease to each of its susceptible neighbors at a rate @xmath15 , but , when this happens , the neighbor first becomes exposed ( or latent ) and can transmit the disease only after a time delay . in other words , such an exposed individual becomes infectious at a rate @xmath18 , and at this moment can start infecting each of its susceptible neighbors at a rate @xmath15 . as before , infectious individuals recover at a rate @xmath16 . as regards to rewiring , it can be modelled differently depending on when an individual starts and stops having a rewiring rate and also depending on which individuals it rewires away from . our model considers three different rewiring rates : a susceptible individual rewires away from each exposed neighbor at a rate @xmath19 , a susceptible individual rewires away from each infectious neighbor at a rate @xmath20 , and an exposed individual rewires away from each infectious neighbor at a rate @xmath6 . it is of course possible to also allow for rewiring from other states , e.g. that susceptible individuals rewire from recovered individuals , but since we are primarily focusing on the initial phase such rewiring would have negligible effects . if infected people are detected only when they become infectious ( e.g. if the latent period is the same as the incubation period ) , and hence individuals rewire away from infected neighbors only when those become infectious and not while they are exposed , this would correspond to @xmath21 and @xmath22 : susceptible individuals are not aware of exposed ( latent ) neighbors being infected and hence do not rewire , and susceptible but also exposed ( latent but unaware ) individuals rewire away from the infectious ones . we could obtain a second scenario if also exposed ( latent ) individuals are known to have been infected ( e.g. by contact tracing or because they show some symptoms ) . we would then have @xmath23 and @xmath7 : susceptible individuals rewire away both from exposed and infectious neighbors but exposed individuals do not rewire away since they know they have already been infected . in all cases , the reconnection probability @xmath1 modulates the fraction of rewirings that are effectively done , and it is assumed to be the same both for susceptibles and for exposeds . therefore , @xmath24 is the effective rewiring rate of individuals in state @xmath25 away from individuals in state @xmath25 . the seir-@xmath0 network epidemic has all the parameters of the sir-@xmath0 epidemic except that the rewiring rate @xmath0 now becomes three different rates : @xmath19 , @xmath20 and @xmath6 , and there is a rate @xmath18 at which latent individuals become infectious . most stochastic epidemic models allow for a branching process approximation of the early stages of an outbreak , an approximation which can be made rigorous as the population size @xmath9 tends to infinity ( e.g. @xcite ) . this applies also to network epidemics we now describe the approximation of the current model . we derive expressions for the basic reproduction number @xmath3 , here denoted by @xmath26 to distinguish its expression from the one obtained using pair approximation . we also derive the exponential growth rate @xmath27 ( the malthusian parameter ) for the situation that @xmath28 , and the average degree of infected individuals . since rewiring is a focus of this paper , we look at both the degree of newly infected individuals as well as on the average of all infectious individuals , the latter expected to be smaller than the former since individuals rewire away from infectious neighbors . recall that @xmath3 is defined as the mean number of new infections caused by a typical infected individual during the early stage of the epidemic . individuals that get infected during the early stage will , at the time of infection , have the size biased degree distribution of neighbors , @xmath29 , where @xmath30 ( e.g. @xcite ) . one of the neighbors is its infector whereas the remaining neighbors , with large probability during the early stage of an outbreak , will be susceptible . when considering the disease progress it is only the @xmath31 susceptibles that are of interest since it is not possible to reinfect the infector . from the derivation above , the mean number of susceptible neighbours a typical new infectee has during the early stages equals @xmath32 ( e.g. @xcite ) . the probability to infect a given such neighbor is obtained by considering the competing events that may happen : there could be an infection ( rate @xmath15 ) , the neighbor could rewire away from the infectee ( rate @xmath0 ) , or the infectee can recover ( rate @xmath16 ) . the probability of infection hence equals @xmath33 . the basic reproduction number equals the mean degree multiplied by the transmission probability , i.e. @xmath34 if there is no rewiring ( @xmath35 ) , the basic reproduction number equals @xmath36 as is well known . therefore , the rewiring reduces @xmath3 , as expected . we note that @xmath37 is independent of @xmath1 , so it has no effect on the beginning of an outbreak if rewired edges are dropped , always attached to new susceptible individuals or a mixture of two . during the early stage and assuming a large population , the number of infectives in the epidemic will asymptotically ( as @xmath38 ) evolve like a branching process . @xmath3 is the corresponding mean offspring distribution . another important quantity associated to this branching process is @xmath39 , the expected birth rate ( rate of new infections ) of an infectee having age @xmath40 , where age corresponds to time since infection . we now derive @xmath39 which in turn will help us derive the exponential growth rate of the epidemic . as derived earlier , the average number of susceptible neighbors upon infection is @xmath41 , and the infectee will infect each neighbor independently . the average rate of infection ( = birth ) for each neighbor is obtained by considering what must be fulfilled for infection to happen . in order to infect a neighbor @xmath40 time units after infection , the infectee must still be infectious , the neighbor should not have rewired , and the infectee should not yet have infected the neighbor . given this , the infection rate equals @xmath15 . since all events are assumed to follow an exponential distribution this gives us the following expression for @xmath42 : @xmath43 the average total number of births ( infections ) is hence @xmath44 as it should be . the mean birth rate @xmath42 also determines the exponential growth rate of the epidemic , i.e. for which @xmath27 , @xmath45 ( cf . this @xmath27 , the malthusian parameter , is given by the solution of the euler - lotka equation @xmath46 for our model this gives us , after a bit of algebra , @xmath47 also the exponential growth rate is independent of @xmath1 . we now turn to the mean degree of infectives during the early stages of the epidemic . we consider two different means . the first one is for newly infected , which in fact has already been shown : during the early stages newly infected individuals will have degree distribution @xmath48 ( when considering the degree distribution we also count the non - susceptible infector ) , so the mean degree of newly infected individuals equals @xmath49 . the second mean , @xmath4 , denotes the average number of neighbors of _ all _ infectives during the early stages , not only that of the newly infected ones . as described above , during the early stage of an outbreak the degree distribution of newly infected equals @xmath48 . however , while still infectious , an individual loses susceptible neighbors by rewiring : each susceptible neighbor is lost at a rate @xmath0 . the probability that a susceptible neighbor @xmath50 is still a neighbor ( i.e. has not rewired ) @xmath40 time units after our individual @xmath51 was infected and given that @xmath51 is still infectious , is obtained by conditioning on the potential infection time of the neighbor ( @xmath50 only rewires if not yet infected ) . so , we have @xmath52 note that we condition on that @xmath51 remains infectious at @xmath40 . when deriving the degree distribution of all infectives during the early stage , we have to take into account both this decrease of degree with age , but also the fact that , in the exponential phase of the epidemic ( recall that @xmath45 ) , young infectives will be over - represented . the ratio of individuals infected @xmath53 time units ago over the number of individuals infected at present equals @xmath54 due to the exponential growth rate . and only a fraction @xmath55 of them are still infectious at present . consequently , the fraction of infectives that were infected @xmath53 time units ago or longer equals @xmath56 , so the age distribution of infectives is exponential with parameter @xmath57 ( this is the so - called _ stable age distribution _ of this branching process , @xcite ) . the mean degree of all infectives during the early stage is obtained by conditioning on their age : @xmath58 where , as before , @xmath59 , and @xmath60 was defined in equation ( [ r - sir ] ) . this mean degree should be valid after a couple of generations and will then change as the depletion of susceptibles will start affecting things . as seen in ( [ e(d_i ) branching ] ) also the mean degree of infectives is independent of @xmath1 . at first this might seem surprising since the degree of infectious individuals are affected by rewiring . however , an infectious individual can loose edges ( which reduces the degree ) due to susceptible neighbours rewiring away from the infective , but it does not affect the degree of the infective whether these links are simply dropped or the rewiring susceptible connects to new individuals . we now study the extended seir model recalling that individuals who get infected are now first latent for an exponentially distributed time with rate parameter @xmath18 , after which they become and remain infectious according to earlier rules . individuals rewire away from infected neighbors . more precisely , a susceptible individual rewires from each exposed ( latent ) neighbor at a rate @xmath19 and from each infectious neighbor at a rate @xmath20 . moreover , exposed individuals rewire away from infectious neighbors at a rate @xmath6 . of course , some of these rewiring intensities may be zero ( cf . sec 2.2 ) . as before , upon each rewiring event the individual reconnects to a randomly chosen susceptible or recovered individual with probability @xmath1 and with the remaining probability the edge is simply dropped . during the early stage of an outbreak , at the time of infection an exposed individual @xmath61 has degree distribution @xmath48 as before , one neighbor @xmath25 being the infector and the remaining neighbors being susceptible . so , @xmath61 has @xmath62 expected susceptible neighbors . however , in the seir model this number can eventually increase by one , provided that @xmath61 can rewire away from its infector @xmath25 ( if not yet recovered ) and reconnect to a susceptible individual that will become a new neighbor . the probability for this to happen is @xmath63 . in consequence , the expected number of susceptible neighbors for @xmath61 is @xmath64 . any such neighbor , say @xmath53 , will get infected if @xmath61 first becomes infectious ( before @xmath53 rewires away from @xmath61 ) and , then , an infection occurs ( before @xmath61 recovers or @xmath53 rewires from @xmath61 ) . hence , the probability for this to happen is the product of the probabilities of these two events , namely , @xmath65 . to conclude , the expected number of neighbors infected by @xmath61 equals : @xmath66 by studying eq . ( [ r0-seir ] ) we make the following observations . if there is no rewiring from any state , @xmath26 reduces to @xmath67 , i.e. the same as for the sir case . further , @xmath26 is decreasing in both @xmath19 and @xmath20 as expected . however , @xmath26 _ increases _ with @xmath1 and @xmath6 . if @xmath21 and @xmath68 , the perhaps most realistic example discussed in sec . 2.2 , then @xmath26 can be increasing in @xmath0 for some parameter set - ups , implying that the quicker individuals rewire the larger epidemic outbreak ! the explanation to this is that , when @xmath69 , the exposed ( latent ) individuals can rewire away from their infector to a susceptible neighbor , with the effect that they may later ( once they become infectious ) infect the new susceptible neighbor . as with the sir-@xmath0 model , in order to compute the malthusian parameter @xmath27 we first derive an expression for @xmath42 , the average rate at which an individual , who was infected during the initial phase of the epidemic , infects new individuals @xmath40 time units after his / her time of infection . at the time when an individual gets infected he has on average @xmath70 neighbors , one being infectious ( its infector ) and the remaining @xmath71 will , with large probability since we are in the beginning of the epidemic , be susceptible . at a rate @xmath15 , the infected individual infects each of the @xmath71 initially susceptible neighbors @xmath40 time units after infection if the following conditions are fulfilled : the infected individual must have terminated the latent period without the neighbor having rewired , and after that the infectious period should still be active , an infection should not yet have taken place , and the neighbor should not have rewired . as mentioned in the previous subsection , it is also possible that the infected individual infects through the link to the infector . this happens with a rate @xmath15 at @xmath40 time units after infection if the following holds : the infected individual rewired from its infector ( and connects to a susceptible neighbor ) while still latent , and after this the infected individual has become infectious , has not yet infected the neighbor nor has the new neighbor rewired . the above reasoning leads to the following expression for @xmath42 : @xmath72 by conditioning on the end of the latency period , the first probability equals @xmath73 the second probability , now conditioning both on the time of first rewiring and the end of the latent period , equals @xmath74 using these expressions in equation ( [ lambda_seir ] ) and solving the integrals results in the following expression for @xmath39 : @xmath75 we now use @xmath42 to derive the exponential growth rate @xmath27 of the epidemic in case it takes off , and also to confirm our expression for @xmath3 . the latter is easy . if we compute @xmath76 using the expression above we get exactly @xmath3 as defined in eq . ( [ r0-seir ] ) , as it should be . as for the malthusian parameter @xmath27 this is given as the solution to the equation @xmath77 . for the expression of @xmath39 above , this can be shown to be equivalent to @xmath78 for @xmath69 , this is a third order equation , but for positive values of @xmath27 ( the relevant values as we assume @xmath28 ) the left hand side is decreasing in @xmath27 , starting from a value larger than 1 when @xmath79 and decreasing to 0 as @xmath80 implying that there is a unique solution to the equation . equation ( [ r - seir ] ) is not explicit but it is still possible to see how various parameters affect the growth rate . for example , @xmath27 is increasing in the infection rate @xmath15 and the mean degree @xmath70 . as regards to the rewiring rates , @xmath27 decreases in @xmath20 and @xmath19 but increases with the `` harmful '' rewiring rate @xmath81 . finally , as we increase the rate to leave the latent state ( i.e. making the latent state shorter ) , the effect depends on other parameter values , but if we increase @xmath18 towards infinity it can be shown that we obtain the expression for @xmath27 of the sir-@xmath0 model ( cf . ( [ r - sir ] ) ) as expected . for the seir-@xmath0 model it is also possible to derive specific features of infected individuals during the initial phase of an epidemic . for example , as we did for the sir-@xmath0 case , we can compute @xmath4 , the average degree of infectives during the early stage of the outbreak . however , other mean quantities might be equally relevant as , for instance , the mean degree @xmath82 of infected but still latent individuals , or the mean degree @xmath83 of either latent or infectives , or for that matter the expected number of susceptible neighbors while in one of these states . for brevity and because some of these quantities have even more complicated expressions , we compute @xmath4 without solving the integrals appearing in the derivation , and indicate how to modify the derivations if we want to compute another mean . to compute @xmath4 , let us pick at random an infectious individual @xmath25 during the early stage of an outbreak and let @xmath84 denote how long ago this individual was infected . we first compute the expected degree of @xmath25 conditional upon @xmath85 , denoted by @xmath86 . this is done by conditioning on the duration of the latent period @xmath87 , which must lie between 0 and @xmath40 since @xmath25 is infectious @xmath40 time units after infection : @xmath88 the second factor is given by @xmath89 . as for the first factor , the individual has @xmath70 expected neighbours at the time of infection , one being infectious and the rest being susceptible . for the susceptible neighbors we compute the probability that they are still neighbors . for the infector , it could have lost a neighbor from this edge only if @xmath25 rewired away from the infector to a susceptible neighbor _ and _ the new neighbor later rewired away from @xmath25 . we hence get @xmath90 the first probability in ( [ cond - mean ] ) is obtained by conditioning on whether the susceptible neighbor was infected or not , and , in the former case , whether the latent period ended before @xmath40 or not : @xmath91 here there should be no @xmath1 in front of @xmath6 because the event concerns an exposed neighbor and it is irrelevant for the degree of the infective whether this neighbor reconnects or not upon rewiring . the second probability in ( [ cond - mean ] ) consists of two events . either @xmath25 rewired away from its infector and dropped the edge while exposed , or else @xmath25 rewired away from infector and reconnected while exposed , and the new neighbor rewired away from @xmath25 . the first event has probability @xmath92 the second event is obtained by conditioning on the time when @xmath25 rewires away from the infector and reconnects to a susceptible neighbor , whether the second rewiring happens during the latency or infectious period of @xmath25 , and in the latter case whether infection takes place or not : @xmath93 note that @xmath1 appears in front of @xmath6 only when it concerns the infective , because then it has to reconnect to make spreading to a new individual possible , whereas there is no @xmath1 when the rewiring refers to exposed neighbors rewiring away from the infective . in the latter case it is irrelevant for the degree of the infective whether or not the exposed individual reconnects upon rewiring . it remains to derive the distribution for @xmath84 , the time since infection . for this we know that due to the exponential growth rate @xmath27 of infectives , there is a fraction @xmath94 to choose from @xmath40 units earlier as compared to present time . however , we also require that the individual is infectious at present , an event which happens with probability @xmath95 . the probability density of @xmath84 is hence proportional to @xmath96 , which after a bit of algebra gives the following density @xmath97 finally , the expected degree of a randomly chosen infective during the early stages is obtained by integrating with respect to this density : @xmath98 in order to compare this predicted @xmath4 with the one obtained from pair approximation ( see next section ) , for each set of values of the parameters we obtain the value of @xmath27 given by the positive solution of and evaluate the resulting expression of the previous integral . if we were to compute e.g. the average degree of a latent individual @xmath82 , we would similarly condition on the time @xmath84 since infection of the randomly chosen latent individual . this individual would have @xmath70 neighbours at the time of infection , one infectious and the rest susceptible , and we need to compute the probability that these neighbours would not have been lost similarly to what we did before . we would then integrate this expected value with respect to the probability density of @xmath84 , which is proportional to @xmath99 . we now derive expressions for @xmath3 and @xmath4 using an alternative deterministic approximation based on the closed - form equations for the dynamics of pairs of disease status , the so - called pairwise models . while the sir-@xmath0 pairwise model was already introduced in @xcite , the seir-@xmath0 pairwise model is a generalization of the one also introduced in @xcite that includes rewiring of exposeds and reconnection rules introduced in accordance with it . this extended model will allow for a better understanding of the impact of rewiring on @xmath3 derived under this approach , here denoted by @xmath100 . let @xmath101 $ ] , @xmath102 $ ] and @xmath103 $ ] be the expected number of susceptible , infectious , and recovered individuals , respectively . moreover , let @xmath104 $ ] be the expected number of non - ordered connected _ @xmath25-@xmath105 pairs _ , i.e. pairs whose individuals are in states @xmath25 and @xmath105 , and let @xmath106 $ ] the expected number of connected non - oriented @xmath25-@xmath105-@xmath107 triples ( @xmath108 ) . so , if @xmath9 is the network size and @xmath109 is the total number of links in the network , then @xmath101+[i]+[r]=n$ ] and @xmath110+[si]+[sr]+[ii]+[ir]+[rr]=l$ ] . the sir-@xmath0 model formulated in terms of triplets can be closed by assuming the statistical independence at the level of pairs which leads to the following approximations for the expected number of the involved triples : @xmath111 \approx ( e(\tilde{d } ) -1 ) 2[ss][si]/(\mu [ s ] ) $ ] , @xmath112 \approx 1/2 \cdot ( e(\tilde{d})-1 ) [ si]^2/(\mu [ s])$ ] , and @xmath113 \approx ( e(\tilde{d})-1 ) [ si][sr]/(\mu [ s])$ ] with @xmath114 being the average degree ( see @xcite for details ) . note that , since we focus our analysis on the early epidemic stage with a very small number of initially infectious nodes , we approximate the expected degree of the susceptible central node of a triple by @xmath115 , the mean degree of a node reached by following a randomly chosen link in a wholly susceptible population at @xmath116 , i.e. , when the degree distribution is the initial one . upon introducing the triple closure into the original model , the initial dynamics of the sir-@xmath0 model is determined by @xmath117 & = & - \beta [ si ] , \nonumber \\ \frac{d}{dt}\,[i ] & = & \beta [ si ] - \gamma [ i ] , \nonumber \\ \frac{d}{dt}\,[si ] & = & \left ( \beta z \left ( \frac{2[ss]}{[s ] } - \frac{[si]}{[s ] } \right ) - \beta - \gamma - \omega \right ) [ si ] , \nonumber \\ \frac{d}{dt}\,[ii ] & = & \beta \left ( 1 + z \frac{[si]}{[s ] } \right ) [ si ] - 2 \gamma [ ii ] , \label{sir - pa}\\ \frac{d}{dt}\,[ss ] & = & \alpha \ , \omega \frac{[s]}{n-[i ] } [ si ] - \beta z \frac{2[ss]}{[s ] } [ si ] , \nonumber \\ \frac{d}{dt}\,[sr ] & = & \left ( \gamma + \alpha \ , \omega \frac{[r]}{n-[i ] } \right ) [ si ] - \beta z \frac{[is]}{[s]}[sr ] , \nonumber \\ \frac{d}{dt}\,[ir ] & = & 2 \gamma [ ii ] - \gamma [ ir ] + \beta z \frac{[is]}{[s]}[sr ] , \nonumber\end{aligned}\ ] ] where @xmath118 . note that susceptible individuals break off ( at a rate @xmath0 ) with infectious neighbours and reconnect ( at a rate @xmath119 ) to other susceptibles or to recovered individuals with probability @xmath101/(n-[i])$ ] and @xmath103/(n-[i])$ ] , respectively , in place of the deleted links . the average number of susceptible individuals around an infective is @xmath120/[i]$ ] . if this quantity stabilizes to a value @xmath121/[i])^*$ ] during the initial exponential growth of an epidemic , then we can use it to compute @xmath3 as ( @xcite ) @xmath122}{[i ] } \right)^ * . \label{r0_pa}\ ] ] moreover , if @xmath123/[i]$ ] and @xmath124/[i]$ ] also stabilize during this initial phase , then the mean degree of the infectives at this stage is given by @xmath125}{[i ] } \right)^ * + \left ( \frac{2[ii]}{[i ] } \right)^ * + \left ( \frac{[ir]}{[i ] } \right)^*. \label{ed_i_pa}\ ] ] the equations for the dynamics of the local densities @xmath120/[i]$ ] , @xmath126/[i]$ ] , and @xmath124/[i]$ ] are obtained from the sir-@xmath0 model by using the standard rules of differentiation and are given by @xmath127}{[i ] } \right ) & = & - \left(\beta + \omega + \beta z \left ( \frac{[si]}{[s ] } - \frac{2[ss]}{[s ] } \right ) + \beta \frac{[si]}{[i ] } \right ) \frac{[si]}{[i ] } \ , , \nonumber \\ \frac{d}{dt } \left(\frac{2[ii]}{[i ] } \right ) & = & 2\beta \left ( 1 + z \frac{[si]}{[s ] } \right ) \frac{[si]}{[i ] } - \left ( \gamma + \beta \frac{[si]}{[i ] } \right ) \frac{2[ii]}{[i ] } \ , , \label{loc - densities}\\ \frac{d}{dt } \left(\frac{[ir]}{[i ] } \right ) & = & \gamma \frac{2[ii]}{[i ] } + \beta z \frac{[si]}{[s ] } \frac{[sr]}{[i ] } - \beta \frac{[si]}{[i ] } \frac{[ir]}{[i]}. \nonumber\end{aligned}\ ] ] now , taking the limit when @xmath128/[s ] \to \mu=2l / n$ ] , and @xmath120/[s ] \to 0 $ ] , i.e. , after the introduction of the first infectious individuals , we obtain the following limit system : @xmath127}{[i ] } \right ) & = & \left(\beta ( e(\tilde{d } ) - 2 ) - \beta \frac{[si]}{[i ] } - \omega \right ) \frac{[si]}{[i ] } \ , , \nonumber \\ \frac{d}{dt } \left(\frac{2[ii]}{[i ] } \right ) & = & 2\beta \ , \frac{[si]}{[i ] } - \left(\gamma + \beta \frac{[si]}{[i ] } \right ) \frac{2[ii]}{[i ] } \ , , \label{limit - loc - densities } \\ \frac{d}{dt } \left(\frac{[ir]}{[i ] } \right ) & = & \gamma \frac{2[ii]}{[i ] } - \beta \frac{[si]}{[i ] } \frac{[ir]}{[i ] } \ , . \nonumber\end{aligned}\ ] ] it is interesting to observe that the reconnection probability @xmath1 does not appear in the system . this means that disconnections from infectious individuals play a role in the early epidemic dynamics , but the way the new connections are created , or even if they occur at all ( @xmath2 ) , does not play any role at this stage . the first equation of is decoupled from the other two and has a unique positive equilibrium @xmath121/[i])^ * = e(\tilde{d } ) - 2 - \omega/\beta$ ] , which is globally asymptotically stable . from this equilibrium and , it immediately follows that @xmath129 which defines the same epidemic threshold @xmath130 as @xmath26 ( cf . eq . ) , but overestimates @xmath3 when it is larger than one . a graphical comparison of the expressions of @xmath3 obtained from each modelling approach is shown in fig . [ r0s ] using @xmath15 as a tuning parameter . the other components of the positive equilibrium of follow upon substituting @xmath121/[i])^*$ ] for @xmath120/[i]$ ] into the last two equations . then , from , we have @xmath131 which is the same expression as the one obtained for @xmath132 ( cf . eq . ) . according to the rewiring processes described in section 2.2 , and using the same notation and the triple closure as before , the equations of the seir model at the initial phase of the epidemic and with the three types of rewiring are @xmath133 & = & -\beta [ si ] , \qquad % \nonumber \\ \frac{d}{dt}[e ] = \beta [ si]-\phi[e ] , \qquad % \nonumber \\ \frac{d}{dt}[i ] = \phi [ e]-\gamma [ i ] , \nonumber \\ \frac{d}{dt}[si ] & = & \left(- \beta z \frac{[si]}{[s]}-\beta -\gamma - \omega_{si}\right)[si ] + \phi [ se ] , \nonumber \\ \frac{d}{dt}[se ] & = & \beta z \left ( \frac{2[ss]}{[s]}-\frac{[se]}{[s ] } \right ) [ si ] - \left(\phi+\omega_{se}\right ) [ se ] + \alpha \ , \omega_{ei } \frac{[s]}{n-[i ] } [ ei ] + \alpha \ , \omega_{si } \frac { \delta [ e]}{n-[i ] } [ si ] , \nonumber \\ \frac{d}{dt}[ss ] & = & \left(-\beta z \frac{2[ss]}{[s ] } + \alpha \ , \omega_{si } \frac{[s]}{f([s],[e],[r])}\right)[si ] + \alpha \ , \omega_{se } \frac{[s]}{[s]+[r ] } [ se ] , \nonumber \\ \frac{d}{dt}[sr ] & = & \left(-\beta z \frac{[sr]}{[s]}+ \gamma + \alpha \ , \omega_{si } \frac{[r]}{f([s],[e],[r])}\right)[si ] + \alpha \ , \omega_{se } \frac{[r]}{[s]+[r ] } [ se ] , \label{seir - pa } \\ \frac{d}{dt}[ei ] & = & \beta \left(z \frac{[si]}{[s]}+1\right)[si]-\left(\phi + \gamma + \omega_{ei}\right)[ei]+2 \phi [ ee ] , \nonumber \\ \frac{d}{dt}[ee ] & = & \beta z \frac{[se]}{[s]}[si]-2 \phi [ ee ] + \alpha \ , \omega_{ei } \frac{[e]}{n-[i ] } [ ei ] , \nonumber \\ \frac{d}{dt}[er ] & = & \beta z \frac{[sr]}{[s]}[si ] + \left ( \gamma + \alpha \ , \omega_{ei } \frac{[r]}{n-[i ] } \right ) [ ei ] - \phi [ er ] , \nonumber \\ \frac{d}{dt}[ii]&=&\phi [ ei]-2\gamma[ii ] , \nonumber \\ \frac{d}{dt}[ir]&= & \phi [ er ] + \gamma\left(2 [ ii]- [ ir]\right ) , \nonumber\end{aligned}\ ] ] with @xmath101+[e]+[i]+[r]=n$ ] , @xmath110+[se]+[si]+[sr]+[ee]+[ei]+[er]+[ii]+[ir]+[rr]=l$ ] , @xmath134 , and @xmath135 the mean degree of the initial degree distribution @xmath14 . moreover , @xmath136,[e],[r])=[s]+[e]+[r]=n-[i]$ ] and @xmath137 if @xmath138 , whereas @xmath136,[e],[r])=[s]+[r]$ ] and @xmath139 otherwise . in the first case , susceptible individuals do not disconnect from exposed individuals and can reconnect to the latter when they rewire away from an infectious neighbor . in the second case ( @xmath140 ) , susceptible individuals recognize exposed ones , rewire away from them , and only reconnect ( with probability @xmath1 ) to other susceptibles or to recovereds ( so , @xmath139 ) . note that , in both cases , @xmath136,[e],[r ] ) \to n$ ] at the early stage of an epidemic . if @xmath141 , exposed individuals who break off a link with an infectious neighbor randomly reconnect to any susceptible , recovered , or exposed individual with a probability @xmath142/(n-[i])$ ] , @xmath143/(n-[i])$ ] , and @xmath144/(n-[i])$ ] , respectively . this corresponds to the situation where latent individuals are asymptomatic and , so , they do not know they have already got the infection . therefore , one can also assume that susceptible individuals do not know the disease status of the exposed neighbours and take @xmath21 . hence , susceptibles who break off a link with an infected neighbor reconnect to any susceptible , exposed or recovered individual with the same probabilities as the exposed ones , namely , @xmath145/(n-[i])$ ] with @xmath146,[e],[r]\}$ ] . in fig . [ r0s2 ] , we show the predicted and observed @xmath3 as a function of the infection rate @xmath15 . the values of the parameters are the same in both panels except for @xmath19 and @xmath6 which are 1 and 0 in the left panel , and 0 and 1 in the right one , respectively . it is illustrative to check that the limit system for the dynamics of the local densities of disease status around an infectious individual ( see appendix ) only contains one term with the reconnection probability @xmath1 , namely , the one with @xmath147 as a prefactor . therefore , the other contributions from the remaining rewiring rates in do not appear when we restrict ourselves at the early stage of an epidemic . in other words , the precise rules for reconnecting susceptibles who have rewired away from an infectious neighbor , even if there is no reconnection at all of those individuals ( @xmath2 ) , does not affect the epidemic dynamics during the initial phase . this claim , however , is not true for exposed individuals . the last term of the equation for @xmath148/[i]$ ] tells us that , when @xmath149 , exposeds enhance the spread of the disease by rewiring away from infectious neighbors ( because they will replace the latter with susceptible individuals ) and , hence , @xmath100 must increase under this rewiring . [ cols="^,^ " , ] finally , it is important to note that , for dynamic heterogeneous networks whose degree distribution evolves in time , the value of @xmath3 does not determine the final epidemic size . while the computation of @xmath3 is based on the initial degree distribution , the final epidemic size depends on the whole evolution of the degree distribution . in particular , since reconnection is assumed to be uniform with respect to the degree of nodes , the variance of the degree distribution decreases over time whenever the initial network is highly heterogeneous ( see @xcite ) . determining an expression for the final epidemic size , using any approximation method , and studying how it depends on model parameters , remains a highly interesting open problem . t.b . is grateful to riksbankens jubileumsfond ( grant p12 - 0705:1 ) for financial support . the research of d.j . and j.s . have been partially supported by the research grants mtm2011 - 27739-c04 - 03 and mtm2014 - 52402-c3 - 3-p of the spanish government ( mineco ) . j.s . and d.j . are members of the research groups 2014sgr-1083 and 2014sgr-555 of the generalitat de catalunya , respectively . the limit system for the local densities involved in the computation of @xmath4 for the seir pairwise model when @xmath128/[s ] \to 2l / n$ ] ( the mean degree ) , @xmath148/[s ] \to 0 $ ] , @xmath120/[s ] \to 0 $ ] , and @xmath150/[s ] \to 0 $ ] , i.e. , at the beginning of an epidemic , is @xmath151}{[i ] } \right ) & = - \left ( \beta + \omega_{si } + \phi \frac{[e]}{[i ] } \right ) \frac{[si]}{[i ] } + \phi \frac{[se]}{[i ] } \\ \frac{d}{dt } \left ( \frac{[se]}{[i ] } \right ) & = \beta ( \bar{q } - 1 ) \frac{[si]}{[i ] } - \left ( \phi \left ( 1 + \frac{[e]}{[i ] } \right ) + \omega_{se } - \gamma \right ) \frac{[se]}{[i ] } + \alpha \ , \omega_{ei } \frac{[ei]}{[i ] } \\ \frac{d}{dt } \left ( \frac{[e]}{[i ] } \right ) & = \beta \frac{[si]}{[i ] } - \left ( \phi \left ( 1 + \frac{[e]}{[i ] } \right ) - \gamma \right ) \frac{[e]}{[i ] } \\ \frac{d}{dt } \left ( \frac{[ei]}{[i ] } \right ) & = \beta \frac{[si]}{[i ] } - \left ( \phi \left ( 1 + \frac{[e]}{[i ] } \right ) + \omega_{ei } \right ) \frac{[ei]}{[i ] } + \phi \frac{2[ee]}{[i ] } \\ \frac{d}{dt } \left ( \frac{2[ii]}{[i ] } \right ) & = 2 \phi \frac{[ei]}{[i ] } - \left ( \phi \frac{[e]}{[i ] } + \gamma \right ) \frac{2[ii]}{[i ] } \\ \frac{d}{dt } \left ( \frac{[ir]}{[i ] } \right ) & = \phi \frac{[er]}{[i ] } + \gamma \frac{2[ii]}{[i ] } - \phi \frac{[e]}{[i ] } \frac{[ir]}{[i ] } \\ \frac{d}{dt } \left ( \frac{2[ee]}{[i ] } \right ) & = - \left ( \phi \left ( 2 + \frac{[e]}{[i ] } \right ) - \gamma \right ) \frac{2[ee]}{[i ] } \\ \frac{d}{dt } \left ( \frac{[er]}{[i ] } \right ) & = - \left ( \phi \left ( 1 + \frac{[e]}{[i ] } \right ) - \gamma \right ) \frac{[er]}{[i ] } + \gamma \frac{[ei]}{[i]}\end{aligned}\ ] ] where @xmath152 . note that only one positive term contains a rewiring rate , namely , the last one in the second equation which has @xmath147 as a prefactor . since @xmath148/[i]$ ] increases with @xmath147 and , in turn , the first equation tells us that @xmath120/[i]$ ] increases with @xmath148/[i]$ ] , the presence of such a positive term means that also @xmath3 increases with @xmath147 . on the other hand , since the first three equations of the limit system are decoupled from the other five when @xmath2 , the same three equations will also govern the dynamics of @xmath120/[i]$ ] when individuals simply drop connections to infectious neighbours ( @xmath2 ) . in particular , this reflects the ( expected ) fact that whether exposed individuals break off with infectious neighbors ( @xmath153 ) or not ( @xmath7 ) does not affect the early dynamics of @xmath120/[i]$ ] as long as they do not replace these links with new connections . for a given value @xmath154 of @xmath155/[i])^*$ ] , the equilibrium equations ( i.e. , the equations obtained by making the right - hand side ( rhs ) of the previous system equal to 0 ) define a linear system for the remaining variables and , hence , the equilibrium can be easily expressed in terms of @xmath154 . a simple inspection of the equations shows that there are two equilibria , @xmath156 and @xmath157 , where the local densities around infectious nodes , @xmath121/[i])^*$ ] , @xmath158/[i])^*$ ] , @xmath159/[i])^*$ ] , and @xmath160/[i])^*$ ] , are zero . moreover , from the third equation we see that there exists an equilibrium with @xmath121/[i])^ * > 0 $ ] if and only if @xmath155/[i])^ * > \gamma/\phi - 1 $ ] , which in turn implies that @xmath161/[i])^*=0 $ ] . consequently , the first four equations at equilibrium do not depend on the last four . expressing @xmath121/[i])^*$ ] , @xmath162/[i])^*$ ] , and @xmath159/[i])^*$ ] in terms of @xmath154 , and replacing them into the second equation , it follows that there will be an equilibrium with positive local densities around infectious nodes if there exists a solution @xmath163 of the equation @xmath164 such that @xmath165 if @xmath166 . note that , for @xmath167 , eq . becomes a quadratic equation in @xmath154 with a positive solution @xmath168 given by ( see @xcite ) @xmath169 with the proviso that @xmath170 . moreover , in this case ( @xmath7 ) , when @xmath166 it follows that @xmath165 if @xmath171 . if @xmath172 and @xmath153 , a necessary and sufficient condition for the existence of a unique solution @xmath168 of such that @xmath165 is that the straight line defined by the left - hand side ( lhs ) of intersects the vertical line @xmath173 above the intersection with this vertical line of the cubic polynomial defined by the rhs of . this polynomial has two negative roots , @xmath174 and @xmath175 , and the third root @xmath176 can be positive or negative depending on the sign of @xmath177 . the resulting condition on the parameters is @xmath178 note that this condition is fulfilled when @xmath21 which was , indeed , what we assumed in the model if @xmath153 . if @xmath179 , then @xmath180 and a sufficient condition for the existence of a @xmath163 is that the intersection of the lhs of with the @xmath181-axis is above the intersection with this axis of the rhs of . this condition leads to @xmath182 since we are assuming @xmath153 , then @xmath21 , and from this inequality it follows a simpler sufficient condition for the existence of @xmath183 , namely , @xmath184 . from , it follows that the only positive value of @xmath168 for which @xmath186 is @xmath187 . after replacing @xmath168 by this value , expression can be rewritten as @xmath188 comparing this expression and that of @xmath3 given by , it follows that @xmath189 if @xmath190 , and that @xmath191 ( and @xmath192 ) if @xmath193 . so , for @xmath194 , both approximations lead to the same epidemic threshold when @xmath7 . 10 diekmann , o. , heesterbeek , j. a. p. , metz , j.a.j . , 1990 . on the definition and the computation of the basic reproduction ratio @xmath3 in models for infectious diseases in heterogeneous populations . 28 , 365382 . lau , j.t.f . , griffiths , s. , choi , k.c . , tsui , h.y . , 2010 . avoidance behaviors and negative psychological responses in the general population in the initial stage of the h1n1 pandemic in hong kong . _ bmc infectious diseases _ 10 , 139 . doi:10.1186/1471 - 2334 - 10 - 139 springborn , m. , chowell , g. , maclachlan , m. , fenichel , e.p . , 2015 . accounting for behavioral responses during a flu epidemic using home television viewing . bmc infectious diseases 15:21 . doi 10.1186/s12879 - 014 - 0691 - 0

the evolution of disk galaxies is significantly influenced by the main features within their disks , notably spirals and bars . in turn , galactic evolution also significantly influences the evolution , morphology , and dynamics of these features . bars at the present epoch can be found in about 70% of massive disk galaxies using near - infrared images ( sellwood & wilkinson 1993 ; knapen et al . 1999 ; eskridge et al . the fraction is somewhat lower on optical images . the gravitational torque method ( gtm , buta & block 2001 ) was developed to quantify bar strength . the idea is to transform a deprojected near - infrared image of a spiral galaxy into a gravitational potential , and then compute the ratio of the tangential force to the mean radial force as a function of position in the plane of the galaxy . the potential is derived from poisson s law after assuming an exponential vertical density law and a constant mass - to - light ratio ( quillen , frogel , and gonzlez 1994 ) . the mean radial force represents the axisymmetric background due to the bulge , disk , and bar . a fully quantitative measure of bar strength can be defined from the maximum of the tangential - to - radial force ratio , as long ago suggested by sanders & tubbs ( 1980 ) and combes & sanders ( 1981 ) . the maximum force ratio is equivalent to the maximum gravitational torque per unit mass per unit square of the circular speed . the gtm has been applied to large samples of galaxies in several recent studies ( block et al . 2001 , 2002 ; laurikainen , salo , & rautiainen 2002 ; laurikainen & salo 2002 ; buta , laurikainen , & salo 2004 , hereafter bls ; laurikianen , salo , & buta 2004a ; laurikainen et al . in general , the gtm has an advantage over other methods previously used to define bar strength because it is based on the forcing due to the bar itself and not just on the bar s apparent shape . the relative bar torque is a good measure of the importance of non - axisymmetric forces , which play a role in gas accretion , spiral arm generation , and overall disk evolution . studies at optical wavelengths suggested that early - type barred galaxies are associated with grand design spiral structure because the bars end near their own corotation resonances , and at this point they have enough torque to drive corotating spiral waves outward into the disk ( e.g. , elmegreen & elmegreen 1985 , 1989 ) . late - type bars could have different bar and spiral pattern speeds , as observed for ngc 925 ( elmegreen , wilcots & pisano 1998 ) . the interaction between bars and spirals is not well understood observationally because most studies have been based on optical images , which are confused by dust and star formation . we have shown , for example , that @xmath3-band spirals often look different from optical spirals with regard to pitch angle , continuity , and symmetry ( block & wainscoat 1991 ; block et al . 1999 ) . because the light distribution at 2.2@xmath4 m emphasizes the mass distribution in the old disk , the near - infrared spirals are more important for bar driving than the optical spirals . buta , block , & knapen ( 2003 , hereafter bbk ) have shown that it is possible to use straightforward fourier techniques to separate the luminosity distributions of bars and spirals ( see also lindblad , lindblad , & athanassoula 1996 ) . the gtm is then used to derive maximum relative torques for each nonaxisymmetric feature alone , thereby allowing us to examine in a quantitative manner the possible correlation between bar and spiral strengths . in this paper , we apply the bbk method to 17 representative disk galaxies covering a range of early to intermediate hubble types and de vaucouleurs family classifications . our goal is to investigate the relation between bar torque and spiral arm amplitude . our sample consists of the 17 galaxies listed in table 1 . these were selected on the basis of hubble type ( range sab - scd ) and inclination ( less than 70@xmath5 ) . the absolute blue magnitudes of the galaxies range from @xmath618.3 to @xmath621.4 , with an average of @xmath620.5@xmath70.9 . thus , these are massive , high luminosity systems , typical of the bright galaxy population . a range of apparent bar strengths covering the de vaucouleurs families sa , sab , and sb was also selected . fifteen of our sample galaxies were observed during a total of five nights ( 2001 september 5 , and 2001 october 58 ) with the ingrid camera ( packham et al . 2003 ) attached to the 4.2-m william herschel telescope ( wht ) . the images were taken in the @xmath3-short ( 2.2@xmath4 m , or @xmath0 ) band , and have a scale of 0.241 arcsec / pix and a field of view approximately 4 arcminutes square . total on - source exposure times average about 59 min , but ranged from 16 to 100 min ( see table 1 ) . the observing techniques are the same as those described in knapen et al . ( 2003 ) . particularly important is the background subtraction , which was performed by interspersing between small blocks of galaxy observations several exposures of the same length on a blank background field . we estimate the typical sky subtraction error to be on the order of 0.1%-0.2% of the total background . to eliminate bad pixels , columns , and ( in the case of the background frames ) foreground stars , dithering was used for both the galaxy and background frames . we required deeper than usual exposures at 2@xmath4 m in order to effectively measure spiral arm torques . the limiting surface brightness and typical signal - to - noise ratio were estimated from our @xmath0 image of ngc 1530 using an independently measured @xmath0 flux listed in the ned . in this image , a 48 min exposure , the signal - to - noise ratio is 2.6 at a surface brightness level @xmath8 = 20.0 mag arcsec@xmath9 . typical @xmath10 background noise is 21.0 magarcsec@xmath9 , down to around 21.3 magarcsec@xmath9 for our longest exposures . in addition to the 15 wht objects , our analysis includes two southern strongly - barred galaxies , ngc 1365 and ngc 1433 , which were part of previously published studies . these were observed with the ctio infrared imager ( cirim ) attached to the 1.5-m telescope of cerro tololo inter - american observatory ( ctio ) and have a scale of 1@xmath1114 pix@xmath12 . details of the observations of ngc 1365 , observed at @xmath3 , are provided by regan & elmegreen ( 1997 ) , while those for ngc 1433 , observed at @xmath13 , are provided by buta et al . the gtm as used in previous studies involves the derivation of the maximum tangential - to - radial force ratio , a single number that characterizes the strength of nonaxisymmetric perturbations in a galaxy . bb01 inspected the two - dimensional ratio ( butterfly " ) maps of tangential force @xmath14 to mean radial force @xmath15 to insure that this maximum was mostly measuring a bar for their 36 galaxies . hence , they called the force ratio maximum @xmath2 . a more automated approach was used by laurikainen , salo , & rautiainen ( 2002 ) and block et al . ( 2002 ) whereby the maximum force ratio is derived from the radial variation of the maximum ratio @xmath16 . depending on the relative importance of the bar and the spiral , the maximum @xmath17 could be in the bar or the spiral region . for this reason , bbk proposed calling the maximum force ratio @xmath18 to remove any ambiguity about what it represents . since @xmath18 in a barred galaxy could be affected by spiral arm torques , we can not use @xmath18 alone to assess whether stronger bars correlate with stronger spirals . instead , we need to use the special technique outlined by bbk , a fourier - based method that separates the torques due to a bar from those due to a spiral . however , unlike estimates of @xmath18 , bar / spiral separation is not automated but depends on an iterative procedure . in this section , we carry out the bbk separation procedure on each sample galaxy , and then compare the results with estimates based on arm - interarm and bar - interbar contrasts . the galaxy images are deprojected using mainly rc3 ( de vaucouleurs et al . 1991 ) orientation parameters , with revisions for three galaxies based on isophotal ellipse fits ( ngc 6951 , 7723 ) or kinematic parameters ( ngc 1530 , regan et al . gravitational potentials are evaluated using the method of quillen , frogel , & gonzlez ( 1994 ) , under the assumptions of a constant near - ir mass - to - light ratio and an exponential vertical scaleheight @xmath19 estimated from the radial scalelength @xmath20 and the type - dependent ratio @xmath21 from de grijs ( 1998 ) . a revised lookup table for the vertical dimension ( see bbk ) was used that did not involve any gravity softening . in the bbk approach , relative fourier intensity amplitudes are derived from the deprojected near - infrared images . in the inner parts of the galaxy , we assume that the nonaxisymmetric amplitudes are due entirely to the bar , while past a radius @xmath22 ( where @xmath23 is the fourier azimuthal index ) the observed amplitudes are due to a combination of the bar and the spiral . we assume that the even relative fourier amplitudes due to the bar decline past the maximum at @xmath22 in the same manner as they rose to that maximum . a fourier image of the bar is constructed and removed from the original image , to give a spiral plus disk image . then the @xmath23=0 fourier image is added back to the bar image to give the bar plus disk image . we convert each of these images into gravitational potentials , and derive maps of the ratio of the tangential force to the mean background radial force , the latter being defined by the @xmath23=0 fourier image . from these ratio maps , we derive the maximum force ratios @xmath2 due to the bar and @xmath1 due to the spiral arms . in practice , we base our analysis entirely on fourier - smoothed images using terms up to @xmath23=20 . bbk fully discuss the pitfalls and uncertainties in this method of separation , but it is probably the most straightforward approach one can use for this purpose . although bbk illustrated the technique using a fairly symmetric barred spiral ( ngc 6951 ) , our present sample includes one asymmetric barred spiral , ngc 7741 , that necessitated special attention to odd fourier amplitudes in the bar . ngc 7741 is a late - type ( sb(s)cd ) barred spiral with strong asymmetry in the bar , a characteristic which is actually fairly typical of sd - sm spirals ( de vaucouleurs & freeman 1972 ) . to deal with it , we took advantage of the fact that ngc 7741 has no bulge , and chose a center for the fourier analysis based on the faintest discernible isophotes of the bar . with such a center , the odd fourier terms in the bar go to zero near the bar ends . in this circumstance , we can define the bar by even and odd fourier terms and allow for its asymmetry . we have not accounted for a slight asymmetry in the bars of two other cases , ngc 972 and ngc 7479 , because these have bright central regions and the effect is less important . the success of bar / spiral separation depends on the quality of the near - infrared images . in our case , the wht images are sufficiently well - exposed that noise levels are minimal in the outer parts of the images . very bright star - forming regions were removed from the images to minimize their impact . these could have a different mass - to - light ratio than the surrounding old disk stars , and could give false torque amplitudes . application of the bbk method is not automatic and requires some iteration for the best choice of the radii @xmath22 . plots of @xmath24 , the amplitude of the @xmath23th fourier component relative to the @xmath23=0 component , were initially made for each galaxy and were evaluated by eye . the most reliable extrapolations were chosen by examining the resulting bar+disk and spiral+disk images . poor choices of extrapolation would sometimes leave artifical depressions in the spiral+disk image , especially near the ends of the apparent bar . the final extrapolations used are those which provided the cleanest - looking separations . bbk estimated the errors on @xmath2 and @xmath1 due to extrapolation uncertainties , and showed that a @xmath710% uncertainty in the choice of @xmath25 , the radius of the maximum of the dominant @xmath23=2 term , can lead to a @xmath74% uncertainty in @xmath2 and @xmath710% uncertainty in @xmath1 , at least for ngc 6951 . we expect that this is a reasonable estimate of the uncertainty for the other galaxies in our sample as well ( see section 3.2 ) . we show fourier extrapolations for the 12 sample galaxies that required bar / spiral separation in figure [ allfourier ] . ( the ones that did not require such separation , ngc 908 , 1058 , 6643 , 7217 , and 7606 , did not show a measurable bar signal . ) in these plots , the curves are the relative intensity amplitudes , and only the lower order terms are shown ( the actual analysis used terms up to @xmath23=20 as in bbk ) . the symbols show the mappings ( interpolations and extrapolations ) used for the bar . except for ngc 1255 , ngc 7723 , and ngc 7741 , the mappings are symmetric around a radius @xmath22 . for @xmath26 , the rising relative amplitudes are used as observed , while for @xmath27 , the amplitudes are extrapolated as the exact reverse of the rise . for ngc 1255 , ngc 7723 , and ngc 7741 , the observed amplitudes already decline past @xmath22 , and in these cases a smaller amount of extrapolation was needed . these curves show some of the different characteristics of the bars in the sample galaxies . for example , both ngc 1365 and ngc 1433 have bars that can be extrapolated with relatively flat - topped @xmath23=2 profiles . although the bar is nothing more than a weak oval in ngc 1255 , it still shows a distinct signature in the @xmath23=2 profile that required little extrapolation . the bar is stronger in ngc 7723 , but it dominates the amplitudes inside @xmath28=25@xmath11 and required little extrapolation past @xmath22 . in ngc 7479 , the bar and strong spiral produce a double - humped @xmath23=2 profile , with one hump corresponding to the bar and the other hump corresponding to the spiral . ngc 7741 shows an asymmetric @xmath23=2 profile . bar / spiral separation is most reliable when the bar has a relatively constant position angle in the disk . figure [ allphases ] shows plots of the @xmath23=2 fourier phase @xmath29 as a function of radius for the same 12 galaxies as in figure [ allfourier ] . the solid vertical lines show the radius , @xmath30 , of the bar maximum from the force ratio maps ( see table 2 ) . these show that the assumption of a constant bar phase is fairly good for the more strongly barred galaxies in the sample , but is less good for the weakly - barred ones . nevertheless , the extrapolations still produce reasonable separations even for the weaker cases . figure [ allrotc ] shows plots of the rotation curves predicted from the light distribution for each sample galaxy except ngc 7217 ( see buta et al . 1995 for a detailed study of this galaxy ) . these curves are derived from the @xmath23=0 component of the gravitational potential as @xmath31 and are normalized to the maximum values . the curves look relatively typical of high luminosity spirals . figures [ allimages ] and [ allratios ] show the analysis images and ratio maps , respectively , for the 12 sample galaxies requiring bar / spiral separation . three images are shown for each galaxy ( following bbk ) : the @xmath23=0 - 20 fourier sum image , the separated bar+disk image , and the separated spiral+disk image . in all cases except ngc 1808 , the bar or oval has been rotated to the approximate horizontal position . for a classical barred spiral like ngc 1365 , figure [ allimages ] shows that the extrapolations in figure [ allfourier ] produce a fairly clean separation of the bar and the strong spiral . the ratio maps for ngc 1365 in figure [ allratios ] show a very regular pattern of alternating force maxima / minima for this galaxy . in the case of ngc 1433 , the extrapolations in figure [ allfourier ] also produced a clean separation , but the separated spiral is an oval ring - shaped pattern . hence , the spiral shows a butterfly pattern similar to the bar . the remaining galaxies each show distinctive characteristics in both the images and ratio maps , but the main point is that the separations appear to be reasonable representations of the galaxies . in the cases of ngc 1255 and 6814 , where the bar is nothing more than an @xmath23=2 oval , the four maximum points lie at about @xmath740@xmath5 - 45@xmath5 to the bar axis , while for the more strongly - barred cases , these points lie much closer to the bar axis . as noted by bb01 , higher order terms in the bar potential cause the maximum points to lie closer to the bar axis . several of the more highly inclined galaxies in our sample , ngc 908 , 972 , 5033 , 6643 , and 7606 , required bulge / disk decompositions in order to minimize the effects of bulge `` deprojection stretch . '' the idea is to model the bulge profile , subtract off the bulge from the 2d images , deproject the residual disk light , and then add the bulge back to the deprojected disk light . since none of these five galaxies is strongly barred , one - dimensional decomposition techniques were adequate for our purposes . we used either the kent ( 1986 ) iterative method or @xmath32 bulge and exponential disk decompositions ( e.g. , kormendy 1977 ) . the bar separation is perhaps less certain for ngc 5033 than for the other galaxies . the bulge decomposition left some residual zones of lower intensity around the galaxy minor axis . a small bar is found in the inner regions , but its ratio map is affected by the decomposition uncertainties . the ratio map also highlights the bar - like nature of the spiral in this galaxy . of the strongly barred spirals with bright centers , the asymmetry in ngc 7479 s bar left a residual pattern in the spiral plus disk image . one sees a low intensity region on one side of the center , and a higher intensity region on the other side . ignoring this asymmetry will not affect @xmath2 too much in this case , since @xmath2 is based on an average of the maximum points in the four quadrants , but will cause us to underestimate the scatter in these maximum points . our refined procedure for ngc 7741 accounts for the asymmetry in the bar very well , but leaves a sharp edge in the area around the bar ends . this has very little impact on @xmath2 , but @xmath1 could have a larger uncertainty because of it . from the separated ratio maps , we derived plots of the maximum force ratio , @xmath33 as a function of radius @xmath28 in the galaxy plane . the way these curves are derived is fully described by bbk . at each radius , we locate the force ratio maximum in four quadrants and then average the results . figure [ qtvsr ] shows the curves for 16 of our sample galaxies , including the ones for which no clear bar signal was detected . in this figure , the radii are normalized to @xmath34 = @xmath35/2 , where @xmath35 is the galactic extinction - corrected face - on blue light isophotal diameter at a surface brightness level of 25.0 mag arcsec@xmath9 ( rc3 ) . the results for ngc 7217 are not shown since little forcing due either to a bar or a spiral was detected in this object . most of the signal seen in the force - ratio map is likely to be due to bulge deprojection stretch , and has @xmath36 @xmath37 0.04 everywhere . also , ngc 7217 is the most bulge - dominated system in our sample ( buta et al . 1995 ) , and our procedure will not account for this reliably since , like bb01 , we have transformed the light distributions into potentials assuming all components have the same vertical scaleheight . for our main analysis we have set @xmath2 @xmath38 @xmath1 = 0 for ngc 7217 . table 2 summarizes the results of the separations . four other sample galaxies , ngc 908 , 1058 , 6643 , and 7606 , also had little or no detectable bar , and we have simply set @xmath2=0 for these . the weak ovals in ngc 1255 and 6814 were easily separated from their spirals . the remaining galaxies have a range of bar strengths up to @xmath2=0.61 in ngc 1530 . owing to a limited field of view , the spiral parameters for ngc 1808 are more uncertain than the listed errors imply . the radii of the maximum relative torques tend to be well inside the standard isophotal radius , for both bars and spirals . for the seven objects in table 2 having @xmath2 @xmath39 0.25 , @xmath40 = 0.24@xmath70.09 ( s.d . ) and @xmath41 = 0.42@xmath70.09 ( s.d . ) . for these same galaxies , @xmath42 = 0.57@xmath70.13 ( s.d . ) , so that the maxima tend to be well - separated on average . the main uncertainties in the gtm are due to the uncertainties in the assumed vertical scaleheight @xmath19 , variations in the mass - to - light ratio ( both due to dark matter or stellar population differences ) , the adopted orientation parameters ( inclination , line of nodes ) , the bulge deprojection stretch , the bar thickness , the sky subtraction , and the galaxy asymmetry . in addition to these , bar / spiral separation involves uncertainties due to the method of extrapolation . for the more highly inclined galaxies in the sample , we have minimized bulge deprojection stretch using bulge / disk decomposition . bls show that the uncertainties in maximum relative torques average about 12% due to uncertainties in the vertical scaleheight . a nonconstant vertical scaleheight in the bar could lead to a further 5% uncertainty in @xmath2 ( laurikainen & salo 2002 ) . uncertainties in galaxy inclinations lead to an inclination - dependence in the uncertainty in @xmath2 or @xmath1 . for an error of @xmath75@xmath5 in inclination @xmath43 , the error in maximum relative torques ranges from 4% for @xmath4435@xmath5 to 20% for @xmath4560@xmath5 . an uncertainty of @xmath74@xmath5 in line of nodes position angle @xmath46 has less of an impact , ranging from 4% for @xmath4435@xmath5 to 11% for @xmath4560@xmath5 . typical sky subtraction uncertainties on the wht images could lead to an additional 3.5% uncertainty in @xmath2 and 5% uncertainty in @xmath1 ( bbk ) . finally , uncertainties in the bar extrapolations and in the symmetry assumption of the relative fourier intensity amplitudes can lead to an additional 4% uncertainty in @xmath2 and an 11% uncertainty in @xmath1 ( bbk ) . table 2 lists the total uncertainty on each parameter from this analysis . for ngc 1365 , the uncertainty in the orientation parameters may be larger than we have assumed . jorsater & van moorsel ( 1995 ) analyzed an hi velocity field of ngc 1365 and concluded that circular motions may be reliable only between radii of 120@xmath11 and 240@xmath11 , where the kinematic position angle and inclination are relatively constant . the kinematic position angle , 220@xmath5 , agrees well with the rc3 value of 212@xmath5 . however , the kinematic inclination of 40@xmath5 significantly disagrees with the nearly 58@xmath5 inclination implied by the rc3 logarithmic isophotal axis ratio of @xmath47=0.26 . since the bar of ngc 1365 is nearly along the galaxy s minor axis , the derived torque parameters will be sensitive to this disagreement . the impact of dark matter on maximum gravitational torques for the osu bright galaxy sample was shown by bls to be fairly small , especially for those galaxies more luminous than @xmath48=2.5@xmath4910@xmath50 @xmath51 in the @xmath52-band ( corresponding to @xmath53=@xmath620.5 ) . because the galaxies in the osu sample are on average luminous , massive systems , the correction to @xmath54 was generally less than 10% with an average of about 5% . our sample here has similar characteristics , with an average luminosity of @xmath55=@xmath48 . the maximum relative gravitational torques in both the bar and the spiral regions tend to lie in the bright inner regions of the galaxies , where the dark halo contribution is small in such luminous galaxies . thus , it is likely that dark matter has only a minimal impact on our results . the stellar mass - to - light ( @xmath56 ) ratio is a separate issue that could impact gravitational torque calculations . bell & de jong ( 2000 ) used simple spectrophotometric evolution models of spiral galaxies to show that stellar @xmath56 variations can , in fact , be significant , even in the @xmath3-band . they present simple relations between color index and the @xmath56 correction for a given passband . however , we can not make a reliable deduction from their analysis of the error committed by ignoring these variations . reddening can be significant in color index maps , and would invalidate any @xmath56 corrections from the bell & de jong relations in some regions , such as bar dust lanes . also , tests we have made with @xmath57 color index maps of ngc 1530 and ngc 7723 indicate that the near - ir spirals are decoupled from the optical spirals ( see section 6 ) and are not necessarily much bluer than bars , implying that the mass - to - light ratios of the spirals might not be very different from those of bars . however , interbar and interarm regions can be bluer than these features , implying that we could be underestimating @xmath58 and @xmath59 slightly . since our sample is defined by intermediate to late - type spirals , it is likely that @xmath58 and @xmath59 are affected by stellar @xmath56 variations in a similar manner for each galaxy , thus largely preserving any relationship between them . with our estimates of separate bar and spiral strengths , we can now check how well these two quantities correlate . figure [ qsqb ] shows @xmath1 versus @xmath2 assuming constant @xmath56 . for @xmath2 @xmath60 0.3 , there is only a weak correlation with @xmath1 , meaning that spirals as strong as @xmath1@xmath380.3 are possible even in the absence of a bar . however , the strongest spirals in our sample are only associated with bars having @xmath2 @xmath61 0.4 . the error bars tend to be large for these stronger cases in part because two of the galaxies having @xmath2 @xmath39 0.4 ( ngc 1365 and 1530 ) have rc3 inclinations close to 60@xmath5 and one ( ngc 7741 ) has considerable asymmetry in the bar . figure [ qsqb ] also suggests that among barred galaxies , stronger bars have stronger spirals . that is , for @xmath62 , @xmath2 and @xmath1 are linearly correlated . in addition to gravitational torques , the relative importance of bars and spirals can be assessed in a variety of different ways . for example , the deprojected ellipticity of a bar is thought to be a good indicator of bar strength ( e.g. , martin 1995 ) , based on the theoretical study by athanassoula ( 1992 ) . abraham & merrifield ( 2000 ) have refined this idea using an automated analysis of two - dimensional surface brightness distributions and a rescaled bar ellipticity parameter called @xmath63 ( whyte et al . however , bar ellipticity is an incomplete measure of bar strength since the latter also depends on the total mass of the bar ( laurikainen , salo , & rautiainen 2002 ) . other estimates of bar importance include the bar - interbar contrast and the amplitude of the @xmath64 fourier component of the bar . similarly , the arm strength can be assessed through the use of arm - interarm contrasts . we used the deprojected images to make these estimates for our sample galaxies . arm classes ( flocculent , multiple arm , and grand design ; elmegreen & elmegreen 1982 ) were assigned to the galaxies according to their deprojected @xmath0-band appearance , as indicated in table 3 ; some assignments differ from their original classification in the @xmath52-band ( e.g. , elmegreen & elmegreen 1987 ) . intensity cuts were made along the major axis of the discernible bars ( using pvector in iraf ) in order to determine their radial profiles . bars with constant surface brightness profiles are labeled `` flat '' in the table , while bars with decreasing profiles are labeled `` exponential '' ( as discussed in elmegreen & elmegreen 1985 ) . the lengths of the bars were determined using a program , sprite , which fits ellipses to isophotes via least - squares ( see buta et al . bar length is taken to be the major axis length at maximum ellipticity , close to where the position angle begins to change . the maximum ellipticities @xmath65 and bar radii @xmath66 ( relative to the extinction - corrected isophotal radius from rc3 ) are given in table 3 . in addition to maximum ellipticities and bar radii from the full images , we also derived these parameters from the _ separated bar images_. this is useful for those more strongly barred spirals where the brightest parts of the arms lie very close the bar ends . these are listed as @xmath67/@xmath68 and @xmath69 in table 3 . comparisons between @xmath30/@xmath68 , @xmath70/@xmath68 , and @xmath66/@xmath68 indicate that @xmath70 @xmath38 @xmath66 in several of the more strongly barred cases where the inner parts of the spirals affect the isophotal bar fit . however , @xmath70 @xmath39 @xmath67 in all cases . the deprojected galaxies were transformed into polar images , with gray scale intensities for radius versus azimuthal angle . the bar - interbar amplitude at 0.7@xmath66 was estimated from intensity profiles parallel and perpendicular to the bar . also , azimuthal intensity cuts with a width of three pixels were made from the polar images at radii chosen to be at 0.25 and 0.50 times the standard isophotal radius . from these profiles , the magnitude differences between the bars and the interbar regions or between the arms and the interarm regions were determined for each galaxy , where @xmath71 for @xmath72 = bar or arm intensity and @xmath73 = interbar or interarm intensity . we label the contrast parameters as @xmath74 , @xmath75 , and @xmath76 , respectively . the results are compiled in table 3 . fourier transforms for the @xmath23=2 components of the intensity cuts at 0.7@xmath66 were made using the equation @xmath77 ^ 2 + [ \sum { i(\theta ) cos ( 2 \theta ) } ] ^2 ] } / \sum(i(\theta))\ ] ] where @xmath78 is the intensity at each azimuthal angle @xmath79 ( ranging from 0@xmath5 to 360@xmath5 ) , and the sums are over all angles . these are also tabulated in table 3 . the results reveal several correlations among these parameters . the bar - interbar contrast @xmath74 scales in a nearly linear fashion with @xmath2 , but with a considerable scatter at large @xmath74 , as shown in figure [ debbie]a . the bar ellipticity also correlates with bar torque , as shown in figure [ debbie]b , although the relation is not linear . this is fully consistent with the results of laurikainen , salo , & rautiainen ( 2002 ) , who derived this same correlation from images of more than 40 galaxies from the two micron all - sky survey ( 2mass ) . for @xmath2 @xmath60 0.4 , the correlation between @xmath65 and @xmath2 is fairly good , while for @xmath2 @xmath39 0.4 , the correlation is weaker in the sense that small changes in @xmath65 can correspond to large changes in @xmath2 . the correlation is only slightly improved when @xmath69 is used . the fourier strength parameter @xmath80 correlates with q@xmath81 ( with some scatter at large torque ) , as shown in figure [ debbie]c . these results indicate that the bar - interbar contrast , ellipticity , @xmath2 , and @xmath23=2 fourier component all provide reasonable measures of bar strength , especially for the weaker bars . the bar - interbar contrast increases with increasing bar radius , as shown in figure [ debbie]d , where @xmath74 is plotted versus the fractional bar radius @xmath82 . furthermore , the bar torque is stronger for flat bars ( average value of @xmath2 is 0.43@xmath83 for 6 objects ) than for exponential bars ( 0.28@xmath84 for 4 objects ) . these results are consistent with previous @xmath3-band studies , which indicated that flat bars are both longer and stronger than exponential bars ( elmegreen & elmegreen 1985 , regan & elmegreen 1997 ) . the previous studies also found that earlier spiral types tend to have flat strong bars and later types tend to have exponential weak bars , although a correlation with hubble type is not obvious in the present small sample . however , correlations of maximum relative gravitational torques with hubble type have been established by laurikainen , salo , & rautiainen ( 2002 ) , bls , and laurikainen , salo , & buta ( 2004 ) . early - type bars and spirals tend to be weaker than late - type bars and spirals owing to the dilution effect of bulges in early - types . in the present sample , the strongest bar is that in ngc 7741 , which also lacks any significant bulge . thus , the scatter in the correlation between bar strength and bar length could be influenced by the importance of the bulge . the arm - interarm contrast @xmath75 scales with the arm - interarm contrast @xmath76 ( not shown ) but is slightly weaker , which is consistent with previous findings that the arm - interarm contrast increases with radius to about mid - disk ( elmegreen & elmegreen 1985 ) . there is a weak correlation ( not shown ) between arm strength and arm class , with grand design galaxies having stronger spirals than flocculent galaxies . the arm - interarm contrast increases linearly with the maximum relative spiral torque @xmath1 , as shown in figure [ debbie]e . the correlation is poorer for @xmath76 . we have applied the dust - penetrated classification scheme of block & puerari ( 1999 ) to each of our sample galaxies . on the basis of deprojected near - infrared images , evolved stellar disks may be grouped into three principal dust penetrated archetypes : those with tightly wound stellar arms characterized by pitch angles at @xmath0 of @xmath85 10@xmath5 ( the @xmath86 class ) , an intermediate group with pitch angles of @xmath85 25@xmath5 ( the @xmath87 class ) and thirdly , those with open spirals demarcated by pitch angles at @xmath0 of @xmath85 40@xmath5 ( the @xmath88 bin ) . to take full cognizance of the duality of spiral structure and decouplings between gaseous and stellar disks , it has been demonstrated ( e.g. block & wainscoat 1991 ) that we require _ two _ classification schemes one for the population i disk , and a separate one for the population ii disk . a near - infrared classification scheme can never _ replace _ an optical one , and vice - versa , because the _ current _ distribution of old stars strongly affects the _ current _ distribution of gas in the population i disk . for dust - penetrated classification , logarithmic spirals of the form @xmath89 ( see danver 1942 ) are employed in the decomposition . the amplitude of each fourier component is given by ( schrder et al . , 1994 ) @xmath90 where @xmath28 and @xmath79 are polar coordinates , @xmath91 is the intensity at position @xmath92 , @xmath23 represents the number of arms or modes , and @xmath93 is the variable associated with the pitch angle @xmath94 of a given @xmath23-mode , defined by @xmath95 . we select the pitch angle of the dominant @xmath23-mode to define the dust - penetrated pitch angle classes described above . of course , the spiral arms in barred galaxies often depart from a logarithmic shape . as noted by block et al . ( 2001 ) , the arms may break at a large angle to the bar and then wind back to the other side , as in a `` pseudoring . '' we minimize the impact of non - constant pitch angle due to rings or pseudorings angles by excluding from our analysis the bar regions of the galaxies in question , and by restricting the fits to a limited range in radius . to illustrate what the fourier method is extracting , we show in figure [ ngc1530 ] @xmath23=2 inverse fourier transform contours superposed on our deprojected @xmath0 image of ngc 1530 . this galaxy has a very open two - armed spiral breaking from near the ends of its bar . the dominant @xmath23-mode is in this region , and the contours are a reasonable representation of the pitch angle of the arms . table 4 summarizes the results of our pitch angle analysis . the radius ranges used for the fits are listed in column 5 , except for ngc 1365 and 1433 whose classifications are from buta & block ( 2001 ) . we continue to find a ubiquity of low order @xmath96 and @xmath64 modes in this sample , consistent with earlier studies ( e.g. block & puerari 1999 ) . the most uncertain classifications are for ngc 1808 and ngc 5033 , where the field of view limits the part of the main spiral that we can see . under very special circumstances , dominant modes with @xmath23 greater than 2 may develop within the modal theory of spiral structure . block et al . ( 1994 ) and bertin ( 1996 ) hinted that in a _ gas - rich _ system , some dominant higher-@xmath23 modes should develop , and this might also induce some response in the stellar disk , for the stronger cases . furthermore , non - linear modes may couple and again give rise to higher-@xmath23 structures : @xmath64 and @xmath96 combine to give @xmath97 , @xmath64 and @xmath64 combined to give @xmath98 , etc . ( elmegreen , elmegreen , & montenegro 1992 ; block et al . , 1994 and g. bertin , private communication ) . we use the terminology h3 and h4 for these third and fourth harmonics , to assist with easy visualization of their evolved disk morphologies . finding examples of true three- and four - armed spiral galaxies in the near - infrared is a great challenge , requiring the investigation of many galaxies to encounter one or two unambiguous cases . a good example of an @xmath98 stellar disk in the present study is ngc 6814 , while a previously published example is eso 566@xmath624 ( buta et al . 1998 ; rautiainen et al . 2004 ) . we also note that early hubble type galaxies ( e.g. ngc 972 , sab ) can present very wide open arms in the near - infrared ; ngc 972 belongs to the @xmath88 class . three other examples are ngc 1808 ( of rc3 type sa but of dust penetrated arm class @xmath88 ) , ngc 1530 and ngc 1365 ( both hubble type sbb ; near - infrared arm class @xmath88 ) . conversely , it is also possible for spirals , classified as late type in the optical regime , to have arms in the near - infrared which are not very wide open ( e.g. ngc 908 and ngc 1058 are of hubble type sc , but both belong to the more tightly wound @xmath87 bin : see table 4 ) . the present study continues to illustrate decouplings between gaseous and stellar disks and the great advantage of dust - penetration . in the optical , ngc 972 is classified as flocculent by elmegreen & elmegreen ( 1987 ; see also table 3 ) . the photograph reproduced by sandage & bedke ( 1994 ) in panel 148 fully supports the flocculent designation . sandage & bedke ( 1994 ) note that the optical image of ngc 972 _ `` is dominated by the heavy dust lanes crossing the near - side of the high - surface - brightness bulge''_. in the near - infrared , ngc 972 presents a principally two - armed spiral ; the dominant harmonic is @xmath23=2 . its grand design , two - armed , evolved stellar disk appears to be completely decoupled from its fleece - like flocculent population i gas disk . while many optically flocculent spirals may still present a flocculent appearance in the near - infrared ( elmegreen et al . 1999 ) , decouplings of the two components , when present , are indeed very striking . ngc 1808 is famous for its dust lanes which appear to radiate almost perpendicularly to the major axis . sandage & bedke ( 1994 ) note in their panel 193 , that _ `` the central region of ngc 1808 provides what appears to be direct evidence of a galactic fountain composed of narrow dust lanes perpendicular to the plane . '' _ the near - infrared imaging successfully penetrates the dust in this saba spiral , yielding an inner pair of grand design , wide - open arms ( the dominant harmonic in ngc 1808 is @xmath23=2 ) . in this case , bar / spiral separation has the added uncertainty that the bar position angle is not constant ( see figure [ allphases ] ) and treats the wide - open spiral as part of the bar . also in this case , there is a strong _ coupling _ between the gas and star - dominated components , although the inner near - infrared morphology is far more regular . we use table 1 of bb01 to define three gravitational torque classes : the total torque class based on @xmath18 , the bar torque class based on @xmath2 , and the spiral torque class based on @xmath1 . the dust - penetrated type ( dp - type ) , written in the form `` harmonic class - pitch angle class - bar torque class '' following bb01 , is given in the last column of table 4 . except for ngc 7741 , _ the highest spiral torque classes all belong to the highest pitch angle class , @xmath88_. six of the galaxies could not be assigned a dp - type . the theory of bar - driven spirals is somewhat controversial . kormendy & norman ( 1979 ) showed that of 33 galaxies having differential rotation and global spiral structure , nearly 80% are barred and likely to have their spirals driven by the bars . the implication was that bars and spirals would have the same pattern speeds . however , sellwood & sparke ( 1988 ) showed that in @xmath99-body numerical simulations where an initially unstable disk forms both a bar and a spiral , the spiral has a lower pattern speed than the bar . sellwood & wilkinson ( 1993 ) suggested that the quadrupole moment of a realistic bar falls off too rapidly to drive the spirals well beyond the ends of the bar . in this circumstance , the spiral could be an independent instability , or a driven response to a resonance interaction between the bar and spiral . tagger et al . ( 1987 ) and yuan & kuo ( 1997 , 1998 ) have shown that resonance interactions between the bar and the spiral can be an efficient mechanism of wave amplification . the question of whether spirals and bars corotate can be answered in some cases by the observed alignment of resonance rings . inner rings and pseudorings in sb galaxies tend to be aligned parallel to bars , while outer rings and pseudorings may be aligned parallel or perpendicular to bars ( kormendy 1979 ; schwarz 1984 ; buta 1986 , 1995 ) . the morphologies of the various ring types suggest they are driven by bars : inner rings often have pointy oval shapes , indicating a connection with the inner 4:1 ultraharmonic resonance ( uhr , buta & combes 1996 ; salo et al . 1999 ) , while outer pseudorings have three distinctive morphologies that suggest a connection to the outer lindblad resonance ( olr ) . these are the r@xmath100 , r@xmath101 , and r@xmath102r@xmath101 morphologies known as the " olr subclasses ( buta 1985 , 1986 , 1995 ; buta & crocker 1991 ) . the first two types were predicted by schwarz ( 1981 ) from test - particle simulations as being possibly due to differences in the gas distribution around the olr , while byrd et al . ( 1994 ) showed that the three morphologies may be connected in an evolutionary manner . since resonance rings probably develop from the secular evolution of spiral patterns ( schwarz 1981 ; byrd et al . 1994 ; rautiainen & salo 2000 ) , the frequent alignments suggest that the spirals that form the rings are driven by the bars at the same pattern speeds . buta & combes ( 1996 ) further argue that the co - existence of several rings in the same galaxy , showing shapes and alignments that are compatible with periodic orbits near resonances ( e.g. , ngc 3081 , buta & purcell 1998 ) , speaks against the existence of several patterns with different pattern speeds . salo et al . ( 1999 ) show that a model that assumes the bar and spiral co - rotate can fit the morphology and the velocity field of the early - type ringed galaxy ic 4214 . the situation is a little different with the central regions of barred galaxies . here it is possible for the primary bar to induce enough mass flow into the central kiloparsec to cause an inner bar instability , leading to formation of a secondary bar with a different pattern speed ( pfenniger & norman 1990 ) . support for the idea of a different pattern speed comes from studies of relative primary bar , secondary bar position angles ( e.g. , buta & crocker 1993 ; wozniak et al . 1995 ) . rautiainen & salo ( 2000 , hereafter rs ) have reconsidered ring formation in @xmath99-body models with dissipatively colliding test particles . these models show the same features as previous test - particle models except that with the self - gravitating stellar disk , additional spiral modes can develop in the same manner as in sellwood & sparke ( 1988 ) . rs consider how these modes , which often have a pattern speed less than that of the bar , affect the morhology and formation of outer rings and pseudorings . the olr subclasses of outer rings and pseudorings still form in these models , but cyclic changes between the morphologies can occur due to the influence of the modes with a lower pattern speed . if the influence of the lower pattern speed modes is significant enough , rs show that misalignments may be possible . observed misalignments between inner and outer rings and bars are rare but do occur ( buta 1995 ; buta , purcell , & crocker 1995 ) . the only prominent ringed galaxies in our sample are ngc 1433 and ngc 7723 . of these , ngc 1433 shows many features in common with test - particle simulations , including a highly elongated and aligned inner ring likely connected with the inner uhr , an r@xmath100 outer pseudoring likely connected with the olr , a misaligned nuclear ring , and two secondary spiral arcs , or plumes , in the vicinity of the outer uhr ( buta et al . elmegreen & elmegreen ( 1985 ) argue that a bar like that in ngc 1433 could still be growing through the slow loss of angular momentum to the stellar spiral . in ngc 7723 , deprojection of a blue - light image indicates that the inner ring is a nearly circular feature from which a prominent multi - armed spiral pattern emerges . this galaxy does not fit into available test - particle simulations nearly as well as does ngc 1433 . we can see in the plots in figure [ qtvsr ] that some of the bar torques in our sample are weak in the spiral region , as predicted by sellwood & wilkinson ( 1993 ) . this is the case for ngc 972 , 1255 , 5033 , and 6814 . this may also be true in the milky way . our galaxy has a bar about 1.5 kpc long which generates the 3 kpc arm and perhaps makes the bulge , which is clearly bar - like and about the right length ( see , e.g , cole & weinberg 2002 ) . corotation is at @xmath383 kpc . the spirals outside this radius , which include the local spiral , the sgr - carina spiral , and the perseus spiral , may be independent of the bar . this means their pattern speed is unrelated to the bar pattern or perhaps related in a complex way through resonant excitation . in either case , the bar strength and the spiral arm strength should not be simply related . for resonance excitation , even a weak perturbation can cause a strong response . another test of the corotation of bars and spirals concerns the place in azimuth where the inner part of the spiral lies relative to the bar . if the pattern speeds are different or the bar and spiral are unrelated , then the distribution of azimuths for these inner - limit spiral points should be uniform . if all spirals end in their inner regions at the same angular distance from the end of the bar , then the bars and spirals are probably related . to provide a preliminary answer to this question , we examined the twelve barred galaxies in our sample and visually estimated @xmath103 , the angle between the bar and the inner - limit spiral points , for each prominent spiral arm in each galaxy . table 3 compiles @xmath104 , the average over the prominent arms , and figure [ thetas ] shows how this average correlates with @xmath105 . the plot shows that for all but two of the galaxies , the spirals appear to begin within 20@xmath5 of the bar axis . the two objects having @xmath104 @xmath39 40@xmath5 are ngc 5033 and 7723 , the latter having three spiral arms forming part of the inner ring . thus , for none of the most strongly barred galaxies in our sample does the near - ir spiral appear to begin at a large intermediate angle to the bar axis . however , in blue light , the situation can be different . for example , both ngc 1530 and 6951 could be interpreted as having serious mismatches between the ends of the bar and the beginnings of the spiral . on the other hand , the weak inner pseudorings in these two objects are made from arms that break from the ends of the bar and wrap around the other ends . from a sample of 17 bright spiral galaxies having a range of bar and spiral morphologies and hubble types , we find the following from @xmath0-band images : \1 . bars and spirals can be effectively separated using a fourier - based image analysis technique ( bbk ) . from this separation , we can quantify the bar and spiral strengths in terms of tangential - to - radial force ratios . in the sample , bar strengths @xmath2 range from 0 to 0.75 while spiral strengths @xmath1 range from 0 to 0.46 . the spiral strength @xmath1 correlates with @xmath2 in a nonlinear fashion : spirals form in non - barred galaxies so @xmath1 is independent of @xmath2 when @xmath105 @xmath60 0.3 , but spiral strengths increase linearly with bar strengths when bars are present , for which @xmath106 . the effect is not an artifact of stellar @xmath56 variations or dark matter in these high luminosity spirals . if bars and spirals grow together in a global disk instability , then the linear relation between their strengths imply they have about the same growth rates . the bar - interbar contrast @xmath74 and the deprojected bar ellipticity @xmath65 correlate with @xmath105 , with a smaller scatter for @xmath65 . the arm - interarm contrast also correlates with @xmath1 , with a better correlation for the contrast measured at a radius of 0.25@xmath107 compared to 0.5@xmath107 . strong bars tend to be longer relative to the galaxy isophotal radius . this length - strength correlation may be the result of bar elongation over time as more and more stellar orbits join the bar potential ( see simulations in combes & elmegreen 1993 ) . dust - penetrated classifications indicate that the strongest bars with the strongest spirals tend to have the most open spiral arms . their pitch angles are 40@xmath5 or more and their pitch angle class is @xmath88 . the openness of a spiral arm depends on the relative size of the toomre ( 1964 ) length , @xmath108 for total disk column density @xmath109 and epicyclic frequency @xmath110 . this length is approximately the separation between the arms . if this length is large compared to the radius , then the arms are very open . the ratio of the toomre length to the galaxy radius is approximately the ratio of the disk mass to the halo mass . thus open spirals have relatively massive disks ( see also bertin et al . 1977 ; elmegreen & elmegreen 1990 ) . our observations imply that the most massive disks have the strongest bars and spirals . this suggests that bars and spirals form together in a global disk instability . the correlations between the maximum relative gravitational torques of the bars and spirals and between a number of other measures of bar and spiral amplitudes , in addition to the small angular displacements between the ends of the bars and the inner parts of the spirals , and the alignments of most resonance rings with their bars , all suggest that the bars and spirals in most of our sample have shared the same pattern speeds for cosmological times . in these cases , the bar and spiral corotation radii are the same . other cases with more irregular ring structures or no evident correlations between bar and spiral strengths or alignments could have their spirals excited either independently of the bar , or excited at a higher order resonance , giving the bars and spirals different pattern speeds and no correlation in structure . the milky way is apparently in this latter category . these conclusions do not necessarily apply to sb0 galaxies , which have strong bars and weak ( or no ) spirals . this difference from our present sample illustrates the changing morphology of bar - spiral patterns over time . our results here suggest that the growth phase of a strong bar - spiral pattern is the result of a combined instability having one pattern speed and one growth rate in a relative massive disk . this growth may be spontaneous or it may follow an encounter with another galaxy . after some time , the bar - spiral growth should slow or stop , but it does this for different reasons in the bar and spiral regions . bars stop growing when they saturate to very large strengths , placing most of the disk stars within the inner 4:1 resonance inside the bar . the inner stellar population is very hot when this happens , because the stellar orbits are highly non - circular . spirals stop growing when the stellar disk near corotation also heats up , but for the spirals , the high stellar velocities are more random than for the bar , removing the spiral pattern . the high stellar dispersion also requires a nearly complete conversion of gas into stars so that gas dissipation is absent and young stars with low dispersions no longer form . the result is a relatively strong , squared - off bar with little dust and star formation structure in the disk and very weak or no spiral arms , making the sb0 class . we thank an anonymous referee for helpful comments on this manuscript . we also thank s. stedman for assistance during the wht observations . rb acknowledges the support of nsf grant ast 0205143 to the university of alabama . bge was partially funded by nsf grant ast-0205097 . dlb is indebted to the anglo american chairman s fund for continued support ; in particular , he wishes to thank m. keeton , h. rix and the board of trustees . 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 nasa . the wht is operated on the island of la palma by the isaac newton group in the spanish observatorio del roque de los muchachos of the instituto de astrofsica de canarias . llccccccc ngc 908 & sa(s)c & 1.78 & 0.36 & 10.35 & 17.8 & @xmath620.9 & wht - ingrid & 73 + ngc 972 & sab & 1.55 & 0.29 & 11.48 & 21.4 & @xmath620.2 & wht - ingrid & 64 + ngc 1058 & sa(rs)c & 1.51 & 0.03 & 11.55 & 9.1 & @xmath618.3 & wht - ingrid & 62 + ngc 1255 & sab(rs)bc & 1.62 & 0.20 & 11.26 & 19.9 & @xmath620.2 & wht - ingrid & 60 + ngc 1365 & sb(s)b & 2.05 & 0.26 & 9.93 & 16.9 & @xmath621.2 & ctio - cirim & 40 + ngc 1433 & ( r@xmath111)sb(r)ab & 1.81 & 0.04 & 10.64 & 11.6 & @xmath619.7 & ctio - cirim & 30 + ngc 1530 & sb(rs)b & 1.72 & 0.28 & 11.42 & 36.6 & @xmath621.4 & wht - ingrid & 48 + ngc 1808 & ( r)sab(s)a & 1.81 & 0.22 & 10.43 & 10.8 & @xmath619.7 & wht - ingrid & 56 + ngc 5033 & sa(s)c & 2.03 & 0.33 & 10.21 & 18.7 & @xmath621.2 & wht - ingrid & 16 + ngc 6643 & sa(rs)c & 1.60 & 0.30 & 11.14 & 25.5 & @xmath620.9 & wht - ingrid & 63 + ngc 6814 & sab(rs)bc & 1.54 & 0.03 & 11.32 & 22.8 & @xmath620.5 & wht - ingrid & 60 + ngc 6951 & sab(rs)bc & 1.68 & 0.08 & 10.71 & 24.1 & @xmath621.2 & wht - ingrid & 60 + ngc 7217 & ( r)sa(r)ab & 1.63 & 0.08 & 10.53 & 16.0 & @xmath620.5 & wht - ingrid & 64 + ngc 7479 & sb(s)c & 1.63 & 0.12 & 11.22 & 32.4 & @xmath621.3 & wht - ingrid & 59 + ngc 7606 & sa(s)b & 1.73 & 0.40 & 10.88 & 28.9 & @xmath621.4 & wht - ingrid & 58 + ngc 7723 & sb(r)b & 1.55 & 0.17 & 11.57 & 23.7 & @xmath620.3 & wht - ingrid & 48 + ngc 7741 & sb(s)cd & 1.65 & 0.17 & 11.43 & 12.3 & @xmath619.0 & wht - ingrid & 100 + lllllllllll ngc 908 & 379 & ( 0.00 ) & 0.28@xmath70.10 & 0.28 & .... & 73.0 & 73.0 & .... & 0.40 & 0.40 + ngc 972 & 372 & 0.22@xmath70.06 & 0.21@xmath70.06 & 0.24 & 9.0 & 29.0 & 8.0 & 0.08 & 0.27 & 0.08 + ngc1058 & 100 & ( 0.00 ) & 0.09@xmath70.02 & 0.09 & .... & 18.0 & 18.0 & .... & 0.19 & 0.19 + ngc1255 & 482 & 0.07@xmath70.01 & 0.15@xmath70.04 & 0.17 & 4.5 & 23.5 & 76.0 & 0.04 & 0.19 & 0.61 + ngc1365 & 1587 & 0.40@xmath70.11 & 0.36@xmath70.11 & 0.49 & 118.5 & 178.0 & 164.0 & 0.35 & 0.53 & 0.49 + ngc1433 & 581 & 0.37@xmath70.06 & 0.23@xmath70.05 & 0.43 & 68.0 & 106.5 & 69.0 & 0.35 & 0.55 & 0.36 + ngc1530 & 777 & 0.61@xmath70.16 & 0.42@xmath70.16 & 0.73 & 45.0 & 65.5 & 49.0 & 0.29 & 0.42 & 0.31 + ngc1808 & 600 & 0.22@xmath70.04 & 0.10@xmath70.02 & 0.24 & 70.5 & 118.0 & 73.5 & 0.36 & 0.61 & 0.38 + ngc5033 & 841 & 0.07@xmath70.02 & 0.12@xmath70.03 & 0.12 & 8.5 & 29.5 & 29.5 & 0.03 & 0.09 & 0.09 + ngc6643 & 383 & ( 0.00 ) & 0.21@xmath70.08 & 0.21 & .... & 28.0 & 28.0 & .... & 0.23 & 0.23 + ngc6814 & 375 & 0.07@xmath70.01 & 0.09@xmath70.02 & 0.10 & 10.0 & 44.5 & 15.0 & 0.10 & 0.43 & 0.14 + ngc6951 & 640 & 0.28@xmath70.04 & 0.21@xmath70.06 & 0.34 & 31.5 & 57.5 & 43.0 & 0.22 & 0.40 & 0.30 + ngc7217 & 566 & ( 0.00 ) & ( 0.00 ) & @xmath600.04 & .... & .... & .... & .... & .... & .... + ngc7479 & 494 & 0.59@xmath70.10 & 0.46@xmath70.12 & 0.71 & 31.0 & 50.0 & 45.0 & 0.24 & 0.39 & 0.35 + ngc7606 & 898 & ( 0.00 ) & 0.08@xmath70.03 & 0.08 & .... & 42.5 & 42.5 & .... & 0.26 & 0.26 + ngc7723 & 508 & 0.30@xmath70.05 & 0.12@xmath70.02 & 0.31 & 16.0 & 33.0 & 16.0 & 0.15 & 0.31 & 0.15 + ngc7741 & 274 & 0.74@xmath70.22 & 0.35@xmath70.07 & 0.77 & 15.0 & 46.0 & 15.0 & 0.11 & 0.34 & 0.11 + ccccccccccl ngc 908 & m & 0.59 & 0.79 & .... & .... & .... & .... & .... & .... & ..... + ngc 972 & f & 0.96 & 1.27 & 0.53 & 0.11 & 0.51 & 0.10 & 0.51 & .... & 3:@xmath71 : + ngc 1058 & f & 0.38 & 0.52 & .... & .... & .... & .... & .... & .... & ..... + ngc 1255 & m & 0.56 & 0.67 & 0.20 & 0.06 & 0.18(exp ) & 0.06 & 0.18 & .... & 18@xmath717 + ngc 1365 & g & .... & 2.10 & 2.05 & 0.53 & 0.76(flat ) & 0.50 & 0.71 & 0.40 & 9@xmath74 + ngc 1433 & g & .... & 1.78 & 1.64 & 0.54 & 0.70(flat ) & 0.37 & 0.65 & 0.33 & 8@xmath72 + ngc 1530 & g & 1.67 & 3.01 & 2.35 & 0.43 & 0.74(flat ) & 0.33 & 0.67 & 0.38 & 15@xmath75 + ngc 1808 & g & 0.39 & 0.96 & 0.69 & 0.37 & 0.52(exp ? ) & 0.37 & 0.52 & 0.06 & ..... + ngc 5033 & m & 0.65 & .... & .... & ... & .... & .... & .... & .... & 67@xmath75 + ngc 6643 & m & 0.73 & 0.67 & .... & ... & .... & .... & .... & .... & ..... + ngc 6814 & m & 0.38 & 0.76 & 0.37 & 0.12 & 0.25(exp ) & 0.16 & 0.27 & .... & 4@xmath73 + ngc 6951 & g & 1.25 & 1.63 & 1.15 & 0.33 & 0.64(flat ) & 0.25 & 0.58 & 0.22 & 12@xmath77 + ngc 7217 & f & 0.19 & 0.34 & .... & .... & .... & .... & .... & .... & ..... + ngc 7479 & g & 2.03 & 1.46 & 1.47 & 0.39 & 0.71(flat ) & 0.23 & 0.68 & 0.93 & 7@xmath77 + ngc 7606 & g & 0.43 & 0.33 & .... & .... & .... & .... & .... & .... & ..... + ngc 7723 & m & 0.54 & 0.57 & 1.02 & 0.18 & 0.61(flat ) & 0.18 & 0.61 & 0.11 & 51@xmath722 + ngc 7741 & f & 1.35 & 0.94 & 1.34 & 0.14 & 0.73(exp ) & 0.14 & 0.73 & 0.32 & 7@xmath74 + cccccccl ngc 908 & 3 & 0 & 3 & 2460 & 2 & 28.07 & h2@xmath870 + & & & & & 3 & 38.66 & + ngc 972 & 2 & 2 & 2 & 619 & 2 & 38.66 & h2@xmath882 + ngc 1058 & 1 & 0 & 1 & 612 & 2 & 25.20 & h2@xmath870 + ngc 1255 & 2 & 1 & 2 & 827 & 2 & 31.61 & h2@xmath871 + ngc 1365 & 5 & 4 & 4 & ..... & 2 & ..... & h2@xmath884 + ngc 1433 & 4 & 4 & 2 & ..... & 2 & ..... & h2@xmath864 + ngc 1530 & 7 & 6 & 4 & 2958 & 2 & 33.69 & h2@xmath887 + ngc 1808 & 2 & 2 & 1 & 1848 & 2 & 69.44 & h2@xmath882 + ngc 5033 & 1 & 1 & 1 & 1024 & 2 & 28.07 & h2@xmath871 + ngc 6643 & 2 & 0 & 2 & 741 & .. & ..... & .... 0 + ngc 6814 & 1 & 1 & 1 & 836 & 4 & 34.82 & h4@xmath881 + & & & & & 2 & 11.31 & + ngc 6951 & 3 & 3 & 2 & 2460 & 2 & 45.00 & h2@xmath883 + ngc 7217 & 0 & 0 & 0 & ..... & .. & ..... & .... 0 + ngc 7479 & 7 & 6 & 5 & 2460 & 2 & 33.69 & h2@xmath886 + ngc 7606 & 1 & 0 & 1 & 1748 & 2 & 31.61 & h2@xmath870 + ngc 7723 & 3 & 3 & 1 & 1024 & .. & ..... & .... 3 + ngc 7741 & 8 & 8 & 4 & 2234 & .. & ..... & .... 8 +

the study of the formation and evolution of large - scale structure is essential to our understanding of the universe . the typical method to measure the distances of faint galaxies is to obtain their spectroscopic redshifts . the spectroscopic method , however , is very time consuming for surveying a large area with a deep limiting magnitude , even when using a multi - object spectrograph . the less accurate photometric redshift technique has been proved to be an efficient and effective way to measure approximately the redshifts of galaxies and to study their statistical properties and evolution ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? basically , the technique is a transformation between a set of observable parameters ( e.g. , magnitudes , colors ) and estimates of the physical properties of galaxies ( e.g. , redshift , luminosity ) . often the photometric redshift technique is applied to multi - color deep imaging surveys over a small area ( e.g. , the hubble deep field , sawicki et al . 1997 and fernndez - soto et al . 1999 ; the canada - france deep fields , @xcite ; the combo-17 survey , @xcite ; the las campanas infrared survey , @xcite ; the goods - s , @xcite ) . these surveys have sufficient depth for measuring objects at high redshift ( @xmath11 ) , but they only probe small cosmological volumes which might not provide a representative sample of the universe due to the limited survey area . with more and more large multi - color imaging surveys , larger and statistically more complete studies of galaxies can be done . although the application of photometric redshift techniques to the sloan digital sky survey ( sdss ) covers a large sky area and provides a huge redshift sample , the photometric data are still too shallow ( @xmath12 @xmath13 21 ) for galaxies at @xmath14 @xcite . the oxford - dartmouth thirty degree survey ( odts , @xcite ) has both sufficient depth and a relatively large area ( @xmath1530 deg@xmath4 ) , but it does not have a large spectroscopic training set to calibrate the photometric redshifts . a deep multicolor survey with large area and proper spectroscopic data will vastly improve on the current results . the red - sequence cluster survey ( rcs , * ? ? ? * ) is a 4m - class imaging survey in the @xmath1 and @xmath0 bands covering @xmath16 deg@xmath4 total in the northern and southern sky . we have obtained follow - up observations in @xmath2 and @xmath3 for a sub - area of the rcs . the four color photometry ( @xmath0 , @xmath1 , @xmath2 , and @xmath3 ) allows us to cover a relatively large sub - area ( 33.6 deg@xmath4 ) which samples a large cosmological volume for galaxies up to redshift @xmath15 1.5 . this sub - area with four - color photometry overlaps with the canadian network for observational cosmology ( cnoc2 ) field galaxy redshift survey ( yee et al . 2000 ; 2005 , in preparation ) which provides several thousand redshifts which can be used for the empirical photometric redshift technique and provide verification of the photometric redshifts obtained . by applying an empirical quadratic polynomial photometric redshift fitting method to the rcs photometric catalogs , a photometric redshift catalog containing about one million galaxies at @xmath17 with relatively small redshift errors can be generated . in this paper we describe details of the rcs follow - up @xmath2 and @xmath3 observations and the techniques used to generate the photometric redshift catalog for the rcs survey . in [ data ] we describe the data we used for the photometric redshift fitting and the photometric calibrations for the @xmath2 and @xmath3 data . in [ trainingset ] we provide a description of the spectroscopic training sets used in the photometric redshift fits . section [ photoz - method ] presents the photometric redshift technique we used . we describe the method used to estimate the photometric redshift error in [ photoz - err ] . we then demonstrate the results and present an example of the catalog in [ result ] . in [ conclusion ] we summarize the advantage of the photometric redshift catalog from the rcs and discuss possible follow - up studies based on the catalog . the rcs @xcite was designed to find clusters in a large volume , at high redshift , and to low masses . the primary goals are to define a large sample of galaxy clusters at @xmath18 , hence providing a measurement of @xmath19 and @xmath20 , and to perform studies of the evolution of galaxy clusters using a large complete sample . two filters ( @xmath0 and @xmath1 ) were chosen to find clusters at @xmath21 by applying the cluster red - sequence method @xcite . there are 19 widely separated patches in both the northern and southern sky , covering @xmath16 deg@xmath4 in total . the northern half of the survey was observed with the cfh12k ccd camera on the 3.6 m canada - france - hawaii telescope ( cfht ) , and the southern half was obtained using the mosaic ii camera on the cerro tololo inter - american observatory ( ctio ) 4 m blanco telescope . for the cfht rcs runs , each patch consists of fifteen cfh12k pointings arranged in a slightly overlapping grid of 3 @xmath22 5 pointings . the cfh12k camera is a 12k @xmath22 8k pixel@xmath23 ccd mosaic camera , consisting of twelve 2k @xmath22 4k pixel@xmath23 ccds ( from mit lincoln laboratories ) arranged in a 6 @xmath22 2 grid . it covers a 42 @xmath22 28 arcminute@xmath4 area for the whole mosaic at prime focus ( f/4.18 ) , corresponding to 0.2059 arcseconds per pixel . we note that the average east - west ( x ) gap between the ccds is 7.8 arcseconds , and the average north - south ( y ) gap between the ccds is 6.8 arcseconds . the observation dates are from may 1999 to january 2001 . the typical seeing is 0.62 arcsec for @xmath0 and 0.70 arcsec for @xmath1 . the integration times are typically 1200s for @xmath0 and 900s for @xmath1 , with average limiting magnitudes of @xmath24 and @xmath25 ( vega ) for point sources , adoping an aperture of diameter 2.7 arcsec ( on average ) . the data and the photometric techniques are described in detail in @xcite . the follow - up observations in @xmath2 and @xmath3 were obtained using the cfh12k camera and cover 33.6 deg@xmath4 ( @xmath26% complete for the original cfht rcs fields ) . the follow - up @xmath2 and @xmath3 observations contain 108 pointings spread over the ten cfht rcs patches . the pointings of the cfht rcs patches are shown in table [ patch ] . for each pointing , a single exposure in each field was taken without dithering . the observation dates are from may 2001 to june 2002 . most data were taken in three runs in queue mode : 01ac44 , 01bc25 , and 02ac25 . the data for another 29 pointings were taken using the observing time ( run 01bc28 ) shared with the explore project @xcite . some data were also taken during two cnoc2 photometric redshift runs using cfh12k in january 2000 and december 2000 . we had 16 photometric nights out of 50 nights in which data were taken . because some data were taken under non - photometric conditions , short exposures were obtained in photometric weather for calibration . the typical seeing is 0.65 arcsec for @xmath2 and 0.95 arcsec for @xmath3 . the average exposure times for @xmath2 and @xmath3 are 480s and 840s , respectively , and the median limiting magnitudes ( vega ) for point sources are 24.5 and 25.0 , respectively , with an adopted aperture of diameter 2.9 arcsec ( on average ) . we present histograms of the @xmath2 and @xmath3 5@xmath27 limiting magnitudes in figure [ vb_lim ] . ccccccc 01 & 0226 + 00 & 02 26 07.0 & + 00 40 35 & 4.81 & 3.99 & cnoc2 patch@xmath28 + 02 & 0351 - 09 & 03 51 20.7 & -09 57 41 & 4.79 & 4.30 & + 03 & 0926 + 37 & 09 26 09.6 & + 37 10 12 & 5.59 & 4.85 & cnoc2 patch@xmath28 + 04 & 1122 + 25 & 11 22 22.5 & + 25 05 55 & 4.78 & 4.72 & + 05 & 1327 + 29 & 13 27 41.9 & + 29 43 55 & 4.54 & 1.34 & pdcs patch@xmath29 + 06 & 1416 + 53 & 14 16 35.0 & + 53 02 26 & 4.53 & 3.04 & groth strip + 07 & 1449 + 09 & 14 49 26.7 & + 09 00 27 & 4.17 & 2.01 & cnoc2 patch@xmath28 + 08 & 1616 + 30 & 16 16 35.5 & + 30 21 02 & 4.26 & 4.16 & + 09 & 2153 - 05 & 21 53 10.8 & -05 41 11 & 3.43 & 2.96 & cnoc2 patch@xmath28 + 10 & 2318 - 00 & 23 18 10.7 & -00 04 55 & 4.84 & 2.23 & + photometric data reduction including object finding , star - galaxy classification , and aperture photometry was performed using the program ppp , described in detail in @xcite and @xcite . the rcs @xmath0 and @xmath1 object finding and photometry were already done as part of the primary cluster finding program , and are described in detail in @xcite . to simplify the photometric measurement procedure , we used the pixel positions of objects in the @xmath1 image to do photometry for @xmath2 and @xmath3 . to achieve this , the position of each object in @xmath2 and @xmath3 must be accurately determined to within two pixels in radius relative to the @xmath1 image , sufficiently close for ppp to do a recentroiding of the object position before performing photometry . however , the follow - up observations in @xmath2 and @xmath3 began a semester after the observations in @xmath0 and @xmath1 were finished . the cfh12k ccd was taken off and re - installed on the prime focus for each run , and the camera was not at the identical position after each re - installation . hence , the @xmath2 and @xmath3 images are rotated slightly relative to the @xmath0 and @xmath1 images for the same field . the telescope pointing is also not identical . to deal with these problems , we match the @xmath2 and @xmath3 images to the @xmath1 image before performing photometry . because we did not dither to make one large image per pointing , we match the images chip by chip . the transformation function between the @xmath2 and @xmath3 images and the @xmath1 images can be easily determined from bright reference objects . however , the new transformed coordinates for each pixel are almost always not integer numbers after applying the transformation function with non - integer shifts and rotation . therefore re - sampling is needed . to avoid degradation of the image quality and to preserve the poisson characteristics of the photon noise in the images , we use the nearest neighbor re - sampling algorithm , which also preserves flux . we use the closest integer coordinate of the transformed non - integer coordinate as the new coordinate for each original pixel . however , two original pixels may produce two transformed non - integer coordinates which may have the same nearest integer coordinate . in other words , one new pixel may be assigned fluxes from two original pixels because of round off in the pixel coordinates . in addition , some new coordinates may have zero flux because they are not the nearest coordinates for any original pixels . the photometry results would not be correct when either condition occurs within a photometry aperture . we try to deal with the problem by subtracting the local background from a pixel containing double fluxes because that pixel will also contain double the background counts . on the other hand , we put the counts of the local background into those pixels with zero flux . these procedures minimize the errors on the photometry . we note that on average , only approximately 500 pixels ( ranging from 0 to 3,000 depending on the rotation angle ) are affected for each chip . given that the total number of objects in each chip is @xmath30 and that the average diameter of the photometry aperture is 14 pixels , we estimate that only @xmath31 pixels in photometry apertures are affected in each chip ; i.e. , only one object out of 100 is affected and the effect has been minimized by the method we described above ( the magnitude difference before and after the transformation for an affected object is less than 0.02 magnitude ) . additionally , the transformed image of each chip often overlaps with several nearby chips of the original image , and the matched regions from other chips are also cropped and included in the transformed image . we note that within each pointing , images from all the chips have been scaled to have the same zeropoint . an example of a re - sampled @xmath2 image is shown in figure [ matching - image ] . the gaps between chips are also filled with counts of the local background . objects falling into the gaps are lost . the final transformed @xmath2 and @xmath3 images are matched with the @xmath1 images for the same fields . the typical shift is between a few pixels to one hundred pixels . the typical rotation angle is about 0.03 degree and the largest angle is less than one degree . after the matching is done , we apply the coordinate file of objects from the @xmath1 image to perform the aperture photometry procedure . since the same coordinate file is used for the data in different filters , the object i d of each object in different filters of the same pointing remains the same . the multi - color photometric catalog can be generated very easily by just combining the photometric results object by object directly . the colors are measured using identical apertures for all four filters , as prescribed in @xcite . the preliminary @xmath2 and @xmath3 photometric calibrations ( zero - point , airmass , and color terms ) were taken directly from the cfht elixir solutions in the image fits file headers . the elixir system at cfht provides real - time data quality assessment , end - of - run detailed calibration analysis , and image pre - processing and meta - data compilation for data distribution . data taken under bad weather conditions are calibrated by short exposures taken in photometric nights . however , we find that the color ( @xmath32 or @xmath33 ) vs. color ( @xmath34 ) diagrams have offsets from pointing to pointing , and sometimes the offsets are as large as @xmath35 magnitude , which implies that the photometric calibrations are not sufficiently accurate . an example of the differences between two pointings on the color - color diagram is shown in figure [ color - color ] . hence , we use the following method to recalibrate the photometry . according to a recent star count study @xcite , although the densities of stars would be different for different fields , the normalized color distributions for bright stars are very similar from field to field as long as the selected fields are sufficiently large . based on this result , we recalibrated the @xmath0 , @xmath2 , and @xmath3 photometry by using the @xmath36 , @xmath32 , and @xmath33 color distributions , assuming that the photometry of @xmath1 calibrated by standard stars is correct @xcite . stars with @xmath37 are selected to determine the magnitude offsets on a pointing - to - pointing and patch - to - patch basis . figure [ color - histo ] represents an example of the difference between the color histograms of two pointings . for the pointing - to - pointing recalibration within each patch , the pointing with the least scatter in the color - color diagrams ( @xmath32 or @xmath33 vs. @xmath34 ) is chosen to be the reference pointing . ( the reference pointings are 0226a3 , 0351c1 , 0926b4 , 1122c5 , 1327b2 , 1416c1 , 1449b2 , 1616c5 , 2153a1 , and 2318b1 for the 10 patches . ) the magnitude offsets in @xmath0 , @xmath2 , and @xmath3 are computed using cross - correlation of the distributions of @xmath36 , @xmath32 , and @xmath33 , respectively , between the reference pointing and the test pointing . the patch - to - patch photometry recalibrations are performed after the pointing - to - pointing recalibrations are done . patch 0926 is chosen to be the reference patch . the same recalibration procedure as the pointing - to - pointing recalibration is applied to the patch - to - patch recalibration , and all the pointings from each patch are used during the patch - to - patch recalibration procedure . we compare the recalibrated photometry of galaxies with the overlapping published sdss database as a check on our recalibration procedure . patches 0226 , 2318 , 0926 , 1416 , 1449 , and 1616 overlap with the sdss data release 3 ( dr3 ; * ? ? ? * http://www.sdss.org/dr3/ ) . we match the objects between rcs and sdss , and use the following equations to determine the relation between the two different photometry systems : @xmath38 where the @xmath39 term is the coefficient of the color term in the transformation , and @xmath40mag , where mag @xmath41 , is the magnitude zeropoint difference between the rcs and sdss photometry systems . we compare rcs magnitudes against sdss model magnitudes . the quality of the recalibration can be verified by checking the consistency of the @xmath40mag values between different rcs patches . in figure [ sdss - offset ] we plot @xmath40mag for @xmath3 , @xmath2 , and @xmath0 for those rcs patches overlapping the sdss , where we have arbitrarily set @xmath40mag @xmath42 for patch 0226 _ after _ recalibration . for these patch - to - patch comparisons we have ignored the slope terms above and simply set @xmath40mag to the median offset between the respective rcs and sdss magnitudes for each patch . the open boxes indicate @xmath40mag using elixir calibrations , and the filled boxes indicate @xmath40mag after recalibration . the recalibrated data have less scatter between patches than the elixir - calibrated data in all the panels , especially for @xmath3 . figure [ sdss - sd ] shows the standard deviations of the @xmath40mag values for @xmath3 , @xmath2 , and @xmath0 among the different pointings within each rcs patch . here we first fit the above transformation equations to each patch as a whole , transform the sdss magnitudes accordingly , and then compute the median offset between rcs magnitudes and transformed sdss magnitudes for each pointing within a patch . the standard deviations of these pointing - by - pointing offsets are plotted in figure [ sdss - sd ] . the open boxes and the filled boxes again indicate the elixir - calibrated results and the recalibrated results , respectively . the pointing - to - pointing recalibrations reduce the standard deviations dramatically for @xmath2 and @xmath3 . this rcs - sdss comparison proves that recalibration using the color distributions of bright stars significantly improves the calibrations of the @xmath2 and @xmath3 photometry over the original elixir - based calibrations , and that most of the original photometric calibrations ( as performed by rcs , @xcite ) for @xmath0 are already good . galactic extinction has to be applied to the data before doing photometric redshift fitting . the extinction values are calculated according to prescriptions from @xcite , @xcite , and @xcite . the galactic extinction values for the four filters for each patch are shown in table [ extinction ] . note that our procedure will tend to result in slightly wrong galactic extinction corrections for the @xmath3 , @xmath2 , and @xmath0 filters , since the stars used in the recalibration procedure already suffer extinction to some extent and the recalibration procedure will have corrected for some of the extinction differences among the patches . however , the error we make is reduced because we use colors in the recalibration and because we use relative shifts among the patches . for example , if we assume that the stars we use already suffer the full amount of galactic extinction , then for patch 2318 , which has the largest @xmath43 value , we would have made only errors of -0.049 , -0.022 , and 0.024 mag for @xmath3 , @xmath2 , and @xmath0 , respectively . moreover , the comparison against sdss data described above and shown in figure [ sdss - offset ] indicates that the extinction of stars does not cause large errors in our recalibration procedure . crrrrrrrrrr e(b - v ) & 0.036 & 0.043 & 0.012 & 0.018 & 0.012 & 0.010 & 0.029 & 0.038 & 0.035 & 0.044 + @xmath44 & 0.140 & 0.165 & 0.048 & 0.068 & 0.046 & 0.037 & 0.112 & 0.146 & 0.135 & 0.167 + @xmath45 & 0.108 & 0.127 & 0.037 & 0.053 & 0.035 & 0.028 & 0.086 & 0.113 & 0.105 & 0.129 + @xmath46 & 0.083 & 0.098 & 0.029 & 0.040 & 0.027 & 0.022 & 0.066 & 0.087 & 0.080 & 0.099 + @xmath47 & 0.053 & 0.063 & 0.018 & 0.026 & 0.018 & 0.014 & 0.043 & 0.056 & 0.052 & 0.064 + since we use an empirical quadratic polynomial fit to estimate photometric redshifts for the rcs data , we need spectroscopic data which overlap with the rcs fields to create a training set . we primarily use the spectroscopic data from the cnoc2 project , but we also include goods / hdf - n data to improve the limiting magnitude and redshift range of the training set . all the data included in our training set are described in detail below . the canadian network for observational cosmology ( cnoc2 ) field galaxy redshift survey ( yee et al . 2000 ; 2005 , in preparation ) is a spectroscopic / photometric survey of faint galaxies . there are four widely separated patches named 0226 , 0926 , 1449 , and 2153 . the data were obtained using the multi - object spectrograph on the cfht . the survey covers over 1.5 deg@xmath4 of sky with a total sample of @xmath48 redshifts with @xmath49 . the cnoc2 survey used a band - limiting filter for the spectroscopic observations to increase the number of objects observed . this produces an effective redshift range for the statistically complete sample of 0.12 - 0.55 , and 0.0 - 0.68 for emission - line galaxies . the nominal statistical completeness magnitude is @xmath50 , but there are objects as faint as @xmath51 in the redshift sample . fifteen rcs pointings , distributed over four rcs patches and including four - filter photometric data , overlap the cnoc2 fields . we match the objects from the two surveys and create a training set containing the photometric data from the rcs and the spectroscopic redshifts from cnoc2 . there are 3,130 objects in this training set . compared to the spectroscopic limit of cnoc2 , the photometric data of rcs are much deeper ( 100% completeness limit @xmath52 for rcs compared to the spectroscopic completeness limit @xmath50 for cnoc2 ) . furthermore , the cnoc2 spectroscopic sample has a nominal redshift limit of @xmath15 0.55 due to the use of a band - limiting filter , and so does not provide a good training set for objects with @xmath53 at @xmath54 . a deeper dataset ( limiting magnitude of @xmath55 ) with higher redshifts ( @xmath56 ) is thus needed in addition to the original training set to improve redshift estimates for high-@xmath57 objects in the rcs . thus , we choose the goods / hdf - n data to be the additional training set for high-@xmath57 and fainter objects . the great observatories origins deep survey ( goods ; * ? ? ? * ) is a survey based on multi - band imaging data obtained with the hubble space telescope ( hst ) and the advanced camera for surveys ( acs ) . it covers two fields ( hdf - n and cdf - s ) with roughly 320 arcmin@xmath4 total area . for our high - redshift , faint - magnitude training set , we use publicly available @xmath58 photometry and spectroscopic redshifts for the goods / hdf - n field . the photometry comes from the ground - based hawaii hdf - n data set of @xcite , which was obtained with the subaru 8.3 m telescope and has limiting magnitudes @xmath59 , @xmath60 , @xmath61 , and @xmath62 measured in 3 " diameter apertures , and typical integration times of 600s , 1200s , 480s , and 180s/240s , respectively . the spectroscopic redshift data for the goods / hdf - n field are from the samples of @xcite and @xcite , obtained using the keck 10 m telescope . after combining and matching objects in the hawaii hdf - n photometric catalog to the goods / hdf - n spectroscopic catalog , a training set containing 1,794 objects is generated . this additional training set is not only deeper than the rcs+cnoc2 set , but it also contains many more objects with higher redshifts ( @xmath63 ) . to combine the cnoc2 and goods training sets , we also need to offset the magnitudes of the goods data to match the zero - points of the rcs data . we did not use stars to derive the magnitude offsets because the field is too small to contain a statistically sufficient number of stars for the purpose . we first apply galactic extinction corrections to the goods / hdf - n data ( @xmath64 , @xmath65 , @xmath66 , and @xmath67 for @xmath0 , @xmath1 , @xmath2 , and @xmath3 ) , and then compute the difference in the color distributions for galaxies with @xmath68 between the rcs and goods / hdf - n samples . by assuming @xmath1 is always correctly calibrated , the magnitude offsets for @xmath0 , @xmath2 , and @xmath3 can be determined . magnitude offsets @xmath69 , @xmath70 , and @xmath71 for @xmath0 , @xmath2 , and @xmath3 , respectively , are applied to the goods / hdf - n data . the final training set created by combining the rcs+cnoc2 and the goods / hdf - n data includes 4924 objects . this combined training set will improve the accuracy of photometric redshifts , especially for objects at fainter magnitudes and higher redshifts . photometric redshift techniques have been developed for decades ( e.g. , * ? ? ? * ) , but there are two primary approaches to estimating the photometric redshift . one way is to compare the photometric data against templates generated from models or from a real spectral energy distribution database ( e.g. , hogg et al . 1998 ; fernndez - soto et al . 1999 ) . the other way is to find the empirical relation between the photometric data and the redshift identified from the spectroscopic data , e.g. , empirical polynomial fitting ( connolly et al . the empirical method is especially effective ( e.g. , see @xcite ) when there is a large spectroscopic redshift data set available , as in the case of our rcs data set . we use empirical quadratic polynomial fitting to estimate photometric redshifts for the rcs data . first we need to generate a subset called a `` training set , '' which includes information on object i d , spectral redshift , and photometric data for each filter ( see section [ trainingset ] ) . we then fit this subset with the following second order equation using least - squares fitting : @xmath72 including the constant term , 15 parameters are derived from the fit . the above formula describes the empirical relation between the photometric data and the spectroscopic redshift . by applying this formula to the photometric data of an object , an estimated photometric redshift for that object is readily obtained . one can use a brute - force single fit for all the data , but fitting all galaxies with a single quadratic formula is not optimal , since different types of galaxies may have different fitting parameters in the quadratic formula . the single fit method gives a result with large scatter and systematic deviations ( see left panel , figure [ photoz_comp ] ) . to improve the photometric redshift results , different types of galaxies should be fit separately . we describe below two methods of separating galaxies into different samples . color is one of the important signatures for identifying the galaxy type . however , it is subject to k - corrections for galaxies at different redshifts . thus , other parameters have to be used to break the color - redshift degeneracy . roughly speaking , more distant galaxies have fainter magnitudes , and for a reasonable range of galaxy types , they also have redder observed colors . hence by using some appropriate boundaries to divide galaxies in the color - magnitude plane , we can mimic very roughly the effect of separating galaxies of different types at different redshifts . figure [ color - mag ] presents a color ( @xmath33 ) vs. magnitude ( @xmath1 ) diagram . the sample data consist of rcs+cnoc2 ( dots ) and goods ( crosses ) galaxies with @xmath6 in our training set ( see [ trainingset ] for a detailed description of the training set ) . there is a gap around @xmath33=1.8 to 2.0 which roughly separates early - type galaxies ( redder ) and late - type galaxies ( bluer ) . we divide the color - magnitude plane into ten parts . regions 1 - 4 are for redder galaxies , and region 5 - 10 are for bluer galaxies . this method produces a result with smaller scatter and systematic deviations than the single fit method ( middle panel , figure [ photoz_comp ] ) . however , the @xmath33 color uses filters that are too blue to make a good separation for galaxies with redshifts larger than 0.6 , where the 4000 break will be shifted to @xmath73 6500 . to solve this problem , a redder filter ( @xmath0 ) has to be used in the separation criteria to produce a better result . in our second method of separating the training set , we add one more color ( @xmath34 ) to the original two - dimensional color - magnitude plane to form a three - dimensional color - color - magnitude space . the kd - tree algorithm , which uses median values to divide up the data points in a k - dimensional space successively @xcite , is used to separate galaxies in our three - dimensional space . we use a kd - tree depth of five , so that the space is separated into 32 cells . each cell contains about 150 objects . the photometric redshift fitting procedure is applied to each cell separately . we tried other numbers of cells , and they give similar results . in general , using a larger number of cells gives slightly better fits , but we choose 32 cells as a compromise so as not to be in danger of overfitting ( i.e. , having too few objects per cell for a 15-parameter fit ) . the three - dimensional kd - tree method gives a better result than the method of cutting the color - magnitude plane into 10 regions ( see right panel , figure [ photoz_comp ] ) . figure [ photoz_comp ] compares the quality of the photometric redshifts obtained using the three different cutting methods described above . the panels from left to right are the photometric redshift vs.spectroscopic redshift diagrams obtained using : ( 1 ) brute force single fit for all data , ( 2 ) cutting into 10 regions in the color - magnitude plane , and ( 3 ) the kd - tree method with 32 cells . both the high-@xmath57 and low-@xmath57 ends are improved with much reduced systematics as we go from method ( 1 ) to method ( 3 ) . the scatter in the differences between photometric and spectroscopic redshifts is reduced as more advanced cutting algorithms are used . hence , we choose the three - dimensional kd - tree algorithm with 32 cells for our final photometric redshift catalog . by examining the distribution of the differences between spectroscopic redshifts and photometric redshifts for the training set , we can estimate the uncertainty in the redshift fits . however , this does not provide a measurement of the photometric redshift error for individual galaxies in the catalog . knowing the confidence limits on the photometric redshift measurements for individual objects is very important for subsequent science analyses . without the confidence limits for each object , the analyses based on the photometric redshift may suffer from catastrophic errors , which happen when the photometric redshift estimates are unknowingly very different from the true spectroscopic redshifts . by considering the photometric redshift error for each object and taking it into account in estimating the errors in a subsequent science analysis , one will obtain more realistic confidence limits in the analysis results . in the following subsections , we describe the photometric redshift error determined by comparing the photometric redshifts to the spectroscopic redshifts in our training set ( empirical error ) , and we also describe the method we use to estimate the photometric redshift error for individual objects ( computed error ) . figure [ zerr ] presents the accuracy of the photometric redshifts by comparing the photometric redshift and the spectroscopic redshift for the objects in the training set . the @xmath74 is the 68th percentile difference between the photometric redshift and the spectroscopic redshift in bins of 100 objects each along the spectroscopic redshift axis . this provides a general estimate of how statistically accurate our photometric redshift measurements are . for objects at @xmath75 , the rms scatter @xmath74 is less than 0.05 . the overall rms scatter is 0.068 . however , most objects at @xmath76 in the training set are from the goods / hdf - n sample , which have a much deeper limiting magnitude . the photometric redshift error for an object in the rcs at high-@xmath57 is thus under - estimated by the @xmath74 of the training set . by adding additional gaussian noise to the goods / hdf - n data , we simulate the photometric redshift error as if the goods / hdf - n data have the same depth and seeing conditions as the rcs data . the results are shown in figure [ zerr1 ] . for objects at @xmath8 , the rms scatter remains roughly the same , but it becomes @xmath77 for an object at @xmath78 . the extremely large rms for objects at @xmath79 is caused by the large systematic deviation of the photometric redshifts from the spectroscopic redshifts . the overall average rms scatter is 0.11 . this result shows the real error levels in the photometric redshift catalog . we also calculate the catastrophic error rate for different redshift ranges in bins of 100 objects each and show the result in figure [ catastrophic_err ] . we define the catastrophic error rate as the ratio of the number of objects with @xmath80 to the total number of objects for each redshift bin . the result is calculated using the rcs / cnoc2 and noise added goods / hdf - n data . note that the histogram has variable redshift bin widths because we choose widths that always include exactly 100 objects per bin . the catastrophic error rate is below 0.03 for @xmath81 . it becomes around 0.05 for @xmath82 due to larger @xmath74 . for redshifts higher than 1.2 , the catastrophic error rate is 0.25 , which is primarily due to the larger photometric redshift systematic error . basically the photometric redshift uncertainty for an individual object comes from two sources . one source is the error in the photometric data themselves , and the other source is the uncertainty in the empirical fitting parameters in the quadratic formula . as described below , we will use the combination of a monte - carlo method and a bootstrap method to estimate the effect of both these error sources and thereby compute the total photometric redshift error for each object . we refer to this error as the `` estimated '' or `` computed '' error . the measurement error of the photometry is estimated from two sources . one is from the uncertainty of the sky measurement ; the other is from the internal error of the magnitude measurement . the photometry errors are given by ppp @xcite . the monte - carlo method provides a way to estimate how the photometric errors statistically propagate into the error on the photometric redshift . by assuming that the distribution of the photometric uncertainties is gaussian , we generate 500 photometric measurements , using a normal distribution with a width equal to the photometric error and centered on the original magnitude in each filter . for the estimates of the uncertainties in the photometric redshift fitting parameters ( equation [ photozeqn ] ) , the training set is bootstrap - resampled 500 times to generate 500 bootstrapped training sets and corresponding solutions . by combining 500 simulated photometric data sets with 500 bootstrapped training sets , 250,000 photometric redshift estimates are produced for each galaxy . the median value of these 250,000 redshifts is chosen to be the photometric redshift of the galaxy , and the 68% width ( @xmath83 if the photometric redshift error distribution is gaussian ) is the estimated photometric redshift error . note that the photometric redshift error distribution is not necessarily gaussian . our computed error estimate is intended to cover the 68% error range , regardless of the shape of the photometric redshift error distribution . to test the quality of our photometric redshift error estimates , we compare the computed errors with the empirical errors for the galaxies in the training set . figure [ err_comp ] shows the comparison of the empirical error and the computed error for the rcs and goods / hdf - n data in our training set . the empirical error plotted is the 68th percentile difference between the photometric redshift and the spectroscopic redshift , in bins of 100 objects each along the 68% computed error axis . the computed error plotted is the median value of the computed errors of objects in the same bin of 100 objects . the computed errors agree very well with the empirical errors , with @xmath84empirical error - computed error@xmath85 . this near exact agreement of the empirical and computed error is fortuitous . when different methods of separating the training set are used , in general there is an offset and/or scaling of the computed error relative to the empirical error . however , in general , regardless of the exact fitting algorithm used , the computed errors are always well - behaved linear functions of the empirical error , showing that our method of estimating individual errors is robust . to produce the final photometric redshift catalog , we use the 32-cell kd - tree cutting method . from figure [ zerr1 ] , for objects at @xmath8 , the rms scatter of @xmath86 is less than 0.06 . for objects at @xmath87 , the rms scatter of @xmath86 is @xmath77 . the overall average rms scatter is 0.11 . it can be seen that objects with lower spectroscopic redshifts are over - estimated while objects with higher spectroscopic redshifts are under - estimated ( figure [ zerr1 ] ) . the bluest filter for the rcs is @xmath3(4500 ) and it is not blue enough for the 4000 break at @xmath88 . this creates a poor redshift estimate for low - z ( @xmath88 ) objects , and they tend to have higher photometric redshifts than real redshifts because their seds and the seds of objects at @xmath89 are very similar . for those objects with redshift higher than 1.2 , the 4000 break is moving out of the reddest filter ( @xmath0 ) . a behavior opposite to the low - z objects is seen for the high-@xmath57 objects , i.e. , they tend to have lower photometric redshifts than real redshifts . we present the computed error vs. photometric redshift in different @xmath1 bins in figure [ esterr_photoz ] . for the panels with @xmath90 , we plot a dot for every 15 objects . for the panel with @xmath91 , we plot a dot for every 20 objects . for the panel with @xmath92 , we plot a dot for every 30 objects , and for the panel with @xmath93 , every 50 objects . the curve in each panel is the median value for all the data points in each @xmath1 bin , not just for the data shown in the figure . note the different scales for the computed error axes in each panel . in general , for galaxies with @xmath94 with photometric redshift less than 0.8 , the errors are below 0.1 . for objects with @xmath92 , the errors become larger but there are still a significant number of galaxies with errors less than 0.1 . for objects with @xmath95 , the photometric redshift errors are greater than 0.3 , which is due to larger photometric errors and also to a higher fraction of late - type galaxies . generally speaking , the photometric redshift error increases with magnitude due to larger photometric uncertainty . although the galaxies are fit separately in several color and magnitude bins , the rms scatter for bluer galaxies is still much larger than the scatter for redder galaxies ( shown in figure [ color - err ] ) . this is because early - type ( redder ) galaxies have some significant spectral features ( e.g. , the 4000 break ) and have similar seds , while late - type ( bluer ) galaxies tend to have a featureless and relatively flat continuum . ongoing star formation , gas absorption / emission , and dust extinction will also complicate the seds of late - type galaxies , and different galaxies have different combinations of these effects . because different late - type galaxies have weaker features and larger variation in their seds than early - type galaxies , the errors in the photometric redshift estimates will be larger for late - type galaxies . we compare the photometric redshift technique we use to other techniques described in @xcite . they use nine different methods including template fits and empirical fits to estimate photometric redshifts for the sdss early data release . from their study , in general , the results using empirical fits have smaller @xmath74 than the ones using template fits , and the best algorithm with the smallest @xmath74 among the empirical fits is also the kd - tree method . however , the kd - tree method they use is a two - dimensional tree to cut their training set in a color - color plane , unlike what we use , which is a three - dimensional tree in a color - color - magnitude space . the addition of magnitudes provides a rough estimate of redshift which makes the separation of the training set more precise than using just colors . combining deeper photometry , a larger training set over a wide redshift range , and a better cutting method for the training set , the limiting magnitude of our photometric redshift result is @xmath96 magnitudes deeper than the sdss result , even with only four wide - band filters ( compared to five filters for sdss ) and a much higher redshift limit . however , qualitatively , the agreement between spectroscopic and photometric redshift is similar between the sdss and the rcs ; e.g. , the very low redshift galaxies tend to have over - estimated photometric redshifts . to test our photometric redshift results , we compare the redshift distributions of our data against the photometric redshift distributions from the canada - france deep fields ( cfdf)@xcite , and the spectroscopic redshift distributions for the goods / hdf - n field @xcite . figure [ zdist ] presents the comparison results . the solid lines indicate the rcs data . the dashed lines represent the cfdf data , and the dotted lines indicate the goods data . the agreement between the different samples is in general very good . the scatter for the goods data in the @xmath97 panels is due to the small area of the goods / hdf - n field ( @xmath13 160 arcmin@xmath4 ) . the agreement for the three data sets is good in the @xmath98 bin . in the @xmath99 bin , the rcs data is somewhat broader presumably due to the larger redshift uncertainty , and it also has a slightly lower average . these comparisons provide a high level of confidence that the rcs photometric redshift measurements , despite the relatively shallow photometry , are statistically robust and reliable to as faint as @xmath100 , whereas for @xmath101 , the photometric redshift catalog should be used with caution . all galaxies with @xmath102 are included in the photometric redshift catalog . the total number of galaxies is more than 1,300,000 . about 168,000 objects ( 13% of the total ) have photometric redshifts at @xmath103 or @xmath104 , which are beyond the redshift range of the training set . negative redshifts are unlikely to be physical , and because the extrapolation of the polynomial fit is very unstable , photometric redshifts @xmath105 can not be trusted . for these objects , we use a value of `` 99 '' to indicate this type of problematic photometric redshift in the catalog . figure [ photoz - ratio ] presents the fraction of good photometric redshifts vs. @xmath1 magnitude . the definition of `` good photometric redshift '' is @xmath106 , where @xmath107 is the computed photometric redshift error . patches 0351 , 1122 , and 1327 have much poorer success rates at the faint end compared to the other patches due to poor @xmath2 and @xmath3 data quality ( the limiting magnitudes for @xmath2 and @xmath3 are about 24 and 23.5 , respectively , compared to the average of 24.5 and 25.0 , see figure [ vb_lim ] ) . for galaxies with @xmath108 in patches 0226 , 1449 , 1616 , and 2153 , the fractions of good redshifts are higher than 90% . for the remaining four patches , the fractions are roughly higher than 90% for galaxies with @xmath109 . for brighter ( @xmath110 ) galaxies , the lower fractions are probably due to a lack of bright galaxies in our training set , improper star - galaxy separation due to saturated stars , saturation of bright galaxy images in one filter or more , and the lack of a sufficiently blue filter for galaxies at low redshift . for fainter galaxies , the fractions are also going down due to larger magnitude errors , especially for redder galaxies because they are even fainter in @xmath2 and @xmath3 magnitudes . according to this figure , an overall reliable completeness magnitude limit ( 90% to 98% ) for the full catalog is @xmath111 , while if the patches 0351 , 1122 , and 1327 are removed , the statistical completeness limit extends to @xmath112 . if we relax the photometric redshift uncertainty limit to @xmath113 , the completeness limit extends to @xmath114 . because ppp star - galaxy classification begins to become less robust at @xmath115 due to the low signal - to - noise ratio for faint objects @xcite , some faint galaxies are classified as stars . to test the effect and extent of misclassification , a `` fake '' photometric redshift catalog for point sources in each patch is also generated . if we force fit photometric redshifts to real stars , it would produce results that systematically do not have the same redshift distribution as the galaxies . figure [ zdist.gal_star ] demonstrates the photometric redshift distributions of ppp classified galaxies and stars in different @xmath116 magnitude bins . the upper panels show the galaxy fraction of all objects ( galaxies + stars ) , and the lower panels show the stellar fraction . for the @xmath117 , @xmath118 , and @xmath119 bins , the redshift distributions for ppp classified galaxies and stars are not similar . the `` fake '' photometric redshifts of stars tend to be separated into two bins at @xmath120 for the bluer disk stars , and @xmath78 for the redder faint halo stars . the peaks for the three brighter bins do not change with magnitude , unlike those for the galaxies . but for the @xmath99 bin , the shape of the redshift distributions in the upper and lower panels are very similar , which implies that a significant fraction of the `` stars '' are actually galaxies . of course , we can not rule out that high - surface brightness , concentrated galaxies ( more likely to be misclassified as stars ) may have a different redshift distribution compared to low - surface brightness , less concentrated galaxies . table [ catalog ] is an example showing the information available for each object in the photometric redshift catalog ; the full photometric redshift catalog ( table 4 ) is available in the electronic version of the paper . listed in the sample catalog are the following rows : @xmath121the rcs object number . the first two digits indicate the patch number , numbered 01 to 10 in the order of the patches in table [ patch ] . the second two digits indicate the pointing number within a patch . the third two digits indicate the ccd chip number within the pointing . the remaining five digits represent the object number within each ccd chip . @xmath122the r.a . in hours and dec . in degrees ( j2000 ) . @xmath123the total magnitudes , magnitude errors , and color errors for @xmath0 , @xmath1 , @xmath2 , and @xmath3 . the magnitude errors are derived from the photon noise in the optimal aperture ( on total magnitude , see @xcite ) for each object . the color errors are the sum in quadrature of the photon errors for each filter in the color aperture . @xmath124the photometric redshift . @xmath125the computed error on the photometric redshift . cccccc ra & 2.435250 & 2.428926 & 2.432872 & 2.429133 & 2.432602 + dec & 0.90261 & 0.90425 & 0.90534 & 0.90611 & 0.90704 + @xmath0 & 20.81 & 18.70 & 20.04 & 19.27 & 19.43 + + /- c of @xmath0 & 0.05 0.04 & 0.01 0.01 & 0.03 0.02 & 0.02 0.01 & 0.01 0.01 + @xmath1 & 21.00 & 19.36 & 20.58 & 19.98 & 20.04 + + /- c of @xmath1 & 0.02 0.01 & 0.01 0.00 & 0.01 0.01 & 0.01 0.01 & 0.01 0.01 + @xmath2 & 21.45 & 20.67 & 21.66 & 21.28 & 21.30 + + /- c of @xmath2 & 0.04 0.02 & 0.01 0.01 & 0.03 0.03 & 0.03 0.02 & 0.02 0.02 + @xmath3 & 22.23 & 22.06 & 22.79 & 22.58 & 22.67 + + /- c of @xmath3 & 0.03 0.03 & 0.03 0.03 & 0.05 0.05 & 0.05 0.04 & 0.04 0.04 + photometric redshift & 0.263 & 0.436 & 0.393 & 0.442 & 0.395 + redshift error & 0.072 & 0.031 & 0.071 & 0.031 & 0.018 + we present a photometric redshift catalog from the rcs data , constructed by applying an empirical polynomial fitting technique using a training set of 4924 objects with redshifts from the cnoc2 and goods / hdf - n samples . a 32-cell kd - tree algorithm is used to divide up our sample to improve the accuracy of the photometric redshift estimates . our catalog includes 1.3 million galaxies in 33.6 deg@xmath4 of sky ( distributed over 10 patches ) with redshifts less than 1.5 . the rms photometric redshift scatter is @xmath7 within the redshift range @xmath8 , and @xmath9 for galaxies with @xmath10 . the computed photometric redshift errors for individual galaxies are also provided . the magnitude limit for completeness of the catalog is @xmath126 ( with @xmath106 ) if the three shallow patches ( 0351 , 1122 , and 1327 ) are excluded . the limit extends to @xmath127 for @xmath113 . we also compare the redshift distribution from our catalog to the cfdf and the goods / hdf - n data , and the agreement between the different samples is in general very good . we are carrying out a number of scientific studies using the photometric redshift catalog . one important example is to provide a large sample of galaxies over a large look - back time to measure the luminosity function and its evolution for field galaxies ( lin et al . 2005 , in preparation ) . moreover , the luminosity function for galaxies in clusters discovered by the rcs can be determined more accurately using photometric redshifts than by using two filters ( @xmath1 and @xmath0 ) . furthermore , we plan to develop cluster finding algorithms using the photometric redshift catalog , creating cluster catalogs for comparison with the result from the rcs technique @xcite . this will allow us to verify the completeness of the two - filter cluster red - sequence finding algorithm . another topic is the search for close galaxy pairs , both in the field and in galaxy clusters , to study the evolution of the merging rate ( hsieh et al . 2005 , in preparation ) . such a study will provide data to test the hierarchical model of galaxy evolution . we are also using the catalog to find sub - structures ( groups ) in clusters , and to study the cluster population as a function of radius . finally , the photometric redshift catalog provides accurate redshift distributions of lens galaxies and source galaxies , which are both very important for weak lensing analysis . with photometric redshift information , many weak lensing analyses , from galaxy - galaxy lensing to cosmic shear , can be improved . for example , knowing the redshifts of the lenses allows one to derive the mass - to - light ratio of galaxies as a function of galaxy luminosity ( hoekstra et al . 2005 , in preparation ) , testing the scaling relations of bayonic and dark matter . we wish to thank the staff at cfht for the observations and reduction of part of the data under the queue system . the research of h. y. is supported by grants from the national science and engineering research council of canada and the university of toronto . and @xmath3 limiting magnitudes . the distributions of point source magnitudes from individual chips for the @xmath2 filter and @xmath3 filter are shown in the left panel and the right panel , respectively . the thick lines indicate the median value of the limiting magnitudes for @xmath2 and @xmath3 , which are 24.54 and 24.99 , respectively . note that a number of pointings have very shallow @xmath3 limits due to observations done in poor seeing . [ vb_lim ] , scaledwidth=90.0% ] is used as the reference image . the central image is the original ( before matching ) image in @xmath2 . the right image is the shifted / rotated image in @xmath2 made by matching to the @xmath1 image . note that the upper part and left part of the right image are copied from the nearby chips to keep the completeness of the @xmath2 image , and the gaps are not filled with any counts here for clarity of presentation . the bad columns in the original @xmath2 image are fixed for the shifted / rotated @xmath2 image by using a linear interpolation algorithm.[matching - image ] , scaledwidth=90.0% ] vs. @xmath34 diagram . the right panel is the @xmath32 vs. @xmath34 diagram . the white dots and black dots indicate extended sources with @xmath68 in 0226a3 and 0226c1 , respectively . the distributions of galaxies in the two pointings in the color - color diagrams do not match , especially for @xmath32 vs. @xmath34 . it is evident that the original photometric calibrations are not sufficiently accurate , and some additional recalibrations have to be performed . [ color - color ] , scaledwidth=90.0% ] , @xmath32 , and @xmath36 , respectively . the solid line indicates the data from 0226a3 and the dashed line indicates the data from 0226c1 . only stars with @xmath37 are used . all histograms are normalized . in the three panels the color histograms for the two pointings are similar in shape but are offset from each other . by assuming that the photometric calibrations for @xmath1 are all correct and by choosing the pointing 0226a3 as the reference pointing , the magnitudes of @xmath3 and @xmath2 should be brighter and the @xmath0 magnitude should be a little fainter for the 0226c1 pointing . shifts of @xmath128 , @xmath129 , and @xmath130 mag need to be applied to the @xmath3 , @xmath2 , and @xmath0 magnitudes , respectively , of the 0226c1 pointing in order to make the color histograms match . [ color - histo ] , scaledwidth=90.0% ] mag between rcs and sdss in @xmath3 , @xmath2 , and @xmath0 ( left to right ) for overlapping patches ; see text for details . we arbitrarily set @xmath40mag @xmath131 for patch 0226 _ after _ recalibration . note that we do not calibrate the rcs photometry to the sdss photometry . the open boxes and filled boxes indicate the elixir - calibrated results and the recalibrated results , respectively . the recalibrated results show less scatter than the elixir - calibrated ones . [ sdss - offset ] , scaledwidth=100.0% ] mag between rcs and sdss for the set of pointings within each rcs patch ; see text for details . @xmath3 , @xmath2 , and @xmath0 are shown ( left to right ) . the open boxes and filled boxes indicate the elixir - calibrated results and the recalibrated results , respectively . the inter - pointing recalibrations for each patch reduce the standard deviations for @xmath2 and @xmath3 significantly . [ sdss - sd ] , scaledwidth=100.0% ] ) vs. magnitude ( @xmath1 ) diagram showing the layout of the regions used for the data separation . the dots indicate the rcs+cnoc2 data and the crosses indicate the goods / hdf - n data in our training set . the galaxies are roughly in two loci on the color - magnitude plane with a gap around @xmath33=1.8 to 2.0 . the early - type galaxies dominate the redder ( upper ) locus while the late - type galaxies dominate the bluer ( lower ) locus . we separate the upper locus into four regions ( 1 , 2 , 3 , and 4 ) and the lower locus into six regions ( 5 , 6 , 7 , 8 , 9 , and 10 ) for the photometric redshift fits . note that this cutting method is not used for the final catalog . [ color - mag],scaledwidth=90.0% ] vs. spectroscopic redshift diagram ( left ) , and median(@xmath86 ) or @xmath74 vs. spectroscopic redshift diagram ( right ) before noise is added . the fitting uses the kd - tree 32-cell cutting of the training set . the dots indicate the rcs / cnoc2 data , the crosses indicate the goods / hdf - n data , and the solid curve in the left panel is the median value of @xmath86 . the same solid curve is also shown in the right panel for clearer presentation . the dashed curve in the right panel is the @xmath74 which indicates the 68th percentile difference between the photometric redshift and the spectroscopic redshift , in bins of 100 objects each along the spectroscopic redshift axis . the @xmath74 for objects at our median redshift is less than 0.05 . the error becomes larger ( @xmath132 ) for objects at @xmath133 . the extremely large error for objects at @xmath79 is due to the large systematic deviations at those high redshifts . [ zerr ] , scaledwidth=90.0% ] but noise has been added to the goods / hdf - n sample ( see text ) . the @xmath74 for objects at our median redshift is similar compared to that in figure [ zerr ] , but the error increases to around 0.2 for objects at @xmath87 , which is much larger than the value from figure [ zerr ] . [ zerr1],scaledwidth=90.0% ] to the total number of objects for each redshift bin . the result is calculated using the rcs / cnoc2 and noise added goods / hdf - n data . note that the histogram has variable redshift bin widths because we choose widths that always include exactly 100 objects per bin . the catastrophic error rate is below 0.03 for @xmath81 . it becomes around 0.05 for @xmath82 due to larger @xmath74 . for redshifts higher than 1.2 , the catastrophic error rate is 0.25 , which is primarily due to the larger systematic error . [ catastrophic_err],scaledwidth=90.0% ] bins . we plot a dot for every 15 and 20 objects for the @xmath90 and @xmath91 bins , respectively . for the panels with @xmath92 and @xmath95 bins , we plot a dot for every 30 and 50 objects , respectively . the curve in each panel is the median value for all data points in each @xmath1 bin , not just for the data shown in the figure . note that we use different scales for the computed error axes in each panel . the error grows with fainter @xmath1 and higher redshift . for the objects at @xmath134 , the errors also become larger because we do not have a sufficiently blue filter for low redshift galaxies . generally speaking , the photometric redshift error increases with magnitude due to larger photometric uncertainty . [ esterr_photoz],scaledwidth=90.0% ] vs. color ( @xmath33 ) diagram for rcs data in our training set . @xmath74 is 0.03 for redder objects ( @xmath135 ) . the width of the error distribution becomes 0.043 for objects with @xmath136 . for objects with @xmath137 , this becomes larger than 0.055 . [ color - err],scaledwidth=90.0% ] , defined as @xmath106 , where @xmath107 is the computed photometric redshift error . each line indicates one patch . patches 0351 , 1122 , and 1327 have much lower fractions at the faint end due to the poorer @xmath2 and @xmath3 data quality ( see text and figure [ vb_lim ] ) . for patches 0226 , 1449 , 1616 , and 2153 , the fractions are higher than 90% for @xmath108 . for the remaining three patches , the fractions are higher than 90% for @xmath109 . the completeness levels increase to @xmath127 mag when the criterion @xmath113 is used . [ photoz - ratio],scaledwidth=90.0% ] bins . the upper four panels represent the galaxy fraction among all objects ( galaxies + stars ) . the lower four panels show the `` stars '' fraction among all objects . only for the @xmath99 bin is the redshift distributions very similar between the upper and lower panels . it implies that many `` stars '' in this magnitude range are actually galaxies . [ zdist.gal_star],scaledwidth=90.0% ]

ricci flow is now a standard tool in geometry and topology , and one of its main uses is to provide a means of deforming a given riemannian metric , hopefully into one that is `` distinguished '' in some sense . on a fixed riemannian manifold , ricci flow may be viewed as a dynamical system on the space of riemannian metrics , modulo diffeomorphisms . in this context , distinguished metrics are einstein or ricci soliton metrics , and we phrase the problem of `` deforming into a distinguished metric '' in terms of dynamical stability . namely , given a stationary solution @xmath0 of ricci flow and some topology on the space of metrics , does there exist a neighborhood @xmath1 of @xmath0 such that all ricci flow solutions with initial data in @xmath1 converge to @xmath0 ? many authors have considered stability for solutions of ricci flow ; see the introduction of @xcite for a description of some of the results and techniques used . the approach to stability that we take in this paper follows the program initiated by guenther , isenberg , and knopf in @xcite , which , roughly speaking , has three steps : 1 . modify ricci flow ( e.g. , by diffeomorphisms and rescaling ) so that it has a suitable class of fixed points . 2 . compute the linearization at a fixed point and prove linear stability . 3 . set up a sequence of tensor spaces with certain interpolation properties and apply a theorem of simonett to get dynamical stability . since its introduction in @xcite , this program has been used to prove a variety of other stability results @xcite . for step 1 , the flow that we wish to study appeared in @xcite . given a manifold @xmath2 , @xmath3 , and a vector field @xmath4 , the _ curvature - normalized ricci flow _ is @xmath5 a ricci soliton @xmath6 , which satisfies @xmath7 is clearly a stationary solution of . ( recall that an einstein metric corresponds to a ricci soliton where the field @xmath4 is a killing field . ) for step 2 , the linearization of the flow was also considered in @xcite . after modification by deturck diffeomorphisms , the linearization is @xmath8 where @xmath9 is the lichnerowicz laplacian acting on symmetric 2-tensors . a stationary solution of is _ strictly _ ( resp . _ weakly _ ) _ linearly stable _ if the operator @xmath10 has negative ( resp . non - positive ) spectrum , that is , if there exists some @xmath11 ( resp . @xmath12 ) such that @xmath13 for all symmetric 2-tensors @xmath14 taken from some appropriate tensor space . we emphasize that the notion of linear stability for a given metric depends on a particular linear operator . for example , koiso studied stability of einstein metrics with respect to the second variation of the total scalar curvature functional @xcite . other authors have studied stability of shrinking ( @xmath15 ) , gradient ( @xmath16 for some @xmath17 ) ricci solitons with respect to the second variation of perelman s @xmath18-functional ; see , for example @xcite . in this paper , we will not be directly concerned with step 3 and the aspects of dynamical stability . indeed , our goal is to study the _ linear stability _ of homogeneous einstein and ricci soliton metrics on non - compact manifolds , with respect to the operator from . ( step 3 is , however , the primary focus of a companion paper by wu and the third author @xcite , see theorem [ thm : dynamical - all ] below . ) the first results for such metrics are found in @xcite , where the authors give a detailed analysis of homogeneous solitons on three spaces : @xmath19 , @xmath20 , and @xmath21 . since that work , much has been discovered about homogeneous solitons . it turns out that the only non - trivial homogeneous ricci solitons occur in the non - compact , expanding ( @xmath22 ) , and non - gradient ( @xmath23 for any @xmath17 ) case ; see section 2 of @xcite for a discussion of this fact . additionally , all known examples can be realized as left - invariant metrics on simply connected solvable lie groups ( or , equivalently , as inner products on solvable lie algebras well use both interpretations ) , such that @xmath24 where @xmath25 is the ricci endomorphism , @xmath26 , and @xmath27 is a derivation of the lie algebra . a metric satisfying this equation is called an _ algebraic soliton _ ; on nilpotent and solvable lie groups , such metrics are called _ nilsolitons _ and _ solvsolitons _ , respectively . an algebraic soliton is in fact a ricci soliton ( see , e.g. , @xcite ) ; conversely , a ricci soliton on a solvable lie group is isometric to a solvsoliton ( on a possibly different lie group ) . also , a solvable lie group admitting a non - flat algebraic soliton is diffeomorphic to @xmath28 . these last two facts appear in @xcite . lauret has also proven strong structural results for algebraic solitons @xcite . essentially , nilsolitons are the nilpotent parts of solvsolitons a solvsoliton implies the existence of a nilsoliton on the nilradical , and conversely , any nilsoliton can be extended in a specific way to a solvsoliton . see theorem [ thm : lauret ] for more details . this new knowledge allows us to expand upon the linear stability work done in @xcite . here is an outline of the paper and a description of the results . in section [ sec : lin - criteria ] , we analyze the linearized ricci flow operator from and derive a sufficient condition for linear stability of soliton metrics ( including einstein metrics ) under the flow . this condition covers a somewhat general class of metrics , but our main focus is the algebraic soliton case , where @xmath29 . [ prop : stab - conds ] let @xmath6 be a ricci soliton with constant scalar curvature . the following condition implies that @xmath30 is strictly linearly stable : @xmath31 for all symmetric 2-tensors @xmath14 . when @xmath30 is @xmath32-einstein , this condition is @xmath33 here , @xmath34 is the action of the riemann curvature tensor on symmetric 2-tensors . we also have criteria based on sectional curvature bounds . [ prop : sec - stab ] let @xmath6 be a ricci soliton with constant scalar curvature , and suppose that the sectional curvature satisfies @xmath35 for some constant @xmath36 . a sufficient condition for @xmath30 to be strictly linearly stable is that @xmath37 we emphasize that these results only give _ sufficient _ conditions for linear stability , as they do not involve the full operator from . the hope is that , in the case of solvable lie groups , one can use algebraic methods to verify these conditions relatively simply . it turns out that this hope is too optimistic , as there are examples of metrics both einstein and non - einstein for which these propositions fail to give stability ; see subsection [ subsec : abelian - einstein ] . however , as we will see in theorem [ thm : stab - all ] , these propositions succeed in many other cases . in section [ sec : extensions ] , we explore the relationship between linear stability of a solvable einstein metric and its associated nilsoliton in the case of codimension - one nilradical . using proposition [ prop : stab - conds ] , we show that stability of the einstein metric implies stability of the nilsoliton when the eigenvalues of the nilsoliton derivation are not too large , relative to the trace of the derivation . [ prop : ext - stab ] let @xmath38 be a solvable lie algebra with codimension - one nilradical @xmath39 $ ] . if a @xmath32-einstein metric on @xmath38 satisfies , then a sufficient condition for the nilsoliton on @xmath40 to be strictly linearly stable is that @xmath41 where @xmath42 are the eigenvalues of the nilsoliton derivation @xmath27 . also in that section we use the continuity of the curvature operators to show that any non - nilpotent solvsoliton satisfying is contained in an open set of stable solvsolitons . this set is open in the moduli space of solvsoliton extensions of the nilradical of the original stable metric . in section [ sec : abelian ] , we consider abelian nilsolitons ( which are trivially stable ) and their solvsoliton extensions . we calculate the sectional curvatures and use proposition [ prop : sec - stab ] to find an open set ( in an appropriate moduli space ) of stable solvsolitons whose nilradicals have codimension one ; see proposition [ prop : abelian - sec ] . in section [ sec:2step ] , we use proposition [ prop : stab - conds ] to derive conditions for stability of two - step nilsolitons . using structural results due to eberlein @xcite , and analyzing the possible ricci tensors of nilsoliton metrics , we prove that all two - step nilsolitons are stable ; see proposition [ prop:2step - stab ] . the space of all such solitons is quite large ; in particular , there are continuous families @xcite . in section [ sec:3step ] , we consider linear stability of solvable einstein metrics whose nilradicals are two - step and codimension - one . while the techniques of section [ sec:2step ] do not work in all cases here , we are able to obtain stability for two countably infinite families of metrics . first are those where the nilradical is a generalized heisenberg group , and second are where the nilradical is a free two - step nilpotent lie group ; see proposition [ prop : stab - solv - heis - free ] . we also briefly discuss spaces of negative curvature and proposition [ prop : sec - stab ] . in particular , we use results of heber @xcite to find high - dimensional families of stable einstein metrics that arise as deformations of certain quaternionic and cayley hyperbolic spaces . in section [ sec : low - dim ] , we consider `` low - dimensional '' examples . namely , we prove linear stability of all nilsolitons in dimension six or less , using the classification found in @xcite . we also prove linear stability for the corresponding one - dimensional solvable einstein extensions , except in the case of the hyperbolic space @xmath43 , which is only weakly linearly stable . this gives almost 100 examples of stable metrics on solvable lie groups . see proposition [ prop : low - dim ] and tables [ table : solitons ] and [ table : solitons2 ] . we show that each metric in a curve of seven - dimensional nilsolitons is strictly linearly stable , and also their one - dimensional einstein extensions . finally , we show that each solvsoliton extension of the three - dimensional heisenberg algebra is strictly linearly stable . here is a summary of the spaces shown to be stable . [ thm : stab - all ] the following algebraic solitons are strictly linearly stable with respect to the curvature - normalized ricci flow : 1 . [ item1 ] every nilsoliton of dimension six or less , and every member of a certain one - parameter family of seven - dimensional nilsolitons ; 2 . every abelian or two - step nilsoliton ; 3 . every four - dimensional solvsoliton whose nilradical is the three - dimensional heisenberg algebra ; 4 . an open set of solvsolitons whose nilradicals are codimension - one and abelian ; 5 . every solvable einstein metric whose nilradical is codimension - one and 1 . found in [ item1 ] and has dimension greater than one , or 2 . a generalized heisenberg algebra , or 3 . a free two - step nilpotent algebra ; 6 . for each @xmath44 , an @xmath45-dimensional family of negatively - curved einstein metrics containing the quaternionic hyperbolic space @xmath46 ; 7 an @xmath47-dimensional family of negatively - curved einstein metrics containing the cayley hyperbolic plane @xmath48 . further , any non - nilpotent solvsoliton @xmath49 satisfying and @xmath50 is contained in an open set of stable solvsolitons . this list is , of course , not complete . one might be tempted to conjecture that all solvsolitons are strictly linearly stable , but this is not clear . as mentioned above , in subsection [ subsec : abelian - einstein ] we give examples of metrics for which both propositions [ prop : stab - conds ] and [ prop : sec - stab ] fail to give stability . these metrics are possibly stable , but our methods are insufficient to prove it . as the focus of this paper is linear stability , we will not directly consider the dynamical aspects of stability . however , using the results of @xcite , we have dynamical stability for each metric in theorem [ thm : stab - all ] . [ thm : dynamical - all ] each metric from theorem [ thm : stab - all ] is dynamically stable in the context of ( * ? ? ? * theorem 1.1 ) for einstein metrics and of ( * ? ? ? * theorem 1.2 ) for non - einstein solitons . this theorem adds to the understanding of long - time behavior of ricci flow solutions . indeed , expanding homogeneous solitons are expected to act as singularity models for ( at least some ) type - iii ricci flow singularities , and this is known to be true in some cases @xcite . theorem [ thm : dynamical - all ] exhibits many more examples of ricci flow solutions that converge to expanding homogeneous solitons . in this section we find conditions that imply linear stability , and doing so requires us to understand the linear operator @xmath10 from . in particular , we analyze the various components that make up the @xmath51 inner product @xmath52 , which include various curvature operators , laplace operators , and lie derivatives . the main results of this section , propositions [ prop : stab - conds ] and [ prop : sec - stab ] , give sufficient conditions for linear stability in the case of constant scalar curvature . the first requires estimating the eigenvalues of @xmath53 , acting on symmetric 2-tensors , and the second requires sectional curvature bounds . the results are also adapted to the case of algebraic solitons . let @xmath54 be a riemannian manifold . the lichnerowicz laplacian , acting on symmetric 2-tensors , has the form @xmath55 here , @xmath56 is the connection laplacian on 2-tensors . the second term is the action of the riemann tensor on symmetric 2-tensors . in local coordinates , this is @xmath57 this operator satisfies @xmath58 , so @xmath59 . when @xmath30 is @xmath32-einstein , @xmath60 , so @xmath32 is an eigenvalue of @xmath34 . additionally , when @xmath30 is @xmath32-einstein , we have @xmath61 , so that it is actually an operator on trace - free symmetric @xmath62-tensors . we will not need this fact , however . the last two terms in are an abuse of notation for the action of the ricci tensor on symmetric 2-tensors . in local coordinates , this is @xmath63 we can combine these two curvature operators to form @xmath64 , the _ weitzenbck curvature operator _ , acting on symmetric 2-tensors - tensors by @xmath65 in an orthonormal frame @xmath66 . here , the action of the curvature tensor @xmath67 on a tensor @xmath68 is @xmath69 one gets a lichnerowicz laplacian on @xmath70 tensors by @xmath71 . ] : @xmath72 note that when @xmath30 is @xmath32-einstein , this becomes @xmath73 . this operator allows us to write as @xmath74 an important formula involving the lichnerowicz laplacian is koiso s bochner formula @xcite , which we adapt here in a slightly more general setting . [ prop : koiso ] on a riemannian manifold @xmath54 , let @xmath14 be a symmetric 2-tensor with compact support and define a 3-tensor @xmath68 by @xmath75 then @xmath76 we sketch the proof . first , it is easy to see that @xmath77 integrate by parts and commute covariant derivatives to obtain @xmath78 now follows from combining these two equations . we saw `` the '' lichnerowicz laplacian above , @xmath79 but there is actually a family of laplacians depending on a parameter @xmath80 , @xmath81 as a special case , let us consider the _ codazzi laplacian _ : @xmath82 the name arises from the fact that the kernel of this operator , acting on symmetric 2-tensors , consists of codazzi tensors . by , we have @xmath83 so this operator is non - positive . consider a ricci soliton @xmath6 satisfying . using the expression of the lie derivative in terms of covariant derivatives , we have @xmath84 where @xmath85 this gives an equivalent soliton equation on the endomorphism level : @xmath86 express the lie derivative of a compactly supported symmetric 2-tensor @xmath14 as @xmath87 writing @xmath88 and integrating by parts gives @xmath89 if @xmath90 is constant ( which happens when @xmath30 has constant scalar curvature ) , this becomes @xmath91 next , consider @xmath92 and @xmath14 on the endomorphism level , so that @xmath93 using this and the endomorphism inner product @xmath94 , we have @xmath95 now , combining this with gives @xmath96 we now use the various formulas above to rewrite the linearized operator @xmath10 , and describe some sufficient conditions for linear stability . define the quadratic form @xmath97 let @xmath6 be a ricci soliton with constant scalar curvature . the following condition implies that @xmath30 is strictly linearly stable : @xmath98 when @xmath30 is @xmath32-einstein , this condition is @xmath99 the linearization is @xmath100 using and , we have @xmath101 the curvature terms are exactly @xmath102 , and since we want @xmath103 , the result follows . in the einstein case , the condition for linear stability is the same as the stability condition considered by koiso and others ; see @xcite or @xcite . recall that a homogeneous metric has constant scalar curvature . furthermore , in the case of an algebraic soliton on a simply connected solvable lie group @xmath104 with lie algebra @xmath38 , we have @xmath105 , where @xmath106 . tracing this and comparing with the traced version of yields @xmath29 . we also note that in the homogeneous case , we only need pointwise estimates . therefore , we do nt need the ( integrated ) @xmath51 inner product , only a single inner product , so it suffices to consider @xmath107 . [ cor : alg - stab - cond ] let @xmath108 be an algebraic ricci soliton . the following condition implies that @xmath30 is strictly linearly stable : @xmath109 in estimating @xmath10 , one may not want to use the codazzi laplacian , in which case we use to get @xmath110 this gives another condition for stability , @xmath111 . it has the advantage of only requiring estimates of the riemann curvature , but the estimates must be better than those for @xmath112 . we also describe conditions for stability when the sectional curvature is bounded . let @xmath6 be a ricci soliton with constant scalar curvature , and suppose that @xmath35 for some constant @xmath36 . a sufficient condition for @xmath30 to be strictly linearly stable is that @xmath113 or @xmath114 in the case of an algebraic soliton . select an orthonormal basis that diagonalizes @xmath14 and consider the ( unintegrated ) operator @xmath102 . @xmath115 now , we use above . @xmath116 we want this to be less than zero , so the result follows . if we take @xmath117 , then the condition reduces to @xmath118 or @xmath119 . also , we can allow for @xmath120 by using the inequality @xmath121 . the condition then becomes @xmath122 but it is not clear how useful this is . simply connected solvable lie groups provide a rich class of non - compact spaces that often admit ricci solitons and einstein metrics . indeed , as mentioned above , all known examples of non - compact homogeneous expanding ricci solitons can be realized as left - invariant metrics on solvable lie groups . building on the work of heber @xcite , lauret has exploited the algebraic structure of these spaces to prove many powerful results regarding these metrics . before we state these results , recall that , given a lie algebra @xmath40 with @xmath123 , the semi - direct product @xmath124 has the bracket structure @xmath125 = 0 , \qquad [ a , b ] = 0 , \qquad [ a , x ] = a(x),\ ] ] for all @xmath126 , @xmath127 . also recall that the _ mean curvature vector _ of a metric lie algebra @xmath124 is @xmath128 such that @xmath129 for all @xmath130 . [ thm : lauret ] * * let @xmath131 be a nilsoliton with ricci operator @xmath132 , @xmath22 and @xmath133 . consider an abelian subalgebra @xmath134 of @xmath135-symmetric derivations of @xmath40 . then the simply connected solvmanifold @xmath104 with lie algebra @xmath136 has an inner product @xmath137 that is a solvsoliton with @xmath105 . this inner product satisfies @xmath138 and the derivation @xmath106 satisfies @xmath139 where @xmath140 is the mean curvature vector of @xmath141 . furthermore , the inner product is einstein if and only if @xmath142 , or equivalently , @xmath143 . * all solvsolitons are of the form described in ( a ) , with @xmath144 $ ] . * let @xmath104 and @xmath145 be two solvsolitons which are isomorphic as lie groups . then @xmath104 is isometric to @xmath145 up to scaling . this essentially says that nilsolitons are the nilpotent parts of solvsolitons . in particular , given a nilsoliton metric on a nilpotent lie algebra @xmath40 , it is always possible to construct a one - dimensional extension of @xmath40 that admits an einstein metric . we can form the solvable lie algebra @xmath146 , where @xmath27 is the nilsoliton derivation , and its einstein metric is the one described in part ( a ) of the theorem . our goal in this section is to relate the linear stability of a left - invariant @xmath32-einstein metric on a solvable lie group @xmath104 with that of the corresponding nilsoliton in the case of a codimension - one nilradical @xmath147 . namely , we suppose that the einstein metric satisfies the condition of proposition [ prop : stab - conds ] : @xmath148 . then we then use this to estimate the spectrum of the operator @xmath112 on the nilradical . this comes down to comparing the eigenvalues of the soliton derivation to its trace . note again that all estimates are pointwise as the fixed - point metrics are homogeneous . also note that the einstein metric will be a fixed point of the flow @xmath149 for metrics on @xmath104 , and the nilsoliton is a fixed point of the flow for metrics on @xmath147 : @xmath150 in terms of notation , geometric quantities on @xmath38 will not have subscripts , but those on @xmath40 or @xmath134 will . suppose that @xmath151 and @xmath152 , and let @xmath153 be an orthonormal basis of @xmath49 and let @xmath154 , @xmath155 , and @xmath156 . in this section , we will use roman indices for elements of @xmath40 , greek indices for elements of @xmath134 . recall that the covariant derivative corresponding to a left - invariant metric is given by @xmath157 where @xmath158 , see @xcite . write @xmath159 for the orthogonal projection of @xmath160 onto @xmath134 . using that @xmath161 , @xmath162 , and @xmath163 , it is easy to see that @xmath164 the riemann curvature tensor satisfies @xmath165 } v , w \rangle.\ ] ] using this and , @xmath166,z),a \rangle \\ r(x , y , z , w ) & = r_{\ensuremath \mathfrak{n}}(x , y , z , w ) \\ & \quad + \langle { \ensuremath \mathrm{pr}}_{\ensuremath \mathfrak{a } } \sigma(x , z),{\ensuremath \mathrm{pr}}_{\ensuremath \mathfrak{a } } \sigma(y , w ) \rangle - \langle { \ensuremath \mathrm{pr}}_{\ensuremath \mathfrak{a } } \sigma(y , z),{\ensuremath \mathrm{pr}}_{\ensuremath \mathfrak{a } } \sigma(x , w ) \rangle . \end{aligned}\ ] ] the ricci curvature ( see also ( * ? ? ? * lemma 1.4 ) or ( * ? ? ? * lemma 4.4 ) ) is @xmath167 since we re ultimately interested in the operators @xmath34 and @xmath112 , we calculate @xmath168 and @xmath169 . the idea is to exploit the product structure of the algebra . first , note that a symmetric 2-tensor has the form @xmath170 and @xmath171 we use capital roman letters to refer to indices ranging over both @xmath172 ( from @xmath40 ) and @xmath173 ( from @xmath134 ) , that is , @xmath174 . then summing over repeated indices , @xmath175 and @xmath176 these types of expansions will also appear later in the paper . let @xmath177 be @xmath32-einstein and suppose that @xmath178 for all symmetric 2-tensors @xmath179 on @xmath38 . by corollary [ cor : alg - stab - cond ] , the metric is strictly linearly stable . suppose further that @xmath38 has a codimension - one nilradical @xmath40 , which admits a nilsoliton by theorem [ thm : lauret ] . we want to show that @xmath180 , where @xmath14 now lives on @xmath40 . such an @xmath14 can be thought of as a certain type of tensor on @xmath38 , where @xmath181 for all @xmath182 , @xmath126 . this will allow us to use the estimate for @xmath168 . we begin with an observation about the quantity @xmath159 . [ lem:1-dim - d ] let @xmath183 be a nilsoliton with @xmath105 . consider the 1-dimensional einstein extension @xmath184 , where @xmath185 . then we have @xmath186 let @xmath187 be a unit vector spanning @xmath134 , so that @xmath188 for some @xmath189 to be determined , and let @xmath127 . the decomposition of the covariant derivative from involves the quantity @xmath159 , which in this case is essentially the second fundamental form , @xmath190 , thought of as a real - valued bilinear map . we compute @xmath191 this means that on the endomorphism level , @xmath192 . to find @xmath193 , we use the form of the metric on @xmath134 as in theorem [ thm : lauret ] ( a ) : @xmath194 and since @xmath195 , we have @xmath196 . the desired formula follows . now , using together with and the lemma , it is not hard to compute @xmath197 similarly , from , , and the lemma we obtain @xmath198 now we can prove the main result of this section . let @xmath38 be a solvable lie algebra with codimension - one nilradical @xmath39 $ ] . if a @xmath32-einstein metric on @xmath38 satisfies @xmath199 for all symmetric 2-tensors @xmath14 , then a sufficient condition for the nilsoliton on @xmath40 to be strictly linearly stable is that @xmath41 where @xmath42 are the eigenvalues of the nilsoliton derivation @xmath27 . our assumption on @xmath137 says that @xmath200 , and we want to show that @xmath201 . using and it is easy to see that @xmath202 therefore , we need to show that @xmath203 select an orthonormal basis of @xmath40 such that @xmath27 is diagonalized , with ( positive ) diagonal entries @xmath204 . then we have @xmath205 the first inequality uses the arithmetic / geometric mean inequality . for the second inequality to hold , it is enough for the following two inequalities to hold : @xmath206 we rewrite as follows : @xmath207 this is equivalent to one the following statements @xmath208 or @xmath209 the second condition in the first statement is never satisfied , since @xmath210 for all @xmath211 , so we consider the second statement . the inequalities there can be rewritten @xmath212 and the second is always satisfied . therefore is equivalent to @xmath213 . now , notice that follows from : @xmath214 here we verify the last statement in theorem [ thm : stab - all ] by applying corollary [ cor : alg - stab - cond ] and observing that @xmath112 and @xmath27 vary continuously as we vary the lie structure of the underlying solvable lie group . in all of the following , we normalize our soliton metric by fixing the soliton constant @xmath32 appearing in the equation @xmath215 . consider a nilpotent lie algebra @xmath40 which admits a nilsoliton metric . as described in theorem [ thm : lauret ] , for any abelian subalgebra @xmath134 of symmetric derivations , the solvable lie algebra @xmath216 admits a solvsoliton . further , the solvsoliton derivation @xmath27 on @xmath38 satisfies @xmath217 where @xmath140 is the mean curvature of @xmath38 and @xmath218 is the soliton derivation of @xmath40 . the metric on @xmath38 is einstein precisely when @xmath219 . to see more precisely how a choice of abelian subalgebra of symmetric derivations effects the resulting solvsoliton , we recall a corollary of lauret s theorem ; see , e.g. , @xcite . let @xmath220 denote symmetric linear maps on @xmath40 , with respect to the nilsoliton metric . define the _ rank _ of the nilpotent lie algebra to be @xmath221 , where @xmath222 is a maximal abelian subalgebra . [ prop : sol - space ] let @xmath223 . the moduli space of @xmath224-dimensional solvsoliton extensions of @xmath40 , up to isometry and scaling , is parametrized by @xmath225 where @xmath226 is the weyl group of @xmath227 . note that when @xmath228 or @xmath229 the moduli space is a point consisting of the original nilsoliton in the former case , or a single einstein metric in the latter . now we consider changing @xmath230 , which amounts to moving within the grassmannian @xmath231 . the mean curvature vector @xmath128 , which satisfies @xmath129 for all @xmath232 , can be alternately described as the unique vector @xmath128 that solves @xmath233 from this and we see that the solvsoliton derivation @xmath27 of @xmath38 depends continuously on @xmath134 . we also observe that the geometric quantities @xmath34 and @xmath25 depend continuously on @xmath134 , since these quantities are completely determined by the structure constants of the lie algebra , which are determined by @xmath134 and @xmath40 ; see subsection [ subsec : solv - curv ] . hence , @xmath112 also depends continuously on @xmath134 . finally , suppose that the soliton on @xmath49 satisfies the condition of corollary [ cor : alg - stab - cond ] . as long as @xmath234 , the condition of corollary [ cor : alg - stab - cond ] is preserved under sufficiently small perturbations of @xmath134 . the result is an open set in @xmath231 of stable metrics . taking the quotient by the finite group @xmath226 , we have an open set of stable metrics in the moduli space of solvsoliton extensions of @xmath40 . consider an abelian nilsoliton @xmath40 , which we identify with @xmath235 through an orthonormal basis @xmath236 . since the metric is flat , @xmath237 and we may choose any @xmath22 as the soliton constant . if we pick @xmath238 , then the soliton derivation is @xmath239 . strict linear stability of these solitons follows from corollary [ cor : alg - stab - cond ] , since @xmath240 when @xmath241 . solvable extensions of abelian nilpotent lie groups are called two - step solvable lie groups . for example , real hyperbolic space is obtained by adjoining to @xmath235 the 1-dimensional algebra spanned by the identity map ( which is a derivation ) . similarly , any other linear map yields a two - step solvable space admitting a solvsoliton . using [ prop : sec - stab ] , we construct an open set of strictly linearly stable two - step solvsolitons . as described in see ( * ? ? ? * section 3.1 ) , the algebra of derivations of @xmath40 is @xmath242 , and with respect to the chosen orthonormal basis , a maximal abelian subalgebra @xmath243 is the @xmath244-dimensional space of all diagonal matrices . a choice of a subalgebra @xmath245 defines a solvsoliton extension of @xmath40 that is unique up to the action of the symmetric group @xmath246 on @xmath247 ( which permutes diagonal entries ) . solvsoliton extensions of @xmath40 are therefore parametrized by the union of the spaces @xmath248 , where @xmath249 is the dimension of @xmath134 . when @xmath250 this space is @xmath251 , and when @xmath252 or @xmath253 this space is a point consisting of only the nilsoliton or an einstein metric , respectively . select a subalgebra @xmath245 of dimension @xmath254 , which we describe with an orthonormal basis @xmath255 consisting of diagonal matrices : @xmath256 ( as before , roman indices will refer to vectors in @xmath257 and greek to vectors in @xmath134 ) . by theorem [ thm : lauret ] , we get a solvsoliton metric on @xmath258 that makes @xmath235 and @xmath134 orthogonal , and on elements of @xmath134 it is given by @xmath259 in particular , we have @xmath260 the lie algebra structure on @xmath261 is the semi - direct product , so we have the bracket relations described in . in terms of structure constants , we have @xmath262 from this we can compute the transposes of the adjoint maps . for example , @xmath263 the mean curvature vector is @xmath128 such that @xmath129 for all @xmath130 . from this it is easy to compute that @xmath264 and so @xmath265 its adjoint satisfies @xmath266 this means @xmath267 is diagonal in any orthonormal basis , and the diagonal entries are @xmath268 or @xmath269 . we also have that @xmath270 we can say more about this quantity . [ lem : abelian - norm - h ] the mean curvature vector corresponding to a two - step solvsoliton satisfies @xmath271 , with equality if and only if the metric is einstein . think of elements of the orthonormal basis @xmath272 as column vectors and put them into a matrix @xmath187 ( which satisfies @xmath273 , by ) . next , complete this basis to an orthonormal basis of the @xmath244-dimensional maximal abelian subalgebra @xmath247 . adding these new columns to @xmath187 , we get a matrix @xmath274 . then @xmath275 and so @xmath276 summing over @xmath211 and @xmath277 we get @xmath278 this implies that @xmath271 . noting that the nilsoliton derivation @xmath279 is represented by the vector @xmath280 the above calculation also gives us the following equivalent statements . @xmath281 by theorem [ thm : lauret ] ( a ) , this happens if and only if the extension is einstein . we want to compute the curvatures of @xmath38 , so recall the formulas from section [ sec : extensions ] . since @xmath282 is abelian , we have that @xmath283 for all @xmath127 . also recall that @xmath284 , and that @xmath159 is the orthogonal projection of @xmath160 onto @xmath134 . using the above formulas , it is easy to see that @xmath285 now , using this and , we calculate the riemann curvatures . @xmath286 now , let @xmath1 and @xmath287 be a pair of orthonormal vectors in @xmath38 . we will compute the sectional curvature @xmath288 in a manner similar to , with similar cancellations . first , we introduce the short - hand notation @xmath289 now , sum over all repeated indices . @xmath290 we will use this to prove the main result of this section . [ prop : abelian - sec ] there is an open set of strictly linearly stable solvsolitons whose nilradicals have codimension one , are abelian , and have dimension two or greater . when @xmath250 , the above curvature formula reduces to @xmath291 where @xmath269 refers to the @xmath134-component of the vectors . clearly , if @xmath292 then the sectional curvature is non - positive . since @xmath293 satisfies @xmath294 , we may think of @xmath187 as an element of the unit sphere @xmath295 . the set of @xmath187 satisfying contains an open set on @xmath295 . as mentioned above , the moduli space of non - isometric solvsolitons is parametrized by @xmath251 . so , after taking the quotient of @xmath295 by several finite groups , we still have an open set of non - positively curved solvsolitons in the moduli space . from theorem [ thm : lauret ] , the soliton derivations satisfy @xmath296 , which is positive in the non - einstein case by lemma [ lem : abelian - norm - h ] . the single einstein metric has negative curvature ; see the next subsection . therefore , each metric in this set is stable by proposition [ prop : sec - stab ] . here we describe some properties of two - step solvable einstein metrics . of particular interest are examples of metrics for which both corollary [ cor : alg - stab - cond ] and proposition [ prop : sec - stab ] fail to give stability . let @xmath152 , with @xmath297 . the solvsoliton extension of @xmath235 is determined by @xmath134 , independent of any basis . in the einstein case , we can always pick a basis of @xmath134 such that @xmath298 ( normalized as in lemma [ lem:1-dim - d ] ) , and @xmath299 for @xmath300 . when @xmath250 , the extension is real hyperbolic space @xmath301 and the matrix we adjoin is @xmath302 . taking this a bit further , we see that every two - step solvable einstein metric is an extension of an abelian algebra by a hyperbolic space . indeed , @xmath303 is an ideal in @xmath261 , so we have a short exact sequence : @xmath304 when @xmath305 , the hyperbolic space @xmath301 has constant negative curvature , so it is stable by proposition [ prop : sec - stab ] . when @xmath306 , one can check that @xmath43 is only weakly linearly stable . compact quotients of this space are dynamically stable , however , and proving this requires more detailed analysis of the null - space of the operator in [ eq : cnrf - lin ] ; see ( * ? ? ? * lemma 5 ) . also , by ( * ? ? ? * theorem 3 ) , a metric with strictly negative curvature must have @xmath307 , so hyperbolic spaces are the only two - step solvable einstein metrics with strictly negative curvature . [ ex : hyp - prod ] when @xmath253 , it is convenient to represent the basis of @xmath134 as a matrix as in the proof of lemma [ lem : abelian - norm - h ] . in this case we take @xmath308 . it is not hard to see that the resulting @xmath309-dimensional space is isometric to the @xmath244-fold riemannian product @xmath310 this space is einstein with @xmath238 . one can compute the eigenvalues of @xmath34 and see that the largest is in fact @xmath311 , so we only have weak linear stability , as with @xmath43 . note that by ( * ? ? ? * proposition 2.1 ) , any two - step solvable einstein metric that has non - positive sectional curvature will be a product of @xmath254 real hyperbolic spaces . let us consider two more concrete metrics . let @xmath312 and @xmath313 , and consider the following matrix @xmath187 describing the basis for @xmath134 ( again as in the proof of lemma [ lem : abelian - norm - h ] ) . @xmath314 the resulting einstein metric does not satisfy the conditions of either corollary [ cor : alg - stab - cond ] or proposition [ prop : sec - stab ] . indeed , using one can see that @xmath315 one can also see easily that the sectional curvatures have mixed sign . we note that taking @xmath187 to be only the first two columns yields a soliton that is stable according to corollary [ cor : alg - stab - cond ] . thinking of the soliton above as a 1-dimensional extension of this one , we see that a converse of proposition [ prop : ext - stab ] can not always hold , at least in the non - nilpotent case . next , let @xmath316 this gives a non - einstein solvsoliton with @xmath238 and @xmath317 so that @xmath318 . again , the condition of corollary [ cor : alg - stab - cond ] is not satisfied , since @xmath319 it is also easy to see that the sectional curvature has mixed sign , so that proposition [ prop : sec - stab ] does not apply . taking any 1-dimensional extension of this example will yield an einstein metric as in example [ ex : hyp - prod ] , meaning it is possible to `` improve '' the spectrum of @xmath112 under extensions . in this section , we prove that all two - step nilsolitons are strictly linearly stable . the idea is to use corollary [ cor : alg - stab - cond ] and extra information from the two - step nilpotent structure . let @xmath40 be a two - step nilpotent lie algebra . suppose that @xmath320 $ ] has dimension @xmath70 , and let @xmath321 be a complementary subspace of dimension @xmath322 , so that @xmath323 and @xmath324 . the dimensions @xmath70 and @xmath322 must satisfy @xmath325 suppose that @xmath40 has an inner product @xmath137 such that @xmath326 . we obtain a homomorphism of @xmath327-algebras , @xmath328 , which is defined implicitly by the relation @xmath329 \rangle\ ] ] for all @xmath330 and @xmath331 . each map @xmath332 is therefore skew - symmetric . the map @xmath328 is called the _ j - map_. [ ex : heis - def ] we call a two - step nilpotent metric lie algebra @xmath40 a _ generalized heisenberg algebra _ if @xmath333 for all @xmath331 , or equivalently if @xmath334 for all @xmath335 . the unique simply connected lie group @xmath147 with lie algebra @xmath40 is then called a _ generalized heisenberg group_. these spaces were introduced by a. kaplan @xcite ; see also @xcite . if @xmath336 is the quadratic form on @xmath337 induced by the inner product ( that is , @xmath338 ) , then condition implies that @xmath339 extends to a representation of the clifford algebra @xmath340 on @xmath321 . it is not hard to see that the converse is true , namely , that every representation of a clifford algebra corresponds to a generalized heisenberg algebra . the classification of clifford algebras therefore provides a classification of generalized heisenberg algebras . [ ex : free-2step - def ] let @xmath341 be a semi - simple lie algebra , and let @xmath342 be a faithful representation . then we can give @xmath343 the structure of a two - step nilpotent metric lie algebra as follows . first , give @xmath344 an inner product such that each @xmath345 is skew - symmetric , give @xmath341 an @xmath346-invariant inner product , and declare that the summands @xmath344 and @xmath341 are orthogonal . now , the bracket structure is determined as in equation . a special case is the usual representation of @xmath347 on @xmath344 . we call the resulting algebra the _ free two - step nilpotent lie algebra on @xmath322 generators _ and write @xmath348 . this means @xmath349 . the ricci tensor of a two - step nilpotent metric lie algebra @xmath350 is _ optimal _ ( in the sense of @xcite ) if @xmath351 for some @xmath352 . it is easy to see that metrics with optimal ricci tensors are nilsolitons . in fact , using the remarks following propositions 7.3 and 7.4 in @xcite , the constant @xmath353 can be eliminated : @xmath354 and @xmath355 . the nilsoliton metrics on generalized heisenberg algebras and free two - step nilpotent algebras all have optimal ricci tensors ; see subsections [ subsec:3step - heis ] and [ subsec:3step - free ] for more information . it turns out that the property of having optimal ricci tensor is equivalent to being a nilsoliton with eigenvalue type @xmath356 , which means that the eigenvalues of the nilsoliton derivation are @xmath311 and @xmath62 ( perhaps after rescaling ) . this property is `` generic '' among two - step nilpotent metric lie algebras in a specific sense ; see ( * ? ? ? * proposition 7.9 ) or ( * ? ? ? * theorem 4 ) . we use the results of eberlein @xcite to describe the riemann curvature tensor on a two - step nilpotent lie group @xmath323 . let @xmath357 be an orthonormal basis for @xmath40 , with @xmath358 and @xmath359 . now we will use roman indices for elements of @xmath40 and greek indices for elements of @xmath337 . the ( non - zero ) riemann , sectional , and ricci curvatures are @xmath360,[w , w ' ] \rangle + { { \textstyle{\frac{1}{4 } } } } \langle [ u , w],[v , w ' ] \rangle -{{\textstyle{\frac{1}{4 } } } } \langle [ u , w'],[v , w ] \rangle \\ r(u , v , x , y ) & = { { \textstyle{\frac{1}{4 } } } } \langle j_x u , j_y v \rangle - { { \textstyle{\frac{1}{4 } } } } \langle j_x v , j_y u \rangle \\ r(u , x , v , y ) & = -{{\textstyle{\frac{1}{4 } } } } \langle j_x v , j_y u \rangle \end{aligned}\ ] ] @xmath361|^2 & & & \operatorname{rc}(u , v ) & = { { \textstyle{\frac{1}{2 } } } } \sum_\alpha \langle j_\alpha^2 u , v \rangle \\ \sec(u , x ) & = { { \textstyle{\frac{1}{4 } } } } |j_x u|^2 & & & \operatorname{rc}(x , y ) & = - { { \textstyle{\frac{1}{4 } } } } \operatorname{tr}(j_x \circ j_y)\end{aligned}\ ] ] for all @xmath362 and @xmath363 . in terms of structure constants , the ricci curvatures are @xmath364 and , recalling that @xmath25 is negative - definite on @xmath321 and positive - definite on @xmath337 , we define for @xmath365 and @xmath331 , @xmath366 [ lem:2step - ric - est ] when @xmath40 is a two - step nilsoliton of type @xmath367 , rescaled so that @xmath368 , we have @xmath369 be a simply - connected lie group with a left - invariant metric . as the group acts by isometries on itself , all of the geometry is encoded in a single tangent space with inner product . we take the lie algebra @xmath370 of @xmath371 to be our tangent space of interest and write it as @xmath372 , where @xmath373 denotes the lie bracket on the underlying vector space @xmath374 . we denote the inner product on this vector space by @xmath375 . ( we choose our identification of @xmath376 with @xmath374 so that the ` standard basis ' is orthonormal relative to our fixed inner product on @xmath376 . ) we study @xmath371 and @xmath376 by viewing the lie bracket @xmath373 as an element of the space @xmath377 of ( anti - symmetric ) algebra structures on @xmath374 . this perspective has proven very effective for working with nilsolitons , cf . @xcite . for the sake of consistency and ease of referencing , we hold closely to the notation used by lauret . take @xmath378 to be the standard basis of @xmath374 and denote by @xmath379 the element of @xmath380 which is dual to @xmath381 . then @xmath382 is a basis for @xmath287 . the element @xmath383 is the element such that , when evaluated on basis vectors of @xmath374 , we have @xmath384 and zero otherwise . we define an inner product on @xmath287 such that the above basis is orthonormal . recall that the vector space @xmath287 comes naturally equipped with an action @xmath385 of @xmath386 defined by @xmath387 where @xmath388 , @xmath389 , and @xmath390 . as we will see , the basis @xmath391 of @xmath287 actually consists of weight vectors for this representation . in the following , we equip @xmath386 with the inner product @xmath392 . then for a diagonal matrix @xmath393 , we have @xmath394 where @xmath395 and @xmath396 is the matrix with a 1 in the @xmath397 position along the diagonal and zero elsewhere . thus the diagonal matrices @xmath398 are the weights of the representation ( relative to the inner product that we have fixed on @xmath386 ) . for a given element @xmath399 , we write @xmath400 . when @xmath373 is a lie algebra structure on @xmath374 , we see that the @xmath401 are the usual structure constants which satisfy @xmath402 . let @xmath399 be a nilpotent lie algebra structure on @xmath374 . as discussed above , the metric lie algebra @xmath403 corresponds to a simply - connected lie group with left - invariant metric which we denote by @xmath404 . denote the ricci and scalar curvatures of @xmath404 by @xmath405 and @xmath406 , respectively . in terms of the norm on @xmath287 given above , we have @xmath407 and , further , @xmath408 where @xmath409 is the moment map of the representation , which is defined by @xmath410 for all @xmath388 . now let @xmath404 be a nilsoliton , so that @xmath411 for some @xmath412 and @xmath27 a derivation of the lie algebra @xmath413 . first observe that @xmath373 is an eigenvector of @xmath405 with eigenvalue @xmath414 . ( this follows from the fact that @xmath415 , as @xmath27 is a derivation , and @xmath416 acts by @xmath417 . ) we may assume that the ricci tensor is diagonal with respect to @xmath378 and , encoding the previous statement using the weights of our representation , we have @xmath418 now applying the definition of the moment map given above , we also see that @xmath419 from which we obtain @xmath420 . if we normalize so that @xmath421 , we see that @xmath422 and so @xmath423 . this yields @xmath424 finally we consider the case of interest , namely when @xmath373 is 2-step nilpotent . from @xcite , we know @xmath405 preserves the commutator subalgebra @xmath425 ( and also its orthogonal complement ) . furthermore , the eigenvalues of @xmath405 on the commutator subalgebra @xmath425 are positive and non - positive on its orthogonal compliment . applying we see that @xmath426 for any eigenvalue @xmath427 of @xmath405 and so we have . for , recall that for any nilsoliton , we have @xmath428 for @xmath429 such that @xmath430 . as the @xmath398 are all of the same length , this implies that @xmath405 is the minimal convex combination of @xmath431 , i.e. the element of minimal length in the convex combination . from this , we see that @xmath432 is greater than or equal to the minimal convex combination of all the @xmath398 such that @xmath433 and @xmath434 . we denote said minimal convex combination by @xmath435 and observe that it satisfies @xmath436 for all @xmath429 such that @xmath433 and @xmath434 . by fixing any two of @xmath437 and varying the third , and applying , we see that @xmath438 now implies @xmath439 as @xmath435 is a convex combination of the @xmath398 , and each @xmath398 has trace @xmath417 , we have @xmath440 from these two equations , we see that @xmath441 and thus @xmath442 . this means @xmath443 our goal is to estimate the eigenvalues of the operator @xmath444 . namely , according to corollary [ cor : alg - stab - cond ] , we want to show that @xmath445 , where @xmath14 is any symmetric 2-tensor . these tensors have the same form as in . for @xmath112 , we will consider the riemann and ricci parts separately . a computation similar to shows that @xmath446 assume that the orthonormal basis of @xmath40 is chosen so that the variation @xmath14 is diagonal on @xmath447 and on @xmath448 . ( we can not assume that @xmath14 has any special form on @xmath449 . ) then the first term is @xmath450|^2 \\ & = { { \textstyle{\frac{3}{4 } } } } \sum_{i } h_{ii}^2 \sum_{j\alpha } ( c_{ij}^\alpha)^2 = { { \textstyle{\frac{3}{4 } } } } \sum_{i } h_{ii}^2 \ , 2 |r_{ii}| \\ & \leq { { \textstyle{\frac{3}{2 } } } } \rho_{- } \sum_{i } h_{ii}^2 \end{aligned}\ ] ] where the first inequality uses the arithmetic / geometric mean inequality . the middle term is @xmath451 and last term is @xmath452 for the ricci part of @xmath112 , we assume that the orthonormal basis of @xmath40 is chosen such that the ricci operator is diagonal . then , as in , @xmath453 noticing that the right - hand expression in is invariant under change of orthonormal basis , we combine the above calculations to obtain @xmath454 we now have another criterion for stability of nilsolitons in the two - step case . [ prop:2step - stab - cond ] a two - step nilsoliton is strictly linearly stable if all of the following hold . @xmath455 we now use this to prove the main result of this section . [ prop:2step - stab ] every two - step nilsoliton is strictly linearly stable . we will use proposition [ prop:2step - stab - cond ] and the results of lemma [ lem:2step - ric - est ] . as in the lemma , we normalize so that @xmath456 . first , we have @xmath457 and we can rewrite the trace of @xmath27 as @xmath458 also , implies that @xmath459 by , for the first two conditions in proposition [ prop:2step - stab - cond ] to be satisfied it is enough that @xmath460 using , it is therefore enough to show that @xmath461 similarly , for third condition in proposition [ prop:2step - stab - cond ] to be satisfied it is enough that @xmath462 again using , it is enough to show that @xmath463 one other condition is the restriction on the dimensions of @xmath321 and @xmath337 from before , @xmath464 which must hold for any two - step nilpotent lie algebra . as @xmath465 , one can rewrite conditions , , and as @xmath466 the curves from the first two conditions intersect at @xmath467 , and the second and third intersect at @xmath468 ; see figure [ fig:2step ] , where the solid line represents the third inequality . from this , it is not difficult to see that all pairs @xmath367 satisfy these three conditions simultaneously , except those of the form @xmath356 , @xmath469 , and @xmath470 . but these satisfy @xmath471 and are covered in section [ sec : low - dim ] . -plane corresponding to conditions , , and , and the three types @xmath356 , @xmath469 , @xmath470 not satisfying those conditions.,scaledwidth=80.0% ] let @xmath38 be a three - step solvable lie algebra . that is , @xmath472 = { \ensuremath \mathfrak{v } } \oplus { \ensuremath \mathfrak{z}}$ ] is two - step nilpotent , and @xmath49 where @xmath134 is abelian . assume that @xmath473 ( as in the previous section ) and that @xmath134 is one - dimensional ( as in the construction following theorem [ thm : lauret ] ) . the goal of this section is to prove that einstein metrics on certain spaces of this form are linearly stable . [ prop : stab - solv - heis - free ] let @xmath38 be a solvable lie algebra whose nilradical is codimension - one , and either a generalized heisenberg algebra or a free two - step nilpotent lie algebra . then the natural einstein metric on @xmath38 is strictly linearly stable . the idea is to use corollary [ cor : alg - stab - cond ] , which means we must understand the operator @xmath34 on the relevant spaces . the analysis similar in spirit to that in the previous section , but more complicated due to the added complexity of the solvable spaces involved . let @xmath475 be an orthonormal basis for @xmath38 , with @xmath358 , @xmath359 , and @xmath130 . we will use capital roman indices for elements of @xmath40 , lowercase roman indices for elements of @xmath321 , greek indices for elements of @xmath337 , and 0 as the index for @xmath134 . with respect to this basis , let us describe the riemann curvature . writing @xmath476 to mean @xmath477 , here are all the curvature components sorted according to the number of indices from @xmath321 , @xmath337 , and @xmath134 . nilpotent lie groups admitting nilsoliton metrics have been completely classified in low dimensions ; a list of nilsolitons in dimensions up to six appears in @xcite ( progress has also been made in dimensions 7 and 8 @xcite ) . in this section , we show that all such nilsolitons are linearly stable , and each has a one - dimensional einstein extension ( described following theorem [ thm : lauret ] ) that is linearly stable as well . see tables [ table : solitons ] and [ table : solitons2 ] for a list of these nilsolitons ( as described in @xcite ) , which includes the dimension , step , soliton constant , trace of the soliton derivation , and largest eigenvalue of the operator @xmath112 . the tables also includes the largest eigenvalue of the operator @xmath34 for each 1-dimensional solvable einstein extension . the method to determine stability is to use corollary [ cor : alg - stab - cond ] and see that the required bounds on the relevant operators are indeed satisfied . using a computer algebra system , it is possible to input the structure constants for a lie algebra and from them calculate all related geometric quantities . in particular , with respect to an orthonormal basis for the space of symmetric 2-tensors , one can describe @xmath34 and @xmath112 as matrices and calculate their respective eigenvalues . while the eigenvalues in tables [ table : solitons ] and [ table : solitons2 ] are only approximate , they can be described to great precision as roots of characteristic polynomials . [ prop : low - dim ] each nilsoliton metric on a nilpotent lie algebra of dimension six or less is strictly linearly stable . each einstein metric on the corresponding 1-dimensional extension is strictly linearly stable . [ ex : mu11 ] here we consider the 6-dimensional algebra listed as # 11 in table [ table : solitons2 ] . this does appear in the classifications in @xcite , but the explicit nilsoliton is incorrect ; see ( * ? ? ? * remark 5.3 ) . what follows is an alternate description . with respect to an orthonormal basis @xmath478 , the non - zero structure constants are @xmath479 & = x_4 & & & [ x_1,x_4 ] & = x_5 & & & [ x_1,x_5 ] & = x_6 \\ [ x_2,x_3 ] & = x_6 & & & [ x_2,x_4 ] & = x_6 & & & \end{aligned}\ ] ] represent this structure as @xmath480 , where @xmath481 is the variety of 6-dimensional nilpotent bracket structures . the ricci tensor is @xmath482 consider @xmath483 so that under the action of @xmath484 on @xmath485 given by @xmath486 we have an isomorphic lie algebra @xmath487 whose non - zero structure constants are @xmath479 & = \frac{3}{5 } \sqrt{\frac{3}{2 } } x_3 + \frac{3}{10 } x_4 & & & [ x_1,x_3 ] & = \frac{3}{\sqrt{10 } } x_5 & & & [ x_1,x_5 ] & = \frac{3}{5 } \sqrt{\frac{3}{2 } } x_6 \\ [ x_2,x_3 ] & = \frac{3}{10 } x_6 & & & [ x_2,x_4 ] & = \frac{3}{5 } \sqrt{\frac{3}{2 } } x_6 \end{aligned}\ ] ] this new bracket representation results in a diagonal ricci tensor , and one can check that the soliton derivation and constant are as follows : @xmath488 the maximum eigenvalues of q ( and of @xmath34 for the one - dimensional einstein extension ) can now be computed , and appear in table [ table : solitons2 ] . [ ex : jorgescurve ] lauret has produced a curve of pairwise non - isometric @xmath489-dimensional einstein solvmanifolds @xcite that depend on a parameter @xmath490 . with respect to an orthonormal basis @xmath491 , the corresponding codimension - one nilradical has lie bracket relations @xmath479 & = ( 1-t)^{1/2 } x_3 & & & [ x_2,x_3 ] & = x_5 \\ [ x_1,x_3 ] & = x_4 & & & [ x_2,x_4 ] & = x_6 \\ [ x_1,x_4 ] & = t^{1/2 } x_4 & & & [ x_2,x_5 ] & = t^{1/2 } x_7 \\ [ x_1,x_5 ] & = x_6 & & & [ x_3,x_4 ] & = ( 1-t)^{1/2 } x_7 \\ [ x_1,x_6 ] & = x_7 & & & & \end{aligned}\ ] ] this is a nilsoliton : @xmath492 note that these quantities do not depend on @xmath493 . we can form the einstein extensions , and compute the curvatures numerically and see that the eigenvalues of @xmath494 are less than @xmath495 for each @xmath490 , so we have strict linear stability . see figure [ fig : curve ] . linear stability of the nilsolitons follows from proposition [ prop : ext - stab ] , since @xmath496 [ ex : nil3-solv - ext ] here we consider a family of non - einstein solvsolitons . recall that for a given nilsoliton there is a moduli space of corresponding solvsoliton extensions , and will has described all such moduli spaces in low dimensions @xcite . we consider the 1-parameter space of non - einstein solvsolitons constructed from @xmath498 , found in ( * ? ? ? * section 4.1 ) . for @xmath499 , we have non - zero bracket relations @xmath479 = 3 \quad [ a , x_1 ] = t x_1 \quad [ a , x_2 ] = \sqrt{1-t^2 } x_2 \quad [ a , x_3 ] = ( t+\sqrt{1-t^2 } ) x_3 \end{aligned}\ ] ] and the solvsoliton metrics are given by @xmath500 one can compute all curvatures and see that for each @xmath493 , @xmath501 for all non - zero @xmath14 , so all solvsolitons are stable . see figure [ fig : nil3-moduli ] .

in a previous paper @xcite , a family of fixed - point algorithms for the numerical approximation of nonlinear systems of the form @xmath0 was introduced . in ( [ mm1 ] ) , @xmath1 is a nonsingular @xmath2 real matrix and @xmath3 is an homogeneous function with degree @xmath4 ( this means that @xmath5 ) . among other applications , systems of this form are very typical in the numerical generation of traveling waves and ground states in water wave problems and nonlinear optics . for the numerical approximation to solutions of ( [ mm1 ] ) , the use of the classical fixed - point algorithm is not suitable . this is due to the fact that if @xmath6 is a solution and @xmath7 stands for the iteration matrix at @xmath6 , then , since @xmath8 is homogeneous of degree @xmath9 then @xmath10 and therefore @xmath11 that is , @xmath9 is an eigenvalue of @xmath12 with magnitude above one . this makes the iteration not convergent in general . as an alternative and based on the petviashvili method , @xcite , the following fixed - point algorithms were considered in @xcite : @xmath13 from @xmath14 and where @xmath15 is a @xmath16 function satisfying the following properties : * a set of fixed points of the iteration operator @xmath17 coincides with a set of fixed points of ( [ mm1 ] ) . this means that : ( a ) if @xmath6 is a solution of ( [ mm1 ] ) then @xmath18 ; ( b ) inversely , if the sequence @xmath19 , generated by ( [ mm2 ] ) , converges to some @xmath20 , then @xmath21 ( and , consequently , @xmath20 is a solution of ( [ mm1 ] ) ) . * @xmath22 is homogeneous with degree @xmath23 such that @xmath24 . the function @xmath22 is called the stabilizing factor of the method , inheriting the nomenclature of the petviashvili method . actually , formula ( [ mm2 ] ) generalizes the petviashvili scheme , which corresponds to the choice @xmath25 the first part of the work , carried out in @xcite ( see also @xcite ) , analyzed the convergence of ( [ mm2 ] ) . the main conclusion was that , compared to the classical fixed - point algorithm , the stabilizing factor acts like a filter of the spectrum of the matrix @xmath12 in the sense that : * the eigenvalue @xmath26 of @xmath12 is transformed to the eigenvalue @xmath27 of the iteration matrix @xmath28 of ( [ iterop ] ) . * the rest of the spectrum of @xmath28 is included into the spectrum of @xmath12 . thus , the convergence of ( [ mm2 ] ) depends of the spectrum of @xmath12 different from @xmath9 . from these conclusions , several results of convergence can be derived , see @xcite for details . the motivation of this paper is two - fold . first , several numerical experiments in the literature show that the petviashvili type algorithms are in sometimes computationally slower than other alternatives . in order to continue to benefit from the easy implementation ( one of the advantages of the methods ) the algorithms should improve their performance with the inclusion of some acceleration technique . a further motivation comes from the known mechanism of some extrapolation methods , @xcite , to transform divergent into convergent cases . the application of this property to these petviashvili type methods may extend their use to compute traveling waves under more demanding conditions , for example in two dimensions or / and in case of highly oscillatory waves . the literature on acceleration techniques is very rich with many different families and strategies , @xcite . this paper will be focused on two types of procedures : the vector extrapolation methods , @xcite and the anderson mixing , @xcite . we think that the first one is the most widely studied group ; in particular , known convergence results for some of these methods will serve us to justify several examples of transformation from divergence to convergence when generating traveling waves iteratively . the second family accelerates the convergence by introducing the strategy of minimization of the residual in some norms at each step . they have been revealed efficient in , for example , electronic structure computations , @xcite ( see also @xcite and references therein ) and , to our knowledge , this is the first time they are applied to the numerical generation of traveling waves . the main purpose of this paper is then exploring by numerical means the application of these two acceleration methods to the generation of traveling waves through several problems of interest and from the petviashvili type methods ( [ mm2 ] ) and their extended versions derived in @xcite . with the case studies presented here we have tried to cover different situations of hard computation of the waves as an attempt to give some guidelines of application . in this sense , the paper provides several conclusions to be emphasized : * the use of acceleration techniques is highly recommended here since it improves the performance in general and allows to extend the application of the methods to computationally harder situations , with especial emphasis on two - dimensional simulations and highly oscillatory wave generation . * by comparing the two families of acceleration techniques considered in this study , the vector extrapolation methods are in general more competitive for these problems compared to the anderson acceleration methods . among the vector extrapolation methods , the polynomial methods provide a better performance in general ( some exceptions can be seen in the experiments below ) . * the main drawback of the anderson acceleration methods concerns the numerical treatment of the associated minimization problem , since most of the difficulties come from ill - conditioning . this might be improved by including suitable preconditioning techniques ( here the methods were implemented in a standard way , @xcite ) . however , it is remarkable that when the anderson acceleration methods work , their performance is in general comparable to that of some vector extrapolation methods . the structure of the paper is as follows . section [ se2 ] is devoted to a description of the two families of acceleration techniques considered in this study . this also includes some comments on the implementation and convergence results . the application of both techniques to the methods ( [ mm2 ] ) is studied in section [ se3 ] through a plethora of numerical experiments involving the computation of different types of traveling waves : ground states , classical and generalized solitary waves as well as periodic traveling waves . the numerical study will be focused on the two main motivations of the paper : the improvement of the efficiency and the extension of application of the methods to computationally harder problems and where the iteration is initially not convergent . finally , section [ se4 ] completes the computational study with some illustrations of the application of the acceleration to the extended versions of the algorithms ( [ mm2 ] ) , treated in @xcite and suitable when the nonlinearity in ( [ mm1 ] ) contains several homogeneous terms with different degree . some concluding remarks are in section [ se5 ] . besides the local character of the convergence , in some cases and compared to other alternatives , fixed point algorithms has the additional disadvantage of a slow performance . in what follows , several techniques of acceleration will be considered and applied to the methods ( [ mm2 ] ) , with the aim of improving their efficiency . furthermore , as in the case of the classical algorithm , @xcite , some cases of divergence will be transformed to convergent iterations . this section introduces two families of acceleration techniques : the vector extrapolation methods ( vem from now on ) and the anderson acceleration methods ( aam ) . we will include a description of the schemes ( including some convergence results ) and some comments on implementation . the first group of acceleration techniques consists of vector extrapolation methods ( vem ) . for a more detailed analysis and implementation of the methods see @xcite and references therein . here we will describe the general features of the procedures and their application to ( [ mm2 ] ) . two families of vem are typically emphasized in the literature . the first one covers the so - called polynomial methods ; they include , as the most widely cited , the minimal polynomial extrapolation ( mpe ) , the reduced rank extrapolation ( rre ) and the modified minimal polynomial extrapolation ( mmpe ) methods , @xcite . the second family consists of the so - called @xmath29-algorithms ; typical examples are the scalar and vector @xmath29-algorithms and the topological @xmath29-algorithm , @xcite . all the methods share of course the idea of introducing the extrapolation as a procedure to transform the original sequence @xmath30 of the involved iterative process by some strategy . the polynomial methods are usually described in terms of the transformation ( @xmath31 ) @xmath32 where * @xmath33 . * @xmath34 ( @xmath35 ) denotes the @xmath36 matrix of columns @xmath37 . * @xmath38 stands for the @xmath36 matrix of some columns @xmath39 with @xmath40 as the adjoint matrix of @xmath38 ( conjugate transpose ) . in ( [ fsec211 ] ) , @xmath41 stands for the moore - penrose generalized inverse of @xmath42 , defined as @xmath43 , @xcite . different choices of the vectors @xmath44 lead to the most widely used polynomial methods : 1 . minimal polynomial extrapolation ( mpe ) : @xmath45 . 2 . reduced rank extrapolation ( rre ) : @xmath46 . 3 . modified minimal polynomial extrapolation ( mmpe ) : @xmath47 , for arbitrary , fixed , linearly independent vectors @xmath48 . the formulation of the vem may follow an alternative approach , @xcite . the transformation ( [ fsec211 ] ) can be computed in the form @xmath49 where the coefficients @xmath50 are obtained from the resolution ( in some sense ) of overdetermined , inconsistent systems @xmath51 for some vectors @xmath52 . different methods emerge by combining different choices of the norm where the residual vector @xmath53 is minimized with suitable vectors @xmath52 . thus , for example , assuming @xmath54 , we have : * rre is obtained by writing ( [ fsec214 ] ) in the form @xmath55 where the @xmath56 solve ( [ fsec215 ] ) with @xmath57 and the euclidean norm with equal weights is used . * mpe is obtained by using ( [ fsec214 ] ) with @xmath58 where @xmath59 and the @xmath60 solve ( [ fsec215 ] ) with @xmath61 in the sense of minimization with the euclidean norm with equal weights . * mmpe is obtained from ( [ fsec214 ] ) but where instead of ( [ fsec215 ] ) a system of the form @xmath62 is used . in ( [ fsec216 ] ) @xmath63 , being @xmath64 . these formulations can be unified by a representation with determinants , @xcite . this writes the extrapolation steps @xmath65 in the form @xmath66 with @xmath67 where the @xmath68 are scalars that depend on the extrapolation method and where the expansion of ( [ fsec218 ] ) is in the sense @xmath69 with @xmath70 the cofactor of @xmath71 in the first row . thus in the case of the numerator in ( [ fsec217 ] ) , formula ( [ fsec219 ] ) is a vector , while in the case of the denominator in ( [ fsec217 ] ) , formula ( [ fsec219 ] ) is a scalar . ( see @xcite for the interpretation in terms of the schur complement of a matrix . ) the three previously mentioned polynomial methods correspond to the following choices of @xmath68 : * mpe : @xmath72 . * rre : @xmath73 . * mmpe : @xmath74 , where @xmath75 are linearly independent vectors in @xmath76 . a second family of vem is called the @xmath29-algorithms . a description of them may start from the scalar @xmath29-algorithm of wynn , @xcite . this scalar extrapolation method can be derived from the representation ( [ fsec217 ] ) , ( [ fsec218 ] ) ( in the scalar case ) with @xmath77 ( which are scalars in the scalar case ) . the corresponding ratio of determinants @xmath78 is called the classical @xmath79- ( or shanks schmidt @xmath80 ) transform , @xcite . this ratio can be evaluated recursively for increasing @xmath81 and @xmath82 without the computation of determinants or schur complements . the corresponding formulation is @xmath83 where @xmath84 and works along diagonals on @xmath85 constant . thus , from ( [ fsec2111 ] ) , formulas ( [ fsec2112 ] ) , ( [ fsec2113 ] ) compute each entry of a triangular array in terms of the previous entries . the extension of the scalar @xmath29-algorithm to the vectorial case was carried out by brezinski , @xcite , and wynn , @xcite , by using different definitions of inverse of a vector , see @xcite . wynn suggests to consider the transpose of the moore - penrose generalized inverse of a vector , @xmath86 leading to the vector @xmath29-algorithm ( vea ) , whose formulas are of the form ( [ fsec2112 ] ) , ( [ fsec2113 ] ) where the scalars are substituted by vectors and ( [ fsec2113 ] ) makes use of ( [ fsec2114 ] ) . this implies that the @xmath79-transform ( [ fsec2111 ] ) is understood in the above described vectorial sense . this is called the generalized shanks schmidt ( gss ) transform , @xcite . on the other hand , brezinski defines the inverse of pair of vectors @xmath87 such that @xmath88 as the pair of vectors @xmath89 where @xmath90 thus , @xmath91 is called the inverse of @xmath92 with respect to @xmath93 and viceversa . this definition leads to the so - called topological @xmath29-algorithm ( tea ) , when an arbitrary vector @xmath20 is fixed and the inverses of @xmath94 and @xmath95 are considered with respect to @xmath20 , that is @xmath96 the recursive formulas are @xmath97 brezinski proved , @xcite , the connection with the gss - transform , showing that @xmath98 for an efficient implementation of ( [ fsec2115 ] ) see @xcite . thus ( tea ) corresponds to take @xmath99 in ( [ fsec217 ] ) , ( [ fsec218 ] ) . the mechanism of working of the vem can be described as follows , see @xcite for details . one starts from assuming an asymptotic expression for the sequence @xmath100 of the form @xmath101 where @xmath102 and @xmath103 , ordered such that @xmath104 if @xmath105 with only a finite number of @xmath106 having the same modulus . the expansion ( [ fsec212 ] ) can be generalized by considering , instead of constant vectors @xmath107 , polynomials @xmath108 in @xmath82 with vector coefficients of the form @xmath109 with @xmath110 linearly independent in @xmath76 and where if @xmath111 then @xmath112 , @xcite . for simplicity , the description below will make use of ( [ fsec212 ] ) . the asymptotic expansion ( [ fsec212 ] ) is considered in a general vector space ( finite or infinite dimensional ) where the iteration is defined . it is understood in the sense that any truncation differs from @xmath100 in less than some power of the next @xmath113 . this means that for any positive integer @xmath8 there are @xmath114 and a positive integer @xmath115 that only depend on @xmath8 such that for every @xmath116 @xmath117 allows to identify @xmath118 as limit or anti - limit of the sequence @xmath100 , according to the size of @xmath119 . ( that is , if @xmath120 then @xmath121 exists and equals @xmath118 . if @xmath122 then @xmath121 does not exist and @xmath118 is called the anti - limit of the sequence @xmath100 . ) under these conditions , several results of convergence for mpe , rre , mmpe and tea are obtained in the literature , @xcite and references therein . for these methods , one can find an extrapolation step @xmath123 such that @xmath124 ) may explain the convergent behaviour of the extrapolation in some cases . if the @xmath113 s are identified as the eigenvalues of the linearization operator of the iteration at the limit ( or anti - limit ) @xmath118 , then the extrapolation has the effect of translating the behaviour of the iteration to an eigenvalue @xmath125 that may be into the unit disk , even if the previous ones are out of it . hence , @xmath118 may be anti - limit for the original iteration and the extrapolation converges to it . these results are extended to the defective linear case with more general polynomials ( [ fsec212b ] ) in @xcite . the integer @xmath123 is related to the concept of minimal polynomial @xmath126 of a matrix @xmath42 with respect to a vector @xmath92 , @xcite ; this is the unique polynomial of least degree such that @xmath127 thus in the case of linear iteration with matrix @xmath42 , @xmath123 is taken to be the degree of the minimal polynomial of @xmath42 with respect to the first iteration @xmath128 . in the case of a nonlinear system written in fixed point form @xmath129 then @xmath123 is theoretically defined as the degree of the minimal polynomial of @xmath130 with respect to @xmath128 , where @xmath6 is a solution of ( [ fsec2110 ] ) . in contrast with the linear case , there is no way to determine @xmath123 in advance for the nonlinear case . this forces to consider several strategies for the choice and the corresponding implementation , see the discussion in @xcite and the comments here below . we also mention that in the linear case , the extrapolation methods mpe and rre are mathematically equivalent to the method of arnoldi , @xcite , and the gmres , @xcite , respectively , see @xcite , while the mmpe is mathematically equivalent to the hessenberg method , @xcite and tea to the method of lanczos , @xcite , see @xcite . efficient and stable implementation of the rre , mpe and mmpe methods by using qr and lu factorizations can be seen in @xcite . for the case of tea , see @xcite and @xcite for vea . the implementation is usually carried out in a cycling mode . a cycle of the iteration is performed by the following steps : consider a method ( [ mm2 ] ) with a stabilizing factor @xmath22 satisfying ( p1 ) , ( p2 ) . given @xmath14 and a width of extrapolation @xmath131 , for @xmath132 , the advance @xmath133 is : * set @xmath134 and compute @xmath135 steps of the fixed - point algorithm : @xmath136 * compute the extrapolation steps ( [ fsec217 ] ) with any of the methods described above and @xmath137 . * set @xmath138 and go to step ( a ) . the cycle ( a)-(b)-(c ) is repeated until the error ( residual or between two consecutive iterations ) is below a prefixed tolerance , a maximum number of iterations is attained or the discrepancy between the stabilizing factor at the iterations and one is below a prefixed tolerance . the width of extrapolation @xmath135 depends on the choice of the technique : @xmath139 for mpe , rre or mmpe and @xmath140 for vea or tea , @xcite . since @xmath123 is generally unknown , some strategy for the implementation must be adopted . in practice , as discussed in @xcite , the methods are implemented with some ( small ) values of @xmath135 and take that with the best performance . it may be also different for each cycle , although quadratic convergence is not expected if @xmath123 is too small . this choice of @xmath123 will be computationally studied in some examples in section [ se3 ] . the hypotheses for the expansion ( [ fsec212 ] ) include the conditions @xmath141 . in many problems for traveling wave generation , @xmath142 appears as eigenvalue of the iterative technique ( of fixed point type ) although under especial circumstances that allow to extend the convergence results in some sense . this especial situation is related to the presence of symmetries in the equations for traveling waves . in order to extend the results of convergence to this case , one has to consider the orbits by the symmetry group of the equations and interpret the convergence in the orbital sense , that is a convergence for the orbits . the description of this orbital convergence can be seen in @xcite . finally , local quadratic convergence is proved in @xcite ( see also @xcite ) for the four methods and vea for a general nonlinear system ( [ fsec2110 ] ) and under the following hypotheses on @xmath143 : * the jacobian matrix @xmath144 does not have @xmath142 as eigenvalue . * @xmath81 is the degree of the minimal polynomial of @xmath144 with respect to @xmath145 . * the algorithm is implemented in the cycling mode where @xmath123 is chosen in the @xmath146-th cycle as the degree of the minimal polynomial of @xmath144 with respect to @xmath147 . this result can be extended to the case where @xmath148 admits a @xmath149-parameter ( @xmath150 ) group of symmetries and , consequently , @xmath142 is an eigenvalue of @xmath144 , by using the reduced system for the orbits of the group and in the orbital sense , @xcite . this would lead to quadratic orbital convergence but linear local convergence . a second family of acceleration techniques considered here is the so - called anderson family or anderson mixing , @xcite . it is widely used in electronic structure computations and only recently it has been analyzed in a more general context , @xcite . ( to our knowledge , this is the first time that aam are applied to accelerate traveling wave computations . ) the main goal of the approach consists of combining the iteration with a minimization problem for the residual at each step . for linear problems , this technique is essentially equivalent to the gmres method , @xcite . the stages for an iteration step are as follows : given @xmath151 , set @xmath152 . for @xmath153 * set @xmath154 * set @xmath155 where @xmath156 * determine @xmath157 that solves @xmath158 * set @xmath159 the resolution of the optimization problem ( [ minimize ] ) is the source of the additional computational work of the acceleration . one way to reduce this extra effort is the so - called multisecant updating @xcite . ( this also clarifies the connection with quasi - newton methods . ) this technique consists of writing the problem in an equivalent form @xmath160 but with a more direct resolution , and determining the acceleration from it . the general step becomes : * set @xmath154 * determine @xmath161 by solving ( [ minip ] ) . * set @xmath162 * set @xmath163 as mentioned in @xcite , if @xmath164 is full - rank , the solution of the minimization problem can be written as @xmath165 and the anderson acceleration has the alternative form @xmath166 ( note that @xmath167 can be viewed as an approximate inverse of the jacobian of @xmath168 ) . the formulation ( [ aaalternative ] ) motivates the generalization of the anderson mixing , @xcite . this is performed replacing @xmath169 in ( [ aaalternative ] ) by some @xmath170 satisfying @xmath171 and ( [ aaalternative ] ) becomes @xmath172 the resulting methods are collected in the so - called anderson s family , @xcite . two particular members are emphasized : the type - i method ( denoted by aa - i from now on ) , which corresponds to @xmath173 in ( [ aaalternative2 ] ) and type - ii method ( or aa - ii ) , which is the original anderson mixing ( [ aaalternative ] ) . to our knowledge , some convergence results can be seen in @xcite . in @xcite the authors identify some anderson methods for linear problems and in some sense with the gmres method and the arnoldi ( fom ) method ; some convergence results can be derived from this identification . in @xcite the equivalence with gmres for linear problems is completely characterized . finally , @xcite gives some proofs of convergence of the anderson acceleration when applied to contractive mappings : @xmath23-linear convergence of the residual for linear problems under certain conditions when @xmath174 and local @xmath175-linear convergence in the nonlinear case . ( these types of convergence are defined in the paper . ) on the other hand , as observed in @xcite , the implementation of aam should be carried out by attending to three main points : a convenient formulation of the minimization problem ( [ minimize ] ) , a numerical method for its efficient resolution and , finally , the parameter @xmath176 , which plays a similar role to that of the extrapolation width @xmath135 in the vem . in our computations below , we have followed the treatment described in @xcite . this is based on the use of the unconstrained form ( [ minip ] ) and its numerical resolution with @xmath177 decomposition . for other alternatives in both problems , see the discussion in @xcite and the references cited there . ( according to our results below , the use of alternative preconditioning techniques might be recommendable in some cases . ) as far as the choice of @xmath176 is concerned , a similar strategy to that of @xmath135 will be used , since our experiments , @xcite , suggest that @xmath176 ( as @xmath135 ) strongly depends on the problem under study and large values are not recommended . finally the codes are implemented by retaining the definition of @xmath178 since other alternatives , @xcite , did not improve the results in a relevant way . presented here is a comparative study on the use of vem and aam as acceleration techniques from the petviashvili type methods ( [ mm2 ] ) in traveling wave computations . the comparison is organized according to two main points : the type of traveling wave to be generated and the elements of each family of methods to be used for the generation . the first point includes the following case studies : 1 . classical solitary waves , generalized solitary waves and periodic traveling waves of the four - parameter boussinesq system , @xcite . localized ground state solutions of nls type equations , @xcite . highly oscillatory solitary waves of the one- and two - dimensional benjamin equation , @xcite . this plethora of waves attempts to discuss and overcome different computational difficulties and with the aim of establishing as more general conclusions as possible . as for the second point , each family of techniques has been represented by the following methods : * mpe and rre standing for polynomial extrapolation methods . * vea and tea standing for @xmath29 algorithms . * aa - i and aa - ii standing for the aam . for simplicity , the acceleration will be applied to the petviashvili method ( [ mm2 ] ) , ( [ mm3c ] ) with @xmath179 . due to the similar behaviour of the methods of the family ( [ mm2 ] ) , illustrated in @xcite , the conclusions from the corresponding results can reasonable serve when the petviashvili method is substituted by any of ( [ mm2 ] ) . in all the cases considered , the traveling wave profiles appeared as solutions of initial value problems of ode s . their discretization to generate approximations to the profiles , was carried out in a common and standard way . the corresponding initial periodic boundary value problem ( on a sufficiently long interval ) was discretized by using fourier collocation techniques , @xcite ; the discretization leads to a nonlinear system of algebraic equations for the approximate values of the profile at the collocation points or for the discrete fourier coefficients of the approximation . this system is iteratively solved with the classical petviashvili method ( [ mm2 ] ) , ( [ mm3c ] ) along with the selected acceleration technique . this will be described in each equation considered below . several stopping criteria for the iterations are implemented : * a maximum number of iterations . * the iteration stops when one of the following quantities are below a prefixed , small tolerance @xmath180 : * * the difference in euclidean norm between two consecutive iterations @xmath181 * * the residual error ( also in euclidean norm ) @xmath182 * * the discrepancy between the stabilizing factor and ( in case of convergence ) its limit one @xmath183 the numerical experiments that form the comparative study are of different type : * for several values of @xmath123 ( in the case of vem ) and @xmath176 ( in the case of aam ) we have computed the number of iterations required by each method to achieve a residual error below the prefixed tolerance . this allows to compare some performance of the methods between the two families , among different techniques within a same family and indeed with the petviashvili method without acceleration . * some eigenvalues of the iteration matrices for the classical fixed point algorithm and the petviashvili method have been computed ( with the corresponding standard matlab function ) in order to explain the behaviour of the second one , @xcite and how the acceleration eventually changes it . * the form of the approximate profiles and some experiments to check their accuracy are also displayed . in this first example we study the numerical generation of traveling wave solutions of the four - parameter family of boussinesq system @xmath184 where @xmath185 and the four parameters @xmath186 satisfy @xmath187 with some constant @xmath188 $ ] , @xcite . system ( [ fsec311 ] ) appears as one of the alternatives to model the bidirectional propagation of the irrotational free surface flow of an incompressible , inviscid fluid in a uniform horizontal channel under the effects of gravity when the surface tension and cross - channel variations of the fluid are assumed to be negligible . if @xmath189 denotes the undisturbed water depth , then @xmath190 stands for the deviation of the free surface from @xmath189 at the point @xmath191 and time @xmath192 , while @xmath193 is the horizontal velocity of the fluid at @xmath191 and at the height @xmath194 ( where @xmath195 corresponds to the channel bottom ) at time @xmath192 . for the derivation of ( [ fsec311 ] ) from the two - dimensional euler equations and the mathematical theory see @xcite . for the modification of ( [ fsec312 ] ) to include the influence of surface tension see @xcite . the boussinesq system ( [ fsec311 ] ) admits different types of traveling wave solutions . first , being an approximation to the corresponding two - dimensional euler equations in the theory of surface waves , it is expected to have classical solitary wave solutions . they are solutions of the initial value problem of ( [ fsec311 ] ) of smooth traveling wave form @xmath196 with some speed @xmath197 and decaying to zero as @xmath198 . substitution into ( [ fsec311 ] ) and after one integration the profiles @xmath199 must satisfy the ode system @xmath200 the problems of existence , asymptotic decay and stability of solutions of ( [ fsec313 ] ) have been analyzed in many references and for particular values of @xmath186 , see @xcite and references therein . furthermore , in the same reference , linearly well - posed systems ( [ fsec313 ] ) may be studied as a first order ode system and , based on normal form theory , a discussion on the values of the parameters leading to boussinesq systems admitting solitary wave solutions with speed @xmath201 is established . according to it , two classes of systems can be distinguished . the first one admits classical ( in the sense above defined ) solitary wave solutions . this group contains the bona - smith system ( @xmath202 ) , @xcite , or the bbm - bbm system ( @xmath203 ) , @xcite . ( the classical boussinesq system , which corresponds to @xmath204 , is also in this group , although it is out of the general discussion of @xcite and has been studied separately , @xcite . ) a second class of boussinesq systems admits generalized solitary wave solutions , that is traveling wave profiles which are not homoclinic to zero at infinity but to small amplitude periodic waves , @xcite . the kdv - kdv system ( @xmath205 ) , @xcite , is an example of this second group . finally , the existence of periodic traveling wave solutions of ( [ fsec311 ] ) is studied in @xcite by applying topological degree theory for positive operator to the corresponding periodic initial value problem posed on an interval @xmath206 and some cnoidal wave solutions of the bbm - bbm system are computed . a smooth periodic traveling wave solution @xmath207 with some speed @xmath197 must satisfy the ode system @xmath208 for some real constants @xmath209 . these are related to the period parameter @xmath210 . the resolution of ( [ fsec313b ] ) involves a modified system for which these constants of integration are set to zero , @xcite . this is briefly described as follows . one first searches for constant solutions @xmath211 of ( [ fsec313b ] ) . this leads to the system @xmath212 which can be solved as a cubic equation for @xmath213 : @xmath214 once @xmath215 is obtained ( there may be more than one solution indeed ) the differences @xmath216 must satisfy @xmath217 this strategy will be considered in the numerical generation of the profiles in ( [ fsec313b ] ) : the system ( [ fsec313d ] ) wil be discretized to compute approximations to the variables @xmath218 and to the variables @xmath219 from them . in order to generate numerically classical and generalized solitary wave solutions of ( [ fsec311 ] ) the corresponding periodic value problem of ( [ fsec313 ] ) on a long enough interval @xmath206 is discretized with a fourier collocation method leading to a discrete system of the form @xmath220 where @xmath221 are approximations @xmath222 to the values of a solution of ( [ fsec313b ] ) at the grid points @xmath223 , @xmath224 is the pseudospectral differentiation matrix , @xcite , @xmath225 is the @xmath2 identity matrix and the nonlinear term @xmath8 , which is homogeneous of degree @xmath226 , involves hadamard products . in the case of periodic traveling waves and as was mentioned above , system ( [ fsec314 ] ) will be substituted in the implementation by @xmath227 for the approximations @xmath228 to the @xmath229 variables at the grid points and where @xmath230 are previously known from the resolution of ( [ fsec313c ] ) , ( [ fsec313c1 ] ) . then @xmath231 . the methods ( [ mm2 ] ) along with the corresponding acceleration technique are then applied to the discrete systems ( [ fsec314 ] ) and ( [ fsec314b ] ) . the implementation is performed in the fourier space ; for example ( [ fsec314 ] ) becomes @xmath232 thus the @xmath233 system ( [ fsec314 ] ) is divided into @xmath234 blocks of @xmath235 systems for the corresponding @xmath9-th discrete fourier coefficients @xmath236 ( for simplicity , we assume that @xmath237 for some @xmath238 . ) alternatively , ( [ fsec314 ] ) can be written in the form @xmath239 where @xmath240 denotes periodic convolution and if @xmath241 , the vectors @xmath242 have discrete fourier coefficients @xmath243 in order to explain the behaviour of the iteration , the size of the eigenvalues of the iteration matrix will be relevant in the numerical study . in this case , the corresponding iteration matrix of the classical fixed point iteration at a solution @xmath244 has the form @xmath245 ( where @xmath246 stands for the diagonal matrix with diagonal entries given by the components of @xmath247 ) . some information on the spectrum of @xmath12 is known . we already have the eigenvalue @xmath248 , corresponding to the degree of homogeneity of the nonlinear part , with @xmath244 as an eigenvector . also , the application of @xmath224 to ( [ fsec314 ] ) leads to @xmath249 which means that @xmath142 is an eigenvalue of @xmath12 and @xmath250 is an associated eigenvector . this corresponds to the translational invariance of ( [ fsec311 ] ) . three particular systems of ( [ fsec311 ] ) will be taken to illustrate the numerical generation of traveling waves . the first one is the classical boussinesq system ( @xmath204 ) , @xcite @xmath251 which is known to have classical solitary wave solutions , @xcite . the second one is the so - called kdv - kdv system ( @xmath252 ) @xmath253 that admits generalized solitary wave solutions , @xcite . finally , in order to illustrate the numerical generation of periodic traveling waves , the bbm - bbm system ( @xmath203 ) , @xmath254 will be taken , @xcite . in the case of system ( [ fsec317 ] ) a first experiment of comparison of the acceleration techniques has been made by taking @xmath255 and a hyperbolic secant profile as initial iteration with @xmath256 and @xmath257 collocation points . the petviashvili method ( [ mm2 ] ) , ( [ mm3c ] ) with @xmath258 was first run , generating approximate @xmath259 and @xmath118 profiles as shown in figures [ figuresw0](a ) and ( b ) while figures [ figuresw0](c ) and ( d ) stand for the corresponding phase portraits of the approximate profiles in ( a ) and ( b ) . ( they show the classical character of the solitary waves , represented as homoclinic to zero orbits with exponential decay , @xcite . ) the accuracy of the iteration is checked in figure [ figuresw1 ] . figure [ figuresw1](a ) illustrates the convergence of the sequence @xmath260 of stabilizing factors , computed with the corresponding formula ( [ mm3c ] ) and the optimal value @xmath258 . the discrepancy ( [ fsec33 ] ) is below the tolerance @xmath261 in @xmath262 iterations , while the first residual error below @xmath180 is @xmath263 at @xmath264 . ( this also happens in the rest of the experiments : when the procedure is convergent , the error @xmath265 achieves the tolerance before the residual error ; therefore , the control on this last one is a harder test and will be adopted as the main one to stop the iteration . ) convergence is also confirmed by table [ tav1b ] . this shows the six largest magnitude eigenvalues of the iteration matrix ( [ fsec316 ] ) ( first column ) and of the iteration matrix ( at the same iterate ) of the petviashvili procedure ( second column ) both at the last computed iterate @xmath266 . the first column reveals the dominant eigenvalues @xmath267 , both simple , while the rest is below one . the filtering effect of the petviashvili method , @xcite , is observed in the second column ; the dominant eigenvalue is filtered to zero ( recall that @xmath258 ) and the rest is preserved . since @xmath268 corresponds to the translational symmetry of ( [ fsec311 ] ) , this guarantees the local convergence of the method ( also in the orbital sense mentioned above ) . .classical solitary wave generation of ( [ fsec317 ] ) . six largest magnitude eigenvalues of the approximated iteration matrix @xmath269 ( first column ) and of the iteration matrix @xmath270 , generated by the petviashvili method ( [ mm2 ] ) , ( [ mm3c ] ) with @xmath258 , both evaluated at the last computed iterate @xmath266 . [ cols="^,^",options="header " , ] in this paper we have studied numerically the use of acceleration techniques applied to fixed point algorithms of petviashvili type to generate numerically traveling waves in nonlinear dispersive wave equations . the comparison has been established between vector extrapolation methods and anderson acceleration methods for different types of traveling waves . from the plethora of numerical experiments , our main conclusions are : * the use of acceleration techniques improves the performance of the petviashvili type methods in all the cases . this improvement is observed in two main points : first , when the petviashvili type method is convergent , the acceleration reduces the number of iterations in a relevant way . ( in some cases , this is really important : in some one - dimensional problems , the reduction is at least of @xmath271 and attains up to @xmath272 . ) on the other hand , the mechanism of acceleration , especially in the case of vem , allows to transform initially divergent sequences into convergent processes . this is particularly relevant in traveling waves with high oscillations . furthermore , acceleration has been shown to be more efficient than other alternatives for some cases , like numerical continuation . * in general , vem provide better results and among them , polynomial methods such as mpe and rre are more efficient than @xmath29-algorithms like vea and tea , although in some convergent cases the acceleration in terms of the number of iterations is very similar among all the methods while in computational time the @xmath29-algorithms work better . * the aam are competitive in some cases but they are mostly affected by ill - conditioning and a more computational effort due to their longer implementation . the best results of these methods are obtained when generating numerically periodic traveling waves in some nonlinear dispersive systems and ground state profiles in nls type equations , while their performance is poor when computing highly oscillatory traveling waves . their application to the e - petviashvili type methods suggests an uncertain behaviour with respect to relevant quantities of the problem through the iteration . the main question to that this comparative study has not been able to answer is in the authors opinion finding a deeper understanding of the way how acceleration techniques ( especially vem ) work on these problems . in particular , we miss some conclusions about the width of extrapolation ( which is related to the extrapolation step for convergence ) to be used a priori ( if it is possible to do that ) . we have observed that this looks to be strongly dependent on the problem under study . this might be though a good starting point for a future research . this research has been supported by project mtm2014 - 54710-p . ablowitzm l. a. abramyan , y. a. stepanyants , the structure of two - dimensional solitons in media with anomalous small dispersion , siv . jetp , 61(5)(1985 ) 963 - 966 . albert , j.l . bona , j.m . restrepo , solitary - wave solutions of the benjamin equation , siam j. appl . , 59 ( 1999 ) 21392161 . j. alvarez , a. duran , petviashvili type methods for traveling wave computations : i. analysis of convergence , j. comp . , 266 ( 2014 ) 39 - 51 . j. alvarez , a. duran , an extended petviashvili method for the numerical generation of traveling and localized waves , commun . nonlinear sci . simulat . , 19(2014 ) 2272 - 2283 . j. alvarez , a. duran , corrigendum to `` petviashvili type methods for traveling wave computations : i. analysis of convergence '' [ j. comput . ( 2014 ) 3951 ] , j. comp . math . , 277(2015 ) 215 - 216 . d. g. anderson , iterative procedures for nonlinear integral equations , j. assoc . , 12 ( 1965 ) 547 - 560 . t. b. benjamin , internal waves of permanent form in fluids of great depth , j. fluid mech . 29 ( 1967 ) 559 - 592 . t. b. benjamin . a new kind of solitary wave , j. fluid mech . , 245 ( 1992 ) 401 - 411 . t. b. benjamin , solitary and periodic waves of a new kind , phil . a , 354 ( 1996 ) 1775 - 1806 . j. l. bona and m. chen , a boussinesq system for two - way propagation of nonlinear dispersive waves , physica d , 116 ( 1998 ) , 191224 . j. l. bona , m. chen , j .- c . saut , boussinesq equations and other systems for small - amplitude long waves in nonlinear dispersive media : i. derivation and linear theory , j. nonlin . 12 ( 2002 ) , 283 - 318 . j. l. bona , m. chen , j .- c . saut , boussinesq equations and other systems for small - amplitude long waves in nonlinear dispersive media : ii . the nonlinear theory , nonlinearity 17 ( 2004 ) , 925 - 952 . j. l. bona , v. a. dougalis , d. e. mitsotakis , numerical solution of kdv - kdv systems of boussinesq equations : i. the numerical scheme and generalized solitary waves , math . simul . , 74(2007 ) 214 - 228 . j. l. bona , v. a. dougalis , d. e. mitsotakis , numerical solution of kdv - kdv systems of boussinesq equations : ii . generation and evolution of radiating solitary waves , nonlinearity 21 ( 2008 ) 2825 - 2848 . j. l. bona and r. smith , a model for the two - way propagation of water waves in a channel , math . 79(1976 ) , 167182 . j. v. boussinesq , thorie des ondes et des remous qui se propagent le long dun canal rectangulaire horizontal , en communiquant au liquide contenu dans ce canal des vitesses sensiblement pareilles de la surface au fond , j. math . pures appl . 17 ( 1872 ) 55 - 108 . , chebyshev and fourier spectral methods , 2nd ed . dover publications , new york , 2000 . c. brezinski , gnralization de la transformation de shanks , de la table de pad et de lepsilon - algorithm , calcolo , 12 ( 1975 ) 317 - 360 . c. brezinski , a general extrapolation algorithm , numer . , 35(1980 ) 175 - 187 . c. brezinski , convergence acceleration during the 20th century , j. comput . , 122 ( 2000 ) 1 - 21 . c. brezinski , a. c. rieu , the solution of systems of equations usiung the vector @xmath29-algorithm and an application to boundary value problems , math . comp . , 28(1974 ) 731 - 741 . c. brezinski , m. redivo zaglia , extrapolation methods , theory and practice , north - holland , amsterdam , 1991 . c. brezinski , m. redivo zaglia , a schur complement approach to a general extrapolation algorithm , linear algebra and appl . , 368 ( 2003 ) 279 - 301 . , spectral methods in fluid dynamics . springer - verlag , new york - heidelberg - berlin , 1988 . s. cabay , l. w. jackson , a polynomial extrapolation method for finding limits and antilimits for vector sequences , siam j. numer . , 13(1976 ) 734 - 752 . m. chen , m. chen , n. v. nguyen , cnoidal wave solutions to boussinesq systems , nonlinearity , 20 ( 2007 ) , 1443 - 1461 . m. chen , n. v. nguyen , s .- sun , solitary - wave solutions to boussinesq systems with large surface tension , disc . systems , ser . s , 2(1)(2009 ) 37 - 53 . m. chen , n. v. nguyen , s .- sun , existence of traveling - wave solutions to boussinesq systems , diff . eq . , 24(2011 ) 895 - 908 . p. daripa , r. k. dash , a class of model equations for bi - directional propagation of capillary - gravity waves . sci . , 41(2003 ) 201 - 218 . j. demmel , applied numerical linear algebra , siam philadelphia , 1997 . v. a. dougalis , a. duran , d. e. mitsotakis , numerical approximation of solitary waves of the benjamin equation , accepted in math . simul . , 2012 . v. a. dougalis , a. duran , d. e. mitsotakis , numerical solution of the benjamin equation , wave motion 52(2015 ) 194 - 215 . v. a. dougalis , d. e. mitsotakis , theory and numerical analysis of boussinesq systems : a review , in : effective computational methods in wave propagation , n. a. kamparis , v. a. dougalis , j. a. ekaterinaris ( eds . ) crc press 2008 , 63 - 110 . h. fang , y. saad , two classes of multisecant methods for nonlinear acceleration , numer . linear algebra appl . , 16 ( 2009 ) 197 - 221 . r. p. eddy , extrapolation to the limit of a vector sequence , in : p.c.c . wang ( ed . ) , information linkage between applied mathematics and industry , academic press , new york , 1979 , 387 - 396 . e. eyert , a comparative study on methods for convergece acceleration of iterative vector sequences , j. comput . , 124(1996 ) 271 - 285 . e. gekeler , on the solution of systems of equations by the epsilon algorithm of wynn , math . , 26(1972 ) 427 - 437 . g. h. golub , ch . f. van loan , matrix computations , j. h. u. press , baltimore , 1996 . k. jbilou , h. sadok , some results about vector extrapolation methods and related fixed point iterations , j. comput . appl . math . , 36(1991 ) 385 - 398 . k. jbilou , h. sadok , lu - implementation of the modified minimal polynomial extrapolation method , i m a j. numer . , 19(1999 ) 549 - 561 . k. jbilou , h. sadok , vector extrapolation methods . applications and numerical comparisons , j. comput . math . , 122 ( 2000 ) 149 - 165 . b. b. kadomtsev , v. i. petviashvili , on the stability of solitary waves in weakly dispersive media , sov . phys . dokl . 15 ( 1970 ) 539 - 541 . b. kim , three - dimensional solitary waves in dispersive wave systems , doctoral dissertation , department of mathematics , mit , 2006 . b. kim , t. r. akylas , on gravity - capillary lumps , j. fluid mech . 540 ( 2005 ) 337 - 351 . b. kim , t. r. akylas , on gravity - capillary lumps . two - dimensional benjamin equation , j. fluid mech . 557 ( 2006 ) 3 + 237 - 256 . , a generalized petviashvili method for scalar and vector hamiltonian equations with arbitrary form of nonlinearity , j. comput . ( 2007 ) 1668 - 1692 . lakoba , j. yang , a mode elimination technique to improve convergence of iteration methods for finding solitary waves , j. comp . phys . 226 ( 2007 ) 1693 - 1709 . c. lanczos , solutions of systems of linear equations by minimized iterations , j. res . , 49(1952 ) 33 - 53 . h. le ferrand , the quadratic convergence of the topological @xmath29-algorithm for systems of nonlinear equations , numer . algorithms , 3 ( 1992 ) 273 - 284 . e. lombardi , oscillatory integrals and phenomena beyond all algebraic orders , with applications to homoclinic orbits in reversible systems , lecture notes in mathematics , springer - verlag , berlin , 2000 . m. mesina , convergence acceleration for the iterative solution of the equations @xmath273 , comput . methods appl . engrg . , 10(1977 ) 165 - 173 . c. d. meyer , matrix analysis and applied linear algebra , siam philadelphia , 2000 . p. ni , anderson acceleration of fixed - point iteration with applications to electronic structure computations , ph . d. thesis , worcester polytechnic institute , worcester , ma , 2009 . r. l. pego , m. i. weinstein , convective linear stability of solitary waves for boussinesq equations , l stud . appl . math . 99(1997 ) 311 - 375 . , convergence of petviashvili s iteration method for numerical approximation of stationary solutions of nonlinear wave equations , siam j. numer . anal . 42 ( 2004 ) 1110 - 1127 . , soviet j. plasma phys . 2 ( 1976 ) 257 - 258 . f. potra , h. engler , a characterization of the behaviour of the anderson acceleration on linear problems , linear algebra and its applications , 438(3)(2013 ) 1002 - 1011 . p. pulay , improved scf convergence , j. comput . , 3 ( 1982 ) 556 - 560 . y. saad , krylov subspace methods for solving large unsymmetric linear systems , math . , 37(1981 ) 105 - 126 . y. saad , m. h. schultz , gmres : a generalized minimal residual algorithm for solving nonsymmetric linear systems , siam j. sci . comput . , 7(1986 ) 856 - 869 . a. sidi , convergence and stability of minimal polynomial and reduced rank extrapolation algorothms , siam j. numer . , 23 ( 1986 ) 197 - 209 . a. sidi , extrapolation vs. projection methods for linear systems of equations , j. comp . math . , 22(1)(1988 ) 71 - 88 . a. sidi , efficient implementation of minimal polynomial and reduced rank extrapolation methods , j. comput . math . , 36(1991 ) 305 - 337 . a. sidi , practical extrapolation methods , theory and applications , cambridge university press , new york , 2003 . a. sidi , j. bridger , convergence and stability analyses for some vector extrapolation methods in the presence of defective iteration matrices , j. comp . math . , 22(1988 ) 35 - 61 . a. sidi , w. f. ford , d. a. smith , acceleration of convergence of vector sequences , siam j. numer . , 23 ( 1986 ) 178 - 196 . j. r. schmidt , on the numerical solution of linear simultaneous equations by an iterative method , phil . , ser . 7 , 32(1941 ) 369 - 383 . d. shanks , non - linear transformations of divergent and slowly convergent sequences , j. math . and phys . , 34(1955 ) 1 - 42 . s. skelboe , computation of the periodic steady - state response of nonlinear networks by extrapolation methods , ieee trans . circuits and systems , 27(1980 ) 161 - 175 . d. a. smith , w. f. ford , a. sidi , extrapolation methods for vector sequences , siam rev . , 29 ( 1987 ) 199 - 233 . r. c. e. tan , implementation of the topologicasl @xmath29.algorithm , siam j. sci . 9 ( 1988 ) 839 - 848 . a. toth , c. t. kelley , convergence analysis for anderson acceleration , siam j. numer . , 53(2015 ) 805 - 819 . j. van iseghem , convergence of vectorial sequences . applications , numer . math . , 68(1994 ) 549 - 562 . v. v. voronovich , v. i. shira , y. a. stepanyants , two - dimensional models for nonlinear voriticy waves in shear flows , stud . appl . math . , 100(1998 ) 1 - 32 . h. f. walker , p. ni , anderson acceleration for fixed - point iterations , siam j. numer . , 49 ( 2011 ) 1715 - 1735 . p. wynn , on a device for computing the @xmath274 transformation , mathematical tables and other aids to computation , 10(1956 ) 91 - 96 . p. wynn , acceleration techniques in numerical analysis , with particular reference to problems in one independent variable , proceedings of the ifip congress 1962 , north - holland , 149 - 156 . p. wynn , general purpose vector epsilon algorithm algol procedures , numer . math . , 6(1964 ) 22 - 36 . p. wynn , on the convergence and stability of the epsilon algorithm , siam j. numer . , 3(1966 ) 91 - 122 . j. yang , _ nonlinear waves in integrable and nonintegrable systems _ , siam , philadelphia , 2010 . j. yang , classification of solitary wave bifurcations in generalized nonlinear schrdinger equation , stud . 129(2012 ) 133162 . c. yang , j. c. meza , b. lee , l. w. wang , kssolv- a matlab toolbox for solving the kohn - sham equations , acm trans . software , 36 ( 2009 ) 1 - 35 .

the bcs to bec crossover problem of a fermi gas with an adjustable attractive interaction has been investigated theoretically for quite some time @xcite . for low temperatures the gas is superfluid and , in the case of s - wave interactions , it exhibits a smooth crossover from the well known bcs regime of weakly bound cooper pairs to the bec regime of tightly bound bosonic dimers with a residual repulsive interaction @xcite . in recent years , this crossover has been realized experimentally using ultracold fermi gases in optical traps , where the interaction can be tuned using feshbach resonances @xcite . in the experimentally relevant case of so called broad feshbach resonances , which is in principle always realized in the dilute limit , the physical properties of the homogeneous gas at equal densities for both spin components are described by only two dimensionless parameters : the interaction strength @xmath3 and the temperature @xmath4 . here , @xmath5 is the @xmath6-wave scattering length which fully characterizes interactions in the dilute , ultracold limit , while the scales for length and energy are determined by the fermi wave number @xmath7 and the fermi energy @xmath8 , respectively , where @xmath9 is the particle density . a particularly interesting regime is located near the feshbach resonance , where the scattering length @xmath5 is infinite . at this point and , more generally , in the so - called _ unitary regime _ where @xmath10 , the dimensionless interaction parameter @xmath11 disappears from the problem . all thermodynamic quantities are therefore universal functions of the dimensionless temperature @xmath4 @xcite . on a microscopic level , the unitary gas exhibits a particular kind of scale invariance , similar to a gas with purely inverse square two - particle interactions @xcite . more generally , as shown by nikolic and sachdev @xcite , universality is not restricted to the unitary regime . it is tied to the fact that the unitary balanced gas at zero density is an unstable fixpoint with only three relevant perturbations . since there is no small expansion parameter , the unitary regime is the most challenging one from a theoretical point of view . in addition , it is in fact precisely this regime which is accessible experimentally ( see e.g. the recent review articles by ketterle and zwierlein @xcite , by bloch _ et al . _ @xcite and by giorgini _ et al . _ @xcite ) . in a recent paper @xcite , we have presented a field theoretic approach for the thermodynamics of the bcs to bec crossover , which is based on the formalism developed by luttinger - ward @xcite and dedominicis - martin @xcite . in the following , this approach for the homogeneous gas is applied to calculate the thermodynamic properties of the trapped fermi gas , using a local density approximation . we compare our results with a recent experiment by luo _ et al . _ @xcite and with recent theories @xcite . in particular , we provide results for the entropy as a function of temperature , which allows to do reliable thermometry for the trapped unitary gas and also gives a precise value for the critical temperature and the associated entropy per particle . in addition , we show that starting well in the superfluid regime , much lower temperatures and entropies can be reached by removing atoms from the edge of the cloud , which carry most of the entropy . in our previous paper we have calculated the thermodynamic quantities for the homogeneous system . at a given particle density @xmath9 , these are the internal energy per particle @xmath12 and the entropy per particle @xmath13 . the fermi wave number @xmath7 and the fermi energy @xmath8 can be used as scale factors in order to make the thermodynamic quantities dimensionless . in the presence of an external confining potential @xmath14 , the particle density @xmath15 is non - uniform . within a local density approximation ( lda ) , thermodynamic quantities like pressure @xmath16 , chemical potential @xmath17 , entropy per particle @xmath18 or the internal energy per particle @xmath19 are then also spatially varying , being determined by the corresponding equilibrium values in the uniform system evaluated at the local density @xmath15 . the local density approximation neglects the dependence of thermodynamic properties on density gradients . at zero temperature , it is essentially a zeroth order semiclassical approximation @xcite . it is valid as long as the local fermi wave number @xmath20 times the oscillator length @xmath21 defined by the characteristic frequency @xmath22 of the confining potential is much larger than one . except near the edge of the cloud , where the density approaches zero , this condition is well justified for most of the experiments because typical fermi energies @xmath23 are of the order of several khz , while the trapping frequencies are around @xmath24 or smaller . specifically , near the trap center @xmath25 , where @xmath26 is the total particle number in a trap . as will be shown below , the finite size corrections to the ground state energy are of relative order @xmath27 in a harmonic trap . they are therefore negligible , at least for global observables , for the typical particle numbers in experiment , where @xmath28 @xcite . this conclusion is also supported by a recent comparison of lda with a numerical solution of the bogoliubov - degennes equations @xcite . within the lda , the global thermodynamic quantities of the trapped fermi gas are obtained by integrating over the whole space . specifically , we define @xmath29 \ , \qquad \label{eq : total_energy } \\ s & = & \int d^3r\ n(\mathbf{r})\ s(\mathbf{r } ) \ , \label{eq : entropy}\end{aligned}\ ] ] as the particle number , the potential energy , the internal energy , the total energy , and the total entropy . here the particle density acts like a distribution function to define averages over the trap . in an optical trap where the laser intensity profiles are gaussian functions , the confining potential @xmath14 is given by an anisotropic gaussian function . close to the center of the trap this potential can be approximated by an anisotropic harmonic function @xmath30 where @xmath31 is the mass of the atoms and @xmath32 , @xmath33 , @xmath34 are the harmonic oscillator frequencies in the three spatial directions . in the following , we use the harmonic potential and neglect the anharmonic terms . since in lda @xmath17 is the chemical potential relative to the potential @xmath14 , in thermal equilibrium the total chemical potential @xmath35 is constant which implies the condition @xmath36 this equation together with the requirement of a constant temperature @xmath37 determines the spatial dependence of all local thermodynamic quantities . the particle density @xmath15 implies a local fermi wave number @xmath20 and a local fermi energy @xmath38 . as a consequence the dimensionless parameters @xmath39 and @xmath40 depend on the local position in the trap . it is convenient to define the weighted radial coordinate @xmath41 by @xmath42 where @xmath43 is the average harmonic frequency . with this definition the confining potential acquires the simple form @xmath44 . as a consequence , the anisotropy of the trap becomes irrelevant because all local quantities @xmath45 etc . only depend on the weighted radial coordinate @xmath41 . rewriting the space integrals in eqs . - in terms of @xmath41 , this applies also to all thermodynamic quantities derived from the local density . a convenient measure for the overall length scale in a trap is provided by the thomas - fermi radius @xmath46 of the confined _ non - interacting _ fermi gas at zero temperature with a given total particle number @xmath26 . similarly , as a characteristic scale for the energy we define the corresponding fermi energy @xmath47 of the non - interacting gas . the thermodynamic quantities can be made dimensionless by considering the ratios @xmath48 , @xmath49 , and @xmath50 . these ratios depend on the space coordinate @xmath51 only implicitly via the dimensionless parameters @xmath39 and @xmath40 . a particular case is the unitary gas at zero temperature , where both parameters vanish identically @xmath52 , @xmath53 . using standard thermodynamic relations , it is straightforward to show that all thermodynamic quantities can be expressed in terms of a _ single _ dimensionless parameter , the so - called bertsch parameter @xmath54 @xcite in the form @xmath55 these relations hold both for an ideal fermi gas , where @xmath56 and also at the unitarity point , where @xmath54 has a nontrivial value @xcite . in our previous work @xcite , we have calculated both ratios independently and obtained @xmath57 and @xmath58 . the first ratio implies a bertsch parameter @xmath59 , while the second ratio implies @xmath60 . these two values differ by about @xmath61 . consequently , the relation @xmath62 , which is valid both for an ideal and a unitary fermi gas , is satisfied up to an error of @xmath61 . inserting the harmonic potential into eq . and using the weighted radial coordinate , we obtain the chemical potential @xmath63 \label{eq : chemical_potential_2}\ ] ] where @xmath64 is the thomas fermi radius of the unitary gas at zero temperature . the dimensionless ratios imply similar functional forms for the other quantities , e.g. @xmath65^{1/2}$ ] for the local fermi wavevector . these expressions are valid only for @xmath66 , because the particle density @xmath15 is nonzero only in this case and zero otherwise . evidently , the thomas - fermi radius @xmath64 of the unitary fermi gas is the only parameter which determines the spatial dependence of the thermodynamic quantities . it is related to the thomas - fermi radius @xmath67 of the ideal fermi gas by the bertsch parameter via @xmath68^{1/4 } = \xi^{1/4 } \ . \label{eq : thomas_fermi_radius}\ ] ] using the dimensionless ratio @xmath57 of our numerical calculations @xcite we obtain the result @xmath69 . this value agrees very well with the most recent field theoretic result @xmath70 for the bertsch parameter , obtained from a borel resummation of an expansion around the upper critical dimension four , carried out to three loop order @xcite . it is somewhat smaller , however , than the result @xmath71 found by using the bertsch parameter @xmath72 that follows from variational monte carlo calculations @xcite or from a theory that includes the gaussian fluctuations around the bcs mean field , extended to arbitrary coupling , where @xmath73 @xcite . next we insert the local internal energy per particle @xmath19 , the potential @xmath74 , and the local particle density @xmath15 into eqs . - in order to calculate the energies of the unitary fermi gas in the harmonic trap . we thus obtain the dimensionless ratios @xmath75^{1/2 } \nonumber\\ & = & 0.449 \ , \label{eq : potential_energy_3 } \\ 2\ , \frac{u}{n e_f } & = & \frac{3}{4}\ , \frac{u(\mathbf{0})}{e_f } = \frac{3}{4}\ , \frac{u(\mathbf{0})}{\varepsilon_f(\mathbf{0 } ) } \left [ \frac{\mu(\mathbf{0})}{\varepsilon_f(\mathbf{0 } ) } \right]^{-1/2 } \nonumber\\ & = & 0.440 \ , \label{eq : internal_energy_3 } \\ \frac{e}{n e_f } & = & \frac{u+e_\mathrm{pot}}{n e_f } = 0.444 \ . \label{eq : total_energy_3}\end{aligned}\ ] ] the explicit numbers are obtained by inserting the ratios @xmath57 and @xmath58 of our numerical calculation @xcite . for a harmonic potential @xmath14 it is well known that the internal energy @xmath76 and the potential energy @xmath77 are related to each other by the virial theorem @xmath78 . as shown by thomas _ et al . _ @xcite , this theorem also holds for the interacting fermi gas in the unitary regime . as a consequence , the results of eqs . - should be equal . by using the dimensionless ratios of the homogeneous system we can express the results in terms of the bertsch parameter @xmath54 according to @xmath79 in practice , our resulting numbers differ by about @xmath61 . this difference is related to the fact , that the exact relation @xmath62 for the unitary gas is satisfied only with an error of @xmath61 in our theory . the result for the ground state energy in the trap is based on using lda and provides the exact leading order contribution in the limit @xmath80 . the question of how large the subleading corrections to this result are has been addressed by son and wingate @xcite . using a gradient expansion of the effective field theory describing the low energy physics of the fermionic superfluid at unitarity , they have determined the @xmath81-corrections to the density response , which is equal to the uniform compressibility @xmath82 at @xmath83 . these corrections give rise to an additional contribution @xcite @xmath84 \ . \label{eq : energies_5}\ ] ] to the ground state energy , which contains the two dimensionless coefficients @xmath85 and @xmath86 which appear beyond the leading coefficient @xmath54 of the uniform system in an expansion up to second order in gradients . the leading correction to lda is thus of relative order @xmath27 ( formally it is @xmath87 ) and describes a _ curvature _ instead of the naively expected _ surface _ correction , which would scale like @xmath88 . the absence of a surface correction also apppears for an ideal fermi gas and is a peculiarity of the harmonic confinement @xcite . the prefactor of the @xmath27 correction in has been determined from an expansion around dimension four by rupak and schfer @xcite and is @xmath89 for an isotropic trap . for experimentally relevant particle numbers @xmath28 , the beyond lda relative corrections to the ground - state energy are therefore only around @xmath90 and thus are clearly below the present experimental accuracy . the final results of this section are the thomas - fermi radius @xmath69 and the total energy @xmath91 of the unitary fermi gas at zero temperature in a harmonic trap , which depend only on the bertsch parameter @xmath59 of the homogeneous system by eqs . and . these results will be compared with experiments and other theories in the next section , where we discuss the situation at finite temperatures . for nonzero temperatures , the local thermodynamic quantities are space dependent via the scale factors @xmath20 , @xmath38 , and also via the dimensionless temperature @xmath40 . in the unitary regime the dimensionless interaction @xmath92 is constant . since the effective temperature increases towards the edge of the cloud , the local particle density @xmath15 does not follow the simple @xmath93^{3/2}$ ] law valid at @xmath2 and eq . must be solved numerically in order to obtain the detailed density profile @xmath45 as a function of the weighted radial coordinate @xmath41 . of the unitary fermi gas in a harmonic trap as a function of the radius @xmath41 for several temperatures @xmath94 ( black solid ) , @xmath95 ( brown short dashed ) , @xmath96 ( red long dashed ) , @xmath97 ( orange short dashed ) , @xmath98 ( green long dashed ) , @xmath99 ( blue dot - dashed ) , @xmath100 ( magenta doubledot - dashed ) , and @xmath101 ( turquoise dot - doubledashed ) . the length scale is @xmath102.,scaledwidth=44.0% ] using the results of our previous numerical calculations @xcite for the homogeneous system , we obtain the density profiles @xmath103 for several values of the temperature @xmath37 which are shown in fig . [ fig : density_profiles ] . the black solid line is the density profile for zero temperature . it is nearly identical to the expected zero temperature thomas - fermi profile . differences are due to the limited numerical accuracy , which are , however , much smaller than the thickness of the line . at zero temperature , the fermi gas is superfluid in the whole trap . the colored and dot - dashed lines represent the density profile for nonzero temperatures . for very low temperatures , changes occur only close to the surface of the cloud where @xmath104 . for this reason , the brown short dashed line is visible only for @xmath105 while otherwise it is nearly on top of the black solid line . the red long dashed line is still very close to the black solid line . at a position in space @xmath51 the fermi gas may be normal fluid or superfluid if the local dimensionless temperature @xmath106 is above or below the critical value @xmath107 . for the brown , red , orange , and green dashed lines there exists a respective weighted radius @xmath108 so that @xmath109 . in these cases the fermi gas is superfluid in the inner region of the trap where @xmath110 and normal fluid in the outer region where @xmath111 . for the blue , magenta , and turquoise dot - dashed lines the dimensionless temperature is @xmath112 for all positions in space @xmath51 so that the fermi gas is normal fluid in the whole trap . since @xmath113 has its lowest value for @xmath114 , the dimensionless temperature at the center of the trap @xmath115 determines the superfluid transition of the confined fermi gas . for @xmath116 the fermi gas is completely normal fluid , while for @xmath117 there exists a superfluid region close to the center . in a homogeneous gas , the normal to superfluid transition is a continuous phase transition of the 3d xy type along the complete bcs to bec crossover , because the broken symmetry is that associated with a complex scalar order parameter . by contrast , our approach @xcite predicts a weak first - order superfluid transition because the superfluid phase of the luttinger - ward theory does not smoothly connect with the normal - fluid phase at a single critical temperature @xmath118 . as a result , there are two slightly different critical temperatures @xmath119 and @xmath120 . the upper value @xmath119 is defined by the condition that for temperatures above this value the superfluid order parameter vanishes , while the lower value @xmath120 is defined by the temperature below which the normal - fluid phase is no longer stable . fortunately , the difference between both temperatures , which should vanish in an exact theory , is rather small over essentially the whole bcs to bec crossover . in particular , at unitarity , the upper and lower values for @xmath121 are @xmath122 and @xmath123 , which is within the present numerical uncertainties in the determination of the critical temperature of the unitary gas . indeed , our critical temperature agrees very well with the most precise calculations of @xmath121 so far by quantum monte carlo calculations , which give @xmath124 for the uniform gas at unitarity @xcite , a value that has been confirmed very recently @xcite . the existence of two different critical temperatures leads to a multivaluedness of thermodynamic quantities , which is an artifact of the first order nature of the superfluid transition within our theory . in order to avoid multivalued local density or entropy profiles , we have connected the normal and superfluid branches with a kink at the point , where they are closest , thus providing an optimal approximation to the exact continuous profiles in a theory which properly accounts for the continuous nature of the transition in the infinite system . this point is related to the upper value of the dimensionless critical temperature @xmath119 . hence , in fig . [ fig : density_profiles ] the brown , red , orange , and green dashed lines have kinks located at @xmath125 , @xmath126 , @xmath127 , @xmath128 , respectively . the inner , superfluid branches of these lines show a bulge in the center of the trap , which is appreciable , in particular for the green curve . the formation of a small bulge upon condensation , which first appears near the trap center , has indeed been observed experimentally @xcite , although the effect in axially integrated profiles is rather small . previous theoretical results for the density profile have been obtained by bulgac _ et al . _ @xcite within a monte - carlo approach . for zero temperature , their result is very close to our black solid line in fig . [ fig : density_profiles ] . they obtain the thomas - fermi radius @xmath129 @xcite . this is larger than our value @xmath130 because of the larger value @xmath131 of the bertsch parameter . for nonzero temperatures ( the same as those in fig . [ fig : density_profiles ] ) , bulgac _ et al . _ obtain single valued density profiles which agree qualitatively with our results but differ somewhat quantitatively . they also observe a superfluid bulge which , however , is smaller . of the unitary fermi gas in a harmonic trap as a function of the radius @xmath41 for several temperatures . the curves are related to those in fig . [ fig : density_profiles ] . an additional curve is included for @xmath132 ( black short dashed).,scaledwidth=44.0% ] a very interesting quantity is the local entropy per particle @xmath133 . from our numerical calculations @xcite of the homogeneous system , and within lda , the resulting profiles @xmath134 of the local entropy per _ volume _ are shown in fig . [ fig : entropy_profiles ] for several values of the temperature @xmath37 . the colors of the curves are related to those in fig . [ fig : density_profiles ] . for high temperatures where the fermi gas is completely normal fluid ( blue , magenta , and turquoise dot - dashed curves ) , the entropy density is distributed over the whole trap with a maximum at the center . for low temperatures ( orange , red , and brown dashed curves ) the center part is superfluid but the outer part is normal fluid . consequently , in these cases the entropy density is minimum in the center of the trap but maximum at a nonzero radius . for very low temperatures ( black dashed line ) , the main contribution of the entropy is located close to the surface of the atom cloud . the width of this surface layer decreases for decreasing temperature and eventually shrinks to zero in the zero temperature limit ( black solid line ) . in fig . [ fig : entropy_profiles ] we have again eliminated multivalued regions which are an artifact of the first - order nature of the transition in our theory for the homogeneous system by using the same criterion as in the density profiles of fig . [ fig : density_profiles ] . indeed , since the normal to superfluid transition is not associated with a latent heat , the local entropy density will be single valued . an interesting observation that is evident from fig . [ fig : entropy_profiles ] , is that sufficiently below the onset of superfluidity in a trapped fermi gas , most of the entropy is located in the normal region near the cloud edge . this observation suggests an efficient way to lower the entropy further by removing atoms in the boundary layer and simultaneously readjusting the trapping potential such that the now smaller system has a radius close to that beyond which atoms have been removed . this idea is - of course - in the same spirit than the standard evaporation cooling in the normal state @xcite , however it is much more efficient . indeed , consider starting with a dimensionless temperature @xmath135 ( red long dashed curve ) or - equivalently - an entropy per particle @xmath136 which is close to that reached in current experiments . removing about @xmath137 percent of the particles in the shell beyond @xmath138 , will lower the entropy per particle by a factor @xmath139 to @xmath140 and the dimensionless temperature by a factor @xmath141 to @xmath142 . in turn , for an initial temperature @xmath143 ( blue dot - dashed curve ) where the cloud is a normal gas , removing the same amount of particles will reduce the entropy only by a factor @xmath144 . removing atoms in the outer shell repeatedly thus provides an effective tool to reach very low entropies and temperatures . in practice , for atoms confined in an optical dipole trap , this may be achieved by lowering the depth of the optical trap , as was done e.g. previously in experiments realizing a condensate of fermionic dimers @xcite . from the local profiles @xmath15 , @xmath19 , and @xmath18 , it is straightforward to calculate the the total energy @xmath1 and the global entropy @xmath145 by evaluating the integrals in eqs . and . in particular , we may eliminate the local value @xmath146 of the dimensionless temperature in the trap center , to obtain the function @xmath147 . our result for the ultracold fermi gas in the unitary regime is shown in fig . [ fig : entropy ] as blue - green - red solid line . to distinguish the superfluid and the normal - fluid parts of the curve , the solid line is shown in blue or red color , respectively . apparently , this curve is continuous and there is no particular feature , which indicates the superfluid transition . in fact , since @xmath148 is just the inverse temperature , this behavior is expected not only in a trap , where none of the thermodynamic functions exhibits a singularity , but even in the homogeneous gas , because the superfluid transition is continuous . as mentioned before , however , our theory predicts a weak first - order transition . the solid line is therefore multivalued in the intervals @xmath149 and @xmath150 , which is indicated by the green section of the line . within these intervals the superfluid transition is located . in practice , evidently , the multivaluedness is so tiny that it can hardly be seen in fig . [ fig : entropy ] . in the multivalued region the blue , green , and red branch of the solid line therefore lie on top of teach other . bulgac _ et al . _ have obtained @xmath0 from their monte - carlo calculation @xcite , which , in fig . [ fig : entropy ] , is shown as orange dashed line . clearly , the agreement with our theory is nearly perfect for high temperatures well above the superfluid transition . however , for lower temperatures close to and below the transition , the results differ . the deviations are largest for zero temperature where @xmath151 . they are essentially due to the fact that the ground - state energy @xmath152 in both approaches differ . indeed , from eq . we obtain @xmath153 while the corresponding value of bulgac _ et al . _ @xmath154 is larger because of the larger value @xmath131 of the bertsch parameter . apart from the deviations in the limit of zero temperature , the theories also differ in their predictions of the behavior near the superfluid transition temperature . in particular , bulgac _ et al . _ @xcite have calculated the critical values @xmath155 and @xmath156 at the superfluid transition . these results are considerably larger than our predictions @xmath157 and @xmath158 , whose uncertainty is due to the multivaluedness near @xmath159 . as a result , the value of the critical temperature @xmath160 of the unitary fermi gas in a trap in the theory of ref . is considerably larger than our prediction @xmath161 . as a function of the total energy @xmath1 for the unitary fermi gas in a harmonic trap : present theory ( blue - green - red solid ) , nsr theory of hu _ et al . _ @xcite ( magenta dot - dashed ) monte - carlo simulation of bulgac _ et al . _ @xcite ( orange dashed ) , and experimental data of luo _ et al . _ @xcite ( green data points).,scaledwidth=44.0% ] now , as pointed out above , our result for the critical temperature of the uniform gas agrees rather well with the most precise numerical calculations of this quantity @xcite . in addition , it is also consistent with the recent calculations of bulgac _ et al . _ @xcite , which indicate that @xmath162 is less or equal to @xmath163 , again in very good agreement with our numbers . this underlines that our self - consistent , conserving theory of the bcs - bec crossover @xcite provides a quantitatively reliable description of the thermodynamics of a balanced fermi gas near unitarity , despite the problems with the first order nature of the transition in our approach and the absence of a small expansion parameter . within lda , which provides a rather accurate description in the relevant regime of particle numbers @xmath164 , our critical temperature @xmath165 for the trapped gas is therefore expected to be close to the exact result . this is consistent with a very recent analysis of the experimental data of ref . , which indicates a critical temperature very close to this value @xcite . previous , much higher values of the critical temperature for both the homogeneous or the trapped gas that were obtained in different extensions of the theory by nozires and schmitt - rink @xcite and also in numerical calculations @xcite are clearly ruled out . an alternative approach to the bcs to bec crossover problem has been developed by hu _ et al . _ @xcite for the homogeneous system and applied to the ultracold fermi gas in a trap @xcite . this theory is an extension of the approach by nozires and schmitt - rink ( nsr ) @xcite to the superfluid region at low temperatures . while the order parameter @xmath166 is determined by the standard mean - field gap equation , the chemical potential @xmath167 is calculated by a particle - density equation , which includes condensed and noncondensed bound pairs . the extended nsr approach is in fact a limiting case of our theory in which the full , self - consistently determined green functions are replaced by their zeroth order form obtained in bcs theory . et al . _ @xcite have calculated @xmath147 and obtain a result which is shown as magenta dot - dashed line in fig . [ fig : entropy ] . again , the agreement is very good for high temperatures in the normal fluid region while , however , for low temperatures in the superfluid region there are deviations . et al . _ @xcite obtain the ground - state energy @xmath168 which is related to the bertsch parameter @xmath73 . moreover , similar to the theory of bulgac _ et al . _ @xcite their critical temperature @xmath169 and entropy @xmath170 are considerably larger than our values . more recently , hu _ et al . _ @xcite have published a self - consistent result for the entropy @xmath0 which they call `` @xmath171 approximation '' . this method is equivalent to our approach for the homogeneous system @xcite and gives results that are nearly identical to the prediction of our present theory including the critical temperature @xmath172 . while in the experimental setup the correct trap potential is gaussian , all the curves shown in fig . [ fig : entropy ] are calculated for the harmonic potential . slight differences would occur in the high - temperature regime where the energies @xmath1 and the entropies @xmath145 are large . however , it is important that all curves are calculated for the same potential so that they all converge to a single line in the high - temperature limit . the entropy versus the total energy has been measured experimentally for ultracold @xmath173 atoms in an optical trap by luo _ et al . _ the data are shown as green points with horizontal and vertical error bars . clearly the data agree with all theories in the high - temperature regime . the slightly larger entropy values of the experimental data for large energies may be due to the fact that the experiment is performed for a gaussian potential while the theoretical curves are calculated for a harmonic potential . apparently , for low temperatures and low entropies the experimental data agree better with the theories of bulgac _ et al . _ @xcite and of hu _ et al . _ @xcite than with our theory . in particular , an extrapolation to zero entropy gives a ground - state energy @xmath174 , considerably higher than our value @xmath175 . it is difficult , however , to quantify the error involved in such an extrapolation because the determination of the entropy from a comparison with its ideal fermi gas limit reached after an adiabatic ramp to magnetic fields @xmath176 far on the bcs side of the crossover becomes increasingly difficult as @xmath145 approaches zero . as pointed out above , a quite sensitive parameter which distinguishes previous theories from our present one is the critical temperature of the superfluid transition and the associated value of the entropy and energy . unfortunately , it is difficult to determine the critical temperature and thus also the corresponding value of the entropy from measurements of @xmath0 . based on the excellent agreement of our value for @xmath159 and the bertsch parameter @xmath54 with the most precise numerical or field - theoretical results and the fact that all thermodynamic relations are obeyed at the few percent level , it is likely that our present theory gives a reliable description of the thermodynamics of the trapped unitary gas , despite the fact that , superficially , the agreement with the data shown in fig . [ fig : entropy ] is not as good as those of previous theories . the most important parameter for thermodynamic properties is - of course - the temperature , which unfortunately can not be measured directly . in principle , this is possible from the density profile @xmath15 which changes as a function of temperature . for balanced gases , however , this method is not very reliable . indeed , for low temperatures , the density converges to the simple thomas - fermi profile . as shown in fig . [ fig : density_profiles ] , the deviations from such a profile at finite temperatures are extremely small for the inner part of the atom cloud ( see the red and brown low - temperature curves , which are nearly identical to the black zero - temperature curve ) . only close to the surface @xmath177 , small variations with the temperature are observed . therefore , the signal to noise ratio will be small in the interesting regime below @xmath178 . a quantity which is much more sensitive to temperature is the entropy density , shown in fig . [ fig : entropy_profiles ] . at low temperatures , it is peaked near the surface of the atom cloud . unfortunately , the local entropy density @xmath179 is not accessible experimentally . we therefore consider the total entropy @xmath145 , which has been measured by @xcite , as discussed above . in fig . [ fig : entropy - temperature ] , we plot the total entropy @xmath145 in a trap as a function for the temperature @xmath37 in units of the fermi temperature @xmath180 of the trapped non - interacting gas . as in fig . [ fig : entropy ] , we indicate the superfluid and the normal - fluid region by blue and red color of the curve , respectively . again , a tiny multivalued region is observed close to the superfluid transition which is indicated by the green section of the line . however , the three branches in the multivalued region are nearly on top of each other . consequently , the entropy @xmath181 is effectively continuous at the superfluid transition as expected . using the curve in fig . [ fig : entropy - temperature ] , the measurements of the entropy by luo _ et al . _ @xcite therefore allow to reliably infer the related temperatures @xmath37 . as a function of the temperature @xmath37 for the unitary fermi gas in a harmonic trap . the blue , red , and green sections represent the superfluid , normal - fluid and multivalued branches of the curve , respectively.,scaledwidth=44.0% ] alternatively , we may consider the excess internal energy density @xmath182 n(r)$ ] where @xmath183 is the ground - state internal energy per particle at zero temperature . the related profiles are similar to those shown in fig . [ fig : entropy_profiles ] . for low temperatures the excess internal energy density is peaked close to the surface of the atom cloud . integrating over the space we obtain the excess internal energy @xmath184 which is related to the excess total energy by the virial theorem @xmath185 . moreover , the total energy @xmath1 is related to the mean square radius of the atom cloud @xmath186 according to @xmath187 . since @xmath186 can be measured in a model - independent way @xcite , the total energy @xmath188 as a function of the temperature @xmath37 provides an alternative method to determine the temperature of the interacting fermi gas . we obtain a curve which is similar like fig . [ fig : entropy - temperature ] . however , the major drawback of this method is that it requires knowledge of the ground state energy @xmath152 , which must be determined with sufficient accuracy . as evident from fig . [ fig : entropy ] , there are significant discrepancies in the ground state energy @xmath152 for different experiments and theories . in the superfluid regime at very low temperatures the fermionic quasiparticles are frozen out and do not contribute to thermodynamic quantities because they have an energy gap which is related to the binding energy of the cooper pairs . however , the spontaneous symmetry breaking implies a gapless goldstone mode which is the bogoliubov - anderson mode . this mode propagates with a constant velocity @xmath189 like phonons . for this reason , the low temperature behavior of the entropy density and the internal energy density is ruled by the well known stefan - boltzmann formulas @xmath190 n(\mathbf{r } ) & = & 2\ , \frac{\sigma(\mathbf{r})}{c(\mathbf{r})}\ , t^4 \label{eq : stefan_boltzmann_energy}\end{aligned}\ ] ] for phonons with one polarization degree of freedom where @xmath191 ^ 2)$ ] is the stefan - boltzmann factor . since the sound velocity @xmath192 depends on the local particle density , @xmath193 and @xmath192 are space - dependent parameters . eliminating the temperature we obtain a relation between the local entropy per particle @xmath18 and the local internal energy per particle @xmath19 , which can be written in the form @xmath194^{3/4 } \ . \label{eq : local_asymptotic_formula}\end{aligned}\ ] ] here @xmath195 is the local fermi velocity , and @xmath196 is the ground - state energy per particle . the ratio @xmath197 is constant and related to the bertsch parameter @xmath59 . in our previous publication @xcite we have calculated all thermodynamic quantities for the homogeneous system . hence , also the function @xmath198 is available , so that we can check the asymptotic formula for the homogeneous system . it turns out that the resulting exponent at low energy is indeed equal to @xmath199 . however the amplitude , which is determined fully by the sound velocity @xmath189 , gives @xmath200 for the unitary fermi gas . this is about a factor of two larger than the expected value @xmath201 from the bertsch parameter . this discrepancy indicates , that the accuracy of our theory at very low temperatures is not sufficient to extract a reliable value of the sound velocity from the entropy . in the trapped case , unfortunately , asymptotic formulas like - do not hold for the total entropy @xmath145 and the total energy @xmath1 . to see this , we integrate eqs . and over the whole space . using eqs . and we obtain well defined results @xmath145 and @xmath202 for the left hand sides , respectively . however , while the temperature @xmath37 is constant , the space dependence on the right - hand sides arise from the factor @xmath203^{-3 } \sim [ v_f(\mathbf{r})]^{-3 } \sim [ n(\mathbf{r})]^{-1}$ ] . this factor is minimum at the center of the trap but maximum at the surface of the atom cloud . thus , the main contribution of the integral arises from the surface of the atom cloud ( see also fig . [ fig : entropy_profiles ] ) . here the fermi gas is a normal fluid , and the dimensionless temperature @xmath204 is large , so that the stefan - boltzmann formulas are not applicable . consequently , an asymptotic formula like can not be derived for @xmath145 and @xmath205 of the whole trap . of the function @xmath0 where @xmath1 is two times the potential energy ( green dashed ) , two times the internal energy ( red dot - dashed ) , and one times the total energy ( blue solid ) . here @xmath152 is the ground - state energy for @xmath2 and @xmath151.,scaledwidth=44.0% ] empirically , it has been found in the experiments by luo _ _ @xcite that the total entropy at low energies varies with an effective power law @xmath206^b \label{eq : asymptotic_formula}\ ] ] with an exponent @xmath207 . in order to check whether such a behavior is consistent with a microscopic theory , we determine the exponent @xmath208 by logarithmic differentiation of the blue - green - red solid curve @xmath147 in fig . [ fig : entropy ] . the result is shown as blue solid curve in fig . [ fig : exponents ] and confirms that @xmath0 in a trap indeed exhibits a power law behavior in the regime near the ground state . we have adjusted the ground - state energy @xmath152 by fine tuning in order to obtain a well defined exponent in the limit @xmath209 . the resulting exponent @xmath210 is surprisingly close to the value @xmath211 of the local asymptotic formula but differs from the value @xmath212 inferred from the experimental fit . in fig . [ fig : exponents ] the superfluid transition is located at the position @xmath213=-0.65 $ ] . the left part of the blue solid curve corresponds to the superfluid region . here the exponent is nearly constant up to the superfluid transition . even though the stefan - boltzmann formulas are not valid , the exponent @xmath210 is remarkably close to the theoretical value @xmath211 . the right part of the blue solid curve corresponds to the normal fluid region . here the logarithmic derivative of the entropy with respect to energy decreases monotonically with increasing energy @xmath1 and thus no well defined exponent can be attributed to the high energy part of the curve . the virial theorem implies the energy relations @xmath214 . in order to check the validity of the virial theorem we consider the entropy functions @xmath215 and @xmath216 and calculate the related exponents @xmath217 and @xmath218 , which are shown in fig . [ fig : exponents ] as green dashed line and as red dot - dashed line , respectively . these lines should be compared with the blue solid line , which represents the exponent @xmath219 . in the normal - fluid region the virial theorem is well satisfied , because the right parts of the curves are lying nearly on top of each other where the small deviations are numerical errors . in the superfluid region the virial theorem is satisfied only approximately because of the modification of the theory in order to have a gapless bogoliubov - anderson mode ( see sec . ii.j in ref . ) . as a consequence , the left parts of the curves deviate from each other . we find three different values for the , exponents which are @xmath220 , @xmath221 , and @xmath222 . these results are related to the three different ground - state energies , , and , respectively . based on our previous results for the bcs to bec crossover problem in a homogeneous gas @xcite , we have calculated density and entropy profiles in a trap within a local density approximation . in addition , we have determined the total entropy @xmath145 and energy @xmath1 in the unitary regime and have compared our results with both experiment and recent theories . for temperatures above the superfluid transition temperature , very good agreement is obtained . however , the value of the critical temperature and the behavior at very low temperatures differ appreciably from the experimental estimates and their theoretical analysis in earlier work . first of all , our value for the bertsch parameter @xmath54 and thus the ground - state energy @xmath152 is about @xmath223 smaller than the results obtained from variational monte carlo calculations or from gaussian fluctuation theories around the bcs ansatz for the ground state . while our value @xmath224 agrees well with the most precise results so far obtained from an @xmath225 expansion @xcite , it differs from those obtained from variational monte carlo calculations @xcite , or from those including gaussian fluctuations around the bcs mean field @xcite , where @xmath72 or @xmath73 , respectively . given the uncertainty in determining @xmath54 experimentally ( which requires an extrapolation to zero temperature ) it is clearly important for theory to make precise predictions for @xmath54 which do not rely on approximations that are apparently limiting all present results . in view of the fundamental importance of this parameter in the context of strongly interacting fermi gases , progress here would be highly desirable . as a second point , our values for both the critical temperature and value of entropy at @xmath159 are appreciably lower than those obtained in the theories of bulgac _ et al . _ @xcite and of hu _ et al . _ @xcite and also those inferred from the original analysis of the experimental data @xcite . now , as is evident from fig . [ fig : entropy ] , a measurement of the function @xmath0 does not provide a sensitive measure of the critical temperature . quantitative results for the critical temperature of the unitary gas have been obtained by shin _ et al . _ they rely on using gases with a finite imbalance @xmath226 , which have always a fully polarized outer shell in a trap . since a single species ultracold fermi gas is noninteracting , its temperature can be reliably determined from cloud profiles . an extrapolation back to zero imbalance gives a critical temperature at unitarity which is close to the value predicted both in our theory and in monte - carlo calculations by burovski _ et al . _ @xcite and bulgac _ et al . _ as pointed out in section iv , there is now evidence that our result @xmath227 for the critical temperature of the unitary gas in a trap is rather precise . together with the entropy - temperature curve shown in fig . [ fig : entropy - temperature ] , this would allow doing precise thermometry for balanced gases , that has not been possible so far . finally , we have shown that for temperatures of order @xmath228 which have been reached experimentally @xcite , an efficient way of further cooling the gas is possible by removing the high entropy outer part of the atomic cloud and readjusting the confining potential . this method is similar in spirit than the standard evaporative cooling idea , but potentially much more efficient . it might open the avenue to reach regimes in which the entropy per particle is much less than @xmath229 , a necessary condition for realizing many of the nontrivially ordered states that are in principle accessible with ultracold fermions @xcite . we would like to thank a. bulgac , p.d . drummond , j.e . drut and j. thomas for useful discussions and for making their results available for comparison with our theory . w.z . is grateful for the hospitality as a visitor at the mit - harvard center for ultracold atoms and for a number of helpful discussions with m. zwierlein . this work was supported by the dfg forschergruppe 801 on ultracold quantum gases . w. ketterle and m.w . zwierlein , in : _ proceedings of the international school of physics `` enrico fermi '' , course clxiv , varenna , 20 - 30 june 2006 _ , edited by m. inguscio , w. ketterle , and c. salomon ( ios press , amsterdam 2008 ) . the possibility to treat the unitary fermi gas problem by a systematic expansion around the upper and lower critical dimensions @xmath230 and @xmath231 was pointed out originally by z. nussinov and s. nussinov , phys . a * 74 * , 053622 ( 2006 ) [ arxiv : cond - mat/0410597 ] , and has been carried out by y. nishida and d. t. son , phys . lett . * 97 * , 050403 ( 2006 ) . w. ketterle and n.j . van druten , in _ advances in atomic , molecular and optical physics _ * 36 * ( 1997 ) , eds . b. bederson and h. walther . for an introduction see also chapter vi of the lectures notes by j. dalibard available from http://www.phys.ens.fr/@xmath232dalibard/notes_de_cours/ dea_atomes_froids.pdf h. hu , x.j . liu , and p.d . drummond , phys . a * 77 * , 061605(r ) ( 2008 ) .

this paper is based on three lectures i gave at the conference on `` renormalization and galois theories '' that was held in luminy , at the centre international de rencontres mathmatiques ( cirm ) , on march 2006 . the purpose of these talks was to give an elementary overview on classical motives ( pure motives ) and to survey on some of the recent developments of this theory in noncommutative geometry , especially following the introduction of the notion of an _ endomotive_. it is likely to expect that the reader acquainted with the literature on motives theory will not fail to notice the allusion , in this title , to the paper @xcite in which p. deligne states that in spite of the lack of essential progresses on the problem of constructing `` relevant '' algebraic cycles , the techniques supplied by the theory of motives remain a powerful tool in algebraic geometry and arithmetic . the assertion on the lack of relevant progresses on algebraic cycles seems , unfortunately , still to apply at the present time , fifteen years after deligne wrote his paper . despite the general failure of testing the standard conjectures , it is also true that in these recent years the knowledge on motives has been substantially improved by several new results and also by some unexpected developments . motives were introduced by a. grothendieck with the aim to supply an intrinsic explanation for the analogies occurring among various cohomological theories in algebraic geometry . they are expected to play the role of a universal cohomological theory by also furnishing a linearization of the theory of algebraic varieties and in the original understanding they were expected to provide the correct framework for a successful approach to the weil s conjectures on the zeta - function of a variety over a finite field . even though the weil s conjectures have been proved by deligne without appealing to the theory of motives , an enlarged and in part still conjectural theory of mixed motives has in the meanwhile proved its usefulness in explaining conceptually , some intriguing phenomena arising in several areas of pure mathematics , such as hodge theory , @xmath0-theory , algebraic cycles , polylogarithms , @xmath1-functions , galois representations etc . very recently , some new developments of the theory of motives to number - theory and quantum field theory have been found or are about to be developed , with the support of techniques supplied by noncommutative geometry and the theory of operator algebras . in number - theory , a conceptual understanding of the main result of @xcite on the interpretation proposed by a. connes of the weil explicit formulae as a lefschetz trace formula over the noncommutative space of adle classes , requires the introduction of a generalized category of motives inclusive of spaces which are highly singular from a classical viewpoint . the problem of finding a suitable enlargement of the category of ( smooth projective ) algebraic varieties is combined with the even more compelling one of the definition of a generalized notion of correspondences . several questions arise already when one considers special types of zero - dimensional noncommutative spaces , such as the space underlying the quantum statistical dynamical system defined by j. b. bost and connes in @xcite ( the bc - system ) . this space is a simplified version of the adles class space of @xcite and it encodes in its group of symmetries , the arithmetic of the maximal abelian extension of @xmath2 . in this paper i give an overview on the theory of endomotives ( algebraic and analytic ) . this theory has been originally developed in the joint paper @xcite with a. connes and m. marcolli and has been applied already in our subsequent work @xcite . the category of endomotives is the minimal one that makes it possible to understand conceptually the role played by the absolute galois group in several dynamical systems that have been recently introduced in noncommutative geometry as generalizations of the bc - system , which was our motivating and prototype example . the category of endomotives is a natural enlargement of the category of artin motives : the objects are noncommutative spaces defined by semigroup actions on projective limits of artin motives . the morphisms generalize the notion of algebraic correspondences and are defined by means of tale groupoids to account for the presence of the semigroup actions . endomotives carry a natural galois action which is inherited from the artin motives and they have both an algebraic and an analytic description . the latter is particularly useful as it provides the data of a quantum statistical dynamical system , via the implementation of a _ canonical time evolution _ ( a one - parameter family of automorphisms ) which is associated by the theory of m. tomita ( _ cf . _ @xcite ) to an initial state ( probability measure ) assigned on an analytic endomotive . this is the crucial new development supplied by the theory of operator - algebras to a noncommutative @xmath3-algebra and in particular to the algebra of the bc - system . the implication in number - theory is striking : the time evolution implements on the dual system a scaling action which combines with the action of the galois group to determine on the cyclic homology of a suitable noncommutative motive associated to the original endomotive , a characteristic zero analog of the action of the weil group on the tale cohomology of an algebraic variety . when these techniques are applied to the endomotive of the bc - system or to the endomotive of the adles class space , the main implication is the spectral realization of the zeroes of the corresponding @xmath1-functions . these results supply a first answer to the question i raised in the title of this paper ( a quoi servent les endomotifs ? ) . an open and interesting problem is connected to the definition of a higher dimensional theory of noncommutative motives and in particular the introduction of a theory of noncommutative elliptic motives and modular forms . a related problem is of course connected to the definition of a higher dimensional theory of geometric correspondences . the comparison between algebraic correspondences for motives and geometric correspondences for noncommutative spaces is particularly easy in the zero - dimensional case , because the equivalence relations play no role . in noncommutative geometry , algebraic cycles are naturally replaced by bi - modules , or by classes in equivariant @xmath4-theory . naturally , the original problem of finding `` interesting '' cycles pops - up again in this topological framework : a satisfactory solution to this question seems to be one of the main steps to undertake for a further development of these ideas . the theory of motives in algebraic geometry was established by a. grothendieck in the 1960s : 1963 - 69 ( _ cf . _ @xcite,@xcite ) . the foundations are documented in the unpublished manuscript @xcite and were discussed in a seminar at the institut des hautes tudes scientifiques , in 1967 . this theory was conceived as a fundamental machine to develop grothendieck s `` long - run program '' focused on the theme of the connections between geometry and arithmetic . at the heart of the philosophy of motives sit grothendieck s speculations on the existence of a universal cohomological theory for algebraic varieties defined over a base field @xmath5 and taking values into an abelian , tensor category . the study of this problem originated as the consequence of a general dissatisfaction connected to an extensive use of topological methods in algebraic geometry , with the result of producing several but insufficiently related cohomological theories ( betti , de - rham , tale etc . ) . the typical example is furnished by a family of homomorphisms @xmath6 connecting the groups of tale cohomology of two ( smooth , projective ) varieties , as the prime number @xmath7 varies , which are not connected , in general , by any sort of ( canonical ) relation . the definition of a contravariant functor ( functor of motivic cohomology ) @xmath8 from the category @xmath9 of projective , smooth , irreducible algebraic varieties over @xmath5 to a semi - simple abelian category of pure motives @xmath10 is also tied up with the definition of a universal cohomological theory through which every other classical , cohomology @xmath11 ( here understood as contravariant functor ) should factor by means of the introduction of a fiber functor ( realization @xmath12-functor ) @xmath13 connecting @xmath10 to the abelian category of ( graded ) vector - spaces over @xmath2 v_k & ^h^ & grv ect _ + & _ h & ^ + & & m_k(v_k ) following grothendieck s original viewpoint , the functor @xmath14 should implement the sought for mechanism of compatibilities ( in tale cohomology ) and at the same time it should also describe a universal linearization of the theory of algebraic varieties . the definition of the category @xmath10 arised from a classical construction in algebraic geometry which is based on the idea of extending the collection of algebraic morphisms in @xmath9 by including the ( algebraic ) _ correspondences_. a correspondence between two objects @xmath15 and @xmath16 in @xmath9 is a multi - valued map which connects them . an algebraic correspondence is defined by means of an algebraic cycle in the cartesian product @xmath17 . the concept of ( algebraic ) correspondence in geometry is much older than that of a motive : it is in fact already present in several works of the italian school in algebraic geometry ( _ cf . _ severi s theory of correspondences on algebraic curves ) . grothendieck s new intuition was that the whole philosophy of motives is regulated by the theory of ( algebraic ) correspondences : `` _ ... jappelle motif sur @xmath5 quelque chose comme un groupe de cohomologie @xmath7-adique dun schema algbrique sur @xmath5 , mais considre comme indpendant de @xmath7 , et avec sa structure entire , ou disons pour linstant sur @xmath2 , dduite de la thorie des cycles algbriques ... _ '' ( _ cf . _ @xcite , lettre 16.8.1964 ) . motives were envisioned with the hope to explain the intrinsic relations between integrals of algebraic functions in one or more complex variables . their ultimate goal was to supply a machine that would guarantee a generalization of the main results of galois theory to systems of polynomials equations in several variables . here , we refer in particular to a higher - dimensional analog of the well - known result which describes the linearization of the galois - grothendieck correspondence for the category @xmath18 of tale , finite @xmath5-schemes @xmath19 by means of the equivalence between the category of artin motives and that of the representations of the absolute galois group . in the following section we shall describe how the fundamental notions of the theory of motives arose from the study of several classical problems in geometry and arithmetic . a classical problem in algebraic geometry is that of computing the solutions of a finite set of polynomial equations @xmath20 with coefficients in a finite field @xmath21 . this study is naturally formalized by introducing the generating series @xmath22 which is associated to the algebraic variety @xmath23 that is defined as the set of the common zeroes of the polynomials @xmath24 . under the assumption that @xmath15 is smooth and projective , the series encodes the complete information on the number of the rational points of the algebraic variety , through the coefficients @xmath25 . the integers @xmath26 supply the cardinality of the set of the rational points of @xmath15 , computed in successive finite field extensions @xmath27 of the base field @xmath21 . intersection theory furnishes a general way to determine the number @xmath26 as intersection number of two algebraic cycles on the cartesian product @xmath28 : namely the diagonal @xmath29 and the graph @xmath30 of the @xmath31-th iterated composite of the frobenius morphism on the scheme @xmath32 : @xmath33 the frobenius endomorphism is in fact an interesting example of _ correspondence _ , perhaps the most interesting one , for algebraic varieties defined over finite fields . as a correspondence it induces a commutative diagram @xmath34 in tale cohomology . through the commutativity of the above diagram one gets a way to express the action of the induced homomorphism in cohomology , by means of the formula @xmath35 where @xmath36 denote the two projection maps . here , ` algebraic ' refers to the algebraic cycle @xmath37 that performs such a correspondence . for particularly simple algebraic varieties , such as projective spaces @xmath38 , the computation of the integers @xmath26 can be done by applying an elementary combinatorial argument based on the set - theoretical description of the space @xmath39 ( @xmath40 any field ) . this has the effect to produce the interesting description @xmath41 this decomposition of the set of the rational points of a projective space was certainly a first source of inspiration in the process of formalizing the foundations of the theory of motives . in fact , one is naturally led to wonder on the casuality of the decomposition , possibly ascribing such a result to the presence of a cellular decomposition on the projective space which induces a related break - up on the set of the rational points . remarkably , a. weil proved that a similar formula holds also in the more general case of a smooth , projective algebraic curve @xmath42 of genus @xmath43 . in this case one shows that @xmath44 these results suggest that and are the manifestation of a deep and intrinsic structure that governs the geometry of algebraic varieties . the development of the theory of motives has in fact shown to us that this structure reveals itself in several contexts : topologically , manifests its presence in the decomposition of the cohomology @xmath45 , whereas arithmetically it turns out that it is the same structure that controls the decomposition of the series as a rational function of @xmath46 : @xmath47 this is in fact a consequence of the description of the integers @xmath26 supplied by the lefschetz - grothendieck trace formula ( _ cf . _ @xcite ) @xmath48 originally , grothendieck proposed a general framework for a so called category of _ numerically effective motives _ @xmath49 over a field @xmath5 and with rational coefficients . this category is defined by enlarging the category @xmath9 of smooth , projective algebraic varieties over @xmath5 ( and algebraic morphisms ) by following the so - called procedure of pseudo - abelian envelope . this construction is performed in two steps : at first one enlarges the set of morphisms of @xmath9 by including ( rational ) algebraic correspondences of degree zero , modulo numerical equivalence , then one performs a pseudo - abelian envelope by formally including among the objects , kernels of idempotent morphisms . let us assume for simplicity that the algebraic varieties are irreducible ( the general case is then deduced from this , by additivity ) . for any given @xmath50 , one works with correspondences @xmath51 which are elements of codimension equal to @xmath52 in the rational graded algebra @xmath53 of algebraic cycles modulo _ numerical equivalence_. we recall that two algebraic cycles on an algebraic variety @xmath15 are said to be numerically equivalent @xmath54 , if @xmath55 for any algebraic cycle @xmath56 on @xmath15 . here , by @xmath57 we mean the degree of the algebraic cycle @xmath58 . the degree defines a homomorphism from the free abelian group of algebraic cycles @xmath59 to the integers . on the components @xmath60 , the map is defined as follows @xmath61 the symbol ` @xmath62 ' in refers to the intersection product structure on @xmath63 , which is well - defined under the assumption of proper intersection . if @xmath64 is proper ( _ i.e. _ @xmath65 ) , then intersection theory supplies the definition of an intersection cycle @xmath66 . moreover , the intersection product is commutative and associative whenever is defined . passing from the free abelian group @xmath63 to the quotient @xmath67 , modulo a suitable equivalence relation on cycles , allows one to use classical results of algebraic geometry ( so called moving lemmas ) which lead to the definition of a ring structure . one then defines intersection cycle classes in general , even when cycles do not intersect properly , by intersecting equivalent cycles which fulfill the required geometric property of proper intersection . it is natural to guess that the use of the numerical equivalence in the original definition of the category of motives was motivated by the study of classical constructions in enumerative geometry , such as for example the computation of the number of the rational points of an algebraic variety defined over a finite field . one of the main original goals was to show that for a suitable definition of an equivalence relation on algebraic cycles , the corresponding category of motives is _ semi - simple_. this means that the objects @xmath68 in the category decompose , following the rules of a _ theory of weights _ ( _ cf . _ section [ fundstructures ] ) , into direct _ factors _ @xmath69 , with @xmath70 simple ( _ i.e. _ indecomposable ) motives associated to smooth , projective algebraic varieties . the importance of achieving such a result is quite evident if one seeks , for example , to understand categorically the decomposition @xmath71 in cohomology , or if one wants to recognize the role of motives in the factorization of zeta - functions of algebraic varieties . grothendieck concentrated his efforts on the numerical equivalence relation which is the coarsest among the equivalence relations on algebraic cycles . so doing , he attacked the problem of the semi - simplicity of the category of motives from the easiest side . however , despite a promising departing point , the statement on the semi - simplicity escaped all his efforts . in fact , the result he was able to reach at that time was dependent on the assumption of the _ standard conjectures _ , two strong topological statements on algebraic cycles . the proof of the semi - simplicity of the category of motives for numerical equivalence ( the only equivalence relation producing this result ) was achieved only much later on in the development of the theory ( _ cf . _ @xcite ) . the proof found by u. jannsen uses a fairly elementary but ingenious idea which mysteriously eluded grothendieck s intuition as well as all the mental grasps of several mathematicians after him . by looking at the construction of ( pure ) motives in perspective , one immediately recognizes the predominant role played by the morphisms over the objects , in the category @xmath49 . this was certainly a great intuition of grothendieck . this idea led to a systematic study of the properties of algebraic cycles and their decomposition by means of _ algebraic projectors _ , that is algebraic cycles classes @xmath72 satisfying the property @xmath73 notice that in order to make sense of the notion of a projector and more in general , in order to define a law of composition ` @xmath74 ' on algebraic correspondences , one needs to use the ring structure on the graded algebra @xmath75 . the operation ` @xmath74 ' is defined as follows . let us assume for simplicity , that the algebraic varieties are connected ( the general case can be easily deduced from this ) . then , two algebraic correspondences @xmath76 ( of degree @xmath77 ) and @xmath78 ( of degree @xmath79 ) compose accordingly to the following rule ( bi - linear , associative ) @xmath80 @xmath81 in the particular case of projectors @xmath82 , one is restricted , in order to make a sense of the condition @xmath83 , to use only particular types of algebraic correspondences : namely those of degree zero . these are the elements of the abelian group @xmath84 . the objects of the category @xmath49 are then pairs @xmath85 , with @xmath86 and @xmath87 a projector . this way , one attains the notion of a @xmath2-linear , pseudo - abelian , monoidal category ( the @xmath12-monoidal structure is deduced from the cartesian product of algebraic varieties ) , together with the definition of a contravariant functor @xmath88 here @xmath89 denotes the motive associated to @xmath15 and @xmath90 means the trivial ( _ i.e. _ identity ) projector associated to the diagonal @xmath29 . more in general , @xmath85 refers to the motive @xmath91 that is cut - off on @xmath92 by the ( range of the ) projector @xmath82 . notice that images of projectors are formally included among the objects of @xmath49 , by the procedure of the pseudo - abelian envelope . the cut - off performed by a projector @xmath87 on the space determines a corresponding operation in cohomology ( for any classical weil theory ) , by singling out the sub - vector space @xmath93 . the category @xmath49 has two important basic objects : @xmath94 and @xmath95 . @xmath96 is the _ unit motive _ @xmath97 this is defined by the zero - dimensional algebraic variety associated to a point , whereas @xmath98 is the so - called _ lefschetz motive_. this motive determines , jointly with @xmath96 , a decomposition of the motive associated to the projective line @xmath99 @xmath100 one can show that the algebraic cycles @xmath101 and @xmath102 on @xmath103 do not depend on the choice of the rational point @xmath104 and that their sum is a cycle equivalent to the diagonal . this fact implies that the decomposition is canonical . more in general , it follows from the knneth decomposition of the diagonal @xmath105 in @xmath106 by algebraic cycles @xmath107 , ( _ cf . _ @xcite and @xcite for the details ) that the motive of a projective space @xmath108 decomposes into pieces ( simple motives ) @xmath109 where @xmath110 , @xmath111 . it is precisely this decomposition which implies the decomposition on the rational points , when @xmath112 ! for ( irreducible ) curves , and in the presence of a rational point @xmath113 , one obtains a similar decomposition ( non canonical ) @xmath114 with @xmath115 , @xmath116 and @xmath117 . this decomposition is responsible for the formula . in fact , one can prove that these decompositions partially generalize to any object @xmath86 . in the presence of a rational point , or more in general by choosing a positive zero - cycle @xmath118 ( here @xmath15 is assumed irreducible for simplicity and @xmath119 ) , one constructs two rational algebraic cycles @xmath120 which determine two projectors @xmath121 in the chow group @xmath122 of rational algebraic cycles modulo rational equivalence . the corresponding classes in @xmath123 ( if not zero ) determine two motives @xmath124 and @xmath125 ( _ cf . _ _ e.g. _ @xcite ) . for the applications , it is convenient to enlarge the category of effective motives by formally adding the tensor product inverse @xmath126 of the lefschetz motive : one usually refers to it as to the _ tate motive_. it corresponds , from the more refined point of view of galois theory , to the cyclotomic characters . this enlargement of @xmath49 by the so - called `` virtual motives '' produces an abelian , semi - simple category @xmath127 of pure motives for numerical equivalence . the objects of this category are now triples @xmath128 , with @xmath129 . effective motives are of course objects of this category and they are described by triples @xmath130 . the lefschetz motive gains a new interpretation in this category as @xmath131 . the tate motive is defined by @xmath132 and is therefore reminissent of ( in fact induces ) the notion of tate structure @xmath133 in hodge theory . in the category @xmath127 , the set of morphisms connecting two motives @xmath128 and @xmath134 is defined by @xmath135 in particular , @xmath136 , one defines the two motives @xmath137 these determine a canonical decomposition of any virtual motive as @xmath138 , where the direct sum of two motives as the above ones is defined by the formula @xmath139 the general definition of the direct sum of two motives requires a bit more of formalism which escapes this short overview : we refer to _ op.cit . _ for the details . the tensor structure @xmath140 and the involution ( _ i.e. _ auto - duality ) which is defined , for @xmath15 irreducible and @xmath119 by the functor @xmath141 ( the general case follows from this by applying additivity ) , determine the structure of a _ rigid _ @xmath12-_category _ on @xmath142 . here , @xmath143 denotes the transpose correspondence associated to @xmath87 ( _ i.e. _ the transpose of the graph ) . one finds , for example , that @xmath144 . in the particular case of the effective motive @xmath92 , with @xmath15 irreducible and @xmath119 , this involution determines the notion of poincar duality @xmath145 that is an auto - duality which induces the poincar duality isomorphism in any classical cohomological theory . a category of pure motives over a field @xmath5 and with coefficients in a field @xmath0 ( of characteristic zero ) is supposed to satisfy , to be satisfactory , several basic properties and to be endowed with a few fundamental structures . in the previous section we have described the historically - first example of a category of pure motives and we have reviewed some of its basic properties ( in that case @xmath146 ) . one naturally wonders about the description of others categories of motives associated to finer ( than the numerical ) equivalence relations on algebraic cycles : namely the categories of motives for homological or rational or algebraic equivalence relations . however , if one seeks to work with a semi - simple category , the afore mentioned result of jannsen tells us that the numerical equivalence is the only adequate relation . the semi - simplicity property is also attaint if one assumes grothendieck s standard conjectures . following the report of grothendieck in @xcite , these conjectures arose from the hope to understand ( read prove ) the conjectures of weil on the zeta - function of an algebraic variety defined over a finite field . it was well - known to grothendieck that the standard conjectures imply the weil s conjectures . these latter statements became a theorem in the early seventies ( 1974 ) , only a few years later the time when grothendieck stated the standard conjectures ( 1968 - 69 ) . the proof by deligne of the weil s conjectures , however , does not make any use of the standard conjectures , these latter questions remain still unanswered at the present time . the moral lesson seems to be that geometric topology and the theory of algebraic cycles govern in many central aspects the foundations of algebraic geometry . the standard conjectures of `` lefschetz type '' and of `` hodge type '' are stated in terms of algebraic cycles modulo homological equivalence ( _ cf . _ @xcite ) . they imply two further important conjectures . one of these states the equality of the homological and the numerical equivalence relations , the other one , of `` knneth type '' claims that the knneth decomposition of the diagonal in cohomology , can be described by means of ( rational ) algebraic cycles . it is generally accepted , nowadays , to refer to the full set of the four conjectures , when one quotes the standard conjectures . in view of their expected consequences , one is naturally led to study a category of pure _ motives for homological equivalence_. in fact , there are several candidates for this category since the definition depends upon the choice of a weil cohomological theory ( _ i.e. _ betti , tale , de - rham , crystalline , etc ) with coefficients in a field @xmath0 of characteristic zero . let us fix a cohomological theory @xmath147 for algebraic varieties in the category @xmath9 . then , the construction of the corresponding category of motives @xmath148 for homological equivalence is given following a procedure similar to that we have explained earlier on in this paper for the category of motives for numerical equivalence and with rational coefficients . the only difference is that now morphisms in the category @xmath149 are defined by means of algebraic correspondences modulo homological equivalence . at this point , one makes explicit use of the axiom `` cycle map '' that characterizes ( together with finiteness , poincar duality , knneth formula , cycle map , weak and strong lefschetz theorems ) any weil cohomological theory ( _ cf . _ @xcite ) . the set of algebraic morphisms connecting objects in the category @xmath9 is enlarged by including multi - valued maps @xmath150 that are defined as a @xmath0-linear combination of elements of the vector spaces @xmath151 where @xmath152 is a subfield . two cycles @xmath153 are _ homologically equivalent _ @xmath154 if their image , by means of the cycle class map @xmath155 is the same . this leads naturally to the definition of a subvector - space @xmath156 generated by the image of the cycle class map @xmath157 . these spaces define the correspondences in the category @xmath148 . if @xmath15 is purely @xmath158-dimensional , then @xmath159 in general , if @xmath15 decomposes into several connected components @xmath160 , one sets @xmath161 . in direct analogy to the construction of correspondences for numerical equivalence , the ring structure ( `` cap - product '' ) in cohomology determines a composition law ` @xmath74 ' among correspondences . the category @xmath149 is then defined as follows : the objects are triples @xmath162 , where @xmath86 , @xmath129 and @xmath163 is an idempotent . the collection of morphisms between two motives @xmath162 , @xmath164 is given by the set @xmath165 this procedure determines a pseudo - abelian , @xmath0-linear tensor category . the tensor law is given by the formula @xmath166 the commutativity and associativity constraints are induced by the obvious isomorphisms @xmath167 , @xmath168 . the unit object in the category is given by @xmath169 . one shows that @xmath149 is a _ rigid _ category , as it is endowed with an auto - duality functor @xmath170 for any object @xmath68 , the functor @xmath171 is left - adjoint to @xmath172 and @xmath173 is right - adjoint to @xmath174 . in the case of an irreducible variety @xmath15 , the internal hom is defined by the motive @xmath175 the standard conjecture of knneth type ( which is assumed from now on in this section ) implies that the knneth components of the diagonal @xmath176 determine a complete system of orthogonal , central idempotents . this important statement implies that the motive @xmath177 has the expected direct sum decomposition ( unique ) @xmath178 the cohomology functor @xmath179 factors through the projection @xmath180 . more in general , one shows that every motive @xmath181 gets this way a @xmath182-_grading _ structure by setting @xmath183 this grading is respected by all morphisms in the category and defines the structure of a _ graduation by weights _ on the objects . on a motive @xmath184 in the category , one sets @xmath185 where @xmath186 is a pure motive of weight @xmath77 . one finds for example , that @xmath96 has weight zero , @xmath187 has weight @xmath188 and that @xmath189 has weight @xmath190 . more in general , the motive @xmath191 has weight @xmath192 . in order to achieve further important properties , one needs to modify the natural commutativity constraint @xmath193 , @xmath194 , by defining @xmath195 we shall denote by @xmath196 the category of motives for homological equivalence in which one has implemented the modification on the tensor product structure . an important structure on a category of pure motives ( for homological equivalence ) is given by assigning to an object @xmath197 a _ motivic cohomology _ @xmath198 . @xmath198 is a pure motive of weight @xmath77 . this way , one views pure motives as a universal cohomological theory for algebraic varieties . the main property of the motivic cohomology is that it defines a universal realization of any given weil cohomology theory @xmath199 . candidates for these motivic cohomology theories have been proposed by a. beilinson @xcite , in terms of eigenspaces of adams operations in algebraic @xmath0-theory _ @xmath200 and by s. bloch @xcite , in terms of higher chow groups _ i.e. _ @xmath201 . the assignment of a weil cohomological theory with coefficients in a field @xmath0 which contains an assigned field @xmath202 is equivalent to the definition of an exact _ realization @xmath12-functor _ of @xmath203 in the category of @xmath0-vector spaces @xmath204 in particular , one obtains the realization @xmath205 which defines the notion of the _ tate twist _ in cohomology . more precisely : in tale cohomology : @xmath206 , where @xmath207 is a @xmath208-vector space of dimension one endowed with the cyclotomic action of the absolute galois group @xmath209 . the `` twist '' ( or torsion ) @xmath210 in tale cohomology corresponds to the torsion in galois theory defined by the @xmath211-th power of the cyclotomic character ( tate twist ) in de - rham theory : @xmath212 , with the hodge filtration defined by @xmath213 , @xmath214 . here , the effect of the torsion @xmath210 is that of shifting the hodge filtration of @xmath215-steps ( to the right ) in betti theory : @xmath216 . the bi - graduation on @xmath217 is purely of type @xmath218 . the torsion @xmath210 is here identified with the composite of a homothety given by a multiplication by @xmath219 followed by a shifting by @xmath220 of the hodge bi - graduation . using the structure of rigid tensor - category one introduces the notion of _ rank _ associated to a motive @xmath221 in @xmath196 . the rank of @xmath68 is defined as the trace of @xmath222 _ i.e. _ the trace of the morphism @xmath223 , where @xmath224 are _ resp . _ the evaluation and co - evaluation morphisms satisfying @xmath225 , @xmath226 . in general , one sets @xmath227 under the assumption of the standard conjectures ( more precisely under the assumption that homological and numerical equivalence relations coincide ) and that @xmath228 ( @xmath229 ) , the tannakian formalism invented by grothendieck and developed by saavedra @xcite , and deligne @xcite implies that the abelian , rigid , semi - simple tensor category @xmath196 is endowed with an exact , faithful @xmath12-fibre functor to the category of graded @xmath0-vector spaces @xmath230 which is compatible with the realization functor . this formalism defines a _ tannakian _ ( _ neutral _ if @xmath231 ) structure on the category of motives . one then introduces the _ tannakian group _ @xmath232 as a @xmath0-scheme in affine groups . through the tannakian formalism one shows that the fibre functor @xmath13 realizes an equivalence of rigid tensor categories @xmath233 where @xmath234 denotes the rigid tensor category of finite dimensional , @xmath202 representations of the tannakian group @xmath235 . this way , one establishes a quite useful dictionary between categorical @xmath12-properties and properties of the associated groups . because we have assumed all along the standard conjectures , the semi - simplicity of the category @xmath236 implies that @xmath235 is an algebraic , pro - reductive group , that is @xmath235 is the projective limit of reductive @xmath202-algebraic groups . the tannakian theory is a linear analog of the theory of finite , tale coverings of a given connected scheme . this theory was developed by grothendieck in sga1 ( theory of the pro - finite @xmath237 ) . for this reason the group @xmath235 is usually referred to as the _ motivic galois group _ associated to @xmath9 and @xmath199 . in the case of algebraic varieties of dimension zero ( _ i.e. _ for artin motives ) the tannakian group @xmath235 is nothing but the ( absolute ) galois group @xmath238 . in any reasonable cohomological theory the functors @xmath239 are deduced by applying standard methods of homological algebra to the related derived functors @xmath240 which associate to an object in @xmath9 a bounded complex of @xmath5-vector spaces , in a suitable triangulated category @xmath241 of complexes of modules over @xmath5 , whose heart is the category of motives . this is the definition of cohomology as @xmath242 under the assumption that the functors @xmath243 are realizations of corresponding motivic functors _ @xmath244 , one expects the existence of a ( non - canonical ) isomorphism in @xmath245 @xmath246.\ ] ] moreover , the introduction of the motivic derived functors @xmath247 suggests the definition of the following groups of _ absolute cohomology _ @xmath248).\ ] ] for a general motive @xmath249 , one defines @xmath250 ) = \text{ext}^i({\bf 1},m).\ ] ] the motives @xmath198 and the groups of absolute motivic cohomology are related by a spectral sequence @xmath251 the first interesting examples of pure motives arise by considering the category @xmath18 of tale , finite @xmath5-schemes . an object in this category is a scheme @xmath252 , where @xmath253 is a commutative @xmath5-algebra of finite dimension which satisfies the following properties . let @xmath254 denote a fixed separable closure of @xmath5 1 . @xmath255}$ ] 2 . @xmath256 , for @xmath257 finite , separable field extensions 3 . @xmath258 $ ] . the corresponding rigid , tensor - category of motives with coefficients in a field @xmath0 is usually referred to as the category of _ artin motives _ : @xmath259 . the definition of this category is independent of the choice of the equivalence relation on cycles as the objects of @xmath18 are smooth , projective @xmath5-varieties of dimension zero . one also sees that passing from @xmath18 to @xmath260 requires adding new objects in order to attain the property that the category of motives is abelian . one can verify this already for @xmath261 , by considering the real quadratic extension @xmath262 and the one - dimensional non - trivial representation of @xmath263 that factors through the character of order two of @xmath264 . this representation does not correspond to any object in @xmath265 , but can be obtained as the image of the projector @xmath266 , where @xmath267 is the generator of @xmath264 . therefore , @xmath268 is a new object . the category of artin motives is a semi - simple , @xmath0-linear , monoidal @xmath12-category . when @xmath269 , the commutative diagram of functors @xmath270 where @xmath271 is the contravariant functor of linearization @xmath272 determines a linearization of the galois - grothendieck correspondence ( gg ) by means of the equivalence of categories @xmath273 . this is provided by the fiber functor @xmath274 and by applying the tannakian formalism . it follows that @xmath275 is @xmath12-equivalent to the category @xmath276 of representations of the absolute galois group @xmath209 . these results were the departing point for grothendieck s speculations on the definition of higher dimensional galois theories ( _ i.e. _ galois theories associated to system of polynomials in several variables ) and for the definition of the corresponding motivic galois groups . the notion of an _ endomotive _ in noncommutative geometry ( _ cf . _ @xcite ) is the natural generalization of the classical concept of an artin motive for the noncommutative spaces which are defined by semigroup actions on projective limits of zero - dimensional algebraic varieties , endowed with an action of the absolute galois group . this notion applies quite naturally for instance , to the study of several examples of quantum statistical dynamical systems whose time evolution describes important number - theoretic properties of a given field @xmath5 ( _ cf . _ _ op.cit_ , @xcite ) . there are two distinct definitions of an endomotive : one speaks of algebraic or analytic endomotives depending upon the context and the applications . when @xmath5 is a number field , there is a functor connecting the two related categories . moreover , the abelian category of artin motives embeds naturally as a full subcategory in the category of algebraic endomotives ( _ cf . _ theorem [ ncartinthm ] ) and this result motivates the statement that the theory of endomotives defines a natural generalization of the classical theory of ( zero - dimensional ) artin motives . in noncommutative geometry , where the properties of a space ( frequently highly singular from a classical viewpoint ) are analyzed in terms of the properties of the associated noncommutative algebra and its ( space of ) irreducible representations , it is quite natural to look for a suitable abelian category which enlarges the original , non - additive category of algebras and in which one may also apply the standard techniques of homological algebra . likewise in the construction of a theory of motives , one seeks to work within a triangulated category endowed with several structures . these include for instance , the definition of ( noncommutative ) motivic objects playing the role of motivic cohomology ( _ cf . _ ) , the construction of a universal ( co)homological theory representing in this context the absolute motivic cohomology ( _ cf . _ ) and possibly also the set - up of a noncommutative tannakian formalism to motivate in rigorous mathematical terms the presence of certain universal groups of symmetries associated to renormalizable quantum field theories ( _ cf . _ _ e.g. _ @xcite ) . a way to attack these problems is that of enlarging the original category of algebras and morphisms by introducing a `` derived '' category of modules enriched with a suitable notion of correspondences connecting the objects that should also account for the structure of morita equivalence which represents the noncommutative generalization of the notion of isomorphism for commutative algebras . the sought for enlargement of the category @xmath277 of ( unital ) @xmath5-algebras and ( unital ) algebra homomorphisms is defined by introducing a new category @xmath278 of cyclic @xmath279-modules . the objects of this category are modules over the _ cyclic category _ @xmath280 . this latter has the same objects as the simplicial category @xmath105 ( @xmath280 contains @xmath105 as sub - category ) . we recall that an object in @xmath105 is a totally ordered set @xmath281 = \{0<1<\ldots < n\}\ ] ] for each @xmath282 , and a morphism @xmath283 \to [ m]\ ] ] is described by an order - preserving map of sets @xmath284 . the set of morphisms in @xmath105 is generated by faces @xmath285 \to [ n]$ ] ( the injection that misses @xmath77 ) and degeneracies @xmath286 \to [ n]$ ] ( the surjection which identifies @xmath79 with @xmath287 ) which satisfy several standard simplicial identities ( _ cf . _ _ e.g. _ @xcite ) . the set of morphisms in @xmath280 is enriched by introducing a new collection of morphisms : the _ cyclic morphisms_. for each @xmath282 , one sets @xmath288 \to [ n]\ ] ] fulfilling the relations @xmath289 \tau_n \sigma_i = \sigma_{i-1 } \tau_{n+1 } & 1 \leq i \leq n , & \tau_n \sigma_0 = \sigma_n \tau_{n+1}^2 \\[3 mm ] \tau_n^{n+1 } = 1_n . & & \end{array}\ ] ] the objects of the category @xmath278 are @xmath5-modules over @xmath280 ( _ i.e. _ @xmath279-modules ) . in categorical language this means functors @xmath290 ( @xmath291 category of @xmath5-modules ) . morphisms of @xmath279-modules are therefore natural transformations between the corresponding functors . it is evident that @xmath278 is an abelian category , because of the interpretation of a morphism in @xmath278 as a collection of @xmath5-linear maps of @xmath5-modules @xmath292 ( @xmath293 ) compatible with faces , degeneracies and cyclic operators . kernels and cokernels of these morphisms define objects of the category @xmath278 , since their definition is given point - wise . to an algebra @xmath294 over a field @xmath5 , one associates the @xmath279-module @xmath295 . for each @xmath296 one sets @xmath297 the _ cyclic morphisms _ on @xmath298 correspond to the cyclic permutations of the tensors , while the face and the degeneracy maps correspond to the algebra product of consecutive tensors and the insertion of the unit . this construction determines a functor @xmath299 _ traces _ @xmath300 give rise naturally to morphisms @xmath301 in @xmath278 . the main result of this construction is the following canonical description of the cyclic cohomology of an algebra @xmath294 over a field @xmath5 as the derived functor of the functor which assigns to a @xmath279-module its space of traces @xmath302 ( _ cf . _ @xcite,@xcite ) . this formula is the analog of , that describes the absolute motivic cohomology group of a classical motive as an ext - group computed in a triangulated category of motives @xmath245 . in the present context , on the other hand , the derived groups @xmath303 are taken in the abelian category of @xmath278-modules . the description of the cyclic cohomology as a derived functor in the cyclic category determines a useful procedure to embed the nonadditive category of algebras and algebra homomorphisms in the `` derived '' abelian category of @xmath279-modules . this construction provides a natural framework for the definition of the objects of a category of noncommutative motives . likewise in the construction of the category of motives , one is faced with the problem of finding the `` motivated maps '' connecting cyclic modules . the natural strategy is that of enlarging the collection of cyclic morphisms which are functorially induced by homomorphisms between ( noncommutative ) algebras , by implementing an adequate definition of ( noncommutative ) correspondences . the notion of an algebraic correspondence in algebraic geometry , as a multi - valued map defined by an algebraic cycle modulo a suitable equivalence relation , has here an analog with the notion of kasparov s bimodule and the associated class in @xmath4-theory ( _ cf . _ @xcite ) . likewise in classical motive theory , one may prefer to work with ( compare ) several versions of correspondences . one may decide to retain the full information supplied by a group action on a given algebra ( _ i.e. _ a noncommutative space ) rather than partially loosing this information by moding out with the equivalence relation ( homotopy in @xmath4-theory ) . there is a natural way to associate a cyclic morphism to a ( virtual ) correspondence and hence to a class in @xmath4-theory . starting with the category of separable @xmath3-algebras and @xmath304-homomorphisms , one enlarges the collection of morphisms connecting two unital algebras @xmath294 and @xmath305 , by including correspondences defined by elements of kasparov s bivariant @xmath0-theory @xmath306 ( @xcite , _ cf . _ also 8 and 9.22 of @xcite ) . more precisely , correspondences are defined by kasparov s bimodules , that means by triples @xmath307 which satisfy the following conditions : @xmath308 is a countably generated hilbert module over @xmath309 @xmath310 is a @xmath304-homomorphism of @xmath311 to bounded linear operators on @xmath308 ( _ i.e. _ @xmath310 gives @xmath308 the structure of an @xmath311-@xmath309 bimodule ) @xmath202 is a bounded linear operator on @xmath308 such that the operators @xmath312 $ ] , @xmath313 , and @xmath314 are compact for all @xmath315 . a hilbert module @xmath308 over @xmath309 is a right @xmath309-module with a _ positive _ , @xmath309-valued inner product which satisfies @xmath316 , @xmath317 and @xmath318 , and with respect to which @xmath308 is complete ( _ i.e. _ complete in the norm @xmath319 ) . notice that kasparov bimodules are morita - type of correspondences . they generalize @xmath304-homomorphisms of @xmath3-algebras since the latter ones may be re - interpreted as kasparov bimodules of the form @xmath320 . given a kasparov s bimodule @xmath321 , that is a @xmath294-@xmath305 hilbert bimodule @xmath308 as defined above , one associates , under the assumption that @xmath308 is a projective @xmath305-module of finite type ( _ cf . _ @xcite , lemma 2.1 ) , a cyclic morphism @xmath322 this result allows one to define an enlargement of the collection of cyclic morphisms in the category @xmath278 of @xmath279-modules , by considering kasparov s projective bimodules of finite type , as correspondences . one then implements the _ homotopy equivalence relation _ on the collection of kasparov s bimodules . two kasparov s modules are said to be homotopy equivalent @xmath323 if there is an element @xmath324\to { { \mathcal}{b}}\,|\ , f \text { continuous } \}\ ] ] which performs a unitary homotopy deformation between the two modules . this means that @xmath325 is unitarily equivalent to @xmath326 or equivalently re - phrased , that there is a unitary in bounded operators from @xmath327 to @xmath328 intertwining the morphisms @xmath329 and @xmath330 and the operators @xmath331 and @xmath332 . here @xmath333 is the evaluation at the endpoints . there is a binary operation on the set of all kasparov @xmath334-@xmath305 bimodules , given by the direct sum . by definition , the group of kasparov s bivariant @xmath0-theory is the set of homotopy equivalence classes @xmath335 of kasparov s modules @xmath336 . this set has a natural structure of abelian group with addition induced by direct sum . this bivariant version of @xmath0-theory is reacher than both @xmath0-theory and @xmath0-homology , as it carries an intersection product . there is a natural bi - linear , associative composition ( intersection ) product @xmath337 for all @xmath338 and @xmath339 separable @xmath3-algebras . this product is compatible with composition of morphisms of @xmath3-algebras . @xmath4-theory is also endowed with a bi - linear , associative exterior product @xmath340 which is defined in terms on the composition product by @xmath341 a slightly different formulation of @xmath4-theory , which simplifies the definition of this external tensor product is obtained by replacing in the data @xmath342 the operator @xmath202 by an _ unbounded _ , regular self - adjoint operator @xmath343 . the corresponding @xmath202 is then given by @xmath344 ( _ cf . @xcite ) . the above construction which produces an enlargement of the category of separable @xmath3-algebras by introducing correspondences as morphisms determines an additive , although non abelian category @xmath345 ( _ cf . _ @xcite 9.22.1 ) . this category is also known to have a triangulated structure ( _ cf . _ @xcite ) and this result is in agreement with the construction of the triangulated category @xmath245 in motives theory , whose heart is expected to be the category of ( mixed ) motives ( _ cf . _ section [ fundstructures ] and @xcite ) . a more refined analysis based on the analogy with the construction of a category of motives suggests that one should probably perform a further enlargement by passing to the pseudo - abelian envelope of @xmath345 , that is by formally including among the objects also ranges of idempotents in @xmath4-theory . in section [ anendo ] we will review the category of analytic endomotives where maps are given in terms of tale correspondences described by spaces @xmath346 arising from locally compact tale groupoids @xmath347 associated to zero - dimensional , singular quotient spaces @xmath348 with associated @xmath3-algebras @xmath349 in view of what we have said in this section , it would be also possible to define a category where morphisms are given by classes @xmath350 which describe sets of equivalent triples @xmath342 , where @xmath351 is given in terms of a bimodule @xmath352 with the trivial grading @xmath353 and the zero endomorphism @xmath354 . the definition of the category of analytic endomotives is more refined because the definition of the maps in this category does not require to divide by homotopy equivalence . the comparison between correspondences for motives given by algebraic cycles and correspondences for noncommutative spaces given by bimodules ( or elements in @xmath4-theory ) is particularly easy in the zero - dimensional case because the equivalence relations play no role . of course , it would be quite interesting to investigate the higher dimensional cases , in view of a unified framework for motives and noncommutative spaces which is suggested for example , by the recent results on the lefschetz trace formula for archimedean local factors of @xmath1-functions of motives ( _ cf . _ @xcite , section 7 ) . a way to attack this problem is by comparing the notion of a correspondence given by an algebraic cycle with the notion of a geometric correspondence used in topology ( _ cf . _ @xcite , @xcite ) . for example , it is easy to see that the definition of an algebraic correspondence can be reformulated as a particular case of the topological ( geometrical ) correspondence and it is also known that one may associate to the latter a class in @xmath4-theory . in the following two sections , we shall review and comment on these ideas . in geometric topology , given two smooth manifolds @xmath15 and @xmath16 ( it is enough to assume that @xmath15 is a locally compact parameter space ) , a _ topological ( geometric ) correspondence _ is given by the datum @xmath355 where : @xmath356 is a smooth manifold @xmath308 is a complex vector bundle over @xmath356 @xmath357 and @xmath358 are continuous maps , with @xmath359 proper and @xmath360 @xmath0-oriented ( orientation in @xmath0-homology ) . unlike in the definition of an algebraic correspondence ( _ cf . _ section [ fundstructures ] ) one does not require that @xmath356 is a subset of the cartesian product @xmath17 . this flexibility is balanced by the implementation of the extra piece of datum given by the vector bundle @xmath308 . to any such correspondence @xmath361 one associates a class in kasparov s k - theory @xmath362 @xmath363 denotes the class of @xmath308 in @xmath364 and @xmath365 is the element in @xmath4-theory which fulfills the grothendieck riemann - roch formula . we recall that given two smooth manifolds @xmath366 and @xmath367 and a continuous oriented map @xmath368 , the element @xmath369 determines the grothendieck riemann roch formula @xmath370 for all @xmath371 , with @xmath372 the todd genus @xmath373 the composition of two correspondences @xmath374 and @xmath375 is given by taking the fibered product @xmath376 and the bundle @xmath377 , with @xmath378 the projections . this determines the composite correspondence @xmath379 . in fact , one also needs to assume a _ condition on the maps @xmath360 and @xmath380 in order to ensure that the fibered product @xmath356 is a smooth manifold . the homotopy invariance of both @xmath381 and @xmath382 show however that the assumption of transversality is ` generically ' satisfied . theorem 3.2 of @xcite shows that kasparov product in @xmath4-theory ` @xmath12 ' agrees with the composition of correspondences , namely @xmath383 in algebraic geometry , the notion of correspondence that comes closest to the definition of a geometric correspondence ( as an element in @xmath4-theory ) is obtained by considering classes of algebraic cycles in algebraic @xmath0-theory ( _ cf . _ @xcite ) . given two smooth and projective algebraic varieties @xmath15 and @xmath16 , we denote by @xmath384 and @xmath385 the projections of @xmath17 onto @xmath15 and @xmath16 respectively and we assume that they are _ proper_. let @xmath386 be an algebraic cycle . for simplicity , we shall assume that @xmath356 is irreducible ( the general case follows by linearity ) . we denote by @xmath387 and @xmath388 the restrictions of @xmath384 and @xmath385 to @xmath356 . to the irreducible subvariety @xmath389 one naturally associates the coherent @xmath390-module @xmath391 . for simplicity of notation we write it as @xmath392 . we use a similar notation for the coherent sheaf @xmath393 , associated to the irreducible subvariety @xmath394 . then , the sheaf pullback @xmath395 has a natural structure of @xmath396-module . the map on sheaves that corresponds to the cap product by @xmath356 on cocycles is given by @xmath397 since @xmath384 is proper , the resulting sheaf is coherent . using , we can write equivalently @xmath398 we recall that the functor @xmath399 is the right adjoint to @xmath400 ( _ i.e. _ @xmath401 ) and that @xmath399 satisfies the grothendieck riemann roch formula @xmath402 using this result , we can equally compute the intersection product of by first computing @xmath403 and then applying @xmath404 . using and we know that we can replace by @xmath405 with the same effect in @xmath0-theory . thus , to a correspondence in the sense of that is defined by the image in @xmath0-theory of an algebraic cycle @xmath406 we associate the geometric class @xmath407 with @xmath387 , @xmath408 and with the bundle @xmath409 . the composition of correspondences is given in terms of the intersection product of the associated cycles . given three smooth projective varieties @xmath15 , @xmath16 and @xmath410 and ( virtual ) correspondences @xmath411 and @xmath412 , with @xmath413 and @xmath414 closed reduced irreducible subschemes , one defines @xmath415 @xmath416 , @xmath417 , and @xmath418 denote , as usual , the projection maps . under the assumption of ` general position ' which is the algebraic analog of the transversality requirement in topology , we obtain the following result [ corrcorr2 ] suppose given three smooth projective varieties @xmath15 , @xmath16 , and @xmath410 and algebraic correspondences @xmath419 given by @xmath420 and @xmath421 given by @xmath422 . assume that @xmath423 and @xmath424 are in general position in @xmath425 . then assigning to a cycle @xmath356 the topological correspondence @xmath426 satisfies @xmath427 where @xmath428 is the product of algebraic cycles and @xmath429 is the kasparov product of the topological correspondences . notice that , while in the topological ( smooth ) setting transversality can always be achieved by a small deformation ( _ cf . _ iii , @xcite ) , in the algebro - geometric framework one needs to modify the above construction if the cycles are not in general position . in this case the formula @xmath430\otimes [ { { \mathcal}{o}}_{t_2}]= [ { { \mathcal}{o}}_{t_1\circ t_2}]\ ] ] which describes the product in @xmath0-theory in terms of the intersection product of algebraic cycles must be modified by implementing tor - classes and one works with a product defined by the formula ( @xcite , theorem 2.7 ) @xmath431\otimes [ { { \mathcal}{o}}_{t_2}]= \sum_{i=0}^n ( -1)^i \left[{\rm tor}_i^{{{\mathcal}{o}}_x } ( { { \mathcal}{o}}_{t_1},{{\mathcal}{o}}_{t_2})\right].\ ] ] to define the category of _ algebraic endomotives _ one replaces the category @xmath432 of reduced , finite - dimensional commutative algebras ( and algebras homomorphisms ) over a field @xmath5 by the category of noncommutative algebras ( and algebras homomorphisms ) of the form @xmath433 @xmath434 denotes a unital algebra which is an inductive limit of commutative algebras @xmath435 . @xmath436 is a unital , _ abelian semigroup _ of algebra endomorphisms @xmath437 moreover , one imposes the condition that for @xmath438 , @xmath439 is an _ idempotent _ of the algebra and that @xmath440 is an isomorphism of @xmath434 with the compressed algebra @xmath441 . the crossed product algebra @xmath442 is defined by formally adjoining to @xmath434 new generators @xmath443 and @xmath444 , for @xmath438 , satisfying the algebraic rules @xmath445 since @xmath436 is abelian , these rules suffice to show that @xmath446 is the linear span of the monomials @xmath447 , for @xmath448 and @xmath449 . because @xmath450 , with @xmath451 reduced , finite - dimensional commutative algebras over @xmath5 , the construction of @xmath442 is in fact determined by assigning a _ projective system _ @xmath452 of varieties in @xmath432 ( @xmath453 is a countable indexing set ) , with @xmath454 morphisms in @xmath432 and with a suitably defined action of @xmath436 . here , we have implicitly used the equivalence between the category of finite dimensional commutative @xmath5-algebras and the category of affine algebraic varieties over @xmath5 . the graphs @xmath455 of the connecting morphisms of the system define @xmath209-invariant subsets of @xmath456 which in turn describe @xmath457 as algebraic correspondences . we denote by @xmath458 the associated _ pro - variety_. the compressed algebra @xmath441 associated to the idempotent @xmath459 determines a subvariety @xmath460 which is in fact isomorphic to @xmath15 , via the induced morphism @xmath461 . the noncommutative space defined by @xmath442 is the quotient of @xmath462 by the action of @xmath436 , _ i.e. _ of the action of the @xmath463 s . the galois group @xmath464 acts on @xmath462 by composition . by identifying the elements of @xmath462 with characters , _ i.e. _ with @xmath5-algebra homomorphisms @xmath465 , we write the action of @xmath464 on @xmath434 as @xmath466 this action commutes with the maps @xmath463 , _ i.e. _ @xmath467 . thus the whole construction of the system @xmath468 is @xmath464-equivariant . this fact does not mean however , that @xmath464 acts by automorphisms on @xmath442 ! moreover , notice that the algebraic construction of the crossed - product algebra @xmath442 endowed with the actions of @xmath464 and @xmath436 on @xmath462 makes sense also when @xmath469 . when @xmath470 , one defines the set of correspondences @xmath471 by using the notion of kasparov s bimodules @xmath472 which are projective and finite as right modules . this way , one obtains a first realization of the resulting category of noncommutative zero - dimensional motives in the abelian category of @xmath279-modules . in general , given @xmath468 , with @xmath452 a projective system of artin motives and @xmath436 a semigroup of endomorphisms of @xmath473 as above , the datum of the semigroup action is encoded naturally by the _ algebraic groupoid _ @xmath474 this is defined in the following way . one considers the grothendieck group @xmath475 of the abelian semigroup @xmath436 . by using the injectivity of the partial action of @xmath436 , one may also assume that @xmath436 embeds in @xmath475 . then , the action of @xmath436 on @xmath15 extends to define a _ partial action _ of @xmath475 . more precisely , for @xmath476 the two projections @xmath477 only depend on @xmath478 and the map @xmath479 defines an isomorphism of reduced algebras . it is immediate to verify that @xmath480 and that @xmath481 . the algebraic groupoid @xmath482 is defined as the disjoint union @xmath483 which corresponds to the commutative direct - sum of reduced algebras @xmath484 the range and the source maps in @xmath482 are given _ resp . _ by the natural projection from @xmath482 to @xmath15 and by its composition with the antipode @xmath485 which is defined , at the algebra level , by @xmath486 . the composition in the groupoid corresponds to the product of monomials @xmath487 . given two systems @xmath468 and @xmath488 , with associated crossed - product algebras @xmath442 and @xmath489 and groupoids @xmath490 and @xmath491 a _ geometric correspondence _ is given by a @xmath492-space @xmath493 , endowed with a right action of @xmath494 which fulfills the following _ tale _ condition . given a space such as @xmath494 , that is a disjoint union of zero - dimensional pro - varieties over @xmath5 , a right action of @xmath494 on @xmath356 is given by a map @xmath495 and a collection of partial isomorphisms @xmath496 fulfilling the following rules for partial action of the abelian group @xmath475 @xmath497 here @xmath498 denotes the partial action of @xmath475 on @xmath499 . one checks that such an action gives to the @xmath5-linear space @xmath500 a structure of right module over @xmath501 . the action of @xmath494 on @xmath356 is tale if the corresponding module @xmath500 is _ finite and projective _ over @xmath501 . given two systems @xmath468 and @xmath502 as above , an _ tale correspondence _ is therefore a @xmath503-space @xmath356 such that the right action of @xmath504 is tale . the @xmath2-linear space of _ ( virtual ) correspondences _ @xmath505 is the rational vector space of formal linear combinations @xmath506 of tale correspondences @xmath507 , modulo the relations arising from isomorphisms and equivalences : @xmath508 . the composition of correspondences is given by the fiber product over a groupoid . namely , for three systems @xmath468 , @xmath509 , @xmath510 joined by correspondences @xmath511 their composition is given by the rule @xmath512 that is the fiber product over the groupoid @xmath513 . finally , a system @xmath468 as above is said to be _ uniform _ if the normalized counting measures @xmath514 on @xmath515 satisfy @xmath516 the category @xmath517 of _ algebraic endomotives _ with coefficients in a fixed extension @xmath0 of @xmath2 is the ( pseudo)abelian category generated by the following objects and morphisms . the objects are uniform systems @xmath518 of artin motives over @xmath5 , as above . the set of morphisms in the category connecting two objects @xmath519 and @xmath520 is defined as @xmath521 the category @xmath259 of artin motives embeds as a full sub - category in the category of algebraic endomotives @xmath522 the functor @xmath523 maps an artin motive @xmath524 to the system @xmath468 with @xmath525 , @xmath526 and @xmath527 . the category of algebraic endomotives is inclusive of a large and general class of examples of noncommutative spaces @xmath528 which are described by semigroup actions on projective systems of artin motives . one may consider , for instance a pointed algebraic variety @xmath529 over a field @xmath5 and a countable , unital , abelian semigroup @xmath436 of _ finite _ endomorphisms of @xmath529 , _ unramified _ over @xmath530 . then , there is a system @xmath531 of artin motives over @xmath5 which is constructed from these data . more precisely , for @xmath532 , one sets @xmath533 for a pair @xmath534 , with @xmath535 , the connecting map @xmath536 is defined by @xmath537 this is an example of a system indexed by the semigroup @xmath436 itself , with partial order given by divisibility . one sets @xmath538 . since @xmath539 , the base point @xmath540 defines a component @xmath541 of @xmath542 for all @xmath532 . the pre - image @xmath543 in @xmath544 is a union of components of @xmath544 . this defines a projection @xmath545 onto an open and closed subset @xmath546 of the projective limit @xmath15 . it is easy to see that the semigroup @xmath436 acts on the projective limit @xmath15 by partial isomorphisms @xmath547 defined by the property @xmath548 the map @xmath549 is well - defined since the set @xmath550 is cofinal and @xmath551 , with @xmath552 . the image of @xmath549 is in @xmath546 , since by definition of @xmath549 : @xmath553 . for @xmath554 , we have @xmath555 . this shows that @xmath549 defines an isomorphism of @xmath15 with @xmath546 , whose inverse map is given by @xmath556 the corresponding algebra morphisms @xmath557 are then given by @xmath558 this class of examples also fulfill an _ equidistribution property _ , making the uniform normalized counting measures @xmath559 on @xmath542 compatible with the projective system and inducing a probability measure on the limit @xmath15 . namely , one has @xmath560 in this section we assume that @xmath5 is a number field . we fix an embedding @xmath561 and we denote by @xmath254 an algebraic closure of @xmath562 in @xmath563 . when taking points over @xmath254 , algebraic endomotives yield @xmath564-dimensional singular quotient spaces @xmath348 , which can be described by means of locally compact tale groupoids @xmath565 and the associated crossed product @xmath3-algebras @xmath566 . this construction gives rise to the category of _ analytic endomotives_. one starts off by considering a uniform system @xmath567 of artin motives over @xmath5 and the algebras @xmath568 the assignment @xmath569 defines an involutive embedding of algebras @xmath570 . the @xmath3-completion @xmath571 of @xmath572 is an abelian af @xmath3-algebra . one sets @xmath573 this is the @xmath3-completion of the algebraic crossed product @xmath574 . it is defined by the algebraic relations with the involution which is apparent in the formulae ( _ cf . _ @xcite,@xcite ) . in the applications that require to work with cyclic ( co)homology , it is important to be able to restrict from @xmath3-algebras such as @xmath575 to canonical dense subalgebras @xmath576 where @xmath577 is the algebra of locally constant functions . it is to this category of smooth algebras ( rather than to that of @xmath3-algebras ) that cyclic homology applies . the following result plays an important role in the theory of endomotives and their applications to examples arising from the study of the thermodynamical properties of certain quantum statistical dynamical systems . we shall refer to the following proposition , in section [ bcsystem ] of this paper for the description of the properties of the `` bc - system '' . the bc - system is a particularly relevant quantum statistical dynamical system which has been the prototype and the motivating example for the introduction of the notion of an endomotive . we refer to @xcite , 4.1 for the definition of the notion and the properties of a state on a ( unital ) involutive algebra . [ fab ] \1 ) the action of @xmath464 on @xmath462 defines a canonical action of @xmath464 by automorphisms of the @xmath3-algebra @xmath578 , preserving globally @xmath571 and such that , for any pure state @xmath579 of @xmath571 , @xmath580 \2 ) when the artin motives @xmath451 are abelian and normal , the subalgebras @xmath581 and @xmath582 are globally invariant under the action of @xmath464 and the states @xmath579 of @xmath575 induced by pure states of @xmath571 fulfill @xmath583\ ] ] on the totally disconnected compact space @xmath15 , the abelian semigroup @xmath436 of homeomorphisms acts , producing closed and open subsets @xmath584 , @xmath498 . the normalized counting measures @xmath514 on @xmath515 define a _ probability measure _ on @xmath15 with the property that the radon nikodym derivatives @xmath585 are locally constant functions on @xmath15 . one lets @xmath586 be the corresponding tale locally compact groupoid . the crossed product @xmath3-algebra @xmath587 coincides with the @xmath3-algebra @xmath349 of the groupoid @xmath482 . the notion of right ( or left ) action of @xmath588 on a totally disconnected locally compact space @xmath589 is defined as in the algebraic case by and . a right action of @xmath588 on @xmath589 gives on the space @xmath590 of continuous functions with compact support on @xmath589 a structure of right module over @xmath591 . when the fibers of the map @xmath592 are discrete ( countable ) subsets of @xmath589 one can define on @xmath590 an inner product with values in @xmath591 by @xmath593 a right action of @xmath588 on @xmath589 is _ tale _ if and only if the fibers of the map @xmath594 are discrete and the identity is a compact operator in the right @xmath3-module @xmath595 over @xmath349 given by . an _ tale correspondence _ is a @xmath596-space @xmath589 such that the right action of @xmath597 is tale . the @xmath2-vector space @xmath598 of linear combinations of tale correspondences @xmath589 modulo the equivalence relation @xmath599 for disjoint unions , defines the space of ( virtual ) correspondences . for @xmath600 , @xmath601 , and @xmath602 , the composition of correspondences @xmath603 is given following the same rule as for the algebraic case , that is by the fiber product over the groupoid @xmath604 . a correspondence gives rise to a bimodule @xmath352 over the algebras @xmath605 and @xmath606 and the composition of correspondences translates into the tensor product of bimodules . [ extartcalg ] the category @xmath607 of _ analytic endomotives _ is the ( pseudo)abelian category generated by objects of the form @xmath600 with the properties listed above and morphisms given as follows . for @xmath600 and @xmath601 objects in the category , one sets @xmath608 the following result establishes a precise relation between the categories of artin motives and ( noncommutative ) endomotives . [ ncartinthm ] the categories of artin motives and algebraic and analytic endomotives are related as follows . 1 . the map @xmath609 determines a tensor functor @xmath610 from algebraic to analytic endomotives . 2 . the galois group @xmath611 acts by natural transformations of @xmath612 . the category @xmath613 of artin motives embeds as a full subcategory of @xmath614 . the composite functor @xmath615 maps the full subcategory @xmath613 of artin motives faithfully to the category @xmath616 of @xmath464-equivariant @xmath4-theory with coefficients in @xmath0 . given two artin motives @xmath617 and @xmath618 and a component @xmath619 of the cartesian product @xmath620 , the two projections turn @xmath500 into a @xmath621-bimodule @xmath622 . if @xmath623 , @xmath624 defines a sum of bimodules in @xmath625 . the composition of correspondences in @xmath613 translates into the tensor product of bimodules in @xmath616 @xmath626 one composes the functor @xmath627 with the natural functor @xmath628 which associates to a @xmath621-bimodule @xmath629 the @xmath630-bimodule @xmath631 . the resulting functor @xmath632 is faithful since a correspondence such as @xmath419 is uniquely determined by the corresponding map of @xmath0-theory @xmath633 . the prototype example of the data which define an analytic endomotive is the system introduced by bost and connes in @xcite . the evolution of this @xmath3-dynamical system encodes in its group of symmetries the arithmetic of the maximal abelian extension of @xmath634 . this quantum statistical dynamical system is described by the datum given by a noncommutative @xmath3-algebra of _ observables _ @xmath635 and by the _ time evolution _ which is assigned in terms of a one - parameter family of automorphisms @xmath636 of the algebra . the action of the ( multiplicative ) semigroup @xmath637 on the commutative algebra @xmath638 is defined by @xmath639 for the definition of the associate endomotive , one considers the projective system @xmath640 of zero - dimensional algebraic varieties @xmath641 , where @xmath642 $ ] is the group ring of the abelian group @xmath643 . the inductive limit @xmath644 $ ] is the group ring of @xmath645 . the endomorphism @xmath646 associated to an element @xmath647 is given on the canonical basis @xmath648 $ ] , @xmath649 , by @xmath650 the artin motives @xmath651 are normal and abelian , so that proposition [ fab ] applies . the action of the galois group @xmath263 on @xmath641 is obtained by composing a character @xmath652 with the action of an element @xmath653 . since @xmath654 is determined by the @xmath655-th root of unity @xmath656 , this implies that the action of @xmath657 factorizes through the cyclotomic action and coincides with the _ symmetry group _ of the bc - system . the subalgebra @xmath658 coincides with the rational subalgebra defined in @xcite . there is an interesting description of this system in terms of a pointed algebraic variety @xmath529 ( _ cf . _ section [ genendo ] ) on which the abelian semigroup @xmath436 acts by finite endomorphisms . one considers the pointed affine group scheme @xmath659 ( the multiplicative group ) and lets @xmath436 be the semigroup of non - zero endomorphisms of @xmath660 . these endomorphisms correspond to maps of the form @xmath661 , for some @xmath282 . then , the general construction outlined in section [ genendo ] determines on @xmath662 the bc system . one considers the semigroup @xmath663 acting on @xmath664 as specified above . it follows from the definition that @xmath665 where @xmath666/(u_n^n-\,1).\ ] ] for @xmath667 the connecting morphism @xmath668 is defined by the algebra homomorphism @xmath669 , @xmath670 with @xmath671 . thus , one obtains an isomorphism of @xmath2-algebras @xmath672\,,\quad \iota(u_n)=\,e_{\frac{1}{n } } .\ ] ] the partial isomorphisms @xmath673 \to { \mathbb{q}}[{\mathbb{q}}/{\mathbb{z}}]$ ] of the group ring as described by the formula correspond under the isomorphism @xmath523 , to those given by on @xmath674 . one identifies @xmath15 with its space of characters @xmath675\to \bar { \mathbb{q}}$ ] . then , the projection @xmath676 is given by the restriction of ( the character associated to ) @xmath677 to the subalgebra @xmath678 . the projection of the composite of the endomorphism @xmath679 of with @xmath680 is given by @xmath681 this projection is non - zero if and only if the restriction @xmath682 is the trivial character , that is if and only if @xmath683 . moreover , in that case one has @xmath684 and in particular @xmath685 for @xmath686 the inclusion of algebraic spaces @xmath687 is given at the algebra level by the surjective homomorphism @xmath688 thus , one can rewrite as @xmath689 this means that @xmath690 by using the formula , one obtains the desired equality of the @xmath440 s of and . this construction continues to make sense for the affine algebraic variety @xmath691 for any field @xmath5 , including the case of a field of positive characteristic . in this case one obtains new systems , different from the bc system . the functor @xmath692 which connects the categories of algebraic and analytic endomotives establishes a significant bridge between the commutative world of artin motives and that of noncommutative geometry . when one moves from commutative to noncommutative algebras , important new tools of thermodynamical nature become available . one of the most relevant techniques ( for number - theoretical applications ) is supplied by the theory of tomita and takesaki for von neumann algebras ( @xcite ) which associates to a suitable state @xmath579 ( _ i.e. _ a faithful weight ) on a von neumann algebra @xmath68 , a one - parameter group of automorphisms of @xmath68 ( _ i.e. _ the modular automorphisms group ) @xmath693 @xmath694 is the modular operator which acts on the completion @xmath695 of @xmath696 , for the scalar product @xmath697 . this general theory applies in particular to the unital involutive algebras @xmath698 of and to the related @xmath3-algebras @xmath575 which are naturally associated to an endomotive . a remarkable result proved by connes in the theory and classifications of factors ( @xcite ) states that , modulo inner automorphisms of @xmath68 , the one - parameter family @xmath699 is _ independent _ of the choice of the state @xmath579 . this way , one obtains a _ canonically defined _ one parameter group of automorphisms classes @xmath700 in turn , this result implies that the crossed product dual algebra @xmath701 and the dual scaling action @xmath702 are _ independent _ of the choice of ( the weight ) @xmath579 . when these results are applied to the analytic endomotive @xmath703 associated to an algebraic endomotive @xmath518 , the above dual representation of @xmath704 combines with the representation of the absolute galois group @xmath464 . in the particular case of the endomotive associated to the bc - system ( _ cf . _ section [ bcsystem ] ) , the resulting representation of @xmath705 on the cyclic homology @xmath706 of a suitable @xmath707-module @xmath708 associated to the thermodynamical dynamics of the system @xmath709 determines the spectral realization of the zeroes of the riemann zeta - function and of the artin @xmath1-functions for abelian characters of @xmath464 ( _ cf . _ @xcite , theorem 4.16 ) . the action of the group @xmath710 on the cyclic homology @xmath711 of the noncommutative motive @xmath708 is analogous to the action of the weil group on the tale cohomology of an algebraic variety . in particular , the action of @xmath704 is the ` characteristic zero ' analog of the action of the ( geometric ) frobenius on tale cohomology . this construction determines a functor @xmath712 from the category of endomotives to the category of ( infinite - dimensional ) representations of the group @xmath410 . the analogy with the tannakian formalism of classical motive theory is striking . it is also important to underline the fact that the whole thermodynamical construction is non - trivial and relevant for number - theoretic applications only because of the particular nature of the factor @xmath68 ( type @xmath713 ) associated to the original datum @xmath714 of the bc - system . it is tempting to compare the original choice of the state @xmath579 ( weight ) on the algebra @xmath294 which singles out ( via the gelfand - naimark - segal construction ) the factor @xmath68 defined as the weak closure of the action of @xmath715 in the hilbert space @xmath716 , with the assignment of a factor @xmath717 ( @xmath718 ) on a pure motive @xmath719 : _ cf . _ . in classical motive theory , one knows that the assignment of a @xmath182-grading is canonical only for homological equivalence or under the assumption of the standard conjecture of knneth type . in fact , the definition of a weight structure depends upon the definition of a complete system of orthogonal central idempotents @xmath720 . passing from the factor @xmath68 to the canonical dual representation carries also the advantage to work in a setting where projectors are classified by their real dimension ( @xmath721 is of type @xmath722 ) , namely in a noncommutative framework of continuous geometry which generalizes and yet still retains some relevant properties of the algebraic correspondences ( _ i.e. _ degree or dimension ) . the process of dualization is in fact subsequent to a thermodynamical `` cooling procedure '' in order to work with a system whose algebra approaches and becomes in the limit , a commutative algebra ( _ i.e. _ @xmath723 ) . finally , one has also to implement a further step in which one filters ( _ i.e. _ `` distils '' ) the relevant noncommutative motive @xmath708 within the derived framework of cyclic modules ( _ cf . _ section [ cyclicmodules ] ) . this procedure is somewhat reminissent of the construction of the vanishing cohomology in algebraic geometry ( _ cf . _ @xcite ) . when the algebra of the bc - system gets replaced by the noncommutative algebra of coordinates @xmath724 of the adle class space @xmath725 of a number field @xmath5 , the cooling procedure is described by a restriction morphism of ( rapidly decaying ) functions on @xmath726 to functions on the `` cooled down '' subspace @xmath727 of idle classes ( _ cf . _ @xcite , section 5 ) . in this context , the representation of @xmath727 on the cyclic homology @xmath728 of a suitable noncommutative motive @xmath729 produces the spectral realization of the zeroes of hecke @xmath1-functions ( _ cf . _ _ op.cit_ , theorem 5.6 ) . the whole construction describes also a natural way to associate to a noncommutative space a canonical set of `` classical points '' which represents the analogue in characteristic zero , of the geometric points @xmath730 of a smooth , projective curve @xmath731 .

the study of the spontaneous organization of microscopic objects such as colloids , amphiphiles or protein clusters into mesoscale structures ( mesoatoms ) is an active research area that offers challenges on many fronts @xcite . one would like to create a rich warehouse of raw materials from which to engineer mesomolecules or bulk materials with novel mechanical , optical or electronic behavior . a fertile source of potential mesoatoms is provided by the self - assembly of micron or smaller scale colloidal particles on _ curved _ interfaces . the interplay between spatial curvature and condensed matter order in these systems has proven to be very rich . spatial curvature leads to novel defect arrays in the ground state that have widespread implications for the fundamental and applied physics of curved phases of matter with crystalline , hexatic or nematic order . in a recent paper kim _ et al _ reported the formation of toroidal micelles from the self - assembly of dumbbell - shaped amphiphilic molecules @xcite . molecular dumbbells dissolved in a selective solvent self - assemble in an aggregate structure due to their amphiphilic character . this process has been observed to yield coexisting spherical and open - ended cylindrical micelles . these structures change slowly over the course of a week to toroidal micelles which thus appear more stable . toroidal geometries also occur in microbiology in the viral capsid of the coronavirus _ torovirus _ @xcite . the torovirus is an rna viral package of maximal diameter between @xmath0 and @xmath1 nm and is surrounded , as other coronaviridae , by a double wreath / ring of cladding proteins . carbon nanotori form another fascinating and technologically promising class of toroidal crystals @xcite with remarkable magnetic and electronic properties . the interplay between the ballistic motion of the @xmath2 electrons and the geometry of the embedding torus leads to a rich variety of quantum mechanical properties including pauli paramagnetism @xcite and aharonov - bohm oscillations in the magnetization @xcite . ring closure of carbon nanotubes by chemical methods @xcite suggest that nanotubes may be more flexible than at first thought and provides another technique of constructing carbon tori . a unified theoretical framework to describe the structure of toroidal crystals is provided by the elastic theory of defects in a curved background @xcite . this formalism has the advantage of far fewer degrees of freedom than a direct treatment of the microscopic interactions and allows one to explore the origin of the emergent symmetry observed in toroidal crystals as the result of the interplay between defects and geometry . the latter is one of the fundamental hallmarks of two - dimensional non - euclidean crystals and leads to universal features observed in systems as different as viral capsids and carbon macromolecules . in this paper we provide a detailed analysis of the structural properties of toroidal crystals . we show that the ground state has excess @xmath3fold disclination defects on the exterior of the torus and @xmath4fold defects on the interior . the precise number of excess disclinations , as well as their arrangement , is primarily controlled by the aspect ratio of the torus . since defective regions are physically distinguished locations they are natural places for biological activity and chemical functionalization . a thorough understanding of the surface structure of crystalline assemblages could represent a significant step towards a first - principles design of entire libraries of nano and mesoscale components with precisely determined valence . such mesoatoms could serve in turn as the building blocks for mesomolecules or bulk materials via spontaneous self - assembly or controlled fabrication . the paper is organized as follows . in sec . [ sec:2 ] we review the elastic theory of defects in two - dimensional curved geometries and derive the ground state energy of a toroidal crystal . in sec . [ sec:3 ] we summarize some fundamentals of the geometry of triangulated tori and discuss how the intrinsically discrete problem of crystallography can be reconciled with the result of the continuous elastic theory presented in the previous section . sec . [ sec:4 ] is devoted to the analysis of the crystalline structure arising from the solution of the elastic problem . in sec . [ sec:5 ] we discuss the results of numerical minimization of the potential energy of a system of classical particles interacting via a short - range potential on the surface of a torus in the light of the results of sec . [ sec:4 ] . finally , in sec . [ sec:6 ] we specialize our analysis to the case of a `` fat torus '' of aspect ratio one and we show how the curvature singularity at the center of the torus is responsible for a remarkable curvature - driven transition to a disordered state . some of the results presented here have been already announced in ref . . a two - dimensional torus of revolution @xmath5 is described in parametric form by : @xmath6 where @xmath7 are the two radii of the torus . the metric tensor @xmath8 ( with determinant @xmath9 ) and the gaussian curvature @xmath10 can be written respectively as : @xmath11 k = \frac{\cos\psi}{r_{2}(r_{1}+r_{2}\cos\psi)}\,.\end{gathered}\ ] ] within the elastic theory of defects on curved surfaces the original interacting particle problem is mapped to a system of interacting disclination defects in a continuum elastic curved background . disclinations are characterized by their topological or disclination charge , @xmath12 , representing the departure of a vertex from a the @xmath13fold coordination of a perfect triangular lattice . thus @xmath14 , where @xmath15 is the coordination number of the @xmath16th vertex . a classic theorem of euler requires the total disclination charge of any triangulation of a two - dimensional riemannian manifold @xmath17 to equal @xmath18 , where @xmath19 is the euler characteristic of @xmath17 . in the case of the torus @xmath20 , and thus disclinations must appear in pairs of opposite disclination charge ( i.e. @xmath3fold and @xmath4fold vertices with @xmath21 and @xmath22 respectively ) in order to ensure disclination charge neutrality . ( color online ) the standard parametrization of a circular torus of radii @xmath23 and @xmath24 . ] the total free energy of a toroidal crystal with @xmath25 disclinations can be expressed as @xmath26 where @xmath27 is the 2d young modulus and @xmath28 is the solution of the following poisson problem with periodic boundary conditions : @xmath29 where @xmath30 is the laplace - beltrami operator @xmath31 and @xmath32 is the total topological charge density @xmath33 of @xmath25 disclinations located at the sites @xmath34 together with the screening contribution due to the gaussian curvature @xmath35 of the embedding manifold . the first term in eq . represents the long - range elastic distortion due to defects and curvature . the second term in eq . is the defect core - energy representing the energy required to create a single disclination defect . this quantity is related to the short - distance cut - off of the elastic theory and is proportional to the square of the topological charge times a constant @xmath36 @xcite . finally @xmath37 is the free energy of a flat defect - free monolayer . the gauss - bonnet theorem @xcite requires the total topological charge to be zero on the torus : @xmath38 the function @xmath28 can be expressed in the green form : @xmath39 where @xmath40 , the laplacian green function on the torus , satisfies the equation : @xmath41 with @xmath42 the dirac delta function on the torus and @xmath43 is the surface area . it is easy to prove that the solution of the traditional green - laplace equation does nt exist on closed manifolds like the torus , from which the extra term @xmath44 appearing in eq . . @xmath44 is the eigenfunction of the laplacian associated with the null eigenvalue ( zero mode ) and is necessary because a pure isolated source ( giving rise to the delta - function on the right - hand side ) has no place to terminate the field on the closed torus . as usual the calculation of the green function can be remarkably simplified by conformally mapping the torus to a domain of the euclidean plane via a suitable system of isothermal coordinates . intuitively the torus is conformally equivalent to a rectangular domain described by a system of cartesian coordinates . to make this explicit , one can equate the metric of the torus in the coordinates @xmath45 to a conformally euclidean metric in the coordinates @xmath46 : @xmath47 where @xmath48 is a positive conformal factor . taking @xmath49 and @xmath50 , the coordinate @xmath51 is determined by the differential equation : @xmath52 where @xmath53 , the aspect ratio of the torus , may be taken greater or equal to one without loss of generality . choosing the plus sign and integrating both sides of eq . we find : @xmath54 taking @xmath55 $ ] , the integral yields : @xmath56 where @xmath57 in the transformed coordinate system @xmath46 the modified green - laplace equation reads : @xmath58 where @xmath59 and @xmath60 are now the euclidean laplacian and delta function . the function @xmath40 can be expressed in the form : @xmath61 where @xmath62 is the laplacian green function on a periodic rectangle and the angular brackets stand for the normalized integral of the function @xmath62 with respect to the dotted variable : @xmath63 analogously the function @xmath64 is given by @xmath65 and ensures the neutrality property : @xmath66 using standard analysis the green function @xmath62 at the points @xmath67 and @xmath68 of a periodic rectangle of edges @xmath69 and @xmath70 can be calculated in the form : @xmath71 + \frac{1}{2\pi}\log\left|\frac{\vartheta_{1}(\frac{z - z'}{p_{1}/\pi}|\frac{ip_{2}}{p_{1 } } ) } { \vartheta_{1}'^{\frac{1}{3}}(0|\frac{ip_{2}}{p_{1}})}\right|\,,\end{gathered}\ ] ] where @xmath72 , @xmath73 and @xmath74 is the jacobian theta function and reflects the double periodicity of the torus @xcite . a pedagogical derivation of the green function @xmath62 is reported in appendix [ app:1 ] . substituting eq . in eq . with @xmath75 and @xmath76 we obtain : @xmath77 where @xmath78 represents the stress field due to the gaussian curvature of the torus and is given by : @xmath79+\frac{r-\sqrt{r^{2}-1}}{r}\,.\ ] ] the function @xmath80 is the stress field at the point @xmath81 arising from the elastic distortion due to a defect at @xmath34 and is given by @xmath82 + \frac{1}{4\pi^{2}r}\log(r+\cos\psi_{k})-\frac{\kappa}{4\pi^{2}}\operatorname{re}\{\operatorname{li_{2}}(\alpha e^{i\psi_{k}})\}\notag\\[7pt ] + \frac{1}{2\pi}\log\left|\vartheta_{1}\left(\frac{z - z_{k}}{\kappa}\bigg|\frac{2i}{\kappa}\right)\right|\ , , \label{eq : gamma_defects}\end{gathered}\ ] ] where @xmath83 is the usual eulerian dilogarithm and @xmath84 a derivation of the functions @xmath78 and @xmath80 is given in appendix [ app:2 ] . integrating the function @xmath28 on the torus gives the elastic energy of an arbitrary collection of disclinations . a detailed analysis of the crystalline structure arising from eqs . , , and is carried out in sec . [ sec:3 ] . before analyzing the defect distribution arising from the elastic energy of eq . , together with eq . , it is necessary to understand the geometry of triangulated tori . reconciling the predictions of a continuum elastic theory with the intrinsically discrete nature of crystallography requires an understanding of the possible lattices that can be embedded on the torus and the associated defects . the problem of classifying the possible triangulations of the @xmath85torus has received considerable attention from mathematicians , physicists and chemists over the past twenty years . lavrenchenko @xcite proved in 1984 that all the triangulations of the torus can be generated from 21 irreducible triangulations by certain sequences of operations called vertex splitting . after the discovery of carbon nanotubes in 1991 and the subsequent theoretical construction ( later followed by the experimental observation ) of graphitic tori , many possible tessellations of the circular torus have been proposed by the community @xcite . in this section we review the construction of a defect - free triangulated torus and we show how the most symmetric defective triangulations can be generally grouped into two fundamental classes corresponding to symmetry groups @xmath86 and @xmath87 respectively . ( color online ) construction of a defect - free triangulation of the torus . on top planar map of the triangulated torus corresponding to the @xmath88 configuration @xmath89 . on the bottom @xmath90 chiral torus . the edges of each one of the six tubular segments has been highlighted in red.,title="fig : " ] + ( color online ) construction of a defect - free triangulation of the torus . on top planar map of the triangulated torus corresponding to the @xmath88 configuration @xmath89 . on the bottom @xmath90 chiral torus . the edges of each one of the six tubular segments has been highlighted in red.,title="fig : " ] for the sake of consistency with the existing literature we adopt here the language developed to describe the structure of carbon nanotubes . the structure of a triangulated cylinder can be specified by a pair of triangular lattice vectors @xmath91 and @xmath92 , called the _ chiral _ and _ translation _ vector respectively , which together define how the planar lattice is rolled up . in the canonical basis @xmath93 and @xmath94 , the vector @xmath91 has the form : @xmath95 the translation vector @xmath92 , on the other hand , can be expressed as an integer multiple @xmath96 of the shortest lattice vector @xmath97 perpendicular to @xmath91 . the vector @xmath97 is readily found to be of the form : @xmath98 where @xmath99 denotes the greatest common divisor of @xmath100 and @xmath101 and enforces the minimal length . the three - dimensional structure of the torus is obtained by connecting the edge @xmath102 of the rectangle in fig . [ fig : triangulation ] ( top ) to @xmath103 and @xmath104 to @xmath105 . the edge @xmath102 is then mapped to the external equator of the torus while the edge @xmath104 to the @xmath106 meridian . the resultant toroidal lattice has characteristic chirality related to the initial choice of the vector @xmath91 . in the nanotubes literature _ armchair _ referes to the lattice obtained by choosing @xmath107 , _ zigzag _ to that obtained for @xmath108 and _ chiral _ to all other lattices . an example of a @xmath88 chiral torus is shown in fig . [ fig : triangulation ] ( bottom ) for the case @xmath109 , @xmath110 and @xmath111 . the chirality is extremely important in graphitic carbon nanotube or nanotori , where it determines whether the electronic behavior of the system is metallic or semiconducting . by euler s theorem one can prove that the number of triangular faces @xmath112 and the number of vertices @xmath113 of a triangular toroidal lattice is given by : @xmath114 denoting @xmath115 the area of the rectangle with edges @xmath91 and @xmath92 and @xmath116 the area of a fundamental equilateral triangle , the number of vertices of a defect - free toroidal triangulation is then : @xmath117 the planar construction reviewed above allows only lattices with an even number of vertices . defect - free toroidal deltahedra with an odd number of vertices are also possible and their construction is generally achieved by assembling congruent octahedral building blocks . an example of this scheme will be briefly discussed in sec . [ sec:5 ] for the case @xmath118 and @xmath119 . we refer the reader to ref . for an comprehensive review of the topic . the embedding of an equal number of pentagonal and heptagonal disclinations in the hexagonal network was first proposed by dunlap in 1992 as a possible way to incorporate positive and negative gaussian curvature into the cylindrical geometry of carbon tubules @xcite . according to the dunlap construction the necessary curvature is incorporated by the insertion of `` knees '' ( straight cylindrical sections of the same diameter joined with a kink ) in correspondence with each pentagon - heptagon pair arising from the junction of tubular segments of different chirality ( see fig . [ fig : knees ] ) . ( color online ) voronoi lattices of a tp@xmath120 prismatic ( top ) and ta@xmath120 antiprismatic ( bottom ) toroids with @xmath121 and @xmath122 . ] in particular , a junction between a @xmath123 and a @xmath124 tube can be obtained by placing a @xmath4fold disclination along the internal equator of the torus and a @xmath3fold disclination along the external equator . since the radii of the two segments of a junction are different by construction , the values of @xmath120 and @xmath125 are commonly chosen to minimize the ratio @xmath126 . by repeating the @xmath127 construction periodically it is possible to construct an infinite number of toroidal lattices with an even number of disclinations pairs and dihedral symmetry group @xmath86 ( where @xmath128 is the total number of @xmath127 pairs , fig [ fig : lucas - torus ] ) . the structure of the lattice is described by the alternation of two motifs with crystalline axes mutually rotated by @xmath129 as a consequence of the connecting disclination . one of the fundamental aspects of dunlap s construction is that all the disclinations are aligned along the two equators of the torus where the like - sign gaussian curvature is maximal . as we will see below , this feature makes these arrangements optimal in releasing the elastic stress due to curvature . ( color online ) dunlap knees obtained by joining two straight tubular segments with @xmath123 and @xmath124 chirality . [ courtesy of a. a. lucas and a. fonseca , facults universitaries notre - dame de la paix , namur , belgium ] . ] ( color online ) five - fold polygonal torus obtained by joining @xmath130 and @xmath131 tubular segments via ten pairs of @xmath127 rings . this structure was originally proposed by the authors of ref . as a possible low - strain configuration for carbon nanotori . [ courtesy of a. a. lucas and a. fonseca , facults universitaries notre - dame de la paix , namur , belgium ] . ] another class of crystalline tori with dihedral antiprismatic symmetry @xmath87 was initially proposed by itoh _ et al _ @xcite shortly after dunlap . aimed at reproducing a structure similar to the c@xmath132 fullerene , itoh s original construction implied ten disclination pairs and the point group @xmath133 . in contrast to dunlap tori , disclinations are never aligned along the equators in antiprismatic tori , instead being staggered at some angular distance @xmath134 from the equatorial plane . hereafter we will use the symbol ta@xmath120 to refer to toroidal deltahedra with @xmath128 disclination pairs and @xmath87 symmetry group . a systematic construction of defected triangulations of the torus can be achieved in the context of planar graphs @xcite . a topological embedding of a graph in a two - dimensional manifold corresponds to a triangulation of the manifold if each region of the graph is bounded by exactly three vertices and three edges , and any two regions have either one common vertex or one common edge or no common elements of the graph . the simplest example of toroidal polyhedra with @xmath87 symmetry group , featuring only @xmath3fold and @xmath4fold vertices , can be constructed by repeating @xmath120 times the unit cell of fig . [ fig : graph1]a . these _ toroidal antiprisms _ have @xmath135 vertices and can be obtained equivalently from the edge skeleton of a @xmath136fold antiprism by attaching at each of the base edges a pentagonal pyramid and by closing the upper part of the polyhedron with @xmath120 additional triangles . by counting the faces one finds @xmath137 from which @xmath135 . the simplest polyhedron of this family has @xmath138 and @xmath139 symmetry group ( see top left of fig . [ fig : toroidal - antiprisms ] ) and corresponds to the `` drilled icosahedron '' obtained by removing two parallel faces of an icosahedron and connecting the corresponding edges with the six lateral faces of an antiprism with triangular base ( i.e. a prolate octahedron ) . starting from this family of toroidal antiprisms a number of associated triangulations having the same defect structure can be obtained by geometrical transformations such as the goldberg inclusion @xcite . such transformations , popularized by caspar and klug for the construction of the icosadeltahedral structure of spherical viruses @xcite , consist in partitioning each triangular face of the original graph into smaller triangular faces in such a way that old vertices preserve their valence and new vertices have valence six . the partition is obtained by specifying two integer numbers @xmath140 which define how the original vertices of each triangle are connected by the new edges so that the total number of vertices is increased by a factor @xmath141 . ( color online ) first six toroidal antiprisms obtained by repeating the unit cell of fig . [ fig : graph1 ] . the first polyhedron on the left is the `` drilled icosahedron '' . ] a general classification scheme for @xmath87 symmetric tori was provided by berger and avron @xcite in 1995 . their scheme is based on the construction of unit graphs comprising triangular tiles of different _ generations_. in each generation , tiles are scaled in length by a factor @xmath142 with respect to the previous generation . this rescaling approximates the non - uniformity of the metric of a circular torus . ( a ) unit cell for toroidal antiprisms . @xmath3fold vertices are circled and @xmath4fold vertices are boxed . ( b ) unit cell of a @xmath87 torus in the berger - avron construction . the graph consists of four generation of tiles and the internal equator of the torus is mapped into the horizontal line passing to the mid - point between the 6th and the 7th vertex . ] ( color online ) unit cells for dunlap toroids of type @xmath143 and @xmath144 according to the classification scheme given here . highlighted regions correspond to the central polygon . ] dunlap toroids can be obtained from unit cells such as those shown in fig . [ fig : graph2 ] . the geometrical properties of these graphs can be described in different ways . a particularly intuitive way , in the spirit of this paper , consists in specifying the distances between @xmath3 and @xmath4fold pairs . one starts by drawing the smallest convex loop passing through defective sites . this identifies a central polygon whose upper vertices ( @xmath145 and @xmath146 in fig.[fig : graph3 ] ) have degree five and lower vertices ( @xmath147 and @xmath148 in fig . [ fig : graph3 ] ) have degree seven . then calling @xmath100 the distance between @xmath3fold vertices @xmath145 and @xmath146 , @xmath101 that between @xmath4fold vertices @xmath147 and @xmath148 and @xmath149 the length of the segment @xmath150 ( @xmath151 for a trapezoid ) , we can express the total number of triangles enclosed by the central polygon as : @xmath152 each @xmath4fold vertex sits at the apex of a diamond - shaped complex of @xmath153 triangles . each @xmath3fold vertex , on the other hand , is at the apex of a triangular region of @xmath154 triangles . the graph is completed by a rectangle of height @xmath155 and base of arbitrary length @xmath156 containing : @xmath157 triangles . the total number of vertices is @xmath158 the final triangulation of the torus is obtained by repeating the prismatic unit cell @xmath159 times and therefore has @xmath160 vertices . this scheme provides direct information on the arrangement of defective sites . thus for instance an @xmath161 unit cell ( see top of fig . [ fig : graph2 ] ) has @xmath3fold vertices separated by two lattice spacings and @xmath4fold vertices by one lattice spacing . on the other hand the integers @xmath120 and @xmath125 giving the chirality of the two segments of the junction @xmath162 are given directly by as : @xmath163 thus the @xmath143 cell of fig . [ fig : graph2 ] is obtained from the junction between a @xmath164 and a @xmath165 tubular segment . ( color online ) central polygon ( right ) and rectangular ( zig - zag ) region in our construction scheme of dunlap s toroids . in this example @xmath161 . ] unit cells for tp(2)@xmath120 and tp(3)@xmath120 toroids . ] dunlap s toroids are not the only examples of defective triangulations of the torus with dihedral prismatic symmetry group @xmath86 . with the help of numerical simulations ( see sec . [ sec:5 ] ) we found two other classes whose unit cell is shown in fig . [ fig : graph4 ] . unlike dunlap s toroids , the @xmath4fold vertices in these prismatic triangulations are not aligned along the internal equator of the torus , but rather grouped in dimers normal to the equatorial plane . @xmath3fold vertices are distributed along the external equator in the graph of fig . [ fig : graph4]a or form a double ring above and below it in the case of the graph fig . [ fig : graph4]b . toroidal deltahedra obtained by embedding the prismatic graphs of fig . [ fig : graph4 ] on a circular torus are shown in fig . [ fig : prismatic - deltahedra ] for the case of a @xmath3fold symmetric toroid with @xmath166 and a @xmath4fold symmetric toroid with @xmath167 . in the rest of the paper we will reserve the symbol tp@xmath120 for dunalp s toroids and refer with tp@xmath120a and tp@xmath120b to the other two classes of toroids with symmetry group @xmath86 and unit cell of as shown in fig . [ fig : graph4]a and fig . [ fig : graph4]b respectively . ( color online ) ( a ) tp5a and ( b ) tp7b toroids with @xmath166 and 49 obtained by repeating the unit cells of fig . [ fig : graph4 ] . @xmath4fold vertices form dimers normal to the equatorial plane while @xmath3fold vertices are ( a ) distributed along the external equator or ( b ) form a double ring above and below the equatorial plane . ] all defective triangulations presented so far are characterized by an even number of disclination pairs . regular tessellations of the torus comprising an odd number of defects pairs are also possible . such tessellations are obtained by combining segments of prismatic and antiprismatic lattices with a consequent loss of dihedral symmetry . [ fig : r_3.3_200 ] shows the voronoi diagram of a toroidal lattice , with @xmath168 and @xmath169 vertices , containing 11 disclination pairs . for an angular length of approximately @xmath170 the lattice is a prismatic @xmath133 toroid while in the remaining @xmath171 the local structure is that of an antiprismatic toroid . the global structure has only bilateral symmetry about a sagittal plane dividing the lattice in two mirror halves and thus point group @xmath172 . ( color online ) voronoi diagram of a toroidal lattice with @xmath168 and @xmath169 vertices . the lattice exhibits 11 disclination pairs and has @xmath172 symmetry group . ] in the past few years , alternative constructions of triangulated tori have been proposed as well as novel geometrical and graph - theoretical methods to express the coordinates of their three - dimensional structures ( see for example kirby @xcite , lszl _ at al _ @xcite , diudea _ et al _ @xcite ) . here we choose to focus on the defect structure associated with the two most important class tp@xmath120 and ta@xmath120 with groups @xmath86 and @xmath87 . in this section we analyze the crystalline structures arising from the solution of the elastic problem and we show how the interplay between the geometry of the embedding torus , the topology of the lattice , and the mechanical properties of the microscopic units , here encoded in the young modulus @xmath27 and the core energy , lead to a rich variety of structures whose phase - diagram is presented at the end of the section . since the free energy of eq . is minimized when disclinations best approximate the continuum gaussian curvature of the torus , it is clear that disclinations are most likely to be found in regions of like - sign gaussian curvature . maximal curvature occurs along the external ( positive ) and internal ( negative ) equators , which thus constitute preferred regions for the appearance of disclinations in the ground state . ( color online ) the screening function @xmath173 for different values of the aspect ratio @xmath174 . ] to analyze the elastic free energy we start by considering the energies of two opposite sign disclinations constrained to lie on the same meridian . the elastic free energy of this system is shown in fig . [ fig : redbluetorus ] as a function of the angular separation between the two disclinations . the energy is minimized for the positive ( @xmath3fold ) disclination on the external equator ( maximally positive gaussian curvature ) and the negative ( @xmath4fold ) disclination on the internal equator ( maximally negative gaussian curvature ) . the picture emerging from this simple test case suggests that a good _ ansatz _ for an optimal defect pattern is a certain number @xmath175 of equally spaced @xmath176 disclinations on the external equator matched by the same number of equally spaced @xmath22 disclinations on the internal equator . we name this configuration with the symbol @xmath177 , where @xmath175 stands for the total number of disclination pairs . @xmath178 where the two pairs of numbers specify the @xmath45 coordinates of the positive and negative disclinations respectively . a comparison of the energy of different @xmath177 configurations , as a function of aspect ratio and disclination core energy , is summarized in the phase diagram of fig . [ fig : phase_diagram1 ] . we stress here that only @xmath177 configurations with @xmath175 even have an embedding on the torus corresponding to lattices of the tp@xmath179 class . nevertheless a comparison with @xmath180odd configurations can provide additional information on the stability of @xmath180even lattices . for small core energies , moreover , thermally excited configurations with a large number of defects and similar @xmath180polar distributions of topological charge are expected to exhibit an elastic energy comparable in magnitude with that of these minimal constructions . ( color online ) elastic energy of a @xmath127 disclination dipole constrained to lie on the same meridian , as a function of the angular separation . in the inset , illustration of a circular torus of radii @xmath7 . regions of positive and negative gaussian curvature have been shaded in red and blue respectively . ] the defect core energy has been expressed here in the form : @xmath181 the core energy @xmath36 of a single disclination depends on the details of the crystal - forming material the corresponding microscopic interactions . a simple phenomenological argument ( see for example ref . ) gives @xmath182 where @xmath100 the lattice spacing . taking @xmath183 , with @xmath184 the area of the torus , yields : @xmath185 for a system of order @xmath186 subunits , then , the dimensionless core energy on the left hand side of eq . is of order @xmath187 . this estimate motivates our choice of the scale for @xmath188 in fig . [ fig : phase_diagram1 ] . ( color online ) phase diagram for @xmath177 configurations in the plane @xmath189 . for @xmath190 $ ] and @xmath191 the structure is given by a @xmath192 configuration with symmetry group @xmath193 . ] for dimensionless core energies below @xmath194 and aspect ratios @xmath174 between 3.68 and 10.12 the ground state structure is the tp5 lattice corresponding to a double ring of @xmath176 and @xmath22 disclinations distributed on the external and internal equators of the torus as the vertices of a regular decagon ( the @xmath192 configuration ) . the tp5 lattice has dihedral symmetry group @xmath193 . that this structure might represent a stable configuration for polygonal carbon toroids has been conjectured by the authors of ref . , based on the argument that the 36@xmath195 angle arising from the insertion of ten pentagonal - heptagonal pairs into the lattice would optimize the geometry of a nanotorus consistently with the structure of the @xmath196 bonds of the carbon network ( unlike the 30@xmath195 angle of the @xmath13fold symmetric configuration originally proposed by dunlap ) . in later molecular dynamics simulations , han @xcite found that a @xmath3fold symmetric lattice , such as the one obtained from a ( 9,0)/(5,5 ) junction ( see fig . [ fig : lucas - torus ] ) , is in fact stable for toroids with aspect ratio less then @xmath197 . the stability , in this case , results from the strain energy per atom being smaller than the binding energy of carbon atoms . irrespective of the direct experimental observation of such disclinated toroidal crystals , which is still open , we have shown here , from continuum elasticity , that a @xmath3fold symmetric lattice indeed constitutes a minimum of the elastic energy for a broad range of aspect ratios and defect core energies . ( color online ) two examples of defect free `` crop circle '' toroids with @xmath198 and @xmath199 ( left ) and @xmath200 ( right ) . ] for small aspect ratios the @xmath3fold symmetric configuration becomes unstable and is replaced by the @xmath201fold symmetric phase @xmath202 . as we mentioned , however , this configuration does nt correspond to a possible triangulation of the torus . it is likely that the ground sate in this regime consists of ten skew disclination pairs as in the antiprismatic ta@xmath120 lattice . the latter can be described by introducing a further degree of freedom @xmath134 representing the angular displacement of defects from the equatorial plane : @xmath203 \left.\left ( ( -1)^{2k}(\pi-\delta\psi),\frac{2\pi k}{n}\right)_{1 \le k \le n } \right\}\end{gathered}\ ] ] a comparison of the tp5 configuration and the ta5 configuration is shown in fig . [ fig : phase_diagram2 ] for different values of @xmath134 . the intersection points of the boundary curves with the @xmath204axis has been calculated by extrapolating the @xmath205 data points in the range @xmath206 $ ] with @xmath207 . for small @xmath134 and @xmath208 $ ] the prismatic tp5 configuration is energetically favored . for @xmath209 , however , the lattice undergoes a structural transition to the ta5 phase . for @xmath210 the prismatic symmetry of the tp5 configuration breaks down again . in this regime , however , the elastic energy of both configurations rapidly rises because of the lower curvature and defects disappear . ( color online ) phase diagram of a @xmath3fold symmetric lattice in the plane @xmath205 . for small @xmath134 and @xmath174 in the range @xmath211 $ ] the prismatic tp5 configuration is energetically favored . for @xmath209 the system undergoes a structural transition to the antiprismatic phase ta5 . ] for aspect ratio @xmath212 the tp5 phase is replaced by a defect free configuration ( @xmath213 in fig . [ fig : phase_diagram1 ] ) so that configurations with defects are no longer energy minima . any toroidal crystal with aspect ratio larger than @xmath214 is than energetically favored to be defect - free . in the thin torus limit the ground state structure is directly related to the simple problem of finding the most efficient packing of congruent equilateral triangles on the torus of a given aspect ratio . given @xmath113 subunits ( vertices ) one seeks the densest packing of equilateral triangles of edge - length @xmath215 on the torus with aspect ratio @xmath174 , such that each vertex has valence six . using the planar construction described in sec . [ sec:3 ] , the optimal choice of the indices @xmath88 , can be translated into the minimization of the following quantity : @xmath216 obtained by equating the magnitude of the chiral vector @xmath91 with that of the sectional circumference of the embedding torus , under the constraints : @xmath217 n,\,m,\,l\in\mathbb{z } \end{array } \right.\,.\ ] ] this construction successfully predicts the structure of the lattices of fig . [ fig : cropcircle ] . so far we have studied the elasticity of toroidal crystals exclusively in terms of interacting topological defects on a rigid toroidal substrate . thus the elastic strain due to defects and curvature takes the form of pure stretching on the tangent plane of the torus and no out - of - plane deformation takes place . in a more realistic scenario , a crystalline torus would undergo both in - plane stretching and out - of - plane bending . the latter implies an energy cost : @xmath218 with @xmath219 the mean curvature ( see ref . ) . the case of defect - free tori is simple enough to incorporate bending in the problem and see what the optimal aspect ratio of a defect - free torus would be as a function of the fppl - von krmn number @xmath220 representing the ratio of the stretching energy scale to the bending rigidity . in absence of defects the only source of stress is given by the curvature . thus @xmath221}{2r^{2}}\notag\\ & \hspace{0.47\columnwidth}+\operatorname{li_{2}}(\alpha^{2})-2\big\}\,.\label{eq : stretching}\end{aligned}\ ] ] summing eq . and and taking the derivative with respect to @xmath174 ( assuming constant area ) , one obtains the following equation for the optimal value of @xmath174 : @xmath222=0\,.\end{gathered}\ ] ] the optimal aspect ratio @xmath174 as obtained form eq . is shown in fig . [ fig : optimalr ] as a function of @xmath223 . for @xmath224 , when the major contribution to the elastic energy is given by the bending , the optimal geometry is given by the clifford torus ( @xmath225 ) . if , on the other hand , the in - plane stretching dominates , a `` skinny '' torus ( large @xmath174 ) is energetically favoured . ( color online ) optimal value of the aspect ratio @xmath174 as function of the fppl - von krmn number @xmath220 . for @xmath224 the clifford torus with @xmath225 is optimal . larger values of @xmath223 favour instead a `` skinnier '' torus . ] in this section we investigate the possibility of a regime of coexistence between isolated disclinations and grain boundary `` scars '' . the existence of scars , first predicted in the context of spherical crystallography @xcite and later observed experimentally in spherical droplets coated with colloids @xcite , has become one of the fundamental signatures of dense geometrically frustrated systems . in the regime of large particle numbers , the amount of curvature required to screen the stress field of an isolated disclination in units of lattice spacing becomes too large and disclinations are unstable to grain boundary `` scars '' consisting of a linear array of tightly bound @xmath127 pairs radiating from an unpaired disclination @xcite . in a manifold with variable gaussian curvature this effects leads to a regime of coexistence of isolated disclinations ( in regions of large curvature ) and scars . in the case of the torus the gaussian curvature inside ( @xmath226 ) is always larger in magnitude than that outside ( @xmath227 ) for any aspect ratio and so we may expect a regime in which the negative internal curvature is still large enough to support the existence of isolated @xmath4fold disclinations , while on the exterior of the torus disclinations are delocalized in the form of positively charged grain boundary scars . to check this hypothesis we compare the energy of the tp5 lattice previously described with that of `` scarred '' configurations obtained by decorating the original toroid in such a way that each @xmath176 disclination on the external equator is replaced by a @xmath228 mini - scar . the result of this comparison is summarized in the phase diagram of fig . [ fig : phase_diagram3 ] in terms of @xmath174 and the number of vertices of the triangular lattice @xmath113 ( the corresponding hexagonal lattice has twice the number of vertices , i.e. @xmath229 ) . @xmath113 can be derived from the angular separation of neighboring disclinations in the same scar by approximating @xmath230 , with @xmath231 the area of a hexagonal voronoi cell and @xmath100 the lattice spacing . when the aspect ratio is increased from 1 to 6.8 the range of the curvature screening becomes shorter and the number of subunits required to destroy the stability of the tp5 lattice decreases . for @xmath232 , however , the geodesic distance between the two equators of the torus becomes too small and the repulsion between like - sign defects takes over . thus the trend is inverted . ( color online ) isolated defects and scar phases in the @xmath233 plane . when the number of vertices @xmath113 increases the range of the screening curvature becomes smaller than one lattice spacing and disclinations appear delocalized in the form of a @xmath228 grain boundary mini - scar . ] in this section we report the result of a numerical minimization of a system of @xmath113 point - like particles constrained to lie on the surface of a torus and interacting via a pair potential of the form @xmath234 where @xmath235 denotes the euclidean distance in @xmath236 . the problem of finding the minimal energy configuration of repulsively interacting points on a 2-manifold has become a standard problem of potential theory and has its paradigm in the classical thomson problem on the sphere . the choice of the cubic potential is motivated here by the so called `` poppy seed bagel theorem '' @xcite , according to which the configuration of points that minimizes the riesz energy @xmath237 on a rectifiable manifold of hausdorff dimension @xmath238 is uniformly distributed on the manifold for @xmath239 . in the case of a torus of revolution this implies that for small @xmath240 the points are mostly distributed on the exterior of the torus ( the interior becomes completely empty in the limit @xmath241 ) . as @xmath240 is increased , however , the points cover a progressively larger portion of the surface . the distribution becomes uniform for @xmath242 . on the other hand , since the number of local minima of the riesz energy increases with @xmath240 , it is practical to choose a value not much larger than two . the choice @xmath243 has the further advantage of modelling a real physical system of neutral colloidal particles assembled at an interface @xcite and is therefore suitable for direct comparison with experiments on colloidal suspensions . to construct low energy configurations we adopt an carefully designed hybrid optimization algorithm named tapping ( ta ) . like other hybrid algorithms , ta consists in a combination of fast local optimizations and global stochastic moves designed to release the system from the local minimum to which it is confined at the end of a local minimization step . a more detailed description of our algorithm is reported in appendix [ app:3 ] . we study four different aspect ratios : @xmath244 and @xmath245 . for each aspect ratio we consider several different particle numbers up to @xmath246 and each simulation is performed for @xmath247 to @xmath248 ta iterations . the lowest energies found , as well as the number of defects in the corresponding configuration , are reported in table [ tab : minima ] for a selected set of systems . the corresponding lattices are shown in fig . [ fig : numeric_tori ] the lattices are best presented using a voronoi construction corresponding to the dual lattice of the delaunay triangulation . here pentagonal faces are colored in red while heptagonal faces are colored in blue . the complete set of data produced in our simulations together with a collection of interactive 3d graphics for each low energy configuration found is available on - line @xcite . for fewer than @xmath249 particles the results of our numerical minimization are in good agreement with the continuum elastic theory . in particular for @xmath250 and @xmath251 and @xmath252 , we always find minimal energy configurations consisting of ten @xmath3fold disclinations on the outside of the torus and ten @xmath4fold disclinations in the inside as predicted by the elastic theory in the regime of @xmath253 . for @xmath198 and @xmath254 we also find the lowest energy configurations to be defect free . for small numbers of particles we do nt expect the continuum approximation to accurately describe the lowest energy structure of the toroidal clusters presented in fig . [ fig : numeric_tori ] . loosely speaking the limit of validity of the elastic theory can be quantified by requiring the average lattice spacing @xmath255^{1/2}$ ] to be much smaller than the radius @xmath24 of the torus . this condition requires @xmath113 to be of order @xmath256 particles for a torus with aspect ratio @xmath257 . remarkably , good agreement between the theory and simulations is found starting from much smaller values of @xmath113 and in some cases ( see the following discussion on the configuration with @xmath257 and @xmath258 ) , we already observe the onset of the ideal behavior predicted by theory for @xmath259 . the occurrence of a ground state configuration with exact prismatic or antiprismatic symmetry , in particular , is only possible when the number of particles @xmath113 belongs to a specific sequence of `` magic numbers '' described in sec . [ sec:3 ] . nevertheless for @xmath113 outside such a sequence it is still possible to observe in the ground state a predominant prismatic or antiprismatic character depending on the aspect ratio . .[tab : antiprism]maximum and minimum aspect ratio for which the toroidal antiprisms are a global minimum . [ cols="^,^,^",options="header " , ] @xmath260 ) an interesting feature is observed at @xmath261 . as in the case of @xmath257 we also find a minimum with @xmath262 symmetry group , but unlike the latter configuration , it belongs to the tp@xmath252a class and has @xmath3fold disclinations along the external equator . a tp@xmath263b configuration is obtained again for @xmath167 . for @xmath264 and @xmath265 the global minimum is achieved by two spectacular antiprismatic configurations with @xmath266 and @xmath267 symmetry group respectively . these toroids can be obtained from the toroidal antiprisms discussed in sec . [ sec:3 ] by splitting in two such that the total number of vertices is increased by one . the two newly created vertices have coordination number @xmath268 , with @xmath149 is the coordination number @xmath269 , while two of the @xmath149 neighbors of @xmath269 gain a bond . ] one or more times the initial set of @xmath3fold vertices . thus starting from a @xmath270fold toroidal antiprism with @xmath271 vertices and splitting all @xmath272 @xmath3fold vertices one obtains @xmath273 vertices . splitting twice all @xmath274 @xmath3fold vertices of a @xmath275fold toroidal antiprism with @xmath276 , on the other hand , we have @xmath277 . for @xmath278 the global minimum is represented by a fascinating lattice of tp@xmath279a type . lattices with @xmath280 and @xmath281 resemble very closely the structure of a tp@xmath279 graph while for @xmath282 the lattice has a more antiprismatic character with ten defect pairs . @xmath119 ) three defect - free configurations are found at @xmath283 and @xmath284 . the case @xmath118 is a particular example of a defect - free lattice that can not be obtained form the planar construction reviewed in sec . [ sec:3 ] . it consists of @xmath285 octahedra connected in the form of a chain . since each octahedron is attached to other two , it contributes with six faces to the total face count . thus @xmath286 and @xmath287 . for @xmath288 we always find configurations with ten disclination pairs as expected from continuum elasticity . for @xmath289 the regime of coexistence between isolated disclinations and scars described in sec . [ sec : scars ] is observed . the delocalization of isolated disclinations into scars , however , does nt take place at each defective site simultaneously and the regime of coexistence between positively charged scars and isolated @xmath4fold disclinations is preceded by a phase with isolated @xmath3 and @xmath4fold disclinations and scars . for tori with aspect ratio as large as @xmath198 we find defect - free ground states every time it is possible to construct a purely @xmath13valent toroidal graph with the same number of vertices @xmath113 . we have seen that disclination defects , forbidden in the lowest energy state of a planar crystal , may be energetically favored on a substrate of non - vanishing gaussian curvature . it is therefore natural to ask whether large curvature can completely destroy crystalline order by driving the proliferation of a sufficiently high density of defects . the resulting state would be amorphous . the problem of generating amorphous structures by tiling a two - dimensional curved space with identical rigid subunits has drawn attention over the years , particularly through the connection to the structure of such disordered materials as supercooled liquids and metallic glasses . since the work of frank @xcite the notion of geometrical frustration arises frequently in investigations of supercooled liquids and the glass transition . a paradigmatic example is represented by the icosahedral order in metallic liquids and glasses which , although locally favored , can not propagate throughout all of three - dimensional euclidean space . a two - dimensional analog , consisting of a liquid of monodisperse hard disks in a 2-manifold of constant negative gaussian curvature ( the hyperbolic plane ) was first proposed by nelson and coworkers in 1983 @xcite . in such a system the impossibility of covering the entire manifold with a 6-fold coordinated array of disks mimics many aspects of the geometrical frustration of icosahedral order in three dimensions . in all these models of geometrical frustration , however , the origin of the disorder is primarily due to the short - range nature of the potential between the subunits . in a more realistic setting , part of the frustration is relieved by the fact that hexagonal unit cells can compress in order to match the underlying geometry . the embedding of a triangular lattice on an axisymmetric torus , provides a particularly suitable playground to study curvature - driven disorder . when @xmath290 the gaussian curvature on the inside of the torus grows like @xmath291 and diverges on the internal equator at @xmath292 . we thus expect a high density of defects in the vicinity of the curvature singularity and a resultant loss of the local @xmath13fold bond orientational order . in this regime the system will have crystalline regions on the outside of the torus and amorphous regions near the curvature singularity . ( color online ) top view of a defect free triangulation of a fat torus with @xmath293 and @xmath294 . the corresponding elastic energy becomes very high in the interior of the torus where the triangles are more compressed to match the reduction of surface area . ] in this section we substantiate this claim analytically based on the elastic theory of continuous distributions of edge dislocations on a `` fat '' torus . our argument is based on the following construction . as a consequence of the curvature singularity the surface area of an arbitrary wedge of angular width @xmath295 becomes smaller and smaller as the sectional angle @xmath296 increases and vanishes at @xmath292 . if a defect - free lattice is embedded on such a wedge , bragg rows will become closer and closer as the singularity is approached with a consequent rise in the elastic energy ( see fig . [ fig : fat_torus ] ) . an intuitive way to reduce the distortion of the lattice is to recursively remove bragg rows as one approaches the point @xmath292 ( see fig . [ fig : dislocation ] ) . this is equivalent to introducing a growing density of edge dislocations . this dislocation `` cloud '' will ultimately disorder the system by destroying the local @xmath13fold bond orientational order . one might therefore view the curvature as playing the role of a local effective temperature which can drive `` melting '' by liberating disclinations and dislocations . in two - dimensional non - euclidean crystals at @xmath297 , however , the mechanism for dislocation proliferation is fundamentally different from the usual thermal melting . while the latter is governed by an entropy gain due to unbinding of dislocation pairs , the amorphization at @xmath297 is due to the adjustment of the lattice to the geometry of the embedding manifold via the proliferation of defects and the consequent release of elastic stress . a similar phenomenon occurs in the disorder - driven amorphization of vortex lattices in type - ii and high-@xmath298 superconductors @xcite . ( color online ) a schematic example of a dislocation pile - up on a square lattice resulting from the shrinking of the area on a regular wedge of a fat torus . ] since the shrinking area per plaquette on the inside of the torus necessitates a high density of dislocations we may approximate the dislocation cloud in this region by a continuous distribution of burgers vector density @xmath299 . minimizing the elastic energy with respect to @xmath299 yields a variational equation from which the optimal dislocation density can be calculated as a function of the ratio @xmath300 between the dislocation core energy @xmath301 and elastic energy scale @xmath302 with @xmath303 . as a starting point , we calculate the green function @xmath40 in the fat torus limit @xmath290 ( i.e. @xmath304 and @xmath305 , see eq . ) . the conformal angle @xmath51 in this limit is : @xmath306 and to leading order of @xmath307 we have : @xmath308 \frac{\kappa}{4\pi^{2}}\operatorname{re}\{\operatorname{li_{2}}(\alpha e^{i\psi})\ } \rightarrow \frac{\kappa}{16\pi^{2}}\,\psi^{2}+\frac{1}{2\pi^{2}}\log\left(\cos\frac{\psi}{2}\right)\,.\end{gathered}\ ] ] to handle the limit of the jacobi theta function we can take @xmath309 , @xmath310 and calculate the limit @xmath311 . this can be done by using the modular transformation properties of jacobi functions@xcite : @xmath312 thus @xmath313 , @xmath314 and @xmath315 , where : @xmath316 this is easily evaluated by means of the expansion : @xmath317 taking the logarithm and neglecting irrelevant constant terms , we obtain : @xmath318 which finally leads to : @xmath319 with @xmath320 . with the green function in hand , we can calculate the effect of the curvature singularity at @xmath292 on the distribution of defects . let @xmath299 be the burgers vector density of the dislocation cloud . hereafter we work in a local frame , so that : @xmath321 with @xmath322 a basis vector in the tangent plane of the torus whose points are specified by the three - dimensional euclidean vector @xmath323 . the quantity @xmath299 has to be such that : @xmath324 with @xmath325 the total burger s vector in a generic domain @xmath326 . because on a closed manifold dislocation lines can not terminate on the boundary , extending the integration to the whole torus we have : @xmath327 since the basis vectors @xmath328 in eq . have the dimension of length , contravariant coordinates @xmath329 have dimensions of an inverse area . assuming all defects to be paired in the form of dislocations ( i.e. @xmath330 everywhere ) , the total energy of the crystal reads : @xmath331 where @xmath301 is the dislocation core energy and @xmath332 the function @xmath28 encoding the elastic stress due to the curvature and the screening contribution of the dislocation cloud obeys @xmath333 where @xmath334 is the levi - civita antisymmetric tensor on the torus : @xmath335 the stress function @xmath28 can be expressed in the form @xmath336 with @xmath337 + 1\,,\label{eq : fat_gamma_screening}\\[7pt ] \frac{\gamma_{s}(\psi,\phi)}{y } = \int d^{2}y\,\epsilon_{k}^{i}\nabla_{i}b^{k}(\bm{y})\,g_{l}(\bm{x},\bm{y})\,.\label{eq : fat_gamma_defects}\end{gathered}\ ] ] taking advantage of the closeness of the torus we can integrate eq . by parts so that : @xmath338 now we want reduce the integral term in eq . to a more friendly functional of @xmath299 , suitable for a variational approach . given the azimuthal symmetry we assume that all dislocations are aligned along @xmath339 . even though not necessarily true , we argue this to be a reasonable work hypothesis as well as a solid starting point to capture the essential physics of the fat limit . in this case @xmath340 can be recast in the form @xmath341\,.\end{gathered}\ ] ] substituting eq . and in eq . and minimizing with respect to @xmath342 we can now write the variational equation : @xmath343 by inverting the order of integration in the integral on the right hand side , eq can be expressed in the form of a fredholm equation of the second kind : @xmath344 where @xmath345 , @xmath346 and the kernel @xmath347 is given by : @xmath348 the function @xmath349 on the right hand side of eq . is given by : @xmath350\sin\psi-2\operatorname{cl_{2}}(\psi+\pi)\bigr\}\end{gathered}\ ] ] where @xmath351 is the clausen function ( see ref . 1005 - 1006 ) defined as : @xmath352 ( color online ) the burgers vector component @xmath353 for different choices of @xmath354 . ] as previously noted the dislocation core energy is @xmath301 is much smaller than the elastic energy scale @xmath302 . eq is then suitable to be solved in powers of the dimensionless number @xmath354 : @xmath355 the corrections to the zero - order term @xmath356 can be calculated recursively by solving a set of fredholm equations of the first kind : @xmath357 the function @xmath356 associated with the burgers vector density of the dislocation cloud in the limit @xmath358 , on the other hand , can be calculated directly from eq . by setting the effective topological charge density on the right hand side to zero : @xmath359 for a torus of revolution the only nonzero christoffel symbols are @xmath360 \gamma_{\phi\phi}^{\psi } = r_{2}^{-1}\sin\psi(r_{1}+r_{2}\cos\psi)\,.\end{gathered}\ ] ] since @xmath361 by assumption , the first term in eq . can be expressed as : @xmath362 & = ( 1+\cos\psi)\partial_{\psi}b^{\phi}-2\sin\psi\,b^{\phi}\,,\end{aligned}\ ] ] and eq . becomes an ordinary differential equation @xmath363 whose solution is given by @xmath364 so that @xmath365 . the burgers vector density @xmath342 obtained by a numerical solution of eq . is shown in fig . [ fig : burgers ] for different values of @xmath354 . the burgers vector density is measured in units of @xmath366 the function @xmath342 has cubic singularities at @xmath367 and is approximately zero on the outside of the torus . the solid blue curve in fig . [ fig : burgers ] represents the zeroth order solution of eq .. now , in the theory of dislocation mediated melting a system at the solid liquid phase boundary is described as a crystalline solid saturated with dislocations . in three - dimensions , in particular , there is a strong experimental evidence of the existence of a critical dislocation density at the melting point @xmath368 where @xmath101 is the length of the length of the smallest perfect - dislocation burgers vector @xcite . several theoretical works have motivated this evidence both for three - dimensional solids and vortex lattices in super conductors @xcite . on the other hand , given the existence of such a critical density , its value can be empirically used to determine whether a system is in a solid or liquid - like phase in the same spirit as the lindemann criterion . with this goal in mind we can calculate the dislocation density by requiring @xmath369 with @xmath370 the density of single lattice spacing dislocations . this yields : @xmath371 solving @xmath372 as a function of @xmath296 and @xmath113 we obtain the diagram of fig . [ fig : melting - diagram ] . as expected the inside of the torus contains an amorphous region whose angular size decreases with the number of vertices @xmath113 as a consequence of the reduction of the lattice spacing . ( color online ) phase diagram for curvature driven amorphization . the inside of the torus contains an amorphous region whose angular size decreases with the number of vertices @xmath113 as a consequence of the reduction of the lattice spacing . ] in this article we reported a comprehensive analysis of the ground state properties of torodial crystals . using the elastic theory of defects on curved substrates we identified the ground state structures of an arbitrary crystalline torus as a function of the aspect ratio @xmath174 and the ratio @xmath188 of the defect core energy to the elastic energy scale . we showed that for a large range of aspect ratios and core energies the minimal energy structure of a toroidal crystal has ten disclinations pairs and symmetry group @xmath193 , as first conjectured by lambin _ et al _ more than ten years ago @xcite . for large system sizes we proved isolated disclinations are unstable to grain - boundary scars consisting of chains of tightly bound @xmath127 pairs radiating form an unpaired disclination . on a torus , where the gaussian curvature on the inside is always larger in magnitude than that on the outside , the occurrence of scars is marked by a state featuring isolated @xmath4fold disclinations in the inside of the torus together with scars on the outside . for a torus of aspect ratio close to one we showed how a diverging gaussian curvature on the internal equator is responsible for the remarkable occurrence of a curvature - driven transition of the system to a disordered , liquid - like , state . the predictions of our elastic theory were compared with the results of a numerical study of a system of point - like particles constrained on the surface of a torus and interacting via a short range potential , with good agreement . from a purely geometrical point of view we have introduced a number of novel toroidal polyhedra as well as a new construction and classification scheme for certain types of prismatic tori . this project was supported by nsf grants dmr-0305407 and dmr-0705105 and by an allocation through the teragrid advanced support program . lg is supported on a graduate fellowship by the syracuse biomaterials institute . we acknowledge amand lucas and antonio fonseca for sharing with us the beautiful photos presented in sec.[sec:3 ] . we also would like to thank eric j. west and m. cristina marchetti for inspiring discussions during the development of sec . [ sec:6 ] . lg is very grateful to dario giuliani who made possible the realization of the _ toroidal lattices database _ @xcite , the use of which has been crucial for obtaining many of the results and the ideas presented in this paper . the modified laplacian green function on a periodic rectangle of edges @xmath69 and @xmath70 can be conveniently calculated in the form : @xmath373 where @xmath374 is the eigenfunction of the laplace operator with periodic boundary conditions : @xmath375 such that : @xmath376 u_{\lambda}(\xi,0 ) = u_{\lambda}(\xi , p_{2 } ) \end{array } \right.\,.\ ] ] in cartesian coordinates the eigenfunctions are simple plane waves of the form : @xmath377 where @xmath378 and @xmath379 are given by : @xmath380 and the eigenvalue @xmath354 is given by : @xmath381 calling for simplicity @xmath382 and @xmath383 , the function @xmath384 is given by : @xmath385 although is simple , it is very useful to rewrite it in terms of elliptic functions . noting that the odd terms in cancel we have : @xmath386 & = -\frac{2}{p_{1}p_{2 } } \left [ \sum_{m=1}^{\infty}\frac{\cos\frac{2\pi m}{p_{2}}(y-\eta)}{\left(\frac{2\pi m}{p_{2}}\right)^{2}}+ \sum_{n=1}^{\infty}\sum_{m=-\infty}^{\infty}\frac{\cos\frac{2\pi n}{p_{1}}(x-\xi)\cos\frac{2\pi m}{p_{2}}(y-\eta ) } { \left(\frac{2\pi n}{p_{1}}\right)^{2}+\left(\frac{2\pi m}{p_{2}}\right)^{2 } } \right]\,.\end{aligned}\ ] ] an equivalent expression can be obtained by isolating the @xmath108 contribution in the sum rather than the @xmath387 term . the first sum in eq . can be evaluated easily by using : @xmath388 thus we have : @xmath389 & = -\frac{1}{2\,p_{1}p_{2}}\left(\frac{p_{2}^{2}}{6}-p_{2}|y-\eta|+|y-\eta|^{2}\right)\ , . \end{aligned}\ ] ] the second sum in eq . can be evaluated with the help of the poisson summation formula : @xmath390 in particular , if we choose : @xmath391 we can write the second sum in eq . as : @xmath392 & = -\frac{2}{p_{1}p_{2 } } \sum_{n=1}^{\infty}\cos\frac{2\pi n}{p_{1}}(x-\xi ) \sum_{k=-\infty}^{\infty}\int_{-\infty}^{\infty}dt\,\frac{\cos(2\pi k+p)t } { \left(\frac{2\pi n}{p_{1}}\right)^{2}+\left(\frac{2\pi t}{p_{2}}\right)^{2}}\end{aligned}\ ] ] the integral can be easily calculated by considering : @xmath393 thus eq . becomes : @xmath394 the sum in @xmath120 can be calculated by noting : @xmath395 where @xmath396 with @xmath397 and arbitrary choice of sign . thus separating positive and negative values of @xmath398 in eq . and using eq . the function @xmath399 takes the final form : @xmath400 + \sum_{k=1}^{\infty}\log\left|1 - 2q^{2k}\cos\frac{2\pi}{p_{1}}(z-\zeta)+q^{4k}\right|\,,\end{gathered}\ ] ] where : @xmath401 the second term in eq . can be expressed in terms of the jacobi theta function @xmath402 defined as : @xmath403 another useful relation can be obtain by taking the derivative of @xmath402 with respect to @xmath404 : @xmath405 thus we have : @xmath406 in this way we can express : @xmath407\,.\ ] ] then taking : @xmath408 and substituting in eq . we obtain : @xmath409 -\frac{\pi}{p_{1}}|y-\eta| + \log\left|\frac{\vartheta_{1}(u , q)}{\vartheta_{1}'^{\frac{1}{3}}(0,q)}\right|\,.\end{gathered}\ ] ] combining eq . with eq . we conclude that : @xmath410 & = \frac{\log 2}{6\pi } - \frac{1}{2\,p_{1}p_{2}}|y-\eta|^{2 } + \frac{1}{2\pi}\log\left|\frac{\vartheta_{1}(u , q)}{\vartheta_{1}'^{\frac{1}{3}}(0,q)}\right|\,.\end{aligned}\ ] ] an alternative notation frequently used for the jacobi theta function is : @xmath411 with this choice we can write the green function in the final form : @xmath412 + \frac{1}{2\pi}\log\left|\frac{\vartheta_{1}(\frac{z-\zeta}{p_{1}/\pi}|\frac{ip_{2}}{p_{1 } } ) } { \vartheta_{1}'^{\frac{1}{3}}(0|\frac{ip_{2}}{p_{1}})}\right|\,.\end{gathered}\ ] ] in this appendix we derive the analytical expression for the stress functions @xmath413 and @xmath80 in eq . . the former is given by the integral : @xmath414\ ] ] it is convenient to keep the green function @xmath62 in the form of eq . . thus we have : @xmath415 which using eq . and carrying out the integrals yields : @xmath416\ ] ] to calculate the second integral in eq . it is convenient to invert the order of integration and use the result of eq . : @xmath417 & = \log\left[\frac{4(r^{2}-1)}{(r+\sqrt{r^{2}-1})^{2}}\right]-\frac{r-\sqrt{r^{2}-1}}{r}\,.\end{aligned}\ ] ] combining eq . and we obtain : @xmath418 + \frac{r-\sqrt{r^{2}-1}}{r}\,.\ ] ] the elastic stress produced at the point @xmath81 by a disclination at @xmath34 is given by the function : @xmath419 to calculated the average @xmath420 one can again start from @xmath62 in the form of a series and use eq . . this yields @xmath421 to calculate the integrals in eq . it is convenient to expand the conformal angle @xmath51 in fourier harmonics : @xmath422 with : @xmath423\ ] ] where @xmath424 . thus we have : @xmath425 with @xmath426 where @xmath427 stands for the real part and @xmath428 is the usual euler s dilogarithm ( see ref . pp 1004 - 1005 ) . the second integral in eq . is give by : @xmath429\,.\end{gathered}\ ] ] to calculate the integral in eq . one uses parseval s identity : @xmath430 where @xmath431 an arbitrary square - integrable function on the interval @xmath432 $ ] with fourier series : @xmath433 thus we have : @xmath434\,.\end{aligned}\ ] ] using eq . , and we conclude that : @xmath435 & -\frac{\kappa}{8\pi}\operatorname{li_{2}}(\alpha^{2 } ) + \frac{\kappa}{4\pi^{2}}\operatorname{re}\{\operatorname{li_{2}}(\alpha e^{i\psi})\}\notag\\[5pt ] & + \frac{1}{4\pi^{2}r}\log\left[\frac{2(r^{2}-1)}{(r+\cos\psi)(r+\sqrt{r^{2}-1})}\right]\,.\end{aligned}\ ] ] the tapping algorithm ( ta ) is a hybrid algorithm designed to find the optimal crystalline structure of particle systems constrained to lie on a curved surface and interacting with a long range potential of the form @xmath436 . hybrid algorithms , such as basin - hopping @xcite and minima - hopping @xcite , have been successfully employed throughout the years to predict the crystalline structure of molecular clusters and proteins . in general they combine fast local minimizations with global moves whose goal is to release the system from the local minimum it is confined at the end of a local minimization step . a typical hybrid optimization routine can be summarized in the following two steps : _ 1 _ ) after all the independent variables have been randomly initialized , a local optimization is performed and a local minimum @xmath437 is determined ; _ 2 _ ) from @xmath437 a new configuration @xmath438 is constructed by applying a global ( generally stochastic ) move . the new configuration @xmath438 is then used as starting point for a new local optimization step . the two steps are iterated until a stopping criterion is satisfied . the goal is thus to explore the largest possible number of local minima and avoid visiting the same minimum too often . the crucial point in designing an effective hybrid algorithm is clearly the choice of the global move . there is no general rule to identify a successful global transformation @xmath439 and physical intuition and prior experience are typically the only guidelines . in the case of basin - hopping , for instance , the global transformation consists in a monte carlo move in which all the particles of the system are randomly displaced in order to construct a new initial configuration from which a new trial minimum is obtained . the step is accepted with probability @xmath440 , where @xmath441 is the energy difference between the new and previous minimum and @xmath442 is an inverse temperature adjusted to obtain a @xmath443 acceptance ratio . in the case of minima - hopping the escape step is performed by a short molecular dynamics simulation by assigning the particles a fixed kinetic energy . the global move adopted in ta is inspired by the process of close packing of spherical objects by tapping and is motivated by the well established role of topological defects in determining the order of two - dimensional non - euclidean crystals as well as the picture of the potential energy surface ( pes ) of such systems as a multi - funnel landscape . consider a system of say spherical objects confined in a two - dimensional box with an initial disordered configuration . a common way to bias the system toward a close - packed configuration is to provide it kinetic energy by gently tapping the box . if the system is populated by locally ordered regions ( i.e. grains ) separated by clusters of defects , the primary effect of tapping is to produce a glide of defects inside the crystals with a subsequent rearrangement of grains . this mechanism can be reproduced numerically in the following way . the algorithm starts with a random distribution of particles and rapidly quenches the system by performing a fast local minimization . once particles are trapped in a local minimum , defects are identified by a delaunay triangulation of the lattice . then the system is tapped by adding to the defect positions a random displacement . the magnitude of the displacement is given by the typical spacing associated with the particles number times a factor @xmath354 which represents the tapping strength . this factor is initially set to @xmath444 . after defects have been moved a new local minimization is performed in order to construct the trial configuration @xmath438 . the energy of this configuration is compared with the energy of the previous minimum and the move is accepted if their difference is larger than some tolerance factor @xmath445 . if , on the other hand , the energy difference is smaller than @xmath445 , the system has relaxed again to the same minimum . in this case the tapping strength is increased of a factor 10 and the process is repeated until the system successfully hops to a new minimum . the tapping strength @xmath354 is then set to its initial value . the process is iterated until the rate of discovery of new global minima drops below some threshold value or a maximum number of iterations is reached . in the current implementation of the algorithm , the local minimization step is performed using the fletcher - reeves conjugate gradient algorithm @xcite . analytic expressions for the energy gradient and the hessian matrix are coded in the program in order to reduce the number of evaluations of the objective function during the relaxation step to one single event . the delaunay triangulation is calculated via the dwyer s divide and conquer algorithm with alternate cuts @xcite , which runs in @xmath446 time , making the identification of the defects particularly fast . the main difference between ta and other hybrid algorithms ( including basin - hopping ) is that the escape move consists of adaptive displacements of defects only , rather than of the entire system . in the case of non - euclidean crystals , where the conformation of the energy landscape is subtly related to the arrangement of topological defects , this mechanism is believed to explore the pes more accurately . in systems as lennard - jones clusters or spin - glasses , the pes is characterized by an exponential number of local minima separated by energy barriers . for this reason the majority of the algorithms are specifically designed to allow the system to overcome a barrier by providing it a significant amount of energy . if the energy landscape , however , is characterized by the presence of multiple narrow funnels , as believed in this case , the previous methods become ineffective . a funnel represents the basin of attraction of a given local minimum . if the global minimum is also located at the bottom of a funnel , an algorithm that is attempting to locate it via a sequence of local minimization steps has a chance to find it exclusively by starting from a configuration already at the muzzle of the funnel . such possibility , however , is ruled out if all the particles are displaced simultaneously during the escape move and the system is abruptly moved to a completely different place in the energy landscape . on the other hand , by adaptively tapping the defects it is possible to achieve a much finer inspection of the pes and possibly locate the funnel associated with the global minimum . a copy of our code is available by request . whitesides and m. boncheva , _ beyond molecules : self - assembly of mesoscopic and macroscopic components _ , proc . usa * * 9**9 , 4769 - 4774 ( 2002 ) [ www.pnas.org/cgi/doi/10.1073/pnas.082065899 ] . a. r. bausch , m. j. bowick , a. cacciuto , a. d. dinsmore , m. f. hsu , d. r. nelson , m. g. nikolaides , a. travesset and d. a. weitz , science * 299 * , 1716 ( 2003 ) . t. einert , p. lipowsky , j. schilling , m. j. bowick and a. bausch , langmuir * 21 * , 12076 ( 2005 ) . d. r. nelson , nano lett . * 2 * , 1125 ( 2002 ) .

a detailed description of flow in porous media is essential for building mathematical models with applications in , for example , water management , oil and gas recovery , carbon dioxide ( co@xmath0 ) sequestration or nuclear waste disposal . in order to set up a reliable numerical model , one needs to have a good knowledge of the problem geometry and input parameters . for example , the flow of water in granite rock , which is a suitable site for nuclear waste disposal , is driven by the complex system of vugs , cavities and fractures with various topology and sizes . these alter the effective permeability , and therefore should be accurately accounted for in the numerical model . there are two main approaches : either the fractures are considered as free - flow regions , or the fractures contain debris and are also modelled as porous media with specific permeabilities . in the first case , a unified approach to modelling free - flow and porous media regions can be provided by the so called stokes - brinkman equation , which reduces to either the stokes or darcy model in certain parameter limits , e.g. , within the multiscale mixed finite - element ( msmfe ) framework @xcite . in this paper , we consider the latter case , and apply darcy s law to the flow in the reservoir and in the fractures as well ; see @xcite for a related approach . in either case , the preferential flow in large geological dislocations and their intersections should be considered as two- and one - dimensional flows , respectively . due to the quite complex structure of the domains , the discretisation is performed using finite element methods ( fem ) . the resulting meshes are therefore unstructured , and they combine different spatial dimensions ( line elements in 1d , triangles in 2d , and tetrahedrons in 3d ) . the systems of linear equations obtained from the fem discretisation are often very large , so that using direct methods is prohibitive and iterative solvers are warranted . the systems are typically also bad - conditioned due to the mixing of spatial dimensions , large jumps in permeability coefficients and presence of elements of considerably different sizes , and so they are challenging for iterative solvers as well . the matrices have a saddle - point structure @xmath1{cc}a & \ol{b}^{t}\\ \ol{b } & -\ol{c } \end{array } \right ] , \label{eq : sp - matrix}\ ] ] where @xmath2 is symmetric positive definite on the kernel of @xmath3 , @xmath4 is symmetric positive semi - definite , and it is positive definite on the kernel of @xmath5 . the ` penalty ' block @xmath6 arises from connecting meshes of different spatial dimensions . the iterative solution of systems with this structure is a frequently studied topic ; see , for example , @xcite , the monographs @xcite or toselli-2005-ddm and the references therein . however , efficient methodologies for solving saddle - point problems are typically problem dependent . in this paper , we develop a robust and scalable solver for linear systems with the saddle - point structure as in ( [ eq : sp - matrix ] ) with the block@xmath7 either zero or nonzero . the solver is tailored to the mixed - hybrid formulation of flow in porous media using the lowest order raviart - thomas ( @xmath8 ) finite elements with combined mesh dimensions ( 1d , 2d and 3d ) . in particular , we adapt the balancing domain decomposition by constraints ( bddc ) method to this type of problems . the bddc method is currently one of the most popular methods of iterative substructuring . it has been proposed independently in @xcite ; see @xcite for the proof of equivalence . even though bddc has been originally formulated for elliptic problems , it has been successfully extended , for example , beyond elliptic cases @xcite and to multiple levels @xcite . an optimal set - up has been studied in @xcite . a closely related bddc preconditioner for vector field problems discretised with raviart - thomas finite elements has been studied in @xcite . we are interested in applications of the bddc method to saddle - point problems . if @xmath9 in ( [ eq : sp - matrix ] ) , one possible approach is to use an algebraic trick and constrain the iterative solution of the indefinite problem into a _ balanced _ subspace , which is sometimes also called _ benign _ , where the operator is positive definite ; see @xcite for the stokes problem , and @xcite for flow in porous media . however , due to the mixed - hybrid formulation and possible coupling of meshes with different spatial dimensions , @xmath6 in general , and we will favour an alternative , _ dual _ approach here . our methodology is as follows . the mixed - hybrid formulation @xcite is used in order to modify the saddle - point problem to one which is symmetric and positive definite by means of iterative substructuring . in particular , we introduce a symmetric positive definite schur complement with respect to interface lagrange multipliers , corresponding to a part of block @xmath4 . the reduced system is solved by the preconditioned conjugate gradient ( pcg ) method , and the bddc method is used as a preconditioner . from this perspective , our work can be viewed as a further extension of @xcite . our main effort here is in accommodating the bddc solver to flows in porous media with combined mesh dimensions . in addition , the presentation of the bddc algorithm is driven more by an efficient implementation , while it is more oriented towards underlying theory in @xcite . we take advantage of the special structure of the blocks in matrix ( [ eq : sp - matrix ] ) studied in detail in @xcite . in particular , the nonzero structure of block @xmath4 resulting from a combination of meshes with different spatial dimensions is considered in @xcite . we describe our parallel implementation of the method and study its performance on several benchmark and real - world problems . another original contribution of this paper is proposing a new scaling operator in the bddc method suitable for the studied problems . we note that if there is no coupling of meshes with different spatial dimensions present in the discretisation , the block @xmath9 in ( [ eq : sp - matrix ] ) and our method is almost identical to the one introduced in @xcite . the paper is organised as follows . in section [ sec : model ] , we introduce the model problem . in section [ sec : fractures ] , we describe the modelling of fractured porous media and combining meshes of different dimensions . in section [ sec : substructuring ] , we introduce the substructuring components and derive the interface problem . in section [ sec : bddc ] , we formulate the bddc preconditioner . in addition , the selection of interface weights for bddc is studied in detail in section [ sec : scaling ] . in section [ sec : solver ] , we describe our parallel implementation , and in section [ sec : numerical ] we report the numerical results and parallel performance for benchmark and engineering problems . finally , section [ sec : conclusion ] provides a summary of our work . our notation does not , for simplicity , distinguish between finite element functions and corresponding algebraic vectors of degrees of freedom , and between linear operators and matrices within a specific basis the meaning should be clear from the context . the transpose of a matrix is denoted by superscript @xmath10 and the energy norm of a vector@xmath11 is denoted by @xmath12 , where @xmath13 is a symmetric positive definite matrix . let @xmath14 be an open bounded polyhedral domain in @xmath15 . we are interested in the solution of the following problem , combining darcy s law and the equation of continuity written as @xmath16 subject to boundary conditions on @xmath17 , where @xmath18 stands for the part of the boundary with _ natural _ ( dirichlet ) boundary condition , and @xmath19 for the part with _ essential _ ( neumann ) boundary condition . in applications , the variable@xmath20 describes the velocity of the fluid and @xmath21 the pressure ( head ) in an aquifer @xmath14 , @xmath22 is a symmetric positive definite tensor of the hydraulic conductivity , @xmath23 is the gravity term , and @xmath24 is the outer unit normal vector of @xmath25 . the term @xmath26 is present due to the fact , that @xmath27 satisfies @xmath28 , where @xmath29 is the piezometric head . for a thorough discussion of application background we refer , e.g. , to monographs @xcite . let @xmath30 be the triangulation of domain @xmath14 consisting of @xmath31 simplicial elements with characteristic size @xmath32 . we introduce a space @xmath33 equipped with the standard norm . let @xmath34 be the space consisting of the lowest order raviart - thomas ( @xmath8 ) functions and let @xmath35 be the space consisting of piecewise constant functions on the elements of the triangulation @xmath30 . we refer , e.g. , to monograph @xcite for a detailed description of the mixed finite elements and the corresponding spaces . in the _ mixed finite element approximation _ of problem ( [ eq : problem-1])([eq : problem-4 ] ) we look for a pair @xmath36 that satisfies @xmath37 in the discrete formulation , we need @xmath38 and @xmath39 only sufficiently regular so that the integrals in the weak formulation ( [ eq : mixed-1])([eq : mixed-2 ] ) make sense , namely @xmath40 , @xmath41 . next , we recall the mixed - hybrid formulation . it was originally motivated by an effort to modify the saddle - point problem ( [ eq : mixed-1])([eq : mixed-2 ] ) to one which leads to symmetric positive definite matrices . nevertheless , this formulation is also suitable for a combination of meshes with different spatial dimensions , which will be described in detail in the next section . let @xmath42 denote the set of inter - element _ faces _ of the triangulation @xmath30 . we now introduce several additional spaces . first , let us define the space @xmath43 by relaxing the condition of continuity of the normal components in the space @xmath44 on inter - element boundaries @xmath42 . more precisely , we define local spaces @xmath45 for each element @xmath46 , @xmath47 , by @xmath48 and put @xmath49 . next , we define the space of lagrange multipliers @xmath50 consisting of functions that take constant values on individual inter - element faces in @xmath42 , @xmath51 in particular , @xmath52 on @xmath25 for any @xmath53 . in the _ mixed - hybrid finite element approximation _ of problem ( [ eq : problem-1])([eq : problem-4 ] ) , we look for a triple @xmath54 that satisfies @xmath55 & \label{eq : hybrid-1}\\ = -\int_{\partial\omega_{n}}p_{n}\mathbf{v}\cdot\mathbf{n}\,ds-\sum_{i=1}^{n_{e}}\int_{t^{i}}v_{z}\,dx,\quad\forall\mathbf{v } & \in\mathbf{v}^{-1},\nonumber\\ -\sum_{i=1}^{n_{e}}\left [ \int_{t^{i}}q\nabla\cdot\mathbf{u}\,dx\right ] & = -\int_{\omega}fq\,dx,\quad\forall q\in q,\label{eq : hybrid-2}\\ \sum_{i=1}^{n_{e}}\left [ \int_{\partial t^{i}\setminus\partial\omega}\mu(\mathbf{u}\cdot\mathbf{n})|_{\partial t_{i}}\,ds\right ] & = 0,\quad\forall\mu\in\lambda . \label{eq : hybrid-3}\ ] ] equation ( [ eq : hybrid-3 ] ) imposes a continuity condition on the normal component of the velocity ( also called _ normal flux _ ) @xmath56 across @xmath42 which guarantees that @xmath57 . this condition also implies the equivalence of the two formulations ( [ eq : mixed-1])([eq : mixed-2 ] ) and ( [ eq : hybrid-1])([eq : hybrid-3 ] ) . we note that the lagrange multipliers @xmath58 can be interpreted as the approximation of the trace of @xmath21 on @xmath42 , see @xcite for details . let us now write the matrix formulation corresponding to ( [ eq : hybrid-1])([eq : hybrid-3 ] ) as @xmath1{ccc}a & b^{t } & b_{\mathcal{f}}^{t}\\ b & 0 & 0\\ b_{\mathcal{f } } & 0 & 0 \end{array } \right ] \left [ \begin{array } [ c]{c}\mathbf{u}\\ p\\ \lambda \end{array } \right ] = \left [ \begin{array } [ c]{c}g\\ f\\ 0 \end{array } \right ] . \label{eq : hybrid - m}\ ] ] it is important to note that @xmath2 is block diagonal with @xmath31 blocks , corresponding to elements @xmath59 , @xmath60 , and each of the blocks is symmetric positive definite , cf . the first term in ( [ eq : hybrid-1 ] ) . it was shown in @xcite that the system of equations ( [ eq : hybrid - m ] ) can be reduced ( twice ) to the schur complement corresponding to the lagrange multipliers @xmath58 and solved efficiently by a direct or iterative solver . here , we will look for an efficient solution of a slightly modified , and in general also block dense , system which is introduced in the next section . in this section , we recall the main ideas of the discrete model of the flow in fractured porous media that is based on connection of meshes of different dimensions as described in @xcite . let us denote the full domain by @xmath61 . next , consider lower - dimensional domains @xmath62 , @xmath63 , such that @xmath64 consists of polygons and @xmath65 consists of line segments . we will also assume that @xmath66 . the first condition requires that a domain of a lower dimension can not poke out of the domain of higher dimension , while the second condition prevents domains of lower dimension from having boundaries in the interior of domains of higher dimension . we impose these conditions to avoid technical difficulties in the analysis . however , numerical evidence suggests that these conditions are not necessary , and in fact , they are not satisfied for the real - world problems presented in section [ sec : results_engineering ] . for every dimension @xmath67 , we introduce a triangulation @xmath68 of the domain @xmath69 that consists of finite elements @xmath70 @xmath71 and satisfies the compatibility conditions @xmath72 for every @xmath73 , @xmath74 , and @xmath63 . this means that elements of a lower dimension match faces of elements of the higher dimension . we consider equations ( [ eq : problem-2])([eq : problem-4 ] ) on the domains @xmath69 , @xmath67 , completed by a slight modification of the darcy s law ( [ eq : problem-1 ] ) : @xmath75 where @xmath76 stands for the velocity integrated over the cross - section for @xmath77 , i.e. the units of @xmath78 , @xmath79 , and @xmath80 are ms@xmath81 , m@xmath82s@xmath81 , and m@xmath83s@xmath81 , respectively . in addition , @xmath84 , @xmath85 is the thickness of a fracture , and @xmath86 is the cross - section of a 1d preferential channel . the effective fluid source @xmath87 on @xmath64 is given as @xmath88 where @xmath89 is the density of external fluid sources , and the normal fluxes from the two faces of the 3d continuum surrounding the fracture are given through the robin ( also called newton ) boundary conditions @xmath90 in the last formula , @xmath91 are the transition coefficients ( cf . @xcite for possible choices ) and @xmath92 , @xmath93 are the traces of pressure @xmath94 on the two sides of the fracture . the effective fluid source @xmath95 on @xmath65 is similar , @xmath96 where @xmath97 is the density of external fluid sources . in the 3d ambient space , the 1d channel can be connected to @xmath98 faces of 2d fractures . thus @xmath99 is the normal flux from the connected fracture @xmath98 , @xmath100 is the transition coefficient , and @xmath101 is the trace of pressure @xmath102 on the face of fracture @xmath98 . in the following , we describe the discrete mixed - hybrid formulation of the problem . the formulation and discussion of the continuous problem can be found in @xcite . let us consider spaces@xmath103 for the definition of the space @xmath50 , we can not follow ( [ eq : lambda_space ] ) directly , since e.g. on @xmath64 , we need to distinguish values of @xmath104 on two sides of a fracture . thus , we introduce a separate value for every non - boundary side of every element : @xmath105 where @xmath106 is defined in the same way as the space @xmath44 but on the domain @xmath69 . further , we identify values on faces / points that are not aligned to the fractures / channels : @xmath107 finally , we redefine @xmath108 . in the _ mixed - hybrid finite element approximation _ of the flow in fractured porous media we seek a triple @xmath109 that satisfies @xmath110 with @xmath111 , \label{eq : weak_term_a}\\ b\left ( \mathbf{u},q\right ) & = -\sum_{d=1}^{3}\sum_{i=1}^{n_{e}^{d}}\left [ \int_{t_{d}^{i}}q_{d}\,\left ( \nabla\cdot\mathbf{u}_{d}\right ) \,dx\right ] , \\ b_{\mathcal{f}}\left ( \mathbf{u},\lambda\right ) & = \sum_{d=1}^{3}\sum_{i=1}^{n_{e}^{d}}\left [ \int_{\partial t_{d}^{i}\setminus\partial \omega_{d}}\lambda|_{\partial t_{d}^{i}}\,\left ( \mathbf{u}_{d}\cdot\mathbf{n}\right ) \,ds\right ] , \\ c\left ( p , q\right ) & = \sum_{d=2}^{3}\sum_{i=1}^{n_{e}^{d}}\left [ \int_{\partial t_{d}^{i}\cap\omega_{d-1}}\sigma_{d}\,\,p_{d-1}\,q_{d-1}\,ds\right ] , \label{eq : weak_term_cbar}\\ c_{\mathcal{f}}\left ( p,\mu\right ) & = -\sum_{d=2}^{3}\sum_{i=1}^{n_{e}^{d}}\left [ \int_{\partial t_{d}^{i}\cap\omega_{d-1}}\sigma_{d}\ p_{d-1}\,\mu_{d}\,ds\right ] , \label{eq : weak_term_cf}\\ \widetilde{c}\left ( \lambda,\mu\right ) & = \sum_{d=2}^{3}\sum_{i=1}^{n_{e}^{d}}\left [ \int_{\partial t_{d}^{i}\cap\omega_{d-1}}\sigma _ { d}\,\,\lambda_{d}\,\mu_{d}\,ds\right ] , \label{eq : weak_term_ctilde}\\ \left\langle g,\mathbf{v}\right\rangle & = -\sum_{d=1}^{3}\sum_{i=1}^{n_{e}^{d}}\int_{\partial t_{d}^{i}\cap\partial\omega_{n}}p_{n}\,\left ( \mathbf{v}\cdot\mathbf{n}\right ) \,ds { -\sum_{d=1}^{3}\sum_{i=1}^{n_{e}^{d}}\int_{t_{d}^{i}}v_{z}\,dx},\\ \left\langle f , q\right\rangle & = -\sum_{d=1}^{3}\int_{\omega } \delta_{d}\,\tilde{f}_{d}\,q_{d}\,dx . \label{eq : weak_term_f}\ ] ] the system ( [ eq : hybrid - frac-1])([eq : hybrid - frac-3 ] ) now leads to the matrix form @xmath1{ccc}a & b^{t } & b_{\mathcal{f}}^{t}\\ b & -c & -c_{\mathcal{f}}^{t}\\ b_{\mathcal{f } } & -c_{\mathcal{f } } & -\widetilde{c}\end{array } \right ] \left [ \begin{array } [ c]{c}\mathbf{u}\\ p\\ \lambda \end{array } \right ] = \left [ \begin{array } [ c]{c}g\\ f\\ 0 \end{array } \right ] . \label{eq : hybrid - m - fractured}\ ] ] we note that ( [ eq : hybrid - m - fractured ] ) is related to ( [ eq : hybrid - frac-1])([eq : hybrid - frac-3 ] ) in the same way as ( [ eq : hybrid - m ] ) is related to ( [ eq : hybrid-1])([eq : hybrid-3 ] ) . the main difference in the structure of the matrices between ( [ eq : hybrid - m - fractured ] ) and ( [ eq : hybrid - m ] ) is the additional block @xmath112{cc}c & c_{\mathcal{f}}^{t}\\ c_{\mathcal{f } } & \widetilde{c}\end{array } \right ] $ ] related to the normal fluxes between dimensions and arising from ( [ eq : newton-1])([eq : newton-2 ] ) and ( [ eq : newton-3 ] ) via ( [ eq : weak_term_cbar])([eq : weak_term_ctilde ] ) . in particular , the modified right - hand side of the continuity equation for two - dimensional and one - dimensional elements , @xmath87 and @xmath95 , incorporates pressure unknowns on 2d and 1d elements , and traces of pressure on 3d and 2d elements at the fracture , which are nothing but the lagrange multipliers on 3d and 2d elements in the mixed - hybrid method . consequently , @xmath113 in ( [ eq : newton-1])([eq : newton-2 ] ) , and @xmath114 in ( [ eq : newton-3 ] ) . assuming @xmath115 bounded and greater than zero , and using the fact that @xmath116 corresponds to a symmetric positive definite matrix , we see from ( [ eq : weak_term_a ] ) that block @xmath2 in ( [ eq : hybrid - m - fractured ] ) is symmetric positive definite . block @xmath4 is symmetric positive semi - definite since @xmath117 .\ ] ] the following theorem is a standard result , e.g. brezzi-1991-mhf . here , we rewrite it in a form suitable for our setting and we verify the assumptions for the specific blocks of the matrix in ( [ eq : hybrid - m - fractured ] ) . we will further denote @xmath118 , @xmath119 , @xmath120 , and @xmath121 . [ th : regular ] let natural boundary conditions ( [ eq : problem-3 ] ) be prescribed at a certain part of the boundary , i.e. @xmath122 for at least one @xmath123 . then the discrete mixed - hybrid problem ( [ eq : hybrid - m - fractured ] ) has a unique solution . let us first investigate the structure of the matrix in ( [ eq : hybrid - m - fractured ] ) more closely . let us number the unknowns within each block of ( [ eq : hybrid - m - fractured ] ) with respect to spatial dimension @xmath123 . the matrix then takes the form of 9@xmath1249 blocks , @xmath125.\ ] ] suppose for a moment that we solve a problem only on domain @xmath69 , @xmath126 ( i.e. @xmath127 for @xmath128 ) . if no natural boundary conditions are imposed , there is a single @xmath129 entry on each row of @xmath130 and a single @xmath131 entry on each row of @xmath132 . since @xmath69 is a simply connected set , the matrix @xmath133 $ ] has a nontrivial nullspace of constant vectors . enforcing natural boundary condition on a part of @xmath134 changes the @xmath131 value on the corresponding row of matrix @xmath132 to @xmath135 , in which case @xmath136 has only a trivial nullspace , i.e. full column rank ( see e.g. ( * ? ? ? * section iv.1 ) or ( * ? ? ? * lemma 3.2 ) ) . the nullspace becomes more complicated for domains with fractures , in which case @xmath69 typically has more simply connected components separated by fractures ( cf . [ fig : schemes ] ) . let us denote them @xmath137 , @xmath138 , regardless of the dimension . in particular , @xmath137 , @xmath139 will be components without natural condition boundary , i.e. @xmath140 , while for @xmath141 we get components with prescribed natural boundary condition . we also denote @xmath142 the characteristic vector of the component @xmath137 that takes value @xmath131 for degrees of freedom associated with elements and faces of the component @xmath137 . with such notation , the basis of the nullspace of the whole matrix @xmath143\ ] ] consists of characteristic vectors @xmath144 , @xmath145 and has dimension @xmath146 . next , we show that matrix @xmath4 is not only symmetric positive semi - definite , as seen from ( [ eq : c_semi_definite ] ) , but also positive definite on @xmath147 . to this end , take @xmath148 , a vector that is piecewise constant on components , having value @xmath149 on the component @xmath137 of dimension @xmath150 for @xmath139 , and value @xmath151 for other components . then @xmath152 implies @xmath153 . indeed , every component @xmath137 of dimension @xmath63 has some component @xmath154 of dimension @xmath155 on its boundary , and therefore all @xmath149 have the same value , cf . ( [ eq : c_semi_definite ] ) . this value is zero , since there is at least one component with natural boundary condition . applying the congruence transformation , @xmath156 = \left [ \begin{matrix } i & \\ \ol{b } a^{-1 } & i \end{matrix } \right ] \left [ \begin{matrix } a & \\ & -\left ( \ol{b}a^{-1}\ol{b}^t + \ol{c } \right ) \end{matrix } \right ] \left [ \begin{matrix } i & a^{-1 } \ol{b}^t \\ & i \end{matrix } \right].\ ] ] matrix @xmath2 is symmetric positive definite ( spd ) from ( [ eq : weak_term_a ] ) and therefore @xmath157 is spd on range @xmath3 , which is the orthogonal complement of the nullspace of @xmath5 . thus @xmath158 is spd on whole @xmath159 . from the sylvester law of inertia , the number of positive , negative and zero eigenvalues is preserved by the congruence transformation . since the block diagonal matrix on the right - hand side of ( [ eq : congruence1 ] ) has only ( strictly ) positive and ( strictly ) negative eigenvalues , the matrix on the left - hand side also must be nonsingular , and problem ( [ eq : hybrid - m - fractured ] ) has a unique solution . example of a two - dimensional problem : even single fracture gives rise to two components of the two - dimensional mesh ( left ) ; example of three ( shaded ) triangles sharing a single lagrange multiplier in 3d ( right).,title="fig:",scaledwidth=25.0% ] example of a two - dimensional problem : even single fracture gives rise to two components of the two - dimensional mesh ( left ) ; example of three ( shaded ) triangles sharing a single lagrange multiplier in 3d ( right).,title="fig:",scaledwidth=30.0% ] for our purposes of combining meshes with different spatial dimensions , we define _ substructures _ as subsets of finite elements in the mesh rather than parts of a physical domain , cf . @xcite . to begin , let us define the combined triangulation @xmath160 as the union of triangulations for each spatial dimension , i.e. , @xmath161 . the triangulation @xmath160 is subsequently divided into substructures@xmath162 . note that in general a substructure can contain finite elements of different dimensions . we define the interface @xmath163 as the set of degrees of freedom shared by more than one substructure . note that @xmath164 for the current setting . thus , let us split the lagrange multipliers@xmath165 into two subsets . first , we denote by@xmath166 the set of multipliers corresponding to the interface@xmath167 . the remaining multipliers , corresponding to substructure interiors , will be denoted by@xmath168 . the interface@xmath169 of substructure@xmath170 is defined as a subset of@xmath167 corresponding to@xmath171 . next , let @xmath172 be defined as the space of lagrange multipliers corresponding to @xmath173 , @xmath174 , and define a space @xmath175 the substructure problems are obtained by assembling contributions of finite elements in each@xmath170,@xmath1{cccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it } & b_{\mathcal{f},\gamma}^{it}\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it } & -c_{\mathcal{f},\gamma } ^{it}\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i } & -\widetilde{c}_{{\gamma}i}^{it}\\ b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i } & -\widetilde{c}_{{\gamma}{\gamma}}^{i}\end{array } \right ] \left [ \begin{array } [ c]{c}\mathbf{u}^{i}\\ p^{i}\\ \lambda_{i}^{i}\\ \lambda_{\gamma}^{i}\end{array } \right ] = \left [ \begin{array } [ c]{c}g^{i}\\ f^{i}\\ 0\\ 0 \end{array } \right ] , \qquad i=1,\dots , n_{s } , \label{eq : hybrid - fractured - substructure}\ ] ] where the blocks @xmath176 are block diagonal with blocks corresponding to finite element matrices , and the blocks @xmath177 only if the substructure @xmath178 contains coupling of elements of different dimensions . we note that the global problem ( [ eq : hybrid - m - fractured ] ) could be obtained from ( [ eq : hybrid - fractured - substructure ] ) by further assembly at the interface . in _ the iterative substructuring _ ( see e.g. @xcite ) , we first reduce the problem to substructure interfaces . in our context we can eliminate normal fluxes , pressure unknowns , and lagrange multipliers at interiors of substructures , and we can define the substructure schur complements@xmath179 , @xmath180 , formally as @xmath181{ccc}b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i}\end{array } \right ] \left [ \begin{array } [ c]{ccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it}\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it}\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i}\end{array } \right]^{-1 } \left [ \begin{array } [ c]{c}b_{\mathcal{f},\gamma}^{it}\\ -c_{\mathcal{f},\gamma}^{it}\\ -\widetilde{c}_{{\gamma}i}^{it}\end{array } \right ] . \label{eq : substructure - schur}\ ] ] however , in an implementation of a krylov subspace iterative method , we only need to compute the matrix - vector product @xmath182 for a given vector@xmath183 . therefore , the matrix is not constructed explicitly , and the multiplication is obtained as follows . [ alg : s - mult ] given @xmath184 , determine @xmath185 in the following two steps : 1 . solve a local _ dirichlet problem _ @xmath1{ccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it}\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it}\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i}\end{array } \right ] \left [ \begin{array } [ c]{c}\mathbf{w}^{i}\\ q^{i}\\ \mu_{i}^{i}\end{array } \right ] = -\left [ \begin{array } [ c]{c}b_{\mathcal{f},\gamma}^{it}\\ -c_{\mathcal{f},\gamma}^{it}\\ -\widetilde{c}_{{\gamma}i}^{it}\end{array } \right ] \lambda_{\gamma}^{i } , \label{eq : discrete - dirichlet - problem - substructure}\ ] ] 2 . perform two sparse matrix - vector multiplications @xmath186{ccc}b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i}\end{array } \right ] \left [ \begin{array } [ c]{c}\mathbf{w}^{i}\\ q^{i}\\ \mu_{i}^{i}\end{array } \right ] \right ) . \label{eq : substructure - schur - multiple}\ ] ] next , let @xmath187 be the space of global degrees of freedom , such that the values of degrees of freedom from @xmath188 coincide on@xmath167 . the vectors of the local substructure degrees of freedom @xmath184 and the vector of the global degrees of freedom @xmath189 are related by @xmath190 where @xmath191 are the restriction operators . more specifically , each @xmath191 is a @xmath135-@xmath192 matrix such that every row contains one entry equal to one , and @xmath193 . the global schur complement @xmath194 can be obtained as @xmath195 equation ( [ eq : schur - assembly ] ) represents the formal assembly of substructure schur complements into the global schur complement . the global schur complement system , which we would like to solve iteratively , reads @xmath196 where vector @xmath197 is obtained from substructure reduced right - hand sides @xmath198{ccc}b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i}\end{array } \right ] \left [ \begin{array } [ c]{ccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it}\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it}\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i}\end{array } \right ] ^{-1}\left [ \begin{array } [ c]{c}g^{i}\\ f^{i}\\ 0 \end{array } \right ] .\label{eq : reduced - rhs}\ ] ] in our implementation , we factor and store the matrices from ( [ eq : discrete - dirichlet - problem - substructure ] ) . the factors are then used to compute the vectors@xmath199 in ( [ eq : reduced - rhs ] ) , and especially in algorithm [ alg : s - mult ] , which is used in connection to formula ( [ eq : schur - assembly ] ) to evaluate @xmath200 within each iteration of a krylov subspace iterative method . this algorithm allows a straightforward parallel implementation . after an approximate solution of ( [ eq : s_hat - system ] ) is found , solution in interiors of substructures , including all primal variables , is recovered from ( [ eq : hybrid - fractured - substructure ] ) using the factors from ( [ eq : discrete - dirichlet - problem - substructure ] ) . there are other ways to derive the interface problem ( [ eq : s_hat - system ] ) . the authors of @xcite use a mixed - hybrid formulation with lagrange multipliers introduced only at the interface @xmath163 as their starting point . while problem ( [ eq : s_hat - system ] ) is equivalent to the interface problems considered in @xcite , the substructure problems therein have a different structure from ( [ eq : hybrid - fractured - substructure ] ) . in particular , there are no blocks corresponding to @xmath201 , and the matrices corresponding to @xmath176 are no longer block diagonal element - wise . next , the authors of @xcite construct explicit schur complement with respect to the whole block of lagrange multipliers @xmath58 and they show that , due to the special structure of @xmath2 , the complement is both sparse and reasonably cheap to construct . if this was performed substructure by substructure , this could be again seen as an intermediate step in obtaining problem ( [ eq : s_hat - system ] ) by additional elimination of the interior lagrange multipliers @xmath202 . however , this would again lead to different substructure problems based on explicit local schur complements . the following result allows an application of the bddc method to problems with fractures . [ th : sspd ] let natural boundary conditions ( [ eq : problem-3 ] ) be prescribed at a certain part of the boundary , i.e. @xmath122 for at least one @xmath123 . then the matrix @xmath203 in ( [ eq : s_hat - system ] ) is symmetric positive definite . using the notation of ( [ eq : blocks_matrix])([eq : congruence1 ] ) , let us introduce a matrix @xmath204 . the matrix @xmath205 is symmetric positive definite by theorem [ th : regular ] ( and its proof ) . applying another congruence transformation to @xmath205 and denoting the rows corresponding to the interface lagrange multipliers by subscript @xmath206 and the interior by @xmath207 , we obtain @xmath208 = \left [ \begin{matrix } i & \\ \mathcal{s}_{\gamma i } \mathcal{s}_{ii}^{-1 } & i \end{matrix } \right ] \left [ \begin{matrix } \mathcal{s}_{ii } & \\ & \mathcal{s}_{\gamma \gamma } - \mathcal{s}_{\gamma i } \mathcal{s}_{ii}^{-1 } \mathcal{s}_{\gamma i}^{t } \end{matrix } \right ] \left [ \begin{matrix } i & \mathcal{s}_{ii}^{-1 } \mathcal{s}_{\gamma i}^{t } \\ & i \end{matrix } \right].\ ] ] since the matrix on the left - hand side is spd , both diagonal blocks @xmath209 and @xmath210 are also symmetric positive definite from the sylvester law of inertia . it remains to note that the schur complement @xmath203 in ( [ eq : s_hat - system ] ) is symmetric positive definite because @xmath211 . theorem [ th : sspd ] allows us to use the conjugate gradient method for the iterative solution of ( [ eq : s_hat - system ] ) . in the next section , we describe the bddc method used as a preconditioner for @xmath203 . in this section , we formulate the bddc method for the solution of ( [ eq : s_hat - system ] ) . the algorithm can be viewed as a generalisation of @xcite . however , we follow the original description from @xcite , which better reflects our implementation . one step of bddc provides a two - level preconditioner for the conjugate gradient method applied to solving problem ( [ eq : s_hat - system ] ) . it is characterised by the selection of certain _ coarse degrees of freedom _ based on primary degrees of freedom at interface @xmath163 . the main coarse degrees of freedom in this paper are arithmetic averages over _ faces _ , defined as subsets of degrees of freedom shared by the same two substructures . in addition , _ corner _ coarse degrees of freedom , defined as any selected lagrange multiplier at the interface , are used . substructure _ edges _ , defined as subsets of degrees of freedom shared by several substructures , may also appear ( see remark [ rem : constraints ] ) . the bddc method introduces _ constraints _ which enforce continuity of functions from @xmath212 at coarse degrees of freedom among substructures . this gives raise to the space @xmath213 , which is given as the subspace of @xmath212 of functions which satisfy these continuity constraints . in particular , @xmath214 [ rem : constraints ] in three spatial dimensions , several triangular elements can be connected at a single lagrange multiplier in a star - like configuration ( cf . fig . [ fig : schemes ] ) . a similar statement holds for line elements considered in two- or three - dimensional space . this fact may lead to the presence of substructure _ edges _ and even _ vertices _ ( defined as degenerate edges consisting of a single degree of freedom ) , and we may prescribe also edge averages as constraints . as mentioned above , we also select _ corners _ as coarse degrees of freedom . while essentially any degree of freedom at the interface @xmath163 can be _ a corner _ , we select them by the face - based algorithm @xcite . this algorithm considers all vertices as corners , and , in addition , it selects three geometrically well distributed degrees of freedom from the interface between two substructures sharing a face into the set of _ although considering corners is not the standard practice with @xmath8 finite elements , in our experience , corners improve convergence for numerically difficult problems , as can be observed for the engineering applications presented in section [ sec : numerical ] . we now proceed to the formulation of operators used in the bddc method . the choice of constraints determines the construction of matrices @xmath215 . each row of@xmath216 defines one coarse degree of freedom at substructure @xmath178 , e.g. , a corner corresponds to a single @xmath192 entry at a row and an arithmetic average to several @xmath192 s at a row . the _ coarse basis functions _ @xmath217 , one per each substructure coarse degree of freedom , are computed by augmenting the matrices from ( [ eq : hybrid - fractured - substructure ] ) with @xmath215 , and solving the augmented systems with multiple right - hand sides @xmath1{ccccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it } & b_{\mathcal{f},\gamma}^{it } & 0\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it } & -c_{\mathcal{f},\gamma } ^{it } & 0\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i } & -\widetilde{c}_{{\gamma}i}^{it } & 0\\ b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i } & -\widetilde{c}_{{\gamma}{\gamma}}^{i } & d^{it}\\ 0 & 0 & 0 & d^{i } & 0 \end{array } \right ] \left [ \begin{array } [ c]{c}x^{i}\\ z^{i}\\ \phi_{i}^{i}\\ \phi_{\gamma}^{i}\\ l^{i}\end{array } \right ] = \left [ \begin{array } [ c]{c}0\\ 0\\ 0\\ 0\\ i \end{array } \right ] , \quad i=1,\dots , n_{s } , \label{eq : hybrid - fractured - augmented - substructure}\ ] ] where @xmath218 is the identity matrix , and @xmath219 , @xmath220 , and @xmath221 are auxiliary matrices not used any further . as shown in @xcite , the _ local coarse matrix _ @xmath222 is obtained as a side product of solving ( [ eq : hybrid - fractured - augmented - substructure ] ) as @xmath223 \left [ \begin{array } [ c]{cccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it } & b_{\mathcal{f},\gamma}^{it}\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it } & -c_{\mathcal{f},\gamma } ^{it}\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i } & -\widetilde{c}_{{\gamma}i}^{it}\\ b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i } & -\widetilde{c}_{{\gamma}{\gamma}}^{i } \end{array } \right ] \left [ \begin{array } [ c]{c}x^{i}\\ z^{i}\\ \phi_{i}^{i}\\ \phi_{\gamma}^{i}\end{array } \right ] = -l^{i}. \label{eq : scci_def}\ ] ] let us define , similarly to ( [ eq : r ] ) , operators@xmath224 that relate vectors of local coarse degrees of freedom@xmath225 to the vector of global coarse degrees of freedom@xmath226 as @xmath227 the _ global coarse matrix _ @xmath228 is then obtained by assembling the local contributions as @xmath229 finally , let us define the scaling operators @xmath230 which are given as diagonal matrices of weights that satisfy @xmath231 more details on the selection of diagonal entries in @xmath232 are given in section [ sec : scaling ] . with this selection of spaces and operators , we are ready to formulate the bddc preconditioner . [ alg : bddc]the bddc preconditioner @xmath233 is defined in the following steps : 1 . compute the local residuals @xmath234 2 . compute the _ substructure corrections _ @xmath235 by solving the local _ neumann problems _ @xmath1{ccccc}a^{i } & b^{it } & b_{\mathcal{f},i}^{it } & b_{\mathcal{f},\gamma}^{it } & 0\\ b^{i } & -c^{i } & -c_{\mathcal{f},i}^{it } & -c_{\mathcal{f},\gamma } ^{it } & 0\\ b_{\mathcal{f},i}^{i } & -c_{\mathcal{f},i}^{i } & -\widetilde{c}_{ii}^{i } & -\widetilde{c}_{{\gamma}i}^{it } & 0\\ b_{\mathcal{f},\gamma}^{i } & -c_{\mathcal{f},\gamma}^{i } & -\widetilde{c}_{{\gamma}i}^{i } & -\widetilde{c}_{{\gamma}{\gamma}}^{i } & d^{it}\\ 0 & 0 & 0 & d^{i } & 0 \end{array } \right ] \left [ \begin{array } [ c]{c}x^{i}\\ z^{i}\\ \eta_{i \delta } ^{i}\\ \eta_{\gamma \delta}^{i}\\ l^{i}\end{array } \right ] = \left [ \begin{array } [ c]{c}0\\ 0\\ 0\\ r_{\gamma}^{i}\\ 0 \end{array } \right ] , \quad i=1,\dots , n_{s}.\label{eq : bddc - substructure - problem}\ ] ] 3 . compute the _ coarse correction _ @xmath236 by collecting the coarse residual @xmath237 solving the global _ coarse problem _ @xmath238 and distributing the coarse correction@xmath239 4 . combine and average the corrections@xmath240 we note that the factorisations of the matrices from ( [ eq : hybrid - fractured - augmented - substructure ] ) are also used for each solution of ( [ eq : bddc - substructure - problem ] ) . in order to apply the existing bddc theory for elliptic problems ( e.g. @xcite ) to the proposed preconditioner , we introduce some additional notation and make a few observations . due to ( [ eq : substructure - schur ] ) , the substructure corrections in ( [ eq : bddc - substructure - problem ] ) can be written equivalently as @xmath1{cc}-s^{i } & d^{it}\\ d^{i } & 0 \end{array } \right ] \left [ \begin{array } [ c]{c}\eta_{\gamma \delta}^{i}\\ l^{i}\end{array } \right ] = \left [ \begin{array } [ c]{c}r_{\gamma}^{i}\\ 0 \end{array } \right ] , \label{eq : bddc - substructure - problem - with - schur}\ ] ] and the construction of coarse basis functions @xmath217 in ( [ eq : hybrid - fractured - augmented - substructure ] ) as @xmath1{cc}-s^{i } & d^{it}\\ d^{i } & 0 \end{array } \right ] \left [ \begin{array } [ c]{c } \phi_{\gamma}^{i}\\ l^{i}\end{array } \right ] = \left [ \begin{array } [ c]{c}0\\ i \end{array } \right ] . \label{eq : bddc - substructure - problem - with - schur - for - coarse - basis}\ ] ] next , let us formally write the operators and vectors in the block form @xmath241{c}\lambda_{\gamma}^{1}\\ \vdots\\ \lambda_{\gamma}^{n_{s}}\end{array } \right ] , \ r=\left [ \begin{array } [ c]{c}r^{1}\\ \vdots\\ r^{n_{s}}\end{array } \right ] , \ w=\left [ \begin{array } [ c]{ccc}w^{1 } & & \\ & \ddots & \\ & & w^{n_{s}}\end{array } \right ] , \ s=\left [ \begin{array } [ c]{ccc}s^{1 } & & \\ & \ddots & \\ & & s^{n_{s}}\end{array } \right ] .\ ] ] by grouping the steps of algorithm [ alg : bddc ] and using ( [ eq : bddc - substructure - problem - with - schur ] ) , the operator of the bddc preconditioner can be formally written as @xmath242 where @xmath243{cc}i_{{\lambda}_{\gamma}^{i } } & 0 \end{array } \right ] \left [ \begin{array } [ c]{cc}-s^{i } & d^{it}\\ d^{i } & 0 \end{array } \right]^{-1 } \left [ \begin{array } [ c]{c}i_{{\lambda}_{\gamma}^{i } } \\ 0 \end{array } \right ] \right ) + \left ( \sum_{i=1}^{n_{s}}r_{c}^{it}\phi_{\gamma}^{it } \right)^t s_{cc}^{-1 } \left ( \sum_{i=1}^{n_{s}}r_{c}^{it}\phi_{\gamma}^{it } \right ) \right\}.\ ] ] the first term in @xmath244 corresponds to substructure corrections and the second term to the coarse correction ( steps 2 and 3 of algorithm [ alg : bddc ] ) , and @xmath245 is the identity matrix in @xmath246 . from ( [ eq : mbddc - formal ] ) and ( [ eq : stilde ] ) , one can readily see that @xmath247 is symmetric . [ as : nulls]let us assume that @xmath248 in order to satisfy assumption [ as : nulls ] , we must prescribe enough coarse degrees of freedom as constraints along with robin boundary conditions ( [ eq : newton-1])([eq : newton-2 ] ) and ( [ eq : newton-3 ] ) at each fracture within substructure@xmath170 . since constrains in @xmath215 are linearly independent , @xmath249 has full column rank . in particular , assumption [ as : nulls ] is satisfied when arithmetic averages are used on each substructure face ( and eventually edge ) as constraints . [ lemma : positive_definite_preconditioner ] the operator @xmath244 in preconditioner @xmath247 is symmetric and positive definite on the space @xmath213 . the space @xmath213 is decomposed into the substructure spaces and the coarse space , @xmath250 to achieve this splitting , each local space @xmath246 is decomposed into subspaces @xmath251 corresponding to the _ substructure space _ @xmath252 , and the _ coarse space _ on substructure @xmath178 , @xmath253 , respectively . to analyse this decomposition , let us recall that @xmath254 is a positive semi - definite matrix and write ( [ eq : bddc - substructure - problem - with - schur - for - coarse - basis ] ) in detail as @xmath255 from ( [ eq : bddc - substructure - problem - with - schur - for - coarse - basis-1 ] ) , @xmath256 which in turn , similarly to ( * ? ? ? * lemma 8) , gives for any @xmath257 and @xmath258 @xmath259 from ( [ eq : bddc - substructure - problem - with - schur - for - coarse - basis-2 ] ) , the matrix @xmath217 has full column rank , and @xmath260 finally , from assumption [ as : nulls ] and ( [ eq : decomposition_of_subdomain_space ] ) , @xmath261 decomposition of the subdomain space ( [ eq : decomposition_of_subdomain_space ] ) implies decomposition of a function @xmath262 to @xmath263 , where @xmath264 , @xmath265 , and @xmath266 by ( [ eq : s - orthogonality2 ] ) . let us first analyse the substructure corrections . following ( * ? ? ? * section 3.3 ) , the matrix of ( [ eq : bddc - substructure - problem - with - schur ] ) is invertible due to assumption [ as : nulls ] . if we define , in addition , a matrix @xmath267 with orthonormal columns forming a basis of @xmath268 , i.e. @xmath269 we have @xmath1{cc}i_{{\lambda}_{\gamma}^{i } } & 0 \end{array } \right ] \left [ \begin{array } [ c]{cc}-s^{i } & d^{it}\\ d^{i } & 0 \end{array } \right]^{-1 } \left [ \begin{array } [ c]{c}i_{{\lambda}_{\gamma}^{i } } \\ 0 \end{array } \right ] = -q^{i}\left(q^{it } s^{i } q^{i}\right)^{-1}q^{it}.\ ] ] the matrix @xmath270 is symmetric positive definite , and consequently , for any @xmath262 , @xmath271 with equality if @xmath272 . next , let us turn towards the coarse correction . formula ( [ eq : scci_def ] ) for@xmath273 can be written equivalently as @xmath274 since the term on the right - hand side is just a ( negative ) galerkin projection of the positive semi - definite matrix @xmath254 , matrix @xmath222 is symmetric negative semi - definite . if at least one substructure is equipped with natural boundary conditions , the matrix @xmath228 assembled by ( [ eq : scc_def ] ) becomes symmetric negative definite , and so is @xmath275 . we have just verified the negative definiteness of the principal parts of @xmath244 , and the desired positive definiteness is obtained through the change of sign in front of the braces in ( [ eq : stilde ] ) . in view of lemma [ lemma : positive_definite_preconditioner ] , the standard condition number bound follows from ( * ? ? ? * lemma 2 ) . [ th : bddc_bound ] the condition number @xmath276 of the preconditioned operator @xmath277 satisfies @xmath278 the norms in ( [ eq : cond - bound - algebraic ] ) are induced by the matrix @xmath279 defined in ( [ eq : block_operators ] ) for all functions @xmath280 . in addition , in the case of a single mesh dimension in either 2d or 3d , and under the assumption of substructure - wise constant hydraulic conductivities , it has been also derived in ( * ? ? ? * lemma 5.5 and theorem 6.1 ) that the condition number bound@xmath281 satisfies @xmath282 where @xmath283 is the characteristic size of geometric substructures . we note that the bound ( [ eq : cond - bound - analytic ] ) implies that , for a fixed relative subdomain size @xmath284 , the condition number is independent of the problem size . it is worth emphasising that theorem [ th : bddc_bound ] is also valid for combined mesh dimensions . however , several simplifications are employed in @xcite to obtain ( [ eq : cond - bound - analytic ] ) , which are not satisfied in the set - up considered in this paper . in particular ( i ) hydraulic conductivity coefficient here is , in general , not substructure - wise constant nor isotropic , ( ii ) it is not clear whether , in presence of fractures , the interpolation operator onto a conforming mesh introduced in @xcite can be constructed and bounded in the @xmath285 norm . let us now discuss the choice of entries in the diagonal weight matrices @xmath232 . these matrices play an important role in the bddc method , both in the theory , cf . theorem [ th : bddc_bound ] or @xcite , and in the computations , cf . three possible choices are also studied numerically in section [ sec : results_engineering ] . the basic choice is presented by the _ arithmetic average _ taken from values at the neighbouring substructures . in this simplest construction , the entry corresponding to lagrange multiplier @xmath286 is given by the inverse counting function as @xmath287 where @xmath288 is the number of substructures in the set @xmath289 of indices of substructures to which @xmath286 belongs . for 2d or 3d meshes without fractures , @xmath290 for raviart - thomas elements . however , since several two - dimensional fractures can meet in our setting , smaller weights can occasionally appear at such regions . while arithmetic average is sufficient for problems with homogeneous coefficients , it is well known that for problems with large variations in material properties along the interface , it is necessary to incorporate their values into the ( weighted ) average to obtain a robust method . this gives rise to the @xmath291-scaling , for which @xmath292 where @xmath293 is a material characteristic for substructure @xmath294 . this choice is robust with respect to jumps in coefficients across the interface , cf . @xcite ; however , coefficients are assumed constant for each substructure . this requirement is very restrictive for practical computations with quickly varying coefficients , and we employ a generalisation which takes into account the material coefficient of the element to which the lagrange multiplier @xmath286 corresponds . in our case , we use @xmath295 , where @xmath123 is the dimension of the element @xmath59 . this value can be seen as a representative hydraulic conductivity on the element . finally , we propose a modification of the popular scaling by diagonal stiffness @xcite . in the usual diagonal stiffness approach , the optimal weight , which is the diagonal entry of the schur complement , is approximated by the diagonal entry of the original substructure matrix . however , this is not directly applicable to the indefinite system ( [ eq : hybrid - fractured - substructure ] ) , as , in general , matrix @xmath296 contains only seldom nonzeros on the diagonal . for this reason , we approximate the diagonal of the schur complement as @xmath297 where the index @xmath98 corresponds to the row in block @xmath176 of the element face to which the lagrange multiplier @xmath286 belongs . using the diagonal stiffness scaling in connection with the standard lagrange finite elements may lead to poor convergence for problems with rough interface @xcite , for which the diagonal stiffness can vary quickly even for smooth problems with constant coefficients on uniform meshes . this is a severe issue for practical computations , in which graph partitioners are typically used for creating substructures . however , this issue is not as pronounced for raviart - thomas elements , for which only one element contributes to the stiffness on the diagonal at an interface degree of freedom , and thus irregularities caused by changing number of elements contributing to an interface weight can not occur . on the other hand , an advantage of the diagonal stiffness scaling is the fact , that unlike the @xmath291-scaling it takes into account the shape and relative sizes of elements , which vary considerably in engineering applications , as well as the effect of @xmath115 introduced in ( [ eq : weak_term_a ] ) and ( [ eq : weak_term_f ] ) . unless stated otherwise , scaling ( [ eq : scaling_diagonal ] ) is used in the computations presented in section [ sec : numerical ] . the basis for an efficient parallel implementation of the method described in previous sections was obtained by combining two existing open - source software packages : the finite element package _ _ flow123d _ _ ( version 1.6.5 ) for underground fluid flow simulations and the bddc - based solver _ _ bddcml _ _ ( version 2.0 ) used for the solution of the resulting system of equations . however , minor changes have been made to both codes to support the specific features , such as the weights ( [ eq : scaling_rho ] ) and ( [ eq : scaling_diagonal ] ) . the _ flow123d _ package has been developed for modelling complex behaviour of underground water flow and pollution transport . however , only the simple flow in a fully saturated porous media described by darcy s law is considered in this paper . to accurately account for fractures in the medium , such as granite rock , the solver allows us to combine finite elements of different dimensions : the three - dimensional elements of porous media are combined with two - dimensional elements modelling planar fractures , which may be in turn connected in one - dimensional elements for channels . raviart - thomas elements are consistently used throughout such discretisation . although the fractures are also modelled as porous media , their hydraulic conductivity is by orders of magnitude higher than that of the main porous material of the domain . in addition , the finite element discretisations are typically not uniform within the domain , and the relative sizes of elements may also vary by orders of magnitude . both these aspects give rise to very poorly conditioned linear systems , which are very challenging for iterative solvers . the _ flow123d _ solver has been developed for over 10 years and it is written in c / c++ programming language with object - oriented design and parallelism through mpi . the _ bddcml _ is a library for solving algebraic systems of linear equations by means of the bddc method . the package supports the adaptive - multilevel bddc method @xcite suitable for very high number of substructures and computer cores , although we only use the standard ( non - adaptive two - level ) bddc method from @xcite for the purpose of this paper . the _ bddcml _ library is typically interfaced by finite element packages , which may provide the division into substructures . this feature is used in our current implementation , in which the division into non - overlapping substructures is constructed within the _ flow123d _ using the _ metis _ ( version 5.0 ) package @xcite . one substructure is assigned to a processor core in the current set - up of the parallel solver , although _ bddcml _ is more flexible in this respect . the library performs the selection of additional corners by the face - based algorithm from @xcite . the _ bddcml _ package is written in fortran 95 and parallelised through mpi . the _ bddcml _ solver relies on a serial instance of the _ mumps _ direct solver @xcite for the solution of each local discrete dirichlet problem ( [ eq : discrete - dirichlet - problem - substructure ] ) as well as for the solution of each local discrete neumann problem ( [ eq : bddc - substructure - problem ] ) . the coarse problem ( [ eq : bddc - coarse - problem ] ) is solved by a parallel instance of _ mumps_. the main difference from using _ bddcml _ for symmetric positive definite problems is the need to use the @xmath298 factorisation of general symmetric matrices for problems ( [ eq : discrete - dirichlet - problem - substructure ] ) , which are saddle - point ( i.e. indefinite ) systems in the present setting . although the original system ( [ eq : hybrid - m - fractured ] ) is indefinite , system ( [ eq : s_hat - system ] ) is symmetric positive definite , which allows the use of the preconditioned conjugate gradient ( pcg ) method . one step of bddc is used as the preconditioner within the pcg method applied to problem ( [ eq : s_hat - system ] ) . the matrix of problem ( [ eq : s_hat - system ] ) is not explicitly constructed in the solver , and only its actions on vectors are computed following ( [ eq : discrete - dirichlet - problem - substructure])([eq : schur - assembly ] ) . in our implementation , we change the sign neither in the action of @xmath299 ( [ eq : substructure - schur - multiple ] ) nor in the action of the preconditioner @xmath247 ( [ eq : bddc - preconditioner ] ) . since both are then strictly negative definite , the product @xmath277 is the same as if both signs were changed , and the pcg method runs correctly . in this way , no changes are necessary in an implementation developed for symmetric positive definite problems . in this section , we investigate the performance of the algorithm and its parallel implementation on several benchmark problems in 2d and 3d , and on two geoengineering problems of existing localities in 3d . for the two benchmark problems without fractures we perform weak scaling tests . for the benchmark problem with fractures and for the geoengineering problems , we perform strong scaling tests with the problem size fixed and increasing number of processor cores . in all cases , the pcg iterations are run until the relative norm of residual @xmath300 . if not stated otherwise , the proposed scaling by diagonal stiffness ( [ eq : scaling_diagonal ] ) is used within the averaging operator of bddc . first , the performance of the solver is investigated on a unit square and a unit cube discretised solely using two - dimensional and three - dimensional finite elements , respectively . for this reason , block @xmath4 in system ( [ eq : hybrid - m - fractured ] ) , which is related to combining elements of different dimension , is zero , and the problem reduces to the standard problem ( [ eq : hybrid - m ] ) . the sequence of unstructured meshes is approximately uniform for both problems , and the problems do not contain any jumps in material coefficients . in figs . [ fig : square2d_problem ] and [ fig : cube3d_problem ] , example meshes and the resulting pressure head and velocity fields are presented . while gravity is present in the 3d case , its effect is not considered in the 2d case . the results of the weak scaling tests are summarised in tables [ tab:2d_square_weak_scaling ] and [ tab:3d_cube_100k_weak_scaling ] . to give a better view , the resulting solution times for different problem sizes are also visualised in fig . [ fig : timing_100k ] . in these tables , @xmath301 denotes the number of substructures and processors , @xmath302 is the size of the global problem ( [ eq : hybrid - m - fractured ] ) , @xmath303 is the size of the interface problem ( [ eq : s_hat - system ] ) , @xmath304 denotes number of faces , @xmath305 denotes number of corners , ` its . ' stands for resulting number of pcg iterations , and ` cond . ' is the approximate condition number computed from the lanczos sequence in pcg . we report separately the time spent in preconditioner set - up , the time spent by pcg iterations , and the total time for the whole solve . in these weak scaling tests , the number of unknowns per core is kept approximately constant around @xmath306 . these weak scaling tests were performed using up to 64 cores of the sgi altix uv supercomputer at the supercomputing centre of the czech technical university in prague . the computer contains twelve intel xeon processors , each with six cores at frequency 2.67 ghz . intel compilers version 12.0 were used . the numbers of pcg iterations and condition number estimates in tables [ tab:2d_square_weak_scaling ] and [ tab:3d_cube_100k_weak_scaling ] confirm the expected numerical scalability of the bddc method , which is well known for symmetric positive definite problems as well as for darcy s flow problems @xcite . the slight irregularities in the condition number in table [ tab:3d_cube_100k_weak_scaling ] are probably caused by using non - nested unstructured meshes . looking at times in these tables and in fig . [ fig : timing_100k ] , we can see almost optimal scaling , with only mild growth of times with number of cores . the numbers of pcg iterations are higher in 3d , and the time spent in pcg iterations grows proportionally , while the time spent in the set - up phase does not differ considerably between two - dimensional and three - dimensional setting and dominates the overall time . example of solution to the model square problem containing only 2d elements , plot of pressure head with mesh ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=47.0% ] example of solution to the model square problem containing only 2d elements , plot of pressure head with mesh ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=50.0% ] example of solution to the model cube problem containing only 3d elements , plot of pressure head with mesh ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=47.0% ] example of solution to the model cube problem containing only 3d elements , plot of pressure head with mesh ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=50.0% ] [ c]|ccc|ccc|cc|ccc| & & & & & & & & + & & & & & & & & set - up & pcg & solve + 2 & 207k & 103k & 155 & 1 & 2 & 7 & 1.37 & 8.3 & 1.6 & 9.9 + 4 & 440k & 110k & 491 & 5 & 10 & 8 & 1.60 & 12.2 & 2.2 & 14.4 + 8 & 822k & 103k & 1.2k & 13 & 26 & 9 & 1.78 & 11.0 & 2.5 & 13.5 + 16 & 1.8 m & 111k & 2.8k & 33 & 66 & 8 & 1.79 & 14.3 & 2.7 & 17.0 + 32 & 3.3 m & 104k & 5.9k & 74 & 148 & 9 & 1.79 & 12.1 & 3.3 & 15.4 + 64 & 7.2 m & 113k & 13.0k & 166 & 332 & 9 & 1.85 & 14.8 & 4.4 & 19.2 + [ c]|ccc|ccc|cc|ccc| & & & & & & & & + & & & & & & & & set - up & pcg & solve + 2 & 217k & 108k & 884 & 1 & 3 & 11 & 2.88 & 11.7 & 2.3 & 14.0 + 4 & 437k & 109k & 2.3k & 6 & 18 & 12 & 3.04 & 11.7 & 2.5 & 14.2 + 8 & 945k & 118k & 5.7k & 21 & 63 & 15 & 12.00 & 15.4 & 4.0 & 19.3 + 16 & 1.6 m & 103k & 12.8k & 56 & 168 & 16 & 6.58 & 12.9 & 4.0 & 17.0 + 32 & 3.4 m & 106k & 29.8k & 132 & 401 & 18 & 10.10 & 15.4 & 5.2 & 20.6 + 64 & 6.1 m & 95k & 59.6k & 307 & 931 & 19 & 16.58 & 13.7 & 6.3 & 20.0 + weak scaling test for the 2d square problem ( left ) , and the 3d cube problem ( right ) , approx . 100k unknowns per core . computational times separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] weak scaling test for the 2d square problem ( left ) , and the 3d cube problem ( right ) , approx . 100k unknowns per core . computational times separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] the next benchmark problem is considerably more complicated . it consists again of a unit cube , which now contains four planar fractures aligned with diagonals of a 2d cross - section . these four planar fractures meet at a 1d channel in the centre of the cross - section . therefore , the problem contains the full possible combination of 3d , 2d and 1d finite elements . the tensor @xmath22 is isotropic , thus it is just a scalar multiple of identity . the corresponding scalar value is set to 10 , 1 , and 0.1 for 1d , 2d , and 3d elements , respectively . we perform a strong scaling test with this problem , keeping the mesh size fixed with approximately 2.1 million elements and 14.6 million degrees of freedom . in fig . [ fig : cube123d_problem ] , the computational mesh and the resulting pressure head and velocity fields are presented . this scaling test was computed on the cray xe6 supercomputer _ hector _ at the edinburgh parallel computing centre . this supercomputer is composed of 2816 nodes , each containing two amd opteron interlagos processors with 16 cores at 2.3 ghz . gnu compilers version 4.6 were employed . results of the strong scaling test are summarised in table [ tab:123d_cube_strong_scaling ] , and the computing times are visualised also in fig . [ fig : timing_strong_123d ] together with the parallel speed - up . the reference value for computing speed - up is the time on 16 cores , and the speed - up on @xmath307 processors is computed as @xmath308 where @xmath309 is the time on @xmath307 processors . we can see that the number of pcg iterations grows with the number of substructures for this problem which is also confirmed by the growing condition number estimate . while the time spent in set - up phase scales very well , the time spent in pcg grows together with the number of iterations . the reason for this growth seem to be related to the larger interface , at which more numerical difficulties appear . this seems to be related to more 1d-2d and 2d-3d connections at the interface and makes this difficult problem a good candidate for using the _ adaptive bddc _ . however , this will be the subject of a separate study . example of solution to the model cube problem containing 1d , 2d , and 3d elements , plot of pressure head with mesh and fractures ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=47.0% ] example of solution to the model cube problem containing 1d , 2d , and 3d elements , plot of pressure head with mesh and fractures ( left ) and velocity vectors ( right).,title="fig:",scaledwidth=50.0% ] [ c]|cc|ccc|cc|ccc| & & & & & & & + & & & & & & & set - up & pcg & solve + 16 & 912k & 47k & 53 & 159 & 26 & 59.3 & 171.6 & 84.5 & 256.2 + 32 & 456k & 65k & 126 & 380 & 48 & 2091.0 & 90.1 & 109.8 & 200.0 + 64 & 228k & 86k & 301 & 914 & 81 & 1436.1 & 36.8 & 77.1 & 114.0 + 128 & 114k & 116k & 689 & 2076 & 109 & 2635.8 & 14.3 & 43.1 & 57.4 + 256 & 57k & 151k & 1436 & 4365 & 164 & 1700.5 & 6.7 & 31.2 & 38.0 + 512 & 28k & 196k & 3021 & 9244 & 254 & 42614.5 & 4.0 & 26.9 & 30.9 + strong scaling test for the cube problem with 1d , 2d , and 3d elements and 14.6 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] strong scaling test for the cube problem with 1d , 2d , and 3d elements and 14.6 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] the performance of the algorithm and its parallel implementation has been investigated on two engineering problems of underground flows within real geologic locations . for both problems , the porous medium is fractured granite rock , with the fractures modelled by two - dimensional elements . the first problem is the _ melechov locality _ , which models one of the candidate sites for a nuclear waste deposit to be build within the czech republic in future . the goal is to model the underground flow and estimate the speed at which an eventual radioactive pollution would spread . the computational mesh contains 2.1 million finite elements resulting in 15 million unknowns . the geometry of the problem with the resulting distribution of piezometric head and the finite element mesh is presented in fig . [ fig : melechov_1 ] . the problem contains vertical two - dimensional fractures visualised in fig . [ fig : melechov_2 ] . the maximal hydraulic conductivity within the fractures is 6.3@xmath31010@xmath311 ms@xmath81 , while the minimal conductivity of the outer material is 6.0@xmath31010@xmath312 ms@xmath81 , the transition coefficient @xmath313 = 1 s@xmath81 , and the effective thickness of fractures @xmath85 = 0.1 m. we perform a strong scaling test for this problem , keeping the problem size fixed and increasing the number of substructures and computing cores . an example of division into 64 substructures is presented in fig . [ fig : melechov_2 ] . the scaling test was computed on the _ hector _ supercomputer . table [ tab : melechov_large_strong_scaling ] summarises the results of this test . we can still see some growth of the number of iterations with the number of substructures , which is however much milder than the growth observed for the unit cube with fractures in table [ tab:123d_cube_strong_scaling ] . correspondingly , the times reported in table [ tab : melechov_large_strong_scaling ] and visualised in fig . [ fig : timing_melechov_large ] show an optimal scaling of the solver over a large range of core counts . graphics / melechov_image1_piezo ( 9,5)@xmath314 ( 9,9)@xmath315 ( 6,11.5)@xmath316 ( -10,18)@xmath3179940 m ( 25,45)@xmath31812100 m ( 87,35)@xmath3191410 m the problem of the _ melechov locality _ ; the system of fractures ( left ) and an example division into 64 substructures ( right).,title="fig:",scaledwidth=40.0% ] the problem of the _ melechov locality _ ; the system of fractures ( left ) and an example division into 64 substructures ( right).,title="fig:",scaledwidth=40.0% ] [ c]|cc|ccc|cc|ccc| & & & & & & & + & & & & & & & set - up & pcg & solve + 16 & 934k & 36k & 32 & 96 & 40 & 53.0 & 131.4 & 144.1 & 275.6 + 32 & 467k & 54k & 76 & 228 & 70 & 878.3 & 47.5 & 112.9 & 160.4 + 64 & 233k & 82k & 186 & 561 & 67 & 202.4 & 17.4 & 50.2 & 67.7 + 128 & 117k & 116k & 528 & 1592 & 69 & 237.6 & 7.9 & 23.1 & 31.1 + 256 & 58k & 155k & 1235 & 3747 & 96 & 5577.0 & 4.0 & 14.7 & 18.8 + 512 & 29k & 207k & 2699 & 8256 & 106 & 1658.1 & 2.2 & 8.3 & 10.5 + 1024 & 15k & 271k & 5711 & 17581 & 119 & 11554.5 & 2.1 & 7.0 & 9.2 + strong scaling test for the problem of the _ melechov locality _ containing 2d and 3d elements and 15 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] strong scaling test for the problem of the _ melechov locality _ containing 2d and 3d elements and 15 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] the second engineering model is the locality around the _ bedichov tunnel_. the main purpose of this 2.1 km long tunnel near the city of liberec in the north of the czech republic to accommodate water pipes , which supply the city by drinking water from a reservoir in the mountains . however , this locality is also a valuable site for experimental geological measurements performed inside the tunnel . the model aims at describing the flow in the granite rock surrounding the tunnel . the computational mesh consists of 1.1 million elements leading to 7.8 million unknowns . the mesh with the plot of resulting piezometric head is presented in fig . [ fig : tunel_1 ] . the system of fractures and an example division into 256 substructures are visualised in fig . [ fig : tunel_2 ] . the hydraulic conductivity of the fractures is 10@xmath320 ms@xmath81 , while the conductivity of the outer material is 10@xmath321 ms@xmath81 , the transition coefficient @xmath313 = 1 s@xmath81 , and the effective thickness of fractures @xmath85 = 1.1 m. although the mesh contains fewer finite elements than the one of the _ melechov locality _ model , this problem is considerably more complicated . this is caused mainly by the presence of relatively very small and irregularly shaped finite elements in the vicinity of the tunnel and near the cross - sections of fractures ( see fig . [ fig : tunel_3 ] ) generated by the mesh generator . the results of a strong scaling test are summarised in table [ tab : tunnel_strong_scaling ] . as before , the times are also plotted in fig . [ fig : timing_tunnel ] . although the number of iterations is not independent of the number of substructures , the growth is still small . consequently , the computing times , and especially the time for set - up , scale very well over a large range of numbers of substructures . the observed super - optimal scaling may be related to faster factorisation of the smaller local problems by the direct solver for indefinite matrices . table [ tab : tunnel_corner_effect ] summarises an experiment performed to analyse the effect of using corners in the construction of the coarse space in bddc . as has been mentioned in section [ sec : bddc ] , using raviart - thomas finite elements does not lead to ` natural ' corners as cross - points shared by several substructures . on the other hand , the notion of corners was generalised to any selected interface degree of freedom , at which continuity of functions from the coarse space is required . such generalisation is important for the well - posedness of the local problems for unstructured meshes e.g. in elasticity analysis @xcite . this is also the default option for _ bddcml _ , in which selection of corners is performed at each face between two substructures . adding corners improves the approximation properties of the coarse space at the cost of increasing the size of the coarse problem . table [ tab : tunnel_corner_effect ] compares the convergence for variable number of substructures without and with constraints at corners . column ` with corners ' in table [ tab : tunnel_corner_effect ] corresponds to results in table [ tab : tunnel_strong_scaling ] , and it is repeated for comparison . we can see that while the effect of corners on convergence is small for smaller number of substructures , the improvement of the coarse problem and the approximation of bddc becomes more significant for higher numbers of cores . looking at times in table [ tab : tunnel_corner_effect ] , the additional time spent in the set - up phase due to higher number of constraints when using corners is compensated by the lower number of pcg iterations , resulting in lower overall times . thus , using additionally selected corners appears beneficial for complicated engineering problems like this one . graphics / tunel_image1_piezo ( 2,12)1754 m ( width ) ( 16,57)2655 m ( length ) ( 98,43)458 m ( height ) the _ bedichov tunnel _ problem ; the system of fractures ( left ) and an example division into 256 substructures ( right).,title="fig:",scaledwidth=48.0% ] the _ bedichov tunnel _ problem ; the system of fractures ( left ) and an example division into 256 substructures ( right).,title="fig:",scaledwidth=48.0% ] example of difficulties in the mesh of the _ bedichov tunnel _ problem ; detail of the tunnel geometry with fine elements ( left ) and enforced refinement at an intersection of fractures ( right).,title="fig:",scaledwidth=45.5% ] example of difficulties in the mesh of the _ bedichov tunnel _ problem ; detail of the tunnel geometry with fine elements ( left ) and enforced refinement at an intersection of fractures ( right).,title="fig:",scaledwidth=48.0% ] [ c]|cc|ccc|cc|ccc| & & & & & & & + & & & & & & & set - up & pcg & solve + 32 & 245k & 20k & 106 & 322 & 112 & 1514.1 & 110.3 & 144.0 & 254.3 + 64 & 123k & 28k & 192 & 597 & 63 & 117.7 & 42.2 & 36.0 & 78.3 + 128 & 61k & 45k & 413 & 1293 & 75 & 194.4 & 13.4 & 16.8 & 30.3 + 256 & 31k & 72k & 902 & 2791 & 119 & 526.7 & 4.2 & 10.9 & 15.1 + 512 & 15k & 110k & 2009 & 6347 & 137 & 1143.4 & 1.8 & 7.1 & 9.0 + 1024 & 8k & 155k & 4575 & 14725 & 173 & 897.0 & 1.6 & 8.0 & 9.7 + strong scaling test for the problem of the _ bedichov tunnel _ containing 2d and 3d elements and 7.8 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] strong scaling test for the problem of the _ bedichov tunnel _ containing 2d and 3d elements and 7.8 m unknowns , computational time ( left ) and speed - up ( right ) separately for set - up and pcg phases , and their sum ( solve).,title="fig:",scaledwidth=48.0% ] [ c]|cc|ccccc|ccccc| & & & + & & & & & & & + & & & & set - up & pcg & solve & & & set - up & pcg & solve + 32 & 245k & 131 & 1789.6 & 107.5 & 175.0 & 282.5 & 112 & 1514.1 & 110.3 & 144.0 & 254.3 + 64 & 123k & 70 & 122.2 & 40.3 & 40.4 & 80.7 & 63 & 117.7 & 42.2 & 36.0 & 78.3 + 128 & 61k & 96 & 208.5 & 10.9 & 21.6 & 32.6 & 75 & 194.4 & 13.4 & 16.8 & 30.3 + 256 & 31k & 139 & 541.9 & 3.7 & 12.5 & 16.2 & 119 & 526.7 & 4.2 & 10.9 & 15.1 + 512 & 15k & 197 & 1418.9 & 1.4 & 10.0 & 11.4 & 137 & 1143.4 & 1.8 & 7.1 & 9.0 + 1024 & 8k & 312 & 3779.4 & 1.0 & 14.5 & 15.6 & 173 & 897.0 & 1.6 & 8.0 & 9.7 + in the final experiment , we compare the effect of different averaging techniques on the convergence of bddc . in table [ tab : tunnel_averaging ] , results of the strong scaling test for arithmetic averaging ( [ eq : scaling_arithmetic ] ) , the modified @xmath291-scaling ( [ eq : scaling_rho ] ) , and the proposed scaling by diagonal stiffness ( [ eq : scaling_diagonal ] ) are summarised . the final column corresponds to the results from table [ tab : tunnel_strong_scaling ] , which are repeated here for comparison . table [ tab : tunnel_averaging ] suggests , that while the simple arithmetic averaging does not lead to satisfactory convergence for this problem , the modified @xmath291-scaling and the diagonal scaling mostly lead to similar convergence . however , while the former provides slightly better convergence for several cases , it also leads to irregularities for certain divisions , for which the bddc method with this averaging converges rather poorly . therefore , the proposed scaling ( [ eq : scaling_diagonal ] ) can be recommended as the most robust choice among the three tested options . [ c]|cc|ccc|cc|cc|cc| & & & & & & & + & & & & & its . & cond . & its . & cond . & its . & cond . + 32 & 245k & 20k & 106 & 322 & 637 & 9811.7 & 110 & 1467.8 & 112 & 1514.1 + 64 & 123k & 28k & 192 & 597 & 618 & 10254.1 & 62 & 115.1 & 63 & 117.7 + 128 & 61k & 45k & 413 & 1293 & 2834 & 1.0e+11 & 206 & 401641.4 & 75 & 194.4 + 256 & 31k & 72k & 902 & 2791 & 799 & 11172.9 & 117 & 512.9 & 119 & 526.7 + 512 & 15k & 110k & 2009 & 6347 & 883 & 15449.6 & 136 & 1160.1 & 137 & 1143.4 + 1024 & 8k & 155k & 4575 & 14725 & n / a & 2.5e+10 & 504 & 99023.6 & 173 & 897.0 + a parallel solver for the mixed - hybrid finite element formulation based on darcy s law has been presented . the software combines an existing package _ flow123d _ developed for problems in geophysics with _ bddcml _ , a parallel library for solution of systems of linear algebraic equations by the bddc method . in geoengineering applications , the mathematical model is applied to geometries with the presence of fractures . in the present approach , the flow in these fractures is also modelled by darcy s law , although the hydraulic conductivity of the porous media is considered by orders of magnitude higher . these fractures are modelled by finite elements of a lower dimension . in the discretised model , 1d , 2d and 3d finite elements are coupled together through robin boundary conditions . these coupling terms lead to a modification of the usual saddle - point matrix of the system , in which a new non - zero block appears on the diagonal . the bddc method is employed for the solution of the resulting system of linear algebraic equations . bddc is based on iterative substructuring , in which the problem is first reduced to the interface among substructures . the schur complement is not built explicitly . instead , only multiplications by the matrix are performed through solving a discrete dirichlet problem on each substructure . in the setting of the mixed - hybrid problem , the interface is built only as a subset of the block of lagrange multipliers , while remaining unknowns belong to interiors of substructures . although the original problem is symmetric indefinite , the system reduced to the interface is symmetric positive definite . this is also shown to hold for the case with fractures in the present paper . consequently , the pcg method is used for the solution of the reduced problem . however , unlike the symmetric positive definite problems , a direct solver for symmetric indefinite matrices needs to be used for the factorisation and repeated solution of local problems on substructures . one step of the bddc method is used as the preconditioner for the pcg method run on the interface problem . a modification of the diagonal stiffness scaling has been introduced . it is motivated by difficult engineering problems , for which it performs significantly better than other two applicable choices the arithmetic averaging and the modified @xmath291-scaling . arithmetic averages over faces between substructures are used as the basic constraints defining the coarse space . in addition , corners are selected from unknowns at the interface using the face - based algorithm . while corners are not required by the theory , they are shown to improve both the convergence and the computational times for complicated problems . the performance of the resulting solver has been investigated on three benchmark problems in 2d and 3d . both weak and strong scaling tests have been performed . on benchmark problems with single mesh dimension , the expected optimal convergence independent of number of substructures has been achieved . correspondingly , the resulting parallel scalability has been nearly optimal for the weak scaling tests up to 64 computer cores . the strong scaling tests were presented for a benchmark problem of a unit cube and for two engineering problems containing large variations in element sizes and hydraulic conductivities , using up to 1024 computer cores and containing up to 15 million degrees of freedom . the convergence for the unit cube problem with all three possible dimensions of finite elements slightly deteriorated by using more substructures , and this translated to sub - optimal parallel performance . however , for the two engineering applications , in which only 3d and 2d elements are combined , the bddc method has also maintained good convergence properties with the growing number of substructures , resulting in optimal , or even super - optimal parallel scalability of the solver . it has been also shown that the proposed modification of the diagonal stiffness scaling plays an important role in achieving such independence for the challenging engineering problems presented in the paper . elman hc , silvester dj , wathen aj . _ finite elements and fast iterative solvers : with applications in incompressible fluid dynamics_. numerical mathematics and scientific computation , oxford university press , 2005 . cros jm . a preconditioner for the schur complement domain decomposition method . _ domain decomposition methods in science and engineering _ , herrera i , keyes de , widlund ob ( eds . ) . national autonomous university of mexico ( unam ) , mxico , 2003 ; 373380 . 14th international conference on domain decomposition methods , cocoyoc , mexico , january 612 , 2002 . fragakis y , papadrakakis m. the mosaic of high performance domain decomposition methods for structural mechanics : formulation , interrelation and numerical efficiency of primal and dual methods . _ computer methods in applied mechanics and engineering _ 2003 ; * 192*:37993830 . mandel j , sousedk b. adaptive selection of face coarse degrees of freedom in the bddc and the feti - dp iterative substructuring methods . _ computer methods in applied mechanics and engineering _ 2007 ; * 196*(8):13891399 . oden j , lee j. dual - mixed hybrid finite element method for second - order elliptic problems . _ mathematical aspects of finite element methods _ , _ lecture notes in mathematics _ , vol . 606 , galligani i , magenes e ( eds . ) . springer berlin heidelberg , 1977 ; 275291 . maryka j , rozlonk m , tma m. schur complement systems in the mixed - hybrid finite element approximation of the potential fluid flow problem . _ siam journal on scientific computing _ 2000 ; * 22*(2):704723 . bezina j , hokr m. mixed - hybrid formulation of multidimensional fracture flow . _ proceedings of the 7th international conference on numerical methods and applications _ , nma10 , springer - verlag : berlin , heidelberg , 2011 ; 125132 . ertkov m , stek j , burda p. on selection of interface weights in domain decomposition methods . _ proceedings of programs and algorithms of numerical mathematics 16 , doln maxov , czech republic , june 38 , 2012_. institute of mathematics as cr , 2013 ; 3544 .

coordinated motions and pattern formation have been studied for a wide range of biological organisms , from bacteria and amoebae to fish , from birds and wildebeest to humans ( ben - jacob et al . , 2000 ; deneubourg et al . , 2001 ; parrish et al . , 2002 ; partridge , 1982 ; wilson , 1975 ) . animal groups often behave as if they have a single mind , displaying remarkable self - organized behavior . at one extreme , the individuals seem to need little information transfer ( e.g. fish schools ) , while at the other end the information exchange occurs in highly integrated ways through long - term associations among the individuals ( e.g. honeybee hives and human communities ) . controlling such an organized behavior in groups of artificial objects , including autonomous underwater vehicles ( leonard et al . , 2006 ) and groups of autonomous agents ( jadbabaie et al . , 2003 ) , has received extensive attention in contemporary control theory . challenges in both natural and engineering settings involve understanding which patterns emerge from the interaction among individual agents . select laboratory experiments have shed some light on the schooling mechanism ( hunter , 1966 ; van olst and hunter , 1970 ; partridge and pitcher , 1979 , 1980 ; pitcher et al . , 1976 ) . it still remains unclear , however , how the individual - level behavior and group - level ( `` macroscopic '' , or coarse - grained ) patterns are related . more precise experiments using three - dimensional tracking of every individual in a population should lead to better understanding of this linkage . an ultimate experimental study with precise control of every relevant detail may not be possible , yet appropriate mathematical models would provide a venue to establish behavioral cause , as one can consider different hypothetical individual - level interaction rules selectively ( see e.g. , flierl et al . , 1999 ) . several different individual - based models have been proposed , which reproduce certain types of collective behavior in animal groups ( e.g. , see aoki , 1982 ; reynolds , 1987 ; deneubourg and goss , 1989 ) . self - organization emerges also in a wide spectrum of physical and chemical systems , some of which ( e.g. , crystals and ferromagnetic materials ) exhibit apparent similarities with emergent patterns observed in animal groups . vicsek et al . ( 1995 ) have introduced a discrete - time model of self - driven particles , or self - propelled particles ( spp ) , based on near - neighbor rules that are similar with those in the ferromagnetic xy model ( kosterlitz and thouless , 1973 ) . the authors analyzed statistical properties of the model , including phase transition and scaling ( vicsek et al . , 1995 ) . a long - range interaction has been incorporated into the spp model ( mikhailov and zanette , 1999 ) , and continuum , `` hydrodynamic '' versions of this model have been introduced ( toner and tu , 1995 , 1998 ; topaz et al . , 2006 ) . recently , couzin et al . ( 2002 , 2005 ) have introduced a model to provide insights into the mechanism of decision making in biological systems , which reproduces many important observations made in the field , and provides new insights into these phenomena . a review for various models can be found in parrish et al . ( 2002 ) and czirk and vicsek ( 2001 ) . the models of couzin et al . ( 2005 ) , and most other such models , incorporate various detailed mechanistic steps . these shed light on the role of leadership and imitation , and produce a number of surprising results , such as the influence that a few `` informed '' individuals can have on large collectives . what is needed now are efforts to simplify those models , and to show especially what properties of the microscopic simulators are essential to explain that behavior . for some models , closure schemes are available ( flierl et al . , 1999 ) ; but more generally , though we may suspect that closures exist , we can not derive explicit expressions for them . in such circumstances , we need methods such as those used in this paper ; we perform the coarse - grained dynamical analysis by circumventing the derivation of governing equations , using an equation - free computational approach ( theodoropoulos et al . , 2000 ; kevrekidis et al . , 2003 ) . a particular goal is to understand how much of the specific spatial detail is fundamental to the behavior . but turning to the kuramoto - type approximation , where the interaction is assumed to be global , we deliberately ignore local effects . to the extent that the models fail to explain observed types of behavior , we will need to turn next to more detailed models . most of previously proposed models concern populations of _ homogeneous _ ( or indistinguishable ) individuals . furthermore , the dynamical analysis in the literature is often limited to a small subset of the entire parameter space , and a systematic classification of possible global dynamics remains elusive . in the current paper , we study the _ coarse - grained _ alignment dynamics of individual - based animal group models . the measurement of the mean angular deviation of fish schools ( e.g. clupeids and scombroids ; see atz , 1953 ; hunter 1966 ) showed that it varies continuously from no alignment to practically perfect alignment . we account for this continuous change by heterogeneity ( `` quenched noise '' ; characterized by parameters of random variables drawn from a prescribed distribution function ) and the coupling strength . our approach is flexible in that the heterogeneity can be introduced in various places in the model , and the way we analyze different heterogeneity cases does not require any significant modification . the rest of the paper is organized as follows : models for homogeneous and heterogeneous animal groups are described in secs . [ model1 ] and [ model_extension ] , and our approach , equation - free polynomial chaos , is explained in secs . [ method1 ] and [ method2 ] . coarse - grained dynamical analysis and its comparison with fine - scale dynamics , for a system of two informed individuals and a large number of heterogeneous uninformed individuals , are presented in sec . [ results ] . the case of two groups of heterogeneous informed individuals is presented in sec . [ results2 ] . we conclude with a brief discussion in sec . [ conclusions ] . we briefly discuss a `` minimal '' model proposed by nabet et al . ( 2006 ) , which we extend in our study . it concerns the alignment dynamics of a homogeneous population of _ indistinguishable _ @xmath0 individuals with two subgroups of informed individuals ( `` leaders '' ) with populations @xmath1 and @xmath2 respectively and @xmath3 uninformed individuals ( `` followers '' ) , where @xmath4 : @xmath5 here @xmath6 characterizes the average direction of the individuals in each of the two informed subgroups for @xmath7 and the average direction of the uninformed individuals for @xmath8 . @xmath9 is the corresponding informed , preferred direction ( @xmath10 can be set to zero without loss of generality ) and @xmath11 is the coupling strength . this minimal model corresponds to the reduced system of the following system of @xmath0 individuals ( nabet et al . , 2006 ) : @xmath12 where the angle @xmath13 characterizes the direction in which the @xmath14th individual is heading ( we will refer to it as `` orientation '' ) . the average direction @xmath6 is defined as the angle of the average of the phasors ( when each individual s dynamical state is considered as a phasor of unit radius and a phase angle ) of the individuals in the @xmath15th subgroup ; @xmath16 is the magnitude of the average of the phasors . formally , this is written as @xmath17 the large population model in eq . ( [ m3 ] ) has a separation of time scales . individuals within each subgroup synchronize quickly , i.e. , @xmath16 quickly converges to 1 . the slow dynamics are described by the reduced system ( eq . ( [ reduced ] ) ) where the variables @xmath6 characterize the _ lumped behavior _ of each of the three subgroups . it is assumed that the alignment ( orientational ) dynamics are independent of the translational counterpart ( sepulchre et al . , 2005 ) ; hence , the dynamical state of an individual can be characterized by its orientation . the functional form for mutual interaction is borrowed from the well - known kuramoto model ( kuramoto , 1984 ) , a prototypical model for coupled nonlinear oscillators . this simplified global interaction model is consistent with an observation that the strongest correlations are observed between the ( speed and ) direction of the individual and the average ( speed and ) direction of the entire school ( partridge , 1982 ) : in the mean - field form of the kuramoto model , the interaction term can be rewritten as @xmath18 where @xmath19 in this alternate expression , the dependence on the difference between the individual direction and the average direction stands out explicitly . in the absence of coupling ( @xmath20 ) , each leader eventually heads for its preferred direction . nontrivial dynamical behavior for the minimal model ( eq . ( [ reduced ] ) ) are studied in nabet et al . ( 2006 ) ; bifurcations are analyzed for the global phase space in the case @xmath21 and @xmath22 . the aforementioned models concern populations of _ homogeneous _ subgroups , where the individuals in each subgroup quickly synchronize , nearly perfectly ( @xmath23 ) , during the initial transients ( nabet et al . , 2006 ) . in the more general case , the mean angular deviation of fish schools is finite ( atz , 1953 ; hunter , 1966 ) , which is not captured in this `` minimal '' model . we extend the model to account for the distribution of directions within schools , assuming it arises from the _ heterogeneity _ among the group members . we introduce the heterogeneity in the following two ways : + ( i ) _ two leaders and many heterogeneous followers _ first we consider the cases when the population consist of two leaders ( which possibly represent lumped behavior of groups of homogeneous leaders ) and @xmath24 followers : where the heterogeneity is accounted for through the tendency to deviate from the average direction , characterized by @xmath26 , an i.i.d . random variable drawn from a prescribed distribution function @xmath27 ( of standard deviation @xmath28 with mean value zero ) . for notational convenience , we drop a subscript of a variable to represent a random variable of a proper length ( _ cf . _ @xmath26 and @xmath29 ) . as @xmath10 can be set to zero without loss of generality , @xmath30 and @xmath31 are control parameters . in the current study , we consider @xmath27 to be gaussian , but our analysis is not limited to this particular choice . + ( ii ) _ two groups of heterogeneous leaders _ secondly , we consider two groups of heterogeneous leaders without any followers , focusing only on the dynamics among leaders . the heterogeneity is accounted for by introducing randomness in the angles preferred by the leaders . the orientations of the leaders in each group are denoted by @xmath32 s and @xmath33 s ( of sizes @xmath1 and @xmath2 ) respectively : where the preferred angles @xmath35 and @xmath36 are randomly drawn from prescribed distributions @xmath37 and @xmath38 ( i.e. , i.i.d . random variables of standard deviations @xmath39 and @xmath40 ) , respectively . we set @xmath41 , and will vary @xmath42 and @xmath43 @xmath44)$ ] as control parameters ( and investigate some cases of different values of @xmath39 and @xmath40 in sec . [ results2b ] ) . direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] direct integration of a system of two leaders ( open circles ; dashed lines indicate preferred angles ) and 300 followers ( dots ) , initialized from uniformly distributed orientations with randomly assigned heterogeneity variable ( i.e. , no initial correlations between @xmath45 and @xmath29 ) , is shown for an initial transient [ ( a ) to ( c ) ] , and for much longer time scales [ ( d ) to ( f ) ] . insets illustrate time evolution of the followers orientations on the @xmath46 plane , where strong correlations develop during a short time @xmath47 . after that , the leaders and followers , the latter effectively as a `` unit '' , slowly drift to the stable steady state . it takes of the order of @xmath48 for the system to asymptotically converges to this final state . ( @xmath49 ) . , title="fig : " ] the kuramoto model , a paradigm for all - to - all , phase - coupled oscillator models , has been extensively studied and used to shed light on many synchronization phenomena ( kuramoto , 1984 ; acebrn et al . , 2005 , and references therein ) . this model has the property that , in the full synchronization regime ( of large enough @xmath31 values ) , phase angles become quickly correlated with ( or sorted " according to ) the natural frequencies during the initial short transients ( moon et al . , similar correlations ( between the angles and heterogeneity random variables ) are expected to arise in the current model ( which is indeed the case , as will be shown later in fig . [ correlation ] ) , since the coupling term is qualitatively similar . as in moon et al . ( 2006 ) , we choose expansion coefficients in wiener polynomial chaos as coarse - grained `` observables '' , to explore low - dimensional , coarse - grained dynamics . wiener(-hermite ) polynomial chaos was introduced by wiener ( 1938 ) , who represented a random process in terms of functional expansions of wiener process ( historically , this method has been termed as polynomial `` chaos '' , because of its initial usage on homogeneous chaos , such as turbulence and brownian motion , rather than the nature of the method ) . ghanem and spanos ( 1991 ) later extended this idea to treat random processes as functional expansions of random variables , or elements in the hilbert space of random functions , in which a spectral representation in terms of polynomial chaos is identified . the projections ( or coefficients ) on the polynomial base then can be determined through a galerkin approach . this method was subsequently applied in uncertainty quantification of various problems ( e.g. , see ghanem , 1999 ; jardak et al . , 2002 ) , and has been extended to general situations using the askey scheme ( xiu et al . , 2002 ; now known as generalized polynomial chaos ) . in this method , dependent random variables ( @xmath45 of the followers for the case ( i ) , and @xmath50 and @xmath51 for the case ( ii ) ) are expanded in polynomials of independent random variables ( @xmath29 , or @xmath52 and @xmath53 ) using appropriately chosen basis functions . details for the two cases are as follows : + ( i ) _ two leaders and many heterogeneous followers _ for convenience , we introduce the unit gaussian random variable @xmath54 . using this newly defined variable , we expand @xmath55 ( i.e. @xmath56 ) in hermite polynomials of @xmath57 [ @xmath58 = 1 , @xmath59 , @xmath60 , @xmath61 : @xmath62 where @xmath63 is the highest order retained in the truncated series , @xmath64 is the @xmath65th hermite polynomial , and the @xmath66 s are the expansion coefficients which will be referred to simply as `` chaos coefficients '' in this paper . wiener polynomial chaos , utilizing hermite polynomials as basis functions , is the appropriate choice for gaussian random variables that we consider in the present study . the probability density function of the gaussian random variables appears as the weighting function of hermite polynomials , and the hermite polynomial expansion is suggested to converge exponentially for gaussian processes ( lucor et al . , 2001 ) . for other random variables , use of different basis functions ( for instance , legendre polynomials for uniform random variables ) has been suggested for fast convergence , which is the basis of the development of the generalized polynomial chaos ( xiu et al . , 2002 ) . we choose the first few nonvanishing chaos coefficients @xmath66 s , as well as the orientations of the leaders ( @xmath67 and @xmath68 ) , to be the coarse - grained `` observables '' . due to symmetry , all the even order @xmath66 s vanish , except for the zeroth order @xmath69 that corresponds to the average direction of the followers . geometrically , @xmath70 and @xmath71 respectively represent a measure for the linear order spread of the angles ( the `` slope '' between @xmath45 and @xmath29 ) and the cubic order measure . in the continuum limit ( @xmath72 ) , the chaos coefficients can be exactly determined using the orthogonality relations for hermite polynomials . however , in the _ finite _ cases of single realization we consider , @xmath73 , those relations hold only approximately , and the coefficients are evaluated using least squares fitting , following moon et al . + ( ii ) _ two groups of heterogeneous leaders _ in the second case , we expand @xmath50 and @xmath51 in terms of @xmath52 and @xmath53 , respectively : @xmath74 where the chaos coefficients and are the coarse `` observables '' of our choice , @xmath64 s are hermite polynomials ( for gaussian @xmath75 and @xmath76 ) , and @xmath77 and @xmath78 are unit gaussian random variables . + a prerequisite to _ coarse - grained _ dynamical analysis ( which is the main goal of the current study ) is , in a traditional sense , an explicit derivation of coarse - grained governing equations . in principle , such equations for chaos coefficients , in the continuum limit ( @xmath72 ) , might be obtained through a stochastic galerkin method ( ghanem and spanos , 1991 ) . in the present study , we do not even attempt to derive such equations . we circumvent their derivation by using an equation - free multiscale computational approach ( theodoropoulos et al . , 2000 ; kevrekidis et al . , 2003 , 2004 ) . this approach enables us to explore the coarse - grained dynamics without the assumption of the continuum limit . the premise of this approach is that coarse - grained governing equations conceptually _ exist _ , but are not explicitly available in closed form . this approach is based on the recognition that short bursts of appropriately initialized microscopic ( fine - scale ) simulations during a time horizon @xmath79 and the projection of the results onto coarse - grained variables , say , result in time - steppers ( mappings ) for those variables @xmath80 ( which is effectively the same as the discretization of unavailable equations ) : @xmath81 one then processes the results of the short simulations to estimate various coarse - grained quantities ( such as time derivatives , action of jacobians , residuals ) to perform relevant coarse - grained level numerical computations , as if those quantities were obtained from coarse - grained governing equations . for instance , one can _ integrate _ unavailable governing equations in time ( coarse projective integration ; see below ) , or compute the steady states of the above coarse time - stepper , by utilizing fixed point algorithms ( such as newton - raphson or newton - gmres ) . equation - free computations consist of the following steps : * identify coarse - grained variables ( `` coarse observables '' ) that sufficiently describe both the micro- and macroscopic dynamics ; in our study , they are @xmath66 s ( and @xmath82 s ) . for convenience , we denote the microscopic ( macroscopic ) descriptions by @xmath83 ( @xmath84 ) . * choose an appropriate _ lifting _ operator @xmath85 , which maps @xmath84 to one ( or more ) consistent description(s ) @xmath83 ( for the purposes of variance reduction and ensemble - averaging ) . in the current study , this can be achieved by using the relations in eqs . ( [ expansion ] ) and ( [ expansion2 ] ) ; once random variables are drawn , these relations are used to obtain corresponding @xmath83 . * starting from lifted initial condition(s ) @xmath86 , run the microscopic simulator to obtain @xmath87 at a later time ( @xmath88 ) . * use an appropriate _ restriction _ operator @xmath89 ( least squares fitting , in the current study ) which maps the microscopic state(s ) to the macroscopic description @xmath90 , which effectively results in time series of coarse observables , or their coarse time - stepper @xmath80 ; @xmath91 . * apply desired numerical techniques using the coarse - grained variables obtained from the step 4 . and repeat some of the above steps as needed . an extensive discussion can be found in kevrekidis et al . ( 2003 , 2004 ) . direct integration of the `` fine - scale '' model of eq . ( [ microeq1 ] ) in the strong coupling regime ( @xmath92 ) , started from randomly assigned orientations and the heterogeneity variable ( the latter is a gaussian random variable ) , illustrates that a strong correlation between @xmath45 and @xmath29 develops during a short , initial transient time ; the orientations of the followers quickly become a monotonically increasing function of their heterogeneity variable ( fig . [ correlation ] ) , after which they slowly drift as a `` unit '' until they settle down in the final steady state . during the latter slow drift , the system can be described as two leaders and a _ single _ `` clump '' of followers , whose coarse - grained states can be successfully described by a small number of chaos coefficients . a similar time scale separation exists in the model of homogeneous populations . in this case , followers quickly collapse asymptotically to the _ same _ direction ( nabet et al . , 2006 ) . ( color online ) accelerated computation of stable steady states via coarse projective integration using five coarse - grained variables , shown here for two different time scales ( @xmath49 ) . initially all the values are assigned to be 0 . both @xmath70 and @xmath71 reach their steady state values relatively quickly ( see ( a ) ) , while the others are slowly varying ( see ( b ) ; they are still varying at @xmath93 ) . dots represent the time intervals during which short direct integration is performed ( and restricted ) , in the course of the projective integration using forward euler method . solid lines represent the trajectories of direct full integration during the entire time . higher efficiency can be achieved by optimally choosing the time horizon for the direct integration , the projection stepsize , and projection method . , title="fig : " ] ( color online ) accelerated computation of stable steady states via coarse projective integration using five coarse - grained variables , shown here for two different time scales ( @xmath49 ) . initially all the values are assigned to be 0 . both @xmath70 and @xmath71 reach their steady state values relatively quickly ( see ( a ) ) , while the others are slowly varying ( see ( b ) ; they are still varying at @xmath93 ) . dots represent the time intervals during which short direct integration is performed ( and restricted ) , in the course of the projective integration using forward euler method . solid lines represent the trajectories of direct full integration during the entire time . higher efficiency can be achieved by optimally choosing the time horizon for the direct integration , the projection stepsize , and projection method . , title="fig : " ] we begin by accelerating the approach to a stable steady state using an equation - free algorithm , the coarse projective integration method ( gear and kevrekidis , 2003 ) . in contrast to a conventional , direct integration of the full fine - scale model during the _ entire _ time ( until sufficient convergence to stable , final states ) , this method exploits smoothness in the coarse variables ( estimated through a direct integration during a _ short _ time ) , in order to extrapolate and take a large projective time - step ( compared to the original integration time - step size ) . this saves computational effort . the procedure consists of ( @xmath94 ) _ lifting _ ( appropriate initialization of the fine - scale simulator , an integrator of eq . ( [ microeq1 ] ) , consistent with prescribed coarse - grained values ) , ( @xmath95 ) _ direct integration _ of the microscopic simulator during a relatively short time interval ( but long enough to accurately estimate local coarse - grained time derivatives ) , ( @xmath96 ) _ restriction _ ( of fine - scale description onto coarse - grained variables ) , and ( @xmath97 ) _ taking a projective step _ ( using a traditional numerical integration scheme such as forward euler ) . the computational payoff of this method depends on the ratio between a short direct integration time interval , the projective time - step size , and the computational effort required for lifting / restriction procedures ( see e.g. , rico - martinez et al . , 2004 ) . more importantly , successful computation of steady states through this method naturally attests to the _ validity _ of the chosen coarse - grained observables in describing _ both _ fine - scale and coarse - grained states . ( left panel ) a bifurcation diagram observed on one leader @xmath68 , computed using auto2000 ( @xmath98 ) . solid ( dashed ) lines represent stable ( unstable ) branches . there exist a few other unstable branches that are not shown here . at some critical value of @xmath30 , an unstable state in the upper branch undergoes a forward pitchfork bifurcation ; two unstable states coincide . the lower branch of `` trivial '' solutions does not exhibit any bifurcation . ( right panels ) snapshots of two ( symmetric ) stable states in the bistable regime ( @xmath99 ) , marked by dots in the left panel . ] [ cols="^,^,^,^,^,^,^",options="header " , ] projective integration using five coarse - grained variables ( @xmath100 , and the first three non - vanishing @xmath66 s ; @xmath69 , @xmath70 , and @xmath71 ) follows virtually the same trajectories of the full , direct integration ( fig . [ pi ] ) , even if @xmath29 is _ newly _ drawn at each lifting ; the agreement is even better if the same @xmath29 were used ( hence the dynamics are fully deterministic ) . both _ lifting _ ( simply using eq . ( [ expansion ] ) ) and _ restriction _ ( least squares fitting ) operations require minimal computational efforts . therefore , the computational efficiency in the present case is nearly exclusively determined by the projective step size , which is a factor of about four in fig . [ pi ] ; with a more sophisticated projection algorithm , a higher efficiency can be obtained . we see that both @xmath70 and @xmath71 reach their steady state values quickly ( @xmath101 ) , showing that the correlation between @xmath45 and @xmath29 are fully developed by then . however , the other chaos coefficient @xmath69 ( representing the average direction ) slowly drifts towards the steady state , and so do @xmath67 and @xmath68 ( note that it is still varying at @xmath93 ) ; the computation of an asymptotic , steady state requires a very long time integration . direct integration ( including projective integration ) can not compute unstable steady states and are inappropriate for stability computations and parametric bifurcation studies . both stable and _ unstable _ steady state values can be systematically ( and much more efficiently than the projective integrations ) computed by applying coarse - grained fixed point algorithms to the steady state condition of eq . ( [ coarse_stepper ] ) , i.e. , @xmath102 , in _ much lower _ dimension than that of individual level . we use the coarse newton - gmres ( kelley , 1995 ) , a matrix - free , method to compute coarse - grained fixed points . we observe that the algorithm accurately converges within a few steps ( tab . [ gmres ] ) ; the converged values are accurately consistent with the restricted values of the fixed point solution of the detailed ( i.e. , ( @xmath0 + 2)-dimensional ) problem , within prescribed convergence tolerance . by combining a coarse fixed point algorithm with pseudo - arclength continuation ( keller , 1987 ) , we numerically compute coarse - grained bifurcation diagrams in the following sections . the computational efficiency of the coarse fixed point algorithm varies with the choice of the initial guess for the iteration . with a totally uneducated guess , it could take even longer than the direct integration ( note that the latter never computes the exact solutions ) ; however , during the continuation computation shown below , a good initial guess is always available from the previous parameter value(s ) , and even several orders of magnitude of computational efficiency can be achieved . ( a ) a bifurcation diagram observed on one ( arbitrarily chosen ) follower , as a function of @xmath30 ( @xmath98 ) , computed using auto2000 ( the same case as in fig . [ micro ] ) . superscript ` f ' has been added to emphasize that this is the orientation of _ a follower_. a few other existing unstable branches are not included here . the upper branch undergoes a pitchfork bifurcation and becomes stable . ( b ) a coarse bifurcation diagram observed on @xmath69 ( average direction ) , obtained by the coarse newton - gmres method with pseudo arc - length continuation . only a blowup around the bifurcation point coarse - grained dynamics exhibit the same structure as in the fine - scale level . filled ( open ) circles represent stable ( unstable ) steady states . , title="fig : " ] ( a ) a bifurcation diagram observed on one ( arbitrarily chosen ) follower , as a function of @xmath30 ( @xmath98 ) , computed using auto2000 ( the same case as in fig . [ micro ] ) . superscript ` f ' has been added to emphasize that this is the orientation of _ a follower_. a few other existing unstable branches are not included here . the upper branch undergoes a pitchfork bifurcation and becomes stable . ( b ) a coarse bifurcation diagram observed on @xmath69 ( average direction ) , obtained by the coarse newton - gmres method with pseudo arc - length continuation . only a blowup around the bifurcation point coarse - grained dynamics exhibit the same structure as in the fine - scale level . filled ( open ) circles represent stable ( unstable ) steady states . , title="fig : " ] we first analyze the detailed @xmath103-dimensional fine - scale model in the full synchronization regime , in order to obtain insights on fine - scale dynamics to be compared with our coarse - grained analysis below . we use auto2000 ( doedel et al . , 2000 ) to compute the fine - scale bifurcation diagrams as functions of @xmath30 at a fixed value of @xmath31 ; only projections for one leader ( @xmath68 ) are shown in fig . [ micro ] and for one follower in fig . [ macroth2 ] ( a ) . all the other followers exhibit essentially the same dynamical behavior as the one shown here ( except for some quantitative differences ) . the interaction between the individuals causes the steady state directions of the leaders to deviate from the preferred angles @xmath104 and @xmath30 , respectively . such deviation can occur in two directions , either toward the region bounded by @xmath105 $ ] ( an `` obvious '' steady state where followers are directed in between the directions of the leaders ; see fig . [ micro ] ( b ) ) or the other way around ( e.g. , fig . [ micro ] ( a ) ) . the analysis shows that for small @xmath30 values only the former state is stable , while for large values , both of these states become stable . the branches for `` obvious '' stable steady states , which correspond to lower straight solid lines , exhibit no bifurcations ( see figs . [ micro ] and [ macroth2 ] ( a ) ) . on the other branches , forward pitchfork bifurcations at some critical value of @xmath30 give birth to another stable branch ( a state on this stable branch is shown in fig . [ micro ] ( a ) ) , as well as two unstable asymmetric solution branches , hence the population becomes bistable . the critical value of @xmath30 for the onset of the bistability depends on @xmath31 ( precisely speaking , @xmath106 ) ; the critical value is @xmath107 at @xmath108 . as @xmath31 decreases further , the critical value monotonically increases until fully synchronized steady states lose stability at some critical value of @xmath31 . we now compute coarse - grained steady state solutions . a coarse - grained bifurcation diagram for @xmath69 ( representing the average direction of the followers ) is compared with the corresponding diagram observed for _ one _ follower , in figs . [ macroth2 ] ( b ) and ( a ) ; ( b ) is a blowup of the region around the bifurcation . both sides of the bifurcation point can be described by the _ same _ set of coarse - grained observables , which clearly summarize group level dynamical behavior of the followers before and after the bifurcation . as @xmath31 decreases in the kuramoto model , oscillators get desynchronized ( kuramoto , 1984 ) , starting with the oscillator with the maximum value of @xmath109 ( the `` extreme '' oscillator ) through a saddle - node ( actually a `` sniper '' ) bifurcation on a limit cycle ( moon et al . , 2006 ) . we expect the same type of bifurcation to occur in this model . however , when we try to compute the coarse - grained steady states as functions of @xmath31 using the previously mentioned five coarse variables ( via coarse newton - gmres method and pseudo arc - length continuation , neither a bifurcation nor an unstable branch is appropriately identified . the computation , initialized at large @xmath31 steady states , accurately follows stable branches down to some critical value of @xmath31 ( where the transition occurs ) , and then fails to converge . our coarse - grained observables are not sufficient to describe the states on the `` other side '' of the bifurcation point , as we will explain below . ( color online ) ( a ) a bifurcation diagram observed on a few followers , including the one with the maximum of @xmath109 ( the `` extreme '' follower ) , as a function of @xmath31 ( @xmath110 ) , computed using auto2000 . a critical value where the extreme individual gets desynchronized corresponds to a saddle - node bifurcation point on a limit cycle ( a `` sniper '' bifurcation ) . except for the extreme follower , stable and unstable branches nearly coincide ( see the inset ) . ( b ) in order to capture the fine - scale bifurcation , the angle of the extreme follower has to be discounted from the chaos expansion and considered as an _ extra _ coarse - grained variable ( see text ) . we distinguish these chaos coefficients ( from the ones used so far ) by adding a prime . it was computed via the coarse newton - gmres method with continuation . , title="fig : " ] ( color online ) ( a ) a bifurcation diagram observed on a few followers , including the one with the maximum of @xmath109 ( the `` extreme '' follower ) , as a function of @xmath31 ( @xmath110 ) , computed using auto2000 . a critical value where the extreme individual gets desynchronized corresponds to a saddle - node bifurcation point on a limit cycle ( a `` sniper '' bifurcation ) . except for the extreme follower , stable and unstable branches nearly coincide ( see the inset ) . ( b ) in order to capture the fine - scale bifurcation , the angle of the extreme follower has to be discounted from the chaos expansion and considered as an _ extra _ coarse - grained variable ( see text ) . we distinguish these chaos coefficients ( from the ones used so far ) by adding a prime . it was computed via the coarse newton - gmres method with continuation . , title="fig : " ] a fine - scale bifurcation diagram ( computed using auto2000 ) obtained by starting from a stable steady state on the lower branch in fig . [ micro ] is shown in fig . [ turningpt ] ( a ) . here the diagrams for two leaders and only a few followers , including the extreme one , are shown . we find that both stable and unstable branches for each angle nearly coincide for all the individuals ( see inset of fig . [ turningpt ] ( a ) ) , except for the extreme one . as the difference between stable and unstable branches ( at the same value of @xmath31 ) is appreciable _ only _ when observed on this extreme oscillator , a smooth mapping between @xmath45 and @xmath29 does not prevail for unstable states , and the previously used chaos coefficients are not appropriate any more . taking these observations into account , it is easy to remedy the situation as follows : the fact that stable and unstable branches nearly coincide , discounting the extreme follower , suggests that all the individuals _ except for the extreme follower _ can be again described by the same set of chaos coefficients . thus we treat the orientation of the extreme one separately ( introducing it as an additional coarse - grained variable ) , and discount it from the polynomial chaos expansion . ( from the fact that the extreme follower gets desynchronized at the transition , one can also intuitively see that followers have to be considered as a combination of a clump of synchronized `` bulk '' and a separate , extreme one . ) we compute the solutions with continuation , using this new set of _ six _ coarse variables , which captures the bifurcation and appropriately describes the unstable steady states ( fig . [ turningpt ] ( b ) ) ; we have analyzed exactly the same realization used in fig . [ turningpt ] ( a ) for direct comparison . when bifurcation diagrams are computed for ensembles of many realizations , an uncertainty will arise in the exact quantification of the bifurcation point , due to the fluctuation of finite - dimensional random variables among realizations , while the results are qualitatively the same as those of a single realization ( xiu et al . , 2005 ) . the coarse bifurcation results shown in fig . [ turningpt ] ( b ) illustrate that the steady state directions of the leaders and the average direction of followers ( @xmath111 , discounting the extreme one ; a prime is added to distinguish it from the previously used notation ) are virtually the same for a range of @xmath31 . only higher order chaos coefficients ( only @xmath112 is shown in fig . [ turningpt ] ) appreciably vary as a function of @xmath31 , which means that individuals spread more widely as @xmath31 decreases , until the extreme one eventually starts to oscillate freely , while the average steady state direction remains the same . lines : a bifurcation diagram observed on @xmath67 in the minimal model ( the first two odes in eq . ( [ reduced ] ) with @xmath113 ) , for varying @xmath30 at a fixed value of @xmath114 , obtained by auto2000 . for large enough preferred angles ( @xmath115 ) , the system becomes bistable through a forward pitchfork bifurcation . circles : a coarse bifurcation diagram near the pitchfork bifurcation , observed on @xmath69 , as a function of @xmath116 , computed by the coarse newton - gmres method . filled ( open ) circles represent stable ( unstable ) steady states . ] here we explore both the fine - scale and coarse - grained dynamics of a model for two groups of _ heterogeneous _ leaders ( with no followers ) shown in eq . ( [ for_many ] ) , and compare the results of the two different scales . one notable difference from the kuramoto model is that `` oscillators '' in eq . ( [ for_many ] ) do not have finite random variables ( natural frequencies ) , hence there is no onset of the synchronization that occurs at a finite value of @xmath31 ( or , they can be alternatively seen as kuramoto - like oscillators of zero natural frequencies , which result in the onset at @xmath117 , hence they get synchronized for all @xmath31 values ) . the analysis of the minimal model ( the first two of eq . ( [ reduced ] ) with @xmath118 ) reveals that for large enough @xmath119 the system exhibits bistability for a certain range of @xmath31 ( nabet et al . , 2006 ) , as in the previous case in sec . [ results ] . here we will vary @xmath43 as the main parameter for two different values of @xmath31 . for large coupling strengths ( @xmath120 ) , the bistability in the minimal model appears through a forward pitchfork bifurcation , when @xmath30 is varied as a parameter ( fig . [ microk2.4 ] ) . this minimal model can be seen as a special case of the current model , where both @xmath52 and @xmath53 are assumed to be delta functions and each group consists of identical individuals . coarse - grained bifurcation diagrams observed on the first two chaos coefficients : ( a ) @xmath69 ; the average direction of the first group of leaders , and ( b ) @xmath70 ; the `` slope '' between @xmath50 and @xmath52 , as functions of @xmath43 . these are blowups of the region around the forward pitchfork bifurcation point in fig . [ microk2.4 ] . , title="fig : " ] coarse - grained bifurcation diagrams observed on the first two chaos coefficients : ( a ) @xmath69 ; the average direction of the first group of leaders , and ( b ) @xmath70 ; the `` slope '' between @xmath50 and @xmath52 , as functions of @xmath43 . these are blowups of the region around the forward pitchfork bifurcation point in fig . [ microk2.4 ] . , title="fig : " ] we begin by asking whether our model for heterogeneous groups exhibits similar types of dynamical behavior . one can also do accelerated computations of steady states using the coarse projective integration , but here we skip such computations and present only the coarse bifurcation analysis results . _ coarse _ bifurcation diagrams obtained through the coarse newton - gmres method ( kelley , 1995 ) and pseudo arc - length continuation ( keller , 1987 ) ( for gaussian distributions of @xmath52 and @xmath53 ; @xmath121 , @xmath122 ) show that the heterogeneous groups indeed exhibit the same qualitative type of coarse dynamical behavior around the pitchfork bifurcation point ( fig . [ macrok2.4 ] ) . as we consider symmetric unimodal distribution functions , all the even order chaos coefficients ( except for @xmath69 and @xmath123 ) virtually vanish . the diagram for @xmath69 of the first group @xmath124 ( average direction ) exhibits reasonably good quantitative agreement with the corresponding diagram for the minimal model , within fluctuations of finite - size random variables , shown in figs . [ microk2.4 ] and [ micromacrok1.8 ] . it is interesting to note that at the critical point , all the followers are headed for the same direction ( @xmath125 , which corresponds to the `` slope '' between @xmath52 and @xmath50 ) ; see fig . [ macrok2.4 ] ( b ) . lines : a bifurcation diagram for the minimal model of two leaders , for varying @xmath30 at a fixed value of @xmath126 , obtained by auto2000 . for large enough preferred angles ( @xmath127 ) , the system becomes bistable , but the nature of the bifurcation is different from that of higher @xmath31 value cases ( a saddle - node vs. a pitchfork bifurcation ; see fig . [ microk2.4 ] ) . circles : a coarse bifurcation diagram observed on the average direction of the first group of leaders ( @xmath69 ) around the saddle - node bifurcation , as a function of @xmath116 , computed via the coarse newton - gmres method with continuation . filled ( open ) circles represent stable ( unstable ) steady states . ] ( color online ) coarse - grained bifurcation diagrams near a turning point in fig . [ micromacrok1.8 ] , for @xmath52 distributions of three different widths ( the standard deviation @xmath39 = 0.1 for circles ; 0.2 for squares ; 0.3 for triangles ) , obtained via the coarse newton - gmres method and continuation . the standard deviation for the second group , @xmath40 , is kept the same at 0.1 ( @xmath128 ) . filled and open symbols represent stable and unstable states , respectively . ( a ) the first chaos coefficients @xmath69 ( average direction of the first group ) are nearly the same for the three cases . the difference between the cases becomes apparent in higher order coefficients that reflect the degree of spreading ; see @xmath70 in ( b ) . , title="fig : " ] ( color online ) coarse - grained bifurcation diagrams near a turning point in fig . [ micromacrok1.8 ] , for @xmath52 distributions of three different widths ( the standard deviation @xmath39 = 0.1 for circles ; 0.2 for squares ; 0.3 for triangles ) , obtained via the coarse newton - gmres method and continuation . the standard deviation for the second group , @xmath40 , is kept the same at 0.1 ( @xmath128 ) . filled and open symbols represent stable and unstable states , respectively . ( a ) the first chaos coefficients @xmath69 ( average direction of the first group ) are nearly the same for the three cases . the difference between the cases becomes apparent in higher order coefficients that reflect the degree of spreading ; see @xmath70 in ( b ) . , title="fig : " ] the hermite polynomial expansion converges so quickly that the expansions can be accurate even when truncated at the third order . due to the reflection symmetry ( about @xmath43/2 ) , coefficients have similar structures as the ones , after proper reflection and translation . only results on are presented here . as the coupling strength decreases across @xmath129 , the nature of the bifurcation changes ( from a pitchfork ) to a saddle - node bifurcation ( fig . [ micromacrok1.8 ] ) at @xmath129 , which also occurs in the model for homogeneous populations ; the nature of the transition between these different bifurcations , a higher codimension bifurcation , has been discussed in nabet et al . ( 2006 ) . so far we have considered statistically similar groups , namely @xmath21 and @xmath130 ; they differed only by average preferred directions . it is natural to ask how the dynamics change as the parameters concerned with the distributions ( for the preferred directions ) are varied . it is readily expected that the essential dynamics of two different - size groups can be reflected in the minimal model using two different coupling strengths , which is considered in nabet et al . here we consider only the cases with varying width of the distributions ( @xmath131 ) , which has no analog in the minimal model . coarse bifurcation diagrams for three different gaussian distributions for @xmath52 ( @xmath39 is varied while @xmath40 is kept at 0.1 ; see fig . [ vary_std ] ) show that the average directions ( @xmath69 s ) hardly vary with the width of the distributions ; the primary parameter that affects on the average direction is the _ group size_. for the distributions of different widths , the fixed point computation with continuation fails to converge at different values of @xmath69 s ; points marked by arrows in fig . [ vary_std ] are the last points the newton - gmres computations converged in each case , when approached from the stable branches . such a failure of convergence can be expected , because the steady states on this unstable branch overlap with another nearby unstable branch ( which is not shown in this figure , but was shown in fig . [ micromacrok1.8 ] ) ; characterizing the distribution with a few wiener chaos coefficients does not provide an accurate description any more . the differences between the three cases ( of different distribution widths ) manifest themselves clearly in higher order chaos coefficients . while the average behavior remains nearly the same ( fig . [ vary_std ] ( a ) ) , individuals in the group spread more widely ( as reflected in @xmath70 and higher order coefficients ; fig . [ vary_std ] ( b ) ) , as the width of the random variable ( distribution ) increases . we have demonstrated a computational venue ( an equation - free polynomial chaos approach ) to study coarse - grained dynamics of individual - based models accounting for the _ heterogeneity _ among the individuals in animal group alignment models . we considered _ finite _ populations of ( i ) two `` leaders '' ( which have direct knowledge on preferred directions ) and @xmath132 uninformed , heterogeneous `` followers '' , and ( ii ) two groups of heterogeneous `` leaders '' . we explored the _ coarse - grained _ , group level ( low - dimensional ) dynamics using the polynomial chaos expansion coefficients as coarse - grained observables ; these observables account for rapidly developing correlations between random variables , and sufficiently specify both fine - scale and coarse - grained ( group - level ) dynamical states . all the analysis in our study was done expressively avoiding the derivation of coarse - grained governing equations , following a _ nonintrusive _ , equation - free computational approach wrapped around the direct system simulator . it should be noted that we have not assumed that @xmath0 is infinitely large ( so - called the `` continuum limit '' ) . our approach can be used for systems of any finite , large number of populations , and it can be equally applied to various types of random variables ( following generalized polynomial chaos ) and/or various heterogeneity . we compared our results with those of minimal models that do not account for heterogeneity among the individuals . they show good agreement in the lowest order ( i.e. , average directions ) , which clearly highlights the correspondence between the individual- and group - level dynamics ( figs . [ microk2.4 ] and [ micromacrok1.8 ] ) . indeed this implies that the results in nabet et al . ( 2006 ) , where no heterogeneity is explicitly accounted for , are more robust than demonstrated in that paper alone . in order to analyze different coarse - grained bifurcations , it became necessary to use different sets of coarse - grained variables , even if the model is _ the same _ in the fine - scale level ( fig . [ turningpt ] ) . this clearly shows that an appropriate choice of coarse - grained observables ( in terms of which one can obtain useful closures ) is an essential step ; different coarse - grained observables are required , as the same fine - scale model closes differently . in the present study , we assumed that the orientational dynamics can be separated from their translational counterpart , and considered the simplest nontrivial cases of all - to - all ( `` all - visible '' ) , sinusoidal coupling . our future work will involve the incorporation of the translational dynamics and more complicated coupling / network topology , including heterogeneous couplings . our work presented here is the first step of our effort toward the development of more detailed ( and biologically more plausible ) models and their _ coarse - graining_. s.j.m . and i.g.k . were financially supported by doe and nsf grant ef0434319 . s.a.l . was supported in part by nsf grant ef0434319 and darpa grant hr00110510057 . b.n . and n.e.l . were supported in part by onr grants n000140210826 and n000140410534 . atz , j. w. , 1953 . orientation in schooling fishes , in proceedings of a conference on orientation in animals . office of naval research , pp . 103 - 130 . department of the navy , washington , d. c. , section 2 . kevrekidis , i. g. , gear , c. w. , hyman , j. m. , kevrekidis , p. g. , runborg , o. , theodoropoulos , c. , 2003 . equation - free coarse - grained multiscale computation : enabling microscopic simulators to perform system - level tasks . sciences 1 ( 4 ) , 715 - 762 ; e - print physics/0209043 . leonard , n. e. , paley , d. , lekien , r. , sepulchre , r. , fratantoni , d. m. , davis , r. , 2006 . collective motion , sensor networks and ocean sampling . proceedings of the ieee , special issue on `` the emerging technology of networked control systems '' , to appear . lucor , d. , xiu , d. , karniadakis , g. , 2001 . spectral representations of uncertainty in simulations : algorithms and applications . in proceedings of the international conference on spectral and high order methods ( icosahom-01 ) , uppsala , sweden . nabet , b. , leonard , n. e. , couzin , i. d. , levin , s. a. , 2006 . leadership in animal group motion : a bifurcation analysis . proceedings of the 17th international symposium on mathematical theory of networks and systems , kyoto , japan , july 2006 . sepulchre , r. , paley , d. , leonard , n. e. , 2005 . collective motion and oscillator synchronization , in cooperative control : a post - workshop volume , 2003 block island workshop on cooperative control , eds . v. kumar , n. leonard and a.s . morse , springer - verlag .