Datasets:

kjappelbaum

/

chemnlp-arxiv

like 1

# Introduction With the next generation of large-volume proton-decay searches and neutrino experiments currently in the R&D phase (in particular, LBNE , LENA  and Hyper-K ) there are good prospects to push the current lower bounds on the proton lifetime to the unprecedented level of \(10^{35}\) years. On the theory side, the new information may be, at least in principle, used for further testing of the grand unification paradigm; however, this would require a very good grip on the proton lifetime predictions supplied by specific GUTs. Unfortunately, the quality of the existing estimates is rather limited even in very simple scenarios, see FIGURE [\[estimates\]](#estimates){reference-type="ref" reference="estimates"}, and it is namely due to the low accuracy of the leading-order methods used in most of the relevant calculations. On the other hand, consistent next-to-leading-order (NLO) proton lifetime estimates are parametrically more difficult: First, at the NLO level, the GUT scale \(M_{G}\) must be determined at two-loops; this, however, requires a detailed understanding of the one-loop theory spectrum. Second, the flavour structure of the relevant baryon-number-violating (BNV) currents must be constrained by the existing data to a maximum attainable degree. Third, one has to account for several classes of almost irreducible uncertainties related to the Planck-scale physics (such as, e.g., gravity smearing of the gauge unification pattern ) which are often comparable to the NLO effects. Therefore, the only foreseeable way to overcome this conundrum is to focus on the simplest possible GUTs. In contrast to the minimal \(SU(5)\) Georgi-Glashow model which was shown to be incompatible with the electroweak data already back in mid 1980's, the history of the minimal \(SO(10)\) GUTs is rather non-linear and even after almost 40 years it is still lively and evolving. Interestingly, this can be partly attributed also to the fact that, in the \(SO(10)\) context, the very meaning of minimality is not entirely agreed upon. This owes namely to the relatively large number of potentially viable symmetry breaking chains in \(SO(10)\) characterized by different effective scenarios emerging at intermediate scales. Let us recall that this is not the case in SU(5) simply because there the need to preserve rank reduces the set of Higgs representations available for the GUT symmetry breaking to just few. In a certain sense, this is not the case in supersymmetric theories either because the rigidity of the MSSM gauge unification pattern calls for a single-step breaking where most of the details of the GUT-scale dynamics remain obscured. Thus, besides very special features like natural R-parity conservation etc., the main distinctive characteristics of many models is namely their flavour structure. Hence, with the spectacular failure  of the simplest potentially realistic renormalizable SUSY \(SO(10)\)  (advocated by many to be even the very minimal SUSY GUT ), and, in particular, with no signs of SUSY at the LHC so far, the community's attention naturally drifts back to non-supersymmetric GUTs. In this review, we shall comment in brief on the status of the simplest non-SUSY \(SO(10)\) scenarios and on the latest developments including, in particular, the new upper limits on the seesaw scale obtained recently in the work  and possible future prospects of accurate proton lifetime calculations in this scenario. # The minimal \(SO(10)\) grand unfication The simplest multiplet that can consistently support spontaneous breaking of the \(SO(10)\) gauge symmetry in the SM direction is the 45-dimensional adjoint representation. Together with either the 16-dimensional spinor or the 126-dimensional self-dual part of the maximally antisymmetric tensor the models based on the combinations \(45\oplus 16\) or \(45\oplus 126\) are often regarded to as the minimal renormalizable realizations of the Higgs mechanism in the \(SO(10)\) GUTs. In this respect, it is important to recall that this is not the case in SUSY where the \(F\)-flatness conditions align the VEV of \(45_{H}\) along that of \(16_{H}\) which, although providing the desired rank reduction, leaves a full SU(5) as an unbroken subgroup. Remarkably enough, in the non-supersymmetric Higgs model based on \(45\oplus 16\) or \(45\oplus 126\) the SU(5) trap can not be entirely avoided either. ## The tree-level curse of the minimal SO(10) GUTs The point is that there are two states in the scalar spectrum of either of the two variants of the minimal model that can be simultaneously non-tachyonic only in a narrow region of the parameter space which, unfortunately, happens to support only SU(5)-like symmetry-breaking chains. Indeed, the masses of the color-octet and the \(SU(2)_{L}\)-triplet components of \(45_{H}\) are at the tree level given by \[\begin{aligned} \label{PGBmasses} M^2(1,3,0)_{45} & = & 2 a_2 (\omega_{BL}-\omega_R) (\omega_{BL} + 2 \omega_R) \,, \\ M^2(8,1,0)_{45} & = & 2 a_2 (\omega_R-\omega_{BL}) (\omega_R + 2 \omega_{BL}) \,,\nonumber \end{aligned}\] where \(a_{2}\) is a coupling in the relevant scalar potential (see, e.g.,  ) and \(\omega_{BL}\) and \(\omega_{R}\) are the two independent SM-compatible VEVs in \(45_{H}\) \[\langle 45_{H} \rangle ={\rm diag}(\omega_{BL},\omega_{BL},\omega_{BL},\omega_{R},\omega_{R})\otimes \tau_{2}\] (with \(\tau_{2}\) denoting the second Pauli matrix) which, if hierarchical enough, break the \(SO(10)\) gauge symmetry along two different symmetry breaking chains \[\begin{aligned} \label{chainXII} SO(10)&\stackrel{\omega_{R}}{\longrightarrow} &SU(4)_C \otimes SU(2)_L \otimes U(1)_R \; \stackrel{}{\longrightarrow}\ldots \longrightarrow \; \mbox{SM}\,,\\ \label{chainVIII} SO(10)&\stackrel{\omega_{BL}}{\longrightarrow} & SU(3)_c \otimes SU(2)_L \otimes SU(2)_R \otimes U(1)_{B-L} \; \stackrel{}{\longrightarrow}\ldots \longrightarrow \; \mbox{SM}\. \end{aligned}\] Given that, it is clear that the right hand sides (RHS) of both equations in ([\[PGBmasses\]](#PGBmasses){reference-type="ref" reference="PGBmasses"}) are positive if and only if \(-2\leq\omega_{BL}/\omega_{R}\leq-1/2\), i.e., when there is essentially no hierarchy between \(\omega_{R}\) and \(\omega_{BL}\), otherwise the corresponding vacuum is unstable. Obviously, in such a case, \(\langle 45_{H} \rangle\) is almost homogeneous and the descent is \(SU(5)\)-like; this, however, conflicts with the gauge unification constraints as in the Georgi-Glashow model. ## The quantum salvation Until recently, the argument above was taken as a no-go for any potential viability of the minimal \(SO(10)\) model including the adjoint scalar as a Higgs field responsible for the initial \(SO(10)\) symmetry breakdown. However, as shown in this was premature because the hierarchy between \(\omega_{BL}\) and \(\omega_{R}\) may be stabilized by quantum effects because loop corrections such as those diplayed in FIGURE [\[loops\]](#loops){reference-type="ref" reference="loops"} provide non-negligible positive contributions to the RHS of eq. ([\[PGBmasses\]](#PGBmasses){reference-type="ref" reference="PGBmasses"}). A thorough effective potential analysis  in the simplest \(45\oplus 16\) variant yields (in the notation of ) \[\begin{aligned} \label{310onthevac} \Delta M^2(1,3,0)_{45}&=& \frac{1}{4\pi^2} \left[ \tau^2 +\beta^2(2\omega_R^2-\omega_R\omega_{BL}+2\omega_{BL}^2) +g^4 \left(16 \omega_R^2+\omega_{BL} \omega_R+19 \omega_{BL}^2\right)\right]+{\rm logs}\,, \;\;\; \nonumber\\ \label{810onthevac} \Delta M^2(8,1,0)_{45}&=& \frac{1}{4\pi^2} \left[ \tau^2 +\beta^2(\omega_R^2-\omega_R\omega_{BL}+3\omega_{BL}^2) +g^4 \left(13 \omega_R^2+\omega_{BL} \omega_R+22 \omega_{BL}^2\right)\right]+{\rm logs}\,,\nonumber \end{aligned}\] where the "logs" denote the typically sub-leading logarithmic corrections. Hence, for small-enough \(a_{2}\) in ([\[PGBmasses\]](#PGBmasses){reference-type="ref" reference="PGBmasses"}) the two problematic states may have non-tachyonic masses even for a large hierarchy between \(\omega_{R}\) and \(\omega_{BL}\), thus avoiding the tree-level SU(5) trap. Let us also note that, up to the obvious differences in the \({\cal O}(1)\) factors, the same dynamical mechanism works in the \(45\oplus 126\) Higgs model # Seesaw scale in the minimal renormalizable \(SO(10)\) GUT However, the vacuum stability was not the only issue that plagued the \(SO(10)\) GUTs for years. The enormous progress in neutrino physics in the last two decades pinned the light neutrino masses into the sub-eV domain with the upper bound (namely, from cosmology and double-beta-decay searches) in the 1 eV ballpark. In the seesaw picture, this typically translates into a lower bound on the scale of the underlying dynamics somewhere in the \(10^{12-13}\) GeV domain. This, however, was long ago claimed to be incompatible with the basic features of the symmetry-breaking pattern in the minimal SO(10) models. ## Seesaw scale in the minimal survival approximation Without any detailed information about the scalar spectrum of a theory under consideration, the best one can do in order to study the relevant gauge coupling unification patterns is to employ the minimum survival hypothesis (MSH) , i.e., to assume that the components of the unified-theory multiplets cluster around the specific symmetry-breaking scales. As rough as this approximation sounds, it often gives a qualitatively good first look at the salient features of the unification pattern. In the non-SUSY \(SO(10)\) framework, the "natural" positions of the seesaw and grand unification scales have been, under this assumption, studied in and later reviewed in. In order to retain a grip on neutrinos and keep the theory well under control, in what follows we shall focus entirely on the \(45\oplus 126\) realization of the Higgs mechanism in the minimal \(SO(10)\) GUT scheme in which the type-I+II seesaw mechanism is supported at the renormalizable level. For more information about this framework an interested reader is referred to the relevant literature . The resulting constraints on the unification and intermediate scales obtained in this scenario (in the minimal-survival approximation) are displayed in FIGURE [\[fig:minimalsurvival\]](#fig:minimalsurvival){reference-type="ref" reference="fig:minimalsurvival"}. Remarkably enough, for both descends of interest there turn out to be stringent upper limits on the seesaw scale in the minimal survival picture well below \(10^{11}\) GeV and, moreover, for the chains passing through the intermediate \(SU(4)_C \otimes SU(2)_L \otimes U(1)_R\) stage, the upper limit for \(M_{G}\) is in the region which tends to be problematic from the proton lifetime perspective. This, however, implies that seesaw scale is far outside the \(10^{12-14}\) GeV domain favoured by the light neutrino masses unless the Dirac neutrino mass terms are artificially suppressed. Although there is nothing a-priori wrong about this option we shall not entertain it here. Rather than that, we shall attempt to do better than the naive estimates above by exploiting the main drawbacks of the minimum survival approach: First, the MSH does not reflect many important features of realistic spectral patterns (such as, e.g., splitting among different components of multiplets below the relevant symmetry-breaking scales). Second, it is totally ignorant of special regions of the parameter space where the scalar spectrum exhibits *unexpected* features such as, e.g., accidentally light states deep in the desert. However, these are exactly the cases when the unification picture can be altered considerably. ## Consistency beyond minimal survival Beyond the minimum-survival approximation, the only guiding principle left for an adventurous parameter-space explorer is the overall consistency of the theory. This has several basic aspects: #### Non-tachyonic scalar spectrum First, all potentially interesting regions of the parameter space should support stable (or at least metastable) vacua. Since the full-featured vacuum stability analysis is very difficult, we shall stick only to the necessary condition, i.e., that there should be no tachyonic states in the scalar spectrum. Let us point out that, for each such vacuum configuration at hand, one can obtain other viable settings by, e.g., rescaling all dimensionful parameters in the scalar potential by a common factor. Similarly, it is clear that fiddling around with the mass of an accidentally light state within a range well below the typical mass-scale of all other heavy states does not destabilize specific vacua either because such variations correspond to only very small shifts in the fundamental parameters of the theory. These two "degrees of freedom" can subsequently be used as an efficient tool for reducing the complexity of the numerical analysis of consistent unification patterns. #### Current proton decay limits Another obvious constraint on the parameter space of the minimal \(SO(10)\) GUT comes from the proton decay; in particular, the current best limit for the \(p\to e^{+}\pi^{0}\) mode from Super-Kamiokande  should be accommodated. In what follows we shall use this together with two assumed future limits that Hyper-Kamiokande (HK)  may reach by 2025 and 2040 (if built): \[\label{protondecay} \tau({\rm SK, 2011})> 8.2 \times 10^{33}\, {\rm years}\,,\quad\tau({\rm HK, 2025}) > 9 \times 10^{34}\, {\rm years}\,,\quad \tau({\rm HK, 2040}) > 2 \times 10^{35}\, {\rm years}\.\] Furthermore, we shall for simplicity neglect all the details related to the flavour structure of the baryon-number-violating currents so that the numbers above translate directly to the bounds on the position of the GUT scale. In the relevant plots (namely, FIGUREs [\[FigureRemnants\]](#FigureRemnants){reference-type="ref" reference="FigureRemnants"} and [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}), the points falling between these limits will be distinguished by a simple colour-code where the light grey is used for proton lifetimes between \(8.2 \times 10^{33}\) and \(9 \times 10^{34}\) years, dark grey corresponds to lifetimes between \(9 \times 10^{34}\) and \(2 \times 10^{35}\) years and black points yield more than \(2 \times 10^{35}\) years. #### Big-bang nucleosynthesis Third, accidentally light coloured states should not be too-long-lived otherwise their late decays may interfere with the highly successful classical big-bang-nucleosynthesis (BBN) account of the light elements' abundances. Actually, as we shall see, this is not a problem here because the accidentally light multiplets in all fully consistent cases originate in \(126_{H}\) and, thus, couple directly to the SM matter fields through the same Yukawa couplings that give rise to, e.g, right-handed neutrino masses. Thus, all the light remnants should decay well before the BBN epoch. #### Consistent unification patterns The simple constraints above are enough to filter out all but two qualitatively different settings with a single accidentally light scalar multiplet well below the \(B-L\) symmetry-breaking scale: a scenario with a very light colour octet \((8,2,+\tfrac{1}{2})\) and another scheme with an intermediate-mass-scale colour sextet \((6,3,+\tfrac{1}{3})\). The typical shapes of the gauge unification patterns in these two cases are shown in FIGURE [\[running\]](#running){reference-type="ref" reference="running"}. The results of a detailed numerical scan over extended regions supporting these solutions are given in FIGURE [\[FigureRemnants\]](#FigureRemnants){reference-type="ref" reference="FigureRemnants"}. Interestingly, the mass range of the octet solution (on the left panel in FIGURE [\[FigureRemnants\]](#FigureRemnants){reference-type="ref" reference="FigureRemnants"}) can stretch as low as to the TeV domain so, in principle, it can even leave its imprints in the LHC searches; however, the sextet is not allowed below roughly \(10^{9}\) GeV. ## Seesaw scale upper limits in consistent scenarios Finally, the allowed ranges for the \(B-L\)-breaking VEV (denoted \(\sigma\)) in these two scenarios are depicted in FIGURE [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}. Remarkably enough, *the naive MSH-based upper bounds on the seesaw scale are in both cases relaxed by as much as four orders of magnitude* as they stretch up to about \(10^{14}\) GeV in the light-octet case and up to almost \(5\times 10^{14}\) GeV in the case of the light sextet. This, however, makes the implementation of the standard seesaw mechanism possible even without resorting to the excessive fine-tuning in the Yukawa sector implied in previous studies. # Conclusions and outlook Even though the minimal SO(10) models have been recently revived as consistent quantum field theories free of inherent tachyonic instabilities, this beautiful and simple framework has never been rehabilitated as a potentially realistic and predictive GUT scheme. This was namely due to the old studies of the relevant gauge unification patterns which revealed a generic tendency for the \(B-L\) symmetry breaking scale to be confined below \(10^{11}\) GeV, apparently too low for a reasonable implementation of the seesaw mechanism for neutrino masses. However, all these early studies based on the minimal survival hypothesis suffer from a generic incapability to account for the effects of accidentally light multiplets with masses detached from any specific symmetry breaking scale. As we argued, a closer inspection of the unification constraints reveals a much wider room for the \(B-L\) breaking VEV stretching up to the \(10^{14}\) GeV ballpark which, in turn, allows the seesaw picture to be implemented without excessive fine-tuning. Besides that, the minimal renormalizable SO(10) scenario has several other interesting features which make it an interesting candidate for a further theoretical scrutiny: #### Possible LHC imprints of the light octet scenario Remarkably enough, the same pair of scalars that we identified as powerful "running helpers" in the minimal SO(10) framework was recently singled out in the work from a totally different perspective. There, the apparent enhancement in the \(H\to \gamma \gamma\) rate indicated by the current LHC data was shown to be attributable to just this couple of states if any of them falls into the vicinity of the TeV scale. However, at the current level of accuracy only the octet \((8,2,+\tfrac{1}{2})\) can be light enough in the minimal \(SO(10)\) to play any role in the Higgs physics because the sextet is not allowed below about \(10^{9}\) GeV, see FIGURE [\[FigureRemnants\]](#FigureRemnants){reference-type="ref" reference="FigureRemnants"}. Nevertheless, the proton-decay limits used in cutting the low-mass-sextet region therein are rather naive and it can happen that a more-sophisticated analysis including flavour effects opens a bigger room for the light sextet too. #### Suppression of the Planck-scale induced unification-smearing effects in the minimal \(SO(10)\) GUTs Besides simplicity, the \(SO(10)\) models in which the GUT-scale symmetry breaking is driven by a VEV of the 45-dimensional adjoint representation have another very interesting feature. This has to do with the general fragility of grand unification with respect to the Planck-scale (\(M_{Pl}\)) effects which, given the proximity of \(M_{G}\) and \(M_{Pl}\), may not be entirely negligible. Concerning their possible impact on, e.g., proton lifetime estimates, the most important of these is namely the Planck-scale induced violation of the canonical normalization of the heavy gauge fields  due to the higher-order corrections to the gauge kinetic form emerging already at the \(d=5\) level: \({\cal L}^{(5)}\ni {\rm Tr} [ F_{\mu\nu}HF^{\mu\nu}]/M_{Pl}\); here \(H\) is any scalar in the theory which can couple to a pair of adjoint representations of a specific GUT symmetry group, i.e., any field appearing in the symmetric part of the decomposition of their tensor product. For a GUT-scale VEV of \(H\), this induces a percent-level effect which, after a suitable redefinition of the gauge fields, leads to similar-size shifts in the GUT-scale matching conditions. Such a "unification smearing effect" can, subsequently, play a significant role in an accurate NLO GUT-scale determination which, in turn, further adds to the existing theoretical uncertainties in the absolute proton lifetime estimates. However, if in the SO(10) the GUT-scale symmetry breaking is triggered by the VEV of \(45_{H}\), this problem is absent because \({\rm Tr} [ F_{\mu\nu}45_{H}F^{\mu\nu}]=0\) due to the fact that 45 is not in the symmetric part of the \(45\otimes 45\) decomposition \[recall that \((45\otimes 45)_{\rm sym}=54\oplus 210\oplus 770\)\]. Thus, the minimal \(SO(10)\) scheme with the adjoint-driven Higgs mechanism is uniquely robust with respect to this class of quantum gravity effects. This makes the symmetry-breaking analysis more reliable and, hence, admits in principle for a strong reduction of this type of theoretical uncertainties in the proton lifetime estimates. #### Proton lifetime at the next-to-leading-order level The simplicity of the minimal \(SO(10)\) scenario advocated in this study, together with its rather unique robustness with respect to the Planck-scale-induced unification smearing effects make this class of models particularly suitable for a possible next-to-leading-order (NLO) proton lifetime analysis. That, however, is far from trivial. To this end, let us just note that the main source of the large theoretical uncertainties in the existing proton lifetime estimates, cf. FIGURE [\[estimates\]](#estimates){reference-type="ref" reference="estimates"}, is the inaccuracy of the GUT-scale determination, partly due to the uncertainties in the low-energy inputs (especially in \(\alpha_{s}(M_{Z})\)) and, in particular, the limited precision of the one-loop approach--given the logarithmic nature of the renormalization-group evolution, both these errors are exponentially amplified in the resulting proton decay amplitude. The only way to keep such uncertainties under control is thus a careful two-loop renormalization-group calculation including, as a prerequisite, the one-loop spectrum of the theory resulting from a dedicated analysis (like, e.g., that in ) together with the proper one-loop matching conditions. In this respect, the minimal renormalizable \(SO(10)\) GUT of our main concern here can be *the* scenario in which a decisive NLO proton lifetime analysis can be just at the verge of tractability. #### GUT-scale baryogenesis The option of a GUT-scale baryogenesis in SO(10) models, recently revived in, is another interesting feature of the model under consideration. On the technical side, however, the amount of thus generated baryon asymmetry depends on the size of the quartic coupling \(\eta_{2}\) in the scalar potential, cf. formula (3) in reference , which, unfortunately, turns out to be one of the most elusive of all the theory parameters--as it was argued in, \(\eta_{2}\) does not enter the tree-level scalar spectrum and, so, the one-loop unification pattern does not impose any constraints on it. However, this coupling does take part in the decomposition of the light Higgs boson in terms of its defining components in \(10_{H}\oplus 126_{H}\) and, thus, it may be constrained indirectly by the flavour structure of the effective theory. Hence, a dedicated proton lifetime analysis advocated above (which, inevitably, must include a detailed account of the flavour structure of the model) may, as one of its by-products, provide also a better grip on the GUT-scale baryogenesis in the minimal \(SO(10)\) GUT.

{'timestamp': '2012-10-16T02:03:14', 'yymm': '1210', 'arxiv_id': '1210.3789', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3789'}

# Introduction Let \(P = \{p_1, p_2,..., p_n\}\) be a set of \(n\) distinct points on a line \(L\). In this paper, we address the problem of determining a unique placement (up to translation and reflection) of the \(p_i\)'s on \(L\), by querying distances between some pairs of points \(p_i\) and \(p_j\), \(1 \leq i, j \leq n\). The resulting queries can be represented by a *point placement graph* (\(ppg\), for short), \(G = (P, E)\), where each edge \(e\) in \(E\) joins a pair of points \(p_i\) and \(p_j\) in \(P\) if the distance between these two points on \(L\) is known and the length of \(e\), \(|e|\), is the distance between the corresponding pair of points. (Note the dual use of \(p_i\) to denote a point on \(L\) as well as a vertex of \(G\).) We will say that \(G\) is *line rigid* or just *rigid* when there is a unique placement for \(P\). Thus, the original problem reduces to the construction of a line rigid \(ppg\), \(G\). Early research on this problem was reported in . In this paper, our first principal reference is, where it was shown that jewel and \(K_{2,3}\) are both line rigid, as also how to build large rigid graphs of density 8/5 (this is an asymptotic measure of the number of edges per point as the number of points go to infinity) out of the jewel. In a subsequent paper, Damaschke proposed a randomized 2-round strategy that needs \((1+o(1))n\) distance queries with high probability and also showed that this is not possible with 2-round deterministic strategies. He also reported the following result: Our second principal reference is the work of  who improved many of the results of. Their principal contributions are the 3-round construction of rigid graphs of density 5/4 from 6-cycles and a lower bound on the number of queries necessary in any 2-round algorithm. They also introduced the idea of a layer graph which is useful in finding the conditions for rigidity of a *ppg* and proved the following result about it: In   we proposed a 2-round algorithm that query \(4n/3 + O(1)\) edges to construct line rigid \(ppg\) on \(n\) points using 6:6 jewels as the basic components. In this paper, we propose a 2-round algorithm that queries \(9n/7 + O(1)\) edges to construct a line rigid \(ppg\) on \(n\) points, using 3 paths of degree two nodes of length 2 each with a common vertex as the basic component, bettering a result of  that uses 5-cycles. More significantly, we improve their lower bound on any 2-round algorithm from \(17n/16\) to \(9n/8\). # A Two Round Algorithm We shall use 3 paths \(p_1q_1r_1s\), \(p_2q_2r_2s\) and \(p_3q_3r_3s\) of degree 2 nodes of length 2 attached to a node \(s\) of degree 3 as the basic component for the point placement (Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}). Other ends \(p_1\), \(p_2\) and \(p_3\) of the 3 paths are made line rigid in the first round. We shall make the remaining 7 points line rigid in the second round. We find a set of sufficient conditions that make the component line rigid by preventing its drawing as a *layer graph* (Theorem [\[lr=lgtheorem:label\]](#lr=lgtheorem:label){reference-type="ref" reference="lr=lgtheorem:label"}). We shall call the component as 3-path and the corresponding *ppg* as 3-path *ppg*. To find the rigidity conditions we consider \((p_1, q_1, r_1, s, r_2, q_2, p_2)\) as a 7-cycle. We shall find conditions that will make the 7-cycle line rigid. Then \(s\) will be unambiguous. Also \(p_1\), \(p_2\) and \(p_3\) are fixed in the first round. Consequently, the distance between \(p_3\) and \(s\) will be fixed. So, we can consider \((p_3, q_3, r_3, s)\) as a 4-cycle. We shall find the condition for rigidity of this 4-cycle. Then the union of these two sets of conditions will comprise the set of rigidity conditions for the whole 3-path *ppg*. We shall attach all the basic components to triplets of points among some constant number of line rigid (in the first round) points \(p_i\). Then for each component there will be extra 7 points and 9 edges. Thus, the density will be \(O(9/7)\). We shall not query the lengths of the edges \(q_1r_1\), \(q_2r_2\) and \(q_3r_3\) in the first round. We shall query them in the second round. So, we shall find a set of sufficient conditions for rigidity for the basic component that does not involve these edges. Then we can satisfy all the rigidity conditions irrespective of the lengths of these edges which will be reported in the second round. The layer graphs of the 7-cycle \((p_1, q_1, r_1, s, r_2, q_2, p_2)\) can be grouped into 6 groups based on the number of edges on each side (Fig. [\[layergraph7cycle:label\]](#layergraph7cycle:label){reference-type="ref" reference="layergraph7cycle:label"}). When different configurations of the chain \(p_3q_3r_3s\) are attached to them, the total number of layer graphs for the 3-path component becomes 42. From them, by Theorem [\[lr=lgtheorem:label\]](#lr=lgtheorem:label){reference-type="ref" reference="lr=lgtheorem:label"}, we get the following 42 conditions for rigidity of the 3-path component: 1. \(|p_1p_2| \ne |q_2r_2|\), \(|p_1p_2| \ne |q_1r_1|\), \(|p_2q_2| \ne |r_2s|\), \(|p_1q_1| \ne |r_1s|\), \(|q_2r_2| \ne |r_1s|\), \(|q_1r_1| \ne |r_2s|\), \(|p_1q_1| \ne |p_2q_2|\). 2. \(||p_1p_2| \pm |p_2q_2|| \ne |r_2s|\), \(||p_2q_2| \pm |q_2r_2|| \ne |r_1s|\), \(||p_1p_2| \pm |p_1q_1| \pm |p_2q_2|| \ne |r_1s|\), \(|p_1q_1| \ne ||r_1s| \pm |r_2s||\), \(|p_1p_2| \ne ||q_1r_1| \pm |r_1s||\), \(||p_1q_1| \pm |q_1r_1|| \ne |p_2q_2|\), \(||p_1q_1| \pm |p_1p_2|| \ne |q_2r_2|\). 3. \(||p_1p_2| \pm |p_1q_1|| \ne |r_1s|\), \(||p_1q_1| \pm |q_1r_1|| \ne |r_2s|\), \(||p_1p_2| \pm |p_1q_1| \pm |p_2q_2|| \ne |r_2s|\), \(|p_2q_2| \ne ||r_1s| \pm |r_2s||\), \(|p_1p_2| \ne ||q_2r_2| \pm |r_2s||\), \(||p_2q_2| \pm |q_2r_2|| \ne |p_2q_2|\), \(||p_2q_2| \pm |p_1p_2|| \ne |q_1r_1|\). 4. \(|p_1p_2| \ne |r_2s|\), \(|p_1p_2| \ne |r_1s|\), \(|p_2q_2| \ne |r_1s|\), \(|p_1q_1| \ne |r_2s|\), \(||p_1q_1| \pm |p_2q_2| \pm |p_1p_2|| \ne ||r_1s| \pm |r_2s||\), \(|p_2q_2| \ne |q_1r_1|\), \(|p_1q_1| \ne |q_2r_2|\). 5. \(|p_2q_2| \ne ||p_1p_2| \pm |r_1s||\), \(|p_1q_1| \ne ||p_1p_2| \pm |r_2s||\), \(|p_1q_1| \ne ||p_1p_2| \pm |r_1s| \pm |r_2s||\), \(|p_2q_2| \ne ||p_1p_2| \pm |r_1s| \pm |r_2s||\), \(|p_1q_1| \ne ||q_2r_2| \pm |r_2s||\), \(|p_2q_2| \ne ||q_1r_1| \pm |r_1s||\), \(|p_1p_2| \ne ||r_1s| \pm |r_2s||\). 6. \(||p_1q_1| \pm |q_1r_1|| \ne ||p_2q_2| \pm |r_2s||\), \(||p_2q_2| \pm |q_2r_2|| \ne ||p_1q_1| \pm |r_1s||\), \(|q_1r_1| \ne ||p_2q_2| \pm |r_2s||\), \(|q_2r_2| \ne ||p_1q_1| \pm |r_1s||\), \(|p_2q_2| \ne ||p_1q_1| \pm |r_1s||\), \(|p_1q_1| \ne ||p_2q_2| \pm |r_2s||\), \(|p_2q_2| \ne ||p_1q_1| \pm |r_1s| \pm |r_2s||\). Among them 20 conditions involve the edges \(q_1r_1\) and \(q_2r_2\) of the 7-cycle that we want to avoid in the conditions. We shall replace each of these conditions by a set of conditions that prevents the 7-cycle from being drawn as the layer graph representation that corresponds to that condition. Collection of all these new conditions and the ones that are not replaced will constitute the rigidity conditions for the 7-cycle. As stated before if the 7-cycle is line rigid then the \((p_3, q_3, r_3, s)\) will be a 4-cycle which can be made line rigid by imposing the condition \(|p_3q_3| \ne |r_3s|\) . This condition together with the rigidity conditions for the 7-cycle will constitute the rigidity conditions for the whole component. As an example of replacing conditions we shall replace the first condition, viz., \(|p_1p_2| \ne |q_2r_2|\), which corresponds to the layer graph representation of the 7-cycle in Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}. To replace the condition we find a set of conditions that prevent the drawing of layer graph of the 7-cycle \((p_1, q_1, r_1, s, r_2, q_2, p_2)\) in the configuration of Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}. For this we draw all the possible configurations of the layer graph of the whole component with the layer graph of the 7-cycle being in the configuration of Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}. This new set of conditions acts as a replacement for the condition \(|p_1p_2| \ne |q_2r_2|\) since that set will prevent the drawing of the layer graph of the 7-cycle \((p_1, q_1, r_1, s, r_2, q_2, p_2)\) in the corresponding configuration in Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}. Since \(p_1\), \(p_2\) and \(p_3\) are made line rigid in the first round they must lie on a line and their positions must be unique (upto translation and reflection) after first round. Since in the present configuration of the 7-cycle (Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}) \(p_1\) and \(s\) are on the same side of the layer graph the edges \(p_3q_3\), \(q_3r_3\) and \(r_3s\) can have 4 distinct configurations giving rise to 4 distinct layer graph representations (Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}) of the whole component with the layer graph of the 7-cycle being in the configuration of Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"}. Thus, in order to be able to draw the layer graph of the 7-cycle \((p_1, q_1, r_1, s, r_2, q_2, p_2)\) in the configuration of Fig. [\[configa:label\]](#configa:label){reference-type="ref" reference="configa:label"} the layer graph of the whole component must have one of the four distinct configurations as shown in Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}. First, we consider the configuration where \(p_3q_3\) and \(r_3s\) are horizontal, and \(q_3r_3\) is vertical (Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}a). The condition \(|p_1p_2| \ne |q_2r_2|\) prevents the 7-cycle from being drawn as a layer graph of present configuration. However, it involves the edge \(q_2r_2\) which we need to avoid. In the present configuration of the layer graph of the component \(p_1, q_1, r_1, s\) and \(r_2\) are on a line which is parallel to \(p_2q_2\) and \(q_3r_3\). So, we must have \(|q_2r_2| = ||p_2p_3| \pm |p_3q_3| \pm |r_3s||\). Using this the condition becomes \(|p_1p_2| \ne ||p_2p_3| \pm |p_3q_3| \pm |r_3s||\). Since \(||p_1p_2| \pm |p_2p_3|| = |p_1p_3|\) the condition reduces to \(|p_1p_3| \ne ||p_3q_3| \pm |r_3s||\). If we ensure this condition then we must have \(|p_1p_2| \ne |q_2r_2|\) in the present configuration of the component. Thus, the component in general and the 7-cycle in particular cannot be drawn as a layer graph in the present configurations of the 7-cycle and the component. Now we consider the case when \(p_3q_3\) and \(q_3r_3\) are vertical, and \(r_3s\) is horizontal (Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}b). In the present configuration of the layer graph of the component \(p_1, q_1, r_1, s\) and \(r_2\) are on a line, and \(p_3q_3\) and \(q_3r_3\) are on a line. Those lines are parallel and they are parallel to \(p_2q_2\). So, we must have \(|q_2r_2| = ||p_2p_3| \pm |r_3s||\). Using this the condition becomes \(|p_1p_2| \ne ||p_2p_3| \pm |r_3s||\). We have \(||p_1p_2| \pm |p_2p_3|| = |p_1p_3|\). Using this the rigidity condition \(|p_1p_2| \ne |q_2r_2|\) becomes \(|p_1p_3| \ne |r_3s|\). Next, we consider the case when \(p_3q_3\) is vertical, and \(q_3r_3\) and \(r_3s\) are horizontal (Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}c). The condition \(||p_1q_1| \pm |q_1r_1| \pm |r_1s| \pm |sr_2|| \ne |p_2q_2|\) prevents the 7-cycle from being drawn as a layer graph of present configuration. However, it involves the edge \(q_1r_1\) which we need to avoid. In the present configuration of the layer graph of the component \(p_1\), \(p_2\) and \(p_3\) are on a line, and \(q_3\), \(r_3\) and \(s\) are on a line. The lines are parallel. So, we must have \(||p_1q_1| \pm |q_1r_1| \pm |r_1s|| = |p_3q_3|\). Using this the condition becomes \(||p_3q_3| \pm |sr_2|| \ne |p_2q_2|\). Finally, we consider the case when \(p_3q_3\) is vertical, \(q_3r_3\) is horizontal and \(r_3s\) is vertical (Fig. [\[configaall4fig:label\]](#configaall4fig:label){reference-type="ref" reference="configaall4fig:label"}d). In the present configuration of the layer graph of the component \(p_1\), \(p_2\) and \(p_3\) are on a line. The line is parallel to \(q_3r_3\). So, we must have \(||p_1q_1| \pm |q_1r_1| \pm |r_1s| \pm |sr_3|| = |p_3q_3|\). Using this the rigidity condition \(||p_1q_1| \pm |q_1r_1| \pm |r_1s| \pm |sr_2|| \ne |p_2q_2|\) becomes \(||p_3q_3| \pm |sr_3| \pm |sr_2|| \ne |p_2q_2|\). The following lemma justifies the replacement (the proof is omitted). As mentioned before, we make triplet of points \((p_1, p_2, p_3)\) of each 3-path component line rigid in the first round. Let \(S\) be the set of points for such triplets. We make the points in \(S\) line rigid in the first round. We make the remaining 7 points of each 3-path component line rigid in the second round. To select triplet of points in \(S\) as \((p_1, p_2, p_3)\) of a component, let us select any point of \(S\) as \(p_1\). Then let us find another point of \(S\), we denote it as \(p_2\), satisfying the conditions on the length \(|p_1p_2|\) mentioned in serial number 1 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}. By Observation 1, at most 8 edges will not satisfy the conditions on \(|p_1p_2|\). We need at least 8 extra points, i.e., we need to have a total of at least 9 more points, other than \(p_1\), in \(S\) as candidate for \(p_2\). After \(p_2\) is selected, let us find another point of \(S\), we denote it as \(p_3\), from the remaining pints of \(S\) such that the conditions on \(|p_2p_3|\) in serial numbers 2 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} are satisfied. By Observation 1, at most 8 edges will not satisfy the conditions on \(|p_2p_3|\). This warrants the set \(S\) to have at least 8 extra points other than \(p_1\), \(p_2\) and \(p_3\). The point \(p_3\) selected this way by satisfying the conditions on \(p_2p_3\) must also have to satisfy the conditions on \(p_3p_1\) mentioned in serial number 3 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}. By Observation 1, at most 8 edges will not satisfy the conditions on \(|p_3p_1|\). This warrants the set \(S\) to have at least 8 more extra points, i.e., 16 extra points, other than \(p_1\), \(p_2\) and \(p_3\). But if \(S\) has only 19 points for the selection of \(p_i\)s it may happen that all the basic components are attached to the same triplets. This hinders our goal of obtaining a better value for \(\alpha\) than previously known. We need to attach the basic components evenly to all the points of \(S\) so that the same number of edges can be attached to each of them in the first round and all of those edges, except for a constant number, are used to attach the basic components. In other words, we need to attach the 3-path components to the points in \(S\) in such a way that the numbers of components attached to any two points differ by at most a constant number. Now we describe our algorithm to select triplets of points in \(S\) to attach components. To attach a basic component we always select a point in \(S\) with the lowest valence as the first point (say \(p_1\)). Of the remaining points of \(S\), at most 8 points may not be acceptable for the second point (say \(p_2\)), because of the conditions on \(p_1p_2\). From among the rest \(|S|-1\) points that satisfy the conditions on \(p_1p_2\) we select the one that has the lowest valence, as \(p_2\). Of the rest \(|S|-2\) points of \(S\), at most 16 may not be acceptable for the last point, say \(p_3\), because of the conditions on \(p_2p_3\) and \(p_3p_1\). From among the rest points that satisfy the conditions on \(p_2p_3\) and \(p_3p_1\) we choose the one that has the lowest valence, as \(p_3\). This method will be follwed to attach each basic component to the points in \(S\). While, we shall attach the basic components sequentially. To specify the number of basic components attached to a point in \(S\) we shall use the term valence. We denote the set of points with valence \(d\) as \(S_d\). The following lemma tells us how big \(S\) must be (the proof is omitted): We make the above set \(S\) of 35 points line rigid in the first round by using jewel of Damaschke  as the *ppg*. We create 6 jewels hanging from a common strut that is incident on 2 points of \(S\). This will make 32 points line rigid. For this we need to query the lengths of 49 edge. We make the remaining 3 points line rigid by using triangle as the *ppg*. For each of these 3 points we query its distance from each of the pair of points that are incident on the strut. There will be 6 more queries for edge lengths. Thus, we shall query a total of 55 edges in the first round to make the 35 points of \(S\) line rigid in that round. The conditions on \(p_1q_1\), \(p_2q_2\) and \(p_3q_3\) in serial numbers respectively 3, 4 and 5 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} will not be satisfied by at most 40, 90 and 122 edges respectively (by Observation 1). In addition to the 122 extra edges needed at each of \(p_i\)'s to satisfy the conditions on \(|p_1q_1|\), \(|p_2q_2|\) and \(|p_3q_3|\) we need 2 more extra edges incident on each of \(p_i\) to accommodate the difference of 2 between the number of basic components that can be attached to the \(p_i\)'s. Thus, we need a total of 124 extra edges incident on each of the points \(p_i, i = 1,..., 35\) of \(S\). We shall attach \(3b\), \(3b+1\) or \(3b+2\) (where \(b\) is a positive integer) number of 3-path components to each point in \(S\). This requires us to have \(3b+124\) edges incident on each of \(p_i\)'s in \(S\). In the worst case there will be at most 18 points in \(S\) with valence \(3b\), no points in \(S\) with valence \(b+1\) and the remaining points with valence \(3b+2\). Thus, we shall be able to construct a total of at least \(3b+11\) number of 3-path components from the edges provided for \(p_iq_i\) at all the \(p_i\)'s in \(S\). Now we describe the algorithm to construct a composite \(ppg\) made up of 3-path components such that all the rigidity conditions listed in Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} are satisfied for each of them. **Algorithm 1.** Let the total number of points be \(n = 245b + 4,419\), where \(b\) is a positive integer. We attach at least \(3b\) and at most \(3b+2\) numbers of 3-path components (Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}) to each of 35 rigid points in \(S\) subject to the condition that the total number of such components being \(35b + 11\). In the first round, we make distance queries represented by the edges of the graph in Fig. [\[querygraph:label\]](#querygraph:label){reference-type="ref" reference="querygraph:label"}. All the nodes \(p_i\) (\(i=1,..., 35\)) in the subgraph enclosed by the rectangle are elements of \(S\) and are made line rigid in the first round by using the jewel of  as the *ppg*. There are 6 jewels attached to a common strut in the subgraph. Residual 3 points are made line rigid by using triangle as the *ppg*. They are attached to the common strut. There are a total of 55 edges in the subgraph. Each of the vertices \(p_i, p_j\), or \(p_k\) (\(i, j, k = 1,..., 35\)) of \(S\) has \(b+124\) leaves to attach \(3b\), \(3b+1\) or \(3b+2\) 3-path components (Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}). Since there will be \(35b+11\) 3-path components we make \(35b+11\) groups of 4 nodes \((r_{il}, r_{jl}, r_{kl}, s_l)\), \((l = 1,..., 35b+11)\). We query the distances \(|r_{il}s_l|\), \(|r_{jl}s_l|\) and \(|r_{kj}s_l|\), \((l = 1,..., 35b+11)\) in the first round. We will make a total of \(210b+4,428\) pairwise distance queries in the first round for the placement of \(n = 245b + 4,419\) points. In the second round, for each 3-link \((r_{il}, r_{jl}, r_{kl}, s_l), l = 1,..., 35b + 11,\) we construct a 3-path component (Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}), satisfying all its rigidity conditions as in Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}. For each such 3-link we select a point \(p_i\), from the subgraph of 35 points of \(S\) that has the lowest valency of 3-path component of Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}. Since all the 35 points \(p_i, i = 1,..., 35,\) are rigid in the first round, for any pair of such fixed points \((p_i, p_j) (i, j = 1,...35; i \ne j)\) we can find the distance \(|p_ip_j|\). So, for each pair of points \((p_i, p_j) (i, j = 1,..., 35; i \ne j)\), we shall use \((p_i, p_j)\) as an edge in the construction of the 3-path component of Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"}. Now from the subgraph of 35 points of \(S\) we select another point \(p_j (j \ne i)\) such that the length \(|p_ip_j|\) satisfies all the 4 conditions of rigidity on it as stated in serial number 1 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} and that it has the lowest valency of 3-path component of Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"} among all such qualifying points. We note that we can always find such point \(p_j\), because there will be at most 8 edges \((p_ip_j)\) whose lengths do not satisfy the rigidity conditions on it (Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}) whereas we have 34 more points for choosing the point \(p_j\). Similarly, from the subgraph of 35 points of \(S\) we select another point \(p_k (k \ne i, k \ne j)\) such that the length \(|p_jp_k|\) satisfies all the 4 conditions of rigidity on it as stated in serial number 2 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} and the length \(|p_kp_i|\) satisfies all the 4 conditions of rigidity on it as stated in serial number 3 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}, and that it has the lowest valency of 3-path component of Fig. [\[3pathfig:label\]](#3pathfig:label){reference-type="ref" reference="3pathfig:label"} among all such qualifying points. We note that we can always find such point \(p_k\), because there will be at most 16 nodes \(p_k\) such that the lengths of the edges \(p_jp_k\) and \(p_kp_i\) do not satisfy the rigidity conditions on them (Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}) whereas we have 33 more points for choosing the point \(p_k\). Then we find an edge \(p_iq_{il}\) rooted at \(p_i\) satisfying the 20 conditions of rigidity on it as stated in serial no. 4 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}, then we find another edge \(p_jq_{jl}\) rooted at \(p_j\) satisfying the 45 conditions on it as stated in serial no. 5 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"} and finally, we find another edge \(p_kq_{kl}\) rooted at \(p_k\) satisfying the 61 conditions on it as stated in serial no. 6 of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}. Then for each \(l, (l = 1,..., 35b + 11)\), we query the distances \(|q_{il}r_{il}|\), \(|q_{jl}r_{jl}|\) and \(|q_{kl}r_{kl}|\) to form a 3-path component \(p_ip_jp_kq_{il}q_{jl}q_{kl}r_{il}r_{jl}r_{kl}s_l\). Its edges will satisfy all the rigidity conditions of Lemma [\[3pathlemma:label\]](#3pathlemma:label){reference-type="ref" reference="3pathlemma:label"}. Thus, all the \(35b + 11\) 3-links will be consumed to construct \(35b + 11\) 3-path components. For this \(105b + 33\) edges will be queried in the second round. There will be unused leaves \(q_{il}\)/\(q_{jl}\)/\(q_{kl}\)) numbering 4,307 in total for the 35 points of \(S\). We use a 4-cycle *ppg*  to fix 4,306 of them and a triangle *ppg* to fix the rest 1 point in the second round. As before, for each pair of points \((p_i, p_j) (i, j = 1,..., 35; i \ne j)\), we shall use \((p_i, p_j)\) as an edge in the construction of the 4-cycle. For each unused point \(q_{il}\) rooted at \(p_i\) we find another point \(q_{jl}\) rooted at \(p_j\) such that \(|p_ip_{il}| \ne |p_jp_{jl}|\). Then the 4-cycle \(p_iq_{il}q_{jl}p_j\) will be line rigid (Observation 2). Then we query the distance \(|q_{il}q_{jl}|\) in the second round to complete the 4-cycle. Note that we can always find a point like \(q_{jl}\). For, after repeated selection of such matching pairs of edges there may remain at most 2 edges \(p_iq_{il}\) rooted at \(p_i\) of length equal to that of the same number of edges rooted at \(p_j\) (Observation 1). In such a situation we switch the matching to match such edges rooted at \(p_i\) with edges other than those same length edge/s rooted at \(p_j\)-this is always possible because there are at most 2 edges rooted at \(p_j\) that have the same length (Observation 1). To make the remaining 1 leave node line rigid we query in the second round its distance from any point of \(S\) other than its parent node. For 4,307 unused points (after the construction of the 3-path components) 2,153 4-cycles and 1 triangle will be constructed. 2,153 edges will be queried to complete the 4-cycles and 1 edge will be queried to construct the triangle. The total number of queries in the second round will be \((105b+33)+2,153+1\), i.e., \(105b+2,187\). 0◻ The number of queries in the first and second rounds are \(210b + 4,428\) and \(105b + 2,187\) respectively. Thus, in 2 rounds a total of \(315b + 6,615\) pairwise distances are to be queried for the placement of \(245b + 4,419\) points. Now, \(315b + 6,615 = (315/245)*(245b + 4419)-(9/7)*4419 + 6615 = 9n/7 + (46305-39771)/7 = 9n/7 + 6534/7\). Thus, we have the following theorem: # Lower Bound for Two Rounds The argument here closely follows the adversarial argument given in the lower bound proof of . Let the set of edges queried in the first and second round be \(E_1\) and \(E_2\) respectively; \(G_1 = (V,E_1)\) is the query graph for the first round, while \(G_2 = (V,E_1\cup E_2)\) is the final query graph after the second round. The length of a maximal path of degree 2 nodes in a graph is the number of degree 2 nodes in the path. We shall call nodes of degree at least 3 as heavy nodes. In the *first round*, the adversary returns edge-lengths according to the following strategy, with the intention of keeping the linear layout of the *ppg* ambiguous: For a maximal path of degree 2 nodes in \(G_2\), as a consequence of \(S_3\) there are limits on the maximum number of edges from \(E_1\) if the path consists of edges from \(E_1\) only (Fig. [\[deg2maxpath-bothEndHeavy:label\]](#deg2maxpath-bothEndHeavy:label){reference-type="ref" reference="deg2maxpath-bothEndHeavy:label"} shows a degree 2 maximal path \(p_1p_2p_3p_4p_5p_6\) in \(G_1\) with both the end nodes \(p_0\) and \(p_7\) being heavy), and on the maximum number of consecutive edges from \(E_1\) if it contains at least one edge from \(E_2\) (Fig. [\[deg2maxpath-notBothEndHeavy:label\]](#deg2maxpath-notBothEndHeavy:label){reference-type="ref" reference="deg2maxpath-notBothEndHeavy:label"} shows some degree 2 maximal paths in \(G_1\) with none of the end nodes being heavy). If both of \(p_0\) and \(p_{k+1}\) are of degree at least three in the first round the adversary sets the above layout in such a way that if, for any \(i\) with \(i = 1\) (mod 3) and \(i < k\), no edge is attached to either \(p_i\) or \(p_{i+1}\) in the second round their positions will be ambiguous. Thus, for this case the length of a maximal path of degree 2 nodes in \(G_2\) containing only the edges in \(E_1\) can be at most 3. If at least one of \(p_0\) and \(p_{k+1}\), say \(p_{k+1}\), is of degree one in the first round the adversary sets the above layout in such a way that if, for any \(i\) with \(i = 1\) (mod 2) and \(i < k\), no edge is attached to either \(p_i\) or \(p_{i+1}\) in the second round, they can be made ambiguous by setting \(|p_{i}p_{i+1}| = |p_{i-1}p_{i+2}|\) in the second round. Thus, for this case the length of a maximal path of degree 2 nodes in \(G_2\) containing only the edges in \(E_1\) can be at most 2. If \(p_{k+1}\) is of degree 1 and no edge is attached to either \(p_{k-1}\) or \(p_k\) in the second round the positions of \(p_{k-1}\) and \(p_k\) can be made ambiguous by setting \(|p_{k-1}p_k| = |p_{k-2}p_{k+1}|\) in that round. The algorithm must attach an edge in \(G_2\) to \(p_{k-1}\) or \(p_k\). Still then there will be at most 2 free nodes at an end of a path of degree 2 nodes if the end node is of degree 1. The algorithm will fix them in the second round. Thus, in a maximal path of degree 2 nodes in \(G_2\) that contains at least one edge from \(E_2\) there can be at most 2 consecutive edges from \(E_1\). The above results together with \(S_2\) and \(S_3\) imply that the following property holds for the *ppg* .

{'timestamp': '2012-10-16T02:04:00', 'yymm': '1210', 'arxiv_id': '1210.3833', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3833'}

# Introduction These notes describe how to average and fit numerical data that you have obtained, presumably by some simulation. Typically you will generate a set of values \(x_i,\, y_i, \cdots,\, i = 1, \cdots N\), where \(N\) is the number of measurements. The first thing you will want to do is to estimate various average values, and determine *error bars* on those estimates. As we shall see, this is straightforward if one wants to compute a single average, e.g. \(\langle x \rangle\), but not quite so easy for more complicated averages such as fluctuations in a quantity, \(\langle x^2 \rangle-\langle x \rangle^2\), or combinations of measured values such as \(\langle y \rangle / \langle x \rangle^2\). Averaging of data will be discussed in Sec. [2](#sec:averages){reference-type="ref" reference="sec:averages"}. Having obtained several good data points with error bars, you might want to fit this data to some model. Techniques for fitting data will be described in the second part of these notes in Sec. [3](#sec:fit){reference-type="ref" reference="sec:fit"} I find that the books on these topics usually fall into one of two camps. At one extreme, the books for physicists don't discuss all that is needed and rarely *prove* the results that they quote. At the other extreme, the books for mathematicians presumably prove everything but are written in a style of lemmas, proofs, \(\epsilon\)'s and \(\delta\)'s, and unfamiliar notation, which is intimidating to physicists. One exception, which finds a good middle ground, is Numerical Recipes  and the discussion of fitting given here is certainly influenced by Chap. 15 of that book. In these notes I aim to be fairly complete and also to derive the results I use, while the style is that of a physicist writing for physicists. I also include scripts in python, perl, and gnuplot to perform certain tasks in data analysis and fitting. For these reasons, these notes are perhaps rather lengthy. Nonetheless, I hope, that they will provide a useful reference. # Averages and error bars {#sec:averages} ## Basic Analysis {#sec:basic} Suppose we have a set of data from a simulation, \(x_i, \, (i = 1, \cdots, N)\), which we shall refer to as a *sample* of data. This data will have some random noise so the \(x_i\) are not all equal. Rather they are governed by a distribution \(P(x)\), *which we don't know*. The distribution is normalized, \[\int_{-\infty}^\infty P(x) \, d x = 1,\] and is usefully characterized by its moments, where the \(n\)-th moment is defined by \[\langle x^n \rangle = \int_{-\infty}^\infty x^n\, P(x) \, d x\,.\] We will denote the average *over the exact distribution* by angular brackets. Of particular interest are the first and second moments from which one forms the mean \(\mu\) and variance \(\sigma^2\), by The term "standard deviation" is used for \(\sigma\), the square root of the variance. In this section we will estimate the mean \(\langle x \rangle\), and the uncertainty in our estimate, from the \(N\) data points \(x_i\). The determination of more complicated averages and resulting error bars will be discussed in Sec. [2.2](#sec:advanced){reference-type="ref" reference="sec:advanced"} In order to obtain error bars we need to assume that the data are uncorrelated with each other. This is a crucial assumption, without which it is very difficult to proceed. However, it is not always clear if the data points are truly independent of each other; some correlations may be present but not immediately obvious. Here, we take the usual approach of assuming that even if there are some correlations, they are sufficiently weak so as not to significantly perturb the results of the analysis. In Monte Carlo simulations, measurements which differ by a sufficiently large number of Monte Carlo sweeps will be uncorrelated. More precisely the difference in sweep numbers should be greater than a "relaxation time". This is exploited in the "binning" method in which the data used in the analysis is not the individual measurements, but rather an average over measurements during a range of Monte Carlo sweeps, called a "bin". If the bin size is greater than the relaxation time, results from adjacent bins will be (almost) uncorrelated. A pedagogical treatment of binning has been given by Ambegaokar and Troyer . Alternatively, one can do independent Monte Carlo runs, requilibrating each time, and use, as individual data in the analysis, the average from each run. The information *from the data* is usefully encoded in two parameters, the sample mean \(\overline{x}\) and the sample standard deviation \(s\) which are defined by[^1] In statistics, notation is often confusing but crucial to understand. Here, an average indicated by an over-bar, \(\overline{\cdots}\), is an average over the *sample of \(N\) data points*. This is to be distinguished from an exact average over the distribution \(\langle \cdots \rangle\), as in Eqs. ([\[xavexact\]](#xavexact){reference-type="ref" reference="xavexact"}) and ([\[sigma\]](#sigma){reference-type="ref" reference="sigma"}). The latter is, however, just a theoretical construct since we *don't know* the distribution \(P(x)\), only the set of \(N\) data points \(x_i\) which have been sampled from it. Next we derive two simple results which will be useful later: 1. The mean of the sum of \(N\) independent variables *with the same distribution* is \(N\) times the mean of a single variable, and 2. The variance of the sum of \(N\) independent variables *with the same distribution* is \(N\) times the variance of a single variable. The result for the mean is obvious since, defining \(X = \sum_{i=1}^N x_i\), \[\langle X \rangle = \sum_{i=1}^N \langle x_i \rangle = N \langle x_i \rangle \ \boxed{ = N \mu\,.} \label{X}\] The result for the standard deviation needs a little more work: To get from Eq. ([\[1\]](#1){reference-type="ref" reference="1"}) to Eq. ([\[2\]](#2){reference-type="ref" reference="2"}) we note that, for \(i \ne j\), \(\langle x_i x_j\rangle = \langle x_i \rangle \langle x_j\rangle\) since \(x_i\) and \(x_j\) are assumed to be statistically independent. (This is where the statistical independence of the data is needed.) If the means and standard deviations are not all the same, then the above results generalize to Now we describe an important thought experiment. Let's *suppose* that we could repeat the set of \(N\) measurements *very many* many times, each time obtaining a value of the sample average \(\overline{x}\). From these results we could construct a distribution, \(\widetilde{P}(\overline{x})\), for the sample average as shown in Fig. [\[Fig:distofmean\]](#Fig:distofmean){reference-type="ref" reference="Fig:distofmean"}. If we do enough repetitions we are effectively averaging over the exact distribution. Hence the average of the sample mean, \(\overline{x}\), over very many repetitions of the data, is given by \[\langle \overline{x} \rangle = {1 \over N} \sum_{i=1}^N \langle x_i \rangle = \langle x \rangle \equiv \mu \,, \label{xav}\] i.e. it is the exact average over the distribution of \(x\), as one would intuitively expect, see Fig. [\[Fig:distofmean\]](#Fig:distofmean){reference-type="ref" reference="Fig:distofmean"}. Eq. [\[xav\]](#xav){reference-type="eqref" reference="xav"} also follows from Eq. [\[X\]](#X){reference-type="eqref" reference="X"} by noting that \(\overline{x} = X/N\). In fact, though, we have only the *one* set of data, so we can not determine \(\mu\) exactly. However, Eq. ([\[xav\]](#xav){reference-type="ref" reference="xav"}) shows that \[\boxed{ \mbox{the best estimate of\ } \mu \mbox{ is } \overline{x},} \label{xbarest}\] i.e. the sample mean, since averaging the sample mean over many repetitions of the \(N\) data points gives the true mean of the distribution, \(\mu\). An estimate like this, which gives the exact result if averaged over many repetitions of the experiment, is said to be We would also like an estimate of the uncertainty, or "error bar", in our estimate of \(\overline{x}\) for the exact average \(\mu\). We take \(\sigma_{\overline{x}}\), the standard deviation in \(\overline{x}\) (obtained if one did many repetitions of the \(N\) measurements), to be the uncertainty, or error bar, in \(\overline{x}\). The reason is that \(\sigma_{\overline{x}}\) is the width of the distribution \(\widetilde{P}(\overline{x})\), shown in Fig. [\[Fig:distofmean\]](#Fig:distofmean){reference-type="ref" reference="Fig:distofmean"}, so a *single* estimate \(\overline{x}\) typically differs from the exact result \(\mu\) by an amount of this order. The variance \(\sigma_{\overline{x}}^2\) is given by \[\sigma_{\overline{x}}^2 \equiv \langle \overline{x}^2 \rangle-\langle \overline{x} \rangle^2 = {\sigma^2 \over N}\,, \label{dxsq}\] which follows from Eq. [\[dXsq\]](#dXsq){reference-type="eqref" reference="dXsq"} since \(\overline{x} =X / N\). The problem with Eq. ([\[dxsq\]](#dxsq){reference-type="ref" reference="dxsq"}) is that **we don't know \(\sigma^2\)** since it is a function of the exact distribution \(P(x)\). We do, however, know the *sample* variance \(s^2\), see Eq. ([\[sigmafromdata\]](#sigmafromdata){reference-type="ref" reference="sigmafromdata"}), and the average of this over many repetitions of the \(N\) data points, is equal to \(\sigma^2\) since To get from Eq. ([\[3\]](#3){reference-type="ref" reference="3"}) to Eq. ([\[4\]](#4){reference-type="ref" reference="4"}), we have separated the terms with \(i=j\) in the last term of Eq. ([\[3\]](#3){reference-type="ref" reference="3"}) from those with \(i \ne j\), and used the fact that each of the \(x_i\) is chosen from the same distribution and is statistically independent of the others. It follows from Eq. ([\[5\]](#5){reference-type="ref" reference="5"}) that \[\boxed{ \mbox{the best estimate of\ } \sigma^2 \mbox{ is } s^2 \, ,} \label{sigmasamp}\] since averaging \(s^2\) over many repetitions of \(N\) data points gives \(\sigma^2\). The estimate for \(\sigma^2\) in Eq. ([\[sigmasamp\]](#sigmasamp){reference-type="ref" reference="sigmasamp"}) is therefore unbiased. Combining Eqs. ([\[dxsq\]](#dxsq){reference-type="ref" reference="dxsq"}) and ([\[sigmasamp\]](#sigmasamp){reference-type="ref" reference="sigmasamp"}) gives \[\boxed{ \mbox{the best estimate of\ } \sigma_{\overline{x}}^2 \mbox{ is } {s^2 \over N}\; \, ,} \label{errorbar}\] We have now obtained, using only information from the data, that the mean is given by \[\boxed{ \mu = \overline{x}\; \pm \; \sigma_{\overline{x}} \, ,}\] where \[\boxed{ \sigma_{\overline{x}} = {s \over \sqrt{N}},, } \label{finalans}\] which we can write explicitly in terms of the data points as \[\boxed{ \sigma_{\overline{x}} = \left[ {1 \over N(N-1)} \, \sum_{i=1}^N (x_i-\overline{x})^2 \right]^{1/2} \,.} \label{finalans2}\] Remember that \(\overline{x}\) and \(s\) are the mean and standard deviation of the (one set) of data that is available to us, see Eqs. ([\[meanfromdata\]](#meanfromdata){reference-type="ref" reference="meanfromdata"}) and ([\[sigmafromdata\]](#sigmafromdata){reference-type="ref" reference="sigmafromdata"}). As an example, suppose \(N=5\) and the data points are \[x_i = 10, 11, 12, 13, 14,\] (not very random looking data it must be admitted!). Then, from Eq. ([\[meanfromdata\]](#meanfromdata){reference-type="ref" reference="meanfromdata"}) we have \(\overline{x} = 12\), and from Eq. ([\[sigmafromdata\]](#sigmafromdata){reference-type="ref" reference="sigmafromdata"}) \[s^2 = {1 \over 4} \, \left[(-2)^2 + (-1)^2 + 0^2 + 1^2 + 2^2\right] = {5 \over 2}.\] Hence, from Eq. ([\[finalans\]](#finalans){reference-type="ref" reference="finalans"}), \[\sigma_{\overline{x}} = {1 \over \sqrt{5}}\, \sqrt{5 \over 2} = {1\over \sqrt{2}}.\] so \[\mu = \overline{x} \pm \sigma_{\overline{x}} = 12 \pm {1\over \sqrt{2}}.\] How does the error bar decrease with the number of statistically independent data points \(N\)? Equation ([\[5\]](#5){reference-type="ref" reference="5"}) states that the expectation value of \(s^2\) is equal to \(\sigma^2\) and hence, from Eq. ([\[finalans\]](#finalans){reference-type="ref" reference="finalans"}), we see that Hence, to reduce the error bar by a factor of 10 one needs 100 times as much data. This is discouraging, but is a fact of life when dealing with random noise. For Eq. ([\[finalans\]](#finalans){reference-type="ref" reference="finalans"}) to be really useful we need to know the probability that the true answer \(\mu\) lies more than \(\sigma_{\overline{x}}\) away from our estimate \(\overline{x}\). Fortunately, for large \(N\), the central limit theorem, derived in Appendix [\[sec:clt\]](#sec:clt){reference-type="ref" reference="sec:clt"}, tells us (for distributions where the first two moments are finite) that the distribution of \(\overline{x}\) is a Gaussian. For this distribution we know that the probability of finding a result more than one standard deviation away from the mean is 32%, more than two standard deviations is \(4.5\%\) and more than three standard deviations is \(0.3\%\). Hence we expect that most of the time \(\overline{x}\) will be within \(\sigma_{\overline{x}}\) of the correct result \(\mu\), and only occasionally will be more than two times \(\sigma_{\overline{x}}\) from it. Even if \(N\) is not very large, so there are some deviations from the Gaussian form, the above numbers are often a reasonable guide. However, as emphasized in appendix [\[sec:clt\]](#sec:clt){reference-type="ref" reference="sec:clt"}, distributions which occur in nature typically have much more weight in the tails than a Gaussian. As a result, the weight in the tails of the distribution *of the sum* can also be much larger than for a Gaussian even for quite large values of \(N\), see Fig. [\[Fig:converge_to_clt\]](#Fig:converge_to_clt){reference-type="ref" reference="Fig:converge_to_clt"}. It follows that the probability of an "outlier" can be much higher than that predicted for a Gaussian distribution, as anyone who has invested in the stock market knows well! ## Advanced Analysis {#sec:advanced} In Sec. [2.1](#sec:basic){reference-type="ref" reference="sec:basic"} we learned how to estimate a simple average, such as \(\mu_x \equiv \langle x \rangle\), plus the error bar in that quantity, from a set of data \(x_i\). Trivially this method also applies to a *linear* combination of different averages, \(\mu_x, \mu_y, \cdots\) etc. However, we often need more complicated, *non-linear* functions of averages. One example is the fluctuations in a quantity, i.e. \(\langle x^2 \rangle-\langle x \rangle^2\). Another example is a dimensionless combination of moments, which gives information about the *shape* of a distribution independent of its overall scale. Such quantities are very popular in finite-size scaling (FSS) analyses since the FSS form is simpler than for quantities with dimension. An popular example, first proposed by Binder, is \(\langle x^4 \rangle / \langle x^2 \rangle^2\), which is known as the "kurtosis" (frequently a factor of 3 is subtracted to make it zero for a Gaussian). Hence, in this section we consider how to determine *non-linear functions* of averages of one or more variables, \(f(\mu_y, \mu_z, \cdots)\), where \[\mu_y \equiv \langle y \rangle \,,\] etc. For example, the two quantities mentioned in the previous paragraph correspond to \[f(\mu_y, \mu_z) = \mu_y-\mu_z^2 \,,\] with \(y=x^2\) and \(z = x\) and \[f(\mu_y, \mu_z) = {\mu_y \over \mu_z^2} \,,\] with \(y = x^4\) and \(z = x^2\). The natural estimate of \(f(\mu_y, \mu_z)\) from the sample data is clearly \(f(\overline{y}, \overline{z} )\). However, it will take some more thought to estimate the error bar in this quantity. The traditional way of doing this is called "error propagation", described in Sec. [2.2.1](#sec:traditional){reference-type="ref" reference="sec:traditional"} below. However, it is now more common to use either "jackknife" or "bootstrap" procedures, described in Secs. [2.2.2](#sec:jack){reference-type="ref" reference="sec:jack"} and [2.2.3](#sec:boot){reference-type="ref" reference="sec:boot"}. At the price of some additional computation, which is no difficulty when done on a modern computer (though it would have been tedious in the old days when statistics calculations were done by hand), these methods automate the calculation of the error bar. Furthermore, the estimate of \(f(\mu_y, \mu_z)\) turns out to have some *bias* if \(f\) is a non-linear function. Usually this is small effect because it is order \(1/N\), see for example Eq. [\[bias\]](#bias){reference-type="eqref" reference="bias"} below, whereas the statistical error is of order \(1/\sqrt{N}\). Since \(N\) is usually large, the bias is generally much less than the statistical error and so can generally be neglected. In any case, the jackknife and bootstrap methods also enable one to eliminate the leading (\(\sim 1/N\)) contribution to the bias in a automatic fashion. ### Traditional method {#sec:traditional} First we will discuss the traditional method, known as error propagation, to compute the error bar and bias. We expand \(f(\overline{y}, \overline{z})\) about \(f(\mu_y, \mu_z)\) up to second order in the deviations: \[f(\overline{y}, \overline{z}) = f(\mu_y, \mu_z) + (\partial_{\mu_y}f)\, \delta_{\overline{y}} + (\partial_{\mu_z}f)\, \delta_{\overline{z}} + {1\over 2}\, (\partial^2_{\mu_y\mu_y}f)\, \delta_{\overline{y}}^2 + (\partial^2_{\mu_y\mu_z}f)\, \delta_{\overline{y}} \delta_{\overline{z}} + {1\over 2}\, (\partial^2_{\mu_z\mu_z}f)\, \delta_{\overline{z}}^2 + \cdots \,, \label{expand}\] where \[\delta_{\overline{y}} = \overline{y}-\mu_y,\] etc. The terms of first order in the \(\delta's\) in Eq. [\[expand\]](#expand){reference-type="eqref" reference="expand"} give the leading contribution to the error, but would average to zero if the procedure were to be repeated many times. However, the terms of second order do not average to zero and so give the leading contribution to the bias. We now estimate that bias. Averaging Eq. [\[expand\]](#expand){reference-type="eqref" reference="expand"} over many repetitions, and noting that \[\langle \delta_{\overline{y}}^2 \rangle = \langle \overline{y}^2 \rangle-\langle \overline{y} \rangle^2 \equiv \sigma_{\overline{y}}^2, \quad \langle \delta_{\overline{z}}^2 \rangle = \langle \overline{z}^2 \rangle-\langle \overline{z} \rangle^2 \equiv \sigma_{\overline{z}}^2, \quad \langle \delta_{\overline{y}} \delta_{\overline{z}} \rangle = \langle \overline{y}\, \overline{z} \rangle-\langle \overline{y} \rangle \langle \overline{z} \rangle \equiv \sigma_{\overline{y}\,\overline{z}}^2,\] we get \[\langle f(\overline{y}, \overline{z})\rangle-f(\mu_y, \mu_z) = {1\over 2}\, (\partial^2_{\mu_y\mu_y}f)\, \sigma_{\overline{y}}^2 + (\partial^2_{\mu_y\mu_z}f)\, \sigma_{\overline{y}\,\overline{z}}^2 + {1\over 2}\, (\partial^2_{\mu_z\mu_z}f)\, \sigma_{\overline{z}}^2 \,. \label{df}\] As shown in Eq. [\[errorbar\]](#errorbar){reference-type="eqref" reference="errorbar"} our estimate of \(\sigma_{\overline{y}}^2\) is \(N^{-1}\) times the sample variance (which we now call \(s_{yy}^2\)), and similarly for \(\sigma_{\overline{z}}^2\). In the same way, our estimate of \(\sigma_{\overline{y}\,\overline{z}}^2\) is \(N^{-1}\) times the sample *covariance* of \(y\) and \(z\), defined by \[s_{y z}^2 = {1 \over N-1}\, \sum_{i=1}^N \left(y_i-\overline{y}\right)\, \left(z_i-\overline{z}\right) \,.\] Hence, from Eq. [\[df\]](#df){reference-type="eqref" reference="df"}, we have \[f(\mu_y, \mu_z) = \langle f(\overline{y}, \overline{z})\rangle-{1\over N}\, \left[{1\over 2}\, (\partial^2_{\mu_y\mu_y}f)\, s_{y y}^2 + (\partial^2_{\mu_y\mu_z}f)\, s_{y z}^2 + {1\over 2}\, (\partial^2_{\mu_z\mu_z}f)\, s_{z z}^2 \right]\,, \label{bias2}\] where the leading contribution to the bias is given by the \(1/N\) term. Note that the bias term is "self-averaging", i.e. the fluctuations in it are small relative to the average (by a factor of \(1/\sqrt{N}\)) when averaging over many repetitions of the data. It follows from Eq. [\[bias2\]](#bias2){reference-type="eqref" reference="bias2"} that if one wants to eliminate the leading contribution to the bias one should \[\boxed{ \mbox{estimate } f(\mu_y,\mu_z)\ \mbox{ from } f(\overline{y}, \overline{z})-{1\over N}\, \left[{1\over 2}\, (\partial^2_{\mu_y\mu_y}f)\, s_{y y}^2 + (\partial^2_{\mu_y\mu_z}f)\, s_{y z}^2 + {1\over 2}\, (\partial^2_{\mu_z\mu_z}f)\, s_{z z}^2 \right].} \label{bias}\] As claimed earlier, the bias correction is of order \(1/N\). Note that it vanishes if \(f\) is a linear function, as shown in Sec. [2.1](#sec:basic){reference-type="ref" reference="sec:basic"}. The generalization to functions of more than two averages, \(f(\mu_y, \mu_z, \mu_w, \cdots)\), is obvious. Next we discuss the leading *error* in using \(f(\overline{y}, \overline{z})\) as an estimate for \(f(\mu_y, \mu_z)\). This comes from the terms linear in the \(\delta\)'s in Eq. [\[expand\]](#expand){reference-type="eqref" reference="expand"}. Just including these terms we have Hence \[\begin{aligned} \sigma_f^2 &\equiv \langle\, f^2(\overline{y}, \overline{z})\, \rangle-\langle f(\overline{y}, \overline{z}) \rangle^2 \nonumber \\ &= (\partial_{\mu_y}f)^2 \, \langle \delta_{\overline{y}}^2 \rangle + 2(\partial_{\mu_y}f)\, (\partial_{\mu_z}f) \, \langle \delta_{\overline{y}} \delta_{\overline{z}} \rangle + (\partial_{\mu_z}f)^2 \, \langle \delta_{\overline{z}}^2 \rangle \,. \end{aligned}\] As above, we use \(s_{y y}^2 / N\) as an estimate of \(\langle \delta_{\overline{y}}^2 \rangle\) and similarly for the other terms. Hence \[\boxed{ \mbox{the best estimate of } \sigma_f^2 \mbox{ is } {1 \over N}\, (\partial_{\mu_y}f)^2 \, s_{y y}^2 + 2(\partial_{\mu_y}f)\, (\partial_{\mu_z}f) \, s_{y z}^2 + (\partial_{\mu_z}f)^2 \, s_{z z}^2 \,.} \label{sigma_f}\] This estimate is unbiased to leading order in \(N\). Note that we need to keep track not only of fluctuations in \(y\) and \(z\), characterized by their variances \(s_{y y}^2\) and \(s_{z z}^2\), but also cross correlations between \(y\) and \(z\), characterized by their covariance \(s_{y z}^2\). Hence, still to leading order in \(N\), we get \[\boxed{f(\mu_y, \mu_z) = f(\overline{y}, \overline{z}) \pm \sigma_f\, ,}\] where we estimate the error bar \(\sigma_f\) from Eq. [\[sigma_f\]](#sigma_f){reference-type="eqref" reference="sigma_f"} which shows that it is of order \(1/\sqrt{N}\). Again, the generalization to functions of more than two averages is obvious. Note that in the simple case studied in Sec. [2.1](#sec:basic){reference-type="ref" reference="sec:basic"} where there is only one set of variables \(x_i\) and \(f =\mu_x\), Eq. [\[bias2\]](#bias2){reference-type="eqref" reference="bias2"} tells us that there is no bias, which is correct, and Eq. [\[sigma_f\]](#sigma_f){reference-type="eqref" reference="sigma_f"} gives an expression for the error bar which agrees with Eq. [\[finalans\]](#finalans){reference-type="eqref" reference="finalans"}. In Eqs. [\[bias\]](#bias){reference-type="eqref" reference="bias"} and [\[sigma_f\]](#sigma_f){reference-type="eqref" reference="sigma_f"} we need to keep track how errors in the individual quantities like \(\overline{y}\) propagate to the estimate of the function \(f\). This requires inputting by hand the various partial derivatives into the analysis program, and keeping track of all the variances and covariances. In the next two sections we see how *resampling* the data automatically takes account of error propagation without needing to input the partial derivatives and keep track of variances and covariances. These approaches, known as jackknife and bootstrap, provide a *fully automatic* method of determining error bars and bias. ### Jackknife {#sec:jack} We define the \(i\)-th jackknife estimate, \(y^J_i\, (i = 1,2, \cdots, N)\) to be the average over all data in the sample *except the point* \(i\), i.e. \[y^J_i \equiv {1 \over N-1}\, \sum_{j \ne i} y_j \,.\] We also define corresponding jackknife estimates of the function \(f\) (again for concreteness we will assume that \(f\) is a function of just 2 averages but the generalization will be obvious): \[f^J_i \equiv f(y^J_i, z^J_i) \,. \label{fJi}\] In other words, we use the jackknife values, \(y^J_i, z^J_i\), rather than the sample means, \(\overline{y}, \overline{z}\), as the arguments of \(f\). For example a jackknife estimate of the Binder ratio \(\langle x^4 \rangle / \langle x^2 \rangle^2\) is \[f^J_i = {(N-1)^{-1} \sum_{j, (j \ne i)} x_j^4 \over \left[(N-1)^{-1} \sum_{j \ne i} x_j^2\right]^2 }\] The overall jackknife estimate of \(f(\mu_ y, \mu_z)\) is then the average over the \(N\) jackknife estimates \(f_i^J\): \[\boxed{ \overline{f^J} \equiv {1 \over N} \sum_{i=1}^N f_i^J \,.} \label{fJ}\] It is straightforward to show that if \(f\) is a linear function of \(\mu_y\) and \(\mu_z\) then \(\overline{f^J} = f(\overline{y},\overline{z})\), i.e. the jackknife and standard averages are identical. However, when \(f\) is not a linear function, so there is bias, there *is* a difference, and we will now show the resampling carried out in the jackknife method can be used to determine bias and error bars in an automated way. We proceed as for the derivation of Eq. [\[bias2\]](#bias2){reference-type="eqref" reference="bias2"}, which we now write as \[f(\mu_y, \mu_z) = \langle f(\overline{y},\overline{z}) \rangle-{A \over N}-{B\over N^2} + \cdots,\] where \(A\) is the term in rectangular brackets in Eq. [\[bias2\]](#bias2){reference-type="eqref" reference="bias2"}, and we have added the next order correction. The jackknife data sets have \(N-1\) points with the same distribution as the \(N\) points in the actual distribution, and so the bias in the jackknife average will be of the same form, with the same values of \(A\) and \(B\), but with \(N\) replaced by \(N-1\), i.e. \[f(\mu_y, \mu_z) = \langle \overline{f^J} \rangle-{A \over N-1}-{B \over (N-1)^2} \cdots \,.\] We can therefore eliminate the leading contribution to the bias by forming an appropriate linear combination of \(f(\overline{y},\overline{z})\) and \(\overline{f^J}\), namely \[f(\mu_y, \mu_z) = N \langle f(\overline{y},\overline{z}) \rangle-(N-1) \langle \overline{f^J} \rangle + O\left({1\over N^2}\right) \,.\] It follows that, to eliminate the leading bias without computing partial derivatives, one should \[\boxed{ \mbox{estimate } f(\mu_y, \mu_z) \mbox{ from } N f(\overline{y},\overline{z})-(N-1) \overline{f^J} \,. } \label{bias_elim}\] The bias is then of order \(1/N^2\). However, as mentioned earlier, bias is usually not a big problem because, even without eliminating the leading contribution, the bias is of order \(1/N\) whereas the statistical error is of order \(1/\sqrt{N}\) which is much bigger if \(N\) is large. In most cases, therefore, \(N\) is sufficiently large that one can use *either* the usual average \(f(\overline{y}, \overline{z})\), or the jackknife average \(\overline{f^J}\) in Eq. [\[fJ\]](#fJ){reference-type="eqref" reference="fJ"}, to estimate \(f(\mu_y, \mu_z)\), since the difference between them will be much smaller than the statistical error. In other words, elimination of the leading bias using Eq. [\[bias_elim\]](#bias_elim){reference-type="eqref" reference="bias_elim"} is usually not necessary. Next we show that the jackknife method gives error bars, which agree with Eq. [\[sigma_f\]](#sigma_f){reference-type="eqref" reference="sigma_f"} but without the need to explicitly keep track of the partial derivatives and the variances and covariances. We define the variance of the jackknife averages by \[\sigma^2_{f^J} \equiv \overline{\left(f^J\right)^2}-\left( \overline{f^J} \right)^2 \,, \label{sigmafJ}\] where \[\overline{\left(f^J\right)^2} = {1 \over N} \sum_{i=1}^N \left(f_i^J\right)^2 \,.\] Using Eqs. [\[fJi\]](#fJi){reference-type="eqref" reference="fJi"} and [\[fJ\]](#fJ){reference-type="eqref" reference="fJ"}, we expand \(\overline{f^J}\) away from the exact result \(f(\mu_y, \mu_z)\). Just including the leading contribution gives \[\begin{aligned} \overline{f^J}-f(\mu_y, \mu_z) &= {1 \over N} \sum_{i=1}^N \left[ (\partial_{\mu_y} f)\, (y_i^J-\mu_y) + (\partial_{\mu_z} f)\, (z_i^J-\mu_z) \right] \nonumber \\ &= {1 \over N(N-1)} \sum_{i=1}^N \left[ (\partial_{\mu_y} f)\, \left\{N(\overline{y}-\mu_y)-(y_i-\mu_y) \right\} + (\partial_{\mu_z} f)\, \left\{N(\overline{z}-\mu_z)-(z_i-\mu_z) \right\} \right] \nonumber \\ &= (\partial_{\mu_y} f)\, (\overline{y}-\mu_y) + (\partial_{\mu_z} f)\, (\overline{z}-\mu_z) \,. \label{fJ-f} \end{aligned}\] Similarly we find \[\begin{aligned} \overline{\left(f^J\right)^2 } &= {1 \over N} \sum_{i=1}^N \left[ f(\mu_y, \mu_z) + (\partial_{\mu_y} f)\, (y_i^J-\mu_y) + (\partial_{\mu_z} f)\, (z_i^J-\mu_z) \right]^2 \nonumber \\ &= f^2(\mu_y, \mu_z) + 2 f(\mu_y, \mu_z) \, \left[ (\partial_{\mu_y} f)\, (\overline{y}-\mu_y) + (\partial_{\mu_z} f)\, (\overline{z}-\mu_z) \right] \nonumber \\ &\quad + (\partial_{\mu_y} f)^2\, \left[(\overline{y}-\mu_y)^2 + {s_{yy}^2 \over N(N-1)}\right] + (\partial_{\mu_z} f)^2\, \left[(\overline{z}-\mu_z)^2 + {s_{zz}^2 \over N(N-1)}\right] \nonumber \\ &\qquad + 2(\partial_{\mu_y} f)(\partial_{\mu_z} f)\,\left[(\overline{y}-\mu_y) (\overline{z}-\mu_z) + {s_{yz}^2 \over N(N-1)}\right] \,. \end{aligned}\] Hence, from Eqs. [\[sigmafJ\]](#sigmafJ){reference-type="eqref" reference="sigmafJ"} and [\[fJ-f\]](#fJ-f){reference-type="eqref" reference="fJ-f"}, the variance in the jackknife estimates is given by \[\sigma^2_{f^J} = {1 \over N(N-1)} \, \left[ (\partial_{\mu_y} f)^2\, s_{yy}^2 + (\partial_{\mu_z} f)^2\, s_{zz}^2 + 2(\partial_{\mu_y} f)(\partial_{\mu_z} f) s_{yz}\right] \,,\] which is just \(1/(N-1)\) times \(\sigma_f^2\), the estimate of the square of the error bar in \(f(\overline{y}, \overline{z})\) given in Eq. [\[sigma_f\]](#sigma_f){reference-type="eqref" reference="sigma_f"}. Hence \[\boxed{ \mbox{the jackknife estimate for } \sigma_f \mbox{ is } \sqrt{N-1} \, \sigma_{f^J}\,.} \label{error_jack}\] Note that this is directly obtained from the jackknife estimates without having to put in the partial derivatives by hand. Note too that the \(\sqrt{N-1}\) factor is in the *numerator* whereas the factor of \(\sqrt{N}\) in Eq. [\[finalans\]](#finalans){reference-type="eqref" reference="finalans"} is in the *denominator*. Intuitively the reason for this difference is that the jackknife estimates are very close since they would all be equal except that each one omits just one data point. If \(N\) is very large, roundoff errors could become significant from having to subtract large, almost equal, numbers to get the error bar from the jackknife method. It is then advisable to group the \(N\) data points into \(N_\text{group}\) groups (or "bins") of data and take, as individual data points in the jackknife analysis, the average of the data in each group. The above results clearly go through with \(N\) replaced by \(N_\text{group}\). To summarize this subsection, to estimate \(f(\mu_y, \mu_z)\) one can use either \(f(\overline{y}, \overline{z})\) or the jackknife average \(\overline{f^J}\) in Eq. [\[fJ\]](#fJ){reference-type="eqref" reference="fJ"}. The error bar in this estimate, \(\sigma_f\), is related to the standard deviation in the jackknife estimates \(\sigma_{f^J}\) by Eq. [\[error_jack\]](#error_jack){reference-type="eqref" reference="error_jack"}. ### Bootstrap {#sec:boot} The bootstrap, like the jackknife, is a resampling of the \(N\) data points Whereas jackknife considers \(N\) new data sets, each of containing all the original data points minus one, bootstrap uses \({N_{\rm boot}}\) data sets each containing \(N\) points obtained by random (Monte Carlo) sampling of the original set of \(N\) points. During the Monte Carlo sampling, the probability that a data point is picked is \(1/N\) irrespective of whether it has been picked before. (In the statistics literature this is called picking from a set "with replacement".) Hence a given data point \(x_i\) will, *on average*, appear once in each Monte Carlo-generated data set, but may appear not at all, or twice, and so on. The probability that \(x_i\) appears \(n_i\) times is close to a Poisson distribution with mean unity. However, it is not exactly Poissonian because of the constraint in Eq. ([\[constraint\]](#constraint){reference-type="ref" reference="constraint"}) below. It turns out that we shall need to include the deviation from the Poisson distribution even for large \(N\). We shall use the term "bootstrap" data sets to denote the Monte Carlo-generated data sets. More precisely, let us suppose that the number of times \(x_i\) appears in a bootstrap data set is \(n_i\). Since each bootstrap dataset contains exactly \(N\) data points, we have the constraint \[\sum_{i=1}^N n_i = N \,. \label{constraint}\] Consider one of the \(N\) variables \(x_i\). Each time we generate an element in a bootstrap dataset the probability that it is \(x_i\) is \(1/N\), which we will denote by \(p\). From standard probability theory, the probability that \(x_i\) occurs \(n_i\) times is given by a binomial distribution \[P(n_i) = {N! \over n_i! \, (N-n_i)!} \, p^{n_i} (1-p)^{N-n_i} \,.\] The mean and standard deviation of a binomial distribution are given by \[\begin{aligned} _{_{\rm MC}} & = N p = 1 \,, \label{nimc} \\ {[ n_i^2 ]_{_{\rm MC}}}-[n_i]_{_{\rm MC}}^2 & = N p (1-p) = 1-{1 \over N} \,, \label{epsi_epsi} \end{aligned}\] where \([ \dots ]_{_{\rm MC}}\) denotes an exact average over bootstrap samples (for a fixed original data set \(x_i\)). For \(N \to\infty\), the binomial distribution goes over to a Poisson distribution, for which the factor of \(1/N\) in Eq. ([\[epsi_epsi\]](#epsi_epsi){reference-type="ref" reference="epsi_epsi"}) does not appear. We assume that \({N_{\rm boot}}\) is sufficiently large that the bootstrap average we carry out reproduces this result with sufficient accuracy. Later, we will discuss what values for \({N_{\rm boot}}\) are sufficient in practice. Because of the constraint in Eq. ([\[constraint\]](#constraint){reference-type="ref" reference="constraint"}), \(n_i\) and \(n_j\) (with \(i \ne j\)) are not independent and we find, by squaring Eq. [\[constraint\]](#constraint){reference-type="eqref" reference="constraint"} and using Eqs. [\[nimc\]](#nimc){reference-type="eqref" reference="nimc"} and [\[epsi_epsi\]](#epsi_epsi){reference-type="eqref" reference="epsi_epsi"}, that \[_{_{\rm MC}}-[ n_i ]_{_{\rm MC}} [ n_j ]_{_{\rm MC}} =-{1 \over N} \qquad (i \ne j)\,. \label{epsi_epsj}\] First of all we just consider the simple average \(\mu_x \equiv \langle x \rangle\), for which, of course, the standard methods in Sec. [2.1](#sec:basic){reference-type="ref" reference="sec:basic"} suffice, so bootstrap is not necessary. However, this will show how to get averages and error bars in a simple case, which we will then generalize to non-linear functions of averages. We denote the average of \(x\) for a given bootstrap data set by \(x^B_\alpha\), where \(\alpha\) runs from 1 to \({N_{\rm boot}}\), *i.e.* \[x^B_\alpha = {1 \over N} \sum_{i=1}^N n_i^\alpha x_i \,.\] We then compute the bootstrap average of the mean of \(x\) and the bootstrap variance in the mean, by averaging over all the bootstrap data sets. We assume that \({N_{\rm boot}}\) is large enough for the bootstrap average to be exact, so we can use Eqs. ([\[epsi_epsi\]](#epsi_epsi){reference-type="ref" reference="epsi_epsi"}) and ([\[epsi_epsj\]](#epsi_epsj){reference-type="ref" reference="epsi_epsj"}). The result is \[\begin{aligned} \label{xb} \overline{x^B} \equiv {1 \over {N_{\rm boot}}} \sum_{\alpha=1}^{N_{\rm boot}} x^B_\alpha & = & {1\over N} \sum_{i=1}^N [n_i]_{_{\rm MC}} x_i = {1\over N} \sum_{i=1}^N x_i = \overline{x} \\ \sigma^2_{x^B} \equiv \overline{\left(x^B\right)^2}-\left(\overline{x^B}\right)^2 & = & {1\over N^2} \left(1-{1\over N}\right) \sum_i x_i^2-{1 \over N^3} \sum_{i \ne j} x_i x_j \,, \label{sigmab} \end{aligned}\] where \[\overline{\left(x^B\right)^2} \equiv {1 \over {N_{\rm boot}}} \sum_{\alpha=1}^{N_{\rm boot}} \left[ \left(x^B_\alpha\right)^2\right]_{_{\rm MC}} \,.\] We now average Eqs. ([\[xb\]](#xb){reference-type="ref" reference="xb"}) and ([\[sigmab\]](#sigmab){reference-type="ref" reference="sigmab"}) over many repetitions of the original data set \(x_i\). Averaging Eq. ([\[xb\]](#xb){reference-type="ref" reference="xb"}) gives \[\langle \overline{x^B} \rangle = \langle \overline{x} \rangle = \langle x \rangle \equiv \mu_x \,.\] This shows that the bootstrap average \(\,\overline{x^B}\,\) is an unbiased estimate of the exact average \(\mu_x\). Averaging Eq. ([\[sigmab\]](#sigmab){reference-type="ref" reference="sigmab"}) gives \[\left\langle \sigma^2_{x^B} \right\rangle = {N-1 \over N^2} \sigma^2 = {N-1 \over N} \sigma^2_{\overline{x}} \,,\] where we used Eq. ([\[dxsq\]](#dxsq){reference-type="ref" reference="dxsq"}) to get the last expression. Since \(\sigma_{\overline{x}}\) is the uncertainty in the sample mean, we see that \[\boxed{\mbox{the bootstrap estimate of }\sigma_{\overline{x}} \mbox{ is } \sqrt{N \over N-1}\, \sigma_{x^B} \,.} \label{sigmaxb}\] Remember that \(\sigma_{x^B}\) is the standard deviation of the bootstrap data sets. Usually \(N\) is sufficiently large that the square root in Eq. ([\[sigmaxb\]](#sigmaxb){reference-type="ref" reference="sigmaxb"}) can be replaced by unity. As for the jackknife, these results can be generalized to finding the error bar in some possibly non-linear function, \(f(\mu_y, \mu_z)\), rather than for \(\mu_x\). One computes the bootstrap estimates for \(f(\mu_y, \mu_z)\), which are \[f^B_\alpha = f(y^B_\alpha, z^B_\alpha) \,.\] In other words, we use the bootstrap values, \(y^B_\alpha, z^B_\alpha\), rather than the sample means, \(\overline{y}, \overline{z}\), as the arguments of \(f\). The final bootstrap estimate for \(f(\mu_y, \mu_z)\) is the average of these, *i.e.* \[\boxed{ \overline{f^B} = {1 \over {N_{\rm boot}}} \sum_{\alpha=1}^{N_{\rm boot}} f^B_\alpha \,.} \label{fb}\] Following the same methods in the jackknife section, one obtains the error bar, \(\sigma_f\), in \(f(\mu_y, \mu_z)\). The result is \[\boxed{\mbox{the bootstrap estimate for } \sigma_f \mbox{ is } \sqrt{N \over N-1} \,\, \sigma_{f^B}}, \label{sigmafb}\] where \[\boxed{ \sigma^2_{f^B} = \overline{\left(f^B\right)^2}-\left(\overline{f^B}\right)^2 \, ,}\] is the variance of the bootstrap estimates. Here \[\overline{\left(f^B\right)^2} \equiv {1 \over {N_{\rm boot}}} \sum_{\alpha=1}^{N_{\rm boot}} \left(f^B_\alpha\right)^2 \,.\] Usually \(N\) is large enough that the factor of \(\sqrt{N/(N-1)}\) is Eq. [\[sigmafb\]](#sigmafb){reference-type="eqref" reference="sigmafb"} can be replaced by unity. Equation ([\[sigmafb\]](#sigmafb){reference-type="ref" reference="sigmafb"}) corresponds to the result Eq. ([\[sigmaxb\]](#sigmaxb){reference-type="ref" reference="sigmaxb"}) which we derived for the special case of \(f = \mu_x\). Again, following the same path as in the jackknife section, it is straightforward to show that the bias of the estimates in Eqs. ([\[fb\]](#fb){reference-type="ref" reference="fb"}) and ([\[sigmafb\]](#sigmafb){reference-type="ref" reference="sigmafb"}) is of order \(1/N\) and so vanishes for \(N\to\infty\). However, if \(N\) is not too large it may be useful to eliminate the leading contribution to the bias in the mean, as we did for jackknife in Eq. ([\[bias_elim\]](#bias_elim){reference-type="ref" reference="bias_elim"}). The result is that one should \[\boxed{\mbox{estimate } f(\mu_y, \mu_z) \mbox{ from } 2 f(\overline{y}, \overline{z})-\overline{f^B} \,.} \label{improved_boot}\] The bias in Eq. ([\[improved_boot\]](#improved_boot){reference-type="ref" reference="improved_boot"}) is of order \(1/N^2\), whereas \(f(\overline{y}, \overline{z})\) and \(\overline{f^B}\) each have a bias of order \(1/N\). However, it is not normally necessary to eliminate the bias since, if \(N\) is large, the bias is much smaller than the statistical error. I have not systematically studied the values of \({N_{\rm boot}}\) that are needed in practice to get accurate estimates for the error. It seems that \({N_{\rm boot}}\) in the range 100 to 500 is typically chosen, and this seems to be adequate irrespective of how large \(N\) is. To summarize this subsection, to estimate \(f(\mu_y, \mu_z)\) one can either use \(f(\overline{y}, \overline{z})\), or the bootstrap average in Eq. [\[fb\]](#fb){reference-type="eqref" reference="fb"}, and the error bar in this estimate, \(\sigma_f\), is related to the standard deviation in the bootstrap estimates by Eq. [\[sigmafb\]](#sigmafb){reference-type="eqref" reference="sigmafb"}. ### Jackknife or Bootstrap? {#sec:jorb} The jackknife approach involves less calculation than bootstrap, and is fine for estimating combinations of moments of the measured quantities. Furthermore, identical results are obtained each time jackknife is run on the same set of data, which is not the case for bootstrap. However, the range of the jackknife estimates is very much smaller, by a factor of \(\sqrt{N}\) for large \(N\), than the scatter of averages which would be obtained from individual data sets, see Eq. [\[error_jack\]](#error_jack){reference-type="eqref" reference="error_jack"}. By contrast, for bootstrap, \(\sigma_{f^B}\), which measures the deviation of the bootstrap estimates \(f^B_\alpha\) from the result for the single actual data set \(f(\overline{y}, \overline{z})\), *is equal to* \(\sigma_f\), the deviation of the average of a single data set from the exact result \(f(\mu_y,\mu_z)\) (if we replace the factor of \(N/(N-1)\) by unity, see Eq. [\[sigmafb\]](#sigmafb){reference-type="eqref" reference="sigmafb"}). This is the main strength of the bootstrap approach; it samples the full range of the distribution of the sample distribution. Hence, if you want to generate data which covers the full range then should use bootstrap. This is useful in fitting, see for example, Sec. [3.6](#sec:resample){reference-type="ref" reference="sec:resample"}. However, if you just want to generate error bars on combinations of moments quickly and easily, then use jackknife. # Fitting data to a model {#sec:fit} A good reference for the material in this section is Chapter 15 of Numerical Recipes . Frequently we are given a set of data points \((x_i, y_i), i = 1, 2, \cdots, N\), with corresponding error bars, \(\sigma_i\), through which we would like to fit to a smooth function \(f(x)\). The function could be straight line (the simplest case), a higher order polynomial, or a more complicated function. The fitting function will depend on \(M\) "fitting parameters", \(a_\alpha\) and we would like the "best" fit obtained by adjusting these parameters. We emphasize that a fitting procedure should not only 1. [\[give_params\]]{#give_params label="give_params"} give the values of the fit parameters, but also 2. [\[give_errors\]]{#give_errors label="give_errors"} provide error estimates on those parameters, and 3. [\[gof\]]{#gof label="gof"} provide a measure of how good the fit is. If the result of part [\[gof\]](#gof){reference-type="ref" reference="gof"} is that the fit is very poor, the results of parts [\[give_params\]](#give_params){reference-type="ref" reference="give_params"} and [\[give_errors\]](#give_errors){reference-type="ref" reference="give_errors"} are probably meaningless. The definition of "best" is not unique. However, the most useful choice, and the one nearly always taken, is "least squares", in which one minimizes the sum of the squares of the difference between the observed \(y\)-value, \(y_i\), and the fitting function evaluated at \(x_i\), weighted appropriately by the error bars since if some points have smaller error bars than others the fit should be closer to those points. The quantity to be minimized, called "chi-squared",[^2] and written mathematically as \(\chi^2\), is therefore \[\boxed{ \chi^2 = \sum_{i=1}^N \left( \, {y_i-f(x_i) \over \sigma_i } \, \right)^2. } \label{chisq}\] Often we assume that the distribution of the errors is Gaussian, since, according to the central limit theorem discussed in Appendix [\[sec:clt\]](#sec:clt){reference-type="ref" reference="sec:clt"}, the sum of \(N\) independent random variables has a Gaussian distribution (under fairly general conditions) if \(N\) is large. However, distributions which occur in nature usually have more weight in the "tails" than a Gaussian, and as a result, even for moderately large values of \(N\), the probability of an "outlier" might be much bigger than expected from a Gaussian, see Fig. [\[Fig:converge_to_clt\]](#Fig:converge_to_clt){reference-type="ref" reference="Fig:converge_to_clt"}. If the errors *are* distributed with a Gaussian distribution, and if \(f(x)\) has the *exact* values of the fit parameters, then \(\chi^2\) in Eq. [\[chisq\]](#chisq){reference-type="eqref" reference="chisq"} is a sum of squares of \(N\) random variables with a Gaussian distribution with mean zero and standard deviation unity. However, when we have minimized the value of \(\chi^2\) with respect to the \(M\) fitting parameters \(a_\alpha\) the terms are not all independent. It turns out, see Appendix [\[sec:NDF\]](#sec:NDF){reference-type="ref" reference="sec:NDF"}, that, at least for a linear model (which we define below), the distribution of \(\chi^2\) at the minimum is that of the sum of the squares of \(N-M\) (not \(N\)) Gaussian random variable with zero mean and standard deviation unity[^3]. We call \(N-M\) the "number of degrees of freedom" (\(N_\text{DOF}\)). The \(\chi^2\) distribution is discussed in Appendix [\[sec:Q\]](#sec:Q){reference-type="ref" reference="sec:Q"}. The formula for it is Eq. [\[chisq-dist\]](#chisq-dist){reference-type="eqref" reference="chisq-dist"}. The simplest problems are where the fitting function is a *linear function of the parameters*. We shall call this a *linear model*. Examples are a straight line (\(M=2\)), \[y = a_0 + a_1 x \,, \label{sl}\] and an \(m\)-th order polynomial (\(M=m+1\)), \[y = a_0 + a_1 x + a_2 x^2 + \cdots + a_m x^m = \sum_{\alpha=0}^m a_\alpha x^m \,, \label{poly}\] where the parameters to be adjusted are the \(a_\alpha\). (Note that we are *not* stating here that \(y\) has to be a linear function of \(x\), only of the fit parameters \(a_\alpha\).) An example where the fitting function depends *non*-linearly on the parameters is \[y = a_0 x^{a_1} + a_2 \,.\] Linear models are fairly simply because, as we shall see, the parameters are determined by *linear* equations, which, in general, have a unique solution that can be found by straightforward methods. However, for fitting functions which are non-linear functions of the parameters, the resulting equations are *non-linear* which may have many solutions or none at all, and so are much less straightforward to solve. We shall discuss fitting to both linear and non-linear models in these notes. Sometimes a non-linear model can be transformed into a linear model by a change of variables. For example, if we want to fit to \[y = a_0 x^{a_1} \,,\] which has a non-linear dependence on \(a_1\), taking logs gives \[\ln y = \ln a_0 + a_1 \ln x \,,\] which is a *linear* function of the parameters \(a'_0 = \ln a_0\) and \(a_1\). Fitting a straight line to a log-log plot is a very common procedure in science and engineering. However, it should be noted that transforming the data does not exactly take Gaussian errors into Gaussian errors, though the difference will be small if the errors are "sufficiently small". For the above log transformation this means \(\sigma_i / y_i \ll 1\), i.e. the *relative* error is much less than unity. ## Fitting to a straight line To see how least squares fitting works, consider the simplest case of a straight line fit, Eq. ([\[sl\]](#sl){reference-type="ref" reference="sl"}), for which we have to minimize \[\chi^2(a_0, a_1) = \sum_{i=1}^N \left({\, y_i-a_0-a_1 x_i\, \over \sigma_i} \right)^2 \,, \label{chisq_sline}\] with respect to \(a_0\) and \(a_1\). Differentiating \(\chi^2\) with respect to these parameters and setting the results to zero gives We write this as where \[\begin{aligned} &\boxed{U_{\alpha\beta} = \sum_{i=1}^N {x_i^{\alpha + \beta}\over \sigma_i^2}, } \quad \mbox{and} \label{Uab} \\ &\boxed{v_\alpha = \sum_{i=1}^N{ y_i\, x_i^\alpha \over \sigma_i^2 \, }. } \label{v} \end{aligned}\] The matrix notation, while an overkill here, will be convenient later when we do a general polynomial fit. Note that \(U_{10} = U_{01}\). (More generally, later on, \(U\) will be a symmetric matrix). Equations ([\[lssl\]](#lssl){reference-type="ref" reference="lssl"}) are two linear equations in two unknowns. These can be solved by eliminating one variable, which immediately gives an equation for the second one. The solution can also be determined from \[\boxed{ a_\alpha = \sum_{\beta=0}^m \left(U^{-1}\right)_{\alpha\beta} \, v_\beta, } \label{soln}\] (where we have temporarily generalized to a polynomial of order \(m\)). For the straight-line fit, the inverse of the \(2\times 2\) matrix \(U\) is given, according to standard rules, by \[U^{-1} = {1 \over \Delta} \, \begin{pmatrix} U_{11} & -U_{01} \\-U_{01} & U_{00} \end{pmatrix} \label{Uinv}\] where \[\boxed{ \Delta = U_{00} U_{11}-U_{01}^2 ,} \label{Delta}\] and we have noted that \(U\) is symmetric so \(U_{01} = U_{10}\). The solution for \(a_0\) and \(a_1\) is therefore given by We see that it is straightforward to determine the slope, \(a_1\), and the intercept, \(a_0\), of the fit from Eqs. ([\[Uab\]](#Uab){reference-type="ref" reference="Uab"}), ([\[v\]](#v){reference-type="ref" reference="v"}), ([\[Delta\]](#Delta){reference-type="ref" reference="Delta"}) and ([\[soln_sl\]](#soln_sl){reference-type="ref" reference="soln_sl"}) using the \(N\) data points \((x_i,y_i)\), and their error bars \(\sigma_i\). ## Fitting to a polynomial Frequently we need to fit to a higher order polynomial than a straight line, in which case we minimize \[\chi^2(a_0,a_1,\cdots,a_m) = \sum_{i=1}^N \left({y_i-\sum_{\alpha=0}^m a_\alpha x_i^\alpha \over \sigma_i} \right)^2 \label{chisq_poly}\] with respect to the \((m+1)\) parameters \(a_\alpha\). Setting to zero the derivatives of \(\chi^2\) with respect to the \(a_\alpha\) gives \[\boxed{ \sum_{\beta=0}^m U_{\alpha\beta}\, a _\beta = v_\alpha ,} \label{lspoly}\] where \(U_{\alpha\beta}\) and \(v_\alpha\) have been defined in Eqs. ([\[Uab\]](#Uab){reference-type="ref" reference="Uab"}) and ([\[v\]](#v){reference-type="ref" reference="v"}). Eq. ([\[lspoly\]](#lspoly){reference-type="ref" reference="lspoly"}) represents \(M = m+1\) *linear* equations, one for each value of \(\alpha\). Their solution is again given by Eq. ([\[soln\]](#soln){reference-type="ref" reference="soln"}), i.e. it is expressed in terms of the inverse matrix \(U^{-1}\). ## Error Bars {#sec:error_bars} In addition to the best fit values of the parameters we also need to determine the error bars in those values. Interestingly, this information is *also* contained in the matrix \(U^{-1}\). First of all, we explain the significance of error bars in fit parameters. We assume that the data is described by a model with a particular set of parameters \(\vec{a}^\text{true}\) which, unfortunately, we don't know. If we were, somehow, to have many real data sets each one would give a different set of fit parameters \(\vec{a}^{(i)}, i = 0, 1, 2, \cdots\), because of noise in the data, *clustered about the true set* \(\vec{a}^\text{true}\). Projecting on to a single fit parameter, \(a_1\) say, there will be a distribution of values \(P(a_1)\) centered on \(a_1^\text{true}\) with standard deviation \(\sigma_1\), see the top part of Fig. [\[Fig:distofa1\]](#Fig:distofa1){reference-type="ref" reference="Fig:distofa1"}. Typically the value of \(a_1\) obtained from our *one actual data set*, \(a_1^{(0)}\), will lie within about \(\sigma_1\) of \(a_1\). Hence we define the error bar to be \(\sigma_1\). Unfortunately, we can't determine the error bar this way because we have only one actual data set, which we denote here by \(y_i^{(0)}\) to distinguish it from other data sets that we will introduce. Our actual data set gives one set of fit parameters, which we call \(\vec{a}^{(0)}\). Suppose, however, we were to generate many *simulated* data sets from of the one which is available to us, by generating random values (possibly with a Gaussian distribution though this won't be necessary yet) centered at the \(y_i\) with standard deviation \(\sigma_i\). Fitting each simulated dataset would give different values for \(\vec{a}\), *clustered now about* \(\vec{a}^{(0)}\), see the bottom part of Fig. [\[Fig:distofa1\]](#Fig:distofa1){reference-type="eqref" reference="Fig:distofa1"}. We now come to an important, but rarely discussed, point: > We assume that the standard deviation of the fit parameters of these simulated data sets about \(\vec{a}^{(0)}\), which we will be able to calculate from the single set of data available to us, is equal to the standard deviation of the fit parameters of real data sets \(\vec{a}\) about \(\vec{a}^\text{true}\). The latter is what we *really* want to know (since it is our estimate of the error bar on \(\vec{a}^\text{true}\)) but can't determine directly. See Fig. [\[Fig:distofa1\]](#Fig:distofa1){reference-type="ref" reference="Fig:distofa1"} for an illustration. In fact we show in the text below that this assumption is correct for a linear model (and for a non-linear model if the range of parameter values is small enough that it can be represented by an effective linear model). Even if the model is non linear, one usually assumes that the two standard deviations are sufficiently close that the difference is not important. Furthermore, we show in Appendices [\[sec:proof\]](#sec:proof){reference-type="ref" reference="sec:proof"} and [\[sec:proof2\]](#sec:proof2){reference-type="ref" reference="sec:proof2"} that if the noise on the data is Gaussian (and the model is linear), the two distributions in Fig. [\[Fig:distofa1\]](#Fig:distofa1){reference-type="eqref" reference="Fig:distofa1"} are also both Gaussian. Hence, as stated above, to derive the error bars in the fit parameters we take simulated values of the data points, \(y_i^S\), which vary by some amount \(\delta y_i^S\) about \(y_i^{(0)}\), i.e. \(\delta y_i^S = y_i^S-y_i^{(0)}\), with a standard deviation given by the error bar \(\sigma_i\). The fit parameters of this simulated data set, \(\vec{a}^S\), then deviate from \(\vec{a}^{(0)}\) by an amount \(\delta \vec{a}^S\) where \[\delta a_\alpha^S = \sum_{i=1}^N {\partial a_\alpha \over \partial y_i}\, \delta y_i^S\,.\] Averaging over fluctuations in the \(y_i^S\) we get the variance of \(a_\alpha^S\) to be \[\left(\sigma_\alpha^S\right)^2 \equiv \langle \left(\delta a_\alpha^S\right)^2 \rangle = \sum_{i=1}^N \sigma_i^2 \, \left( {\partial a_\alpha \over \partial y_i} \right)^2 \,, \label{sigma_alpha}\] since \(\langle \left(\delta y_i^S\right)^2 \rangle = \sigma_i^2\), and the data points \(y_i\) are statistically independent. Writing Eq. [\[soln\]](#soln){reference-type="eqref" reference="soln"} explicitly in terms of the data values, \[a_\alpha = \sum_\beta \left(U^{-1}\right)_{\alpha\beta} \sum_{i=1}^N { y_i\, x_i^\beta \over \sigma_i^2 \, } \,,\] and noting that \(U\) is independent of the \(y_i\), we get \[{\partial a_\alpha \over \partial y_i} = \sum_\beta \left(U^{-1}\right)_{\alpha\beta} {x_i^\beta \over \sigma_i^2} \,.\] Substituting into Eq. [\[sigma_alpha\]](#sigma_alpha){reference-type="eqref" reference="sigma_alpha"} gives \[\left(\sigma_\alpha^S \right)^2 = \sum_{\beta, \gamma} \left(U^{-1}\right)_{\alpha\beta} \left(U^{-1}\right)_{\alpha\gamma} \left[ \sum_{i=1}^N {x_i^{\beta + \gamma} \over \sigma_i^2} \right] \,.\] The term in rectangular brackets is just \(U_{\beta\gamma}\), and so, noting that \(U\) is given by Eq. [\[Uab\]](#Uab){reference-type="eqref" reference="Uab"} and is symmetric, the last equation reduces to \[\left(\sigma_\alpha^S \right)^2 = \left(U^{-1}\right)_{\alpha\alpha} \,. \label{error_params_s}\] Recall that \(\sigma_\alpha^S\) is the standard deviation of the fitted parameter values about the \(\vec{a}^{(0)}\) when constructing simulated data sets from the one set of data that is available to us. However, the error bar is defined to be the standard deviation the fitted parameter values would have relative to \(a_\alpha^\text{true}\) if we could average over many actual data sets. To determine this quantity we simply repeat the above calculation with \(\delta y_i = y_i-y_i^\text{true}\) in which \(y_i\) is the value of the \(i\)-th data point in one of the actual data sets. The result is identical to Eq. [\[error_params_s\]](#error_params_s){reference-type="eqref" reference="error_params_s"}, namely \[\boxed{ \sigma_\alpha^2 = \left(U^{-1}\right)_{\alpha\alpha} \, ,} \label{error_params}\] in which \(U\) is the *same* in Eq. [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"} as in Eq. [\[error_params_s\]](#error_params_s){reference-type="eqref" reference="error_params_s"} because \(U\) is a constant, for a linear model, independent of the \(y_i\) or the fit parameters \(a_\alpha\). Hence \(\sigma_\alpha\) in Eq. [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"} is the error bar in \(a_\alpha\). In addition to error bars, we also need a parameter to describe the quality of the fit. A useful quantity is the probability that, given the fit, the data could have occurred with a \(\chi^2\) greater than or equal to the value found. This is generally denoted by \(Q\) and is given by Eq. [\[Q_expression\]](#Q_expression){reference-type="eqref" reference="Q_expression"} assuming the data have Gaussian noise. Note that the effects of *non-Gaussian* statistics is to increase the probability of outliers, so fits with a fairly small value of \(Q\), say around \(0.01\), may be considered acceptable. However, fits with a *very* small value of \(Q\) should not be trusted and the values of the fit parameters are probably meaningless in these cases. For the case of a straight line fit, the inverse of \(U\) is given explicitly in Eq. ([\[Uinv\]](#Uinv){reference-type="ref" reference="Uinv"}). Using this information, and the values of \((x_i, y_i, \sigma_i)\) for the data in Fig. [\[fig:slinefit\]](#fig:slinefit){reference-type="ref" reference="fig:slinefit"}, the fit parameters (assuming a straight line fit) are \[\begin{aligned} a_0 &= 0.84 \pm 0.32, \\ a_1 &= 2.05 \pm 0.11, \end{aligned}\] in which the error bars on the fit parameters on \(a_0\) and \(a_1\), which are denoted by \(\sigma_0\) and \(\sigma_1\), are determined from Eq. ([\[error_params\]](#error_params){reference-type="ref" reference="error_params"}). The data was generated by starting with \(y = 1 + 2x\) and then adding some noise with zero mean. Hence the fit should be consistent with \(y = 1 +2x\) within the error bars, and it is. The value of \(\chi^2\) is 7.44 so \(\chi^2/N_\text{DOF} = 7.44 / 9 = 0.866\) and the quality of fit parameter, given by Eq. [\[Q_expression\]](#Q_expression){reference-type="eqref" reference="Q_expression"}, is \(Q = 0.592\) which is good. We call \(U^{-1}\) the "*covariance matrix*". Its off-diagonal elements are also useful since they contain information about correlations between the fitted parameters. More precisely, one can show, following the lines of the above derivation of \(\sigma_\alpha^2\), that the correlation of fit parameters \(\alpha\) and \(\beta\), known mathematically as their "covariance", is given by the appropriate off-diagonal element of the covariance matrix, \[\text{Cov}(\alpha, \beta) \equiv \langle \delta a_\alpha \, \delta a_\beta \rangle = \left(U^{-1}\right)_{\alpha\beta} \,. \label{Covab}\] The correlation coefficient, \(r_{\alpha\beta}\), which is a dimensionless measure of the correlation between \(\delta a_\alpha\) and \(\delta a_\beta\) lying between \(-1\) and 1, is given by \[r_{\alpha\beta} = {\text{Cov}(\alpha, \beta) \over \sigma_\alpha \sigma_\beta} \,. \label{rab}\] A good fitting program should output the correlation coefficients as well as the fit parameters, their error bars, the value of \(\chi^2/N_\text{DOF}\), and the goodness of fit parameter \(Q\). For a linear model, \(\chi^2\) is a quadratic function of the fit parameters and so the elements of the "*curvature matrix*"[^4], \((1/2)\, \partial^2 \chi^2 / \partial {a_\alpha}\partial {a_\beta}\) are constants, independent of the values of the fit parameters. In fact, we see from Eqs. [\[Uab\]](#Uab){reference-type="eqref" reference="Uab"} and [\[chisq_poly\]](#chisq_poly){reference-type="eqref" reference="chisq_poly"} that \[{1\over 2}\, { \partial^2 \chi^2 \over \partial {a_\alpha} \partial {a_\beta}} = U_{\alpha \beta} \,, \label{curv}\] so *the curvature matrix is equal to \(U\)*, given by Eq. [\[Uab\]](#Uab){reference-type="eqref" reference="Uab"} for a polynomial fit. If we fit to a *general* linear model, writing \[f(x) = \sum_{\alpha=1}^M a_\alpha \, X_\alpha(x), \label{general_lin}\] where \(X_1(x), X_2(x), \cdots, X_M(x)\) a fixed functions of \(x\) called basis functions, the curvature matrix is given by \[\boxed{ U_{\alpha\beta} = \sum_{i=1}^N {X_\alpha(x_i)\, X_\beta(x_i) \over \sigma_i^2} \,.} \label{Uab_general}\] Similarly, the quantities \(v_\alpha\) in Eq. [\[v\]](#v){reference-type="eqref" reference="v"} become \[\boxed{ v_\alpha = \sum_{i=1}^N {y_i\, X_\alpha(x_i) \over \sigma_i^2} \, ,} \label{v_general}\] for a general set of basis functions, and best fit parameters are given by the solution of the \(M\) linear equations \[\boxed{ \sum_{\beta=1}^M U_{\alpha\beta}\, a_\beta = v_\alpha \,, } \label{lin_eq}\] for \(\alpha= 1, 2, \cdots, M\). Note that for a linear model the curvature matrix \(U\) is a constant, independent of the fit parameters. However, \(U\) is not constant for a non-linear model. ## Fitting to a non-linear model {#sec:nlmodel} As for linear models, one minimizes \(\chi^2\) in Eq. [\[chisq\]](#chisq){reference-type="eqref" reference="chisq"}. The difference is that the resulting equations are non-linear so there might be many solutions or non at all. Techniques for solving the coupled non-linear equations invariably require specifying an initial value for the variables \(a_\alpha\). The most common method for fitting to non-linear models is the Levenberg-Marquardt (LM) method, see e.g. Numerical Recipes . Implementing the Numerical Recipes code for LM is a little complicated because it requires the user to provide a routine for the derivatives of \(\chi^2\) with respect to the fit parameters as well as for \(\chi^2\) itself, and to check for convergence. Alternatively, one can use the fitting routines in the `scipy` package of `python` or use `gnuplot`. But see the comments in Appendix [\[sec:ase\]](#sec:ase){reference-type="ref" reference="sec:ase"} about getting the error bars in the parameters correct. This applies when fitting to linear as well as non-linear models. Gnuplot and scipy scripts for fitting to a non-linear model are given in Appendix [\[sec:scripts\]](#sec:scripts){reference-type="ref" reference="sec:scripts"}. One difference from fitting to a linear model is that the curvature matrix, defined by the LHS of Eq. [\[curv\]](#curv){reference-type="eqref" reference="curv"}, is not constant but is a function of the fit parameters. Hence it is no longer true that the standard deviations of the two distributions in Fig. [\[Fig:distofa1\]](#Fig:distofa1){reference-type="ref" reference="Fig:distofa1"} are equal. However, it still generally assumed that the difference is small enough to be unimportant and hence that the covariance matrix, which is now defined to be the inverse of the curvature matrix *at the minimum of \(\chi^2\)*, still gives information about error bars on the fit parameters. This is discussed more in the next two subsections, in which we point out, however, that a more detailed analysis is needed if the model is non-linear and the spread of fitted parameters is sufficiently large that it cannot be represented by an effective linear model, i.e. \(\chi^2\) is not well fitted by a parabola over the needed range of parameter values. As a reminder: - The *curvature matrix* is defined in general by the LHS of Eq. [\[curv\]](#curv){reference-type="eqref" reference="curv"}, which, for a linear model, is equivalent to Eq. [\[Uab_general\]](#Uab_general){reference-type="eqref" reference="Uab_general"} (Eq. [\[Uab\]](#Uab){reference-type="eqref" reference="Uab"} for a polynomial fit.) - The *covariance matrix* is the inverse of the curvature matrix at the minimum of \(\chi^2\) (the last remark being only needed for a non-linear model). Its diagonal elements give error bars on the fit parameters according to Eq. [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"} (but see the caveat in the previous paragraph for non-linear models) and its off-diagonal elements give correlations between fit parameters according to Eqs. [\[Covab\]](#Covab){reference-type="eqref" reference="Covab"} and [\[rab\]](#rab){reference-type="eqref" reference="rab"}. ## Confidence limits {#sec:conf_limits} In the last two subsections we showed that the diagonal elements of the covariance matrix give an error bar on the fit parameters. In this section we extend the notion of error bar to embrace the concept of a "confidence limit". There is a theorem  which states that, for a linear model, if we take simulated data sets assuming Gaussian noise in the data about the actual data points, and compute the fit parameters \(\vec{a}^{S(i)}, i = 1, 2, \cdots\) for each data set, then the probability distribution of the \(\vec{a}^S\) is given by the multi-variable Gaussian distribution \[\boxed{ P(\vec{a}^S) \propto \exp\left(-{1 \over 2} \, \sum_{\alpha, \beta} \delta a_\alpha^S\, U_{\alpha\beta}\, \delta a_\beta^S \right) \, ,} \label{theorem}\] where \(\delta \vec{a}^S \equiv \vec{a}^{S(i)}-\vec{a}^{(0)}\) and \(U\), given by Eq. [\[Uab_general\]](#Uab_general){reference-type="eqref" reference="Uab_general"}, is the curvature matrix which can also be defined in terms of the second derivative of \(\chi^2\) according to Eq. [\[curv\]](#curv){reference-type="eqref" reference="curv"}. A proof of this result is given in Appendix [\[sec:proof\]](#sec:proof){reference-type="ref" reference="sec:proof"}. It applies for a linear model with Gaussian noise, and also for a non-linear model if the uncertainties in the parameters do not extend outside a region where an effective linear model could be used. (In the latter case one still needs a non-linear routine to *find* the best parameters). Note that for a non-linear model, \(U\) is not a constant and is the curvature *at the minimum* of \(\chi^2\). From Eq. [\[curv\]](#curv){reference-type="eqref" reference="curv"} the change in \(\chi^2\) as the parameters are varied away from the minimum is given by \[\Delta \chi^2 \equiv \chi^2(\vec{a}^{S(i)})-\chi^2(\vec{a}^{(0)}) = \sum_{\alpha, \beta} \delta a_\alpha^S\, U_{\alpha\beta}\, \delta a_\beta^S \,, \label{Dchisq}\] in which the \(\chi^2\) are all evaluated from the single (actual) data set \(y_i^{(0)}\). Equation [\[theorem\]](#theorem){reference-type="eqref" reference="theorem"} can therefore be written as \[P(\vec{a}^S) \propto \exp\left(-{1 \over 2} \Delta \chi^2 \right) \,. \label{P_dalpha}\] We remind the reader that we have assumed the noise in the data is Gaussian and that either the model is linear or, if non-linear, the uncertainties in the parameters do not extend outside a region where an effective linear model could be used. Hence the probability of a particular deviation, \(\delta \vec{a}^S\), of the fit parameters in a simulated data set away from the parameters in the *actual* data set, depends on how much this change increases \(\chi^2\) (evaluated from the actual data set) away from the minimum. In general a "confidence limit" is the range of fit parameter values such that \(\Delta \chi^2\) is less than some specified value. The simplest case, and the only one we discuss here, is the variation of *one* variable at a time, though multi-variate confidence limits can also be defined, see Numerical Recipes . We therefore consider the change in \(\chi^2\) when one variable, \(a_1^S\) say, is held at a specified value, and all the others \((\beta = 2, 3,\cdots, M)\) are varied in order to minimize \(\chi^2\). Minimizing \(\Delta \chi^2\) in Eq. [\[Dchisq\]](#Dchisq){reference-type="eqref" reference="Dchisq"} with respect to \(a_\beta^S\) gives \[\sum_{\gamma=1}^M U_{\beta\gamma}\, \delta a_\gamma^S = 0, \qquad (\beta = 2, 3, \cdots,M) \,.\] The corresponding sum for \(\beta = 1\), namely \(\sum_{\gamma=1}^M U_{1\gamma}\, \delta a_\gamma^S\), is not zero because \(\delta a_1\) is fixed. It will be some number, \(c\) say. Hence we can write \[\sum_{\gamma=1}^M U_{\alpha\gamma}\, \delta a_\gamma^S = c_\alpha, \qquad (\alpha = 1, 2, \cdots,M) \,,\] where \(c_1 = c\) and \(c_\beta = 0\, (\beta \ne 1)\). The solution is \[\delta a_\alpha^S = \sum_{\beta=1}^M \left(U^{-1}\right)_{\alpha\beta} c_\beta \,. \label{aalpha}\] For \(\alpha = 1\) this gives \[c = \delta a_1^S / \left(U^{-1}\right)_{11} \,. \label{c}\] Substituting Eq. [\[aalpha\]](#aalpha){reference-type="eqref" reference="aalpha"} into Eq. [\[Dchisq\]](#Dchisq){reference-type="eqref" reference="Dchisq"}, and using Eq. [\[c\]](#c){reference-type="eqref" reference="c"} we find that \(\Delta \chi^2\) is related to \(\left(\delta a_1^S\right)^2\) by \[\Delta \chi^2 = {(\delta a_1^S)^2 \over \left(U^{-1}\right)_{11} }. \label{Dchi2}\] (Curiously, the coefficient of \((\delta a_1)^2\) is one over the \(11\) element of the inverse of \(U\), rather than \(U_{11}\) which is how it appears in Eq. [\[Dchisq\]](#Dchisq){reference-type="eqref" reference="Dchisq"} in which the \(\beta \ne 1\) parameters are free rather than adjusted to minimize \(\chi^2\).) From Eq. [\[P_dalpha\]](#P_dalpha){reference-type="eqref" reference="P_dalpha"} we finally get \[P(a_1^S) \propto \exp\left(-{1 \over 2} \, {(\delta a_1^S)^2 \over\sigma_1^2}\right) \,, \label{Pa1S}\] where \[\sigma_1^2 = \left(U^{-1}\right)_{11} \,.\] As shown in Appendices [\[sec:proof\]](#sec:proof){reference-type="ref" reference="sec:proof"} and [\[sec:proof2\]](#sec:proof2){reference-type="ref" reference="sec:proof2"}, Eqs. [\[theorem\]](#theorem){reference-type="eqref" reference="theorem"},  [\[P_dalpha\]](#P_dalpha){reference-type="eqref" reference="P_dalpha"} and [\[Pa1S\]](#Pa1S){reference-type="eqref" reference="Pa1S"} also apply, under the same conditions (linear model and Gaussian noise on the data) to the probability for \(\delta a_1 \equiv a_1^\text{true}-a_1^{(0)}\), where we remind the reader that \(a_1^{(0)}\) is the fit parameter obtained from the actual data, and \(a_1^\text{true}\) is the exact value. In other words the probability of the true value is given by \[\boxed{ P(\vec{a}^\text{true}) \propto \exp\left(-{1 \over 2} \Delta \chi^2 \right) \, ,} \label{P_dalphatrue}\] where \[\Delta \chi^2 \equiv \chi^2(\vec{a}^\text{true})-\chi^2(\vec{a}^{(0)}) \,,\] in which we remind the reader that both values of \(\chi^2\) are evaluated from the single set of data available to us, \(y_i^{(0)}\). Projecting onto a single parameter, as above, gives \[\boxed{ P(a_1^\text{true}) \propto \exp\left(-{1\over 2}\, {(\delta a_1)^2 \over \sigma_1^2}\right) \,, } \label{Pa1}\] so \(\langle \left(\delta a_1\right)^2 \rangle = \sigma_1^2 = \left(U^{-1}\right)_{11}\), in agreement with what we found earlier in Eq. [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"}. We emphasize that Eqs. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"} and  [\[Pa1\]](#Pa1){reference-type="eqref" reference="Pa1"} assumes Gaussian noise on the data points, and either the model is linear or, if non-linear, that the range of uncertainty in the parameters is small enough that a description in terms of an effective linear model is satisfactory. However we have done more than recover our earlier result, Eq. [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"}, by more complicated means since we have gained *additional* information. From the properties of a Gaussian distribution we now know that, from Eq. [\[Pa1\]](#Pa1){reference-type="eqref" reference="Pa1"}, the probability that \(a_\alpha\) lies within one standard deviation \(\sigma_\alpha\) of the value which minimizes \(\chi^2\) is 68%, the probability of its being within two standard deviations is 95.5%, and so on. Furthermore, from Eq. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"}, we see that > *if a single fit parameter is one standard deviation away from its value at the minimum of \(\chi^2\) (the other fit parameters being varied to minimize \(\chi^2\)), then \(\Delta \chi^2 = 1\).* This last sentence, and the corresponding equations Eqs. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"} and [\[Pa1\]](#Pa1){reference-type="eqref" reference="Pa1"}, are not valid for a non-linear model if the uncertainties of the parameters extends outside the range where an effective linear model can be used. In this situation, to get confidence limits, is is necessary to do a bootstrap resampling of the data, as discussed in the next subsection. However, if one is not able to resample the data we argue that it is better to take the range where \(\Delta \chi^2 \le 1\) as an error bar for each parameter rather than the error bar determined from the curvature of \(\chi^2\) at the minimum, see Fig. [\[fig:chi2\]](#fig:chi2){reference-type="ref" reference="fig:chi2"}. The left hand plot is for a linear model, for which the curve of \(\Delta \chi^2\) against \(\delta a_1\) is exactly a parabola, and the right hand plot is a sketch for a non-linear model, for which it is not a parabola though it has a quadratic variation about the minimum shown by the dashed curve. For the linear case, the values of \(\delta a_1\) where \(\Delta \chi^2 = 1\) are the *same* as the values \(\pm \sigma_1\), where \(\sigma_1\) is the standard error bar obtained from the *local* curvature in the vicinity of the minimum. However, for the non-linear case, the values of \(\delta a_1\) where \(\Delta \chi^2 = 1\) are *different* from \(\pm \sigma_1\), and indeed the values on the positive and negative sides, \(\sigma_1^+\) and \(\sigma_1^-\), are not equal. For the data Fig. [\[fig:chi2\]](#fig:chi2){reference-type="ref" reference="fig:chi2"}, it is clear that the value of \(a_1\) is more tightly constrained on the positive side than the negative side, and so it is better to give the error bars as \(+\sigma_1^+\) and \(-\sigma_1^-\), obtained from the range where \(\Delta \chi^2 \le 1\), rather the symmetric range \(\pm \sigma_1\). However, if possible, in these circumstances error bars and a confidence limit should actually be obtained from a bootstrap resampling of the data as discussed in the next section. ## Confidence limits by resampling the data {#sec:resample} More work is involved if one wants to get error bars and a confidence interval in the case where the model is non-linear and the range of parameter uncertainty extends outside the region where an effective linear model is adequate. Even for a linear model, we cannot convert \(\Delta \chi^2\) into a confidence limit with a specific probability if the noise is non-Gaussian. To proceed in these cases, one can bootstrap the individual data points as follows. Each data point \((x_i, y_i)\) has error bar \(\sigma_i\), which comes from averaging over \(N\) measurements, say. Generating bootstrap datasets by Monte Carlo sampling the \(N\) measurements, as discussed in Sec. [2.2.3](#sec:boot){reference-type="ref" reference="sec:boot"}, the distribution of the mean of each bootstrap dataset has a standard deviation equal to the estimate of standard deviation on the mean of the actual data set, see Eq. [\[sigmafb\]](#sigmafb){reference-type="eqref" reference="sigmafb"} (replacing the factor of \(\sqrt{N/(N-1)}\) by unity which is valid since \(N\) is large in practice). Hence, if we generate \(N_\text{boot}\) bootstrap data sets, and fit each one, the scatter of the fitted parameter values will be a measure of the uncertainty in the values from the *actual* dataset. Forming a histogram of the values of a single parameter we can obtain a confidence interval within which 68%, say, of the bootstrap datasets lie (16% missing on either side) and interpret this range as a 68% confidence limit for the actual parameter value. The justification for this interpretation has been discussed in the statistics literature, see e.g. the references in Ref.  but I'm not able to go into the details here. Note that this bootstrap approach could also be applied usefully for a *linear* model if the noise is not Gaussian. Unfortunately, use of the bootstrap procedure to get error bars in fits to non-linear models does not yet seem to be a standard procedure in the statistical physics community. Another possibility for a non-linear model, if one is confident that the noise is close to Gaussian, is to generate *simulated* data sets, assuming Gaussian noise on the \(y_i\) values with standard deviation given by the error bars \(\sigma_i\). Each simulated dataset is fitted and the distribution of fitted parameters is determined. This corresponds to the analytical approach in Appendix [\[sec:proof\]](#sec:proof){reference-type="ref" reference="sec:proof"} but without the assumption that the model can be represented by an effective linear one over of the needed parameter range. ## A tale of two probabilities. When can one rule out a fit? {#sec:lin_or_quad} If the noise on the data is Gaussian, which we will assume throughout this subsection, we have, so far, considered two different probabilities. Firstly, as discussed in Appendix [\[sec:Q\]](#sec:Q){reference-type="ref" reference="sec:Q"}, the value of \(\chi^2\) is typically in the range \(N_\text{DOF} \pm \sqrt{2 N_\text{DOF}}\). The quality of fit parameter \(Q\) is the probability that, *given the fit*, the data could have this value of \(\chi^2\) or greater, and is given mathematically by Eq. [\[Q_expression\]](#Q_expression){reference-type="eqref" reference="Q_expression"}. It varies from unity when \(\chi^2 \ll N_\text{DOF}-\sqrt{2 N_\text{DOF}}\) to zero when \(\chi^2 \gg N_\text{DOF} + \sqrt{2 N_\text{DOF}}\). We emphasize that Secondly, in the context of error bars and confidence limits, we have discussed, in Eqs. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"} and [\[Pa1\]](#Pa1){reference-type="eqref" reference="Pa1"}, the probability that a fit parameter, \(a_1\) say, takes a certain value relative to the optimal one. Equation [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"} becomes very small when \(\Delta \chi^2\) varies by much more than unity. Note that Eqs. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"} and [\[Pa1\]](#Pa1){reference-type="eqref" reference="Pa1"} refer to the At first, it seems curious that the probability \(Q\) remains significantly greater than zero if \(\chi^2\) changes by an amount of order \(\sqrt{N_\text{DOF}}\), whereas if a fit parameter is changed by an amount such that \(\chi^2\) changes by of order \(\sqrt{N_\text{DOF}}\), the probability of this value becomes extremely small, of order \(\exp(-\text{const.}\,\sqrt{N_\text{DOF}})\), in this limit, see Eqs. [\[P_dalphatrue\]](#P_dalphatrue){reference-type="eqref" reference="P_dalphatrue"}. While there is no mathemematical inconsistency, since the two probabilities refer to different situations (one is the probability of the data given the fit and the other is the relative probability of two fits given the data), it is useful to understand this difference intuitively. We take, as an example, a problem where we want to know whether the data can be modeled by a straight line, or whether a quadratic term needs to be included as well. A set of data is shown in Fig. [\[fig:lin_or_quad\]](#fig:lin_or_quad){reference-type="ref" reference="fig:lin_or_quad"}. Looking at the left figure one sees that the data more or less agrees with the straight-line fit. However, one also sees systematic trends: the data is too high for small \(x\) and for high \(x\), and too low for intermediate \(x\). Chi-squared just sums up the contributions from each data point and is insenstive to any systematic trend in the deviation of the data from the fit. Hence the value of \(\chi^2\), in itself, does not tell us that this data is unlikely to be represented by a straight line. It is only when we add another parameter in the fit which corresponds to those correlations, that we realize the straight-line model is relatively very unlikely. In this case, the extra parameter is the coefficient of \(x^2\), and the resulting parabolic fit is shown in the right figure. The qualitative comments in the last paragraph are made more precise by the parameters of the fits. The straight-line fit gives \(a_0 = 0.59 \pm 0.26, a_1 = 2.003 \pm 0.022\) with \(Q = 0.124\), whereas the parabolic fit gives \(a_0 = 2.04 \pm 0.40, a_1 = 1.588 \pm 0.090, a_2 = 0.0203 \pm 0.0042\) with \(Q = 0.924\). The actual parameters used to generate the data are \(a_0 = 2, a_1 = 1.6, a_2 = 0.02\), and there is Gaussian noise with standard deviation equal to \(0.8\). Although the quality of fit factor for the straight-line fit is reasonable, the quadratic fit strongly excludes having the fit parameter \(a_2\) equal to zero, since zero is five standard deviations away from the best value. For a Gaussian distribution, the probability of a five-sigma deviation or greater is \(\text{erfc}(5/\sqrt{2}) \simeq 6 \times 10^{-7}\). The difference in \(\chi^2\) for the quadratic fit, between the best fit and the fit forcing \(a_2 = 0\), is \((0.0203 / 0.0042)^2 \simeq 23\) according to Eqs. [\[Dchi2\]](#Dchi2){reference-type="eqref" reference="Dchi2"} and [\[error_params\]](#error_params){reference-type="eqref" reference="error_params"}. We conclude that, in this case, the straight-line model is unlikely to be correct. The moral of this tale is that a reasonable value of \(Q\) does not, in itself, ensure that you have the right model. Another model might be very much more probable. [^1]: The factor of \(N-1\) rather than \(N\) in the expression for the sample variance in Eq. ([\[sigmafromdata\]](#sigmafromdata){reference-type="ref" reference="sigmafromdata"}) needs a couple of comments. Firstly, the final answer for the error bar on the mean, Eq. [\[finalans2\]](#finalans2){reference-type="eqref" reference="finalans2"} below, will be independent of how this intermediate quantity is defined. Secondly, the \(N\) terms in Eq. ([\[sigmafromdata\]](#sigmafromdata){reference-type="ref" reference="sigmafromdata"}) are not all independent since \(\overline{x}\), which is itself given by the \(x_i\), is subtracted. Rather, as will be discussed more in the section on fitting, Sec. [3](#sec:fit){reference-type="ref" reference="sec:fit"}, there are really only \(N-1\) independent variables (called the "number of degrees of freedom" in the fitting context) and so dividing by \(N-1\) rather than \(N\) has a rational basis. However, this is not essential and many authors divide by \(N\) in their definition of the sample variance. [^2]: \(\chi^2\) should be thought of as a single variable rather than the square of something called \(\chi\). This notation is standard. [^3]: Although this result is only valid if the fitting model is linear in the parameters, it is usually taken to be a reasonable approximation for non-linear models as well. [^4]: It is conventional to include the factor of \(1/2\).

{'timestamp': '2012-10-16T02:03:08', 'yymm': '1210', 'arxiv_id': '1210.3781', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3781'}

# Introduction {#sec1} Interference Alignment (IA) has been a focus of intense research on Gaussian interference channels (GICs) in the recent years on the account of the capacity of interference channels being unknown in general and the potential of IA to get close to the sum-capacity for a broad class of interference channels as the signal to noise ratio (\(SNR\)) tends to infinity. The scaling of the sum-capacity with \(log ~SNR\) is known as the sum degrees of freedom (DoF) of the GIC. The sum-DoF is known to be \(K/2\), with probability \(1\), for the K-user GIC when all the nodes are equipped with a single antenna and time-varying channel gains are assumed. The result was proved by using linear precoding at the transmitters over an arbitrarily large number of symbol extensions and zero forcing at the receivers. In a later work, the sum-DoF is shown to be \(K/2\) with probability \(1\) even in the case of constant channel-coefficients (that are drawn from a continuous distribution), i.e., the channel gains do not vary with time, using a non-linear IA technique. When all the nodes are equipped with \(M\) antennas, the sum-DoF was shown to be \(3M/2\) for the \(3\)-user GIC. This result was proved by linear precoding over the transmit antennas without the use of symbol extensions and holds true even in the case of constant channel coefficients. Later, with the assumption of constant channel coefficients and using a non-linear IA technique, the sum-DoF when \(K\geq \frac{M+N}{\gcd(M,N)}\) was found to be equal to \(\frac{MN}{M+N} K\) where, \(\gcd(M,N)\) denotes the greatest common divisor of \(M\) and \(N\), with \(M\) being the number of antennas at each transmitter and \(N\) being the number of receive antennas at each receiver. All the works cited above assumed full channel state information at all the transmitters (CSIT) and receivers (CSIR). The notion of sum-DoF involves scaling of sum-rate as \(log ~SNR\) at high SNR and therefore, Gaussian input alphabets or lattice codes are always used in the study of sum-DoF. However, in all practical scenarios finite constellations like \(M\)-QAM and \(M\)-PSK are used at the inputs. *With the constraint of finite constellation inputs, it is not known whether IA is optimal in some sense*. Linear precoding for optimizing the mutual information between the input and the output has been studied for the single user MIMO channel with finite constellation inputs in-. Constellation rotation for optimizing the sum-capacity for SISO Multiple Access Channel (MAC) with finite constellation inputs has been examined in and linear precoding for weighted sum-rate maximization in MIMO MAC with finite constellation inputs has been studied in. Note that linear precoding for the SISO MAC corresponds to constellation rotation at the transmitter. Recently, there has been some progress on the analysis of finite constellation effects in \(2\)-user SISO GIC ,. In, constellation rotation was found to increase the constellation constrained sum-capacity of \(2\)-user SISO Gaussian strong interference channel, and in, a metric to find the optimum angle of rotation was proposed. In this paper, we examine achievable rate-tuples with linear precoding for \(K\)-user MIMO Gaussian Interference Channel (GIC) with finite constellation inputs. Specifically, we treat interference as noise, i.e., each transmitter reveals its codebook only to its intended receiver. The maximum rate achievable under such a circumstance for transmitter-\(i\) (Tx-\(i\)) is given by mutual information between the input generated by Tx-\(i\) and the output at receiver Rx-\(i\). The channel conditions and values of SNR under which the decoding scheme of treating interference as noise with Gaussian alphabet inputs is sum-capacity optimal was found for the \(2\)-user SISO GIC in-, for the \(K\)-user SISO GIC in and for the \(2\)-user MIMO GIC in. For given values of channel gains, with Gaussian input alphabets, as the SNR tends to infinity, treating interference as noise is not sum-capacity optimal. *With the constraint of finite constellation inputs, it is not clear whether treating interference as noise is optimal in some sense.* First, we need to define a notion of optimality under the constraint of fixed finite constellation inputs and then analyse decoding and transmit schemes with that notion of optimality. Consider a scenario where each transmitter-\(i\) (Tx-\(i\)), requires to send \(d_i\) independent complex symbols per channel use that take values from fixed finite constellations with uniform distribution, for \(i=1,2,\cdots ,K\), to receiver-\(i\) (Rx-\(i\)). Throughout this paper, we assume that none of the direct channel gains are zero. For a \(K\)-user MIMO GIC with finite constellation inputs, as a measure of optimality of linear precoding in the high SNR sense, we introduce the notion of *Constellation Constrained Saturation Capacity* (CCSC) which is defined as follows. For the ease of exposition, throughout the paper, we assume that the constellations used for the symbols are all the same at all the transmitters, and is of cardinality \(M\). Hence, the CCSC for Tx-\(i\) is given by \(log_2M^{ d_i}\). In this paper, with the assumption of constant \(K\)-user MIMO GIC with full global knowledge of channels gains, and finite constellation inputs, we derive a set of necessary and sufficient conditions on the precoders under which treating interference as noise at Rx-\(i\) will achieve a rate for Tx-\(i\) that tends to CCSC for Tx-\(i\), for all \(i\), as SNR tends to infinity. Precoders satisfying these necessary and sufficient conditions exist for all direct and cross channel gains, and are termed as *CCSC optimal precoders*. Hence, in the case of finite constellation inputs with the use of appropriate precoders, the rate tuples obtained by treating interference as noise tend to values that are independent of the channel gains. For a \(K\)-user SISO GIC, this result is in contrast with the Gaussian input alphabet case where the rate tuples obtained by treating interference as noise tend to values dictated by the channel gains, as the SNR tends to infinity. Interestingly, the precoders that achieve IA, if feasible, are also CCSC optimal precoders. However, finding precoders that align interference is known to be NP-hard in general whereas, the precoders that satisfy the derived necessary and sufficient conditions are easy to find for any given channel-coefficients. Since finite SNR is of more practical interest, we propose a gradient-ascent based algorithm to optimize the precoders for the sum-rate achieved by treating interference as noise. The contributions of the paper are summarized below. - For a constant \(K\)-user MIMO GIC using finite constellation inputs with precoding, a high SNR approximation for the rate tuples achieved by treating interference as noise at the receivers is derived (see Theorem [\[thm1\]](#thm1){reference-type="ref" reference="thm1"} in Section [3](#sec3){reference-type="ref" reference="sec3"}). Based on this approximation, we derive a set of necessary and sufficient conditions under which the precoders are CCSC optimal (see Theorem [\[thm2\]](#thm2){reference-type="ref" reference="thm2"} Section [3](#sec3){reference-type="ref" reference="sec3"}). These conditions are satisfied with probability \(1\) when the entries of the precoders are chosen from any continuous distribution. It is observed that the precoders that achieve IA, if feasible, are CCSC optimal.\ - For the finite SNR case, we propose a gradient-ascent based algorithm to improve the sum-rate achieved by treating interference as noise using finite constellation inputs with precoding. Simulation studies indicate considerable improvement in the ergodic sum-rate with precoders obtained using the proposed algorithm for a \(3\)-user MIMO GIC with \(2\) antennas at all the nodes, \(d_i=1\), for all \(i\), and QPSK inputs, at low and moderate SNRs over that obtained using the IA solution of. The paper is organized as follows. The system model is formally introduced in Section [2](#sec2){reference-type="ref" reference="sec2"}. In Section [3](#sec3){reference-type="ref" reference="sec3"}, a set of necessary and sufficient conditions for CCSC optimal precoders is derived. In Section [4](#sec4){reference-type="ref" reference="sec4"}, our gradient-ascent based algorithm for optimizing the sum-rate using precoders and treating interference as noise at the receivers is given, and a simulation result comparing its performance with respect to that of precoders satisfying IA is presented. Section [5](#sec5){reference-type="ref" reference="sec5"} concludes the paper. *Notations:* For a random variable \(X\) which takes value from the set \(\cal X\), we assume some ordering of its elements and use \(x^i\) to represent the \(i\)-th element of \(\cal X\). Realization of the random variable X is denoted as \(x\). The notation \(diag(V_1,V_2,\cdots,V_n)\) denotes a block diagonal matrix formed by the matrices \(V_i\), \(i=1,2,\cdots,n\). The \(i^{\text{th}}\) coordinate of a complex vector \(X\) is denoted by \(X(i)\). The \(2\)-norm of a complex vector \(X\) is denoted by \(||X||\). The cardinality of a set \({\cal X}\) is denoted by \(|\cal X|\). For a complex number \(a\), \(\Re\{a\}\) and \(\Im\{a\}\) denote the real and imaginary parts of \(a\) respectively. For two complex numbers \(a\) and \(b\), the notation \(a>b\) denotes that \(|\Re\{a\}|>|\Re\{b\}|\) and \(|\Im\{a\}|>|\Im\{b\}|\). The notation \(\underline{0}\) represents the zero vector whose size will be clear from the context. All the logarithms in the paper are to the base \(2\). # System Model {#sec2} Tx-\(i\) intends to communicate with Rx-\(i\), for \(i=1,2,\cdots, K\), as shown in Fig. [\[fig_sys_model\]](#fig_sys_model){reference-type="ref" reference="fig_sys_model"}. Without loss of generality, let \(P\) denote the power constraint at all the transmitters. The signal received at Rx-\(j\) is given by \[\begin{aligned} Y_j=\sum_{i=1}^{K}\sqrt{P}H_{ij}V_iX_i + N_j \end{aligned}\]where, \(H_{ij}\) denotes the constant channel matrix from Tx-\(i\) to Rx-\(j\), \(V_i\) denotes the precoder at Tx-\(i\), \(X_i\) denotes the complex symbol vector generated at Tx-\(i\), \(N_j\) denotes the noise random vector whose coordinates represent independent and identically distributed zero mean unit variance circularly symmetric complex Gaussian random variables. The sizes of the matrices \(H_{ij}\), \(V_i\), \(X_i\), and \(N_j\) are given by \(n_{r_j}\times n_{t_i}\), \(n_{t_i}\times {d_i}\), \({d_i} \times 1\), and \(n_{r_j}\times 1\) respectively, where \(n_{r_j}\) and \(n_{t_i}\) denote the number of receive and transmit antennas at Rx-\(j\) and Tx-\(i\) respectively, and \({d_i}\) denotes the number of independent complex symbols per channel use that Tx-\(i\) wants to transmit to Rx-\(i\). These complex symbols are assumed to take values from finite constellations with uniform distribution over its elements. For simplicity of exposition, we assume that the finite constellations used are all the same at all the transmitters and are of cardinality \(M\). The results of this paper apply with simple modifications when this is not the case. The finite constellation used is denoted by \({\cal S}\) and is of unit power. # CCSC Optimal Precoders {#sec3} In this section, we shall derive a set of necessary and sufficient on the precoders for CCSC optimality which is taken to be a measure of optimality for linear precoding in the high SNR regime for the finite constellation input case. Rate achievable for Tx-\(i\) by treating interference as noise at Rx-\(i\), for all \(i\), is given by \(R_i<I[X_i;Y_i]\). Our focus will be on the boundary point given by \(I[X_i;Y_i]\), for all \(i\). Let \(X=[X_1 ~X_2 ~\cdots ~X_K]^T\). The effective channel matrix from all the transmitters to Rx-\(i\) is given by \(H_i=[H_{1i} ~H_{2i} \cdots ~H_{Ki}]\). Define \[\begin{aligned} V = diag(V_1,V_2,\cdots,V_K). \end{aligned}\] Using the chain rule for mutual information, \[\begin{aligned} I[X_i;Y_i]=I[X_1,X_2,\cdots,X_K;Y_i]-I[X_1,X_2,\cdots,X_K;Y_i|X_i]. \end{aligned}\]With uniform distribution assumed over the elements of the constellation, the expression for \(I[X_1,X_2,\cdots,X_K;Y_i]\) is given by ([\[MI_exp\]](#MI_exp){reference-type="ref" reference="MI_exp"}) (at the top of the next page) which is derived in a similar way as in. Define the matrix \[\begin{aligned} A^{k_1,k_2}=H_iV\left(x^{k_1}-x^{k_2} \right). \end{aligned}\] The following theorem gives a high SNR approximation for \(I[X_1,X_2,\cdots,X_K;Y_i]\). Let \(X_{\not{\hspace{0.05cm}i}}=[X_1 ~X_2 ~\cdots ~X_{i-1} ~X_{i+1} ~\cdots ~X_K]^T\). The channel matrix from all the transmitters, with the exclusion of Tx-\(i\), to Rx-\(i\) is given by \[\begin{aligned} H_{\not{\hspace{0.05cm}i}}=[H_{11} ~H_{12} ~\cdots~ H_{i-1, i} ~H_{i+1, i} ~\cdots~ H_{Ki}]. \end{aligned}\]Define the matrices \[\begin{aligned} &V_{\not{\hspace{0.05cm}i}}=\text{diag}(V_1,V_2,\cdots,V_{i-1},V_{i+1},\cdots,V_K).\\ &B^{i_1,i_2}=H_{\not{\hspace{0.05cm}i}}V_{\not{\hspace{0.05cm}i}}\left(x_{\not{\hspace{0.05cm}i}}^{i_1}-x_{\not{\hspace{0.05cm}i}}^{i_2} \right), \end{aligned}\]for \(i_1,i_2=0,1,\cdots, M^{\sum_{j \neq i}d_j}-1\). Similar to ([\[eqn_approx\]](#eqn_approx){reference-type="ref" reference="eqn_approx"}), we have the following approximation for \(I[X_1,X_2,\cdots,X_K;Y_i|X_i]\) at high \(P\). \[\begin{aligned} \label{eqn_approx2} &log~M^{\sum_{\substack{{j\neq i}\\{j=1}}}^{K}d_i} \\ \nonumber &-\frac{1}{ M^{\sum_{j \neq i}d_j}}\hspace{-0.2cm}\sum_{i_1 = 0}^{ M^{\sum_{j \neq i}d_j}-1}\left[ log \left( \sum_{i_2 = 0}^{ M^{\sum_{j \neq i}d_j}-1}e^{-\left|\left|\sqrt P B^{i_1,i_2}\right|\right|^2 }\right) \right] \end{aligned}\]Hence, a high SNR approximation for \(I[X_i;Y_i]\) is given by \[\begin{aligned} \nonumber &I[X_i;Y_i] \approx log~M^{d_i} \\ \label{actual_approx} &-\frac{1}{M^{\sum_{i=1}^{K}d_i}}\hspace{-0.2cm}\sum_{k_1 = 0}^{M^{\sum_{i=1}^{K}d_i}-1}\left[ log \left( \sum_{k_2 = 0}^{M^{\sum_{i=1}^{K}d_i}-1}e^{-\left|\left|\sqrt P A^{k_1,k_2}\right|\right|^2 }\right) \right].\\ \nonumber &+\frac{1}{ M^{\sum_{j \neq i}d_j}}\hspace{-0.2cm}\sum_{i_1 = 0}^{ M^{\sum_{j \neq i}d_j}-1}\left[ log \left( \sum_{i_2 = 0}^{ M^{\sum_{j \neq i}d_j}-1}e^{-\left|\left|\sqrt P B^{i_1,i_2}\right|\right|^2 }\right) \right]. \end{aligned}\] Now, define the set \[\begin{aligned} &{\cal B}^{i_1}=\left\{ i_2 \neq i_1 \mid B^{i_1,i_2}=0 \right\}. \end{aligned}\] The following theorem gives a set of necessary and sufficient conditions under which the above approximation tends to \(log~M^{d_i}\) as \(P\) tends to infinity and hence, gives a set of necessary and sufficient conditions under which the precoders are CCSC optimal. # Gradient Ascent Based Algorithm for Finite-SNR {#sec4} In the previous section, we studied the rate achieved for each transmitter by treating interference as noise at every receiver as \(P\) tends to infinity. In this section, we focus on the finite SNR case. Specifically, the aim is to maximize the sum-rate achieved by treating interference as noise at every receiver with respect to the precoders, i.e., \[\begin{aligned} \max f(V_1,\cdots,V_K)= \max \sum_{i=1}^{K}I[X_i;Y_i] \text{ with } Tr (V_iV_i^H) \leq 1. \end{aligned}\]This is a non-concave problem in general and difficult to solve. Hence, we propose a gradient-ascent based algorithm to improve the sum-rate starting from some random initialization of precoders. Define the MMSE matrix at Rx-\(j\) by \[\begin{aligned} E_j = \mathbb E[(X-\mathbb E[X|Y_j])(X-\mathbb E[X|Y_j])^H] \end{aligned}\]where, \(\mathbb E\) represents the expectation operator. Define the MMSE matrix at Rx-\(j\) with the exclusion of Tx-\(j\)'s signal by \[\begin{aligned} E_{{\not{\hspace{0.02cm}j}}} = \mathbb E[(X_{\not{\hspace{0.02cm}j}}-\mathbb E[X_{\not{\hspace{0.02cm}j}}|Y_j-H_{jj}X_j])(X_{\not{\hspace{0.02cm}j}}-\mathbb E[X_{\not{\hspace{0.02cm}j}}|Y_j-H_{jj}X_j])^H] \end{aligned}\]The gradient of the sum-rate with respect to the precoder \(V_i\) given by \[\begin{aligned} \nonumber &\nabla_{V_i} f(V_1,\cdots,V_K)= \nabla_{V_i} \sum_{j=1}^{i=K}I[X_j;Y_j]\\ \nonumber &=\nabla_{V_i} \sum_{j=1}^{i=K}I[X_1,X_2,\cdots,X_K;Y_j]-I[X_1,X_2,\cdots,X_K;Y_j|X_j]\\ \label{eqn-grad2} &=log~e\sum_{j=1}^{i=K} H^{H}_{ij}H_jE_j~I_{\sum_{k=1}^{i-1}d_k+1:\sum_{k=1}^{i}d_k}\\ \nonumber &-log~e\sum_{\substack{{j=1}\\{j\neq i}}}^{i=K} H^{H}_{ij}H_{\not{\hspace{0.02cm}j}}E_{\not{\hspace{0.02cm}j}}~I_{\sum_{k=1}^{i-1}d_k-d_j{\cal I}(i-j)+1:\sum_{k=1}^{i}d_k-d_j{\cal I}(i-j)} \end{aligned}\]where, ([\[eqn-grad2\]](#eqn-grad2){reference-type="ref" reference="eqn-grad2"}) follows from the relation between the gradient of mutual information and the MMSE matrix obtained in. The matrices \[\begin{aligned} &I_{\sum_{k=1}^{i-1}d_k+1:\sum_{k=1}^{i}d_k} \text{ and }\\ &I_{\sum_{k=1}^{i-1}d_k-d_j{\cal I}(i-j)+1:\sum_{k=1}^{i}d_k-d_j{\cal I}(i-j)} \end{aligned}\]select the column numbers from \(\sum_{k=1}^{i-1}d_k+1\) to \(\sum_{k=1}^{i}d_k\) of \(E_j\) and \(\sum_{k=1}^{i-1}d_k-d_j{\cal I}(i-j)+1\) to \(\sum_{k=1}^{i}d_k-d_j{\cal I}(i-j)\) of \(E_{\not{\hspace{0.02cm}j}}\) respectively, where \[{\cal I}(i-j)=\left\{ \begin{array}{ll} 1& ~i>j \\ 0 & ~i<j. \end{array} \right.\]Define \(V=diag(V_1,V_2,\cdots,V_K)\). The gradient ascent based algorithm for optimizing \(f(V_1,\cdots,V_K)\) with respect to the precoders is given in Algorithm \(1\). During every iteration, whose number is denoted by \(n\), all the precoders are updated as given in Line \(10\) of Algorithm \(1\) where, \(\nabla_V f|_{V=V^{(n-1)}}\) represents \(diag\left(\nabla_{V_1} f|_{V_1=V_1^{(n-1)}},\nabla_{V_2} f|_{V_2=V_2^{(n-1)}},\cdots,\nabla_{V_K} f|_{V_K=V_K^{(n-1)}}\right)\) and \(\nabla_{V_i} f|_{V_i=V_i^{(n-1)}}\) denotes the gradient \(\nabla_{V_i} f\) evaluated at \(V_i=V_i^{(n-1)}\). If the power constraint for any transmitter Tx-\(i\) is violated then, \(V_i^{(n)}\) is projected onto the feasible set with \(Tr(V_i^{(n)}{V_i^{(n)}}^H)\leq 1\) (see Line \(12\) of Algorithm \(1\)). The condition in Line \(15\) of the algorithm ensures that there is sufficient increase in the objective function. The step size \(t\) of the algorithm is chosen by back-tracking line search with parameters \(\alpha\) and \(\beta\) whose typical values lie between \((0.01,0.3)\) and \((0.1,0.8)\). The proposed algorithm stops when either the number of iterations performed is equal to \(max\_iterations\) or \(f^{(n-1)}-f^{(n-2)}<\epsilon\) (see Line \(5\) of Algorithm \(1\)), for some fixed \(\epsilon\). In, for sum-DoF analysis zero-forcing matrices are used at every receiver. Here, we have treated interference as noise and have not used any zero-forcing matrices as using them can only reduce the rate \(I[X_i;Y_i]\) because of data-processing inequality. # Conclusion {#sec5} The paper discussed linear precoding for \(K\)-user MIMO GIC with finite constellation inputs. We showed that, for constant MIMO GIC with finite constellation inputs, CCSC for every transmitter can be achieved just by using a naive scheme of treating the interference as noise at every receiver, at high SNR. This result is in contrast with the Gaussian alphabet case where, at high SNR, the scheme that treats interference as noise saturates to a value determined by the channel gains for the SISO case. A set of necessary and sufficient conditions for CCSC optimal precoders were derived. It was observed that IA precoders fall under the class of CCSC optimal precoders. However, CCSC optimal precoders are easy to obtain for any given value of channel gains unlike obtaining IA precoders. Finally, a gradient ascent based algorithm was proposed to improve the sum-rate achieved starting from any random initialization of precoders. An example simulation with QPSK input alphabets showed a significant improvement in the ergodic sum-rate achieved by the precoders obtained from the proposed algorithm compared to the ergodic sum-rate obtained from IA precoders, treating interference as noise at the receivers, at low and moderate SNRs.

{'timestamp': '2012-10-16T02:03:33', 'yymm': '1210', 'arxiv_id': '1210.3819', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3819'}

# Introduction "Reinforced random walk\" (RRW) is a remarkable example of non-Markovian random walk, which has been extensively studied in the last twenty years, see the survey  . A good story to help understand RRW appeared in  . Imagine a person getting acquainted with a new city. She walks about the area near the hotel somewhat randomly, but tends to traverse the same blocks over and over as they become familiar. To model this, RRW is defined on the vertices of an undirected graph in such a way that the probability of a transition from one vertex to another depends on the number of previous transitions along the neighboring edges (respectively, vertices). We refer to such models as "edge-reinforced random walk\" (ERRW) (respectively, "vertex-reinforced random walk\" (VRRW)).\ RRWs can be defined on all kinds of graphs, either finite or infinite. It also can be defined by very different reinforcement mechanisms through the weight functions, which could lead to quite different phenomena to occur. A question of interest is the recurrence or transience of RRW. For example, Angel et al.   showed that the linearly ERRW on any graph with bounded degrees is recurrent for sufficiently small initial weights. Another phenomenon of interest is localization. For example, Pemantle and Volkov   showed that VRRW on \(\mathbb{Z}\) has finite range, and then Tarrés   showed that VRRW on \(\mathbb{Z}\) eventually gets stuck on five points. Phase transitions are also of interest. For instance, Pemantle   showed that the ERRW on infinite binary tree can vary from transient to recurrent, depending on the value of an adjustable parameter measuring the strength of the feedback (reinforcement), and Volkov   showed that VRRW on \(\mathbb{Z}\) with weight function \(k^\alpha\) will just visit \(2, 5, \infty\) sites after a large time \(T_0\) when \(\alpha>1,=1\) or \(<1\) respectively.\ So far, almost all the considered RRW models are one particle' self-interacting with positive (attractive) feedback (reinforcement), i.e. the edges (or vertices) already being visited more are more likely to be visited in the future. In analyzing such models, four main methods are commonly used: exchangeability    , branching process embedding        , stochastic approximation via martingale methods    , and dynamical system approach   . Here we want to say several more words about the history of the dynamical system approach, which has been used in the first half of this paper. The dynamical system approach is an approach to analyze stochastic approximation process (defined later) by its approximating ODE. This approach has been extensively studied since it was introduced by Ljung  . From the 90s, Benaïm and his collaborators have developed this approach, which enables to analyze stochastic approximation process with much more general dynamics.\ One direction to generalize RRW is to consider multi-particle RRW. Kovchegov   showed for the two particles' edge reinforced process on \(\mathbb{Z}\), the two particles meet infinitely often a.s.. In his model, each particle's transition probability is positively reinforced (determined) by both particles' visits on the edges. Except his paper, the generalization of RRW to multi-particle RRW models with more general reinforced interacting mechanism hasn't appeared yet so far. In 2010, Itai Benjamini proposed a class of new multi-particle reinforced interacting random walk, and finally our paper is a result of exploration on this model.\ The organization of this paper is as follows: Section 2 is devoted to the definition of the model; Section 3 is devoted to state our main result; Section 4 is the proof of main result; Some further open problems are proposed in Section 5. # The model In this section, we define a class of two-particle vertex-reinforced interacting random walk on a connected graph. Denote the two particles by \(X\) and \(Y\), and the graph by \(G=(V,E)\). At each time unit, both \(X\) and \(Y\) particle will jump to some vertices on \(V\). Set \(X,Y\)'s locations at time \(k\) on \(V\) as random variables \(X_k, Y_k\), and the number of \(X,Y\)'s visits to vertex \(v\) by time \(n\) as random variables \(N(X,v,n),N(Y,v,n)\). In our model, we assume \(X,Y\) visit each vertex initially. Then we have \[N(X,v,n)=1+\sum_{k=1}^n 1_{X_k=v}, \quad N(Y,v,n)=N(Y,v,0)+\sum_{k=1}^n 1_{Y_k=v}, \quad \forall v\in V.\] Denote the natural filtration generated by \(\{X_k,0\le k \le n\}\) and \(\{Y_k,0\le k \le n\}\) by \(\mathcal{F}_n(n\in \mathbb{N})\). Then the stochastic process \((X_n, Y_n)\)'s transition probability is defined \[\label{Xmove} \mathbb{P}(X_{n+1}=v|\mathcal{F}_n)=\frac{1_{v\sim X_n}\cdot w_{N (Y,v,n)}}{\sum_{p\sim X_n} w_{N (Y,p,n)}}\] and \[\label{Ymove} \mathbb{P}(Y_{n+1}=v|\mathcal{F}_n)=\frac{1_{v\sim Y_n}\cdot w_{ N (X,v,n)}}{\sum_{p\sim Y_n} w_{N (X,p,n)}},\] where "\(\sim\)\" generally represents some vertex relation on \(G\) (e.g nearest-neighbor), and \(w_k(k\in \Z_+)\) is a fixed sequence of positive numbers, referred to as "weights\". Like the classical one-particle RRW, one natural weight sequence to work with is \(w_k=k^{-\alpha}\) for some \(\alpha>0\). On a finite graph \(G\) of size \(N_0\), one can set \[\label{randomprocess} x_v(n)=\frac{N(X,v,n)}{n+N_0},\quad y_v(n)=\frac{N(Y,v,n)}{n+N_0}, \forall v\in V\] as \(X\) and \(Y\)'s *empirical occupation measure* on \(V\) by time \(n\). Then the \(2N_0\) dimensional vector \(\left[(x_v(n))_{v\in V}, (y_v(n))_{v\in V}\right]\) is a random process living in a bounded subset of Euclidean space \(\R^{2N_0}\). If \(w_k\) (like \(k^{-\alpha}\)) is homogenous in \(k\), we can rewrite ([\[Xmove\]](#Xmove){reference-type="ref" reference="Xmove"}) and ([\[Ymove\]](#Ymove){reference-type="ref" reference="Ymove"}) as \[\label{measureXmove} \mathbb{P}(X_{n+1}=v|\mathcal{F}_n)=\frac{1_{v\sim X_n}\cdot \varpi(y_v(n))}{\sum_{p\sim X_n} \varpi(y_p(n))}\] and \[\label{measureYmove} \mathbb{P}(Y_{n+1}=v|\mathcal{F}_n)=\frac{1_{v\sim Y_n}\cdot \varpi( x_v(n))}{\sum_{p\sim Y_n} \varpi(x_p(n))},\] where \(\varpi(x)\) is a function \(\varpi: [0,1] \to \R^+\). Hence, \(X,Y\)'s transition probabilities are defined by their occupation measure. In this paper, we will work with the above defined model ([\[measureXmove\]](#measureXmove){reference-type="ref" reference="measureXmove"}) and ([\[measureYmove\]](#measureYmove){reference-type="ref" reference="measureYmove"}) on finite complete graph. The vertex relation "\(\sim\)\" is that the graph distance between two vertices is less than or equal to 1 and \(\varpi(x)=\left[\delta 1_{x\le \delta}+x 1_{x>\delta}\right]^{-\alpha}\) for some small positive \(\delta\). For convenience, we label the vertex set \(V\) of a finite complete graph of size \(d\) by the set \(S=\{1,\ldots, d\}\). Then for any \(i,j\in S\), ([\[measureXmove\]](#measureXmove){reference-type="ref" reference="measureXmove"}) and ([\[measureYmove\]](#measureYmove){reference-type="ref" reference="measureYmove"}) becomes \[\label{completemeasureXmove} \mathbb{P}(X_{n+1}=i|\mathcal{F}_n)=\frac{ \left[\delta 1_{y_i(n)\le \delta}+y_i(n) 1_{y_i(n)>\delta}\right]^{-\alpha}}{\sum_{k=1}^d \left[\delta 1_{y_k(n)\le \delta}+y_k(n) 1_{y_k(n)>\delta}\right]^{-\alpha}}\] and \[\label{completemeasureYmove} \mathbb{P}(Y_{n+1}=j|\mathcal{F}_n)=\frac{ \left[\delta 1_{x_j(n)\le \delta}+x_j(n) 1_{x_j(n)>\delta}\right]^{-\alpha}}{\sum_{k=1}^d \left[\delta 1_{x_k(n)\le \delta}+x_k(n) 1_{x_k(n)>\delta}\right]^{-\alpha}},\] Set \(x(n)=(x_i(n))_{i\in S}\), \(y(n)=(y_i(n))_{i\in S}\) and \(z(n)=(x(n),y(n))\). Then in this paper we will prove \(z(n)\)'s asymptotic behavior. Notice that \(z(n)\) is a Markov chain living in \(\R^{2d}\). Here we want to mention that when \(\min_{i,j\in S} \{x_i(n),y_j(n)\}>\delta\), ([\[completemeasureXmove\]](#completemeasureXmove){reference-type="ref" reference="completemeasureXmove"}) and ([\[completemeasureYmove\]](#completemeasureYmove){reference-type="ref" reference="completemeasureYmove"}) are equivalent to the following formulas: \[\label{XXmove} \mathbb{P}(X_{n+1}=j|\mathcal{F}_n)=\frac{N (Y,j,n)^{-\alpha}}{\sum_{k=1}^d N (Y,k,n)^{-\alpha}}, \quad \forall i, j\in S;\] \[\label{YYmove} \mathbb{P}(Y_{n+1}=j|\mathcal{F}_n)=\frac{N (X,j,n)^{-\alpha}}{\sum_{k=1}^d N (X,k,n)^{-\alpha}}, \quad \forall i,j \in S,\] which can be easily seen as \(X,Y\)'s *repelling* interacting random walk with nonlinear reinforcement. # The main result For the same \(\delta\) appearing in \(\varpi(x)\), we define \(S^\delta=\{(x,y)\in D: \sum_{i=1}^d x_iy_i<2d^{1/(\alpha+1)}\delta\}\). Notice that \(2d^{1/(\alpha+1)}\) is a constant independent of \(\delta\). Then we have the following theorem for any fixed \(\delta>0\). # Proof of the main result In Subsection [4.1](#sectionstochastic){reference-type="ref" reference="sectionstochastic"}, we will conclude that \(z(n)\) belongs to a class of *stochastic approximation algorithms*. In Subsection [4.2](#interpolated){reference-type="ref" reference="interpolated"}, we will introduce the *interpolated process* which has the same limit set as \(z(n)\). In Subsection [4.3](#ODE){reference-type="ref" reference="ODE"}, we will introduce the deterministic ODE associated with \(z(n)\) and construct a non-rigorous Lyapunov function. In Subsection [4.4](#somelemmas){reference-type="ref" reference="somelemmas"}, we will state Benaïm's theorem which addresses the relation between the interpolated process and the ODE. In Subsection [4.5](#mainlemma){reference-type="ref" reference="mainlemma"}, we will give our main lemma, which describes the so-called *chain-recurrent set* for our specific semiflow induced by the ODE. In Subsection [4.6](#unstableanalysis){reference-type="ref" reference="unstableanalysis"}, we will use Pemantle's theorem to show that \(z(n)\) has probability 0 to converge to some isolated unstable equilibrium. In Subsection [4.7](#proofofmaintheorem){reference-type="ref" reference="proofofmaintheorem"}, we finally give the proof of Theorem [\[maintheorem1\]](#maintheorem1){reference-type="ref" reference="maintheorem1"}. ## Stochastic approximation algorithms {#sectionstochastic} Stochastic approximation algorithm basically is a discrete time random process defined by a recursive formula. The theory of stochastic approximations was born in the early 50s with the works of Robbins and Monro   and Kiefer and Wolfowitz  . Now to show that that \(z(n)\) is stochastic approximation algorithm, we start with ([\[completemeasureXmove\]](#completemeasureXmove){reference-type="ref" reference="completemeasureXmove"}) and ([\[completemeasureYmove\]](#completemeasureYmove){reference-type="ref" reference="completemeasureYmove"}). Notice that the right hand sides of ([\[completemeasureXmove\]](#completemeasureXmove){reference-type="ref" reference="completemeasureXmove"}) and ([\[completemeasureYmove\]](#completemeasureYmove){reference-type="ref" reference="completemeasureYmove"}) is a map \(M:D \to D\) and this map is continuous on the whole compact \(D\), but not differentiable near the boundary of \(D\) (denoted by \(\partial D\)) according to the definition of \(\varpi(x)\). Set a map \(\pi=(\pi_i)_{i\in S}: \Delta\to \Delta\): \[\label{stationary} \pi_i(x)=\frac{\varpi(x_i)}{\sum_{k=1}^d \varpi(x_k)}, \quad \forall x\in \Delta.\] Then we can derive a *difference equation* for \(x_i(n)\) \[\begin{aligned} \label{differenceeqn} x_i(n+1)-x_i(n) &=& \frac{N(X,i,n)+1_{X_{n+1}=i}}{n+1+N_0}-\frac{N(X,i,n)}{n+N_0} \nonumber \\ &=& \frac{-x_i(n)+1_{X_{n+1}=i}}{n+1+N_0} \nonumber \\ &=&\frac{-x_i(n)+\pi_i(y(n))}{n+1+N_0}+\frac{1_{X_{n+1}=i}-\pi_i(y(n))}{n+1+N_0} \end{aligned}\] Similarly, we can derive a difference equation for \(y_i(n)\). Set \[\label{xerror} \eps_i^X(n)= \frac{1_{X_{n+1}=i}-\pi_i(y(n))}{n+1+N_0}, \quad \eps^X(n)=\left(\eps_i^X(n)\right)_{i\in S}\] and \[\label{yerror} \eps_i^Y(n)= \frac{1_{Y_{n+1}=i}-\pi_i(x(n))}{n+1+N_0} \quad \eps^Y(n)=\left(\eps_i^Y(n)\right)_{i\in S}.\] It follows from ([\[completemeasureXmove\]](#completemeasureXmove){reference-type="ref" reference="completemeasureXmove"}) and ([\[completemeasureYmove\]](#completemeasureYmove){reference-type="ref" reference="completemeasureYmove"}) that \[\label{zeroerror} \mathbb{E}(\eps_i^X(n))|\mathcal{F}_n)=0,\quad \mathbb{E}(\eps_i^Y(n))|\mathcal{F}_n)=0, \quad \forall i\in S.\] Then we can write the difference equation for the random process \((x(n),y(n))\) in the following vector form: \[\label{xdifference} x(n+1)-x(n)=\frac{1}{n+1+N_0}(-x(n)+\pi(y(n)))+\eps^X(n)\] \[\label{ydifference} y(n+1)-y(n)=\frac{1}{n+1+N_0}(-y(n)+\pi(x(n)))+\eps^Y(n).\] Set \(\eps(n)=\left(\eps^X(n),\eps^Y(n)\right)\) and \(F(z(n))=F(x(n),y(n))=(-x(n)+\pi(y(n)),-y(n)+\pi(x(n)))\). One can also write ([\[xdifference\]](#xdifference){reference-type="ref" reference="xdifference"}) and ([\[ydifference\]](#ydifference){reference-type="ref" reference="ydifference"}) as \[\label{stochasticapproximation} z(n+1)-z(n)=\frac{1}{n+1+N_0} F(z(n))+\eps(n).\] Then ([\[stochasticapproximation\]](#stochasticapproximation){reference-type="ref" reference="stochasticapproximation"}) implies that \(z(n)\) is stochastic approximation algorithm. The stochastic approximations in the form of ([\[stochasticapproximation\]](#stochasticapproximation){reference-type="ref" reference="stochasticapproximation"}) with \(\mathbb{E}(\eps(n)|\mathcal{F}_n)=0\) is also called *Robbins-Monro stochastic approximation algorithm*. So our random process \(z(n)\) is Robbins-Monro stochastic approximation algorithm. As mentioned before, we will use the dynamical system approach to analyze ([\[stochasticapproximation\]](#stochasticapproximation){reference-type="ref" reference="stochasticapproximation"}). The dynamical system approach mainly says one can introduce the averaged ordinary differential equation (denoted by ODE) \[\label{vectorODE} \frac{dz(t)}{dt}=F(z(t))\] and then describe the asymptotic behavior of ([\[stochasticapproximation\]](#stochasticapproximation){reference-type="ref" reference="stochasticapproximation"}) by the semiflow (solution) induced by ([\[vectorODE\]](#vectorODE){reference-type="ref" reference="vectorODE"}). An application of the dynamical system approach in VRRW is given in  . By the limit set theorem proved in  , the main difficulty to apply this approach usually boils down to analyze the deterministic ODE, which is nontrivial in many occasions. Lyapunov function method is an ideal way. However, there is no general method to construct such a function. ## The interpolated process {#interpolated} In this section, we define stochastic approximation algorithm \(z(n)\)'s interpolated process. Set \(\tau_0=0, \sigma_n = 1/(k+N_0), \tau_n=\sum_{k=1}^n \sigma_k\). The interpolated process in Definition [\[definitioninterpolated\]](#definitioninterpolated){reference-type="ref" reference="definitioninterpolated"} can be written in equations as follows: \[\label{xinterpolated} z(t)=\sum_{n\ge 0}\left[(z(n+1)-z(n))\frac{(t-\tau_n)}{1/(n+1+N_0)}+z(n)\right]\cdot 1_{[\tau_n,\tau_{n+1})}(t)\] Set \(z(t)=(x(t),y(t))\) where \(x(t),y(t)\in \Delta\). For any fixed \(t\in [\tau_n,\tau_{n+1})\), from ([\[xdifference\]](#xdifference){reference-type="ref" reference="xdifference"}) and ([\[xinterpolated\]](#xinterpolated){reference-type="ref" reference="xinterpolated"}), we can write the interpolated process of \(x(n)\) as follows \[\begin{aligned} \label{xcalculation} x(t) &=& (x(n+1)-x(n))\frac{(t-\tau_n)}{1/(n+1+N_0)}+x(n) \nonumber \\ &=& (x(n+1)-x(n))\frac{(t-\tau_n)}{1/(n+1+N_0)}+\sum_{k=1}^n (x(k)-x(k-1))+x(0) \nonumber \\ &=& \left[(-x(n)+\pi(y(n)))+(n+1+N_0)\eps^X(n)\right](t-\tau_n) \nonumber \\ & &+\sum_{k=1}^n \left[(-x(k-1)+\pi(y(k-1)))+(k+N_0)\eps^X(k-1)\right](\tau_{k-1}-\tau_{k})+x(0). \nonumber \\ & & \end{aligned}\] Set \(\gamma(n)=(n+1+N_0)\eps^X(n)(n\ge 0)\). Then according to Notation [\[stepfunction\]](#stepfunction){reference-type="ref" reference="stepfunction"}, ([\[xcalculation\]](#xcalculation){reference-type="ref" reference="xcalculation"}) can be written in integral forms \[\label{xintegral} x(t)-x(0)=\int_0^t(-x^0(s)+\pi(y^0(s)))ds+\int_0^t \gamma^0(s)ds.\] Let \(\eta(n)=(n+1+N_0)\eps^Y(n)(n\ge 0)\). Similarly we have \[\label{yintegral} y(t)-y(0)=\int_0^t(-y^0(s)+\pi(x^0(s)))ds+\int_0^t \eta^0(s)ds.\] Set \(z^0(t)=(x^0(t),y^0(t))\) and \(\theta^0(t)=(\gamma^0(t),\eta^0(t))\). Then we can rewrite the integral equations ([\[xintegral\]](#xintegral){reference-type="ref" reference="xintegral"}) and ([\[yintegral\]](#yintegral){reference-type="ref" reference="yintegral"}) in a more unified form \[\label{interpolatedintegral} z(t)-z(0)=\int_0^tF(z^0(s))ds+\int_0^t \theta^0(s)ds.\] Let \[\label{originallimit} L\left(\{z(n)\}_{n\ge 0}\right)=\bigcap_{k\ge 0}\overline{\{z(n):n\ge k\}},\] and \[\label{interpolatedlimit} L\left(\{z(t)\}_{t\ge 0}\right)=\bigcap_{s\ge 0}\overline{\{z(t):t\ge s\}}\] denote the omega limit set of \(z(n)\) and \(z(t)\) respectively. Then we have the following lemma due to Benaïm. ## The deterministic ODE {#ODE} Set \(f(x)=\delta 1_{x\le \delta}+x 1_{x>\delta}\), i.e. \(f(x)=\varpi^{-\frac{1}{\alpha}}\). According to the above analysis, the deterministic (approximating) ODE in our case is: \[\label{DerODE} \begin{cases}\frac{du_i(t)}{dt}=-u_i(t)+\frac{f(v_i)^{-\alpha}}{\sum_{k=1}^d f(v_k)^{-\alpha}} & i=1,\ldots, d \\ \frac{dv_i(t)}{dt}=-v_i(t)+\frac{f(u_i)^{-\alpha}}{\sum_{k=1}^d f(u_k)^{-\alpha}} & i=1,\ldots, d \\ \end{cases}\] or in vector form \[\label{DerODEvector} \begin{cases}\frac{du(t)}{dt}=-u(t)+\pi(v(t)) \\ \frac{dv(t)}{dt}=-v(t)+\pi(u(t)) \\ \end{cases} \quad \text{or} \quad \frac{d\Xi(t)}{dt}=F(\Xi(t))\] where \(\Xi(t)=(u(t),v(t))\in D\). We call the right hand side of ([\[DerODE\]](#DerODE){reference-type="ref" reference="DerODE"}) and ([\[DerODEvector\]](#DerODEvector){reference-type="ref" reference="DerODEvector"}) as the ODE's vector field. In our case, the vector field \(F\) is defined as a map \(F:\Delta\times \Delta \to T(\Delta\times \Delta)\), i.e. \(F:D \to TD\), where \[TD=T(\Delta\times \Delta)=\left\{(u,v)\in \mathbb{R}^{2d}, \sum_{i\in S}u_i=0, \sum_{i\in S}v_i=0 \right\}.\] Notice that the function \(f(x)\) satisfies \(f(x)\ge \delta\) and \(1/f(x)\ge 1\). Hence by the fact that composition of Lipschitz functions is Lipschitz, \(F\) is globally Lipschitz. By theorem of maximal interval of existence of ODE's solution (see e.g.  ), it follows that the vector field \(F\) generates a unique *semiflow* in \(D\) \[\Phi:\mathbb{R}_+\times D \to D,\] \[(t,a)\to \Phi_t(a),\] such that the solution to the initial value problem \(d\Xi(t)/dt=F(\Xi(t))\) with initial condition \(\Xi(0)=a \in D\) is the curve \(t\to \Phi_t(a)\). Here to make things clear, we give the following definition. In this paper, we always work with semiflow. It is easy to check that \(p=(U,U)\) is a critical point of the vector field \(F\), i.e. \(F(p)=0\). After the following definition, we will classify what type of critical point it is. Let \(T\) be the linear approximation to \(F\) near a critical point \(p\) so that \(F(p+w)=T(w)+O(|w|^2)\), then Define the function \(H(u,v)=\sum_{i=1}^d u_i v_i, \left((u,v)\in D \right)\). In fact, \(H(u,v)\) plays the role of a Lyapunov function in our proof. However, from Lemma [\[derivativenegative\]](#derivativenegative){reference-type="ref" reference="derivativenegative"} below, one can see that it is not monotone in the whole domain, and hence it is not exactly a Lyapunov function according to the classical definition. Set the function H(t) associated with the semiflow \(\Phi\) induced by ([\[DerODE\]](#DerODE){reference-type="ref" reference="DerODE"}) as \[\label{Lyapunovfunction} H(t)=H(u(t),v(t))=\sum_{i=1}^d u_i(t) v_i(t).\] Then \[\begin{aligned} \label{derivativeformula} \frac{dH}{dt}&=&\sum_{i=1}^d v_i(t)\frac{du_i(t)}{dt}+ \sum_{i=1}^d u_i(t) \frac{dv_i(t)}{dt} \nonumber \\ &=&\sum_{i=1}^d v_i\left(-u_i+\frac{f(v_i)^{-\alpha}}{\sum_{k=1}^d f(v_k)^{-\alpha}}\right)+ \sum_{i=1}^d u_i \left(-v_i+\frac{f(u_i)^{-\alpha}}{\sum_{k=1}^d f(u_k)^{-\alpha}}\right) \nonumber\\ &=&-2\sum_{i=1}^d u_i v_i+\frac{\sum_{i=1}^d u_if(u_i)^{-\alpha}}{\sum_{k=1}^d f(u_k)^{-\alpha}}+\frac{\sum_{i=1}^d v_if(v_i)^{-\alpha}}{\sum_{k=1}^d f(v_k)^{-\alpha}}. \end{aligned}\] Notice that ([\[DerODE\]](#DerODE){reference-type="ref" reference="DerODE"}) and ([\[Lyapunovfunction\]](#Lyapunovfunction){reference-type="ref" reference="Lyapunovfunction"}) are autonomous, so is ([\[derivativeformula\]](#derivativeformula){reference-type="ref" reference="derivativeformula"}). Hence, the right side of ([\[derivativeformula\]](#derivativeformula){reference-type="ref" reference="derivativeformula"}) is often considered as a continuous function of \((u,v)\) on \(D\), which is independent of time \(t\). We have the following lemma about ([\[derivativeformula\]](#derivativeformula){reference-type="ref" reference="derivativeformula"}). ## Benaïm's theorems {#somelemmas} The exact relationship between the deterministic ODE and the asymptotic behavior of the interpolated process is addressed by the limit set theorem in. We start a definition introduced in  . Then we have the following lemma. Before stating Theorem [\[Benaimconvergence\]](#Benaimconvergence){reference-type="ref" reference="Benaimconvergence"}, we need to give the definition of chain-recurrence. Suppose \(D\) is a metric space with a metric \(\mu\). In our case, \(\mu\) is the metric induced by \(L^1\) norm. We denote by \(CR(\Phi)\) the set of chain-recurrent points. The set \(CR(\Phi)\) has the property to be closed and invariant. The following theorem, which is due to Benaïm, demonstrates the relation between the asymptotic behavior of asymptotic pseudotrajectories and the dynamics of \(\Phi\). The theorem can be viewed as a deterministic statement. ## The main lemma {#mainlemma} Now it comes to our main lemma to prove Theorem [\[maintheorem1\]](#maintheorem1){reference-type="ref" reference="maintheorem1"}. This lemma describes the chain recurrent set for our specific dynamic \(\Phi\). In the proof of this lemma, Lemma [\[derivativenegative\]](#derivativenegative){reference-type="ref" reference="derivativenegative"} will be heavily used. Recall \(S^\delta=\{(x,y)\in D: \sum_{i=1}^d x_iy_i<2d^{1/(\alpha+1)}\delta\}\). \ \ \ Now we can finish the proof of this lemma. From Step 2 and 3, we have \(CR_1 \subset M_1\) and \(CR_2 \subset (U,U)\), and hence \(CR(\Phi) \subset M_1 \cup (U,U)=S^\delta \cup(U,U)\). ◻ Before applying Theorem [4.6](#unstableanalysis){reference-type="ref" reference="unstableanalysis"} to our case, we need the following lemma, which is to check the noise terms \(\epsilon(n)\) in our case satisfies Condition (3) in Theorem [4.6](#unstableanalysis){reference-type="ref" reference="unstableanalysis"}. Now we can apply Theorem [4.6](#unstableanalysis){reference-type="ref" reference="unstableanalysis"} to our case. ## Proof of Theorem [\[maintheorem1\]](#maintheorem1){reference-type="ref" reference="maintheorem1"} {#proofofmaintheorem} # Further questions For our model defined in ([\[Xmove\]](#Xmove){reference-type="ref" reference="Xmove"}) and ([\[Ymove\]](#Ymove){reference-type="ref" reference="Ymove"}), little is known so far. On \(\Z^2\), we even are not able to prove that each vertex will eventually be visited by at least one of \(X,Y\) almost surely. Another question of interest is

{'timestamp': '2012-10-16T02:03:17', 'yymm': '1210', 'arxiv_id': '1210.3795', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3795'}

# Introduction The interest in new scintillation materials is promoted by increasing number of new applications in medicine, science, and homeland security, which require ramp-up of material production. The most perspective scintillators are bromides and iodides doped with Ce\(^{3+}\) and Eu\(^{2+}\) ions, such as SrI\(_{2}\)-Eu and LaBr\(_{3}\)-Ce. These crystals have high light outputs (up to 100000 photons/MeV for SrI\(_{2}\)-Eu), good energy resolution, and high proportionality. Disadvantages of these scintillators are high hygroscopic and price. In addition, SrI\(_{2}\)-Eu has temperature instability of light output observed by [@Alekhin11]. For the most applications a cheaper NaI-Tl scintillator has quite properties (light output about 45000 photons/MeV cited in [@Berkley]). Therefore, one of the way in development of new scintillators is to find new materials with similar to NaI-Tl properties but no hygroscopic. In this way advanced materials for new scintillators are alkali-earth fluorides doped with rare earth ions. Theoretical limit of light output for them is up to 50000 photons/MeV. If an efficient energy transfer is provided then alkali-earth fluorides will be promising scintillators. A real light output of CaF\(_2\)-Eu is 18000-22000 photons/MeV, but BaF\(_2\) and BaF\(_2\)-Ce crystals demonstrate lower light output at about 10000 photons/MeV. Scintillation properties of SrF\(_2\) crystals are almost not investigated. Light output of SrF\(_2\) was estimated about 10000-12000 photons/MeV by [@Schotanus87]. However, potential light output of SrF\(_2\) will be higher. Also SrF\(_2\) crystals doped with Ce\(^{3+}\) and Pr\(^{3+}\) have a temperature stability of light output in wide range (20 \(^\circ\)C to 200 \(^\circ\)C). Therefore, SrF\(_2\) can be high-potential scintillator for well-logging. So, scintillator properties of strontium fluoride crystals are among the least studied of fluorides crystals, but these crystals have potential applications. Thus, the investigations of scintillation properties of strontium fluorides are topical today. This paper describes the scintillation properties of pure and cerium doped strontium fluorides crystals, a newly discovered inorganic scintillator. # Experimental methodology Growing with addition of CdF\(_{2}\) as an oxygen scavenger, oxygen-free crystals of pure SrF\(_2\) and doped with different concentrations of Ce\(^{3+}\) ions were grown in a graphite crucible by the Stockbarger method. We applied several experimental techniques in measurement of scintillation properties of the crystals. To determine light output, pulsed-height spectra under \(^{137}\)Cs 662 KeV gamma source excitation were measured with PMT FEU-39A, a homemade preamplifier and an Ortec 570A spectrometric amplifier. The crystals of 10x10 mm dimensions were polished and then covered with several layers of ultraviolet reflecting Teflon tape (PTFE tape). The shaping time of Ortec 570 spectrometric amplifier was set at 10 \(\mu\)s to collect much light from scintillator. X-ray excited luminescence was performed using x-ray tube with Pd anode operating at 35 kV and 0.8 mA. The spectra were recorded at photon-counting regime using PMT FEU-39A and vacuum grating monochromator VM-4. Scintillation decay time profiles under \(^{137}\)Cs E=662 KeV gamma source excitation were recorded by 200 MHz oscilloscope Rigol DS-1202CA. To register decay curves in wide time interval, we used oscilloscope input resistance set (50 \(\Omega\) and 2.8 k\(\Omega\)). # Experimental results and discussion Figure [\[spectra\]](#spectra){reference-type="ref" reference="spectra"} shows spectra of x-ray luminescence of pure SrF\(_2\), NaI-Tl, SrF\(_2\)-0.3 mol.% Ce\(^{3+}\), and CaF\(_2\)-0.1 mol.% Eu\(^{2+}\). In the spectrum of SrF\(_2\) a wide band at 280 nm is attributed to self-trapped exciton (STE) emission. In SrF\(_2\) doped with Ce\(^{3+}\) ions STE luminescence is quenched and vanished at concentrations Ce\(^{3+}\) ions higher than 1 mol. %. The most intense bands in x-ray luminescence spectra of SrF\(_2\)-Ce\(^{3+}\) crystal at 310 and 325 nm correspond to 5d-4f emission of Ce\(^{3+}\) ions. Luminescence spectrum of CaF\(_2\)-Eu\(^{2+}\) is shown in Fig. [\[spectra\]](#spectra){reference-type="ref" reference="spectra"}, curve 4. Its emission band is centered at 425 nm. This luminescence is due to 4f\(^6\)`<!-- -->`{=html}5d\(^1\)-4f\(^7\) transitions in the Eu\(^{2+}\) ion. Dependence of integral intensity of Ce\(^{3+}\) ions emission bands on Ce concentration is shown in the inset to Figure [\[spectra\]](#spectra){reference-type="ref" reference="spectra"}. The highest light output is demonstrated by SrF\(_2\)-0.3 mol.% Ce\(^{3+}\). X-ray excited luminescence output measured by integral intensity is compared with the one of NaI-Tl crystal (Table [\[table1\]](#table1){reference-type="ref" reference="table1"}). Light output of NaI-Tl crystals is approximately 43000 photons/MeV. Therefore, light output of the measured samples can be estimated. The data are shown in Table [\[table1\]](#table1){reference-type="ref" reference="table1"}. Light output of CaF\(_2\)-0.1 mol.% Eu\(^{2+}\) is about 21500 photons/MeV that is in according with known data for CaF\(_2\)-Eu crystals given by [@Berkley]. Pure SrF\(_2\) has light output about 20640 photons/MeV, doped with 0.3 mol.% and 1 mol.% crystals have the ones about 34000 photons/MeV and 18500 photons/MeV, respectively. All integral intensities and light outputs are presented without correction for spectral response of registration channel. The spectral response curve is shown in Figure [\[spectra\]](#spectra){reference-type="ref" reference="spectra"}, dashed line. The sensitivity of the PMT and monochromator system in SrF\(_2\) and SrF\(_2\)-Ce luminescence spectral range is lower than in NaI-Tl and CaF\(_2\)-Eu emission region. After the correction, light output of pure SrF\(_2\) luminescence is about 40000 photons/MeV, and SrF\(_2\) doped with 0.3 mol. % and 1 mol. % crystals have light outputs about 50000 photons/MeV and 28000 photons/MeV, respectively. Temperature dependences of x-ray excited luminescence intensity of SrF\(_2\)-Ce are given in Figure [\[temp\]](#temp){reference-type="ref" reference="temp"}. Emission intensity does not depend on temperature in the range between-50 \(^\circ\)C and +50 \(^\circ\)C. At higher temperatures decrease of 5d-4f luminescence intensity is observed. At 170 \(^\circ\)C temperature light output of SrF\(_2\) doped with 0.01 mol.% and 0.1 mol% of Ce\(^{3+}\) ions decreases to 30 %. 25 % and 15 % decreases in integral emission are found in the SrF\(_2\) crystals doped with 0.3 mol. % Ce\(^{3+}\) and 1 mol. %, respectively. Crystals of SrF\(_2\)-0.3 mol. % Ce\(^{3+}\) and SrF\(_2\)-1 mol. % Ce\(^{3+}\) demonstrate high temperature stability of light output in the region between-50 \(^\circ\)C and 170 \(^\circ\)C in comparison with NaI-Tl (see Fig. [\[temp\]](#temp){reference-type="ref" reference="temp"}, solid line). For this reason, the SrF\(_2\)-Ce\(^{3+}\) crystals would be perspective scintillators for well-logging applications. Figure [\[pulsed\]](#pulsed){reference-type="ref" reference="pulsed"} shows pulse height spectra of SrF\(_2\), SrF\(_2\)-0.3 mol. % Ce\(^{3+}\) and NaI-Tl. The photopeak corresponding to the \(^{137}\)Cs energy photon is seen in each curve in the Figure [\[pulsed\]](#pulsed){reference-type="ref" reference="pulsed"}. Light output of SrF\(_2\) and SrF\(_2\)-Ce crystals was measured by comparing these response to 662 KeV energy to the response of NaI-Tl crystal with known characteristics under the same conditions. The photopeak in pure SrF\(_2\) is centered at a pulse height that is 42% of the 662 keV photopeak pulse height in NaI-Tl. Using the NaI-Tl light output of 43000 photons/MeV, this implies that the light output of pure SrF\(_2\) is approximately 18000 photons/MeV that is similar to x-ray emission light output (see Table [\[table1\]](#table1){reference-type="ref" reference="table1"}). The full width at half maximum (FWHM) in pure SrF\(_2\) of the 662 keV photopeak is 10 %. FWHM of NaI-Tl is 6.7 %, which is consistent with known results given by scintillators database of [@Berkley]. Dependence of SrF\(_2\)-Ce relative light output versus concentration is demonstrated in the inset to Figure [\[pulsed\]](#pulsed){reference-type="ref" reference="pulsed"}. Best light output of the Ce-doped crystals is found for SrF\(_2\)-0.3 mol. % Ce\(^{3+}\). Its photopeak is centered at a pulse height that is 32 % of the 662 keV photopeak pulse height in NaI-Tl, this implies that the light output of SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) is approximately 14000 photons/MeV. FWHM of the SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) is approximately 9.3 %, which is lower than FWHM of pure SrF\(_2\). This fact means that in spite of worse light output SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) crystal has better energy resolution in comparison with pure strontium fluoride. All light outputs without any corrections are given in Table [\[table1\]](#table1){reference-type="ref" reference="table1"}. Bearing in mind that spectral sensitivity of S20 photocathode (PMT FEU 39A) is higher at CaF\(_{2}\)-Eu and NaI-Tl emission bands than in SrF\(_{2}\) and SrF\(_{2}\)-Ce luminescence region. Corrected light outputs of pure SrF\(_2\) is about 80% (36000 photons/MeV) of NaI-Tl and SrF\(_2\)-0.3 mol. % of Ce\(^{3+}\)--about 60% (26000 photons/MeV) of NaI-Tl. Scintillation decay time profile of SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) is shown in Figure [\[decay\]](#decay){reference-type="ref" reference="decay"}. Resistance of oscilloscope input was 2.6 K\(\Omega\) for registration long time decay components in Ce\(^{3+}\) emission. The decay time is described by a sum of exponents. First component (2.8 \(\mu\)s) in Ce\(^{3+}\) decay is integrated short components. Lifetime of the shortest one equals 130 ns at 50 \(\Omega\) input resistance in SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) and it becomes longer with decrease of Ce concentration. Decay constants of this component in dependence on Ce concentration are presented in the inset of the Figure [\[decay\]](#decay){reference-type="ref" reference="decay"}. Contribution of slow components to scintillation time profile is estimated. In the figure [\[decay\]](#decay){reference-type="ref" reference="decay"} exponential components of total decay curve are shown separately. There are two long time components in SrF\(_2\)-0.3 mol.% Ce\(^{3+}\) emission. 20% of the light is emitted with a 9 \(\mu\)s time constant, and 25% of the light is emitted with a 280 \(\mu\)s time constant. Emission decay time of cerium doped alkaline-earth fluorides under optical excitation at lowest energy absorption bands is estimated about 30 ns. Under vacuum ultraviolet excitation at exciton and higher energies regions the decay of Ce-doped fluorides became nonexponential. Under x-ray excitation the decay curves is also nonexponential (Fig. [\[decay\]](#decay){reference-type="ref" reference="decay"}). Whole decay curve can be described by several processes. Fast stage could be ascribed to resonance energy transition in nearest pairs of exciton and cerium ion. In the inset of the figure [\[decay\]](#decay){reference-type="ref" reference="decay"} concentration dependence of these decay constants is presented. Note that the decay becomes shorter with increasing Ce\(^{3+}\) ions concentration due to reduction of distance between exciton and activator ion with increase of cerium ions concentration. In SrF\(_{2}\)-Ce crystals thermoluminescence glow peaks at 200-250 K were found. Broad glow peaks at higher temperatures are shifted to lower temperature with increasing concentration of Ce\(^{3+}\) ions in the SrF\(_{2}\) crystal. Long stages in scintillation time profile can be attributed to thermoactivated processes related to electron or hole delayed transfer to the activator ion. A similar energy transfer mechanism has been observed in SrF\(_{2}\) doped with Pr. Therefore, the difference in light outputs measured from x-ray luminescence and pulse height spectra is explained by presence of intensive slow components in cerium ions luminescence (see Fig. [\[decay\]](#decay){reference-type="ref" reference="decay"}). They give a large contribution to total light output. Shaping time of pulse height spectrum measurement is 10 \(\mu\)s, and a large part of emitted light is not registered whereas a rate of x-ray excited luminescence spectra registration is amount about 1 s\(^{-1}\) that leads to much more light registration. Light output of SrF\(_2\)-Ce\(^{3+}\) samples can be increased by decreasing of slow component contribution in Ce\(^{3+}\) luminescence. It might be possible by co-doping these crystals with Ga\(^{3+}\), In\(^{3+}\) or Cd\(^{2+}\) ions to change band gap in SrF\(_{2}\)-Ce crystal and, thence, reducing the role of traps in scintillation energy transfer. This idea has been made in garnets by [@Fasoli11]. However, Cd\(^{2+}\) ions bring to STE suppressing in alkali-earth fluorides. Consequently, Cd\(^{2+}\) co-doping leads to suppression of efficient resonance exciton energy transfer mechanism and following light output reducing. Therefore, Cd\(^{2+}\) co-doping is not eligible way for increasing light output of SrF\(_{2}\)-Ce scintillator. A role of Ga\(^{3+}\) and In\(^{3+}\) impurities in STE suppressing has not yet been investigated and follow-up study of the crystals doped with Ga\(^{3+}\) and In\(^{3+}\) ions is required. # Conclusion The SrF\(_2\) crystal are well suited for use as gamma radiation detector. It has a higher (4.18 g/cm\(^{3}\)) than NaI-Tl (3.67 g/cm\(^{3}\)) density, comparable light output, and it is no hygroscopic. Taking into account that crystals SrF\(_{2}\)-0.3 mol.% Ce\(^{3+}\) and SrF\(_{2}\)-1 mol.% Ce\(^{3+}\) have a high temperature stability of light output in the temperature interval from-50 \(^\circ\)C to 200 \(^\circ\)C, these materials can be applied in well-logging scintillation detectors. Summarizing the experimental results we conclude that strontium fluoride crystals would be useful as newly perspective scintillator.

{'timestamp': '2012-10-16T02:02:44', 'yymm': '1210', 'arxiv_id': '1210.3755', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3755'}

# Introduction A blind search for pulsars along the Galactic plane covering 10% of the region between Galactic longitude 45\(^{\circ} <\) l \(<\) 135\(^{\circ}\) and Galactic latitude 0\(^{\circ} < |\)b\(| <\) 5\(^{\circ}\) was recently carried out with the Giant Metrewave Radio Telescope (GMRT). It was named 'GMRT Galactic Plane Pulsar and Transient Survey'. The survey was carried out at 325 MHz with a bandwidth of 16 MHz, divided into 256 filterbank channels. Each field (circular region of radius \(\sim\) 1\(^{\circ}\)) was observed for 1800 s with a sampling time of 256 \(\mu\)s.\ #### The advantage of observing at 325 MHz was the wide field of view (\(\sim\) 4 deg\(^2\)) paired with a sensitivity of 0.6 mJy. The observations were carried out using the incoherent array (IA) mode of the GMRT. The data were written to magnetic tapes. They were extracted to network attached storage (NAS) disks of the high performance computing cluster having 64 dual core nodes. Pulsar search was done using SIGPROC[^1] with extensive RFI mitigation algorithms written by one of the authors. The trial DM range used for the search was 0\(-\)`<!-- -->`{=html}1200 pc-cm\(^{-3}\). The analysis was parallelised such that the DM search for a particular field was divided among the nodes. The results were written back to the NAS disks from where, they could be copied to other machines. The candidate plots thus generated, were manually scrutinised for identifying good candidates.\ #### The follow-up timing observations continued with a new software back-end at GMRT. This provided 512 filterbank channels across a bandwidth of 33 MHz with 122.88 \(\mu\)s sampling. The integration time was 1800 s. The timing analysis was done using TEMPO2[^2] (). # Results PSR J1839+15 came out as a strong candidate and was successfully confirmed in later follow-up observations. The accumulated profile from 23 detections on the pulsar is shown in Figure [\[accprof\]](#accprof){reference-type="ref" reference="accprof"}. It shows a very narrow peak. A weaker second component can be seen just before the main pulse. The region before this component may also have another, still weaker component buried in the noise. Overall, it may consist of two or more components although polarization studies are required to confirm the same. The timing solution obtained so far is given in Table [1](#param){reference-type="ref" reference="param"}\ Numbers in brackets indicate 2\(\sigma\) errors as reported by TEMPO2 in the last digit of the given value. The error on the DM comes from local search done on the time series data. [\[param\]]{#param label="param"} ::: #### During the follow-up timing observations, the pulsar could not be detected for 278 days from 30\(^{th}\) August 2012 to 13\(^{th}\) June 2012. It could be detected regularly then onwards until on 2\(^{nd}\) September 2012, when it was again not detected. The estimated mean flux densities for the detections are plotted in Figure [\[fluvar\]](#fluvar){reference-type="ref" reference="fluvar"}. As can be clearly seen, the 8\(\sigma\) upper limits on non-detections are below the 98% confidence limit on the expected flux density. Thus, we are fairly confident that PSR J1839+15 is an intermittent pulsar with an OFF time scale of roughly 278 days. We are currently working on calculating the Ṗ in the ON and OFF states. Two different values of Ṗ would confirm this as an intermittent pulsar. There are only three more such pulsars known currently. PSR B1931+24 shows a quasi periodic ON-OFF cycle of about 30-40 days (), PSR J1841\(-\)`<!-- -->`{=html}0500 shows an OFF time scale of 580 days () while J1832+0029 remained OFF for 650 days and 850 days in the two sampled OFF states (). The reason for this behavior is not understood. Further investigations and discoveries of new members of this class may shed some light on the underlying physics. # Discussion Intermittent pulsars are a rare breed of pulsars showing very long period nulls. The cause of these nulls is a mystery. To add to the already puzzling phenomenon, it was reported (, and ) that the spin-down rate is considerably higher in the ON state than in the OFF state indicating that the particle flow forming the magnetospheric currents, is different in the ON and OFF states. Another puzzling fact is that these pulsars belong to the normal pulsar population in the P\(-\)Ṗ diagram. The cause of cessation of radio emission altogether may be attributed to the reduced particle flow or it may just be a failure of coherent emission. These pulsars certainly challenge the current emission models and if well studied, may provide vital inputs for coming up with more realistic emission mechanisms.\ #### Given the ON-OFF nature of these pulsars, they provide great motivation for extending the existing blind searches and embarking on new, sensitive blind searches even in the previously searched areas of sky. Assuming a typical duration of a big survey as 3 to 4 years and given the OFF state time scales of these pulsars, a rough estimate of the intermittent pulsar population may go up considerably, thus opening up the possibility of discovering many more of such objects.

{'timestamp': '2012-10-17T02:04:11', 'yymm': '1210', 'arxiv_id': '1210.3784', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3784'}

# Introduction One of the many definitions of the Alexander polynomial of a knot is through state sums. Kauffman has described and studied a state sum model for the Alexander polynomial in great detail. In an earlier paper Crowell has described another state sum model for the Alexander polynomial for the subclass of alternating knots (, Theorem 2.12). In the next section we will recall the definition of Crowell states and examine some of their properties. In Section [3](#sec:proof){reference-type="ref" reference="sec:proof"} we will prove that This theorem is similar in nature to the Clock Theorem of Kauffman which states that any two Kauffman states differ by a finite sequence of clockwise and counterclockwise moves, which was also proven in the language of graphs in section 4 of. This work is independent of those mentioned because of the simple reason that Kauffman states and Crowell states do not correspond to each other in any natural way as observed from the fact that the space of Crowell states do not form a lattice in general (see Proposition [\[prop:notlattice\]](#prop:notlattice){reference-type="ref" reference="prop:notlattice"}). In section [4](#sec:application){reference-type="ref" reference="sec:application"}, as an application we will give an alternative proof that \((2,2n+1)\) torus knots are characterized by their Alexander polynomials among alternating knots, which was originally proven by Ozsváth and Szabó (Proposition 4.1 in ). # A state model {#sec:states} In this section we will review the definition of the state sum model for alternating links given by Crowell and investigate some properties of the states. Given a knot \(K\) and an oriented alternating diagram \(D\) of \(K\) with \(n\) crossings we obtain a weighted labeled directed planar graph \(G(D)\) as follows: replace a small neighborhood of each crossing by a degree \(4\) vertex according to the following figure (\(k\) is the vertex label): Choose a vertex \(v_0\) of \(G(D)\). Spanning trees rooted at \(v_0\) (edges are directed away from \(v_0\)) will be called states. Let \(Tr(v_0)\) be the space of states and \(w(T)\) be the product of weights of all edges in a state \(T\). According to we get the renormalized Alexander polynomial as a sum of monomials corresponding to each state by \[\displaystyle \Delta_K(t) = (-t)^m \cdot \sum_{T\in Tr(v_0)}w(T) \label{eq:statesum}\] where \(m\) is chosen so that the term with the least power of \(t\) is a positive constant. The state sum in Equation [\[eq:statesum\]](#eq:statesum){reference-type="ref" reference="eq:statesum"} resembles the state sum defined by Kauffman. Kauffman has studied an operation called *clock move* that transforms a state to another that differ only at two crossings and showed that all states differ from one another by a sequence of clock moves. With that in mind we define the following operation for reduced alternating diagrams: Edge exchange gives a partial order on the set of states by defining the covering relation of the partial order as \(T_1\) comes immediately before \(T_2\) if \(T_2\) is obtained from \(T_1\) by one positive edge exchange. Comparing these states with the black trees in Kauffman states, even though states are rooted spanning trees in both models, in Kauffman states the orientations on the edges are chosen after a spanning tree of the black graph is chosen, so the same edge can inherit different orientations in different states. Furthermore, consider the graph whose vertices are Crowell states and any two vertices are connected by an edge if there is a terminal edge exchange that takes one state to the other. The following proposition shows that edge exchanges do not correspond to clock moves under any bijection between the Crowell and Kauffman states since Kauffman states form a distributive lattice. # Proof of Theorem [\[thm:connected\]](#thm:connected){reference-type="ref" reference="thm:connected"} {#sec:proof} In this section we will assume that \(K\) is a prime knot and \(D\) is a reduced alternating projection for \(K\). Choose a root vertex \(v_0\) in \(G(D)\). We will provide an algorithm to go from one rooted spanning tree \(T_1\) to another rooted spanning tree \(T_2\) through a sequence of edge exchanges. We will label vertices \(v\) of \(G(D)\) with distinct integers. An initial segment \(IS(v,T)\) of a rooted spanning tree \(T\) is the sequence of vertices on the unique rooted path from the root to \(v\) in \(T\). For \(v\neq v_0\), let \(\phi(v,T)\) denote the vertex that points to \(v\) through an edge not in \(T\). Let \(Bel(v,T)\) be a small neighborhood of the set of vertices below \(v\) in \(T\), i.e., those that can be reached from \(v\) via directed paths in \(T\), the edges between them (not necessarily in \(T\)) and the elementary regions surrounded by those edges. Let \(Bel_1(w,T)\) be the connected component of \(Bel(w,T)\) containing the successor of \(w\) in \(T\) with the smaller label. When the tree \(T\) is obvious from the context, we will suppress \(T\) from these notations. The rooted meet of two rooted trees \(T_1\) and \(T_2\) is the connected component of the root in \(T_1 \cap T_2\) and will be denoted by \(\rootedmeet\). In order to prove Theorem [\[thm:connected\]](#thm:connected){reference-type="ref" reference="thm:connected"}, we will show that for any given two states \(T_1\) and \(T_2\), we can persistently enlarge \(\rootedmeet\). # An application to \((2,2n+1)\) torus knots {#sec:application} In this section we will provide a different proof of the following result originally proved by Ozsváth and Szabó: Let \(D\) be a reduced alternating projection for a knot \(K\) with \(\Delta_K(t)=1+(-t)+(-t)^2+...+(-t)^{2n}\). Since all coefficients of powers of \(-t\) are \(+1\), each state has a different weight and \(\Delta_K(t)\) is not a product of two alternating knot polynomials (c.f. ), hence \(K\) is prime. Let \(T_0\) be the state with the least \(t\) power. Since \(K\) is prime and the fact that each edge exchange changes the power of \(-t\) by \(\pm1\), using Theorem [\[thm:connected\]](#thm:connected){reference-type="ref" reference="thm:connected"} we get a linear ordering on the \(2n+1\) states starting at \(T_0\), reaching each next state by exchanging an edge of weight \(+1\) with an edge of weight \(-t\). According to Proposition [\[prop:terminaledges\]](#prop:terminaledges){reference-type="ref" reference="prop:terminaledges"} and due to this linear order, \(T_0\) and the top state \(T_{2n}\) have only one terminal edge each, hence they have no branching, whereas intermediate states have \(2\) terminal edges. Since Proposition [\[prop:statewithprescribedterminalvertex\]](#prop:statewithprescribedterminalvertex){reference-type="ref" reference="prop:statewithprescribedterminalvertex"} tells that each edge \(v\) (except the two that point to \(v_0\)) can be extended to a state having \(v\) as a terminal edge, and since we can reach that state from \(T_0\) by positive edge exchanges, we see that all edges in \(T_0\) have weight \(+1\). We conclude that \(T_0\) has \(2n\) edges since each edge of weight \(+1\) is used only once in an edge exchange and no new edges emerge with weight \(+1\) as we go from \(T_0\) to \(T_{2n}\). Edge orientations and weights do not depend on the choice of the root vertex, hence, after moving the root from \(v_0\) to \(v_1\), we still get a space of states with the same properties, in particular, there will be a new state \(T'_0\) containing a linear directed chain of \(2n\) vertices starting at \(v_1\), ending at \(v_0\). Hence we get a cycle of length \(2n+1\) of edges of weight \(+1\). Similarly, all remaining edges have weight \(-t\), form a loop and are used in \(T_{2n}\), except the one pointing at the root. This information tells us that if there is an incoming edge of weight \(-t\) at a vertex \(v\), the next edge of weight \(-t\) has to be on the same side of the loop of \(+1\) edges due to the cyclic alternating orientation of edges at a vertex. Since these edges with weight \(-t\) form a loop as well, they have to go between consecutive vertices. This gives us the diagram for the \((2,2n+1)\) torus knot. 0◻ I would like to thank Bedia Akyar for the invitation to give a talk at Dokuz Eylül University, during which time period I started exploring the properties of the Crowell state space. Most of the work was done during my time at Ferris State University. I would also like to thank Mahir Bilen Can and Mohan Bhupal for providing feedback.

{'timestamp': '2012-10-16T02:03:21', 'yymm': '1210', 'arxiv_id': '1210.3798', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3798'}

# Introduction {#introduction .unnumbered} The continuing financial crisis has focused particular attention to systemic risks related to interbank network. The corresponding literature includes papers analyzing real interbank networks, theoretical discussions and modelling and discussion of prudential measures. At the conceptual level the research in this area is based on a theory of complex networks, see e.g.. The main goal of the present study is to examine the structure of the Russian interbank network and the corresponding systemic risks related to possible default of one of the banks and the volume of contagion triggered by this event. In our analysis we use the data on overnight interbank transactions of 767 banks from August 1 2011 till November 3 2011. The choice was made in such a way that each bank had at least one transaction within the considered time period. Only transactions corresponding to borrowing (lending) money without any collateral were taken into account. # Network structure {#network-structure .unnumbered} Let us turn to a more detailed description of the global properties of the Russian interbank networks. As has been already mentioned, the database includes all 767 banks having at least one transaction within the considered period of 69 days. The network of interbank interactions is defined as follows. The nodes of a network stand for banks. A (directed) link between two nodes describes an interbank interaction involving two parties, a borrower and a lender. In what follows we use a standard definition where the link is directed from a borrower to a lender. In network terms lending money to a counterparty creates an outgoing link and borrowing money-an incoming one. In addition, each link is characterized by the amount of money borrowed (lent). An interbank network is thus fully characterized by a directed weighted graph \(G^{W}=(N,W)\), where \(N\) is the number of the nodes and \(W=\{w_{ij}\}\) is an \(N\times N\) matrix of interbank exposures where \(w_{ij}>0\) is a total obligation of the bank \(i\) to the bank \(j\). Let first analyze the gross geometrical features of the interbank network under consideration. The simplest characteristics of a network is a probability \(p\) of having a link which, for a network with \(N\) vertices and \(K\) links can be estimated as \[p \; = \; \frac{2K}{N(N-1)}\] For the Russian interbank network under consideration the average value of \(p\) is \(\langle p \rangle \sim 0.0037\). Another important characteristics of a graph is a clustering coefficient \(C\)[^1] which can conveniently be defined as a ratio of a number of actually existing links between the \(z\) nearest neighbors of a vertex and their total possible number \(z(z-1)/2\). It is clear that for a totally random graph one has \(C=p\). For the network under consideration the averaged clustering coefficients for incoming and outgoing clusters \(C^{In}=0.035\) and \(C^{Out}=0.012\) respectively showing a significant amount of clustering. Let us now turn to the global characteristics of the trading pattern corresponding to a characteristic interbank network. The simplest characteristics of an overall activity is the mean number of banks that are active or, equivalently, the mean number of banks that are passive on a given day. The corresponding values are 470 and 297 respectively, so that on a typical day we have a network of 470 active banks. These latter can be active in a different fashion. At two opposite poles are pure lenders, i.e. vertices with incoming links only-on average, 299 vertices per day and pure borrowers, i.e. vertices with outgoing links only-on average, 92 vertices. The remaining 79 vertices serve as connectors between borrowers and lenders, i.e. borrow and lend at the same time. A more detailed description of an average daily \"in-out\" pattern is presented in the Table [1](#tab:decomposition){reference-type="ref" reference="tab:decomposition"}. ::: The columns in Table [1](#tab:decomposition){reference-type="ref" reference="tab:decomposition"} correspond to an average daily number of vertices of given type (In, Out, etc.) satisfying some certain condition. In parentheses we show the corresponding standard deviations. As has been already mentioned, the systemic risk associated with an interbank network refers to propagation of defaults triggered by the default of one or several banks (vertices) and propagating along outgoing links. In this context the properties of the Russian interbank market characterized by the Table [1](#tab:decomposition){reference-type="ref" reference="tab:decomposition"} lead us to the following observations: 1. The number of pure lenders is almost twice as large as that of pure borrowers. This feature creates specific systemic risks because a default of any borrower may lead to defaults of several lenders. 2. Of special interest are those 29 banks which are characterized by large (\(k_{\rm out}>10\)) values of their out-degree. These banks are clearly especially important sources of systemic risk. Let us note that these banks accumulate \(63\%\) of the total systemic debt. 3. Let us also point at those \(14\) banks that have more than \(10\) incoming links. From the network perspective these banks play a role of hubs absorbing potential shocks due to their loan's diversification. These banks control \(37\%\) of the total loan. An important generic feature of a directed network are the probability distributions \(P(k^{\rm \; in})\) and \(P(k^{\rm out})\) for the number of incoming and outgoing links for a vertex. The majority of networks discussed in the literature, see e.g., are the so-called scale-free ones, i.e. have powerlike tails \(P(k)\sim {\rm const} / k^{\gamma}\). The corresponding marginal in-and out-degree distributions for the Russian interbank market are shown in Figs. [\[ris:MargInDistr\]](#ris:MargInDistr){reference-type="ref" reference="ris:MargInDistr"} and [\[ris:MargOutDistr\]](#ris:MargOutDistr){reference-type="ref" reference="ris:MargOutDistr"} respectively. We see that the network is scale-free for both distributions with \(\gamma^{\rm \; in}=1.92\) and \(\gamma^{\rm out}=2.64\). Propagation of contagion in a network is crucially dependent on its connectivity. The simplest corresponding characteristics is an average number of in-and out-links \(z_2^{\rm \; in}\) and \(z_2^{\rm \; out}\) of the nearest neighbors of a vertex. The bigger are \(z_2^{\rm \; in}\) and \(z_2^{\rm \; out}\) in comparison with the mean number of in-and out-links for a vertex \(z_1 = z_1^{\rm \; in} = z_1^{\rm \; out}\), the easier is contagion propagation along the corresponding cluster. For the Russian interbank network we have \(z_1=1.41\), \(z_2^{\rm \; in}=9\) and \(z_2^{\rm \; out}=28\). This the condition \(z_2/z_1>2\) for an existence of a giant component holds both for in-and out-clusters. A more detailed information on the network connectivity is given by a conditional probability distribution \(P \left( k_2^{\rm \; in},k_2^{\rm \; out} \vert k_1^{\rm \; in},k_1^{\rm \; out} \right)\) characterizing probability for a nearest neighbor of a vertex with \(k_1^{\rm \; in}\) incoming and \(k_1^{\rm \; out}\) outgoing to have \(k_2^{\rm \; in}\) and \(k_2^{\rm \; out}\) incoming and outgoing links respectively. In Fig. [\[ris:DegreeCorrelation\]](#ris:DegreeCorrelation){reference-type="ref" reference="ris:DegreeCorrelation"} we show two lowest moments of its marginal distributions, namely \(\langle k_2^{\rm \; out} \rangle (k_1^{\rm \; out})\), Fig. [\[ris:DegreeCorrelation\]](#ris:DegreeCorrelation){reference-type="ref" reference="ris:DegreeCorrelation"} (a), and \(\langle k_2^{\rm \; in} \rangle (k_1^{\rm \; out})\), Fig. [\[ris:DegreeCorrelation\]](#ris:DegreeCorrelation){reference-type="ref" reference="ris:DegreeCorrelation"} (b). Both plots show pronounced assortiativity at small \(k_1^{\rm \; out}\) for both \(k_2^{\rm \; in}\) and \(k_2^{\rm \; out}\). # Contagion effect {#contagion-effect .unnumbered} Let us start with formulating the model of default contagion spreading in a bank network we are using in the present study. The original source of risk are banks (vertices) that have loans and default on their payment. The banks immediately affected by such a default are the nearest neighbors of this vertex reachable via outgoing links attached to it. If one of the nearest neighbors also defaults, the process can spread further. A probability of infection depends on the number of incoming and outgoing links of a vertex. In the present study we use a simple stylized model of bank balance sheets from. However, at difference with the analysis of and similarly to, we are working with the real day-by-day topologies of the interbank market. In this model a representative bank has a simple balance sheet structure with interbank and illiquid assets on the assets side and capital, deposits and interbank obligations on the liabilities side shown in Fig. [\[fig:BalanceSheet\]](#fig:BalanceSheet){reference-type="ref" reference="fig:BalanceSheet"}. The corresponding condition for the bank \(i\) to be solvent is \[(1-\phi)A_{i}^{IB}+qA_{i}^{M}-L_{i}^{IB}-D_i > 0, \label{eq:SolvencyCondition}\] where \(A_{i}^{IB}\) denotes interbank assets of bank \(i\), \(A_{i}^{M}\)--its illiquid assets, \(L_{i}^{IB}\)--its interbank liabilities, \(\phi\) is a fraction of banks having obligations with respect to bank \(i\) that have defaulted, and \(q\) is the discount for fire-selling illiquid assets. It is assumed that interbank claims and liabilities of a particular bank are uniformly distributed across its borrowers and creditors so that \(\phi=\frac{1}{j_i}\) where \(j_i\) is the number of borrowers. This assumption highlights an importance of the incoming degree for each bank as reflecting its risk diversification. Banks with a high value of the incoming degree have lower probability to go bankrupt due to contagion effect. At the same time banks with a high value of the outgoing degree can be the sources of contagion. For simplicity neglect the fire discount, i.e. assume \(q=1\) in Eq. [\[eq:SolvencyCondition\]](#eq:SolvencyCondition){reference-type="ref" reference="eq:SolvencyCondition"}. Therefore one can rewrite the solvency condition [\[eq:SolvencyCondition\]](#eq:SolvencyCondition){reference-type="ref" reference="eq:SolvencyCondition"} as follows: \[\frac{K_i}{A_{i}^{IB}}>\frac{1}{j_i}, \label{eq:DefaultCondition}\] where the capital buffer \(K_i\) is defined as \[K_i \; = \; A_{i}^{IB}+qA_{i}^{M}-L_{i}^{IB}-D_i\] Our main goal will be to study the impact of the capital buffer size on the number of banks which default due to contagion. The procedure we use is as follows: 1. We set the values of relevant parameters. We assume that \(A_{i}^{IB}\) makes 20% of the balance sheet and study a range of capital buffer values from 4 to 10 % of the balance sheet. 2. Taking real structure of the overnight interbank market we default each bank and determine the size of the default cluster by checking, using Eq. [\[eq:DefaultCondition\]](#eq:DefaultCondition){reference-type="ref" reference="eq:DefaultCondition"}, whether some of its nearest neighbors that can be reached from the defaulted vertex via outgoing links are infected, etc. For each initial bank \(i\) a default cluster for is the number of banks defaulted due to the default of \(i\) as a result of contagion process. 3. Finally, for each value of the capital buffer we calculate an average over default cluster. The results of this simulation are presented in Figs. [\[fig:ProbDistrib\]](#fig:ProbDistrib){reference-type="ref" reference="fig:ProbDistrib"} and [\[fig:DefClusDistrib\]](#fig:DefClusDistrib){reference-type="ref" reference="fig:DefClusDistrib"} in which probability dsitribution of default cluster sizes (Fig. [\[fig:ProbDistrib\]](#fig:ProbDistrib){reference-type="ref" reference="fig:ProbDistrib"}) and a dependence of the average default cluster on the value of capital buffer (Fig. [\[fig:DefClusDistrib\]](#fig:DefClusDistrib){reference-type="ref" reference="fig:DefClusDistrib"}) are shown. From the distribution in Fig. [\[fig:ProbDistrib\]](#fig:ProbDistrib){reference-type="ref" reference="fig:ProbDistrib"} we can make quantitative statements on the significance of systemic network-related risks. For example, for the capital buffer size of \(4\%\) of total balance sheet there is a \(1\%\) probability for more than \(8\) banks go bankrupt. The main conclusion that can be drawn from Fig. [\[fig:DefClusDistrib\]](#fig:DefClusDistrib){reference-type="ref" reference="fig:DefClusDistrib"} is that the average default cluster size is rapidly decaying with growing capital buffer. # Conclusion {#conclusion .unnumbered} In this study we have analyzed some systemic network-related properties of the Russian overnight interbank market. A detailed analysis will be published in. [^1]: For simplicity in computing \(C\) we treat the graph as an undirected one.

{'timestamp': '2012-10-16T02:03:31', 'yymm': '1210', 'arxiv_id': '1210.3814', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3814'}

# Introduction Quantum statistics of Raman scattering were discussed from various points of view in a number of papers (see and references therein, Section 10.4 of and for reviews). In this paper we describe the Raman scattering process with a completely quantum mechanical Hamiltonian. The model is capable to include the stimulated, spontaneous and partially spontaneous Raman process. We use the second order short-time approximation for solution of the Heisenberg equation of motion corresponding to this Hamiltonian. The solution is then used to relate the nonclassical properties of photons and phonons in these processes (i.e. stimulated, spontaneous and partially spontaneous Raman process). To be precise, nonclassical characteristics of photon and phonon modes generated in these processes are exhibited through a number of properties, including squeezing of vacuum fluctuations, quantum entanglement of modes, sub-shot noise and wave variances. Further, joint photon-phonon number, integrated-intensity distributions, conditional and difference number distributions are also found useful to illustrate the observed nonclassicalities. Remaining part of the paper is arranged as follows. In Section [\[sec:The-model-Hamiltonian\]](#sec:The-model-Hamiltonian){reference-type="ref" reference="sec:The-model-Hamiltonian"} we briefly describe the fully quantum Hamiltonian of Raman process and several criteria of nonclassicalities that are used in the present study. In Section [\[sec:The-model-Hamiltonian\]](#sec:The-model-Hamiltonian){reference-type="ref" reference="sec:The-model-Hamiltonian"} we have briefly described the model Hamiltonian that provides a completely quantum description of the Raman scattering process and also describe a normal-ordered characteristic function in the Gaussian form for the same system. Criteria for testing of nonclassicalities are also introduced. In Section [\[sec:Phonon-mode-is-cohrent\]](#sec:Phonon-mode-is-cohrent){reference-type="ref" reference="sec:Phonon-mode-is-cohrent"} we investigate nonclassical character of Raman process for coherent scattering by means of intermodal entanglement, single mode and intermodal squeezing of vacuum fluctuations, sub-shot noise and wave variances. In Section [\[sec:Phonon-mode-is-chaotic\]](#sec:Phonon-mode-is-chaotic){reference-type="ref" reference="sec:Phonon-mode-is-chaotic"} the same nonclassical characteristics are investigated for the chaotic scattering. The observed nonclassicalities are further illustrated through joint photon-phonon number and wave distribution in Section [\[sec:Joint-photon-phonon-number\]](#sec:Joint-photon-phonon-number){reference-type="ref" reference="sec:Joint-photon-phonon-number"}. In Section [\[sec:Difference-and-conditional\]](#sec:Difference-and-conditional){reference-type="ref" reference="sec:Difference-and-conditional"} we study difference and conditional number distributions associated with the Raman process. Finally, Section [\[sec:Conclusion\]](#sec:Conclusion){reference-type="ref" reference="sec:Conclusion"} is dedicated to conclusion. # The model Hamiltonian and the criteria of nonclassicality[\[sec:The-model-Hamiltonian\]]{#sec:The-model-Hamiltonian label="sec:The-model-Hamiltonian"} A fully quantum description of the Raman scattering process can be provided by the following Hamiltonian : \[H=\sum_{j=L,S,A,V}\hbar\omega_{j}a_{j}^{\dagger}a_{j}-\left(\hbar ga_{L}a_{S}^{\dagger}a_{V}^{\dagger}+\hbar\chi^{*}a_{L}a_{V}a_{A}^{\dagger}+{\rm h.c.}\right),\label{eq:hamiltonian}\] where \({\rm h.c.}\) stands for Hermitian conjugate and the subscripts \(L,S,A\) and \(V\) correspond to pump (laser), Stokes, anti-Stokes and vibration (phonon) modes respectively, \(\omega_{j},\, a_{j}\) and \(a_{j}^{\dagger}\) are frequency, annihilation operator and creation operator in the \(j\) th mode, \(g\) and \(\chi\) are the Stokes and anti-Stokes coupling constants. Using the Hamiltonian ([\[eq:hamiltonian\]](#eq:hamiltonian){reference-type="ref" reference="eq:hamiltonian"}) we can construct a set of Heisenberg equations and solve them in short-time approximation. A second order short-time approximated solution was already reported. It is interesting to note that using the short-time approximated solution, we can obtain a normal-ordered characteristic function in the Gaussian form. Such a characteristic function can completely characterize the Raman process, and it can be analytically expressed as \[\begin{array}{lcl} C_{N}(\beta_{L},\beta_{S},\beta_{A},\beta_{V},t) & = & \left\langle \exp\left\{ \sum_{j=L,S,A,V}\right.\left[-B_{j}(t)|\beta_{j}|^{2}\right.\right.\\ & + & \left.\left(\frac{1}{2}C_{j}^{*}(t)\beta_{j}^{2}+{\rm c.c.}\right)+\beta_{j} & + & \left.\left.\sum_{j<k}\left(D_{jk}(t)\beta_{j}^{*}\beta_{k}^{*}+\bar{D}_{jk}(t)\beta_{j}\beta_{k}^{*}+{\rm c.c.}\right)\right\} \right\rangle ,\end{array}\label{eq:characteristic function}\] where \({\rm c.c.}\) stands for complex conjugate terms, the set \((L,\, S,\, A,\, V)\) is assumed to be ordered and \[\begin{array}{lcl} B_{L}(t) & = & |\chi|^{2}t^{2}| B_{S}(t) & = & |g|^{2}t^{2}| B_{A}(t) & = & 0,\\ B_{V}(t) & = & |g|^{2}t^{2}| C_{L}(t) & = &-g^{*}\chi t^{2} C_{S}(t) & = & 0,\\ C_{A}(t) & = & 0,\\ C_{V}(t) & = &-g\chi t^{2} D_{LS}(t) & = &-\frac{1}{2}|g|^{2}t^{2} D_{LA}(t) & = &-\frac{1}{2}|\chi|^{2}t^{2} D_{LV}(t) & = & \left[i\chi t D_{SA}(t) & = &-\frac{1}{2}g\chi^{*}t^{2} D_{SV}(t) & = & \left(igt D_{AV}(t) & = &-\frac{1}{2}|\chi|^{2}t^{2} \bar{D}_{LS}(t) & = &-\frac{1}{2}g\chi^{*}t^{2} all other \(\bar{D}_{jk}=0;\) \( coherent complex amplitudes. As the above characteristic function is Gaussian consequently (\ref{eq:BCD and Dbar}) can be used to obtain normal fluctuation quantities (variances) \)\left\langle \left(\Delta W_{j}\right)^{2}\right\rangle _{N}\( and \)\langle\Delta W_{j}\Delta W_{k}\rangle_{N}\(, which are defined as: \begin{equation} \begin{array}{lcl} \left\langle \left(\Delta W_{j}\right)^{2}\right\rangle _{N} & = & \left\langle a_{j}^{\dagger2}(t)a_{j}^{2}(t)\right\rangle-\left\langle a_{j}(t)a_{j}(t)\right\rangle ^{2}\\ & = & \left\langle B_{j}^{2}+|C_{j}|^{2}+2B_{j}| and \begin{equation} \begin{array}{lcl} \langle\Delta W_{j}\Delta W_{k}\rangle_{N} & = & \left\langle a_{j}^{\dagger}(t)a_{k}^{\dagger}(t)a_{j}(t)a_{k}(t)\right\rangle-\left\langle a_{j}^{\dagger}(t)a_{j}(t)\right\rangle \left\langle a_{k}^{\dagger}(t)a_{k}(t)\right\rangle \\ & = & \left\langle |D_{jk}|^{2}+|D_{jk}|^{2}+\left[D_{jk} Brackets on the right-hand side in (\ref{eq:characteristic function}, \ref{eq:variance} and \ref{eq:cross variance}) mean an average over the initial amplitudes. Equations (\ref{eq:BCD and Dbar})-(\ref{eq:cross variance}) provide us with sufficient mathematical framework required for the study of the nonclassical character of stimulated and spontaneous Raman process. This is so because the criteria for various nonclassical phenomena can be conveniently expressed in terms of the quantities described in (\ref{eq:BCD and Dbar})-(\ref{eq:cross variance}). For example, we may note the criteria for principle squeezing of vacuum fluctuations in single mode (\)j)\( and compound mode (\)ij),\( which are: \begin{equation} \lambda_{j}=1+2(B_{j}-|C_{j}|)<1\label{eq:single-mode squuezing}\] and \[\lambda_{ij}=1+B_{i}+B_{j}-2{\rm Re}\bar{D}_{ij}-|C_{i}+C_{j}+2D_{ij}|<1\label{eq:squeezing compund mode}\] respectively. From the above two criteria it is clear that ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}) provides us sufficient input for analytic study of the principle squeezing of vacuum fluctuations, both in single modes and in compound modes. Similarly, condition for entanglement is in general \[\left(K_{ij}\right)_{\pm}=(B_{i}\pm|C_{i}|)(B_{j}\pm|C_{j}|)-\left(|D_{ij}|\mp|\bar{D}_{ij}|\right)^{2}<0,\label{eq:entanglement}\] and condition for sub-shot noise is \[C_{ij}=B_{i}^{2}+B_{j}^{2}+|C_{i}|^{2}+|C_{j}|^{2}-2|D_{ij}|^{2}-2|\bar{D}_{ij}|^{2}<0;\label{eq:subshot noise}\] further the condition for nonclassical sum-or difference-variance is \[\left\langle \left(\Delta W_{ij}\right)^{2}\right\rangle _{\pm}=\left\langle \left(\Delta W_{i}\right)^{2}\right\rangle _{N}+\left\langle \left(\Delta W_{j}\right)^{2}\right\rangle _{N}\pm\langle\Delta W_{j}\Delta W_{k}\rangle_{N}<0.\label{eq:varianceplus and minus}\] Present work aims to rigorously investigate the presence of different nonclassicalities in the Raman process in the second order short-time approximation. To begin with we will discuss intermodal entanglement in the next section. Before we present our analytic results it is important to note that for the convenience of understanding the process we have introduced following two scaled quantities: \)gt=\tau\( and \)\frac{|\chi|}{|g|}=p\(. The time evolution of various nonclassical characteristics can now be expressed with respect to dimensionless quantity \)gt=\tau\( and the ratio between the Stokes and anti-Stokes coupling constants \)p\(. Further we have used \)I_{j}=| for the incident stimulating intensities and the phases of the complex amplitude\( that are combined as \[\phi_{L}-\phi_{V}-\phi_{S}=\theta_{2}\] and \[\phi_{A}-\phi_{L}-\phi_{V}=\theta_{1},\] where \)\theta_{2}\( and \)\theta_{1}\( can be visualized as the mismatch phases in Stokes (\)\omega_{S}=\omega_{L}-\omega_{V})\( and in anti-Stokes (\)\omega_{A}=\omega_{L}+\omega_{V})\( transitions, respectively. In the following the coupling constants \)g\( and \)\chi$ are assumed to be real. # Phonon mode is coherent[\[sec:Phonon-mode-is-cohrent\]]{#sec:Phonon-mode-is-cohrent label="sec:Phonon-mode-is-cohrent"} In the above discussion, all the modes including the phonon mode are coherent. In such a situation we may investigate the existence of different kind of nonclassicalities by using ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"})-([\[eq:varianceplus and minus\]](#eq:varianceplus and minus){reference-type="ref" reference="eq:varianceplus and minus"}). The same is done in the following subsections, where intermode entanglement, single mode and intermode squeezing, sub-shot noise and variances are studied in detail. ## Intermodal entanglement[\[sec:Intermodal-entanglement\]]{#sec:Intermodal-entanglement label="sec:Intermodal-entanglement"} Substituting ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}) in ([\[eq:entanglement\]](#eq:entanglement){reference-type="ref" reference="eq:entanglement"}) we obtain \[\left(K_{LV}\right)_{+}=\left(K_{LV}\right)_{-}=-p^{2}\tau^{2}I_{A}=E_{LV}\label{eq:k-lv}\] and \[\left(K_{SV}\right)_{+}=\left(K_{SV}\right)_{-}=-\tau^{2}I_{L}=E_{SV}.\label{eq:k-sv}\] \(\left(K_{ij}\right)_{\pm}=0\) for all the other cases. Since we are using a second order short-time approximated solution we cannot conclude anything about the separability of those four modes**[ ]{style="color: blue"}**for which \(\left(K_{ij}\right)_{\pm}=0\). But we can clearly see that in stimulated Raman process (where \(I_{A}\neq0\), \(I_{L}\neq0\), \(I_{S}\neq0\), \(\, I_{V}\neq0\) ) the vibration-phonon mode is entangled with the pump-mode and the Stokes mode and it does not depend on \(I_{S}\) and \(I_{V}\). Consequently if we think of a partially spontaneous Raman process with \(I_{A}\neq0,\, I_{L}\neq0,\, I_{S}=0\), \(I_{V}=0,\) then also we will observe both type of photon-phonon entanglement that we have observed in stimulated Raman process. Interestingly in the completely spontaneous process (where \(I_{A}=0\), \(I_{L}\neq0\), \(I_{S}=0\),  \(I_{V}=0\)) we can also observe entanglement between Stokes mode and phonon mode, but in such situation we cannot conclude about the separability of the pump mode and the phonon mode. ## Single mode and intermodal squeezing[\[sec:Single-mode-and-squuezing\]]{#sec:Single-mode-and-squuezing label="sec:Single-mode-and-squuezing"} Substituting ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}) in ( [\[eq:single-mode squuezing\]](#eq:single-mode squuezing){reference-type="ref" reference="eq:single-mode squuezing"}) and ([\[eq:squeezing compund mode\]](#eq:squeezing compund mode){reference-type="ref" reference="eq:squeezing compund mode"}) we obtain in the interaction picture reflecting the compensation of exponential function in ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}) in a homodyne detection \[\begin{array}{lcl} \lambda_{L} & = & 1+2p\tau^{2}| \lambda_{LA} & = & 1+p^{2}\tau^{2}| \lambda_{SA} & = & 1+I_{L}\tau^{2}(1-p),\\ \lambda_{AV} & = & 1+\tau^{2}\left(I_{L}+p^{2}I_{A}-p^{2}| \lambda_{LV} & = & 1+2p^{2}\tau^{2}I_{A}+\tau^{2}I_{L}-2\left[p^{2}\tau^{2}I_{A}+\frac{(1+p^{2})^{2}}{4}\tau^{4}I_{L}I_{V}\right]^{\frac{1}{2}}-2p\tau^{2}| & \approx & 1+2p^{2}\tau^{2}I_{A}+\tau^{2}I_{L}-2p\tau I_{A}-2p\tau^{2}| \lambda_{LS} & = & 1+\tau^{2}\left(p^{2}I_{A}+I_{L}-| & \approx & 1+\tau^{2}I_{L},\\ \lambda_{SV} & = & 1+2\tau^{2}I_{L}+p^{2}\tau^{2}I_{A}-2\left[\tau^{2}I_{L}+\left(\frac{1}{2}\tau^{2}| & \approx & 1+2\tau^{2}I_{L}+p^{2}\tau^{2}I_{A}-2\tau| From the above equations one can easily obtain the following conditions: \begin{enumerate} \item Since \(| if \)|g|>|\chi|,\( otherwise it is squeezed if \)p| which is expected to be satisfied in most cases. \item \(\lambda_{LA}<1\) in stimulated Raman process if \(| which is the natural case. So intermodal squeezing between pump and anti-Stokes mode can be observed in stimulated Raman process. In spontaneous Raman process \)\lambda_{LA}=1\( so squeezing is not observed, but in partial spontaneous Raman process with \)| squeezing can be observed. \item \(\lambda_{SA}<1\) iff \(p>1,\) i.e. if anti-Stokes coupling is stronger than the Stokes coupling. If \(p>1\) then the intermodal squeezing in Stokes and anti-Stokes modes is observed for both stimulated and spontaneous Raman processes. \item For a completely spontaneous process \(\lambda_{AV}\approx1+\tau^{2}I_{L}\) is always greater than 1. However, also in the stimulated process the term \(\tau^{2}I_{L}\) will be dominant. The same is the\textcolor{red}{{} }case\textcolor{red}{{} }for \(\lambda_{LS}.\) \item For a very short time the linear term in \(\lambda_{LV}\) would dominate and consequently, during that time \(\lambda_{LV}\approx1-2p\tau I_{A}\) will be less than unity and consequently squeezing will be observed in stimulated and partially spontaneous process. \item For a very short time the linear term will dominate in \(\lambda_{SV}\) and consequently\textcolor{blue}{, }during that time \(\lambda_{SV}\approx1-2\tau| would indicate intermodal squeezing in both stimulated and spontaneous process. \end{enumerate} Variation of \)\lambda_{ij}-1\( with respect to \)p\( and \)\tau\( are shown in Fig. \ref{fig:Variation-of-lamij}, which clearly depicts the above observations. \subsection{Sub-shot noise\label{sec:Sub-shot-noise} } When we substitute (\ref{eq:BCD and Dbar}) in (\ref{eq:subshot noise}) we find that \begin{equation} C_{LV}=-2p^{2}\tau^{2}I_{A}=2(K_{LV})_{\pm}\label{eq:clv}\] and \[C_{SV}=-2\tau^{2}I_{A}=2(K_{SV})_{\pm}\label{eq:csv}\] and \)C_{ij}=0\( in the remaining four cases. Since in this particular system \)C_{LV}\( and \)C_{SV}\( are directly proportional to \)(K_{LV})_{\pm}\( and \)(K_{SV})_{\pm}$, wherever we have seen intermodal entanglement we, can also observe sub-shot noise in those cases. ## Variances[\[sec:Variances\]]{#sec:Variances label="sec:Variances"} Using ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"})-([\[eq:cross variance\]](#eq:cross variance){reference-type="ref" reference="eq:cross variance"}) and ([\[eq:varianceplus and minus\]](#eq:varianceplus and minus){reference-type="ref" reference="eq:varianceplus and minus"}) we can obtain the analytic expressions for intermodal variances in the following forms \[\begin{array}{lcl} \left\langle \left(\Delta W_{AV}\right)^{2}\right\rangle _{+} & = & 2\tau^{2}I_{V}\left(I_{L}-p| \left\langle \left(\Delta W_{SA}\right)^{2}\right\rangle _{+} & = & 2\tau^{2}I_{L}\left(I_{S}-p| \left\langle \left(\Delta W_{LA}\right)^{2}\right\rangle _{+} & = &-2p\tau^{2}I_{L}| \left\langle \left(\Delta W_{LS}\right)^{2}\right\rangle _{+} & = & 2p\tau^{2}I_{L}| \left\langle \left(\Delta W_{LV}\right)^{2}\right\rangle _{+} & = & 2p^{2}\tau^{2}\left(3I_{L}I_{A}+3I_{A}I_{V}+I_{A}-I_{L}I_{V}\right)\\ & + & 2p\tau^{2}| \left\langle \left(\Delta W_{SV}\right)^{2}\right\rangle _{+} & = & 2\tau^{2}\left(3I_{L}I_{S}+3I_{L}I_{V}+p^{2}I_{A}I_{V}+I_{L}-I_{S}I_{V}\right)\\ & + & 2p\tau^{2}| and \begin{equation} \begin{array}{lcl} \left\langle \left(\Delta W_{AV}\right)^{2}\right\rangle _{-} & = & 2\tau^{2}I_{V}\left(I_{L}+2p^{2}I_{A}-p| \left\langle \left(\Delta W_{SA}\right)^{2}\right\rangle _{-} & = & 2\tau^{2}I_{L}\left(I_{S}+p| \left\langle \left(\Delta W_{LA}\right)^{2}\right\rangle _{-} & = & 2p\tau^{2}I_{L}\left(2pI_{A}-| \left\langle \left(\Delta W_{LS}\right)^{2}\right\rangle _{-} & = & 2\tau^{2}I_{L}\left(p^{2}I_{A}+2I_{S}-3p| \left\langle \left(\Delta W_{LV}\right)^{2}\right\rangle _{-} & = &-2p^{2}\tau^{2}\left(I_{L}I_{A}+I_{A}I_{V}+I_{A}-I_{L}I_{V}\right)+4\tau^{2}I_{L}I_{V}\\ &-& 6p\tau^{2}| \left\langle \left(\Delta W_{SV}\right)^{2}\right\rangle _{-} & = &-2\tau^{2}\left(I_{L}I_{S}+I_{L}I_{V}-p^{2}I_{A}I_{V}+I_{L}-I_{S}I_{V}\right)\\ &-& 2p\tau^{2}| Negativity of intermodal variances \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{\pm}\) implies nonclassicality. Analytic expressions for intermodal variances \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{+}\) and \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{-}\) for all the possible combinations of modes in the stimulated Raman process are provided in (\ref{eq:wplus}) and (\ref{eq:wminus}), respectively. It is difficult to conclude directly about the presence of nonclassicality from these general analytic expressions of \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{\pm}.\) Thus to visualize the existence of nonclassicality we have plotted the analytic expressions provided in (\ref{eq:wplus}) and (\ref{eq:wminus}). The plots are given in Fig. \ref{fig:wijplus} and Fig. \ref{fig:wijminus} and it is easy to see that both \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{+}\) and \(\left\langle (\Delta W)_{ij}^{2}\right\rangle _{-}\) depicts nonclassical behavior for a) pump and phonon mode and b) pump and anti-Stokes mode, c) Stokes and phonon mode. However for pump and Stokes mode only \(\left\langle (\Delta W)_{LS}^{2}\right\rangle _{+}\) shows the existence of nonclassicality. For the chosen values of \(| but a negative value is possible if \)p| a very strong anti-Stokes coupling (compared to Stokes coupling) may yield nonclassical variance for Stokes and anti-Stokes mode. This is consistent with the appearance of intermodal squeezing where \(\lambda_{AS}-1=I_{L}\tau^{2}(1-p)\) is negative only when \(p>1,\) that is when anti-Stokes coupling is stronger than Stokes coupling. Now from (\ref{eq:wminus}) we can easily observe that for a completely spontaneous Raman process \(\left\langle (\Delta W)_{SV}^{2}\right\rangle _{-}=-2\tau^{2}I_{L}\) is always negative which indicates intermodal nonclassical behavior between phonon mode and Stokes mode. We have already shown that these two modes show intermodal entanglement, sub-shot noise behavior and squeezing of vacuum fluctuations in the spontaneous Raman process. Thus as far as the nonclassicalities in spontaneous Raman process are concerned these two modes play the most important role. From (\ref{eq:wplus}) and (\ref{eq:wminus}) we can see that for very small values of rescaled time \(\tau\) the term linear in \(\tau\) is expected to dominate in \(\left\langle (\Delta W)_{LV}^{2}\right\rangle _{\pm}\) and in \(\left\langle (\Delta W)_{SV}^{2}\right\rangle _{\pm}\); \(\left\langle (\Delta W)_{LV}^{2}\right\rangle _{\pm}\) varies with \(\theta_{1}\), which is exhibited in Fig. \ref{fig:Variation-of-wlv-shorttime}. Further the linear term in \(\left\langle (\Delta W)_{SV}^{2}\right\rangle _{+}\) is very weak and the nonclassical behavior can be seen only for a very small values of \(\tau\). This is why in Fig. \ref{fig:wijplus} we have plotted \(\left\langle (\Delta W)_{SV}^{2}\right\rangle _{+}\) for a very short time only. \section{Phonon mode is chaotic\label{sec:Phonon-mode-is-chaotic} } In this case we perform the average over the initial phonon amplitude in (\ref{eq:characteristic function}) with a Gaussian distribution in the Gaussian approximation. Assuming that the phonon mode is chaotic with average phonon number \(\langle n_{V}\rangle,\) then the coefficients in the interaction picture described in (\ref{eq:BCD and Dbar}) get modified as \begin{equation} \begin{array}{lcl} B_{L} & = & |\chi|^{2}t^{2}| B_{S} & = & |g|^{2}t^{2}| B_{A} & = & |\chi|^{2}t^{2}| B_{V} & \approx & \langle n_{V}\rangle,\\ C_{L} & = &-g^{*}\chi t^{2} C_{S} & = & 0,\\ C_{A} & = & 0,\\ C_{V} & = &-g\chi t^{2} D_{LS} & = &-\frac{1}{2}|g|^{2}t^{2} D_{LA} & = &-\frac{1}{2}|\chi|^{2}t^{2} D_{SA} & = &-\frac{1}{2}g\chi^{*}t^{2} D_{SV} & = & igt D_{LV} & = & i\chi t D_{AV} & = & 0,\\ \bar{D}_{LS} & = &-g\chi^{*}t^{2} \bar{D}_{LA} & = &-g\chi^{*}t^{2} \bar{D}_{SA} & = & 0,\\ \bar{D}_{SV} & = & 0,\\ \bar{D}_{LV} & = & igt \bar{D}_{AV} & = & i\chi t Now using equations (\ref{eq:variance}), (\ref{eq:cross variance}) and (\ref{eq:chaotic-bcdbar}) we can obtain for single-mode variances \begin{equation} \begin{array}{lcl} \langle\left(\Delta W_{L}\right)^{2}\rangle_{N} & = & 2p^{2}\tau^{2}I_{A}I_{L}\left(\langle n_{V}\rangle+1\right)+2\tau^{2}I_{S}I_{L}\langle n_{V}\rangle-2p\tau^{2}| \langle\left(\Delta W_{S}\right)^{2}\rangle_{N} & = & 2\tau^{2}I_{L}I_{S}\left(\langle n_{V}\rangle+1\right),\\ \langle\left(\Delta W_{A}\right)^{2}\rangle_{N} & = & 2p^{2}\tau^{2}I_{L}I_{A}\langle n_{V}\rangle,\\ \langle\left(\Delta W_{V}\right)^{2}\rangle_{N} & \approx & \langle n_{V}\rangle^{2},\end{array}\label{eq:chaotic w-square}\] and for correlation fluctuations \[\begin{array}{lcl} \left(\Delta W_{L}\Delta W_{A}\right)_{N} & = &-\tau^{2}p^{2}I_{L}I_{A}\left(2\langle n_{V}\rangle+1\right)+2p\tau^{2}I_{L}| \left(\Delta W_{L}\Delta W_{S}\right)_{N} & = &-\tau^{2}I_{L}I_{S}\left(2\langle n_{V}\rangle+1\right)+2p\tau^{2}I_{L}| \left(\Delta W_{S}\Delta W_{A}\right)_{N} & = &-\tau^{2}pI_{L}| Analytic expressions of the other cross-correlations are not of interest as all variances that involve phonon mode will always be positive because of the dominance of \(\langle\left(\Delta W_{V}\right)^{2}\rangle\approx\langle n_{V}\rangle^{2}\) term. Now substituting equations (\ref{eq:chaotic-bcdbar}), (\ref{eq:chaotic w-square}) and (\ref{eq:chaotic cross correln}) in the criteria of nonclassicalities introduced in (\ref{eq:single-mode squuezing})-(\ref{eq:varianceplus and minus}) we can investigate the nonclassical character of stimulated and spontaneous Raman process when the phonon mode is chaotic and then compare the results with the similar results obtained in the coherent case. This is done in the following subsections. \subsection{Intermodal entanglement} Substituting (\ref{eq:chaotic-bcdbar}) in (\ref{eq:entanglement}) we obtain \begin{equation} \begin{array}{lcl} \left(K_{SV}\right)_{\pm} & = &-\tau^{2}I_{L}\left(\langle n_{V}\rangle+1\right),\\ \left(K_{LV}\right)_{\pm} & = &-p^{2}\tau^{2}I_{A}\left(\langle n_{V}\rangle+1\right)\mp2p\tau^{2}| All other \(\left(K_{ij}\right)_{\pm}=0\). We can conclude: \begin{enumerate} \item From (\ref{eq:chaotic-kplusminus}) it is clear that the phonon mode is always entangled with Stokes mode. The same characteristic was also observed in coherent case but if we consider \(\left(K_{SV}\right)_{\pm}\) as a measure of amount of entanglement, then the amount of entanglement in chaotic case is increased by a factor of \(\left(1+\langle n_{V}\rangle\right)\) and it is more announced. \item Similarly the phonon mode can be entangled with the pump mode. It is straightforward to see that \(\left(K_{LV}\right)_{+}=-p^{2}\tau^{2}I_{A}\left(\langle n_{V}\rangle+1\right)-2p\tau^{2}| exhibits intermodal entanglement. \item But interestingly \)\left(K_{LV}\right)_{-}$ does not show signature of intermodal entanglement. \item Stokes mode and phonon mode are entangled for completely spontaneous Raman process also but the present calculation is non-conclusive about entanglement of pump and phonon mode. \end{enumerate} \subsection{Single mode and intermodal squeezing } By substituting (\ref{eq:chaotic-bcdbar}) in (\ref{eq:single-mode squuezing}) and (\ref{eq:squeezing compund mode}) we obtain \begin{equation} \begin{array}{lcl} \lambda_{L} & = & 1+2\tau^{2}\left[p^{2}I_{A}\left(\langle n_{V}\rangle+1\right)+I_{S}\langle n_{V}\rangle-p| \lambda_{LA} & \approx & 1+p^{2}\tau^{2}| \lambda_{LS} & \approx & 1+\tau^{2}| \lambda_{SA} & = & 1+\tau^{2}I_{L}(1-p)\left(\langle n_{V}\rangle+1-p\langle n_{V}\rangle\right).\end{array}\label{eq:pustak1}\] We see that: ## Sub-shot noise By substituting ([\[eq:chaotic-bcdbar\]](#eq:chaotic-bcdbar){reference-type="ref" reference="eq:chaotic-bcdbar"}) in ([\[eq:subshot noise\]](#eq:subshot noise){reference-type="ref" reference="eq:subshot noise"}) we obtain \[\begin{array}{lcl} C_{AV} & = & \langle n_{V}\rangle^{2}\left(1-2p^{2}\tau^{2}I_{L}\right),\\ C_{LV} & = & \langle n_{V}\rangle^{2}-2p^{2}\tau^{2}I_{A}\left(\langle n_{V}\rangle+1\right)^{2}-2\tau^{2}I_{S}\langle n_{V}\rangle^{2},\\ C_{SV} & = & \langle n_{V}\rangle^{2}-2\tau^{2}I_{L}\left(\langle n_{V}\rangle+1\right)^{2},\end{array}\label{eq:pustak2}\] and all other \(C_{ij}=0.\) For stimulated Raman process, sub-shot noise is observed in the above three cases. In coherent case subshot noise behavior was not observed for anti-Stokes and phonon mode. Further, negativity of \(C_{AV}\) and \(C_{SV}\) will be observed for spontaneous Raman process too. But in the spontaneous Raman process sub-shot noise behavior will not be observed for pump and phonon modes. However, we can observe it for partially spontaneous process (\(\langle n_{V}\rangle\neq0,\, I_{L}\neq0,\, I_{A}=0\) and \(I_{S}=0\).) ## Variances By substituting ([\[eq:chaotic w-square\]](#eq:chaotic w-square){reference-type="ref" reference="eq:chaotic w-square"}) and ([\[eq:chaotic cross correln\]](#eq:chaotic cross correln){reference-type="ref" reference="eq:chaotic cross correln"}) in ([\[eq:varianceplus and minus\]](#eq:varianceplus and minus){reference-type="ref" reference="eq:varianceplus and minus"}) we obtain \[\begin{array}{lcl} \langle\left(\Delta W\right)_{SA}^{2}\rangle_{\pm} & = & 2p^{2}\tau^{2}I_{L}I_{A}\langle n_{V}\rangle+2\tau^{2}I_{L}I_{S}\left(\langle n_{V}\rangle+1\right)\\ & \mp & 2\tau^{2}pI_{L}| \langle\left(\Delta W\right)_{LS}^{2}\rangle_{\pm} & = & 2p^{2}\tau^{2}I_{A}I_{L}\left(\langle n_{V}\rangle+1\right)+2\tau^{2}I_{L}I_{S}\left(2\langle n_{V}\rangle+1\right)(1\mp1)\\ &-& 2p\tau^{2}I_{L}| \langle\left(\Delta W\right)_{LA}^{2}\rangle_{\pm} & = & 2p^{2}\tau^{2}I_{A}I_{L}\left(\left(2\langle n_{V}\rangle+1\right)\mp\left(\langle n_{V}\rangle+1\right)\right)+2\tau^{2}I_{S}I_{L}\langle n_{V}\rangle\\ &-& 2p\tau^{2}| From (\ref{eq:chaotic variance}) we observe following: \begin{enumerate} \item \(\langle\left(\Delta W\right)_{LS}^{2}\rangle_{+}=2p\tau^{2}| Thus negative variance can be seen for \)p| Since \(| mode and Stokes mode will be observed for small values of mean phonon number \)\langle n_{V}\rangle.\( \item In the analytic expression of \)\langle\left(\Delta W\right)_{AS}^{2}\rangle_{\pm}\( if we assume \)I_{A}\ll I_{S}\( then we obtain \[ \langle\left(\Delta W\right)_{AS}^{2}\rangle_{\pm}=2\tau^{2}I_{L}| which would show nonclassicality if \)p\frac{| This implies \(p| with the assumption \)I_{A}\ll I_{S}\(. Thus if \)I_{A}\ll I_{S}\( then we do not observe nonclassical variance in anti-Stokes and Stokes modes. \item If we assume \)I_{A}\gg I_{S}\( and consider the complete analytic expression of \)\langle\left(\Delta W_{AS}\right)^{2}\rangle_{\pm},\( then the condition \)p\frac{| will serve as necessary but not sufficient condition of nonclassicality. Now if we assume that for some choice of \(p,I_{A},I_{S},\langle n_{V}\rangle\) we observe nonclassical intermodal variance for Stokes and anti-Stokes modes, then we can show that for that situation \(\langle\left(\Delta W_{LS}\right)^{2}\rangle_{+}\) will not show nonclassicality. The proof is simple. First we assume that both \(\langle\left(\Delta W_{LS}\right)^{2}\rangle_{+}\) and \(\langle\left(\Delta W_{AS}\right)^{2}\rangle_{\pm}\) are negative. Therefore, \(\frac{1}{\langle n_{V}\rangle+1}>p\frac{| which implies \)\left(2\langle n_{V}\rangle+1\right)>\left(\langle n_{V}\rangle+1\right)^{2}\( or \)\langle n_{V}\rangle^{2}<0\(. Thus by reductio ad absurdum we have shown that intermodal nonclassical variance cannot be seen simultaneously in a) Stokes and anti-Stokes mode and b) Stokes and pump mode. \item For the compound mode \)(LA)\( one could observe sub-shot noise provided that \)|\chi|| \end{enumerate} \section{Joint photon-phonon number and wave distribution\label{sec:Joint-photon-phonon-number}} We can illustrate the above results for nonclassical behavior of modes in Raman scattering by joint photon-phonon number and integrated-intensity distributions along the lines given in (and references therein) in Gaussian approximation. For simplicity we consider scattering by phonon vacuum (in optical region and for room temperature \(\langle n_{V}\rangle\approx0\)) for compound modes \((SV\) ) and \((LV)\), which exhibit quantum entanglement up to the second order in \(t\). From (\ref{eq:BCD and Dbar}) we see that \(K_{SV}=-B_{S}=-B_{V}=-|g|^{2}t^{2}I_{L}\) provided that we consider spontaneous scattering \((I_{A}=I_{S}=0;\) in this case \(C_{S}=C_{V}=\bar{D}_{SV}=0).\) In principle we can also consider partially stimulated scattering with \(I_{S}\neq0,\) when using shifted distributions in \(W_{S}\) along \(I_{S}\) to adopt spontaneous process. Thus \(K_{SV}+B_{S}=K_{SV}+B_{V}=0\) and from the formulae given in we obtain the joint photon-phonon number distribution \begin{equation} p(n_{S},n_{V})=\frac{(B_{S})^{n_{S}}}{(1+B_{S})^{1+n_{S}}}\delta_{n_{S},n_{V}},\label{eq:pustak3}\] i.e. it is diagonal expressing a pairwise structure of photon-phonon process in this case. It is shown in Fig. [\[fig:pnsnv\]](#fig:pnsnv){reference-type="ref" reference="fig:pnsnv"}a. The corresponding \(s\)-order quasidistribution of integrated-intensities is \[P_{s}\left(W_{S},W_{V}\right)=\frac{1}{\pi B_{Ss}}{\rm e}^{-\frac{W_{S}+W_{V}}{2B_{Ss}}}\frac{\sin\left(\frac{W_{S}-W_{V}}{\sqrt{-K_{VS,s}}}\right)}{W_{S}-W_{V}},\label{eq:pustak4}\] where \(B_{Ss}=B_{S}+\frac{1-s}{2}\) and \(K_{SV,s}=K_{SV}+(1-s)B_{S}+\frac{(1-s)^{2}}{4};\) \(s\) is ordering parameter. For the threshold value of the ordering parameter we have \(s_{th}=1+B_{S}+B_{V}-\sqrt{(B_{S}+B_{V})^{2}-4K_{SV}}\approx1+2B_{S}-2\sqrt{B_{S}}.\) Choosing \(|g|t=0.1,\, I_{L}=10,\) we have \(s_{th}=0.57.\) So we calculate the quasi-distribution for \(s=0.7;\) in this case \(B_{S}=0.1,\) \(B_{Ss}=0.25,\) \(K_{VS,s}=-0.048.\) This quasi-distribution is shown in Fig. [\[fig:pnsnv\]](#fig:pnsnv){reference-type="ref" reference="fig:pnsnv"}b. It takes on negative values exhibiting nonclassical oscillations and behavior. Similarly, we can treat the compound mode \((LV)\) considering again photon vacuum scattering with partial stimulation \(I_{A}\neq0\) and \(I_{S}=0.\) Shifting distribution in \(W_{L}\) along \(I_{L}\) and neglecting short-time terms as above we obtain from ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}), \(K_{LV}=-|\chi|^{2}t^{2}I_{A}=-B_{L}\) and \(B_{V}=|g|^{2}t^{2}I_{L}+|\chi|^{2}t^{2}I_{A},\) i.e. \(K_{LV}+B_{L}=0\) and \(K_{LV}+B_{V}=|g|^{2}t^{2}I_{L}>0\) (\(C_{L}=C_{V}=\bar{D}_{LV}=0\)). Thus for the joint photon-phonon number distribution we obtain \[p(n_{L},n_{V})=\frac{n_{V}!}{n_{L}!(n_{V}-n_{L})!}\frac{(B_{V}+K_{LV})^{n_{V}-n_{L}}}{(1+Bn_{V})^{n_{V}+1}},\, n_{V}\geq n_{L}.\label{eq:pustak5}\] For \(n_{V}<n_{L},\) the distribution is zero. Its quantum behavior is illustrated in Fig. [\[fig:pnlnv\]](#fig:pnlnv){reference-type="ref" reference="fig:pnlnv"}a, showing one-side behavior along the diagonal compared to the earlier cases. For the threshold values of the ordering parameter we have \(s_{th}\approx1+|g|^{2}t^{2}I_{L}-2|\chi|t\sqrt{I_{A}}.\) Assuming for simplicity \(|\chi|=|g|,\, I_{A}=1\) and \(I_{L}=10,\) we have \(s_{th}=0.9\) and nonclassical behavior of wave quasi-distribution is illustrated by the Glauber-Sudarshan quasi-distribution of integrated-intensities for \(s=1\): \[P_{N}(W_{L},W_{V})=\frac{1}{\pi\sqrt{B_{L}B_{V}}}{\rm e}^{-\frac{W_{L}}{2B_{L}}-\frac{W_{V}}{2B_{V}}}\frac{\sin\left[\frac{\sqrt{\frac{B_{V}}{B_{L}}}W_{L}-\sqrt{\frac{B_{L}}{B_{V}}}W_{V}}{\sqrt{B_{L}}}\right]}{\sqrt{\frac{B_{V}}{B_{L}}}W_{L}-\sqrt{\frac{B_{L}}{B_{V}}}W_{V}},\label{eq:pustak6}\] as shown in Fig. [\[fig:pnlnv\]](#fig:pnlnv){reference-type="ref" reference="fig:pnlnv"}b. The existence of nonclassical character is clearly visible through the negative values of \(P_{N}(W_{L},W_{V}).\) # Difference and conditional number distributions [\[sec:Difference-and-conditional\]]{#sec:Difference-and-conditional label="sec:Difference-and-conditional"} In this section we can further illustrate the observed nonclassicalities via difference and conditional number distributions. For example, nonclassical character associated with a mode can be illustrated using conditional Fano factor, which is defined as \[F_{i,C}=\frac{\left\langle (\Delta n_{i})^{2}\right\rangle _{C}}{\left\langle n_{i}\right\rangle },\] for mode \(i\). Corresponding condition for nonclassicality is \(F_{i,C}<1.\) Analytic expressions for conditional Fano factor are obtained here for modes of interest (i.e. for \(F_{L,C}\) and \(F_{V,C}\)) as follows: \[F_{L,C}=1-\frac{B_{L}}{B_{V}},\label{eq:Fano-l}\] and \[F_{V,C}=\frac{(n_{L}+1)\left(\frac{1+B_{V}}{1+B_{L}}\right)^{2}-1}{(n_{L}+1)\left(\frac{1+B_{V}}{1+B_{L}}\right)-1}-1.\label{eq:Fano-V}\] It is now easy to observe from ([\[eq:BCD and Dbar\]](#eq:BCD and Dbar){reference-type="ref" reference="eq:BCD and Dbar"}) that \(B_{L}\) and \(B_{V}\) are always positive, consequently the conditional Fano factor \(F_{L,C}\) is always less than unity. Thus conditional Fano factor always depicts nonclassicality in pump mode. However, in phonon mode the presence of nonclassicality (i.e. \(F_{V,C}<1\)) is not directly visible from the expression, but the same is shown in the Fig. [\[fig:Conditional-Fano-factor\]](#fig:Conditional-Fano-factor){reference-type="ref" reference="fig:Conditional-Fano-factor"}. The corresponding number distributions are obtained as \[\begin{array}{lcl} p_{C}\left(n_{L};n_{V}\right) & = & \frac{n_{V}!}{n_{L}!\left(n_{V}-n_{L}\right)!}\left(1-\frac{B_{L}}{B_{V}}\right)^{n_{V}}\left(\frac{B_{L}}{B_{V}-B_{L}}\right)^{n_{L}},\\ p_{C}\left(n_{V};n_{L}\right) & = & \frac{n_{V}!}{n_{L}!\left(n_{V}-n_{L}\right)!}\frac{1+B_{L}}{1+B_{V}}\left(\frac{B_{V}-B_{L}}{1+B_{V}}\right)^{n_{V}}\left(\frac{1+B_{L}}{B_{V}-B_{L}}\right)^{n_{L}}.\end{array}\label{eq:conditional-distribution}\] These conditional number distributions are plotted in the Fig. [\[fig:Conditional-number-distributions\]](#fig:Conditional-number-distributions){reference-type="ref" reference="fig:Conditional-number-distributions"}. Difference number distribution can be obtained as \[p_{-}(n)=\frac{\left(B_{V}-B_{L}\right)^{n}}{\left(1+B_{V}-B_{L}\right)^{n+1}},\label{eq:difference-distribn}\] \[\langle\left(\Delta n\right)^{2}\rangle_{-}=\left(B_{V}-B_{L}\right)\left(1+B_{V}-B_{L}\right)\] and Poissonian distribution for the same two modes is \[p_{Pois}(n)=\frac{\left(B_{V}+B_{L}\right)^{n}}{n!}{\rm e}^{-\left(B_{V}+B_{L}\right).}\label{eq:Poiss-distribn}\] A joint plot of \(p_{-}(n)\) and \(p_{pois}(n)\) is provided in Fig. [\[fig:Difference-number-distribution\]](#fig:Difference-number-distribution){reference-type="ref" reference="fig:Difference-number-distribution"}, which clearly shows subpoissonian character in \(p_{-}(n)\). Thus a nonclassical difference number distribution is observed. For the sub-shot noise parameter \(R=\frac{\langle\left(\Delta n_{ij}\right)^{2}\rangle}{\left(\langle n_{i}\rangle+\langle n_{j}\rangle\right)}\) we have \(R\thickapprox1-\frac{2B_{L}}{B_{L}+B_{V}}=0.83<1.\) # Conclusion[\[sec:Conclusion\]]{#sec:Conclusion label="sec:Conclusion"} We have observed different type of nonclassicalities in the stimulated, completely spontaneous and partially spontaneous Raman process. The observations that are discussed in detail in Section [\[sec:Intermodal-entanglement\]](#sec:Intermodal-entanglement){reference-type="ref" reference="sec:Intermodal-entanglement"} are summarized in Table [2](#tab:summary){reference-type="ref" reference="tab:summary"} for coherent scattering. We see that in general various nonclassical features of the process can or cannot be directly related, only for combined modes \((LV)\) and \((SV)\) all of them occurs simultaneously. We have not restricted ourselves to the study of coherent scattering alone. In Section [\[sec:Phonon-mode-is-chaotic\]](#sec:Phonon-mode-is-chaotic){reference-type="ref" reference="sec:Phonon-mode-is-chaotic"} we have investigated various nonclassical characters of Raman process when the phonon mode is chaotic. Finally we have illustrated our results by joint photon-phonon number and wave distributions. ::: **Acknowledgment:** A. P. thanks Department of Science and Technology (DST), India for support provided through the DST project No. SR/S2/LOP-0012/2010. He also thanks the Operational Program Education for Competitiveness-European Social Fund project CZ.1.07/2.3.00/20.0017 of the Ministry of Education, Youth and Sports of the Czech Republic. J. P. and J. K. thank the Operational Program Research and Development for Innovations-European Regional Development Fund project CZ.1.05/2.1.00/03.0058 of the Ministry of Education, Youth and Sports of the Czech Republic.

{'timestamp': '2012-10-16T02:03:08', 'yymm': '1210', 'arxiv_id': '1210.3779', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3779'}

# Introduction A *phylogenetic tree* \(T\) is an unrooted tree whose leaves are bijectively mapped to label set \(\ensuremath{\mathcal{L}}(T)\). Labels represent species and phylogenetic trees represent evolutionary history of species. Let \(\ensuremath{\mathcal{P}}\) be a collection of phylogenetic trees. We call \(\ensuremath{\mathcal{P}}\) a *profile*. We denote \(\bigcup_{T \in \ensuremath{\mathcal{P}}}\ensuremath{\mathcal{L}}(T)\) by \(\ensuremath{\mathcal{L}}(\ensuremath{\mathcal{P}})\). A *supertree* of \(\ensuremath{\mathcal{P}}\) is a phylogenetic tree whose label set is \(\ensuremath{\mathcal{L}}(\ensuremath{\mathcal{P}})\). Let \(S\) be a phylogenetic tree. For any \(Y \subseteq \ensuremath{\mathcal{L}}(S)\) let \(S|Y\) denote the tree obtained by suppressing any degree two vertices in the minimal subtree of \(S\) connecting the labels in \(Y\). Let \(T\) be a phylogenetic tree where \(\ensuremath{\mathcal{L}}(T) \subseteq \ensuremath{\mathcal{L}}(S)\). We say that \(S\) *displays* \(T\) if \(T\) can be derived from \(S|\ensuremath{\mathcal{L}}(T)\) by contraction of edges. Given a profile \(\ensuremath{\mathcal{P}}\) of trees, the *tree compatibility* problem asks if there exists a supertree of \(\ensuremath{\mathcal{P}}\) which displays all the trees in \(\ensuremath{\mathcal{P}}\). If such a supertree \(S\) exists, we say that \(\ensuremath{\mathcal{P}}\) is compatible and \(S\) is a compatible tree of \(\ensuremath{\mathcal{P}}\). The tree compatibility problem is NP-complete  but fixed parameter tractable when parametrized by number of trees . An instance of the tree compatibility problem can be transformed to an instance of character compatibility by representing each input tree by its set of splits . A *quartet* is a binary phylogenetic tree with exactly four leaves. An instance of tree compatibility can also be transformed into an instance of quartet compatibility . Grunewald et al.  characterized quartet compatibility problem in terms of finding an unification sequence for a structure called the quartet graph. Vakati and Fernández-Baca characterized the tree compatibility problem in terms of finding a legal triangulation  of the display graph of a profile, a graph introduced by Bryant and Lagergren . Gysel et al. introduced the edge label intersection graph for a profile of phylogenetic trees and used this graph to characterize tree compatibility as a chordal sandwich problem . Here we show that the latter characterization translates to a characterization in terms of minimal cuts of display graphs. We also show how such cuts are closely related to the splits of the compatible supertree. Finally, we show how these two characterizations relate to the legal triangulation characterization given in . # Preliminaries For every nonnegative integer \(m\), we denote the set \(\{1, \dots, m\}\) by \([m]\). Let \(G\) be a graph. We represent the vertices and edges of \(G\) by \(V(G)\) and \(E(G)\) respectively. For any \(U \subseteq V(G)\), \(G-U\) represents the graph derived by removing vertices of \(U\) and their incident edges from \(G\). Similarly, for any \(F \subseteq E(G)\), \(G-F\) represents the graph with vertex set \(V(G)\) and edge set \(E(G) \setminus F\). For any vertex \(v \in V(G)\), we denote the set \(\{x: \{x, v\} \in E(G)\}\) by \(N_G(v)\). For any two nonadjacent vertices \(a\) and \(b\) of \(G\), an \(a\)-\(b\) \(separator\) \(U\) of \(G\) is a set of vertices whose removal disconnects \(G\) and \(a\) and \(b\) are in different connected components of \(G-U\). An \(a\)-\(b\) separator \(U\) is *minimal* if for any \(U' \subset U\), \(U'\) is not an \(a\)-\(b\) separator. A set \(U \subseteq V(G)\) is a *minimal separator* if \(U\) is a minimal \(a\)-\(b\) separator for some nonadjacent vertices \(a\) and \(b\) of \(G\). Two minimal separators \(U\) and \(U'\) are *parallel* if \(G-U\) contains at most one component \(H\) where \(V(H) \cap U' \neq \emptyset\). We represent the set of all minimal separators of graph \(G\) by \(\triangle_G\). Assume that \(G\) is connected. A *cut* is a set of edges \(F \subseteq E(G)\) whose removal disconnects \(G\). A cut \(F\) is *minimal* if there does not exist \(F' \subset F\) where \(G-F'\) is disconnected. Note that if \(F\) is minimal, there will exactly be two connected components in \(G-F\). Two minimal cuts \(F\) and \(F'\) are *parallel* if \(G-F\) has at most one connected component \(H\) where \(E(H) \cap F' \neq \emptyset\). A *chord* is an edge between two nonadjacent vertices of a cycle. A graph \(H\) is *chordal* if and only if every cycle of length four or greater in \(H\) has a chord. A chordal graph \(H\) is a *triangulation* of graph \(G\) if and only if \(V(G) = V(H)\) and \(E(G) \subseteq E(H)\). The edges in \(E(H) \setminus E(G)\) are called *fill-in* edges of \(G\). A *clique tree* of a chordal graph \(H\) is a pair \((T, B)\) where (i) \(T\) is a tree, (ii) \(B\) is a bijective function from vertices of \(T\) to maximal cliques of \(H\), and (iii) \((T, B)\) satisfies the *coherence* property; i.e., for every vertex \(v \in H\), the set of all vertices \(x\) of \(T\) where \(v \in B(x)\) induces a subtree in \(T\). Let \(G\) be a graph and let \(\ensuremath{\mathcal{F}}\) be a collection of subsets of \(V(G)\). We represent by \(G_{\ensuremath{\mathcal{F}}}\) the graph derived from \(G\) by making the set of vertices of \(X\) a clique in \(G\) for every \(X \in \ensuremath{\mathcal{F}}\). Let \(T\) be a phylogenetic tree over label set \(\ensuremath{\mathcal{L}}(T)\). Since there exists a bijective function from leaves of \(T\) to \(\ensuremath{\mathcal{L}}(T)\), we will represent leaves of \(T\) by their labels. Let \(\ensuremath{\mathcal{P}} = \{T_1, T_2, \cdots, T_k\}\) be a profile of \(k\) phylogenetic trees. The *display graph* of profile \(\ensuremath{\mathcal{P}}\), denoted by \(G(\ensuremath{\mathcal{P}})\), is a graph whose vertex set is \(\bigcup_{i \in [k]} V(T_i)\) and edge set is \(\bigcup_{j \in [k]} E(T_j)\). An example of a display graph is given in Fig. [\[fig:example\]](#fig:example){reference-type="ref" reference="fig:example"}. A vertex \(v\) of \(G(\ensuremath{\mathcal{P}})\) is a *leaf* vertex if \(v \in \ensuremath{\mathcal{L}}(\ensuremath{\mathcal{P}})\). Every other vertex of \(G(\ensuremath{\mathcal{P}})\) is an *internal* vertex. An edge of \(G(\ensuremath{\mathcal{P}})\) is an *internal* edge if both its endpoints are internal; otherwise, it is a *non-internal* edge. Let \(H\) be a subgraph of \(G(\ensuremath{\mathcal{P}})\). We represent by \(\ensuremath{\mathcal{L}}(H)\), the set of all leaf vertices of \(H\). A triangulation \(G'\) of \(G(\ensuremath{\mathcal{P}})\) is *legal* if it satisfies the following conditions. 1. For every clique \(C\) of \(G'\), if \(C\) contains an internal edge, then it cannot contain any other edge of \(G(\ensuremath{\mathcal{P}})\). 2. There does not exist a fill-in edge with a leaf vertex as an endpoint. The *edge label intersection* graph (see Fig. [\[fig:example\]](#fig:example){reference-type="ref" reference="fig:example"}) of a profile \(\ensuremath{\mathcal{P}}\), denoted \(L(\ensuremath{\mathcal{P}})\), is the graph whose vertex set is the set of all edges of input trees of \(\ensuremath{\mathcal{P}}\) where there is an edge between vertices \(e\) and \(e'\) if and only if \(e \cap e' \neq \emptyset\) . (Note that in , \(L(\ensuremath{\mathcal{P}})\) is the modified edge label intersection graph.) It can be verified that \(L(\ensuremath{\mathcal{P}})\) is the line graph of \(G(\ensuremath{\mathcal{P}})\) ; i.e., the vertices of \(L(\ensuremath{\mathcal{P}})\) are the edges of \(G(\ensuremath{\mathcal{P}})\) and two vertices in \(L(\ensuremath{\mathcal{P}})\) are adjacent if the corresponding edges in \(G(\ensuremath{\mathcal{P}})\) share a common endpoint. A fill-in edge of \(L(\ensuremath{\mathcal{P}})\) is *valid* if both its endpoints are not from \(L(T)\) for some input tree \(T\). A triangulation \(H\) of \(L(\ensuremath{\mathcal{P}})\) is *restricted* if every fill-in edge of \(H\) is valid. For rest of the paper we will assume that for any profile \(\ensuremath{\mathcal{P}}\), \(G(\ensuremath{\mathcal{P}})\) is connected. Otherwise, there exists a partition \(\ensuremath{\mathcal{P}}_{part}\) of \(\ensuremath{\mathcal{P}}\) such that for every \(P \in \ensuremath{\mathcal{P}}_{part}\), \(G(P)\) is connected and \(P\) is maximal such set. Then, \(\ensuremath{\mathcal{P}}\) is compatible if and only if \(P\) is compatible for every \(P \in \ensuremath{\mathcal{P}}_{part}\). Note that, if \(G(\ensuremath{\mathcal{P}})\) is connected, then \(L(\ensuremath{\mathcal{P}})\) is also connected. # Characterization using minimal cuts Let \(\ensuremath{\mathcal{P}} = \{T_1, T_2, \cdots, T_k\}\) be a profile of phylogenetic trees. A minimal separator \(F\) of \(L(\ensuremath{\mathcal{P}})\) is *legal* if for every \(F' \subseteq F\) such that \(F' \subseteq V(L(T))\) for some input tree \(T\), \(F'\) is a clique in \(L(T)\). For any vertex \(u\) of an input tree, \(\hat{K}(u)\) represents the set of all vertices of \(L(\ensuremath{\mathcal{P}})\) where for every \(e \in \hat{K}(u)\), \(u \in e\). A cut \(F\) of display graph \(G(\ensuremath{\mathcal{P}})\) is *legal*, if it satisfies the following: 1. For every tree \(T \in \ensuremath{\mathcal{P}}\), the edges of \(T\) in \(F\) are incident on a common vertex. 2. There is at least one edge in each of the connected components of \(G(\ensuremath{\mathcal{P}})-F\). We make use of the following simple observation in some of the proofs in this section. The next lemma follows from the definition of restricted triangulation and is from . A set \(\ensuremath{\mathcal{F}}\) of pairwise parallel legal minimal cuts of \(G(\ensuremath{\mathcal{P}})\) is *complete*, if for every input tree \(T \in \ensuremath{\mathcal{P}}\) and for every internal edge \(e\) of \(T\), there exists a cut \(F \in \ensuremath{\mathcal{F}}\) where \(e\) is the only edge of \(T\) in \(F\). Lemmas [\[lm:maximal_seps_equals_maximal_cuts\]](#lm:maximal_seps_equals_maximal_cuts){reference-type="ref" reference="lm:maximal_seps_equals_maximal_cuts"} and [\[lm:maximal_cuts_equals_maximal_seps\]](#lm:maximal_cuts_equals_maximal_seps){reference-type="ref" reference="lm:maximal_cuts_equals_maximal_seps"} imply the following theorem. The next result follows from Theorems [\[thm:edge_label\]](#thm:edge_label){reference-type="ref" reference="thm:edge_label"} and [\[thm:maximal_seps_equals_maximal_cuts\]](#thm:maximal_seps_equals_maximal_cuts){reference-type="ref" reference="thm:maximal_seps_equals_maximal_cuts"}. An analogue of Theorem [\[thm:cuts_equal_compatibility\]](#thm:cuts_equal_compatibility){reference-type="ref" reference="thm:cuts_equal_compatibility"} can be derived for the edge label intersection graph \(L(\ensuremath{\mathcal{P}})\) as follows. A set \(\ensuremath{\mathcal{F}}\) of legal minimal separators of \(L(\ensuremath{\mathcal{P}})\) is *complete*, if for every internal edge \(e\) of an input tree \(T\), there exists a separator \(F \in \ensuremath{\mathcal{F}}\) where \(e\) is the only vertex of \(L(T)\) in \(F\). The next theorem follows from Theorems [\[thm:cuts_seps\]](#thm:cuts_seps){reference-type="ref" reference="thm:cuts_seps"} and [\[thm:cuts_equal_compatibility\]](#thm:cuts_equal_compatibility){reference-type="ref" reference="thm:cuts_equal_compatibility"}, and Lemma [\[lm:parallel\]](#lm:parallel){reference-type="ref" reference="lm:parallel"}. # Relationship to splits compatibility A *split* of a label set \(L\) is a bipartition of \(L\). We denote a split \(\{X, Y\}\) by \(X|Y\). Let \(T\) be a phylogenetic tree. Consider an internal edge \(e\) of \(T\). Deletion of \(e\) breaks \(T\) into two subtrees \(T_1\) and \(T_2\). Let \(L_1\) and \(L_2\) denote the set of all labels in \(T_1\) and \(T_2\) respectively. Set \(\{L_1, L_2\}\) is a split of \(\ensuremath{\mathcal{L}}(T)\). We denote the split corresponding to edge \(e\) of \(T\) by \(\Sigma_e(T)\) and we denote by \(\Sigma(T)\) the set of all splits corresponding to all internal edges of \(T\). A tree \(T\) *displays* a split \(X\) if there exists an internal edge \(e\) of \(T\) where \(\Sigma_e(T) = X\). Then, we also say \(T\) is *compatible* with \(X\). A set of splits is compatible if there exists a tree which displays all the splits in the set. Two splits \(A_1|A_2\) and \(B_1|B_2\) are compatible if and only if at least one of \(A_1 \cap B_1\), \(A_1 \cap B_2\), \(A_2 \cap B_1\) and \(A_2 \cap B_2\) is empty . By the Splits Equivalence Theorem , a collection of splits is compatible if and only if every pair is compatible. For any legal minimal cut \(F\) of \(G(\ensuremath{\mathcal{P}})\), we denote by \(\Sigma(F)\) the split of \(\ensuremath{\mathcal{L}}(\ensuremath{\mathcal{P}})\) induced by \(F\). If \(\ensuremath{\mathcal{F}}\) is a set of legal minimal cuts of \(G(\ensuremath{\mathcal{P}})\), then we denote \(\bigcup_{F \in \ensuremath{\mathcal{F}}}\Sigma(F)\) by \(\Sigma(\ensuremath{\mathcal{F}})\). # Relationship to legal triangulation Let \(\ensuremath{\mathcal{P}}\) be a profile of phylogenetic trees. Theorems [\[thm:lt\]](#thm:lt){reference-type="ref" reference="thm:lt"} and [\[thm:cuts_equal_compatibility\]](#thm:cuts_equal_compatibility){reference-type="ref" reference="thm:cuts_equal_compatibility"} together imply that if \(G(\ensuremath{\mathcal{P}})\) has a complete set \(\ensuremath{\mathcal{F}}\) of pairwise parallel legal minimal cuts, there also exists a legal triangulation of \(G(\ensuremath{\mathcal{P}})\). As shown in , a legal triangulation of \(G(\ensuremath{\mathcal{P}})\) can be derived from a compatible tree of \(\ensuremath{\mathcal{P}}\). In this section, we show how to derive a legal triangulation of \(G(\ensuremath{\mathcal{P}})\) directly from \(\ensuremath{\mathcal{F}}\) without building a compatible tree. This shows the relationship between complete sets of pairwise parallel legal minimal cuts and legal triangulations of display graphs. By Theorems [\[thm:edge_label\]](#thm:edge_label){reference-type="ref" reference="thm:edge_label"} and  [\[thm:cuts_seps\]](#thm:cuts_seps){reference-type="ref" reference="thm:cuts_seps"} and Lemma [\[lm:parallel\]](#lm:parallel){reference-type="ref" reference="lm:parallel"}, this also shows the relationship between restricted triangulations of edge label intersection graphs and legal triangulations of display graphs. A complete set \(\ensuremath{\mathcal{F}}\) of pairwise parallel legal minimal cuts of \(G(\ensuremath{\mathcal{P}})\) is *minimal* if no proper subset of \(\ensuremath{\mathcal{F}}\) is also complete. Let \(\ensuremath{\mathcal{F}}\) be a minimal complete set of pairwise parallel legal minimal cuts of \(G(\ensuremath{\mathcal{P}})\). At a high level, we construct a legal triangulation of \(G(\ensuremath{\mathcal{P}})\) from \(\ensuremath{\mathcal{F}}\) as follows. Consider any cut \(F \in \ensuremath{\mathcal{F}}\). We build a pair \(D_F=(X,Y)\) where \(X\) and \(Y\) are subsets of \(E(F)\) and are vertex separators of \(G(\ensuremath{\mathcal{P}})\). Let \(A\) and \(B\) be the connected components of \(G(\ensuremath{\mathcal{P}})-F\). Also, let \(A'\), \(B'\) be the subgraphs induced in \(G(\ensuremath{\mathcal{P}})\) by the vertex sets \(V(A) \cup \{X \cap Y\}\) and \(V(B) \cup \{X \cap Y\}\) respectively. To legally triangulate \(G(\ensuremath{\mathcal{P}})\) we first triangulate the subgraph of \(G(\ensuremath{\mathcal{P}})\) induced by the vertex set \(X \cup Y\) and then triangulate the subgraphs \(A'\) and \(B'\). To triangulate either of those subgraphs, we again use vertex separators built from endpoints of a different cut. We make sure that, for every set \(D_I\) for some \(I \in \ensuremath{\mathcal{F}}\) built after \(D_F\), both the sets of \(D_I\) are subsets of either \(V(A')\) or \(V(B')\) but not both. We now give the details of our construction. We consider the elements of \(\ensuremath{\mathcal{F}}\) in some arbitrary, but fixed order, and use a set \(W\) to record all such cuts \(F \in \ensuremath{\mathcal{F}}\) for which \(D_F\) has already been constructed. Initially \(W\) is empty. For each successive cut \(F\) in \(\ensuremath{\mathcal{F}}\), we do the following. Let \(F' \subseteq F\) be the set of all internal edges \(e \in F\) such that \(e\) is the only edge of the tree containing \(e\) that is in \(F\). Let \(A\) and \(B\) be the two connected components of \(G(\ensuremath{\mathcal{P}})-F\). Let \(X = V(A) \cap V(F')\) and \(Y= V(B) \cap V(F')\). For every edge \(e\) of \(F'\) whose endpoints are in different sets of some set \(D_I\) where \(I \in W\), we do the following. Let \(Q\) be the connected component of \(G(\ensuremath{\mathcal{P}})-I\) where \(E(Q) \cap F \neq \emptyset\). Note that \(Q\) is the only such component of \(G(\ensuremath{\mathcal{P}})-I\). Let \(v\) be the vertex of \(e\) in \(Q\). Replace the endpoints of \(e\) in sets \(X\) and \(Y\) by \(v\). For every non internal edge \(f \in F\) where \(f\) is the only edge of the tree containing \(f\) that is in \(F\), add the internal vertex of \(f\) to both sets \(X\) and \(Y\). If there exists a tree \(T\) where more than one edge of \(T\) is in \(F\), add the common endpoint of all the edges of \(T\) in \(F\) to both sets \(X\) and \(Y\). Set \(D_F\) to \((X, Y)\). Add \(F\) to \(W\). For every \(F \in \ensuremath{\mathcal{F}}\), let \(O_F\) be the set defined as follows. Let \(D_F=(X,Y)\) and let \(X=\{x_1, \cdots, x_m, z_1,\cdots, z_p\}\) and \(Y = \{y_1, \cdots, y_m, z_1, \cdots, z_p\}\), where \(m > 0\), \(p \geq 0\) and for every \(i \in [m]\), \(\{x_i, y_i\}\) is an internal edge of \(G(\ensuremath{\mathcal{P}})\). Then, \(O_F\) consists of sets \(\{x_1, \cdots, x_j, y_j, \cdots, y_m, z_1, \cdots, z_p\}\) for every \(j \in [m]\). Let \(G'\) be the graph derived from \(G(\ensuremath{\mathcal{P}})\) as follows. For every cut \(F \in \ensuremath{\mathcal{F}}\) where \(D_F=(X,Y)\), add edges to make each of the sets \(X\) and \(Y\) a clique. For every cut \(F \in \ensuremath{\mathcal{F}}\) and for every \(Y \in O_F\), add edges to make \(Y\) a clique. For every leaf \(\ell\), make the vertices of \(N_{G(\ensuremath{\mathcal{P}})}(\ell)\) a clique. To prove Theorem [\[thm:lt_convert\]](#thm:lt_convert){reference-type="ref" reference="thm:lt_convert"} we first prove few useful lemmas. For every cut \(F \in \ensuremath{\mathcal{F}}\) where \(D_F=(X,Y)\), we denote the sets \(X \cup Y\), \(X \cap Y\) by \(F_\cup\) and \(F_\cap\) respectively. For any internal edge \(e\), we call the cut \(F \in \ensuremath{\mathcal{F}}\) a *differentiating* cut of \(e\) if \(e\)'s endpoints are in different sets of \(D_F\). Note that, since \(\ensuremath{\mathcal{F}}\) is minimal, every cut in \(\ensuremath{\mathcal{F}}\) is a differentiating cut of some internal edge. A clique of \(G'\) is *illegal* if it contains a fill-in edge with a leaf vertex as an endpoint or if it contains an internal edge along with any another edge of \(G(\ensuremath{\mathcal{P}})\). Graph \(G'\) is a legal triangulation if and only if \(G'\) does not contain an illegal clique. # Conclusion We have shown that the characterization of tree compatibility in terms of restricted triangulations of the edge label intersection graph transforms into a characterization in terms of minimal cuts in the display graph. We have also shown how these two characterizations relate to the characterization in terms of legal triangulations of the display graph . It remains to be seen whether any of these characterizations can be exploited to derive an explicit fixed parameter algorithm for the tree compatibility problem when parametrized by number of trees. Grunewald et al.  use quartet graphs to characterize when a collection of quartets define and identify a compatible supertree. An interesting question is whether a similar characterization can be derived for collections of phylogenetic trees using display graphs or edge label intersection graphs.

{'timestamp': '2012-10-16T02:02:51', 'yymm': '1210', 'arxiv_id': '1210.3762', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3762'}

# Introduction {#p3:intro} Observations in our Galaxy and in the Local Group of Galaxies (LGGs) have revealed that Giant Molecular Clouds (GMCs) have masses of about \(10^4\) to 10\(^7\) M\(_{\odot}\) and sizes of 50 to several hundred of parsecs, which are often found associated with Hii regions and OB associations, tracers of massive star formation (SF). Massive stars, which are commonly found clustered in OB associations, are short-lived (less than a few 10 Myr), and thus spend most of their lifetime within the cluster in which they were formed. This suggests that GMCs are the principal sites of massive SF. However, how SF occurs in a GMC, the so-called "GMC evolution", has remained poorly understood because it has been difficult to perform molecular gas observations with a combination of high resolution, sensitivity and wide field to enable studies of a sufficiently large sample of GMCs. Such studies in LGGs have the potential to reveal the density distribution and kinematics at GMC scales, as well as its relation to ongoing SF. Several studies focusing on GMC evolution in the Large Magellanic Clouds (LMC) have been performed. The GMCs in the LMC are classified into three types according to their associated massive SF activities: Type i showing no signature of massive SF, Type ii being associated with relatively small Hii regions, and Type iii with both Hii regions and young (less than 10 Myr old) stellar clusters. While in general the positions of young clusters are correlated with GMCs, the clusters older than 10 Myr have a weaker or no correlation with the GMCs. This classification was interpreted to be a template for the GMC evolutionary stages, and the typical lifetime of GMCs with masses larger than \(5 \times 10^4\) M\(_{\odot}\) is estimated to be 20--30 Myr . Analyses of the kinetic temperature and density of several GMCs in LMC shows that they increase generally as the GMCs evolve. This confirms that the denser and warmer molecular gas is directly linked to SF activities. Because these studies are only limited to several specific GMCs in one irregular galaxy, it is still necessary to investigate the properties of GMCs in the different environments present in other galaxies, in order to complete our current knowledge of templates for GMC evolution. M33 is one of the best laboratories to study GMC evolutions under different environments due to its proximity, favorable inclination, as well as the existence of spiral structure and massive SF regions. M33 is the second closest spiral galaxy after M31, but M33 is less inclined, which is ideal to resolve gas components or individual stars with little contamination along the line of sight. Also, studying GMCs over a disk galaxy enable us to compare all GMCs at the same distance, unlike in our Galaxy. Surveys of \(^{12}\)CO(\(J=1-0\)) and \(^{12}\)CO(\(J=2-1\)) emission in M33 to investigate the relation between molecular gas and massive SF have been conducted by several authors. have found that more than two-thirds of the GMCs in M33 are associated with Hii regions within 50 pc, which is similar to the proportion in LMC. On the other hand, the correlation between the molecular gas surface densities traced by CO(\(J=1-0\)) emission and that of SF rate (SFR), the so-called Kennicutt--Schmidt Law in a pixel-to-pixel analysis is found to be weaker at 80 pc resolutions than at lower resolutions. This suggests that the GMCs traced by CO(\(J=1-0\)) are well correlated with the massive SF sites, but their peaks are offset from each other. Recently, the K-S law using CO(\(J=3-2\)) in M33 has shown a tight correlation at 100 pc resolution, unlike CO(\(J=1-0\)) (Onodera et al. 2013, PASJ, accepted). In fact, with higher resolution data, have found that the CO(\(J=3-2\)) emission is spatially better correlated to massive SF sites than CO(\(J=1-0\)) in the most luminous giant Hii region (GHR) in M33, NGC 604. This supports that the dense and warm gas traced by CO(\(J=3-2\)) is more closely linked to the SF sites, as previously argued in four Virgo galaxies in. These results are interpreted as higher kinetic temperatures and densities in GMCs found close to GHRs, in agreement with large velocity gradient (LVG) analysis with multi-\(J\) CO transition data by and. As part of the Nobeyama Radio Observatory (NRO) M33 All-disk survey of Giant Molecular Clouds (MAGiC) project, here we present a catalog of M33 GMCs using new CO(\(J=3-2\)) maps with \(25\arcsec\) (corresponding to 100 pc) resolution. In order to investigate the relation of the GMCs with massive SF, we also present a "young stellar group (YSG)" catalog. In this work, YSGs can be clusters (typical sizes of 15 pc), OB associations (typical sizes of \(\sim\)`<!-- -->`{=html}15--100 pc), and star complexes (sizes of a few hundred pc). There are a number of catalogs for clusters, OB associations, and star complexes in M33 published so far (e.g., [@2007AJ....134..447S], hereafter SM; [@2007AJ....134.2168P]; [@2010ApJ...720.1674S]; [@1980ApJS...44..319H]; [@2005PASRB...4...75I]). However, the number of star clusters whose ages have been estimated is limited, and the methods to derive the ages are not homogenous (SM and references therein; [@2009ApJ...700..103P], hereafter PPL; [@2009ApJ...699..839S], hereafter SSGH). We also aim to estimate their ages in a homogeneous manner using the stellar photometry catalog provided by. The outline of this paper is as follows. The observations and data reduction are summarized in Section [2](#p3:obs){reference-type="ref" reference="p3:obs"}. In Section [3](#p3:res){reference-type="ref" reference="p3:res"} we explain the general procedure for identifying GMCs, as well as the identification of YSGs and the estimation of their ages. We classify GMCs as a function of the age of YSGs associated with them. We also quantify the SF activities and dense gas fractions in the GMCs using the extinction-corrected H\(\alpha\) data and the CO(\(J=3-2\))/CO(\(J=1-0\)) line ratio, \(R_{3-2/1-0}\), and then show the variety of physical states of the identified GMCs. In Section [4](#p3:disc){reference-type="ref" reference="p3:disc"}, we discuss the relationship between the properties of these clouds and their evolutionary stages. Finally, we interpret our results in the context of a continuous SF in GMCs. The summary of this paper is provided in Section [5](#p3:sum){reference-type="ref" reference="p3:sum"}. # Observations and Data Reduction {#p3:obs} We describe the data that is used throughout this paper, including the CO(\(J=3-2\)) data to identify the GMCs, the CO(\(J=1-0\)) data to derive \(R_{3-2/1-0}\), and the optical and infrared data to identify the YSG and Hii regions in M33. ## ASTE CO(\(J=3-2\)) Data {#p3:sec321} We chose eight regions covering the northern and southern spiral arms as well as the galaxy center, where the CO(\(J=1-0\)) emission is prominent. The CO(\(J=3-2\)) (rest frequency: 345.796 GHz) mapping observations were performed between June and November 2011, using the Atacama Submillimeter Telescope Experiment 10-m dish equipped with the 345-GHz side band separating (2SB) SIS receiver CATS345. An XF-Type Digital spectrometer MAC was used to cover a velocity width of 445 km s\(^{-1}\) with a velocity resolution of 2.5 km s\(^{-1}\) at 345 GHz. The On-The-Fly (OTF) mapping technique, based on scanning smoothly and rapidly along one direction across a rectangular map, was employed to obtain the CO(\(J=3-2\)) data. The observed area covers in total about 80 % of the CO(\(J=1-0\)) based molecular gas mass in. The observations and the data calibration were performed for each region as follows. The half power beam width (HPBW) of the ASTE 10-m telescope is 22\(\arcsec\) at 345 GHz. The chopper-wheel technique was employed to calibrate the antenna temperature \(T_A^{\ast}\) and the final data has been converted into units of main beam brightness temperature (\(T_{\rm mb} \equiv T_A^{\ast}/\eta_{\rm mb}\)), where \(\eta_{\rm mb}\) is the main beam efficiency, measured to be \(0.6\,\pm\,0.1\) using the standard source Orion KL at (\(\alpha_{\rm J2000}\), \(\delta_{\rm J2000}\))=(05\(^{\rm h}\)`<!-- -->`{=html}35\(^{\rm m}\)`<!-- -->`{=html}14, 05). Hereafter, all the CO intensity measurements are specified in \(T_{\rm mb}\). The typical system temperatures in a single side band were usually less than 200 K. The absolute pointing accuracy was verified by observing the CO(\(J=3-2\)) emission of o-Cet at 1--1.5 hr intervals, and it was kept better than 4\(\arcsec\) rms. The absolute intensity stability was also monitored using o-Cet during the different observing runs and was found to be stable within \(\sim16\) %. The data reduction was carried out with the Nobeyama OTF Software Tools for Analysis and Reduction package. The "scanning noise" was removed by combining scans using the PLAIT algorithm as described by. The data have been convolved with a Bessel-Gauss function in the spectral domain to create the data cube. The final grid spacing is 8 and the angular resolution is 25. The final data cubes for the eight regions are characterized by an rms noise of \(\sigma_{\rm ch}\) = 16--32 mK in a velocity resolution of 2.5 km s\(^{-1}\), achieved after 87 hr of total integration. We briefly summarize in Table [\[p3:tab1\]](#p3:tab1){reference-type="ref" reference="p3:tab1"} the observational and data reduction parameters for the eight observed regions, including their central positions, sizes, integration times, and rms for a 2.5 km s\(^{-1}\) channel. Figure [\[p3:fig2\]](#p3:fig2){reference-type="ref" reference="p3:fig2"} shows the obtained integrated intensity map over the velocity range \(V_{\rm LSR}=-274\) km s\(^{-1}\) to \(-94\) km s\(^{-1}\), where emission is \(>\) 2 \(\sigma_{\rm ch}\), overlaid on the H\(\alpha\) image (§ [2.3](#p3:2.2){reference-type="ref" reference="p3:2.2"}). The rms noise level of the integrated intensity map (\(\sigma_{\rm mom}\)) is 0.38 K km s\(^{-1}\) on average. The noise level of the CO(\(J=3-2\)) integrated intensity map is calculated using \(\sigma_{\rm mon}\equiv\sigma_{\rm ch}\sqrt{N}\delta v\), where \(N\) is the number of integrated channels and \(\delta v\) is the velocity resolution of a channel (2.5 km s\(^{-1}\)). The observed fields cover the molecular disk up to the galactic radius of \(\sim\)`<!-- -->`{=html}5 kpc. Note that we also present CO integrated intensity maps for each individual GMC in Section [3.1.1](#sec311){reference-type="ref" reference="sec311"} and the Appendix. ## NRO CO(\(J=1-0\)) Data {#p3:sec322} The CO(\(J=1-0\)) emission was observed in a \(31\arcmin \times 36\arcmin\) area (corresponding to 7.6 kpc \(\times\) 8.8 kpc) toward the disk of M33 using the NRO 45-m telescope. The data reduction was also carried out with the NOSTAR reduction package. The angular resolution of the final map was 19 and the grid spacing 7. Further processing was done with MIRIAD, in a manner similar to the CO(\(J=3-2\)) map. We have convolved the CO(\(J=1-0\)) map to a common angular resolution of 25\(\arcsec\) and re-gridded it to 8 per pixel. The convolved map had an rms noise of 87 mK for the velocity resolution of 2.5 km s\(^{-1}\). We have also created CO(\(J=1-0\)) integrated intensity maps for each GMC (see Appendix), integrating over the same velocity range as used for the individual CO(\(J=3-2\)) integrated intensity maps. ## Optical and Infrared Data {#p3:2.2} We have used the photometric optical dataset from the *UBVI* ground-based survey of local star-forming galaxies with the Kitt Peak National Observatory 4 m telescope presented in, and available on their ftp site[^1]. It contains a total of 146,622 stars in a field of 0.8 deg\(^2\) centered on M33, which fully covers the regions that were observed in CO(\(J=3-2\)) and CO(\(J=1-0\)). The photometry in *BVI* bands is less than 1 % above 21.1 mag and even less than 10 % at about 23 mag. Refer to for further details about the observation, data reduction and photometric analysis. We have also compiled available Hii region catalogs. The H\(\alpha\) image from is used to check the accuracy of the positions and extents of the M33 individual Hii regions in these catalogs. Refer to for more details on the observations and reduction process. Finally, we have used the 24 \(\micron\) *Spitzer* data presented in to correct the extinction in the H\(\alpha\) emission. The final combined mosaic image is approximately \(1.1\times1.2\) deg\(^{2}\), and both the point-spread-function FWHM and grid sizes are set to \(5\farcs7\). For more details on the reduction process, refer to. Note that have performed a pixel-by-pixel analysis and have not applied a local background subtraction to create the extinction-corrected data by combining the H\(\alpha\) image and the \(24\,\micron\) data. The lack of background subtraction may cause a systematic shift in the extinction-corrected H\(\alpha\) luminosity. In this paper, instead, we present a new approach in which the H\(\alpha\) and \(24\,\micron\) luminosities are measured with circular apertures of several 100 pc radius and with local background subtraction applied (see Section [3.3](#p3:association){reference-type="ref" reference="p3:association"}). We use the catalog of \(24\,\micron\) sources in, which is the only reference listing 24 \(\micron\) sources to date, for the sake of comparing the positions of Hii regions. # Results {#p3:res} ## The CO(\(J=3-2\)) GMC Catalog {#sec31} The GMC catalog is obtained from our CO(\(J=3-2\)) data using the CPROPS package described in. Briefly, we search for compact emission in adjacent pairs of channels with \(T_{\rm mb} > 4\,\sigma_{\rm rms}\), the so-called *cores*, and then for emission with \(T_{\rm mb} > 2\,\sigma_{\rm rms}\) connected to such cores. CPROPS attempts to account for the amount of flux below the 2 \(\sigma_{\rm rms}\) cutoff by linearly extrapolating the emission profile to the \(0\, {\rm K\,km\,s}^{-1}\) intensity level. Finally, the emission identified with CPROPS must have a minimum velocity width of 2 channels (\(5\,{\rm km\,s}^{-1}\)) with a minimum peak of \(T_{\rm mb} > 4\,\sigma_{\rm rms}\), and an equivalent size of at least the spatial resolution of 25\(\arcsec\) (\(\sim\)`<!-- -->`{=html}100 pc). Therefore, GMCs with a narrower velocity width than 2.5 km s\(^{-1}\) (if any) will not be identified by this analysis. As a result, we have identified a total of 71 GMCs, whose properties are summarized in Table [\[p3:co32cprops\]](#p3:co32cprops){reference-type="ref" reference="p3:co32cprops"}. The columns indicate the cloud ID, the cloud position (R.A., Decl.), the velocity, the peak intensity, the FWHM of the velocity width (\(\Delta V_{\rm FWHM}\)), the major/minor axes and orientation of the GMCs, the CO(\(J=3-2\)) luminosity (\(L^{\prime}_{\rm CO(3-2)}\)), virial mass (\(M_{\rm vir}\)) and cross identifications with the CO(\(J=1-0\)) and CO(\(J=2-1\)) GMC catalogs of previous papers. The cloud position and velocity are derived from an intensity-weighted mean over all the pixels of a GMC. The major/minor diameters (\(A_{\rm maj}, A_{\rm min}\)) without beam deconvolution are also indicated. The virial mass, calculated under the assumption that each cloud is spherical and can be parametrized by a density profile \(\rho \propto r^{-\beta}\), is given by \(M_{\rm vir} = 189\,\Delta V_{\rm FWHM}^2 R_{\rm deconv}\,M_{\odot}\), where \(R_{\rm deconv}\) is the deconvolved radius given by deconvolving the beam size (\(\theta_{\rm beam}\)) from the GMC size, i.e., \(R_{\rm deconv}=\sqrt{[A^2_{\rm maj}-\theta^2_{\rm beam}]^{1/2}[A^2_{\rm min}-\theta^2_{\rm beam}]^{1/2}}\). Note that the virial masses of 27 GMCs are not shown because their minor diameters are smaller than the beam size. Among the 71 GMCs in our catalog, only one GMC (GMC-50) is a new identification. The 70 GMCs have been previously identified in lower-\(J\) CO transition catalogs. GMC-50 is located outside of the observed field in and has a marginal detection in the CO(\(J=1-0\)) integrated intensity map in, but was not identified in their catalog. A total of 10 GMCs in our catalog are composed of more than two smaller GMCs in previous catalogs, due to their higher resolution data. On the other hand, a few single GMCs in the 's catalog appear to be composed of more than two GMCs in our catalog. This is because of our 8--15 times better sensitivity data, since decomposition of neighboring clouds is done based on peak differences exceeding 2 \(\sigma_{\rm rms}\). The ID numbers in these cases are marked with an asterisk in Table [\[p3:co32cprops\]](#p3:co32cprops){reference-type="ref" reference="p3:co32cprops"}. Since 6 GMCs (GMC-18, GMC-28, GMC-47, GMC-51, GMC-58, and GMC-71) are at the edge of the observed fields and their extents are uncertain, they are not included in further analysis and discussion. Histograms of the physical properties of 65 GMCs are presented in Figure [\[p3:co32gmcprop\]](#p3:co32gmcprop){reference-type="ref" reference="p3:co32gmcprop"}. The \(\Delta V_{\rm FWHM}\), \(R_{\rm deconv}\), \(L^{\prime}_{\rm CO(3-2)}\), and \(M_{\rm vir}\) are in the range of 1.2--17.3 km s\(^{-1}\), 12--157 pc, (0.8--25.8)\(\times10^4\) K km s\(^{-1}\) pc\(^2\), and (0.2--89.2) \(\times\)`<!-- -->`{=html}10\(^5\,M_{\odot}\), respectively. A lower limit for the virial mass is \(1.2\times10^5\,M_{\odot}\), assuming an effective radius of \(\sim\)`<!-- -->`{=html}50 pc and \(\Delta V_{\rm FWHM} =2.5\,{\rm km\,s}^{-1}\), which are close to the instrumental limits. Average values are \(\Delta V_{\rm FWHM} \sim 8.3\,\)km s\(^{-1}\), \(R_{\rm deconv}\la73\) pc, \(L_{\rm CO(3-2)} \sim 5.1\times10^4\) K km s\(^{-1}\) pc\(^2\), and \(M_{\rm vir}\la 7.7\times10^5\, M_{\odot}\). Note that we have used the non-deconvolved radius for the GMCs whose minor axes were smaller than the beam size, and thus the averaged \(R_{\rm deconv}\) and \(M_{\rm vir}\) should be regarded as upper limits. These averaged values are within the typical range as traced by CO(\(J=1-0\)) in LGGs. Compared to the high-resolution CO(\(J=3-2\)) observations towards clumps in LMC, the \(R_{\rm deconv}\) and \(M_{\rm vir}\) of M33 GMCs are larger than the typical values of LMC clumps, i.e. \(R_{\rm deconv}=1.1\)--12.4 pc and \(M_{\rm vir}=4.6\times10^3\)--\(2.2\times10^5\,M_{\odot}\). This is likely only due to our more limited spatial resolution, because the \(\Delta V_{\rm FWHM}\) of M33 GMCs is similar to that of the clumps at high resolution, \(\Delta V_{\rm FWHM}\) of \(\sim\)`<!-- -->`{=html}7 km s\(^{-1}\). This suggests that the M33 GMCs obtained in this work are composed of a single or a few small clumps at a similar velocity, but not a large amount of clumps. ### CO(\(J=3-2\)) and CO(\(J=1-0\)) distribution and \(R_{3-2/1-0}\) {#sec311} The CO(\(J=3-2\)) and CO(\(J=1-0\)) integrated intensity maps and also the CO intensity ratio (\(R_{3-2/1-0}\)) map for each individual GMC, with a common 25\(\arcsec\) resolution, are shown in panels (a)--(c) of Figures [\[fig6-gmc-60\]](#fig6-gmc-60){reference-type="ref" reference="fig6-gmc-60"}--[\[fig6-gmc-16\]](#fig6-gmc-16){reference-type="ref" reference="fig6-gmc-16"} and Figures 18--78 in the Appendix. The integrated intensity maps of CO(\(J=3-2\)) emission have been created by using the channel maps where the emission is above 2 \(\sigma_{\rm ch}\). The integrated intensity maps of CO(\(J=1-0\)) emission have been binned over the same velocity ranges as that of CO(\(J=3-2\)) emission. We find a general trend that the distribution of CO(\(J=3-2\)) emission and peaks are similar to that of CO(\(J=1-0\)) (e.g., GMC-3 in Figure 19, GMC-5 in Figure 21). However, in some cases (\(\sim 28\,\%\)), the CO(\(J=3-2\)) emission for some GMCs shows more compact distribution (e.g., GMC-1 in Figure [\[fig6-gmc-1\]](#fig6-gmc-1){reference-type="ref" reference="fig6-gmc-1"}, GMC-8 in Figure [\[fig6-gmc-8\]](#fig6-gmc-8){reference-type="ref" reference="fig6-gmc-8"}). In other GMCs (\(\sim 11\,\%\)), the CO(\(J=3-2\)) emission peak is offset from the CO(\(J=1-0\)) emission peak by over half HPBW (e.g., GMC-1 in Figure [\[fig6-gmc-1\]](#fig6-gmc-1){reference-type="ref" reference="fig6-gmc-1"}, GMC-7 in Figure 23). A further description of the spatial comparison between CO emission and massive SF sites is presented later in Section [3.5](#sec3.4.1){reference-type="ref" reference="sec3.4.1"}. We have calculated \(R_{3-2/1-0}\) by dividing the CO(\(J=1-0\)) map by the CO(\(J=3-2\)) map, after masking the region where the CO(\(J=1-0\)) emission is below 3 \(\sigma\) in its integrated intensity map. Note that each \(R_{3-2/1-0}\) map is masked except inside the boundary of the GMC, to avoid confusion with neighboring GMCs. The \(R_{3-2/1-0}\) values across GMCs vary greatly. Some GMCs are found to have a gradient of \(R_{3-2/1-0}\), while others have a relatively constant ratio. The averaged \(R_{3-2/1-0}\) is measured inside the boundary of GMCs and is provided in Column (9) of Table [\[p3:co32cprops\]](#p3:co32cprops){reference-type="ref" reference="p3:co32cprops"}. The CO(\(J=3-2\)) emission for four GMCs (GMC-49, 56, 57 and 70) have been detected with significant signal-to-noise ratio but the CO(\(J=1-0\)) emission has not been detected at the same position. In case that the \(R_{3-2/1-0}\) for these GMCs is uncertain, we calculate the lower limit of \(R_{3-2/1-0}\) at the CO(\(J=3-2\)) emission peak (see captions in each figure). Figure [\[p3:co32gmcprop_hist\]](#p3:co32gmcprop_hist){reference-type="ref" reference="p3:co32gmcprop_hist"} shows the histogram of \(R_{3-2/1-0}\) for the 65 GMCs. The averaged \(R_{3-2/1-0}\) ranges from 0.18 to 0.89, with a mean of 0.43. High ratios, \(R_{3-2/1-0}\) \(> 0.6\), are found at four GMCs: GMC-1, GMC-7, GMC-34 and GMC-42, which are located around the GHRs (NGC 604, NGC 595) and the vicinity of the galaxy center. ## Identification of YSGs and their Ages {#p3:age} The color magnitude diagram (CMD) is a powerful tool to provide insight into the age of the stellar component in a region of interest. Next we describe the method to identify YSGs as well as how to estimate their ages using the CMD. The results are then used to investigate the spatial correlation between the YSGs and the GMCs identified in Section [3.1](#sec31){reference-type="ref" reference="sec31"}. ### Extraction of Young Stars {#p3:4321} We have created the \(V\) magnitude-(\(B-V\)) CMD (i.e., \(M_V\)-(\(B-V)_0\) diagram) for the disk of M33 using the photometry catalog described in Section [2.3](#p3:2.2){reference-type="ref" reference="p3:2.2"}. To obtain intrinsic absolute magnitudes and then accurate ages, the photometric measurements must be corrected for two sources of reddening: the foreground extinction from our Galaxy and the internal reddening due to the interstellar matter in the disk of M33. We adopt the foreground galactic extinctions in the \(B\) and \(V\) bands of 0.181 and 0.139 mag, respectively. The reddening \(E(B-V)\) has been measured towards some individual clusters in M33, and ranges from 0.06 to 0.3 mag, with a typical value of 0.10. Due to the different density structures of the GMCs it is necessary to consider a variable internal reddening correction. This is especially important if the stars are embedded or lie behind the GMC, as the visual extinction is then expected to be higher. The exact correction for a given line of sight is rather uncertain, but we have used the available \(E(B-V)\) values obtained for some clusters, and the typical value of \(E(B-V)=0.10\) for the remaining stellar groups. The ratio of total to selective extinction \(R_V = A_V / E(B-V) \simeq 3.1\) in our Galaxy is used to obtain the visual extinction \(A_V\). We have used the theoretical isochrones in the Padova stellar population synthesis models to estimate the age of the stars. The adopted distance to M33 corresponds to a distance modulus (\(m-M)_0 = 24.6\). Note that the metallicity (\(Z\)) varies from \(0.6\,Z_{\odot}\) in the central region to 0.4 \(Z_{\odot}\) at 5 kpc galaxy radius. Figure [\[p3:fig3e\]](#p3:fig3e){reference-type="ref" reference="p3:fig3e"}(a) shows a \(M_V\)-(\(B-V)_0\) diagram for the stars in M33 with five stellar isochrones from the Padova models, assuming \(Z=0.5\,Z_{\odot}\), ages of 6, 10, 20, 30 and 100 Myr, and an average internal extinction correction of \(A_V \sim 0.22\). have estimated that \(\sim 40\) % of stars have \((B-V)_0\) between \(0.3-1.0\), but \(M_V\) in the range of 14.6--19.6 are likely foreground objects. Taking into account photometric errors and foreground galactic extinction, the criterion of \((B-V)_0 < 0.3\) is chosen to avoid the foreground stars. The faint limit of the \(V\)-band magnitude of \(M_V < 21.5\) is set to select only young massive stars (\(<\,100\) Myr). The minimum (initial) stellar masses at \(M_V\sim21.5\) for ages of 6, 10, 20, 30, 100 Myr correspond to 14, 13, 10, 8, and \(5\,M_{\odot}\), respectively. The orange boundary in Figure [\[p3:fig3e\]](#p3:fig3e){reference-type="ref" reference="p3:fig3e"}(a) indicates our selection criteria, \((B-V)_0 < 0.3\) and \(M_V < 21.5\), an area that should represent young stars. The selected young stars are plotted in Figure [\[p3:fig3e\]](#p3:fig3e){reference-type="ref" reference="p3:fig3e"}(b), where the point size in the plot is proportional to the stellar brightness in \(V\)-band. These young stars (\(<\) 100 Myr) are found to be distributed in a structure with two main arms and several weak multiple arms, which is in good agreement with. ### Young Stellar Number Density Map, Identification of YSGs and their Ages {#p3:5322} We describe the method to search for YSGs and how to estimate their ages. The term "YSGs" refers to the concentration of young stars selected using the criteria in Section [3.2.1](#p3:4321){reference-type="ref" reference="p3:4321"}. These will include both (young) star clusters as well as OB associations. Generally, a star cluster is defined to be a gravitationally bound system of several stars or more, whose concentration is larger than that of the surrounding stellar background, with a typical size of a few parsec to tens of parsecs. The definition of an OB association is usually a single, loosely bound or unbound concentration of early-type luminous stars, typically extending several tens to over hundred parsecs. The typical masses of young (\(<10\) Myr) clusters and associations are 10--10\(^4\,M_{\odot}\) and 10\(^3\)--10\(^6\,M_{\odot}\), respectively. Here we do not aim to distinguish between them because our interest is to investigate the collective properties of YSGs associated with GMCs. In previous studies, clusters or associations have been selected to be as stellar groups just by visual inspection, and therefore the selection may have been subjective. Recently, a more objective approach has been carried out by, who have performed a stellar cluster identification in NGC 6822 and LMC using the stellar surface density map, and have shown the hierarchical structure of blue stellar clusters. To select YSGs in an objective manner for M33 as well, we define them as an excess in the number of young stars per unit of area (\(n^{\ast}\)) following 's approach. We set the unit of area to the same scale as the pixel size of CO(\(J=3-2\)) map, a \(8\arcsec\times8\arcsec\) box (corresponding to \(\sim30\,{\rm pc}\times 30\,{\rm pc}\)), which is comparable to the typical GMC size. The pixel size of the stellar number density map is set to half of the pixel size of the CO(\(J=3-2\)) map, 15 pc. Figure [\[p3:fig3d\]](#p3:fig3d){reference-type="ref" reference="p3:fig3d"} shows the number density map of young stars per each \(30\,{\rm pc}\times 30\,{\rm pc}\) area in M33 disk. It becomes more clear than in Figure [\[p3:fig3e\]](#p3:fig3e){reference-type="ref" reference="p3:fig3e"}(b) that most of the regions with \(n^{\ast} > 2\) stars per unit area are distributed along the spiral arms of M33. The maximum value of stellar number density in M33 is found in the center of NGC 604, \(n^{\ast} = 28\) stars per area. Over-density regions (\(n^{\ast} > 15\) stars per area) coincide with prominent Hii regions such as NGC 604, NGC 595, NGC 592, IC 135, IC 139, IC 140, and the galaxy center, as labeled in Figure [\[p3:fig3d\]](#p3:fig3d){reference-type="ref" reference="p3:fig3d"}. Next we compare the number density map with other wavelength data. Figure [\[p3:fig4\]](#p3:fig4){reference-type="ref" reference="p3:fig4"} shows the M33 disk in CO(\(J=1-0\)), \(24\,\micron\), H\(\alpha\), and FUV. The background-subtracted FUV map is a reproduction from. Interestingly, the distributions of the molecular gas traced by CO(\(J=1-0\)) in Figure [\[p3:fig4\]](#p3:fig4){reference-type="ref" reference="p3:fig4"}(a) and the young stars in Figure [\[p3:fig3d\]](#p3:fig3d){reference-type="ref" reference="p3:fig3d"} exhibit slightly different patterns between the northern and the southern sides of the disk: molecular gas is more abundant in the northern side of the disk, while the opposite is true for young stars. Note that this asymmetrical pattern of the young stellar distributions is not because the photometric depth is inhomogeneous over the disk, as the completeness limits are similar (below \(V\sim21.5\) mag) both in the northern and southern disk. The distributions of \(24\,\micron\) and H\(\alpha\) emissions show multiple spiral arms similarly to the density map of young stars, but their spatial distribution patterns are not always similar: the \(24\,\micron\) and H\(\alpha\) emissions are enhanced at GHRs such as NGC 604 and NGC 595 compared to other disk regions, while the surface density map of young stars is more pronounced along the spiral arms (Figure [\[p3:fig4\]](#p3:fig4){reference-type="ref" reference="p3:fig4"}b, c). The distribution of the young star number density is remarkably similar to that of the FUV emission in Figure [\[p3:fig4\]](#p3:fig4){reference-type="ref" reference="p3:fig4"}(d), which traces the spiral structure and reproduces the north-south asymmetry. This is consistent with FUV being a diagnostic of the younger (30--100 Myr) stellar population. In this work, we adopt 5 young stars per \(30\,{\rm pc}\times 30\,{\rm pc}\) area as the minimum surface stellar number required to be classified as a YSG, and then we estimate its age from the CMD, assuming that stars are coeval. First, we search a region with a peak of \(n^{\ast} > 5\) stars per area around each GMC, then determine the extent of the YSG at a level of \(n^{\ast} = 2\) stars per unit area. A Gaussian is then fitted to the radial profile of the young stellar distribution down to \(n^{\ast} > 2\), and the extent of YSGs is set as the FWHM of that Gaussian fit assuming a circular shape (\(r_{\rm cl}\)). Second, we have created the \(M_V\)-(\(B-V)_0\) diagrams for the stars contained within the radius of \(r_{\rm cl}\). Finally we estimate their age using the Padova models. As an example, Figure [\[p3:fig3a\]](#p3:fig3a){reference-type="ref" reference="p3:fig3a"}(a) shows the number density map of young stars, overlaid on the CO(\(J=3-2\)) map of the NGC 604 region. Three GMCs are identified, GMC-34, GMC-1 and GMC-27, from north to south in this region. We find three peaks of \(n^{\ast} > 5\) stars per area around these GMCs, at position (\(\alpha_{J2000}\), \(\delta_{J2000}\)) = (\(1\h34\m31\fs4\), \(30\arcdeg47\arcmin48\farcs0\)), (\(1\h34\m33\fs4\), \(30\arcdeg47\arcmin03\farcs5\)) and (\(1\h34\m33\fs0\), \(30\arcdeg46\arcmin23\farcs8\)) from north to south, which we refer to as YSG-71, YSG-73 and YSG-72, respectively. The radii of these YSGs are measured to be \(r_{\rm cl}=9.5\arcsec\) (39 pc), \(21.2\arcsec\) (86 pc) and \(10.4\arcsec\) (42 pc), respectively. Next, we derive ages for each YSG by fitting theoretical isochrones to the CMDs of resolved stars. Figure [\[p3:fig3b\]](#p3:fig3b){reference-type="ref" reference="p3:fig3b"} shows the CMD of the three identified YSGs above, YSG-71, YSG-72 and YSG-73. Dot symbols represent the stars within the radius of \(r_{\rm cl}\) and the lines represent the Padova theoretical isochrone tracks, assuming a metallicity of 0.5 \(Z_{\odot}\). Solid lines represent the upper limit of the estimated ages, regarding the magnitude and color of the brightest main-sequence (MS) star as MS turnoffs for a young star group. In this case, the more massive stars are assumed to have already evolved from the MS lines at (\(B-V)_0\sim0\) and less massive stars are still in the MS. Dashed lines represent the lower limit of the estimated ages, considering that all stars in the YSGs are still in the MS at (\(B-V)_0\sim0\), without any more massive stars in the group evolving to red giant branches. Although chances are low that these massive stars are detected at redder colors, because the more massive the stars, the faster they evolve, no detection of stars at (\(B-V)_0 > 0\) can suggest that more massive stars have not existed before. Note that in either case, we have used only stars at (\(B-V)_0 < 0.3\) to avoid foreground stars (Section [3.2.1](#p3:4321){reference-type="ref" reference="p3:4321"}). The measurable age limit of these YSGs using the Padova models is 3 Myr. In this way, the estimated ages of the three YSGs YSG-71, YSG-72 and YSG-73 are in the range of 3--28 Myr, 3--28 Myr and 3--5 Myr, respectively. The first two YSGs include a smaller number of stars (11 and 21 O stars for YSG-71 and YSG-72, respectively) and mostly less bright stars (\(M_V\ga19\) mag), which makes a larger uncertainty in the estimations. Figure [\[p3:fig3a\]](#p3:fig3a){reference-type="ref" reference="p3:fig3a"}(b) shows the distribution of the young stars, overlaid on the H\(\alpha\) map (grayscale). The size of the blue circle representing each star is proportional to the brightness in \(V\)-band, that is, the larger, the brighter. Many bright young stars are found in YSG-73 and concentrated in the center of the Hii region. In fact, NGC 604 is known to be the most luminous and massive Hii region in M33, containing more than 200 O-type stars. The other two YSGs (YSG-71 and YSG-72) do not have cross identifications in previous cluster catalogs, but are included in previous star complex and OB association catalogs. ### The Catalog of YSGs We have identified 75 YSGs in M33 and have estimated their ages as described in Section [3.2.2](#p3:5322){reference-type="ref" reference="p3:5322"}. Note that only YSGs associated with the identified GMCs are included (see Section [3.3](#p3:association){reference-type="ref" reference="p3:association"}). Figures [\[p3:cmd1\]](#p3:cmd1){reference-type="ref" reference="p3:cmd1"}--[\[p3:cmd3\]](#p3:cmd3){reference-type="ref" reference="p3:cmd3"} show the CMDs of 75 YSGs with the adopted Padova theoretical isochrone tracks. Similarly to Figure [\[p3:fig3b\]](#p3:fig3b){reference-type="ref" reference="p3:fig3b"}, solid and dashed lines represent the upper and lower limits of the estimated ages. Blue dots represent the stars within the \(r_{\rm cl}\) of the YSGs. In Figures [\[p3:cmd1\]](#p3:cmd1){reference-type="ref" reference="p3:cmd1"}, [\[p3:cmd2\]](#p3:cmd2){reference-type="ref" reference="p3:cmd2"}, and [\[p3:cmd3\]](#p3:cmd3){reference-type="ref" reference="p3:cmd3"}, we show the CMDs of 18 YSGs younger than at most 10 Myr, 17 YSGs older than at least 10 Myr, and the remaining 40 YSGs, respectively. Table [\[p3:ys\]](#p3:ys){reference-type="ref" reference="p3:ys"} shows the properties of the YSGs identified in our catalog, including positions, size (\(r_{\rm cl}\)), number of O stars, \(V\)-magnitude of the brightest O star in the YSG, applied reddening correction, estimated age and cross identifications. The number of O stars are counted only for those within our criterion on the \(M_V\)-(\(B-V)_0\) diagrams, i.e., \((B-V)_0 < 0.3\) and \(M_V < 21.5\). Note that this stellar age provides just an upper limit for the YSGs, partly because reddening correction is uncertain. Twenty-three YSGs do not coincide with previously identified clusters, but are included in star complex and OB association catalogs. The remaining are found to have cross identifications with previous cluster catalogs. The radii of the YSGs range from \(6\farcs6\) to \(22\farcs4\) (27--91 pc) and the average is 11\(\arcsec\) (46 pc). The number of O stars in a YSG ranges from 9 to 169 stars. The total stellar mass of the YSGs are estimated to be \(10^{3.5}-10^{4.7}\,M_{\odot}\), assuming that the YSGs are characterized by a Salpeter IMF (slope-2.35), with masses spanning 0.1--58 \(M_{\odot}\). These sizes and masses are comparable to the typical values of OB associations . It is possible to estimate the stellar spectral type of the primary ionizing star contained in a YSG from the brightest \(V\)-magnitude and \((B-V)_0\) color. The most massive stars are found in YSG-21 and YSG-73 and their masses are estimated to be 51 \(M_{\odot}\), corresponding to O3.5V. The brightest stars in YSG-39 and YSG-60 are less massive than in any other YSGs, 8 \(M_{\odot}\) (corresponding to B2V). The latter is comparable to small clusters in our Galaxy and LMC that can be ionized by a single mid-O or B0 star. The ages of the YSGs are estimated to be in the range between \(4\,\pm\,1\) Myr and \(31\,\pm\,19\) Myr, with an average of \(12\,\pm\,5\) Myr. We can compare the ages with other catalogs. Although the YSGs in our catalog do not fully overlap with the clusters whose ages have been previously determined (SM; PPL; SSGH), only 5 YSGs coincide, YSG-1, YSG-13, YSG-41, YSG-63, and YSG-67. The previously derived ages are 25\(^{+38}_{-14}\) Myr (PPL), 13 Myr (SM), 16\(^{+4}_{-3}\) Myr (PPL) or 21 Myr (SM), 5 Myr (SM) or 13\(^{+7}_{-5}\) Myr (PPL), and 40\(^{+60}_{-24}\) Myr (PPL), respectively, while they are \(9\,\pm\,2\) Myr, \(11\,\pm\,5\) Myr, \(9\,\pm\,2\) Myr, \(5\,\pm\,2\) Myr and \(14\,\pm\,11\) Myr in our estimation. Compared to the results for YSG-1, YSG-41, YSG-63 and YSG-67 in PPL who have used a similar method, we notice that our estimations of ages are slightly younger than them but consistent within the uncertainties. This slight difference is mostly because they have used a small member star selection radius (within 1--2\(\arcsec\)) of a compact stellar cluster and also because they have applied different theoretical isochrones in the Padova models, corresponding to \(Z=0.2\,Z_{\odot}\). Besides, even though a cluster in their catalogs is associated with one of our YSGs, in general it only represents a small portion of the YSG, and thus it is not clear in principle whether they represent exactly the same entity. For the ages of the other two YSGs (clusters) in known GHRs (NGC 595 and NGC 604), YSG-2 and YSG-73, they had been previously estimated to be 4--6 Myr and 3--5 Myr, respectively, using optical spectroscopy combined with an instantaneous starburst model. The ages of these young clusters are in good agreement with our results, \(6\,\pm\,3\) Myr and \(4\,\pm\,1\) Myr, respectively. ## Hii Regions and YSGs Associated with GMCs {#p3:association} In this section, we investigate the associations of H\(\alpha\), 24 \(\micron\) sources, and the identified YSGs, with the GMCs. These associations are listed in Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"}. Hii regions are tracers of the younger stellar population (\(<\) 10 Myr), while 24 \(\micron\) sources represent massive SF sites embedded in dust. When the extent of Hii regions, 24 \(\micron\) sources and YSGs are confined within the boundary of a GMC, they are treated as being associated with the GMC. The Hii region catalogs include those Hii regions with H\(\alpha\) luminosities larger than \(10^{34.4}\) erg s\(^{-1}\). Therefore, the sensitivity of the survey is high enough to detect the equivalent to a single O9V type star, assuming a standard relation between ionizing photon rates and H\(\alpha\) luminosities. For the 24 \(\micron\) sources, the faint end of the 24 \(\micron\) luminosity in the catalog is \(\sim10^{37.7}\) erg s\(^{-1}\), corresponding just to an Hii region illuminated by a single B1.5V star. Note that the 24 \(\micron\) source catalog includes discrete sources such as Hii regions, supernovae remnants and planetary nebulae, but most of the 24 \(\micron\) sources are classified as Hii regions, which contain embedded young stellar clusters with ages of 3--10 Myr. The 24 \(\micron\) sources have generally corresponding H\(\alpha\) emitting sources, but there are some cases of 24 \(\micron\) sources without corresponding Hii source (see the last column in Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"}). All GMCs except GMC-60 are found to be associated with Hii regions and/or 24 \(\micron\) sources, while 51 GMCs are found to be associated with YSGs. In 10 GMCs no Hii region exists but 24 \(\micron\) sources have been identified, while in 16 GMCs there are Hii regions without corresponding 24 \(\micron\) sources. The H\(\alpha\) and 24 \(\micron\) luminosities, \(L({\rm H}\alpha)\) and \(L(24\,\micron)\), and SFR for each GMC are also listed in Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"}, where the SFR for each GMC is calculated from the \(L({\rm H}\alpha)\) and \(L(24\,\micron)\) using the relation between the extinction-corrected H\(\alpha\) emission and the SFR. The \(L({\rm H}\alpha)\) and \(L(24\,\micron)\) for each GMC, are measured with a circular aperture of size equal to that of the associated GMC (100 to 400 pc diameter). These aperture photometry measurements include local background subtraction. Determination of the local background is done by fitting the emission over the aperture with Gaussian functions. These aperture sizes are similar or larger than a typical Hii region (1 pc--100 pc) and thus they may potentially contain several Hii regions. A more careful local background subtraction for these Hii regions would just result in even larger differences in luminosities between a small Hii region and a large one because the local background contamination tends to be systematically large at the faint end. The \(L({\rm H}\alpha)\) and \(L(24\,\micron)\) range from \(2.49\times10^{35}\) to \(7.67\times10^{39}\) erg s\(^{-1}\), and from \(1.67\times10^{37}\) to \(4.04\times10^{40}\) erg s\(^{-1}\), respectively. The most luminous source both in H\(\alpha\) and 24 \(\micron\) emission is found at GMC-1, corresponding to NGC 604. The SFR of the M33 GMCs is characterized by a large scatter, ranging from \(8.89\times10^{-5}\) to 0.216 \(M_{\odot}\,{\rm yr}^{-1}\). We estimate the averaged H\(\alpha\) attenuation over a GMC (i.e., a few 100 pc scale) using \(L(24\,\micron)/L(\rm H\alpha)\) ratio as a proxy. The H\(\alpha\) attenuation is given by \(A_{\rm H\alpha}=2.5\, {\rm log} [1+0.038\,L(24\,\micron)/L({\rm H\alpha})]\). We list \(A_{\rm H\alpha}\) in Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"}. The \(A_{\rm H\alpha}\) is generally low (less than 1 mag) and in good agreement with previous studies. The averaged \(A_{\rm H\alpha}\) is 0.4 mag (\(A_V\sim0.5\)) and is consistent with the adopted extinction for individual YSGs in Section [3.2.1](#p3:4321){reference-type="ref" reference="p3:4321"}. Relatively highly obscured sources (\(A_{\rm H\alpha} >\) 2 mag) are found in three GMCs: GMC-6, 15 and 68. The ages for the YSGs associated with these GMCs may be underestimated. Note that highly obscured regions (\(A_{\rm H\alpha} >\) 3 mag) have been found in other nuclei of galaxies, but we can discard that possibility in M33. ## GMC Types {#p3:sec534} We classify the GMCs into four types according to the age of the associated YSGs and Hii regions:\ Type A: GMCs are not associated with Hii regions nor YSGs;\ Type B: GMCs are associated with Hii regions, but not with any YSG;\ Type C: GMCs are associated with Hii regions and young (\(<10\) Myr) stellar groups;\ Type D: GMCs are associated with Hii regions and relatively old (10--30 Myr) stellar groups. Note that averaged values between the upper and lower limits of the stellar ages are used for this classification. If a GMC is associated with several YSGs of different ages, the classification is done based on the youngest stellar group. Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"} lists the classification of each M33 GMC into one of these four types. Out of the 65 GMCs, 1 (1 %), 13 (20 %), 29 (45 %), and 22 (34 %) are found to be Types A, B, C, and D, respectively. Table [\[p3:class\]](#p3:class){reference-type="ref" reference="p3:class"} summarizes these statistics. The Type C GMCs are the majority among all GMC types. The estimated age of some YSGs may be uncertain. We select 44 GMCs with accurate age estimations for their associated YSGs in order to check how robust the relative percentage is among the different types. In this regard, only YSGs whose age upper limit is smaller than 10 Myr (see Figure [\[p3:cmd1\]](#p3:cmd1){reference-type="ref" reference="p3:cmd1"}) and lower limit larger than 10 Myr (see Figure [\[p3:cmd2\]](#p3:cmd2){reference-type="ref" reference="p3:cmd2"}) are used. With this condition we obtain that 1 (2 %), 13 (30 %), 19 (43 %), and 11 (25 %) are classified as Types A, B, C, and D, respectively. The 44 selected GMCs are indicated in Table [\[p3:gmctype\]](#p3:gmctype){reference-type="ref" reference="p3:gmctype"}. We confirm the trend in the relative percentages, being Type C GMCs the majority among all GMC types. A caveat in the Type B and C classifications is that the percentages may be partly influenced by the instrumental sensitivity in 's survey, which may limit the detection of some YSGs. Given the instrumental sensitivity of 's survey, some of Type B sources possibly contain YSGs that have not been detected. However, the sensitivity restricts the contained YSGs to be less massive than \(13\,M_{\sun}\) (\(\sim\)B1V star) (Section [3.2.1](#p3:4321){reference-type="ref" reference="p3:4321"}), and Type C (YSGs with ages \(< 10\) Myr) are typically more massive than that. The three GMC classifications in the LMC, Types i, ii and iii correspond to Types A, B and C in our classification, respectively (see also Table [\[p3:class\]](#p3:class){reference-type="ref" reference="p3:class"}). In the LMC, Type ii (Type B) represents the largest number among all GMC types, while it is Type C in M33. The percentage of Type C GMCs in M33 (40--45 %) is also larger than in LMC (26 %). Note that the clusters with ages of 10--30 Myr in LMC are likely far away from their parent GMCs and thus the relative percentage is largely unknown. In addition, these discrepancies can be partly explained by the different spatial resolution and observed line, as the classification of LMC GMCs have been done with 40 pc resolution in CO(\(J=1-0\)). First, the GMCs in M33 are typically double the size of those in LMC. If a M33 GMC is composed of two different types of GMCs, for example Types B and C, whose sizes are comparable to those in LMC, then the GMC would have been classified as Type C. Therefore, the classifications will be affected. Second, regarding the different tracers used, the M33 GMCs identified from CO(\(J=3-2\)) data represent the denser and/or warmer molecular gas which is more directly linked to massive SF, while the LMC GMCs identified from CO(\(J=1-0\)) data trace the bulk of the molecular gas. In fact, in the case of M33 GMCs identified by using CO(\(J=1-0\)) data, more than two-thirds have associated Hii regions, which is a similar proportion in LMC. ## Physical Properties of GMC Types {#sec3.4.1} Figure [\[p3:typeprop_fig\]](#p3:typeprop_fig){reference-type="ref" reference="p3:typeprop_fig"} shows the distributions of the line width (\(\Delta V_{\rm FWHM}\)), deconvolved radius (\(R_{\rm deconv}\)), CO(\(J=3-2\)) luminosity (\(L^{\prime}_{\rm CO(3-2)}\)), averaged line ratio (\(R_{3-2/1-0}\)) and SFR of the 65 GMCs. Each row represents each GMC type. The shaded areas represent the histograms for the 44 selected GMCs with more accurate classifications (Section [3.4](#p3:sec534){reference-type="ref" reference="p3:sec534"}). Table [\[p3:typeprop_tab\]](#p3:typeprop_tab){reference-type="ref" reference="p3:typeprop_tab"} summarizes the mean and standard deviation of these variables for each GMC type. The \(R_{3-2/1-0}\) distributions of Type C and D show a peak over 0.4, while those of Type B are always smaller than the average value of the former types (i.e., \(R_{3-2/1-0}\)\(< 0.3\)). The SFR distributions of Type C and D show a peak over \(2\times10^{-2}\,M_{\odot}\,{\rm yr}^{-1}\), while those of the other three types are smaller. We have used Kolmogorov-Smirnov (K-S) tests to check whether the distributions for the different types differ. We cannot discard that the \(\Delta V_{\rm FWHM}\), \(R_{\rm deconv}\) and \(L^{\prime}_{\rm CO(3-2)}\) distributions for these three types arise from the same distribution. However, the test indicates that \(R_{3-2/1-0}\) and SFR distributions of Type C and D are different to those of Type B (p-value \(<\) 0.01). The selected GMCs with more accurate classification, as shown in the shaded histograms (Figure [\[p3:typeprop_fig\]](#p3:typeprop_fig){reference-type="ref" reference="p3:typeprop_fig"}), indicate that the average values of all quantities for Type C are the largest, followed by Type D, Type B, and Type A, in this order. This suggests that the Type C GMCs are relatively large and show the most active SF among all types. Note that all quantities of the single Type A object are smaller than for the other three types. However, additional Type A sources need to be identified to properly characterize the properties of this class of objects. In the following subsections, we present a close-up view of the CO(\(J=3-2\)), CO(\(J=1-0\)), \(R_{3-2/1-0}\) and H\(\alpha\) maps of the largest GMC in each GMC type, in order to illustrate the different properties among the four GMC types. Figures [\[fig6-gmc-60\]](#fig6-gmc-60){reference-type="ref" reference="fig6-gmc-60"}--[\[fig6-gmc-16\]](#fig6-gmc-16){reference-type="ref" reference="fig6-gmc-16"} show for four GMCs representative of each type: (a) the CO(\(J=3-2\)) integrated intensity map, (b) CO(\(J=1-0\)) map, (c) the \(R_{3-2/1-0}\) map, and (d) the H\(\alpha\) images. GMC-60 is chosen as a representative of Type A, GMC-8 of Type B, GMC-1 of Type C, and GMC-16 of Type D. The positions of Hii regions, the 24 \(\micron\) sources, and the surface density map of young stars are also marked in these figures. The CO(\(J=3-2\)) emission peaks are in general slightly offset from the CO(\(J=1-0\)) emission peak as we mentioned in Section [3.1.1](#sec311){reference-type="ref" reference="sec311"}, but they are closer to the location of massive SF sites such as YSGs and Hii regions. These differences of spatial distributions necessarily results in a relatively high \(R_{3-2/1-0}\) area in a GMC around such massive SF sites, and the gradient of \(R_{3-2/1-0}\) that gradually increase in the direction toward such massive SF sites. ### GMCs without Massive SF Regions (Type A) {#p3:gmc60} Figure [\[fig6-gmc-60\]](#fig6-gmc-60){reference-type="ref" reference="fig6-gmc-60"} shows the images of GMC-60 of Type A. No large variations in \(R_{3-2/1-0}\) (\(\sim 0.28\) on average over the GMC) or bright H\(\alpha\) emission spots in this GMC are found. Besides it is a relatively small GMC compared to the average size and mass of all identified GMCs (Table [\[p3:typeprop_tab\]](#p3:typeprop_tab){reference-type="ref" reference="p3:typeprop_tab"}). Since only one GMC is classified as Type A, it is unclear whether the Type A source is representative of a class of similar sources or just a statistical outlier. Additional Type A sources need to be identified to properly characterize the properties of this class of objects. ### GMCs with Hii Regions but not YSGs (Type B) {#p3:gmc8} One-fifth of the GMCs in our catalog are classified as Type B. We cannot assure significant differences in \(\Delta V_{\rm FWHM}\), size and \(L^{\prime}_{\rm CO(3-2)}\) compared to those of Type C and Type D (see Figure [\[p3:typeprop_fig\]](#p3:typeprop_fig){reference-type="ref" reference="p3:typeprop_fig"}). However, the \(R_{3-2/1-0}\) is smaller than the latter two types, with \(R_{3-2/1-0} \sim 0.3\) on average, and never exceeding 0.5. In addition, the SFR is also lower than the latter two types. We find that all Type B GMCs share a common trend, in the sense that higher \(R_{3-2/1-0}\) inside GMCs are preferentially found close to Hii regions and 24 \(\micron\) sources, although the peak of \(R_{3-2/1-0}\) is not always coincident with the central position of Hii regions. No large spatial variation in \(R_{3-2/1-0}\) over the GMC is found. The associated Hii regions and 24 \(\micron\) sources show a relatively random distribution over the GMC. In addition, these sources are relatively small both in extent and in luminosity. In fact, we find that they do not usually exceed \(L({\rm H\alpha})\sim 10^{38}\) erg s\(^{-1}\) or \(L({\rm 24\,\micron})\sim 10^{38.5}\) erg s\(^{-1}\), which corresponds to a single O7.5V star, assuming the conversion factor from ionizing photon rates and extinction-corrected H\(\alpha\) luminosities in. This suggests that the associated Hii regions are ionized primarily by a star as massive as a O7.5V star, or later type of OB stars accompanied by several other less massive stars. As an example of Type B GMCs, we present a close-up view of GMC-8 in Figure [\[fig6-gmc-8\]](#fig6-gmc-8){reference-type="ref" reference="fig6-gmc-8"}. This GMC is known as the most massive GMC as traced by CO(\(J=1-0\)), but it is just identified in our CO(\(J=3-2\)) catalog as the eighth brightest GMC. The CO(\(J=3-2\)) emission shows a more compact distribution than that of CO(\(J=1-0\)) and the emission peaks coincide. The \(R_{3-2/1-0}\) slightly increases at the center of the GMC or close to Hii regions and 24 \(\micron\) sources, but no large variation is found all over the GMC. ### GMCs with Hii Regions and YSGs (Type C) {#p3:gmc1} About half of the GMCs in our catalog are classified as Type C. The \(\Delta V_{\rm FWHM}\), size and \(L^{\prime}_{\rm CO(3-2)}\) of this type are similar to those in other types, but \(R_{3-2/1-0}\) and SFR are the largest among all (Figure [\[p3:typeprop_fig\]](#p3:typeprop_fig){reference-type="ref" reference="p3:typeprop_fig"} and Table [\[p3:typeprop_tab\]](#p3:typeprop_tab){reference-type="ref" reference="p3:typeprop_tab"}). There are gradients of \(R_{3-2/1-0}\) around massive SF sites. Peaks of \(R_{3-2/1-0}\) do not coincide in general with the positions of these SF sites, as also found in Type B GMCs. The average \(R_{3-2/1-0}\) of this type, \(\sim\)`<!-- -->`{=html}0.5, is usually larger than that of Type B, and even exceeds over 1.0 at its peak. In addition, unlike Type B, we find that the associated Hii regions usually exceed the equivalent \(L({\rm H\alpha})\) of a single O7.5 star. As an example of Type C GMCs, we show the maps for GMC-1 in Figure [\[fig6-gmc-1\]](#fig6-gmc-1){reference-type="ref" reference="fig6-gmc-1"}. The CO(\(J=3-2\)) emission shows a more compact distribution than that of CO(\(J=1-0\)). GMC-1 overlaps with two YSGs, YSG-73 (corresponding to the GHR NGC 604) and YSG-72. The emission peak of CO(\(J=3-2\)) is offset from that of CO(\(J=1-0\)) and is located northward, closer to the largest YSG, YSG-73, and the center of the GHR. A high \(R_{3-2/1-0}\) (\(>0.8\)) area is found on the southern side of YSG-73, where compact Hii regions are located. High \(R_{3-2/1-0}\) (\(>0.8\)) preferentially around the GHRs suggests that these areas may be excited by young massive stars. A gradient of \(R_{3-2/1-0}\) extends from its peak to the south. Panel (d) shows that the concentrations of many bright (massive) stars are located at the heart of the GHR. Note that it is not certain whether the YSG-72 is physically associated with this GMC and NGC 604 because there is no spatial correlation between the YSG, Hii region, and \(R_{3-2/1-0}\). ### GMCs with Hii Regions and 10--30 Myr YSGs (Type D) {#p3:gmc16} About one-third of the GMCs in our catalog are classified as Type D. Generally, the physical properties of this type is similar to Type C (see Figure [\[p3:typeprop_fig\]](#p3:typeprop_fig){reference-type="ref" reference="p3:typeprop_fig"} and Table [\[p3:typeprop_tab\]](#p3:typeprop_tab){reference-type="ref" reference="p3:typeprop_tab"}). We also find that the associated Hii regions are slightly less bright than those in Type C, but larger than those in Type B. Figure [\[fig6-gmc-16\]](#fig6-gmc-16){reference-type="ref" reference="fig6-gmc-16"} shows the maps of GMC-16, a representative of Type D GMCs. The CO(\(J=3-2\)) emission shows a similar distribution to CO(\(J=1-0\)), and both CO peaks coincide. Again, the peak of \(R_{3-2/1-0}\) (\(\sim0.4-0.6\)) is preferentially enhanced in the vicinity of the associated massive SF sites but not always coincident with them. Such relatively high \(R_{3-2/1-0}\) area is elongated from north to south and there is a gradient of \(R_{3-2/1-0}\) from the \(R_{3-2/1-0}\) peak in the middle of the GMC to the both west and east edges of the GMC. # Discussion {#p3:disc} ## Evolution of the GMCs In this section, we discuss based on our results the possible evolution of a GMC by focusing on the dense gas formation occurring around the first generated stars, which presumably leads to SF. We also estimate the typical lifetime of a GMC in M33. Since there is no correlation between Hii regions or YSGs, we infer that Type A is at an evolutionary stage before massive SF. In the case of LMC with 40 pc resolution and using the CO(\(J=1-0\)) line, have shown that about a quarter of all GMCs are Type A. Possible reasons that so few GMCs in M33 are classified as Type A are spatial resolution and sensitivity limit and the choice of the molecular gas tracer, as explained in Section [3.4](#p3:sec534){reference-type="ref" reference="p3:sec534"}. Type B GMCs, which are associated with Hii regions but not with YSGs, are interpreted as being at an evolutionary stage just after the formation of massive stars, such as a single O7.5V star or later OB type stars, that are still partly embedded to be seen in the optical as YSGs. We suggest that Type C GMCs being associated both with Hii regions and very young stellar groups (less than 10 Myr) are at an evolutionary stage with active massive star formation, as massive as earlier types than O7.5V stars. Type D GMCs associated with Hii regions and relatively old (10--30 Myr) stellar groups are considered to be at an evolutionary stage where they have been continuously forming massive stars for over at least 10 Myr. We find that \(R_{3-2/1-0}\) is enhanced around the massive SF sites for Types B, C and D GMCs (Section [3.5.2](#p3:gmc8){reference-type="ref" reference="p3:gmc8"}--[3.5.4](#p3:gmc16){reference-type="ref" reference="p3:gmc16"}). In order to roughly quantify the offset between them, we plot in Figure [\[p3:mindist\]](#p3:mindist){reference-type="ref" reference="p3:mindist"} the number of pixels above a certain \(R_{3-2/1-0}\) (0.1, 0.3, 0.5, 0.7 and 0.9) as a function of the distance to its closest YSG. The average sizes of the 75 YSGs are shown in the same figure, for reference. The dashed lines represent the frequency distribution expected if the same number of YSGs are distributed at random in the observed area. It shows that the distance to the nearest YSG becomes smaller as \(R_{3-2/1-0}\) increases, but it must be noted that the \(R_{3-2/1-0}\) peaks and the nearest YSGs do not coincide in general. We can assume that \(R_{3-2/1-0}\) is a good tracer of density when \(R_{3-2/1-0} < 0.7\), if kinetic temperature is low (\(\sim\)`<!-- -->`{=html}20 K), as explained in. This suggests that dense molecular gas fraction is enhanced in the vicinity of previously formed YSGs. For regions with high \(R_{3-2/1-0} > 0.8\), the molecular gas can have a higher kinetic temperature due to the nearest GHRs and YSGs (within \(\sim\)`<!-- -->`{=html}50 pc). This is in good agreement with the suggestion by that the GHRs may raise the kinetic temperature of the neighboring (\(<\,100\,\)pc) molecular gas. Therefore, the enhancement of dense molecular gas fraction around the previously formed massive stars imply that once SF starts, dense gas formation is favored around the SF sites, which will then lead to the next generation of SF. Next we estimate a lower limit of the lifetime of a GMC from the age of the oldest stellar group contained within the GMC, assuming that the GMCs and YSGs are being formed in a nearly steady evolution. If the classification is done based on the oldest stellar group within a GMC, the frequency distribution of each type is 1 (2 %), 13 (20 %), 19 (29 %), and 32 (49 %) out of 65 selected GMCs. The YSGs associated with Type C and Type D GMCs are \(8\,\pm\,4\) Myr and \(18\,\pm\,10\) Myr old on average, respectively. Therefore, we estimate that the timescale for each evolutionary stage is \(8\,\pm\,4\) and \(10\,\pm\,6\) Myr in Type C and Type D, from the age difference between the associated YSGs. If we further assume that the timescale for each evolutionary stage is proportional to the number of GMCs, then Types A, B, C, and D are estimated to be \(\sim\)`<!-- -->`{=html}1, 3--7, 5--10, and 8--17 Myr, respectively. As a result, the typical lifetime (from Type A to Type D) of a GMC with masses of \(\ga10^5\,M_{\odot}\) is roughly 20--40 Myr. Note that these results do not change significantly when we use only 44 GMCs with accurate ages for their YSGs (Section [3.4](#p3:sec534){reference-type="ref" reference="p3:sec534"}). In LMC, the lifetime of a GMC with mass as small as \(5\times10^4\,M_{\odot}\) was estimated to be 20--30 Myr. The timescales of Types A, B and C in LMC (corresponding to Types i, ii and iii) are 6 Myr, 13 Myr, and 7 Myr, respectively. The timescales for Type B in M33 are smaller than that in LMC, and for Type C they are similar or slightly larger. It is not appropriate to compare the lifetime for Type A in M33 and in LMC because of the limited number of GMCs in this type. No classification exists for LMC GMCs after Type C, thus we cannot compare (see Section [3.4](#p3:sec534){reference-type="ref" reference="p3:sec534"} for a discussion). Taking into account that M33 GMCs are a few times larger than the LMC GMCs, M33 Type C and D GMCs are possibly composed of smaller clouds. If that is the case, these small clouds in M33 Type C and D may not be coeval. This can explain the fact that the timescale of Type C GMCs in M33 are slightly larger than in LMC. As a conclusion, we propose a scenario as follows. First, dense gas formation occurs at a certain place in a GMC. After the molecular gas become dense enough to form stars (Type A), massive SF occurs in such dense regions, forming Hii regions (Type B). About 4--8 Myr later, the first formed YSGs are not embedded anymore and become visible (Type C). Subsequent dense molecular gas formation occurs due to the effect of previously formed YSGs and the next generation stars are born in such a dense gas (Type D). In this way, the SF propagates from initially generated stars, until the reservoir of molecular gas is exhausted. Such a continuous SF can last at least 10--30 Myr in a GMC with a typical mass of \(\ga10^5\,M_{\odot}\) during their lifetime of 20--40 Myr (after the dense gas formation). # Summary {#p3:sum} We present a GMC catalog using wide field (121 arcmin\(^2\) in total) and high sensitivity (1 \(\sigma=16\)--32 mK in \(T_{\rm mb}\) for a velocity resolution of 2.5 km s\(^{-1}\)) CO(\(J=3-2\)) maps of the nearby spiral galaxy M33 taken with the ASTE 10-m dish, and a complementary new catalog of YSGs for which we have estimated the ages. We summarize our main results as follows: 1. We identify 71 CO(\(J=3-2\)) GMCs from the CO(\(J=3-2\)) data. We discard from the analysis 6 GMCs which are at the edges of the observed field. The remaining 65 GMCs are characterized by the deconvolved sizes in the range of 12--157 pc in radius and virial masses of \(1.5\times10^4\)--\(8.9\times10^6\) \(M_{\odot}\), which are comparable to those of GMCs in our Galaxy. 2. We identify 75 YSGs from the excess of the surface density map of young stars that are associated with the identified GMCs. The total number of O stars in a YSG spans from 9--169, corresponding to total stellar masses of 10\(^{3.5}\)--10\(^{4.7}\,M_{\odot}\), assuming that they have a Salpeter IMF. The YSG ages are also estimated using the Padova model, and are found to be in the range of 4--31 Myr (12 Myr on average). These values are comparable to those of typical OB associations. 3. We compare the GMCs with the distribution of YSGs and classical Hii regions from a compilation in the literature. The identified 65 GMCs are successfully classified into four categories: Type A showing no sign of massive SF, Type B being associated only with relatively small Hii regions, Type C with both Hii regions and young (\(<\) 10 Myr) stellar groups, and Type D with both Hii regions and relatively old (10--30 Myr) stellar groups. Out of the 65 GMCs, 1 (1 %), 13 (20 %), 29 (45 %), and 22 (34 %) are found to be Types A, B, C, and D, respectively. 4. From a comparison of the distributions of the CO(\(J=3-2\))/CO(\(J=1-0\)) intensity ratio, \(R_{3-2/1-0}\), with the YSGs and Hii regions, we find that the \(R_{3-2/1-0}\) peaks within a GMC are slightly offset from the SF sites, but preferentially located around them. 5. We interpret the four types as representing an evolutionary sequence of the GMCs. Assuming that the number of the GMC types is proportional to the timescale of each evolutionary stage, we roughly estimate that they are \(\sim\)`<!-- -->`{=html}1, 3--7, 5--10, and 8--17 Myr, respectively for Types A, B, C, and D. This yields that the lifetime of a GMC with mass of \(\ga10^5\,M_{\odot}\) is 20--40 Myr, similar or slightly larger than the lifetime of LMC clouds. 6. Finally, we propose a scenario where after molecular gas becomes dense enough to form stars, massive SF occurs in such dense part. Then the molecular gas around the first generation stars becomes dense and forms the new generation of stars. In this way, the SF propagates from the initially generated massive stars and continues during the GMC lifetime. We thank the anonymous referee for a very useful report. We gratefully acknowledge the contributions of the ASTE staff to the development and operation of the telescope. The ASTE project is driven by Nobeyama Radio Observatory (NRO) and ALMA-J/Chile Observatory, branches of National Astronomical Observatory of Japan (NAOJ), in collaboration with University of Chile, and Japanese institutes including University of Tokyo, Nagoya University, Osaka Prefecture University, Ibaraki University, Hokkaido University, Joetsu University of Education, Keio University, and Kyoto University. Observations with ASTE were in part carried out remotely from Japan by using NTT's GEMnet2 and its partner R&E (Research and Education) networks, which are based on AccessNova in a collaboration between the University of Chile, NTT Laboratories, and NAOJ. *Facilities:* . [^1]: ftp://ftp.lowell.edu/pub/massey/lgsurvey/

{'timestamp': '2012-10-19T02:03:30', 'yymm': '1210', 'arxiv_id': '1210.3801', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3801'}

# Introduction This paper discusses progress in the development of a lightguide-based system of light collection in liquid argon, primarily for use in a LArTPC experiment such as LAr1  or LBNE . Light collection systems exploit the fact that charged particles traversing liquid argon produce copious amounts of ultraviolet scintillation light. The scintillation photon yield is tens of thousands per MeV of energy deposited , with depending on electric field, local ionization density and impurity concentrations at the parts-per-billion level. The light has a wavelength of 128 nm and is produced via two distinct scintillation pathways with different time constants: a prompt component with lifetime of \(\tau= 6\) ns and a slow component with \(\tau = 1500\) ns . An intermediate component with \(\tau = 40\) ns has also been reported by some groups . The relative normalization of early to late scintillation light depends upon the ionization density in the argon, and has been utilized as a particle identification variable in some dark matter searches . This paper benchmarks improvements to a recently reported lightguide detector design for light collection in LArTPCs . It also provides information that is useful for developing monte carlo simulations of lightguides, for development of future projects. In Sec. [\[sec:motivation\]](#sec:motivation){reference-type="ref" reference="sec:motivation"}, we begin with a brief discussion of the motivation for light collection. Then, Sec. [\[sec:formula\]](#sec:formula){reference-type="ref" reference="sec:formula"} provide the recipes for two coatings that have been studied. In Sec. [\[sec:bar\]](#sec:bar){reference-type="ref" reference="sec:bar"}, we describe characteristics of the acrylic bars used in this study. Next, in Sec. [\[sec:apparatus\]](#sec:apparatus){reference-type="ref" reference="sec:apparatus"}, we provide a description of the apparatus used to test the lightguide response, followed by a discussion of the waveform analysis (Sec. [\[sec:analysis\]](#sec:analysis){reference-type="ref" reference="sec:analysis"}). Sec. [\[sec:castacryl\]](#sec:castacryl){reference-type="ref" reference="sec:castacryl"} presents the study of the lightguide response, and lastly, Sec. [\[sec:conclusion\]](#sec:conclusion){reference-type="ref" reference="sec:conclusion"} summarizes our results. Throughout the discussion, we identify points where further R&D are likely to produce substantial advances. # Light collection in LArTPCs[\[sec:motivation\]]{#sec:motivation label="sec:motivation"} ## Liquid argon time projection chamber (LArTPC) In development of LArTPC detectors, most of the attention has focused on collection of charge to reconstruct tracks to very high precision. When charged particles traverse the detector, ionization electrons are liberated from the argon atoms by particles traversing a body of liquid argon. These electrons are drifted by an electric field and measured with crossed wire planes to form a 3D image of the charge deposits left along the tracks of neutrino interaction products. The coordinates of the charge deposits perpendicular to the wire planes is obtained from the drift time for the ionization electrons to reach the wires. To obtain the absolute drift time, and hence the perpendicular coordinate of the event in the detector volume, the time that the primary interaction occurred, \(t_0\), is used. ## Motivation Recently, attention has turned to establishing light collection systems. By collecting and measuring the argon scintillation light we can record the time structure of the event with few-nanosecond precision and determine the real \(t_0\) of the event. This provides a method for establishing \(t_0\) in cases where beam timing cannot be used. This is a much faster and more broadly applicable method than using the attenuation of charge of known MIP particles to establish the drift start-location. We must know \(t_0\) and hence the absolute drift distance for several important reasons. Firstly, liquid argon TPCs are commonly run in pulsed neutrino beams, with beam spills on the order of a microsecond. Surface based TPCs such as MicroBooNE and some LBNE options expect to be bombarded with a high rate of cosmic rays and secondary cosmogenic particles such as spallation neutrons which can mimic neutrino interactions. It is therefore vital to determine the interaction time with microsecond precision to veto the events which occur outside of the beam window. Once it has been determined that an interaction occurred within the beam window, the TPC can be triggered and read out. The recorded TPC image will in general contain several interactions: some corresponding to cosmogenic particles and one being the neutrino event we are seeking. Determining which interaction corresponds to an incident neutrino involves utilizing geometrical information provided by the optical systems in combination with information about the event topology from the TPC. During charge drift, diffusion and recombination of the ionization charge will occur. Hence there are drift distance dependent corrections which must be made to correctly measure the \(dE/dx\) of a track. Assuming neutrino events from the beam can be identified, \(t_0\) is known from the beam structure and the necessary corrections can be applied. However, for physics searches which involve events without a known time of arrival, such as proton decay and supernova neutrino searches, \(t_0\) must be measured by the optical system in order to apply the required corrections and make an accurate track energy measurement. Triggering on the information from the optical system has other practical benefits from a technical point of view. A typical TPC neutrino detector will have a tremendous amount of channels, and forming trigger logic on such a large set is a complicated procedure. In contrast, a PMT based optical system can achieve coverage of the volume with tens of elements, and forming a trigger becomes more straightforward. Finally, there are possible applications of the optical information for particle ID by pulse shape discrimination. The fast to slow scintillation yield ratio can reveal information about the local ionization of a track, which may be particularly helpful for performing particle ID on very short tracks where a TPC based dE/dx determination is either unreliable or impossible. ## Light Collection Using PMTs Cryogenic photomultiplier tubes (PMTs), which have a photocathode with platinum undercoating, can be used for light collection at LAr temperatures (87 K). However 128 nm scintillation light cannot penetrate any glass windows. Also, typical bi-alkali PMTs are only sensitive to visible light, not the 128 nm scintillation light from LAr. Therefore, in a PMT-based system, the light must be shifted to longer wavelengths. The favored solution in LAr detectors for shifting the 128 nm light has been to use a tetraphenyl-butadiene (TPB) layer between the detector and the PMTs. This fluorescent wavelength-shifter absorbs in the UV and emits in the visible with a peak at \(425\pm20\) nm , which is a favorable wavelength for detection by bi-alkali PMTs. Many detectors have used PMTs directly coated with TPB on windows, applied as either an evaporative coating (ICARUS ) or embedded within a polystyrene (PS) matrix ( WArP ). The MicroBooNE design separates the coating from the PMT by applying a TPB-PS mixture to an acrylic plate positioned directly in front of the PMT. Generally the light collection systems of the large active volume detectors have favored the use of large PMTs, such as the Hamamatsu 5912-02mod 8-inch PMT used in MicroBooNE , which are sparsely distributed for economic reasons. These tubes are located in the field-free region of the detectors, typically behind of the TPC anode wire planes. ## Lightguides as an Alternative System In Ref. , we presented the first detection of scintillation light in liquid argon using a lightguide system. That paper discussed how coated acrylic bars, arranged side-by-side as a paddle, and bent to guide light adiabatically to a single 2 inch cryogenic PMT, could provide a flat-profile light detection system that could potentially be inserted into dead regions between LArTPC wire planes. This potentially could provide an economical light collection system to collect light, if the design is sufficiently efficient. The lightguides utilized a TPB-based coating with an index of refraction that was chosen to match acrylic bars. Acrylic was chosen as the substrate because it is resilient to cryogenic cycling and can be easily bent to the required form. Some of the visible light that is emitted when UV photons hit the TPB coating will undergo total internal reflection because the acrylic has an index of refraction for blue light (\(n=1.49\)) that is higher than that of liquid argon (\(n=1.23\)) . The reported guides used a polystyrene-TPB coating, as suggested in Ref. , mixed in a 3:1 mass ratio. The mixture was dissolved in toluene for application as a liquid. This was the highest mass ratio that could be achieved without the TPB crystallizing on the surface of the guide. Crystallization must be avoided in lightguide coatings because the white crystals cause absorption and scattering of visible light as it is reflected along the bars, reducing the attenuation length. With this design, we were able to demonstrate light collection, albeit with fairly low efficiency. # Lightguide Coatings[\[sec:formula\]]{#sec:formula label="sec:formula"} We use two coatings in this study. The first was the same coating as used the study of Ref. , called PS25%. The second coating is a new recipe using UVT acrylic, called UVT33%. We will show below that the responses of the two coatings are nearly equivalent, and so either can be used in future detector designs. ## PS25% Coating This coating consists of: - 1:3 TPB:polystyrene ratio by mass - 50 ml of toluene for every 1 g of polystyrene This is the coating used for our first lightguides. ## UVT33% Coating This coating consists of: - 1:2 TPB:acrylic ratio by mass - 50 ml of toluene for every 1 g of acrylic - 1:5 ethyl alcohol: toluene ratio in volume This coating uses UVT acrylic and produces a clear, high quality coating for use on waveguides. The TPB and UVT pellets are first dissolved in the toluene, then the ethanol is added. The coating is applied to the acrylic guide in a single brush stroke. From the amount of solution used to coat, we estimate approximately \(5.5\times 10^{-5}\) g/cm\(^2\) of TPB is deposited on the surface. However this cannot be related to the performance of the coated plate easily, because other factors, including coating method and surface condition, are equally important. Addition of ethyl alcohol improves attenuation length because acrylic dissolved directly in toluene produces a rougher coating, more likely to scatter the light in the guides. Ethyl alcohol thins the coating as well as smoothing it. However, addition of ethyl alcohol may introduce negative effects, such as self-absorption. As we show later (Figure [\[fig:SpectrometerAttachments\]](#fig:SpectrometerAttachments){reference-type="ref" reference="fig:SpectrometerAttachments"}), the guided light at the TPB emission peak (436 nm) shows fairly uniform response against the incident spectrum. However, a few drops of ethyl alcohol introduces slight overall reduction (\(<\)`<!-- -->`{=html}10%) and a small dip at 270 nm. We have varied the fraction of alcohol to find the optimum ratio to Toluene. ## Tests of Coatings in a Vacuum Monochrometer Figure [\[vacuumspec\]](#vacuumspec){reference-type="ref" reference="vacuumspec"} shows the setup of the vacuum monochrometer test. Tests were performed in a McPherson 234 vacuum monochrometer using a McPherson model 632 UV Deuterium Lamp. Measurements were taken at a pressure of 11 \(\pm\) 4 mTorr at room temperature, though results at 215 nm and 250 nm are consistent with measurements at atmospheric pressure. A PMT located outside of the vacuum region is used to see the forward emission of the sample plate. PMT 1 in figure [\[vacuumspec\]](#vacuumspec){reference-type="ref" reference="vacuumspec"} was used for this measurement. The data in Figure [\[avemono\]](#avemono){reference-type="ref" reference="avemono"} represent average measurements and are normalized with respect to average measurements of evaporatively coated plates. The evaporative samples contain pure TPB evaporated onto a plate in vacuum, and thus the effects of the different matrices can be compared to the pure TPB response. The evaporative samples used in this study have coating thicknesses of 1.87 \(\mu\)m. We expect these to be somewhat similar in performance to the plates used in reference , which had a thickness of 1.5 \(\mu\)m. The errors on each plate are associated with statistical errors including testing different samples of each coating type and systematic errors associated with the vacuum monochrometer. The overall efficiency including the error on the evaporative measurement at 128 nm of our two types of coatings relative to the previously described evaporatively coated plates are 0.25 \(\pm\) 0.03 and 0.13 \(\pm\) 0.02 for the PS25% and UVT33% coated acrylic respectively. These results on transmission of coated plates at room temperature are reproducible, as is reflected by the small errors. However, they contradict the results we will present in Sec. [\[sec:castacryl\]](#sec:castacryl){reference-type="ref" reference="sec:castacryl"}, where we will show very little difference in response between the PS25% and UVT33% coated lightguide performance in LAr. Indeed, in that section, we will show that UVT33% is slightly favored over PS25%. From this, we must conclude that either temperature plays an unexpected role in the relative emission between the two coatings, or that there is a difference in the capture efficiency between the coatings. The indices of refraction are 1.59 and 1.49 for the polystyrene and acrylic matrices, respectively, while the index of refraction of the cast bar is 1.49. Capture inefficiency could be exacerbated by the fact that the polystyrene has a smaller coefficient of expansion than the cast acrylic bar, while the acrylic coating and the bar are an exact match. So we suspect that the PS25% performance is degraded in the LAr, becoming very similar to the UVT33%. However, this is still under investigation. ## Emission Spectra From Light Guides The guided emission spectra from several short light guide segments of up to 10 cm in length were measured using a fluorimeter. A tunable monochromatic beam, produced using a grating and a xenon lamp, is normally incident upon the TPB coated surface of the light guide. The guided spectrum is measured at 90 degrees to the incident beam through of the end of the light guide using a second tunable grating and a photomultiplier tube. The spectrometer attachment used to implement this configuration is shown in figure [\[fig:SpectrometerAttachments\]](#fig:SpectrometerAttachments){reference-type="ref" reference="fig:SpectrometerAttachments"} (left). The lamp spectrum and PMT response are automatically accounted for by the spectrometer software, and the device is routinely re-calibrated using two standard samples: a rhodamine dye sample to characterize the emission grating and PMT response, and a diffuse glass cuvette to characterize the excitation grating and xenon lamp spectrum. The guided spectrum was measured between 200 and 600 nm for incident wavelengths between 200 and 700nm. For segments of length 6 cm, 8cm and 10 cm, no differences in the shape of the emission or excitation spectra were seen. For incident light above 400 nm, no wavelength shifting behavior is observed, so we omit this region from the reported plots. Figure [\[fig:GuidedEmExSpectra\]](#fig:GuidedEmExSpectra){reference-type="ref" reference="fig:GuidedEmExSpectra"}, top shows the two dimensional emission-excitation spectrum as a contour plot. We also show the emission spectrum at 250nm (Fig. [\[fig:GuidedEmExSpectra\]](#fig:GuidedEmExSpectra){reference-type="ref" reference="fig:GuidedEmExSpectra"}, bottom left), and the wavelength shifting capability at the TPB emission peak wavelength of 436 nm (Fig. [\[fig:GuidedEmExSpectra\]](#fig:GuidedEmExSpectra){reference-type="ref" reference="fig:GuidedEmExSpectra"}, bottom right). These one dimensional plots can be interpreted as a single horizontal and vertical slice from the two dimensional contour map, respectively. The two dimensional excitation-emission spectrum for a backward emission was also measured. In this setup, a short light guide section is illuminated at 45 degrees to the surface with a monochromatic beam. The emitted light at 90 degrees to the incident beam is detected. The spectrometer attachment used to implement this arrangement is shown in figure [\[fig:SpectrometerAttachments\]](#fig:SpectrometerAttachments){reference-type="ref" reference="fig:SpectrometerAttachments"} (right). The two dimensional backward excitation-emission spectrum for the light guide coating is shown in figure [\[fig:SurfaceEmExSpectrum\]](#fig:SurfaceEmExSpectrum){reference-type="ref" reference="fig:SurfaceEmExSpectrum"}. # Bars Under Study[\[sec:bar\]]{#sec:bar label="sec:bar"} In the discussion below we consider coatings applied to extruded and cast acrylic bars to form light guides. Cast acrylic is taken to be the standard in our study. Our cast bars are purchased from Altec Plastics and are polished on the ends. This replaces the extruded acrylic bars used in Ref.  purchased from McMaster-Carr that are substantially more economical, however, have significant imperfections. ## Attenuation Length Measurements of Uncoated Acrylic Bars [\[attennocoat\]]{#attennocoat label="attennocoat"} We measured the attenuation length of uncoated acrylic bars by illuminating the end of rods cut to various lengths and recorded the photo-current observed by a Si photodiode (PD). We used acrylic manufactured in two different ways: a cast UVA acrylic and an extruded UVA acrylic. The setup can be seen in figure [\[intsphere_setup\]](#intsphere_setup){reference-type="ref" reference="intsphere_setup"}. The detector used to measure the light emerging from the end of the bars was an OSI Optoelectronics 3 mm Si PD (P/N: OSD15-0). In this configuration we are overfilling the detector and therefore only sampling a small fraction of the end of the bar. In addition, a \(420\pm10\) nm narrow band filter was placed in front of the detector to restrict the attenuation length measurement to 420 nm. The illumination was performed with a high power \(420\pm10\) nm LED (FutureLED P/N: FL-LED-440-420) driven at a constant current of 187 mA. The LED light was input into an integrating sphere to diffuse the light, thus illuminating the rod end with a uniform illumination pattern. The PD was mounted so that it views only the center of the bar through the narrow band filter. If the LED light is passed directly down the bar, the illumination pattern at the output of the bar is nonuniform and slight misalignments between the LED, light guide and PD can lead to systematics in the measurement of the attenuation length. Uniform illumination from an integrating sphere, on the other hand, will reduce this systematic significantly. This systematic would be minimized in a LAr detector design that under fills the PMT photo-cathode area. Two sets of ten bars, one set for each type of acrylic, were measured with lengths increasing by 7.62 cm (3 inches) between 7.62 cm and 76.2 cm (30 inches). The \(0.64 \times 2.54\) cm\(^2\) (1/4 inch \(\times\) 1 inch) cast UVA acrylic bars, purchased from Altec Plastics, were cut out of large acrylic 0.64 cm thick sheets. The cut edges on the 0.64 \(\times\) 76.2 cm\(^2\) face were polished by Altec Plastics using a polishing machine. The quality of measurement of attenuation length depends strongly on how well the sides are polished. The \(0.48 \times 2.54\) cm\(^2\) (3/16 inch \(\times\) 1 inch) extruded UVA acrylic bar, purchased from McMaster-Carr, didn't require polishing on the 0.48 \(\times\) 76.2 cm\(^2\) face, because the bars were extruded at the 2.54 cm width and the factory edge was suitable for making this measurement. In both cases the ends of the bars were polished in the lab using a diamond tipped fly cutter after they were cut to length. Fig. [\[atteniu\]](#atteniu){reference-type="ref" reference="atteniu"} shows natural log of the normalized photo-current versus the bar length for both the extruded and cast UVA acrylic. A fit to these data resulted in a measured attenuation length of 50\(\pm\)`<!-- -->`{=html}2 cm for the extruded UVA acrylic and 38\(\pm\)`<!-- -->`{=html}1 cm for cast UVA acrylic. ## Preparation and Storage of Coated Bars All bars used in the tests below are 60 cm in length and have polished ends. The cast bars are delivered with polished ends, while the ends of the extruded bars are polished in the lab. Before coating, bars are cleaned with ethyl alcohol. The coatings are applied with one brushstroke using an acid brush, depositing about \(5.5\times 10^{-5}\) g/cm\(^2\) of TPB There is clear evidence that TPB coatings degrade with even modest exposure to laboratory fluorescent lights and sunlight . Our studies show a 30% loss of response after a single day of exposure and 80% degradation after one month. Therefore the bars are handled in a laboratory with UV filters installed on the fluorescent lights and windows. Furthermore, whenever possible, bars are kept covered with light-blocking cloths or are stored in opaque containers. The previous study of lightguides predated the demonstration of the detrimental effect of UV light. Therefore, the precautions described above were not taken with the lightguides discussed in Ref. . We think that this is the primary explanation for why the outputs of the PS25% lightguides presented in this paper are nearly a factor of three higher than that reported in Ref. . There is some evidence that TPB coatings degrade by about 10% due to exposure to humidity in the laboratory . This can be mitigated by desiccating the bars. The effect of humidity is relatively small and the evidence for improvement modest at best. Nevertheless, we store the bars used in these studies in containers with desiccant packets. # Apparatus for LAr Tests[\[sec:apparatus\]]{#sec:apparatus label="sec:apparatus"} The lightguide test stand was described in detail in Ref. . The test stand is constructed from an open-top glass dewar which is 100 cm tall and 14 cm inner diameter into which a holder containing the lightguide is inserted. The holder has a 7725-mod Hamamatsu, 10-stage PMT  with a custom cryogenic base attached at the bottom. One PMT was used for all tests reported here. In the LAr, the PMT floats up against the lightguide which is fixed in the holder, making a good optical connection. An improvement to the holder of Ref.  guides the PMT so that the relative position between PMT face and bar is reproducible. Lightguides are tested with 5.3 MeV \(\alpha\) particles produced from a \(^{210}\)Po source mounted in a plastic disk . The source is electroplated onto foil that is recessed into a 3mm "well," of the plastic disk. In an improvement over Ref. , the disk is held in a holder with a 5.1 mm diameter hole, leading to a well of 4.8 mm depth. The \(\alpha\)s emitted into the well traverse \(\sim\)`<!-- -->`{=html}50 \(\mu\)m in LAr. Scintillation light is then isotropically emitted. The well occludes most of the light; the solid angle acceptance for light at the bar is 7%. The holder is immersed in LAr for 30 minutes before data is taken. Readout of the PMT is performed using an Alazar Tech ATS9870 digitizer. The Waveform is recorded on a scale of 200 mV/128 ADC counts and a trigger is produced by a negative pulse with an amplitude that exceeds 17 ADC counts, corresponding to a \(-27\) mV threshold. When a trigger is produced, 128 pre-trigger samples and 384 post-trigger samples are recorded at a sampling rate of 1 giga-sample per second, leading to a total recorded profile of width 0.512 \(\mu\)s. Triggers from the \(\alpha\) source occurred at a rate of about 300 to 400 Hz, which was consistent with expectation given the short lifetime of the polonium. Runs were taken with no \(\alpha\) source in order to measure the cosmic ray rate, which was found to be about 8 Hz. The dark rate, measured in runs with no light guide with the above threshold, was \(<\)`<!-- -->`{=html}1 Hz. Our studies use four batches of industrial grade LAr. Industrial grade LAr is CGA certified to contain \(<\)`<!-- -->`{=html}20 ppm nitrogen and \(<5\) ppm oxygen . Running with "ultra-pure grade" which has has \(<\)`<!-- -->`{=html}5 ppm nitrogen and \(<1\) ppm oxygen , was also an option, however we found no difference in the results of the studies below with the industrial versus pure LAr. This surprising result may arise for two reasons. First, the process of filling the open dewar system may lead to roughly equal contamination levels of the ultra-pure argon and the industrial argon. Second, the studies reported here focus on the quantity of early light, but not the quantity of late light. Impurities most dramatically affect the quantity of late light . With this said, further investigation with purified LAr is warranted. In order to limit contamination by air when filling, the test stand is first filled with argon gas. The LAr is then poured through the gas into the dewar. We create the initial gas layer in the dewar by half-filling the warm dewar and allowing this LAr to evaporate. # Analysis of the Waveforms[\[sec:analysis\]]{#sec:analysis label="sec:analysis"} Fig. [\[pulses\]](#pulses){reference-type="ref" reference="pulses"} shows the waveforms of some example events from runs with the \(\alpha\) source. One expects to see an initial peak corresponding to many photons followed by pulses from late light that correspond to a single UV photon hitting the bar distributed in time. The lifetime of the early light is 6 ns, and so one expects 95% of the photons to be produced within the first 18 ns of the pulse. In fact, the pulses appear somewhat wider due to fact that some late light also populates the initial peak, although this component is highly quenched due to the impurities in the LAr . As a result, the initial pulse is predominantly early light and we will refer to the initial pulse as early light in the discussion below. ## The Pulse-finding Algorithm and Variables We apply a pulse-finding algorithm that identifies pulses and records the maximum pulse-height from the baseline and the integrated charge of each pulse. The trigger is defined as when the signal drops to 17 or more ADC counts below the baseline. The pulse is then integrated from 30 ns before the trigger to 120 ns after the trigger. This defines total charge, \(Q_{tot}\), which is our primary observable. The long integration period accommodates large initial pulses from cosmic rays as well as the smaller initial pulses from the \(\alpha\) source. The pulse-height, \(Q_{max}\), is defined as the difference between the baseline and the minimum ADC count during the integration period. The baseline is recalculated using a 10 ns time window before each pulse, for both early and late light pulses. The start point for late light pulses is defined as when the signal drops to 2 or more ADC counts below the baseline. ## The \(Q_{tot}\) Distributions of Early Light After corrections for attenuation along the lightguide, the \(Q_{tot}\) distribution of the early pulses can be correlated to the energy of the \(\alpha\) in the event. Examples of \(Q_{tot}\) distributions for the early light are shown in Fig. [\[Qtot\]](#Qtot){reference-type="ref" reference="Qtot"}. One can see that the distributions form a Landau distribution. In the analysis below, we will study variations in these distributions as a function of position and batch of LAr. Thus, for illustrative purposes, we provide examples of distributions with the source at 10 cm (black), 20 cm (magenta), 30 cm (green), 40 cm (blue) and 50 cm (red) for various lightguides and LAr batches. In the analyses below, the \(Q_{tot}\) distributions are taken from runs of \(>70,000\) events. Using \(Q_{tot}\) is a departure from the study in Ref. , which employed pulse-heights, \(Q_{max}\). However, with the improved efficiency for the lightguides, we find that \(Q_{tot}\) is a better representation of the number of photons in an event. The issue arises when the photon arrivals are distributed over long time periods, leading to multiple peaks. An example of such a distortion is seen in the bottom left of Fig [\[pulses\]](#pulses){reference-type="ref" reference="pulses"}. This problem was addressed by the "multipeak analysis" of Ref. , however employing \(Q_{tot}\) is a simpler and more accurate solution. ## Late Light: A Single Photon Sample As can be seen from Fig. [\[pulses\]](#pulses){reference-type="ref" reference="pulses"}, the late light is sparse, however the pulses are well-formed and uniform. The late light is particularly valuable, because it allows measurement of the \(Q_{tot}\) distribution for one and only one photon arriving at the PMT. Therefore, we utilize the late light to calibrate our system. Our late light sample is acquired in the range \(> 400\) ns after the initial pulse producing the trigger. This is sufficiently late in time that considering the yield and time constant of the light, we can be assured that the pulses which arrive at the lightguide correspond to only one UV photon. It has been shown that, on average, 1.3 visible photons are emitted from an evaporative TPB coating per one incident UV photon; which is to say that occasionally TPB will produce multiple photons rather than one However, the acceptance of the lightguides, is only 5%, so there is a negligible probability of multiple photons arriving at the PMT. Most methods of calibration, such as using low intensity LEDs, involve a Poisson distribution of photons arriving at the PMT which is then used to find the 1 PE response. In contrast, this calibration method is assured to sample exactly one photon hitting the PMT. As a result, one expects the \(Q_{tot}\) distribution of the late light will simply reflect the statistics of the early stages of the dynode chain, which is expected to be represented by a Gaussian to a good approximation. As expected, the peak position varies with PMT voltage. However, for a given PMT, set at a specific voltage, the peak position is always located at 40 counts\(\times\)ns, regardless of the lightguide being tested. Also, we find that, for a specific lightguide-and-source set up, if the PMT high voltage is always adjusted such that the 1 PE response peak is at 40 counts\(\times\)ns, then the prompt light response is reproducible. We demonstrate this in Fig. [\[overlaylate\]](#overlaylate){reference-type="ref" reference="overlaylate"}, where we overlay the late light distributions from five example runs. The examples cover four bars and four LAr batches and are typical of all the late light data sets. To allow comparison, the distributions are normalized in the range \(Q_{tot}>50\) counts\(\times\)ns, where background is expected to be low. One sees that the peaks of the late light are in good agreement, but the \(Q_{tot}<30\) counts\(\times\)ns can vary due to the relative strength of the background under the different environment. However this does not affect the location of the 1 PE peak, as Figure [\[overlaylate\]](#overlaylate){reference-type="ref" reference="overlaylate"} show. The source of background is under investigation. The black squares show the mean of ten late light distributions. Using the expectation for PMTs with large (\(>4\)) secondary electron emission at the first dynode and high collection efficiency by the first few dynodes, we find a Gaussian fit to \(Q_{tot}>30\) counts\(\times\)ns yields a single PE peak at 39.3\(\pm\)`<!-- -->`{=html}1.0 counts\(\times\)ns. Thus, in the analysis that follows, the conversion from \(Q_{tot}\) to observed PE is obtained by dividing the results by 39.3 counts\(\times\)ns/PE. This was the same technique as was employed in Ref. . # Cast Acrylic Bars with Acrylic Coatings[\[sec:castacryl\]]{#sec:castacryl label="sec:castacryl"} The following tests are performed on cast acrylic light guides with UV33% coatings brushed onto cast acrylic bars. ## Results of Tests in LAr [\[results\]]{#results label="results"} The studies presented here involve four batches of LAr and four bars, and were performed over about 60 days in order to study a variety of sources of variation of response. Fig. [\[fourbars\]](#fourbars){reference-type="ref" reference="fourbars"} presents measurements for each of the four light guides under study, with a different symbol associated with each bar. The colors indicate the LAr batch in the test stand for a given measurement, where the time-order was: blue, red, magenta and green. Measurements are taken at up to five locations along the bar (10, 20, 30, 40 and 50 cm), and often multiple times. The source is removed and replaced between every measurement, even in the case where a measurement is taken multiple times at the same location. The exact location is varied within \(\sim 1\) cm. The purpose of this is to sample multiple areas of coating at each distance. Within a given location indicated on the plots, the measurements are presented in time order. Every measurement has more than 70,000 entries, and so the error bars are not visible on the plots. Fig. [\[fourbarsoneplot\]](#fourbarsoneplot){reference-type="ref" reference="fourbarsoneplot"} transfers information from Fig. [\[fourbars\]](#fourbars){reference-type="ref" reference="fourbars"}, maintaining the meanings of the symbols for each bar and colors for each batch. In this figure, time information has been removed and the time integrated ADC counts, \(Q_{tot}\), are plotted as a function of location of the source. The mean at each location is indicated by the black hexagon with the error bar. The deviation of measurements from the mean is such that a 16% systematic spread encompasses 68% of the data points. The following conclusions can be drawn from Fig. [\[fourbars\]](#fourbars){reference-type="ref" reference="fourbars"} and Fig. [\[fourbarsoneplot\]](#fourbarsoneplot){reference-type="ref" reference="fourbarsoneplot"}: - At a given source location, the spread of measured values is very large compared to the statistical error. - The spread in measurements at each location is, to a good approximation, the same fraction (16%) of the measured value. - The spread is not due to LAr batch variations. The results from multiple batches are quite consistent within the spread. - There is no evidence of systematic degradation with time due to an external source such as UV light. The spread in measurements seems likely to be due to variations in the quality of the coating within the region sampled at each location. In principle, one expects an exponential attenuation along the bar. Fig. [\[means\]](#means){reference-type="ref" reference="means"} presents the mean of the measurements in LAr (black points) as a function of location along the bar on a semi-log plot. The point at 10 cm lies significantly higher than the expectation for an exponential. The green curve, which includes the 10 cm point in the fit, results in an attenuation length of 44 cm, while the blue curve, which excludes the 10 cm point, indicates an attenuation length of 79 cm. These results are in qualitative agreement with the warm, uncoated acrylic bar measurements presented in Sec. [\[attennocoat\]](#attennocoat){reference-type="ref" reference="attennocoat"}), where the overall fit gave 38\(\pm\)`<!-- -->`{=html}1 cm attenuation length. However, the warm, uncoated bars gave better agreement with a straight line on a semi-log plot. It is possible that the deviations observed in the LAr measurements come from the imperfections in coating. The source will be pursued in the future. To establish the performance of the coating, apart from the attenuation, we use the 10 cm point as a benchmark. This is likely to be a conservative estimate as some attenuation may be occurring over the 10 cm. The \(Q_{tot}\)-to-PE conversion yields 21.9 PE (7.2 PE) at 10 cm (50 cm). This result can be compared with the previously reported light guides that reported \(\sim 7\) PE at 10 cm. ## Comparison of Acrylic vs Polystyrene Coatings. The cast bars with UVT33% coating are compared to cast bars and extruded bars with PS25% coating in Fig. [\[psbars\]](#psbars){reference-type="ref" reference="psbars"}. The acrylic coating is consistently better than the polystyrene coating by a factor of about 1.3. # Conclusions[\[sec:conclusion\]]{#sec:conclusion label="sec:conclusion"} This paper has presented information useful to simulating lightguide paddles in future LArTPC experiments. We have presented a comparison of response of lightguides prepared with two coatings, UVT33% and PS25%. We find that the response levels are similar for lightguides in LAr, although the transmission tests at room temperature indicated that PS25% was the better coating. At 10 cm, typical response of the light guides is about 21 PE, which is a factor of three improvement over previously reported lightguides. The authors thank the National Science Foundation (NSF-PHY-084784) and Department Of Energy (DE-FG02-91ER40661). We thank Dr. A. Pla of Fermi National Accelerator Laboratory for helpful discussions and use of equipment.

{'timestamp': '2012-10-16T02:03:17', 'yymm': '1210', 'arxiv_id': '1210.3793', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3793'}

# Introduction {#introduction .unnumbered} Groups in which all elements belong to the conjugacy class of their inverses are called *real groups*. It is well known that all entries in the character tables of finite real groups are real. However real groups may admit complex representations which are not realizable over \(\R\). Such representations are called *symplectic*. The complex representation \[1 \mapsto \left(\begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array}\right), ~~a \mapsto \left(\begin{array}{cc} i & 0 \\ 0 &-i \end{array}\right), ~~b \mapsto \left(\begin{array}{cc} 0 & 1 \\-1 & 0 \end{array}\right), ~~c \mapsto \left(\begin{array}{cc} 0 & 1 \\ i & 0 \end{array}\right)\] of the Quaternion group \(Q_2 = \langle-1, a, b, c: (-1)^2 = 1, a^2 = b^2 = c^2 = abc =-1 \rangle\) is symplectic. Real groups which do not admit any symplectic representation are called *totally orthogonal*. In 1985, Gow proved that groups \(\operatorname{O}_n(q)\) of special isometries of quadratic forms are totally orthogonal. It was already known by then that every element of \(\operatorname{O}_n(q)\) is strongly real. *Strongly real* elements in a group are those which can be expressed as a product of two self-inverses. A group is called *strongly real* if all its elements are strongly real. It is straightforward to see that a group \(G\) is strongly real if for every element \(x\) of \(G\), there exists an element \(y\) in \(G\) such that \(y^2 = 1\) and \(x^{-1}=y^{-1}xy\).\ Like \(\operatorname{O}_n(q)\) there are plenty of groups which are both totally orthogonal and strongly real. According to a conjecture of Tiep finite simple groups are totally orthogonal if and only if they are strongly real. However there is no general class of groups known which exhibits one of these properties and not the other. In this article we exhibit an infinite class of groups, namely the class of special \(2\)-groups, for which none of the notions of strongly reality and total orthogonality imply the other. This generalizes an example of a strongly real group admitting a symplectic representation.\ The plan of the article is as follows: In §1 we recall basics on quadratic forms over \(\F_2\), where \(\F_2\) denotes the field with two elements. Then in §2 we record special \(2\)-groups, quadratic maps, second cohomology groups and their interconnection. The understanding of this interconnection gives us a characterization of strongly real special groups in terms of their associated quadratic maps. This characterization is used in next section to show that all extraspecial \(2\)-groups, except the quaternion group \(Q_2\), are strongly real. This requires a classification of extraspecial \(2\)-groups as central products of \(D_4\) and \(Q_2\), where \(D_4\) denotes the dihedral group of order \(8\). In the same section we convert this classification in the language of quadratic forms over \(\F_2\). In last section we give examples of groups which are strongly real but not totally orthogonal, and vice-versa. A characterization of strongly real groups (Th. [\[strongly-real-criterion\]](#strongly-real-criterion){reference-type="ref" reference="strongly-real-criterion"}) and a characterization of totally orthogonal groups play key roles in the construction of these examples. # Quadratic forms in characteristic \(2\) Let \(\F\) be a field of characteristic two and \(V\) be an \(n\) dimensional vector space over \(\F\). A map \(b: V\times V\rightarrow \F\) is called *\(\F\)-bilinear* if it satisfies the following properties: 1. \(b(\alpha v_1+\beta v_2, w)=\alpha b(v_1,w)+\beta b(v_2,w)\) for all \(v_1, v_2, w \in V\) and \(\alpha, \beta \in \F\). 2. \(b(v, \alpha w_1+\beta w_2)=\alpha b(v,w_1)+\beta b(v,w_2)\) for all \(v, w_1, w_2 \in V\) and \(\alpha, \beta \in \F\) A map \(q:V\rightarrow \F\) is called *quadratic form* if 1. \(q(\alpha v)=\alpha^2 q(v)\) for all \(v \in V\) 2. The map \(b_q: V\times V\rightarrow \F\) given by \(b_q(v,w):=q(v+w)-q(v)-q(w)\) is \(\F\)-bilinear. For a quadratic form \(q\) the map \(b_q\) is called the *polar map* of \(q\). The pair \((V,q)\) is called the *quadratic space* over \(\F\). A bilinear map \(b_q\) is called *alternative* if \(b_q(v,v)=0\) for all \(v \in V\).\ If \(B = \{e_1,e_2,\cdots,e_n\}\) is a basis of \(V\) then any \(n \times n\) matrix \(Q\) satisfying \(q(x)=x^t Qx\), where \(x\in V\) is indeterminate column vector and \(x^t\) denotes the transpose of \(x\), is called a *matrix of \(q\) with respect to basis \(B\)*. Every matrix of \(q\) with respect to same basis is of form \(Q + A\), where \(A\) is an alternating matrix and \(Q\) is the unique upper triangular matrix of \(q\) with respect to basis \(B\). If we change the basis and \(T\) is transition matrix for the change of basis then upper triangular matrix of \(q\) with respect to new basis is \(T^t QT\).\ Two \(n\)-dimensional quadratic spaces \((V_1, q_1)\) and \((V_2, q_2)\) over \(\F\) are said to be *isometric* if there exists an \(\F\)-linear isomorphism \(T:V_1\rightarrow V_2\) such that \(q_1(v)=q_2(T(v))\) for all \(v\in V\). Isometry between two quadratic spaces is denoted by \((V_1, q_1)\cong(V_2, q_2).\) A quadratic space \((V, q)\) is called the *orthogonal sum* of \((V_1, q_1)\) and \((V_2, q_2)\) if \(V = V_1\oplus V_2\) and \(q(v)=q_1(v_1)+q_2(v_2)\), where \(v = (v_1, v_2) \in V\) is an arbitrary element of \(V\). In this case we write \(q=q_1 \bot q_2\). Conversely, let \((V,q)\) be a quadratic space and \(V_i,~1 \leq i\leq m\) be subspaces of \(V\) such that \(V = V_1\oplus \cdots \oplus V_m\) and \(b_q(v_i,v_j) = 0\) for \(v_i\in V_i, v_j\in V_j, i\neq j\). Then \(q = q_1 \bot \cdots \bot q_m\) where \(q_i\) denotes the restriction of \(q\) to \(V_i\).\ The subspace \(\operatorname{rad}(V) := \{w\in V: b_q(v,w)=0 ~\forall v\in V\}\) is called the *radical* of \((V,q)\). The quadratic space \((V,q)\) is called *regular* if \(\operatorname{rad}(V)=0\).\ The following theorem is analogous to the diagonalization of quadratic spaces over the field of characteristic different from \(2\).\ More explicitly, the orthogonal decomposition \(V = U \oplus \operatorname{rad}(V)\) is such that there exists a basis \(\{e_i, f_i, g_j, 1\leq i\leq r, 1\leq j\leq s\}\) of \(V\) where \(2r+s=\dim(V)\) and elements \(a_i, b_i, c_j \in \F,1\leq i\leq r, 1\leq j\leq s\) such that for all \(\displaystyle v=\sum_{i = 1}^r (x_i e_i + y_i f_i)+\sum_{i = 1}^s z_j g_j,\) we have \[q(v)=\sum_{i = 1}^r (a_i x_{i}^{2}+x_i y_j+ b_i y_{i}^{2})+\sum_{i = 1}^s c_j z_{j}^{2}.\] In this case we say that \([a_1,b_1]\bot \cdots \bot [a_r, b_r]\bot \langle c_1,\cdots,c_s \rangle\) is the *normalized form* of \(q\) and \([a_1,b_1]\bot \cdots \bot [a_r, b_r]\) is the *regular part* of \(q\). A quadratic form \(q\) is said to be *non-singular* if \(s=0\). It is called *singular* if \(s\neq0\). In addition, if \(r=0\) then \(q\) is called *totally singular*. If \(s>0\) then regular part of quadratic form is generally not determined uniquely up to isometry, whereas the part \(\langle c_1,\dots,c_2\rangle\) is always determined uniquely up to isometry. For example \([1,1]\oplus \langle 1 \rangle \cong [0,0]\oplus \langle 1 \rangle\) but \([1,1]\cong [0,0]\) holds if and only if the quadratic equation \(x^2+x+1=0\) has a solution in \(\F\).\ It immediately follows from above Proposition that every regular quadratic form over a field of characteristic two is even dimensional.\ A quadratic from is said to be *isotropic* if there exists \(0\neq v\in V\) such that \(q(v)=0\), otherwise it is called *anisotropic*. Quadratic form \([0,0]\) is the only \(2\)-dimensional isotropic quadratic form up to isometry. It is called *hyperbolic plane* and is denoted by \(H\). A quadratic space is said to be a *hyperbolic space* if it is orthogonal sum of hyperbolic planes.\ The following result is an analogue in characteristic 2 to the usual Witt decomposition in characteristic not equal to 2. The anisotropic part \(q_r ~\bot~ q_s\) of \(q\) will be denoted by \(q_{an}\) in short. Two quadratic forms \(q_1\) and \(q_2\) are called *Witt equivalent* (denoted \(q_1 \sim q_2\)) if \({q_1}_{an}\cong {q_2}_{an}\). If \(q_1\) and \(q_2\) are non-singular then \(q_1 \sim q_2\) if \(q_1 ~\bot~-q_2\) is hyperbolic. The set \(W_q(\F)\) of Witt equivalence classes of regular quadratic forms over \(\F\) forms an abelian group under the operation of orthogonal sum of quadratic spaces. It is called the *Witt group of quadratic forms*.\ As all regular quadratic forms are of even dimension, the *dimension invariant* \(e_0: W(\F) \to \frac{\mbox{\)\mathbb Z\(}}{2\mbox{\)\mathbb Z\(}}\) given by \(q \mapsto \dim(q)\mod 2\) is trivial. In \(\operatorname{char}(\F) = 2\) case an invariant at next level is Arf invariant which is a substitute of Discriminant in \(\operatorname{char}(\F) \neq 2\) case. It was defined by Arf in his classical paper. Let \(q: V \rightarrow \F\) be a regular \(2n\)-dimensional quadratic form. As \(b_q\) is an alternating form, the space \((V, b_q)\) has a symplectic basis \(\{e_i, f_i, 1\leq i\geq n\}\). Let \(\wp(\F)=\{x^2+x: ~x\in \F\}\). The set \(\wp(\F)\}\) is a subgroup of \((\F,+)\). In the quotient \(\F/\wp(\F)\), the class of element \(\sum_{i = 1}^n q(e_i)q(f_i)\) is called the *Arf invariant of \(q\)* and is denoted by \(\operatorname{Arf}(q)\). It is independent the choice of symplectic basis (see ). Moreover, it defines a homomorphism \(\operatorname{Arf}: W_q(\F) \to \F/\wp(\F)\) on the Witt group of \(\F\). More explicitly if \(q=[a_1,b_1]~\bot~ \cdots ~\bot~ [a_n,b_n]\) then \(\operatorname{Arf}(q) = a_1b_1 + \cdots + a_nb_n \in \F/\wp(\F)\).\ # Special \(2\)-groups and quadratic forms Let \(G\) be a finite group. Let \(\Omega (G)=\langle x\in G~:~x^2=1\rangle\) and \(G^{\prime} = [G, G]\), the derived subgroup of \(G\). Let \(Z(G)\) denote the center of \(G\) and \(\Phi(G)\) denote the Frattini subgroup of \(G\). Recall that for a non-trivial group the intersection of its index \(2\) subgroups is called its *Frattini subgroup*. We record that \(\Phi(G) =\langle x^2: x\in G \rangle\).\ A \(2\)-group \(G\) is called *special \(2\)-group* if \(G\) is non-commutative and \[G^{\prime} = \Phi (G) = Z(G) = \Omega (Z(G)).\] Moreover, if \(G\) is a special \(2\)-group such that \(|Z(G)|=2\) then \(G\) is called an *extraspecial \(2\)-group*.\ From now onwards \(\F_2\) will denote the field of two elements and \(V\) and \(W\) will be vector spaces over \(\F_2\). A map \(c: V \times V \rightarrow W\) is called a *normal \(2\)-cocycle* on \(V\) with coefficients in \(W\) if for all \(v,v_1,v_2,v_3\in V\) it satisfies the following conditions: 1. \(c(v_2,v_3)-c(v_1+v_2,v_3)+c(v_1,v_2+v_3)-c(v_1,v_2)=0\). 2. \(c(v,0)=(0,v)=0\).\ We denote the set of normal \(2\)-cocycles on \(V\) with coefficients in \(W\) by \(Z^2(V,W)\) and consider it as an abelian group under pointwise addition. Let \(\lambda:V\rightarrow W\) be a linear map such that \(\lambda(0)=0\). Then the map \(c_{\lambda}: V \times V \to W\) defined by \(c_{\lambda}(v_1,v_2)=\lambda(v_2)-\lambda(v_1+v_2)+\lambda(v_1)\) is a normal \(2\)-cocycle. Such \(2\)-cocycles are called *normal \(2\)-coboundaries* and their collection is denoted by \(B^2(V,W)\). The set \(B^2(V,W)\) forms a subgroup of \(Z^2(V,W)\). The quotient \(H^2(V,W)=\frac {Z^2(V,W)}{B^2(V,W)}\) is called the *second cohomology group of \(V\) with coefficients in \(W\)*.\ We consider \(V\) and \(W\) as groups under the operation of vector space addition. A short exact sequence of groups \(1\rightarrow V\rightarrow G\rightarrow W\rightarrow 1\) is called a *central extension of \(W\) by \(V\)* if \(V\subseteq Z(G)\). The set of isomorphism classes of central extension of \(W\) by \(V\) is in one to one correspondence with \(H^2(V,W)\) . We record that the central extension of \(W\) by \(V\) corresponding to a cocycle class \([c]\in H^2(V,W)\) is isomorphic to the group \(V\dot{\times} W\), where the underlying set of the group \(V\dot{\times} W\) is just the cartesian product \(V \times W\) and its group operation is defined by for all \(v, v^{\prime} \in V\) and \(w, w^{\prime} \in W\). The identity element this group is \((0,0)\) and the inverse of \((v,w)\) is \((v, c(v,v)+w)\).\ A map \(q: V \to W\) is called a *quadratic map* if \(q(\alpha v) = \alpha^2 q(v)\) for all \(v \in V\), \(\alpha \in \F\) and the map \(b_q: V \times V \to W\) defined by \(b_q (v, w) = q(v+w)-q(v)-q(w)\) is bilinear. We denote by \(\langle b_q(V \times V) \rangle\) the subgroup of \(W\) generated by the image of \(b_q\). Let \(\operatorname{Quad}(V,W)\) denote the set of quadratic maps from \(V\) to \(W\). We consider it as a group under the group operation of pointwise addition of maps. The following proposition gives correspondence between \(H^2(V,W)\) and \(\operatorname{Quad}(V,W)\). If dimension of \(V\) is finite then above proposition and the correspondence of elements of \(H^2(V,W)\) with central extension of \(V\) by \(W\) gives a useful correspondence between \(\operatorname{Quad}(V,W)\) and central extension of \(V\) by \(W\). Note that the quadratic map \(q\) in the above theorem is regular and the image of \(b_q\) generates \(Z(G)\). This quadratic map \(q\) is called the *quadratic map associated to the special \(2\)-group \(G\)*. The special \(2\)-group \(G\) in the above theorem is called *the group associated to regular quadratic form \(q\)*. We recall the definition of central product of a two groups. Let \(G_1\) and \(G_2\) two groups, \(Z(G_1)\) and \(Z(G_2)\) be their centers and \(\theta: Z(G_1) \to Z(G_2)\) be a group isomorphism. Let \(N\) denote the normal subgroup \(\{(x, y) \in Z(G_1) \times Z(G_2): \theta(x)y = 1\}\) of \(Z(G_1) \times Z(G_2)\). The *central product* of \(G_1\) and \(G_2\) is the quotient of direct product \(G_1 \times G_2\) by \(N\). The following lemma relates orthogonal sum of quadratic maps to central products. We need to show that \(G = G_1 \circ G_2\). By definition \(G_1 \circ G_2\) is the quotient of \(G_1 \times G_2\) by \(\ker(f)\) where \(f:W\times W\rightarrow W\) is the homomorphism defined by \(f(w,w^{\prime})=w+w^{\prime}\); \(w,w^{\prime}\in W\). Define \(\phi :G_1\times G_2\rightarrow G\) by where \((v_1,w)\in G_1,(v_2,w^{\prime})\in G_2\). We notice that \(\phi\) is a group homomorphism as \[\begin{aligned} &\quad \quad \phi(((v_1,w_1),(v_2,w_1^{\prime}))+((v_1^{\prime},w_2),(v_2^{\prime},w_2^{\prime}))) \\ &=\phi(((v_1,w_1)+(v_1^{\prime},w_2)),((v_2,w_1^{\prime})+(v_2^{\prime},w_2^{\prime})))\\ &=\phi((v_1+v_1^{\prime},c_1(v_1,v_1^{\prime})+w_1+w_2),(v_2+v_2^{\prime},c_2(v_2,v_2^{\prime})+w_1^{\prime}+w_2^{\prime}))\\ &=((v_1+v_1^{\prime}, v_2+v_2^{\prime}), c_1(v_1,v_1^{\prime})+c_2(v_2,v_2^{\prime})+w_1+w_1^{\prime}+w_2+w_2^{\prime})\\ &=((v_1,v_2) + (v_1^{\prime},v_2^{\prime}), c((v_1,v_2),(v_1^{\prime},v_2^{\prime}))+w_1+w_1^{\prime}+w_2+w_2^{\prime})\\ &=((v_1,v_2),w_1+w_1^{\prime})((v_1^{\prime},v_2^{\prime}),w_2+w_2^{\prime})\\ &=\phi((v_1,w_1),(v_2,w_1^{\prime}))\phi((v_1^{\prime},w_2),(v_2^{\prime},w_2^{\prime})). \end{aligned}\] where \((v_1,w_1),(v_1^{\prime},w_2)\in G_1\) and \((v_2,w_1^{\prime}),(v_2^{\prime},w_2^{\prime})\in G_2\). The homomorphism \(\phi\) is surjective because for an arbitrary \(((v_1,v_2),w)\in G\) we have \(\phi((v_1,0),(v_2,w))=((v_1,v_2),w)\). Now identifying \(W\times W\) with \((0\dot{\times}W)\times (0\dot{\times}W)\subset G_1\times G_2\), it follows that \(\ker{\phi}\) gets identified with \(\ker{f}\) and we finally have \(G \simeq \frac{G_1\times G_2}{\ker(\phi)} = \frac{G_1\times G_2}{\ker(f)} = G_1 \circ G_2\). \(\square\)\  ◻ where \([c] \in H^{2}(V,W)\) and \(q(x)=c(x,x)\). Let \(x = (v, w) \in V \dot{\times} W = G\). By hypothesis there exists \(a \in V\) such that \(q(a) = q(a-v) = 0\). We take \(y = (a-v, 0) \in G\). Since \[(a-v, 0) + (a-v, 0) = (2(a-v), c(a-v,a-v)) = (2(a-v), q(a-v)) = (0, 0)\] it follows that \(y^2 = 1\). Moreover \[\begin{aligned} (a-v, 0)(v,w)(a-v, 0) &=(2a-v, c(a-v,v)+c(a,a-v)+w)\\ &=(v, c(v,v) + c(a,a)+w) \\ &=(v, q(a) + c(v,v) + w) \\ &=(v, c(v,v) + w) = (v,w)^{-1}. \end{aligned}\] Therefore \(yxy = x^{-1}\). Further since \(y^2 = 1\), we conclude that \(yxy^{-1} = x^{-1}\) and \(G\) is strongly real. \(\square\)\  ◻ In the rest of this article, \(D_4\) will denote the dihedral group of order \(8\) and \(Q_2\) will denote the quaternion group of order \(8\). Presentations of these groups are where \(v,v^{\prime} \in V\), \(\alpha,\alpha^{\prime}\in \F_2\) and \(c_1\in Z^2(V,\F_2)\) is the normal \(2\)-cocycle such that \(q_1(v)=c_1(v,v)\) for all \(v\in V\).\ Define \(\psi:D_4\rightarrow V \dot{\times} \F_2\) by \(\psi(a)=((1,1),1)\) and \(\psi(b)=((1,0),1)\), where \(a\) and \(b\) are generating elements of \(D_4\) as in the presentation of \(D_4\). We check that \(\psi\) is an isomorphism of groups. Note that both \(D_4\) and \(V_{1}\dot{\times}\F_2\) are groups of order \(8\). It is easy to see that the orders of \(((1,1),1)\) and \(((1,0),1)\) are \(4\) and \(2\), respectively. Moreover, \[\begin{aligned} ((1,0),1)((1,1),1)^{-1}&=((1,0),1)((1,1),c_1((1,1),(1,1))+1)\\ &= ((1,0),1)((1,1),q_1(1,1))+1)\\ &= ((1,0),1)((1,1),0)\\ &= ((1,0),c_1((1,0),(1,1))+1)\\ &= ((0,1),1)\\ &= ((1,1),1)((1,0),1) \end{aligned}\] Hence \(\psi\) is an isomorphism of groups. On similar lines it can be shown that group associated with quadratic form \(q_{2} = [1,1]\) is \(Q_{2}\). The isomorphism is given by \(\psi^{\prime}: Q_2\rightarrow V_{2}\dot{\times}\F_2\), where \(\psi^{\prime}(c)=((1,1),1)\), \(\psi^{\prime}(d)=((1,0),1)\) and \(c\), \(d\) denote generators of group \(Q_2\) as in the given presentation of \(Q_2\). \(\square\) ◻ This completes the classification of extraspecial \(2\)-groups. Their classification is also given in using group theoretic methods. We learn from that the classification of extraspecial \(2\)-group using quadratic forms is known. However we do not know of any reference where it is mentioned in as much detail.\ We shall denote the extraspecial 2-group \(D_{4}\circ D_{4}\circ \cdots \circ D_{4}\) (\(n\) copies of \(D_4\)) by \(D_4^{(n)}\) and the extraspecial \(2\)-group \(Q_{2}\circ D_{4}\circ \cdots\circ D_{4}\) (\(n-1\) copies of \(D_4\)) by \(Q_2D_4^{(n-1)}\). We now study strong reality of these groups. To show that \(Q_{2}\circ D_{4}\) is strongly real using the criterion of Th. [\[strongly-real-criterion\]](#strongly-real-criterion){reference-type="ref" reference="strongly-real-criterion"}, for each \(v \in V\) we have to exhibit some \(a \in V\) such that \(q(a) = q(a-v) = 0\). The following table demonstrates that it is indeed possible. Therefore it follows that groups \(Q_2D_4^{(n-1)}\) for \(n \geq 2\) are strongly real. The only extraspecial \(2\)-group which is left out it \(Q_2\), which is not strongly real. This is because in \(Q_2\) there is only one element of order \(2\), which is central. \(\square\)\  ◻ If \(G\) is an extraspecial \(2\)-group with associated quadratic form \(q\) then \(\operatorname{Hom}_{\F_2}(W,\F_2)\) consists only of identity map and by Th. [\[totally-orthogonal-criterion\]](#totally-orthogonal-criterion){reference-type="ref" reference="totally-orthogonal-criterion"} the group \(G\) is totally orthogonal if and only if \(\operatorname{Arf}(q)\) is trivial.\ For all \(n\in \N\), extraspecial \(2\)-groups \(D_4^{(n)}\) are totally orthogonal because the Arf invariant of the quadratic form \(q= [0,0] \perp [0,0] \cdots \perp [0,0]\) associated with to the group \(D_4^{(n)}\) is trivial. Extraspecial \(2\)-groups \(Q_2 D_4^{(n-1)}\) are not totally orthogonal because the Arf invariant of associated the quadratic form \(q=[1,1] \perp [0,0] \cdots \perp [0,0]\) is not trivial.\ Therefore to sum up, all extraspecial \(2\)-groups \(Q_2 D_4^{(n-1)}\), \(n\geq 2\) are examples of strongly real groups which are not totally orthogonal. In fact these groups have exactly one symplectic representation which is of degree \(2^n\) (see ). Further, the least order of a strongly real finite group which are not totally orthogonal is \(32\) and \(Q_2 \circ D_4\) the only group of order \(32\) with this property. We have checked it by computer algebra system GAP. Next order in which an example of strongly real group with symplectic representations is found is \(64\). This is the only special group of order \(64\) which is strongly real and not totally orthogonal. Another group of order \(64\) which is strongly real and not totally orthogonal is \(\mathcal G = \mu_2 \times( Q_2 \circ D_4)\), where \(\mu_2\) is the group of order \(2\). The group \(\mathcal G\) is not special. We have checked using GAP that \(G\) and \(\mathcal G\) are the only strongly real groups of order \(64\) which are not totally orthogonal.\ The polar form associated with \(q\) is \(b_q: V \times V \to W\), \[b_q((w_1,x_1,y_1,z_1), (w_2,x_2,y_2,z_2))=(w_1x_2+x_1w_2+y_1z_2+z_1y_2,w_1y_2+y_1w_2,x_1y_2+y_1x_2)\] where \((w_1,x_1,y_1,z_1), (w_2,x_2,y_2,z_2)\in V\). It is straightforward to check that \(\operatorname{rad}(b_q)=0\) and \(\langle b_q(V\times V)\rangle = W\). Hence by Th. [\[special-2-group-of-a-quad-map\]](#special-2-group-of-a-quad-map){reference-type="ref" reference="special-2-group-of-a-quad-map"} there exists a unique special \(2\)-group \(G\) such that \(V=Z(G)\) and \(\frac{G}{Z(G)} = W\). The order of this group is \(|V| \times |W| = 128\). We shall make explicit computations to show that \(G\) is not strongly real but it is totally orthogonal. For strong reality take, for example, \(v = (1,1,1,1) \in V\). Then we claim that for every \(a \in V\) with \(q(a) = 0\) we have \(q(v-a) \neq 0\). We first identify all \(a \in V\) such that \(q(a) = 0\). Let \(a = (w,x,y,z) \in V\) be one such vector. Then \(q(w,x,y,z) = 0\) will imply \[wx + yz = wy = xy = 0\] If \(y \neq 0\) then the above condition forces \(x = w = z = 0\). This gives \(a = (0, 0, 1, 0)\). If \(y = 0\) then to ensure the above condition one must have either \(w = 0\) or \(x = 0\). Therefore we conclude that \(a \in \{(0,0,0,0),(0,1,0,0), (1,0,0,0),(0,0,0,1),(0,1,0,1), (1,0,0,1)\}\). For each of these possibilities for \(a\) we compute \(q(a-v)\). This table confirms that for \(v = (1,1,1,1)\) there is no \(a \in V\) such that \(q(a) = q(a-v) = 0\) and from Th. [\[strongly-real-criterion\]](#strongly-real-criterion){reference-type="ref" reference="strongly-real-criterion"} we conclude that \(G\) is not strongly real.\ Now we show that the special \(2\)-group \(G\) associated to the quadratic form \(q\) as in ([\[qform-example\]](#qform-example){reference-type="ref" reference="qform-example"}) is totally orthogonal. Since \(\dim_{\F_2}(W,\F_2)=3\), there exist exactly \(7\) non-zero \(\F_2\)-linear maps from \(W\) to \(\F_2\), which are following \[s_n(x,y,z)=ix+jy+kz; \quad \quad(x,y,z)\in W, 1\leq n \leq 7,\] where \(n = 4i+2j+k\) is the binary expansion of \(n \in \{1, 2, \cdots, 7\}\). We write various transfer maps of \(q\) \[\begin{aligned} s_{1_*}(q)(w,x,y,z)&=xy\\ s_{2_*}(q)(w,x,y,z)&=wy\\ s_{3_*}(q)(w,x,y,z)&=wy+xy\\ s_{4_*}(q)(w,x,y,z)&=wx+yz\\ s_{5_*}(q)(w,x,y,z)&=wx+yz+xy\\ s_{6_*}(q)(w,x,y,z)&=wx+yz+wy\\ s_{7_*}(q)(w,x,y,z)&=wx+yz+wy+xy, \end{aligned}\] where \((w,x,y,z)\in V\). By suitable linear changes of variables each of above quadratic forms is isometric to either \(q_1:V\rightarrow \F_2\) defined by \(q_1(w,x,y,z)=wy\) or \(q_2:V\rightarrow \F_2\) defined by \(q_1(w,x,y,z)=wx+yz\).\ Now \(\operatorname{rad}(b_{q_1}) = \langle (0,1,0,0),(0,0,0,1) \rangle\) and \(\frac{V}{\operatorname{rad}(b_{q_1})} = \langle (1,0,0,0),(0,0,1,0) \rangle\) so \(q_1\) induces regular quadratic form \(q^{\prime}_1:\frac{V}{\operatorname{rad}(b_{q_1})}\rightarrow \F_2\) defined by \(q^{\prime}_1(\alpha,\beta)=\alpha\beta\), where \((\alpha,\beta)\in \frac{V}{\operatorname{rad}(b_{q_1})}\). Now \(\operatorname{Arf}(q_1) = \operatorname{Arf}(q^{\prime}_1)=0\). On the similar lines, since \(\operatorname{rad}(b_{q_2}) = 0\) the quadratic form \(q_2\) is regular and \(\operatorname{Arf}(q_2)=0\).\ As a consequence, for all \(s \in \operatorname{Hom}_{\F_2}(W,\F_2)\) the Arf invariant of the transfer \(q_s\) is trivial and by Th. [\[totally-orthogonal-criterion\]](#totally-orthogonal-criterion){reference-type="ref" reference="totally-orthogonal-criterion"} the group \(G\) is totally orthogonal.\ #### **Remark** We remark that the smallest order in which a totally orthogonal special \(2\)-group which is not strongly real exists is \(128\). We have checked using GAP that the smallest totally orthogonal group which is not strongly real is of order \(64\). That group, though, is not a special \(2\)-group.

{'timestamp': '2012-10-16T02:03:15', 'yymm': '1210', 'arxiv_id': '1210.3790', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3790'}

# Introduction Recently, it was shown that an SL(2,R) hidden conformal symmetry  could be realized in a scalar wave equation around the Schwarzschild black hole when taking the near-region and low-energy limits . We refer to the geometry modified in this way as the *subtracted geometry* which shows that it has the same near-horizon properties as the original Schwarzschild black hole, but different asymptotes at infinity . Actually, it does not have an asymptotically flat spacetime implied by the Schwarzschild black hole but it has an asymptotically anti-de Sitter (AdS) spacetime. Importantly, it was claimed that a hidden conformal symmetry has been used to derive purely imaginary quasinormal frequencies (QNFs) by using the operator method. However, we wish to clarify that the Rindler space is the truly near-horizon geometry of the Schwarzschild black hole, while the near-region and low-energy limits of the wave equation around the Schwarzschild spacetime corresponds to the wave equation around the AdS segment of the AdS\(_2\times S^2\), which is inspired by the near-extremal Reissner-Nordström (RN) black hole . Here the near-horizon geometry (region) means \(r\approx r_+\), while both the near-region (\(\omega r \ll1\)) and the low-energy limit (\(\omega r_+ \ll 1\))  are necessary to develop a hidden conformal symmetry in a scalar wave equation. The AdS segment shrinks to the Rindler space for the non-extremal Schwarzschild black hole, whereas it grows to the AdS space for the near-extremal RN black hole. On the other hand, it seems difficult to derive QNFs of a scalar propagating on the Schwarzschild black hole by using a hidden conformal symmetry solely. As is well known, quasinormal modes will be determined by solving a scalar wave equation around the Schwarzschild black hole as well as imposing the boundary conditions: ingoing waves at the horizon and outgoing waves at infinity. We stress that any restriction on the frequency \(\omega\) is not allowed for deriving quasinormal modes. Purely imaginary QNFs were found by employing the operator method developed in the *subtracted geometry* . In this case, the ingoing waves at the horizon were guaranteed, but quasinormal modes do not satisfy the outgoing wave-boundary condition because these modes were developed using the near-region and low-energy limits  where the frequency \(\omega\) should satisfy inequalities of \(\omega \ll 1/r\) and \(\omega \ll 1/r_+\). Furthermore, it was suggested that purely imaginary QNFs (\(\omega_n\)) in Eq. ([\[impq\]](#impq){reference-type="ref" reference="impq"}) yield the correct leading behavior of \(\omega\) in Eq. ([\[nqn\]](#nqn){reference-type="ref" reference="nqn"}) for large damping with large overtone number \(n\). However, we observe from Eq. ([\[impq\]](#impq){reference-type="ref" reference="impq"}) that the important differences are the absence of real part, the non-large overtone number \(n\), and the appearance of angular momentum number \(l\) in the imaginary part. In view of these, purely imaginary QNFs may not be acceptable as QNFs describing a scalar perturbation around the Schwarzschild black hole because the outgoing boundary condition is not imposed at infinity. A promising case was found in the RN black hole. We have derived the purely imaginary QNFs of the RN black hole by making use of a hidden conformal symmetry developed in the near-region and low-energy limits of the scalar equation . We are aware that the operator approach has a limitation because the \(\omega\)-dependent potential is approximated by a single term of a hidden conformal symmetry potential. This means that developing a hidden conformal symmetry in the near-horizon region means neglecting the large \(r>r_+\) behavior of the potential in the whole RN black hole spacetime and thus, leading to the *subtracted geometry*. Fortunately, it was known that purely imaginary QNFs could be also obtained from a scalar perturbation around the near-extremal RN black hole without any modifications . This implies that the *subtracted geometry* \[=near-extremal RN black hole\] is the solution to the original Einstein-Maxwell theory which provides the RN black hole. It is worth noting that both the near-horizon and asymptotic geometries (whole information) are necessary to derive QNFs, even though the entropy counting for a black hole may require the near-horizon geometry only . In fact, it shows different utilities of the hidden conformal symmetry in deriving between QNFs and entropy. Hence we propose that if one obtains QNFs (purely imaginary QNFs) using the hidden conformal symmetry based on the *subtracted geometry*, one has to find the corresponding black hole (near-extremal RN black hole) where the same QNFs will be found by solving the scalar equation as well as imposing two boundary conditions without taking any limits. One counter example is the Schwarzschild black hole. In this case, the *subtracted geometry* \[=near-extremal RN black hole\] is not the solution to the original Einstein theory which provides the Schwarzschild black hole. We would like to stress that the Rindler spacetime is the near-horizon solution to the Einstein theory which gives us the Schwarzschild black hole. In this work, we will show that purely imaginary QNFs obtained using a hidden conformal symmetry are not suitable for describing largely damped modes around the Schwarzschild black hole. # Hidden conformal symmetry in the subtracted geometry Let us begin with the Schwarzschild metric given by [\[Smetric\]]{#Smetric label="Smetric"} ds\^2\_Sch=-f(r)dt\^2++r\^2d\^2_2, where \(f(r)=1-r_+/r\) with \(r_+=2M\). Here, \(M\) is the ADM mass and the surface gravity is ==2T_H with the Hawking temperature \(T_H\). The Klein-Gordon equation for a massless scalar is given by [\[SKG\]]{#SKG label="SKG"} \|\_Sch= 0. Expanding in eigenmodes of [\[ansatz\]]{#ansatz label="ansatz"} (t,r,,) = e\^-itY\^l_m(,) , and using the tortoise coordinate defined by [\[tor\]]{#tor label="tor"} r\_\*= = r+r\_+(-1), the radial part of Eq. ([\[SKG\]](#SKG){reference-type="ref" reference="SKG"}) becomes the Schrödinger-type equation R(r) +R(r)=0, where the potential is given by [\[pot\]]{#pot label="pot"} V\_Sch(r)=f(r). On the other hand, the approximated Klein-Gordon equation could be expressed in terms of the eigenvalue equation  [\[hcs-sch\]]{#hcs-sch label="hcs-sch"} H\^2&=& (r\^2f(r)\^2_r+2(r-M)\_r -\^2_t)\ &=&l(l+1)which is designed for describing a scalar propagating on the Schwarzschild spacetime in the near-region and low-energy limits. Note that the SL(2,R) Casimir operator \({\cal H}^2\) is given by \^2=-H\^2_0+(H_1H\_-1+H\_-1H_1), where three operators H\_ &=& i e\^t/4M(\^1/2\_r 4M(r-M)\^-1/2\_t),\ H_0 &=&-4 i M \_t   with  =r\^2f(r) obey the SL(2,R) commutation relations: \[\label{sl-alg} \Big[H_0,H_{\pm 1}\Big]=\mp i H_{\pm 1},~~\Big[H_1,H_{-1}\Big]=2iH_{0}.\] In order to investigate what happens when the near-region and low-energy limits are taken into account, we rewrite the approximated Klein-Gordon equation ([\[hcs-sch\]](#hcs-sch){reference-type="ref" reference="hcs-sch"}) by using the tortoise coordinate \(r_*\) as R(r) +R(r)=0, where the \(\omega\)-dependent potential V\_(r)=\^2(1-) + V\_Sch(r) has an additional term to the original potential \(V_{\rm Sch}(r)\). Since the potential \(V_{\rm Sch}(r)\) contains all information for the Schwarzschild black hole spacetime where a scalar propagates on, the appearance of \(\omega^2\)-dependent term is unusual and thus, it reflects the near-region and low-energy limits. We observe that the \(\omega^2\)-dependent term arises from when replacing five terms in \(\omega^2r^4/\Delta\) with a single term \(\omega^2r^4_+/\Delta\) , which is regarded as a key step to develop the hidden conformal symmetry in the Schwarzschild spacetime. This replacement is done when taking the near-region limit (\(\omega r \ll 1\)) and low-energy limit (\(\omega r_+\ll 1\)). If one does not take these limits, the \(r_+\) in the \(\omega^2\)-dependent term goes back to \(r\) and thus, the \(\omega^2\)-dependent term disappears. This implies that \(V_\omega(r)\) leads to \(V_{\rm Sch}(r)\) for \(r \to r_+\). In Fig. [\[fig.1\]](#fig.1){reference-type="ref" reference="fig.1"}, we depict the four potentials which consist of \(V_{\rm Sch}(r)\) and three \(V_\omega(r)\) for \(\omega=0.1,~0.2,~0.3\). The figure shows that in the near-horizon limit (\(r\to r_+=2\)), all the potentials have the nearly same form, irrespective of \(\omega^2\) but for \(r>r_+\), they have different forms depending on \(\omega^2\). Hence, we observe that for \(r\approx r_+\), \(V_\omega(r) \approx V_{\rm Sch}(r)\) for any \(\omega\). When \(\omega\) approaches zero, \(V_\omega(r)\) recovers the original potential \(V_{\rm Sch}(r)\). However, it is worth noting that the near-horizon region of \(r \approx r_+\) is not enough to derive the quasinormal modes because two boundary conditions are required. Definitely, one observes a difference between \(V_{\rm Sch}(r) \sim 0\) and \(V_\omega(r)\sim \omega^2\) for large \(r\). In order to describe the near-horizon region well, it is convenient to introduce a new coordinate defined by [\[rho\]]{#rho label="rho"} -. In terms of \(\rho\), the event horizon \(r_+\) is mapped into \(\rho\rightarrow\infty\), while the spatial infinity \(r\rightarrow\infty\) into \(\rho\rightarrow 0\): \(r \in[r_+,\infty]\) is inversely mapped into \(\rho \in[\infty,0]\). Using the coordinate ([\[rho\]](#rho){reference-type="ref" reference="rho"}), the Schwarzschild metric becomes [\[rho-metric\]]{#rho-metric label="rho-metric"} ds\^2\_ = -()dt\^2 +\^-1()()\^2 with ()=e\^-2. Then, making use of the ansatz (t,,,) = e\^-itR()Y\^l_m(,) , the massless scalar propagating in the spacetime ([\[rho-metric\]](#rho-metric){reference-type="ref" reference="rho-metric"}) satisfies [\[radial\]]{#radial label="radial"} ()\^2R() +R()=0. This can be rewritten as the Schrödinger-type equation [\[sch2\]]{#sch2 label="sch2"} R() +R()=0. Here the \(\omega\)-dependent potential is given by [\[pot2\]]{#pot2 label="pot2"} \_()= \^2+. We note a useful relation between \(r\) and \(\rho\) ()\^4 =16\^2()\^4(). In the limits of the near-region (\(\rho\rightarrow\infty\)) and low-energy (\(\omega \to 0\)), the square bracket in ([\[pot2\]](#pot2){reference-type="ref" reference="pot2"}) is approximated to be zero as \^2\~ \~0. Thus, in these limits, the \(\omega\)-dependent potential \(\tilde{V}_\omega\) ([\[pot2\]](#pot2){reference-type="ref" reference="pot2"}) reduces to a single term [\[pot3\]]{#pot3 label="pot3"} V\_HCS() = . We note that this potential is not the original potential \(\tilde{V}_\omega\). This is just an approximated potential to explore the hidden conformal symmetry in the near-horizon region of the Schwarzschild spacetime. Hence, we have the *subtracted geometry* when using \(V_{\rm HCS}\) instead of \(\tilde{V}_\omega\). Its near-horizon and asymptotic forms take \(V_{\rm HCS}(\rho\to \infty)\sim e^{-2\kappa \rho}\) and \(V_{\rm HCS}(\rho\to 0)\sim \frac{1}{\rho^2}\), which show that \(V_{\rm HCS}(\rho)\) is similar to the potential of a scalar field around the AdS-black hole. This suggests that its asymptote is changed from a flat spacetime implied by the Schwarzschild black hole to an AdS spacetime. In order to exhibit the hidden conformal structure, we construct three vectors defined as H\_1/-1&=&e\^t ,\ H_0 &=&-\_t, which satisfy the SL(2,R) algebra ([\[sl-alg\]](#sl-alg){reference-type="ref" reference="sl-alg"}). Then, the SL(2,R) Casimir operator is given by [\[hcs-sch1\]]{#hcs-sch1 label="hcs-sch1"} H\^2=-\^2\^2_t +\^2\^2\_. As a result, the Schrödinger equation ([\[sch2\]](#sch2){reference-type="ref" reference="sch2"}) with \(V_{\rm HCS}(\rho)\) in Eq. ([\[pot3\]](#pot3){reference-type="ref" reference="pot3"}) instead of \(\tilde{V}_\omega(\rho)\) is equivalent to the eigenvalue equation [\[eigenv\]]{#eigenv label="eigenv"} H\^2=l(l+1). Now, one may use the hidden conformal symmetry to derive QNFs of the Schwarzschild black hole represented by the *subtracted geometry*. We start with the primary state given by H_0\^(0)=ih\^(0) which satisfies the highest weight condition [\[hwc\]]{#hwc label="hwc"} H_1\^(0)=0. Then, using the ansatz \^(0)=e\^-i_0 tR\^(0)()Y\^l_m(,), one has a conformal weight [\[h1\]]{#h1 label="h1"} h=i=i . On the other hand, the SL(2,R) Casimir operator satisfies [\[caop\]]{#caop label="caop"} H\^2\^(0)=h(h+1)\^(0). Comparing Eq. ([\[caop\]](#caop){reference-type="ref" reference="caop"}) with Eq. ([\[eigenv\]](#eigenv){reference-type="ref" reference="eigenv"}), one has h=\[1(2l+1)\], and thus, one finds \_0=-i\[1(2l+1)\]. Since the QNFs are purely imaginary \(\omega_I>0~(\omega=\omega_R-i\omega_I)\) with \(\omega_R=0\), we choose the upper sign as \_0=-i(l+1). All the descendants can be constructed by acting the operator \(H_{-1}\) on \(\Phi^{(0)}\) \^(n)=(-iH\_-1)\^n\^(0) so that we have \^(n)=e\^-i_n tR\^(n)()Y\^l_m(,). Here the QNFs could be read off as [\[impq\]]{#impq label="impq"} \_n=\_0-in=-i,   nZ\^+ which are purely imaginary. On the other hand, numerical computations of the QNFs for the Schwarzschild black hole in the limit of large damping is given by  [\[nqn\]]{#nqn label="nqn"} = (1.098612) T\_H-i 2T\_H (n+), where the real part approaches a constant of \(\ln(3)\) , while the imaginary part becomes equally spaced with large \(n\) (\(\omega_I\to \infty\)). Comparing ([\[impq\]](#impq){reference-type="ref" reference="impq"}) with the highly damped frequencies ([\[nqn\]](#nqn){reference-type="ref" reference="nqn"}), one may see apparently that the leading behavior of \(i2\pi T_H n\) for large damping comes out correctly. However, we observed from ([\[nqn\]](#nqn){reference-type="ref" reference="nqn"}) that the important differences are the absence of real part, the non-large \(n\), and the appearance of \(l\) in the imaginary part. These show that the QNFs \(\omega_n\) are not suitable for describing the scalar perturbation absorbing into the black hole. Moreover, the \(n\)-th radial eigenfunction \(R^{(n)}(\rho)\) is constructed as R\^(n)()&=&()\^-n (-i\_n-1()-())\ && (-i\_n-2()-())\ && (-i_0()-())R\^(0)(), which may be regarded as the \(n\)-th radial quasinormal modes. We note that \(\Phi^{(n)}\) forms a principally discrete highest weight representation of the SL(2,R) H_0\^(n)=i(h+n)\^(n). On the other hand, the highest weight condition ([\[hwc\]](#hwc){reference-type="ref" reference="hwc"}) provides the radial solution [\[R0sol\]]{#R0sol label="R0sol"} R\^(0)()=C\^i. Near the horizon of \(\rho\to \infty~(r_*\to-\infty)\), the solution ([\[R0sol\]](#R0sol){reference-type="ref" reference="R0sol"}) behaves as R\^(0) \~e\^i_0 , showing that this is the outgoing mode (\(\rightarrow\)) into the horizon which is equivalent to the ingoing mode (\(\leftarrow\)) at the horizon when using the tortoise coordinate \(r_*\) in Eq. ([\[tor\]](#tor){reference-type="ref" reference="tor"}). We observe that \(R^{(0)}(0)=0\), which may show that it is not the outgoing wave at infinity but it satisfies the Dirichlet boundary condition at the infinity of AdS spacetime. Furthermore, the first radial eigenfunction \(R^{(1)}(\rho)\) can be explicitly constructed as R\^(1)() = which show that at infinity \(R^{(1)}\) vanishes, while near the horizon \(R^{(1)}\) remains to be the outgoing mode. One can easily show that the \(n\)-th radial eigenfunction \(R^{(n)}(\rho)\) behaves as the same way as \(R^{(1)}\) by induction. In order to obtain the truly QNFs, we have to impose the two boundary conditions: outgoing mode near the horizon (\(\rho \to \infty\)) and ingoing mode at infinity (\(\rho \to 0\)). However, we point out that \(R^{(n)}(\rho)\) do not satisfy the ingoing boundary condition at infinity because they are the solutions which were obtained by considering the *subtracted geometry*. Curiously, we observe that \(R^{(n)}(0)=0\) which implies that in this framework, one could not impose the ingoing boundary condition at infinity. In this sense, we could not regard ([\[impq\]](#impq){reference-type="ref" reference="impq"}) as the truly QNFs which describe the largely damped modes around the Schwarzschild black hole because we have considered the *subtracted geometry*. # QNFs of scalar around AdS-segment In this section, we show explicitly that the approximated potential ([\[pot3\]](#pot3){reference-type="ref" reference="pot3"}) comes from the AdS\(_2\)-segment inspired by the AdS\(_2\times S^2\) of the near-extremal RN black hole, but not from the Rindler space which is the genuine near-horizon geometry of the Schwarzschild spacetime. This implies that we will no longer use the *subtracted geometry* to derive the QNFs of scalar around the Schwarzschild black hole. First of all, we mention that the near-horizon geometry of the Schwarzschild spacetime is a product of the Rindler space and a two sphere \(S^2\). This can be easily realized when using the coordinate transformations [\[nearS\]]{#nearS label="nearS"} =,   r=r\_+(1+). The Schwarzschild spacetime ([\[Smetric\]](#Smetric){reference-type="ref" reference="Smetric"}) is turned out to be the Rindler spacetime [\[Snear\]]{#Snear label="Snear"} ds\^2\_RS=r\^2\_+(-\^2d\^2+d\^2+d\^2_2) in the near-horizon region. Note that it is impossible to develop a hidden conformal symmetry in the Rindler spacetime as will be shown in the next section. On the other hand, the AdS\(_2\)-segment  is introduced to be [\[ads2\]]{#ads2 label="ads2"} ds\^2\_AdS=Q\^2(-\^2  d\^2 +d\^2+d\^2_2), which describes the near-extremal RN black hole. We note here that for \(r_0=1\) , ([\[ads2\]](#ads2){reference-type="ref" reference="ads2"}) reduces to the near-horizon geometry (AdS\(_2\times S^2\), Bertotti-Robinson spacetime) of the extremal RN black hole  [\[ads2li\]]{#ads2li label="ads2li"} ds\^2\_ERN=Q\^2(-\^2  d\^2 +d\^2+d\^2_2), which is a solution to the Einstein-Maxwell theory. In this extremal case, its Hawking temperature is zero and its QNFs are not defined properly . The spacetime is described by Ricci scalars R\_AdS_2=-,  R\_S\^2=,  R=R\_AdS_2+R\_S\^2=0. It is clear that ([\[ads2li\]](#ads2li){reference-type="ref" reference="ads2li"}) could not describe the near-extremal RN black hole because it contains only \(Q\) when comparing to ([\[ads2\]](#ads2){reference-type="ref" reference="ads2"}) with two parameters \(Q\) and \(r_0\). We are interested in obtaining QNFs of scalar around the black hole with non-zero Hawking temperature. Hence, we require \(r_0 \not=1\). In fact, in the limit of \(\eta\rightarrow 0\) together with \(r_0= 1\) and \(Q=r_+\), ([\[ads2\]](#ads2){reference-type="ref" reference="ads2"}) reduces to the Rindler metric ([\[Snear\]](#Snear){reference-type="ref" reference="Snear"}). In order to make a further connection to the near-extremal RN black hole, we introduce the coordinate transformations \^2=\^2-r\^2_0,   =. Then, one can obtain the near-extremal RN black hole with \(r_0=(r_+-r_-)/2\ll1\) [\[nearRN\]]{#nearRN label="nearRN"} ds\^2\_NERN=-dt\^2+d\^2+Q\^2d\^2 from the metric ([\[ads2\]](#ads2){reference-type="ref" reference="ads2"}). Here \(\tilde\rho\in[r_0,\infty]\). For this near-extremal RN black hole, the surface gravity is computed to be [\[dikappa\]]{#dikappa label="dikappa"} =. Introducing the tortoise coordinate defined by \_\*=(),  \_\*, the Klein-Gordon equation for the radial coordinate becomes the Schrödinger-type equation R(\_\*) + R(\_\*)=0, where the near-extremal RN potential is given by [\[neblp\]]{#neblp label="neblp"} V\_NERN(\_\*)=. This becomes the same form of \(V_{\rm HCS}(\rho)\) in Eq. ([\[pot3\]](#pot3){reference-type="ref" reference="pot3"}) when replacing \(\tilde{\kappa}\) and \(\rho_*\) by \(\kappa\) and \(\rho\). An important fact to point out is that the near-extremal RN potential is valid for whole spacetimes outside the horizon because we consider the near-extremal RN black hole itself. Next, in order to exhibit a hidden conformal structure of the spacetime ([\[nearRN\]](#nearRN){reference-type="ref" reference="nearRN"}), we construct three vectors as [\[hcs-ads\]]{#hcs-ads label="hcs-ads"} \_1/-1&=& e\^t (\_t (\^2-r\^2_0)\^1/2\_),\ \_0 &=&-\_t satisfying the SL(2,R) algebra ([\[sl-alg\]](#sl-alg){reference-type="ref" reference="sl-alg"}). Then, the SL(2,R) Casimir operator is given by [\[casimir1\]]{#casimir1 label="casimir1"} \^2= . As a result, the radial equation can be rewritten as the eigenvalue equation [\[eigenv-ads\]]{#eigenv-ads label="eigenv-ads"} \^2=l(l+1), which is valid for whole space of \(\tilde\rho \in [r_0,\infty]\). Here, using the hidden conformal symmetry ([\[hcs-ads\]](#hcs-ads){reference-type="ref" reference="hcs-ads"}) as before, we can easily find the QNFs as [\[qnfa\]]{#qnfa label="qnfa"} \_n=\_0-i n=-iwith different surface gravity \(\tilde{\kappa}\) in ([\[dikappa\]](#dikappa){reference-type="ref" reference="dikappa"}). The \(n\)-th radial eigenfunction \(R^{(n)}(\tilde{\rho})\) is now obtained as R\^(n)()&=& \^-n (- +(\^2-r\^2_0)\^1/2)\ && (- +(\^2-r\^2_0)\^1/2)\ && (- +(\^2-r\^2_0)\^1/2)R\^(0)(). On the other hand, by solving the highest weight condition \_1\^(0)=0, we have R\^(0)()=D(\^2-r\^2_0)\^-. In terms of \(\rho_*\), it can be expressed as R\^(0)(\_\*)=D \^i . Near the horizon of \(\rho_*\to \infty(\tilde\rho\rightarrow r_0)\), it behaves as R\^(0)(\_\*)\~e\^i_0\_\*, which is the outgoing mode (\(\rightarrow\)) into the horizon. At infinity (\(\rho_* \to 0,~\tilde\rho\rightarrow\infty\)), we observe that R\^(0)(\_\*)\~0, which satisfies the Dirichlet boundary condition at the spatial infinity of the AdS\(_2\) spacetime. Furthermore, the first radial eigenfunction \(R^{(1)}(\tilde\rho)\) can be explicitly constructed as R\^(1)() =- (\^2-r\^2_0)\^-. Expressing it in terms of \(\rho_*\) and then, checking the boundary conditions at both sides leads to R\^(1)(\_\*) = It shows that at infinity \(R^{(1)}(\rho_*)\) vanishes, while near the horizon, \(R^{(1)}(\rho_*)\) remains to be the outgoing mode. One can easily show that the \(n\)-th radial eigenfunction \(R^{(n)}(\rho_*)\) satisfies the same AdS-boundary condition as \(R^{(1)}(\rho_*)\) by induction. # QNFs in the Rindler spacetime First of all, we investigate the boundary conditions for QNFs in the Rindler spacetime. For that purpose, let us introduce the tortoise coordinate u=-, which maps \(\eta\in [0,\infty]\) inversely into \(u\in [\infty,0]\). The Schrödinger-type equation can be written as R+ R=0. Here the Rindler potential is given by V\_RS(u)= l(l+1)\^2 e\^-2u, which is not the form of the potential \(V_{\rm HCS}(\rho)\) in Eq. ([\[pot3\]](#pot3){reference-type="ref" reference="pot3"}) when replacing \(u\) by \(\rho\). Near the horizon (\(u\rightarrow \infty\)), the Rindler potential shows the same behavior as \(V_{\rm AdS}(\rho)\), while it shows different behavior at infinity (\(u \to 0\)). Thus, for quasinormal mode-boundary condition, one requires an outgoing mode as \(R(u)\sim e^{i\omega u}\) (ingoing mode \(R(\eta)\sim e^{-i\omega \eta}\) in terms of \(\eta\)) near the horizon. Since the potential height is finite as \(l(l+1) \kappa^2\) at infinity (\(u\rightarrow 0)\), one may require the ingoing mode (outgoing mode expressed in terms of \(\eta\)). However, it turned out that there is no QNFs in the Rindler spacetime satisfying such QNFs boundary conditions  because the modes satisfying near horizon boundary condition fail to fulfill the boundary condition at infinity at the same time. Moreover, for the Rindler spacetime which is the genuine near-horizon limit of the Schwarzschild black hole spacetime, the SL(2,R) algebra is not closed. Explicitly, three vectors obtained from Eq. ([\[hcs-ads\]](#hcs-ads){reference-type="ref" reference="hcs-ads"}) in the small \(\eta\) limit as H'\_1/-1=-i e\^ (\_t\_),   H'\_0 =-i_t give the following commutation relation =iH'\_,  \[H'\_1,H'\_-1\] =0, whose last term does not satisfy the SL(2,R) algebra ([\[sl-alg\]](#sl-alg){reference-type="ref" reference="sl-alg"}). # Conclusion We have shown that the near-horizon limit of the Schwarzschild black hole is the Rindler spacetime but not the near-region and low-energy limits of the Schwarzschild black hole known as the *subtracted geometry*. It was shown that the hidden conformal symmetry is developed for the *subtracted geometry* only, but the *subtracted geometry* is not sufficient to derive the QNFs. The purely imaginary QNFs ([\[impq\]](#impq){reference-type="ref" reference="impq"}) developed on the *subtracted geometry* without imposing outgoing boundary condition at infinity have also been obtained as ([\[qnfa\]](#qnfa){reference-type="ref" reference="qnfa"}) from the quasinormal modes which satisfy two boundary conditions of AdS\(_2\) spacetime: ingoing mode at the horizon and Dirichlet boundary condition at infinity. The latter boundary condition is designed for the near-extremal RN black hole and thus, is sharply contrasted to the outgoing mode-boundary condition imposed at asymptotically flat spacetime of the Schwarzschild black hole. Hence, purely imaginary QNFs obtained from the *subtracted geometry* could not describe the largely damped modes of scalar perturbation around the Schwarzschild black hole. When comparing ([\[impq\]](#impq){reference-type="ref" reference="impq"}) with ([\[nqn\]](#nqn){reference-type="ref" reference="nqn"}), we have observed that the important differences are the absence of real part, the non-large overtone number \(n\), and the appearance of angular momentum number \(l\) in the imaginary part. This shows clearly that the QNFs \(\omega_n\) ([\[impq\]](#impq){reference-type="ref" reference="impq"}) are not suitable for describing the scalar perturbation absorbing into the Schwarzschild black hole with temperature \(T_H=1/8\pi M\). The purely imaginary QNFs \(\omega_n\) is suitable for the scalar absorbing into the near-extremal RN black hole with temperature \(\tilde{T}_H=r_0/2\pi Q^2\).

{'timestamp': '2012-10-16T02:02:50', 'yymm': '1210', 'arxiv_id': '1210.3760', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3760'}

# Introduction Understanding the characteristic structure of fermion masses and flavor mixings is one of the major outstanding problems in particle physics. Up-type quarks, down-type quarks, charged leptons and neutrinos have distinct hierarchical mass patterns from each other. Moreover, the observed flavor mixing is small for quarks but large for leptons. This situation is in sharp contrast to a naive expectation from quark--lepton unification. We can disentangle ourselves from this discordance by considering the state-mixings between quarks (leptons) and extra particles beyond the minimal supersymmetric standard model (MSSM). In fact, it was shown in the context of \(SU(6) \times SU(2)_R\) string-inspired model, which contains massless particles beyond the MSSM, that we are able to explain characteristic patterns of the observed mass spectra and mixing matrices of quarks and leptons. In the model the Froggatt--Nielsen (F--N) mechanism plays an important role. In Ref. it was predicted that the absolute value of (1, 3) element of the MNS matrix \(|U_{e3}| = |\sin \theta_{13}|\) lies between \(\lambda\) and \(\lambda^2\), where \(\lambda = 0.23\). Recently, new data on neutrino mixings become available. The observed value of mixing angle \(|\sin \theta_{13}|\) turned out to be \(\sim \lambda^{1.3}\), which supports this prediction. In this paper we carry out more detailed study of fermion mass hierarchies and flavor mixings in the \(SU(6) \times SU(2)_R\) string-inspired model. The hierarchical structure due to the F--N mechanism comes out not only in the effective Yukawa couplings but also in the R-handed neutrino Majorana mass matrix. In the neutrino sector, where the seesaw mechanism is at work, the hierarchical factors (the F--N factors) in the Dirac mass matrix are faced with the inverse of those in the Majorana mass matrix. This situation in the neutrino sector brings about a significant feature peculiar to the MNS matrix, on which main emphasis is placed in this paper. As will be shown later, the neutrino mass matrix takes the form \[{\cal M}_{\nu} \propto \varLambda_{\kappa} \, \varSigma \, \varLambda_{\kappa} \label{neutrino}\] in the mass diagonal basis for charged leptons with \[\begin{aligned} \varLambda_{\kappa} & = & {\rm diag}\left( \kappa_1, \ \kappa_2, \ 1 \right), \\ \varSigma & \simeq & \left( \begin{array}{ccc} 1 & 0 & 0 \\ -\sigma_1 & 1 & 0 \\ \sigma_3 & -\sigma_4 & 1 \end{array} \right) \times N^{-1} \times \left( \begin{array}{ccc} 1 & -\sigma_1 & \sigma_3 \\ 0 & 1 & -\sigma_4 \\ 0 & 0 & 1 \end{array} \right), \label{Sgm} \end{aligned}\] where the parameters \(\kappa_i \ (i = 1, \ 2)\) and \(\sigma_i \ (i = 1, \ 3, \ 4)\) are \({\cal O}(1)\). The above form of \({\cal M}_{\nu}\) is derived as a consequence of the fact that the F--N factors appearing in the charged lepton masses cancel out in large part through the seesaw mechanism. The matrix \(N\) in Eq.([\[Sgm\]](#Sgm){reference-type="ref" reference="Sgm"}), in which the F--N factors are eliminated from R-handed neutrino Majorana mass matrix, has no hierarchical structure and \(\det N = 1\). As an example of parameter choice, taking a simple case that \(N = \varLambda_{\kappa} = {\bf 1}\) and putting \(\sigma_1 = \sigma_4 = 2.2\) and \(\sigma_3 = 1.2\), we obtain a large mixing solution of the MNS matrix \[V_{\rm MNS} = \left( \begin{array}{ccc} \ 0.865 \ & \ 0.471 \ & \ 0.172 \\ \ -0.455 \ & \ 0.594 \ & \ 0.663 \\ \ 0.211 \ & \-0.652 \ & \ 0.728 \end{array} \right)\] for the normal hierarchy. All elements of \(V_{\rm MNS}\) offered here are accommodated to the present experimental data within 20% provided that \(\delta_{CP} = 0\). In addition, eigenvalues of \(\varSigma\) lead to \[\left( \frac{\varDelta m_{32}^2}{\varDelta m_{21}^2} \right)^{1/2} = 5.57,\] which also coincides with the observed value. It is also found that a simple choice of the parameter values is consistent with all observed values of fermion masses and mixings. This paper is organized as follows. In section 2 we briefly review the \(SU(6) \times SU(2)_R\) string-inspired model together with the F--N mechanism. Taking the mass matrix of up-type quarks as an example, we illustrate the whole scheme of the present model. For comparison with the case of the lepton sector, we study the mass matrix of down-type quarks in section 3. State-mixings occur between down-type quarks and colored Higgsinos with even R-parity. We give the explicit form of the CKM matrix, which proves to exhibit small mixing. Similarly to the case of down-type quarks, state-mixings take place between leptons and \(SU(2)_L\)-doublet Higgsinos. The mass matrix in the charged lepton sector is studied in section 4. Section 5 deals with the neutrino sector in which state-mixings enter into the seesaw mechanism. The present approach provides a phenomenological framework which enable us to analyze many experimental data. Numerical analysis of the MNS matrix and fermion spectra is given in section 6. Section 7 is devoted to summary. # Model and Froggatt--Nielsen mechanism In this study we choose \(SU(6) \times SU(2)_R\) as a unification gauge symmetry at the underlying string scale \(M_S\), which can be derived from the perturbative heterotic superstring theory via the flux breaking. In terms of \(E_6\) we set matter superfields which consist of three families and one vector-like multiplet, i.e., \[3 \times {\bf 27}(\Phi_{1,2,3}) + ({\bf 27}(\Phi_0)+\overline{\bf 27}({\overline \Phi})).\] The superfields \(\Phi\) are decomposed into two multiplets of \(G= SU(6) \times SU(2)_R\) as \[\Phi({\bf 27})=\left\{ \begin{array}{lll} \phi({\bf 15},{\bf 1}) &: & \quad \mbox{\(\{Q,L,g,g^c,S\}\)}, \\ \psi(\overline{\bf 6},{\bf 2}) &: & \quad \mbox{\(\{(U^c,D^c),(N^c,E^c),(H_u,H_d)\}\)}, \end{array} \right.\] where \(g\), \(g^c\) and \(H_u\), \(H_d\) represent colored Higgs and \(SU(2)_L\)-doublet Higgs superfields, respectively. Doublet Higgs and color-triplet Higgs fields belong to different representations of \(G\) and this situation is favorable to solve the triplet-doublet splitting problem. The superfields \(N^c\) and \(S\) are R-handed neutrinos and \(SO(10)\)-singlets, respectively. Although \(D^c\) and \(g^c\) as well as \(L\) and \(H_d\) have the same quantum numbers under the standard model gauge group \(G_{\rm SM} = SU(3)_c \times SU(2)_L \times U(1)_Y\), they belong to different irreducible representations of \(G\). We assign odd R-parity to the superfields \(\Phi_{1,2,3}\) and even to \(\Phi_0\) and \(\overline{\Phi}\), respectively. Since ordinary Higgs doublets have even R-parity, they are contained in \(\Phi_0\). It is assumed that R-parity remains unbroken down to the electroweak scale. The gauge symmetry \(G\) is spontaneously broken in two steps at the scales \(\langle S_0\rangle=\langle \overline{S} \rangle\) and \(\langle N_0^c\rangle=\langle \overline{N^c} \rangle\) as \[G = SU(6) \times SU(2)_R \buildrel \langle S_0 \rangle \over \longrightarrow SU(4)_{\rm PS} \times SU(2)_L \times SU(2)_R \buildrel \langle N^c_0 \rangle \over \longrightarrow G_{\rm SM},\] where \(SU(4)_{\rm PS}\) represents the Pati-Salam \(SU(4)\). Hereafter it is supposed that the symmetry breaking scales are \(\langle S_0 \rangle = 10^{17-18}\)GeV and \(\langle N^c_0 \rangle = 10^{15-17}\)GeV. We have two types of gauge invariant trilinear couplings \[\begin{aligned} (\phi ({\bf 15},{\bf 1}))^3 & = & QQg + Qg^cL + g^cgS, \nonumber \\ \phi ({\bf 15},{\bf 1})(\psi (\overline{\bf 6},{\bf 2}))^2 & = & QH_dD^c + QH_uU^c + LH_dE^c + LH_uN^c \\ {}& & \qquad + SH_uH_d + gN^cD^c + gE^cU^c + g^cU^cD^c \nonumber \end{aligned}\] in the superpotential \(W\). From the viewpoint of the string unification theory, it is probable that the hierarchical structure of Yukawa couplings is attributed to some kind of flavor symmetries at the string scale \(M_S\). If the flavor symmetry contains Abelian groups, the F--N mechanism works for the interactions among quarks, leptons and Higgs fields. Accordingly, effective Yukawa interactions arise from non-renormalizable terms which respect the flavor symmetry. Here we consider the effective Yukawa interactions for up-type quarks given by \[W_U = \sum_{i,j=1}^{3} \, {\cal M}_{ij} Q_i U^c_j H_{u0},\] where the subscripts \(i\) and \(j\) are the generation indices. Due to the F--N mechanism, the matrix \({\cal M}\) takes the form \[{\cal M} = f_M \, \varGamma_1 M \varGamma_2. \label{Mnoh}\] Our basic assumption is that the hierarchical structure of all \(3 \times 3\) mass matrices is attributed to the F--N factors \(\varGamma_1\) and/or \(\varGamma_2\). Hence, hierarchy of \({\cal M}\) stems only from \(\varGamma_1\) and \(\varGamma_2\), and \(M\) contains no hierarchical structure. Here we put a factor \(f_M\) in order to set \(\det M = 1\). The F--N factors \(\varGamma_1\) and \(\varGamma_2\) are described as \[\varGamma_1 = {\rm diag}( x^{\alpha_1}, \ x^{\alpha_2}, \ 1), \qquad \varGamma_2 = {\rm diag}( x^{\beta_1}, \ x^{\beta_2}, \ 1),\] where \(x\) is given by \[x = \frac{\langle S_0 \rangle \, \langle \overline{S} \rangle}{M_S^2} < 1\] and \((S_0{\overline S})\) is a \(G\)-invariant with a nonzero flavor charge. The exponents \(\alpha_1, \ \alpha_2, \ \beta_1, \ \beta_2\) in the F--N factors are determined by assigning flavor charges to the matter fields. Even if \(x\) by itself is not a very small number, physical parameters can be very small if they depend on high powers of \(x\). We assume the hierarchical patterns \[x^{\alpha_1} \ll x^{\alpha_2} \ll 1, \qquad x^{\beta_1} \ll x^{\beta_2} \ll 1\] by suitably chosen flavor charges. In this paper we ignore the phase factors of vacuum expectation values (VEV's). The mass matrix \({\cal M}\) is diagonalized via bi-unitary transformation as \[{\cal V}_u^{-1} {\cal M} \,{\cal U}_u = \varLambda_u, \qquad v_{u0} \varLambda_u = {\rm diag}(m_u, \ m_c, \ m_t)\] with \(v_{u0} = \langle H_{u0} \rangle\). Up-type quark masses are given by \[(m_u, \ m_c, \ m_t) \, \simeq \, v_{u0} \, f_M \times \left( \frac{1}{\overline{m}_{11}} \, x^{\alpha_1 + \beta_1}, \ \frac{\overline{m}_{11}}{m_{33}} \, x^{\alpha_2 + \beta_2}, \ m_{33} \right), \label{umass}\] where \(m_{ij} = (M)_{ij}\) and \(\overline{m}_{ij} = \varDelta(M)_{ij}^*\). \(\varDelta(M)_{ij}\)'s are the cofactors of \(M\). Diagonalization matrix is described in terms of column vectors \(\bw_i^{(u)}\) \((i = 1,2,3)\) as \[{\cal V}_u = ( \bw_1^{(u)}, \ \bw_2^{(u)}, \ \bw_3^{(u)}),\] where \(\bw_i^{(u)}\)'s are eigenvectors of \({\cal M} {\cal M}^{\dag}\) and expressed as \[\bw_1^{(u)} = N_1^{(u)} \left( \begin{array}{c} 1 \\ u_1^{(u)} \\ v_1^{(u)} \end{array} \right), \quad \bw_2^{(u)} = N_2^{(u)} \left( \begin{array}{c} u_2^{(u)} \\ 1 \\ v_2^{(u)} \end{array} \right), \quad \bw_3^{(u)} = N_3^{(u)} \left( \begin{array}{c} u_3^{(u)} \\ v_3^{(u)} \\ 1 \end{array} \right)\] with \[\begin{array}{ll} u_1^{(u)} \simeq x^{\alpha_1-\alpha_2} \ \displaystyle\frac{\overline{m}_{21}}{\overline{m}_{11}}, \qquad \qquad & v_1^{(u)} \simeq x^{\alpha_1} \ \displaystyle\frac{\overline{m}_{31}}{\overline{m}_{11}}, \\[4mm] u_2^{(u)} \simeq-(u_1^{(u)})^*, \qquad \qquad & v_2^{(u)} \simeq-x^{\alpha_2} \ \displaystyle\frac{m_{23}^*}{m_{33}^*}, \\[4mm] u_3^{(u)} \simeq x^{\alpha_1} \ \displaystyle\frac{m_{13}}{m_{33}}, \qquad \qquad & v_3^{(u)} \simeq-(v_2^{(u)})^* \end{array}\] and \[\qquad N_i^{(u)} = \left( 1 + \left| u_i^{(u)} \right|^2 + \left| v_i^{(u)} \right|^2 \right)^{-1/2}, \quad (i = 1, \ 2, \ 3).\] # The CKM matrix For comparison with the case of the lepton sector, we study the mass matrix of down-type quarks in this section. At energies below the scale \(\langle N_0^c \rangle\) there appear mixings between \(D^c\) and \(g^c\) which are \(SU(2)_L\)-singlets.Effective Yukawa interactions among down-type colored fields are of the forms \[W_D = \sum_{i,j=1}^{3} \, \left[ {\cal Z}_{ij} \ g_i g^c_j S_0 + {\cal M}_{ij} \left( g_i D^c_j N^c_0 + Q_i D^c_j H_{d0} \right) \right],\] where \({\cal Z} = f_Z \, \varGamma_1 Z \varGamma_1\) and \(\det Z = 1\). As explained in the previous section, there is no hierarchical structure in \(Z\). The mass matrix of down-type colored fields is given by the \(6 \times 6\) matrix \[\begin{array}{r@{}l} \vphantom{\bigg(} & \begin{array}{ccc} \quad \ g^c & \quad \ D^c & \end{array} \\ \widehat{{\cal M}}_d = \begin{array}{l} g \\ D \\ \end{array} & \left( \begin{array}{cc} \rho_S {\cal Z} & \rho_N {\cal M} \\ 0 & \rho_d {\cal M} \end{array} \right) \end{array} \label{Md}\] in unit of \(M_S\), where \(\rho_S = \langle S_0 \rangle /M_S\), \(\rho_N = \langle N^c_0 \rangle /M_S\) and \(\rho_d = \langle H_{d0} \rangle /M_S = v_{d0} /M_S\). The above mass matrix \(\widehat{{\cal M}}_d\) is diagonalized via bi-unitary transformation as \[\widehat{\cal V}_d^{-1} \widehat{{\cal M}}_d \, \widehat{\cal U}_d = {\rm diag}(\varLambda_d^{(0)}, \ \epsilon_d \, \varLambda_d^{(2)}),\] where \(\epsilon_d = \rho_d / \rho_N = v_{d0} / \langle N^c_0 \rangle = {\cal O}(10^{-15})\). To solve the eigenvalue problem, it is convenient to take \(\widehat{{\cal M}}_d \widehat{{\cal M}}_d^\dag\) and express it as \[\widehat{{\cal M}}_d \widehat{{\cal M}}_d^{\dag} = \left( \begin{array}{cc} A_d + B_d & \epsilon_d B_d \\ \epsilon_d B_d & \epsilon_d^2 B_d \end{array} \right),\] where \(A_d = |\rho_S|^2 \, {\cal Z} {\cal Z}^{\dag}\) and \(B_d = |\rho_N|^2 \, {\cal M} {\cal M}^{\dag}\). Since \(\epsilon_d\) is a very small number, we can carry out our calculation by using perturbative \(\epsilon_d\)-expansion. Among six eigenvalues three of them represent heavy modes with the GUT scale masses. The remaining three, corresponding to down-type quarks, are derived from diagonalization of the \(\epsilon_d^2 \, (A_d^{-1} + B_d^{-1})^{-1}\), namely \[\epsilon_d^2 \, {\cal V}_d^{-1} (A_d^{-1} + B_d^{-1})^{-1} {\cal V}_d = (\epsilon_d \, \varLambda_d^{(2)})^2.\] Down-type quark masses are given by \(M_S \, \epsilon_d \, \varLambda_d^{(2)} = {\rm diag}(m_d, \ m_s, \ m_b)\). Explicit forms are \[(m_d, \ m_s, \ m_b) \, \simeq \, v_{d0} \, f_M \, x^{\beta_1} \times \left( \frac{1}{\sqrt{a}} \, x^{\alpha_1}, \ \sqrt{\frac{a}{b}} \, x^{\alpha_2}, \ \sqrt{\frac{b}{c}} \right), \label{dmass}\] where \[\begin{aligned} a & = & b \ = \ \label{qadc} c & = & & & \end{aligned}\] Here we use the notations \(z_{ij} = (Z)_{ij}\), \(\overline{z}_{ij} = \varDelta(Z)_{ij}^*\) and \[\begin{aligned} & & {\bm}_i = ( m_{1i}, \ m_{2i}, \ m_{3i})^T, \qquad \ \overline{\bm}_i = ( \overline{m}_{1i}, \ \overline{m}_{2i}, \ \overline{m}_{3i})^T, \nonumber \\ & & {\bz}_i = ( z_{1i}, \ z_{2i}, \ z_{3i})^T, \qquad \qquad \ \overline{\bz}_i = ( \overline{z}_{1i}, \ \overline{z}_{2i}, \ \overline{z}_{3i})^T. \nonumber \end{aligned}\] Diagonalization matrix \({\cal V}_d\) is of the form \[{\cal V}_d = ( \bw_1^{(d)}, \ \bw_2^{(d)}, \ \bw_3^{(d)}),\] \[\bw_1^{(d)} = N_1^{(d)} \left( \begin{array}{c} 1 \\ u_1^{(d)} \\ v_1^{(d)} \end{array} \right), \quad \bw_2^{(d)} = N_2^{(d)} \left( \begin{array}{c} u_2^{(d)} \\ 1 \\ v_2^{(d)} \end{array} \right), \quad \bw_3^{(d)} = N_3^{(d)} \left( \begin{array}{c} u_3^{(d)} \\ v_3^{(d)} \\ 1 \end{array} \right),\] where \[\begin{array}{ll} u_1^{(d)} \simeq x^{\alpha_1-\alpha_2} \ \displaystyle\frac { { & v_1^{(d)} \simeq x^{\alpha_1} \ \displaystyle\frac { { u_2^{(d)} \simeq-(u_1^{(d)})^*, & v_2^{(d)} \simeq-x^{\alpha_2} \ \displaystyle\frac{(\overline{\bz}_1 \times \overline{\bm}_1)_2} {(\overline{\bz}_1 \times \overline{\bm}_1)_3}, \\[4mm] u_3^{(d)} \simeq x^{\alpha_1} \ \displaystyle\frac{(\overline{\bz}_1 \times \overline{\bm}_1)_1^*} {(\overline{\bz}_1 \times \overline{\bm}_1)_3^*}, & v_3^{(d)} \simeq-(v_2^{(d)})^* \end{array}\] and \[\qquad N_i^{(d)} = \left( 1 + \left| u_i^{(d)} \right|^2 + \left| v_i^{(d)} \right|^2 \right)^{-1/2}, \quad (i = 1, \ 2, \ 3).\] We are now in a position to calculate the CKM mixing matrix as \[V_{\rm CKM} = {\cal V}_u^{-1} \,{\cal V}_d = {\cal V}_u^{\dag} \,{\cal V}_d. \label{VCKM}\] Thus \[(V_{\rm CKM})_{ij} = {\bw}_i^{(u)*} \cdot {\bw}_j^{(d)}.\] More explicitly, we have \[\begin{aligned} V_{us} = (V_{\rm CKM})_{12} & \simeq & x^{\alpha_1-\alpha_2} \ \frac{ {\overline{m}_{11}^* \, a}, \nonumber \\[-4mm] V_{cb} = (V_{\rm CKM})_{23} & \simeq & x^{\alpha_2} \ \frac{\overline{m}_{11}^* \, ({\bm}_3 \cdot \overline{\bz}_1^*)} {m_{33} \, (\overline{\bz}_1 \times \overline{\bm}_1)_3^* }, \nonumber \\[-2mm] \label{ckm} V_{cd} = (V_{\rm CKM})_{21} & \simeq &-(V_{us})^*, \\[2mm] V_{ts} = (V_{\rm CKM})_{32} & \simeq &-(V_{cb})^*, \nonumber \\[2mm] V_{td} = (V_{\rm CKM})_{31} & \simeq & (V_{us} \, V_{cb})^*, \nonumber \\[-2mm] V_{ub} = (V_{\rm CKM})_{13} & \simeq & x^{3\alpha_1-2\alpha_2} \ \frac{ {\overline{m}_{11}^* \, \left| (\overline{\bz}_1 \times \overline{\bm}_1)_3 \right|^2}. \nonumber \end{aligned}\] All off-diagonal elements of the \(V_{\rm CKM}\) contain the F--N factors. This means that there is little difference between the diagonalization matrices for up-type quarks and for down-type quarks in \(SU(2)_L\)-doublets. Further, \(V_{ub}\) is zero in the leading order but non-zero in the next-to-leading order, which implies that the element \(V_{ub}\) is naturally suppressed compared to \(V_{td}\). # Charged lepton mass matrix Effective Yukawa interactions among charged leptons are described as \[W_E = \sum_{i,j=1}^{3} \, \left[ {\cal H}_{ij} H_{di} H_{uj} S_0 + {\cal M}_{ij} \left( L_i H_{uj} N^c_0 + L_i E^c_j H_{d0} \right) \right],\] where \({\cal H} = f_H \, \varGamma_2 H \varGamma_2\) with \(\det H = 1\). The matrix \(H\) has no hierarchical structure. In the lepton sector, mixings occur between \(L\) and \(H_d\) which are \(SU(2)_L\)-doublets. Consequently, the charged lepton mass matrix is expressed in terms of the \(6 \times 6\) matrix \[\begin{array}{r@{}l} \vphantom{\bigg(} & \begin{array}{ccc} \quad \ H_u^+ & \quad E^{c+} & \end{array} \\ \widehat{{\cal M}}_l = \begin{array}{l} H_d^- \\ L^- \\ \end{array} & \left( \begin{array}{cc} \rho_S \, {\cal H} & 0 \\ \rho_N {\cal M} & \rho_d {\cal M} \end{array} \right) \end{array}\] in unit of \(M_S\). The study of the charged lepton mass matrix is parallel to that of the down-type quark mass matrix in the previous section. The matrix \(\widehat{{\cal M}}_l\) is diagonalized via bi-unitary transformation as \[\widehat{\cal V}_l^{-1} \widehat{{\cal M}}_l \, \widehat{\cal U}_l = {\rm diag}(\varLambda_l^{(0)}, \ \epsilon_d \, \varLambda_l^{(2)}).\] Among six eigenvalues three of them represent heavy modes with the GUT scale masses. The remaining three, corresponding to charged leptons, are derived from the diagonalization \[\epsilon_d^2 \, {\cal V}_l^{-1} (A_l^{-1} + B_l^{-1})^{-1} {\cal V}_l = (\epsilon_d \, \varLambda_l^{(2)})^2,\] where \(A_l = |\rho_S|^2 \, {\cal H}^{\dag} {\cal H}\) and \(B_l = |\rho_N|^2 \, {\cal M}^{\dag} {\cal M}\). Charged lepton masses are given by \(M_S \, \epsilon_d \, \varLambda_l^{(2)} = {\rm diag}(m_e, \ m_{\mu}, \ m_{\tau})\). Explicit forms are \[(m_e, \ m_{\mu}, \ m_{\tau}) \, \simeq \, v_{d0} \, f_M \, x^{\alpha_1} \times \left( \frac{1}{\sqrt{a'}} \, x^{\beta_1}, \ \sqrt{\frac{a'}{b'}} \, x^{\beta_2}, \ \sqrt{\frac{b'}{c'}} \right), \label{chlma}\] where \[\begin{aligned} a' & = & b' \ = \ \label{ladc} c' & = & |({\bh}_3^{'*} \cdot \overline{\bm'}_1)|^2 \, + \, |({\bm}_3^{'*} \cdot \overline{\bh}'_1)|^2, \\[-1mm] & & \, = \, \end{aligned}\] Here we use the notations \(\overline{m'}_{ij} = \varDelta(M)_{ji} = \overline{m}_{ji}^*\), \(h'_{ij} = (H^{\dag})_{ij}\), \(\overline{h'}_{ij} = \varDelta(H)_{ji}\) and \[\begin{aligned} & & {\bm}'_i = ( m_{i1}, \ m_{i2}, \ m_{i3})^{\dag}, \qquad \ \overline{\bm'}_i = ( \overline{m}_{i1}, \ \overline{m}_{i2}, \ \overline{m}_{i3})^{\dag}, \nonumber \\ & & {\bh}'_i = ( h'_{1i}, \ h'_{2i}, \ h'_{3i})^T, \qquad \qquad \ \overline{\bh'}_i = ( \overline{h'}_{1i}, \ \overline{h'}_{2i}, \overline{h'}_{3i})^T. \nonumber \end{aligned}\] The diagonalization matrix is of the form \[{\cal V}_l = ( {\bw}_1^{(l)}, \ {\bw}_2^{(l)}, \ {\bw}_3^{(l)}), \label{Vchal}\] \[{\bw}_1^{(l)} = N_1^{(l)} \left( \begin{array}{c} 1 \\ u_1^{(l)} \\ v_1^{(l)} \end{array} \right), \quad {\bw}_2^{(l)} = N_2^{(l)} \left( \begin{array}{c} u_2^{(l)} \\ 1 \\ v_2^{(l)} \end{array} \right), \quad {\bw}_3^{(l)} = N_3^{(l)} \left( \begin{array}{c} u_3^{(l)} \\ v_3^{(l)} \\ 1 \end{array} \right).\] Each element is given by \[\begin{array}{ll} u_1^{(l)} \simeq \sigma_1 \ x^{\beta_1-\beta_2}, \qquad \qquad \ & v_1^{(l)} \simeq \sigma_2 \ x^{\beta_1}, \\[2mm] u_2^{(l)} \simeq-(u_1^{(l)})^*, & v_2^{(l)} \simeq \sigma_4 \ x^{\beta_2}, \\[2mm] u_3^{(l)} \simeq \sigma_3^* \ x^{\beta_1}, & v_3^{(l)} \simeq-(v_2^{(l)})^* \end{array}\] and \[\qquad N_i^{(l)} = \left( 1 + \left| u_i^{(l)} \right|^2 + \left| v_i^{(l)} \right|^2 \right)^{-1/2} \quad (i = 1, \ 2, \ 3),\] where \[\begin{array}{ll} \sigma_1 = \displaystyle\frac { { \sigma_2 = \displaystyle\frac { { \sigma_3 = \displaystyle\frac{(\overline{\bh'}_1 \times \overline{\bm'}_1)_1} {(\overline{\bh'}_1 \times \overline{\bm'}_1)_3}, & \sigma_4 =-\displaystyle\frac{(\overline{\bh'}_1 \times \overline{\bm'}_1)_2} {(\overline{\bh'}_1 \times \overline{\bm'}_1)_3}. \end{array}\] # Neutrino mass matrix In the neutral lepton sector we have the \(15 \times 15\) mass matrix \[\begin{array}{r@{}l} \vphantom{\bigg(} & \begin{array}{cccccc} \quad \, H_u^0 & \quad \ H_d^0 & \quad \ L^0 & \quad \ \ N^c & \quad \ S & \end{array} \\ \widehat{{\cal M}}_{NS} = \begin{array}{l} H_u^0 \\ H_d^0 \\ L^0 \\ N^c \\ S \\ \end{array} & \left( \begin{array}{ccccc} 0 & \rho_S \, {\cal H} & \rho_N {\cal M}^T & 0 & \rho_d {\cal M}^T \\ \rho_S \, {\cal H} & 0 & 0 & 0 & \rho_u {\cal M}^T \\ \rho_N {\cal M} & 0 & 0 & \rho_u {\cal M} & 0 \\ 0 & 0 & \rho_u {\cal M}^T & {\cal N} & 0 \\ \rho_d {\cal M} & \rho_u {\cal M} & 0 & 0 & {\cal S} \\ \end{array} \right), \end{array} \label{NM}\] where \(\rho_u = \langle H_{u0} \rangle /M_S = v_{u0} /M_S\) and \({\cal N}, \ {\cal S}\) stand for Majorana mass matrices for the superfields \(N^c\) and \(S\) with odd R-parity. This mass matrix comes from the effective Yukawa interactions \[\begin{aligned} W_{NS} & = & \sum_{i,j=1}^{3} \, {\cal H}_{ij} H_{di} H_{uj} S_0 + \ \sum_{i,j=1}^{3} \, {\cal M}_{ij} \left( L_i H_{uj} N^c_0 + L_i N^c_j H_{u0} \right) \nonumber \\[-5mm] & & \qquad \qquad + \ \sum_{i,j=1}^{3} \, {\cal M}_{ij} \left( S_i H_{uj} H_{d0} + S_i H_{dj} H_{u0} \right) \end{aligned}\] and from Majorana mass terms for \(N^c\) and \(S\). Here we assume (the electroweak scale) \(\ll\) (\({\cal N}\) scale) \(\ll\) (\({\cal S}\) scale). The matrix \({\cal N}\) has the form \[{\cal N} = f_N \, \varGamma_2 N \varGamma_2,\] in which \(N\) contains no hierarchical structure and \(\det N = 1\). Mixings in the lepton sector occur between \(SU(2)_L\)-doublet fields \(L\) and \(H_d\). When we diagonalize the charged lepton mass matrix, the neutral leptons in \(SU(2)_L\)-doublet undergo the same transformation as the diagonalization matrix for charged leptons. In addition, the seesaw mechanism is at work. Hence, neutrino mass matrix for light modes becomes \[{\cal M}_{\nu} = M_S \, \epsilon_u^2 \, \varLambda_l^{(2)} \, {\cal V}_l^{\dag} \, {\cal N}^{-1} \, {\cal V}_l^* \, \varLambda_l^{(2)} \label{Mn}\] in the diagonal mass basis for charged leptons, where \(\epsilon_u = \rho_u / \rho_N = v_{u0} / \langle N^c_0 \rangle = {\cal O}(10^{-15})\). By diagonalizing \({\cal M}_{\nu}\) we obtain neutrino masses \[{\cal V}_{\nu}^T \, {\cal M}_{\nu} \, {\cal V}_{\nu} = {\rm diag}(m_{\nu_1}, \ m_{\nu_2}, \ m_{\nu_3})\] and the MNS matrix \[V_{\rm MNS} = {\cal V}_{\nu}^T.\] The matrix \({\cal M}_{\nu}\) is rewritten as \[{\cal M}_{\nu} = \frac{M_S \, \epsilon_u^2}{f_N} \ Y^T \, N^{-1} \, Y,\] where \[Y = \varGamma_2^{-1} \, {\cal V}_l^* \, \varLambda_l^{(2)}.\] As seen in Eq.([\[Vchal\]](#Vchal){reference-type="ref" reference="Vchal"}), \({\cal V}_l\) is of the form \[{\cal V}_l \simeq \left( \begin{array}{ccc} 1 &-\sigma_1^* \,x^{\beta_1-\beta_2} & \sigma_3^* \,x^{\beta_1} \\ \sigma_1 \,x^{\beta_1-\beta_2} & 1 &-\sigma_4^* \,x^{\beta_2} \\ \sigma_2 \,x^{\beta_1} & \sigma_4 \,x^{\beta_2} & 1 \end{array} \right).\] In addition, from Eq.([\[chlma\]](#chlma){reference-type="ref" reference="chlma"}) the matrix \(\varLambda_l^{(2)}\) is described as \[\varLambda_l^{(2)} \propto \varGamma_2 \times {\rm diag}\left( \frac{1}{\sqrt{a'}}, \ \ \sqrt{\frac{a'}{b'}}, \ \ \sqrt{\frac{b'}{c'}}\right).\] Hence, \(Y = ( {\by}_1, \ {\by}_2, \ {\by}_3)\) is given by \[\begin{aligned} {\by}_1 & \simeq & \rho_N \, f_M \ x^{\alpha_1} \, \frac{1}{\sqrt{a'}} \, \left( \begin{array}{c} 1 \\ \sigma_1^* \,x^{2(\beta_1-\beta_2)} \\ \sigma_2^* \,x^{2\beta_1} \end{array} \right), \nonumber \\ {\by}_2 & \simeq & \rho_N \, f_M \ x^{\alpha_1} \, \sqrt{\frac{a'}{b'}} \, \left( \begin{array}{c} -\sigma_1 \\ 1 \\ \sigma_4^* \,x^{2\beta_2} \end{array} \right), \\ {\by}_3 & \simeq & \rho_N \, f_M \ x^{\alpha_1} \, \sqrt{\frac{b'}{c'}} \, \left( \begin{array}{c} \sigma_3 \\ -\sigma_4 \\ 1 \end{array} \right), \nonumber \end{aligned}\] where \(\sigma_i\)'s \((i = 1 \sim 4)\) are \({\cal O}(1)\). We note that in the upper triangular elements of \(Y\) the F--N factors \(x^{\beta_1}\) and \(x^{\beta_2}\) cancel out. Introducing the notation \[\varLambda_{\kappa} = {\rm diag}\left( \sqrt{\frac{c'}{a' \,b'}}, \ \ \frac{\sqrt{c' \, a'}}{b'}, \ \ 1 \right) = {\rm diag}\left( \kappa_1, \ \kappa_2, \ 1 \right), \label{Lk}\] we have \[{\cal M}_{\nu} = \frac{v_{u0}^2 \, f_M^2}{M_N} \, x^{2 \alpha_1} \, \frac{b'}{c'} \times \varLambda_{\kappa} \, \varSigma \, \varLambda_{\kappa} \label{MNe}\] with \(M_N = f_N \, M_S\), which represents the typical scale of the R-handed neutrino Majorana masses. In the present model the F--N factors \(x^{2 \beta_1}, \ x^{2 \beta_2}\) and \(x^{2 (\beta_1-\beta_2)}\) are sufficiently small compared to 1. Therefore, \(\varSigma\) is approximated as \[\varSigma \simeq \left( \begin{array}{ccc} 1 & 0 & 0 \\ -\sigma_1 & 1 & 0 \\ \sigma_3 & -\sigma_4 & 1 \end{array} \right) \times N^{-1} \times \left( \begin{array}{ccc} 1 & -\sigma_1 & \sigma_3 \\ 0 & 1 & -\sigma_4 \\ 0 & 0 & 1 \end{array} \right). \label{Pi}\] # Numerical analysis In a simple setting \(N = {\bf 1}\) and \(\varLambda_{\kappa} = {\bf 1}\), the matrix \(V_{\rm MNS}\) is nothing but diagonalization matrix of \(\varSigma\). Further, when we put \(\sigma_1 = \sigma_4 = 2.2\) and \(\sigma_3 = 1.2\) in Eq.([\[Pi\]](#Pi){reference-type="ref" reference="Pi"}), the eigenvalues of \(\varSigma\) turn out to be 0.039, 2.122 and 11.959. It follows that the neutrino mass ratio is \[\left( \frac{\varDelta m_{32}^2}{\varDelta m_{21}^2} \right)^{1/2} = 5.57\] for the normal hierarchy, which is consistent with experimental data: \(5.56 \pm 0.20\). The observed absolute value of \(\varDelta m_{32}^2 \simeq m_{\nu_3}^2\) is obtained by taking \(M_N = {\cal O}(10^9){\rm GeV}\). The MNS matrix is estimated as \[V_{\rm MNS} = \left( \begin{array}{ccc} \ 0.865 \ & \ 0.471 \ & \ 0.172 \\ \ -0.455 \ & \ 0.594 \ & \ 0.663 \\ \ 0.211 \ & \-0.652 \ & \ 0.728 \end{array} \right).\] Experimental data provided that \(\delta_{CP} = 0\) show \[V_{\rm MNS} = \left( \begin{array}{ccc} \ 0.824 \ & \ 0.547 \ & \ 0.145 \\ \ -0.499 \ & \ 0.583 \ & \ 0.641 \\ \ 0.266 \ & \ -0.601 \ & \ 0.754 \end{array} \right).\] Although our setting \(N = {\bf 1}\) and \(\varLambda_{\kappa} = {\bf 1}\) is very simple, all elements offered here are consistent with the observed values within 20% including \(|(V_{\rm MNS})_{13}| = |\sin \theta_{13}|\). Finally we discuss fermion mass spectra and the CKM matrix. The present setting \(\varLambda_{\kappa} = {\bf 1} \ (\kappa_1 = \kappa_2 = 1)\) means \[m_e: m_{\mu}: m_{\tau} = x^{\beta_1}: x^{\beta_2}: 1. \label{chlm}\] From experimental values of masses we have \(x^{\beta_1} = \lambda^{5.6}\) and \(x^{\beta_2} = \lambda^{1.9}\) with \(\lambda = 0.23\). As to the other parameters, we set \(x^{\alpha_1} = \lambda^{3.3}\), \(x^{\alpha_2} = \lambda^{2.3}\) for the F--N factor \(\varGamma_1\) and \(\begin{equation} f_Z = \frac{\rho_N}{\rho_S} \, f_M \, \lambda^{2.5}, \qquad f_H = \frac{\rho_N}{\rho_S} \, f_M \, \lambda^{2.7}, \end{equation} respectively. In this setting of parameter values, each first term in Eq.(\ref{qadc}) for\)a\(and\)c\(, and Eq.(\ref{ladc}) for\)a'\(and\)c'\(is dominant. From Eqs.(\ref{umass}), (\ref{dmass}) and (\ref{ckm}), we have \begin{equation} \frac{m_s}{m_u} \, V_{us} \simeq \frac{v_{d0}}{v_{u0}}, \end{equation} which is independent of values of\)\_i\(and\)\_i\[(i = 1,  2)\(. The observed value of the left hand side leads to\)(v\_d0/v\_u0) = \^-1.6\(, which is reverse to the usual solution with large\)= v_u/v_d\(. This fact suggests that Higgs fields other than\)H\_u0\(and\)H\_d0\(develop their VEV's. It is expected that there exist rich spectra of Higgs fields beyond those of the MSSM at the TeV scale. We take\)(v\_d0/v\_u0) = \^-1.6\(as an input of the present analysis. In the present setting most of fermion mass hierarchies are attributed to the F--N factors. In order to reproduce observed mass spectra and the CKM matrix precisely, we adjust the other parameters as \begin{eqnarray} & & m_{33} = \lambda^{0.7}, \qquad z_{33} = \lambda^{-0.3}, \nonumber \\ & & \overline{m}_{11} = \lambda^{0.5}, \qquad \overline{z}_{11} = \lambda^{-0.3}, \qquad \overline{h'}_{11} = \lambda^{2.9}, \nonumber \\ & & |(\overline{\bz}_1 \times \overline{\bm}_1)_3| = \lambda^{0.2}, \qquad |({\bz}_3^* \cdot \overline{\bm}_1)| = \lambda^{0.2}, \qquad |({\bm}_3 \cdot \overline{\bz}_1^*)| = \lambda^{0.3}, \\ & & |(\overline{\bh'}_1 \times \overline{\bm'}_1)_3| = \lambda^{0.8}, \qquad |({\bh'}_3^* \cdot \overline{\bm'}_1)| = \lambda^{0}. \nonumber \end{eqnarray} As seen in Tables 1 and 2, our results are in good agreement with observed values. Although there are many parameters, it is noteworthy that these parameters other than\)\_11\(are\)O(1)\(. Note that this parameter\)\_11\(stands for the cofactor\)(H)\_11\(. \begin{table}[t] \begin{center} \caption{Quark and lepton masses divided by top quark mass (\)\lambda = 0.23\()} \vspace{2mm} \begin{tabular}{|c|c|c|} \hline \vphantom{\LARGE I} mass ratio & \quad our result \quad & observed values \\ \hline \vphantom{\LARGE I}\)m_c/m_t\(&\)\^3.3\(&\)\^3.34\(\\ \vphantom{\LARGE I}\)m_u/m_t\(&\)\^7.7\(&\)\^7.65\(\\ \hline \vphantom{\LARGE I}\)m_b/m_t\(&\)\^2.5\(&\)\^2.54\(\\ \vphantom{\LARGE I}\)m_s/m_t\(&\)\^5.1\(&\)\^5.11\(\\ \vphantom{\LARGE I}\)m_d/m_t\(&\)\^7.1\(&\)\^7.14\(\\ \hline \vphantom{\LARGE I}\)m\_/m_t\(&\)\^3.1\(&\)\^3.12\(\\ \vphantom{\LARGE I}\)m\_/m_t\(&\)\^5.0\(&\)\^5.04\(\\ \vphantom{\LARGE I}\)m_e/m_t\(&\)\^8.7\(&\)\^8.67\(\\ \hline \end{tabular} \end{center} \end{table} \vspace{3mm} \begin{table}[t] \begin{center} \caption{Elements of the CKM matrix} \vspace{2mm} \begin{tabular}{|c|c|c|} \hline \vphantom{\LARGE I}\)(V\_CKM)\_ij\(& \quad our result \quad & observed values \\ \hline \vphantom{\LARGE I}\)V\_us\(&\)\^1.0\(&\)\^1.02\(\\ \vphantom{\LARGE I}\)V\_cb\(&\)\^2.2\(&\)\^2.17\(\\ \vphantom{\LARGE I}\)V\_td\(&\)\^3.2\(&\)\^3.23\(\\ \vphantom{\LARGE I}\)V\_ub\(&\)\^3.9\(&\)\^3.87\(\\ \hline \end{tabular} \end{center} \end{table} We may change the above values of the parameters\)\_i\(and\)\_i\](i = 1,  2)\(. As another setting of the F--N factors\)\_1\(and\)\_2\(, we put \begin{equation} x^{\alpha_1} = \lambda^{4.6}, \qquad x^{\alpha_2} = \lambda^{2.6}, \qquad x^{\beta_1} = \lambda^{5.6}, \qquad x^{\beta_2} = \lambda^{1.9}. \end{equation} In this case the relation \begin{equation} m_d: m_s: m_b = x^{\alpha_1}: x^{\alpha_2}: 1 \end{equation} holds together with Eq.(\ref{chlm}). Further, we take \begin{equation} f_Z = \frac{\rho_N}{\rho_S} \, f_M \, \lambda^{1.6}, \qquad f_H = \frac{\rho_N}{\rho_S} \, f_M \, \lambda^{2.5}, \end{equation} corresponding to\) \[\begin{aligned} & & m_{33} = \lambda^{1.23}, \qquad z_{33} = \lambda^{0.03}, \nonumber \\ & & \overline{m}_{11} = \lambda^{1.27}, \qquad \overline{z}_{11} = \lambda^{0.87}, \qquad \overline{h'}_{11} = \lambda^{2.17}, \nonumber \\ & & |(\overline{\bz}_1 \times \overline{\bm}_1)_3| = \lambda^{1.13}, \qquad |({\bz}_3^* \cdot \overline{\bm}_1)| = \lambda^{0}, \qquad |({\bm}_3 \cdot \overline{\bz}_1^*)| = \lambda^{0.7}, \\ & & |(\overline{\bh'}_1 \times \overline{\bm'}_1)_3| = \lambda^{0.83}, \qquad |({\bh'}_3^* \cdot \overline{\bm'}_1)| = \lambda^{-0.7}. \nonumber \end{aligned}\] In this choice of the parameter values our results remain the same as given in Tables 1 and 2. # Summary We have studied flavor mixings, especially the MNS matrix, in detail in the \(SU(6) \times SU(2)_R\) model, in which the F--N mechanism plays an important role. Due to the F--N mechanism both effective Yukawa couplings and R-handed neutrino Majorana mass matrix have hierarchical structure, which is described in terms of the F--N factors. In this model the \(D^c\)--\(g^c\) and \(L\)--\(H_d\) mixings as well as generation mixings occur and affect both fermion mass spectra and flavor mixings. In the \(D^c\)--\(g^c\) mixings, since \(D^c\) and \(g^c\) are both \(SU(2)_L\)-singlets, the disparity between the diagonalization matrices for up-type quarks and down-type quarks in \(SU(2)_L\)-doublets is rather small. Accordingly, \(V_{\rm CKM}\) exhibits small mixing. On the other hand, in the \(L\)--\(H_d\) mixings, since \(L\) and \(H_d\) are both \(SU(2)_L\)-doublets, there appears no disparity between the diagonalization matrices for charged leptons and neutrinos unless the seesaw mechanism does not take place. As a matter of fact, however, the seesaw mechanism is at work and an additional transformation is required to diagonalize the neutrino mass matrix. This additional transformation matrix yields nontrivial \(V_{\rm MNS}\). In the present model the neutrino mass matrix has characteristic structure as seen in Eqs.([\[MNe\]](#MNe){reference-type="ref" reference="MNe"}) and ([\[Pi\]](#Pi){reference-type="ref" reference="Pi"}). This is due to the fact that the F--N factors cancel out in the upper triangular elements of \(Y\). As a consequence, there is no hierarchical structure in \({\cal M}_{\nu}\) and eventually \(V_{\rm MNS}\) exhibits large mixing. The characteristic structure of fermion spectra is attributed to the hierarchical effective Yukawa couplings due to the F--N mechanism and also to the \(D^c\)--\(g^c\) and \(L\)--\(H_d\) mixings. In particular, the difference of mass hierarchies among up-type quarks, down-type quarks and charged leptons has its origin in \(D^c\)--\(g^c\) and \(L\)--\(H_d\) mixings. In the neutrino sector we have to incorporate the Majorana mass hierarchy of R-handed neutrinos with the seesaw mechanism. Numerical results are consistent with all observed values of fermion masses and mixings. The present model provides a unified description of mass spectra and flavor mixings. In order to determine the F--N factors and the magnitude of \(f_i\)'s \((i = M, \ Z, \ H, \ N)\) theoretically, we need an appropriate flavor symmetry and also the flavor charge assignment to matter fields. In the previous works we made an attempt to find several solutions. The detailed study of this issue is the subject of future works. Further, in this paper we ignored the phase factors of VEV's for matter fields and assumed \(\delta_{CP} = 0\). The study of the CP-violation will be carried out elsewhere.

{'timestamp': '2012-10-16T02:02:51', 'yymm': '1210', 'arxiv_id': '1210.3763', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3763'}

# Introduction AdS/CFT belongs to a class of dualities in which the dualization procedure is not explicitly introduced. The reason for that is simple: implementing the procedure is subtle and difficult. Evidently, however, understanding the procedure must be essential for the first principle derivation of AdS/CFT. There has been progress in this direction (see, e.g., and refs therein), and one of the goals in this work is to further that progress. While doing so, we report on two conceptual/technical advances: the first is a new Kaluza-Klein reduction procedure based on ADM decomposition and the other is the IIB realization of a braneworld. One of the distinctions of AdS/CFT type dualities is that the dualization and inverse dualization seem very different at low energy field theory level. We refer to the procedure in which one gets the closed string/gravity degrees of freedom from open string/gauge theory as *forward dualization*. The *inverse dualization* refers to the reverse procedure. It was proposed in that it should be the quantum/strong coupling effects that must be behind the forward dualization. As commented in (footnote 15), the inverse dualization must be initiated by a spontaneous symmetry breaking of the supergravity system. (Here we are using the term 'spontaneous symmetry breaking' in a general sense that is associated with expanding an action around a solution.[^1]) In the context of \((A)dS/CFT\) type dualities, it is natural to view the \((A)dS_5\) as a foliation of \((A)dS_4\) along a direction of \((A)dS_5\). (See the figure.) One of the leaves can serve as our braneworld with a certain gravity localization mechanism that we discuss later. We focus on the \(dS\) case henceforth whenever possible. Figure 1 depicts \(dS_5\) as foliation of \(dS_4\) hypersurfaces. One may describe the \(dS_5\) through bulk gravity setup. On the other hand, it seems plausible to describe the bulk \(dS_5\), or at least some aspects of it, through collective dynamics of the hypersurfaces. Of course, the existence of these \"dual\" descriptions must be what is behind gauge/gravity correspondence. (The bulk dynamics would include the dynamics associated with the \"radial\" direction which is typically associated with renormalization group flow. Therefore, in general, the collective surface dynamics would not cover the entire bulk dynamics. There are various levels of equivalence that the term \"duality\" describes. Ideally, the term should be reserved only for the cases where the two theories under consideration are fully equivalent. For example, a canonical transformation can be viewed as a precise duality: it maps to a theory that is fully dual to the original theory. However, even if the full equivalence is not obvious, the term \"duality\" is often used in some string theory contexts for the cases in which the two theories capture substantial aspects each other.) What procedure could lead to the surface degrees of freedom starting from the bulk theory? As anticipated in, it should be a procedure initiated by a spontaneous symmetry breaking. It is also likely that the procedure should involve a certain dimensional reduction scheme, conventional or unconventional. Although it should be possible to deduce the hypersurface degrees of freedom through the conventional Kaluza-Klein reduction (see, e.g., for a relatively recent discussion), we will pave our way through an unconventional reduction scheme. This procedure of acquiring surface degrees of freedom should be viewed as a novel Kaluza-Klein (KK) reduction-what we call ADM reduction. What is unusual about this scheme is that the reduced lagrangian is not a gravity system: the *dynamical* fields are the worldvolume (i.e., the selected hypersurface) gauge fields. This phenomenon, unusual in the Kaluza-Klein context, must be what triggers the inverse dualization of the AdS/CFT type dualities. An essential computational ingredient for obtaining the hypersurface degrees of freedom from a spontaneous symmetry breaking was obtained in a remarkable series of papers,. The authors showed that the Hamilton-Jacobi equation of the gravity system under consideration admits a solution of the worldvolume theory form. We will apply the technique of to a specific setup of the 5D gravity that can be obtained by reducing IIB supergravity on a 5D hyperboloidal space \({\cal H}^{p,q}, p+q=5\) considered in (and also on \(S^5\)). Once the 5D (A)dS gravity is obtained by reducing IIB supergravity on a 5D hyperboloidal space \({\cal H}^5\) (\(S_5\)), a canonical transformation can be performed on the system to convert it into an equivalent, therefore dual, formulation that still takes the form of a supergravity. Following, one can show that the dual system admits, in the case of \(S^5\) reduction, a worldvolume action as solution of the Hamilton-Jacobi equation of the gravity system. Obviously, the resulting worldvolume action will capture some aspects of the original gravity theory, and can be viewed as a dual pair at least in a wider sense of the term mentioned previously. In essence, the holographic dualities must have their roots in that one may adopt two different approaches to describe the geometry. In the first approach, one can adopt the conventional degrees of freedom, the metric, to describe the bulk physics. In the other approach, one slices the bulk into a set of hypersurfaces, i.e., focus on each leaf of the foliated geometry. Gravity is not needed as *dynamical* degrees of freedom to describe the hypersurface, although it serves as a background for the gauge degrees of freedom. We will elaborate on this in the main body of this paper. Below we consider both the ADM reduction and the standard toroidal-type reduction. In the ADM reduction scheme, one employs the HJ procedure, and a worldvolume effective action will appear as a solution of the HJ equation. The ADM reduction scheme can be viewed as \"emergent gauge theory\" in the sense that a worldvolume gauge field emerges from the symmetry breaking. In the standard toroidal reduction, dependence of one of the coordinates (\"\(r\)\") will be removed. The model that we obtain below has a scalar that can be viewed as an inflaton field in four dimensions. We comment on the potential phenomenological value of our model in the main body postponing the full analysis for the near future. In the related literature, usually, an explicit coupling between gravity and various brane sources is employed followed by Calabi-Yau compactification in order to obtain a de Sitter space in the lower dimensions. One drawback of the Calabi-Yau compactification is the implicit nature of the analysis involved. Moreover, the original motivation for considering the Calabi-Yau manifold as opposed to a maximally symmetric manifold has diminished with better understanding of supersymmetry breaking effects of D-branes. The organization of the paper is as follows: in sec 2, we carry out reduction of IIB supergravity on a manifold denoted by \({\cal M}^5\) that we take either \({\cal M}^5={\cal H}^5\) or \({\cal M}^5=S^5\) (5-sphere) in the subsequent sections. For \({\cal M}={\cal H}^5\), we obtain a 5D de Sitter gravity. The ADM reduction of the 5D theory obtained thereby is carried out in sec 3. In particular, we elaborate on the appearance of the worldvolume gauge field. We discuss the implications of our results on braneworld realization and black hole information physics. In sec 4, we obtain a domain-wall solution for the 5D system obtained in sec 2 in the case \({\cal M}^5=S^5\). Keeping the minimal set of fields, the system admits an AdS vacuum solution. We comment on possibility of obtaining a dS solution with the form fields added. In another direction, we carry out toroidal reduction to 4D, and obtain an action that may have phenomenological value for inflationary physics. In the conclusion, the results are summarized and future directions are suggested. We also comment on the potential cosmological/black hole applications of our results. # Spherical/hyperboloidal reduction in Einstein frame to 5D [\[hr\]]{#hr label="hr"} Although mathematically elegant, the usual Calabi-Yau compactification has one shortcoming: the requirement of the internal manifolds to be of CY is not sufficiently restrictive, an aspect that can be seen from the fact that there exists many moduli. Starting with simple compactification such as compactification on a maximally symmetric space could be more effective. In the KKLT type compactification, one introduces and/or adds (anti)-branes to lift up the moduli. This can be viewed as a narrowing-down to special sectors of the moduli space. Therefore, this approach ultimately might not be more general than the present approach where one restricts to a certain class of special internal manifolds from the beginning. In this section, we consider Kaluza-Klein reduction on an inhomogeneous[^2] hyperboloidal space \({\cal H}^{p,q}\), a manifold considered in. The authors of showed that reduction on \({\cal H}^{p,q}\) leads to a ghost-free dS gravity in the lower dimensions. The ansatz of led to a 5D potential that has a saddle shape. In the reduction that we carry out in this section, we consider an ansatze that is an analogue of those of keeping three scalars for 5D system: the dilaton, the axion and a breathing mode from the 10D metric. Only the breathing mode generates the potential for 5D theory as we will see below.[^3] The bosonic part of type IIB supergravity action takes the following form in Einstein frame \[I= \fr{1}{2\k_{10}^2}\int d^{10}X \sqrt{-G_E}\Big[\left( R-\fr12 \pa_M \Phi \pa^M\Phi -\fr1{2\cdot 3!}e^{-\Phi}(H_{\sst{(3)}})^2 \right)\] \[-\fr12 e^{2\Phi}\pa_M C_{\sst{(0)}} \, \pa^M C_{\sst{(0)}}-\fr1{2\cdot 3!} e^{\Phi}(\tilde{F}_{\sst{(3)}})^2-\fr1{4\cdot 5!}(\tilde{G}_{\sst{(5)}})^2+\mathcal{L}_{PST} \Big]\] \[-\fr{1}{4\k_{10}^2}\int_{\mathcal{M}^{10}} C_{\sst{(4)}}\wedge H_{\sst{(3)}}\wedge F_{\sst{(3)}} \label{IIBLagEq0}\] Let us consider the following reduction ansatz[^4], \[\begin{aligned} ds_{10}^2 &=& e^{2\tilde{\rho }(\hat{x})}h_{\underline{m} \underline{n}}(\hat{x})d\hat{x}^{\underline{m}} d\hat{x}^{\underline{n}}+e^{-6\tilde{\rho}(\hat{x})/5}d\Omega_5\equiv e^{2\tilde{\rho }(\hat{x})} h_{\underline{m} \underline{n}}(\hat{x})d\hat{x}^{\underline{m}} d\hat{x}^{\underline{n}}+e^{-6\tilde{\rho}(\hat{x})/5}g_{ij}dy^i dy^j \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \Phi (X)&=& \f(\hat{x}) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ B_{\sst{(2)}} (X)&=& \fr12 B_{\underline{m} \underline{n}}(\hat{x})d\hat{x}^{\underline{m}} \wedge d\hat{x}^{\underline{n}} \equiv B_{\sst{(2)}} (\hat{x}) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ C_{\sst{(0)}} (X)&=& \chi(\hat{x}) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ C_{\sst{(2)}} (X)&=& \fr12 C_{\underline{m} \underline{n}}(\hat{x})d\hat{x}^{\underline{m}} \wedge d\hat{x}^{\underline{n}} \equiv C_{\sst{(2)}} (\hat{x}) \label{ans1q0} \end{aligned}\] and \[\begin{aligned} C_{\sst{(4)}} (X) &=& \fr{1}{4!}D_{\underline{m} \underline{n} \underline{k} \underline{l}}(\hat{x})d\hat{x}^{\underline{m}} \wedge d\hat{x}^{\underline{n}} \wedge d\hat{x}^{\underline{k}} \wedge d\hat{x}^{\underline{l}} +\fr1{4!}k E_{i_1i_2i_3i_4} dy^{i_1} \wedge dy^{i_2} \wedge dy^{i_3} \wedge dy^{i_4} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &\equiv & D_{\sst{(4)}}(\hat{x})+kE_{\sst{(4)}} (y) \label{ans2q0} \end{aligned}\] The \(y\)-coordinates describe either \({\cal H}^5\) or \(S^5\): \[\begin{aligned} {\cal M}^5= {\cal H}^5\quad \mbox{or} \;\;\;S^5 \end{aligned}\] \({\cal M}^5\) is later decided to be \(S^5\) for the discussion in sec 3. The field \(E\) satisfies \(5 \pa_{[i_1}E_{i_2i_3i_4i_5]}=(1/\sqrt{g})\epsilon_{i_1i_2i_3i_4i_5}\). (Our form conventions are summarized in Appendix A.) Substituting , into 10D equation of motion, one can show after some algebra that the reduced field equations follow from the following 5D action: \[I= \fr{1}{2\k_{5}^2}\int d^{5}\hat{x} \,\sqrt{-h}\Big[\left( R^{\sst{(5)}}-\fr{24}{5} \pa_{\underline{m}} \tilde{\rho} \pa^{\underline{m}}\tilde{\rho}-\fr12 \pa_{\underline{m}}\phi \pa^{\underline{m}}\phi -\fr1{2\cdot 3!}e^{-\phi-4\tilde{\rho}}(H_{\sst{(3)}})^2 \right)\] \[-\fr12 e^{2\phi}\pa_{\underline{m}} \chi \, \pa^{\underline{m}} \chi-\fr1{2\cdot 3!} e^{\phi-4\tilde{\rho}}(\tilde{F}_{\sst{(3)}})^2-\fr1{2\cdot 5!}e^{-8\tilde{\rho}}(\tilde{G}_{\sst{(5)}})^2+e^{16\tilde{\rho}/5}R^{{\cal M}_5} \Big]\, \label{5DLagEins}\] Above \[\tilde{F}_{\sst{(3)}}=F_{\sst{(3)}}-\chi H_{\sst{(3)}}\equiv dC_{\sst{(2)}} (\hat{x})-\chi(\hat{x}) \wedge dB_{\sst{(2)}}(\hat{x}) \.\] \[\begin{aligned} \tilde{G}_{\sst{(5)}} &=& G_{\sst{(5)}}-C_{\sst{(2)}}\wedge H_{\sst{(3)}},\quad H=\fr12 \pa_{[\underline{m}}B_{\underline{n}\underline{k}]}d\hat{x}^{\underline{m}} \wedge d\hat{x}^{\underline{n}} \wedge d\hat{x}^{\underline{k}}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ G_{\sst{(5)}} &=& \fr1{4!} \pa_{[\underline{m}_1}D_{\underline{m}_2\cdots \underline{m}_5]}d\hat{x}^{\underline{m}_1} \wedge \cdots \wedge d\hat{x}^{\underline{m}_5} \. \end{aligned}\] With rescaling \[\tilde{\rho}=-\fr14 \sqrt{\fr53}\rho \label{rho5d}\] the action becomes the following form with canonical kinetic terms: \[I= \fr{1}{2\k_{5}^2}\int d^{5}\hat{x}\, \sqrt{-h}\Big[\left( R^{\sst{(5)}}-\fr12 \pa_{\underline{m}} {\rho} \pa^{\underline{m}}{\rho}-\fr12 \pa_{\underline{m}}\phi \pa^{\underline{m}}\phi -\fr1{2\cdot 3!}e^{-\phi+\sqrt{\fr53}{\rho}}(H_{\sst{(3)}})^2 \right)\] \[-\fr12 e^{2\phi}\pa_{\underline{m}} \chi \, \pa^{\underline{m}} \chi-\fr1{2\cdot 3!} e^{\phi+\sqrt{\fr53}{\rho}}(\tilde{F}_{\sst{(3)}})^2-\fr1{2\cdot 5!}e^{2\sqrt{\fr53}{\rho}}(\tilde{G}_{\sst{(5)}})^2+e^{-\sqrt{\fr{16}{15}} \rho}R^{{\cal M}_5} \Big] \label{5DLagEins1}\] Rescaling the \(\rho\) field further in \[\rho \rightarrow \sqrt{\fr53}\rho\] we arrive at an alternative form of the action: \[\begin{aligned} I_5 &=& \fr1{2\k_5^2}\int d^5-\fr12(\pa \f)^2-\fr1{2\cdot 3!} e^{-\phi}e^{\fr53 \rho} H_{\sst{(3)}}^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && -\fr12 e^{2\phi}(\pa\chi)^2-\fr1{2\cdot 3!}e^{\phi}e^{\fr53 \rho}\tilde{F}^{2}_{\sst{(3)}}-\fr1{2\cdot 5!}e^{\fr{10}3 \rho}\tilde{G}^2_{\sst{(5)}}+e^{-\fr43\rho} R^{{\cal M}_5}\Big] \label{5dactRhoEinq} \end{aligned}\] # ADM Reduction from 5D to 4D [\[ADMred\]]{#ADMred label="ADMred"} In section 4, we will obtain a solution of a certain three brane configuration. As a matter of fact, the 5D system admits a whole class of D3-brane solutions as we will see. The first step is to obtain the Hamilton-Jacobi (HJ) equation pertaining to through a series of manipulations following the works of. The fact that a class of solutions of the HJ system of takes a form of a DBI action has deep physical implications. For example, the steps for obtaining the worldvolume form solution to the HJ equation should be viewed as a realization of a reduction scheme. The reason is that the field equations that follow from the worldvolume action can be viewed as outcome of substituting an appropriately constructed Kaluza-Klein gravity ansatze into the 5D Hamilton-Jacobi equation. Therefore the whole procedure is in the usual spirit of Kaluza-Klein reduction; hence it can legitimately be called an ADM \"reduction\" scheme. Also the resulting D3 action should be a dual description, at least in the wider sense of the term \"dual\". ## Converting to 5D \"string\" type frame [\[5Dsf\]]{#5Dsf label="5Dsf"} In the next section, we consider the HJ equation of the 5D gravity system obtained in the previous section. It turns out more convenient for the purpose at hand to cast into another frame which we call 5D \"string\" frame. To that end, let us consider \[\begin{aligned} (h_{Ein})_{\underline{m}\underline{n}}=e^{-\fr12 \f+\fr{5}{6} \rho}(h_{str})_{\underline{m}\underline{n}}\,; \end{aligned}\] With this, now takes[^5] \[\begin{aligned} I&=& \fr{1}{2\k_{5}^2}\int d^{5}\hat{x}\, \sqrt{-h}\Big[e^{-\fr34 \f+\fr{{5}}{4}\rho}\Big( R^{\sst{(5)}}-{\fr{{10}}{3}}\nabla^2\rho-\fr{35}{12} \pa_{\underline{m}} {\rho} \pa^{\underline{m}}{\rho}+2\nabla^2\f \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr54 \pa_{\underline{m}}\phi \pa^{\underline{m}}\phi+\fr52 \pa_{\underline{m}} \f \pa^{\underline{m} }\rho-\fr1{2\cdot 3!}(H_{\sst{(3)}})^2 \Big) -\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\underline{m}} \chi \, \pa^{\underline{m}} \chi +\fr1{ 3!} \tilde{F}_{\sst{(3)}}^2 +\fr1{ 5!}\tilde{G}_{\sst{(5)}}^2\bigg)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\hspace{2in}+e^{-\fr54 \f+\fr34 \rho}R^{{\cal M}_5} \Big] \label{5DLagst2q} \end{aligned}\] After partial integration, one finds \[I= \fr{1}{2\k_{5}^2}\int d^{5}\hat{x}\, \sqrt{-h}\Big[e^{-\fr34 \f+\fr{{5}}{4}\rho}\left( R^{\sst{(5)}}+\fr{5}{4} \pa_{\underline{m}} {\rho} \pa^{\underline{m}}{\rho}+\fr14 \pa_{\underline{m}}\phi \pa^{\underline{m}}\phi -\fr52 \pa_{\underline{m}}\rho \pa^{\underline{m}}\phi-\fr1{2\cdot 3!}H_{\sst{(3)}}^2 \right)\] \[-\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\underline{m}} \chi \, \pa^{\underline{m}} \chi +\fr1{ 3!} \tilde{F}_{\sst{(3)}}^2 +\fr1{ 5!}\tilde{G}_{\sst{(5)}}^2\bigg)+e^{-\fr54 \f+\fr34 \rho}R^{{\cal M}_5} \Big] \label{5DLagst2q}\] Let us split the index \(\underline{m}\): \[\begin{aligned} \underline{m}=\m,r \end{aligned}\] Carrying out ADM decomposition and adding GH boundary terms yields (see, e.g.,) \[\begin{aligned} \int\, d^5\hat{x}\,{\cal L}_{bulk+bd} &=&\int \, dr d^4 x\,\sqrt{-g}\, n \left[\fr{}{}\right. e^{-\fr34 \f+\fr{{5}}{4}\rho} \bigg(-K_{\mu\nu}^2+K^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{3}{2n}[\pa_r \phi-n^{\mu}\pa_{\mu}\phi]K +\fr{5}{2n}[\pa_r \rho-n^{\mu}\pa_{\mu}\rho]K \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&+\frac{1}{n^2}\bigg\{\frac{1}{4}[\pa_r \phi-n^{\mu}\pa_{\mu}\phi]^2 +\frac{5}{4}[\pa_r \rho-n^{\mu}\pa_{\mu}\rho]^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr52[\pa_r \phi-n^{\mu}\pa_{\mu}\phi][\pa_r \rho-n^{\n}\pa_{\n}\rho]-\frac{1}{4}[H_{r\mu\nu}-n^{\lambda}H_{\lambda\mu\nu}]^2 \bigg\}\bigg)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\frac{1}{2n^2}e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg\{[\pa_r \chi-n^{\mu}\pa_{\mu}\chi]^2 +\frac{1}{2}[\tilde{F}_{r\mu\nu}-n^{\lambda}\tilde{F}_{\lambda\mu\nu}]^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&+\frac{1}{24}[\tilde{G}_{r\mu\nu\lambda\rho}-n^{\sigma}\tilde{G}_{\sigma\mu\nu\lambda\rho}]^2 \bigg\} +{\cal L}^{\sst{(4)}} \;\;\bigg], \label{fstotq} \end{aligned}\] where \(r\) is one of the spatial coordinates and will play the role of \"time\" in the next section, and \[\begin{aligned} {\cal L}^{\sst{(4)}} &\equiv & e^{-\fr34\f+\fr54 \rho}\Big(R^{\sst{(4)}} +\fr32 \nabla_\m\nabla^\m \f-\fr52 \nabla_\m\nabla^\m \rho-\fr78\pa_\m \f \pa^\m \f-\fr{15}{8}\pa_\m \rho\pa^\m \rho+\fr54 \pa_\m \f \pa^\m \rho-\fr1{12}H_{\m\n\rho} H^{\m\n\rho}\Big) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && -\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\m} \chi \, \pa^{\m} \chi +\fr1{ 3!} \tilde{F}^{\m\n\rho}\tilde{F}_{\m\n\rho} \bigg)+e^{-\fr54 \f+\fr34 \rho}R^{{\cal M}_5} \end{aligned}\] With \(r\) playing the role of \"time\", the total \"hamiltonian\" of the system[^6] \[\begin{aligned} {\cal H}&=& \pi^{\m\n}\pa_r g_{\m\n}+\pi_{\phi}\pa_r \phi +\pi_{\rho}\pa_r \rho+\pi_B^{\mu\nu} \pa_r B_{\mu\nu} +\pi_{\chi}\pa_r \chi +\pi_C^{\mu\nu} \pa_r C_{\mu\nu} +\pi_D^{\mu\nu\lambda\rho} \pa_r D_{\mu\nu\lambda\rho}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\hspace{3in}-{\cal L}_{bulk+bd} \label{pqdq} \end{aligned}\] can be written as (see, e.g., ) \[\begin{aligned} {\cal H}\equiv nH+n_{\mu}H^{\mu}+B_{r\mu}Z_B^{\mu}+C_{r\mu}Z_C^{\mu} +D_{r\mu\nu\lambda}Z_D^{\mu\nu\lambda} \label{hamildefq} \end{aligned}\] Here \(n,\; n_{\mu}, \; B_{r\mu}\), \(C_{r\mu}\) and \(D_{r\mu\nu\lambda}\) behave like Lagrange multipliers, giving the following set of constraints \[\begin{aligned} H=0, \;\;\; H^{\mu}=0, \;\;\; Z_B^{\mu}=0, \;\;\; Z_C^{\mu}=0, \;\;\; Z_D^{\mu\nu\lambda}=0. \label{constraintsq} \end{aligned}\] One can show \[\begin{aligned} H&=& e^{\fr34 \f-\fr54 \rho}\bigg(-\pi_{\m\n}^2-\fr12\pi_\m^\m \pi_\f +\fr12\pi_\m^\m \pi_\rho-\fr12\pi_\f^2-\fr{3}{10}\pi_\rho^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \hspace{2in}-(\pi_{B\m\n}+\chi \pi_{C\m\n}+6C^{\lambda\rho}\pi_{D\m\n\lambda\rho})^2 \bigg)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-e^{-\fr54 \f-\fr54 \rho}\bigg(\fr12 \pi_\chi^2+\pi_{C\m\n}^2+12\pi_{D\m\n\lambda\rho}^2 \bigg)-{\cal L}^{\sst{(4)}} \end{aligned}\] and \[\begin{aligned} H^\m &=&-2\nabla_\n \pi^{\m\n}+\pi_\f \pa^\m \f+\pi_\rho \pa^\m \rho+\pi_{B \n\lambda}H^{\m\n\lambda} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && +\pi_{\chi}\pa^\m \chi+\pi_{C\n\lambda}F^{\m\n\lambda}+ \pi_{D\n\lambda\rho\s}(G^{\m\n\lambda\rho\s}-4C^{\m\n}H^{\lambda\rho\s}) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ Z_B^\m &=& 2\nabla_\n \pi_B^{\m\n} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ Z_C^\m &=& 2\nabla_\n \pi_C^{\m\n}+4\pi_D^{\m\n\lambda\rho}H_{\n\lambda\rho} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ Z_D^{\m\n\lambda}&=& 4 \nabla_\rho \pi_D^{\m\n\lambda\rho} \end{aligned}\] Define \(\bar{g}_{\mu\nu}(x,r)\) to be a classical solution to the field equation associated with with the following boundary condition[^7] \[\begin{aligned} \bar{g}_{\mu\nu}(x,r_0)&=&g_{\mu\nu}(x)\quad,\quad \pi^{\mu\nu}(x)=\bar{\pi}^{\mu\nu}(x,r=r_0) \label{bcq} \end{aligned}\] The boundary configurations for other fields are similarly defined. The standard procedure of the HJ formalism yields, \[\begin{aligned} &&\pi^{\mu\nu}(x)=\frac{1}{\sqrt{-g(x)}} \frac{\delta S}{\delta g_{\mu\nu}(x)},\;\;\; \label{pianddelSq} \end{aligned}\] and similarly for other fields. The HJ equation reads \[\begin{aligned} &&e^{\fr34 \f-\fr54\rho} \left[-\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}}\bigg)^2-\fr{1}{2}\frac{g_{\mu\nu}}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}} \frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \phi}-\fr1{2}\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \f}\bigg)^2 \right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \hspace{.5in} -\fr{3}{10}\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \rho}\bigg)^2 +\fr1{2}\frac{g_{\mu\nu}}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}} \frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \rho}\right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \hspace{1.5in}-\frac{1}{(\sqrt{-g})^2}\bigg(\frac{\delta S_0}{\delta B_{\mu\nu}} +\chi\frac{\delta S_0}{\delta C_{\mu\nu}} +6C_{\lambda\rho} \frac{\delta S_0}{\delta D_{\mu\nu\lambda\rho}}\bigg)^2\right] \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{ e^{-\fr54\f-\fr54 \rho}}{(\sqrt{-g})^2}\left[\bigg(\frac{1}{2}\frac{\delta S_0}{\delta \chi}\bigg)^2 +\bigg(\frac{\delta S_0}{\delta C_{\mu\nu}}\bigg)^2 +12\bigg(\frac{\delta S_0}{\delta D_{\mu\nu\lambda\rho}} \bigg)^2\right]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\Big[e^{-\fr34\f+\fr54 \rho}\Big(R^{\sst{(4)}} +\fr32 \nabla_\m\nabla^\m \f-\fr52 \nabla_\m\nabla^\m \rho-\fr78\pa_\m \f \pa^\m \f-\fr{15}{8}\pa_\m \rho\pa^\m \rho+\fr54 \pa_\m \f \pa^\m \rho \Big) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr1{12} H_{\m\n\rho}H^{\m\n\rho}-\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\m} \chi \, \pa^{\m} \chi +\fr1{ 3!} \tilde{F}^{\m\n\rho}\tilde{F}_{\m\n\rho} \bigg)+e^{-\fr54\f+\frac{3}{4}\rho}R^{{\cal M}_5}\Big]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&=0 \label{HJfullst} \end{aligned}\] The action does admit a DBI form of once we assume that the fields are constant on the fixed \"time\" surface as in. (In that case, only the \(R^{{\cal M}_5}\) term contributes to the HJ equation among the terms in \({\cal L}^{\sst{(4)}}\).) We will discuss this in the next subsection. In the remainder of this section, we address the issue of the appearance of the gauge field. Also, we discuss the HJ procedure keeping all the terms in \({\cal L}^{\sst{(4)}}\). An interesting toggle between \(S_5\) and \({\cal H}_5\) will be noted depending on whether one uses constant field approximation. The appearance of a gauge field through a spontaneous symmetry breaking should be a general phenomenon independent of coefficients in \(H\). Indeed, it is a general phenomenon as proved by the following observation. The solution \(S\) that appears in can be viewed as a functional of an antisymmetric \"moduli field\", \(F_{\m\n}\), \[\begin{aligned} S=S[F_{\m\n}] \end{aligned}\] To be able to view \(F_{\m\n}\) as a field strength of a gauge field, closure of \(F_{\m\n}\) must be established. Let us take a covariant derivative on \[\begin{aligned} &&0=\nabla_{[\k}\pi^{\mu\nu]}(x)=\nabla_{[\k}\frac{1}{\sqrt{-g(x)}} \frac{\delta S}{\delta g_{\mu\nu]}(x)} \label{dpimn} \end{aligned}\] This implies \[\begin{aligned} \nabla_{[\k} F_{\m\n]}=0, \label{jaco} \end{aligned}\] and therefore \(F_{\m\n}\) can be taken as the field strength of a gauge field, \(A_\m\).[^8] For the solution, one can perform the type of derivative expansion considered, e.g., in. Let us set \[\begin{aligned} S_0\equiv S_0^{\sst{(0)}}+S_0^{\sst{(1)}}+\cdots \label{sersol} \end{aligned}\] where \(S_0^{\sst{(0)}}\) (\(S_0^{\sst{(1)}}\)) represents the leading (next) order term in the derivative expansion. Let us work out \(S_0^{\sst{(0)}}\) and \(S_0^{\sst{(1)}}\). In the leading order, the HJ equation reads \[\begin{aligned} &&\fr{e^{\fr34 \f-\fr54\rho}}{(\sqrt{-g})^2} \left[-\bigg(\frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}}\bigg)^2-\fr{1}{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}} \frac{\delta S_0^{\sst{(0)}}}{\delta \phi}-\fr1{2}\bigg(\frac{\delta S_0^{\sst{(0)}}}{\delta \f}\bigg)^2 \right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \qquad\qquad-\fr{3}{10}\bigg(\frac{\delta S_0^{\sst{(0)}}}{\delta \rho}\bigg)^2 +\fr1{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}} \frac{\delta S_0^{\sst{(0)}}}{\delta \rho}\right] \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{ e^{-\fr54\f-\fr54 \rho}}{(\sqrt{-g})^2}\left[ 12\bigg(\frac{\delta S_0^{\sst{(0)}}}{\delta D_{\mu\nu\lambda\rho}} \bigg)^2\right]-e^{-\fr54\f+\frac{3}{4}\rho}R^{{\cal M}_5}=0 \label{HJzeroth} \end{aligned}\] Although is not a solution once the terms with derivatives in \({\cal L}^{\sst{(4)}}\) are included, the similar types of the terms that would appear when is expanded should appear in the solution. Guided by this let us try the following ansatz \[\begin{aligned} S_0^{\sst{(0)}}=\b_{\sst{(0)}} \int d^4x\,\sqrt{-g}\; e^{-\f+\rho} \label{Szz} \end{aligned}\] Substituting into leads to \[\begin{aligned} \fr{1}{5}\b_{\sst{(0)}}^2=R^{{\cal M}_5} \end{aligned}\] This indicates that one should take \({\cal M}_5=S_5\). At the next order, the HJ equation takes \[\begin{aligned} &&\fr{e^{\fr34 \f-\fr54\rho}}{(\sqrt{-g})^2} \left[-2 \frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}}\frac{\delta S_0^{\sst{(1)}}}{\delta g_{\rho\s}}g_{\m\rho}g_{\n\s}-\fr{1}{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}} \frac{\delta S_0^{\sst{(1)}}}{\delta \phi}-\fr{1}{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(1)}}}{\delta g_{\mu\nu}} \frac{\delta S_0^{\sst{(0)}}}{\delta \phi} \right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \hspace{.5in}-\frac{\delta S_0^{\sst{(0)}}}{\delta \f}\frac{\delta S_0^{\sst{(1)}}}{\delta \f}-\fr{3}{5}\frac{\delta S_0^{\sst{(0)}}}{\delta \rho}\frac{\delta S_0^{\sst{(1)}}}{\delta \rho} +\fr1{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(0)}}}{\delta g_{\mu\nu}}\frac{\delta S_0^{\sst{(1)}}}{\delta \rho} \right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \hspace{.5in} +\fr1{2}{g_{\mu\nu}}\frac{\delta S_0^{\sst{(1)}}}{\delta g_{\mu\nu}}\frac{\delta S_0^{\sst{(0)}}}{\delta \rho}-\bigg(\frac{\delta S_0^{\sst{(1)}}}{\delta B_{\mu\nu}} +\chi\frac{\delta S_0^{\sst{(1)}}}{\delta C_{\mu\nu}} +6C_{\lambda\rho} \frac{\delta S_0^{\sst{(0)}}}{\delta D_{\mu\nu\lambda\rho}}\bigg)^2\right] \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{{e^{-\fr54\f-\fr54 \rho}}}{(\sqrt{-g})^2}\left[ \bigg(\frac{\delta S_0^{\sst{(1)}}}{\delta C_{\mu\nu}}\bigg)^2 \right]-\Big[e^{-\fr34\f+\fr54 \rho}\Big(R^{\sst{(4)}} +\fr32 \nabla_\m\nabla^\m \f-\fr52 \nabla_\m\nabla^\m \rho-\fr78\pa_\m \f \pa^\m \f \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{15}{8}\pa_\m \rho\pa^\m \rho+\fr54 \pa_\m \f \pa^\m \rho -\fr{1}{12} H_{\m\n\rho}H^{\m\n\rho}\Big) -\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\m} \chi \, \pa^{\m} \chi +\fr1{ 3!} \tilde{F}^{\m\n\rho}\tilde{F}_{\m\n\rho} \bigg)\Big]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&=0 \label{HJ1st} \end{aligned}\] With the zeroth order solution substituted, this reduces to \[\begin{aligned} &&\fr{e^{-\fr14 \f-\fr14\rho}}{\sqrt{-g}}\b_{\sst{(0)}} \left[ \fr{2}{5} \frac{\delta S_0^{\sst{(1)}}}{\delta \rho}-\fr{e^{\f-\rho}}{\b_{\sst{(0)}} \sqrt{-g}}\bigg(\frac{\delta S_0^{\sst{(1)}}}{\delta B_{\mu\nu}} +\chi\frac{\delta S_0^{\sst{(1)}}}{\delta C_{\mu\nu}} \bigg)^2\right]-\fr{e^{-\fr54\f-\fr54 \rho}}{(\sqrt{-g})^2} \bigg(\frac{\delta S_0^{\sst{(1)}}}{\delta C_{\mu\nu}}\bigg)^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\Big[e^{-\fr34\f+\fr54 \rho}\Big(R^{\sst{(4)}} +\fr32 \nabla_\m\nabla^\m \f-\fr52 \nabla_\m\nabla^\m \rho-\fr78\pa_\m \f \pa^\m \f \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\fr{15}{8}\pa_\m \rho\pa^\m \rho+\fr54 \pa_\m \f \pa^\m \rho -\fr{1}{12} H_{\m\n\rho}H^{\m\n\rho}\Big) -\fr12 e^{\fr54\phi+\fr{{5}}{4}\rho}\bigg(\pa_{\m} \chi \, \pa^{\m} \chi +\fr1{ 3!} \tilde{F}^{\m\n\rho}\tilde{F}_{\m\n\rho} \bigg)\Big]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && =0 \label{HJ1stred} \end{aligned}\] One can show that admits the following form of the solution, \[\begin{aligned} S_0^{\sst{(1)}}&=& \fr1{\b_{\sst{(0)}}}\int d^4x \sqrt{-g} e^{-\fr12\f+\fr32\rho} \Big[\fr53 R^{\sst{(4)}}+\fr{55}{24} (\pa \f)^2+\fr{25}{8}(\pa \rho)^2-\fr52 (\pa \f\cdot \pa \rho) \Big]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&+\fr{5}{6\b_{\sst{(0)}}} \int d^4x \sqrt{-g}\;e^{-\f+\rho} (\pa \chi)^2 -\fr{1}{10}\b_{\sst{(0)}} \int d^4x \sqrt{-g}e^{\fr32\f+\fr32\rho} {\cal F}_{\m\n}{\cal F}^{\m\n} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \end{aligned}\] It should be possible to find \(S_0^{n}, n>1\) in a similar way. ## Reduction that leads to DBI form solution to the HJ equation Let us verify that the HJ equation associated with admits a DBI form solution. In order to make our analysis slightly more general, note the following freedom. Suppose we use \[\begin{aligned} \tilde{\f}\rightarrow a\hat{\phi}+b\hat{\r} \qquad \tilde{\rho}\rightarrow c\hat{\phi} +d\hat{\r} \end{aligned}\] in the ansatze . Then one should get where \(\tilde{\f},\tilde{\rho}\) are replaced by \(a\hat{\phi}+b\hat{\r},\hat{\phi} +d\hat{\r}\) respectively. Let us utilize this freedom and introduce the following linear combinations of \(\f, \rho\), \[\begin{aligned} \f &\equiv& a\hat{\phi}+b\hat{\r} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \rho &\equiv & c\hat{\phi}+d \hat{\r} \end{aligned}\] Now Eq. can be rewritten \[\begin{aligned} &&\int\, d^5\hat{x}\,{\cal L}_{bulk+bd}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ =&&\int \, dr d^4 x\,\sqrt{-g}\, n \left[\fr{}{}\right. e^{-\fr34 (a\hat{\phi}+b\hat{\r})+\fr{{5}}{4}(c\hat{\phi}+d \hat{\r})} \bigg(-K_{\mu\nu}^2+K^2 +\fr{u_1}{n}f(\hat{\phi})K +\fr{u_2}{n}f(\hat{\r})K \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&+\frac{1}{n^2}\bigg\{u_3f(\hat{\phi})^2 +u_4f(\hat{\r})^2+u_5f(\hat{\phi})f(\hat{\r})-\frac{1}{4}[H_{r\mu\nu}-n^{\lambda}H_{\lambda\mu\nu}]^2 \bigg\}\bigg)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-\frac{1}{2n^2}e^{\fr54(a\hat{\phi}+b\hat{\r})+\fr{{5}}{4}(c\hat{\phi}+d \hat{\r})}\bigg\{[\pa_r \chi-n^{\mu}\pa_{\mu}\chi]^2 +\frac{1}{2}[\tilde{F}_{r\mu\nu}-n^{\lambda}\tilde{F}_{\lambda\mu\nu}]^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&+\frac{1}{24}[\tilde{G}_{r\mu\nu\lambda\rho}-n^{\sigma}\tilde{G}_{\sigma\mu\nu\lambda\rho}]^2 \bigg\} +{\cal L}^{\sst{(4)}} \;\;\bigg], \label{fstotq2} \end{aligned}\] where \[\begin{aligned} f(\hat{\phi})&\equiv & \pa_r \hat{\phi}-n^{\mu}\pa_{\mu}\hat{\phi} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ f(\hat{\r})&\equiv & \pa_r \hat{\r}-n^{\mu}\pa_{\mu}\hat{\r} \end{aligned}\] and the \(u\)'s are related to \(a,b,c,d\) \[\begin{aligned} && u_1\equiv-\fr32 a+\fr52 c, \qquad u_2 \equiv -\fr32 b+\fr52 d,\qquad u_3\equiv \fr14 a^2+\fr54 c^2-\fr{5}{2}ac,\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && u_4\equiv \fr14 b^2+\fr54 d^2-\fr{5}{2}bd,\qquad u_5 \equiv \fr12 ab+\fr52 cd-\fr52 bc-\fr52 ad \label{us} \end{aligned}\] After some algebra, one can show \[\begin{aligned} \!\!\!\!\!\!\!\!\!H&=& e^{\fr34 (a\hat{\phi}+b\hat{\r})-\fr54 (c\hat{\phi}+d\hat{\r})}\bigg(-\pi_{\m\n}^2+w_1(\pi_\m^\m)^2+w_2\pi_{\hat{\r}}^2+w_3\pi_{\hat{\phi}}^2 +w_4\pi_\m^\m \pi_{\hat{\phi}} +w_5\pi_\m^\m \pi_{\hat{\r}}+w_6\pi_{\hat{\phi}} \pi_{\hat{\r}} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \hspace{2in}-(\pi_{B\m\n}+\chi \pi_{C\m\n}+6C^{\lambda\rho}\pi_{D\m\n\lambda\rho})^2 \bigg)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-e^{-\fr54 (a\hat{\phi}+b\hat{\r})-\fr54 (c\hat{\phi}+d\hat{\r})}\bigg(\fr12 \pi_\chi^2+\pi_{C\m\n}^2+12\pi_{D\m\n\lambda\rho}^2 \bigg)-{\cal L}^{\sst{(4)}} \label{Hmix} \end{aligned}\] The parameters \(w_1...,w_6\) are related to \(u\)'s that appear in by \[\begin{aligned} w_1 &\equiv& \fr{u_2^2 u_3+u_1^2 u_4-4 u_3 u_4-u_1u_2u_5+u_5^2}{D} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ w_2 &\equiv & \fr{u_1^2-3u_3}{D}\quad w_3 \equiv \fr{u_2^2-3u_4}{D}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ w_4 &\equiv & \fr{2u_1u_4-u_2u_5}{D}\quad w_5 \equiv \fr{2u_2u_3-u_1u_5}{D}\quad w_6 \equiv -\fr{2u_1u_2-3u_5}{D} \end{aligned}\] with \[\begin{aligned} D\equiv 4u_2^2 u_3+4u_1^2 u_4-12 u_3 u_4-4u_1u_2u_5+3u_5^2 \end{aligned}\] As in, we assume that the fields are constant on the fixed \"time\" surface. Due to this assumption, only the \(R^{{\cal M}_5}\) term contributes to the HJ equation among the terms in \({\cal L}^{\sst{(4)}}\); substituting into the hamiltonian constraint, one finds the following HJ equation: \[\begin{aligned} &&e^{\fr34 (a\hat{\phi}+b\hat{\r})-\fr54(c\hat{\phi}+d\hat{\r})} \left[-\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}}\bigg)^2+w_1\bigg(\frac{g_{\m\n}}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}}\bigg)^2 +w_4\frac{g_{\mu\nu}}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}} \frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\phi}} \right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. +w_3\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\phi}}\bigg)^2 +w_2\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\r}}\bigg)^2 +w_6 \frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\phi}}\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\r}} +w_5 \frac{g_{\mu\nu}}{\sqrt{-g}}\frac{\delta S_0}{\delta g_{\mu\nu}} \frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \hat{\r}}\right. \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\left. \hspace{1.5in}-\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta B_{\mu\nu}} +\chi\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta C_{\mu\nu}} +6C_{\lambda\rho}\frac{1}{\sqrt{-g}} \frac{\delta S_0}{\delta D_{\mu\nu\lambda\rho}}\bigg)^2\right] \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&-e^{-\fr54(a\hat{\phi}+b\hat{\r})-\fr54 (c\hat{\phi}+d\hat{\r})}\left[\bigg(\frac{1}{2}\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta \chi}\bigg)^2 +\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta C_{\mu\nu}}\bigg)^2 +12\bigg(\frac{1}{\sqrt{-g}}\frac{\delta S_0}{\delta D_{\mu\nu\lambda\rho}} \bigg)^2\right]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && = e^{-\fr54(a\hat{\phi}+b\hat{\r})+\frac{3}{4}(c\hat{\phi}+d\hat{\r})}R^{{\cal M}^5} \label{HJmod} \end{aligned}\] Following, let us examine whether admits the following form of the solution, which is a slight modification of the corresponding solution in: \[\begin{aligned} S_0=S_c+S_{DBI}+S_{WZ} \label{szero} \end{aligned}\] where \[S_c=\alpha \int d^4x \sqrt{-g}\, e^{\z_3\hat{\phi}+\z_4\hat{\r}}\] \[S_{DBI}=\beta \int d^4x\, e^{\z_1\hat{\phi}+\z_2\hat{\r}} \sqrt{-\det (g_{\mu\nu}+{\cal F}_{\mu\nu})}\] \[\begin{aligned} S_{WZ} = \gamma \int d^4x \;\varepsilon^{\mu\nu\lambda\rho} \left( \frac{1}{24}D_{\mu\nu\lambda\rho} +\frac{1}{4}C_{\mu\nu}{\cal F}_{\lambda\rho} +\frac{1}{8}\chi {\cal F}_{\mu\nu}{\cal F}_{\lambda\rho} \right), \label{ScSBISWZ66} \end{aligned}\] where \[\begin{aligned} {\cal F}_{\m\n}\equiv-B_{\m\n}+F_{\m\n} \end{aligned}\] Inspection of the terms' structures reveals that the presence of \((\pi_\m^\m)^2\) would require a major modification of ; let us impose \[\begin{aligned} w_1=0 \label{wconstr} \end{aligned}\] Detailed computation implies that would admit a solution of the form once the following conditions are imposed in addition to the previous condition : \[\begin{aligned} && 2w_4\z_1+2w_5 \z_2=1,\quad \b^2(w_3\z_1^2+w_2\z_2^2+w_6\z_1\z_2)=-\fr12 \g^2,\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \z_1=-a,\quad \z_2=-b,\quad \z_3=-a+c,\quad \z_4=-b+d \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \a^2(-1+2w_4 \z_3+w_3 \z_3^2+w_2 \z_4^2+w_6 \z_3\z_4+2w_5 \z_4)=R^{{\cal M}_5},\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && w_4\z_3+w_5\z_4=1,\quad 2w_4\z_1+2w_3\z_1\z_3+2w_2\z_2\z_4+w_6\z_2\z_3+w_6\z_1\z_4+2w_5\z_2=0 \label{wconstr2}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \end{aligned}\] The constraints and amount to 4 constraints among \(a,b,c\) and \(d\). In other words, one can first use the second line to replace \(\z\)'s by the corresponding expressions on the right-hand sides of the second line. One can then solve \[\begin{aligned} \label{weq} && w_1=0,\quad 2w_4\z_1+2w_5 \z_2=1,\quad w_4\z_3+w_5\z_4=1 \\ && 2w_4\z_1+2w_3\z_1\z_3+2w_2\z_2\z_4+w_6\z_2\z_3+w_6\z_1\z_4+2w_5\z_2=0 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document} \end{aligned}\] where \(\z\)'s should take the explicit expression in terms of \((a,b,c,d)\). Once the solutions are determined, they can be substituted into the remaining two equations \[\begin{aligned} && \b^2(w_3\z_1^2+w_2\z_2^2+w_6\z_1\z_2)=-\fr12 \g^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \a^2(-1+2w_4 \z_3+w_3 \z_3^2+w_2 \z_4^2+w_6 \z_3\z_4+2w_5 \z_4)=R^{{\cal M}^5} \label{crucialeq} \end{aligned}\] and these equations will determine the relations between \(\a,\b,\g\). Interestingly, it turns out that the four equations are automatically satisfied. This implies that one can freely choose \((a,b,c,d)\); as far as the remaining equations in , \[\begin{aligned} && \b^2(w_3\z_1^2+w_2\z_2^2+w_6\z_1\z_2)=-\fr12 \g^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \z_1=-a,\quad \z_2=-b,\quad \z_3=-a+c,\quad \z_4=-b+d \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \a^2(-1+2w_4 \z_3+w_3 \z_3^2+w_2 \z_4^2+w_6 \z_3\z_4+2w_5 \z_4)=R^{{\cal M}_5}, \label{remeq}\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \end{aligned}\] are satisfied, the reduction will be consistent. The first and third equation in becomes \[\begin{aligned} \b^2&=& \g^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \fr15 \a^2 &=& R^{{\cal M}_5} \end{aligned}\] regardless of values of \((a,b,c,d)\), and therefore one must take \[\begin{aligned} {{\cal M}_5}={S_5} \end{aligned}\] The case we have considered in the previous subsection corresponds to[^9] \[\begin{aligned} a=d=1,\quad b=c=0 \end{aligned}\] which leads to \[\begin{aligned} u_1=-\fr32,\quad u_2=\fr52,\quad u_3=\fr14,\quad u_4=\fr54,\quad u_5=-\fr52 \end{aligned}\] and \[\begin{aligned} w_1=0,\quad w_2=-\fr3{10},\quad w_3=-\fr12,\quad w_4=-\fr12,\quad w_5=\fr12,\quad w_6=0 \end{aligned}\] ## Implications In the previous subsection, a gauge theory action was obtained after the HJ procedure. The HJ principal function \(S\) is nothing but the lagrangian with \(r\) playing the role of time. As mentioned in the introduction, the gauge action can be interpreted as a dual action to the 4D gravity action. The 4D gravity system itself is a (gauged) dS supergravity. ### on realization of \"braneworld\" Let us ponder whether the \"braneworld\" is realized by the current procedure. First of all, we should note that the current procedure implies a qualitatively different braneworld from the conventional Randall-Sundrum type in that the only dynamical degrees of freedom are those of the gauge multiplet after integrating out the gravitational degrees of freedom. The situation is analogous to the usual QFT procedure where instantons become dynamical degrees of freedom that are \"dual\" to the original gauge theory. One has instanton moduli and fluctuation degrees of freedom in the path integral once one expands around an instanton solution. After one integrates out the fluctuation degrees of freedom, one finds an instanton action that can be viewed as \"dual\" to the original action (See and for related discussions.) The HJ procedure is a solution-finding procedure, and we saw the moduli field \(F_{\m\n}\) enter for the case at hand. One should then integrate out the 4D gravitational degrees of freedom (i.e., all the other degrees of freedom than the moduli field), and eventually find an action of the moduli field. There will also be the gravitational part of the background, therefore, the ultimate action of the moduli fields would be in that gravity background.[^10] For the braneworld realization, it would be required to check whether a brane solution of localizes at some value of \(r\). As stated above, the current procedure leads to a qualitatively different braneworld. There still exists a feature within the current setup that might be an indication of the localization of all the degrees of freedom: \[\begin{aligned} \frac{\pa S}{\pa r_0}=0. \end{aligned}\] ### new paradigm for black hole physics As discussed in, the black hole information paradox is an amenable problem in string theory context. It is in the 4D pure Einstein gravity where the paradox becomes more subtle. The present work may have an application in black hole physics; in particular, in the aspect associated with the information paradox in the 4D pure Einstein gravity. In the usual approach of QFT in a curved spacetime, the geometry enters as a background whereas the matter fields are treated on the quantum level. It is almost evident that geometry as a non-dynamic background would be inadequate for describing physics in which the back-reaction plays a crucial role. The information paradox should lie in the classical treatment of the geometry (see, e.g., the recent discussion in ). The geometry is strictly classical in the conventional approach because the matter quantum fields in the usual approach do not directly describe the fluctuations of the *geometry*.[^11] The best solution for this status of matter would be the full quantization of the Einstein-Hilbert action. Given the unavailability of such an apparatus, the second best solution would be to have a *semi-classical* treatment of geometry. The dual gauge action obtained through the ADM reduction of this paper should provide the needed semi-classical tool. In the usual approach, the matter fields are not intrinsically gravitational degrees of freedom. In contrast, the gauge action obtained as a result of ADM reduction provides degrees of freedom that are intrinsic to the original gravity system. The matter field equations are solved in some background metric in the conventional approach. Since it is not the proper full coupled equations between matter and metric that are solved, the result is bound to be without a proper account of back-reaction from the metric that gets deformed by the matter. In the proposed ADM reduction approach, one gets the \"matter\" system, i.e., the YM field after the spontaneous symmetry breaking. In other words, the appearance of the \"matter fields\" is built into the formulation. One can then try to solve those equations associated with YM field. However, the interpretation is now very different: the gauge field equations directly, although semi-classically, describe the fluctuations of the geometry. We will have more on this as well as other speculative issues in the conclusion. # Domain-wall solution, toroidal compactification and \"inflaton\" In this section, we analyze two more aspects of the 5D action that has been obtained by the sphere reduction in sec 2. One thing to note is that even if we are using the notation \(r\), it is not necessarily a radial coordinate; it is one of the spatial coordinates. ## Domain-wall solution One may use either a 5D Einstein-type frame or a \"string\"-type frame to find a solution. In this section, we use an Einstein type frame. (It should be possible to find the corresponding solution in a 5D \"string-type\" frame.) Consider which we quote here for convenience, \[\begin{aligned} I_5 &=& \fr1{2\k_5^2}\int d^5-\fr12(\pa \f)^2-\fr1{2\cdot 3!} e^{-\phi}e^{\fr53 \rho} H_{\sst{(3)}}^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && -\fr12 e^{2\phi}(\pa\chi)^2-\fr1{2\cdot 3!}e^{\phi}e^{\fr53 \rho}\tilde{F}^{2}_{\sst{(3)}}-\fr1{2\cdot 5!}e^{\fr{10}3 \rho}\tilde{G}^2_{\sst{(5)}}+e^{-\fr43\rho} R^{{\cal M}_5}\Big] \label{5dactRhoEinqq} \end{aligned}\] and the reduced field equations setting \(\chi=H=\f=F=0\). In this section, we take \[\begin{aligned} R^{{\cal M}_5}= R^{S^5} \end{aligned}\] Below we will set \(G=0\) as well because only that case admits a relatively simple solution. It follows from the action given in with \(F=\chi=H=\f=0\) that \[\nabla_{\underline{m}}\left(e^{\fr{10}{3} \rho}\, \tilde{G}_{\sst{(5)}}^{\;\;\underline{m} \underline{n}_1...\underline{n}_4}\right)=0, \label{Gtfeq}\] \[\nabla^2\rho-\fr1{120}e^{\fr{10}3\rho}\tilde{G}^2_{\sst{(5)}}-\fr4{5} e^{-\fr43\rho} R^{{\cal M}_5}=0 \label{rhofeq}\] \[\begin{aligned} R^{\sst{(5)}}_{\underline{m}\underline{n}}-\fr5{6}\pa_{\underline{m}} \rho \,\pa_{\underline{n}} \rho-\fr{ 1}{4\cdot 4!}e^{\fr{10}3 \rho}\, \tilde{G}_{\underline{m}\underline{p}\underline{q}\underline{r}\underline{s}}{\tilde{G}_{\underline{n}}}^{~\underline{p}\underline{q}\underline{r}\underline{s}}-h_{\underline{m}\underline{n}}\left(\fr12 R^{\sst{(5)}}-\fr{5}{12}(\pa \rho)^2+\fr12 e^{-\fr43\rho}R^{{\cal M}_5} \right)=0\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \label{Einsfeq} \end{aligned}\] Let us try the following metric ansatz, \[\begin{aligned} ds_5^2= e^{2A}dr^2+e^{2C(r)}ds_{dS_4}^2 \label{ds5sol2q} \end{aligned}\] The \(G_5\) field equation implies \[\begin{aligned} G_5^{m_1...m_5} &=& \fr{k}{\sqrt{-h}}e^{-\fr{10}{3}\rho}\epsilon^{m_1...m_5} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ G_5^2 &=&-5!k^2 \;e^{-\fr{20}{3}\rho} \end{aligned}\] \[\tilde{G}_{\underline{m} \underline{n}_1 \dots \underline{n}_4}{\tilde{G}_{\underline{k}}}^{\,\,\,\underline{n}_1\dots \underline{n}_4}=\fr15 h_{\underline{m}\underline{k}}\tilde{G}^2_{\sst{(5)}} \. \label{Gtrel5q}\] Consider \((rr)\) and \((11)\) components of : \[R^{\sst{(5)}}-\fr{5}{6}h^{rr}(\pa_r \rho)(\pa_r \rho)+ e^{-\fr43\rho}R^{{\cal M}_5}-\fr2{h_{rr}}\Big[R^{\sst{(5)}}_{rr}-\fr{5}{6}(\pa_r \rho)(\pa_r \rho)-\fr{1}{4\cdot 4!} e^{\fr{10}3 \rho}\tilde{G}_{r\underline{p}\underline{q}\underline{r}\underline{s}}{\tilde{G}_{r}}^{~\underline{p}\underline{q}\underline{r}\underline{s}}\Big]=0 \label{Einsrr}\] \[R^{\sst{(5)}}-\fr{5}{6}h^{rr}(\pa_r \rho)(\pa_r \rho)+ e^{-\fr43\rho}R^{{\cal M}_5}-\fr2{h_{11}}\Big[R^{\sst{(5)}}_{11}-\fr{1}{4\cdot 4!} e^{\fr{10}3 \rho} \tilde{G}_{1\underline{p}\underline{q}\underline{r}\underline{s}}{\tilde{G}_{1}}^{~\underline{p}\underline{q}\underline{r}\underline{s}}\Big]=0 \label{Eins11}\] Combining and , one gets \[\begin{aligned} \fr1{h_{rr}}\Big[R^{\sst{(5)}}_{rr}-\fr5{6}\pa_{r} \rho \,\pa_{r} \rho \Big] =\fr1{h_{11}}R^{\sst{(5)}}_{11} \label{difone} \end{aligned}\] Using the result, e.g., in appendix B of, one can show \[\begin{aligned} R_{11}^{\sst{(5)}}&=& R_{11}^{\sst{(4)}}+g_{11}e^{2C-2A}(-4 \nabla_r C\nabla_r C-\nabla_r \nabla_r C +\nabla_r A \nabla_r C) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ R_{rr}^{\sst{(5)}}&=& 4(\nabla_r A\nabla_r C-\nabla_r C \nabla_r C-\nabla_r\nabla_r C) \end{aligned}\] With these, yields[^12] \[\begin{aligned} && 4\nabla_r A \nabla_r C-4\nabla_r C \nabla_r C-4\nabla_r \nabla_r C-\fr56 (\pa_r\rho)^2 \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ =&&e^{2A-2C}(\L-4 e^{2C-2A}\nabla_r C\nabla_r C-e^{2C-2A}\nabla_r \nabla_r C +e^{2C-2A}\nabla_r A \nabla_r C ) \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \label{difone2} \end{aligned}\] Let us take \(h^{mn}\) on : \[\begin{aligned} R^{\sst{(5)}}-\fr{5}{6}h^{rr}(\pa_r \rho)(\pa_r \rho)+\fr53 e^{-\fr43\rho}R^{{\cal M}_5} +\fr{1}{6\cdot 4!} e^{\fr{10}3\rho}G_5^2=0 \label{EinsRicci} \end{aligned}\] Combining and and setting \(G=0\)[^13], one gets \[\begin{aligned} -\fr23 e^{-\fr43 \rho}R^{{\cal M}_5}=2 e^{-2C}(\L^{\sst{(4)}}-4e^{2C}\nabla_r C \nabla_r C-e^{2C}\nabla_r\nabla_r C) \label{ricci11} \end{aligned}\] Then implies \[\begin{aligned} C=\fr23 \rho\quad,\quad \L^{\sst{(4)}}=3q_2^2-\fr13 R^{{\cal M}_5} \end{aligned}\] The \(\rho\)-eq takes \[\begin{aligned} \fr{d^2 \rho}{dr^2}+(4 \fr{dC}{dr}-\fr{dA}{dr})\fr{d\rho}{dr}-\fr45 e^{2A-\fr43\rho}R^{{\cal M}_5}=0 \label{rhoeq4} \end{aligned}\] In the absence of \(G_{\sst{(5)}}\), one can show that takes \[\begin{aligned} && R^{\sst{(5)}}-\fr56 e^{-2A}(\fr{d\rho}{dr})^2+ e^{-\fr43 \rho}R^{{\cal M}_5}=\fr{2}{h_{11}}R_{11}^{\sst{(5)}} \label{Eins112} \end{aligned}\] Let us consider the following set of ansatze: \[\begin{aligned} && A=0\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&\rho=p\ln(q_1+q_2\, r)\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ &&C=\fr23 p\ln(q_1+q_2\, r) \label{rhoc} \end{aligned}\] Using \[\begin{aligned} && R_{\sst{(5)}}= e^{-2C}R_{\sst{(4)}}+4\Big[ -2 \nabla_r\nabla_r C-5 \nabla_r C \nabla_r C\Big]\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ \end{aligned}\] takes \[\begin{aligned} && e^{-2C}R^{\sst{(4)}}-4(2\nabla_r\nabla_r C+5\nabla_r C\nabla_r C)+e^{-\fr43 \rho}R^{{\cal M}_5}-\fr56 e^{-2C}(\pa_r \rho)^2\nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ =&& 2e^{-2C}(\L-4 e^{2C}\nabla_r C\nabla_r C-e^{2C}\nabla_r \nabla_r C) \label{R11eq} \end{aligned}\] Substituting into and , one can see that \(p=\fr32\) and \[\begin{aligned} && R^{{\cal M}_5}=\fr{45}{8}q_2^2\quad \mbox{from \rf{rhoeq4}} \nonumber} \def\bd{\begin{document}} \def\ed{\end{document}\\ && \L=\fr{9}{8}q_2^2 \quad \mbox{from \rf{difone2}} \end{aligned}\] where \(\L\) (\(\equiv \L^{\sst{(4)}}\)) is defined by \(R_{\m\n}^{\sst{(4)}}=\L g_{\m\n}\). Eq. also produces a consistent result: \[\begin{aligned} && 2\L+R^{{\cal M}_5}=\fr{63}{8}q_2^2 \end{aligned}\] ## Toroidal compactification and \"inflaton\" Carrying out a conventional dimensional reduction on the 5D action will be worthwhile because it will yield a 4D gravity theory with various gauge fields with positive cosmological constant. The theory has been obtained from IIB supergravity, and provides a potentially interesting inflationary model. Let us consider the \(\rho\)-rescaled form . One can easily carry out dimensional reduction keeping as many fields in as one wished. Focusing on the perspective of 4D inflatonary physics, we illustrate the case with the metric and \(\rho\): \[\begin{aligned} I&=& \fr1{2\k_5^2}\int d^5 +e^{-\fr{4}{\sqrt{15}}\rho} R^{{\cal M}^5}\Big] \label{5dactRhoEinqred} \end{aligned}\] One can consider a simple dimensional reduction given by \[\begin{aligned} \rho=\rho(x^\m)\quad,\quad ds_5^2=dr^2+g_{\m\n}dx^\m dx^\n \end{aligned}\] The resulting 4D action takes \[\begin{aligned} I&=& \int d^4 x \sqrt{-g}\Big[ R^{\sst{(4)}}-\fr12(\pa\rho)^2 +e^{-\fr{4}{\sqrt{15}}\rho} R^{{\cal M}^5}\Big] \label{4dinf} \end{aligned}\] The gauge fields \(\chi ,C, D\) and \(B\) can easily be accommodated. It would be interesting to investigate whether one could construct a dS solution with addition of various form fields. If so, that would be in line with the observation made in the KKLT type approaches. We leave the resulting model's phenomenological study for the future. # Conclusion In sec 2, we have carried out reduction of IIB supergravity on \({\cal M}_5 (={\cal H}_5\;\; \mbox{or}\;\; S_5 )\). The resulting 5D gravity lagrangian has been analyzed in several different directions. In one direction, we have shown that it admits a 4D curved domain-wall solution. In another direction, we have performed, following, the Hamilton-Jacobi procedure of canonical transformation and have obtained another gravity description. As shown in sec 3, the Hamilton-Jacobi equation admits a class of solutions that take a form of a gauge theory action.[^14] The way the gauge field strength \(F_{\m\n}\) appears is intriguing. The worldvolume gauge fields emerge as \"moduli fields\": regardless of the values that the gauge fields take, the gauge action satisfies the Hamilton-Jacobi equation of the gravity system, and the gauge fields describe the fluctuations of the moduli space. They must be an inequivalent set of extremum solutions, and the inequivalence must stem from different patterns of the brane fluctuations. Those patterns are parameterized through the field \(F_{\m\n}\). Interestingly, the appearance of the dual degrees of freedom as a form of moduli fields was observed before in the context of forward duality: the strong coupling limit of a DBI action admits a class of solutions that can be collectively interpreted as a closed string action. One of the potentially powerful implications is the fact that getting non-gravitational degrees of freedom through ADM reduction would work for the pure 4D Einstein-Hilbert action. We believe that for the proper treatment of the black hole information two conditions are required for the QFT tool adopted to tackle the paradox. The first is that the adopted QFT should *directly* describe the fluctuations of the geometry. It is necessary to use a formulation that is self-consistent or \"closed\" under the forward and backward dualizations. The second is that the QFT interactions in their precise forms must be included. The ADM reduction would, in principle at least, determine the precise form of the interactions of the resulting gauge theory. In these regards, the ADM reduction approach should provide a proper paradigm for black hole physics. There are multiple future directions: One is an obvious direction of studying the supersymmetry aspect of the 5D/4D theories. Several other directions are associated with a better understanding of the ADM reduction itself. One may try to extend the program of sec 3.1 to higher orders in the derivative expansion. Another direction would be to address the following question: what at the full string theory level would be responsible for the appearance of a gauge field from a gravity system? The appearance of a gauge field seems to be a general phenomenon that occurs in a low energy theory that may not have embedding in a string theory. However, it would be still interesting to see the full stringy mechanism that is behind for the theories that do have stringy embedding. (See our speculation below.) There are other related issues that require further study. One of them is the issue of whether the emerging gauge field would be abelian or non-abelian. The appearance of an abelian gauge moduli field is straightforward. The real question is if there could be a (relatively simple) way to introduce non-abelian degrees of freedom. At this point we can only state what we anticipate and should postpone a better answer until further research. Presumably, the abelian vs non-abelian issue would depend on whether one uses a collection of D3 branes or, say D1 branes to describe the bulk physics. As observed, e.g., in, a higher dimensional abelian brane can be described by a non-abelian lower dimensional branes. As stated in sec 2, applications of the ADM reduction approach to black hole information should be interesting as well. The other directions concern phenomenological aspects and applications. One may take with other form fields as a starting point, and study the resulting 4D Friedmann-Lemaitre-Robertson-Walker eqs in the presence of D3 (or even D7). It will be interesting to make a connection this way with the KKLT and related compactification scenarios. Finally remarks on more speculative aspects are in order: The *forward dualization* mentioned in the introduction should be associated with endpoints of an open string sticking together and becoming a closed string. By the same token, the appearance of gauge degrees of freedom should presumably be associated with a closed string opens up and becomes an open string on the closed string theory level. It would be very interesting if one could make this more precise and quantitative. The following question was raised in. The DBI action contains all \(\a'\)-order terms[^15] but it still appears as a solution of reduced supergravity that is just the leading \(\a'\) action of a closed string. Perhaps the answer lies in the following. The gauge form solution represents excitations of massless open string modes. The higher \(\a'\) corrections to the IIB supergravity may be associated with a massive gauge action that has all the *massive open string modes* appearing explicitly at first and then subsequently integrated out in the open string context, therefore, deforming the massless gauge field. (The integrating out procedure should be done using the full string theory setup which of course would be a hard step in practice.) Differently put, the massless closed strings viewed as a composite open string state should be massive, and apparently they seem sufficient to account for the massless gauge theory modes. AJN acknowledges partial support from the Joint DFFD-RFBR Grant \# F40.2/040. Part of this work was carried out during IP's visit to CQUeST, Sogang university. IP thanks B. H. Lee, J. H. Park and C. Rim for their hospitality. # Appendix A: Differential form conventions {#appendix-a-differential-form-conventions .unnumbered} We use the following conventions on differential forms. The flat space metric signature is mostly positive, so \(\det \eta_{ab}=-1\) in any space-time dimensions. The Levi-Civita tensor \(\epsilon^{a_1\dots a_D}\) is defined such that \[\epsilon^{01\dots D-1}=1,\quad \epsilon_{01\dots D-1}=-1, \label{LCD}\] hence \[\epsilon^{a_1\dots a_D}\epsilon_{a_1\dots a_D}=\epsilon_{a_1\dots a_D}\eta^{a_1a^\prime_1}\dots \eta^{a_D a^\prime_D}\epsilon_{a^\prime_1\dots a^\prime_D}=\det \eta \cdot D!=-D! \label{LCLCD}\] Generalization of to a curve background is given by \[\epsilon^{m_1 \dots m_D}\epsilon_{m_1 \dots m_D}=\det g \cdot D! \label{LCLCcD}\] For a \(p\)-form we choose \[\Omega_{(p)}=\fr1{p!}\Omega_{m_1\dots m_p} \, dx^{m_1}\wedge \dots \wedge dx^{m_p}, \label{pform}\] and \[d\Omega_{(p)}=\fr1{p!}\pa_{[m_{p+1}}\Omega_{m_1\dots m_p]}\, dx^{m_{p+1}}\wedge dx^{m_1}\wedge \dots dx^{m_p}\] \[=\fr1{p!}\pa_{[m_1}\Omega_{m_2\dots m_{p+1}]}\, dx^{m_1}\wedge dx^{m_2}\wedge \dots dx^{m_{p+1}}. \label{dpform}\] The external derivative \(d\) acts from the left, i.e., \[d(\Omega_{(p)} \wedge \Omega_{(q)})=d\Omega_{(p)}\wedge \Omega_{(q)}+(-)^p \Omega_{(p)} \wedge d\Omega_{(q)}. \label{dpq}\] We define the Hodge star as \[\ast \left(dx^{n_1}\wedge \dots dx^{n_p} \right)=\fr1{(D-p)!} \fr1{\sqrt{-g}} \, {\epsilon_{{m_1}\dots m_{D-p}}}^{n_1 \dots n_p} \, dx^{m_1}\wedge \dots dx^{m_{D-p}}, \label{astD}\] so the dual to \(\Omega_{(p)}\) form \(\ast \Omega_{(p)}\) is defined by \[\ast{\Omega}_{(p)}=\fr1{(D-p)!p!}\fr1{\sqrt{-g}}\, {\epsilon_{{m_1}\dots m_{D-p}}}^{n_1 \dots n_p} \, \Omega_{n_1\dots n_p} \, dx^{m_1}\wedge \dots dx^{m_{D-p}}. \label{dualpD}\] On account of the latter expression and eq. , one gets \[\ast^2_p=(-)^{Dp+p+1}. \label{ast2}\] To get , the following relation should be used: \[{\epsilon_{{m_1}\dots m_{D-p}}}^{n_1 \dots n_p} {\epsilon^{{m_1}\dots m_{D-p}}}_{k_1 \dots k_p}=\det g \cdot (D-p)! p! \d^{[n_1}_{[k_1} \dots \d^{n_p]}_{k_p]}\equiv \det g \cdot (D-p)! p! \d^{n_1 \dots n_p}_{k_1 \dots k_p} \. \label{eepD}\] Also, we have \[dx^{m_1}\wedge \dots \wedge dx^{m_D}=d^Dx \, \epsilon^{m_1 \dots m_D}, \label{dxVD}\] Taking into account , and , this implies \[\fr1{2 \cdot p!} \int \, d^D x \, \sqrt{-g} \left( F_{(p)} \right)^2=(-)^{Dp+p+1}\, \fr12 \int_{\mathcal{M}^D} \, F_{(p)} \wedge \ast F_{(p)}. \label{Fp2D}\] In our notation \[\int \, d^D x \, \sqrt{-g} \equiv \int_{\mathcal{M}^D} \, \mathbf{1} \,, \label{VD}\] and, therefore, \[\int \, d^D x \, \sqrt{-g} R \equiv \int_{\mathcal{M}^D} \, \mathbf{1} \cdot R \,. \label{EHdiffD}\] [^1]: The symmetries that are broken and the breaking patterns will not be pursued in this work. [^2]: Here inhomogeneity refers to the fact that the space is not a coset manifold. [^3]: The authors of considered IIB action in string frame. The Einstein frame ansatze that we consider below are not connected to the ansatze considered in, therefore should belong to a different class of ansatze. (\(S_5\) vs \({\cal H}_5\) does not matter for this matter.) As a matter of fact, one can show that the precise forms of (2.4), (2.5) and (2.6) of used in *Einstein* frame also lead to consistent reduction. (The precise forms of (2.4-6) of and our ansatze , lead to the same 5D action up to a numerical rescaling, namely, .) This seems rather unusual, and must be attributed to the simplicity of the ansatze. (Our earlier false accusations of the work of were made in part by this subtlety.) In general, if an ansatz leads to consistent reduction in one frame, it would not in the other frame. In one of the footnotes in section 3, we point out another related aspect of the two ansatze. [^4]: In a typical Kaluza-Klein reduction, it is usually a scalar sector that makes the procedure complicated. An example of sphere reduction with many scalars can be found in. [^5]: In the earlier version of this work, the cross term \(\fr52 \pa_{\underline{m}} \f \pa^{\underline{m}} \rho\) was missed. (This has been pointed by Sato and Tsuchiya.) [^6]: We choose the Hamiltonian as follows: \[S=\int \, d^5 \hat{x} \,{{\cal L}}_{bulk+bd}=\int \, dr d^4 x \, \sqrt{-g}\left(\bf{P}\cdot \pa_r{\bf Q}-{\cal H} \right)\,,\] where \(\bf{Q}\) and \(\bf{P}\) are the "coordinates" and the corresponding "momenta" \[{\bf{P}}=\fr1{\sqrt{-g}}\fr{\d S}{\d \pa_r \bf{Q}} \,.\]. [^7]: The \(r=r_0\) surface should not be taken as a genuine boundary. If it were a genuine boundary, the metric on the boundary would be a constant as implied by the Dirichlet boundary condition. (Recall that the GH boundary terms were introduced for Dirichlet boundary conditions for the metric.) Rather it should be taken as a device that bridges the bulk description and the hypersurface description. [^8]: There are some subtleties here. Strictly speaking, \(S=S[{\cal F}_{\m\n}]\) with \({\cal F}_{\m\n}\equiv-B_{\m\n}+F_{\m\n}\). One may simply consider the \(B_{\m\n}=0\) case to avoid unnecessary complications. One can go a little further by the following observation: the right hand side of vanishes due to the (anti)symmetry property. Therefore, \({\cal F}_{\m\n}\) is closed, and this implies the closure of \(B_{\m\n}\). (Remember that one can consider the solution of HJ equation with the vanishing moduli field \(F_{\m\n}=0\).) In other words, only a closed \(B_{\m\n}\) can be a solution of HJ equation. Another subtlety is a question whether \(F_{\m\n}\) would be abelian or non-abelian. We comment on this and related issues in the conclusion. [^9]: One can also consider the case \[\begin{aligned} a=\fr83,\quad d=1,\quad b=c=0 \end{aligned}\] This choice casts the exponential factors in into the forms that were considered in. One finds, in this case, \[\begin{aligned} u_1=-4,\quad u_2=\fr52,\quad u_3=\fr{16}9,\quad u_4=\fr54,\quad u_5=-\fr{20}3 \end{aligned}\] Some of these coefficients are different from those that appeared in , and should be an indication that and belong to a different class of ansatze than those of. Even though the two ansatze are different, they admit the same DBI solutions; we take this as certain robustness of the DBI form solution. [^10]: The action in the curved background might be identified as including quantum and none-perturbative effects. Such an identification was made in for example. [^11]: It is the matter fields that represent the fluctuating degrees of freedom in the conventional approach, and unlike the current ADM reduction approach, the matter fields are extrinsic to the geometric degrees of freedom. One may say that the matter fields indirectly describe the fluctuations of the geometry since they are coupled to the metric. However, this would be so only when the geometry degrees of freedom are quantized as well. [^12]: where \[\begin{aligned} \nabla_r \nabla_r= \nabla^2 =e^{-A}\pa_r (e^{-A}\pa_r) \end{aligned}\] [^13]: If one keeps \(G\), two different types of \(e^{\rho}\) factors are present, and this eliminates possibility of any simple solution. [^14]: The fact that the original lagrangian admits a domain-wall solution and its Hamilton-Jacobi equation admits a worldvolume action as a solution should be related although we will not pursue this aspect on a deeper level than is apparent. [^15]: Of course, this is true only in the leading derivative expansion in \({\cal F}_{\m\n}\); once the subleading terms such as \(\pa {\cal F}\), \(\pa\pa {\cal F}\), etc. are taken into account, new terms would appear.

{'timestamp': '2012-11-05T02:02:49', 'yymm': '1210', 'arxiv_id': '1210.3825', 'language': 'en', 'url': 'https://arxiv.org/abs/1210.3825'}

# Introduction Molecular clouds belong to the densest and coldest parts of the Milky Way interstellar medium. Shielded from interstellar radiation fields, they provide conditions necessary for star formation to take place. Observationally, molecular clouds exhibit a complicated, irregular, and filamentary morphology, and (sub)millimeter-line observations of molecular clouds suggest that the gas in the clouds is moving supersonically. Consensus has not been reached concerning the origin and nature of molecular clouds. It must be recognized that molecular gas is just one of the phases of the Milky Way interstellar medium, and its evolution is determined by many processes that occur in the disk. To understand it, we must also look into the large-scale structure of the multi-phased interstellar medium, and understand the cloud evolution within this context. Both observational and theoretical approaches have been taken in this direction. Observationally, the distribution of molecular gas in nearby galaxies can be accessed through millimeter line mapping, from which structures such as spiral arms, filaments, and spurs can be identified. Theoretically, the structure of the multi-phased interstellar medium in a galactic disk has been studied through simulating the whole disk with different approaches, complemented by analytical calculations. It was found that filaments or spurs can be created through the combination of gravitational instability of a galactic disk, galactic shear, and frequent encounters/agglomeration between molecular clouds. In the Milky Way, studies of the structure of molecular gas have been confined spatially to the molecular cloud scale or have been limited to the structure of the spiral arms. This is partly due to the complicated morphology of molecular gas and partly due to the superposition of the emission of molecular gas from different structures along the line of sight. In spite of these difficulties, it is of both observational and theoretical interest to identify large, coherent molecular structures in the Milky Way apart from the spiral arms, since these structures are natural tracers of the large-scale gas circulation in the Milky Way disk. In contrast to the extragalactic case, where we are limited by the resolution and sensitivity of the telescopes (and the filtering of interferometers), for our Milky Way it is possible to study the kinematics of the molecular gas and the associated star formation with in more detail. The Milky Way interstellar medium has long been thought to be dynamic. Shells and rims are generally found in the disks of the Milky Way and other galaxies. It has been proposed that the expansion of H ii\(\,\) regions, which creates shells and rims, can collect the interstellar medium into a gravitationally unstable state, and trigger star formation. The expansion of the bubbles can also energize the interstellar medium of the Milky Way efficiently. In this work, we present an observational study of the region at \(49.5^{\circ}<l<52.5^{\circ}\) in the Milky Way. The molecular gas in the region exhibits a high degree of coherence, and forms a filamentary gas wisp (gas filament) with a length of \(\sim 3^{\circ}\). The eastern part of the filamentary gas wisp sits at the edge of a bubble and is located at \(\sim 0.75^{\circ}\) above the galactic plane. This eastern part is listed in the context of infrared bubbles as one of the "favorites of the Milky-Way-Project volunteers", and it was studied in terms of the G52L nebula by, who claimed that it may be the largest single H ii region in the Milky Way. Based on several estimations , the filamentary gas wisp has a distance of \(9.77\;\rm kpc\), which implies a physical length of \(\gtrsim 500\) pc. This is \(\sim 5\) times longer than the Nessie Nebula reported by [^1]. The physical length of the filamentary gas wisp exceeds by much the size of a molecular cloud, and this filamentary gas wisp is by far the largest coherent molecular structure identified in the Milky Way. It exhibits a coherent velocity structure, and is composed of several molecular structures, including two molecular clouds and one expanding bubble structure. We present observations and an analysis of the region (Sect. [2](#sec:obs){reference-type="ref" reference="sec:obs"}, [3](#sec:theregion){reference-type="ref" reference="sec:theregion"}), followed by a detailed discussion focusing on the implications on the dynamics of the Milky Way interstellar medium and the life cycle of molecular gas (Sect. [4](#sec:discussion){reference-type="ref" reference="sec:discussion"}). In Sect. [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"} we conclude. # Archival data {#sec:obs} We obtained 3.6 \(\mu\)m and 8 \(\mu\)m data from the GLIMPSE project, which is a fully sampled, confusion-limited, four-band near-to-mid infrared survey of the inner Galactic disk. We obtained 24 \(\mu\)m data from the MIPSGAL project, which is a survey of the Galactic disk with the MIPS instrument on *Spitzer* at 24 \(\mu\)m and 70 \(\mu\)m. We obtained \(^{13}\)CO(1-0) molecular line data (\(\nu_0=110.2\rm\; GHz\)) from the Galactic Ring Survey, which is a survey of the Milky Way disk with the SEQUOIA multipixel array on the Five College Radio Astronomy Observatory 14 m telescope, and covers a longitude range of \(18 ^{\circ}<l<55.7^{\circ}\) and a latitude range of \|b\|\<\(1^{\circ}\) with a spatial resolution of \(46\arcsec\). # Results {#sec:theregion} ## Region Figure [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"} shows the *Spitzer* three-color image of the region from \(49.5^{\circ}<l<52^{\circ}\). The overlaid contours show the molecular gas traced by \(^{13}\)CO(1-0). The CO emission in all the panels is integrated within \(-5.0 \;{\rm km\;s^{-1}}<v_{\rm lsr}<17.4\;{\rm km\;s^{-1}}\). Several features can be identified. At \(51.5^{\circ}<l<52.5^{\circ}\), there is a bubble with a radius of \(\sim 1^{\circ}\) [G52L nebula, @2012ApJ...759...96B]. The molecular gas is situated to the north of the bubble. At \(51^{\circ}<l<52.5^{\circ}\), the molecular gas is organized in the form of two molecular clouds (G052.24+00.74 and G051.69+00.74). The cloud G052.24+00.74 has a roundish shape. This cloud is connected with another molecular cloud, G051.69+00.74. This cloud has a more elongated geometry, and star formation occurs only at its eastern part. At \(49.5^{\circ}<l<50^{\circ}\), there is noticeable contamination from gas with \(5.2 {\;\rm km\;s^{-1}}<v_{\rm lsr}<7.2 \;\rm km\;s^{-1}\) (see the red arrows in Fig. [\[fig:channel\]](#fig:channel){reference-type="ref" reference="fig:channel"}). The contaminating gas has an extremely narrow line width (\(\lesssim 0.5\;\rm km\;s^{-1}\)) and tends to spread along the spatial direction. This makes it easily distinguishable from the emission from the gas filament. This narrow line width implies that the emission comes from a close-by cloud. This is supported by the fact that the contaminating gas has a more diffuse morphology (see Appendix [6](#sec:app1){reference-type="ref" reference="sec:app1"} for \(^{13}\)CO(1-0) channel maps of the region). This distinction is similar to the supernova remnant G016.05-0.57 studied in. The two molecular clouds (G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74) have a similar velocity and velocity dispersion: the cloud G052.24\(+\)`<!-- -->`{=html}00.74 has \(v_{\rm lsr}\sim 4.6\;\rm km\;s^{-1}\) and \(\delta v \sim 2.6 \;\rm km\;s^{-1}\) and the cloud G051.69\(+\)`<!-- -->`{=html}00.74 has \(v_{\rm lsr}\sim 3.6\;\rm km\;s^{-1}\) and \(\delta v \sim 3.6 \;\rm km\;s^{-1}\). In the position-position and position-velocity space, they are connected with some wispy gas filaments (at \(l\sim 52^{\circ}\) and \(b\sim 0.8 ^{\circ}\) of the top and middle panels of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"}). The similarity of the velocity dispersions and the proximity of the clouds in position-velocity space imply that the two clouds are physically connected. It can be readily seen from the \(^{13}\)CO(1-0) emission that this double-cloud system belongs to a large filament (Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"} and Appendix [6](#sec:app1){reference-type="ref" reference="sec:app1"}). The filamentary gas wisp is coherent in both the spatial and the velocity direction, which makes it distinguishable from other molecular structures, for instance the \(\sim 50\;\rm km\;s^{-1}\) clouds (Appendix [7](#sec:app2){reference-type="ref" reference="sec:app2"}). Seen from the middle panel of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"}, the filamentary gas wisp extends from \(l=49.5^{\circ}\) to \(l=52.5^{\circ}\), which implies an angular extent of \(\gtrsim 3^{\circ}\). Seen from the bottom panel of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"}, the filamentary gas wisp has a limited velocity range of \(\sim 22 \;\rm km\;s^{-1}\) (\(-5.0 \;{\rm km\;s^{-1}}<v_{\rm lsr}<17.4\;{\rm km\;s^{-1}}\)). Similar to the double-cloud system, all the molecular gas in the filamentary gas wisp has a similar velocity dispersion. At \(49.5^{\circ}<l<50.5^{\circ}\), the filamentary gas wisp seems to be split in both the position-position and position-velocity maps (middle and bottom panels of Fig. [\[figure:bubble\]](#figure:bubble){reference-type="ref" reference="figure:bubble"}). This coincides with the presence of a bubble in the infrared band. To summarize, the filamentary gas wisp is composed of two molecular clouds and one bubble. Star formation occurs in different parts of the filamentary gas wisp. Star formation in molecular clouds can be conveniently traced by 24 \(\mu\)m emission, which originates from the dust heated by newly-born stars. In the *Spitzer* image, this appears as red regions (Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"} top and Fig. [\[fig:double\]](#fig:double){reference-type="ref" reference="fig:double"}). In the cloud pair G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74, several star-forming sites can be identified (Fig. [\[fig:double\]](#fig:double){reference-type="ref" reference="fig:double"}) based on the *Spitzer* 24 \(\mu\)m emission, three of which are currently hosting compact H ii regions. At \(51.5^{\circ}<l<52.5^{\circ}\), all the star-forming sites are located at the edge of the G52L bubble. ## Distance and size of the filament The distance to the region has been estimated by several authors. Without trigonometric parallaxes, the distance to the region can be determined with the kinematic method. One key step in determining the kinematic distance is to resolve the kinematic distance ambiguity. There are different ways to resolve the ambiguity. Distance of the filamentary gas wisp can be determined by studying the distance to molecular clouds and H ii regions that belong to the filament. Using  self-absorption, found that the molecular clouds G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74 are located at the far distance. Using the H\(_2\)CO absorption line, found that the H ii region G52.23+0.74 is located at the far distance. Recently, and studied the distance to the H ii regions G052.201\(+\)`<!-- -->`{=html}0.752 and G052.259\(+\)`<!-- -->`{=html}0.700 with  emission/absorption method, and again found that they are at the far distance. Therefore we conclude that the filamentary gas wisp is located at the far distance, which is approximately 9.8 kpc. This suggests a galactocentric distance of 8.2 kpc, and the filamentary gas wisp probably resides in or around the Perseus arm. Accordingly, the filamentary gas wisp we identified has a spatial extent of \(\sim 500\) pc. If the filamentary gas wisp follows the spiral structure, it is probably angled \(\sim 45^{\circ}\) to our line of sight, and therefore probably has a deprojected length a factor of \(\sqrt{2}\) longer. Therefore we conclude that the filamentary gas wisp has a length of \(\gtrsim 500\) pc. It is one the of the largest coherent molecular structures in the Milky Way apart from the spiral arms and the molecular ring. The total mass of the filamentary gas wisp can be estimated using the \(^{13}\)CO(1-0) emission. To do this, we integrated over the region with the line-of-sight integrated flux \(I>3.5\rm\; K\;km\;s^{-1}\). This corresponds to the first contour in the upper panel of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"}. This mass estimate should be considered as a lower limit since by selecting this threshold we only take the region with a high column density (\(N_{H_2}>1.75\times 10^{21}\;\rm cm^{-2}\)) into account. Using Eq. 1--3 of and assuming an excitation temperature of \(10\;\rm K\), we obtain a total mass of \(\sim 1\times 10^{5} M_{\odot}\) for the whole gas wisp (\(49.5^{\circ}<l<52.5^{\circ}\)). According to, the derived mass is only weakly sensitive to this choice of excitation temperature, and in our case an excitation temperature of \(20\;\rm K\) gives a mass of \(\sim 0.6 \times 10^{5} M_{\odot}\). The two clouds at the eastern part of the filamentary gas wisp have \(b\sim 0.74^{\circ}\). Using a kinematic distance of \(9.77\;\rm kpc\), the double-cloud system is \(\sim 130\;\rm pc\) above the Galactic plane. At a Galactocentric distance of \(\sim 8\;\rm kpc\), the molecular disk of the Milky Way has a FWHM thickness of \(90-180\;\rm pc\) [at 7--8 kpc the FWHM is \(\sim 90\) pc and at 8--9 kpc the FWHM is 186 pc, @2006PASJ...58..847N]. This corresponds to an e-folding height of 38--80 pc. Therefore the height of the double-cloud system is about 1.5--4 times the e-folding height of the Galactic disk. The double-cloud system is a unique cloud system that is located far above the Galactic plane. According to, one possible explanation is that the material of the system has been displaced by the expansion of the G52L nebula. ## The bubble structure at \(l\sim 50^{\circ}\) Figure [\[figure:bubble\]](#figure:bubble){reference-type="ref" reference="figure:bubble"} shows the bubble structure at \(l\sim 50^{\circ}\). Its boundary is visible in both the 8 \(\mu\)m emission, which traces polycyclic aromatic hydrocarbon (PAHs) and in the \(^{13}\)CO(1-0) emission. The bubble is not easily visible at 24 \(\mu\)m, which traces hot dust heated by a central star. Because of the apparent absence of the 24 \(\mu\)m emission, the bubble structure does not seem to be driven by the expansion of a H ii region. This is also supported by the absence of a diffuse H ii region in the VGPS continuum image. It is more likely that the bubble structure is driven by the expansion of a supernova. The *Spitzer* image of the bubble resembles that of several supernova remnants in the catalog. From panel (a) of Fig. [\[figure:bubble\]](#figure:bubble){reference-type="ref" reference="figure:bubble"}, using the kinematic distance of 9.8 kpc, we estimate a diameter of \(1^{\circ}\sim180\;\rm pc\), and from panel (b) of Fig. [\[figure:bubble\]](#figure:bubble){reference-type="ref" reference="figure:bubble"} we estimate a total expansion velocity of \(10\;\rm km \;s ^{-1}\). These give an age of \(50\;\rm Myr\). The energy of a possible supernova explosion can be estimated through the Sedov-Taylor solution: \(E\sim r^3 v^2 \rho \sim 0.16 \times 10^{51}\;\rm erg\). The energy is consistent with a supernova explosion. Here, a density of \(10^{-24} \; \rm g\; cm^{-3}\) is used, which is typical of warm neutral medium. # Discussion {#sec:discussion} ## Morphology of the filamentary gas wisp This giant molecular structure is among the largest molecular structures studied in the Milky Way (\(\gtrsim 500\;\rm pc\)). The physical size of the gas filamentary gas wisp is much larger than that of a typical molecular cloud. The velocity dispersion along a single line of sight in the filamentary gas wisp is not significantly different from that of ordinary molecular clouds. The molecular gas in the filamentary gas wisp is concentrated in the vertical (Galactic-latitude) direction and elongated along the horizontal (Galactic-longitude) direction. At different locations, the filamentary gas wisp exhibits a different width. At \(49.5^{\circ}<l<50.5^{\circ}\), the filamentary gas wisp is split in the map, which makes it difficult to define its width. From the map, the cloud G052.24\(+\)`<!-- -->`{=html}00.74 appears to be more extended in the vertical direction than the cloud G051.69\(+\)`<!-- -->`{=html}00.74. We therefore used its vertical extent as an estimate of the width of the filamentary gas wisp. The vertical extent of the cloud G051.69\(+\)`<!-- -->`{=html}00.74 is measured for the region with \(I> 3.5\;\rm K\;km\;s^{-1}\) (\(N_{\rm H_2}>1.75\times 10^{21}\;\rm cm^{-2}\)). This corresponds to the first contour in the upper panel of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"}. We found that the cloud extends from \(b\sim 0.65^{\circ}\) to \(b\sim 0.82^{\circ}\). From this we estimated a diameter of 30 pc, which implies an aspect ratio of \(\sim 600/30=20\) for the gas wisp. The gas wisp is one of the most elongated molecular structures found in the Milky Way [see also the Nessie nebula, @2010ApJ...719L.185J]. The width of the filamentary gas wisp is narrower than the FWHM thickness of the molecular disk of the Milky Way, which is about 90--180 \(\;\rm pc\) at a Galactocentric distance of \(\sim 8\) kpc. Similar large-scale molecular structures have been observed in other galaxies. In spiral galaxies, elongated gas condensations are frequently observed. They can be seen as narrow dark lanes that extend perpendicular to the spiral arms. The exact definitions of dust lanes or spurs differ in the literature. However, in most cases spurs refer to the objects whose widths are similar to that of spiral arms. In our case, the filamentary gas wisp should not be termed a spur because its width is narrower than the width of the spiral arms of a typical galaxy, which is \(\sim 500\) pc. In our case, the filamentary gas wisp is about one or two orders of magnitudes longer than ordinary molecular clouds, but is still much narrower than the spurs in galaxies. Therefore we propose that the filamentary gas wisp is a new object that is yet to be classified. Because of this, we termed it a gas wisp in this work to emphasize its elongated morphology. Even though the thickness of the filamentary gas wisp is similar to the resolution of the PAWS survey of M51, filamentary gas wisps of this size would not be detected. This is because the survey is only sensitive to objects with a mass \(\gtrsim 1.2 \times 10^5 M_{\odot}\) and the clouds in the filamentary gas wisp are only \(\sim 10^{4} M_{\odot}\). However, similar large-scale gas structures in nearby galaxies are probably suitable targets for ALMA thanks to its improved sensitivity. Nearby face-on galaxies are expected to be excellent sites for studying these gas wisps since line-of-sight confusion can be avoided. A future project at ALMA targeting at the molecular gas in nearby face-on galaxies is expected to resolve similar gas condensations and provide a more complete picture of the structure of molecular gas in galaxies. ## Implications for the formation of molecular clouds The formation and evolution of molecular clouds is one of the fundamental problems in interstellar medium studies. To account for the short formation timescale of molecular clouds, two scenarios have been proposed. The first scenario involves colliding flows. In this scenario, molecular clouds form from diffuse gas (warm neutral medium) collected into a dense phase (cold neutral medium/molecular medium) through colliding flows. The molecular gas can form quickly in the converging flows because of dynamically-triggered thermal instability. The second scenario has been proposed by and. In this scenario, the gas is already relatively dense and cold prior to becoming a giant molecular cloud. According to, there is expected to be copious cold gas in the inter-arm regions since the circulation of the molecular gas is a process that occurs at the disk scale. To describe the global circulation of the gas, we divided it into the in-arm phase in which the gas is situated inside the spiral arms, and the inter-arm phase in which the gas is situated in the inter-arm regions. As discussed in, in the inter-arm phase, the cold gas exists in the form of wisps, and the gas in these wisps will show up as giant molecular clouds during the spiral-arm phase. This has been largely confirmed by the simulations of, which track the evolution of single molecular clouds. These authors found that molecular clouds begin to disperse as they leave the spiral arm. Due to differential shear, the molecular clouds are transformed into filamentary gas wisps in the inter-arm region. Since the shear occurs at a large scale, we expect to see gas wisps whose physical scale exceeds the thickness of the Milky Way disk. In our case, the physical length of the filamentary gas wisp (\(\gtrsim 500\) pc) is much larger than the scale-height of the Milky Way molecular disk. This is consistent with the cloud-formation scenario by and. Such large-scale structures are also observed in other numerical simulations of galactic disks [@2000ApJ...536..173T; @2002ApJ...570..132K; @2006MNRAS.367..873D; @2006MNRAS.371.1663D; @2006ApJ...647..997S; @2009ApJ...700..358T; @2012MNRAS.420.3490C]. On the other hand, it is difficult to understand the filamentary gas wisp in the converging flow scenario. In this scenario, molecular gas forms from the converging  gas through the dynamically-triggered thermal instability. As summarized in, the sources of the converging flows can be stellar winds or supernovae, turbulence in the interstellar medium, spiral shocks, and gravitational instability. In our case, the filamentary gas wisp cannot be created by converging stellar winds, supernovae, or turbulence in the interstellar medium, since these mechanisms are local and cannot create structures that are larger than the thickness of the Milky Way disk. Spiral shocks and gravitational instability might create conditions favorable for converging flows to occur. However, to access these possibilities we need to simulate converging flows in a galactic context and properly quantify the role of the dynamically-triggered thermal instability in the formation of molecular gas. This task has not been achieved yet. Either the converging flow scenario is unable to explain how filamentary gas wisps form, or our current understanding of converging flows in a galactic context is incomplete. ## Star formation in the molecular cloud pair G0524.2\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74 It is unclear to what extent molecular clouds are gravitationally bound. Gravity is important at a variety of physical scales during star formation. According to, the clouds G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74 have virial parameters of \(0.29\) and \(0.33\), respectively. This means that both clouds are gravitationally bound [^2]. Star formation takes place in several sites, which is traced by the 24 \(\mu m\) emission in the *Spitzer* image. As indicated in Fig. [\[fig:double\]](#fig:double){reference-type="ref" reference="fig:double"}, all these sites seem to be located at the edge of a bubble (G52L nebula). The connection between the location of the star-forming sites and the edge of the bubble agrees with the statistical study of, in which a significant overdensity of young stellar objects toward the edges of the bubbles was found. This is consistent with the collect-and-collapse scenario of triggered star formation. # Conclusions {#sec:conclusion} We studied a giant coherent molecular structure (a filamentary gas wisp) at \(49.5^{\circ}<l<52.5^{\circ}\). The eastern part of the filamentary gas wisp is located \(\sim 130\;\rm pc\) above the Galactic disk (which corresponds to 1.5--4 e-folding scale-heights), and the total mass of the gas wisp is \(\gtrsim 1 \times 10^5 M_{\odot}\). Apart from the spiral arms and the molecular ring, this is among the largest coherent molecular structures identified in the Milky Way. The velocity structure of the filamentary gas wisp is coherent and smooth at \(50.5^{\circ}<l<52.5^{\circ}\), and at \(49.5^{\circ}<l<50.5^{\circ}\), the gas wisp is disturbed by a bubble structure. This might be caused by a supernova. The overall velocity structure of the filamentary gas wisp can be understood as a quiescent filamentary gas wisp disturbed by the expansion of a bubble. The eastern part of the filamentary gas wisp is composed of a system of two molecular clouds (G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74) and is located \(\sim 130\;\rm pc\) above the Galactic plane. Star formation already takes place in several parts of this filamentary gas wisp. In the cloud pair G052.24\(+\)`<!-- -->`{=html}00.74 and G051.69\(+\)`<!-- -->`{=html}00.74, nearly the entire star formation occurs at the edge of a bubble [G52L nebula, @2012ApJ...759...96B]. This is consistent with the collect-and-collapse scenario of triggered star formation, and can be understood in the statistical context of. The discovery of this filamentary gas wisp, whose length exceeds the thickness of the molecular disk of the Milky Way, suggests that the formation and evolution of molecular clouds is a phenomenon that occurs at the disk scale. The large physical extent is consistent with the cloud-formation scenario by and, in which the gas that constitutes the molecular clouds is already relatively cold prior to the cloud formation. We are currently unable to answer how representative this filamentary gas wisp is in the Milky Way disk. One reason is that we are restricted by the line-of-sight confusion, and the filamentary gas wisp is fragile in nature. In our case, at \(49.5^{\circ}<l<50.5^{\circ}\), the filamentary gas wisp is already being destroyed by the expansion of a bubble structure. It is possible that a significant fraction of the gas in the Milky Way exists in this form during at least part of its lifetime. Another difficulty is to properly quantify the coherence of molecular structures beyond the cloud scale. A position-velocity plot of the \(^{13}\)CO(1-0) data from the same region shown in the bottom panel of Fig. [\[fig:region\]](#fig:region){reference-type="ref" reference="fig:region"} reveals filamentary structures at \(v_{\rm lsr}\sim 50\;\rm km\;s^{-1}\). From a visual inspection we found that these structures are not as coherent as the filamentary gas wisp (see Appendix [7](#sec:app2){reference-type="ref" reference="sec:app2"} for a comparison). In general one cannot yet quantify the coherence of molecular structures in the Milky Way. More studies of the morphology of the molecular gas in both the Milky Way and other galaxies with improved observations and analyses are needed to fully understand the circulation of molecular gas at large scales. [^1]: Note that report a much larger length of "many hundreds of pc" for the Nessie nebula. See also http://milkywaybones.org/ for more details. [^2]: Here and in the virial parameter \(\alpha\) of a molecular cloud is the ratio of its virial mass \(M_{\rm vir}\) to its mass.

{'timestamp': '2013-10-14T02:11:32', 'yymm': '1310', 'arxiv_id': '1310.3267', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3267'}

# Introduction In recent years, great advances have been made in the study of the extremal and structural properties of sparse random sets. For example, the threshold functions for many classical theorems, such as Szemerédi's theorem on arithmetic progressions and Turán's theorem in extremal graph theory, have been determined (see ), and in a few cases sharp thresholds have been shown to exist (see, e.g., ). In this paper we will determine the sharp threshold for the maximum sum-free subset problem in an arbitrary even-order abelian group. Our main theorem improves some recent results of Balogh, Morris and Samotij , who resolved the case \(G = \mathbb{Z}_{2n}\), and obtained weaker bounds in the general setting. Given an abelian group \(G\), we say that a subset \(A \subset G\) is *sum-free* if \(A \cap (A+A) = \emptyset\), or, equivalently, if there is no solution to the equation \(x + y = z\) with \(x,y,z \in A\). The study of such sets was introduced by Schur  in 1916, and their extremal and structural properties have been extensively studied over the past several decades (see, e.g., ). For example, it is easy to see that if \(|G| = 2n\) then the largest sum-free subset of \(G\) has size \(n\) (consider the odd coset of a subgroup of index 2), and in 1969 Diananda and Yap  extended this simple fact by solving the extremal problem whenever \(|G|\) has a prime divisor \(q\) with \(q \not\equiv 1 \pmod 3\). Nevertheless, more than 30 years passed before the classification was completed by Green and Ruzsa . The structure of a typical sum-free subset of an even-order abelian group was determined by Lev, Łuczak and Schoen  and Sapozhenko , and similar results in the set \(\{1,\ldots,n\}\) were obtained by Green  and Sapozhenko . We refer the reader to  for some more recent sparse refinements of these results. The study of sparse random analogues of classical extremal and Ramsey-type results was introduced for graphs by Frankl and Rödl  and Babai, Simonovits and Spencer , and for additive structures by Kohayakawa, Łuczak and Rödl , and notable early progress was made by Rödl and Ruciński . The first result of this type for sum-free sets was obtained by Graham, Rödl and Ruciński , who determined the threshold function for Schur's theorem. More precisely, they showed that if \(p \gg 1/\sqrt{n}\), then almost every \(p\)-random[^1] subset \(A \subset \mathbb{Z}_n\) has the following property: every 2-colouring of \(A\) contains a monochromatic Schur triple, i.e., a triple with \(x + y = z\). On the other hand, if \(p \ll 1/\sqrt{n}\) then with high probability there exist 2-colourings of \(A\) that avoid such triples. In this paper we consider the extremal version of this question, that is, how large is a maximum-size sum-free set in a \(p\)-random subset of an abelian group? For the group \(\mathbb{Z}_{2n}\), this problem was resolved (asymptotically) by Conlon and Gowers  and Schacht , who determined the following threshold: \[\label{eq:CGS} \max\big\{ |B| \,:\, B \subset A = (\mathbb{Z}_{2n})_p \textup{ is sum-free} \big\} \,=\, \left\{ \begin{array} {c@{\quad \textup{if} \quad}l} \big( 1+o(1) \big) \cdot 2pn & p \ll 1/\sqrt{n} \\[+1ex] \big( 1/2 + o(1) \big) \cdot 2pn & p \gg 1/\sqrt{n} \end{array}\right.\] with high probability as \(n \to \infty\). More precisely, one can show using the methods of  (see ), and also using those of , that (with high probability) the maximum-size sum-free subsets of \(A\) contain only \(o(pn)\) even numbers. Moreover, a corresponding result holds for any even-order abelian group. This fact will be a key tool in the proof below. We will be interested in the following more precise question, which was first studied by Balogh, Morris and Samotij . Given an even-order abelian group \(G\), note that the maximum-size sum-free subsets of \(G\) are exactly the odd cosets of subgroups of index 2, and that a \(p\)-random subset \(A \subset G\) has a sum-free subset of (expected) size \[\label{eq:lowerbound} \max\big\{ |A \cap \mathcal{O}| \,:\, \mathcal{O} \textup{ is the odd coset of a subgroup of index 2} \big\} \, \geqslant \, \left( \frac{1}{2} + o(1) \right) p |G|.\] For which functions \(p = p(n)\) is it true that, with high probability, the size of the largest sum-free subset of \(A\) is equal to the left-hand side of [\[eq:lowerbound\]](#eq:lowerbound){reference-type="eqref" reference="eq:lowerbound"}? In other words, for which densities does the *exact* extremal result in \(G\) transfer to the sparse random setting? Proving such exact extremal results is often extremely difficult; for example, the threshold for Mantel's theorem was determined only very recently by DeMarco and Kahn . Nevertheless, it was shown in  that the threshold for this property is \(\big( \frac{\log n}{n} \big)^{1/2}\) for every even-order[^2] abelian group, and moreover that there is a sharp threshold at \(\big( \frac{\log n}{3n} \big)^{1/2}\) in the group \(\mathbb{Z}_{2n}\). In other words, writing \(\textup{SF}(A)\) for the collection of maximum-size sum-free subsets of \(A\), and \(\mathcal{O}_{2n}\) for the set of odd numbers in \(\mathbb{Z}_{2n}\), they proved that for every \(\varepsilon>0\), \[\mathbb{P}\Big( \textup{SF}\big( (\mathbb{Z}_{2n})_p \big) = \big\{ (\mathbb{Z}_{2n})_p \cap \mathcal{O}_{2n} \big\} \Big) \to \left\{ \begin{array} {c@{\quad \textup{if} \quad}l} 0 & p \leqslant \big(1-\varepsilon \big) \sqrt{ \frac{\log n}{3n} } \\[+1ex] 1 & p \geqslant \big(1 + \varepsilon \big) \sqrt{ \frac{\log n}{3n} } \end{array}\right.\] as \(n \to \infty\). For more on the general theory of the existence of (sharp) thresholds, we refer the reader to , and to  for an example involving monochromatic triangles. Since Balogh, Morris and Samotij  were able to prove such a sharp threshold for the group \(\mathbb{Z}_{2n}\), but only a weaker threshold result for other even-order abelian groups, it is natural to ask whether one can also obtain a more precise result in the general setting. In this paper we answer this question in the affirmative, by determining the sharp threshold for every even-order abelian group. In order to state our main theorem, we shall need the following function, which determines the location of the sharp threshold. We encourage the reader to think of \(\delta\) as a function going to zero slowly, and \(n\) as a function going to infinity much faster. The following theorem is our main result. Here, as usual, \(o(1)\) denotes a function that tends to zero as \(n \to \infty\). We shall refer to the two bounds as the 0-and 1-statements respectively. The proof of Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"} uses the method of , but we will require several substantial new ideas in order to overcome various obstacles which do not occur in the case \(G = \mathbb{Z}_{2n}\). Many of these arise from the fact that \(\textup{SF}(G)\) can be quite large (as big as \(|G|\) in the case of the hypercube), which means that we must obtain much stronger bounds than in  if we wish to apply the union bound. For the 0-statement we shall do this using a recent concentration inequality of Warnke , which allows us to deduce for almost all \(\mathcal{O} \in \textup{SF}(G)\) that, with very high probability, the set \(A \cap \mathcal{O}\) is not a maximal sum-free set. For the 1-statement, however, such a straightforward strategy is not feasible, since the threshold for the event that \(A \cap \mathcal{O}\) is maximal for *every* odd coset \(\mathcal{O} \in \textup{SF}(G)\) is not given by \(\lambda^{(\delta)}(G)\). In order to avoid this problem, we need to show that \(A \cap \mathcal{O}\) is a maximal sum-free set for each \(\mathcal{O} \in \textup{SF}(G)\) such that \(|A \cap \mathcal{O}|\) is maximal. Unfortunately, conditioning on the size of \(A \cap \mathcal{O}\) introduces significant dependence between odd cosets, and our first attempts to prove the 1-statement failed as a consequence. We resolve this issue by fixing the *number* of elements of \(A\) (i.e., coupling with the hypergeometric distribution), which essentially eliminates the positive correlation between the quantities \(|A \cap \mathcal{O}|\) for different cosets. A third issue involves the analysis of the Cayley graphs \(\mathcal{G}_S\) for each \(S \subset \mathcal{E}\), where \(\mathcal{E}\) is a subgroup of index \(2\), \(V(\mathcal{G}_S) = \mathcal{O}\) (the corresponding odd coset) and \(xy \in E(\mathcal{G}_S)\) if either \(x + y \in S\) or \(x-y \in S\). Although counting the edges in these graphs precisely is not entirely trivial, we are fortunate that we can absorb most of the resulting mess into an error term. However, we still need to do some rather careful (and delicate) counting of the number of sets \(S\) that contain a given number of edges of \(\mathcal{H}_W\), the Cayley graph of the set \(W = \{ a + a: a \in \mathcal{O} \}\), since this controls the size of \(e(\mathcal{G}_S)\), see Section [3](#sec:edge:counts){reference-type="ref" reference="sec:edge:counts"}. The remainder of the paper is organized as follows. In Section [2](#sec:preliminaries){reference-type="ref" reference="sec:preliminaries"}, we recall the structural version of [\[eq:CGS\]](#eq:CGS){reference-type="eqref" reference="eq:CGS"} for even-order abelian groups from , and collect some probabilistic tools and simple group-theoretic facts that will be needed later. In Section [3](#sec:edge:counts){reference-type="ref" reference="sec:edge:counts"} we analyse the Cayley graph \(\mathcal{G}_S\) for each set \(S \subset \mathcal{E}\), where \(\mathcal{E}\) is a subgroup of index 2, and count the number of such sets \(S\) whose Cayley graph has fewer edges than expected. In Section [4](#sec:zero:statement){reference-type="ref" reference="sec:zero:statement"} we deduce the 0-statement from Warnke's concentration inequality (see Section [2.2](#sec:prob:tools){reference-type="ref" reference="sec:prob:tools"}), together with some of the more straightforward bounds from Section [3](#sec:edge:counts){reference-type="ref" reference="sec:edge:counts"}. Finally, in Section [5](#sec:one:statement){reference-type="ref" reference="sec:one:statement"} we prove the 1-statement of Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"} using the method of , combined with the coupling argument and careful counting described above. We end the paper with a short Appendix, which contains a somewhat technical calculation involving the hypergeometric distribution. # Preliminaries {#sec:preliminaries} In this section we shall lay the groundwork necessary for the proof of our main theorem. In particular, we will recall the asymptotic stability version of Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"}, which was proved in  using the method of . We will also recall the FKG inequality and the concentration inequalities of Warnke and Janson, and state some simple facts about abelian groups that will be useful in the proof. ## Sparse stability for sum-free sets We begin by recalling the following theorem from , which determines the asymptotic structure of the maximum-size sum-free subsets in a \(p\)-random subset of an even-order abelian group. The theorem follows by either the method of Conlon and Gowers , or that of Schacht  (as modified by Samotij ), or that of Balogh, Morris and Samotij  and Saxton and Thomason , in each case using results of Lev, Łuczak, and Schoen  and Green and Ruzsa . We refer the reader to Sections 2 and 3 of  for the details. We remark that the probability of failure in Theorem [\[thm:approx:stability\]](#thm:approx:stability){reference-type="ref" reference="thm:approx:stability"} is exponentially small in \(pn\). ## Probabilistic tools {#sec:prob:tools} Our main tool for the 0-statement will be the following concentration inequality, recently proved[^3] by Warnke . We also recall two well-known probabilistic inequalities: Janson's inequality and the FKG inequality. We refer the reader to  for various more general statements and their proofs. ## Group-theoretic facts In order to avoid repetition, we shall assume throughout the paper that \(G\) is a finite abelian group of order \(2n\). Given a subset \(X \subset G\), we write - \(R(X)\) for the collection of elements \(x \in X\) for which \(x=-x\), and \(r(X) = |R(X)|\). - \(m(X)\) for number of two-element subsets of \(X\) that are of the form \(\{x,-x\}\). We will need a few basic facts about finite abelian groups. The first one is well-known. The second fact we need is a characterization of the index \(2\) subgroups of \(G\). Note that Fact [\[mainfact\]](#mainfact){reference-type="ref" reference="mainfact"} implies that \(G\) has exactly \(r(G)-1\) index 2 subgroups. Finally, we make a simple but useful observation. # Edge counts in Cayley graphs {#sec:edge:counts} In order to bound the probability of the event "\(A \cap \mathcal{O} \in \textup{SF}(A)\)\" for some fixed maximum-size sum-free set \(\mathcal{O} \in \textup{SF}(G)\) and its corresponding set of evens \(\mathcal{E} = G \setminus \mathcal{O}\), we will need to consider events of the form \[\textup{``\(\big( (A \cap \mathcal{O}) \cup S \big) \setminus T\) is sum-free"}\] where \(S \subset A \cap \mathcal{E}\), \(T \subset A \cap \mathcal{O}\) and \(|S| \geqslant |T|\). This event is contained in the event that \((A \cap \mathcal{O}) \setminus T\) is an independent set in the Cayley graph \(\mathcal{G}_S\), defined below, and to bound its probability we will need to analyse carefully the number of edges in this Cayley graph for each such set \(S\) of evens. In particular, there may be an exceptional collection of sets \(S\) with too few edges for our purposes (that is, for our application of the union bound over all sets \(S\)), and we will need to bound the size of this collection. Let us begin by stating precisely the main results we will prove in this section. We fix throughout an arbitrary \(\varepsilon > 0\), a sufficiently small \(\delta > 0\) and a sufficiently large \(n \in \mathbb{N}\).[^4] We also fix an abelian group \(G\) of order \(2n\), an odd coset \(\mathcal{O} \in \textup{SF}(G)\), and its corresponding set of evens \(\mathcal{E} = G \setminus \mathcal{O}\), which is a subgroup of \(G\) of index 2. For each set \(S \subset \mathcal{E}\), we define the Cayley graph \(\mathcal{G}_S\) of \(S\) to have vertex set \(\mathcal{O}\) and edge set \[E(\mathcal{G}_S) \, = \, \bigg\{ \{y,z\} \in \binom{\mathcal{O}}{2} \,:\, y+z \in S \textup{ or } y-z \in S \bigg\},\] where (for simplicity) we do not permit \(\mathcal{G}_S\) to have loops. Recall that we write \(r(X)\) for the number of order 2 elements in \(X \subset G\), and \(m(X)\) for the number of pairs \(\{x,-x\} \subset X\). We will prove the following propositions. When \(r(G) \geqslant \delta n\) the edge counts are slightly different. In order to prove Propositions [\[prop:edge:counts\]](#prop:edge:counts){reference-type="ref" reference="prop:edge:counts"} and [\[prop:edge:counts:bigR\]](#prop:edge:counts:bigR){reference-type="ref" reference="prop:edge:counts:bigR"}, we will first count edges in \(\mathcal{G}_x = \mathcal{G}_{\{x\}}\) for each \(x \in \mathcal{E}\), and then study the intersections between these graphs. These will depend on the parameter \(r(S)\), as the reader can see from the statement. However, they will also depend on the intersection of \(S\) with the set \[W = \{a+a: a \in \mathcal{O}\},\] and with its Cayley graph. We will use several times the fact that \(|W| = n / r(\mathcal{E})\). ## Edge counts in \(\mathcal{G}_x\) We begin with the relatively simple task of counting the edges in the Cayley graph of a single vertex \(x\). To be precise, we will prove the following lemma. Lemma [\[lem:count:Gx\]](#lem:count:Gx){reference-type="ref" reference="lem:count:Gx"} has the following simple consequence, which we shall use several times. Before continuing to the proof of Proposition [\[prop:edge:counts\]](#prop:edge:counts){reference-type="ref" reference="prop:edge:counts"}, let us note how to obtain (heuristically) the function \(\lambda^{(\delta)}(G)\) from Lemma [\[lem:count:Gx\]](#lem:count:Gx){reference-type="ref" reference="lem:count:Gx"}. We call an element \(0 \ne x \in \mathcal{E}\) *safe* if \((A \cap \mathcal{O}) \cup \{x\}\) is sum-free, and let \(S^\mathcal{E}(A)\) denote the collection of safe elements in \(\mathcal{E}\). Note that an element \(x \in \mathcal{E}\) is safe if[^5] and only if \(A \cap \mathcal{O}\) is an independent set in \(\mathcal{G}_x\). We need one more definition, whose slightly odd appearance will be motivated by the lemmas below. The next lemma says that almost all index \(2\) subgroups are nice. We now prove the following bound on the expected number of safe elements, which we will use in the proof of the 0-statement of Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"}. Fortunately, however, \(|W| = n / r(\mathcal{E}) = O(1/\delta)\). We can therefore easily deduce a lower bound on \(\mathbb{E}[|S^\mathcal{E}(A)|]\) for nice subgroups. Indeed, since \(r(\mathcal{O}) = r(\mathcal{E}) = \beta(G) n/2\), and again using the unimodality of \(x e^{-c x^2}\), it follows from Table [1](#table:thetable){reference-type="ref" reference="table:thetable"} above that \[\label{eq:safe:hypercube} \mathbb{E}\big[ |S^\mathcal{E}(A)| \big] \, \geqslant \, \sum_{x \in R(\mathcal{E})} \big( 1-p^2 \big)^{e(\mathcal{G}_x)} = \, \Omega \Big( r(\mathcal{E}) e^{-p^2 ( n-r(\mathcal{O})/2 )} \Big) \, \gg \, \frac{\log n}{p},\] as required, where the last step follows since \(1-\big( 1-\beta(G)/4 \big) \lambda^{(\delta)}(G) = 1/2\). ◻ Moreover, we can bound the size of each intersection. Let us denote by \(\mathcal{H}_W\) the graph on vertex set \(\mathcal{E}\) with edge set \(\{xy: x+y \in W \}\), and note that we have \(\Delta\big( \mathcal{H}_W \big) \leqslant d\), where \(d := |W| = n/r(\mathcal{E})\). By Observations [\[obs:GxGy:exists\]](#obs:GxGy:exists){reference-type="ref" reference="obs:GxGy:exists"} and [\[obs:GxGy:size\]](#obs:GxGy:size){reference-type="ref" reference="obs:GxGy:size"}, we have \[\label{eq:EGxEGy} \sum_{x,y \in S, \, x \ne y} \big| E(\mathcal{G}_x) \cap E(\mathcal{G}_y) \big| \, \leqslant \, 2 \cdot r(\mathcal{E}) \cdot e(\mathcal{H}_W[S])\] for every \(S \subset \mathcal{E}\) with \(m(S) = 0\). Since, by Lemma [\[lem:count:Gx\]](#lem:count:Gx){reference-type="ref" reference="lem:count:Gx"}, we have good bounds on the sum of \(e(\mathcal{G}_x)\) over \(x \in S\), the following lemma is all we need to complete the proof of Proposition [\[prop:edge:counts\]](#prop:edge:counts){reference-type="ref" reference="prop:edge:counts"}. We are now ready to prove the two propositions. # Proof of the 0-Statement {#sec:zero:statement} In this section we will prove that if \(A \subset G\) is a \(p\)-random set and \[\label{eq:pbounds:zerostatement} \frac{\log n}{n} \, \ll \, p \leqslant \big(1-\varepsilon \big) \sqrt{ \lambda^{(\delta)}(G) \frac{\log n}{n} },\] then \(A \cap \mathcal{O} \not\in \textup{SF}(A)\) for every \(\mathcal{O} \in \textup{SF}(G)\) with high probability as \(n \to \infty\). The main step will be proving the following proposition.[^6] Recall also that at most \(O(1/\delta)\) of the index 2 subgroups of \(G\) are not nice. We will use the following simple-sounding lemma to deal with these subgroups. The proof of Lemma [\[lem:M:nice\]](#lem:M:nice){reference-type="ref" reference="lem:M:nice"}, although not difficult, is surprisingly technical, and so we shall postpone it to the appendix. Note that the 0-statement in Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"} follows from Proposition [\[prop:zero:statement\]](#prop:zero:statement){reference-type="ref" reference="prop:zero:statement"} and Lemma [\[lem:M:nice\]](#lem:M:nice){reference-type="ref" reference="lem:M:nice"} by taking a union bound over nice subgroups. Recall that an element \(x \in \mathcal{E}\) is called *safe* if \((A \cap \mathcal{O}) \cup \{x\}\) is sum-free, and that \(S^\mathcal{E}(A)\) denotes the collection of safe elements in \(\mathcal{E}\). We will bound the probability of the event \(A \cap \mathcal{O} \in \textup{SF}(A)\) by the probability that there exists no safe element \(x \in A \cap \mathcal{E}\). Since the random variable \(S^\mathcal{E}(A)\) is independent of the set \(A \cap \mathcal{E}\), it follows that \[\label{eq:zero:Chernoff} \mathbb{P}\bigg( \Big( A \cap \mathcal{O} \in \textup{SF}(A) \Big) \cap \bigg( |S^\mathcal{E}(A)| \geqslant \frac{3 \log n}{p} \bigg) \bigg) \, \leqslant \, \big( 1-p \big)^{(3 \log n) / p} \, \leqslant \, \frac{1}{n^3},\] and so it is enough to consider the event that \(|S^\mathcal{E}(A)| \leqslant (3 \log n) / p\). We will bound the probability of this event using Warnke's concentration inequality, which was stated in Section [2.2](#sec:prob:tools){reference-type="ref" reference="sec:prob:tools"}. The first step--showing that \(|S^\mathcal{E}(A)|\) has large expected value--was already carried out in the previous section. Indeed, we have \[\label{eq:safe:logn} \mathbb{E}\big[ |S^\mathcal{E}(A)| \big] \, \gg \, \frac{\log n}{p}\] whenever \(p \in (0,1)\) satisfies [\[eq:pbounds:zerostatement\]](#eq:pbounds:zerostatement){reference-type="eqref" reference="eq:pbounds:zerostatement"}, by Lemma [\[lem:expected:safe\]](#lem:expected:safe){reference-type="ref" reference="lem:expected:safe"}. Our main task will be to prove the following lemma, which shows that \(|S^\mathcal{E}(A)|\) is concentrated around its expected value. We will prove Lemma [\[lem:Xconcentration\]](#lem:Xconcentration){reference-type="ref" reference="lem:Xconcentration"} by applying Warnke's inequality to the function \(A \mapsto |S^\mathcal{E}(A)|\). In order to do so, we need to define an event \(\Gamma \subset \mathcal{P}(\mathcal{O})\), and prove the 'typical Lipschitz condition' \[\label{eq:typLips} \big| |S^\mathcal{E}(A)|-|S^\mathcal{E}(B)| \big| \, \leqslant \, \left\{ \begin{array} {c@{\quad}l} c(\mathcal{E},p) := n^{-(1/4 + \delta)} \cdot \mathbb{E}\big[ S^\mathcal{E}(A) \big] & \textup{if } A \in \Gamma, \\ n & \textup{otherwise} \end{array}\right.\] for every \(A,B \subset \mathcal{O}\) with \(| A \triangle B | = 1\) (note that \(c(\mathcal{E}, p) \gg 1\), by [\[eq:safe:logn\]](#eq:safe:logn){reference-type="eqref" reference="eq:safe:logn"}). We define the event \(\Gamma\) so that [\[eq:typLips\]](#eq:typLips){reference-type="eqref" reference="eq:typLips"} holds by definition: \[\label{def:gamma} \Gamma \, := \, \Big\{ A \subset \mathcal{O} \,:\, \max\big\{ \big| |S^\mathcal{E}(A)|-|S^\mathcal{E}(B)| \big|: | A \triangle B | = 1 \big\} \leqslant c(\mathcal{E},p) \Big\}.\] We would like to show that \(\mathbb{P} \big( A \not\in \Gamma \big) \leqslant n^{-5}\), since this will imply the desired upper bound on the probability of the event \(\mathcal{B}\) given by Warnke's inequality. The main technical step in the proof of Lemma [\[lem:Xconcentration\]](#lem:Xconcentration){reference-type="ref" reference="lem:Xconcentration"} is proving such a bound on the probability that \(A \not\in \Gamma\). To do so, note first that if \(A \notin \Gamma\) then there exists \(u \in \mathcal{O}\) such that \(\big||S^\mathcal{E}(A)|-|S^\mathcal{E}(A \Delta \{u\})|\big|>c(\mathcal{E},p)\). Let \(\Gamma^c(u)\) be the set of choices of \(A\) for which this property holds, so that \(\Gamma^c = \bigcup_{u \in \mathcal{O}}\Gamma^c(u)\), and note that, by symmetry,[^7] \[\label{eq:xnotinA} \mathbb{P}\big( A \in \Gamma^c(u) \,\big|\, u \in A \big) \, = \, \mathbb{P}\big( A \in \Gamma^c(u) \,\big|\, u \not\in A \big).\] We will bound \(\mathbb{P}\big(A \in \Gamma^c(u) \big)\) for each fixed \(u \in \mathcal{O}\), and then sum over \(u\). Motivated by [\[eq:xnotinA\]](#eq:xnotinA){reference-type="eqref" reference="eq:xnotinA"}, let us fix \(u \in \mathcal{O}\), assume that \(u \not\in A\), and write \[Y_u^\mathcal{E}(A) \, = \, S^\mathcal{E}(A) \setminus S^\mathcal{E}(A \cup \{u\}).\] Observe that \(A \in \Gamma^c(u)\) if and only if \(|Y_u^\mathcal{E}(A)| > c(\mathcal{E},p)\). We will prove the following lemma. Note that the first inequality follows from the comments above and Markov's inequality. The intuition behind the second inequality is based on our expectation that \(|Y_u^\mathcal{E}(A)| =\Theta(p\big|S^\mathcal{E}(A)|\big)\), and that the events \(\big\{ z \in Y_u^\mathcal{E}(A): z \in \mathcal{E} \big\}\) are more or less independent of one another. We expect \(|Y_u^\mathcal{E}(A)|\) to take roughly this value since \(Y_u^\mathcal{E}(A) \subset S^\mathcal{E}(A)\), and moreover for each \(z \in Y_u^\mathcal{E}(A)\) there is a \(v \in \mathcal{O}\) with \(uv \in E(\mathcal{G}_z)\) such that \(v \in A\). In order to make this argument precise, the following notion will be crucial. Fix \(u \in \mathcal{O}\), and say that a set \(0 \ne Z \subset \mathcal{E}\) is *covered* by \(Y \subset \mathcal{O}\) if for each \(z \in Z\) there is a \(y \in Y\) such that \(uy \in E(\mathcal{G}_z)\). Say that \(Z\) is *cover-maximal* if \(|Y| \geqslant |Z|\) for every set \(Y\) that covers \(Z\), and for each \(Z \subset \mathcal{E}\) choose a maximum-size cover-maximal subset \(g(Z) \subset Z\). Note that since any singleton in \(Z\) is cover-maximal, \(g(Z)\) is non-empty. The following lemma is key. We also need the following simple observation, which follows easily from the definition. We are ready to prove Lemma [\[lem:prob:gamma\]](#lem:prob:gamma){reference-type="ref" reference="lem:prob:gamma"}. It is now straightforward to deduce Lemma [\[lem:Xconcentration\]](#lem:Xconcentration){reference-type="ref" reference="lem:Xconcentration"}, and hence Proposition [\[prop:zero:statement\]](#prop:zero:statement){reference-type="ref" reference="prop:zero:statement"}. The 0-statement now follows immediately. # Proof of the 1-statement {#sec:one:statement} In this section we will prove that if \(A \subset G\) is a \(p\)-random set and \[p \geqslant \big(1 + \varepsilon \big) \sqrt{ \lambda^{(\delta)}(G) \frac{\log n}{n} },\] then every \(B \in \textup{SF}(A)\) is equal to \(A \cap \mathcal{O}\) for some \(\mathcal{O} \in \textup{SF}(G)\), with high probability as \(n \to \infty\). The proof has three steps: an application of Theorem [\[thm:approx:stability\]](#thm:approx:stability){reference-type="ref" reference="thm:approx:stability"} to obtain an asymptotic version, an argument for a given odd coset \(\mathcal{O} \in \textup{SF}(G)\), using the method of  (see Lemma [\[lem:BMSmethod\]](#lem:BMSmethod){reference-type="ref" reference="lem:BMSmethod"}), and a comparison with the hypergeometric distribution, which allows us to a partition the odd cosets depending on the size of \(A \cap \mathcal{O}\) (see Lemma [\[le:existsb\]](#le:existsb){reference-type="ref" reference="le:existsb"}). Recall throughout that we have already fixed an arbitrary \(\varepsilon>0\), a sufficiently small \(\delta >0\) and a sufficiently large \(n\in\mathbb{N}\). We begin by proving the statement we will require for a given odd coset \(\mathcal{O} \in \textup{SF}(G)\). For each \(k \in \mathbb{N}\), let \(\mathcal{B}_k^\mathcal{O}(A)\) denote the event that there exist sets \(S \subset A \cap \mathcal{E}\) and \(T \subset A \cap \mathcal{O}\), with \(|S| = k \geqslant |T|\), such that \(\big( (A \cap \mathcal{O} ) \cup S \big)\setminus T\) is sum-free. Let us denote by \(\mathbb{P}_{p^\pm} = \mathbb{P}^\mathcal{O}_{p^\pm}\) the probability distribution in Lemma [\[lem:BMSmethod\]](#lem:BMSmethod){reference-type="ref" reference="lem:BMSmethod"}, in which each element of \(\mathcal{O}\) is chosen (independently) with probability \((1-\delta)p\) and each element of \(\mathcal{E}\) is chosen with probability \((1+\delta)p\). Note that the event \(\mathcal{B}^\mathcal{O}_k(A)\) is increasing in \(A \cap \mathcal{E}\) and decreasing in \(A \cap \mathcal{O}\), so \(\mathbb{P}_p\big(\mathcal{B}^\mathcal{O}_k(A) \big) \leqslant \, \mathbb{P}_{p^\pm}\big(\mathcal{B}^\mathcal{O}_k(A) \big)\) for every \(\delta \geqslant 0\). In order to deduce the 1-statement in Theorem [\[thm:even:abelian\]](#thm:even:abelian){reference-type="ref" reference="thm:even:abelian"} from Lemma [\[lem:BMSmethod\]](#lem:BMSmethod){reference-type="ref" reference="lem:BMSmethod"}, we cannot simply apply the union bound over odd cosets \(\mathcal{O} \in \textup{SF}(G)\), since an even-order abelian group \(G\) can have as many as \(|G|\) distinct maximum-size sum-free subsets. On the other hand, Lemma [\[lem:BMSmethod\]](#lem:BMSmethod){reference-type="ref" reference="lem:BMSmethod"} (together with Theorem [\[thm:approx:stability\]](#thm:approx:stability){reference-type="ref" reference="thm:approx:stability"}) does imply that the maximum-size sum-free subset of \(A\) contains (with high probability) only \(O(1)\) even elements, and moreover that any *given* collection of \(n^{o(1)}\) odd cosets are all likely to be 'locally' maximal. Motivated by these observations, it is natural to attempt to partition the odd cosets into two classes, depending on whether or not \(|A \cap \mathcal{O}|\) is within \(O(1)\) of \(\max_{\mathcal{O}'} |A \cap \mathcal{O}'|\). However, the random variables \(\{ |A \cap \mathcal{O}'|: \mathcal{O}' \in \textup{SF}(G) \}\) are highly correlated with one another, due to the large (size \(n/2\)) overlap between different odd cosets, and for this reason the maximum is not easy to control.[^13] We resolve this problem by coupling with the hypergeometric distribution, for which the positive correlation between the variables \(|A \cap \mathcal{O}|\) is greatly diminished. (In fact, these variables are roughly pairwise independent of one another.) For each \(0 \leqslant m \leqslant 2n\), let \(\mathbb{P}_m\) denote the probability measure on subsets of \(G\) obtained by choosing each subset of size \(m\) with equal probability. Note that, since any pair of distinct subgroups \(\mathcal{E},\mathcal{E}' \subset G\) of index \(2\) intersect in a subgroup of index \(4\), the information that \(|A \cap \mathcal{O}| \geqslant a\) (and therefore \(|A \cap \mathcal{E}| \leqslant m-a\)) has very little influence on the probability that \(|A \cap \mathcal{O}'| \geqslant a\). This crucial property of the hypergeometric distribution is captured by the following lemma. Given \(k \in \mathbb{N}\) and an odd coset \(\mathcal{O} \in \textup{SF}(G)\), define \(M^\mathcal{O}_k(A)\) to be the event that \(|A \cap \mathcal{O}| \geqslant k\), and let \[X_k(A) \, := \, \sum_{\mathcal{O} \in \textup{SF}(G)} \mathbbm{1}\big[ M^\mathcal{O}_k(A) \big]\] denote the number of odd cosets \(\mathcal{O} \in \textup{SF}(G)\) for which \(|A \cap \mathcal{O}| \geqslant k\). The proof of Lemma [\[le:existsb\]](#le:existsb){reference-type="ref" reference="le:existsb"} involves some straightforward but technical approximations of binomial coefficients, and so we defer it to an Appendix. Let us denote by \(\mathcal{C}_k^\mathcal{O}(A)\) the event that \(|A \cap \mathcal{O}'| < |A \cap \mathcal{O}| + k\) for every \(\mathcal{O}' \in \textup{SF}(G)\). We are now ready to complete the proof of our main theorem. [^1]: The \(p\)-random subset of a set \(X\), often denoted \(X_p\), is obtained by including each element with probability \(p\), independently of all other elements. [^2]: In fact Theorem 1.1 of  is more general: it determines the threshold for any abelian group whose order has a (fixed) prime factor \(q\) with \(q \equiv 2 \pmod 3\). Here, as before, we set \(|G| = qn\). [^3]: In fact the theorem stated here is only a special case of Warnke's inequality; for the sake of simplicity, we have chosen to state only the version we need. [^4]: We think of \(\delta\) as a function of \(n\) which tends to zero sufficiently slowly as \(n \to \infty\). [^5]: This is only true if we ignore sums of the form \(x = y + y\). However, such sums will never play a significant role in any of the calculations below. [^6]: We remark that the bound \(1/n^2\) could easily be replaced by \(1/n^C\) for any \(C > 0\). [^7]: Indeed, if \(B = A \Delta \{u\}\) then \(A \in \Gamma^c(u) \Leftrightarrow B \in \Gamma^c(u) \Leftrightarrow \big| |S^\mathcal{E}(A)|-|S^\mathcal{E}(B)| \big| > c(\mathcal{E},p)\). [^8]: When \(s = \ell\), we trivially bound the number of sets \(Z\) such that \(e(\mathcal{G}_Z) \geqslant \ell \left(n-\frac{r(\mathcal{O})}{2}\right)\) by \(n^\ell\). [^9]: Alternatively, we may simply carry this factor of \(C^{2k}\) through the proof, and perform an easy but tedious calculation later on. [^10]: If \(C \geqslant 2/3\) then simply note that the previous line is decreasing in \(C\), since \(3(k-m)-r \geqslant 2k-m \geqslant k\). [^11]: The case \(i = O(1)\) was already covered by the proof in part \((a)\). [^12]: When \(s = k\), we trivially bound the number of sets \(Z\) such that \(e(\mathcal{G}_{\hat{S}}) \geqslant k \left(n-\frac{r(\mathcal{O})}{2}\right)\) by \(n^k\). [^13]: The behaviour of the random variable \(\max_{\mathcal{O}'} |A \cap \mathcal{O}'|\) is in fact somewhat mysterious, and we believe that it merits further investigation. [^14]: Note that \(p \geqslant C / \sqrt{n}\) since \(n \geqslant n_0(\varepsilon,\delta)\) is sufficiently large.

{'timestamp': '2014-07-22T02:01:48', 'yymm': '1310', 'arxiv_id': '1310.3236', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3236'}

# Introduction Different approaches to quantum gravity such as String Theory\[1-5\], non commutative geometry\[6\], loop quantum gravity\[7\], and Doubly Special Relativity predict the existence of a minimal measurable length or a maximal observable momentum\[8,9\]. These theories argue that near the Planck scale, the Heisenberg Uncertainty Principle should be replaced by the so called Generalized Uncertainty principle(GUP)\[10,11\]. In these paper we want to study the effects of this GUP on the statistical mechanics of ideal gas. In ordinary standard statistical mechanics, it is impossible to define the position of a representative point in the phase space of the given system more accuracy than the situation which is given by \((\Delta x \Delta p)\geq\hbar\). It means that around any point (q,p) of the (two dimensional) phase space, there exist an area of the order of \(\hbar\) which can't say where the exact position of the particle is. In \(2N\) dimensional phase space, the corresponding volume of uncertainty around any point would be of order \(\hbar^{N}\). So we can assume approximately that the phase space is made of elementary cells of volume \(\hbar^{N}\). These cells have one to one correspondence with the quantum mechanics states of the given system\[12\]. Not that an elegant formulation of statistical mechanics of multi-dimensional Cantor sets based on fractal nature of space-time has been provided by El Naschie\[13\]. Although, this issue has been studied by some authors \[14\], but they ignored this fact that a minimal measurable length essentially requires the existence of a maximal momentum encoded in the duality of position-momentum spaces or uncertainty principle. Existence of a maximal measurable momentum for a test particle modifies the results of the mentioned studies considerably. In which follow we consider the mentioned GUP, as our primary input. We consider a micro-cononical ensemble and then study the thermodynamical properties of an ideal gas composed of monatomic non-interacting particles within mentioned GUP. For this reason, we study the total number of microstates of the given system and find the novel and unusual results. One may ask about the possible detection of these extra ordinary effects. Up to now, there is no direct experimental or observational scheme for detecting of these novel effects. Nevertheless, since the basis of our calculations come back to the GUP, possible experimental schemes for gravity predictions have been proposed\[15\]. Therefore any search for quantum gravity signals provides possible indirect test of generalized statistical mechanics which we want to consider. It should be mentioned that these results are only important in the limit of high momentum and without these consideration, we have the results of normal statistical mechanics. # Thermodynamics of classical ideal gas in the presence of GUP In this section we use an ideal gas that composed of monatomic non-interacting particles within GUP that admits both minimal length and maximal momentum. We suppose this ideal gas with the mentioned situations, in a micro-canonic ensemble. In micro-canonical ensemble, the macro state of the system is given by the number of molecules \(N\), the physical volume \(V\), and energy \(E\) of the system. In this ensemble, the volume \(\omega\) of the phase space to the representative points of the system where lie anywhere within a hypershell defined by the boundary condition \(E-\frac{\Delta}{2}\leq H(q,p)\leq E+\frac{\Delta}{2}\) is given by \[\omega=\int' d\omega=\int'\int' (d^{3N}q)(d^{3N}p)\] where \(\omega\equiv\omega(N,V,E;\Delta)\). In this equation the primed symbol means that integral could extends only over that part of the phase space which agrees to the above condition. Since the Hamiltonian of the classical ideal gas is only function of the \(ps\), the integration over the \(qs\) can be written as \(V^{N}\). The remaining integral \[\int' (d^{3N}p)\] should be calculated under the following condition: \[2m[E-\frac{\Delta}{2}]\leq \Sigma_{k=1}^{3N}p_{i}^{2}\leq 2m[E+\frac{\Delta}{2}]\] This condition is suitable only in standard framework. In this paper, with considering a GUP in the presence of both minimal measurable length and maximal momentum, the hypershell situation should be modified. This GUP can be written as follows \[\Delta x\Delta p\geq\frac{\hbar}{2}\bigg[1+\Big(\frac{\beta}{\sqrt{\langle p^2\rangle}}+4\beta^{2}\Big)(\Delta p)^2+4\beta^{2}\langle p\rangle^2-2\beta\sqrt{\langle p^2\rangle}\bigg].\] Since \((\Delta p)^2=\langle p^2\rangle-\langle p\rangle^2\), by setting \(\langle p\rangle=0\) for simplicity, we find \[\Delta x\Delta p\geq \frac{\hbar}{2}\bigg(1-\beta(\Delta p)+2\beta^{2}(\Delta p)^2\bigg).\] which \(\beta\) is a positive quantity.\ It is easy to show how this setup leads to a maximal momentum. To show this end, we note that the absolute minimal measurable length in our setup is given by, \(\Delta x_{min}(\langle p\rangle=0)=\frac{2\sqrt{2}-1)}{2}\hbar\beta\). Due to duality of position and momentum operators, it is reasonable to assume \(\Delta x_{min}\propto\Delta p_{max}\). Now, saturating the inequality in relation (5), we find \[2(\Delta x\Delta p)=\hbar\bigg(1-\beta(\Delta p)+2\beta^{2}(\Delta p)^2\bigg).\] This results in \[(\Delta p)^2-\frac{(2\Delta x+\beta\hbar)}{2\hbar\beta^{2}}(\Delta p)+\frac{1}{2\beta^{2}}=0.\] So, we find \[(\Delta p_{max})^2-\frac{(2\Delta x_{min}+\beta\hbar)}{2\hbar\beta}(\Delta p_{max})+\frac{1}{2\beta^{2}}=0.\] Now using the value of \(\Delta x_{min}\), we find \[(\Delta p_{max})^2-\frac{\sqrt{2}}{\beta}(\Delta p_{max})+\frac{1}{2\beta^{2}}=0.\] The solution of this equation is \[\Delta p_{max} =\frac{\sqrt{2}}{2\beta }.\] So, there is an upper bound on particle's momentum uncertainty. As a nontrivial assumption, we assume that this maximal uncertainty in particle's momentum is indeed the maximal measurable momentum. This is of the order of Planck momentum. We note that neglecting a factor of \(\frac{1}{2}\) for simplicity in our forthcoming arguments, the GUP formulated as (5) gives the following generalized commutation relation \[= i\hbar\Big(1-\beta p + 2\beta^{2} p^2\Big).\] With comparing this equation with its standard form we can interpret this as a generalization of \(\hbar\). This generalization has important results. One of them is increasing the volume of phase space, it means that the standard volume of phase space \(\hbar^{N}\) changes to \([\hbar(1-\beta p+2\beta^{2}p^{2})]^{N}\). It is clear that the nature result of increasing the volume of the phase space is decreasing of the number of accessible microstates for the given system. Now with using this GUP we should modified the hypershell boundary condition. This issue firstly has been considered by Kalyana Rama. He has discussed the effect of GUP on various thermodynamical quantities in grand canonical ensemble\[16\]. Then some authors considered the effects of GUP on thermodynamics of ideal gas in micro-canonical ensemble\[17\]. But they didn't consider the effects of maximal momentum in their calculations. In this paper, we want to study this results in the presence of GUP that formulated as (5). Within this GUP framework, particle's momentum should be generalized. This generalized momentum is given by \[p^{GUP}\simeq p(1-\frac{1}{2}\beta p +\frac{2}{3}\beta^{2}p^{2}).\] On the other hand, when the momentum be generalized, energy will generalize too \[E^{GUP}\simeq E(1-\frac{1}{2}\beta E +\frac{2}{3}\beta^{2}E^{2}).\] Now with using equations (12) and (13) we should rewrite the standard hypershell equation. So we have \[2m\bigg[E(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2} E^{2})-\frac{\Delta}{2}\bigg]\leq\sum_{i=1}^{3N}p_{i}^{2}(1-\beta p_{i}+\frac{4}{3}\beta^{2} p_{i}^2)\leq 2m\bigg[E(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2} E^{2})+\frac{\Delta}{2}\bigg]\] In this situations, integral(2) is equal to the volume of a 3N-dimensional hypershell, bounded by two hyperspheres of radii \[\sqrt{2m\bigg[E(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2} E^{2})-\frac{\Delta}{2}\bigg]}\] and \[\sqrt{2m\bigg[E(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2} E^{2})+\frac{\Delta}{2}\bigg]}.\] So we can write \[\int'...\int'\prod_{i=1}^{3N}dp_{i}=B_{3N}\Bigg(\sqrt{2m[E^{GUP}+\frac{\Delta}{2}]} \Bigg)^{3N}:=K\] This equation has written from the following condition \[0\leq\sum_{i=1}^{3N} p_{i}^2(1-\beta p_{i}+\frac{4}{3}\beta^{2}p_{i}^2)\leq 2m[E^{GUP}+\frac{\Delta}{2}]\] This equation gives half of the volume of the phase space, so we should multiply our final result to the factor of \(2\). Here \(B_{3N}\) is a constant which depends only on the dimensionality of the given phase space. Now, we can calculate the volume element \(dK\) as follows \[dK=\frac{3}{2}N B_{3N}(\sqrt{2m})^{3N}\Bigg[\sqrt{E^{GUP}+\frac{\Delta}{2}}\Bigg]^{3N-2} dE^{GUP}\] To evaluate \(B_{3N}\), we use the following integral formula \[\int_{-p_{p}}^{+p_{p}} \exp(-p^{2}+\beta p^{3}-\frac{4}{3}\beta^{2}p^{4})dp=R(\beta)\] As we see, when we use GUP that admits both minimal measurable length and maximal momentum, the boundary condition of the integral, should be restricted. In this equation \(p_{p}\) is the Planck momentum and \(R(\beta)\) is the result of integral. Because of the complication of the solution of integral, first we expand the exponential function up to eighth order and then calculate the integral. This calculation give us \[R(\beta)=2p_{p}-\frac{2}{3}p_{p}^{3}+\frac{1}{2}\beta p_{p}^{4}+\frac{2}{5}\bigg(\frac{1}{2}-\frac{4}{3}\beta^{2}\bigg)p_{p}^{5}-\frac{1}{3}\beta p_{p}^{6}\] \[+\frac{2}{7}\bigg(\frac{11}{6}\beta^{2}-\frac{1}{6}\bigg)p_{p}^{7}+\frac{1}{4}\bigg(\frac{1}{2}\beta-\frac{4}{3}\beta^{3}\bigg)p_{p}^{8}+...\] For simplicity, we supposed this calculation equal to \(R(\beta)\). Another way to calculate the above integral is considering only first order of \(\beta\), so we have \[R(\beta)=\beta-\beta p_{p}^{2}e^{-p_{p}^{2}}+\sqrt{\pi} erf(p_{p})\] where \(erf(x)\) is the error function and defined as \[erf(x)=\frac{2(\int_{0}^{x}e^{-t^{2}}dt)}{\sqrt{\pi}}\] With multiplying \(3N\) to such integrals, one for each of variables \(p_{i}\), we have \[\int_{-p_{p}}^{+p_{p}}...\int_{-p_{p}}^{+p_{p}}\exp\Bigg(-\sum_{i=1}^{3N}p_{i}^{2}(1-\beta p_{i}+\frac{4}{3}\beta^{2}p_{i}^2)\Bigg)\prod_{i=1}^{3N} dp_{i}=[R(\beta)]^{3N}\] Therefore, it follows that: \[\int_{-p_{p}}^{+p_{p}}\exp\Bigg(-2m(E^{GUP}+\frac{\Delta}{2})\bigg)dK=[R(\beta)]^{3N}\] Now we put Eq.(17) to the above equation and find \[\int_{-p_{p}}^{+p_{p}}\frac{3}{2}N B_{3N} (\sqrt{2m})^{3N}\bigg[\sqrt{E^{GUP}+\frac{\Delta}{2}}\bigg]^{3N-2}\exp\bigg(-2m(E^{GUP}+\frac{\Delta}{2})\bigg)dE^{GUP}=[R(\beta)]^{3N}\] In the next step, with calculating this integral, we can find \(B_{3N}\). So we have: \[B_{3N}=\frac{2[R(\beta)]^{3N}\exp(m\Delta)}{3N(2m)^{\frac{3N}{2}}\int_{-p_{p}}^{+p_{p}}(E^{GUP}+\frac{\Delta}{2})^{\frac{3N-2}{2}}\exp(-2mE^{GUP})dE^{GUP}}\] For \(\Delta\ll E^{GUP}\), this equation reduces to: \[B_{3N}=\frac{2[R(\beta)]^{3N}}{3N(\frac{3N}{2}-1)!}\] Now, from equation (15), we find \[\int...\int\prod_{i=1}^{3N}dp_{i}\equiv\frac{2[R(\beta)]^{3N}(2mE^{GUP})^{\frac{3N}{2}}\bigg[1+\frac{3N\Delta}{4E^{GUP}}\bigg]}{3N(\frac{3N}{2}-1)!}\] As we now ,for thermodynamical systems, \(N\gg1\). So we can rewrite the above equation as follow \[\int...\int\prod_{i=1}^{3N}dp_{i}\simeq \frac{\Delta}{2E^{GUP}}\frac{[R(\beta)]^{3N}}{(\frac{3N}{2}-1)!}\bigg(2mE^{GUP}\bigg)^{\frac{3N}{2}}\] So, with these situations, the total volume of the phase space enclosed within hypershell is given by \[w\simeq\frac{\Delta}{E^{GUP}}V^{N}\frac{(2[R(\beta)]^{2}m E^{GUP})^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\] Now we want to find the number of microstates. To find this end, first we should find the fundamental volume \(\omega_{0}\) in the presence of the minimal length and maximal momentum. It is given by \[\omega_{0}=(\Delta q \Delta p)^{3N}=\bigg[\hbar(1-\beta p+2\beta^{2}p^{2})\bigg]^{3N}\equiv\hbar'^{3N}\] In this equation, for simplicity we supposed that\(\bigg[\hbar(1-\beta p+2\beta^{2}p^{2})\bigg]\equiv \hbar'\). With compering Eq.(28) with its standard form, we can consider it as generalized of \(\hbar\). In other words, \(\hbar\longrightarrow \hbar(1-\beta p+2\beta^{2}p^{2})\). It is obvious that, with dividing total volume phase space to fundamental phase space, means \(\frac{\omega}{\omega_{0}}\), we can find total number of microstates within hypershell and it has shown by \(\Omega\).\ So we have \[\Omega=\frac{V^{N}}{\hbar'^{3N}}\frac{\Delta}{E^{GUP}}\frac{(2[R(\beta)]^{2}m E^{GUP})^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\] Obviously, within GUP in the presence of minimal length and maximal momentum, due to increased fundamental volume \(\omega_{0}\), the number of total microstates decreases.\ The complete thermodynamics of the given system would then given by, \[S(N,V,E^{GUP})=k\ln\Omega=k\ln\Big(\frac{V^{N}}{\hbar'^{3N}}\frac{\Delta}{E^{GUP}}\frac{(2[R(\beta)]^{2}m E^{GUP})^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\Bigg)\] Where \(S\) is the entropy of the given system. In the absence of quantum gravity effects, when \(\beta\longrightarrow0\), we have the usual standard entropy. Now, after finding entropy, we can find various thermodynamical quantities. it is obvious from above equations that the reduction of total number of accessible microstates in high momentum regime cause reduction of entropy. It seems that thermodynamical system in very short distances have an unusual thermodynamics. # Thermodynamics of extreme relativistic gas in the presence of GUP In this part, we want to calculate thermodynamics of an ultra relativistic monatomic non interacting gaseous system in the presence of GUP that admit both minimal measurable length and maximal momentum. With using of arguments in previous section, the hypershell equation for ultra relativistic gaseous system is given by the following equation \[\frac{1}{c}\bigg[E\big(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2}E^{2}\big)-\frac{\Delta}{2}\bigg]\leq\sum_{i=1}^{3N}p_{i}\big(1-\frac{1}{2}\beta p_{i}+\frac{2}{3}\beta^{2}p_{i}^{2}\big)\leq \frac{1}{c}\bigg[E\big(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2}E^{2}\big)+\frac{\Delta}{2}\bigg]\] In this situation, \(\int(d^{3N}p)\) is equal to the volume of a \(3N\) dimensional hypershell, bounded by two hypershell of radii\ \[\sqrt{\frac{1}{c}\bigg[E\big(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2}E^{2}\big)-\frac{\Delta}{2}\bigg]}\] and \[\sqrt{\frac{1}{c}\bigg[E\big(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2}E^{2}\big)+\frac{\Delta}{2}\bigg]}\] The number of microstates for the system is proportional to the volume of this hypershell. Similarly to the previous section, we have \[\int'...\int'\prod_{i=1}^{3N}dp_{i}=A_{3N}\Big(\sqrt{\frac{1}{c}[E^{GUP}+\frac{\Delta}{2}]} \Big)^{3N}:=F\] This equation follows the following condition \[0\leq\sum_{i=1}^{3N} p_{i}\bigg(1-\frac{1}{2}\beta p_{i}+\frac{2}{3}\beta^{2}p_{i}^{2}\bigg)\leq \frac{1}{c}\bigg[E^{GUP}+\frac{\Delta}{2}\bigg]\] where \(E^{GUP}\) is equal to \(E\big(1-\frac{1}{2}\beta E+\frac{2}{3}\beta^{2}E^{2}\big)\). This equation shows half of the volume of the phase space. So we should multiply the final result with a factor of \(2\). Now we want to find \(dF\): \[dF=\frac{3}{2}N A_{3N}(\frac{1}{c})^{3N}\bigg[\sqrt{E^{GUP}+\frac{\Delta}{2}} \bigg]^{3N-2} dE^{GUP}\] To evaluate \(A_{3N}\), we use the following integral formula: \[\int_{-p_{p}}^{+p_{p}} \exp\bigg(-p+\frac{1}{2}\beta p^2-\frac{2}{3}\beta^{2} p^{3}\bigg)dp=H(\beta)\] Here \(H(\beta)\) is the result of the integral. In above equation because of the complicated form of sentences, first we expand them up to eighth order, and then calculate the integral, so we have: \[H(\beta)=2p_{p}-p_{p}^{2}+\frac{2}{3}(\frac{1}{2}+\frac{1}{2}\beta)p_{p}^{3}+\frac{1}{2}(-\frac{2}{3}\beta^{2}-\frac{1}{2}\beta-\frac{1}{6})p_{p}^{4}\] \[+\frac{2}{5}(\frac{19}{24}\beta^{2}+\frac{1}{4}\beta+\frac{1}{24})p_{p}^{5}+\frac{1}{3}(-\frac{1}{3}\beta^3-\frac{11}{24}\beta^2-\frac{1}{12}\beta-\frac{1}{120})p_{p}^{6}\] \[+\frac{2}{7}(\frac{2}{9}\beta^{4}+\frac{17}{48}\beta^{3}+\frac{25}{144}\beta^{2}-\frac{1}{48}\beta+\frac{1}{720})p_{p}^{7}+\frac{1}{4}(-\frac{11}{36}\beta^{4}-\frac{3}{16}\beta^{3}\] \[-\frac{7}{144}\beta^{2}-\frac{1}{240}\beta-\frac{1}{5040})p_{p}^{8}+...\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\ \] Similarly to the previous section, another way to calculate the above integral is considering only first order of \(\beta\). With this situation, we have \[H(\beta)=2+2\beta-2e^{-p_{p}}(1+\beta+\beta p_{p}+\frac{1}{2}\beta p_{p}^{2})\] Now with multiplying \(3N\) to this integral, for each of variable \(p_{i}\), we obtain \[\int_{-p_{p}}^{+p_{p}}...\int_{-p{p}}^{+p_{p}}\exp\Big(-\sum_{i=1}^{3N}p_{i}(1-\frac{1}{2}\beta p_{i}+\frac{2}{3}\beta^{2}{p_{i}^{2}})\Big)\prod_{i=1}^{3N}dp_{i}=[H(\beta)]^{3N}\] Since, it follows that \[\int_{-p_{p}}^{+p_{p}} \exp\bigg(-\frac{1}{c}(E^{GUP}+\frac{\Delta}{2})\bigg)dF=[H(\beta)]^{3N}\] Now with putting Eq.(34) in the above equation, we obtain \[\int_{-p_{p}}^{+p_{p}}\frac{3}{2}N A_{3N}(\frac{1}{c})^{\frac{3N}{2}}\bigg[E^{GUP}+\frac{\Delta}{2}\bigg]^{\frac{3N-2}{2}}\exp\bigg(-\frac{1}{c}(E^{GUP}+\frac{\Delta}{2})\bigg)dE^{GUP}=[H(\beta)]^{3N}\] Therefore we can easily find \(A_{3N}\) \[A_{3N}=\frac{2[H(\beta)]^{3N}\exp(\frac{\Delta}{2c})}{3N(\frac{1}{c})^{\frac{3N}{2}}\int_{-p_{p}}^{+p_{p}}\big[E^{GUP}+\frac{\Delta}{2}\big]^{\frac{3N-2}{2}}\exp\big(-\frac{1}{c}E^{GUP}\big)dE^{GUP}}\] For \(\Delta\ll E^{GUP}\), this equation equals to \[A_{3N}=\frac{2[H(\beta)]^{3N}}{3N (\frac{3N}{2}-1)!}\] Now from Eq.(32), we can obtain \[\int...\int\prod_{i=1}^{3N}dp_{i}=\frac{2[H(\beta)]^{3N}(\frac{1}{c}E^{GUP})^{\frac{3N}{2}}\big[1+\frac{3N\Delta}{4E^{GUP}}\big]}{3N (\frac{3N}{2}-1)!}\] On the other hand, when \(N\gg1\), we can write \[\int...\int\prod_{i=1}^{3N}dp_{i}\simeq \frac{\Delta}{2E^{GUP}}\frac{[H(\beta)]^{3N}}{(\frac{3N}{2}-1)!}(\frac{1}{c}E^{GUP})^{\frac{3N}{2}}\] So the total volume of the phase space enclosed within hypershell is given by \[\omega\simeq\frac{\Delta}{E^{GUP}}V^{N}\frac{\big(\frac{1}{c}[H(\beta)]^{2}E^{GUP}\big)^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\] As we know, the total number of microstates is given by \(\Omega= \frac{\omega}{\omega_{0}}\), and from Eq.(28), we have the \(\omega_{0}\) value in the GUP that admit both minimal length and maximum momentum. So we have \[\Omega=\frac{V^{N}}{\hbar'^{3N}}\frac{\Delta}{E^{GUP}}\frac{\big(\frac{1}{c}[H(\beta)]^{2}E^{GUP}\big)^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\] Now, after finding total number of microstates, it is easy to find thermodynamical quantities of the given system. In this step, the first thermodynamical quantity we find is entropy. So \[S(N,V,E^{GUP})=k\ln\Omega=k\ln\Big(\frac{V^{N}}{\hbar'^{3N}}\frac{\Delta}{E^{GUP}}\frac{\big(\frac{1}{c}[H(\beta)]^{2}E^{GUP}\big)^{\frac{3N}{2}}}{(\frac{3N}{2}-1)!}\Big)\] As we know, entropy is directly depend on the total accessible number of phase space, so decreasing of the total number of microstates cause reduction of entropy, and this is the unusual behavior of thermodynamical quantities near the planck scale. In standard situation, when \(\beta\longrightarrow 0\), we have the results of ordinary statistical mechanics. Various thermodynamical quantities can then be calculated with using of Eq.(46).\ # Summary and conclusion In this paper we consider a GUP that admits both minimal measurable length and maximal momentum. In the presence of these GUP we studied some thermodynamical properties of an ideal gas both in the classical and ultra relativistic limit. We have found that in the presence of this mentioned GUP, the fundamental volume of the phase space increased. We can interpret this increasing as a generalized of the Planck's constant. This increasing of the volume caused reduction of the accessible number of phase space and because of this reduction, some important thermodynamical quantities such as entropy of ideal gas decreased. It was an unusual behavior of thermodynamics in very short distances.\ **Acknowledgment**\ I would like to thank Prof. Kourosh Nozari for insightful comments and discussion.

{'timestamp': '2013-10-14T02:11:07', 'yymm': '1310', 'arxiv_id': '1310.3243', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3243'}

# Introduction[\[sec:intro\]]{#sec:intro label="sec:intro"} One of the long-standing issues in solar physics is the problem of the origin and even the very existence of a high degree of linear polarization of some chromospheric spectral lines observed in solar flares. Over the last few decades, several authors have reported the detection of linear polarization signals in solar flares, with fractional polarization amplitudes at the level of several percent. The observed spectral lines include Ca ii K, Na D\(_2\), and the hydrogen lines H\(\alpha\) and H\(\beta\). This polarization is typically not observed in the center of the flare ribbons but rather at their boundaries with the surrounding chromosphere and it is often found to be perpendicular (i.e., radial) or parallel (i.e., tangential) to the nearest solar limb. The standard interpretation of this polarization is in the terms of impact polarization of the atomic levels due to non-thermal particle beams and/or the neutralizing return currents, even though polarization can also be observed during the gradual phase of a flare. On the other hand, other authors have reported no observable linear polarization above the 0.1 % level. Up to now, most of the interpretations of the observed polarization have been done either in the so-called last scattering approximation or using one-dimensional (1D) models taking into account the effects of radiation transfer. A specific version of impact polarization due to evaporative upflows has been proposed by. In this letter, we propose a different perspective on the interpretation of such observations. The chromospheric lines of interest are formed under non-LTE conditions and they are susceptible to resonant scattering polarization. The 1D plane-parallel approximation of the flare ribbons must fail at the ribbon edges where the strong horizontal gradients of the physical conditions are found. From the point of view of the radiative transfer theory, the ribbons are a remarkable example of a multi-dimensional medium where the anisotropy and symmetry breaking of the radiation field may be very significant. This can result in enhanced emission of linearly polarized radiation in spectral lines due to resonance scattering. For a basic investigation of resonance line polarization and the Hanle effect in horizontally inhomogeneous stellar atmospheres see. One of the most familiar manifestations of scattering polarization in the Sun is the so-called second solar spectrum. In the vicinity of a solar flare, the cylindrical symmetry of the atmosphere is broken and multi-dimensional calculations are necessary. In this letter, we don't consider the possible role of impact polarization and we solve a two dimensional (2D) radiative transfer problem using a simple model of a flaring atmosphere to study the role of scattering polarization and the Hanle effect on the emergent spectral line polarization. # Formulation of the problem[\[sec:formul\]]{#sec:formul label="sec:formul"} ## Resonance scattering polarization[\[ssec:polar\]]{#ssec:polar label="ssec:polar"} We describe the polarization state of radiation by means of the vector of Stokes parameters \((I,Q,U)^{\rm T}\), where \(I\) denotes the specific intensity, and \(Q\) and \(U\) quantify the linear polarization. In our calculation, we do not consider the circular polarization due to the Zeeman effect. Following the usual convention, we define the positive \(Q\) direction to be parallel to the nearest solar limb. In this work, we apply the theory of spectral line polarization described in. The description of the mean intensity, anisotropy, and symmetry properties of the radiation field at a given point of the atmosphere can be done using the irreducible representation of the radiation field tensors, \(J^K_Q\), with \(K=0,1,2\), and \(Q=-K,\dots,K\). The physical meaning of the individual \(J^K_Q\) components becomes apparent from their definition. In particular, \(J^0_0\) corresponds to the familiar mean intensity of the radiation and it is the only non-zero radiation tensor if the field is isotropic. In order to find the atomic excitation state at every point within the model atmosphere, the \(J^K_Q\) tensors need to be calculated at every such point by solving the radiative transfer equation for the Stokes parameters. The radiation field tensors enter the equations of statistical equilibrium whose solution provides the density matrix of the atomic levels, \(\rho^K_Q\), where \(K=0,\dots,2J\) and \(Q=-K,\dots,K\) for the level with the total angular momentum \(J\). In this representation, \(\rho^0_0\) is proportional to the population of the level and \(K>0\) components contain information on the polarization state of the level. In our model, only the \(K=2\) components (alignment) need to be accounted for, while the \(K=1\) components (orientation) are identically zero. It follows that levels with angular momentum \(J<1\) can only hold population but not the atomic polarization. After a self-consistent solution of the non-LTE problem, one can synthesize the emergent polarized spectra. In the case of a cylindrically symmetric plane-parallel atmosphere, all the coherence components \(\rho^K_{Q\neq 0}\) are identically equal to zero and only the \(Q=0\) components remain. Such a description is adequate in the case of 1D unmagnetized models of the solar atmosphere in which the emergent polarization is either radial or tangential with respect to the nearest limb. If the cylindrical symmetry is broken due to the presence of an inclined magnetic field and/or due to horizontal inhomogeneities of the thermal structure of the plasma, a general description taking into account all the \(\rho^K_Q\) elements is necessary. If magnetic field is present in the atmosphere, scattering line polarization can be modified via the Hanle effect. The Hanle effect typically leads to rotation of the polarization direction and to decrease of the polarization degree of the emergent radiation. The order of magnitude of the magnetic field strength at which the Hanle effect becomes significant for a particular spectral line can be quantified by the so-called critical Hanle field, \(B_{\rm H}=1.137\times 10^{-7}/(t_{\rm life}g_J)\), where \(t_{\rm life}\) is the lifetime of an involved atomic level and \(g_J\) is the Landé factor. For chromospheric lines, \(B_{\rm H}\) is usually in the interval from milligauss to few deca-gauss. If the field is significantly stronger than \(B_{\rm H}\) (the so-called Hanle effect saturation regime), the quantum coherence in the reference frame in which the magnetic field is parallel to the quantization axis, is destroyed and the polarization direction of the emitted radiation is either parallel or perpendicular to the plane defined by the magnetic field vector and the photon propagation direction. ## The model atom[\[ssec:atom\]]{#ssec:atom label="ssec:atom"} Given that the goal of this letter is to point out the general mechanism of a possible creation of the linear polarization in solar flares, we do not particularize the model atom to any specific chemical species. Instead, we choose a generic two-level model atom with a resonant transition at \(\lambda_0=5\,000\) Å, the Einstein coefficient of the spontaneous emission of \(A_{u\ell}=10^8\,{\rm s}^{-1}\), and the atomic weight of hydrogen. The angular momenta of the lower and upper levels are \(J_\ell=0\) and \(J_u=1\), respectively. The line absorption profile is a Voigt profile with a damping parameter \(a=10^{-2}\), constant throughout the atmosphere. We neglect the effect of stimulated emission. We assume the approximation of complete frequency redistribution (CRD). According to Sect. [\[ssec:polar\]](#ssec:polar){reference-type="ref" reference="ssec:polar"}, the lower atomic level with zero angular momentum can only hold population but it cannot be polarized. The only polarizable level is thus the upper level whose non-zero density matrix elements are the population \(\rho^0_0\) and the five complex components \(\rho^2_Q\) of the alignment. The Landé factor of the upper level is \(g_u=1\) and the critical Hanle field of our line is \(B_{\rm H}=11\) G. ## The model atmosphere[\[ssec:atmosphere\]]{#ssec:atmosphere label="ssec:atmosphere"} We consider a simple isothermal 2D model atmosphere with the kinetic temperature \(T=6\,000\) K representing the solar chromosphere. It corresponds to a vertical slice perpendicular to the flare ribbon with \(z\) being the vertical and \(x\) being the horizontal axis (see below). The atmosphere is exponentially stratified with the number density of atoms given by \(n(z)=n_0{\rm e}^{-z/H}\), where \(n_0=10^{15}\,{\rm cm}^{-3}\) and \(H=75\) km. We assume periodic boundary conditions in the \(x\)-direction. The computational domain extends horizontally from \(x=-20\) Mm to \(x=20\) Mm. At these boundaries, the solution is virtually identical to the corresponding 1D solution for a given vertical column of the isothermal atmosphere. This justifies the use of periodic boundary conditions. Along the \(z\) axis, the model extends from \(z=0\) Mm, which corresponds to the photospheric level where the line is thermalized, to \(z=2\) Mm, which corresponds to the optically thin upper chromosphere. The thermal collisions are characterized by the collisional destruction probability \(\epsilon = C^{\rm th}_{u\ell}/(A_{u\ell}+C^{\rm th}_{u\ell})=10^{-2}\), where \(C^{\rm th}_{u\ell}\) is the collisional de-excitation rate due to the thermal collisions. We do not consider any collisional depolarization in this work. We model a single flare ribbon by adding the non-thermal collisional excitation rate in the central part of the model atmosphere. According to the theory of non-thermal beam propagation, particle beams deposit part of their energy via non-thermal excitation of the atoms. This process is most efficient in the middle and upper chromosphere. In our model, the region of enhanced non-thermal excitation is restricted in both \(z\) and \(x\) directions in order to mimic the presence of a spatially localized ribbon. Using the sigmoid function \(\sigma(w,d,x)=[1+e^{-(x-d)/w}]^{-1}\), we consider the total collisional excitation rate (i.e., thermal plus non-thermal), to be given by an ad hoc expression \[C_{\ell u}(x,z) = C^{\rm th}_{\ell u} \left[ 1 + 10\, {\rm e}^{-(z-z_0)^2/w_z^2}\sigma(-w_x,d,x)\sigma(w_x,-d,x) \right]\,, \label{eq:clu}\] where \(z_0=1.55\) Mm is a \(z\)-coordinate of the maximum of collisional rates, \(w_z=0.3\) Mm controls the vertical extension of the region of enhanced non-thermal collisions, \(w_x=3\) Mm is the half-width of the ribbon in the horizontal direction, and \(d=0.15\) Mm determines the horizontal gradient of collisional rates at the ribbon edges. The spatial distribution of collisional excitation rates is plotted in Fig [\[fig:colls\]](#fig:colls){reference-type="ref" reference="fig:colls"}. The above increase of collisional rates is chosen by analogy with more realistic models. Note that our choice is rather conservative since the non-thermal collisional rates can exceed the thermal ones by much more than the factor of ten considered here in Eq. ([\[eq:clu\]](#eq:clu){reference-type="ref" reference="eq:clu"}). Only the collisional de-excitation due to the thermal electrons must be considered while the de-excitation by the high-energy non-thermal collisions can be neglected. The parameter \(\epsilon\) is therefore constant in the model atmosphere and the only spatially dependent rate is \(C_{\ell u}\). Note that \(C_{\ell u}\) is the population transfer rate and that we do not consider impact polarization in our model. Magnetic field at the flare-loop footpoints can be assumed to be roughly vertical with a strength of a few hundred gauss. That is well above the Hanle effect saturation of most of the spectral lines of interest and also above the saturation of our generic line. In the following section, we study both non-magnetized and magnetized solutions of the radiative transfer problem. # Results[\[sec:res\]]{#sec:res label="sec:res"} We have solved the non-LTE problem described in the previous section using the radiative transfer code PORTA applied to a 2D grid with \(N_x\times N_z = 256\times 100\) grid points. In Fig. [\[fig:int\]](#fig:int){reference-type="ref" reference="fig:int"}, we compare the line intensity profile of the 'quiet' region of the atmosphere with the profile obtained at the center of the ribbon. The later resembles the characteristic emission line profiles in solar flares. ## Non-magnetized atmosphere[\[ssec:nonmag\]]{#ssec:nonmag label="ssec:nonmag"} In Fig. [\[fig:jkq\]](#fig:jkq){reference-type="ref" reference="fig:jkq"}, we show the self-consistent values of the \(J^K_Q\) tensors within the model atmosphere. The anisotropy of the line radiation vanishes everywhere below \(z\approx 1.5\) Mm, where the optical thickness of the line becomes significantly larger than unity. Far away from the ribbon, at \(|x|\gtrsim 7\) Mm, the atmosphere can be accurately modelled using the plane-parallel approximation. There we can find the height variation of \(J^2_0\) anisotropy which is typical in the isothermal atmospheres. In accordance with our expectations, we have found that the anisotropy of the radiation is highest at the edges of the ribbon with \(|x|\gtrsim 3\) Mm and \(|x|\lesssim 7\) Mm (see Fig. [\[fig:jkq\]](#fig:jkq){reference-type="ref" reference="fig:jkq"}), i.e., mainly in the regions of the atmosphere which are not directly affected by the beam itself. All the \(J^2_Q\) tensorial components at the height of formation of the line center, i.e., above \(z=1.69\) Mm, are affected by the strong anisotropic emission of the ribbon. It is in these boundary regions where the \(J^2_1\) and \(J^2_2\) components, which are due to the breaking of the plane-parallel approximation, are nonzero. Anisotropy in the center of the ribbon is relatively small. This is due to the fact that in the center of the ribbon, the horizontal gradients of physical quantities are smaller than at the ribbon edges. However, in case of a narrow ribbon, one can expect a polarization even in the ribbon center due to different radiation intensities along and perpendicularly to the ribbon orientation. In Fig. [\[fig:stokes\]](#fig:stokes){reference-type="ref" reference="fig:stokes"}, we show the emergent Stokes profiles of the line for an inclined line of sight (LOS). The maximum fractional polarization is found at the ribbon edges and the total degree of linear polarization reaches \(P\approx 8\) %. The degree and orientation of polarization depends on both inclination and azimuth of the LOS, i.e., on the position and orientation of the ribbon on the solar disk. Polarization degree in the central part of the ribbon is significantly reduced with respect to the edges due to the lower radiation field anisotropy in the ribbon center and due to the increased role of the collisional excitation. ## Magnetized atmosphere and the Hanle effect Interestingly enough, in many spectropolarimetric observations, the orientation of polarization vector is either perpendicular or parallel to the nearest solar limb regardless of the ribbon orientation at the solar surface. This fact is usually used for advocating the role of impact polarization due to the particle beams of various energies propagating along the vertical magnetic field lines. However, the magnetic field itself can directly affect the line polarization via the Hanle effect. We include this mechanism in the calculations presented in this section. We have kept the model atmosphere and the model atom as in Sect. [\[ssec:nonmag\]](#ssec:nonmag){reference-type="ref" reference="ssec:nonmag"} but we have included a uniform vertical magnetic field of strength \(B=500\) G, i.e., well above the Hanle-effect saturation of our line. Given the fact that the radiation field is not cylindrically symmetric at the ribbon edges, the Hanle effect of the vertical field modifies the atomic polarization. The atomic coherences \(\rho^2_{Q\neq 0}\) are practically removed by the action of the vertical magnetic field in the saturation regime and only the population \(\rho^0_0\) and alignment \(\rho^2_0\) of the upper atomic level remain.[^1] Similarly to the case of a cylindrically symmetric plane-parallel atmosphere, in which only the \(\rho^0_0\) and \(\rho^2_0\) components remain, polarization of the emitted photons is either radial or tangential. We demonstrate this in Fig. [\[fig:stokesb\]](#fig:stokesb){reference-type="ref" reference="fig:stokesb"} which has been calculated for the same LOS as Fig. [\[fig:stokes\]](#fig:stokes){reference-type="ref" reference="fig:stokes"} but taking into account the action of the magnetic field. The \(U/I\) signal, which quantifies a deviation of the polarization vector from the radial and tangential directions, is negligible with respect to \(Q/I\) in the regions of interest. In case of the LOS considered in our example, the orientation of the polarization is radial. # Concluding comments[\[sec:concl\]]{#sec:concl label="sec:concl"} We have demonstrated that the horizontal inhomogeneities of the atmosphere may lead to the creation of significant scattering polarization signatures in spectral lines which are consistent with some of the most common observational findings. We can summarize our results as follows: 1. The scattering polarization is largest at the ribbon edges due to scattering of the anisotropic ribbon emission. The degree of such polarization can be of the order of several percent. Sufficient spatial resolution of the observations is necessary for detection of this polarization because the region of enhanced polarization is small. The non-detection of may be due to the low spatial resolution of the observations which was \(10''\times 10''\). 2. Polarization in the center of the ribbon is smaller than polarization at the edges due to the lower anisotropy of radiation and due to stronger inelastic collisions. The role of collisional depolarization in the ribbon center due to collisions with thermal electrons and protons may also play a role in the case of the hydrogen Balmer lines but it is questionable for the lines of other species (Na D\(_2\), Ca ii K) which are mainly depolarized by collisions with neutral hydrogen whose density actually decreases in the ribbons. 3. In the presence of a strong vertical magnetic field, the relaxation of atomic coherences due to the Hanle effect causes emission of predominantly radial or tangential spectral line polarization, depending the particular spectral line, on the LOS, and on the anisotropy of the radiation field. 4. Polarization can be expected during both the impulsive and gradual phases of a flare. The condition for creation of the polarization is a presence of a localized flare ribbon breaking the plane-parallel symmetry of the atmosphere. Our model is simple and we have neglected several important facts: We have assumed that the ribbon can be approximated by a 2D structure; we have neglected a possible influence from a different nearby ribbon; we have assumed that the ribbon is symmetric with respect to the \(x=0\) plane. The last of these approximations is probably most severe because it ignores the different physical conditions in the inner ribbon edges that are affected by the presence of the post-flare loops. We have chosen the non-thermal collisional rates so that the resulting emission line intensity profile resembles those of. Reduction of the non-thermal collisional rates or reduction of their horizontal gradient leads to decrease of the radiation anisotropy and, consequently, to reduction of the emergent fractional polarization. The scattering polarization signals are further altered by number of physical quantities such as the horizontal gradients of the temperature of the plasmas, the magnetic field direction, collisional depolarizing rates, and the particular spectral line under consideration. Different combinations of these parameters may lead to a large variety of the observed polarization signals. Exploration of such parameter space should be a subject of future detailed investigations. We are grateful to Javier Trujillo Bueno (IAC) and Stanislav Gunár (AIAS) for their valuable comments to the manuscript. Financial support by the Grant Agency of the Czech Republic through grants P209/12/P741 and P209/12/1652, and by the project RVO:67985815 is gratefully acknowledged. [^1]: Note however, that \(J^2_Q\) with \(Q\neq 0\) are significantly affected by the presence of magnetic field.

{'timestamp': '2013-10-15T02:00:30', 'yymm': '1310', 'arxiv_id': '1310.3284', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3284'}

# Introduction In his seminal papers that gave birth to the field of information theory, Shannon showed that the capacity of a channel could be achieved using codes whose codewords are random bit strings. Despite this optimality, random codes have no practical use because we do not know how to decode them efficiently, i.e. in a time polynomial in the number of encoded bits. In 1978, Berlekamp, McEliece, and van Tilborg (see also ) showed that decoding a classical linear code is an \(\NPComplete\) problem, which strongly indicates that no efficient algorithm will ever be found to decode generic classical codes. A central problem in coding theory therefore consists in designing codes that retain the essential features of random codes, but yet have enough structure to be efficiently (approximately) decoded. Quantum information science poses additional challenges to coding theory. While the stabilizer formalism establishes many key parallels between classical and quantum coding, important distinctions remain. On the one hand, quantum code design is obstructed by the additional burden that check operators must mutually commute. For that reason, it has proven difficult to quantize some of the best families of classical codes, such as low density parity check (LDPC) codes and turbo codes. On the other hand, quantum codes can be *degenerate*, which means that distinct errors can have the same effect on all codewords. This changes the nature of the decoding problem, and our goal here is to explore how degeneracy impacts the computational complexity of the decoding problem. The conceptually simplest method to decode a stabilizer code is to ignore error degeneracy and to proceed as with classical linear codes: amongst all errors consistent with the error syndrome, find the one with the highest probability. We call this decoding method *Quantum Maximum Likelihood Decoding* (\(\QMLD\)). It was shown in that \(\QMLD\) is \(\NPComplete\). In the presence of degeneracy however, errors naturally fall into equivalence classes, with all errors in the same class having the same effect on all codewords. The optimal decoding method searches over all equivalence classes of errors that are consistent with the error syndrome, the one with the largest probability. The probability of a class of error is simply the sum of the probabilities of all errors it contains. We call this decoding method *Degenerate Quantum Maximum Likelihood Decoding* (\(\DQMLD\)). Our main result is the following theorem. (Informally) \(\DQMLD \in \sharpPComplete\) *up to polynomial-time Turing reduction.* We need Turing reduction since decoding is not a counting problem, while problems in \(\sharpP\) consist in counting the number of solutions to a decision problem in \(\NP\). Our result can be understood intuitively from the fact that in order to compute the probability associated to an equivalence class, \(\DQMLD\) must determine how many errors of a given weight belong to an equivalence class, hence the need to count. Our proof uses a reduction from the problem of evaluating the *weight enumerator* of a classical (binary) linear code, which was shown by Vyalyi to be \(\sharpPComplete\). The rest of this paper is organized as follows. For self-containment, the next two sections provide elementary introductions to computational complexity and to the stabilizer formalism. Sec. ([4](#decoding-problem){reference-type="ref" reference="decoding-problem"}) formally defines the decoding problem with a particular emphasis on the role of degeneracy. This section also contains an informal discussion on the importance of error degeneracy and how it impacts the decoding problem in specific settings. Sec. ([5](#sec-hardness){reference-type="ref" reference="sec-hardness"}) presents the main result, which is proved in Sec. ([6](#sec-red-over){reference-type="ref" reference="sec-red-over"}); the expert reader can jump directly to these two sections. The conclusion proposes possible extensions of the present work. # Computational Complexity {#complexity-intro} Two key resources required to solve any problem in computer science are space and time. One way of classifying problems is based on the runtime of a corresponding algorithm for that problem. This time is expected to depend on the size of the input, \(n\), to the problem. Instead of precisely denoting the runtime as a function \(f(n)\), which would depend on the underlying hardware of the solver, it is largely classified by its *limiting behavior* as: \(\mathcal{O}(\log_{2}n), \mathcal{O}(n^{k}), \mathcal{O}(2^{n})\) and so on. Consequently, the class of problems for which there exists an algorithm which runs in time \(\mathcal{O}(n^{k})\), on an input of size \(n\), \(k\) being a constant independent of \(n\), is called \(\P\). Note that any problem in \(\P\) is a decision problem, i.e, one for which the solution space is binary. If the problem is formulated to produce an output string, then the existence of a polynomial time algorithm classifies this problem in the class \(\FP\). There are problems to which any witness or certificate for a solution can be verified in polynomial time. These fall into the class called \(\NP\) and clearly, \(\P \in \NP\). A problem \(P_{1}\) is *at least as hard as* \(P_{2}\) if we could use the solver for \(P_{1}\) to solve \(P_{2}\). This is formalized by the notion of a *reduction*, which enables the classification of problems that are harder than a particular class, thereby introducing \(\NPHard\) and \(\NPComplete\), the latter being a class of hardest problems in \(\NP\). Besides decision problems, a category of problems involve enumerating the elements of a set. When this set is in \(\NP\), the corresponding enumeration problem is classified as \(\sharpP\). Hence, a function in \(\sharpP\) counts the number of solutions to a corresponding decision problem in \(\NP\). One can also compare the relative hardness of two counting problems as is done for the \(\NP\) case, using the notion of counting reductions. That is, \(f(x)\) can be computed in polynomial-time with polynomially many queries to an oracle for \(g(x)\), for any \(x\). Subsequently, analogous to \(\NPComplete\), a notion of reduction for counting functions defines the hardest problems in \(\sharpP\), as the class \(\sharpPComplete\). The last criterion can also be replaced with: \(g\thinspace \in\thinspace \FP^{f}\), for some \(g\thinspace \in\thinspace \sharpPComplete\). We will prove \(\sharpPComplete\)-ness of the problem of our concern, by adhering to such a recursive definition. The level of hardness of a \(\sharpPComplete\) problem can be appreciated by a consequence of a result shown by Toda, stating that a polynomial time algorithm that is allowed a single access to a \(\sharpP\) oracle, can solve any problem in \(\PH\), i.e, \(\P^{\sharpP} = \NP \cup \NP^{\NP} \cup \NP^{\NP^{\NP}} \dots\). A particular example of a counting function that will be of interest in the coming sections is the *weight enumerator* function for a linear code. The corresponding decision problem, which is to determine if there exists a codeword in \(\mathcal{C}\) of weight \(i\), is known to be in \(\NPComplete\). This immediately implies that \(\WE_{i}(\mathcal{C}) \in \sharpP\), and furthermore, it is known to be complete for this class. Hence to show that a problem of computing \(f\) is in \(\sharpPComplete\), in addition to its membership in \(\sharpP\), it suffices to show that for any \((n,k)\) linear code \(\mathcal{C}\), \(\WE_{i}(\mathcal{C})\), \(i\in o(\textsf{polylog}(n))\) can be computed in polynomial-time by allowing at most polynomially many queries to an oracle for computing \(f\). # Stabilizer codes {#sec-stab-codes} In this section, we will provide the necessary background material on stabilizer codes, see for more complete introductions. In the setting of quantum coding, information is encoded into \(n-\)qubit states in \(\mathcal{H}^{n}_{2} = (\mathbb C^2)^{\otimes n}\). Errors are operators acting on this space, modeled as elements of the *Pauli group* \({\mathcal{G}}_{n}\). Though the definition of the Pauli group in Def. ([\[def-pauli-group\]](#def-pauli-group){reference-type="ref" reference="def-pauli-group"}) contains the scalars \(\{\pm 1, \pm i\}\), they are often considered unimportant for error detection or correction as they do not affect the error syndrome nor the error correction. This enables us to define the effective Pauli group by identifying operators that are related by a multiplicative constant in \(\{\pm 1, \pm i\}\), denoted by \(\overline{\mathcal{G}}_{n} = \mathcal{G}_{n}/{\{\pm 1, \pm i\}}\). The number of qubits affected by an error \(E\in\overline{\mathcal{G}}_{n}\) is the number of non-identity components in the tensor product form of \(E\), and is called the *weight* of the error, denoted by \(\wt(E)\). A quantum stabilizer code is defined by a set of check operators, which are also elements of the Pauli group. Stabilizer codes are by far the most widely studied codes in the quantum setting. ## Operator basis When \(\mathcal{S}\) is generated by \(n-k\) independent *stabilizer generators* \(\{S_j\}_{j=1\ldots n-k}\), the subspace \(\mathcal{Q}\) has dimension \(2^k\), so it encodes \(k\) logical qubits, and we denote its parameters by \([[n,k]]\). All the elements in \(\overline{\mathcal{G}}_{n}\) that leave the code \(\mathcal{Q}\) globally invariant belong to the *normalizer* of \(\mathcal{S}\) in \(\overline{\mathcal{G}}\), denoted by \(\mathcal{N}(\mathcal{S})\). The normalizer \(\mathcal{N}(\mathcal{S})\) forms a group generated by \(n+k\) generators and \(\mathcal{S} \subset \mathcal{N}(\mathcal{S})\). However, not all operators in \(\mathcal{N}(\mathcal{S})\) necessarily fix every state in \(\mathcal{Q}\). Since the action of \(\mathcal{S}\) is trivial on all code states, we define a subset of operators in \(\mathcal{N}(\mathcal{S})\) called the *logical operators* that represent the quotient space \(\mathcal{L} = \mathcal{N}(\mathcal{S})/\mathcal{S}\), each of which have a distinct action on individual states in \(\mathcal{Q}\). For a shorthand notation, we will identify these representative with elements of \(\mathcal{L}\). This group has \(2\thinspace k\) *canonical logical generators* \(\{\overline X_j,\overline Z_j\}_{j=1\ldots k}\), such that all generators mutually commute except the pairs \(\overline X_j\) and \(\overline Z_j\) that anti-commute. The smallest weight of a nontrivial logical operation is the distance of the code, \[\begin{gathered} d = \min_{E\in\mathcal{N}(\mathcal{S})\backslash\mathcal{S}}\wt(E). \end{gathered}\] Note that the operators defined so far, \(\{S_j\}_{j=1\ldots n-k}\) and \(\{\overline X_j,\overline Z_j\}_{j=1\ldots k}\) do not generate \(\overline{\mathcal{G}}_{n}\). To complete the basis, we need to define the group of *pure errors* \(\mathcal{T} = \mathcal{N}(\mathcal{L})/\mathcal{S}\), which is also Abelian. We can always find a set of *canonical pure error generators* \(\{T_j\}_{j=1\ldots n-k}\) such that \(T_j\) commutes with all other pure error generators, all logical generators, and all stabilizer generators except \(S_j\) with which it anti-commutes. To summarize, we have the canonical basis of the Pauli group \(\{S_i, T_i, \overline X_j, \overline Z_j\}_{i=1\ldots n-k, j=1\ldots k}\) with all commutation relations trivial, except \[T_iS_i =-S_iT_i \quad {\rm and}\quad \overline X_j\overline Z_j =-\overline Z_j\overline X_j. \label{commutations}\] Any Pauli operator \(E \in \bar{\mathcal{G}}_{n}\) can be expressed as a product of elements in these respective groups: \[\begin{gathered} E = T\cdot L\cdot S \label{decomp-pauli-op} \qquad (\text{where \(T\in\mathcal{T}\), \(L\in\mathcal{L}\) and \(S\in\mathcal{S}\)}). \end{gathered}\] Decomposition in this basis will be particularly useful to formulate the decoding problem. ## Degenerate and non-degenerate errors {#sec-deg-err} Two errors \(E\) and \(E'\) are called *degenerate* if they have an identical effect on all code states, i.e. \(E\ket\psi = E'\ket\psi\thinspace, \thinspace \forall \ket\psi\in\mathcal{Q}\). Given the decomposition in Eq. ([\[decomp-pauli-op\]](#decomp-pauli-op){reference-type="ref" reference="decomp-pauli-op"}) and the definition of the code in Eq. ([\[eq-stabilize\]](#eq-stabilize){reference-type="ref" reference="eq-stabilize"}), we see that this is only possible if the two errors are related by an element of \(\mathcal{S}\), i.e., \(E' = E\cdot S\) for some \(S\in \mathcal{S}\). This naturally leads to an equivalence relation between errors, with two errors belonging to the same equivalence class if they are related by an element of \(\mathcal{S}\). In other words, the set of equivalence classes is the quotient space \(\overline{\mathcal{G}}_{n}/\mathcal{S} \sim \mathcal{L}\cdot \mathcal{T}\). We can thus label the equivalence classes by \(L,T\) with \(L\in \mathcal{L}\) and \(T\in \mathcal{T}\). For a fixed \(T\), the different classes labelled by \(L\in\mathcal{L}\) are referred to as *logical classes* and each class is of size \(2^{n-k}\). The class labelled by \(L = \mathbb{I}\) and \(T = \mathbb{I}\) is \(\mathcal{S}\) itself. Any other logical class can be expressed as a coset of \(\mathcal{S}\), but however they are not groups by themselves. Note that degeneracy is unique to the quantum error correction setting. When a bit flip pattern \(e\) is applied to a classical bit string \(x\), the resulting string \(y = x+e\) always differs from the original one, except for the trivial error \(e=0^n\). Consequently, each logical class contains only one element. Our main result in this paper shows that accounting for degenerate errors in decoding stabilizer codes greatly increases its computational complexity. ## Symplectic representation {#sec-symplectic} There is a one-to-one correspondence between *classical symplectic linear codes* in \(\mathbb{Z}^{2n}_{2}\) and stabilizer codes, which follows from a mapping \(\eta\) of pauli operators in \(\overline{\mathcal{G}}_{n}\) into binary strings of length \(2n\). The mapping is performed by first expressing every \(\textsf{M}\in\overline{\mathcal{G}}_{n}\) in the form: \[\begin{gathered} \textsf{M} = \left(\textsf{A}_{1}\otimes \dots \otimes \textsf{A}_{n}\right)\cdot\left(\textsf{B}_{1}\otimes\dots\otimes \textsf{B}_{n}\right) \thickspace, \thickspace\thickspace \textsf{A}_{i}\in\{Z,I\} \thickspace, \thickspace\thickspace \textsf{B}_{j}\in\{X,I\}\thickspace, \thickspace \thickspace 1\leq i,j\leq n \thickspace \label{general-stabilizer-element-X-Z} \end{gathered}\] and substituting \(X\) and \(Z\) by \('1'\) and \(\mathbb{I}\) by \('0'\). For instance, \(\eta(X\otimes Z\otimes I\otimes Y)=0 1 0 1 1 0 0 1\). This maps \(\overline{\mathcal{G}}^n\) into its *symplectic representation* in \(\mathbb{Z}^{2n}_{2}\). Moreover, any two mutually commuting operators \(P,Q \in \mathcal{G}_n\) are mapped into binary strings \(x,y \in \mathbb{Z}^{2n}_{2}\) that are orthogonal under the *symplectic product*, defined by \((x,y) = x \Lambda y^{\text{T}}\) with \(\Lambda = I_n\otimes X\). This immediately implies that \(\eta(\mathcal{N}(\mathcal{S}))\) is a vector space (code) which is the kernel of \(\eta(\mathcal{S})\) under the symplectic product. The parity check matrix for this classical code is \(\eta(\{S_{i}\}_{i=1}^{n-k})\) and its generator matrix is \(\eta(\{\overline{X}_{j},\overline{Z}_{j}\}_{j=1}^{k},\{S_{i}\}_{i=1}^{n-k})\). For a \(2n\)-bit string \(b = (z|x)\) expressed as the concatenation of two \(n\)-bit stings, we denote the inverse mapping \(\eta^{-1}(b) = Z^z \cdot X^x\) where we use the shorthand notation \(Q^x = Q^{x_1}\otimes Q^{x_2}\otimes\cdots\otimes Q^{x_n}\). The mapping \(\eta\) implicitly indicates that we consider a \(Y-\)type Pauli operation as a \(Z-\)type operation followed by a \(X-\) type operation, in other words, as an error of weight two. Consequently, we define the *symplectic-weight* of the Pauli error as \(\wt_{2}(E) = |\eta(E)|\) with \(|\cdot|\) denoting the usual Hamming weight. Clearly, \(2\wt(E)\geq \wt_{2}(E) \geq \wt(E)\). This correspondence with classical linear codes is the key to most of the complexity results involving stabilizer codes, including ours. In many cases it can be used to build parallels to already known results for classical (linear) symplectic codes. ## Error model An error model assigns probabilities to various errors, which are then used by the decoder to statistically infer what recovery is most likely given the error syndrome. We restrict ourselves to errors from \(\overline{\mathcal{G}}_{n}\), so the corresponding error model is referred to as a *Pauli channel*. There are multiple type of Pauli channels which are often used in studying error correcting codes. Here, we will further assume that errors act independently on each qubit. In general, the probabilities \(q_{i, E_{i}}\) can be different for all qubits. One important feature of memoryless Pauli channels is that they can be efficiently specified (to finite accuracy), i.e. using \(\mathcal{O}(n)\) bits of information. An obvious simplification of the above channel is made by supposing that the noise rates are the same for all \(X,Y,Z\) type errors on each qubit. This specifies to a *depolarizing channel*, and can be expressed as \[\begin{gathered} \Prob(E) = \left(\dfrac{p}{3}\right)^{\wt(E)}\times (1-p)^{n-\wt(E)}. \label{eq-depol-err} \end{gathered}\] Alternatively, one can assume that each qubit is first affected by \(X\)-type errors, and a then by \(Z\)-type errors, and moreover, these errors are independent of each other. A \(Y\)-type error occurs only when both a \(Z-\)type as well as an \(X-\)type error affect a qubit. This specifies a *independent \(X-Z\) channel*, which we will use to prove our main result. One key feature of both the depolarizing channel and the independent \(X-Z\) channel is that the probability of an error depends only on its weight (either \(\wt\) or \(\wt_2\)). As a consequence, evaluating the probability of a logical class can be done by counting the number of its elements of a given weight, which puts the problem in \(\sharpP\). Notice also that \(\Prob(E)\) is monotonically decreasing with its weight for \(p\in[0,\slfrac{1}{2}]\), implying that in such a range, a minimum weight error has the maximum probability. We will often refer to \(p\) to as the error-rate per qubit or the physical noise-rate. # The decoding problem {#decoding-problem} In this section we define the decoding problem more formally, and explain how it is affected by the existence of degenerate errors as defined in Sec. ([3.2](#sec-deg-err){reference-type="ref" reference="sec-deg-err"}). The qubits are prepared in a code state \(\ket\psi \in \mathcal{Q}\) and are subject to the memoryless Pauli channel (def. [\[def-pauli-ch\]](#def-pauli-ch){reference-type="ref" reference="def-pauli-ch"}). The received state is \(\ket\phi = E \ket\psi\) where \(E \in \overline{\mathcal{G}}_n\) is an unknown error chosen from the distribution \(\Prob(E)\). *Decoding* refers to the operation performed by the receiver in recovering the state \(|\psi\rangle\) from the state \(|\phi\rangle = E|\psi\rangle\). Since all Pauli operators square to the identity, it suffices for the receiver to determine \(E\) and apply it to the system to recover the original state, i.e. \(E\ket\phi = E^{2}\ket\psi = \ket\psi\). The error, being an element of the Pauli group, can either commute or anti-commute with each of the stabilizer generators. Thus, upon measurement of each stabilizer generator, we obtain an eigenvalue \(+1(-1)\) indicating that \(E\) commutes (anti-commutes) with the stabilizer generator: \[S_j\ket\phi = S_{j}\cdot E\ket\psi = \left\{ \begin{array}{llll} E\cdot S_{j}\ket\psi &= E\ket\psi &= +\ket\phi & {\rm if\ } E\cdot S_{j} = S_{j}\cdot E \\-E\cdot S_{j}\ket\psi &= E\ket\psi &=-\ket\phi & {\rm if\ } E\cdot S_{j} =-S_{j}\cdot E. \end{array} \right.\] Each \(\pm 1\) measurement outcome \(m_{j}\) is encoded into a bit \(s_{j}\) such that \(m_{j} = (-1)^{s_{j}}\). The outcomes of measuring all the check operators is encoded as a \((n-k)\) bit vector \(\vec{s}\) called the *error syndrome*. ## Non-degenerate decoding The decoding problem consists in identifying \(E\) conditioned on knowledge of the error syndrome. As in the classical case, the conceptually simplest strategy is to choose \(E\) that has the highest probability amongst all errors consistent with the measured syndrome. This is called *Quantum Maximum Likelihood Decoding* (\(\QMLD\)), and can be formulated mathematically as: \[E_{\QMLD}(\vec s) = \argmax_{E\in \bar\mathcal{G}_n} \Prob(E\thickspace |\thickspace \vec s)\] Using the decomposition Eq. ([\[decomp-pauli-op\]](#decomp-pauli-op){reference-type="ref" reference="decomp-pauli-op"}), we can view the error probability \(\Prob(E)\) as a joint probability over the group \(\mathcal{T}\), \(\mathcal{L}\), and \(\mathcal{S}\) in a natural way: \[\Prob(T,L,S) = \Prob(E = T\cdot L\cdot S), \quad {\rm with \ } T\in \mathcal{T},\ L\in \mathcal{L}, \rm{ \ and \ } S\in \mathcal{S}.\] Using the commutation relations given at Eq. ([\[commutations\]](#commutations){reference-type="ref" reference="commutations"}), it follows that the knowledge of \(\vec s\) is equivalent to knowledge of \(T\), since \(T_j\) is the only element of this basis that anti-commutes with \(S_j\). Thus, we have \(T_{\vec s} = \prod_{j} T_{j}^{s_{j}}\), and all errors in \(T_{\vec{s}}\cdot\mathcal{N}(\mathcal{S})\) are consistent with the syndrome \(\vec{s}\). Hence, at this stage a *best guess* for the elements in \(\mathcal{S}\) and \(\mathcal{L}\) needs to be employed. This involves finding \(L\cdot S\) that *maximizes the likelihood* of \(E = T_{\vec{s}}\cdot L\cdot S\), implying an equivalent definition of \(\QMLD\): \[E_{\QMLD}(\vec s) = T_{\vec s} \cdot \argmax_{L\in \mathcal{L}, S\in \mathcal{S}} \Prob(L,S | T_{\vec s}) \label{maxQMLD}\] where the conditional probability is given by Bayes' rule \(\Prob(L,S|T_{\vec{s}}) = \Prob(L,S,T_{\vec{s}})/\Prob(T_{\vec{s}})\) with the marginal defined as usual \(\Prob(T_{\vec{s}}) = \sum_{L,S} \Prob(L,S,T_{\vec{s}})\). Informally speaking, \(\QMLD\) addresses the problem of determining the element of \(\mathcal{L}\cdot \mathcal{S}\), whose probability is maximum, given an error rate and a syndrome. For the special cases of the depolarizing channel Eq. ([\[eq-depol-err\]](#eq-depol-err){reference-type="ref" reference="eq-depol-err"}) and the independent \(X-Z\) channel Def. ([\[def-XZ-ch\]](#def-XZ-ch){reference-type="ref" reference="def-XZ-ch"}), the search for an operator with maximum probability is synonymous to the search for an operator of minimum weight (with two possible notions of weight, one over the \(\overline{\mathcal{G}}_{n}\) and the other over \(\mathbb{Z}_{2}^{2n}\)). Consequently, \(\QMLD\) is also known as a *minimum-weight decoder*. One subtlety arises in case where the maximum probability is a close tie, as we cannot expect a decoder to discriminate probabilities within an arbitrary accuracy. Thus, we can define \(\QMLD\) as the problem of identifying the optimal couple \(L, S\), but we tolerate that it fails when more than one choice have probabilities that are within a small distance \(\Delta\) from the optimal. The standard way of formalizing this notion is with a *promise gap*, where we just assume in the definition of the problem that there is no close tie. As mentioned above, for the depolarizing channel and the independent \(X-Z\) channel \(\QMLD\) is formally equivalent to a minimum-weight decoder, in which case the promise gap is irrelevant and can be set to \(0\). ## Degenerate decoding {#sec-deg-dec} We will now explain how degeneracy changes the decoding strategy. As explained in Sec. ([3.2](#sec-deg-err){reference-type="ref" reference="sec-deg-err"}), errors can be classified into equivalence classes labelled by \(L,T\), with all errors within a class having the same effect on the code and therefore all being correctable by the same operation. As a consequence, we see that \(\QMLD\) is a suboptimal decoding strategy---in the sense that it does not reach the maximum probability of correctly decoding---because it fails to recognize the equivalence between degenerate errors. Instead of searching the most likely error, the optimal decoder seeks for the most likely equivalence class of errors, with the probability of a class of errors equal to the sum of the probability of the errors it contains. Since all errors in an equivalence class are related by an element of \(\mathcal{S}\) and their \(\mathcal{T}\) component is fixed by the syndrome, we can write the probability of a class conditioned on syndrome \(\vec{s}\) as \[\Prob(L|\vec s) = \sum_{S\in \mathcal{S}} \Prob(L,S|T_{\vec s}),\label{eq:PLS}\] where we use standard Bayesian calculus as above. The *Degenerate Maximum Likelihood Decoding* (\(\DQMLD\)) problem can be formulated mathematically as determining an error in the most probable logical class, for a given syndrome, Eq. ([\[eq:PLS\]](#eq:PLS){reference-type="ref" reference="eq:PLS"}). Note that the promise gap can be expressed as a relative gap or an additive gap, the former being the right one in our setting. This is because the relative promise gap can be related to the failure probability of the code under optimal decoding. Consider a large promise gap \(\Delta = 1-4^{-k}\epsilon\). Rewriting the promise as \(P(L|T_{\vec s}) \leq 4^{-k} \epsilon P(L^*|T_{\vec s})\) and summing over all \(L\neq L^*\) (of which there are \(4^k-1\)), we arrive at \(P(L^*|T_{\vec s}) \geq 1-2\epsilon\), which simply says that the probability that the error that occurred is not equivalent to \(L^*\)---and hence that the decoder fails---is at most \(2\epsilon\). Note also that for a fixed \(T\), the probabilities \(\Prob(L,S|T)\) and \(\Prob(L,S,T)\) differ only by a constant, so we can perform the optimization in Def. ([\[def-DQMLD\]](#def-DQMLD){reference-type="ref" reference="def-DQMLD"}) on the joint probability instead of the conditional probability. The sum appearing in the this probability Eq. ([\[eq:PLS\]](#eq:PLS){reference-type="ref" reference="eq:PLS"}), being over \(2^{n-k}\) terms, forbids a polynomial-time direct computation of its value. However, for i.i.d. Pauli channels such as the depolarizing channel and the independent \(X-Z\) channel, by grouping terms of equal weight in the sum, we can express the sum in Eq. ([\[eq-logical-prob\]](#eq-logical-prob){reference-type="ref" reference="eq-logical-prob"}) more succinctly. For the case of the independent \(X-Z\) channel Def. ([\[def-XZ-ch\]](#def-XZ-ch){reference-type="ref" reference="def-XZ-ch"}), the above joint probability can be expressed as \[\begin{gathered} \Prob(T_{\vec{s}},L,S) = \left(1-\dfrac{p}{2}\right)^{2n}\sum_{S\in\mathcal{S}}\tilde{p}^{\wtt(T_{\vec{s}} \cdot L\cdot S)} \label{eq-logical-prob} \end{gathered}\] where \(\tilde{p} = \slfrac{p}{(2-p)}\). By grouping terms of equal weight in the sum, we arrive at a sum involving only \(n+1\) terms \[\begin{gathered} \Prob(T_{\vec{s}},L,S) = \left(1-\dfrac{p}{2}\right)^{2n} \sum_{i = 0}^{n}A_{i}(\vec{s},L) \thinspace \tilde{p}^{i} \label{dqmld-we-coeff} \\ \text{where }A_{i}(\vec{s},L) = \left| \{S\in\mathcal{S}\thickspace: \thickspace \wtt(E = T_{\vec{s}}\cdot L\cdot S) = i\}\right|, \text{ and } \sum_{i=0}^{n}A_{i}(\vec{s},L) = 2^{n-k}. \label{we-coeff-sum} \end{gathered}\] The coefficients \(\{A_{i}(\vec{s},L)\}_{i=0}^{n}\) are called the *weight enumerators of the coset* associated to \(\vec s\) and \(L\). Note that the coset weight enumerators play a very important role in estimating the decoder performances of both \(\QMLD\) as well as \(\DQMLD\). The sum in Eq. ([\[dqmld-we-coeff\]](#dqmld-we-coeff){reference-type="ref" reference="dqmld-we-coeff"}) is now over polynomially many terms unlike its previous form Eq. ([\[eq-logical-prob\]](#eq-logical-prob){reference-type="ref" reference="eq-logical-prob"}). Computing such a sum for each logical operator and subsequently optimizing over their values would solve \(\DQMLD\). An \([[n,k]]\) stabilizer code has \(|\mathcal{L}| = 4^{k}\), implying that even if the weight enumerators can be computed efficiently, a polynomial-time optimization cannot be performed over the different cosets labeled by \(L\), unless \(k \in \mathcal{O}(\log(n))\), which is the regime that we are interested in. Furthermore, we believe \(\DQMLD\) is at least as hard otherwise, i.e for \(k \in \Omega(\log(n))\). At this stage, we like to remark that though \(\QMLD\) and \(\DQMLD\) are stated as decision problems in, the problem of practical interest is one of determining a Pauli operator that maximizes the respective probabilities in Defs.([\[defn-QMLD\]](#defn-QMLD){reference-type="ref" reference="defn-QMLD"}) and ([\[def-DQMLD\]](#def-DQMLD){reference-type="ref" reference="def-DQMLD"}). To enable a decision problem formulation in, both \(\QMLD\) and \(\DQMLD\) are defined to take as input an additional constant \(c\). Subsequently, \(\DQMLD\) would be made to decide the existence of an \(L\in\mathcal{L}\) whose probability Eq. ([\[eq-logical-prob\]](#eq-logical-prob){reference-type="ref" reference="eq-logical-prob"}) is at least \(c\). This appears to be achieving more than what is really expected from of \(\DQMLD\) (or \(\QMLD\)) in practice. In particular, by varying this constant \(c\), along with the input in Def. ([\[def-DQMLD\]](#def-DQMLD){reference-type="ref" reference="def-DQMLD"}), one could use the oracle to not only learn the optimal correction but also the probability of its equivalence class. The latter is not necessary to perform error correction and thus gives more power to the decoder oracle than it should. This is why we formulated \(\DQMLD\) as a function problem that does not explicitly reveal the probability of the equivalence class of the optimal correction, and therefore we consider it to be closer to the real world decoding problem. ## Importance of degeneracy Before addressing the computational complexity of degenerate decoding, we close this section with a discussion of its practical relevance. The two decoders \(\QMLD\) and \(\DQMLD\) will provide different answers whenever the most likely equivalence class does not contain the error with the largest probability. Consider the hypothetical scenario where the class of error \(L_1\) contains a single error of low weight \(a\) and \(2^{n-k}-1\) errors of high weight \(b \gg a\), and that the class \(L_2\) contains \(2^{n-k}\) errors of intermediate weight c, \(a\ll c\ll b\). The \(\QMLD\) would chose the error from the class \(L_1\), because it is the most likely error. On the other hand, the probabilities of these two classes are given by \[\Prob(L_1) \propto \tilde p^a + (2^{n-k}-1) \tilde p^b \approx \tilde p^a \ {\rm and} \ P(L_2) \propto 2^{n-k} \tilde p^c.\] Thus, we see that \(\DQMLD\) would provide a different answer if \(P(L_1) < P(L_2)\), or equivalently \(\tilde p > 2^{-\frac{n-k}{c-a}}\), and that the two decoders would agree otherwise. Thus, for sufficiently low error rates, and in particular in the extreme limit \(p \leq 2^{k-n}\), degeneracy does not affect the decoding problem. More importantly, degeneracy becomes unimportant when the *failure rate* of the code is very low, which doesn't necessarily require a low physical noise rate. Remember Sec. ([4.2](#sec-deg-dec){reference-type="ref" reference="sec-deg-dec"}) that the failure rate \(2\epsilon\) of the code is related to the \(\DQMLD\) promise gap \(\Delta = 1-4^{-k}\epsilon\). The following lemma, whose proof is presented in App. ([\[app-large-gap\]](#app-large-gap){reference-type="ref" reference="app-large-gap"}), shows that for \(\epsilon = 2^{k-n}\), \(\QMLD\) provides the same answer as \(\DQMLD\), which in turn implies that \(\DQMLD\) is in \(\NP\) with such a large gap. In the light of these observations, one might imagine that in general, \(\DQMLD\) can at most offer a marginal improvement over \(\QMLD\), i.e. that for a given code and noise rate, the decoding error probability of \(\QMLD\) is upper bounded by a function of the decoding error probability of \(\DQMLD\). There is strong evidence that this is not the case however. Monte Carlo simulations have shown that \(\QMLD\) and \(\DQMLD\) achieve different error thresholds with Kitaev's topological code. (This statement is equivalent to the fact that the critical disorder strength separating the ordered from the disordered phase in the random bound Ising model decreases below the Nishimori temperature.) Thus, for noise rates \(p\) falling between these two thresholds, the failure probability of \(\DQMLD\) tends to 0 as the number of qubit \(n\) increases, while the failure probability of \(\QMLD\) tends to \(1-4^{-k}\) (the failure probability of a random guess), so the performances of both decoders can be significantly different. Degeneracy can also severely impact the performances of certain decoding algorithms. Due to degeneracy of quantum errors, the ability to correct errors does not imply that the a posteriori marginal error probability over individual qubits is sharply peaked, in contrast to the classical setting. This is the case in particular for low density parity check (LDPC) codes. These codes have the property of admitting a set of stabilizer generators \(S_j\) of weight bounded by a constant. As a consequence, the weight of equivalent errors \(E\) and \(E'\) related by a stabilizer generator \(E' = ES_j\) will differ at most by a small constant. Thus, we expect degeneracy to play an important role: each equivalence class contains many errors of roughly the same weights. As discussed in, conventional decoding algorithms for LDPC codes (belief propagation ) are *marginal decoders* in the sense that they optimize the probability of error independently for each qubit. But in the presence of degeneracy, we can have a probability sharply peaked over a single equivalence class of errors---ensuring the success of \(\DQMLD\)---but yet have a very broad marginal distribution over individual qubits---leading to a failure of a marginal decoder. So the existence of a good, general purpose decoder for quantum LDPC codes, playing a role analogous to belief propagation in the classical setting, remains an outstanding open question. This situation is best illustrated with Kitaev's topological code. In this code, errors correspond to strings on a regular square lattice, and error syndromes are located at the endpoints of the strings. The weight of an error is equal to the length of the corresponding string. Lastly, the different equivalence classes of errors correspond to the homology classes of the lattice. For a syndrome configuration shown at Fig. ([\[fig-toric-code-marginal\]](#fig-toric-code-marginal){reference-type="ref" reference="fig-toric-code-marginal"}), all the short paths have the same homology, so the probability is sharply peaked over one equivalence class. But there are several distinct strings of the same length compatible with this syndrome, so the marginal error probability over individual qubits is very broad. Lastly, we note that to achieve the true capacity of certain quantum channels (as opposed to the single-shot capacity), it is necessary to encode the information in a degenerate code. In other words, there are certain channels that could not be used to send any quantum information if a non-degenerate code was used, but can reliably do so at a finite rate with degenerate codes. We do not know however if degeneracy also needs to be taken into account during the decoding process to realize this. In particular, the example in uses a generalization of Shor's code, for which \(\QMLD\) and \(\DQMLD\) always yield the same output. We know of only a few examples of codes for which \(\DQMLD\) can be computed efficiently, namely concatenated codes, convolutional codes, and Bacon-Shor codes. There exist heuristic methods to take degeneracy into account in topological codes and turbo codes. # Complexity of the decoding problem {#sec-hardness} The one-to-one correspondence between \([n,k]\) stabilizer codes and \((2n,k)\) symplectic linear codes is used in to show that a solution for \(\QMLD\) can be used to decide an \(\NPComplete\) problem in polynomial time. Consequently, \(\QMLD\in\NPComplete\). \(\DQMLD\) was shown to be \(\NPHard\) for the case of an independent \(X-Z\) channel in and depolarizing channel in, using a reduction from a \(\NPComplete\) problem pertaining to classical linear code. We now state our main result, which establishes that \(\DQMLD\) is in fact much harder than what there previous results anticipated. In the following sections, we will show that for a classical binary linear code \(\mathcal{C}\) and \(\lambda \in [0,n]\), the problem of computing \(\WE_{i}(\mathcal{C})\) for \(i=0,1,2,\ldots, \lambda\) is polynomial time Turing reducible to \(\DQMLD\) on an independent \(X-Z\) channel, for \(\Delta \leq 2[2 + n^{\lambda}]^{-1}\). In the case of a general memoryless Pauli channel, it is not always possible to express the probability of a logical class as a weight enumerator sum Eq. ([\[dqmld-we-coeff\]](#dqmld-we-coeff){reference-type="ref" reference="dqmld-we-coeff"}) with polynomially many terms in the sum. Hence, in such cases, the containment of \(\DQMLD\) in \(\sharpP\) is not known and we can only claim that it is \(\sharpPHard\). However, whenever one can express \(\Prob(L|\vec{s})\) as a sum with polynomially many terms, each of which is a \(\sharpP\) function, then \(\DQMLD\) can be put into \(\sharpP\). For the independent \(X-Z\) channel, the containment in \(\sharpP\) is straightforward. # Reduction {#sec-red-over} In this section, we present a polynomial time algorithm that accepts as input a classical linear code \(\mathcal{C}\) and outputs \(\{\WE_{i}(\mathcal{C})\}_{i=0}^{n}\) as in Def. ([\[WE-def-code\]](#WE-def-code){reference-type="ref" reference="WE-def-code"}) by querying a \(\DQMLD\) oracle with an independent \(X-Z\) noise model and a promise gap \(\Delta\) which is \(\slfrac{1}{\textsf{quasi-polynomial}(n)}\). ## Reduction overview The correspondence between symplectic linear codes and stabilizer codes discussed in Sec. ([3.3](#sec-symplectic){reference-type="ref" reference="sec-symplectic"}) implies that with an independent \(X-Z\) channel, the probability of a logical class is related to the weight enumerator polynomial of a classical linear code. In particular, with the trivial syndrome \(\vec s = \vec 0\) and at very low noise rate \(p\ll \slfrac{1}{2}\), the most likely equivalence class is always the trivial one \(L = \mathbb{I}\). Hence, in this setting, the probability of the trivial logical class for a quantum code with stabilizers \(\mathcal{S}\) can be expressed, similar to Eq. ([\[dqmld-we-coeff\]](#dqmld-we-coeff){reference-type="ref" reference="dqmld-we-coeff"}), in terms of a corresponding classical code \(\mathcal{C}_{\mathcal{S}}\) as follows \[\begin{gathered} \Prob(\mathbb{I}\thinspace|\thinspace\vec{0}) = \left(1-\dfrac{p}{2}\right)^{2n}\sum_{i=0}^{2n}\WE_{i}(\mathcal{C}_{\mathcal{S}})\tilde{p}^{i} \label{prob-I-WE} \end{gathered}\] Since this probability is a polynomial of degree \(2n\), determining its value for \(2n+1\) distinct values of the physical noise rate \(p\) would enable us to determine the weight enumerator of the corresponding classical code. There are two caveats to this approach. First, this approach only works for classical codes \(\mathcal{C}_{\mathcal{S}}\) whose generator matrix corresponds to the symplectic representation of a quantum stabilizer code. Second, this approach requires knowledge of the probability of a certain equivalence class, while \(\DQMLD\) only outputs the equivalence class with the largest probability; it does not reveal the value of the corresponding probability. The first caveat can easily be circumvented, for instance by padding the \(n\)-bit classical code with \(n\) additional \(0\)s, thus obtaining a valid symplectic representation of a \(n\)-qubit quantum code (one whose stabilizer generators contain only \(Z\) operators). Consequently, all stabilizers have weights between 0 and \(n\) and the probability of the trivial logical class can be expressed as Eq. ([\[prob-I-WE\]](#prob-I-WE){reference-type="ref" reference="prob-I-WE"}), where the range of sum is up to \(n\). To circumvent the second caveat, we need to use the \(\DQMLD\) oracle to obtain equality constraints on the weight enumerator of \(\mathcal{C}\). This is done by varying the physical noise rate \(p\), always keeping the syndrome trivial. As mentioned above, at very low noise rate the optimal logical class is \(\mathbb I\). Increasing the noise rate \(p\), we will reach a *crossing point* \(p_{1}\) where the \(\DQMLD\) output changes from \(\mathbb{I}\) to \(L^{\star} \neq \mathbb{I}\). At this point, the promise gap condition is violated, i.e, \(\left|\Prob(\mathbb{I}|\vec s)-\Prob(L^{\star}|\vec s)\right| \leq \Delta \Prob(\mathbb{I}|\vec{s})\), which can be expressed in terms of weight enumerators as: \[\sum_{i=0}^{n} \WE_{i}(\mathcal{C})\tilde{p}_{1}^{i}-\sum_{i=0}^{n} B_{i} \tilde{p}_{1}^{i} \leq \Delta \sum_{i=0}^{n} \WE_{i}(\mathcal{C})\tilde{p}_{1}^{i} \label{eq-crossing-I-L}\] where \(B_{i}\) are the weight enumerators of an affine code. In the case where \(\Delta = 0\), this crossing point provides an equality condition between two polynomials, which is what we are seeking. But since these are polynomials with integer coefficients, knowing the location of a crossing point within a finite accuracy, which translate into a finite promise gap \(\Delta\), is enough to determine the exact crossing point, see Lemma. ([\[gap-size-qpoly\]](#gap-size-qpoly){reference-type="ref" reference="gap-size-qpoly"}). As we will show, there exists a fixed range of \(p\) that provably contains a unique crossing point, enabling a polynomial-time accurate determination of the crossing point using a binary search procedure. This gives us one potential equality, but introduces more unknown coefficients \(B_{i}\). To get additional linear constraints, we modify the code in a very special way, described in Sec. ([6.2](#sec-stab-cons){reference-type="ref" reference="sec-stab-cons"}). This modification requires adding one *tunable* qubit and one stabilizer generator. By varying the noise rate on the tunable qubit over a range of values, we can change the location of the crossing point, and thus obtain new linear constraints relating \(\{\WE_{i}(\mathcal{C})\}_{i=0}^{n}\) and the \(\{B_{i}\}_{i=0}^{n}\). Repeating this procedure \(2n+2\) times and making sure that all the linear constraints are linearly independent Sec. ([6.4](#sec-indep-cons){reference-type="ref" reference="sec-indep-cons"}) enable us to determine the weight enumerator coefficients. While the ability to change the noise rate of the tunable qubit gives us more linear constraints, it breaks the requirement that the noise model be the independent \(X-Z\) channel with the same strength on all qubits. We will fix this problem in App. ([\[app-concatenate\]](#app-concatenate){reference-type="ref" reference="app-concatenate"}) by showing that the required channel can be simulated by concatenating the code with a Shor code. In fact, we will use this technique repeatedly in our proof. ## Stabilizer code construction {#sec-stab-cons} Let \(G\) be the \(k\times n\) generator matrix of an \((n,k)\) classical linear code \(\mathcal{C} = \{x\in {\mathbb Z}_2^n: x = yG, \ y\in {\mathbb Z}_2^k \}\). Denote \(\{g_i\}_{i=1,\ldots k}\) the rows of \(G\) and let \(\{g_i\}_{i=k+1\ldots n}\) be a generating set of the complement of the row space of \(G\), i.e. in such a way that \(\{g_i\}_{i=1\ldots n}\) span \(\mathbb Z_2^n\). Construct a matrix \(\tilde G\) with rows\(\{g_i\}_{i=1\ldots n}\). This matrix is full rank, and therefore has an inverse \(H\) that can be computed efficiently, and obeys \(\tilde GH^{T} = \mathbb I\). Denote the rows of \(H\) by \(\{h_i\}_{i=1}^{n}\). We define a \([[2n-k-1,1]]\) quantum code with stabilizer generators and logical operators given by \[\begin{aligned} S_i &= Z^{g_i} & {\rm for }\ i=1,\ldots,k \\ S_{k+i} &= Z^{g_{k+i}} \otimes Z_{n+i} & {\rm for }\ i=1,\ldots,n-k-1 \\ S_{n-1+i} &= X^{h_{k+i}} \otimes X_{n+i} & {\rm for }\ i=1,\ldots,n-k-1 \\ \overline Z &= Z^{ g_n} &\\ \overline X &= X^{h_n}, & \end{aligned}\] where it is implicitly assumed that operators are padded to the right by identities to be elements of \(\mathcal{G}_{n+k-1}\). The validity of the resulting code can be verified from the fact that 1. There are in total \(2n-k-1\) qubits. 2. The \(2n-k-2\) stabilizer generators are independent. This follows from the linear independence of the \(h_i\) and the linear independence of the \(g_i\), together with the fact that \(X\)-type operators are linearly independent of \(Z\) type generators. 3. The stabilizer generators mutually commute. This is trivial among the first \(n-1\) generators as they contain only \(Z\) operators and similarly among the last \(n-k-1\) last generators. Between these two sets, the commutation follows from the fact that \(h_i \cdot g_j = 0\) except when \(i=j\), in which case the presence of additional \(X\) and \(Z\) on the \(n+i\)th qubit ensures commutation. 4. The logical operators commute with the stabilizer generators. This follows from the fact that \(h_i\cdot g_j = 0\) for \(i\neq j\), and the fact that \(X\)-type operators commute among themselves and similarly for \(Z\)-type operators. As discussed in Sec. ([4.2](#sec-deg-dec){reference-type="ref" reference="sec-deg-dec"}), the probability of the trivial logical class \(\mathbb I\) given a trivial syndrome \(\vec s = \vec 0\) is simply the sum of the probabilities of all stabilizer group elements. Suppose now that the last \(n-k-1\) qubits are error-free, while the other qubits are subject to an independent \(X-Z\) channel. Then, the probability of an element of \(\mathcal{S}\) is zero if it contains a generator \(S_i\) from the above list with \(i>k\). Otherwise, this element of \(\mathcal{S}\) can be written as \(S = Z^x \otimes {\mathbb I}\) for some \(x\in {\mathbb Z}_2^n\) and its probability is \((\slfrac{p}{2})^{|x|}(1-\slfrac{p}{2})^{2n-|x|}\). We conclude that the probability of the trivial logical class is given by Eq. ([\[prob-I-WE\]](#prob-I-WE){reference-type="ref" reference="prob-I-WE"}) with \(\mathcal{C}\) the classical code defined by the generating matrix \(G\). Constraints on the weight enumerator polynomial will be obtained by finding crossing points where \(\Prob({\mathbb I}|\vec 0) \approx \Prob(L|\vec 0)\) with \(L \neq {\mathbb I}\). For technical reasons, we would like to be able to choose which \(L\) will be the one realizing the crossing. This is because we want to force the crossing to happen with the same \(L\) every time. To do this, we will modify the stabilizer by adding an extra qubit and making the transformations \[\begin{aligned} S_i &\rightarrow S_i\otimes {\mathbb I} \\ \overline Z &\rightarrow \overline Z \otimes {\mathbb I} \\ \overline X &\rightarrow \overline X \otimes X \end{aligned}\] to the stabilizers and logical operators, and adding the following stabilizer generator \[S_{2n-k} = \overline Z \otimes Z.\] This defines an \([[2n-k,1]]\) stabilizer code, and its validity can easily be verified given the commutation relations worked out above. Moreover, if we assume that the added \((2n-k)\)th qubit is also error-free, then the probability of the trivial logical class \(\mathbb I\) given a trivial syndrome \(\vec s = \vec 0\) is unchanged, and moreover the only other logical class with non-zero probability is the one associated to \(\overline Z\), i.e. \(\Prob(\overline X|\vec 0) = \Prob(\overline Y|\vec 0) = 0\). We need to perform one last modification to the code in order to be able to tune the crossing point, and hence obtain linearly independent equalities between \(\Prob({\mathbb I}|\vec 0)\) and \(\Prob(\overline Z|\vec 0)\). This transformation is quite similar to the previous one, and given by \[\begin{aligned} S_i &\rightarrow S_i\otimes {\mathbb I} \\ \overline Z &\rightarrow \overline Z \otimes Z \\ \overline X &\rightarrow \overline X \otimes \mathbb I \end{aligned}\] and adding the following stabilizer generator \[S_{2n-k+1} = \overline X \otimes X.\] For this last qubit, we will assume a noise model where \(p_X = p_Y = 0\), \(p_Z = q\), and \(p_{\mathbb I} = 1-q\) with \(q\) being a tunable parameter. With this last choice, the only two non-zero probabilities of logical class conditioned on the trivial syndrome are given by \[\begin{aligned} \Prob({\mathbb I}|\vec 0) &= \frac 1\mathcal{Z} (1-q) \sum_{i=0}^{n} \WE_{i}(\mathcal{C})\tilde{p}_{1}^{i} \label{stab-cons-I} \\ \Prob(\overline Z|\vec 0) &= \frac1\mathcal{Z} q\sum_{i=0}^{n} B_{i} \tilde{p}_{1}^{i} \label{stab-cons-Z}, \end{aligned}\] where \(\tilde{p} = \slfrac{p}{(2-p)}\) as above, \(\mathcal{Z}\) is a suitable normalization factor, and \(B_i\) are the weight enumerators of the affine code associated to the \(\overline Z\) logical class \[B_i = \big| \{ x \in \mathcal{C}+g_n: |x| = i\} \big|.\] A crossing point is observed when \[v \sum_{i=0}^{n} \WE_{i}(\mathcal{C})\left(\dfrac{\tilde{p}_{1}}{2}\right)^{i}-\sum_{i=0}^{n} B_{i} \left(\dfrac{\tilde{p}_{1}}{2}\right)^{i} \leq \Delta v\sum_{i=0}^{n} \WE_{i}(\mathcal{C})\left(\dfrac{\tilde{p}_{1}}{2}\right)^{i} \label{CPv}\] where \(v = (1-q)/q\) is a tunable parameter over the positive reals. Changing the value of \(v\) will change the crossing point between these two logical classes, and provide linear constraints between two degree \(n\) polynomials. If we can identify \(2n+1\) such crossing points, it would provide enough information to retrieve the two polynomials, and hence solve the weight enumerator problem. ## Finding crossing points {#sec-crossing} At this point, we have a deterministic procedure that, given any classical linear code \(\mathcal{C}\), can be used to generate linear constrains on its weight enumerator coefficients. Clearly, the overhead in the runtime of this procedure is the time required to spot a crossing point. A crossing point can potentially be observed at a physical noise rate \(p\) anywhere between 0 and 1. One obvious indication of a crossing point is the switch in the output of the \(\DQMLD\) oracle as we move \(p\) across the crossing point. However if we move \(p\) across two crossing points, we will not notice any net switch in the outputs of the \(\DQMLD\) oracle. For this reason, we now want to restrict the values of \(p\) to a range where we can prove that there is at most one crossing point. This will be possible by restricting the tunable parameter \(v\) to a small interval near 0. where the last term is added because we have changed the range of summation. It only remains to show that the expression above is strictly negative. Clearly, this cannot be true for all \(p \in [0,1]\). However, when \(p\leq \slfrac{1}{n}\), we know that the first term in Eq. ([\[eq:rand\]](#eq:rand){reference-type="ref" reference="eq:rand"}) is non-negative. Indeed, all of its terms are positive by definition, except the one in square bracket which is non-negative when \(p\leq \slfrac{1}{n}\). We claim now that when \(p\leq \slfrac{1}{n}\), the second term of Eq. ([\[eq:rand\]](#eq:rand){reference-type="ref" reference="eq:rand"}) is negative and greater in norm than the first one, so the entire expression is strictly negative. This can be observed by the following inequality: Hence, we see that indeed the probability of the trivial logical class is strictly decreasing when \(0\leq p \leq \slfrac{1}{n}\). Since there are only two logical classes in our setting, it is clear that the probability associated to the other class is increasing in that interval. We have identified an interval where the probabilities are monotonic, and what remains to be shown is that there is indeed a crossing point inside this interval when the parameter \(v\) is chosen carefully. Intuitively, we can see that decreasing the value of \(v\) (and hence of the trivial logical class) will decrease the value of \(p\) where the first crossing point occurs. We will now set an upper bound \(p_{\max}\) on the value of \(p\) where the first crossing point occurs. The first crossing point will occur at the latest when \(v\Prob(\mathbb{I}|\vec{0})-\Prob(\overline{Z}|\vec{0}) = \Delta \Prob(\mathbb{I}|\vec{0})\), so equivalently \(v(1-\Delta) = \slfrac{\Prob(\overline{Z} | \vec{0})}{\Prob(\mathbb{I}|\vec{0})}\). On the other hand, the ratio \(\slfrac{\Prob(\overline{Z} | \vec{0})}{\Prob(\mathbb{I}|\vec{0})}\) is lower-bounded by \([p/(2-p)]^d\) since each word in \(x\in \mathcal{C}\) is mapped onto a word \(y=x+g_{n}\) of weight at most \(|x|+|g_{n}|\) in \(\mathcal{C}+g_{n}\). Hence, the first crossing occurs for a value of \(p\) lower or equal to the point where \[v = (1-\Delta)^{-1} \left[\dfrac{p}{2-p}\right]^{d}, \ {\rm or\ equivalently\ } p_{\max} = \dfrac{2}{1 + (1-\Delta)^{\frac{1}{d}}\thickspace v^{-\frac{1}{d}}}.\] By choosing \[\begin{gathered} p_{\max} \leq \dfrac{1}{n}, \thickspace \ {\rm or\ equivalently\ } \thickspace v\leq (1-\Delta)^{-1}\thickspace n^{-d}, \label{eq-max-crossing} \end{gathered}\] we are sure that the first crossing point occurs in the monotonic region, and hence that the interval \(0\leq p \leq \slfrac{1}{n}\) contains a single crossing point. ◻ Though it is clear that we must have \(2n+1\) linearly independent constraints on \(\{\WE_{i}(\mathcal{C}),B_{i}\}_{i=0}^{n}\) along with a promise gap of \(\Delta \leq 2[2 + n^{\lambda}]^{-1}\), it is not immediately clear how many distinct crossing points need to be located to obtain these constraints, i.e. some of the crossing points associated to distinct pairs \((v_l,p_l)\) may result in linearly dependent constraints Eq. ([\[CPv\]](#CPv){reference-type="ref" reference="CPv"}). The following lemma ensures that we can always efficiently find linearly independent constraints. The last ingredient we need is a bound on the distance between crossing points. Remember that we are only able to locate the value of a crossing probability \(p_k\) to exponential accuracy. Thus, it is necessary that changing the tunable parameter \(v\) has a significant effect on the value of the crossing point in order to generate linearly independent constraints (with significantly different values of \(p_k\)). Combining the restriction on the values of the tunable parameter in Eq. ([\[eq-max-crossing\]](#eq-max-crossing){reference-type="ref" reference="eq-max-crossing"}) with Lemma. ([\[indep-cons\]](#indep-cons){reference-type="ref" reference="indep-cons"}) immediately tells that the smallest change in the tunable parameter will be at least \((1-\Delta)n^{-n}8^{-1}n^{-2}\). This naturally implies a minimum separation between two crossing points, as the lemma below addresses: Since \(\gamma \geq (1-\Delta)n^{-n}8^{-1}n^{-2}\), we find: \(\delta \geq (1-\Delta)n^{-n}8^{-1}n^{-3}4^{1-n} \geq 4^{-n\log_{2}n}\). ◻ # Conclusion We will close with mentioning a few open problems which we were not able to address in this paper. In the course of this paper we have addressed the optimal decoding problem on an independent \(X-Z\) channel. However, the same can be done for a depolarizing channel by introducing the notion of a generalized weight described in. Hence, Sec. ([6.2](#sec-stab-cons){reference-type="ref" reference="sec-stab-cons"}) of the paper will undergo certain modifications when choosing to address a depolarizing channel. The key problem turns out to be the classification of the *parametrized* complexity of the decoding with the promise gap parameter Def. ([\[def-DQMLD\]](#def-DQMLD){reference-type="ref" reference="def-DQMLD"}), denoted by \(\Delta\). We know the complexities for two extreme cases, namely \(\DQMLD\) is \(\sharpPComplete\) when \(\Delta = \slfrac{1}{\textsf{quasi-polynomial}(n)}\), c.f. Thm. ([\[main-theorem\]](#main-theorem){reference-type="ref" reference="main-theorem"}) and in \(\NP\) when \(\Delta = 1-2^{-n-k}\), c.f. Lemma. ([\[lemma-large-gap\]](#lemma-large-gap){reference-type="ref" reference="lemma-large-gap"}). However, for a vast intermediate range of \(\Delta\), complexity of \(\DQMLD\) remains open. As described in Sec. ([4.2](#sec-deg-dec){reference-type="ref" reference="sec-deg-dec"}), for a code encoding a single qubit, the promise gap \(\Delta\) is related to the decoding failure probability \(\epsilon\) as \(\Delta = 1-\epsilon\). Thus, the two extreme cases considered above correspond respectively to optimal decoding in a very noisy regime (failure probability approaching unity) and optimal decoding with an exponentially small failure probability. Unfortunately, the case of practical interest falls somewhere in between. The complexity of any problem is only a highlight of the runtime of any algorithm on the worst case instance of the problem. Hence it is of practical interest to know the runtime of the algorithm for any *typical* instance. This could refer for instance to a typical syndrome or a random code. It is well known however that random codes are non-degenerate, so the decoding is not expected to be affected by the degeneracy in errors, so our result is probably not relevant in this setting. However, for the practically relevant class of *sparse codes*, the complexity of optimal decoding strategy remains an important open question. Lastly, our analysis has focused on stabilizer codes over Pauli channels. This is particularly convenient due to the discrete nature of the resulting decoding problem. This setting could be generalized in two obvious ways. First, we could consider codes that are not stabilizer codes, defined from a set of commuting projectors. There exists a growing interest for those codes, particularly in the setting of topological quantum order. In this setting, we could study for instance the decoding problem for systems that support non-Abelian anyons. Second, we could consider errors that are not described by Pauli operators. This problem is of practical importance because no real-world device undergoes a Pauli channel; for instance the physical process of relaxation is not described by a Pauli channel.

{'timestamp': '2013-10-14T02:10:58', 'yymm': '1310', 'arxiv_id': '1310.3235', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3235'}

# Distances and moments of inertia {#distances} Let us describe (an highly idealized) measurement which we expect to work for weak pulsars. A proper introduction, §[2](#intro){reference-type="ref" reference="intro"}, and discussion of the theory, §[3](#arist){reference-type="ref" reference="arist"}, follow. From the Fermi catalog we take the following: - period \(P\); - period derivative \(\dot{P}\); - bolometric flux \(f\) (\([f]={{\rm erg}\over {\rm cm}^2{\rm s}})\); - photon energy cutoff \(E_{\rm cut~obs}\); - bolometric lightcurve \(l_{\rm obs}(\phi )\), where \(\phi\) is the pulse phase; \(l_{\rm obs}(\phi )\) is proportional to the bolometric flux at a given phase, and normalized by \(l_{\rm obs~max}=1\). For weak pulsars, the 'theory' gives the following: - Photon cutoff energy is \[\label{cut} E_{\rm cut~th}=e(\theta, \chi)L_{34}^{3/8}P_{\rm ms}^{-1/4}{\rm GeV}.\] Here \(\theta\) is the spin-dipole angle, \(\chi\) is the observer angle (the angle between the spin axis and the direction to observer), \(L_{34}\) is the spin-down power in units of \(10^{34}\)erg/s, \(P_{\rm ms}\) is the period in ms. The dimensionless function \(e(\theta, \chi)\) is currently known to some 10% accuracy for an axisymmetric pulsar (\(e(0, \chi)\) drops from about \(5\) at \(\chi \approx 90^\circ\) to about \(2.5\) at \(\chi \approx 65^\circ\)) and to yet unclear accuracy for generic \(\theta\). - The normalized bolometric lightcurve is \[\label{curve} l_{\rm th}(\phi )=l_{\rm th}(\phi; \theta ,\chi).\] The accuracy of \(l_{\rm th}(\phi; \theta ,\chi)\) is yet unclear. - Bolometric efficiency \[\label{eff} \epsilon =\epsilon(\theta ,\chi),\] defined as the ratio of the pulsed bolometric luminosity (as seen at observation angle \(\chi\)) to the spin-down power. The dimensionless function \(\epsilon(\theta ,\chi)\) is currently known to some 10% relative accuracy for an axisymmetric pulsar (\(\epsilon (0, \chi)\) drops from about \(10\) at \(\chi \approx 90^\circ\) to about \(1\) at \(\chi \approx 85^\circ\), to about \(0.1\) at \(\chi \approx 65^\circ\)) and to yet unclear accuracy for generic \(\theta\). The measurement procedure is then straightforward: - Use Eq.([\[curve\]](#curve){reference-type="ref" reference="curve"}) to fit the lightcurve, thereby measuring both \(\theta\) and \(\chi\). The lightcurves (both observational and theoretical) are rich enough (at least in some cases with many local maxima, etc.), and this may work. - Use Eq.([\[cut\]](#cut){reference-type="ref" reference="cut"}) to measure the spin-down power \(L_{34}\), and then, knowing \(P\) and \(\dot{P}\), deduce the moment of inertia of the pulsar. - Use Eq.([\[eff\]](#eff){reference-type="ref" reference="eff"}) to calculate the bolometric luminosity, and then, knowing the bolometric flux \(f\), deduce the distance to the pulsar. # Pulsar Theory {#intro} A first-principle, i.e., using no arbitrary parameters, computation of pulsar spectra and lightcurves has been presented in. The theoretical results are supposedly exact, or close to exact, at least in principle (although the current numerical accuracy is poor). Only weak pulsars have been treated. The Fermi pulsar catalog supposedly contains many weak pulsars; and perhaps most Fermi pulsars can be usefully approximated as weak (for strong pulsars, the averaged efficiency must be much smaller than the Fermi's median value of about 15%). The non-weak pulsar problem (significant, as compared to Goldreich-Julian per rotation, pair production near the light cylinder) does not seem to be insurmountably more difficult either, and is expected to be solved in the near future. Once the pulsar theory delivers, the procedure outlined in §[1](#distances){reference-type="ref" reference="distances"} will become feasible. A logical questions is why don't we do it here. We are currently computing a library of magnetospheres with different spin-dipole angles \(\theta\), and we will attempt the lightcurve fitting. The results will be published, regardless of whether we fail or succeed. (It is already clear from that we cannot fail too miserably.) We want to stress that the core of the theory, Aristotelian Electrodynamics (AE, §[3](#arist){reference-type="ref" reference="arist"}), appears, by virtue of near-triviality, to be unassailable; our potential failure can only come from bad numerics and/or failure of the calculation recipe and/or non-weak pulsar effects. These problems, if they indeed occur, should be temporary. It seems very likely that AE is capable of fully solving the pulsar (only at high energies, of course). The purpose of this note is to invite computational effort of other researchers. # Aristotelian Electrodynamics {#arist} AE (numerical) calculation of the pulsar gives the electromagnetic field and positron and electron densities everywhere in the magnetosphere. The electromagnetic field is computed, starting from zero, by Maxwell equations. To solve Maxwell equations, one needs to know the electric current. The electric current inside the star is given by the standard Ohm's law (plus permanent current responsible for the magnetization). The electric current outside the star is \[\label{cur} {\bf j}=\rho_+{\bf v}_+-\rho_-{\bf v}_-,\] where \(\rho_\pm\) and \({\bf v}_\pm\) are the (positron charge normalized) number densities and velocities of positrons and electrons. In AE, \[\label{ae} {\bf v}_{\pm}={{\bf E}\times {\bf B}\pm(B_0{\bf B}+E_0{\bf E})\over B^2+E_0^2}.\] Here the scalar \(E_0\) and the pseudoscalar \(B_0\) are the proper electric and magnetic fields defined by \[B_0^2-E_0^2=B^2-E^2,~ B_0E_0={\bf B}\cdot {\bf E},~ E_0\geq 0.\] Eq.([\[ae\]](#ae){reference-type="ref" reference="ae"}) must be valid simply because this is the only possible Lorentz covariant expression for the velocity in terms of the local electromagnetic field; and it is clear that, at least where they radiate, the charges move at near the speed of light, and, due to strong radiation overdamping, the charge velocity depends only on the local values of the electric and magnetic fields (see for further details). It remains to calculate the densities \(\rho_\pm\). This is almost straightforward, as we know how the charges move: \[\label{con} \dot{\rho_{\pm}}+\nabla \cdot (\rho_{\pm}{\bf v}_{\pm})=Q.\] The only subtlety is the pair production rate \(Q\). For weak pulsars, one postulates pair production near the star with an almost arbitrary prescription which keeps \(Q\) positive and large (again compared to Goldreich-Julian per rotation) so long as the proper electric field \(E_0\) does not drop well below its typical vacuum value. For non-weak pulsars, one needs to add pair production in the radiation zone. Here pairs are produced in photon collisions. The necessary calculation is also clear (see Conclusions of arXiv:1310.1894) and appears doable.

{'timestamp': '2013-10-14T02:11:21', 'yymm': '1310', 'arxiv_id': '1310.3261', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3261'}

# Generation of Hot Electrons ## Background The interaction between high intensity, short pulse lasers and solid targets has been a subject of significant research over the last 15 years. The reason for this interest is the experimental observation (and successful computer simulation) of the emission of "hot electrons", that is, electrons that have energies far in excess of that corresponding to emission from a thermal source. This hot electron production has become an important research topic not only because it represents a useful source of high energy electrons itself, but because it is the mechanism for many important subsequent processes, including X ray , \(\gamma\) ray , and positron production , and seeding even further processes such as ion acceleration . Consequently there have been numerous efforts to increase the number and energy of the hot electrons as well as to increase the efficiency of the conversion of laser energy into hot electron energy. Most of these studies have emphasized the role of the laser pulse energy, duration, and intensity ; in addition, there is a robust literature describing the effect on electron energy of a "pre-plasma" on the front of the target . ## Target Design This paper discusses another means of enhancing the energy and directionality of the laser generated hot electrons, quite apart from using higher energy, more intense lasers. We propose, and verify through 3D PIC simulation, target designs that include specific modifications of the front surface geometry that not only substantially increase the number of relativistic electrons from a target, but can greatly decrease the solid angle of emission, all while maintaining or improving the conversion efficiency of overall laser to electron energy. There is a substantial body of experimental work on the effect of surface roughness of targets on laser-plasma coupling, including the laser absorption and the production efficiency and energy spectrum of the generated x-rays or ions [.]{style="color: blue"} Various simulation works attribute the observed improvement to a surface area increase and the local field enhancement introduced by the roughness . Recently, several authors have shown that larger scale structures can give rise to enhanced production of hot electrons and/or high energy ions . Kluge and Gaillard shot microcone targets on the cone walls and observed a significant increase in electron and proton energies which they attributed to direct laser acceleration of electrons along the cone walls . Zheng et al. proposed and simulated a "slice cone" target attributing the accelerated electrons to a similar mechanism. 2D simulations on similar shaped nanobrush targets have also been published . Our target design employs a similar acceleration mechanism to that of the previous work on microcones, but changes the structures to periodic regular arrays. This change has the virtue of a target for which the analysis and PIC simulations are easier to understand. Additionally, this target configuration has the added feature of being easier to align, in principle being no more difficult than a standard flat foil. In the simulations described here we find that hot electrons start off being extracted from the structures in a manner not dissimilar to that reported by some previous 2D simulations . However, as we discuss in Sec. [[3](#sec:2Dvs3D){reference-type="ref" reference="sec:2Dvs3D"}]{style="color: blue"}, fully 3D simulations are required to reveal accurately the trajectories of hot electrons. ## General Characteristics We use 3D PIC simulations to study the laser absorption and electron spectrum changes due to front surface target structure. Fig. [[\[fig:intro_cartoon\]](#fig:intro_cartoon){reference-type="ref" reference="fig:intro_cartoon"}]{style="color: blue"} is a schematic showing a general overview of the results from a 3D PIC simulation of a specific geometric shape placed on the front of a target compared to a regular flat target with pre-plasma. We have chosen these shapes not only to facilitate the discussion of the electron acceleration, but to highlight the limitations of 2D PIC simulations as well. The colored cones on the backside of the target represent the angular distributions of electrons with different energies. Blue, green, yellow and red cones indicate electron energies from low to high. For a regular flat target with pre-plasma (shown at the bottom), electrons usually have divergence angles between \(30-60^{\circ}\) . For a target with tower structures on the front (shown at the top), there are considerably more higher energy electrons generated. The highest energy hot electrons form into two narrow cones lining in the y direction (red), whereas the laser is polarized along x. The inset on the top left shows a typical 2D simulation setup, where only x and z dimensions can be modeled. # Simulation Setup Our simulations use the 3D PIC code LSP . Three types of targets are studied, and are indicated by Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(a)]{style="color: blue"}, [(b)]{style="color: blue"} and [(c)]{style="color: blue"}. [\[sec:Our-simulations-use\]]{#sec:Our-simulations-use label="sec:Our-simulations-use"}Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(a)]{style="color: blue"} shows a normally flat target with an exponentially decaying pre-plasma on the front. The scale length of the pre-plasma is \(1\mu m\). The target in Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(b)]{style="color: blue"} is chosen to be an example of a target that apparently lends itself to a 2D simulation; we refer to this as a slab target. The slabs are \(10\mu m\) deep in the laser z direction, and \(1\mu m\) wide in the perpendicular laser polarization direction. The spacing between the slabs is \(2\mu m\). The third type is shown in[(c)]{style="color: blue"}, what we have labeled as a tower target. The towers are also \(10\mu m\) deep and \(1\mu m\) wide. Here again the transverse spacings are \(2\mu m\). Both structures in [(b)]{style="color: blue"} and [(c)]{style="color: blue"} have sharp interfaces. The base of each target is \(11\mu m\) on a side. The material is Al, initialized as singly ionized, but subsequently treated by ADK ionization model; collisions are not included. (2D simulations using Spitzer cross-sections with collision frequency capped to a maximum value of \(2\times10^{16} s^{-1}\) do not show significant changes in hot electron distributions either spectrally or angularly.) The simulation box is made of \(120\times120\times600\) cells with mesh sizes \(\Delta x=\Delta y=0.1\mu m\), \(\Delta z=0.05\mu m\). (We checked the appropriateness of these parameters using short simulations with \(120\times120\times1200\) cells and \(\Delta z=0.025\mu m\). We found little difference in either the energies or the trajectories of the hot electrons compared to the coarser grid.) The boundary conditions are absorbing, and the time step is 0.03fs. In the simulation, a Gaussian laser pulse is used with a wavelength of \(800nm\), \(15J\) energy, \(30fs\) FWHM pulse duration and a focal spot diameter of \(2.9\mu m\), yielding \(5\times10^{21}W/cm^{2}\) peak intensity. Electron energy and spatial distributions are measured at a plane \(5\mu m\) inside the target. In all of our simulations reported here, we do not take into account changes to the electron spectrum due to target charging when the hot electrons leave the target, but we note that charging will have a minimal effect on the simulation for the higher energy (\(>100\) MeV) hot electrons that are of interest here . Finally, we employed a direct-implicit advance with an energy conserving particle push which greatly reduces numerical heating. # 2D vs. 3D Simulations {#sec:2Dvs3D} While we have found it important to use full 3D PIC simulations in our study, the use of 3D simulations is not common in the literature of hot electron production. Largely because of the often prohibitively large computational demands of 3D PIC simulations, 2D PIC simulations of hot electron production have been the de facto standard. However, in our studies we observe multiple features of the hot electrons' behavior in 3D that simply cannot be addressed in 2D. For example, note that our slab (Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(b)]{style="color: blue"}) and tower structures (Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(c)]{style="color: blue"}) *cannot be differentiated* using a 2D simulation; while in fact, we find that using 3D simulations the electron acceleration dynamics are significantly different for the two geometries. These differences show up most clearly and significantly in the predicted angular spectrum. It is only with a fully 3D simulation that we discover the remarkably narrow angular divergence pattern discussed in [[4.2](#angular distribution){reference-type="ref" reference="angular distribution"}]{style="color: blue"}. The use of 2D results as a proxy to predict 3D electron energy spectra and angular distributions requires significant and in many cases suspect assumptions. Fundamentally, the conversion is only possible when the desired quantity depends only on aspects of the target-laser interaction such as material, degree of pre-pulse etc. that have no transverse asymmetry. For a linearly polarized laser pulse striking a plane, unstructured target, the low energy part of the 2D simulated energy spectrum of hot electrons can be generalized to 3D by appropriately symmetrizing the 2D spectrum through 360 degrees. This method can be useful for calculating quantities that are dominated by low energy electrons, such as \(K_{\alpha}\) radiation, but not simulating the high energy electrons in the kind of structured targets we discuss here. In this study our main interest is in the highest energy electrons. As we show later, these are produced by direct laser acceleration. The inset in the upper left of Fig. [[\[fig:intro_cartoon\]](#fig:intro_cartoon){reference-type="ref" reference="fig:intro_cartoon"}]{style="color: blue"} shows a setup for both towers and slabs if 2D simulations were to performed. Clearly this configuration fails the test for meaningful 2D simulation, for the target geometry is not cylindrically symmetric within the laser spot dimensions and the laser-plasma dynamics clearly do not depend only on local conditions. There are other, more subtle reasons why this specific problem cannot be treated in 2D: The background plasma density generated by the short pulse laser itself is higher in 2D than in 3D because of the translational invariance in the virtual-y dimension. As we discuss below, the maximum electron energy from DLA is quite sensitive to the background plasma density. In addition, because the transverse dimensions of our structures are on the order of the laser wavelength, they tend to act as waveguides which changes the laser phase velocity. The DLA mechanism is strongly affected by the phase velocity of the laser relative to the accelerating hot electron, so that only a 3D simulation can capture this physics (see [[5.1](#DLA){reference-type="ref" reference="DLA"}]{style="color: blue"}). # 3D Simulation Results ## Energy Spectrum The electron energy spectrum for the 3 types of targets is shown in Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}(d)]{style="color: blue"}. These simulations are for the laser conditions outlined in [[\[sec:Our-simulations-use\]](#sec:Our-simulations-use){reference-type="ref" reference="sec:Our-simulations-use"}]{style="color: blue"}. The blue curve is for the simple unstructured flat target. Pre-plasma, modeled as an exponential with a scale length of \(1\mu m\), is included in all of our unstructured flat target simulations because most experimental situations involve generation of some pre-plasma and because it provides a more interesting comparison since pre-plasma increases the laser absorption and coupling to the target. We calculate a \(17.9\%\) conversion efficiency from laser energy to fast electron kinetic energy (electrons with energies \(>1MeV\)). For the slab type target [(b)]{style="color: blue"}, the conversion efficiency is enhanced to \(23.0\%\), with substantially more electrons generated above \(50MeV\). The tower target yields yet another, different spectrum. With a conversion efficiency of \(16.6\%\), comparable to that of the flat target with pre-plasma, the spectrum shows a large reduction in the yield of low energy electrons and a dramatic increase in the yield of high energy electrons. Specifically, the conversion efficiencies for fast electrons with energies above \(50MeV\) are \(1.5\%\), \(5.2\%\) and \(7.9\%\) for the flat, slab and tower targets respectively. This is a significant re-shaping of the energy spectrum with respect to that from a flat target such that the low energy portion of the electron spectra is shifted to higher energies. These target geometries substantially increase the number of electrons with the highest energies, while maintaining the overall efficiency of coupling. This result constitutes one of the two primary findings of our work. The other concerns the electron angular distribution. ## Angular Distribution {#angular distribution} The simulations show a striking modification of the angular distribution by our targets, as is shown in Fig. [[\[fig:ang_distrib_3D\]](#fig:ang_distrib_3D){reference-type="ref" reference="fig:ang_distrib_3D"}]{style="color: blue"}. The top row is electron number distribution as a function of kinetic energy and angle. The angle indicates the direction of electron velocities in a solid angle, \(\Delta\Omega=2\pi sin\theta\Delta\theta\), where \(\theta=tan^{-1}(\frac{\sqrt{p_{x}^{2}+p_{y}^{2}}}{p_{z}})\). The bottom row shows the 2D projected divergence map of fast electrons (\(>1MeV\)). The three graphs are the fast electron number distributions as a function of \(\theta_{x}\) and \(\theta_{y}\), where \(\theta_{x}=\pm tan^{-1}(\frac{p_{x}}{p_{z}})\), \(\theta_{y}=\pm tan^{-1}(\frac{p_{y}}{p_{z}})\). Subplots [(a1,a2)]{style="color: blue"}, [(b1,b2)]{style="color: blue"} and [(c1,c2)]{style="color: blue"} correspond to targets [(a)]{style="color: blue"}, [(b)]{style="color: blue"} and [(c)]{style="color: blue"} in Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}]{style="color: blue"} respectively. Comparing [(a1)]{style="color: blue"}, [(b1)]{style="color: blue"} and [(c1)]{style="color: blue"}, in general the electron divergence reduces with increasing energy. However, for the flat target, the decrease is small compared to the two types of structured targets. While some collimation is seen in the high energy portion of the electrons generated by the slab target [(b1)]{style="color: blue"}, the tower target [(c1)]{style="color: blue"} shows a significant improvement in the collimation. This result is clearer in the bottom graphs of Fig. [[\[fig:ang_distrib_3D\]](#fig:ang_distrib_3D){reference-type="ref" reference="fig:ang_distrib_3D"}]{style="color: blue"}. The flat target [(a1)]{style="color: blue"} shows a cylindrically symmetric angular distribution, with a large divergence angle of about \(60^{\circ}\). Both the shape of slab and tower breaks the rotational symmetry, so that, as expected, the corresponding angular distributions do not maintain such symmetry. For the slab target [(b1)]{style="color: blue"}, the distribution is wider in the x direction than in the y direction. For the tower target [(c1)]{style="color: blue"}, although the target shape itself has \(90^{\circ}\) rotational symmetry, the angular distribution does not. The distribution shows 2 peaks centered at \(\theta_{y}\approx\pm4-5^{\circ}\), each peak is about \(4-5^{\circ}\) FWHM. Remarkably, in the case of the tower target, the highest energy electrons tend to preferentially fall into these 2 small cones. For electrons \(>100MeV\), approximately \(30\%\) of the electrons (or \(5.7\times10^{9}\) in number) fall into the two \(5^{\circ}\) full angle cones. We find that for the tower targets the fast electron bunch has a pulse duration of \(42fs\). More importantly, for electrons \(>100MeV\), the pulse is even shorter, about \(13fs\) leading to an average current of \(70kA\). Assuming a source diameter of approximately \(3\mu m\), this electron source brightness at \(100MeV\) is on the order of \(10^{23}s^{-1}mm^{-2}mrad^{-2}(0.1\%\ bandwidth)\). # Discussion We attribute the enhancement in the high energy electron spectrum to direct laser acceleration (DLA) . Electrons undergoing DLA gain energy directly from the laser fields. This mechanism requires the accelerated electrons to be injected into the pulse at the right phase, and then to travel along with the pulse over a significant length, and then be injected into the target at the critical surface. ## Direct Laser Acceleration {#DLA} In DLA the highest energies are achieved by electrons which are optimally injected: injected at a point as far from the critical surface as allowed by the laser and evolving plasma and at an ideal phase of the laser. By its nature, DLA is exquisitely sensitive to the initial velocity and position of the accelerated electron relative to the phase of the laser electric field. If a low energy electron is injected into a weak plane wave at a node of the electric field, the electron will gain energy in the first half of the laser cycle and then return it to the pulse during the second half of the cycle. Net energy gain is possible if the electron can remain in the accelerating half-cycle of the laser field long enough and, crucially, can subsequently escape before experiencing the decelerating half-cycle. When the laser intensity is well into the relativistic regime, the Lorentz force accelerates the electron's longitudinal velocity to roughly the speed of light within the first quarter cycle of the wave. Provided the phase velocity of the laser pulse is close to \(c\) as well, the appropriately phased moving electron will experience a positive acceleration from the wave over an extended distance and will gain substantial energy. For example, an electron starting at rest introduced into a node of a plane wave in vacuum with an intensity of \(5\times10^{21}W/cm^{2}\) will accelerate for a distance of \(272\mu m\), gaining a peak energy of \(1.14GeV\) before it begins to decelerate. The purpose of the structure on the front surface of the target is to provide a means for inserting the electrons into the plane wave (see below) as well as permit an extended acceleration length followed by an abrupt exit (at the relativistic critical surface). It is important to note that the vast majority of the electrons, even in the structured targets we propose, do not fulfill the requirement of DLA . This is consistent with the observation that only a small proportion of the laser-plasma electrons are observed to have high energy . In a plasma the phase velocity of the laser pulse depends upon the plasma density and is given by \(v_{ph}=c/\sqrt{1-(\frac{\omega}{\omega{}_{p}})^{2}}\), where \(\omega_{p}^{2}=\frac{n_{e}e^{2}}{m\epsilon_{0}}\); thus the de-phasing length for an electron undergoing DLA is also sensitive to plasma density. As an example, for an intensity of \(5\times10^{21}W/cm^{2}\), if \(n_{e}=0.02n_{cr}\), where \(n_{cr}\) is the critical density (\(=\frac{m\epsilon_{o}\omega^{2}}{e^{2}})\), the acceleration length is about \(25\mu m\) and the maximum energy is \(170MeV\); however, increasing the density by a factor of 5, the acceleration length reduces to \(7\mu m\) with a maximum energy of \(70MeV\). When we compare the 2D simulation results for the slabs with 3D simulations for both the slabs and the towers, we find the electron cut-off energy for 3D towers to be much higher. This is because the 3D simulation of the towers has the most vacuum space thus lowest background electron density. With lower background density the acceleration length is longer. As discussed in [[3](#sec:2Dvs3D){reference-type="ref" reference="sec:2Dvs3D"}]{style="color: blue"}, this is one reason why 3D simulations are required to accurately predict the production mechanism of hot electrons on non-flat targets. Although 3D slabs have a lower background density than 2D slabs, the cut-off energy is lower due to a different effect that we discuss in [[5.4](#3D_confinement){reference-type="ref" reference="3D_confinement"}]{style="color: blue"}. ## Electron Injection While the DLA mechanism makes it possible in principle to accelerate electrons to significant energies, the practical issue that arises is the need for placing, or injecting, the electrons into the laser field at the proper position and time relative to the oscillating laser fields. A successful injection technique must fulfill four conditions: 1) A significant number of electrons need to be injected; 2) The position where the electrons are injected needs to be far enough from (relativistic) critical density so that there is a sufficient acceleration length; 3) The point of injection should be at a position where the laser intensity is high; 4) Over the course of the acceleration length there must be a channel with a relatively low electron density so the laser can propagate with a phase velocity on the order of, but not significantly larger than, \(c\). With these conditions in mind, we now address DLA using Fig. [[\[fig:physics_2D\]](#fig:physics_2D){reference-type="ref" reference="fig:physics_2D"}]{style="color: blue"}[(a)]{style="color: blue"}, which shows the electron density from a 2D simulation. (Below we compare this 2D simulation to 3D next where we find that the injection process is similar but the electron confinement to be very different). The laser is incident from the left and the plot is at a time when the laser pulse is roughly half way down the length of the structures. The target and laser parameters are the same as in the 3D simulations: the transverse gap size is \(2\mu m\), the structures are \(1\mu m\) in width, and \(10\mu m\) in depth. The electrons are pulled out of the structures by the laser E field, as is indicated by the bunches along the structure surfaces. There are several noteworthy characteristic features: The bunches are located in such a way that they are separated by one laser wavelength on one side and are \(\pi\)-phase shifted on opposite facing sides, corresponding to regions where \(\vec{E}\cdot\hat{n}<0\) (\(\hat{n}\) is the surface normal). At a given time, the bunching maximizes at the point of the electric field anti-nodes. For this intensity, \(5\times10^{21}W/cm^{2}\), each bunch has an electron density of more than \(10^{22}/cm^{3}\).The electron density in the center of the gaps is considerably lower such that the pulse has a phase velocity approximately equal to the speed of light. These bunches constitute the injected electrons into the DLA acceleration channel: the final energy of any bunch depends upon when it is formed relative to the laser pulse and the position of its original location. The most energetic electrons originate from the tip of the structures, as long as the structure is shorter than the de-phasing length. We have found through trial and error that the maximum number of the highest energy electrons occurs when the spacing between the structures is slightly smaller than the laser focal spot, and the laser intensity at the structure surface is sufficiently large to ionize electrons while simultaneously not increasing the electron density in the acceleration channel to the point that the acceleration lengths are reduced due to increased phase velocity of the laser pulse. A full optimization of our structures would require numerous 3D simulations, and these have not been done here. In fact we know that the target size parameters are certainly *not* optimum: in order to keep the grid size small, and the 3D simulation manageable, we chose a small focal spot so we could have a small spacing between the structures. However, the problem is not scalable: \(2\mu m\) spacing between the structures results in some wave-guiding of the \(0.8\mu m\) wavelength laser (with its attendant modification of the phase velocity) and the background density between the structures fills too rapidly to preserve the acceleration length. ## Electron Confinement in 2D {#subsec:e_confine} In our 2D simulations we observe that electrons are guided by the structure shape. This observation is consistent with other recent 2D simulations . The responsible physics based upon the 2D simulation is diagrammed in Fig. [[\[fig:physics_2D\]](#fig:physics_2D){reference-type="ref" reference="fig:physics_2D"}(b)]{style="color: blue"}. Once electrons are pulled out into the pulse, some of them with the right initial conditions accelerate with the laser pulse. They form a current density \(j_{hot}\) near the slab surface, leaving positive charges at the surface of the structures. These positive charges draw return currents just inside the structure with a current density of \(j_{return}\) in the opposite direction of \(j_{hot}\). Now these two surface currents are uniform in the virtual-y direction in a 2D xz coordinate system, giving rise to a \(B_{y}\) field also uniform in the y direction. At the same time, the charge separation induces an electric field pointing away from the structure surface. These two fields, \(B_{y}\) and \(E_{x}\), can be quite large; in our simulations, they are approximately 0.3 of the peak laser B and E fields. When the electron velocity in the z direction is approximately the speed of light the Coulomb force from \(E_{x}\) and the magnetic force from \(cB_{y}\) are equal and opposite along the surface of the slab, resulting in electrons being guided forward along the structure. Because the slab targets are invariant in y near the central region where the laser is incident, one might expect them to be essentially well modeled by a 2D simulation and it is tempting to accept the physical picture of electron trapping in the slabs presented above. But, as we discussed earlier, the simulation problem is in fact, not 2D because the laser fields provide forces on the electrons that have an essential 3D element, specifically forces in the y dimension. Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"} (a)]{style="color: blue"} shows the trajectories of the high energy hot electrons from the 3D simulation; in contrast to the 2D simulation, they do not propagate along the structure surfaces. ## Electron Confinement in 3D {#3D_confinement} Fig. [[\[fig:physics_fin\]](#fig:physics_fin){reference-type="ref" reference="fig:physics_fin"}(a)]{style="color: blue"} indicates the responsible physics: After the electrons are pulled out into the gap, in addition to a \(B_{y}\) field, there is also a \(B_{x}\) field around the hot electron current. This field pinches the hot electron current in the y direction, leading to an increase in the current density \(j_{hot}\) as it propagates. A plot of the B field in the simulation is shown in Fig. [[\[fig:physics_fin\]](#fig:physics_fin){reference-type="ref" reference="fig:physics_fin"}(b)]{style="color: blue"}. It is an xy section of \(B_{\perp}\), the B field normal to z. The color map shows the magnitude while the black arrows indicate the directions. While the return current remains the same as in the 2D simulation, here the \(B_{y}\) field stemming from both \(j_{hot}\) and \(j_{return}\) grows such that the Lorentz force exceeds the Coulomb force. Electrons are thus pulled away from the slab surface, and since the spacing between the slabs is much smaller than the length of the slabs, electrons can easily hit the slab on the other side and again bounce back into the gap. In general, the 3D simulation shows that the electrons are pinched in the y dimension, and bounce back and forth in the x dimension. For the electrons, each bounce off of a surface means a sudden reversal of sign in \(v_{x}\), and the DLA force, \(-ev_{x}B_{laser}\), becomes a decelerating force, effectively terminating the acceleration length and thus restraining the energy gain in the forward direction. This physics is responsible for the fact that the cutoff energy found in our 3D simulations of the slabs is consistently smaller than the cutoff energy found in our 2D simulations. In addition, these bounces also widen the angular distribution of the emitted electrons in the x direction while, in the y direction, the angular spread is determined by the pinching. This explains the asymmetry in Fig. [[\[fig:ang_distrib_3D\]](#fig:ang_distrib_3D){reference-type="ref" reference="fig:ang_distrib_3D"}(b2)]{style="color: blue"} Because the 3D simulation of the slab targets suggests that the acceleration length is constrained by the collisions with the walls, an obvious improvement is to reduce the probability of collision with the walls by replacing the slabs with towers (see Fig. [[\[fig:3struct_fE\]](#fig:3struct_fE){reference-type="ref" reference="fig:3struct_fE"}]{style="color: blue"}). A field analysis of the tower target structure is shown in Fig. [[\[fig:physics_spk\]](#fig:physics_spk){reference-type="ref" reference="fig:physics_spk"}]{style="color: blue"}. The green arrows in Fig. [[\[fig:physics_spk\]](#fig:physics_spk){reference-type="ref" reference="fig:physics_spk"}(a)]{style="color: blue"} are a cartoon of the B field directions in the xy plane consistent with the directions of \(j_{hot}\) and \(j_{return}\). Fig. [[\[fig:physics_spk\]](#fig:physics_spk){reference-type="ref" reference="fig:physics_spk"}(b)]{style="color: blue"} is an xy field map taken from the simulation indicating \(B_{\perp}\) at the plane where the laser peak is located. Electrons start off by being pulled out by the laser E field in the x direction, where they spill into the gaps in the XZ plane between the towers. Inside these gaps, because the E field points mainly in the charge separation direction x, its force on the electrons can balance that due to \(cB_{y}\) when the electrons are along the tower surfaces. However, the \(cB_{x}\) component of the Lorentz force causes the electrons to move in the y direction towards y=0. The general trend of the high energy fast electron trajectories is shown by the black curves in Fig. [[\[fig:physics_spk\]](#fig:physics_spk){reference-type="ref" reference="fig:physics_spk"}(a)]{style="color: blue"}. Electrons are mostly confined in the yz plane (red plane in Fig. [[\[fig:physics_spk\]](#fig:physics_spk){reference-type="ref" reference="fig:physics_spk"}(a)]{style="color: blue"}) close to the towers. They are bent towards the center in the y direction, where the \(B_{y}\) field is small and plays no important role in the electron trajectories. This explains the two peaks along the y direction in the angular distribution map Fig. [[\[fig:ang_distrib_3D\]](#fig:ang_distrib_3D){reference-type="ref" reference="fig:ang_distrib_3D"}(c2)]{style="color: blue"}. Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"}]{style="color: blue"} shows the trajectories of a sampling of electrons above \(120MeV\) in simulations for both slab and tower targets. Electrons originating from \(X>0\) (\(X<0\)) are labeled red (blue). For the slab-structured target, it is clear that electrons bounce between the slab walls looking down the y axis (Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"}(a2)]{style="color: blue"}), and pinch looking down the x axis (Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"}(a3)]{style="color: blue"}). For the tower-structured target, the confinement is in the XZ plane (Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"}(b2)]{style="color: blue"}) and the pinching is in the YZ plane (Fig. [[\[fig:track_GRT_SPK\]](#fig:track_GRT_SPK){reference-type="ref" reference="fig:track_GRT_SPK"}(b3)]{style="color: blue"}). # Conclusion We have proposed front surface target structures (towers and slabs) for use in the generation of high energy, collimated electrons. We have shown that these targets cannot be adequately studied in 2D PIC simulations, but require 3D. This will generally be the case for structured targets. Compared to regular flat targets with \(1\mu m\) pre-plasma, the yield of electrons at the high energy end can be improved by several orders of magnitude, while the FWHM divergence angle of the most energetic hot electrons can be greatly reduced to \(<\)`<!-- -->`{=html}5\(^{\circ}\). For electrons above \(100MeV\), the average current can be as high as \(70kA\). For high energy electrons, e.g. electrons at \(100MeV\), the brightness is on the order of \(10^{23}s^{-1}mm^{-2}mrad^{-2}(0.1\%\ bandwidth)\). The conversion efficiency from laser to fast electrons using towers is comparable to that using a flat target with pre-plasma, while using slabs the conversion efficiency is improved by \(28\%\). For electrons with energies above \(50MeV\), the conversion efficiencies are \(1.5\%\), \(5.2\%\) and \(7.9\%\) for the three targets respectively. This makes front surface targets ideal for applications requiring collimated high-energy electrons. We have shown that the hot electrons generated using our structured targets are accelerated through the direct laser acceleration mechanism (DLA) and are guided by surface fields. Since the guiding effect is a sensitive function of the details of the target shape, a precise description of the target shape in 3D is required not only to determine the correct guiding fields, but is also important in terms of getting correct background plasma density and wave guiding effects, both of which can drastically change the electron energy.

{'timestamp': '2013-10-15T02:00:27', 'yymm': '1310', 'arxiv_id': '1310.3283', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3283'}

# Introduction

{'timestamp': '2013-10-24T02:00:40', 'yymm': '1310', 'arxiv_id': '1310.3258', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3258'}

# Introduction In recent years there has been a strong interest in identifying electromagnetic (EM) counterparts to loud gravitational wave (GW) events. Apart from the intrinsic information that EM waves carry about the source, EM signals will also help localize the source on the sky. Knowledge of the precise location of the source on the sky eliminates degeneracies and results in improved parameter estimation from GWs . In addition to being strong sources of GWs, compact binaries, such as binary black hole--neutron stars (BHNSs), and binary neutron star--neutron stars (NSNSs) are also promising sources of "precursor" and "aftermath" EM signals. Here, precursor (aftermath) means before (after) merger has taken place. For example, BHNS or NSNS mergers may provide the central engine that powers a short-hard gamma-ray burst. Moreover, during merger neutron-rich matter can be ejected that can shine as a "kilonova" due to the decay of r-process elements. While some studies have been performed in Newtonian gravitation or the conformal flatness approximation of general relativity (GR), only a fully GR calculation can reliably determine the amount of ejected mass and its distribution, as well as the GW signature. Equation of state effects, mass ejection, effects of cooling and finite temperature, as well as waveforms from the inspiral and merger of BHNSs and NSNSs, have been computed in full GR via hydrodynamic simulations (see e.g. for BHNSs and for NSNSs), and magnetohydrodynamic (MHD) simulations (see e.g. for BHNS mergers and for NSNSs). In all of these earlier simulations of magnetized neutron stars, the magnetic field was confined within the stellar interior, partly due to the inability of existing ideal MHD schemes to deal with magnetic fields exterior to the star where the matter magnetization can become very high. However, (spinning) neutron stars are believed to be endowed with dipole magnetic fields extending into the exterior, which comprises a force-free magnetosphere . Thus, toward the end of a BHNS or NSNS inspiral electromagnetic interactions can give rise to detectable EM pre-merger signals, e.g. either via establishing a unipolar inductor DC circuit, via magnetospheric interactions or via emission of magnetic dipole radiation . As these mechanisms operate in strongly curved, dynamical spacetimes, numerical relativity simulations are necessary to reliably determine the amount of EM output. Modelling these effects requires to first order either a GR resistive MHD computational scheme (e.g. ) or a scheme that matches the ideal MHD interior of the NS to the exterior force-free magnetosphere, such as those presented in. Simulations in GR attempting to model these effects are still in their infancy. Only recently have simulations begun to explore the viability of these mechanisms and calculate the total EM output (see for NSNSs and for BHNSs). In this paper we present the details and tests of our GR force-free electrodynamics formalism and new code, and our new scheme for matching ideal MHD to its force-free limit. This code has already been used and briefly described in. We demonstrate the robustness of our new force-free code in a series of 1D flat spacetime and 3D black-hole spacetime tests, and we test our new matching scheme by reproducing the force-free aligned rotator solution for a rotating magnetized star. The paper is structured as follows. In Sec. [2](#sec:conventions){reference-type="ref" reference="sec:conventions"} we discuss the general spacetime and EM field conventions. In Sec. [3](#sec:basic_eqns){reference-type="ref" reference="sec:basic_eqns"} we review the standard formulation of force-free electrodynamics, discuss some subtleties arising in so-called electrovacuum solutions, and derive for the first time some new identities emerging in this formulation. We also present the force-free formulation we adopt, and derive several new useful identities arising in this formulation. In Sec. [4](#sec:numerical){reference-type="ref" reference="sec:numerical"} we present our methods for numerically evolving the GR force-free electrodynamics (GRFFE) equations and matching them to ideal MHD stellar interiors. Sec. [5](#tests){reference-type="ref" reference="tests"} reviews the tests we adopt to demonstrate the robustness of our new code, as well as the results from our simulations. We conclude in Sec. [6](#sec:summaryandfuturework){reference-type="ref" reference="sec:summaryandfuturework"} with a summary and discussion of future work. # 3+1 Decomposition and General Conventions {#sec:conventions} In this section we describe the general conventions we use in our MHD/Force-Free formalism. Throughout we use geometrized units, setting \(c=1=G\). Latin indices denote spatial components (1--3) and Greek indices denote spacetime components (0--3). The signature of the spacetime metric is (-+++). ## 3+1 spacetime decomposition We use a 3+1 decomposition of spacetime in which the line element becomes (see e.g. ) \[ds^2 =-\alpha^2 dt^2 + \gamma_{ij} (dx^i + \beta^i dt) (dx^j + \beta^j dt),\] where \(\gamma_{ij}\) is the induced three-metric in 3D spatial hypersurfaces of constant time \(t\), \(\alpha\) is the lapse function and \(\beta^i\) the shift vector. The full (4D) spacetime metric \(g_{\mu \nu}\) is related to the three-metric \(\gamma_{\mu \nu}\) by \(\gamma_{\mu \nu} = g_{\mu \nu} + n_{\mu} n_{\nu}\), where is the future-directed, timelike unit vector normal to 3D spatial hypersurfaces. ## Maxwell 's equations and electromagnetic stress tensor The basic equations of ideal GRMHD and their implementation in a 3+1 spacetime decomposition has been treated in a number of papers (see e.g. ) and textbooks (e.g. ), but we review them here to set the stage for our applications below. The Faraday tensor \(F^{\mu \nu}\) can be decomposed into the 3+1 form \[F^{\mu \nu} = n^\mu E^\nu-n^\nu E^\mu-\epsilon^{\mu \nu \alpha \beta} B_\alpha n_\beta,\] where \(\epsilon^{\mu \nu \alpha \beta}\) is the Levi-Civita tensor. The electric and magnetic fields measured by normal observers are defined as \[\begin{aligned} E^\mu &=& n_\nu F^{\mu \nu} \label{def:Emu} \\ B^\mu &=& \frac{1}{2} \epsilon^{\mu \nu \alpha \beta} n_\nu F_{\beta \alpha} = n_\nu {}^*F^{\nu \mu}, \label{def:Bmu} \end{aligned}\] where \[{}^*F^{\mu \nu} = \frac{1}{2} \epsilon^{\mu \nu \alpha \beta} F_{\alpha \beta}\] is the dual of \(F^{\mu \nu}\). Note that \(n_\mu E^\mu = n_\mu B^\mu=0\). Hence both \(E^\mu\) and \(B^\mu\) are purely spatial. It is convenient to introduce the following variables: Here \(j^\mu\) is the 4-current density. With these new definitions, and \[{}^*\mathcal F^{\mu \nu} =-n^\mu \mathcal B^\nu + n^\nu \mathcal B^\mu-\mathcal E_\alpha n_\beta \epsilon^{\mu \nu \alpha \beta}. \label{dec:Fsab}\] Straightforward calculations yield \[\mathcal F^{\mu \nu} \mathcal F_{\mu \nu} = 2 (\mathcal B^2-\mathcal E^2) \ \ \ \mbox{and} \ \ \ {}^*\mathcal F^{\mu \nu} \mathcal F_{\mu \nu} = 4 \mathcal E_\mu \mathcal B^\mu, \label{eq:dotproducts}\] where \(\mathcal B^2 = \mathcal B_\mu \mathcal B^\mu = \mathcal B_i \mathcal B^i\) and \(\mathcal E^2 = \mathcal E_\mu \mathcal E^\mu = \mathcal E_i \mathcal E^i\). Maxwell's equations can be expressed in terms of the new variables as \[\nabla_\mu \mathcal F^{\mu \nu} =-\mathcal J^\nu \ \ \, \ \ \ \nabla_{[\alpha} \mathcal F_{\beta \gamma]} = 0. \label{eq:maxwell}\] It follows from the antisymmetric property of \(\mathcal F_{\mu \nu}\) that \(\nabla_{[\alpha} \mathcal F_{\beta \gamma]} = 0\) can be written as \[\nabla_\alpha \mathcal F_{\beta \gamma} + \nabla_\beta \mathcal F_{\gamma \alpha} + \nabla_\gamma \mathcal F_{\alpha \beta} = 0. \label{eq:maxwellabc}\] In addition, Hence \(\nabla_{[\alpha} \mathcal F_{\beta \gamma]} = 0\) is equivalent to The stress-energy tensor associated with the EM field is \[T_{\rm EM}^{\mu \nu} = \mathcal F^\mu{}_\lambda \mathcal F^{\nu \lambda}-\frac{1}{4} g^{\mu \nu} \mathcal F^{\lambda \sigma} \mathcal F_{\lambda \sigma}, \label{def:Tem}\] where \(g_{\mu \nu}\) is the spacetime metric. Straightforward calculation yields where \(\gamma_{\mu \nu} = g_{\mu \nu} + n_\mu n_\nu\) is the spatial metric on 3D hypersurfaces of constant time. It follows from Eq. ([\[def:Tem\]](#def:Tem){reference-type="ref" reference="def:Tem"}) and Maxwell's equations that \[\nabla_\nu T_{\rm EM}^{\mu \nu} =-\mathcal F^{\mu \nu} \mathcal J_\nu. \label{eq:divTem}\] The Poynting vector is defined as \[S^\mu =-n_\nu T_{\rm EM}^{\mu \nu} = \frac{\mathcal B^2+\mathcal E^2}{2} n^\mu-\epsilon^{\mu \nu \alpha \beta} n_\nu \mathcal E_\alpha \mathcal B_\beta. \label{def:Poynting}\] It follows that \[\mathcal B_\mu S^\mu = 0.\] In the flat spacetime limit (\(g_{\mu \nu} = \eta_{\mu \nu}\)) we obtain the familiar results \[S^0 = \frac{\mathcal B^2+\mathcal E^2}{2} \ \ \, \ \ \ S^i = \epsilon^{ijk} \mathcal E_j \mathcal B_k. \label{eq:Poynting_SR}\] ## Ideal MHD Condition {#sec:MHD} The ideal MHD condition is \[u_\mu \mathcal F^{\mu \nu} = 0, \label{mhd_cond}\] where \(u^\mu\) is a unit timelike vector (\(u_\mu u^\mu =-1\)) equal to the plasma 4-velocity in ideal MHD, and may be regarded as the plasma 4-velocity in the force-free limit. Contracting Eq. ([\[mhd_cond\]](#mhd_cond){reference-type="ref" reference="mhd_cond"}) with \(n_\nu\) and using \(\mathcal E^\mu = n_\nu \mathcal F^{\mu \nu}\) yields \(u_\mu \mathcal E^\mu = 0\). Comparing Eq. ([\[mhd_cond\]](#mhd_cond){reference-type="ref" reference="mhd_cond"}) with Eq. ([\[def:Emu\]](#def:Emu){reference-type="ref" reference="def:Emu"}), one may interpret the ideal MHD condition as the vanishing electric field measured by an observer with four-velocity \(u^\mu\). These observers include the one comoving with the plasma as well as others boosted with respect to this observer in a direction parallel to the B-field (i.e., \(u_\perp^\mu = u_{\perp,\rm comoving}^{\mu}\)). The magnetic field measured by such an observer is \[b^\mu = u_\nu {}^*\mathcal F^{\nu \mu}. \label{def:bmu}\] The Faraday and electromagnetic stress tensors can be decomposed by \(u^\mu\) and \(b^\mu\) by analogy to the decomposition with \(n^\mu, B^{\mu}, E^\mu\) presented in the previous section, i.e., \[\begin{aligned} \mathcal F^{\mu \nu} &=& \epsilon^{\mu \nu \alpha \beta} u_\alpha b_\beta \label{mhd:Fab} \\ {}^*\mathcal F^{\mu \nu} &=& b^\mu u^\nu-u^\mu b^\nu \label{mhd:Fsab} \\ \mathcal F^{\mu \nu} \mathcal F_{\mu \nu} &=& 2b^2 \label{mhd:FdotF} \\ {}^*\mathcal F^{\mu \nu} \mathcal F_{\mu \nu} &=& 0 \label{mhd:FdotFs} \\ T_{\rm EM}^{\mu \nu} &=& b^2 u^\mu u^\nu + \frac{b^2}{2} g^{\mu \nu}-b^\mu b^\nu. \label{mhd:Tem} \end{aligned}\] Eqs. ([\[mhd:FdotFs\]](#mhd:FdotFs){reference-type="ref" reference="mhd:FdotFs"}) and ([\[eq:dotproducts\]](#eq:dotproducts){reference-type="ref" reference="eq:dotproducts"}) yield \[\mathcal E_\mu \mathcal B^\mu = 0.\] and combining Eq. ([\[mhd:Fsab\]](#mhd:Fsab){reference-type="ref" reference="mhd:Fsab"}) with \(\mathcal B^\nu = n_\mu {}^*\mathcal F^{\mu \nu}\) yields \[\mathcal B^\nu = u^\nu n_\alpha b^\alpha-b^\nu n_\alpha u^\alpha.\] It is straightforward to show that \[b^\mu = \frac{P^\mu{}_\nu \mathcal B^\nu}{-n_\alpha u^\alpha} = \frac{P^\mu{}_\nu \mathcal B^\nu}{\gamma_v}, \label{eq:bmufromBmu}\] where \[P^\mu{}_\nu = \delta^\mu{}_\nu + u^\mu u_\nu.\] is the projection tensor, \(\gamma_v=-n_\alpha u^\alpha=\alpha u^0\) is the Lorentz factor corresponding to the relative velocity of \(u^\mu\) with respect to a normal observer \(n^\mu\). It follows from Eq. ([\[eq:bmufromBmu\]](#eq:bmufromBmu){reference-type="ref" reference="eq:bmufromBmu"}) that \[b^2 = b^\mu b_\mu = \frac{P_{\mu \nu} \mathcal B^\mu \mathcal B^\nu}{\gamma_v^2} = \frac{\mathcal B^2 + (u_\mu \mathcal B^\mu)^2}{\gamma_v^2}.\] Hence \(b^2\) is positive-definite, and \(b^2=0\) if and only if \(\mathcal B^\mu=0\), which also implies \(b^\mu=0\) and \(\mathcal F^{\mu \nu} =0\) from Eqs. ([\[eq:bmufromBmu\]](#eq:bmufromBmu){reference-type="ref" reference="eq:bmufromBmu"}) and ([\[mhd:Fab\]](#mhd:Fab){reference-type="ref" reference="mhd:Fab"}). By use of Eqs. ([\[mhd:FdotF\]](#mhd:FdotF){reference-type="ref" reference="mhd:FdotF"}), ([\[eq:dotproducts\]](#eq:dotproducts){reference-type="ref" reference="eq:dotproducts"}) and the condition \(b^2\geq 0\) we have \[\mathcal F^{\mu \nu} \mathcal F_{\mu \nu} \geq 0 \ \ \ \mbox{and} \ \ \ \mathcal B^2 \geq \mathcal E^2.\] The equality holds if and only if \(\mathcal F^{\mu \nu}=0\) or, equivalently, \(\mathcal B^\mu=\mathcal E^\mu=0\). Therefore, the ideal MHD condition forbids the (vacuum EM wave) solution \(B^2=E^2\) with \(B^2 > 0\). # Force-Free Electrodynamics (FFE) {#sec:basic_eqns} In this section we present the FFE conditions and briefly review the two most popular formulations of FFE. The first one uses the electric and magnetic fields as the fundamental dynamical variables, and the second one replaces the electric field by the Poynting vector. We include derivations of several key equations in order to clarify subtle points, correct typos in the literature, and to present the basis of our approach. ## FFE conditions {#sec:ffeaxi} The force-free conditions are \[\begin{aligned} \mathcal F^{\mu \nu} \mathcal J_\nu &=& 0, \label{ffe:FdotJ} \\ {}^*\mathcal F^{\mu \nu} \mathcal F_{\mu \nu} &=& 0, \label{ffe:FdotFs} \\ \mathcal F^{\mu \nu} \mathcal F_{\mu \nu} & > & 0. \label{ffe:FdotF} \end{aligned}\] The above conditions can be regarded as axioms of FFE (in addition to the Maxwell and Einstein equations). Physically, these conditions are expected to apply when the magnetic fields dominate over the inertia of the matter. In terms of the 3+1 variables, Eqs. ([\[ffe:FdotJ\]](#ffe:FdotJ){reference-type="ref" reference="ffe:FdotJ"})--([\[ffe:FdotF\]](#ffe:FdotF){reference-type="ref" reference="ffe:FdotF"}) become \[\begin{aligned} \rho \mathcal E^i + \epsilon^{ijk} J_j \mathcal B_k &=& 0, \label{ffe:FdotJ31} \\ \mathcal E_i \mathcal B^i &=& 0, \label{ffe:FdotFs31} \\ \mathcal B^2 & > & \mathcal E^2. \label{ffe:FdotF31} \end{aligned}\] These can be regarded as the FFE axioms in terms of \(\mathcal E\) and \(\mathcal B\) fields, where \(\epsilon^{ijk}=n_\mu \epsilon^{\mu ijk}\) is the Levi-Civita tensor associated with the spatial metric \(\gamma_{ij}\), and the 4-current density has been decomposed into the 3+1 form \[\mathcal J^\mu = \rho n^\mu + J^\mu\] with \[\rho =-n_\mu \mathcal J^\mu \ \ \, \ \ \ J^\mu = \gamma^\mu{}_\nu \mathcal J^\nu.\] Contracting Eq. ([\[ffe:FdotJ\]](#ffe:FdotJ){reference-type="ref" reference="ffe:FdotJ"}) with \(n_\mu\) and using \(\mathcal E^\mu = n_\nu \mathcal F^{\mu \nu}\) gives \[\mathcal J_\mu \mathcal E^\mu = \mathcal J_i \mathcal E^i = 0. \label{ffe:JdotE}\] The conditions ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) and ([\[ffe:FdotF\]](#ffe:FdotF){reference-type="ref" reference="ffe:FdotF"}) are properties of the ideal MHD condition, and as it was first shown in, the ideal MHD condition is contained in the force-free conditions. In particular, it can be shown that if the conditions ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) and ([\[ffe:FdotF\]](#ffe:FdotF){reference-type="ref" reference="ffe:FdotF"}) are satisfied, there exists a one-parameter family of timelike unit vectors \(\{U^\mu \}\) so that \(u_\nu \mathcal F^{\mu \nu}=0\) for any \(u^\mu \in \{U^\mu \}\). This one-parameter family of unit timelike vectors is given by \[u^\mu_L = \sqrt{\frac{\mathcal B^2}{\mathcal B^2(1-L^2)-\mathcal E^2} } \left( n^\mu-\frac{\epsilon^{\mu \beta \gamma \delta} n_\beta \mathcal E_\gamma \mathcal B_\delta}{\mathcal B^2} + L \frac{\mathcal B^\mu}{\mathcal B} \right) \label{eq:umuLtext}\] where the \(L\) parameter is restricted by \[|L| < \sqrt{ \frac{\mathcal B^2-\mathcal E^2}{\mathcal B^2}}. \label{eq:Lrestricttext}\] In Appendix [\[timelikevectors_exist\]](#timelikevectors_exist){reference-type="ref" reference="timelikevectors_exist"} we present a proof of Eqs. [\[eq:umuLtext\]](#eq:umuLtext){reference-type="eqref" reference="eq:umuLtext"}, [\[eq:Lrestricttext\]](#eq:Lrestricttext){reference-type="eqref" reference="eq:Lrestricttext"} using standard 3+1 notation. As was pointed out in in this family of unit timelike vectors, the one that has the *minimum* Lorentz factor is given by \(L=0\), i.e. \(u^\mu\) is orthogonal to \(\mathcal B^\mu\). The corresponding \(u^\mu\) is \[u_{(m)}^\mu = \sqrt{\frac{\mathcal B^2}{\mathcal B^2-\mathcal E^2} } \left( n^\mu-\frac{\epsilon^{\mu \beta \gamma \delta} n_\beta \mathcal E_\gamma \mathcal B_\delta}{\mathcal B^2} \right), \label{eq:ummu}\] or \[\begin{aligned} u_{(m)}^0 &=& \frac{1}{\alpha} \sqrt{\frac{\mathcal B^2}{\mathcal B^2-\mathcal E^2} } \label{eq:um0} \\ v_{(m)}^i &=& \frac{u_{(m)}^i}{u_{(m)}^0} = \alpha \frac{\epsilon^{ijk} \mathcal E_j \mathcal B_k}{\mathcal B^2}-\beta^i = \alpha \frac{\gamma^{ij} S_j}{\mathcal B^2}-\beta^j. \label{eq:vmi} \end{aligned}\] In the flat spacetime limit, \(u_{(m)}^\mu\) reduces to \[u_{(m)}^0 = \gamma_v = \sqrt{\frac{\mathcal B^2}{\mathcal B^2-\mathcal E^2} } \ \ \, \ \ \ v^i_{(m)} = \frac{u_{(m)}^i}{u_{(m)}^0} = \frac{\epsilon^{ijk} \mathcal E_j \mathcal B_k}{\mathcal B^2}.\]\ The three-velocity \(v^i_{(m)}\) appearing in this last equation is also known as the drift velocity. Finally, by use of Eqs. ([\[ffe:FdotJ\]](#ffe:FdotJ){reference-type="ref" reference="ffe:FdotJ"}) and ([\[eq:divTem\]](#eq:divTem){reference-type="ref" reference="eq:divTem"}), we obtain \[\nabla_\nu T_{\rm EM}^{\mu \nu} = 0. \label{ffe:divTem}\] Hence, FFE can be regarded as a limiting case of the MHD in which the plasma has negligible inertia. It is this property that motivates our scheme for matching ideal MHD to its force-free limit, which we present in Sec. [4.2](#machingMHDFFE){reference-type="ref" reference="machingMHDFFE"}. ## On the \(\ve{\mathcal E_i \mathcal B^i=0}\) Condition {#sec:EBcondition} In some literature (e.g. ), it is claimed that Eq. ([\[ffe:FdotFs31\]](#ffe:FdotFs31){reference-type="ref" reference="ffe:FdotFs31"}) follows from Eq. ([\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="ref" reference="ffe:FdotJ31"}). We argue that this is not true. Taking a dot product of Eq. ([\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="ref" reference="ffe:FdotJ31"}) with \(\mathcal B^i\) gives \(\rho \mathcal E_i \mathcal B^i=0\), while taking the cross product of Eq. ([\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="ref" reference="ffe:FdotJ31"}) with \(\mathcal E^i\) and using Eq. [\[ffe:JdotE\]](#ffe:JdotE){reference-type="eqref" reference="ffe:JdotE"} gives \(J^k (\mathcal E_i \mathcal B^i)=0\). Hence, from a mathematical point of view \(\mathcal E_i \mathcal B^i=0\) follows only if \(\mathcal J^\mu \neq 0\). Hence, the condition \(\mathcal E_i B^i=0\) can be violated in regions where \(\mathcal J^\mu=0\), if one uses only Eqs. [\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="eqref" reference="ffe:FdotJ31"} and [\[ffe:FdotF31\]](#ffe:FdotF31){reference-type="eqref" reference="ffe:FdotF31"} as the FFE conditions. One simple example is the initial data \(\mathcal E^i = \mathcal E_0^i/\sqrt{\gamma}\,\) and \(\mathcal B^i=\mathcal B_0^i/\sqrt{\gamma}\,\) with \(\mathcal E_0^i\) and \(\mathcal B_0^i\) being constant vectors and \(\mathcal E_{0i} \mathcal B_0^i \neq 0\) and \(|\mathcal B_o^i| > |\mathcal E_0^i|\). Clearly the initial data satisfy the Maxwell constraints \(D_i \mathcal E^i = \rho\) and \(D_i \mathcal B^i =0\) for \(\mathcal J^\mu=0\), as well as the remaining force-free constraints [\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="eqref" reference="ffe:FdotJ31"} and [\[ffe:FdotF31\]](#ffe:FdotF31){reference-type="eqref" reference="ffe:FdotF31"}. Hence they are valid EM initial data but not valid force-free initial data. Moreover, Eq. ([\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="ref" reference="ffe:FdotJ31"}) holds while Eq. ([\[ffe:FdotFs31\]](#ffe:FdotFs31){reference-type="ref" reference="ffe:FdotFs31"}) does not. The situation \(\mathcal J^\mu=0\) and \(T^{\mu \nu}=T_{\rm EM}^{\mu \nu}\) everywhere in the spacetime is known as the *electrovacuum*. In the electrovacuum, both \(\mathcal E_i \mathcal B^i=0\) and \(\mathcal B^2 > \mathcal E^2\) conditions can be violated. Examples of electrovacuum solutions that are not force-free include the Kerr-Newmann black holes (\(\mathcal E^2>\mathcal B^2\) and \(\mathcal E_i \mathcal B^i \neq 0\)), and Wald's electrovacuum solution in rotating black holes (\(\mathcal E_i \mathcal B^i \neq 0\)) . One may therefore choose to replace the condition ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) by \(\mathcal J^\mu \neq 0\). However, doing this will exclude some of the electrovacuum solutions that are also force-free under the condition ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}). One example is Wald's electrovacuum solution in Schwarzschild spacetime, which has been used to test GRFFE codes (see  and Sec. [\[wald\]](#wald){reference-type="ref" reference="wald"} below) or even a nonrotating star with a dipole magnetic field. Therefore, we suggest that the condition ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) should be kept in favor of \(\mathcal J^\mu \neq 0\). One may also define FFE as a limiting case of ideal MHD, as was done in . In that case, the condition ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) is inherited from the ideal MHD conditions. The advantage of the axiomatic approach we adopt is the ability to formulate FFE without reference to the 4-velocity \(u^\mu\) \[see also in , where the ideal MHD condition \(u_\mu \mathcal F^{\mu \nu}=0\) is replaced by ([\[ffe:FdotFs\]](#ffe:FdotFs){reference-type="ref" reference="ffe:FdotFs"}) and ([\[ffe:FdotF\]](#ffe:FdotF){reference-type="ref" reference="ffe:FdotF"})\]. While it may come as a surprise that there exist electrovacuum solutions (no matter present) that are also FFE solutions (tenuous plasma present) this is not a contradiction. As a model of physical reality, force-free electrodynamics applies to cases where a highly-conducting tenuous plasma is involved. Hence, physically, force-free environments cannot be the same as an electrovacuum environment. However, mathematically, any electrovacuum solution satisfying the force-free conditions [\[ffe:FdotJ\]](#ffe:FdotJ){reference-type="eqref" reference="ffe:FdotJ"}-[\[ffe:FdotF\]](#ffe:FdotF){reference-type="eqref" reference="ffe:FdotF"}, will also be a force-free solution. For example an electrovacuum solution in which \(\mathcal E^i=0\) and \(\mathcal B^i \neq 0\), is simultaneously a force-free solution. ## Evolution Equations for \(\ve{\mathcal E}\) and \(\ve{\mathcal B}\) {#sec:formalism1} Perhaps the most popular formulation of FFE uses the \(\ve{\mathcal E}\) and \(\ve{\mathcal B}\) fields as dynamical variables. As shown in , without any assumption of MHD the general Maxwell equations ([\[eq:maxwell\]](#eq:maxwell){reference-type="ref" reference="eq:maxwell"}) can be brought into the 3+1 form: \[\begin{aligned} D_i \mathcal E^i & = & \rho \label{divE} \\ \partial_t \mathcal E^i & = & \epsilon^{ijk} D_j ( \alpha \mathcal B_k) -\alpha J^i + \alpha K \mathcal E^i + \mathcal L_{\bf \beta} \mathcal E^i \label{Edot} \\ D_i \mathcal B^i & = & 0 \label{divB} \\ \partial_t \mathcal B^i & = &-\epsilon^{ijk} D_j (\alpha \mathcal E_k) + \alpha K \mathcal B^i + \mathcal L_{\bf \beta} \mathcal B^i \label{Bdot}, \end{aligned}\] where \(D_i\) is the covariant derivative associated with the spatial metric \(\gamma_{ij}\), \(K=K^i{}_i\) is the trace of the extrinsic curvature, and \(\mathcal L_{\bf \beta}\) is the Lie derivative along the shift vector \(\beta^i\). The general set of Maxwell Eqs. [\[divE\]](#divE){reference-type="eqref" reference="divE"}-[\[Bdot\]](#Bdot){reference-type="eqref" reference="Bdot"}, coupled to the general fluid equations for the matter \[\(\nabla_{\mu} (T_{\rm matter}^{\mu\nu} + T_{\rm EM}^{\mu\nu}) = 0\)\], reduce to the equations of ideal MHD (e.g., Eqs. (5.168)-(5.175) in ) whenever collision timescales are sufficiently short for the plasma to behave as an isotropic fluid and the magnetic Reynolds number is sufficiently large that resistivity can be ignored. To apply Eqs. [\[divE\]](#divE){reference-type="eqref" reference="divE"}-[\[Bdot\]](#Bdot){reference-type="eqref" reference="Bdot"} for FFE, an expression for the 3-current density \(J^i\) is needed. It is useful to decompose \(J^i\) into a component perpendicular to \(\mathcal B^i\) and a component parallel to \(\mathcal B^i\): \[J^i = J_\perp^i + J_\parallel \frac{\mathcal B^i}{\mathcal B^2}, \label{eq:Jidecomp}\] with \[J_\parallel \equiv \mathcal B_i J^i = \mathcal B_\mu \mathcal J^\mu \ \ \ \mbox{and} \ \ \ J_\perp^i \equiv J^i-(\mathcal B_k J^k) \frac{\mathcal B^i}{\mathcal B^2}. \label{def:Jpara_Jperp}\] The perpendicular component (after contracting Eq. ([\[ffe:FdotJ31\]](#ffe:FdotJ31){reference-type="ref" reference="ffe:FdotJ31"}) with \(\epsilon_i{}^{lm} \mathcal B_l\) and taking the cross product with \(\mathcal B^i\)) is given by \[J_\perp^i = \rho \frac{\epsilon^{ijk} \mathcal E_j \mathcal B_k}{\mathcal B^2} = (D_m \mathcal E^m) \frac{\epsilon^{ijk} \mathcal E_j \mathcal B_k}{\mathcal B^2}. \label{eq:Jperpi}\] The parallel component can be determined by demanding that the evolution equations preserve the constraint \(C_{EB}=\mathcal E_i \mathcal B^i=0\), i.e. \(\partial_t C_{EB}=0\) (see e.g. ), eventually giving \[J_\parallel = \epsilon^{ijk} (\mathcal B_i D_j \mathcal B_k-\mathcal E_i D_j \mathcal E_k)-2 \mathcal E^i \mathcal B^j K_{ij}. \label{eq:Jpara}\] Combining the results yields Equation ([\[ffe:Ji\]](#ffe:Ji){reference-type="ref" reference="ffe:Ji"}) is known as the Ohm's law in dissipationless GRFFE. In the flat spacetime limit, it reduces to the well-know expression (see, e.g. ) \[\ve{J} = \frac{(\ve{\mathcal E}\times \ve{\mathcal B}) (\ve{\nabla}\cdot \ve{\mathcal E}) + \ve{\mathcal B} [ \ve{\mathcal B}\cdot (\ve{\nabla}\times \ve{\mathcal B})-\ve{\mathcal E} \cdot (\ve{\nabla} \times \ve{\mathcal E})]}{\mathcal B^2}.\] Note that since \(K_{ab} =-\nabla_{(a} n_{b)}-n_{(a} a_{b)}\) (\(a_b = n^c \nabla_c n_b\)), \(-2\mathcal E^i \mathcal B^j K_{ij} = \mathcal E^\mu \mathcal B^\nu (n_{\mu;\nu} + n_{\nu;\mu})\). Hence Eq. ([\[eq:Jpara\]](#eq:Jpara){reference-type="ref" reference="eq:Jpara"}) agrees with the expression in Eq. (25) of . However, the \(-2\mathcal E^i \mathcal B^j K_{ij}\) term is missing in Eq. (80) of . While the constraint \(C_{EB}\equiv \mathcal E_i \mathcal B^i =0\) is preserved due to the FFE Ohm's law ([\[eq:Jpara\]](#eq:Jpara){reference-type="ref" reference="eq:Jpara"}), the constraint \(C_{dB}\equiv D_i \mathcal B^i=0\) is preserved by Eq. ([\[Bdot\]](#Bdot){reference-type="ref" reference="Bdot"}). To see this, consider \[\partial_t (\sqrt{\gamma}\, C_{dB}) = \partial_t [ \partial_i (\sqrt{\gamma}\, \mathcal B^i) ] = \partial_i (\mathcal B^i \partial_t \sqrt{\gamma}\, + \sqrt{\gamma}\, \partial_t \mathcal B^i),\] where \(\gamma\) is the determinant of the spatial metric \(\gamma_{ij}\). It follows from the Arnowitt-Deser-Misner equations \[see e.g. Eq. (2.136) in \] that \[\partial_t \sqrt{\gamma}\, = \sqrt{\gamma}\, (-\alpha K + D_i \beta^i).\] Using the identities \[D_i [\epsilon^{ijk} D_j(\alpha E_k)] = \epsilon^{ijk} D_i D_j(\alpha E_k) = 0,\] and, for any spatial vector \(w^i\), \[D_{[i} D_{j]} w^i = {}^{(3)}R_{ij}{}^i{}_k w^k = {}^{(3)}R_{jk} w^k,\] where \({}^{(3)}R_{ijkl}\) is the Riemann tensor associated with \(\gamma_{ij}\), we find after some algebra \[\begin{aligned} \partial_t (\sqrt{\gamma}\, C_{dB}) &=& \partial_j (\sqrt{\gamma}\, C_{dB} \beta^j). \label{eq:CdBdot} \end{aligned}\] Equation ([\[eq:CdBdot\]](#eq:CdBdot){reference-type="ref" reference="eq:CdBdot"}) shows that if \(C_{dB}=0\) initially, the evolution equations preserve the constraint. Equations ([\[Edot\]](#Edot){reference-type="ref" reference="Edot"}) and ([\[Bdot\]](#Bdot){reference-type="ref" reference="Bdot"}), combined with Eq. ([\[ffe:Ji\]](#ffe:Ji){reference-type="ref" reference="ffe:Ji"}) are the evolution equations for \(\mathcal E^i\) and \(\mathcal B^i\), which are subject to the three constraints \(C_{dB}=D_i \mathcal B^i=0\), \(C_{EB}=\mathcal E_i \mathcal B^i=0\) and \(\mathcal B^2 > \mathcal E^2\). The first two constraints are preserved by the evolution equations but not the last one. Violation of \(\mathcal B^2 > \mathcal E^2\) indicates the breakdown of FFE, which typically occurs in current sheets. Mathematically, if violation occurs the initial value problem for Eqs. [\[Edot\]](#Edot){reference-type="eqref" reference="Edot"}, [\[Bdot\]](#Bdot){reference-type="eqref" reference="Bdot"} with [\[ffe:Ji\]](#ffe:Ji){reference-type="eqref" reference="ffe:Ji"} becomes ill-posed. Moreover, the constraint [\[divE\]](#divE){reference-type="eqref" reference="divE"} is automatically satisfied, if one uses it to compute the charge density. Finally, Eqs. [\[eq:maxwell\]](#eq:maxwell){reference-type="eqref" reference="eq:maxwell"} also imply charge conservation, as it is straightforward to see that which can be used as an evolution equation for the charge density. Perhaps the greatest advantage of using Eqs. [\[divE\]](#divE){reference-type="eqref" reference="divE"}-[\[Bdot\]](#Bdot){reference-type="eqref" reference="Bdot"}, is that they are general and can be used to find both force-free solutions \[as long as the current is given by Eq. [\[ffe:Ji\]](#ffe:Ji){reference-type="eqref" reference="ffe:Ji"}\], and electrovacuum solutions (as long as one sets \(\mathcal J^\mu = 0\)). The disadvantage is that this formulation is not straightforward to use in conjunction with the well-known constrained transport methods for preserving the Maxwell constraints see e.g.. Thus, common numerical implementations of this formulation usually resort to divergence cleaning methods to maintain the Maxwell constraints (see e.g. ). ## Evolution Equations for \(\ve{S_i}\) and \(\ve{\mathcal B^i}\) {#sec:formalism2} Instead of evolving \(\ve{\mathcal E}\) and \(\ve{\mathcal B}\), Refs.  and suggest the adoption of \(S_i\) and \(\mathcal B^i\) as dynamical variables. These are the fundamental variables we adopt. It follows from Eq. ([\[def:Poynting\]](#def:Poynting){reference-type="ref" reference="def:Poynting"}) that \(S_i = \epsilon_{ijk} \mathcal E^j \mathcal B^k\). Taking the cross product of \(S_i\) with \(\mathcal B^i\) and using \(\mathcal E_i \mathcal B^i=0\) gives \[\epsilon^{ijk} \mathcal B_j S_k = \epsilon^{ijk} \epsilon_{klm} \mathcal B_j \mathcal B^m \mathcal E^l =\mathcal B^2 \mathcal E^i.\] The condition \(\mathcal B^2 > \mathcal E^2\) guarantees that \(\mathcal B^2>0\). Hence \[\mathcal E^i = \frac{\epsilon^{ijk} \mathcal B_j S_k}{\mathcal B^2}. \label{eq:E-SB}\] The above equation can be rewritten using the identities \(S_\mu =-n_\nu T_{\rm EM}^{\nu}{}_\mu\) and \(\epsilon^{\mu \nu \alpha} = n_\beta \epsilon^{\beta \mu \nu \alpha}\) as \[\mathcal E^\alpha =-\frac{\epsilon^{\alpha \beta \gamma \delta} \mathcal B_\beta S_\gamma n_\delta}{\mathcal B^2} = \frac{\epsilon^{\alpha \beta \gamma \delta} \mathcal B_\beta T^{\mu}_{\rm EM}{}_\gamma n_\mu n_\delta}{\mathcal B^2}, \label{eq:E-SBn}\] (see also Eq. (10) in ). Note that the constraint \(C_{EB}=\mathcal B_i \mathcal E^i =0\) is automatically satisfied by Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}). Contracting Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}) with \(\mathcal E_i\) gives where \(C_{SB}\equiv \mathcal B^i S_i=0\) and In this formulation the condition \(\mathcal B^2 > \mathcal E^2\) is expressed through Eq. [\[E2S2B2\]](#E2S2B2){reference-type="eqref" reference="E2S2B2"} as \[\bar{S}^2 < \mathcal B^4. \label{eq:S2ltB4}\] If we define the densitized magnetic field \[\tilde{\mathcal B}^i \equiv \sqrt{\gamma}\, \mathcal B^i = \sqrt{\gamma}\, n_\nu {}^*\mathcal F^{\nu i} = \alpha \sqrt{\gamma}\, {}^*\mathcal F^{i0},\] the time component of the Maxwell Eq. [\[nablastarFmunu\]](#nablastarFmunu){reference-type="eqref" reference="nablastarFmunu"} yields the constraint equation \[\partial_i \tilde{\mathcal B}^i = 0, \label{eq:divBeq0}\] whereas the spatial components of Eq. [\[nablastarFmunu\]](#nablastarFmunu){reference-type="eqref" reference="nablastarFmunu"} give the equation \[\partial_t \tilde{\mathcal B}^i + \partial_j (\beta^i \tilde{\mathcal B}^j-\beta^j \tilde{\mathcal B}^i + \alpha \sqrt{\gamma}\, \epsilon^{ijk} \mathcal E_k) = 0,\] where Eqs. ([\[dec:Fsab\]](#dec:Fsab){reference-type="ref" reference="dec:Fsab"}) and [\[normalvector\]](#normalvector){reference-type="eqref" reference="normalvector"} have been used. Substituting \(\mathcal E_k\) using Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}) yields the induction equation \[\partial_t \tilde{\mathcal B}^i + \partial_j \left( \alpha S_k \frac{\tilde{\mathcal B}^i \gamma^{jk}-\tilde{\mathcal B}^j \gamma^{ik}}{\mathcal B^2} +\beta^i \tilde{\mathcal B}^j-\beta^j \tilde{\mathcal B}^i \right) = 0. \label{eq:induction3}\] Introducing a 3-vector \(v^i\) defined as \[v^i = \alpha \frac{\gamma^{ij}S_j}{\mathcal B^2}-\beta^i. \label{def:vi}\] Then the induction equation ([\[eq:induction3\]](#eq:induction3){reference-type="ref" reference="eq:induction3"}) takes the familiar form \[\partial_t \tilde{\mathcal B}^i + \partial_j (v^j \tilde{\mathcal B}^i-v^i \tilde{\mathcal B}^j) = 0. \label{eq:induction}\] The induction equation clearly preserves the constraint \(C_{dB}=0\): \[\partial_t (\sqrt{\gamma}\, C_{dB}) = \partial_t (\partial_i \tilde{\mathcal B}^i) =-\partial_i \partial_j (v^j \tilde{\mathcal B}^i-v^i \tilde{\mathcal B}^j) = 0.\] The evolution equation for \(S_i\) can be derived from Eq. ([\[ffe:divTem\]](#ffe:divTem){reference-type="ref" reference="ffe:divTem"}), which gives \(\nabla_{\nu} T_{\rm EM}^{\nu}{}_i=0\) or \[\partial_t \tilde{S}_i + \partial_j (\alpha \sqrt{\gamma}\, T_{\rm EM}^j{}_i ) = \frac{1}{2} \alpha \sqrt{\gamma}\, T_{\rm EM}^{\mu \nu} \partial_i g_{\mu \nu}, \label{eq:Sidot}\] where is the densitized spatial Poynting vector, and the EM stress-energy tensor can be expressed in terms of \(\mathcal B^i\) and \(S_i\) via Eqs. ([\[eq:Tem31\]](#eq:Tem31){reference-type="ref" reference="eq:Tem31"}), ([\[def:Poynting\]](#def:Poynting){reference-type="ref" reference="def:Poynting"}) and the first equality of Eq. ([\[eq:E-SBn\]](#eq:E-SBn){reference-type="ref" reference="eq:E-SBn"}). The quantities \(S_0\), \(S^\mu\), \(\mathcal B^0\) and \(\mathcal B_\mu\) can be expressed in terms of \(\mathcal B^i\) and \(S_i\) using Eq. ([\[def:Poynting\]](#def:Poynting){reference-type="ref" reference="def:Poynting"}) and \(n_\mu \mathcal B^\mu=0\) as \[\mathcal B^0 = 0 \ \, \ \ \mathcal B_0 = \gamma_{ij} \beta^i \mathcal B^j \ \, \ \ \mathcal B_i = \gamma_{ij} \mathcal B^j,\] and Note that the time component of Eq. ([\[ffe:divTem\]](#ffe:divTem){reference-type="ref" reference="ffe:divTem"}) also implies \(\nabla_{\nu} T_{\rm EM}^{\nu}{}_0=0\), which gives the energy equation \[\partial_t (\sqrt{\gamma}\, S_0) + \partial_j (\alpha \sqrt{\gamma}\, T_{\rm EM}^j{}_0 ) = \frac{1}{2} \alpha \sqrt{\gamma}\, T_{\rm EM}^{\mu \nu} \partial_t g_{\mu \nu}. \label{eq:S0dot}\] However, Eqs. ([\[eq:induction3\]](#eq:induction3){reference-type="ref" reference="eq:induction3"}) and ([\[eq:Sidot\]](#eq:Sidot){reference-type="ref" reference="eq:Sidot"}) already provide a complete system of evolution equations for \(S_i\) and \(B^i\), which can be used to calculate \(\mathcal E^i\) using Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}). Thus, the energy equation ([\[eq:S0dot\]](#eq:S0dot){reference-type="ref" reference="eq:S0dot"}) is either a constraint or redundant and it must be able to be derived from Eqs. ([\[eq:divBeq0\]](#eq:divBeq0){reference-type="ref" reference="eq:divBeq0"}), ([\[eq:induction3\]](#eq:induction3){reference-type="ref" reference="eq:induction3"}), ([\[eq:Sidot\]](#eq:Sidot){reference-type="ref" reference="eq:Sidot"}) and ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}). We show that the energy is indeed redundant and not a constraint in Appendix [\[energy_redundant\]](#energy_redundant){reference-type="ref" reference="energy_redundant"}. Finally, the Maxwell equation \(\nabla_\alpha \mathcal F^{\nu \alpha}=\mathcal J^\nu\) implies that one of the force-free conditions \(\mathcal F^{\mu \nu} \mathcal J_\nu=0\) is also enforced by the evolution equations and the constraint \(D_i \mathcal B^i=0\). The evolution equations ([\[eq:induction3\]](#eq:induction3){reference-type="ref" reference="eq:induction3"}) and ([\[eq:Sidot\]](#eq:Sidot){reference-type="ref" reference="eq:Sidot"}) consist of a system of 6 coupled partial differential equations for 6 variables \(\mathcal B^i\) and \(S_i\), which contain the same number of equations as Eqs. ([\[Edot\]](#Edot){reference-type="ref" reference="Edot"}) and ([\[Bdot\]](#Bdot){reference-type="ref" reference="Bdot"}) in § [3.3](#sec:formalism1){reference-type="ref" reference="sec:formalism1"}. The system of partial differential equations in § [3.3](#sec:formalism1){reference-type="ref" reference="sec:formalism1"} are subject to two constraints: \(C_{dB}=D_i \mathcal B^i=0\) and \(C_{EB}=\mathcal E_i \mathcal B^i=0\). In the present system, the constraint \(C_{dB}=0\) remains, but \(C_{EB}=0\) is automatically satisfied by Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}). This fact was also pointed out in, but, another constraint that arises in this formulation was ignored: A simple change of variables cannot change the number of constraints in a dynamical system. The constraint that replaces \(C_{EB}=0\) in the \(\ve{S}\)-\(\ve{\mathcal B}\) formulation of GRFFE is \(C_{SB}\equiv\mathcal B^i S_i=\mathcal B^\mu S_\mu=0\). It can be shown that the evolution equations ([\[eq:induction3\]](#eq:induction3){reference-type="ref" reference="eq:induction3"}) and ([\[eq:Sidot\]](#eq:Sidot){reference-type="ref" reference="eq:Sidot"}) preserve the constraint \(C_{SB}=0\) as long as both \(C_{dB}=0\) and \(C_{SB}=0\) initially (see Appendix [\[evolCsb\]](#evolCsb){reference-type="ref" reference="evolCsb"}). As the \(\ve{S}\)-\(\ve{\mathcal B}\) formulation is equivalent to the \(\ve{E}\)-\(\ve{\mathcal B}\) formulation of GRFFE, one can use Eq. ([\[ffe:Ji\]](#ffe:Ji){reference-type="ref" reference="ffe:Ji"}) to compute the 4-current density \(\mathcal J^\mu = \rho n^\nu + J^\mu\). It is possible to prove that in the \(\ve{S}\)-\(\ve{\mathcal B}\) formalism the FFE current density is given by the same equation as in the \(\ve{E}\)-\(\ve{\mathcal B}\) formalism (see Appendix [\[FFE_current\]](#FFE_current){reference-type="ref" reference="FFE_current"}). # Numerical Method {#sec:numerical} Here we summarize the formulation and numerical methods we use to solve the equations of GRFFE and our new scheme for matching the ideal MHD to its force-free limit. ## Evolution scheme for the GRFFE equations {#sec:GREFFEalgorithm} The greatest advantage of the \(\ve{S}\)-\(\ve{\mathcal B}\) formulation is that it is straightforward to implement numerically, if one has already developed a GRMHD code. There are at least two more reasons for adopting the \(\ve{S}\)-\(\ve{\mathcal B}\) formulation: a) the evolution equations for \(S_i\) and \(\mathcal B^i\) are basically the same as their MHD counterparts. This already hints that the same evolution equations can be used to match ideal MHD domains to force-free domains. b) The constrained-transport method can be used to enforce the \(D_i \mathcal B^i=0\) constraint as in the MHD case. The remaining constraint \(S_i \mathcal B^i=0\), which was ignored in, is algebraic and can be enforced by replacing \(S_i \rightarrow S_i-(S_j \mathcal B^j)\mathcal B^i/\mathcal B^2\) after each evolution timestep, i.e., in the same way the \(\mathcal E_i \mathcal B^i\) constraint is enforced in the \(\ve{\mathcal E}\)-\(\ve{\mathcal B}\) formulation see e.g.. See also for other alternatives for enforcing the \(\mathcal E_i \mathcal B^i\) constraint. So, to transform a GRMHD high-resolution shock capturing code to a force-free code all one has to do is to remove from the GRMHD code all terms related to the perfect fluid stress tensor (i.e. the matter is ignored), and add a new algorithm for the primitives recovery. To summarize, the complete set of evolution equations are the induction and momentum equations \[\begin{aligned} \partial_t \tilde{B}^i + \partial_j (v^j \tilde{B}^i-v^i \tilde{B}^j) &=& 0 \label{eq:inductionrecap} \\ \partial_t \tilde{S}_i + \partial_j (\alpha \sqrt{\gamma}\, T_{\rm EM}^j{}_i) &=& \frac{1}{2} \alpha \sqrt{\gamma}\, T_{\rm EM}^{\mu \nu} \partial_i g_{\mu \nu} \label{eq:dtSiforcefree}, \end{aligned}\] with \[v^i = 4\pi \alpha \frac{\bar{S}^i}{B^2}-\beta^i = 4\pi \alpha \frac{\gamma^{ij} \tilde{S}_j}{\sqrt{\gamma}\, B^2}-\beta^i, \label{eq:vifromcons}\] where \(\bar{S}^i = \gamma^{ij} S_j\). Note that the factor \(4\pi\) has reappeared since our GRMHD code uses \(B^i\) instead of \(\mathcal B^i\). The evolution equations can be made to look even more similar to the MHD equations by introducing the unit timelike 4-vector \(u^\mu\) as where \(\tilde{S}^2=\gamma^{ij} \tilde{S}_i \tilde{S}_j\). Note that this is exactly the same as \(u^\mu_{(m)}\) in Eq. ([\[eq:ummu\]](#eq:ummu){reference-type="ref" reference="eq:ummu"})-the unit timelike 4-vector that satisfies \(u_\mu F^{\mu \nu}=0\) with the minimum Lorentz factor \(\gamma_v =-n_\mu u^\mu = \sqrt{B^2/(B^2-E^2)}\). The EM stress-energy tensor is given by Eq. ([\[mhd:Tem\]](#mhd:Tem){reference-type="ref" reference="mhd:Tem"}) \[T_{\rm EM}^{\mu \nu} = b^2 u^\mu u^\nu + \frac{b^2}{2} g^{\mu \nu}-b^\mu b^\nu, \label{eq:Tembmu}\] where \(b^\mu\) can be computed from \(B^\mu\) and \(u^\mu\) using Eq. ([\[eq:bmufromBmu\]](#eq:bmufromBmu){reference-type="ref" reference="eq:bmufromBmu"}) \[b^\mu = \frac{P^\mu{}_\nu B^\nu}{\sqrt{4\pi} \gamma_v}. \label{eq:bmufromBmu2}\] Equations ([\[eq:inductionrecap\]](#eq:inductionrecap){reference-type="ref" reference="eq:inductionrecap"})--([\[eq:bmufromBmu2\]](#eq:bmufromBmu2){reference-type="ref" reference="eq:bmufromBmu2"}) give the complete evolution equations for \(B^i\) and \(S_i\). We embed this GRFFE formulation in the conservative ideal GRMHD, high-resolution shock capturing infrastructure we have presented and tested in , and in which we preserve the \(\partial_i \tilde B^i = 0\) constraint via a vector potential formulation which is equivalent to the standard staggered-mesh constrained-transport scheme in uniform-resolution grids . To close the system we choose the generalized Lorenz gauge we developed and used in . The evolution ("conservative") variables are \(\tilde{B}^i\) and \(\tilde{S}_i\). The "primitive" variables are \(B^i\) and \(v^i\), as in the MHD case. Reconstructions are done on the primitive variables. The inversion from conservative to primitive variables is trivial in GRFFE: \(B^i = \tilde{B}^i/\sqrt{\gamma}\,\) and \(v^i\) from Eq. ([\[eq:vifromcons\]](#eq:vifromcons){reference-type="ref" reference="eq:vifromcons"}). The electric field \(E^i\) is not needed for evolution but may be computed from Eq. ([\[eq:E-SB\]](#eq:E-SB){reference-type="ref" reference="eq:E-SB"}) \[E^i = 4\pi \frac{\epsilon^{ijk} B_j S_k}{B^2}.\] Inversion fails whenever the condition \(B^2 > E^2\) is violated as it leads to superluminal velocity (i.e. \(\gamma_v\) becomes purely imaginary). Thus, the condition for the primitive inversion to yield a physical solution is \[\tilde{S}^2 < \frac{\tilde{B}^4}{16\pi^2 \gamma}. \label{eq:stineq}\] It should be noted that the inequality ([\[eq:stineq\]](#eq:stineq){reference-type="ref" reference="eq:stineq"}) should be checked after removing the component of \(\tilde{S}\) along the magnetic field, i.e. imposing the constraint \(\tilde{B}^i \tilde{S}_i=0\) by the procedure \(\tilde{S}_i \rightarrow \tilde{S}_i-(\tilde{S}_j \tilde{B}^j) \tilde{B}^i/\tilde{B}^2\). It is straightforward to show that removing the \(\tilde{B}^i\) component from \(\tilde{S}_i\) always leads to smaller \(\tilde{S}^2\). The inequality ([\[eq:stineq\]](#eq:stineq){reference-type="ref" reference="eq:stineq"}) may be imposed by specifying a maximum Lorentz factor \(\gamma_{\rm max}\) and requiring that \(\gamma_v = \alpha u^0 \leq \gamma_{\rm max}\). It follows from Eq. ([\[eq:u0ui\]](#eq:u0ui){reference-type="ref" reference="eq:u0ui"}) that the condition \(\gamma_v \leq \gamma_{\rm max}\) is equivalent to \[\tilde{S}^2 \leq (1-\gamma_{\rm max}^{-2}) \frac{\tilde{B}^4}{16\pi^2 \gamma}.\] Define a factor \[f \equiv \sqrt{ (1-\gamma_{\rm max}^{-2}) \frac{\tilde{B}^4}{16\pi^2 \gamma \tilde{S}^2}}.\] The inequality ([\[eq:stineq\]](#eq:stineq){reference-type="ref" reference="eq:stineq"}) can be imposed by setting \[\tilde{S}_i \rightarrow \tilde{S}_i \min(1,f).\] Imposing the condition \(B^2>E^2\) when the FFE is supposed to break down (as in e.g. a current sheet) is effectively to add artificial dissipation to the fields and remove energy immediately to bring the fields back to the FFE regime. We typically set \(\gamma_{\rm max} = 2000\). In addition, as was proposed in in current sheets we null the inflow velocity normal to the current sheet, i.e., if \(\tilde n^i\) is the normal to the current sheet we set within an infinitesimal region above and below the current sheets covered by four zones. For a discussion motivating this approach and of its possible shortcomings we refer the interested reader to. ## Matching ideal MHD to its force-free limit {#machingMHDFFE} Force-free magnetospheres appear in many occasions in astrophysical environments, e.g., including neutron stars. The interior of a NS is highly conducting and the assumption of perfect conductivity is well-justified. As a result ideal MHD applies to the NS interior. However, existing high-resolution shock capturing MHD schemes cannot deal with high magnetizations and as a result they cannot typically deal with magnetic fields exterior to the highly conducting matter. On the other hand NSs are typically endowed with a force-free magnetosphere and since force-free electrodynamics can be regarded as the limit of ideal MHD in which the magnetic fields dominate the inertia of the matter, there must exist ways of making this transition from the ideal MHD interior to the force-free exterior. Such a scheme for matching ideal GRMHD to its force-free limit was first proposed in using the \(\ve{\mathcal E}\)-\(\ve{\mathcal B}\) formulation, but the implementation required the introduction of new variables and coding of additional evolution equations, as well as prescribing a penalty function based on the rest-mass density for transitioning from the interior to the exterior. Our scheme for matching ideal MHD interiors to force-free exteriors utilizes the fact that the magnetic field is frozen-in and is simply advected with the fluid for sufficiently weak magnetic fields. So, we propose that the frozen-in condition be enforced in the dense interior of the star and the surface values for the B-field and the Poynting vector then provide the boundary conditions for the exterior FFE evolution using the \(\ve{S}\)-\(\ve{\mathcal B}\) formalism we outlined in Sec. [3.4](#sec:formalism2){reference-type="ref" reference="sec:formalism2"}. We point out here that our matching scheme does not allow for any back-reaction of the exterior magnetic field onto the interior matter. This back-reaction potentially may become important in a thin layer near the surface of a star. However, resistive MHD studies of NSs, which include magnetic field back-reaction, indicate that neglecting it leads only to small errors. ### Matching when the fluid rest-mass density and four velocity are given {#easy_matching} First we will consider the case where we are evolving the EM field of a star with a well-defined surface, and that both the interior fluid four-velocity \(u_\mu\) and the rest-mass density distribution \(\rho_0\) are known and given for all times (e.g. a stationary rotating star with a weak interior field). Physically, the rest-mass density in a force-free magnetosphere *cannot* be zero. However, the equations of FFE *ignore* the existence of matter, and for *numerical* purposes we can safely set the rest-mass density exterior to the star equal to zero. Therefore, in our algorithm the stellar surface is defined as the 2-surface where the rest-mass density transitions from \(\rho_0^{(\rm num)}=\rho_0 \neq 0\) to \(\rho_0^{(\rm num)} = 0\), i.e., the *numerical* magnetosphere has zero density. - Interior to the star \(\rho_0^{(\rm num)} \neq 0\), and the frozen-in condition is enforced. We evolve the induction equation for the A-field with any convenient EM gauge choice to determine the scalar potential \(\Phi\), setting the three-velocity \(v^j\) equal to the given fluid velocity. In the continuum limit this truly enforces the frozen-in condition, while in the discrete limit small deviations from the frozen-in condition are expected. These converge away with increasing resolution. Given the A-field, we then determine the B-field, and compute the E-field using the ideal MHD condition [\[eq:udotF=0\]](#eq:udotF=0){reference-type="eqref" reference="eq:udotF=0"} where again the fluid 3-velocity is used. We set the interior \(\ve{\tilde S}^{\rm (in)}\) equal to where \(T_{\mu\nu} = T^{(\rm matter)}_{\mu\nu} + T^{(\rm EM)}_{\mu\nu}\) where \(T^{(\rm EM)}_{\mu\nu}\) is given by Eq. [\[eq:Tembmu\]](#eq:Tembmu){reference-type="eqref" reference="eq:Tembmu"}. Notice that as we approach the stellar surface the matter inertia contribution becomes subdominant: \(T_{\mu\nu} \approx T^{(\rm EM)}_{\mu\nu}\), and Eq. [\[Siinter\]](#Siinter){reference-type="eqref" reference="Siinter"} smoothly becomes the densitized Poynting vector \(\tilde S_i\) of Eq. [\[densSi\]](#densSi){reference-type="eqref" reference="densSi"} \[see also Eq. [\[def:Poynting\]](#def:Poynting){reference-type="eqref" reference="def:Poynting"}\]. This approach provides valid boundary conditions for \(\tilde S_i\) and \(\tilde B^i\) for the exterior force-free evolution. - Exterior to the star, \(\rho_0^{(\rm num)} = 0\), and the force-free limit applies. In the exterior we again evolve both the induction equation [\[dtAi\]](#dtAi){reference-type="eqref" reference="dtAi"} and the Poynting vector [\[eq:dtSiforcefree\]](#eq:dtSiforcefree){reference-type="eqref" reference="eq:dtSiforcefree"}, only now the 3-velocity is given by Eq. [\[eq:vifromcons\]](#eq:vifromcons){reference-type="eqref" reference="eq:vifromcons"}, and the evolution methods are those described in Sec. [4.1](#sec:GREFFEalgorithm){reference-type="ref" reference="sec:GREFFEalgorithm"}. Note that the same EM gauge has to be used in the interior and exterior to ensure that the magnetic field will smoothly join from the ideal-MHD regime to its force-free limit. The method we have just described applies to cases where we can treat the numerical magnetosphere as if it has no matter. We have used this method successfully in where we studied BHNS magnetospheres. In this paper, we demonstrate the validity of our approach by reproducing the aligned rotator solution in Sec. [\[aligned_rotator\]](#aligned_rotator){reference-type="ref" reference="aligned_rotator"}. In all these cases a dynamical GRMHD evolution of the matter is redundant, because the fluid four-velocity is known, and an unambiguous definition of the stellar surface is possible. This method is ideally suited for studying the dependence on the orbital separation of the total EM output generated from compact binaries endowed with force-free magnetospheres. This study can be performed by using a sequence of quasiequilibrium initial data for the fluid and the spacetime and running simulations similar to those we presented in, but at multiple orbital separations. Moreover, the approach we described in is ideal for preparing relaxed EM initial data for binary inspiral simulations, i.e., a dynamical evolution in full GR. Important studies (such as those in ), could be enhanced by adopting relaxed initial exterior EM fields, thereby avoiding the initial transient behavior associated with unrelaxed fields. In cases where a dynamical evolution is required, such as merging binary BHNSs or NSNSs, our scheme is also applicable, but with some modifications. ### Matching when the fluid rest-mass density and four velocity is determined dynamically through an evolution Now we will consider the case where we require a dynamical evolution of a star. Here neither is its surface sharply defined (because most high-resolution-schock-capturing schemes require a tenuous atmosphere and because a dynamical evolution will cause the stellar surface to oscillate) nor do we know a priori the fluid four-velocity. First, the stellar surface must be defined. We propose that the ratio \(\rho_0/b^2\) be used to determine the transition from the dense MHD interior to the tenuous force-free exterior: this ratio indicates how dominant the magnetic field is with respect to the inertia of the matter. For example, for an ideal gas the condition for EM dominance \(T_{\mu\nu}^{(\rm matter)} \ll T_{\mu\nu}^{(\rm EM)}\), generally implies \(\rho_0/b^2 \ll 1\) near the stellar surface where \(P\ll \rho_0 \leq \rho\). Therefore, if \(\rho_0/b^2\lesssim \mbox{few} \%\), then the environment is practically force-free and the exterior velocity should be recovered using Eq. [\[eq:vifromcons\]](#eq:vifromcons){reference-type="eqref" reference="eq:vifromcons"}. The remaining MHD primitive variable \(\rho_0\) can be recovered given the exterior 4-velocity, magnetic field and Poynting vector setting a floor value to prevent it from becoming too small. If \(\rho_0/b^2\gtrsim \mbox{few} \%\), then the environment is sufficiently dense and the primitives recovery can be performed the usual way, e.g. see. This scheme has not been fully implemented, yet and we will report on it in the near future. # Code Test Problems {#tests} In this section we test our new methods for evolving the GRFFE equations and for matching ideal GRMHD to its force-free limit. We test our force-free implementation with a robust suite of standard 1D solutions in Minkowski spacetime and 3D solutions in BH spacetimes, and finally we test our new matching method by reproducing the aligned rotator solution. For the tests shown in this section, the GRFFE equations are evolved by a high-resolution shock-capturing technique that employs the PPM  reconstruction, coupled to the Harten, Lax, and van Leer approximate Riemann solver . ## One-Dimensional Tests in Minkowski Spacetime These 1D tests are based on those considered in . We now present the grid setup, initial data for the vector potential, and, for comparison and completion the magnetic and electric field initial data. We do so in part to correct the initial data presented in the literature or to use slightly modified values. All these tests are evolved using the generalized Lorenz gauge and on uniformly-spaced spatial grids using three resolutions. A standard Runge-Kutta 4th order time integration scheme is employed with the Courant factor set equal to 0.5. Results from these simulations are shown in Fig. [\[1Dtests\]](#1Dtests){reference-type="ref" reference="1Dtests"}, where it is demonstrated that our code reproduces the exact solutions. All these plots show results from our "medium resolution" runs. ### Fast wave {#fast_wave} The initial configuration is defined by [^1] \[E^x(0,x) = 0.0 \, \ E^y(0,x) = 0.0 \, \ E^z(0,x) =-B^y(0,x).\] The initial data for \(v^i\) can be computed using Eq. ([\[def:vi\]](#def:vi){reference-type="ref" reference="def:vi"}), which, in Minkowski spacetime, reduces to \[\ve{v}=\frac{\ve{E}\times \ve{B}}{B^2}.\] A vector potential generating these \(B^i\) initial data is The fast wave travels to the right with speed \(\mu=1\). Hence the solution at time \(t\) is given by \[Q(t,x) = Q(0,x-t),\] where \(Q\) denotes \(B^i\), \(E^i\), or \(v^i\). We perform this test in a domain \(x\in [-0.5,1.5]\) using low, medium and high resolutions covering the domain with 160, 320, 640 zones, respectively. ### Alfvén wave {#alfven_wave} The initial data at the wave frame are where \(f(x)=1+\sin (5\pi x)\). \[E'^{x'}(x') =-B'^z(0,x') \ \, \ \ E'^y(x') = 0.0 \ \, \ \ E'^z(x') = 1.0 .\] The above data are taken from . The initial data in the grid frame are given by simple Lorentz boost where \(\mu\) is the wave speed relative to the grid frame and \(\gamma_\mu = (1-\mu^2)^{-1/2}\). Note that the Lorentz contraction \(x'=\gamma_\mu x\) has been taken into account in the above transformation. The value of \(\mu\) can be anything between \(-1\) and 1, and is set to \(-0.5\) for this test. A vector potential that generates the initial \(B^i\) is where \(g(x) = \cos (5\pi \gamma_\mu x)/\pi\). The solution at time \(t\) is given by \[Q(t,x) = Q(0,x-\mu t).\] We perform this test in a domain \(x\in [-1.5,1.5]\) using low, medium and high resolutions covering the domain with 200, 400, 800 zones. ### Degenerate Alfvén wave {#deg_alfven_wave} The initial data in the wave frame are where \[\phi(x') = \left \{ \begin{array}{lll} 0.0 & \mbox{if} & x' \leq-0.1 \\ 2.5\pi (x'+0.1) & \mbox{if} &-0.1 \leq x' \leq 0.1 \\ 0.5 \pi & \mbox{if} & x' \geq 0.1 \end{array} \right..\] The grid frame \(B^i\) and \(E^i\) can be obtained by Eqs. ([\[eq:Bboost\]](#eq:Bboost){reference-type="ref" reference="eq:Bboost"}) and ([\[eq:Eboost\]](#eq:Eboost){reference-type="ref" reference="eq:Eboost"}) with arbitrary \(\mu \in (-1,1)\). For this test, \(\mu\) is set to 0.5. A vector potential that generates the initial \(B^i\) is where \(h_1(x)= \cos [ 2.5\pi (\gamma_\mu x+0.1) ]\), \[A_z = \left \{ \begin{array}{lll}-2(\gamma_\mu x + 0.1) & \mbox{if} & x \leq-0.1/\gamma_\mu \\ -(0.8/\pi) h_2(x) & \mbox{if} &-0.1/\gamma_\mu \leq x \leq 0.1/\gamma_\mu \\ -0.8/\pi & \mbox{if} & x \geq 0.1/\gamma_\mu \end{array} \right..\] where \(h_2(x)=\sin [ 2.5\pi (\gamma_\mu x+0.1) ]\). The Alfvén speeds are given by (see ) \[\mu_a^{\pm} = \frac{B_z E_y-B_y E_z \pm \sqrt{ B_x^2 (B^2-E^2)}}{B^2}.\] For the initial data set considered here, \(\mu_a^+ = \mu_a^-=\mu\), hence the Alfvén wave is said to be degenerate. The solution at time \(t\) is \[Q(t,x) = Q(0,x-\mu t).\] We perform this test in a domain \(x\in [-1.5,1.5]\) using low, medium and high resolutions covering the domain with 200, 400, 800 zones. ### Three waves {#three_waves} For this test, the initial discontinuity at \(x=0\) splits into two fast discontinuities and a stationary Alfvén wave. The initial data are \[\ve{B}(0,x) = \left \{ \begin{array}{lll} (1.0,1.5,3.5) & \mbox{if} & x<0 \\ (1.0,3.0,3.0) & \mbox{if} & x>0 \end{array} \right., \label{eq:3wavesB}\] \[\ve{E}(0,x) = \left \{ \begin{array}{lll} (-1.0,-0.5,0.5) & \mbox{if} & x<0 \\ (-1.5,2.0,-1.5) & \mbox{if} & x>0 \end{array} \right.. \label{eq:3wavesE}\] A vector potential that generates the initial \(B^i\) is where \(H\) is the Heaviside step function. Note that this set of initial data is not the same as that in . The initial data in  are not adopted here because they do not satisfy the \(\ve{E} \cdot \ve{B}=0\) constraint. The initial data ([\[eq:3wavesB\]](#eq:3wavesB){reference-type="ref" reference="eq:3wavesB"}) and ([\[eq:3wavesE\]](#eq:3wavesE){reference-type="ref" reference="eq:3wavesE"}) are composed of three waves: where corresponding to a stationary Alfvén wave, corresponding to the right-going fast wave, and corresponding to the left-going fast wave. The solution at \(t\) is given by \[Q(t,x) = Q_a(0,x) + Q_+(0,x-t) + Q_-(0,x+t).\] We perform this test in a domain \(x\in [-1.,1.]\) using low, medium and high resolutions covering the domain with 160, 320, 640 zones. ### FFE breakdown test {#ffe_breakdown} The initial data are where \(z(x) =-10.0x+1.0\). A vector potential that generates the initial \(B^i\) is According to the simulation reported in , \(B^2-E^2\) decreases in time and approaches 0 at \(t \stackrel{>}{ _{\sim}} 0.02\), leading to the breakdown of FFE. We perform this test in a domain \(x\in [-0.4,0.6]\) using low, medium and high resolutions covering the domain with 200, 400, 800 zones. In Fig. [\[1Dtests\]](#1Dtests){reference-type="ref" reference="1Dtests"} we show the solution obtained with our code. The only difference between our solution and the one shown in Fig. 5 of , is due to the fact that we plot \(B^2-E^2\), while \((B^2-E^2)/B^2\) is plotted in . When we plot \((B^2-E^2)/B^2\) our results are in excellent agreement with . However, we prefer to show \(B^2-E^2\) as was done in , with which our results also agree. The important aspect of this problem is to demonstrate that \(B^2-E^2 =0\) occurs at \(t\approx 0.02\). As can be seen in Fig. 1 our code reproduces the solution. ## Multidimensional, Black-Hole Spacetime Tests These tests are based on the 3D BH tests considered in , only that we perform them here in Cartesian coordinates, corresponding to shifted Kerr-Schild (KS) coordinates, i.e., the radial coordinate on our grid is \(r = r_{\rm KS}-r_0\), where \(r_{\rm KS}\) is the KS radial coordinate and \(r_0\) is a constant by which we shift the coordinate. This choice is convenient because it excludes the BH singularity from our domain, as \(r=0\) corresponds to \(r_0\) in KS coordinates. The transformation from shifted KS spherical coordinates to Cartesian is done in the usual way. We now describe these tests, present the grid setup, and the results of our simulations which all reproduce the expected solutions. ### Split monopole {#split_monopole} The split monopole solution is derived from the Blandford-Znajek monopole solution by inverting the solution in the lower hemisphere. The Blandford-Znajek monopole solution is an approximate solution for small black-hole spin \(a_{*} = a/M= J/M^2 \ll 1\). The derivation can be found in . The solution in  is given in spherical Kerr-Schild coordinates and is the one we use to perform the test. The 4-vector potential is given by (dropping the subscript "KS" in r) \[\begin{aligned} A_r &=&-\frac{aC}{8}\cos\theta \left( 1 + \frac{4M}{r} \right) \sqrt{ 1 + \frac{2M}{r}} + O(a_{*}^{3}), \hfill \label{eq:Ar-bzmono} \\ A_\phi &=& M^2 C [1-\cos \theta + a^2 f(r) \cos \theta \sin^2 \theta] + O(a_{*}^4), \qquad \label{eq:Aphi-bzmono} \\ {\cal A}_t &=&-\frac{a}{8M^2} A_\phi + O(a_{*}^3), \label{eq:At-bzmono} \end{aligned}\] where \(C\) is a constant and \(f\) is the radial function given by Eq. (41) of . \[\begin{aligned} f(r) &=& \frac{r^2(2r-3M)}{8M^3} L \left( \frac{2M}{r}\right) + \frac{M^2+3Mr-6r^2}{12M^2}\ln \frac{r}{2M} \cr && + \frac{11}{72} + \frac{M}{3r} + \frac{r}{2M}-\frac{r^2}{2M^2}. \end{aligned}\] where \(L\) is the dilogarithm function defined as \[L(x) = {\rm Li}_2(x) + \frac{1}{2} \ln x \ln (1-x) \ \ \ \mbox{for } 0<x<1 \label{def:Lx01}\] and \({\rm Li}_2\) is defined as \[{\rm Li}_2(x) =-\int_0^1 \frac{\ln(1-tx)}{t} dt = \sum_{k=1}^\infty \frac{x^k}{k^2}. \label{def:Li2}\] Note that our Eq. [\[eq:Aphi-bzmono\]](#eq:Aphi-bzmono){reference-type="eqref" reference="eq:Aphi-bzmono"} is not exactly the same as in  ; the term \(CM^2\) has been added to the original expression given in  to prevent the Cartesian components of \(A_i\) from diverging on the upper \(z\)-axis. The magnetic field is given by Eqs. (47)--(49) of . However, there is a factor of \(\alpha\) different between their definition of \(B^i\equiv {}^*F^{it}\) and the \(B^i=n_{\nu} {}^*F^{\nu i}=\alpha {}^*F^{it}\) adopted here. So, we have The Faraday tensor \(F_{\mu \nu}\) is given by Eqs. (27)--(31) of . They can be used to compute the electric field: \[E_i = n^{\mu} F_{i\mu} =-\frac{1}{\alpha} F_{ti} + \frac{\beta^j}{\alpha} F_{ji},\] where, for the Kerr-Schild metric in spherical coordinates, and where \(\Sigma = r^2+a^2\cos^2 \theta\). Finally, \[\begin{aligned} E_r &=&-\frac{C a^3}{8\alpha M^3} f'(r) \cos \theta \sin^2 \theta \\ E_\theta &=&-\frac{Ca}{8\alpha}[\sin \theta + a^2 f(r) \sin \theta (2 \cos^2 \theta-\sin^2 \theta) ] \cr &&-\beta^r \sqrt{\gamma}\, \frac{a C}{8 r^2}\left( 1+\frac{4M}{r}\right) \\ E_\phi &=& \frac{\beta^r}{\alpha M} Ca^2 f'(r) \cos \theta \sin^2 \theta. \end{aligned}\] Note that \(f(r) \sim-r^2 \ln r/4\) as \(r \rightarrow \infty\), invalidating the solution at large \(r\) (because \(B^2 < E^2\) at sufficiently large \(r\)). Following   we drop terms involving \(f(r)\) and \(f'(r)\), making the solution accurate only to first order in \(a_*\). To perform the split monopole the constant \(C\) is changed to \(-C\) in the lower hemisphere (\(\theta > \pi/2\)), in the expressions of \(B^i\) and \(E^i\). The vector potential for the split monopole test then can be written as \[\begin{aligned} A_r &=&-\frac{Ca}{8} |\cos \theta| \left( 1 + \frac{4M}{r} \right) \sqrt{ 1 + \frac{2M}{r}} + O(a_{*}^3) \qquad \\ A_\phi &=& C (1-|\cos \theta|) + O(a_{*}^2). \end{aligned}\] As pointed out in, the split-monopole configuration is sensible only if there exists a conducting disc in the equatorial plane of the black hole to sustain it. Otherwise, the equatorial current sheet cannot be stable--the magnetic field lines will reconnect and be pushed away. If one assumes that the equatorial current sheet is stable, because it is sustained by a disk, then no reconnection is expected to take place. We can model both scenarios by turning off and on our resistivity prescription, i.e., the nulling of the inflow velocity in the current sheet. If we do null the inflow velocity into the equatorial current sheet, no reconnection takes place and our solution is in agreement with the solution found in, as expected. The results of this test without the resistivity prescription are shown in Fig. [\[split_mon\]](#split_mon){reference-type="ref" reference="split_mon"}, and are in good agreement with the ones obtained in. We perform this test setting \(a_{*} = 0.1\), and chose \(r_0 = 1.0M\), so that the BH horizon corresponds to \(r \approx 0.995M\) in the shifted KS radial coordinate. We perform this test on a fixed-mesh-refinement grid hierarchy with 6 levels of refinement setting the outer boundary at 100M. The half-side length of the refinement levels is \(3.125\times 2^{6-n}M, \ \ n=1,2,\ldots, 6\), \(n=1\) indicating the coarsest level in the hierarchy. The resolution of each level is \(\Delta x_{\min}\times 2^{6-n}, \ \ n=1,2,\ldots, 6\), where \(\Delta x_{\min}\) is the resolution of the finest level. We use 3 resolutions \(\Delta x_{\min}=M/8\), \(\Delta x_{\min}=M/16\), and \(\Delta x_{\min}=M/24\). The Courant factor is set to \(0.03125\times 2^{n-1}, \ n=1,\ldots, 3\) and 0.5 for \(n=4,5,6\). We use the generalized Lorenz gauge to run the test with damping parameter \(solution shown in Fig.~\ref{split_mon} corresponds to our low resolution run, and the results of all other resolutions are almost overlapping, indicating that the resolutions used are sufficiently high. \subsubsection{The Wald solution} \label{wald} The EM field of the solution to Maxwell's equations in the electrovacuum about a black hole is generated by the 4-vector potential \begin{equation} {\cal A}_\mu = \frac{B_0}{2} (\phi_\mu + 2 a t_\mu), \end{equation} where\)B_0\(is a constant,\)\^= (/)\^\(and\)t\^= (/t)\^\(. In the Schwarzschild black hole case (\)a=0\(), this electrovacuum solution is also a force-free solution, which will be the case considered in this test. The 4-vector potential in this case is \begin{equation} {\cal A}_\mu = \frac{B_0}{2} \phi_\mu. \end{equation} The 3-vector potential is\)A_i = B_0 \_i/2= B_0 g\_i/2\(. In Kerr-Schild metric written in spherical coordinates, the only nonvanishing component is \labeq{Aphiwald}{ A_\phi = \frac{B_0}{2} g_{\phi \phi} = \frac{B_0}{2} r^2 \sin^2 \theta. } The magnetic field is given by \begin{equation} B^i = \epsilon^{ijk} \partial_j A_k = \frac{[ijk]}{\sqrt{\gamma}} \partial_j A_k, \end{equation} where\)\[ijk\]\(denotes the antisymmetric permutation symbol. Hence the components of\)B\^i\(in spherical Kerr-Schild coordinates are \begin{eqnarray} B^r &=& \frac{1}{\sqrt{\gamma}\,} \partial_\theta A_\phi = B_0 \left( 1+ \frac{2M}{r}\right)^{-1/2} \cos \theta \\ B^\theta &=&-\frac{1}{\sqrt{\gamma}\,} \partial_r A_\phi =-\frac{B_0}{r} \left( 1+ \frac{2M}{r}\right)^{-1/2} \sin \theta \\ B^\phi &=& 0 \\ B_r &=& B_0 \cos \theta \sqrt{1+ \frac{2M}{r}} \\ B_\theta &=&-B_0 r \sin \theta \left( 1+ \frac{2M}{r}\right)^{-1/2} \\ B_\phi &=& 0 \\ B^2 &=& B_0^2 \left( 1-\frac{2M}{r+2M} \sin^2 \theta \right). \end{eqnarray} As\)r \(,\)B\^i\(becomes a uniform vector with magnitude\)B_0\(and points in the\)z\(-direction. To compute\)E\^i\(, first calculate \labeq{Ei}{ E_i = n^\nu F_{i\nu} = \frac{1}{\alpha} (F_{i0}-\beta^j F_{ij}). } Given that\) F\_i0 = \_i A\_0-\_t A\_i = \_i \_0 = \_i g\_t = 0 \(and\)F\_ij=\_i A_j-\_j A_i\(. The nonzero\)F\_\(components are \labeq{Fmnunuwald}{ \begin{split} F_{r\phi} = &\-F_{\phi r} = B_0 r \sin^2 \theta, \\ F_{\theta \phi} = &\-F_{\phi \theta} = B_0 r^2 \sin \theta \cos \theta. \end{split} } By use of Eq. \eqref{Fmnunuwald}, Eq. \eqref{Ei} yields \begin{equation} E_r = E_\theta = 0 \ \ \, \ \ \ E_\phi = 2M B_0 \left( 1+ \frac{2M}{r}\right)^{-1/2} \sin^2 \theta, \end{equation} and \begin{eqnarray} E^r &=& E^\theta=0 \ \ \, \ \ \ E^\phi = \frac{2MB_0}{r^2} \left( 1+ \frac{2M}{r}\right)^{-1/2} \\ E^2 &=& \frac{4M^2 B_0^2 \sin^2 \theta}{r^2} \left( 1+ \frac{2M}{r}\right)^{-1}. \end{eqnarray} The electric field vanishes as\)r\(. It is now straightforward to see that\)**E**=0\(and\)E\^2 \< B\^2\(. Hence, this electrovacuum solution is also a force-free solution. The 3-velocity can be calculated by \labeq{vi}{ v^i = \alpha \frac{\epsilon^{ijk} E_j B_k}{B^2}-\beta^i = \frac{[ijk] E_j B_k}{B_0^2 r \sin \theta (r+2M\cos^2 \theta)}-\beta^i. } and we find \begin{eqnarray} v^r &=&-\frac{2M\cos^2 \theta}{r+2M\cos^2 \theta} \\ v^\theta &=& \frac{M \sin 2\theta}{r (r+2M \cos^2 \theta)} \\ v^\phi &=& 0. \end{eqnarray} For this test we arbitrarily chose\)r_0 = 0.4M\(, so that the BH horizon corresponds to\)r = 1.6M\(in the shifted KS radial coordinate. We perform this test on the same fixed-mesh-refinement grid hierarchy as the split-monopole test, using the same 3 resolutions and EM gauge. In Fig.~\ref{vac_wald} we show the Poloidal field lines at\)t=0M\(and\)t=5M\(for the low resolution run-the two overlap and cannot be distinguished by eye. Since this test is the only smooth 3D exact solution in the testbeds we consider we use it to also show that our code is convergent. Our convergence test study is presented in section Sec.~\ref{convergence}. \subsubsection{Magnetospheric Wald Problem} \label{magnetospheric_wald} This again is a force-free problem. The initial data for the magnetic field are given by the same spatial vector potential as the Wald's solution, i.e., \begin{equation} A_i = \frac{B_0}{2} (\phi_i + 2 a t_i) = \frac{B_0}{2} (g_{i\phi} + 2a g_{ti}). \end{equation} However, the electric field is set to 0 initially, as in~. Hence\)S_i=0\(and\)v\^i=-\^i\(. There is no analytic solution to this problem. The evolution of the initial data is expected to reach a steady state similar to the one reported in~. Following, we perform this test setting\)a\_\*=0.9\(. We also set\)r_0 = 0.4359M\(, so that the BH horizon lies at\)r 1.0M\(on our grid. We perform this test on the same fixed-mesh-refinement grid hierarchy as the other BH tests, using the same 3 resolutions and EM gauge. In Fig.~\ref{fig:wald} we plot the poloidal magnetic field lines at\)t=126M\(at which point the solution has reached steady state and is very similar to that obtained in. \subsection{Force-free aligned rotator} \label{aligned_rotator} Here we reproduce the aligned rotator, force-free solution in flat spacetime. However, instead of applying the boundary condition on the NS surface, we use our new matching technique, which we described in Sec.~\ref{easy_matching}, to evolve the magnetic field both interior and exterior to the star. In this approach the density profile of the star can be anything, as the magnetic field does not back-react onto the matter and an integration of the ideal MHD fluid equations is not performed. Instead, the density and velocity are evolved simply by ``rotating'' their initial values as described in ~. The density profile serves only as a proxy for locating the surface of the star. We endow the star with a uniform rotational three-velocity \labeq{}{ \mathbf{v} = \Omega \mathbf{e}_z \times \mathbf{r}, } where\)\_z\(is the unit vector in the z-direction, and\[is the stellar angular velocity. As in we choose a spherical star and set\]such that the theoretically expected location of the light cylinder radius,\)R\_LC\(, is 5 stellar radii away from the stellar center, i.e.,\)= 1/5R\_NS\(. In the exterior, the 3-velocity is set to 0. The electric field is set according to\) =-\(everywhere and the Poynting vector is calculated using Eq. \eqref{def:Poynting}. The star and its magnetosphere are endowed with a magnetic field corresponding to a dipole determined by the toroidal vector potential \labeq{}{ A_\phi = \frac{\mu\varpi^2}{r^3}, } where\)= B_p R\_NS\^3 /2\(is the magnetic dipole moment, the cylindrical radial coordinate\)\^2 = x\^2+y\^2\(, and\)r=\(is the radial coordinate. We perform the test using 8 levels of refinement and set the outer boundary at\)`<!-- -->`{=html}35.3R\_LC\(. The length of each refinement box is\)`<!-- -->`{=html}2.94R\_NS\^8-n, n=1,2,...,8\(, where\)n=8\(corresponds to the finest refinement level. We use 3 resolutions: the low, medium and high resolutions cover the stellar radius with 34, 68 and 87 zones, respectively. In Fig. \ref{rot_star} we show the poloidal magnetic field lines in the x-z plane, where it is clear that our code successfully captures the standard features of the pulsar magnetosphere: 1) the formation of a Y-point at the expected location of the light cylinder, 2) open field lines above the equatorial current sheet and beyond the light cylinder, and 3) dipole magnetic field structure within the light cylinder. In addition, the evolved interior field remains ``frozen in'' to the rotating matter. The right panel of the figure shows the structure of the magnetic field in the interior and the immediate exterior of the star, demonstrating that our matching technique is smooth. The expected spin-down luminosity of an aligned rotator is\)L = (1)\^2 \^4\(. We have calculated the outgoing EM radiation using the Poynting flux and the Penrose scalar\)\_2\((see e.g.~), and we find that it converges to a value within\)`<!-- -->`{=html}4%\(of\)\^2 \^4\(, and hence in good agreement with previous studies. We plot the angular frequency of the magnetic field lines in the exterior \labeq{}{ \Omega_F(r,\theta) = \frac{F_{tr}}{F_{r\phi}} = \frac{F_{t\theta}}{F_{\theta \phi}}, } on the x-z plane at\)r = 0.3R\_LC\(and\)r = 0.7R\_LC\(as a function of the polar angle\)\(. The result after\)\~3\(periods of evolution is shown in Fig. \ref{OmegF} (cf. who perform axisymmetric high-resolution simulations). It is clear that the magnetosphere within the light cylinder corotates with the star and that the higher the resolution, the closer is the magnetosphere to corotation. \subsection{Convergence} \label{convergence} The Wald vector potential which generates the stationary magnetic field is itself time independent provided the proper electromagnetic gauge choice is made. A straightforward calculation demonstrates that \labeq{ucrossB}{ \begin{split} \epsilon_{ijk} v^j B^k = &\ \epsilon_{ijk}\bigg(\alpha \frac{\epsilon^{j\ell m}E_\ell B_m}{B^2}-\beta^j\bigg) B^k \\ = &\-\alpha E_i-\epsilon_{ijk}\beta^j B^k \\ = &\-\alpha E_i-\tilde\epsilon_{ijk}\tilde\epsilon^{k\ell m}\beta^j \partial_\ell A_m \\ = &\-\alpha E_i-\beta^j( \partial_i A_j-\partial_j A_i) \\ = &\-\alpha E_i-\beta^j F_{ij} = 0, \end{split} } where in the first line we used Eq. \eqref{vi}, in the second line we used the degeneracy constraint (\)**E** = 0\(), in the second and fourth lines we used the property\)\_ijk\^jm = (\_i\^\_k\^m-\_i\^m_k\^)\(and in the third line we used the definition\)B\^k=\^km\_A_m\(. The last equality in the fifth line holds true because of Eq. \eqref{Ei} and\)F\_i0 = 0\(. This result implies that\)\_t **B** = 0\(from the magnetic induction equation \eqref{eq:induction}, but also has an interesting consequence regarding the typical electromagnetic gauges we use in our code: The evolution equation for the vector potential is given by Eq. \eqref{dtAi}. In the original algebraic electromagnetic gauge\)= \^j A_j\(. Hence, the evolution equation \eqref{dtAi} preserves the initial\)A_i\(field (\)\_t A_i = 0\(). In fact, a straightforward calculation using the Wald vector potential shows that\)\^i A_i = 0\(, which implies that the right-hand-side of Eq. \eqref{dtAi} must be \labeq{walddtAi}{ \partial_t A_i =-\partial_i(\alpha \Phi) } Thus, any electromagnetic gauge condition, for which\)\_i()=0\(will preserve the initial vector potential. Thus, the gauge\)=const.\(also preserves the initial A-field. This is the case in the (generalized Lorenz) gauge we have developed, as we now demonstrate. The generalized Lorenz gauge is \labeq{}{ \partial_t (\sqrt{\gamma}\Phi) + \partial_j(\alpha \sqrt{\gamma}A^j)-\partial_i (\sqrt{\gamma}\beta^i \Phi) =-} where\) hence, This means that if \(\Phi = 0\) initially, \(\partial_t \Phi = 0\). Thus, even the generalized Lorenz gauge will preserve the Wald A-field [\[Aphiwald\]](#Aphiwald){reference-type="eqref" reference="Aphiwald"}, as long as the initial value for \(\Phi\) is 0. Of course due to truncation error the right-hand-side of the evolution equation for \(A_i\) will not be exactly 0, but will converge to 0 at second order which is the accuracy of our scheme. A test of this convergence is shown in Fig. [\[vac_wald_conv\]](#vac_wald_conv){reference-type="ref" reference="vac_wald_conv"}, where we plot the L2 norm of the difference between the numerical and analytic solution for the vector potential defined as where \(A_i^{\rm num},\ A_i^{\rm ex}\) designate the numerical and exact solutions, respectively. This norm should converge to 0 with increasing resolution. Fig.  [\[vac_wald_conv\]](#vac_wald_conv){reference-type="ref" reference="vac_wald_conv"} demonstrates that our code is second-order convergent. # Summary {#sec:summaryandfuturework} Neutron stars either in isolation or in compact binaries are likely to be endowed with a force-free magnetosphere. For inpiralling binaries, the GWs in the premerger regime can be accompanied by detectable "precursor" electromagnetic signals propagating through this magnetosphere. To study these effects numerical relativity simulations are necessary and require a scheme that matches the ideal MHD interior of the NS to the exterior force-free magnetosphere. Here we present a new method for matching general relativistic ideal MHD to its force-free limit. We have tested out force-free code using a series of 1D flat spacetime tests, as well as 3D stationary black hole tests. We confirmed the validity of our new matching scheme by reproducing the well-known aligned rotator solution. We demonstrated the robustness of our algorithms and code and new techniques and we have shown that for smooth solutions our new code is second-order convergent. This new method has already been used in, where we presented the first GR simulations of a binary black hole-neutron star magnetosphere. We plan to use this code to simulate other complicated dynamical spacetime scenarios involving neutron stars and their magnetospheres. In a future paper we also plan to extend our code to handle dynamical scenarios to extend our study to the inspiral of compact binaries involving neutron stars. It is a pleasure to thank Yuk Tung Liu, Zachariah B. Etienne, Roman Gold, and Milton Ruiz for useful discussions. This paper was supported in part by NSF Grants PHY-0963136 and PHY-1300903 as well as NASA Grants NNX11AE11G and NN13AH44G at the University of Illinois at Urbana-Champaign. VP gratefully acknowledges support from a Fortner Fellowship at UIUC. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by NSF grant number OCI-1053575. [^1]: Note that  give the initial data for \(\mathcal B^i\) and \(\mathcal E^i\). However, since the FFE equations are invariant if \(\mathcal B^i\) and \(\mathcal E^i\) are multiplied by a constant factor, initial data with \(B^i\) and \(E^i\) having the same values as the ones with \(\mathcal B^i\) and \(\mathcal E^i\) are equally valid and the subsequent evolution will be exactly the same as the old set of initial data after multiplying an appropriate factor. Therefore, the initial data listed in this note are not multiplied by the factor \(\sqrt{4\pi}\).

{'timestamp': '2013-10-22T02:00:57', 'yymm': '1310', 'arxiv_id': '1310.3274', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3274'}

# Introduction Discontinuous Galerkin (DG) methods are known to be flexible and efficient solvers for a wide range of partial differential equations. Among their advantages, when applied to second order elliptic problems, we emphasize that they are locally conservative, they can handle general meshes with hanging nodes and they allow the use of different polynomial degrees in each element. DG methods can be coupled with the boundary element method (BEM) in different ways. In it was shown that it is possible to benefit from the features highlighted above when approximating non-homogeneous (and even nonlinear ) exterior elliptic problems if a local discontinuous Galerkin method (LDG) is used as an interior solver in combination with the BEM. The symmetric LDG-BEM formulation is obtained by rewriting locally the elliptic problem in mixed form and considering a Calderón identity on the boundary. In this way, one ends up with a system of two variational equations in the interior domain (involving both the potential and the flux as independent variables) and a system of two boundary integral equations relating the Cauchy datum of the problem on the coupling interface. In the first LDG-BEM formulation, the coupling between the two systems is performed by using Costable's approach. From the DG point of view, this amounts to using the normal derivative of the solution on the coupling boundary as a Neumann datum when defining the numerical fluxes for the LDG method. In the resulting coupled scheme, the normal derivative becomes an independent unknown and the other BEM variable (the discrete trace) must match the discrete potential that comes from the LDG method. The problem is that these unknowns are of different nature: the restriction of the LDG approximation of the potential to the coupling interface is discontinuous while the BEM discretization is conforming and produces a continuous and piecewise polynomial approximation of this variable. This inconvenience is addressed in by introducing a further unknown that acts as a Lagrange multiplier and enforces weakly the imposition of the missing transmission condition. A later paper eliminated the need of the Lagrange multiplier by demanding that the discontinuous piecewise polynomial functions that approximate the potential in the LDG method be continuous at the coupling interface. Here, the normal derivative is the only boundary unknown, which reduces the number of unknown functions by two with respect to the first version. However, in order to deal properly with this formulation in practice a Lagrange multiplier must come again into play. Moreover, this formulation imposes for the BEM the mesh inherited from the interior partition of the domain, which reduces much of the flexibility provided by the discrete Galerkin method near the coupling boundary. Finally, we point out that recently non-symmetric couplings of DG with BEM have also been studied, cf. and the references therein. In this paper, following, we take advantage of the fact that the flux variable is an LDG active unknown (as in the traditional mixed formulation) and consider a dual approach: we define the numerical fluxes by considering the trace of the solution on the coupling boundary as Dirichlet datum. Hence, as opposed to the former strategy, the trace of the solution is an independent variable while the LDG normal flux and the normal derivative must be merged on the coupling boundary. Notice that in this case both variables are (naturally) nonconforming and no Lagrange multiplier or special restriction is needed to match them. Consequently, the resulting numerical scheme enjoys all the good properties of a typical DG method and allows for using an independent boundary mesh. Moreover, one can employ both a conforming or a nonconforming approximation on the boundary. In this paper, we take advantage of the results from to deal with a DG finite element method on the boundary, the resulting scheme will be referred to as the LDG-FEM/DG-BEM method. To our knowledge, this is the first FEM-BEM scheme that combines DG approximations on the boundary and in the interior. Technical difficulties that already arose in, oblige us to consider conforming and quasi-uniform families of triangulations on the coupling boundary. Following the technique from, this can be relaxed to meshes that are conforming and quasi-uniform on planar sub-surfaces of the coupling interface. However, for simplicity, the technical details for such an extension are omitted here and we will consider globally conforming and quasi-uniform boundary meshes. Fortunately, restrictions on the boundary mesh have no negative impact on the triangulation of the interior domain since the two meshes are related by a mild local condition, see [\[localUnif\]](#localUnif){reference-type="eqref" reference="localUnif"} below. Finally, we analyze the scheme that is obtained by using a conforming rather than non-conforming BEM on the interface. The resulting scheme will be referred to as the LDG-FEM/BEM method. The paper is organized as follows. In Section [2](#s2){reference-type="ref" reference="s2"} we present our model problem and recall some basic properties of boundary integral operators. For simplicity of exposition we will restrict our interest to a three-dimensional problem posed in the whole space. In Section [3](#s3){reference-type="ref" reference="s3"} we derive the LDG-FEM/DG-BEM scheme and prove that it admits a unique solution. Stability and a priori error estimates are proved in Section [4](#s4){reference-type="ref" reference="s4"}. In Section [5](#s5){reference-type="ref" reference="s5"} we show that the same technical arguments provide (without the quasi-uniformity requirement for the meshes on the coupling boundary) a convergence result for the LDG-FEM/BEM scheme. Finally, numerical experiments are reported in Section [6](#s6){reference-type="ref" reference="s6"}. Given a real number \(r\geq 0\) and a polyhedron \(\mathcal O\subset \mathbb R^d\), \((d=2,3)\), we denote the norms and seminorms of the usual Sobolev space \(H^r(\mathcal O)\) by \(\|\cdot \|_{r,\mathcal O}\) and \(|\cdot|_{r,\mathcal O}\) respectively (cf. ). We use the convention \(L^2(\mathcal O):= H^0(\mathcal O)\) and let \((\cdot,\cdot)_{\mathcal O}\) be the inner product in \(L^2(\mathcal{O})\). We recall that, for any \(t \in [-1,\: 1 ]\), the spaces \(H^{t}(\partial \mathcal O)\) have an intrinsic definition (by localization) on the Lipschitz surface \(\partial \mathcal O\) due to their invariance under Lipschitz coordinate transformations. Moreover, for all \(0< t\leq 1\), \(H^{-t}(\partial\mathcal O)\) is the dual of \(H^{t}(\partial\mathcal O)\) with respect to the pivot space \(L^2(\partial \mathcal{O})\). Also, \(\dual{\cdot, \cdot}_{\partial \mathcal{O}}\) denotes both the \(L^2(\partial \mathcal{O})\) inner product and its extension to the duality pairing of \(H^{-t}(\partial \mathcal{O})\times H^{t}(\partial \mathcal{O})\). # The model problem {#s2} Let \(\Omega\subset \mathbb{R}^3\) be a bounded polyhedral domain with a Lipschitz boundary \(\Gamma\). We denote by \(\mathbf{n}\) the unit normal vector on \(\Gamma\) that points towards \(\Omega^e:= \mathbb{R}^3\setminus \bar \Omega\). For the sake of simplicity, we assume that \(\Omega^e\) is connected. We consider the transmission problem \[\label{ModelProblem} \begin{array}{rcll}-\Delta u &=& f &\text{in \(\Omega\)}\\[2ex] u &=& u^e + g_0 &\text{on \(\Gamma\)}\\[2ex] \nder{u} &=& \nder{u^e} + g_1 &\text{on \(\Gamma\)}\\[2ex] -\Delta u^e &=& 0 &\text{in \(\Omega^e\)}\\[2ex] u^e &=& O(\displaystyle\frac{1}{\abs{\boldsymbol{x}}}) &\text{as \(\abs{\boldsymbol{x}}\to \infty\)}, \end{array}\] where \(f\in L^2(\Omega)\), \(g_0\in H^{1/2}(\Gamma)\) and \(g_1\in L^2(\Gamma)\) are given functions. We can write the problem in \(\Omega\) by introducing the flux \(\boldsymbol{\sigma}\) as a new variable: \[\label{inOmega} \begin{array}{rcll} \boldsymbol{\sigma} &=& \nabla u &\text{in \(\Omega\)},\\[2ex]-\textrm{div} \boldsymbol{\sigma} &=& f &\text{in \(\Omega\)}. \end{array}\] With the notation \[\lambda := \nder{u^e}\quad \text{and} \quad \psi = u^e|_\Gamma,\] the transmission conditions are \[\label{trans} \begin{array}{rcll} u &=& \psi + g_0 &\text{on \(\Gamma\)}\\[2ex] \boldsymbol{\sigma} \cdot \boldsymbol{n} &=& \lambda + g_1 &\text{on \(\Gamma\)}. \end{array}\] Using the integral representation of the harmonic function \(u^e\) in \(\Omega^e\) gives \[\label{IntegralRep} u^e = \Psi_{DL}(\psi)-\Psi_{SL}(\lambda) \quad\text{in \(\Omega^e\)}\] where \[\Psi_{SL}( \quad \text{and} \quad \Psi_{DL}(\varphi)(x) := \int_{\Gamma} \displaystyle\frac{\partial E(\abs{\boldsymbol{x}-\boldsymbol{y}})}{\partial \boldsymbol{n}(\boldsymbol{y})}\, \varphi(\boldsymbol{y})\, dS(\boldsymbol{y})\] are the single and double layer potentials, respectively, and \(E(\abs{\boldsymbol{x}}):= \frac{1}{4\pi}\frac{ 1}{\abs{\boldsymbol{x}}}\) is the fundamental solution of the Laplace operator. The jump properties of the single and double layer potentials across \(\Gamma\) provide the following integral equations relating the Cauchy data on this boundary: \[\begin{aligned} \label{inteq1} \psi &= ( \frac{\text{id}}{2} + K) \psi-V \lambda \\ \label{inteq2} \lambda &=-W\psi + (\frac{\text{id}}{2}-K') \lambda \end{aligned}\] where \(V\), \(K\), \(K'\) are the boundary integral operators representing the single, double and adjoint of the double layer, respectively, and \(W\) is the hypersingular operator. Let us recall some important properties of the boundary integral operators, see for details. The boundary integral operators are formally defined at almost every point \(\boldsymbol{x} \in \Gamma\) by \[V \displaystyle\frac{\partial E(\abs{\boldsymbol{x}-\boldsymbol{y}})}{\partial \boldsymbol{n}(\boldsymbol{y})}\, \varphi(\boldsymbol{y})\, dS(\boldsymbol{y}),\] \[K' \displaystyle\frac{\partial E(\abs{\boldsymbol{x}-\boldsymbol{y}})}{\partial \boldsymbol{n}(\boldsymbol{x})}\, W \varphi (\boldsymbol{x}) :=-\displaystyle\frac{\partial}{\partial \boldsymbol{n}(\boldsymbol{x})} \int_\Gamma \displaystyle\frac{\partial E(\abs{\boldsymbol{x}-\boldsymbol{y}})}{\partial \boldsymbol{n}(\boldsymbol{y})}\, \varphi(\boldsymbol{y})\, dS(\boldsymbol{y}).\] They are bounded as mappings \(V:\, H^{-1/2}(\Gamma)\to H^{1/2}(\Gamma)\), \(K:\, H^{1/2}(\Gamma) \to H^{1/2}(\Gamma)\) and \(W:\, H^{1/2}\to H^{-1/2}(\Gamma)\). The single layer operator is coercive, there exists \(C_0>0\) such that \[\label{eq-coercive-1} \dual{\chi, V\chi}_\Gamma \,\ge\, C_0\, \norm{\chi}_{-1/2,\Gamma}^2\quad \forall\, \chi \,\in\, H^{-1/2}(\Gamma)\] and \[\label{eq-coercive-2} \dual{W \varphi, \varphi}_\Gamma \,+\, \left(\int_\Gamma \varphi \, \right)^2 \,\ge\, C_0\,\norm{\varphi}_{1/2,\Gamma}^2\quad \forall\,\varphi \,\in\, H^{1/2}(\Gamma).\] Moreover, \(V:\, H^{s-1}(\Gamma)\to H^s(\Gamma)\) is bounded for any \(0\leq s \leq 1\). We recall that the operators \(V\) and \(W\) are related by \[W = \textrm{curl}_\Gamma V \mathbf{curl}_\Gamma\] where \(\mathbf{curl}_\Gamma\) is the surface curl operator and \(\textrm{curl}_\Gamma\) is its adjoint operator, cf.. Consequently, \[\dual{W\psi, \varphi}_\Gamma = \dual{\mathbf{curl}_\Gamma \psi, V \mathbf{curl}_\Gamma \varphi}_\Gamma, \qquad \forall \psi, \varphi \in H^{1/2}(\Gamma).\] # The LDG-FEM/DG-BEM formulation {#s3} We denote by \(\mathcal{T}_h\) a subdivision of the domain \(\bar \Omega\) into shape regular tetrahedra \(K\) of diameter \(h_K\) and unit outward normal to \(\partial K\) given by \(\boldsymbol{n}_K\). We point out that the partition \(\mathcal{T}_h\) is not necessarily a conforming mesh of \(\bar \Omega\). We also introduce a shape regular conforming quasi-uniform triangulation \(\mathcal{G}_h:=\set{T}\) of the interface \(\Gamma\) into triangles \(T\) of diameter \(h_T\). The set of edges of \(\mathcal{G}_h\) is denoted by \(\mathcal{E}_h\). The parameter \(h\) represents the mesh size, i.e., \(h:= \max_{K\in \mathcal{T}_h;\; T\in\mathcal{G}_h} \{h_K, h_T\}\). Henceforth, given any positive functions \(A_h\) and \(B_h\) of the mesh parameter \(h\), the notation \(A_h \lesssim B_h\) means that \(A_h \leq C B_h\) with \(C>0\) independent of \(h\) and \(A_h \simeq B_h\) means that \(A_h \lesssim B_h\) and \(B_h \lesssim A_h\). We say that a closed subset \(F\in \overline{\Omega}\) is an interior face if \(F\) has a positive 2-dimensional measure and if there are distinct elements \(K\) and \(K'\) such that \(F = K\cap K'\). A closed subset \(F\in \overline{\Omega}\) is a boundary face if there exists \(K\in \mathcal{T}_h\) such that \(F\) is a face of \(K\) and \(F = K\cap \Gamma\). We consider the set \(\mathcal{F}_h^0\) of interior faces and the set \(\mathcal{F}_h^\partial\) of boundary faces and introduce \[\mathcal{F}_h = \mathcal{F}_h^0\cup \mathcal{F}_h^\partial.\] For any element \(K\in \mathcal{T}_h\), we introduce the set \[\mathcal{F}(K):= \set{F\in \mathcal{F}_h;\quad F\subset \partial K}\] of faces composing the boundary of \(K\). Similarly, for any \(T\in \mathcal{G}_h\), we introduce the set \[\mathcal{E}(T):= \set{e\in \mathcal{E}_h;\quad e\subset \partial T}.\] We also consider for any \(T\in \mathcal{G}_h\), \[\mathcal{F}(T):= \set{F\in \mathcal{F}_h^\partial;\quad F\cap T\not=\emptyset}.\] In what follows we assume that \(\mathcal{T}_h\cup\mathcal{G}_h\) is locally quasi-uniform, i.e., there exists \(\delta>1\) independent of \(h\) such that \(\delta^{-1} \leq \frac{h_K}{h_{K'}} \leq \delta\) for each pair \(K\), \(K'\in \mathcal{T}_h\) sharing an interior face and \(\delta^{-1} \leq \frac{h_K}{h_{T}} \leq \delta\) for each pair \(K\in \mathcal{T}_h\), \(T\in \mathcal{G}_h\) with \(K\cap T\not=\emptyset\). This assumption implies that the sets \(\mathcal{F}(K)\) and \(\mathcal{F}(T)\) have uniformly bounded cardinalities and that there exists a constant \(C>0\) independent of \(h\) such that \[\label{localUnif} h_F \leq h_K \leq C \delta h_F \quad \forall F\in \mathcal{F}(K) \quad \text{and} \quad h_F \leq h_T \leq C \delta h_F \quad \forall F\in \mathcal{F}(T),\] where \(h_F\) stands for the diameter of the face \(F\). For any \(s\geq 0\), we consider the broken Sobolev spaces \[H^s(\mathcal{T}_h) := \prod_{K\in \mathcal{T}_h} H^s(K), \qquad \mathbf{H}^s(\mathcal{T}_h) := \prod_{K\in \mathcal{T}_h} H^s(K)^3,\] \[H^s(\mathcal{G}_h) := \prod_{T\in \mathcal{G}_h} H^s(T), \qquad \mathbf{H}^s(\mathcal{G}_h) := \prod_{T\in \mathcal{G}_h} H^s(T)^3.\] For each \(v:=\set{v_K}\in H^s(\mathcal{T}_h)\), \(\boldsymbol{\tau}:= \set{\boldsymbol{\tau}_K}\in \mathbf{H}^s(\mathcal{T}_h)\) and \(\varphi:= \set{\varphi_T}\in \mathbf{H}^s(\mathcal{G}_h)\), the components \(v_K\), \(\boldsymbol{\tau}_K\) and \(\varphi_T\) represent the restrictions \(v|_K\), \(\boldsymbol{\tau}|_K\) and \(\varphi|_T\). When no confusion arises, the restrictions of these functions will be written without any subscript. The spaces \(H^s(\mathcal{T}_h)\) and \(\mathbf{H}^s(\mathcal{T}_h)\) are endowed with the Hilbertian norms \[\norm{v}_{s,\mathcal{T}_h}^2 := \sum_{K\in \mathcal{T}_h} \norm{v_K}^2_{s,K} \qquad \norm{\boldsymbol{\tau}}_{s,\mathcal{T}_h}^2 := \sum_{K\in \mathcal{T}_h} \norm{\boldsymbol{\tau}_K}^2_{s,K}.\] The corresponding seminorms are denoted by \[|v|_{s,\mathcal{T}_h}^2 := \sum_{K\in \mathcal{T}_h} |v_K|^2_{s,K} \qquad |\boldsymbol{\tau}|_{s,\mathcal{T}_h}^2 := \sum_{K\in \mathcal{T}_h} |\boldsymbol{\tau}_K|^2_{s,K}.\] Similarly, the norms and the seminorms on \(H^s(\mathcal{G}_h)\) are given by \[\norm{\varphi}_{s,\mathcal{G}_h}^2 := \sum_{T\in \mathcal{G}_h} \norm{\varphi_T}^2_{s,T} \qquad |\varphi|_{s,\mathcal{G}_h}^2 := \sum_{T\in \mathcal{G}_h} |\varphi_T|^2_{s,T}.\] Identical definition for the norms and the seminorms are considered on the vectorial counterpart of \(H^s(\mathcal{G}_h)\). We use the convention \(H^0(\mathcal{T}_h)= L^2(\mathcal{T}_h)\) for all the spaces defined previously. We will also need the spaces given on the skeletons of the triangulations \(\mathcal{T}_h\) and \(\mathcal{G}_h\) by \[L^2(\mathcal{F}_h):= \prod_{F\in \mathcal{F}_h} L^2(F), \qquad L^2(\mathcal{F}_h^0):= \prod_{F\in \mathcal{F}_h^0} L^2(F), \qquad L^2(\mathcal{E}_h):= \prod_{e\in \mathcal{E}_h} L^2(e)\] \[\mathbf{L}^2(\mathcal{F}_h):= \prod_{F\in \mathcal{F}_h} L^2(F)^3, \qquad \mathbf{L}^2(\mathcal{F}_h^0):= \prod_{F\in \mathcal{F}_h^0} L^2(F)^3, \qquad \mathbf{L}^2(\mathcal{E}_h):= \prod_{e\in \mathcal{E}_h} L^2(e)^3.\] Similarly, the components \(\mu_F\) and \(\boldsymbol{\beta}_F\) of \(\mu := \set{\mu_F}\in L^2(\mathcal{F}_h)\) and \(\boldsymbol{\beta}:= \set{\boldsymbol{\beta}_F} \in \mathbf{L}^2(\mathcal{F}_h)\) coincide with the restrictions \(\mu|_F\) and \(\boldsymbol{\beta}|_F\) and the components \(\varphi_e\) and \(\boldsymbol{\psi}_e\) of \(\varphi := \set{\varphi_e}\in L^2(\mathcal{E}_h)\) and \(\boldsymbol{\psi}:=\set{\boldsymbol{\psi}_e}\in \mathbf{L}^2(\mathcal{E}_h)\) are given by the restrictions \(\varphi|_e\) and \(\boldsymbol{\psi}|_e\) respectively. We introduce the inner products \[\dual{\lambda, \mu}_{\mathcal{F}_h} := \sum_{F\in \mathcal{F}_h} \dual{\lambda_F, \mu_F}_F, \quad \dual{\lambda, \mu}_{\mathcal{F}^0_h} := \sum_{F\in \mathcal{F}^0_h} \dual{\lambda_F, \mu_F}_F \quad \text{and} \quad \dual{\psi, \varphi}_{\mathcal{E}_h} := \sum_{e\in \mathcal{E}_h} \dual{\psi_e, \varphi_e}_e.\] and the corresponding norms \[\norm{\mu}^2_{0,\mathcal{F}_h}:= \dual{\mu, \mu}_{\mathcal{F}_h}, \qquad \norm{\mu}^2_{0,\mathcal{F}^0_h}:= \dual{\mu, \mu}_{\mathcal{F}^0_h}\quad \text{and} \quad \norm{\varphi}^2_{0,\mathcal{E}_h}:= \dual{\varphi, \varphi}_{\mathcal{E}_h}\] on \(L^2(\mathcal{F}_h)\), \(L^2(\mathcal{F}^0_h)\) and \(L^2(\mathcal{E}_h)\) respectively. Given \(v\in H^1(\mathcal{T}_h)\), we define averages \(\mean{v}\in L^2(\mathcal{F}_h^0)\) and jumps \(\jump{v}\in \mathbf{L}^2(\mathcal{F}_h^0)\) by \[\mean{v}_F := 1/2(v_K + v_{K'}) \quad \text{and} \quad \jump{v}_F := v_K \boldsymbol{n}_K + v_{K'}\boldsymbol{n}_{K'} \quad \forall F \in \mathcal{F}(K)\cap \mathcal{F}(K').\] For vector valued functions \(\boldsymbol{\tau}\in \mathbf{H}^1(\mathcal{T}_h)\), we define \(\mean{\boldsymbol{\tau}}\in \mathbf{L}^2(\mathcal{F}_h^0)\) and \(\jump{\boldsymbol{\tau}}\in L^2(\mathcal{F}_h^0)\) by \[\mean{\boldsymbol{\tau}}_F := 1/2(\boldsymbol{\tau}_K + \boldsymbol{\tau}_{K'}) \quad \text{and} \quad \jump{\boldsymbol{\tau}}_F := \boldsymbol{\tau}_K\cdot \boldsymbol{n}_K + \boldsymbol{\tau}_{K'}\cdot\boldsymbol{n}_{K'} \quad \forall F \in \mathcal{F}(K)\cap \mathcal{F}(K').\] Similarly, given \(\varphi\in H^1(\mathcal{G}_h)\), we define averages \(\mean{\varphi}\in L^2(\mathcal{E}_h^0)\) and jumps \(\jump{\varphi}\in \mathbf{L}^2(\mathcal{E}_h)\) by \[\mean{\varphi}_e := 1/2(\varphi_T + \varphi_{T'}) \quad \text{and} \quad \jump{\varphi}_e := \varphi_T \boldsymbol{t}_e + \varphi_{T'} \boldsymbol{t}_{e'} \quad \forall e \in \mathcal{E}(T)\cap \mathcal{E}(T').\] Here, \(\boldsymbol{t}_e\) is the tangent unit vector along the edge \(e\) given by \(\boldsymbol{t}_e = (\boldsymbol{n} \times \boldsymbol{n}_{\partial F})|_e\), where \(\boldsymbol{n}_{\partial F}\) is the outward unit normal vector to the boundary of the face \(F\) in the hyperplane defined by \(\boldsymbol{n}|_F\). Hereafter, given an integer \(k\geq 0\) and a domain \(D\subset \mathbb{R}^3\), \(\mathcal{P}_k(D)\) denotes the space of polynomials of degree at most \(k\) on \(D\). We consider the linear spaces \[\mathcal{P}_0(\mathcal{F}_h):= \prod_{F\in \mathcal{F}_h} \mathcal{P}_0(F) \quad \text{and} \quad \mathcal{P}_0( \mathcal{F}_h^0):= \prod_{F\in \mathcal{F}_h^0} \mathcal{P}_0(F),\] and for any \(m\geq 1\), we introduce the finite element spaces \[V_h := \prod_{K\in \mathcal{T}_h} \mathcal{P}_m(K) \quad \text{and} \quad \boldsymbol{\Sigma}_h := \prod_{K\in \mathcal{T}_h} \mathbf{RT}_m(K),\] where \[\mathbf{RT}_m(K):=\set{\mathcal{P}_{m-1}(K)^3 + \boldsymbol{x}\mathcal{P}_{m-1}(K)}\] is the finite element of Raviart-Thomas of order \(m-1\). We consider the following formulation in the bounded domain \(\Omega\): find \((\boldsymbol{\sigma}_h, u_h) \in \boldsymbol{\Sigma}_h\times V_h\) such that for each \(K\in \mathcal{T}_h\) there holds \[\label{formT1} \begin{array}{rcll} (\boldsymbol{\sigma}_h, \boldsymbol{\tau})_K-(\nabla u_h, \boldsymbol{\tau})_K + \dual{u_h-\bar u, \boldsymbol{\tau} \cdot \boldsymbol{n}_K}_{\partial K} &=& 0 & \forall \boldsymbol{\tau} \in \boldsymbol{\Sigma}_h\\[2ex] \displaystyle (\boldsymbol{\sigma}_h, \nabla v)_K-\dual{ \bar \boldsymbol{\sigma} \cdot \boldsymbol{n}_K, v}_{\partial K} &=& (f, v)_K & \forall v \in V_h. \end{array}\] Before defining the numerical traces \(\bar u\) and \(\bar \boldsymbol{\sigma}\) let us consider the finite element approximation of the boundary integral equations [\[inteq1\]](#inteq1){reference-type="eqref" reference="inteq1"} and [\[inteq2\]](#inteq2){reference-type="eqref" reference="inteq2"}. We consider the operator \(\boldsymbol T\) defined for any \(\varphi \in H^1(\mathcal{G}_h)\) by \[(\boldsymbol T\varphi)|_e := (V \mathbf{curl}_h \varphi)|_e \quad (e\in \mathcal{E}_h),\] where \(\mathbf{curl}_h\) stands for the element-wise \(\mathbf{curl}\) operator: \[(\mathbf{curl}_h \varphi)|_F := \mathbf{curl}_F(\varphi|_F), \qquad \forall F\in \mathcal{G}_h.\] We consider two sequences of boundary element spaces \[\begin{aligned} \Lambda_h &:= \set{\boldsymbol{\tau}\cdot \boldsymbol{n};\quad \boldsymbol{\tau} \in \boldsymbol{\Sigma}_h } \subset H^{-1/2}(\Gamma),\\ \Psi_h &:= \prod_{T\in \mathcal{G}_h} \mathcal{P}_m(T) \cap L^{2}_0(\Gamma) \subset H_0^{1/2}(\mathcal{G}_h):=H^{1/2}(\mathcal{G}_h)\cap L^2_0(\Gamma) \end{aligned}\] with \(L^{2}_0(\Gamma):=\set{\varphi\in L^{2}(\Gamma); \quad \dual{1,\varphi}_\Gamma = 0}\). We then replace [\[inteq1\]](#inteq1){reference-type="eqref" reference="inteq1"}, [\[inteq2\]](#inteq2){reference-type="eqref" reference="inteq2"} by the Galerkin equations: find \(\psi_h\in \Psi_h\), \(\lambda^\star_h\in \Lambda_h\) such that \[\label{formB} \begin{array}{rcll} \dual{ \psi_h, \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma} &=& \dual{( \frac{\text{id}}{2} + K) \psi_h, \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma}- \dual{V (\lambda^\star_h-g_1), \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma} &\forall \boldsymbol{\tau} \in \boldsymbol{\Sigma}_h\\[2ex] \dual{(\lambda^\star_h-g_1), \varphi}_{\Gamma} &=&-d(\psi_h,\varphi) + \dual{(\frac{\text{id}}{2}-K') (\lambda^\star_h-g_1), \varphi}_{\Gamma} &\forall \varphi \in \Psi_h. \end{array}\] Here we used the transmission condition for \(\lambda\), [\[trans\]](#trans){reference-type="eqref" reference="trans"}, and \(\lambda^\star_h\) will be an approximation to \(\lambda+g_1=\boldsymbol{\sigma}\cdot\boldsymbol{n}\). Furthermore, \[\label{d} d(\psi, \varphi) := \dual{V\mathbf{curl}_h \psi, \mathbf{curl}_h\varphi }_\Gamma + \dual{\boldsymbol T\psi, \jump{\varphi}}_{\mathcal{E}_h} -\dual{\jump{\psi}, \boldsymbol T \varphi}_{\mathcal{E}_h} + \dual{\nu\jump{\psi}, \jump{\varphi}}_{\mathcal{E}_h}\] and \(\nu\in \prod_{e\in \mathcal{E}_h} \mathcal{P}_0(e)\) is a piecewise constant function such that \[\label{nu} \nu\simeq 1.\] Let \(\alpha\in \mathcal{P}_0(\mathcal{F}_h)\), and \(\boldsymbol{\beta} \in \mathcal{P}_0( \mathcal{F}_h^0)^3\) be given piecewise constant functions satisfying \[\label{coef} \max_{F\in \mathcal{F}_h^0} |\boldsymbol{\beta}_F| \lesssim 1 \quad \text{and} \quad h_\mathcal{F}\, \alpha \simeq 1,\] where \(h_\mathcal{F}\in \mathcal{P}_0(\mathcal{F}_h)\) is defined by \(h_\mathcal{F}|_F := h_F\) \(, \forall F \in \mathcal{F}_h\). We substitute \(\bar u\) and \(\bar \boldsymbol{\sigma}\) given by \[\bar u_F = \begin{cases} \mean{u_h}_F + \boldsymbol{\beta}_F\cdot \jump{u_h}_F &\text{if \(F\in \mathcal{F}_h^0\)}\\[2ex] \psi_h+g_0 &\text{if \(F\in \mathcal{F}_h^{\partial}\)} \end{cases}\] and \[\bar \boldsymbol{\sigma}_F = \begin{cases} \mean{\boldsymbol{\sigma}_h}_F-\jump{\boldsymbol{\sigma}_h}_F \boldsymbol{\beta}_F-\alpha_F \jump{u_h}_F &\text{if \(F\in \mathcal{F}_h^0\)}\\[2ex] \boldsymbol{\sigma}_h|_F-\alpha_F (u_h|_F-\psi_h-g_0) \boldsymbol{n}|_F &\text{if \(F\in \mathcal{F}_h^{\partial}\)} \end{cases}\] in [\[formT1\]](#formT1){reference-type="eqref" reference="formT1"} and add the equations over \(K\in \mathcal{T}_h\) to obtain the following LDG formulation of the problem in \(\Omega\): find \((\boldsymbol{\sigma}_h,u_h)\in \boldsymbol{\Sigma}_h\times V_h\) such that \[\label{ldgOmega} \begin{array}{rcll} (\boldsymbol{\sigma}_h, \boldsymbol{\tau})_{\Omega}-\left\{ (\nabla_h u_h, \boldsymbol{\tau})_{\Omega}-S(u_h, \boldsymbol{\tau}) \right\} -\dual{\psi_h, \boldsymbol{\tau}\cdot \boldsymbol{n}}_\Gamma &=& \dual{g_0, \boldsymbol{\tau}\cdot \boldsymbol{n}}_\Gamma \\[2ex] \left\{(\nabla_h v, \boldsymbol{\sigma}_h)_{\Omega}-S(v, \boldsymbol{\sigma}_h)\right\} + \boldsymbol{\alpha}_0(u_h, v) +\dual{\alpha (u_h-\psi_h), v}_\Gamma &=& (f,v)_{\Omega} + \dual{\alpha g_0, v}_\Gamma, \end{array}\] for all \(\boldsymbol{\tau} \in \boldsymbol{\Sigma}_h\) and \(v\in V_h\), where \(\nabla_h\) stands for the element-wise gradient and \[S(u, \boldsymbol{\tau}) := \dual{\jump{u}, \mean{\boldsymbol{\tau}}-\jump{\boldsymbol{\tau}}\boldsymbol{\beta}}_{\mathcal{F}_h^0} + \dual{u, \boldsymbol{\tau}\cdot \boldsymbol{n}}_\Gamma, \qquad \forall u\in H^1(\mathcal{T}_h),\, \forall \boldsymbol{\tau} \in \mathbf{H}^1(\mathcal{T}_h),\] \[\boldsymbol{\alpha}_0(u,v) = \dual{ \alpha \jump{u}, \jump{v}}_{\mathcal{F}_h^0} \qquad \forall u,v\in H^1(\mathcal{T}_h),\] In order to simplify the notations, let us denote by \(\hat{u}_h=(u_h,\psi_h)\) and \(\hat{v}:=(v,\varphi)\) couples of elements from \(V_h \times \Psi_h\). We also consider \[\jump{\hat{u}_h}: = \begin{cases} \jump{u_h}_F & \text{if \(F\in \mathcal{F}_h^0\)}\\ (u_h-\psi_h)\boldsymbol{n}|_F & \text{if \(F\in \mathcal{F}_h^{\partial}\)} \end{cases} \quad \text{and} \quad \jump{\hat{v}}: = \begin{cases} \jump{v}_F & \text{if \(F\in \mathcal{F}_h^0\)}\\ (v-\varphi)\boldsymbol{n}|_F & \text{if \(F\in \mathcal{F}_h^{\partial}\)} \end{cases}.\] We now couple [\[formB\]](#formB){reference-type="eqref" reference="formB"} and [\[ldgOmega\]](#ldgOmega){reference-type="eqref" reference="ldgOmega"} by identifying \(\lambda^\star_h=\boldsymbol{\sigma}_h\cdot\boldsymbol{n}\) and by approximating the transmission condition for the traces in [\[trans\]](#trans){reference-type="eqref" reference="trans"} by \[\dual{\alpha (u_h-\psi_h), \varphi}_\Gamma = \dual{\alpha g_0, \varphi}_\Gamma \quad \forall \varphi \in \Psi_h.\] A combination of [\[formB\]](#formB){reference-type="eqref" reference="formB"} and [\[ldgOmega\]](#ldgOmega){reference-type="eqref" reference="ldgOmega"} then yields our LDG-FEM/DG-BEM coupling: find \((\boldsymbol{\sigma}_h, \hat{u}_h)\in \boldsymbol{\Sigma}_h\times (V_h \times \Psi_h)\) such that \[\label{ldg-FemBem} \begin{array}{rcll} a(\boldsymbol{\sigma}_h,\boldsymbol{\tau}) + b(\boldsymbol{\tau}, \hat{u}_h) &=& \dual{g_0, \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma} + \dual{Vg_1, \boldsymbol{\tau} \cdot \boldsymbol{n}}_{\Gamma} &\forall \boldsymbol{\tau}\in \boldsymbol{\Sigma}_h\\[2ex] -b(\boldsymbol{\sigma}_h, \hat{v})+ c(\hat{u}_h, \hat{v}) &=& (f,v)_{\Omega} + \dual{\alpha g_0, v-\varphi}_{\Gamma} +\dual{( \frac{\text{id}}{2} + K')g_1, \varphi}_{\Gamma} & \forall \hat{v}\in V_h\times \Psi_h. \end{array}\] Here, \[a(\boldsymbol{\sigma}_h,\boldsymbol{\tau}):= (\boldsymbol{\sigma}_h, \boldsymbol{\tau})_{\Omega} + \dual{\boldsymbol{\tau}\cdot \boldsymbol{n}, V (\boldsymbol{\sigma}_h\cdot \boldsymbol{n})}_{\Gamma}, \quad c(\hat{u}_h, \hat{v}) := \dual{\alpha \jump{\hat{u}_h},\jump{\hat{v}}}_{\mathcal{F}_h} + d(\psi_h, \varphi)\] and \[b(\boldsymbol{\tau}, \hat{v}) :=-(\nabla_h v, \boldsymbol{\tau})_{\Omega} + \dual{\boldsymbol{\tau}\cdot \boldsymbol{n}, ( \frac{\text{id}}{2}-K) \varphi}_{\Gamma} +\dual{\jump{v}, \mean{\boldsymbol{\tau}}-\jump{\boldsymbol{\tau}}\boldsymbol{\beta}}_{\mathcal{F}_h^0} + \dual{ \jump{\hat{v}}, \boldsymbol{\tau}}_{\Gamma}.\] Problem [\[ldg-FemBem\]](#ldg-FemBem){reference-type="eqref" reference="ldg-FemBem"} can be rewritten in the more compact form as follows: Find \(\boldsymbol{\sigma}_h\in \boldsymbol{\Sigma}_h\) and \(\hat{u}_h=(u_h, \psi_h)\in V_h\times \Psi_h\) such that \[\label{compactForm} A(\boldsymbol{\sigma}_h, \hat{u}_h; \boldsymbol{\tau}, \hat{v}) = F(\boldsymbol{\tau}, \hat{v}),\] by setting \[\label{A} A(\boldsymbol{\sigma}_h, \hat{u}_h; \boldsymbol{\tau}, \hat{v}):= a(\boldsymbol{\sigma}_h,\boldsymbol{\tau}) + b(\boldsymbol{\tau}, \hat{u}_h)-b(\boldsymbol{\sigma}_h, \hat{v}) + c(\hat{u}_h, \hat{v})\] and \[F(\boldsymbol{\tau}, \hat{v}):= (f,v)_{\Omega}+ \dual{Vg_1 + g_0, \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma} + \dual{\alpha g_0, v-\varphi}_{\Gamma} +\dual{( \frac{\text{id}}{2} + K')g_1, \varphi}_{\Gamma}.\] We end this section by proving that our LDG-FEM/DG-BEM scheme is consistent. # Convergence analysis {#s4} In this section, we develop the error analysis of the LDG-FEM/DG-BEM scheme [\[ldg-FemBem\]](#ldg-FemBem){reference-type="eqref" reference="ldg-FemBem"}. We first introduce a series of technical results that are used in the proof the Céa's error estimate provided by Theorem [\[cea\]](#cea){reference-type="ref" reference="cea"}). Then, we use well-known interpolation error estimates to obtain the main convergence result stated in Theorem [\[main\]](#main){reference-type="ref" reference="main"}. ## Technical results The following discrete trace inequality is standard,. The \(H^{1/2}(\mathcal{G}_h)\)-ellipticity of the bilinear form \(\dual{V\mathbf{curl}_h \psi, \mathbf{curl}_h\varphi }_\Gamma\) in \(\Psi_h\) is essential for the stability of our method. The main difficulty that we had to deal with in our analysis is that this bilinear form is not uniformly bounded on \(\Psi_h\) with respect to this broken-norm. The following estimate is a Poincaré-Friedrichs inequality for piecewise polynomial functions. Finally, the following bound for \(\boldsymbol T\) can be found in. ## Stability of the LDG-FEM/DG-BEM method For all \(K\in \mathcal{T}_h\), we introduce the \(L^2(K)\)-orthogonal projector \(\Pi_K\) onto \(\mathcal{P}_m(K)\). Moreover, we consider on each \(T\in \mathcal{G}_h\) the usual triangular Lagrange finite element of order \(m\) \((m\geq 1)\) and denote by \(\tilde\pi_T:\, \mathcal{C}^0(T)\to \mathcal{P}_m(T)\) the corresponding Lagrange interpolation operator. We will also use the Raviart-Thomas interpolation operator \(\boldsymbol{\Pi}_K\) in \(\mathbf{RT}_m(K)\), see. The global operators \(\Pi:\, L^2(\mathcal{T}_h) \to V_h\), \(\boldsymbol{\Pi}:\, \mathbf{H}^1(\mathcal{T}_h) \to \boldsymbol{\Sigma}_h\) and \(\tilde\pi:\, \mathcal{C}^0(\Gamma) \to \Psi_h\cap \mathcal{C}^0(\Gamma)\) are given by \[(\Pi v )|_K := \Pi_K (v_K), \quad (\boldsymbol{\Pi}\boldsymbol{\tau} )|_K := \boldsymbol{\Pi}_K (\boldsymbol{\tau}_K)\,\, \forall K \in \mathcal{T}_h \quad \text{and}\quad (\tilde\pi \varphi)|_T := \tilde\pi_T (\varphi_T)\,\, \forall T\in \mathcal{G}_h\] respectively. For all \(\boldsymbol{\tau} \in \mathbf{H}^1(\mathcal{T}_h)\) and \(\hat{v}:= (v,\varphi)\in H^1(\mathcal{T}_h)\times H^{1}(\mathcal{G}_h)\), we introduce the semi-norms \[\norm{(\boldsymbol{\tau}, \hat{v})} := \left(\norm{\boldsymbol{\tau}}^2_{0,\Omega} + \norm{\boldsymbol{\tau}\cdot\boldsymbol{n}}^2_{-1/2,\Gamma} + |\varphi|^2_{1/2,\mathcal{G}_h} + \norm{\alpha^{1/2} \jump{\hat{v}}}^2_{0,\mathcal{F}_h} + \norm{\nu^{1/2}\jump{\varphi}}^2_{0, \mathcal{E}_h}\right)^{1/2},\] \[\begin{gathered} \norm{(\boldsymbol{\tau}, \hat{v})}_{\#} := \Big(\norm{\boldsymbol{\tau}}^2_{0,\Omega} + \norm{\boldsymbol{\tau}\cdot\boldsymbol{n}}^2_{-1/2,\Gamma} + \norm{\alpha^{1/2} \jump{\hat{v}}}^2_{0,\mathcal{F}_h} + \\[1ex] \norm{\mathbf{curl} _h\varphi}^2_{-1/2,\Gamma}+ \norm{\nu^{1/2}\jump{\varphi}}^2_{0, \mathcal{E}_h}\Big)^{1/2} \end{gathered}\] and for all \(\boldsymbol{\tau} \in \mathbf{H}^1(\mathcal{T}_h)\) and \(\hat{v}:= (v,\varphi)\in H^1(\mathcal{T}_h)\times H^{1}(\Gamma)\), we introduce \[\begin{gathered} \norm{(\boldsymbol{\tau}, \hat{v})}_*:= \Big(\norm{(\boldsymbol{\tau}, \hat{v})}^2+ \sum_{K\in \mathcal{T}_h}\norm{\alpha^{-1/2} \boldsymbol{\tau}\cdot \boldsymbol{n}_K}^2_{0,\partial K}+ \norm{ \boldsymbol{\tau}\cdot \boldsymbol{n}}^2_{0,\Gamma}+ \sum_{K\in \mathcal{T}_h} \norm{\alpha^{1/2} v}^2_{0,\partial K} +\\[1ex] \norm{\varphi}^2_{1/2,\Gamma}+ h^{-1} \norm{\varphi}^2_{0,\Gamma}+h |\varphi|^2_{1,\Gamma}\Big)^{1/2}. \end{gathered}\] It is clear that \[\label{rn1} \norm{(\boldsymbol{\tau}, \hat{v})} \leq \norm{(\boldsymbol{\tau}, \hat{v})}_*\quad \forall (\boldsymbol{\tau}, \hat{v}) \in \mathbf{H}^1(\mathcal{T}_h)\times ( H^1(\mathcal{T}_h)\times H^{1}(\Gamma) ).\] Moreover, taking into account [\[eqnorm\]](#eqnorm){reference-type="eqref" reference="eqnorm"}, we deduce that \[\label{rn2} \norm{(\boldsymbol{\tau}, \hat{v})} \lesssim \norm{(\boldsymbol{\tau}, \hat{v})}_{\#} \quad \forall (\boldsymbol{\tau}, \hat{v}) \in \mathbf{H}^1(\mathcal{T}_h)\times ( H^1(\mathcal{T}_h)\times H^{1}(\mathcal{G}_h)).\] In the following we abbreviate \[\pi_{\boldsymbol{\sigma}} := \boldsymbol{\sigma}-\boldsymbol{\Pi}\boldsymbol{\sigma}, \quad \pi_u := u-\Pi u,\quad \tilde\pi_\psi := \psi-\tilde\pi\psi \quad\text{and}\quad \pi_{\hat{u}} := (u-\Pi u, \psi-\tilde\pi \psi).\] Let us introduce the errors \[e_{\boldsymbol{\sigma}} := \boldsymbol{\sigma}-\boldsymbol{\sigma}_h, \quad e_u := u-u_h,\quad e_\psi := \psi-\psi_h \quad \text{and} \quad e_{\hat{u}} := (u-u_h, \psi-\psi_h).\] We notice that, under the regularity hypothesis of Proposition [\[consistency0\]](#consistency0){reference-type="ref" reference="consistency0"}, we have the following Galerkin orthogonality \[\label{consistency} A(e_{\boldsymbol{\sigma}},e_{\hat{u}}; \boldsymbol{\tau}, \hat{v} ) = 0 \quad\forall (\boldsymbol{\tau}, \hat{v})\in \boldsymbol{\Sigma}_h \times (V_h\times \Psi_h).\] ## Asymptotic error estimates In this section we need to handle functions that are piecewise smooth on the boundary \(\Gamma\) of the polyhedron \(\Omega\). Let \(\set{\Gamma_1, \cdots, \Gamma_N}\) be the open polygons, contained in different hyperplanes of \(\mathbb{R}^3\), such that \(\Gamma= \cup_{j= 1}^N \overline{\Gamma}_j\). For any \(t\geq 0\), we consider the broken Sobolev space \(H^t_{\mathfrak{b}}(\Gamma):= \prod_j H^t(\Gamma_j)\) endowed with the graph norm \[\norm{\varphi}^2_{t,\mathfrak{b},\Gamma}: = \sum_{j= 1}^N \norm{\varphi}^2_{H^t(\Gamma_j)}.\] Let us recall some well-known approximation properties related with the (local and global) projection and interpolation operators. # Conforming approximation on the boundary {#s5} With little more effort we can provide the convergence analysis for a Galerkin scheme based on a conforming BEM-approximation. To this end, we introduce \[\begin{aligned} \widetilde \Psi_h &:= \prod_{T\in \mathcal{G}_h} \mathcal{P}_m(T) \cap H^{1/2}_0(\Gamma) \subset H_0^{1/2}(\Gamma):=H^{1/2}(\Gamma)\cap L^2_0(\Gamma) \end{aligned}\] and consider the problem: find \((\boldsymbol{\sigma}_h, \hat{u}_h)\in \boldsymbol{\Sigma}_h\times (V_h \times \widetilde \Psi_h)\) such that \[\label{ldg-FemConfBem} \begin{array}{rcll} a(\boldsymbol{\sigma}_h,\boldsymbol{\tau}) + b(\boldsymbol{\tau}, \hat{u}_h) &=& \dual{g_0, \boldsymbol{\tau}\cdot \boldsymbol{n}}_{\Gamma} + \dual{Vg_1, \boldsymbol{\tau} \cdot \boldsymbol{n}}_{\Gamma} &\forall \boldsymbol{\tau}\in \boldsymbol{\Sigma}_h\\[2ex] -b(\boldsymbol{\sigma}_h, \hat{v})+ c(\hat{u}_h, \hat{v}) &=& (f,v)_{\Omega} + \dual{\alpha g_0, v-\varphi}_{\Gamma} +\dual{( \frac{\text{id}}{2} + K')g_1, \varphi}_{\Gamma} & \forall \hat{v}\in V_h\times \widetilde\Psi_h. \end{array}\] Note that the restriction of the bilinear form \(d(\cdot, \cdot)\), used in the definition of \(c(\cdot, \cdot)\) and introduced in [\[d\]](#d){reference-type="eqref" reference="d"}, reduces to \[d(\psi, \varphi) := \dual{V\mathbf{curl}_\Gamma \psi, \mathbf{curl}_\Gamma\varphi }_\Gamma\] for functions \(\psi\) and \(\varphi\) in \(\widetilde\Psi_h\). This will simplify considerably the analysis of the scheme. All the other bilinear forms in [\[ldg-FemConfBem\]](#ldg-FemConfBem){reference-type="eqref" reference="ldg-FemConfBem"} remain unchanged. Apart from the fact that the shape regular conforming triangulation \(\mathcal{G}_h:=\set{T}\) is no longer needed to be quasi-uniform, in the sequel, we will use the same hypothesis on the triangulations and we will also use the same notations introduced in previous sections. The well-posedness and the consistency of the scheme [\[ldg-FemConfBem\]](#ldg-FemConfBem){reference-type="eqref" reference="ldg-FemConfBem"} follow by straightforward simplifications of the arguments used in the proofs of Propositions [\[wellposed\]](#wellposed){reference-type="ref" reference="wellposed"} and [\[consistency0\]](#consistency0){reference-type="ref" reference="consistency0"}. Reexamining carefully the proof of Lemma [\[boundA\]](#boundA){reference-type="ref" reference="boundA"} we obtain the following stability property for scheme [\[ldg-FemConfBem\]](#ldg-FemConfBem){reference-type="eqref" reference="ldg-FemConfBem"}. It is straightforward to deduce from [\[eq-coercive-1\]](#eq-coercive-1){reference-type="eqref" reference="eq-coercive-1"} and [\[eq-coercive-2\]](#eq-coercive-2){reference-type="eqref" reference="eq-coercive-2"} that \[\label{coerConf} A(\boldsymbol{\tau}, \hat{v};\boldsymbol{\tau}, \hat{v}) \gtrsim (\norm{(\boldsymbol{\tau}, \hat{v})}^c)^2 \quad \forall \boldsymbol{\tau}\in \mathbf{H}^1(\mathcal{T}_h), \quad \forall \hat{v}=(v,\varphi) \in H^1(\mathcal{T}_h)\times H^{1/2}_0(\Gamma).\] Combining [\[coerConf\]](#coerConf){reference-type="eqref" reference="coerConf"} with Lemma [\[boundAconf\]](#boundAconf){reference-type="ref" reference="boundAconf"} yields the following error estimate. In the conforming BEM case, we can also provide an estimate of the error \(u-u_h\) in the \(L^2(\Omega)\)-norm. To this end, we follow and use a duality argument. For any \(\rho \in L^2(\Omega)\) we consider the exterior problem \[\label{auxiliary} \begin{array}{rcll} \Delta w &=& \tilde\rho &\text{in \(\mathbb{R}^3\)},\\[2ex] w &=& O(\displaystyle\frac{1}{\abs{\boldsymbol{x}}}) &\text{as \(\abs{\boldsymbol{x}}\to \infty\)}, \end{array}\] where \(\tilde \rho\) is the extension by zero of \(\rho\) outside \(\Omega\). It is well known from the theory of regularity of elliptic problems that \(w\in H^2(\Omega)\) and there exists \(C_{reg}>0\) such that \[\label{reg} \norm{w}_{2,\Omega}\leq C_{reg} \norm{\rho}_{0,\Omega}.\] # Numerical results {#s6} In this section we present a numerical experiment confirming the theoretical error estimate obtained for the LDG-FEM/DG-BEM scheme [\[ldg-FemBem\]](#ldg-FemBem){reference-type="eqref" reference="ldg-FemBem"} and the LDG-FEM/BEM scheme [\[ldg-FemConfBem\]](#ldg-FemConfBem){reference-type="eqref" reference="ldg-FemConfBem"}. For simplicity we consider our model problem in two dimensions. The corresponding theory and results from three dimensions apply with trivial modifications. We choose \(\Omega=(0,1)^2\) and select the data so that the exact solution is given by \[u(x_1, x_2)=\sin(10 x_1+3 x_2)\quad \text{in \(\Omega\)}\quad \text{and}\quad u^e(x_1, x_2)=\frac{x_1+x_2-1}{(x_1-0.5)^2+(x_2-0.5)^2}\quad \text{in \(\Omega_e\)}.\] We consider uniform triangular meshes \(\mathcal{T}_h\) on \(\Omega\) and inherited meshes \(\mathcal{G}_h\) on \(\Gamma\) (and for simplicity denote \(h\) to be the length of the shortest edge). Lowest order discrete spaces are taken, i.e. \(m=1\), so that \[V_h = \prod_{K\in \mathcal{T}_h} \mathcal{P}_1(K), \quad \boldsymbol{\Sigma}_h = \prod_{K\in \mathcal{T}_h} \mathbf{RT}_1(K),\] \[\Psi_h= \set{\varphi \in L_0^{2}(\Gamma); \quad \varphi|_T\in P_1(T)\quad \forall T\in \mathcal{G}_h}\] and \[\tilde\Psi_h= \set{\varphi \in H_0^{1/2}(\Gamma); \quad \varphi|_T\in P_1(T)\quad \forall T\in \mathcal{G}_h}.\] Moreover, we select \(\boldsymbol{\beta}\) to be normal on the interior edges (in a certain direction) with \(|\boldsymbol{\beta}|=1\), and \(\alpha=h_{\mathcal{F}}^{-1}\). In this case Theorem [\[main\]](#main){reference-type="ref" reference="main"} proves the behaviors \(\norm{\psi-\psi_h}_{1/2,\mathcal{G}_h}= O(h)\), \(\norm{\boldsymbol{\sigma}-\boldsymbol{\sigma}_h}_{0,\Omega} + \norm{(\boldsymbol{\sigma}-\boldsymbol{\sigma}_h)\cdot\boldsymbol{n}}_{-1/2,\Gamma}=O(h)\) and \(\norm{\jump{\hat u_h}}_{0,\mathcal{F}_h} = O(h^{3/2})\). In Figure [\[fig\]](#fig){reference-type="ref" reference="fig"} the errors \(\norm{\psi-\psi_h}_{[0,1],1/2,\mathcal{G}_h}\), \(\norm{\boldsymbol{\sigma}-\boldsymbol{\sigma}_h}_{0,\Omega}\) and \(\norm{\jump{\hat u_h}}_{0,\mathcal{F}_h}\) are labeled , and respectively and they are depicted versus the total number of unknowns on a double-logarithmic scale. Here, \[\norm{\psi-\psi_h}_{[0,1],1/2,\mathcal{G}_h} := \Bigl(\norm{\psi-\psi_h}_{0,\Gamma}^2 + \sum_{T\in\mathcal{G}_h} \|\psi-\psi_h\|_{0,T} |\psi-\psi_h|_{1,T}\Bigr)^{1/2}\] which, by interpolation, is an upper bound for \(\norm{\psi-\psi_h}_{1/2,\mathcal{G}_h}\) up to a constant factor. The curves \(h\) and \(h^{3/2}\) are also given multiplied by appropriate factors to shift them closer to the corresponding curves. The numerical experiment confirms the convergence rates \(\norm{\boldsymbol{\sigma}-\boldsymbol{\sigma}_h}_{0,\Omega}=O(h)\), \(\norm{\jump{\hat u_h}}_{0,\mathcal{F}_h}=O(h^{3/2})\) and suggests the stronger convergence \(\norm{\psi-\psi_h}_{1/2,\mathcal{G}_h}=O(h^{3/2})\). In Figure [\[fig0\]](#fig0){reference-type="ref" reference="fig0"}, the errors \(\norm{\psi-\psi_h}_{[0,1],1/2,\Gamma}\), \(\norm{\boldsymbol{\sigma}-\boldsymbol{\sigma}_h}_{0,\Omega}\), \(\norm{u-u_h}_{0,\Omega}\) and \(\norm{\jump{\hat u_h}}_{0,\mathcal{F}_h}\) are labeled , , and respectively and they are represented again versus the total number of unknowns on a double-logarithmic scale. Here, \[\norm{\psi-\psi_h}_{[0,1],1/2,\Gamma} := \Bigl(\norm{\psi-\psi_h}_{0,\Gamma}^2 + \|\psi-\psi_h\|_{0,\Gamma} |\psi-\psi_h|_{1,\Gamma}\Bigr)^{1/2}\] which, by interpolation, is an upper bound for \(\norm{\psi-\psi_h}_{1/2,\Gamma}\) up to a constant factor. The numerical results are in agreement with the convergence rates \(\norm{\boldsymbol{\sigma}-\boldsymbol{\sigma}_h}_{0,\Omega}=O(h)\), \(\norm{u-u_h}_{0,\Omega}=O(h^2)\) and \(\norm{\jump{\hat u_h}}_{0,\mathcal{F}_h}=O(h^{3/2})\) obtained in Theorem [\[mainConf\]](#mainConf){reference-type="ref" reference="mainConf"}, and indicate the stronger convergence \(\norm{\psi-\psi_h}_{1/2,\Gamma}=O(h^{3/2})\).

{'timestamp': '2013-10-14T02:09:50', 'yymm': '1310', 'arxiv_id': '1310.3201', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3201'}

# Introduction, Results and Synopsis {#s:IRS}

{'timestamp': '2013-10-24T02:00:42', 'yymm': '1310', 'arxiv_id': '1310.3256', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3256'}

# Introduction Neutrinoless double-\(\beta\) decay (\(0\nu\beta\beta\)), if observed, would prove that neutrinos are Majorana fermions, an important milestone in the search for physics beyond the Standard Model. In addition, one could extract more information about the nature of the decay mechanism and possibly determine the light neutrino mass hierarchy and the lightest neutrino mass, provided that the associated nuclear matrix elements (NME) are calculated with good accuracy. There are many possible mechanisms that could contribute to the \(0\nu\beta\beta\) decay process, and some of the associated matrix elements were investigated by using several approaches, including the quasiparticle random phase approximation (QRPA), the interacting shell model, the interacting boson model, the generator coordinate method, and the projected Hartree-Fock Bogolibov model. With the exception of the QRPA, all other methods entail using the closure approximation . Some older and more recent analyses suggest that the deviation of the NME for the light neutrino-exchange mechanism from the closure approximation result should be small, but a full analysis of this deviation within the shell model is not yet available. In addition, the QRPA analysis is affected by uncertainties due to the \(g_{pp}\) factor used to tune the residual interaction. For example, results from Ref. indicate a deviation of about up to 10% between closure and nonclosure NME, but its magnitude and sign depend on the choice of \(g_{pp}\). The only shell-model analysis going beyond the closure approximation that we are aware of was done in Ref. for \({}^{48}\)Ca using a model space consisting of only the \(f_{7/2}\) orbital. This model space is known to be insufficient for a good description of the NME due to the missing spin-orbit partner orbital \(f_{5/2}\), which significantly reduces the Gamow-Teller strength. The authors of Ref. report very small changes of the NME from closure to nonclosure, and in most cases the magnitude of the nonclosure results is slightly smaller than the magnitude of the closure result. In this paper we analyze and compare the closure and nonclosure NME for the \(0\nu\beta\beta\) decay of \({}^{48}\)Ca using a shell-model approach in the full \(pf\) shell. For the analysis we used the GXPF1A interaction. This analysis requires knowledge of a large number of one-body transition densities connecting the ground states of the initial and final states of \({}^{48}\)Ca and \({}^{48}\)Ti, respectively, with states of the intermediate nucleus \({}^{48}\)Sc. The total number of states in \({}^{48}\)Sc with angular momentum smaller than \(J=7\) is about 100000. This is still an unmanageable task. However, we show that using only a few hundred states of each \(J\) suffices to get accurate NME. In order to validate our results we also analyzed the \(0\nu\beta\beta\) NME of the "fictitious\" decays of \({}^{44}\)Ca and \({}^{46}\)Ca, for which a full account of all relevant states in the intermediate nucleus \({}^{48}\)Sc is possible. We find that the nonclosure NME always increases relative to its closure value by about 10%. The paper is organized as follows. Section II gives a brief description of the light neutrino exchange \(0\nu\beta\beta\) NME relevant for the distinction between the nonclosure approach and the closure approximation. Section III provides a brief description of the closure approximation. Section IV describes the approach we use to obtain the nonclosure results and outlines new mixed methods that use the closure approach to accelerate the convergence. In Sect. V we analyze the numerical results, and Sec. VI is devoted to conclusions and outlook. Details of the calculations are shown in the appendices. # The nuclear matrix element The decay rate for a \(0\nu\beta\beta\) decay process, under the assumption that the light neutrino-exchange mechanism dominates, can be written as \[\left[ T^{0\nu}_{1/2} \right]^{-1} = G^{0\nu} | M^{0\nu} |^2 \left(\frac{\langle m_{\beta \beta}\rangle}{m_e}\right )^2.\] Here \(G^{0\nu}\) is the phase-space factor, \(M^{0\nu}\) is the nuclear matrix element, and the effective neutrino mass \(\langle m_{\beta \beta}\rangle\) is defined by the neutrino mass eigenvalues \(m_k\) and the elements of neutrino mixing matrix \(U_{ek}\), \[\langle m_{\beta \beta}\rangle = \left| \sum_k m_k U^2_{ek} \right|.\] The nuclear matrix element \(M^{0\nu}\) is usually presented as a sum of Gamow-Teller (GT), Fermi (F), and Tensor (T) nuclear matrix elements (see, for example, Ref. ), \[\label{nme1} M^{0\nu} = M^{0\nu}_{GT}-\left( \frac{g_{V}}{g_{A}} \right)^2 M^{0\nu}_{F} + M^{0\nu}_{T},\] where \(g_{V}\) and \(g_{A}\) are the vector and axial constants correspondingly; in our calculations we use \(g_{V}=1\) and \(g_{A}=1.254\). The nuclear matrix elements in Eq. ([\[nme1\]](#nme1){reference-type="ref" reference="nme1"}) describe the transition from an initial nucleus \(|i\rangle=|0^+_i\rangle\) to a final nucleus \(|f\rangle=|0^+_f\rangle\), and they can be presented as a sum over intermediate nuclear states \(| \kappa \rangle=|J^\pi_\kappa \rangle\) with certain angular momentum \(J_\kappa\), parity \(\pi\), and energy \(E_\kappa\) \[\label{nme2} M^{0\nu}_{\alpha}=\sum_{\kappa} \sum_{1234} \langle 1 3 | {\cal O}_{\alpha} | 2 4\rangle \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 | \kappa \rangle \langle \kappa | \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle,\] where operators \({\cal O}_{\alpha}\), \(\alpha=\{ GT, F, T \}\), contain neutrino potentials, spin and isospin operators, and the explicit dependence on the intermediate state energy \(E_\kappa\). They are given by \[\begin{aligned} \label{op} {\cal O}_{GT} = & \tau_{1-} \tau_{2-} \; (\mbox{\boldmath{\(\sigma\)}}_1 \cdot \mbox{\boldmath{\(\sigma\)}}_2) \; H_{GT}(r, E_\kappa), \\ {\cal O}_{F} = & \tau_{1-} \tau_{2-} \; H_{F}(r, E_\kappa), \\ {\cal O}_{T} = & \tau_{1-} \tau_{2-} \; S_{12}\; H_{T}(r, E_\kappa), \end{aligned}\] with \(S_{12}=3(\mbox{\boldmath{\)\sigma\(}}_1 \cdot {\bf n})(\mbox{\boldmath{\)\sigma\(}}_2 \cdot {\bf n})-(\mbox{\boldmath{\)\sigma\(}}_1 \cdot \mbox{\boldmath{\)\sigma\(}}_2)\), \({\bf r}={\bf r}_1-{\bf r}_2\), \(r= | {\bf r} |\), and \({\bf n}={\bf r}/r\). The neutrino potentials, \(H_{\alpha}(r, E_\kappa)\), are integrals over the neutrino exchange momentum, \(q\), \[\label{pot} H_{\alpha}(r, E_\kappa)=\frac{2 R}{\pi} \int_0^\infty \frac{ f_{\alpha}(q r) h_{\alpha}(q^2) q d q}{q+E_\kappa-(E_i + E_f)/2},\] where \(f_{GT, F}(q r)=j_0(q r)\) and \(f_{T}(q r)=j_2(q r)\) are spherical Bessel functions. The nuclear radius \(R=1.2\times A^{1/3}\,{\rm fm}\) was introduced to make the neutrino potentials dimensionless (and since the phase-space factor \(G^{0\nu}\) contains \(1/R^2\) the final transition probability does not depend on \(R\)). The form factors \(h_{\alpha}(q^2)\) are defined in Appendix [\[ap1\]](#ap1){reference-type="ref" reference="ap1"} and they include vector and axial nucleon form factors that take into account nucleon size effects. Calculation details for two-body matrix elements, \(\langle 1 3 | {\cal O}_{\alpha} | 2 4\rangle\), are discussed in Appendix [\[ap4\]](#ap4){reference-type="ref" reference="ap4"}. Let us note that the two-body wave functions in the matrix elements ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) are not antisymmetrized, as one would expect for nuclear two-body matrix elements. They should be understood as \[|2 4\rangle = |2\rangle \cdot |4\rangle\; \mbox{ and } \; |1 3\rangle = |1\rangle \cdot |3\rangle,\] where 1, 2, 3, and 4 represent single-nucleon quantum numbers (for example, \(1=\{ \tau_{1 z},n_1,l_1,j_1,\mu_1 \}\) and so on). Appendices [\[ap2\]](#ap2){reference-type="ref" reference="ap2"}, [\[ap3\]](#ap3){reference-type="ref" reference="ap3"}, and [\[ap4\]](#ap4){reference-type="ref" reference="ap4"} provide expressions for the nuclear matrix elements ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) by considering rotational symmetry and isospin invariance. # The closure approximation If one replaces the energies of the intermediate states in Eq. ([\[pot\]](#pot){reference-type="ref" reference="pot"}) by an average constant value one gets the closure approximation, \[\label{eq8} \left[ E_\kappa-(E_i + E_f)/2 \right] \rightarrow \langle E \rangle.\] The operators \({\cal O}_\alpha \rightarrow \tilde{{\cal O}}_\alpha \equiv {\cal O}_\alpha(\langle E \rangle)\) become energy independent and the sum over the intermediate states in the nuclear matrix element ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) can be taken explicitly by using the completeness relation \[\label{tbtd} \sum_\kappa \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 | \kappa \rangle \langle \kappa | \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle = \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle.\] The advantage of this approximation is significant, because it eliminates the need of calculating a very large number of states in the intermediate nucleus, which could be computationally challenging, especially for heavy systems. One needs only to calculate the two-body transition densities ([\[tbtd\]](#tbtd){reference-type="ref" reference="tbtd"}) between the initial and the final nuclear states. This approximation is very good because the values of \(q\) that dominate the matrix elements are of the order of \(100-200\) MeV, while the relevant excitation energies are only of the order of 10 MeV. The obvious difficulty related to this approach is that we have to find a reasonable value for this average energy, \(\langle E \rangle\), which can effectively represent the contribution of all the intermediate states. This average energy needs to account also for the symmetric part of the two-body matrix elements, \(\langle 1 3 | {\cal O}_{\alpha} | 2 4\rangle\), in Eq. ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}). Indeed, the two-body wave functions \(|1 3\rangle\) and \(|2 4\rangle\) are not antisymmetric; by replacing the energies of the intermediate states with a constant, only the antisymmetric part of these matrix elements is taken into account. The uncertainty in the value of the nuclear matrix elements is related to our inability to derive the average energy, \(\langle E \rangle\), associated with the closure approximation. Fortunately, the nuclear matrix elements are not very sensitive to the value of this average energy (with the uncertainty being estimated to be about 10%; see, for example, ). Such weak dependence on the average energy originates from the large value of typical momentum of the virtual neutrino \[see Eq. ([\[pot\]](#pot){reference-type="ref" reference="pot"})\], which is \(\sim 1\,{\rm fm}^{-1}\) (\(\sim 200\,{\rm MeV}\)), i.e., much larger than the typical nuclear excitations. # Nonclosure and mixed methods In the nonclosure approach one needs to calculate the sum in Eq. ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) explicitly, which is an obvious challenge due to the large number of intermediate states \(|\kappa \rangle\). For the case of \({}^{48}\)Ca in the \(fp\) model space there are about \(10^5\) intermediate states; it is extremely difficult to find and include all these states. Let us introduce a cutoff energy \(E\) to investigate the convergence of the sum over \(\kappa\) in Eq. ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) (where here and below the sum over repeated indices \(\{1,2,3,4\}\) is omitted): \[\label{nme3} M^{0\nu}_{\alpha}(E)=\sum_{E_\kappa < E} \langle 1 3 | {\cal O}_{\alpha} | 2 4\rangle \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 | \kappa \rangle \langle \kappa | \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle.\] Alternatively, we can use a cutoff on the number of states, \(N\), calculating the sum only for \(\kappa < N\). At the limit of large cutoff energies \(M^{0\nu}_{\alpha}(E)\) approaches the exact value of the nuclear matrix element ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}). The difference between the closure and nonclosure calculations originates mainly from the low-lying excitation energies. The intermediate and higher energies cannot produce much of a difference, because with increase of the excitation energy the one-body matrix elements rapidly become very small. Based on this observation, we will use the nonclosure approach for low energies, which we can manage within the framework of the standard shell model. For the higher excitation energies, we will use the closure approximation, which is also manageable. To proceed further we introduce the sum similar to Eq. ([\[nme3\]](#nme3){reference-type="ref" reference="nme3"}) for the closure approximation: \[\label{nme4} {\cal M}^{0\nu}_{\alpha}(E)=\sum_{E_\kappa < E} \langle 1 3 | \tilde{{\cal O}}_{\alpha} | 2 4\rangle \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 | \kappa \rangle \langle \kappa | \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle.\] The difference between Eqs. ([\[nme3\]](#nme3){reference-type="ref" reference="nme3"}) and ([\[nme4\]](#nme4){reference-type="ref" reference="nme4"}) is that for the nonclosure approach the operators \({\cal O}_\alpha\) in Eq. ([\[op\]](#op){reference-type="ref" reference="op"}) are functions of the excitation energy \(E_\kappa\), while for the closure approximation the same operators \(\tilde{{\cal O}}_\alpha\) are functions of the average energy \(\langle E \rangle\) \[see the energy substitution given by Eq. ([\[eq8\]](#eq8){reference-type="ref" reference="eq8"})\]. At large cutoff energies, \(E\rightarrow \infty\), \[\label{eq12} {\cal M}^{0\nu}_{\alpha}(E) \rightarrow {\cal M}^{0\nu}_{\alpha}(\infty)= \langle 1 3 | \tilde{{\cal O}}_{\alpha} | 2 4\rangle \langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle,\] we get an "exact value\" in the framework of the closure approximation. To avoid disadvantages of both approaches we propose an interpolation method which combines both the nonclosure and closure approaches, by introducing the mixed NME \[\label{nme5} {\bar M^{0\nu}_{\alpha}}(E)= {M}^{0\nu}_{\alpha}(E)-{\cal M}^{0\nu}_{\alpha}(E)+{\cal M}^{0\nu}_{\alpha}(\infty).\] We expect that this mixed NME, \({\bar M^{0\nu}_{\alpha}}(E)\), will converge much faster with the cutoff energy than the nonclosure, \(M^{0\nu}_{\alpha}(E)\), and closure, \({\cal M}^{0\nu}_{\alpha}(E)\), matrix elements separately. At higher excitation energies these two NME will behave similarly, and the energy dependence will cancel out. We also expect that the mixed NME, Eq. ([\[nme5\]](#nme5){reference-type="ref" reference="nme5"}), will have much weaker dependence on the average energy \(\langle E \rangle\) than the pure closure NME; at least this dependence should weaken when the cutoff energy increases. It should be also mentioned that calculating \({\cal M}^{0\nu}_{\alpha}(E)\) and \({\bar M^{0\nu}_{\alpha}}(E)\) does not require more computational effort than calculating the energy-dependent nonclosure NME, \(M^{0\nu}_{\alpha}(E)\), for a given energy cutoff. \({\cal M}^{0\nu}_{\alpha}(\infty)\) can be calculated by using Eq. ([\[eq12\]](#eq12){reference-type="ref" reference="eq12"}) (the details of which are described in Ref. ). # Results Figures [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} and [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} present the closure NME \({\cal M}^{0\nu}_{\alpha}(E)\) for the fictitious \(0\nu\beta\beta\) decay cases of \({}^{44}\)Ca and \({}^{46}\)Ca. We calculated NME for these two cases only to demonstrate the convergence of the corresponding nuclear matrix elements with the increase of the cutoff energy. We could check our code by comparing with the NME calculated with a totally different method. The one-body transition densities (\(\langle f | \hat{c}^\dagger_{3} {\hat{c}}_4 | \kappa \rangle\) and \(\langle \kappa | \hat{c}^\dagger_{1} {\hat{c}}_2 | i \rangle\)) were calculated with the NUSHELLX code, and we developed our code for the two-body matrix elements. We used the GXPF1A two-body interaction in the \(pf\) model space. In the calculations we used \(\langle E \rangle=7.72\) MeV, and we also included the short-range correlations (SRC) parametrization based on the AV18 potential and the standard nucleon finite-size effects. The horizontal lines represent the "exact values\", \({\cal M}^{0\nu}_{\alpha}(\infty)\). One can see how the NME converge to their exact values: for \({}^{46}\)Ca it is enough to take into account about 50 states (instead of \(\sim 20\,000\)) and for \({}^{44}\)Ca about 25 states are needed to obtain an accuracy better than 1% for the total NME. We should also mention that for \({}^{44}\)Ca and \({}^{46}\)Ca we were able to include all the states in the intermediate nucleus, and we got the same results as using the traditional nonclosure approach \[see, e.g., Eq. ([\[tbtd\]](#tbtd){reference-type="ref" reference="tbtd"})\]. Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} and Table [1](#tbl1){reference-type="ref" reference="tbl1"} present the comparison of the results for the nonclosure approach, Eq. ([\[nme4\]](#nme4){reference-type="ref" reference="nme4"}), with the closure NME, for the decay of \({}^{48}\)Ca. In these calculations we use \[\label{eq14} \left[ E_\kappa-(E_i + E_f)/2 \right] \rightarrow 1.9\,{\rm MeV} + E^{*}_\kappa,\] where \(E^{*}_\kappa\) is the excitation energy of the intermediate nucleus \({}^{48}\)Sc, the harmonic oscillator parameter \(b_{osc}=1.989\,{\rm fm}\), and for the closure approximation the average energy was \(\langle E \rangle = 7.72\,{\rm MeV}\). Here, we also used the AV18 SRC parametrization. In Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} the nonclosure NME are represented by solid black and gray bars and the closure NME are the dashed bars, shown for various angular momenta \(J_\kappa\) of intermediate states \(|\kappa\rangle\). The Gamow-Teller matrix elements are all positive (upper part), and the Fermi matrix elements are all negative (bottom part). The main difference between closure and nonclosure comes from the GT nuclear matrix element corresponding to the intermediate angular momentum \(J_\kappa = 1\). The reason is that the transitions from an initial \(0^+\) state to an intermediate \(1^+\) state occur most naturally via the \(\mbox{\boldmath{\)\sigma\(}} \tau_-\) operator. For the other types of operators and for the intermediate spins different from \(J_\kappa=1\), we have to expand the form factors over the neutrino momentum \(q\), which makes the nuclear matrix element insensitive to low excitation energies, and therefore does not contribute to the difference between closure and nonclosure NME. This decomposition of the matrix elements, which is often provided by QRPA calculations (see, e.g., Fig. 3 of Ref. ) is presented for the first time here as a result of a shell-model analysis. As mentioned in Ref., there are no contributions from the negative-parity states of the intermediate nucleus when the model space is restricted to one major harmonic oscillator shell. Figure [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} represents another possible way to decouple the nuclear matrix elements. In this approach we consider two-body matrix elements \(\langle 1 3 | {\cal O}_\alpha | 2 4 \rangle\) where the single-particle states \(|1\rangle\) and \(|3\rangle\) (proton states) and the states \(|2\rangle\) and \(|4\rangle\) (neutron states) are coupled to certain common spin \({\cal J}\), so that the total NME can be presented as \(M_\alpha = \sum_{\cal J} M_\alpha({\cal J})\). The details of such decoupling are in Appendix [\[ap2\]](#ap2){reference-type="ref" reference="ap2"}. The nonclosure NME in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} are represented with solid black and gray bars and the closure NME are the dashed bars. In contrast to the intermediate spin decoupling, where all the spins \(J_\kappa\) contribute coherently (see Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}), in the \({\cal J}\)-decoupling scheme we see a significant cancellation between \({\cal J}=0\) and \({\cal J}=2\). Such a cancellation is responsible for the small matrix element of the double magic nucleus \({}^{48}\)Ca. Similar effects have been observed in seniority-truncation studies of the NME of \({}^{48}\)Ca (see also Ref. for effects of higher seniority in shell model calculations). QRPA results are available for heavier nuclei (see, e.g., Fig. 1 of Ref. ), for which the \({\cal J}=0\) and \({\cal J}=2\) contributions are still dominant, but the cancellation effect is significantly reduced. Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} presents the convergence of the total nuclear matrix element for \({}^{48}\)Ca to its final value, \(100\% \times \delta M/M\), as a function of the cutoff energy. The solid line defined by Eq. ([\[nme3\]](#nme3){reference-type="ref" reference="nme3"}) represents the nonclosure approach. We see that the matrix elements approach their final values (with the central shaded region corresponding to \(\pm 1\%\)) quite fast. In order to calculate the sum over the intermediate states in Eq. ([\[nme2\]](#nme2){reference-type="ref" reference="nme2"}) within an accuracy better than 1% it is enough to include only the first 100 states for each \(J_\kappa\). We conclude that if we restrict the sum over intermediate states to about 100 states of each spin, the uncertainty we introduce into the calculation by this restriction would be of the order of 1%. The dotted and dashed lines in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} represent the mixed method, where the NME are defined by Eq. ([\[nme5\]](#nme5){reference-type="ref" reference="nme5"}). The dotted lines show the total matrix element, which includes all possible intermediate spins \(J_\kappa\). It converges much faster than the pure nonclosure matrix element. To get an accuracy of about 1% using this method we have to take into account only states of up to 7 MeV in excitation energy (about 20 states per each \(J_\kappa\)). The hope is that using this mixed method we can achieve the desirable accuracy significantly faster (with a lower number of intermediate states) than using a pure nonclosure approach. To obtain the NME of heavier nuclei, for which the dimensions are extremely high, such a decrease in computational demands can be crucially important. The main contribution to the NME originates from the intermediate states with spin \(J_\kappa=1\) (see Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}). This observation can be used to decrease the number of intermediate states required for a given accuracy. The dashed lines in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} represent the NME when the intermediate sates with \(J_\kappa=1\) are only taken into account. The difference between dotted and dashed lines is only 2%, which means that if we include only the first 20 states with \(J_\kappa=1\) we already achieve an accuracy of 3%. This allows us to avoid calculation of all the intermediate states with \(J_\kappa \ne 1\) and still get the NME with good accuracy. Table [1](#tbl1){reference-type="ref" reference="tbl1"} summarizes the difference between the total matrix elements calculated within the closure approximation and the nonclosure approach. We found about an 11% percent difference for the GT matrix element, which is quite noticeable. For the total matrix element this difference decreases to 10%. The nonclosure results can be obtained in the closure approximation if one uses an appropriate value for \(\langle E \rangle\) and not \(\langle E \rangle = 7.72\) MeV as suggested by QRPA calculations . For CD-Bonn and AV18 SRC parametrizations (see Table [2](#tbl2){reference-type="ref" reference="tbl2"}) this appropriate energy is found to be about \(\langle E \rangle = 0.5\) MeV, but its value may be different for different model spaces, interactions, or SRC parametrizations. ::: Finally, Table [2](#tbl2){reference-type="ref" reference="tbl2"} presents the nonclosure \(^{48}\)Ca NME calculations performed with different SRC parametrization sets. # Conclusions and Outlook In conclusion, we investigated the closure versus nonclosure approach of the \(0 \nu \beta \beta\) NME for \(^{48}\)Ca using for the first time shell-model techniques in the realistic \(pf\) shell valence space. We found that the closure approximation always gives smaller NME, \(M^{0 \nu}_{\nu}\), by about 10%. A similar comparison of closure versus nonclosure NME for heavy nuclei, such as \(^{76}\)Ge, \(^{96}\)Zr, \(^{100}\)Mo, and \(^{130}\)Te, was done within the QRPA method in Ref. (see, e.g., its Fig. 4), where the authors came to the same conclusion, namely, that the nonclosure NME are about 10% larger than the closure NME. In addition, we were able to obtain for the first time a decomposition of the shell-model NME versus the total spin \(J\) of the intermediate states, and we found that for the case of \(^{48}\)Ca the \(J=1\) states provide the largest contribution. We have also found that most of the additional difference between closure and nonclosure comes from the transitions to the \(1^+\) states in the intermediate nucleus. By combining the nonclosure and closure approaches together we propose a new method of calculating the \(0\nu\beta\beta\) NME, which converges very quickly using only a very small number of states in the intermediate nucleus. This result suggests that one can apply this method to obtain the shell-model nonclosure NME for \(0\nu\beta\beta\) decay of heavier nuclei, such as \({}^{76}\)Ge or \({}^{82}\)Se. It would be also interesting to go beyond the closure approximation for the NME corresponding to other mechanisms that may contribute to the \(0\nu\beta\beta\) decay rates. Finally, it is worth mentioning that the nonclosure approach does not constrain the states of the intermediate nucleus to be in the same model space used for the initial and the final state, as is the case for the closure approximation (see, e.g., Ref. ). For example, it was recently shown that the two-neutrino double-\(\beta\) decay NME, which need to be calculated using a nonclosure approach, could change if the model space used for the intermediate \(1^+\) states is enlarged. This effect could be considered in future studies. Here, we use for the nonclosure approach the same constraint as that imposed by the closure approximation. RAS is grateful to N. Auerbach and V. Zelevinsky for constructive discussions. Support from the NUCLEI SciDAC Collaboration under U.S. Department of Energy Grant No. DE-SC0008529 is acknowledged. MH also acknowledges U.S. NSF Grant No. PHY-1068217.

{'timestamp': '2013-12-17T02:17:42', 'yymm': '1310', 'arxiv_id': '1310.3254', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3254'}

# Introduction {#Introd_sect} The *block maxima* (BM) approach in extreme value theory (EVT), consists of dividing the observation period into nonoverlapping periods of equal size and restricts attention to the maximum observation in each period \[see, e.g., \]. The new observations thus created follow---under domain of attraction conditions, cf. [\[mda_cond\]](#mda_cond){reference-type="eqref" reference="mda_cond"} below---approximately an extreme value distribution, \(G_\gamma\) for some real \(\gamma\). Parametric statistical methods for the extreme value distributions are then applied to those observations. In the *peaks-over-threshold* (POT) approach in EVT, one selects those of the initial observations that exceed a certain high threshold. The probability distribution of those selected observations is approximately a generalized Pareto distribution. In the case of the POT method, exact conditions under which the statistical method is justified can be described by a second-order term \[see, e.g., and, Section 2.3\]. In the case of block maxima, usually it is taken for granted that the maxima follow very well an extreme value distribution. In this paper, we take this misspecification into account by quantifying it in terms of a second-order expansion; cf. Condition [\[2ndordcond\]](#2ndordcond){reference-type="ref" reference="2ndordcond"} below. Since \(G_\gamma\) is not the exact distribution for those observations, a bias may appear. The POT method picks up all "relevant" high observations. The BM method on the one hand misses some of these high observations and, on the other hand, might retain some lower observations. Hence the POT seems to make better use of the available information. There are practical reasons for using the BM method: - The only available information may be block maxima, for example, yearly maxima with long historical records or long range simulated data sets. - The BM method may be preferable when the observations are not exactly independent and identically distributed (i.i.d.). For example, there may be a seasonal periodicity in case of yearly maxima or, there may be short range dependence that plays a role within blocks but not between blocks; cf. for example, Katz, Parlange and Naveau () and for further discussion. - The BM method may be easier to apply since the block periods appear naturally in many situations \[, van den Brink, Können and Opsteegh (), \]. On the other hand, the POT method allows for greater flexibility in many cases since it might be difficult to change the block size in practice. When working with BM, there are two sets of estimators that are widely used: the maximum likelihood (ML) estimators \[e.g., \] and the probability weighted moment (PWM) estimators. Recently, has proved consistency of the former. The present paper concentrates on the latter. Our work has given rise to the paper on the multivariate case. The PWM estimators under the \(G_\gamma\) model are very popular, for example, in applications to hydrologic and climatologic extremes, because of their computational simplicity, good performance for small sample sizes and robustness even for location and scale parameters \[, Katz, Parlange and Naveau (), \]. The relative merits of POT and BM have been discussed in several papers, all based on simulated data: states that for \(\gamma=0\) and ML estimators, the POT estimate for a high quantile is better only if the number of exceedances is larger than 1.65 times the number of blocks; writes that POT is as efficient as BM model for high quantiles, based on PWM estimators; Madsen, Pearson and Rosbjerg () and Madsen, Rasmussen and Rosbjerg () write that POT is preferable for \(\gamma>0\), whereas for \(\gamma<0\), BM is more efficient, again with the number of exceedances larger than the number of blocks; state that the gains (when using historical data) with the BM model are in the range of the gains with the POT model, based on ML estimators; in a vast simulation study writes that with POT samples having an average of two or more observations per block, the estimates are more accurate than the corresponding BM estimates, and with more than 200 years of data the accuracies of the two approaches are similar and rather good, based on several estimators including the PWM and ML estimators. From all these studies, some even with mixed views, the following two features seem dominant. First, POT is more efficient than BM in many circumstances, though needing, on average, a number of exceedances larger than the number of blocks. Secondly, POT and BM often have comparable performances, for example, for large sample sizes. Our theoretical comparison shows that BM is rather efficient. The asymptotic variances of both extreme value index and quantile estimators are always lower for BM than for POT. Moreover, the approximate minimal mean square error is also lower for BM under usual circumstances. The optimal number of exceedances is generally higher than the optimal number of blocks. The paper is organized as follows. In Section [2](#AsympN_sect){reference-type="ref" reference="AsympN_sect"}, we state exact conditions to justify the BM method, along with the asymptotic normality result for the PWM estimators including high quantile estimators. In Section [3](#blockmaxPOT_sect){reference-type="ref" reference="blockmaxPOT_sect"}, we provide a theoretical comparison between the two methods, BM and POT. The analysis is based on a uniform expansion of the relevant quantile process given in Section [2.1](#Asymptnorm_sect){reference-type="ref" reference="Asymptnorm_sect"}. This expansion also provides a basis for analysing alternative estimators besides the PWM estimator. Proofs are postponed to Section [4](#proofs_sect){reference-type="ref" reference="proofs_sect"}. Throughout the paper, we assume that the observations are i.i.d. In future work, we shall extend the results to the non-i.i.d. case and to the maximum likelihood estimator. # The estimators and their properties {#AsympN_sect} Let \(\tilde X_1, \tilde X_2, \ldots\) be i.i.d. random variables with distribution function \(F\). Define for \(m=1,2,\ldots\) and \(i=1,2,\ldots,k\) the block maxima \[\label{blockmaxima_def} X_i=\max_{(i-1)m<j\leq im} \tilde X_j.\] Hence, the \(m\times k\) observations are divided into \(k\) blocks of size \(m\). Write \(n=m\times k\), the total number of observations. We study the model for large \(k\) and \(m\), hence we shall assume that \(n\to\infty\); in order to obtain meaningful limit results, we have to require that both \(m=m_n\to\infty\) and \(k=k_n\to\infty\), as \(n\to\infty\). The main assumption is that \(F\) is in the domain of attraction of some extreme value distribution \[G_\gamma(x)=\exp \bigl(-(1+\gamma x)^{-1/\gamma} \bigr),\qquad \gamma \in \R, 1+\gamma x>0,\] that is, for appropriately chosen \(a_m>0\) and \(b_m\) and all \(x\) \[\begin{aligned} \label{mda_cond} \lim_{m\to\infty} P \biggl( \frac{X_i-b_m}{a_m}\leq x \biggr)&=&\lim_{m\to \infty} F^m (a_mx+b_m ) \nonumber \\[-8pt] \\[-8pt] \nonumber &=&G_\gamma(x),\qquad i=1,2,\ldots,k. \end{aligned}\] This can be written as \[\lim_{m\to\infty} \frac{1} m \frac{1}{-\log F (a_mx+b_m )}= (1+\gamma x)^{1/\gamma},\] which is equivalent to the convergence of the inverse functions: \[\lim_{m\to\infty}\frac{V(mx)-b_m}{a_m}=\frac{x^\gamma-1}{\gamma},\qquad x>0,\] with \(V= (-1/\log F )^{\leftarrow}\). Hence, \(b_m\) can be chosen to be \(V(m)\). This is the first-order condition. For our analysis, we also need a second-order expansion as follows. Let \(X_{1,k},\ldots,X_{k,k}\) be the order statistics of the block maxima \(X_1,\ldots,X_k\). The statistics \(\beta_0=k^{-1}\sum_{i=1}^k X_{i,k}\) and \[\beta_r=\frac{1} k \sum_{i=1}^k \frac{(i-1)\cdots(i-r)}{(k-1)\cdots (k-r)} X_{i,k},\qquad r=1,2,3,\ldots, k>r,\] are unbiased estimators of \(EX_1F^{rm}(X_1)\) \[\]. The PWM estimators for \(\gamma\), as well as the location \(b_m\) and scale \(a_m=a([m])\), are simple functionals of \(\beta_0\), \(\beta_1\) and \(\beta_2\). The estimator \(\hat\gamma_{k,m}\) for \(\gamma\) is defined as the solution of the equation \[\begin{aligned} \frac{3^{\hat\gamma_{k,m}}-1}{2^{\hat\gamma_{k,m}}-1}&=&\frac{3\beta _2-\beta_0}{2\beta_1-\beta_0},\nonumber \\ \quad\hat a_{k,m}&=&\frac{\hat\gamma_{k,m}}{2^{\hat\gamma_{k,m}}-1} \frac {2\beta_1-\beta_0}{\Gamma (1-\hat\gamma_{k,m} )}\quad \mbox{and}\quad \hat b_{k,m}=\beta_0+\hat a_{k,m}\frac{1-\Gamma (1-\hat\gamma _{k,m} )}{\hat\gamma_{k,m}}, \end{aligned}\] where \(\Gamma(x)=\int_0^\infty t^{x-1} e^{-t} \,dt\), \(x>0\) \[\]. The rationale behind the estimator of \(\gamma\) becomes clear when checking the statement of Theorem [\[AsymptNgab\]](#AsymptNgab){reference-type="ref" reference="AsymptNgab"} below. ## Asymptotic normality {#Asymptnorm_sect} The following theorem is the basis foranalysing estimators in the BM approach. Let \(\lceil u \rceil\) represent the smallest integer larger than or equal to \(u\). Note that \(\Gamma'(1-\gamma)=\int_0^\infty u^{-\gamma}e^{-u}(\log u) \,du\) and \(\Gamma''(1)=1.97811\). ## High quantile estimation {#Asymptnormq_sect} Our estimator for \(x_n=F^\leftarrow (1-p_n )= V (1/ (-\log (1-p_n ) ) )\), with \(p_n\) small, is \[\hat x_{k,m}=\hat b_{k,m}+\hat a_{k,m} \frac{ (mp_n )^{-\hat \gamma_{k,m}}-1}{\hat\gamma_{k,m}}.\] # Theoretical comparison between BM and POT methods {#blockmaxPOT_sect} In this section, we develop a theoretical comparison between the BM and POT methods, by comparing the two PWM estimators for the two methods \[Hosking and Wallis () and Hosking, Wallis and Wood (), resp., for POT and BM\]. First, we introduce the PWM-POT estimators for \(\gamma\) and \(a(n/k)\), where \(k\) is the number of selected order statistics, \(\{\tilde X_{n-i,n}\}_{i=0}^{k-1}\), from the original sample \(\tilde X_1,\tilde X_2,\ldots,\tilde X_n\). The statistics \[P_n= \frac{1}{k} \sum _{i=0}^{k-1} \tilde X_{n-i,n}-\tilde X_{n-k,n} \quad\mbox{and}\quad Q_n= \frac{1}{k} \sum _{i=0}^{k-1} \frac{i}{k} (\tilde X_{n-i,n}-\tilde X_{n-k,n} )\] are estimators for \(a(n/k)(1-\gamma)^{-1}\) and \(a(n/k) (2(2-\gamma )^{-1} )\), respectively. Consequently, the PWM estimators are \[\hat\gamma_{k,n}=1-\biggl(\frac{P_n}{2 Q_n}-1 \biggr)^{-1} \quad\mbox{and}\quad \hat a(n/k)=P_n \biggl( \frac{P_n}{2 Q_n}-1 \biggr)^{-1}.\] The quantile estimator is \[\hat x_{k,n}=\tilde X_{n-k,n}+\hat a(n/k)\frac{ ({k}/{(np_n)} )^{\hat\gamma_{k,n}}-1}{\hat\gamma_{k,n}}.\] Asymptotic normality under conditions equivalent to the ones in Theorems [\[AsymptNgab\]](#AsymptNgab){reference-type="ref" reference="AsymptNgab"} and [\[AsymptNquantile\]](#AsymptNquantile){reference-type="ref" reference="AsymptNquantile"} holds \[see, e.g., \], if \(\rho\in[-1,0]\) with a caveat for \(\rho =-1\) \[for certain cases the functions \(A\) in the corresponding second-order conditions may not be the same asymptotically resulting in different values of \(\lambda\) in the limiting distributions; cf. \]. For BM, \(k\) is defined as the number of blocks and, for POT, \(k\) is defined as the number of selected top order statistics. Hence, in both cases \(k\) means the number of selected observations. For the theoretical comparison, we confine ourselves to the range \(\rho\in [-1,0]\) and \(\gamma\in[-1,1/2)\), a usual range in many applications. ## Extreme value index estimators {#extreme-value-index-estimators .unnumbered} - First, we compare asymptotic variance and bias for a common value of \(k\): The asymptotic variances of the two \(\gamma\) estimators are shown in Figure [\[VarBMPOT.fig\]](#VarBMPOT.fig){reference-type="ref" reference="VarBMPOT.fig"}: the curve from BM is always below the other one, meaning lower values for the asymptotic variance for all values of \(\gamma\). The asymptotic biases are compared in Figure [\[biasRATIO.fig\]](#biasRATIO.fig){reference-type="ref" reference="biasRATIO.fig"}, through the ratio "bias BM/bias POT". Recall that the bias depends on both first-and second-order parameters \(\gamma\) and \(\rho\). Contrary to what is observed for the variance, the bias of BM is always larger but for \(\rho=0\) they are the same regardless the value of \(\gamma\), equal to 1 \[or \(\lambda\) if one takes into account the asymptotic contribution of \(\sqrt k A(n/k)\) to the biases\]. - Next, we compare asymptotic mean square errors for the "optimal choice" of \(k\) (i.e., that value that makes the limiting mean square error of \(\hat\gamma-\gamma\) minimal), which is different in the two cases: An asymptotic expression of the "asymptotic minimal mean square error" (MINMSE in the sequel) is obtained in the following way. Suppose \(\rho<0\). First we find for each estimator the optimal \(k\) in the sense of minimizing the approximate asymptotic mean square error. Denote by \(\sigma^2_i=\sigma^2_i(\gamma)\) and \(B^2_i=B^2_i(\gamma,\rho )\) (\(i=1,2\); "1" refers to PWM-BM and "2" refers to PWM-POT) the asymptotic variance and squared bias of the estimators. Under Condition [\[2ndordcond\]](#2ndordcond){reference-type="ref" reference="2ndordcond"}, we can write \(A^2(t)=\int_t^\infty s(u) \,du\) with \(s(\cdot)\) decreasing and \(2\rho-1\) regularly varying. The limiting mean square error is, approximately, \[\label{minmse} \inf_k \biggl(\frac{\sigma_i^2}k+A^2(n/k)B_i^2 \biggr)\] or, writing \(r\) for \(n/k\), \(\inf_r ( (r/n)\sigma_i^2+B_i^2\int_r^\infty s(u) \,du )\). Setting the derivative equal to zero and using properties of regularly varying functions one finds for the optimal choice of \(r\), \(r_0^{(i)}\sim (1/ s )^\leftarrow(n) (B_i^2/\sigma _i^2 )^{1/(1-2\rho)}\) and, in terms of \(k\), \[k_0^{(i)}\sim\frac{n}{ ({1} / s )^{\leftarrow}(n)} \biggl(\frac{\sigma_i^2}{B_i^2} \biggr)^{1/(1-2\rho)}.\] Note that the optimal \(k_0^{(i)}\) is different but of the same order for both methods. Next, inserting \(k_0^{(i)}\) in [\[minmse\]](#minmse){reference-type="eqref" reference="minmse"}, after some manipulation we get the following asymptotic expression for MINMSE, \[\frac{1-2\rho}{-2\rho}\frac{ ({1}/ s )^{\leftarrow}(n)}n \bigl(B_i^2 \bigr)^{1/(1-2\rho)} \bigl(\sigma_i^2 \bigr)^{-2\rho/(1-2\rho)}.\] It follows that MINMSE(BM)/MINMSE(POT) is, approximately, \[\biggl(\frac{B_1^2(\gamma,\rho)}{B_2^2(\gamma,\rho)} \biggr)^{1/(1-2\rho )} \biggl( \frac{\sigma_1^2(\gamma)}{\sigma_2^2(\gamma)} \biggr)^{-2\rho /(1-2\rho)},\] which does not depend on n, just on \(\gamma\) and \(\rho\). The contour plot of "MINMSE(BM)/MINMSE(POT)" is represented in Figure [\[ratiocontour.fig\]](#ratiocontour.fig){reference-type="ref" reference="ratiocontour.fig"}. It can be seen that the BM has lower MINMSE for a large range of \((\gamma,\rho)\) combinations. Note that this range includes \(\gamma\) negative and \(\gamma\) positive close to zero which seem to be common values in many practical situations, for example, in hydrologic and climatologic extremes. Only for \(\gamma>0.2\) approximately, MINMSE for POT can be lower depending on \(\rho\). Finally, comparing the optimal sample sizes (cf. Figure [\[contourk0RATIO.fig\]](#contourk0RATIO.fig){reference-type="ref" reference="contourk0RATIO.fig"} with contour plot of the ratio of the optimal values of \(k\)), one sees that POT requires systematically larger optimal sample size even when the approximate MINMSE is smaller for POT than BM. ## Quantile estimators {#quantile-estimators .unnumbered} We repeat the previous analysis for the quantile estimators: - The asymptotic variances of the two estimators are compared in Figure [\[VarQBMPOT.fig\]](#VarQBMPOT.fig){reference-type="ref" reference="VarQBMPOT.fig"}: again the curve from BM is always below the other one meaning lower values for the asymptotic variance for all values of \(\gamma\). In Figures [\[biasQBM.fig\]](#biasQBM.fig){reference-type="ref" reference="biasQBM.fig"} and [\[biasQPOT.fig\]](#biasQPOT.fig){reference-type="ref" reference="biasQPOT.fig"}, the asymptotic bias is represented for each case separately. Note that for \(\gamma\) negative, the bias for BM approaches zero when \(\rho\uparrow0\) whereas in the POT case it escapes to \(-\infty\). - The contour plot for the ratio "MINMSE(BM)/MINMSE(POT)" is represented in Figure [\[contourQRATIO.fig\]](#contourQRATIO.fig){reference-type="ref" reference="contourQRATIO.fig"}. Again the BM method has lower MINMSE for a large range of \((\gamma,\rho)\) combinations. The "irregularity" around \(\gamma \approx-0.2\) is due to a change of sign in the bias in the POT case. Finally, Figure [\[contourk0QRATIO.fig\]](#contourk0QRATIO.fig){reference-type="ref" reference="contourk0QRATIO.fig"} gives the contour plot for the ratio of the optimal values of \(k\), which is smaller than one when \(\gamma\) is small and \(\rho\) is closer to zero. *In conclusion*, for both the extreme value index and quantile PWM estimators, the ones from the BM method have always lower asymptotic variances. Moreover, at an optimal level the BM gives lower MINMSE, thus being more efficient, under many practical situations. This is in agreement with some of Sofia Caires' () conclusions, for example, that for equal sample sizes or with more than 200 years of data the uncertainty or the error of the estimates are lower for BM than for POT. # Proofs {#proofs_sect} Throughout this section, \(Z\) represents a unit Fréchet random variable, that is, one with distribution function \(F(x)=e^{-1/x}\), \(x>0\), and \(\{ Z_{i,k} \}_{i=1}^k\) are the order statistics from the associated i.i.d. sample of size \(k\), \(Z_1,\ldots,Z_k\). Similarly, \(\{ X_{i,k} \}_{i=1}^k\) represents the order statistics of the block maxima \(X_1,\ldots,X_k\) from [\[blockmaxima_def\]](#blockmaxima_def){reference-type="eqref" reference="blockmaxima_def"} and, \(X_{\lceil u\rceil,k}=X_{r,k}\) for \(r-1<u\leq r\), \(r=1,\ldots,k\). Recall the function \(V\) from Section [2](#AsympN_sect){reference-type="ref" reference="AsympN_sect"}. The following representation will be useful: \[\label{XVZrepresentation} X=^d V(mZ).\] We start by formulating a number of auxiliary results. The following is an easily obtained variant of Theorem B.3.10 of. Note that \[H_{\gamma,\rho}(x)=\cases{ \displaystyle\frac{1}{\rho} \biggl( \frac{x^{\gamma+\rho}-1}{\gamma+\rho}-\frac {x^{\gamma}-1}{\gamma} \biggr), & \quad \(\rho\neq0\neq\gamma,\) \vspace*{2pt}\cr \displaystyle\frac{1}{\gamma} \biggl(x^\gamma\log x-\frac{x^{\gamma}-1}{\gamma} \biggr) , &\quad \(\rho=0\neq\gamma,\) \vspace*{2pt}\cr \displaystyle\frac{1}{\rho} \biggl(\frac{x^{\rho}-1}{\rho}-\log x \biggr), &\quad \(\rho\neq0=\gamma,\) \vspace*{2pt}\cr \displaystyle\frac{1}{2} (\log x)^2, &\quad \(\rho= 0=\gamma.\)}\]

{'timestamp': '2014-12-31T02:07:42', 'yymm': '1310', 'arxiv_id': '1310.3222', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3222'}

# Introduction The invention of laser and evaporative cooling of atoms has led to significant new research, including new frequency standards, precision spectroscopy, quantum computation, studies of cold collisions and new degenerate quantum gases. The first laser cooling experiments were done with highly "closed cycle" species (alkali and alkaline-earth atoms), that show little spontaneous decay into metastable states. In subsequent work, species with more complex internal structure were cooled and trapped, including those with significant "leaks" into off-resonant states. Notable examples include Yb, Cr, Tm, Er and Dy. The latter four have large magnetic dipole moments, suitable for the study of systems with long-range dipole-dipole interactions and the testing of fundamental theories, whereas Yb has been used in quantum simulation studies. Other possible uses for ultracold non \(S\)-state atoms could include the creation of exotic quantum phases and quantum magnetism. Quantum many-body and quantum information research has driven the search for new species, including polar molecules. The additional degrees of freedom in molecules offer physical effects not found in atoms. Ultracold diatomic polar molecules are proposed candidates for novel quantum information and simulation experiments and new ultracold chemistry. The specific molecule necessary for an experiment can depend strongly on the desired physics to be studied. A method to provide a variety of molecules at mK temperatures and at a high phase-space density is desired, but is unavailable at present. One possible approach to useful sources of some (ultra-)cold molecules is laser cooling, including magneto-optical trapping. Recent experiments employing hydrodynamically enhanced cryogenic buffer gas beam sources (CBGBs) have reported laser cooling of molecules that have highly diagonal Frank-Condon factors: the optical slowing and cooling of SrF; the creation of a two-dimensional magneto-optical trap for YO; and the laser slowing of CaF in a supersonic jet. Work is ongoing in several groups to create MOTs for these molecules. In this paper, we report the successful use of a two-stage, slower CBGB for loading MOTs, including species that have leaky optical cycling transitions. This is a step toward a simple approach to loading polar diatomic molecules into a MOT. We demonstrate the method with lanthanide atoms. Employing a two-stage helium CBGB, we create cold, slow atomic beams of lanthanide atoms and load them directly into a magneto-optical trap. The effects of the collisions between residual helium from the CBGB and the trapped atoms are characterized. The feasibility of loading a MOT for molecules using this source is studied. The low mean forward velocity of our beam source is \(\sim\) 65 m/s, which renders a Zeeman slower unnecessary and thus allows for a direct loading of the MOT. This demonstrates a possible experimental path to a MOT for molecules; this beam is slower than the hydrodynamic CBGB used in other laser cooling experiments and thus offers a much shorter slowing path for the molecules, before the MOT. The initial cooling stage of our experiment relies only on collisions with He atoms (and thus is not dependent on the internal structure of the species). We demonstrate a unique flexibility of this system with regard to the choice of species by creating MOTs for the elements Yb, Tm, Er and Ho in the same apparatus with no hardware change except tuning the MOT lasers, and no additional slowing light. In the case of Yb, we load MOTs for the isotopes of mass 170, 171, 172, 174 and 176 by changing one parameter only, the detuning of the MOT laser frequency. The total number of atoms in common MOTs for Yb using the 400 nm transition is limited to \(\sim 4-5\times10^6\) due to spontaneous decay into metastable states during the loading phases of these traps (\(\sim 1-10\) s). We overcome this limitation with our high instantaneous flux beam source, allowing us to fully load the MOT in only a few milliseconds. # Description of Atomic Species To most simply explain our method, we focus the discussion on only one species, Ho. Miao et al.  have recently reported laser cooling and trapping of Ho. We also refer the reader to work on MOTs of the other species: Yb, Tm and Er. Ho has only one stable nuclear isotope, \(^{165}\)Ho. With its nuclear spin quantum number of \(I=7/2\), it has the largest number of hyperfine states of any element, which makes it potentially useful for qubit implementations. Like many other lanthanides, it has a large magnetic dipole moment (\(9\,\mu_\textrm{\footnotesize B}\)). As shown in the level scheme in (c), the transition at 410.5 nm with a linewidth of \(2\pi\times 32.5\) MHz connecting the \(4f^{11}6s^2 ( J=15/2, F=11)\) and \(4f^{11}6s6p (J=17/2, F=12)\) states was used for laser cooling. As with previous work on other lanthanides (see e.g. Ref. ), no additional repumper was necessary to trap Ho atoms in our MOT, despite the existence of possible decay channels into other hyperfine levels of the ground state manifold. # Experimental Method The apparatus is shown in . A detailed study of the two-stage buffer gas beam source is described elsewhere. Briefly, the cell, operating at 2.5 K, uses a combination of hydrodynamic extraction and a second slowing stage to produce a cold and effusive-like beam with a peak forward velocity of \(\sim 60-70\) m/s and FWHM of \(\sim 70\) m/s. A considerable fraction of atoms move below the capture velocity of the atomic MOTs, which is estimated to be \(\lesssim 30\) m/s for all species we trap here. Atoms are introduced into the gas phase by laser ablation of solid precursor targets with 4 ns long pulses and \(\sim 14\) mJ energy from a 532 nm Nd:YAG laser. The atoms thermalize translationally via collisions with the cold He buffer gas inside the cell to a temperature of around 2 to 4 K and leave the cell to form an atomic beam. We use He flow rates between 0.2-4 sccm (standard cubic cm per minute). The beam travels from the cryogenic section, through a room temperature beam region, past a mechanical shutter, and into the MOT section. Since the atoms to be trapped are present only for a few ms after the ablation pulse, the shutter keeps residual He buffer gas from entering the MOT region when the atomic species of interest is not present in the beam. For the MOT measurements reported here, the shutter is opened for 10 ms (with 3 ms delay after ablation) during the MOT loading phase. In order to keep the amount of residual buffer gas background to a minimum, each section of the chamber is pumped by a turbomolecular pump to maintain steady-state low pressures of \(\sim 1\times 10^{-6}\) Torr in the beam region and \(\sim 2\times 10^{-7}\) Torr in the MOT regions. We detect atoms in the MOT by imaging fluorescence onto a calibrated photomultiplier tube (PMT). To reduce background due to scattered laser light, we spatially filter the collected fluorescence using an objective that focuses the MOT image through a variable intermediate aperture with a minimal diameter of \(\sim 600\mu\)m. Simulation of the imaging system using commercial raytracing software, determines the collection efficiency to be \((2.5\pm0.5)\times 10^{-3}\). MOT laser beams for each atomic species are derived from the same frequency-doubled Ti:Sapphire ring laser. Each MOT beam has \(\sim 15\) mW power with a \(1/e^2\) diameter of \(9.8 \pm 0.5\) mm. The Ti:Saph laser is locked to a HeNe laser via a transfer cavity, providing \(<5\) MHz laser linewidth, which is used to determine the error in the lifetime measurements. The quadrupole field for the MOT is generated from a pair of water-cooled coils. The coils produce an axial (radial) field gradient of \(3.8 \times 10^{-3}\) T cm\(^{-1}\) (\(1.9 \times 10^{-3}\) T cm\(^{-1}\)) for the Yb MOT and \(1.7 \times 10^{-3}\) T cm\(^{-1}\) (\(0.8 \times 10^{-3}\) T cm\(^{-1}\)) for the Tm, Er and Ho MOTs, which is estimated by a finite-element software package. # Results ## Model for the Pulsed Loading Process The loading process of the MOT can be described by a phenomenological differential equation for the number of trapped particles \(n(t)\), \[\begin{aligned} \label{eq:loading_dgl} \frac{dn}{dt} = R(t)-\alpha n(t)-\beta n(t)^2\quad, \end{aligned}\] where \(R(t)\) is the loading rate, \(\alpha\) the loss due to background gas collisions and \(\beta\) the intra-particle two-body loss. In our measurements, no evidence for two-body effects is observed and, thus, the corresponding term will henceforth be neglected (\(\beta = 0\)). The loading rate is time-dependent due to the pulsed nature of the loading process. We find that approximating the loading pulse by a Gaussian function as \[\begin{aligned} \label{eq:pulsed_load_func} R(t) = \frac{n_\textrm{\footnotesize tot}}{\sqrt{2\pi} w} \cdot e^{-\frac{(t-t_0)^2}{2 w^2}} \end{aligned}\] yields a very good agreement with our measured data. The total number of atoms \(n_\textrm{\footnotesize tot}\) is defined by the normalization \(\int_{-\infty}^{\infty} dt R(t) = n_\textrm{\footnotesize tot}\), where \(t_0\) is the pulse arrival time and \(w\) the pulse width. The solution to the loading equation is \[\begin{aligned} \label{eq:pulsed_load_fit} \nonumber n(t) &=& \frac{1}{2}\cdot n_\textrm{\footnotesize tot} \cdot e^{\frac{1}{2} \alpha \left(-2(t-t_0) + \alpha w^2\right) } \cdot\\* &&\left( \textrm{erf}\left(\frac{t_0+\alpha w^2}{\sqrt{2}w}\right)-\textrm{erf}\left(\frac{-t + t_0+\alpha w^2}{\sqrt{2}w}\right) \right)\quad, \end{aligned}\] with the error function defined as \(\textrm{erf}(x) = \frac{2}{\sqrt{\pi}} \int_0^x dt e^{-t^2}\). Example of the measured time traces of the MOT loading process are shown in along with a fit to the pulsed loading model of . We focus on the case of Ho,  (c). Note that only the tail of the buffer gas beam signal between 20-30 ms is loaded into the MOT. While the beam signal represents only a velocity class corresponding to a certain detuning, the much larger MOT signal stems from particles from a range of velocities which have been actively cooled and trapped. \(^a\) Previous work: 23(11) s\(^{-1}\) from Ref. . \(^b\) Previous work: 22(6) s\(^{-1}\) from Ref. . \(^c\) Previous work: 1695(43) s\(^{-1}\) from Ref. . \(^d\) These measurements were taken with a repetition rate of the ablation laser of \(0.6\) Hz. \(^f\) This measurement employs a longitudinal single-frequency slowing laser. ::: The resulting peak values for the observed number of atoms loaded into the MOT are given for each element in . Note that the variation among the different species is not only caused by the cooling efficiency given by their corresponding natural linewidths of the cooling transition, but also by the ablation yields, which strongly vary between the specific atoms. The number of trapped atoms can be increased by applying low repetition rates of the ablation laser, leading to smaller heat loads on the buffer gas cell and more atoms below the capture velocity, and by using a single-frequency slowing laser. We demonstrate this with one of the elements, Yb; we observe a total of \(1.3(0.7) \times 10^8\) trapped Yb atoms with an average loading rate of \(n_\textrm{\footnotesize tot}/(2w) = 2.0(1.0) \times 10^{10}\) atoms/s. ## Limitation of the MOT Lifetime due to the Helium Buffer Gas Two processes compete with the trap loading and limit the MOT lifetime, namely, collisions with the background gas and decay into metastable states. The first process is dominated by residual He buffer gas in the MOT region. After cold He hits the room-temperature walls of the vacuum chamber, it can bounce back and hit the trapped atoms, leading to losses. In the case of Yb, for example, a head-on collision of an atom at rest with a room-temperature He atom increases its speed by \(\sim 50\) m/s, which means the atom cannot be recaptured into the trap. The measured decrease of MOT lifetimes with varying He flow is shown in . All measurements were taken at the laser detuning which results in the maximum MOT fluorescence signal. Each point is the result of averaging over five ablation shots. The highest lifetimes for each element were observed for \(\sim 0.3\) sccm He flow and are listed in . ## Decay into Metastable States The large differences among the species are due to the second lifetime-limiting effect, namely the presence of decay channels to metastable states. Once the atom decays after a number of photon scattering events into such a dark state, it escapes the cooling and trapping process. In order to fully describe this decay, for example in the case of Er, 110 intermediate states have to be considered. As a result, theoretical modeling of such complex atoms is daunting. Instead, we apply a simplified model which considers a three-level system including trap losses (). In this model, the lifetime \(1/\alpha\) scales with the fractional population in the excited state and depends on the laser detuning as \[\begin{aligned} \label{eq:lifetime_detuning_function} \alpha(\Delta) = \Gamma_0 + \Gamma_1 \cdot \left( \frac{s_{0, \textrm{\footnotesize eff}}/2}{1+s_{0, \textrm{\footnotesize eff}}+4\Delta^2/\Gamma^2} \right)\quad, \end{aligned}\] where \(\Gamma_0\) represents the loss rate due to background collisions, \(\Gamma_1\) is the decay rate into intermediate metastable states, \(s_{0, \textrm{\footnotesize eff}} = \kappa \cdot s_0\) is the effective saturation parameter, \(\Delta\) is the laser detuning from the resonance, and \(\Gamma = 2\pi\times \Delta\nu\), where \(\Delta\nu\) is the natural linewidth. The effective saturation parameter takes averaging over Zeeman substates and random light polarization in the MOT region into account, where \(\kappa \approx 3 \cdot \frac{2 F + 1}{2 F + 3}\) and \(F\) corresponds to the hyperfine quantum number of the ground state. This simplified model omits any population which gets recycled back into the ground state after the decay to the metastable state reservoir, but allows for the determination of a lower limit of the decay rate into the intermediate states. The measurements for each species are shown in . Again, we focus on the case of Ho,  (c), as an example. We observe the largest MOT fluorescence signal with a red-detuned laser at \(\approx-1.3\,\Gamma\). Although this is the optimal detuning for creating high trapped atom numbers, the longest MOT lifetime is observed at detunings further to the red. This effect is explained by the lower population in the excited state and, consequently, a lower probability of decaying into a metastable dark state. Increasing the excited state population by moving the laser frequency closer to the transition line center decreases the MOT lifetime almost by a factor of three. The strength of this effect is governed by the decay rate \(\Gamma_1\). The measurement results obtained for each atomic species are summarized in . For the fitting, \(\Gamma_{0}\) and \(\Gamma_{1}\) were free parameters. Our measurements indicate an increase of the decay rates with an increase of the He flow rate for all species. In the table, we quote values of the decay rates which are extrapolated to zero He flow. The collision rate with the background gas \(\Gamma_0\) can be estimated as \(\Gamma^{est}_0 = n_\textrm{\tiny He} \cdot \bar{v}_\textrm{\tiny He} \cdot \sigma_\textrm{\tiny He-Atom}\), where \(n_\textrm{\tiny He} = \frac{P}{k_\textrm{\tiny B} \cdot T}\) is the He gas density and \(k_\textrm{\tiny B}\) is the Boltzmann constant. We assume that He at room temperature (\(T = 300\,\)K) is the dominant part of the background gas. With a cross-section of \(\sigma_\textrm{\tiny He-Atom} \approx 10^{-14}\) cm\(^2\), the velocity \(\bar{v}_\textrm{\tiny He} \approx 1120\) m/s, we find that \(\Gamma^{est}_0\) ranges from 4 to 36 s\(^{-1}\) for steady-state pressures between \(P = 10^{-7}\) and \(10^{-6}\) Torr, which is consistent with our measurements. The results for \(\Gamma_1\) for Yb, Tm and Er are in agreement within the errors with previously reported work. For Ho, both our measurement and the result of the Saffman group are currently the only experimental estimates of the decay rate to metastable states. At present, no theoretical prediction is available. # Conclusion In conclusion, we demonstrated, using Yb, Tm, Er and Ho, the direct loading of magneto-optical traps from a cryogenic two-stage He buffer gas beam source. Despite the presence of the He background gas, we observed lifetimes of up to 80 ms (for Yb). We report a decay rate to metastable states for Ho of \(1510(203)\) s\(^{-1}\). For further applications, we envision an experiment where the buffer gas source and the trapping region are separated by a differential pumping stage to further reduce the residual He background pressure near the MOT. Moreover, for atoms or molecules with magnetic ground states, a magnetic guide could aid in this process by redirecting only the species of interest into a ultra-high vacuum region. The unprecedentedly high loading rate of \(2.0(1.0) \times 10^{10}\) atoms/s, which we demonstrated with Yb, will allow for loading a high number of atoms with loss mechanisms that have a time-scale comparable to the trap loading time into a consecutive trap. As an example, the number of atoms in a blue Yb MOT is limited by spontaneous decay into metastable states. This is typically overcome by using an intermediate MOT with a narrow intercombination line which has no optical leaks. As an alternative, our approach would surpass the need of such an intermediate step and simplify experimental requirements. If developed further, it could produce quantum gases with larger atom number, at a higher repetition rate. This is particularly enticing as the cryogenic setup used is very straightforward. The great flexibility and low intrinsic temperature of this atomic and molecular beam source make it a possible tool for implementing MOTs for any species that precludes the common approach of a combination of a high-temperature oven and Zeeman slower. In particular, recent experiments on slowing and cooling diatomic molecules, all of which successfully employ buffer gas beam sources, reflect this potential advantage. Moreover, our approach constitutes a straightforward testbed for laser cooling and trapping of other atomic species which have not been studied yet. In particular, elements with low vapor pressures at temperatures which are feasible with common oven sources are interesting candidates since laser ablation can produce high densities of such elements in the gas phase. Since the initial cooling stage of our beam source relies solely on collisions with He, only a single laser providing the optical force of the magneto-optical trap has to be adapted to the wavelength of the corresponding species. At the same time, one can imagine the simultaneous loading of multiple species out of the same source. This is a major advantage over regular setups where a separate Zeeman slower is needed for each species loaded into a magneto-optical trap. We would like to thank Jun Ye, Mark Yeo, Matthew Hummon, Alejandra Collopy, Benjamin Stuhl, Hsin-I Lu and Nicholas R. Hutzler for helpful discussions. We also acknowledge the contributions of Matthew J. Wright to the initial phase of this experiment. This work was supported by the NSF and the ARO. We thank Mark Saffman for helpful discussions in the preparation of this manuscript.

{'timestamp': '2014-06-09T02:10:43', 'yymm': '1310', 'arxiv_id': '1310.3239', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3239'}

# Introduction One of the outstanding problems in solar physics is to identify the physical mechanism that gives rise to solar flares. The electron beam heating scenario has been a popular mechanism to explain X-ray emissions in solar flares. Under this mechanism, electrons accelerated high in the corona stream down a coronal loop towards the chromosphere, depositing their energy through collisions, and emitting high-energy X-rays via electron-ion bremsstrahlung with the ambient ions. As they deposit their energy, the footpoints heat up and chromospheric ablation drives material into the corona, filling the loop and producing lower energy X-rays via thermal bremsstrahlung. Solar flares are commonly classified in terms of their peak SXR flux in the 1-8 Å band, as observed by the Geostationary Operational Environmental Satellite (GOES) X-ray flux monitor at terrestrial distance. The largest flares, X-class, have peak fluxes greater than 10\(^{-4}\) W m\(^{-2}\), with smaller flares classified by a decrease in peak flux by factors of 10 as M, C, B, and A-class. The flux in this GOES passband is determined primarily by a combination of thermal and non-thermal bremsstrahlung, and is therefore connected intimately with the properties of the beam of accelerated electrons. A large number of flares (but not all) show the Neupert effect, which states that the fluence of the hard X-rays (HXRs) is proportional to the flux of the SXRs, or equivalently, that the HXR flux is proportional to the time derivative of the SXR flux. In the thick-target model, the deposition of energy by the electron beam induces the observed HXR bursts, subsequently ablating material into the corona and heating the plasma to produce the more gradual SXR light curves. demonstrate that the Neupert effect, combined with this model of chromospheric ablation, should lead to a linear proportionality between the non-thermal energy and the SXR flux; in contradiction to this, show that the SXR flux (as seen by GOES) should be proportional to non-thermal energy \(E^{1.75}\). In the last decade the Reuven Ramaty High-Energy Solar Spectroscopic Imager (RHESSI; @lin2002) and more recently the Fermi Gamma-Ray Burst Monitor (GBM; @meegan2009) have provided excellent coverage of soft and hard X-ray spectra in flares. The spectra obtained with these instruments can be used to derive the properties of electron beams through various inversion methods. The basic method involves inverting the equation relating an observed intensity to parameters of the electron beam (see Equation [\[thicktargetbrem\]](#thicktargetbrem){reference-type="ref" reference="thicktargetbrem"}, below) to solve for the mean electron distribution function (see for example, @holman2003 [@kontar2003; @piana2003]). evaluate four different methods of inversion, concluding that all of the methods recover the general magnitude of a given distribution but have trouble recovering sharp features. Electron beams are primarily characterized by three parameters: the energy flux of the electrons, the spectral index of the electron distribution, and the low-energy cut-off of that distribution. The primary hindrance in determining the electron distribution accurately is the low-energy cut-off. At low energies, thermal bremsstrahlung masks the non-thermal component, rendering accurate determination of the cut-off difficult. Further, most studies assume a simple power-law with a sharp cut-off for the electron distribution, although it is not clear that this model is correct. Numerous models have been developed to study heating processes in solar flares. developed a 1D hydrodynamic model of loops heated by a thermal conduction front to study the formation of SXR emission in flares. Later models, using the heating function derived by, explored the hydrodynamic response of the solar atmosphere to a non-thermal beam of electrons (@nagai1984 [@macneice1984; @mariska1989]). Most of these studies assumed heating lasted for less than a minute, with fixed beam parameters. However, using results from RHESSI, we can now develop a model combining time-dependent beam properties with an advanced hydrodynamic model to forward model spectra and study the effects of an electron beam on observed spectra directly. In this paper, we examine the sensitivity of the GOES classification to the beam properties. We use a numerical model to explore the effect of varying each of the beam parameters on light curves as measured by GOES. In Section [2.1](#hydrad){reference-type="ref" reference="hydrad"} we briefly describe the numerical code used to perform the simulations and several improvements that make it suitable for the studies undertaken here. In Section [2.2](#bhf){reference-type="ref" reference="bhf"} we describe the beam heating function and the specific assumptions we have made in implementing it, and in Section [2.3](#brems){reference-type="ref" reference="brems"} we describe the calculation of bremsstrahlung emissions and the synthesis of GOES light curves from the simulations. In Section [3](#results){reference-type="ref" reference="results"} we present the results of 60 numerical experiments to explore the correlation of GOES classification to each beam parameter derived for two flares observed by RHESSI. Finally, we summarize the results and discuss future work in Section [4](#conclusions){reference-type="ref" reference="conclusions"}. # Numerical Modeling ## The HYDRAD code {#hydrad} The work presented here has been performed using the HYDRAD code, which solves the hydrodynamic equations for an isolated magnetic flux tube and a multi-fluid plasma. The equations (conservation of mass, momentum, and energy) and assumptions are detailed in the appendix of. However, there have been a number of important improvements to the code. First, the pre-flare atmosphere is now based on the VAL model C of the photosphere and chromosphere, allowing for more realistic temperature and density profiles in the lower solar atmosphere. We recalculated the density distribution for hydrostatic equilibrium to be consistent with the average particle mass chosen for our model (\(m_i = 2.171 \times 10^{-24}\) g), which accounts for the relative abundances of hydrogen, helium, and heavier elements (roughly 90% hydrogen and 10% helium). The initial transition region and corona are derived by integrating the hydrostatic equations from the top of the VAL C atmosphere to the apex of the coronal loop. Second, the code has been modified to allow for the presence of neutrals in the lower atmosphere. The two fluids in the code are now electrons and hydrogen atoms (which includes ions and neutrals), with trace amounts of electrons due to ionization of heavier elements. The energy equations have been modified in a manner similar to to include the ionization of hydrogen to account for the effects of neutrals on energy balance between the two fluids (e.g., via collisions and thermal conduction of neutrals, @orrall1961). Finally, the chromospheric radiative energy balance is now based on the recipe derived by, which accounts for cooling from optically thick lines and continua, heating due to the same processes, and heating due to coronal radiation (back-warming from the hot corona). We also used their ionization balance to calculate the electron number density (\(n_e\)), for consistency with this radiation calculation. Figure [\[loweratm\]](#loweratm){reference-type="ref" reference="loweratm"} shows the lower atmosphere density and temperature profiles at the start of Run 1. The loop geometry is semi-circular, along the field-aligned direction. The initial temperature and density profiles were found by integrating the hydrostatic equations from the chromosphere (VAL model C, @vernazza1981) to the apex of the coronal loop (as done, for example, in @aschwanden2001). The electron and ion populations were assumed to be in thermal equilibrium at the beginning of simulations. The loops are cool and tenuous, before significant heating drives chromospheric ablation (compare, *e.g.*, @macneice1984, @nagai1984, @mariska1989). ## Beam Heating Function {#bhf} We have implemented a modified form of the beam heating function derived by. Following and, we add a low-energy knee to the beam electron distribution while keeping the total beam energy constant (also compare the simulations of @warren2006). As those authors note, this form of the electron distribution produces smoother temperature and density variations than a sharp cut-off, without significantly altering the X-ray spectrum (see Sections 3.3 and 3.4 of @holman2011 for comparisons of different shapes of low-energy cut-offs). The distribution function becomes: \[\mathfrak{F}(E_{0}, t) = \frac{4 (\delta-2) F_{0}(t)}{(\delta + 2) E_{c}^{2} } \begin{cases} (\frac{E_{0}}{E_{c}})^{2} & \text{if } E_{0} \leq E_{c} \\ (\frac{E_{0}}{E_{c}})^{-\delta} & \text{if } E_{0} \geq E_{c} \end{cases} \label{distfunc}\] where \(\delta\) is the spectral index above the cut-off energy, \(E_{c}\), and \(F_{0}(t)\) is the energy flux of the beam as a function of time (erg sec\(^{-1}\) cm\(^{-2}\)). The usual caveats apply to this distribution: the number of electrons accelerated is extremely high, there may be instabilities due to Langmuir wave generation, and there may be a return current generated (see @brown1977 for more information). Following, the heat input as a function of position and time is: \[H_{e}(s, t) = \frac{4 \pi e^{4} n_{H} \gamma (\delta-2)}{\mu_{0} (\delta + 2)} \frac{F_{0}(t)}{E_{c}^{2}} \begin{cases} z^{-\delta / 2} B[\frac{\delta}{2}, \frac{2}{4 + \beta}] & \text{if } z > 1 \\ z \int_{z}^{1} \frac{ dy }{y^{2} (1-y)^{2/3}} + z^{-\delta / 2} \int_{0}^{z} y^{\delta / 2-1} (1-y)^{-2 / 3} dy & \text{if } z \leq 1 \end{cases}\] where \(n_{H}\) is the hydrogen density, \(\mu_{0}\) the cosine of the pitch angle of injection, \(\gamma = x \Lambda + (1-x) \Lambda^{\prime}\), \(\beta = [2 x \Lambda + (1-x) \Lambda^{\prime \prime}] / [\Lambda^{\prime} + x (\Lambda-\Lambda^{\prime})]\) (see appendix [\[coulomb\]](#coulomb){reference-type="ref" reference="coulomb"} for explanation of the different Coulomb logarithms \(\Lambda\) used here), \(x\) is the ionization fraction of hydrogen, \(z = N / N_{c}\) for \(N\) the column density and \(N_{c} = \mu_{0} E_{c}^{2} / [ (2 + \beta / 2) \gamma 2 \pi e^{4}]\) the stopping depth of an electron with energy \(E_{c}\), and finally \(B\) is the (complete) beta function. Following the ideas of, we also generalize this expression to non-uniform ionization, which is important for recovering spectral breaks in observed spectra. As noted by those authors, the heating rate is fairly insensitive to the quantity \(\beta\), but it must be constant to integrate Equation 26 of. Therefore, we assume that \(\beta = 2\) (corresponding to a fully ionized plasma, although we do *not* assume \(x=1\) in general) and integrate. Second, we modify the column depth to take into account the varying ionization structure (an equivalent ionized column depth, \(N^{*}(N) = \int_{0}^{N} \frac{\gamma}{\Lambda} dN^{\prime}\)). The heating function in this form is then evaluated numerically as a function of time and position, given the beam parameters. ## Bremsstrahlung emissions {#brems} An important element of this work is predicting the GOES class of our model flares and so we must evaluate both thermal and non-thermal bremsstrahlung as a function of time. We note the following important properties of GOES relevant in evaluating this emission. (1) GOES observes in the bands 1-8 Å (1.55 to 12.41 keV) and 0.5-4 Å (3.10 to 24.82 keV), with flares being classified by the peak flux in the former. (2) GOES light curves are not spatially resolved; *i.e.* GOES does not distinguish coronal sources from foot-points. (3) GOES light curves are a summation of thermal and non-thermal bremsstrahlung, as well as background emission. We evaluate both thermal and non-thermal bremsstrahlung everywhere along the flaring loop as a function of time and sum the emission to construct a light curve in the native units of GOES (W m\(^{-2}\)). To evaluate the thermal bremsstrahlung, we use Equations 1, 2, and 3 of, derived for solar flares as observed at terrestrial distance. These equations include contributions from both free-free and free-bound emission. We approximate the emission measure by \(\int_{V} n^2 dV \approx n_{e} n_{H} A \Delta s\), where \(A\) is an estimate of the area based on observations and \(\Delta s\) is the length of a given loop segment. Making these changes, we have, in units of photons sec\(^{-1}\) cm\(^{-2}\) keV\(^{-1}\): \[I_{thermal} = 3.6 \times 10^{-39} \overline{Z^{2}} k_{B}^{0.3} T_{e}^{-0.2} \epsilon^{-1.3} \exp{\Bigg( \frac{-\epsilon}{k_{B} T_{e}}\Bigg)} \Bigg[1-\Big(\frac{\epsilon}{88.0}\Big)^{k_{B} T_{e} / 3}\Bigg]^{-1} n_{e} n_{H} A \Delta s\] where \(\overline{Z^{2}} = 1.4\) is the average charge-squared of ions and \(\epsilon\) is the emitted photon energy (in keV). Note that this function is only valid for photon energies \(1.5 \leq \epsilon \leq 15\) keV, which encompasses the full 1-8 Å range, and is only valid for temperatures below about 20 MK. Outside of this range, we use the following expression, derived by combining Equations 1 and 1a of their paper: \[I_{thermal} = 3.6 \times 10^{-39} \overline{Z^{2}} T_{e}^{-0.5} \epsilon^{-1.0} \exp{\Bigg( \frac{-\epsilon}{k_{B} T_{e}}\Bigg)} n_{e} n_{H} A \Delta s\] These functions will give the emitted thermal bremsstrahlung at a given spatial location along the loop at a given time. Since GOES has no spatial resolution, the total emission is found by spatially integrating along the loop and over the time period of interest. Non-thermal bremsstrahlung is more difficult to evaluate. It requires knowledge of the beam electron distribution function and the appropriate cross-section of interactions. Following either or, thick-target bremsstrahlung observed at terrestrial distance is given in general by (in photons sec\(^{-1}\) cm\(^{-2}\) keV\(^{-1}\)): \[I_{thick} = \frac{A}{4 \pi R^{2}} \int_{\epsilon}^{\infty} \mathfrak{F}(E_{0},t) \nu(\epsilon, E_{0}) dE_{0} \label{thicktargetbrem}\] where \(R = 1.496 \times 10^{13}\) cm (1 AU), \(A\) is the loop area (cm\(^{2}\)), \(\mathfrak{F}(E_{0},t)\) is the electron flux spectrum of injected electrons (as before, although converted to units of electrons sec\(^{-1}\) cm\(^{-2}\) keV\(^{-1}\)), \(E_{0}\) is the initial electron energy, and \(\nu\) (below) essentially gives the photon yield for a given electron: \[\nu(\epsilon, E_{0}) = \int_{E_{0}}^{\epsilon} \frac{ n_{H} v Q({\epsilon, E}) dE}{dE/dt} \label{nu}\] for \(v\) the electron velocity, \(Q({\epsilon, E})\) the cross-section of the interaction (cm\(^{2}\) keV\(^{-1}\)), and \(\frac{dE}{dt}\) the energy lost by the electron per unit time (keV sec\(^{-1}\)). Note that \(\nu\) has units of photons electron\(^{-1}\) keV\(^{-1}\). Following, we take \[\frac{dE}{dt} = n_{H} v \frac{dE}{dN} = n_{H} v \Bigg( \frac{-2 \pi e^{4}}{E} \Big[x \Lambda + (1-x) \Lambda^{\prime}\Big] \Bigg) \label{dedt}\] where we now evaluate over the column density, allowing for non-uniform ionization (with variables as defined in previous sections). Note that for unit consistency, \(e^4\) must be in units of cm\(^{2}\) keV\(^{2}\) with \(E\) in keV. We can reverse the order of integration in Equation [\[thicktargetbrem\]](#thicktargetbrem){reference-type="ref" reference="thicktargetbrem"}: \[I_{thick} = \frac{A}{4 \pi R^{2}} \int_{\epsilon}^{\infty} \frac{ n_{H} v Q({\epsilon, E}) dE}{dE/dt} \int_{E}^{\infty} \mathfrak{F}(E_{0},t) dE_{0} \label{reversebrem}\] where the inside integral is now analytic, allowing for easier numerical integration regardless of what form of cross-section is chosen. Evaluation of this integral is explained in detail in Appendix [\[reverse\]](#reverse){reference-type="ref" reference="reverse"}. We must now decide what cross-section is appropriate for non-thermal bremsstrahlung emissions. At the energies we are considering (1.55 to 12.41 keV and 3.10 to 24.82 keV), an appropriate choice is the Bethe-Heitler cross-section, with the Elwert correction factor, given by: \[\begin{aligned} Q(\epsilon, E) &=& \frac{16 \overline{Z^{2}} r_{0}^{2} \alpha}{3} \frac{m_{e}^{2} c^{4}}{\epsilon E (E + 2m_{e}c^{2})} \ln{\Bigg[ \frac{1 + \Big( \frac{(E-\epsilon)(E-\epsilon + 2m_{e}c^{2})}{E (E+2m_{e}c^{2})} \Big)^{1/2}}{1-\Big( \frac{(E-\epsilon)(E-\epsilon + 2m_{e}c^{2})}{E (E+2m_{e}c^{2})} \Big)^{1/2}} \Bigg]} \\ && \times \frac{[E(E+2m_{e}c^{2})]^{1/2} [E-\epsilon + m_{e}c^{2}] \Big[1-\exp{\Big(-\frac{2 \pi \alpha [E + m_{e}c^{2}]}{[E(E+2m_{e}c^{2})]^{1/2}} \Big)} \Big] } {[(E-\epsilon)(E-\epsilon+2m_{e}c^{2})]^{1/2} [E + m_{e}c^{2}] \Big[1-\exp{\Big(-\frac{2 \pi \alpha [E-\epsilon + m_{e}c^{2}]}{[(E-\epsilon)(E-\epsilon +2m_{e}c^{2})]^{1/2}} \Big)} \Big] } \nonumber \label{cross} \end{aligned}\] with \(\alpha\) the fine-structure constant and \(r_{0}\) the classical electron radius. We can then numerically integrate the thick-target emissions with the use of Gauss-Laguerre and Gauss-Legendre quadrature, as appropriate. To facilitate comparisons to observational data, the response of the GOES instrument must be accounted for in the bremsstrahlung calculations. describe the response of the X-ray sensor (XRS) aboard GOES-8 in depth. Note that the sensitivity of the GOES XRS has remained nearly constant for each spacecraft. We convolve the response function (given by Equation 1 and Figure 3 of @hanser1996) with all bremsstrahlung calculations to improve our estimates of GOES classification for the flares studied in this paper. # Results We performed 60 numerical experiments exploring the correlation of the GOES classification to the beam properties. We study two flares observed with RHESSI and base our simulations on their parameters. The results are split into two sets. The first set is based on beam parameters for the 2002-04-15 M1.2 flare (Table [1](#20020415sim){reference-type="ref" reference="20020415sim"}) derived by, with a cut-off energy \(E_{c} = 24\) keV, a spectral index \(\delta(t)\) ranging between about 6 and 10, and a beam power ranging from approximately \(6 \times 10^{25}\) to \(5 \times 10^{27}\) erg sec\(^{-1}\) (see Figure 6 of their paper). The second set is based on 2002-07-23 X4.8 flare (Table [2](#20020723sim){reference-type="ref" reference="20020723sim"}) derived by, with a cut-off \(E_{c}(t)\) ranging from about 18 to 43 keV, spectral index \(\delta(t)\) ranging from about 2.5 to 8, and a beam power between \(7 \times 10^{26}\) and \(2 \times 10^{29}\) erg sec\(^{-1}\) (see Figure 3 of their paper). In both cases, the beam parameters were derived from RHESSI observations of the flares. The geometry of the 2002-04-15 flare loop (that is, length and cross-sectional area) is based on the estimate of. The length of the 2002-07-23 flare loop is estimated from the foot-point locations in; the area is estimated from, decreased slightly to \(7.0 \times 10^{18}\) cm\(^{2}\) to calibrate the base run (\# 31) to the observed GOES classification. In each group of simulations, we vary one beam parameter per simulation in the following way (holding all other values constant). First, we multiply the beam flux \(F_{0}(t)\) by a factor of \[\(\frac{1}{10},\frac{1}{5},\frac{1}{3},\frac{1}{2},2,3,5,10\)\]. Second, we change the spectral index of the electron distribution \(\delta(t)\) by \[\(-3,-2,-1,-\frac{1}{2},\frac{1}{2},1,2,3\)\] (although we limit \(\delta\) to 2.01 since it must be strictly greater than 2). Finally, we change the cut-off energy \(E_{c}(t)\) by \[\(-20,-15,-10,-5,5,10,15,20,30,40,50,75,100\)\] keV (limiting \(E_{c}\) to a minimum of 1 keV). Table [1](#20020415sim){reference-type="ref" reference="20020415sim"} summarizes the results for the set of simulations concerning the 2002-04-15 flare. Table [2](#20020723sim){reference-type="ref" reference="20020723sim"} summarizes the results for the set of simulations concerning the 2002-07-23 flare. In both cases, we list the run number along with the changes in the beam parameters, the GOES class (in both channels), the maximum temperature (in MK), and the maximum apex density (cm\(^{-3}\)). For example, consider Figure [\[20020415base\]](#20020415base){reference-type="ref" reference="20020415base"}, which shows the results of Run 1. The beam lasts 640 seconds, for a loop of length \(2L = 77\) Mm. The upper panels show the electron temperature (reaching a peak of 27.3 MK) and density (with a maximum apex value near \(6 \times 10^{10}\) cm\(^{-3}\)) as functions of position, at different times as indicated. The temperature rises for around 300 seconds, when the beam begins to weaken. There is significant radiative cooling in the chromosphere; in the corona, thermal conduction drives energy losses initially, gradually transitioning to radiative cooling as the density rises and temperature falls. The density in the corona rises due to chromospheric ablation throughout the duration of heating, and the maximum intensity occurs when the heating ceases (a few minutes after it reaches its maximum temperature). The center left panel similarly shows the heat deposition (which includes background heating). As the loop fills due to chromospheric ablation, the heat deposition occurs higher and higher in the corona as the mean-free path of beam electrons decreases. Although the beam continues for 640 seconds, the beam flux begins to fall around 300 seconds, and thus the heat deposition begins to fall. The center right figure shows the spectrum calculated at one of the footpoints at a few selected times. Initially (\(t = 150\) sec), the emission is entirely non-thermal bremsstrahlung, but as the loop begins to heat and fill, thermal bremsstrahlung eventually becomes the dominant component of the emission at lower energies (e.g., at 600 seconds, the thermal emission at 1 keV is more than 100 times stronger than the non-thermal emission). Finally, the predicted GOES light curve (with 1-8 Å in red and 0.5-4 Å in blue) is shown at bottom right, with the relative amounts of thermal and non-thermal bremsstrahlung indicated. The light curve (1-8 Å) peaks at \(1.1 \times 10^{-5}\) W m\(^{-2}\), corresponding to M1.1, while the 0.5-4 Å component peaks at \(3.1 \times 10^{-6}\) W m\(^{-2}\) (C3.1). The actual (background subtracted) GOES light curves, starting at 15 April 2002 23:07 UT, are overlaid for comparison (peaking at M1.1 and C2.0, respectively). The light curve is initially non-thermal bremsstrahlung, but as the loops heats and fills up, thermal bremsstrahlung increases until it becomes the dominant component in the passband. Similarly, Figure [\[20020723base\]](#20020723base){reference-type="ref" reference="20020723base"} shows the evolution of Run 31. The beam lasts for 1220 seconds, for a loop of length of \(2L = 36\) Mm. The predicted light curve in this case peaks at \(3.6 \times 10^{-4}\) W m\(^{-2}\) (X3.6) in the 1-8 Å channel and \(1.1 \times 10^{-4}\) W m\(^{-2}\) (X1.1) in the 0.5-4 Å channel. As before, the temperature rises sharply due to the energy deposition, which then triggers chromospheric ablation. As the loop fills and heats up, the contribution of thermal bremsstrahlung to the light curve rises sharply. In this case, the loop begins to cool and drain before the beam ceases. In both cases, the GOES class is in approximate agreement with the observations. Note that the results are fairly insensitive to the initial density profile. If we had assumed an initial coronal density of \(10^{10}\) cm\(^{-3}\), over a loop length of 16 Mm, we would find that only electrons with energy \(\lesssim 12\) keV would be stopped in the corona, and thus the majority of energy would still be deposited in the chromosphere (using the estimate of stopping depth from @nagai1984). We now turn our attention towards altering the beam parameters one at a time, to examine their effect on the GOES class. First, holding all other values constant, we multiply the beam flux by \[\(\frac{1}{10},\frac{1}{5},\frac{1}{3},\frac{1}{2},2,3,5,10\)\] for both flares (equivalent to multiplying the total non-thermal energy by the same amount). There is a clear correlation between the GOES class and the beam flux (see the first plot in figure [\[goesbeam\]](#goesbeam){reference-type="ref" reference="goesbeam"}). For Runs 1-9, we find that the GOES class \(\Psi\) (in the 1-8 Å channel) is proportional to the beam energy \(E\), such that \(\Psi \propto E^{\alpha}\) for \(\alpha = 1.71 \pm 0.084\); for Runs 31-39, we find \(\alpha = 1.77 \pm 0.060\). In the 0.5-4 Å channel, we find a similar trend: for Runs 1-9, we find \(\Psi \propto E^{\alpha}\) for \(\alpha = 1.47 \pm 0.103\) and for Runs 31-39 \(\alpha = 1.60 \pm 0.077\). In Runs 2-5 and 32-33 (and 34 in the 0.5-4 Å channel), the loops were not heated enough to produce significant thermal bremsstrahlung, and thus their GOES classes were primarily determined by non-thermal emissions. Note that non-thermal emissions are linearly proportional to the beam flux (Equation [\[thicktargetbrem\]](#thicktargetbrem){reference-type="ref" reference="thicktargetbrem"}); if we fit a line to just Runs 2-5, we find an exponent \(\alpha = 0.98 \pm.023\) in the 1-8 Å channel and \(\alpha = 1.00 \pm 0.014\) in the 0.5-4 Å channel, confirming that the non-thermal emission is proportional to the mean beam energy. In all the other runs, the peak of the light curves was a combination of thermal and non-thermal emissions, with thermal emissions being indirectly related to the beam flux. If we remove the simulations without thermal emissions and recalculate, we find exponents \(\alpha = 1.94 \pm 0.026\) (1-8 Å) and \(\alpha = 2.04 \pm 0.054\) (0.5-4 Å) for Runs 1 and 6-9, and \(1.91 \pm.014\) (1-8 Å) for Runs 31 and 34-39 and \(1.84 \pm 0.050\) (0.5-4 Å) for Runs 31 and 35-39. Thus, the maximum GOES class \(\Psi\) is related to the total beam energy \(E\), \(\Psi \propto E^{\alpha}\) for \(\alpha \approx 1.7\) (1-8 Å) and \(\alpha \approx 1.6\) (0.5-4 Å). There are two reasons why the values differ slightly for each flare: different loop lengths and different cut-off energies, both of which affect the amount of heating and thus thermal bremsstrahlung produced. This result should be compared to that of, who showed that \(\Psi \propto E^{1.75}\) (1-8 Å) and \(\Psi \propto E^{2.25}\) (0.5-4 Å) using simple analytic considerations. Our results are in agreement for the 1-8 Å channel, while they are in contrast in the higher energy channel. If the emission were entirely non-thermal, then we would find a linear relationship (\(\Psi \propto E\), Equation [\[thicktargetbrem\]](#thicktargetbrem){reference-type="ref" reference="thicktargetbrem"}), regardless of which channel is being considered. Since the 0.5-4 Å channel is more sensitive to higher energies, we would expect it to be closer to linear than the lower energy channel as it would be more sensitive to the non-thermal emissions, and thus the exponent \(\alpha\) should be lower. The reason for this contrast is perhaps their use of the hydrostatic scaling laws derived for the corona, whereas the non-thermal emissions will be primarily at chromospheric depths. It should also be noted that our result, as well as that of, are in contrast to what is expected from the Neupert effect. As shown by by using the Neupert effect relation, the maximum flux in the SXRs should be linearly proportional to the energy deposited by the electron beam. suggest that this difference is due to the peak SXR flux being dependent upon other flare parameters, to which we now turn our attention. Next, holding all the other values constant again, we add \[\(-3,-2,-1,-\frac{1}{2},\frac{1}{2},1,2,3\)\] to the spectral index \(\delta(t)\). Varying the spectral index of the electron distribution has only a small effect on the GOES class (see the second plot in Figure [\[goesbeam\]](#goesbeam){reference-type="ref" reference="goesbeam"}). For Runs 1 and 10-17, with a change of \(\delta \pm 3\), the GOES flux (1-8 Å) varies by a factor of 2, and about a factor of 3 in the higher energy channel; for Runs 31 and 40-47, the flux varies by less than a factor of 2 in both channels. In all cases, though, an increase in the spectral index does increase the GOES class slightly in the 1-8 Å channel and decreases it slightly in the 0.5-4 Å channel. The reason is that an increase in the index reduces the proportion of high-to low-energy electrons, so that more energy is deposited higher up in the atmosphere, increasing the temperature and thus the amount of thermal bremsstrahlung, while slightly reducing the non-thermal emissions. Thus, the channel which is more sensitive to thermal emissions increases slightly, and the channel more sensitive to non-thermal emissions decreases slightly. The immediate question is whether this result is due to the assumed form of the beam electron distribution (*i.e.* the importance of the low-energy knee electrons). Runs 1, 10-17, 31, and 40-47 were repeated using a sharp low-energy cut-off (\(\mathfrak{F}_{0}(E_{0},t) = 0\) for \(E_{0} < E_{c}\)), but equivalent amount of total beam energy (see Figure [\[indexsharp\]](#indexsharp){reference-type="ref" reference="indexsharp"}). Although the slopes are larger in this case, they remain small. The conclusion is unaltered: the spectral index only marginally affects the GOES class. The primary reason that the spectral index minimally affects the GOES flux is because the total beam energy was held constant. Changing the spectral index is equivalent to slightly altering the proportion of high-to low-energy electrons, which in turn affects the mean location of energy deposition. Increasing the index decreases the number of high-energy electrons in the beam, which then causes deposition of energy slightly higher up the loop (and vice versa). Compare, who found that the photon spectral index \(\gamma\) is not correlated to the peak HXR flux (and in turn, \(\delta = \gamma + 1\), so that the relation holds for the electron index as well). The cut-off energy of the electron distribution can alter the GOES class significantly (see the third plot in Figure [\[goesbeam\]](#goesbeam){reference-type="ref" reference="goesbeam"}). Slight decreases in the cut-off increase the GOES flux, but large decreases will significantly decrease the GOES flux. On the other side, as the cut-off increases more and more, the energy is deposited deeper in the dense lower atmosphere, where there is a greater heat capacity and the energy is radiated away more quickly, leading to less heating, less thermal bremsstrahlung, and thus a lower GOES class in the 1-8 Å channel. In the 0.5-4 Å channel, increases in the cut-off increase the amount of non-thermal bremsstrahlung, so the flux in this channel tends to increase with higher cut-off energies. At extremely high cut-offs, the energy is deposited too low to heat the loop, so no thermal bremsstrahlung will be emitted; however, there will still be non-thermal emissions as the beam traverses the loop, which will be determined primarily by the energy flux in the beam. Eventually, as the cut-off is increased to large enough values, the emission in both channels will be entirely non-thermal bremsstrahlung, to which the higher energy channel is more sensitive. As the cut-off decreases, more heat is deposited higher up, leading to a higher temperature and more thermal emission. If it decreases too much though, the beam is then essentially composed of thermal electrons, leading to less non-thermal bremsstrahlung emission and a lower GOES class (in both channels). The cut-off energy is directly related to the location of energy deposition of the beam electrons. give the mean stopping column density of an electron with energy \(E_{c}\) as \(\approx 10^{17} [E_{c}\) (keV)\(]^{2}\) cm\(^{-2}\), which can then be used to find the location of maximal energy deposition. In Figure [\[deploc\]](#deploc){reference-type="ref" reference="deploc"}, we show the predicted and actual location of maximal energy deposition as a function of time for Runs 1 and 20 in one half of the loop. In Run 20, it should be noted, the coronal density reached higher values and so the deposition location is in general higher than in Run 1, leading to more heating and more thermal bremsstrahlung, and thus a higher GOES class. There is also a clear relation between the maximum temperature of a flare and the maximum GOES flux. Figure [\[classtemper\]](#classtemper){reference-type="ref" reference="classtemper"} shows the maximum electron temperature versus the maximum GOES flux found in the 60 simulations. There is a clear correlation between the two, although there are two separate trends for each group of 30 simulations. This is due to differences in the spectral index. The index was much higher at all times in the M1.2 flare, and thus more energy was deposited higher in the atmosphere, leading to consistently higher temperatures despite lower GOES classes. The spectral index therefore determines whether a flare will be thermally driven or beam dominated. Note, however, that in each case there is a horizontal branch extending to lower temperatures with roughly constant GOES flux. These branches are the cases where there are extremely high cut-off energies, which deposit their energy too low to heat the loop, but still produce roughly constant non-thermal emissions. These results should be compared to Figure 6 and Equation 6 of, who found a similar trend, but significantly lower temperatures. These lower temperatures result from the fact that they were measuring the temperature at the time of peak emission of the Fe XXV or Ca XIX channels of the Bent Crystal Spectrometer on SMM, which occurs after the time of maximum temperature (the density continues to rise). For example, in Run 1, the maximum temperature was 260 seconds into the flare, while the maximum intensity of the Ca XIX line occurred 500 seconds into the flare, a 4-minute discrepancy, by which time the loop had cooled by 7 MK (see @bradshaw2011 for explanation of the forward modeling of spectral lines such as the Ca XIX line). There is some observational evidence for cold flares, which have detectable GOES emission but low temperatures. For example, report a C-class flare with a maximum temperature of 6 MK. The authors suggest that the low temperature is due to a weak beam flux, which leads to little heating, so that the GOES emission would then be composed primarily of non-thermal bremsstrahlung. This explanation agrees with our results, where small beam fluxes produce low temperatures, but emission detectable by GOES (see the lower left of Figure [\[classtemper\]](#classtemper){reference-type="ref" reference="classtemper"}). One other possibility, which seems less likely, is that the flare had a large cut-off energy so that the energy was deposited deep in the chromosphere, once again leading to minimal heating but some emission in the GOES passbands nonetheless (see the horizontal branches in Figure [\[classtemper\]](#classtemper){reference-type="ref" reference="classtemper"}). # Summary and Conclusions {#conclusions} We have investigated the effect of individual beam heating parameters on the GOES classification of solar flares. Using two independent sets of observationally determined parameters, 60 numerical experiments have been performed, from which GOES light curves have been synthesized. We have found clear trends between the beam parameters and the GOES flux. First, the GOES classification strongly depends on the total beam energy (\(\Psi \propto E^{\alpha}\) with \(\alpha\) around 1.7 in the 1-8 Å channel and around 1.6 in the 0.5-4 Å channel). This parameter dominates: the SXR spectrum is primarily determined by the amount of heat deposited in the solar atmosphere. Second, the spectral index of the beam electron distribution does not significantly alter the GOES class (whether the cut-off is sharp or a knee). There is a small positive correlation, though: an increase in \(\delta\) slightly increases the GOES class. Finally, the cut-off energy, which determines the mean location of energy deposition of the beam, affects the amount of both the thermal and non-thermal bremsstrahlung produced. Thus, changes to the cut-off can either increase or decrease the GOES flux, but it is not a simple relation, unlike the other two parameters. As noted earlier, the Neupert effect predicts a linear relation between non-thermal energy and SXR flux (\(\Psi \propto E\)), which is in disagreement with the present results and those of. In this paper we have shown that the GOES flux strongly depends on the cut-off energy of the electron beam as well. A change in the cut-off energy can alter the relation between non-thermal energy and SXR flux. Consider the case of an extremely high cut-off energy where there is little heating and thus little thermal bremsstrahlung: all the emission will be non-thermal, and thus a linear relation would follow (Equation [\[thicktargetbrem\]](#thicktargetbrem){reference-type="ref" reference="thicktargetbrem"}). Although it seems unlikely that such extremely high cut-off energies occur, it is clear that the cut-off energy can affect the relation. We can constrain the parameters as well. For example, Runs 38 and 39 have a GOES class larger than any flare ever observed, and thus can be discarded as unlikely to occur, setting an effective upper limit on the non-thermal energy budget in flares (the largest ever seen with GOES was around X40, @brodrick2005). The total non-thermal energy in the 2002-07-23 flare was \(2.6 \times 10^{31}\) erg, so there is an effective upper limit around \(\approx 7.5 \times 10^{31}\) erg (for similar cut-off energies). The extremely high cut-off energies used in some of the runs were also unrealistic. For example, Runs 55-60 are X-class flares with maximum temperatures beneath 10 MK, and densities too low to be observed, suggesting that such high cut-offs are not realistic (at least for the entire duration of the flare, there is some evidence that they may reach high cut-offs temporarily: @warmuth2009). The SXR spectrum is only one facet of emissions produced by solar flares, though. In future work, we will predict extreme ultraviolet (EUV) spectra to further constrain the model. The properties of EUV lines, such as widths, Doppler shifts, and intensities, can be predicted from the model and directly compared with observations to help pin down signatures of the fundamental driving mechanism of solar flares. For example, point out that while the 4 November 2003 flare had the largest GOES flux ever measured, the 28 October 2003 flare was more intense in the EUV. The model developed here needs to be used to explain these differences as well. One of the authors (JWR) was supported by NASA Headquarters under the NASA Earth and Space Science Fellowship Program-Grant NNX11AQ54H. We thank the anonymous referee for detailed comments that substantially improved the paper. We also thank our colleagues at the SHINE and SPD conferences this summer for their helpful discussions and suggestions.

{'timestamp': '2013-10-14T02:11:07', 'yymm': '1310', 'arxiv_id': '1310.3242', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3242'}

{'timestamp': '2013-11-27T02:02:18', 'yymm': '1310', 'arxiv_id': '1310.3210', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3210'}

# Introduction {#Intro} The universal rate for the conversion of gas into stars in galaxies peaks at redshifts z \(\sim 1.5-3\) (e.g. [@2006ApJ...651..142H]). At this epoch star forming galaxies drive powerful galactic winds which can transport a significant fraction of the gas away from the central galaxy making it temporarily unavailable for star formation (e.g. [@2000ApJ...528...96P; @2003ApJ...588...65S; @2009ApJ...692..187W; @2010ApJ...717..289S; @2011ApJ...733..101G; @2012ApJ...752..111N; @2012ApJ...761...43N]). Spatially resolved high-redshift observations indicate that these winds are launched directly from the sites of the-typically strongly clustered-star formation. The estimated outflow rates \(\dot{M}_{\mathrm{out}}\) can be several times higher than the star formation rates (SFR). The resulting high mass loading \(\eta = \dot{M}_{\mathrm{out}}/\textrm{SFR}\) indicates that-even at the peak epoch of cosmic star formation-the amount of gas expelled from the galaxies is comparable to the amount of gas converted into stars inside the galaxies. The direct observational evidence for inefficient conversion of gas into stars is supported by indirect constraints from halo abundance matching techniques. Here the galaxy formation efficiency can be defined as the fraction of the stellar mass of a galaxy to the total available baryonic mass of its host dark matter halo. In a concordance \(\Lambda\)-CDM cosmology this efficiency peaks-almost independent of redshift-for galaxies in halos of about \(10^{12}\hbox{\)\mathrm{\thinspace M_{\odot}}\(}\) and never exceeds \(\sim 20-25 \%\) (e.g. [@2010ApJ...710..903M; @2010ApJ...717..379B; @2010MNRAS.404.1111G; @2013MNRAS.428.3121M]). Therefore at least \(3/4\) (significantly more in halos of higher as well as lower mass than \(10^{12}\hbox{\)\mathrm{\thinspace M_{\odot}}\(}\)) of the baryonic material is never converted into stars, eventually due to powerful galactic winds. It is plausible that the main physical processes responsible for driving the outflows also regulate the efficiency with which the available gas is converted into the stellar components of galaxies in the Universe. In connection to the evolution of stellar populations a number of physical processes are-in principle-energetic enough to expel gas from star forming galactic disks. Besides AGN for high mass galaxies, type II supernovae have long been considered the most promising candidates, in particular for lower mass galaxies. Although the amount of energy per event is significant the thermal energy is mainly deposited at the sites of star formation, i.e. dense molecular clouds. Here the cooling times are very short and the energy can be efficiently radiated away making it difficult but not impossible to drive large scale galactic winds (for recent discussions see e.g. [@2011MNRAS.415.1051B; @2012MNRAS.426..140D]). However, even before the supernova explosions, the momentum and energy input from massive stars in the form of stellar winds and stellar luminosity is significant and might support the wind driving. In this paper we focus on a separate mechanism: the formation of large scale magnetized galactic winds driven by cosmic rays. As supernovae drive strong shocks into the interstellar medium some fraction of the explosion energy is consumed to accelerate ionized particles to relativistic energies which are then injected into the ISM as cosmic rays (CR). This relativistic fluid is coupled to the galactic magnetic field and-in particular the hadronic component-is less prone to energy losses than the gaseous component of the ISM. Analytic estimates and numerical experiments without or only approximate inclusions of galactic magnetic fields clearly indicate that CRs can help driving large scale galactic winds. However, CRs are strongly coupled to magnetic fields whose evolution should be followed in a self-consistent way. have shown that CRs promote buoyancy effects in the interstellar medium, leading to the break-out of magnetic fields from galactic disks and, at the same time, to magnetic field amplification by CR-driven dynamo action. Plausibly, such processes are also relevant for star forming galaxies at high redshift which are observed to have significant magnetic fields at the level of tens of \(\mu G\).. Recent observations even demonstrate the existence of large magnetic fields up 50 kpc away from the galaxy indicating strong large-scale magnetized winds. In this letter we present a three-dimensional full MHD simulation of a massive gas-rich disk galaxy (section [2](#numerical_setup){reference-type="ref" reference="numerical_setup"}) and follow the formation of large scale magnetized winds as a dynamic response to the injection of CRs (section [3](#simulations){reference-type="ref" reference="simulations"}). This is considered as a proof of principle for the importance of this physical process. The most important implications are presented in section [4](#discussion){reference-type="ref" reference="discussion"}. # Numerical setup {#numerical_setup} For the simulations we use the PIERNIK MHD code, a grid-MHD code based on the Relaxing TVD (RTVD) scheme by [-@jin-xin-95] and [-@2003ApJS..149..447P]. The induction equation, including the Ohmic resistivity term, is integrated with a constraint transport (CT) algorithm. The original scheme is extended to include dynamically independent, but interacting fluids: thermal gas and a diffusive CR gas, described within the fluid approximation.. We incorporate selfgravity of interstellar gas and gravitational potential is obtained by solving the Poisson equation inside the computational domain with an iterative, multi-grid solver combined with a multipole solver to properly treat the gravitational potential at 'isolated' boundaries. We assume a fixed gravitational field due to the stellar disk and the dark matter halo and compute the gravitational potential using the model of with \(M_{\textrm{halo}} = 8 \cdot 10^{11} \hbox{\)\mathrm{\thinspace M_{\odot}}\(}\) within \(R_{\textrm{cutoff}} = 100{\rm\thinspace kpc}\) and \(M_{\textrm{disk}} =8.6 \cdot 10^{10} \hbox{\)\mathrm{\thinspace M_{\odot}}\(}\). We neglect the contribution of a central bulge. Fresh gas is supplied to the disk at a fixed rate of \(\dot{M}_{\textrm{in}} = 100 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1}\) following the initial gas density distribution. To simplify the setup the gas is added directly at the disk plane. This is clearly a simplified model. However, another simplified way of treating the gas supply by spherical accretion is much more difficult to control. Even in this case we expect that the highly collimated winds forming in our simulation would punch through spherically accreting gas not changing our conclusions. The more realistic alternative of gas accretion along filaments would provide fresh gas, carrying high angular momentum, at disk peripheries and is not expected to change the wind properties significantly. We construct a three-dimensional gas distribution \(\rho_0(x,y,z)\) (e.g. [@1998ApJ...497..759F]). At every time step we add gas at the given inflow rate \(\Delta \rho (x,y,z) = \dot{M}_{\textrm{in}}/M_{\textrm{0}} \rho_0(x,y,z) \Delta t\) within the disk volume. The disk collects gas until it becomes gravitationally unstable. We assume that star formation is controlled by a star formation efficiency parameter \(\epsilon \le 1\). In the actual models we assume \(\epsilon = 0.1\). The star formation rate per unit volume is computed as \[\rho_{\textrm{SFR}} = \epsilon \sqrt{\frac{G\rho^3}{32\pi}}\] provided that gas density exceeds some threshold density \(\rho_{\textrm{thr}}\), which we treat as a free parameter. Its value (\(\simeq 600\textrm{ H atoms }{\rm\thinspace cm}^{-3}\)) is adopted to obtain highly localized star formation, and to regulate the overall SFR of the galaxy. We locally deplete the gaseous ISM at the same rate at every timestep. We assume that one supernova occurs per \(100\hbox{\)\mathrm{\thinspace M_{\odot}}\(}\) of gas forming new stars, and that \(10\%\) of the explosion energy is used to accelerate CRs. Individual CR particles propagate at relativistic speeds, however fast streaming of CRs along magnetic field lines leads to streaming instabilities, the generation of small-scale turbulence, and subsequently the scattering of CRs on self-excited turbulence. This implies that a CR pressure gradient term has to be taken into account in the gas equation of motion. The bulk motion of CRs is considered as a combination of diffusive and advective propagation and can be described by the diffusion-advection transport equation. In our model CRs diffuse preferentially along magnetic field lines, while diffusion perpendicular to the magnetic field is significantly much less efficient. A more elaborated picture of CR propagation should include additional effects, such as energy conversion from CRs to waves, energy sinks for MHD waves due to ion-neutral collisions and nonlinear Landau damping (for a detailed discussion see e.g. and references therein). The above mentioned authors adopt constant CR diffusion coefficients ranging from \(K= 10^{27} {\rm\thinspace cm}^2 {\rm\thinspace s}^{-1}\) up to \(10^{30} {\rm\thinspace cm}^2 {\rm\thinspace s}^{-1}\) in their 1-D numerical models. The assumption of constant diffusion coefficient implies that the large-scale diffusion velocity \(v_\mathrm{diff} =-K\nabla e_\mathrm{CR}/e_\mathrm{CR}\) may become larger than the Alfvén speed in regions of steep gradients of the CR energy density. This can occur in the disk, and especially around CR production regions, where molecular and neutral gas components dominate. Since Alfvén waves are efficiently dissipated there by ion-neutral damping, the most important contribution to the random magnetic field in the disk is induced by supernova explosions. Therefore ion-neutral collisions should reduce the amplitude of Alfvén waves, leading to enlargement of the mean free path of the CR particles, and therefore to higher diffusion coefficients. The numerical algorithm of the anisotropic CR propagation, within the framework of staggered mesh MHD code, has been described in . The values of CR diffusion coefficients, parallel and perpendicular to the magnetic fields adopted for the actual models are \(K_\parallel = {3\cdot{10}^{28}}{\centi\meter\squared\per\second}\), \(K_\perp= {3\cdot {10}^{26}}{\centi\meter\squared\per\second}\). Initially the toroidal magnetic field pervading the disk has a strength of \(3 \mu G\) and a uniform magnetic diffusivity \(\eta={3\cdot{10}^{25}}{\centi\meter\squared\per\second}\), corresponding to a standard value of turbulent diffusivity of the ISM. The simulations have been performed at a resolution of \(512^3\) grid cells, distributed among equal-sized MPI blocks, in the Cartesian domain spanning a volume of \({100^3}{{\rm\thinspace kpc}^3}\). The disk is placed at the centre of the domain, and the disk plane is parallel to \(x-y\) plane of the coordinate system. We impose outflow boundary conditions for the gas component at all domain boundaries. Fixed boundary conditions (\(e_{\rm CR}=0\)) on external domain boundaries are assumed for the CR component. # Simulations Initially the gaseous disk collects gas at the presumed global infall rate \(\dot{M}_{\textrm{in}}\) until it becomes locally gravitationally unstable. Supernovae start to explode and deposit CRs in the ISM after the gas density exceeds the critical value. After about \(t\simeq 300{\rm\thinspace Myr}\) the disk reaches an equilibrium state with a star formation rate at a level of \(\mathrm{SFR} \simeq 40\hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace Myr}^{-1}\). A typical snapshot of the system after \(600{\rm\thinspace Myr}\) of evolution is shown in Fig. [\[fig:den_cr\]](#fig:den_cr){reference-type="ref" reference="fig:den_cr"}. Most of the supernovae activity is confined to isolated regions in kpc-sized dense gas clouds (upper right panel). These regions can be also identified as spots of high CR energy density apparent as dark brown and black patches in the face-on map (lower right panel of Fig. [\[fig:den_cr\]](#fig:den_cr){reference-type="ref" reference="fig:den_cr"}). One can identify about \(10-12\) discrete star formation regions with CR energy densities exceeding \(\simeq100 {\rm\thinspace eV} {\rm\thinspace cm}^{-3}\) dropping to \(1 {\rm\thinspace eV} {\rm\thinspace cm}^{-3}\) at larger distances away from the disk (lower panels of Fig. [\[fig:den_cr\]](#fig:den_cr){reference-type="ref" reference="fig:den_cr"}). The distribution of the CR energy density in the galactic halo is highly non-uniform. Sharp edges of CR-populated regions can be identified with similar structures in the maps of vertical mass flux and vertical magnetic field component shown in Fig. [\[fig:vz_mz_bz\]](#fig:vz_mz_bz){reference-type="ref" reference="fig:vz_mz_bz"}. The vertical streams of rarefied gas visible in gas density distribution (upper panels of Figs. [\[fig:den_cr\]](#fig:den_cr){reference-type="ref" reference="fig:den_cr"} and [\[fig:vz_mz_bz\]](#fig:vz_mz_bz){reference-type="ref" reference="fig:vz_mz_bz"}) are accelerated, by CRs, to high velocities (several \(10^3 {\rm\thinspace km\thinspace s}^{-1}\)). The streams can extend several tens of \({\rm\thinspace kpc}\) above and below the disk plane and significant fraction of the outflowing gas has velocities above escape velocity and will be able to leave the galaxy altogether. Maps of the mass flux \(f_{z} = \rho v_z\) (mid panels of Fig. [\[fig:vz_mz_bz\]](#fig:vz_mz_bz){reference-type="ref" reference="fig:vz_mz_bz"}) show the bimodal nature of the outflow perpendicular to the disk plane with peak values up to \(0.2 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1} {\rm\thinspace kpc}^{-2}\) (the color scale of the mass flux panels is saturated at only \(0.02 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1} {\rm\thinspace kpc}^{-2}\) to show the wind structure far from the disk plane). Streams of gas emanating from a single star forming region have a large cross-section, visible at the horizontal slice of \(\rho v_z\) at \(z=2{\rm\thinspace kpc}\). Individual SF regions generate outflows of \(5\hbox{\)\mathrm{\thinspace M_{\odot}}\(}/{\rm\thinspace yr}\) on average and form streams of about \(3{\rm\thinspace kpc}\) in radius on both sides of the disk. In some cases the streams from neighboring SF regions merge. The horizontal area of a single stream is a few \(\sim 10 {\rm\thinspace kpc}^2\). This area multiplied by \(0.2 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1} {\rm\thinspace kpc}^{-2}\) gives a number consistent with \(\sim 5\hbox{\)\mathrm{\thinspace M_{\odot}}\(}/{\rm\thinspace yr}\) through the stream cross-section area, even though the average flux at \(z=2{\rm\thinspace kpc}\) may be half of the peak value \(\simeq 0.1 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1} {\rm\thinspace kpc}^{-2}\) for two outflows on both sides of the disk. The mass loaded wind is enriched with fresh CRs and is highly magnetized. Field strengths exceeding \(1 \mu G\) can be reached naturally at distances of \(20-50{\rm\thinspace kpc}\) away from the central galaxy. Magnetic flux tubes coherent over such large distances would be directly detectable with Faraday rotation measurement techniques. To quantify the vertical structure of the CR-driven wind we plot in Fig. [\[fig:mflux\]](#fig:mflux){reference-type="ref" reference="fig:mflux"} the total vertical mass flux as a function of distance from the galactic disk from \(t=400 {\rm\thinspace Myr}\) to \(t=700{\rm\thinspace Myr}\). In general the flow patterns are symmetric and outflow dominates inflow by an order of magnitude. Total outflow rates decline from \(\sim \pm 60 \hbox{\)\mathrm{\thinspace M_{\odot}}\(}/{\rm\thinspace yr}\) close to the disk plane to \(< 20 \hbox{\)\mathrm{\thinspace M_{\odot}}\(}/{\rm\thinspace yr}\) at \(50 {\rm\thinspace kpc}\). A more detailed inspection of wind density-velocity distribution (Fig. [\[fig:vd_histogram\]](#fig:vd_histogram){reference-type="ref" reference="fig:vd_histogram"}) shows that the density of the CR driven wind gas is typically below \(0.1 {\rm\thinspace cm}^{-1}\) and the high-velocity gas (\(v_z \geq 500 {\rm\thinspace km\thinspace s}^{-1}\)) escapes at densities \(\leq 0.003\) H atoms \({\rm\thinspace cm}^{-3}\). In Fig. [\[fig:moots\]](#fig:moots){reference-type="ref" reference="fig:moots"} we show the evolution of the star formation rate and the integrated outflow rates at different altitudes at \(z= 2 {\rm\thinspace kpc}\), \(10{\rm\thinspace kpc}\), and \(49 {\rm\thinspace kpc}\) above the disk plane. After about 400 Myr the star formation rate settles to a value of \(\dot{M}_\mathrm{SFR} \sim 40 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1}\). The mean surface density of disk gas within the radius 10 kpc reaches an equilibrium value of \(\sim 100 \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace pc}^{-2}\) and SFR surface density is \(\sim 10^{-1} \hbox{\)\mathrm{\thinspace M_{\odot}}\(} {\rm\thinspace yr}^{-1} {\rm\thinspace kpc}^{-2}\). These values are in good agreement with nearby highly star forming galaxies and typical massive high-redshift disks. At about \(10 {\rm\thinspace kpc}\) away from the disk plane the mass outflow rate matches the star formation rate and the galaxy average mass-loading \(\eta = \dot{M}_{\mathrm{out}}/{\rm SFR}\) becomes of order unity. Closer to the disk plane the mass-loading is higher \(\eta \sim 1.5\) and further away from the disk plane it is still significant. It only decreases to values of \(\eta \sim 0.5\). # Discussion and conclusions {#discussion} We have demonstrated as a proof of principle that the injection of only \(10\%\) of SN energy in the form of CRs and neglecting the thermal and kinetic energy input is sufficient to drive a large scale galactic wind in a gas-rich and highly star forming disk with properties similar to typical star forming high-redshift galaxies. The additional pressure gradient of the relativistic fluid-which in contrast to heated dense gas cannot easily dissipate its energy away-drives the formation of a strong bi-polar galactic wind with velocities exceeding \(10^3 {\rm\thinspace km\thinspace s}^{-1}\). CRs can easily escape far from dense regions with almost negligible energy losses and deposit their energy and momentum in rarefied medium. This process is supported by the CR driven break-out of field lines whose vertically open structure is maintained by the wind. Cosmic rays can rapidly diffuse along these field lines far into the galactic halo. To check consequences of our assumption of constant diffusion coefficients we examined the ratio of CR diffusion speed to the Alfvén speed over the simulation volume. As shown in Fig.[\[fig:den_cr\]](#fig:den_cr){reference-type="ref" reference="fig:den_cr"} the CR energy distribution is very smooth everywhere except the close to the SF regions and some apparent shock-like structures, especially when compared to the gas density distribution. Vertical stratification of the CR energy distribution can be observed only far from star formation regions and the corresponding scale height is of the order of a few kpc. The smooth distribution of CRs is a consequence of a highly irregular vertical magnetic field component, guiding the field aligned CR diffusion in the vertical direction. We have found that \(v_\mathrm{diff}/v_\mathrm{A} \leq 1\) in a major part of the simulation volume. Consequently, \(v_\mathrm{diff}/v_\mathrm{A} > 1\) in regions of galactic disk with high gas density. As we mentioned already, we consider the limitation of the diffusion velocity by Alfvén waves less restrictive in the disk volume, because neutrals are supposed to reduce the level of CR self-excited Alfvénic turbulence. Moreover, we note that steep gradients of CR energy density around star formation regions provide a specific feedback mechanism. High production rates of CRs implies fast expansion of overpressured regions, and subsequently enhanced CR advective expansion leads to a reduction of the CR energy density gradient. The fast expansion of CR overpressured bubbles increases the Alfvén speed locally, leading to a reduction of the ratio \(v_\mathrm{diff}/v_\mathrm{A}\). The CR driving is so significant that the mass outflow rate can become of the same order as the star formation rate in the galactic disk, even in our simplified setup where the disk plane is more or less treated as an inner boundary condition and thermal as well as kinetic feedback from stellar evolution and supernovae have been neglected entirely. Based on our simulations we can conclude that relativistic particles accelerated in supernova remnants in combination with a strong magnetic fields (typical for high-redshift galaxies ) provide a natural and efficient mechanism to help explaining the ubiquitously observed mass-loaded galactic winds in high-redshift galaxies (e.g. ) as well as the highly magnetized medium surrounding these galaxies. The efficiency-in terms of mass loading-of this wind driving process appears to be comparable to momentum and energy driving from stellar evolution and supernovae explosions (e.g. [@2012MNRAS.421.3522H; @2012MNRAS.426..140D; @2013ApJ...770...25A]) and requires further investigation. [^1]: <http://yt-project.org/>

{'timestamp': '2013-10-16T02:07:38', 'yymm': '1310', 'arxiv_id': '1310.3273', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3273'}

# [\[sec:Introduction\]]{#sec:Introduction label="sec:Introduction"}Introduction The dynamic behavior of colloidal particles at fluid interfaces is relevant to a vast field of applications ranging from drug delivery to synthesis of nanostructured materials. Owing to recent developments in the synthesis of nanoparticles, complex geometries and heterogeneous surfaces (e.g., patchy particles) can be engineered to fully exploit physical phenomena such as surface activity and colloidal self-assembly . Theoretical models based on continuum thermodynamics are well developed for idealized particle geometries such as spheres, ellipsoids, or cylinders. Recent studies, however, point out significant limitations of such models in the presence of microscale features that are much smaller than the particle and are not considered by idealized geometric representations of the macroscale morphology. These microscale features of physical or chemical nature can cause metastability, which can lead to a very slow relaxation to equilibrium, or even the jamming of single particles at a non-equilibrium state. Analytical models considering the transitions between metastable states predict that colloidal particles at fluid interfaces can remain in a given (non-equilibrium) position and angular orientation over unexpectedly long times; this can be highly desirable or undesirable depending on the specific application. A better understanding of dynamic effects produced by microscale features and surface heterogeneities could enable novel technical applications in which it is desirable to prescribe the translational and rotational motion of the particle. In the present study, we employ Molecular Dynamics (MD) techniques to investigate the effect of localized features, smaller than the particle size, on the rotational dynamics of nanoparticles that straddle a liquid-liquid interface. The article is organized as follows. In Sec. [\[sec:Method\]](#sec:Method){reference-type="ref" reference="sec:Method"}, we describe the particle geometries studied and the methodology employed to compute the system free energy as a function of the angular orientation of the particle. In Sec. [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, we report main findings and discuss the effects of localized geometric features on the rotational free energy. The computed free energies reveal that the atomistic nature of the studied nanoparticles produces nontrivial surface morphologies described by an effective diameter that varies as the particle rotates. MD simulations show the presence of long-lived metastable states where the particle remains "locked" at angular orientations where certain localized features lie at the interface. The observed lifetime of these metastable states compares well with predictions from Kramers' escape theory. In Sec. [\[sec:CR\]](#sec:CR){reference-type="ref" reference="sec:CR"}, we discuss the relevance of our findings for diverse technical applications (e.g., nanoparticle synthesis and self-assembly, reactive emulsions) and possible strategies to exploit the studied effects.

{'timestamp': '2013-10-15T02:00:34', 'yymm': '1310', 'arxiv_id': '1310.3287', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3287'}

# Introduction Although the blood flow in the left heart is inherently related to the heart function, this link is yet to be understood. The general organization of the flow can be found in textbooks, but the details of this three-dimensional unsteady flow in a moving/deforming domain are still subjected to intense research. The improvement of medical imaging and numerical simulation techniques has given the physicians and the researchers new insights in the left heart flow over the last years. As a complement to medical imaging techniques like phase-contrast magnetic resonance imaging (PC-MRI) or echocardiography, we developed a numerical technique to perform numerical simulations of the left heart flow using morphological images. The numerical procedure is presented in. The video briefly explains the different steps of the method, which are summarized here: 1. A medical exam consisting of 4-D morphological images of a patient-specific left heart is indispensable for our method. CT-scan and MRI exams have already been successfully used. 2. The images are segmented and a model of the left heart is constructed at one instant of the heart cycle. In the case considered, it contains the left atrium and the left ventricle, the end of the four pulmonary veins and the aortic root. 3. This computational domain is discretized into an unstructured grid. 4. By image registration, the unstructured grid is deformed so that the computational domain follows the medical images. Interpolation is used to reconstruct the geometry between instants where medical images are available. From the knowledge of the wall displacements, the unstructured grid is deformed along the cardiac cycle using a mesh deformation algorithm. A time-evolving computational patient-specific computational domain is thus available. 5. Because of limitations in space-time resolution of medical data, valves cannot be accounted for in the same way and they are thus modeled. From medical images, we extract the location of the valvular annuli and the main characteristics of the mitral valve (length of the leaflets; position of the center, small and large radii of the best ellipse modelling the cross section area; periods of time when the valve is closed/open). A simple geometrical model fed by these data is then used to assess the position of the leaflets over time and their effect on the blood flow is then accounted for thanks to an immersed boundary method. 6. Mass flow rates are imposed at the four pulmonary veins (top of figure 1). We assume that either the mitral valve or the aortic valve is closed along the cardiac cycle, so that blood always enters in a closed domain. Mass conservation arguments can then be used to obtain the incoming mass flow rate, evenly distributed between the pulmonary veins. 7. Blood is modeled as a Newtonian fluid and Navier-Stokes equations are solved in a moving domain with prescribed boundary motion, using an Arbitrary Lagrangian-Eulerian framework. In the present [video](http://www.math.univ-montp2.fr/~yales2bio/IMAGES/APS_DFD2013/apsdfd102246_lowformat.mp4) (see also the [video in HD](http://www.math.univ-montp2.fr/~yales2bio/IMAGES/APS_DFD2013/apsdfd102246.mp4)), a CT-scan exam is used. The computational grid consists of 3 million tetrahedral cells, with a typical grid size of \(0.8\) mm. The maximum Reynolds number at the mitral valve during diastole and at the aortic valve during systole is of order of 5000. The heart rate is 60 beats per minute. Twenty-five heart cycles are computed and phase averages are gathered over 15 cycles. The YALES2BIO solver ([http://www.math.univ-montp2.fr/\(\sim\)yales2bio/](http://www.math.univ-montp2.fr/~yales2bio/)) is used to perform the present large-eddy simulations. The non-dissipative 4th-order finite-volume numerical method, inherited from the YALES2 solver (<http://www.coria-cfd.fr/index.php/YALES2>), enables to detect turbulent spots along the cardiac cycle. The simulation recovers the main characteristics of the left heart flow, as reported for instance in medical exams or *in vitro* experiments. However, the simulation shows that, in addition of the usual picture of the left heart flow, turbulence may be present at specific regions and specific periods along the heart cycle. Two regions in the left heart mainly show turbulence activity: the upper part of the atrium and the lower part of the ventricle, near the apex. In both places, the flow destabilizes when the jets filling the cavities decelerate. In the ventricle, this occurs in the second half of diastole, after the impact of the vortex ring generated during the E-wave. Note that the A-wave is rather weak in this heart, which most probably impacts the development of turbulence in the ventricle. An effective way of visualizing the flow features in the left heart is the volume rendering of vorticity magnitude, as shown in figure 1 and in the fluid dynamics videos.

{'timestamp': '2013-10-14T02:09:48', 'yymm': '1310', 'arxiv_id': '1310.3199', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3199'}

# Introduction One of the outstanding issues in astrophysics is characterising the birth environment of the Solar System. In particular, understanding whether the Sun is an 'average' star in terms of its formation and evolution is important for assessing how likely the formation of a quiescent, habitable Solar System is when placed in the context of other planetary systems. A strong constraint on the formation of our Solar System appears to be the presence of short-lived radioactive isotopes in meteorites originating from the epoch of planet formation. The short half-life and abundance of such isotopes (inferred from their stable daughter products) argues for their rapid inclusion in meteorites during the early phases of the Solar System. Short-lived radiogenic isotopes may also be the dominant heat source for forming planetesimals in protoplanetary discs. This could affect the survival of volatile elements in the inner region of the Solar System and have implications for planet habitability. Several short-lived isotopes with half-lives ranging from tens of days to several Myr are present in meteorites, but two--\(^{26}\)Al and \(^{60}\)Fe--are very difficult to produce without nucleosynthesis in massive stars. It is possible to produce \(^{26}\)Al through spallation or from evolved asymtotic giant branch (AGB) stars, but the presence of \(^{60}\)Fe points toward enrichment from a supernova explosion [^1]. Following the discovery of \(^{26}\)Al in meteorites, suggested that the Sun could have formed when a supernova explosion triggered the collapse of a star-forming giant molecular cloud. This scenario requires that the supernova explosion does not destroy the GMC altogether. Other authors have suggested that whereas \(^{60}\)Fe may be delivered from a supernova, \(^{26}\)Al can also be produced in the winds of evolved massive stars, and that the isotope enrichment occurs in a sequential star formation process. Firstly, \(^{60}\)Fe is delivered to the nearby interstellar medium (ISM) by multiple supernovae from the first generation of star formation. These supernovae then trigger a second generation of star formation in which \(^{26}\)Al is delivered into the ISM by the wind of a single massive star. The Sun is then born in a third generation of star formation within the shell of contaminated ISM material. Finally, \(^{26}\)Al and \(^{60}\)Fe can be delivered directly to the disc from which the Solar System formed. In this scenario, the massive star is assumed to form coevally with the Sun, but it evolves faster and the resultant core collapse supernova occurs before the protoplanetary disc has begun to coalesce and form large planetesimals. To obtain the correct enrichment levels, and suggest that the Sun's protoplanetary disc must have been between 0.1 and 0.3 pc from the supernova (at distances less than 0.1 pc, the supernova is likely to strip away too much of the disc, and beyond 0.3 pc the yield of radioactive isotopes is too low, [@Adams11]). If the meteorite enrichment occurs during a single supernova explosion, then a 25 M\(_\odot\) star is most likely to deliver the relative isotopic abundances. At first sight, a 25 M\(_\odot\) star in close proximity to the Sun may seem unlikely; most stars form in clusters or associations and there is a relation between the most massive star that can form in a cluster and the mass of the cluster (cf. number of stars) itself. show that when randomly sampling an initial mass function (IMF) a 25 M\(_\odot\) star is likely to form in a cluster with at least \(N = 2000 \pm 100\) other stars (the exact number of stars depends on the adopted IMF). This moderately high expectation value for the number of stars that form in the company of a 25 M\(_\odot\) star, coupled with the fact that embedded clusters typically have radii less than several pc, suggests that the birth environment of the Solar System could be rather dense and therefore hostile. UV radiation from massive stars, which would evaporate or truncate the protoplanetary disc, and dynamical interactions during close encounters with intermediate and low-mass stars could inhibit planet formation in such an environment. Several authors have estimated the maximum number of stars in the Sun's natal cluster that would allow the formation of a 25 M\(_\odot\) star, but also not be too hostile for the formation and evolution of the Solar System. For example, calculate collisional cross sections for the Solar System to undergo disruptive interactions with passing stars and calculates the likely encounter rates for Sun-like stars in two different types of star forming region; an extremely dense cluster versus a more diffuse OB association. In general these authors find that a cluster with \(N = 10^3-10^4\) stars would enable the formation of a 25 M\(_\odot\) star without dynamical interactions prohibiting the formation of the Solar System--provided that the cluster quickly disperses. However, it remains unclear whether the evolution of a 'typical' cluster which forms a 25 M\(_\odot\) star does result in supernova enrichment of G-dwarf stars like the Sun, without those G-dwarfs suffering dynamical interactions which would hinder or disrupt planet formation. Previous work on this topic has assumed that if the cluster contains a 25 M\(_\odot\) star then enrichment is virtually guaranteed, and instead focuses on whether the encounter history of Sun-like stars in the cluster may be the prohibitive factor in deciding whether the Solar System could form and/or survive. Here, we simultaneously combine the two approaches and determine how many G-dwarfs are within 0.1--0.3 pc of the supernova but do not suffer disruptive dynamical interactions during the course of the cluster's evolution. In this paper, we use \(N\)-body simulations to model the evolution of star clusters that contain exactly one 25 M\(_\odot\) star in a cluster with \(\sim\)`<!-- -->`{=html}2000 other stars and determine the number of G-dwarfs that experience the necessary supernova enrichment, where the 25 M\(_\odot\) star is expected to go supernova at \(\sim 6.6\) Myr. Of those G-dwarfs, we also determine their interaction history within the cluster. We model four different initial cluster set-ups to cover a large range of potential star formation scenarios, but keep the stellar population constant so that stochastic differences in the clusters' evolution can be identified. The paper is organised as follows: in Section [2](#typical){reference-type="ref" reference="typical"} we describe our method of sampling the IMF to obtain a 'typical' cluster, in Section [3](#method){reference-type="ref" reference="method"} we describe the \(N\)-body simulations, in Section [4](#results){reference-type="ref" reference="results"} we describe the results for four different types of cluster initial conditions. We focus on the dynamical histories of enriched G-dwarfs from a representative simulation in Section [5](#history){reference-type="ref" reference="history"}, we provide a discussion in Section [6](#discuss){reference-type="ref" reference="discuss"} and we conclude in Section [7](#conclude){reference-type="ref" reference="conclude"}. # A 'typical' cluster {#typical} There are two distinct methods for populating a (model) star cluster with stars from an IMF; random sampling versus sorted sampling. In the first scenario, the mass of the cloud from which stars form is the only upper limit for stellar masses--for example, in very rare scenarios a 100 M\(_\odot\) cluster could (mathematically) produce a 100 M\(_\odot\) star . In the second senario, there is a direct physical dependence between the cluster mass and the most massive star that can form. claim that the latter scenario is supported by the observation that many clusters follow a relation that is consistent with sorted sampling. Such a relation would not be fundamental if massive stars could be definitively shown to form in (relative) isolation, and recent work by, and have shown many tens of O-type stars to be apparently isolated. However, the issue of random versus sorted sampling is still the subject of much debate, with claiming that it is statistically impossible to prove one scenario over the other. The discussion is relevant here because sorted sampling implies a minimum number of stars (\(\sim\)`<!-- -->`{=html}2000) is required for a cluster to host a 25 M\(_\odot\) star. Any G-dwarf enriched by the supernova could then be subject to dynamical interactions in this populous cluster. If a cluster was populated randomly from the IMF, then a 25 M\(_\odot\) star could form with very few companions--in this scenario the low probability of this cluster forming could then outweigh the probability of a G-dwarf not suffering perturbing interactions in a more populous cluster. In this paper we will not consider the dynamical histories of clusters with 'unusual' IMFs from random sampling. Instead, we use the results of Monte Carlo experiments from who randomly sampled a IMF to examine the distribution in cluster mass of clusters that contain only one massive star (\(>\)`<!-- -->`{=html}17.5 M\(_\odot\)). The median cluster mass from \(10^4\) realisations of a cluster where the most massive star is 25 M\(_\odot\) corresponds to \(\sim\)`<!-- -->`{=html}2100 stars, which is also consistent with values expected from the sorted sampling method advocated by --compare Figs. 4 and 5 in. # Cluster models {#method} We conduct pure \(N\)-body simulations of four different dynamical scenarios for cluster evolution, characterised by the initial virial ratio \(\alpha_{\rm vir} = T/|\Omega|\), where \(T\) and \(|\Omega|\) are the total kinetic energy and total potential energy of the stars, respectively. We will first discuss clusters in virial equlibrium (\(\alpha_{\rm vir} = 0.5\)) with a smooth radial profile. We will then discuss substructured clusters with three different initial virial states; subvirial ('cool'--\(\alpha_{\rm vir} = 0.3\)), virial ('tepid'--\(\alpha_{\rm vir} = 0.5\)) and supervirial ('warm'--\(\alpha_{\rm vir} = 0.7\)). To obtain an idea of the stochasticity in the simulations, we model 20 different realisations of the same cluster. We retain the same stellar population, but set the positions and velocities of the stellar systems with a different random number each time. ## Dynamical evolution We evolve the clusters for 10 Myr using the `kira` integrator in the Starlab package. This follows the clusters until they have started to dissolve and hence contribute to the Galactic field population. found that at 10 Myr a significant fraction (20--40 per cent) of stars are unbound in similar clusters to those modelled here, and noted that the more substructured a cluster is, the more likely it is to evaporate on timescales less than 10 Myr. We do not impose an external Galactic tidal field on the clusters, as this will have only a minimal effect in the first 10 Myr. We implement stellar and binary evolution by using the `SeBa` code, also within Starlab, which updates the evolutionary state of stars more frequently than the timestep of the \(N\)-body integrator. The combination of `kira` and `SeBa` enables us to model the clusters as fully collisional systems with accurate stellar evolution (including stellar mergers and binary evolution). ## Smooth clusters in virial equilibrium We model smooth clusters in virial equilibrium using a Plummer sphere, according to the prescription in @Aarseth74. We force the most massive star in the cluster to be at the cluster centre, as mass segregation is observed in several large clusters, and smooth, virialised clusters cannot mass segregate dynamically on short timescales. The Plummer spheres have a half-mass radius of 0.4 pc. ## Substructured clusters We set up substructured clusters using the fractal prescription in. This has the advantage that the substructure is described by just one parameter, the fractal dimension \(D\). In three dimensions, a highly substructured cluster has a fractal dimension \(D = 1.6\), and a uniformly spherical cluster has \(D = 3.0\). We set up clusters with a moderate level of substructure, with \(D = 2.0\). The velocities of the stellar systems (be they single or binary) are correlated according to the substructure; stars that are close have similar velocities, whereas distant stars can have very different velocities. We refer the reader to for a fuller description of this cluster set-up method. The fractals have a radius of 1 pc. We then vary the initial virial ratio of the stars and scale the velocities of the individual stars to the desired virial ratio. In one suite of simulations the clusters are subvirial (\(\alpha_{\rm vir} = 0.3\)), which results in cool collapse during the first 1 Myr. In this set-up, the most massive stars are placed at random in the fractal--they may subsequently mass-segregate so that the massive stars sink to the cente of the cluster. Another suite of simulations are initially in virial equilibirum (\(\alpha_{\rm vir} = 0.5\))--the initial substructure is subsequently erased through dynamical interactions, but the cluster is not expected to form a central core which is as dense as in the cool-collapse clusters. Finally, we run a suite of simulations where the stars are initially supervirial (\(\alpha_{\rm vir} = 0.7\))--to determine whether supernova enrichment of G-dwarfs can occur if the birth cluster is globally unbound. ## Stellar systems We place the same population of stellar systems in each cluster to establish that any differences in the evolution of the clusters is due to the random differences in system velocity or position, rather than total mass or different binary properties. The majority of G-type stars in the Galactic field have a binary companion. We include binaries in our simulations for the reason that a Sun-like star that experiences the necessary amount of supernova enrichment may not be a suitable Solar System analogue if it is in a close (\(< 100\) au) binary system. We set the most massive star in the cluster to be 25 M\(_\odot\) and then draw the remaining *primary* masses randomly from a IMF of the form \[\frac{dN}{dM} \propto \left\{ \begin{array}{ll} M^{-1.3} \hspace{0.4cm} m_0 < M/{\rm M_\odot} \leq m_1 \,, \\ M^{-2.3} \hspace{0.4cm} m_1 < M/{\rm M_\odot} \leq m_2 \,, \end{array} \right.\] where \(m_0\) = 0.1 M\(_\odot\), \(m_1\) = 0.5 M\(_\odot\), and \(m_2\) = 20 M\(_\odot\), so that we do not have any other \(>\)`<!-- -->`{=html}20 M\(_\odot\) (O-type) stars in the clusters. ### Binary systems We set stellar systems up with the binary fraction and orbital parameters observed in the Galactic field. Note that the field is probably a dynamically evolved population; the primordial binary fraction was likely higher, and the period and eccentricity distributions will also have evolved. In principle, it is possible to 'reverse engineer' the initial binary population by comparing the observed binary properties in young clusters with simulated clusters. However, for the purposes of this paper, we simply wish to impose a lower limit on the number of G-dwarfs that reside in binary systems initially and the field population is a suitable lower-limit; we will discuss this assumption in Section [6](#discuss){reference-type="ref" reference="discuss"}. The field binary fraction decreases as a function of primary mass. Primary masses in the range 0.1 \(\leq M/{\rm M}_\odot~<\) 0.47 are M-dwarfs, with a binary fraction of 0.42. K-dwarfs have masses in the range 0.47 \(\leq~M/{\rm M}_\odot\) \(<\) 0.84 with a binary fraction of 0.45, and G-dwarfs have masses from 0.84 \(\leq~M/{\rm M}_\odot~<\) 1.2 with a binary fraction of 0.57. All stars more massive than 1.2 M\(_\odot\) are grouped together and assigned a binary fraction of unity, as massive stars have a much larger binary fraction than low-mass stars . If a random number exceeds the binary fraction of the primary mass, a secondary mass is drawn from a flat mass ratio distribution. The periods of binary systems in the field are observed to have a log-normal distribution of the form \[f\left({\rm log_{10}}P\right) \propto {\rm exp}\left \{ \frac{-{({\rm log_{10}}P-\overline{{\rm log_{10}}P})}^2}{2\sigma^2_{{\rm log_{10}}P}}\right \},\] where \(\overline{{\rm log_{10}}P} = 4.8\), \(\sigma_{{\rm log_{10}}P} = 2.3\) and \(P\) is in days. The eccentricities of binary stars are drawn from a thermal distribution of the form \[f_e(e) = 2e.\] In the sample of, close binaries (with periods less than 10 days) are almost exclusively on tidally circularised orbits. We account for this by reselecting the eccentricity of a system if it exceeds the following period-dependent value: \[e_{\rm tid} = \frac{1}{2}\left[0.95 + {\rm tanh}\left(0.6\,{\rm log_{10}}P-1.7\right)\right].\] We combine the primary and secondary masses of the binaries with their semi-major axes and eccentricities to determine the relative velocity and radial components of the stars in each system.\ We continue this procedure until the cluster has 2100 stars, which corresponds the median cluster mass from \(10^4\) realisations of a cluster where the most massive star is 25 M\(_\odot\). The next nine most massive stars in this cluster range from 7--13 M\(_\odot\), and the cluster contains a total of 96 G-dwarfs. The single stars and binaries are then placed randomly at a system position in the fractal or Plummer sphere. # Cluster evolution {#results} In this Section, we consider four different sets of initial conditions for star cluster formation and follow the subsequent dynamical evolution for 10 Myr. For each cluster, we determine the number of G-dwarfs, \(N_{\rm enrich}\), that are within 0.1--0.3 pc of the the supernova and therefore experience the required levels of isotope enrichment observed in Solar System meteorites . Of these \(N_{\rm enrich}\) G-dwarfs, we then determine how many are either in close (\(<\)`<!-- -->`{=html}100 au) binary systems, or suffer a close (\(<\)`<!-- -->`{=html}100 au) encounter that could affect the outer regions of the Solar System. We label the number of these enriched 'singletons' \(N_{\rm enrich, sing}\). Finally, we might expect that a large velocity perturbation could disrupt planet formation (or a young system of planets). We therefore count the number of enriched singletons that do not suffer a velocity kick greater than 1km s\(^{-1}\) (the typical velocity dispersion in a bound embedded cluster) as being dynamically unperturbed, \(N_{\rm enrich, sing, unp}\). In Section [5](#history){reference-type="ref" reference="history"} we will show the dynamical histories of several enriched G-dwarfs in a representative simulation, but in this Section we focus on whether the numbers of enriched G-dwarfs depend on the different adopted initial conditions for star formation. ## Smooth, virial clusters We show the typical morphology of a smooth Plummer-sphere cluster at 0 Myr (Fig. [\[plummer_ic\]](#plummer_ic){reference-type="ref" reference="plummer_ic"}) and at the supernova time (6.63 Myr--Fig. [\[plummer_snapshot\]](#plummer_snapshot){reference-type="ref" reference="plummer_snapshot"}). The black triangle indicates the position of the supernova progenitor at each time. The cumulative distributions of distances from the supernova for all 96 G-dwarfs for 10 clusters is shown in Fig. [\[plummer_SN_dist\]](#plummer_SN_dist){reference-type="ref" reference="plummer_SN_dist"}. The 0.1--0.3 pc 'Goldilocks zone' for enrichment of our own Solar System is between the two vertical dashed lines. As one might expect, the smooth Plummer-sphere clusters in virial equilibrium follow very similar evolutionary patterns. They all retain a smooth, centrally concentrated morphology. However, the number of enriched G-dwarfs \(N_{\rm enrich}\) does vary between clusters, as does the number of enriched, unperturbed singletons \(N_{\rm enrich, sing, unp}\). In Fig. [\[plummer_enrich\]](#plummer_enrich){reference-type="ref" reference="plummer_enrich"} we show the distribution of \(N_{\rm enrich}\) in all twenty simulations by the open histogram, the distribution of \(N_{\rm enrich, sing}\) by the grey histogram and the distribution of \(N_{\rm enrich, sing, unp}\) by the black histogram. Note that the simulations are sorted by \(N_{\rm enrich, sing, unp}\), then \(N_{\rm enrich, sing}\) and then \(N_{\rm enrich}\). Firstly, 5/20 clusters do not contain any enriched G-dwarfs. This is mainly due to the supernova progenitor interacting with the other massive stars in the cluster and being ejected; however, in one cluster (number (xx) in Table [1](#Plummer_data){reference-type="ref" reference="Plummer_data"}) the 25 M\(_\odot\) star merges with a 5 M\(_\odot\) star and the product does not explode as a supernova before the end of the simulation. In the 15 simulations where we do have supernova enrichment, the number of enriched G-dwarfs, \(N_{\rm enrich}\), varies between 2 and 13 (from a total of 96 G-dwarfs). 14/20 clusters contain between 1 and 6 enriched stars that are singletons and 11/20 clusters contain 1, 2, or 3 enriched singletons that are unperturbed. We summarise the results in Table [1](#Plummer_data){reference-type="ref" reference="Plummer_data"}. ::: \(^a\)No G-dwarfs between 0.1--0.3 pc from the supernova due to fractal evolving into a diffuse association. [\[warm_fractal_data\]]{#warm_fractal_data label="warm_fractal_data"} # Dynamical histories {#history} In the previous section we showed the number of G-dwarf stars that were enriched at a distance between 0.1--0.3 pc from the supernova (\(N_{\rm enrich}\)) as a function of initial cluster conditions. Of those \(N_{\rm enrich}\), we determined the number \(N_{\rm enrich, sing}\) that were singletons, meaning that they were never in a close binary system, nor did they suffer a sub-100 au encounter. Either scenario would most likely preclude the formation and stable evolution of our Solar System. Finally, we applied a stricter criterion that an enriched singleton suffering a strong velocity perturbation (i.e.  in excess of the typical velocity dispersion in a cluster) could also be prohibited from forming a stable Solar System. The number of systems that are enriched singletons that do not suffer such perturbations is \(N_{\rm enrich, sing, unp}\). Here, we focus on the dynamical histories of 3 enriched G-dwarfs in a representative simulation where the cluster is initially substructured and subvirial \[number (vi) in Table [2](#cool_fractal_data){reference-type="ref" reference="cool_fractal_data"}\]. Because substructure is erased on short timescales (\(<\)`<!-- -->`{=html}2 Myr) in subvirial and virial clusters, the subsequent dynamical evolution of the cluster is similar, irrespective of the assumed initial morphology. The only exception is a supervirial (unbound) cluster, which expands and so does not have a well-defined 'centre'. However, when supernova enrichment does occur in supervirial clusters, the parameters we concentrate on here (nearest neighbour distance to the enriched G-dwarf, velocity perturbations) are similar for all initial morphologies and virial states. In the chosen simulation, \(N_{\rm enrich} = 3\) G-dwarfs were enriched by the supernova. In Fig. [\[fractal_NN_dist\]](#fractal_NN_dist){reference-type="ref" reference="fractal_NN_dist"} we show the distance to the nearest neighbour of each enriched G-dwarf as a function of time. One of the G-dwarfs is in a close (\(\sim\)`<!-- -->`{=html}10 au) binary (the black line), and therefore is not a 'singleton' and cannot be an analogue of our own Solar System. The remaining two enriched G-dwarfs have occasional encounters that are of order 500 au. Encounters of this magnitude have been suggested as potential mechanisms to create the high eccentricities of some Edgeworth--Kuiper Belt Objects, such as Sedna, and so the fact that the G-dwarfs in our simulations undergo sub--1000 au encounters should not hinder planet formation and subsequent orbitial stability. In Fig. [\[fractal_CC_dist\]](#fractal_CC_dist){reference-type="ref" reference="fractal_CC_dist"} we show the distance from the cluster centre for each of the \(N_{\rm enrich, sing} = 2\) singletons in the simulation. Due to dynamical mass segregation, the supernova progenitor has sunk to the cluster centre before the explosion, and both G-dwarfs are required to be on cluster-centric orbits to enable enrichment. This has two implications for our two enriched singletons. Firstly, they must pass through the cluster centre at least once--during this time their discs could be subject to photoevaporation from other massive stars, which are also likely to reside in the cluster centre. Inspection of Fig. [\[fractal_CC_dist\]](#fractal_CC_dist){reference-type="ref" reference="fractal_CC_dist"} shows that one of our singletons, shown by the red line, passes though the inner 0.5 pc of the cluster centre much more often than the other (shown by the green line). Secondly, at the time of the supernova (shown by the dotted vertical line in all panels), the singletons just happen to be passing through the cluster centre at that instant--they are "in the right place at the right time". Finally, we show the change in velocity magnitude for the enriched singletons in Fig. [\[fractal_Delta_V\]](#fractal_Delta_V){reference-type="ref" reference="fractal_Delta_V"}. One of these singletons experiences velocity kicks in excess of 1 km s\(^{-1}\) (the red line), which we suggest could disrupt planet formation and/or evolution. We note that this enriched singleton also passes through the cluster centre more often, and therefore it is not surprising that it has a more hostile dynamical history than the other enriched singleton in this cluster. # Discussion {#discuss} In Sections [4](#results){reference-type="ref" reference="results"} and [5](#history){reference-type="ref" reference="history"} we have presented the results of \(N\)-body simulations of star cluster evolution in which we have investigated the numbers of Solar-type (G-dwarf) stars that could be enriched in short-lived isotopes by ejecta from the supernova of a 25 M\(_\odot\) star. In this scenario a supernova enriches the Sun's protoplanetary disc with the levels of \(^{26}\)Al and \(^{60}\)Fe found in meteorites from the epoch of planet formation. In order to experience the enrichment levels required without stripping too much of the disc away, the G-dwarf(s) must be within 0.1--0.3 pc of the supernova explosion. We have varied the initial conditions of the star cluster in an attempt to cover as much parameter space as possible for the initial conditions of star-forming regions that are likely to produce at least one 25 M\(_\odot\) star. In our first simulation, we adopt a smooth, virialised morphology; this model is unlikely to be representative of the initial conditions of star forming regions, which exhibit a high degree of substructure. However, we expect the evolution of Plummer spheres to be less stochastic than fractal clusters , so these simulations provide a useful benchmark comparison to the fractal simulations. In the remaining three suites of simulations we have created clusters with primordial substructure, and also varied the initial virial ratio. Observations of stars in star-forming regions have shown them to have subvirial (cool) velocities, which in tandem with primordial substructure, facilitates a violent relaxation process resulting in a dense spherical cluster. Such initial conditions have been successful in explaining the Orion Nebula Cluster, but do not lead to the formation of unbound associations. Indeed, massive unbound associations (e.g. Sco Cen, Carina) have been suggested as the more likely birthplace of the Solar System, and such regions are observed to contain short-lived radioactive isotopes from nucelosynthesis. For this reason, we also ran simulations of substructured clusters with virialised (tepid) and supervirial (warm) velocities to investigate whether an unbound association could be a likely birthplace of the Sun. Somewhat surprisingly, the assumed initial virial ratio does not greatly affect the results. In the initially substructured models, 6/20 cool clusters and 5/20 tepid clusters contain enriched, unperturbed singleton G-dwarfs. The warm, substructured clusters also host a low number of potential Solar System analogues (6/20 clusters contain enriched unperturbed singletons). However, in this case the low densities achieved by these expanding associations are the cause of the low number, rather than hostile dynamical interactions. As an example, supervirial cluster number (xviii) in Table [4](#warm_fractal_data){reference-type="ref" reference="warm_fractal_data"} contains 0 enriched, unperturbed singletons. However, if we assume the supernova exploded at 4 Myr instead of 6.64 Myr, then the cluster contains 6 unperturbed singletons (from a total of 12 enriched G-dwarfs). Adding substructure to the clusters does appear to influence the results. In the Plummer models, we find that 15/20 clusters contain enriched G-dwarfs, but applying our constraints that the star must be a singleton and not suffer a perturbing velocity kick we find that 11/20 clusters contain several enriched, unperturbed singleton G-dwarfs. This is roughly a factor of two higher than the number of substructured clusters that contain enriched, unperturbed singletons, and is likely due to the fact that Plummer spheres are relaxed potentials, whereas the fractals take more than 1 Myr to relax and are therefore less quiescent environments. One interesting aspect of our results is that the numbers of enriched, unperturbed singletons in the clusters are rather uniform (albeit subject to low-number statistics) compared to the distribution of the number of enriched G-dwarfs. Taking two substructured, tepid clusters as an example, cluster number (i) in Table [3](#tepid_fractal_data){reference-type="ref" reference="tepid_fractal_data"} contains 5 enriched G-dwarfs, 4 of which are unperturbed singletons, whereas cluster (ii) contains 15 enriched G-dwarfs, but only 3 unperturbed singletons. This highlights two processes within the numerical simulations; firstly, the evolution of the clusters is highly stochastic (even the Plummer spheres contain a wide spread in \(N_{\rm enrich}\)) and secondly, large numbers of enriched stars equates to the supernova occuring in a particularly dense region. The higher chances of one of the 96 G-dwarfs experiencing enrichment in a dense cluster must be offset by the fact that the G-dwarfs will likely suffer more dynamical interactions and close encounters. We caution that the number of enriched, unperturbed singletons from our simulations may be overestimated because we have assumed a field-like primordial binary fraction for G-dwarfs in the clusters. Current estimates suggest that the binary fraction in the field is \(\sim\)`<!-- -->`{=html}46 percent which may be significantly lower than the primordial binary fraction, due to dynamical processing. If we were to increase the binary fraction of the G-dwarfs, we would expect fewer enriched G-dwarfs to be 'singletons' throughout the evolution of the cluster. In addition to the supernova time, the other constraint for enrichment is that the supernova explodes between 0.1 and 0.3 pc from the G-dwarf (a balance between stripping away too much of the protosolar nebula, and injecting enough radioactive isotopes into the disc). Panel (c) in Figs. [\[plummer_cluster\]](#plummer_cluster){reference-type="ref" reference="plummer_cluster"} -- [\[warm_frac_cluster\]](#warm_frac_cluster){reference-type="ref" reference="warm_frac_cluster"} indicates that a relaxing of this upper bound would imply enrichment of more G-dwarfs; however, the boundary at 0.3 pc already assumes highly efficient injection and it is unlikely that this is underestimated. Our simulations have shown that--whilst it is possible to have multiple enriched, unperturbed singletons in a cluster--50 to 75 per cent of the clusters we model do not contain any. Furthermore, the late-stage injection of \(^{26}\)Al and \(^{60}\)Fe, which occurs at 6.6 Myr for a supernova with progenitor mass 25 M\(_\odot\), has been cited by as being too late in the disc evolution to be homogeneously included in the meteorites. This could in principle be alleviated if the supernova progenitor formed first, and the lower-mass stars formed several Myr later, although this would require firm evidence of age spreads in star-forming regions. Evidence for and against such age spreads is currently the subject of much debate. Recently, however, suggested that the ages of star-forming regions and open clusters may be underestimated by a factor of two; if this is the case, then protoplanetary discs are longer-lived and the arguments against late-stage injection become weaker. We are then left with the straightforward question: is late-stage enrichment too uncommon to be a feasible delivery mechanism for the \(^{26}\)Al and \(^{60}\)Fe levels in the early Solar System? Models of Solar System formation triggered by a supernova naturally account for the \(^{26}\)Al, but require that \(^{60}\)Fe abundances in the early Solar System be similar to the background ISM levels. Indeed, recently proposed a three-stage star formation scenario where the supernovae of several stars produces the \(^{60}\)Fe, then triggers the formation of a massive star which produces \(^{26}\)Al during its Wolf--Rayet phase. The Sun is then born in a third generation of star formation with the correct enrichment levels . Such a scenario would require age spreads to be observed in star-forming regions (like the disc enrichment model), but of the order \(\sim\)`<!-- -->`{=html}15 Myr, and is perhaps more unlikely than the age spread of \(\sim\)`<!-- -->`{=html}3 Myr that would alleviate drawbacks in the disc enrichment model. If we assume that all the clusters in our simulations eventually evaporate into the Galactic field, then the fraction of enriched, unperturbed singletons can be derived by dividing the sum of \(N_{\rm enrich, sing, unp}\) from all clusters by the sum of all G-dwarfs (i.e.  96 \(\times\) 20 simulations, a total of 1920 G-dwarfs). The sum of \(N_{\rm enrich, sing, unp}\) ranges from 19 (the smooth, tepid clusters) to 6 (the substructured, warm clusters). The fraction is therefore of order 1 per cent (smooth, tepid), or 0.3 per cent (substructured, warm) and of order 0.5 per cent for the substructured, cool/tepid clusters. If we relax the constraint that the enriched singletons must not suffer velocity perturbations, the enriched singleton fraction is between 2.2 (smooth, tepid) and 0.4 per cent (substructured, warm). Finally, we note that the dynamical evolution of star clusters is highly stochastic. Some clusters eject the supernova progenitor before enrichment, and even when enrichment occurs the numbers of enriched singleton G-dwarfs varies between 0 and 7, simply through subtle differences in the evolution of our (initially) statistically identical model clusters. Such stochasticity is impossible to characterise observationally without prior information on the star formation process. # Conclusions {#conclude} Evidence of short-lived isotopes in meteorites suggests that the Sun was in close proximity to a 25 M\(_\odot\) star which went supernova at the epoch of planet formation in the Solar Sytem. If these isotopes are delivered to the Sun's protoplanetary disc, the supernova must have occurred at a distance between 0.1--0.3 pc of the Sun. We have conducted \(N\)-body simulations of the dynamical evolution of star clusters with \(N = 2100\) members, which would be expected to form at least one 25 M\(_\odot\) star under the assumption of a normal IMF. We have determined the number of G-dwarfs that experience the necessary levels of enrichment, and then determined their dynamical histories to ascertain whether they could potentially be Solar System analogues. Our conclusions are the following: (i) Typically, between 50 per cent and 75 per cent of clusters contain supernova-enriched G-dwarfs at a distance of between 0.1 and 0.3 pc from the supernova. The number of enriched G-dwarfs is in the range of several to over 10 (from a total of 96). (ii) If we consider only 'singletons'--G-dwarfs that are never in close binaries, or suffer disruptive encounters, then only \(\sim\)`<!-- -->`{=html}25 per cent of clusters contain G-dwarfs that are enriched, unperturbed singletons. Usually these clusters contain only one or two such objects. (iii) The assumed initial conditions for star formation have little impact on the results; there is little difference in the numbers of enriched, unperturbed singletons between substructured clusters which are subvirial (cool collapse), virial (tepid and static) or supervirial (warm and expanding). The only caveat is that if the supernova were to explode earlier, the expanding supervirial clusters would have a higher occurrence of enriched, uperturbed singletons. (iv) Summing together all the G-dwarfs from each suite of simulations, the global fraction of G-dwarfs that are enriched, unperturbed singletons is of order 0.5--1 per cent. (v) The cluster evolution and numbers of enriched stars is highly stochastic; statistically identical clusters can enrich over 10 G-dwarfs, or only several, or none at all--differences are due to the inherently chaotic nature of star cluster evolution.\ At first sight the \(N\)-body models suggest that supernova enrichment of unperturbed singleton G-dwarfs like our Sun is a rare occurrence. However, it does occur in a significant fraction of clusters, and sometimes to more than one G-dwarf in the same cluster. Future investigation of the assumed cluster parameters (IMF, morphology, density, virial state), and the time of the supernova explosion, would be beneficial to investigate whether the probability of disc enrichment can be raised by changing one or more of these parameters.

{'timestamp': '2013-10-15T02:00:12', 'yymm': '1310', 'arxiv_id': '1310.3270', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3270'}

# Introduction {#introduction .unnumbered} Recent studies have shown that brain networks obtained from functional Magnetic Resonance Imaging (fMRI) recordings are altered in patients with severe disorder of consciousness (DOC). DOC can result from severe brain injury and is characterized by an absence of awareness of the self and the environment, either with preserved or disrupted sleep-awake cycle. DOC encompasses a wide spectrum of clinical conditions with different levels in the content of conscious awareness, ranging from the coma state (CS, patients who have a disrupted sleep-awake cycle and don't wake up), vegetative state (VS, who preserve sleep-awake cycle but are unaware of themselves and the environment), minimally consciousness state (MCS, patients who are unable to reliably communicate but show reproducible albeit fluctuating behavioral evidence of awareness), to lock-in syndrome (LI, patients who are fully conscious but are completely paralyzed except for small movements of the eyes or eyelids). For the prognosis of these patients, the clinical practice scores this graduation in DOC response by the Glasgow Coma Scale (GCS), or as we will use in this paper, by an alternative scale such as the JFK Coma Recovery Scale-Revised (CSR-R) . This scale encodes the neurological and behavioural state of the DOC patient providing a number ranging from 0 to 23, 0 for the deepest coma state, 23 for the fully recovered one. Despite the existence of such scales, there is a need for more reliable methods that based on brain neuroimaging can provide better characterization of the large-scale disturbances of brain function in DOC. Ultimately these approaches should help in understanding and eventually predicting coma outcome. The resting state functional Magnetic Resonance Imaging (rs-fMRI) accounts for the spontaneous brain activity occurring in the high-amplitude ultra-slow (0.1 Hz) fluctuations in the Blood-Oxygen-Level-Dependent (BOLD) signal, defining networks of correlated spontaneous activity of brain Functional Connectivity (FC). The interaction between these distributed networks as well as subcortical modules is considered critical for conscious processing, and has been shown to be disrupted in DOC state. Furthermore, the rs-fMRI paradigm is a very suitable strategy for DOC patients, since they are not able to efficiently perform specific tasks. The present study addresses the question of whether the FC obtained from the rs-fMRI is altered at different brain regions as a consequence of consciousness disturbances. To this end, we investigate the FC obtained by two different measures: the Partial Correlation (PC) and the Transfer Entropy (TE), in two different groups: healthy adults and DOC patients. Information theory offers an arsenal of different measures, complementing the linear correlation estimations of FC. These information tools are typically built as extensions of the Shannon Entropy, quantify the interactions between variables by measuring the information which is shared or transferred between them. In the last decade, the transfer entropy (TE) method is growing in popularity as it can account for directed interactions between time-series variables. When applied to neuroimaging time-series, TE is a data-driven measure that assesses the functional connectivity between brain areas even for non-linear interactions. Unlike the correlations, TE reveals directionality in the interactions, allowing for determining a *directed* FC between areas. We hypothesize that FC would be reduced in DOC patients since consciousness implies functional integration. We anticipate that PC and TE would show different behaviors in patients with increasing level of consciousness, provided that they can be related to different mechanisms of information processing in the brain. The paper is organized as follow: in Material and Methods, we give details on the the data acquisition and preprocessing and define the two measures PC and TE to compute FC patterns. The next section is dedicated to present the results of the analysis. The paper closes with a discussion on some consequences of the alteration of the FC patterns in DOC patients. # Material & Methods {#material-methods .unnumbered} ## Subjects {#subjects .unnumbered} Seventeen healthy subjects (**Group 1**) aged 25 \(\pm\) 5 year old (8 men, 9 women), with no history of neurological or psychiatric problems, participated in this study as a control group. The Edinburgh Handedness Inventory was used to assess handedness, resulting in thirteen subjects right-handed and four left-handed. Eleven DOC patients (**Group 2**) were scanned (age range, 17-44 years; 6 men, 5 women). Data from two patients were subsequently excluded because of unacceptable degrees of head and body movements. The coma severity for each patient was clinically assessed using the Revised Coma Recovery Scale (CRS-R, ): scores range from 0 (meaning deep coma state) to 23 (full recovery). The patients were scanned the first time between 2 to 6 months after major acute brain injury, and a second time between 3 to 6 months after the first scan (Table [1](#Table1){reference-type="ref" reference="Table1"}). For better comparison, group 2 was subdivided into 2 subgroups: **Group 2a** (n = 12) is composed by all scans of DOC patients who had a corresponding CRS-R scale. **Group 2b** (n = 4) includes the second scans of the four patients who recovered consciousness before the second session (marked with asterisks in Table [1](#Table1){reference-type="ref" reference="Table1"}). The study protocol was approved by the Institutional Review Board of the Institute of Neurological Research FLENI. Informed consent was directly obtained from healthy participants and from the next kin of each of the patients. ::: In summary, TE analysis exposes alterations in the FC exhibited by DOC patients. In particular, TE within hemispheres and between hemispheres is smaller, although no difference was found when looking at homologue areas. In contrast to the results obtained in the PC analysis, the differences found uphold irrespective of the Euclidean distance separating ROIs pairs, although when considering LL TE, a slight decrease in the statistical p value can be observed. ## Between-homologue inter-hemispheric PC and left intra-hemispheric TE {#between-homologue-inter-hemispheric-pc-and-left-intra-hemispheric-te .unnumbered} The results show that for all analyzed areas the best two discriminators are the between-homologue inter-hemispheric (HIH) PC (Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}a-d) and the left intra-hemispheric (LL) TE (Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}e-h). Here, colors denote group differences: black (G1), blue (G2), green (G2a) and magenta (G2b). For both PC and TE the thickness of links and arrows is proportional to the PC and TE values. For PC there is a manifest anatomical disparity in the correlations pattern: it can be observed that homologue areas that are closer to each other show stronger correlations than farther ones (i.e. thicker connections at shorter distances in comparison with thinner connections at longer distances). To disentangle the behavior of the neural correlations regarding to a spatial factor, we look at the Euclidean distances between the centroids of homologue areas. For G1 the areas close to each other presented a high correlation, and beyond a threshold distance of 20 mm, correlations decreased, although the values remained high. Interestingly, the same behavior was found in G2. However, the correlation values there were shifted down, with lower mean value for areas closer than 20 mm, and decreasing for increasing distances. Thus, for ROIs areas distance-separated smaller than 20mm, differences between G1 and G2 were smaller compared to areas separated at long distances, distance separation \(<\) 20mm pval=\(10^{-6}\), distance \(>\)`<!-- -->`{=html}40mm pval=\(10^{-14}\). When inspecting G2a and G2b subgroups, there were no observable differences for anatomically closer areas, whilst it could be detected a higher correlation of some of the anatomically further areas for G2b. Regarding to the TE, not only the mean values of TE in LL areas were different between groups (Table [3](#Table3){reference-type="ref" reference="Table3"}), but the number of significant values of TE, i.e., the number of arrows plotted in Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}e-h varies across different groups. This number was more than 9 times bigger in G1 compared with G2 (G1 \# links=47, Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}e ; G2 \# links=5, Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}f). When comparing with group G2b, this number doubled the one in group G1 (\# links=99, Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}h), possibly indicating a \"transient\" brain state in the pattern of information flows in group G2b in comparison with control. ## Correlation between fMRI measures and CRS-R scores {#correlation-between-fmri-measures-and-crs-r-scores .unnumbered} We then asked if the two fMRI measures, between-homologue inter-hemispheric PC and left intra-hemispheric TE were correlated with the neurological and behavioural scale given by the CRS-S. This is represented in Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}i-k. For homologue inter-hemispheric pairs we found that TE gave the biggest correlation with the corresponding value in the communication function scale. For left intra-hemispheric pairs, TE had 0.73 correlations with oromotor/verbal function scale, 0.73 with the communication function scale and 0.73 with the total CRS-R (marked as \"JFK\" in Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}i-k). # Discussion {#discussion .unnumbered} In this study we have investigated whether the functional connectivity is altered as a consequence of consciousness disturbances. We have applied the Partial Correlation and the Transfer Entropy approaches to analyze the FC from resting-state fMRI data. We have compared two groups, healthy subjects and Disorder of Consciousness patients. The analysis was done over the 90 anatomical brain areas, defining regions of interest from the AAL atlas. We have grouped the different pairs of ROIs in inter-hemispheric homologue regions, inter-hemispheric, left intra-hemispheric, right intra-hemispheric and total (all regions). We have found two particular markers that account for the large-scale disturbance of patients brain function: the PC calculated over homologue inter-hemispheric (HIH) regions and the TE calculated over the left intra-hemispheric (LL) ROIs. The PC in HIH regions was found to be notably larger for control compared to DOC patients. This results holds also when comparing G1 with the recovered G2b group. The same comparison but done over the total average of the 90 regions did not shown significant differences. Thus, one relevant result of our analysis is the finding that only by the calculation of the PC in the proposed grouping of brain regions, it was possible to detect a significant marker for the patients disturbance, results that is hidden when we looked at the PC of the total AAL brain regions. In the case of TE, the total score did not show any significant difference either, but the brain subdivision revealed that the intra-hemispheric influences were different in control respect DOC. This happened for both LL and RR, although the TE in LL discriminated better than in RR. This is a very novel finding whose origin is still unclear and deserves further investigation. ## Methodological issues {#methodological-issues .unnumbered} The PC is a straightforward measure able to eliminate for each specific ROIs pair, the contribution to the correlations coming from common neighbors, preserving *effective* correlations between two time series. Unlike the PC which is a symmetrical measure, the TE quantifies interaction between ROIs in a directed form, i.e., region A influences to region B but the opposite is not necessary true. In concrete, TE quantifies information bits (uncertainty reduction) flowing from one ROI to the future of the other. For the case of Gaussian data, the information bits measured by the TE coincide with the Granger causality measured from time series; however for non Gaussian data, TE and causality might result in different measures. TE emerges as a very suitable measure for the study of temporal causality in brain fMRI activity in parallel to the advantage of an accurate spatial resolution. TE assessment in a population of patients with disorder of consciousness provides the opportunity of gaining insight into brain mechanisms of information processing and the finding of possible predictors of coma outcome. Regarding to the calculation of TE, it is well-known that the computation of the entropies with small data sets introduces some a bias. Because we are performing groups comparison with the same data size in each group (i.e., the time series in each subject have the same data points), such a bias will be the same in the two groups, thus not affecting the validity of the groups comparison. Nevertheless, as far as we understand there is not any reported study analyzing either information reduction (i.e. TE) or causality in fMRI data from DOC patients. ## Inter-relation between PC and TE in DOC patients {#inter-relation-between-pc-and-te-in-doc-patients .unnumbered} To exhibit high correlations is different from having high TE between two time series. This can be clearly understood by a counter-example; two fully correlated time series have zero TE as to compute the uncertainty reduction in the future of \(i\), conditioning on the two pasts \(i\) and \(j\) is not adding any further information to the situation of solely adding the past of \(i\), i.e. the two terms in the right-hand side in Eq. ([\[TE\]](#TE){reference-type="ref" reference="TE"}) are equal. As a consequence of this, the observation of having high PC for HIH pairs in healthy subjects implies to have high isolation of the information within hemispheres; thus, the TE values in both LL and RR are significantly higher than the corresponding values in HLR and HRL. Interestingly, we found that while PC is reduced in DOC patients between inter-hemispheric homologue areas, TE shows an altered pattern at the level of general inter-hemispheric interactions. In the control group we observe that despite the coherence is high between homologue areas, their TE is low. Conversely, while PC between hemispheres is low, LR and RL TE are high. The DOC patients show the same trend, although the LR and RL TE is significantly lower than in controls. This supports the notion that consciousness arises from long-range modulation of neural activity. A disruption in long-range communication could affect mechanisms such as increase of stimulus' salience, facilitation of propagation across sparsely connected networks, and selective routing, mechanisms that are related to conscious processing. ## rs-fMRI inter-hemispheric correlations and gamma rhythms {#rs-fmri-inter-hemispheric-correlations-and-gamma-rhythms .unnumbered} Recently it has been shown that the inter-hemispheric correlations in the rs-fMRI dynamics correlate with the inter hemispheric coherence exhibited by electrophysiological recordings in human sensory cortex, mainly with the slow modulation of the gamma rhythms in Local Field Potentials. Other studies have also found such modulation in high-level cognition tasks. Thus, one could conjecture that at the functional level, a breakdown in the inter-hemispheric rs-fMRI correlations in DOC patients could be an indication of a similar deficit in the gamma power coherence. One possibility is that low-frequency oscillatory activity is related to an underlying neuronal mechanism allowing for maintenance and consolidation of neural events across wide sections of the brain, and for the handling of incoming stimuli. Although increasing evidence points toward a property of the brain relevant for conscious processing, Vidal et al. point out that gamma-amplitude correlation would also be reflecting the parallel organization of the brain, where neural networks interact for purposeful processing of information. ## Comparison with previous results {#comparison-with-previous-results .unnumbered} As fas as we know, a single study have reported that DOC patients in comparison with healthy subjects manifest a strong reduction in the inter-hemispheric correlations in the rs-fMRI time series. The authors in did not use any atlas to compute inter-hemispheric correlations; instead they investigated specific areas such as pre-and post-central gyrus and the intra-parietal sulcus. Among other reasons, the authors selected those areas for being well separated each from the other (arguing the existence of less noise in the signal). This is consistent with our finding that DOC patients kept more similar correlations to control for ROIs separation below 20 mm. In addition to this, our study adds the novelty of having analyzed the FC obtained by the TE. ## TE density to measure consciousness alteration {#te-density-to-measure-consciousness-alteration .unnumbered} We have shown in Figs. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"} e-h how the number of TE connections can account not only for the differences between control (G1) and DOC (G2) but for the transitory brain state in the group G2b: the patients that awaked and became fully conscious at the second fMRI acquisition. Thus, we have found that the number of TE connections were 47 (G1), 5 (G2) and 99 (G2b). In a similar spirit, Seth and colleagues defined the causal density for measuring consciousness in brain states as the number of Granger-causality connections flowing in and out per each specific area. Interestingly, a similar behavior has been reported during recovery from anesthesia, where an increment in functional connectivity above the normal wakeful baseline is found. ## DOC impairment at specific brain areas {#doc-impairment-at-specific-brain-areas .unnumbered} The aim of this analysis is not to work at the level of an individual DOC patient but to search for rs-fMRI markers that can account for groups differences in DOC patients. We have not studied yet any measure that can account for DOC impairment at specific brain areas. To this end, one could study in principle the FC graphs obtained by either PC or TE using complex networks analysis, or any other kind of graph exploration methods. In a much simpler spirit (just to illustrate that this approach is plausible), we have chosen to plot the PC values per area comparing group G2 versus G1. This is illustrated in the Suppl. Fig. S1. The decorrelation index per area is plotted, (corrG1-corrG2)/corrG1. Colored in blue, the five biggest decorrelation indices correspond to the following areas: Fusiform, Insula, Parietal Superior, Precentral and Temporal Superior, revealing that those areas had the major DOC impairment. Conversely the areas with less DOC impairment (colored in red) were the Cingulum Anterior, Cingulum Middle, Frontal Superior Orbital, Superior Motor Area and Temporal Inferior. ## Limitations of the study {#limitations-of-the-study .unnumbered} One of the important limitation of studying DOC patients is the great amount of involuntary movements they exhibit, leading to potential artifacts in the fMRI acquisition. Techniques to overcome this issue include affine transformations to the time series creating a head-motion parameter matrix which can be used to regress out and remove the spurious variances they introduce. Although these methods can correct signals from movements spanning the dimensions of up to 3 to 4 voxels, recent work suggest that no technique could remove completely the effects of these artifacts over the FC. Thus especial care is necessary to tackle these problems and, eventually, discard the entire scan. ## Future directions {#future-directions .unnumbered} In this study PC and TE measures were used to assess for the assessment of functional connectivity in unconscious patients. In particular we characterized their disruptions at an anatomical level, in the basis of distances between homotopic areas. Other questions that can be explored, include the integrity of FC between the areas that constitute hubs in the brain network, between areas with high *rich-clubness*, or between associative vs sensory areas.

{'timestamp': '2013-10-14T02:10:13', 'yymm': '1310', 'arxiv_id': '1310.3217', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3217'}

# Introduction Star formation is one the most important processes in galaxies, yet our understanding of it is far from satisfactory. While it is commonly recognised that the evolution of the large-scale structure of the Universe is linked to that of dark matter, which is driven by gravitation, baryonic physics is much more challenging. Having a good understanding of star formation would be a great piece to put in the puzzle of galaxy formation and evolution. The first step is to be able to measure accurately the amount of star formation itself for a large number of galaxies. This means we need to be able to build statistical samples with observables that are linked to the recent star formation activity. One of the easiest way to perform this is to consider rest-frame ultraviolet (UV) selected samples, as the emission of galaxies in this range of the spectrum is dominated by young, short-lived, massive stars. Thanks to the combination of various observatories, building UV-selected samples is now feasible over most of the evolution of the Universe, from \(z\sim10\) to \(z=0\) [e.g. @Bouwens_2012; @Ellis_2013; @Martin_2005; @Reddy_2012b]. There is however one drawback to this approach, which is that the attenuation by dust is particularly efficient in the UV. As the absorbed energy is re-emitted in the far-infrared (FIR) range of the spectrum, it is necessary to combine both of these tracers to get the complete energy budget of star formation. The current observational facilities however are such that it is much easier to build large samples from the restframe UV than from the restframe IR over a wide redshift range. It is then useful to look at the FIR properties of UV-selected galaxies as a function of redshift in order to understand the biases inherent to a UV selection, to characterise for instance the galaxy populations probed by IR and UV selections, determine the amount of total cosmic star formation rate probed by a rest-frame UV selection, or the link between the level of dust attenuation [as probed by the ratio of IR to UV luminosities, @Gordon_2000] and physical properties. This approach has been successful by combining UV selections and *Spitzer* data at \(z\la 1\) to study the link between dust attenuation and UV luminosity or stellar mass, as well as correlation with galaxy colors. By measuring the ratio between the cosmic star formation rate density estimated from IR and UV selections, showed that the fraction of the cosmic star formation rate probed by a UV selection, without correction for dust attenuation, decreases from 50 per cent to 16 per cent between \(z=0\) and \(z=1\). At \(z>1.5\), *Spitzer* data probe the mid-IR range of the spectrum, which can lead to an overestimation of the IR luminosity. At these redshifts, *Herschel* data become particularly valuable for such projects. extended this kind of study by stacking \(z=2\) Lyman Break Galaxies (LBGs) in *Herschel*/PACS images to investigate their dust attenuation properties: they estimated that typical UV-selected galaxies at these epochs have infrared luminosities similar to Luminous Infrared Galaxies (LIRGs, \(10^{11}<L_{\rm IR}/\textrm{L}_{\odot}<10^{12}\)). combined the measurements at \(0<z<4\) of the UV and IR restframe luminosity functions to infer the redshift evolution of the total (UV+IR) cosmic star formation rate and dust attenuation. In a previous study based on a stacking analysis of UV-selected galaxies at \(z\sim1.5\) in *Herschel*/SPIRE images, we showed that using a UV-selection at \(z\sim1.5\) with a proper correction for dust attenuation enables us to recover most of the total cosmic star formation activity at that epoch. It is also necessary to investigate the link between dust attenuation and a number of galaxy properties, in order to be able to accurately correct for dust attenuation, by providing empirical relations for instance. One of the most commonly used empirical relation in this context is based on the correlation between the slope of the UV continuum and the dust attenuation. Such correlation has been observed for star-forming galaxies from high to low redshifts [e.g. @Buat_2005; @Burgarella_2005; @Heinis_2013; @Reddy_2010; @Seibert_2005]. However, the common assumption that the relation derived from local starbursts [@Calzetti_2001; @Meurer_1999] is universal is questionable [@Heinis_2013; @Hao_2011] as the extinction curve is dependent on the dust geometry and dust properties. Moreover, the UV slope of the continuum encodes partly the star formation history of the galaxies, and the observed relation between the UV slope and the dust attenuation is also selection-dependent. It is then useful to turn towards other observables, which might provide better ways to correct for dust attenuation in a statistical sense. Dust attenuation is for instance not really well correlated with observed UV luminosity [e.g. @Buat_2009; @Heinis_2013; @Xu_2007]. On the other hand, the correlation with stellar mass is tighter [e.g. @Buat_2012; @Finkelstein_2012; @Garn_2010; @Pannella_2009; @Xu_2007]. This is somewhat expected as the dust production is linked to the star formation history, through heavy elements production, and stellar mass in this context can be seen as a crude summary of star formation history. Investigating the link between dust attenuation and stellar mass is interesting by itself, but getting a direct estimate of the IR luminosity implies that we can also derive the star formation rate (SFR) accurately. This means that we are able for instance to characterise the relation between the SFR and the stellar mass. By considering galaxy samples based on star-formation activity, we are actually expecting to deal with objects belonging to the so-called 'Main Sequence' of galaxies. A number of studies pointed out that there is a tight relation between the SFR and the stellar mass of galaxies, from high to low redshift [@Bouwens_2012; @Daddi_2007; @Elbaz_2007; @Noeske_2007; @Wuyts_2011b]. Galaxies on this Main Sequence are more extended than starbursts [@Elbaz_2011; @Farrah_2008; @Rujopakarn_2013], the latter representing only a small contribution, in terms of number density, to the global population of star forming galaxies. The relation between SFR and stellar mass also seems to be independent of the environment of the galaxies. While there is debate on the slope and scatter of this relation, it is definitely observed at various redshifts, with its amplitude decreasing with cosmic time [@Iglesias-Paramo_2007; @Martin_2007; @Noeske_2007; @Wuyts_2011b]. The mere existence of this relation raises a number of issues for galaxy formation and evolution, as it implies that galaxies experience a rather smooth star formation history. In this paper, we take advantage of the combination of the multiwavelength data available within the COSMOS field, with the *Herschel*/SPIRE observations obtained in the framework of the Herschel Multi-Tiered Extragalactic Survey key program[^1]. We are assuming here that the rest-frame FIR emission we measure originates from the dust responsible for the UV/optical attenuation. Indeed, the wavelength range covered by SPIRE is dominated by the emission of dust heated by stars, the contribution from dust heated by Active Galactic Nuclei being significantly lower at these wavelengths. Moreover, our UV-selection biases against galaxies dominated by old stellar populations, hence the FIR emission we measure is mostly due to the dust heated by young stellar populations. We focus on three UV-selected samples at \(z\sim4,3\), and 1.5. We revisit the relations between dust attenuation and UV luminosity as well as stellar mass, over this wide redshift range, using homogeneous selections and stellar mass determination. Our aim is to disentangle the link between dust attenuation and these two physical quantities, by directly measuring their IR luminosities thanks to *Herschel*/SPIRE data. We also put new constraints on the SFR-stellar mass relations from \(z\sim4\) to \(1.5\), and use our results to discuss the star formation histories of Main Sequence galaxies. This paper is organised as follows: in Sect. [2](#sec_data){reference-type="ref" reference="sec_data"} we present the UV-selected samples we build from the multiwavelength data available in the COSMOS field. As most of the galaxies of these samples are not detected individually with *Herschel*/SPIRE, we perform a stacking analysis, and describe the methods we use in Sect. [3](#sec_stacking){reference-type="ref" reference="sec_stacking"}. We present our results in Sect. [4](#sec_results){reference-type="ref" reference="sec_results"}: we detail the relations between dust attenuation and UV luminosity (Sect. [4.1.2](#sec_irx_luv){reference-type="ref" reference="sec_irx_luv"}) and between dust attenuation and stellar mass (Sect. [4.2](#sec_irx_mass){reference-type="ref" reference="sec_irx_mass"}). We present in Sect. [4.4](#sec_sfr_mass){reference-type="ref" reference="sec_sfr_mass"} the SFR-stellar mass relations for UV-selected samples we obtain at \(z\sim1.5, 3\) and 4. We also investigate the link between dust attenuation and UV luminosity and stellar mass jointly (Sect. [4.3](#sec_irx_luv_mass){reference-type="ref" reference="sec_irx_luv_mass"}). We discuss these results in Sect. [5](#sec_discussion){reference-type="ref" reference="sec_discussion"} and present our conclusions in Sect. [6](#sec_conclusion){reference-type="ref" reference="sec_conclusion"}. Throughout this paper, we make the following assumptions: we use a standard cosmoslogy with \(\Omega_{\rm m} = 0.3\), \(\Omega_{\Lambda} = 0.7\), and \(H_0 = 70\,\)km s\(^{-1}\) Mpc\(^{-1}\); we denote far-UV (FUV) and IR luminosities as \(\nu L_{\nu}\); use AB magnitudes, and consider a Initial Mass Function (IMF). When comparing to other studies, we consider that no conversion is needed for SFR and stellar mass estimates between and IMFs. When converting from IMF to IMF, we divide \(M_{*, \rm Salpeter}\) by 1.74, and SFR\(_{\rm Salpeter}\) by 1.58. # Data samples {#sec_data} ## Photometric redshifts and stellar masses We base this study on the photometric redshift catalogue built from the COSMOS data by. This catalogue is based on an \(i\)-band detection, down to 0.6\(\,\sigma\) above the background. These estimates benefit from new near-infrared imaging in the \(Y\), \(J\), \(H\), and \(K_s\) bands obtained with the VISTA telescope as part of the UltraVISTA project. In the redshift range \(1.5<z<3.5\), the precision on the photometric redshifts (defined as the scatter of the difference with spectroscopic redshifts, in 1+\(z\)) is around 3 per cent. This value is given by for objects with \(K_s<24\), and has been obtained by comparing to zCOSMOS faint sample (\(I_{med} =23.6\)) and faint DEIMOS spectroscopic redshifts (\(I_{mean}=23.5\)). At \(z \sim4\), the spectroscopic redshifts available (\(I_{med} =24.4\)) yield a precision of 4 per cents, and suggest that the contamination from low redshift galaxies is negligible. On the other hand, this spectroscopic sample at \(z\sim4\) is not likely to be representative of our sample at the same redshifts (see Table [1](#tab_samples){reference-type="ref" reference="tab_samples"}). The actual photometric redshift error for our samples might be larger than this, as we are dealing with fainter objects. We also quote in Table [1](#tab_samples){reference-type="ref" reference="tab_samples"} as an alternative the mean photometric redshift error, in (1+z), estimated from the PDF of the photometric redshifts derived by . showed that the error measured from the PDF is a robust estimate of the accuracy as measured with respect to spectroscopic objects. At \(z\sim3\), the mean error from the PDF is 0.1, and 0.17 at \(z\sim4\). We also consider in this paper the stellar masses estimates of. Briefly, the stellar masses are derived from SED fitting to the available photometry, assuming single stellar population templates, an exponentially declining star formation history, and the IMF. showed that the assumption of an exponentially declining star formation history does not have a strong impact on the stellar masses estimates. ## UV-selected samples {#sec_uv_sel} We consider three UV-selected samples at \(z\sim1.5\), \(z\sim 3\), and \(z\sim 4\). The sample at \(z\sim1.5\) has already been presented in. We detail here how we build the samples at \(z\sim 3\), and \(z\sim 4\). We use optical imaging of the COSMOS field from in \(r^+\) and \(i^+\), both from Subaru. We cross-match single band catalogues built from these images with the photometric redshift catalogue of. Ninety-nine per cent of the objects with \(u^*<26\) have a counterpart in the catalogue of, while 92 per cent of objects with \(r^+<26\) have a counterpart. In the \(i^+\)-band, we use directly the catalog of, as it is based on an \(i^+\)-band detection. We then build UV-selected samples, at \(z\sim 3\) and \(z\sim 4\). We detail in Table [1](#tab_samples){reference-type="ref" reference="tab_samples"} the main characteristics of the three samples we consider here. All these samples probe the FUV rest-frame range of the spectrum, with rest-frame effective wavelengths within the range \(1570-1620\) Å at the mean redshifts of the samples (see Table [1](#tab_samples){reference-type="ref" reference="tab_samples"}). We will perform stacking at 250, 350 and 500 \(\mu\)m as a function of FUV luminosity, \(L_{\rm FUV}\), and stellar mass \(M_*\). We derive \(L_{\rm FUV}\) from the observed magnitude as follows: \[\label{eq_m2luv} L_{\nu} = \frac{4\pi D_{L}^2(z)10^{-0.4(48.6+m)}}{1+z}\] where \(D_{L}(z)\) is the luminosity distance at \(z\), and \(m\) is the observed magnitude: we use \(u^*\) at \(z\sim1.5\), \(r^+\) at \(z\sim 3\), and \(i^+\) at \(z\sim 4\). We then compute the UV luminosity at 1530 Å. We estimate a reliability limit in stellar mass for each sample the following way. We compute, as a function of \(M_*\), the fraction of objects with 3.6 \(\mu\)m flux measurements fainter than the 80 per cent completeness limit. We choose the reliability limit as the minimum \(M_*\) value where this fraction is lower than 0.3. In other words, above this value of \(M_*\), the fraction of objects that have a flux at 3.6 \(\mu\)m larger than the 80% completeness limit is \(\ge0.7\). Note that we do not impose a cut on 3.6 \(\mu\)m fluxes. The stellar mass is also estimated for objects with 3.6 \(\mu\)m flux fainter than 2.5 \(\mu\)Jy, however this estimate is less robust than for brighter objects. We quote the reliability limits for each sample in Table [1](#tab_samples){reference-type="ref" reference="tab_samples"}. # Stacking measurements {#sec_stacking} We base our study on the *Herschel*/SPIRE imaging of the COSMOS field obtained within the framework of the HerMES key program. Most of the objects from our UV-selected samples are not detected individually in these images, so we rely on a stacking analysis. We use the same methods as those presented in to measure flux densities using stacking[^10]. We recall here only the main characteristics of the methods. We perform stacking using the IAS library [@Bavouzet_2008; @Bethermin_2010][^11]. We use mean stacking, without cleaning images from detected sources. We showed in that using our method or median stacking with cleaning images from detected sources, yields similar results. We correct the stacking measurements for stacking bias, using extensive simulations of the detection process of the sources. We perform these simulations by injecting resolved artificial sources in the original images, and keeping track of the recovered sources. We then use the stacking of these artificial sources to correct the actual measurements. We also correct for the clustering of the input catalogue by taking into account the angular correlation function of the input sample. We derive errors on the stacking flux densities by bootstrap resampling. We use hereafter the ratio of the stacking flux density over its error as a measurement of signal-to-noise ratio. For each stacking measurement, we obtain a flux density at 250, 350 and 500 \(\mu\)m. We derive an infrared luminosity \(L_{\rm IR}\) by adjusting these fluxes to the templates, using the SED-fitting code CIGALE[^12]. The templates have been shown to be a reasonable approximation of the SEDs of *Herschel* sources. We consider \(L_{\rm IR}\)  as the integration of the SED over the range \(8<\lambda<1000\,\mu\)m. CIGALE estimates the probability distribution function of \(L_{\rm IR}\). We consider the mean of this distribution as our \(L_{\rm IR}\)  value, and the standard deviation as the error on \(L_{\rm IR}\). We use as redshift the mean redshift of the galaxies in the bin. Hereafter, we perform stacking as a function of \(L_{\rm FUV}\)  and \(M_*\) separately in Sect. [4.1.1](#sec_lir_luv){reference-type="ref" reference="sec_lir_luv"}, [4.1.2](#sec_irx_luv){reference-type="ref" reference="sec_irx_luv"}, [4.2](#sec_irx_mass){reference-type="ref" reference="sec_irx_mass"}, and [4.4](#sec_sfr_mass){reference-type="ref" reference="sec_sfr_mass"}, and we also perform stacking as a function of both \(L_{\rm FUV}\)  and \(M_*\) in Sect. [4.3](#sec_irx_luv_mass){reference-type="ref" reference="sec_irx_luv_mass"}. We characterise each bin by the mean value of \(L_{\rm FUV}\)  and/or \(M_*\). We derive the errors on the mean \(L_{\rm FUV}\)  using mock catalogues. These mock catalogues are only used to estimate errors on mean \(L_{\rm FUV}\)  and \(M_*\). We build 100 mock catalogues, with new redshifts for each object, drawn within the probability distribution functions derived by. We can then assign new \(L_{\rm FUV}\)  using eq. [\[eq_m2luv\]](#eq_m2luv){reference-type="ref" reference="eq_m2luv"}. For a given stacking measurement including a given set of objects, we compute the mean of \(L_{\rm FUV}\)  for each mock catalogue. The error on the mean \(L_{\rm FUV}\) is then the standard deviation of the means obtained from all mock catalogues. We derive errors on the mean \(M_*\) in a similar way, using the stellar mass probability distribution functions derived by. # Results {#sec_results} We first show results of the stacking as a function of \(L_{\rm FUV}\); we look at the relation between the average \(L_{\rm IR}\)  and \(L_{\rm FUV}\)  (Sect. [4.1.1](#sec_lir_luv){reference-type="ref" reference="sec_lir_luv"}) and then at the relation between the dust attenuation, probed by the IR to UV luminosity ratio, and \(L_{\rm FUV}\)  (Sect. [4.1.2](#sec_irx_luv){reference-type="ref" reference="sec_irx_luv"}). We further turn to results we obtain by stacking as a function of stellar mass, looking at the relation between dust attenuation and stellar mass (Sect. [4.2](#sec_irx_mass){reference-type="ref" reference="sec_irx_mass"}). We also investigate the joint dependence between \(L_{\rm FUV}\), \(M_*\), and dust attenuation (Sect. [4.3](#sec_irx_luv_mass){reference-type="ref" reference="sec_irx_luv_mass"}). As we obtain estimates of \(L_{\rmn{IR}}\), we derive a total star formation rate by combining with the observed UV luminosity, and look at the relation between star formation rate and stellar mass in our samples (Sect. [4.4](#sec_sfr_mass){reference-type="ref" reference="sec_sfr_mass"}). ## Stacking as a function of \(L_{\rm FUV}\)  ### \(L_{\rm IR}\) -\(L_{\rm FUV}\)  relation from \(z\sim4\) to \(z\sim1.5\) {#sec_lir_luv} In Fig. [\[fig_stack_lfuv_vs_lir\]](#fig_stack_lfuv_vs_lir){reference-type="ref" reference="fig_stack_lfuv_vs_lir"}, we show the \(L_{\rm IR}\)  measured by stacking as a function of \(L_{\rm FUV}\)  at \(z\sim1.5, 3,\) and 4. At \(z\sim1.5\), for galaxies with \(3\times10^{9}<L_{\rm FUV}/\textrm{L}_{\odot}<8\times10^{9}\), \(L_{\rm IR}\)  is roughly constant at \(L_{\rm IR}\sim 4\times10^{10}\textrm{L}_{\odot}\). For \(L_{\rm FUV}\)  brighter than \(8\times10^{9}\,\textrm{L}_{\odot}\), \(L_{\rm IR}\)  is increasing with \(L_{\rm FUV}\), with a power law slope of \(1.1\pm0.2\). This shows that in this range of UV luminosities at \(z\sim1.5\), \(L_{\rm IR}\)  and \(L_{\rm FUV}\)  are well correlated. At \(z\sim3\) and \(z\sim4\), the situation is quite different. At these redshifts, we explore a smaller dynamic range of UV luminosities, \(10^{10}<L_{\rm FUV}/\textrm{L}_{\odot}<10^{11}\). At these epochs, we do not measure any statistically significant trend of \(L_{\rm IR}\)  with \(L_{\rm FUV}\)  in UV-selected samples. We find that \(L_{\rm IR}\)  is roughly constant at \(L_{\rm IR} \sim 4\times 10^{11}\textrm{L}_{\odot}\). ### Dust attenuation as a function of \(L_{\rm FUV}\)  from \(z\sim4\) to \(z\sim1.5\) {#sec_irx_luv} In Fig. [\[fig_stack_lfuv\]](#fig_stack_lfuv){reference-type="ref" reference="fig_stack_lfuv"}, we show the relations between the \(L_{\rm IR}/L_{\rm FUV}\) ratio, a proxy for dust attenuation, and \(L_{\rm FUV}\)  at \(z\sim3\) and \(z\sim4\). We also show for comparison the results we obtained at \(z\sim1.5\). We indicate the equivalent dust attenuation in the FUV, \(A_{\rm FUV}\), derived from the \(L_{\rm IR}\)  to \(L_{\rm FUV}\)  ratio using: \[\begin{aligned} \label{eq_afuv} A_{\rm FUV} & = &-0.0333\rm{IRX}^3 + 0.3522\rm{IRX}^2+1.1960\rm{IRX} \nonumber\\ & + & 0.4967\\ \rm{IRX} & = & \log\left(\frac{L_{\rm IR}}{L_{\rm FUV}}\right). \nonumber \end{aligned}\] In the ranges of UV luminosity we probe, the relations between dust attenuation and \(L_{\rm FUV}\)  change from \(z\sim4\) to \(z\sim1.5\). At \(z\sim1.5\), the dust attenuation is mostly independent of \(L_{\rm FUV}\). At \(z\sim3\) and \(z\sim4\), we observe that the dust attenuation on average decreases with \(L_{\rm FUV}\). This decrease is linked to the fact that \(L_{\rm FUV}\)  is not well correlated with \(L_{\rm IR}\), as suggested by Fig. [\[fig_stack_lfuv_vs_lir\]](#fig_stack_lfuv_vs_lir){reference-type="ref" reference="fig_stack_lfuv_vs_lir"}. Our results also show that at given \(L_{\rm FUV}\), dust attenuation is larger at \(z\sim3,4\) than at \(z\sim1.5\) for galaxies with \(L_{\rm FUV} <4\times 10^{10} L_{\odot}\). We show later that this effect is actually linked to the stellar mass of the galaxies (see Sect. [4.3](#sec_irx_luv_mass){reference-type="ref" reference="sec_irx_luv_mass"}). ## Dust attenuation as a function of stellar mass {#sec_irx_mass} We investigate here the relation between dust attenuation and stellar mass. We show in Fig. [\[fig_stack_M\*\]](#fig_stack_M*){reference-type="ref" reference="fig_stack_M*"} our measurements of the ratio of IR to UV luminosity as a function of stellar mass, at \(z\sim4\), \(z\sim3\), and \(z\sim1.5\). The link between dust attenuation and stellar mass is strikingly different from the link between dust attenuation and UV luminosity. At all the redshifts we consider here, there is a clear correlation, on average, between dust attenuation and stellar mass. The results in Fig. [\[fig_stack_M\*\]](#fig_stack_M*){reference-type="ref" reference="fig_stack_M*"} show that the \(L_{IR}/L_{FUV}\) ratio is much better correlated with stellar mass than with UV luminosity. Within the same samples, the \(L_{IR}/L_{FUV}\) ratio varies by a factor of two at most as a function of \(L_{\rm FUV}\), while it varies by one order of magnitude as a function of \(M_*\). Our results also suggest that there is no significant evolution with redshift of the dust attenuation at a given stellar mass, between \(z\sim4\) and \(z\sim1.5\). There is a possible trend at the high mass range (\(M_*>10^{11}\textrm{M}_{\odot}\)) that dust attenuation decreases between \(z\sim3\) and \(z\sim1.5\). The statistics is however low for these mass bins, and the fraction of UV-selected objects directly detected at SPIRE wavelengths is the highest. Assuming that the relation between the \(L_{\rm IR}\)  to \(L_{\rm FUV}\)  ratio and \(M_*\) can be parameterised as: \[\textrm{IRX} = \alpha\log\left(\frac{M_*}{10^{10.35}}\right) + \textrm{IRX}_0\] we obtain as best fits parameters at \(z\sim1.5\) \(\alpha=0.72\pm0.08\) and \(\textrm{IRX}_0 = 1.32\pm0.04\). This relation is valid at \(z\sim1.5, 3,\) and 4 for \(10^{10}<M_*/\textrm{M}_{\odot}<10^{11}\). We compare our results with previous estimates of the relation between dust attenuation and stellar mass for UV-selected samples. At \(z\sim 1.5\), our results are in reasonable agreement with those from, derived from SED fitting, based on UV-selected objects with spectroscopic redshifts and photometry from the restframe UV to the restframe FIR. Our results are also in good agreement with those from at \(1.<z<1.5\), who studied a mass-selected sample of star-forming galaxies. Our findings are also consistent with those from, who observed that the ratio of SFRs derived from the IR and the UV increases with total SFR (\(=\)SFR\(_{\rmn{IR}} +\)SFR \(_{\rmn{UV}}\)) and \(M_*\). While we observe a higher amplitude at a given mass, our measurements show a slope of the IRX\(-M_*\) relation similar to the one derived by, whose results are derived from SED fitting applied to a sample of BX galaxies at \(z\sim 2.3\), using photometric redshifts, and UV/optical restframe data. We also compare our results at \(z\sim4\) with the measurements of, who studied the link between the slope of the UV continuum, \(\beta\), and the stellar mass. We converted their measurements of \(\beta\) to \(A_{\rm FUV}\) assuming the relation, which has been claimed to be valid at \(z=4\). The measurements of probe a lower mass range than ours, making a direct comparison difficult. Our measurement in the lowest mass bin we probe at \(z\sim4\) is in formal agreement with theirs, however it has a low signal to noise ratio, and may suffer from significant incompleteness in mass as well. Nevertheless, the extrapolation of the relation observed by at higher masses does not match our measurements. We also compare our results with the relation derived by at \(z=2\), from radio stacking of a sample of BzK-selected galaxies. This relation would significantly overpredict the dust attenuation for a UV-selected sample when compared to our results. These different relations between dust attenuation and stellar mass for UV and BzK-selected samples coud be due to the fact that the BzK selection is less sensitive to dust attenuation, and probes galaxies that are dusty enough to be missed by UV selections. We note the more recent results from are in better agreement with our measurements. ## Dust attenuation as a function of stellar mass and UV luminosity {#sec_irx_luv_mass} The results presented in Sects. [4.1.2](#sec_irx_luv){reference-type="ref" reference="sec_irx_luv"} and [4.2](#sec_irx_mass){reference-type="ref" reference="sec_irx_mass"} show that dust attenuation is on average well correlated with stellar mass, and that this correlation is tighter than the correlation between dust attenuation and \(L_{\rm FUV}\). However, dust attenuation is not *completely* independent of \(L_{\rm FUV}\) : while at \(z\sim1.5\), dust attenuation is mostly constant for \(5 \times 10^{9}<L_{\rm FUV}/\textrm{L}_{\odot}<5 \times 10^{10}\) it increases for fainter UV luminosities. On top of this, dust attenuation is higher at \(z\sim3\) than at \(z\sim1.5\) at the same \(L_{\rm FUV}\), but is found to be decreasing with \(L_{\rm FUV}\). It seems then that dust attenuation depends *both* on \(L_{\rm FUV}\)  and \(M_*\), and that we need to investigate what is the link between dust attenuation and these two quantities. We performed stacking as a function of \(L_{\rm FUV}\)  and \(M_*\) at \(z\sim1.5, 3,\) and 4, using binnings of \(\left(\Delta \log(L_{\rm FUV}/\textrm{L}_{\odot}), \Delta \log(M_*/\textrm{M}_{\odot})\right) = (0.3,0.3), (0.3,0.4)\), and \((0.4,0.4)\) respectively. We show in Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"} the result of the stacking as a function of UV luminosity and stellar mass. Note that filled cells indicate bins where the stacking measurements have \(S/N>3\) in all SPIRE bands, hatched cells bins where there is at most two SPIRE band with \(S/N>3\), and other cells are kept empty. These empty cells indicate that there is no robust stacking detection in these bins. The measurements in Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"} clearly show that dust attenuation depends both on \(L_{\rm FUV}\)  and \(M_*\). Dust attenuation increases with \(M_*\) at a given \(L_{\rm FUV}\), while it decreases with \(L_{\rm FUV}\)  at a given \(M_*\). We already observed an increase of the dust attenuation for faint UV galaxies at \(z\sim1.5\) [also observed previously by @Buat_2009; @Buat_2012; @Burgarella_2006]. Indeed, galaxies with large stellar masses and strong dust attenuation exhibit faint UV luminosities, which is true for all redshifts we study here. The results in Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"} also show that the range of dust attenuation values over the stellar mass range decreases with \(L_{\rm FUV}\), as suggested in a previous study. We also represent in Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"} the location of the mean stellar mass for each UV luminosity bin. The results at \(z\sim 1.5\) in particular show that lines of constant dust attenuation follow lines roughly parallel to this relation. This explains the global lack of dependence of dust attenuation with \(L_{\rm FUV}\) at \(z\sim1.5\). At \(z\sim3\) and \(z\sim4\), there is only a weak correlation between \(M_*\) and \(L_{\rm FUV}\). This implies that bins in \(L_{\rm FUV}\)  are mostly dominated by low mass galaxies in these samples. As shown in Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"}, the dust attenuation at a given mass decreases with \(L_{\rm FUV}\), which is exactly what we observe when stacking as a function of \(L_{\rm FUV}\)  only. The relation between dust attenuation and \((L_{\rm FUV},M_*)\) also depends on redshift. Indeed, at a given mass and \(L_{\rm FUV}\), the attenuation is higher at \(z\sim3\) than at \(z\sim1.5\). For instance, galaxies with \(10^{10}\la L_{\rm FUV}/L_{\odot}\la 10^{10.35}\) have a dust attenuation roughly 0.2 dex larger at a given mass at \(z\sim3\) with respect to galaxies at \(z\sim1.5\). This, combined with slightly different \(L_{\rm FUV}\) -\(M_*\) relations explains why the dust attenuation for this range of UV luminosities is larger at \(z\sim3\) compared to \(z\sim1.5\) (see Fig. [\[fig_stack_lfuv\]](#fig_stack_lfuv){reference-type="ref" reference="fig_stack_lfuv"}). We can use the results presented above in order to provide empirical recipes to estimate dust attenuation as a function of \(M_*\) and \(L_{\rm FUV}\). We detail those in Appendix [\[app_emp_irx\]](#app_emp_irx){reference-type="ref" reference="app_emp_irx"}. ## Star formation rate-stellar mass relations from \(z\sim4\) to \(z\sim1.5\) {#sec_sfr_mass} \ The fits are performed assuming that \(\textrm{SFR} = \textrm{SFR}_0M_{*}^{\alpha}\). The measurements presented above yield average estimates of \(L_{\rm IR}\)  as a function of stellar mass at \(z\sim1.5, 3,\) and 4. We can combine these measurements with those of the observed, uncorrected UV luminosities to obtain a total star formation rate as: \[\rm{SFR} = SFR_{\rm IR} + SFR_{\rm UV}\] with \[\begin{aligned} \rm SFR_{\rm IR}[\textrm{M}_{\odot} \rmn{yr}^{-1}] & = & 1.09\times10^{-10} L_{\rm IR} [\rm L_{\odot}]\\ \rm SFR_{\rm UV}[\textrm{M}_{\odot} \rmn{yr}^{-1}] & = & 1.70\times10^{-10} L_{\rm FUV} [\rm L_{\odot}] \end{aligned}\] where we use the factors from that we converted from a to a IMF. We show in Fig. [\[fig_stack_sfr\]](#fig_stack_sfr){reference-type="ref" reference="fig_stack_sfr"} the average SFR-mass relations we obtain at \(z\sim1.5\), \(z\sim3\), and \(z\sim4\), along with best fits from a number of previous studies (references on the figure). We find that there are well defined average SFR-mass relations in our UV-selected samples at the epochs we focus on. The SFR-mass relations at \(z\sim4\) and \(z\sim3\) are similar to each other, while at a given \(M_*\) the SFR is around 4 times lower at \(z\sim1.5\). We note that SFR is here equivalent to \(L_{\rm IR}\)  for \(M_* \ga 10^{10}\,\textrm{M}_{\odot}\), the UV contribution to the SFR being negligible, as \(L_{\rm IR}/L_{\rm FUV} >10\) in this range of masses (see Fig. [\[fig_stack_M\*\]](#fig_stack_M*){reference-type="ref" reference="fig_stack_M*"}). Fig. [\[fig_stack_sfr\]](#fig_stack_sfr){reference-type="ref" reference="fig_stack_sfr"} shows that UV-selected samples do probe the ULIRGs regime at \(z\sim3,4\) for \(M_* \ga10^{10}\,\textrm{M}_{\odot}\) as a SFR of \(100\,\textrm{M}_{\odot}\)yr\(^{-1}\) correspond roughly to \(L_{\rm IR}=10^{12}\,\textrm{L}_{\odot}\). This is different from what is suggested by Figs. [\[fig_stack_lfuv_vs_lir\]](#fig_stack_lfuv_vs_lir){reference-type="ref" reference="fig_stack_lfuv_vs_lir"} and [\[fig_stack_lfuv\]](#fig_stack_lfuv){reference-type="ref" reference="fig_stack_lfuv"}. The origin of this difference is the underlying relations between \(L_{\rm FUV}\), \(L_{\rm IR}\), and \(M_*\). When stacking as a function of \(M_*\), ULIRGs are recovered in a UV selection. There are on the other hand not recovered while stacking as a function of \(L_{\rm FUV}\), because they are mixed with other galaxies which have fainter \(L_{\rm IR}\). This shows that \(L_{\rm FUV}\)  is not well correlated with \(L_{\rm IR}\)  and \(M_*\). The SFR-mass relations we observe are well described by power laws with an average slope of 0.7; we provide fits for these relations in Table [2](#tab_sfr_mass_fits){reference-type="ref" reference="tab_sfr_mass_fits"}. Note nevertheless that at \(z\sim1.5\), the SFR-mass relation we observe is better described by a broken power law, with a slope of \(\sim 0.85\) for \(M_*<10^{10.5}\,\textrm{M}_{\odot}\) and a shallower slope \(\sim 0.5\) for higher masses. We compare our results with previous determinations of the SFR-mass at various redshifts. At \(z=1\), the average relation from, derived from a restframe optical selection and using \(24\,\mu\)m observations to constrain the amount of dust attenuation, has a lower amplitude than ours. Our results at \(z\sim3, 4\) and \(z\sim1.5\) bracket those at \(z=2\) of and. based their study on a \(K-\)band selection and \(24\,\mu\)m observations, while at the same redshift used optical selections and a combination of FIR observations (including *Herschel*/PACS) and SED fitting for dust attenuation. At \(z=3\), derived a SFR\(-M_*\) relation for LBGs with IRAC observations, and correcting for dust attenuation using the UV slope of the continuum. Our results at \(z\sim3\) agree with theirs at the high mass end, but have a higher amplitude in the lower mass range we explore. On the other hand, our measurements are in good agreement with those from @Bouwens_2012 [based on a LBG sample, and using the slope of the UV continuum to correct for dust attenuation] at \(z\sim4\) in the range of masses where they overlap, as well as if we extrapolate them at higher masses. In summary, the SFR\(-M_*\) relations we obtain are in good agreement with these other studies. ## Intrinsic and observed relations between dust attenuation and \(M_*\) for UV-selected galaxies {#sec_irx_bias} We investigate here the impact of the faint UV population on the recovery of the relation between dust attenuation and stellar mass. We follow the approach of to create a mock catalogue, which has the following properties: \(L_{\rm FUV}\), \(L_{\rm IR}\), SFR, and \(M_*\). Our goal here is to model the intrinsic relation between dust attenuation and stellar mass, by taking into account galaxies fainter than the detection limit. We focus here on the \(z\sim1.5\) case, but show in Appendix [\[app_simul_irx_mass\]](#app_simul_irx_mass){reference-type="ref" reference="app_simul_irx_mass"} results for \(z\sim3\) and \(z\sim4\). In practice, we consider the best fit of the UV luminosity function at \(z\sim1.5\) we determined for our sample, down to \(L_{\rm FUV}=10^{8}\rm{L}_{\odot}\). We build a mock catalogue by assigning UV luminosities according to this luminosity function. Then we assign a FIR luminosity to each object of this catalogue. We assume that the distribution of \(\log(L_{\rm IR}/L_{\rm FUV})\) is a Gaussian. We use as mean of this distribution the stacking results from, and as dispersion, the dispersion required to reproduce the few per cents of UV-selected objects detected at SPIRE wavelength. We only have measurements for objects brighter that \(L_{\rm FUV} = 10^{9.5}\rm{L_{\odot}}\). For fainter objects, we assume that \(\log(L_{\rm IR}/L_{\rm FUV})\) is constant, as well as its dispersion, using the results from. The values of these constants are \(\log(L_{\rm IR}/L_{\rm FUV})_{\rmn{faint}} = 0.94\), and \(\sigma(\log(L_{\rm IR}/L_{\rm FUV}))_{\rmn{faint}} = 0.73\). The value \(\log(L_{\rm IR}/L_{\rm FUV})_{\rmn{faint}}\) is higher than the average value for the sample \(\langle \log(L_{\rm IR}/L_{\rm FUV})\rangle = 0.84\pm0.06\), but consistent with the values measured at the faint end of the sample. We determined this value in such that the IR luminosity function of a UV selection recovers the IR luminosity function of a IR selection. Given the limited constraints on the latter, the assumption that \(\log(L_{\rm IR}/L_{\rm FUV})\) and its dispersion are constant for \(L_{\rmn {FUV}}\) fainter that the limit of our sample is necessary. The conclusions we draw from this modeling exercise would differ if the average IR to UV luminosity ratio for galaxies fainter than the limit of our sample is similar to that of the galaxies of the sample, which is unlikely given the available data. Having now a mock catalogue with \(L_{\rm FUV}\)  and \(L_{\rm IR}\), we can assign a SFR to each of the objects by adding the IR and UV contributions. We finally assign a stellar mass by assuming the average SFR-mass relation we observe at \(z\sim1.5\), and assuming a dispersion of 0.15 dex. Note that this value might underestimate the actual dispersion of the SFR-mass relation, but this does not have a strong impact on our results here. We also checked that there is no impact of incompleteness in UV on the SFR-mass relation we observe (see Appendix [\[app_sfr_mass\]](#app_sfr_mass){reference-type="ref" reference="app_sfr_mass"}). We show in Fig. [\[fig_simul_irx\]](#fig_simul_irx){reference-type="ref" reference="fig_simul_irx"} our modeled intrinsic IR to UV luminosity ratio as a function of stellar mass and per bins of \(L_{\rm FUV}\) from this mock catalogue. Note that we attempt to model the intrinsinc distribution, but that our mock catalogue is also self-consistent as we recover the observed dust attenuation-stellar mass relation for galaxies with \(L_{\rm FUV}>10^{9.5}\,\textrm{L}_{\odot}\). The results from Fig. [\[fig_simul_irx\]](#fig_simul_irx){reference-type="ref" reference="fig_simul_irx"} show that fainter objects in UV have smaller stellar masses and higher dust attenuation. Our mock catalogue suggests that we observe a relation between the IR to UV luminosity ratio and \(M_*\) partly because we are probing a limited range of \(L_{\rm FUV}\). We note also that we observe that the dispersion in dust attenuation is larger for fainter galaxies [see @Heinis_2013 and also Fig. [\[fig_stack_mstar_luv\]](#fig_stack_mstar_luv){reference-type="ref" reference="fig_stack_mstar_luv"}]. Our mock catalogue shows that this dispersion actually originates from the \(L_{\rm IR}/L_{\rm FUV}-M_*\) relation. Our previous results also suggest that galaxies fainter than the current sensitivity levels in UV restframe luminosity (i.e. down to \(u^*\sim 30\)) are dustier. If that is the case, this suggests then that the actual average relation between \(L_{\rm IR}/L_{\rm FUV}\) and stellar mass has a higher amplitude than the one we are observing, and also that the actual dispersion in dust attenuation at a given stellar mass is much higher, because of faint UV galaxies. # Discussion {#sec_discussion} ## Impact of UV-selection on SFR-Mass relations {#sec_sel_uv_mass} We derive here average SFR-\(M_*\) relations for UV-selected samples from \(z\sim4\) to \(z\sim1.5\). While the relations we obtain are not strongly sensitive to incompleteness in the UV, our results are not drawn from a mass selection. We investigate here whether this has any impact on our results. We note first that we derive SFR-\(M_*\) relations which have slopes consistent with 0.7 from \(z\sim4\) to \(z\sim3\), which is shallower than the value of \(\sim1\) derived by a number of studies [@Elbaz_2007; @Daddi_2007; @Magdis_2010; @Wuyts_2011b], but in agreement with. This shallower slope might be caused by the fact that we are selecting galaxies by their UV flux, and hence missing objects which have low star formation rates. To further examine this, we compare in Fig. [\[fig_mass_functions\]](#fig_mass_functions){reference-type="ref" reference="fig_mass_functions"} the mass function of our sample at \(z\sim1.5\) with mass functions derived from a mass-selected sample, based on \(3.6\,\mu\)m data[^13]. This comparison shows that the mass function of our UV-selected sample is similar to the total mass function of star-forming galaxies only at the low mass end, and is otherwise lower. also divided their sample into high activity and intermediate activity star forming galaxies, based on the restframe \(NUV-R\) color. Fig. [\[fig_mass_functions\]](#fig_mass_functions){reference-type="ref" reference="fig_mass_functions"} shows that the mass function of UV-selected galaxies at \(z\sim1.5\) is similar to that of high activity star-forming galaxies at \(M_*<10^{10.5}\,\textrm{M}_{\odot}\), while it is larger above this mass. On the other hand, the mass function of UV-selected galaxies at \(z\sim1.5\) is lower than that of intermediate star-forming galaxies at \(M_*>10^{10.5}\,\textrm{M}_{\odot}\). This comparison suggests that the UV-selection at \(z\sim1.5\) is likely to probe the full population of highly star-forming galaxies, while it may miss roughly half the number density of intermediate star-forming ones at \(M_*>10^{10.5}\,\textrm{M}_{\odot}\). We note that at \(z\sim3\) and \(z\sim4\) UV-selected samples also miss a significant fraction of high stellar mass galaxies. This shows that the amplitudes of our SFR-Mass relations might be overestimated, and also that there might be an impact on the slope of these relations, if these high stellar mass galaxies we are missing have high SFR and large dust attenuation. On the other hand, we can also in this context compare our results to those from, who perform radio stacking on a mass-selected sample. They derive SFR-mass relations which have an amplitude at most 2 times lower than ours, and a similar slope. Note that measure SFRs from stacking in VLA-radio data. While some contamination by AGN is possible, we consider here for the comparison their results from star-forming galaxies, which are not expected to be dominated by radio-AGNs [@Hickox_2009; @Griffith_2010]. ## Impact of star formation history on conversion from observed UV and IR luminosities to SFR The values of the factors commonly used to convert from UV or IR luminosities to SFR assume that the star formation has been constant over timescales of around 100 Myrs. While useful, this assumption is not correct for galaxies with other star formation histories. The impact of the star formation history on the conversion from \(L_{\rm FUV}\)  or \(L_{\rm IR}\)  to SFR has been studied by various authors [including @Kobayashi_2013; @Reddy_2012b; @Schaerer_2013]: in the early phases of star formation (\(t<10\,\)Myr), the actual conversion factors are larger than the values (implying that the SFR values are underestimated when adopting the conversion factor from ), while for later phases there are lower. The amplitude of the difference depends on the star formation history, with faster evolutions yielding larger differences. In our case, if we assume that our SFR values are overestimated, this means that the bulk of our samples is a population of galaxies in later phases of star formation, with rapidly declining star formation histories, like starbursts for instance. It is beyond the scope of this paper to characterise precisely the star formation histories of the galaxies in our samples. We can however base our argumentation on the results of SED fitting of dropouts at \(3<z<6\) from. They found that the currently available data is suggesting that these galaxies experienced either exponentially declining or delayed star formation histories. They also note in particular that, assuming their SED fitting, the SFR would be slighty *underestimated* if the conversion factors would have been used. Moreover, showed by backtracing galaxies using different star formation histories that the declining star formation scenario does not enable to reproduce the number densities of star-forming galaxies between \(z=4\) and \(z=0\). In summary, given the state-of-the art SED fitting, we believe that the impact of star formation histories different from that assumed by is negligible on our results. ## Evolution of specific star formation rate with redshift Our measurements show that the amplitude of the SFR-\(M_*\) relation is similar between \(z\sim4\) and \(z\sim3\), and then decreases significantly from \(z\sim3\) to \(z\sim1.5\). Another way to look at these results is to consider the specific star formation rate \(\rm{sSFR} = \rm{SFR}/M_*\) which is an indicator of star formation history, in the sense that it is the inverse of the time needed for a galaxy to double its mass if it has a constant SFR. We show in Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"} the evolution with redshift of the specific star formation rate for three mass bins: \(10^{10}\,\textrm{M}_{\odot}\), \(10^{10.5}\,\textrm{M}_{\odot}\), and \(10^{11}\,\textrm{M}_{\odot}\). We compute the average SFR for our samples by stacking galaxies in bins of stellar mass centered on these values, with sizes of 0.2 dex at \(z\sim1.5\) and \(z\sim3\), and a size of 0.4 dex at \(z\sim4\). We compare our results to the measurements of @Daddi_2007 [@VanDokkum_2013; @Karim_2011; @Magdis_2010; @Noeske_2007; @Wuyts_2011b]. At \(z>4\), there are basically no results yet in the mass range we explore. We show here an extrapolation of the results from and. give values of sSFR at \(M_* = 5\times10^9\,\textrm{M}_{\odot}\) corrected from dust attenuation (based on the UV slope of the continuum), using their own sample at \(z=4\), and the results from and at higher redshifts. derive sSFRs also at \(M_* = 5\times10^9\,\textrm{M}_{\odot}\) at \(4<z<7\), taking into account the impact of emission lines on the measure of stellar masses, and correcting from dust attenuation using the slope of the UV continuum. We extrapolate results from both studies in our mass range assuming that there is a power law relation between SFR and stellar mass at \(z>4\), and that the slope of this relation is between 0.7 (the value measured at \(z=4\) by, also consistent with our results) and 1 [closer to the value observed at lower redshifts by other studies like @Wuyts_2011b]. Our results are in overall agreement with previous measurements at \(z\sim1.5\). Note that all measurements are significantly higher than those of, who derived the star formation history of Milky Way-like galaxies (see Sect. [5.4](#sec_sfh_ms){reference-type="ref" reference="sec_sfh_ms"} for further discussion). At \(z\sim3\), our measurements are quite high compared to the values from previous studies, in particular at \(M_*=10^{10}\,\textrm{M}_{\odot}\). In this mass bin, our estimates are larger than the measurements from and, but they are consistent at \(1.2\,\sigma\) and \(0.3\,\sigma\) respectively. In other word, our sSFR results represent the upper range of available measurements. Note however that our results are in very good agreement with those of at \(z\sim3\) for \(M_*=10^{10.5.}\,\textrm{M}_{\odot}\) and \(M_*=10^{11.}\,\textrm{M}_{\odot}\). At \(z\sim4\), our results agree with those from and at \(z=4\). Our results are also in agreement with the sSFR being constant at \(3<z<4\), while the results of suggest that the sSFR is increasing at higher redshifts (\(z>5\)). We compare our results with a few models, from, ,, , and. These models are quite different and give a sample of various simulation techniques available. We briefly describe all of them. @Boissier_2000 [see also and ] built an analytical model which predicts the chemical and spectrophotometric evolution of spiral galaxies over the Hubble time. This model reproduces a large number of present properties of the Milky Way and local spiral galaxies (such as: color-magnitude diagrams, luminosity-metallicity relationship, gas fractions, as well as color and metallicity gradients). based their model under the assumption that the gas accretion in galaxies is mostly driven by the growth of dark matter haloes. They also assume that the gas accretion efficiency decreases with cosmic time, and is only efficient for dark matter haloes of masses \(10^{11}<M_h/\textrm{M}_{\odot}<1.5\times10^{12}\). ran hydrodynamical simulations which include galactic outflows, implementing several models for winds; we show on Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"} the range of sSFR spanned by these models, including the model without winds. compared the predictions from three semi-analytical models, namely those of. All three models are based on the combination of dark matter simulations complemented by empirical relations for baryonic physics. All these models include supernovae and AGN feedback. We show on Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"} the range of sSFR spanned by these three models. studied the mass assembly of galaxies using abundance matching models, by matching observed stellar mass functions simultaneously at various redshifts. The comparison in Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"} of observations and models shows that models match the observations roughly well at low redshift [\(z<0.5\), see also e.g. @Damen_2009], underestimate the sSFR up to \(z=4\), and are potentially in better agreement at higher redshifts. An interesting point is that the models we consider here are quite different in terms of implementation and assumptions; however they all predict a similar evolution which does not match the observations for \(0.5\la z\la4\). At \(M_*=10^{10}\textrm{M}_{\odot}\), the model of and the compilation of models from are the closest to the observations among the ones we consider here. Still, these models do not reproduce the high sSFR we observe at \(z\sim3\). At \(M_*=10^{10.5}\textrm{M}_{\odot}\), the model of presents the same level of agreement with our measurements, while the discrepancies between the compilation of and the observations are more important. We note also that all these models are actually more or less consistent with the redshift evolution expected according to the cold gas accretion scenario. This scenario predicts that the baryonic accretion onto galaxies follows directly the dark matter accretion onto dark matter haloes, and evolves as \(\dot{M} \propto (1+z)^{2.25}\). Our results show that this scenario is in agreement with the observations for \(0< z \la 1.5\), but is less efficient at reproducing galaxies properties at \(1.5\la z\la3\). There has been some attempts to reconcile model predictions with the observations of the redshift evolution of the sSFR. noted that a number of observations suggest that the IMF is not universal and could evolve with redshift, in the sense that it would be weighted towards more massive stars at high redshift. Such an IMF would imply that SFRs as derived here are *overestimated* with respect to using an evolving IMF, by a factor that increases with redshift, being around 4 at \(z=4\). Whether the IMF is universal, or evolves with redshift, remains to date a controversial subject. Indeed recent studies suggest in contrary to that there is observational evidence for bottom-heavy IMF at high redshift [see e.g. @VanDokkum_2012]. considered a number of modifications to semi-analytical models in order to match the observed redshift evolution of the sSFR. They found that models can match the observations at \(z>4\) if there is either strong stellar feedback at high redshift at all masses, or inefficient star formation. At \(z=2-3\), where the models underpredict the sSFR, the feedback could drop, or gas which was prevented to form stars earlier could be at that time available for star formation. We provide new and improved observational constraints to test these scenarios. Future observations of the gas content of high redshift galaxies will also enable to discriminate between those. ## The star formation histories of Main sequence galaxies {#sec_sfh_ms} Our measurements bring new constraints at high redshift on the sSFR of the Main Sequence galaxies. We can use these results to derive the star formation history of galaxies staying on the Main Sequence. We first recall that galaxies can not remain on the Main Sequence from high redshift to \(z=0\), given the stellar masses and SFR they would have in the local Universe. We then give estimates of the timescale galaxies can stay on the Main Sequence before quenching of the star formation. We consider here a parameterised form of the dependence with redshift and stellar mass of the sSFR of the Main Sequence. We follow the approach of, and we assume that: \[\begin{aligned} \label{eq_ssfr_z} \rmn{sSFR}_{\rmn{MS}}(z, M_*) & = & \rmn{sSFR_{\rmn{MS, 0}}}\times \left(\frac{M_{*}}{10^{11}\,\textrm{M}_{\odot}}\right)^{\beta_{\rmn{MS}}}\\ \nonumber & & \times\left(1+\min(z, z_{\rm evo})\right)^{\gamma_{\rmn{MS}}}\rmn{,} \end{aligned}\] where \(\rmn{sSFR_{\rmn{MS, 0}}}\) is the sSFR of the Main Sequence at \(z=0\) for galaxies of \(M_* = 10^{11}\textrm{M}_{\odot}\), \(\beta_{\rmn{MS}}\) is the slope of the sSFR-\(M_*\) relation, and \(\gamma_{\rmn{MS}}\) encodes the power-law redshift evolution of the amplitude of the sSFR-\(M_*\) relation. We modify the values of these parameters to match our measurements as well as the measurements at lower redshifts from: \(\rmn{sSFR}_{\rmn{MS, 0}}=10^{-10.66} \rmn{yr}^{-1}\), \(\beta_{\rm MS}=-0.33\), \(z_{\rm evo} = 2.16\), and \(\gamma_{\rm MS}=4.4\). We show the resulting sSFR evolution using these parameters as a red line on Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"}. We note that eq. [\[eq_ssfr_z\]](#eq_ssfr_z){reference-type="ref" reference="eq_ssfr_z"} can also be written as \[\begin{aligned} \label{eq_ssfr_t} \frac{1}{1-R}\frac{1}{M_*}\frac{\rmn{d}M_*}{\rmn{d}t} & = & \rmn{sSFR_{MS, 0}}\times \left(\frac{M_{*}}{10^{11}\,\textrm{M}_{\odot}}\right)^{\beta_{\rm MS}}\\ \nonumber & & \times\left(1+\min\left(z(t), z(t_{\rm evo})\right)\right)^{\gamma_{\rm MS}} \end{aligned}\] where \(t_{\rm evo}\) is the lookback time corresponding to \(z_{\rm evo}\). We wrote the SFR in terms of the derivative of \(M_*\) with respect to time assuming that \[\frac{\rmn{d}M_*}{\rmn{d}t} = \rmn{SFR}(1-R)\rmn{.}\] \(R\) is the return fraction, that we set to \[R = 0.05\ln\left(1+\frac{\Delta t}{0.03\rmn{Myr}}\right)\] where \(\Delta t\) is the time elapsed since the formation of stars. We can then use the fact that eq. [\[eq_ssfr_t\]](#eq_ssfr_t){reference-type="ref" reference="eq_ssfr_t"} is a differential equation for \(M_*(z)\). We obtain \(M_*(z)\), and from this SFR\((z)\). This procedure requires boundary conditions of stellar mass at a given redshift. In other words, we can start the integration of eq. [\[eq_ssfr_t\]](#eq_ssfr_t){reference-type="ref" reference="eq_ssfr_t"} at any redshift, but we need to choose an initial stellar mass at this redshift. This means that we are making galaxies 'enter' on the Main Sequence at these stellar mass and redshift. We are considering here only the mean location of the Main Sequence. This means that, prior to entering the Main Sequence in the sense of this simple model, galaxies could for instance be lower in the SFR-Mass plane, but still within the Main sequence at redshifts higher than this initial redshift. We consider here the result of, who derive the star formation history of Milky Way-like galaxies, by studying up to \(z=2.5\) galaxies with the same number density as galaxies with the stellar mass of Milky Way at \(z=0\). derive the redshift evolution of the stellar mass of such galaxies. We use their fit to get initial stellar mass at a given redshift[^14]. We integrate eq. [\[eq_ssfr_t\]](#eq_ssfr_t){reference-type="ref" reference="eq_ssfr_t"} down to \(z=0\), starting from various initial redshifts, which we consider between \(z=3\) and \(z=0.5\). We show on Fig. [\[fig_sfr_mass_integration\]](#fig_sfr_mass_integration){reference-type="ref" reference="fig_sfr_mass_integration"} the evolution of the stellar mass and SFR for galaxies which remain on the Main Sequence and have the same stellar mass as the Milky Way at these initial redshifts. Doing so we look at the star formation history of galaxies which have the same stellar mass as the Milky Way at these initial redshifts, and stay on the Main Sequence until \(z=0\)[^15]. Assuming that a galaxy is on the Main Sequence for \(1\la z<3\) leads to much higher SFR and stellar mass than the Milky Way at \(z=0\). On the other hand, if we assume that the Milky Way is on the Main Sequence between \(z=0.5\) and \(z=0\), we obtain a stellar mass similar to the Milky Way at \(z=0\), and a SFR around 2 times higher. Note that galaxies with \(M_* \sim 10^{10}\,\textrm{M}_{\odot}\) at \(z\sim 2\) would have \(M_* \sim 2\times10^{11}\,\textrm{M}_{\odot}\) at \(z=0\). This is in strong disagreement with measurements of the redshift evolution of the stellar mass functions of star forming galaxies which show little evolution between \(z=2\) and \(z=0\) at the high mass end. The star formation histories on Fig. [\[fig_sfr_mass_integration\]](#fig_sfr_mass_integration){reference-type="ref" reference="fig_sfr_mass_integration"} are actually quite different from that expected for the Milky Way (dotted line on bottom panel), even though we assumed the observed stellar mass of Milky Way-like galaxies at various redshifts as boundary conditions. This is actually due to the fact that the Milky Way is not on the mean location of the Main Sequence for \(1<z<2\) (see crosses showing the measurements of on Fig. [\[fig_ssfr_z\]](#fig_ssfr_z){reference-type="ref" reference="fig_ssfr_z"}). Assuming the values from and the results of for the distribution of galaxies in the \((\rmn{SFR},M_*)\) plane suggests that the Milky Way is rather on the lower enveloppe of the Main Sequence for \(0<z<2\). Our results suggest on the other hand that the sSFR of star-forming galaxies is quite high at \(z=3,4\), which yields a high SFR peak in the derived star formation histories. The results shown on Fig. [\[fig_sfr_mass_integration\]](#fig_sfr_mass_integration){reference-type="ref" reference="fig_sfr_mass_integration"} suggests that the assumption that galaxies remain on the Main Sequence until \(z=0\) is not correct. The consequence is that the Main Sequence is built of different star-forming galaxies at various redshifts. These results raise the question of the amount of time galaxies can stay on the Main Sequence. In order to determine this time, we need to define a criterion to determine the epoch when galaxies exit the Main Sequence. We use here the 'quenching mass' (\(M_{\rm Q}\)) as defined by. We used the same method as above to investigate this. We consider once again eq. [\[eq_ssfr_t\]](#eq_ssfr_t){reference-type="ref" reference="eq_ssfr_t"}, but this time we stop the integration, i.e. we make galaxies exit the Main Sequence, at the redshift when their stellar mass is larger than the quenching mass at the same time. Galaxies experiencing quenching of star formation exit the Main Sequence by going down in the \((\rmn{SFR},M_*)\) plane at a given M\(_*\). We do not consider here starbursts galaxies as they represent a significantly smaller number density. We follow and assume that the quenching mass is the mass where the number density of quiescent galaxies is maximum. We consider the measurements from of the mass function of quiescent galaxies (available for \(0.5<z<3\)) and complement them at \(z=0\) by the measurement of. The evolution with redshift of the quenching mass can be adjusted to the following form: \[M_{\rm Q}(z)[\textrm{M}_{\odot}] = 3.7\times10^{10}\times(1+z)^{0.53}\] We make the galaxies enter the Main Sequence at redshifts \(1<z<4\), and at masses in the range \(10^8<M_*/\textrm{M}_{\odot}<10^{10}\). We show the time galaxies stay on the Main Sequence in Fig. [\[fig_time_MS\]](#fig_time_MS){reference-type="ref" reference="fig_time_MS"}. We perform the integration only until \(z=0\); in other words, we do not derive times larger than the time to \(z=0\) for galaxies that have not reached \(M_{\rm Q}\) at \(z=0\). This means that galaxies that are still on the Main Sequence at \(z=0\) are represented by locations on the dashed line on the top panel of Fig. [\[fig_time_MS\]](#fig_time_MS){reference-type="ref" reference="fig_time_MS"}, or at \(z=0\) in the bottom panel. Given our assumptions, our results show that galaxies which enter the Main Sequence at \(z<4\) stay on it at least 1 Gyr. As expected, at a given entrance redshift on the Main Sequence, less massive galaxies spend more time on the Main Sequence to reach the quenching mass. Galaxies entering on the Main Sequence at \(2.5<z<4\) stay around 1 Gyr on it. At lower redshifts, the quenching mass decreases, but the average sSFR also decreases, which in turn yields that galaxies stay longer on the Main Sequence. For instance, with the scenario we consider here, galaxies with masses \(10^8<M_*/\textrm{M}_{\odot}<10^{10}\) which enter the Main Sequence at \(z<1.2\) stay on the Main Sequence until \(z=0\). and reach similar conclusions regarding the star formation histories of Main Sequence galaxies at \(z<2\). We assumed here that the sSFR is constant for \(z>2.16\). Assuming that the sSFR increases with \(z\) from \(z\sim3\) would mean faster evolution for high redshift galaxies, implying: stronger disagreement for the evolution of the Milky Way as discussed here, and shorter times on the Main sequence for high redshift galaxies. We note that the simplistic calculation presented here requires to be tested against the redshift evolution of the stellar mass functions of quiescent and star forming galaxies, which is beyond the scope of this paper, and will be the subject of forthcoming work. # Conclusions {#sec_conclusion} We studied the FIR properties of large samples of UV-selected galaxies at \(1.5<z<4\), by combining the COSMOS multiwavelength dataset with the HerMES/*Herschel* SPIRE imaging. We measured by stacking the average IR luminosity as a function of UV luminosity, stellar mass, and both. Our results can be summarised as follows: 1. At \(z\sim1.5\), there is a good correlation between \(L_{\rm IR}\)  and \(L_{\rm FUV}\) (\(8\times10^{9}<L_{\rm FUV}/L_{\odot}<5\times10^{10}\)), while at \(z\sim3\) and \(z\sim4\), \(L_{\rm IR}\)  and \(L_{\rm FUV}\)  are not well correlated. 2. Consequently, the ratio \(L_{\rm IR}\) \(/\[L_{\rm FUV}\)  at \(z\sim3,4\) is decreasing with \(L_{\rm FUV}\). 3. The average dust attenuation (as traced by the \(L_{\rm IR}\) \(/\]L_{\rm FUV}\)  ratio) is well correlated with stellar mass at \(1.5<z<4\), and does not show significant evolution in this redshift range, in the range of masses we explore. 4. We investigated the joint dependence of dust attenuation with stellar mass and \(L_{\rm FUV}\). While well correlated with stellar mass, dust attenuation also shows secondary dependence on \(L_{\rm FUV}\). At a given stellar mass, dust attenuation decreases with \(L_{\rm FUV}\); at a given \(L_{\rm FUV}\), dust attenuation increases with stellar mass. We also provide empirical relations between dust attenuation, \(M_*\), and \(L_{\rm FUV}\), at \(z\sim1.5\) and \(z\sim3\). 5. The average SFR-\(M_*\) relations for UV-selected samples at \(1.5<z<4\) are well approximated by a power law, with a slope of around 0.7. At a given stellar mass, the average SFR is similar at \(z\sim3\) and \(z\sim4\), but is 4 times higher than at \(z\sim1.5\). 6. Our results provide new constraints on the sSFR at \(1.5<z<4\). Current models of galaxy formation and evolution do not reproduce accurately the sSFR evolution we observe, in particular at \(z\sim3\) and \(z\sim4\), where standard models underpredict the observations. 7. We use our results for the evolution of the sSFR with redshift to characterise the star formation histories of Main Sequence galaxies. We find that galaxies would have too large stellar masses if they stay on the Main Sequence from high redshift to \(z=0\). Assuming that galaxies exit the Main Sequence when their stellar mass is equal to the 'quenching mass', we determine the time galaxies stay on the Main Sequence. This suggests that galaxies stay around 1 Gyr on the Main Sequence at high redshift (\(2.5<z<4\)), while they stay longer on the Main Sequence at lower redshifts. For instance, Main Sequence galaxies (with \(10^8<M_*/\textrm{M}_{\odot}<10^{10}\)) at \(z=1\) stay until \(z=0\) on the Main Sequence, as they do not reach the quenching mass.

{'timestamp': '2013-10-14T02:10:44', 'yymm': '1310', 'arxiv_id': '1310.3227', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3227'}

# Introduction Diversity serves as one of the major solutions to combat channel impairment caused by random fading in wireless environments. Recently, cooperative communication has emerged as a promising technique of achieving spatial diversity in a distributed fashion. A variety of cooperation schemes such as opportunistic relaying and space-time coded cooperation have been proposed to provide full cooperative diversity (CD) in multi-relay networks. Among these schemes, opportunistic relaying achieves full CD by selecting the "best" relay to support transmission. Moreover, it is outage-optimal under an aggregate power constraint, and can be implemented with low complexity, hence it attracts much attention. On the other hand, it is well known that multiuser diversity (MUD) constitutes an inherent resource of diversity in a multiuser network. Since many users experience independent fading, the probability that the "best" user has a "strong" channel is very high. Therefore, by allowing only the user with the highest instantaneous signal-to-noise ratio (SNR) to transmit, MUD can be obtained to improve the outage probability and/or capacity performance. In multiuser cooperative networks, it is potentially feasible to achieve both CD and MUD, and there have been some studies focusing on the combination of CD and MUD. More specifically, and discussed this combination in some specific cooperative networks from the capacity perspective, other literature investigated the reliability performance of the combined use of CD and MUD. The authors of established a multiuser cooperative network model where each user transmits with the aid of one exclusive relay, and analyzed the diversity order for both the amplify-and-forward (AF) and the decode-and-forward (DF) protocols. Furthermore, in, they extended the analysis of to a more generalized multiuser network model in which each user has multiple exclusive relays. However, the assumption of exclusive relay might not be realistic although it brings convenience to theoretical analysis. The authors of considered a more practical scenario where all the users share all the relays, and proposed an optimal "user-relay" pair selection strategy to achieve CD and MUD simultaneously. Nevertheless, global channel state information (CSI) is needed to perform such "user-relay" pair selection. Namely, in an \(N\)-user \(M\)-relay network, the CSI of all the \(N\left(M+1\right)\) links in the network is required for a single "user-relay" pair transmission. This requirement makes the complexity of selection excessively high for large \(N\) and \(M\). To reduce the complexity, the authors of proposed a two-step selection scheme while still obtaining both CD and MUD. To elaborate a little further, firstly, the "best" user with the highest direct-link channel quality is selected to transmit, then a "best" relay is chosen to support the transmission. In this way, only the CSI of the \(N\) direct links and the \(2M\) links related to the relays are needed for the user selection and the relay selection, respectively. The existing studies are based on the time-resource allocation (TRA) framework that two time slots (TSs) are allocated for each transmission request. In the first TS the selected user broadcasts its information, and then in the second TS the selected relay forwards its observation. However, considering the two TSs as a whole, the framework is essentially the same as those in the traditional non-cooperative systems. In this framework, using two TSs together to serve one user causes a degradation of spectrum efficiency. Recently, a two-phase TRA framework (TP-TRA) is exploited to improve the spectrum efficiency. In TP-TRA the whole transmission is divided into two phases: the broadcast phase and the relay phase. Firstly, the users broadcast their messages in the broadcast phase, and then the relays assist in transmission in the relay phase. showed that all the users can achieve a diversity order of two by transmitting a network coding (NC) combined packet within one relay TS in single DF-relay aided systems. studied the cooperative schemes in general networks with multiple users and multiple DF relays. More specifically, proposed a Galois field NC based scheme to achieve full CD, and developd a criterion for binary field NC to guarantee full CD. In, the authors showed that the diversity gain of NC based cooperation comes from selection, and based on this revelation, they further proposed a user selection strategy in the relay phase to achieve full CD for both AF and DF networks. Attributing to the TP-TRA framework, these full-CD schemes can improve the spectrum efficiency significantly. However, the problem of jointly exploiting CD and MUD in multiuser multi-relay cooperative networks has not been studied yet under the TP-TRA framework. In this paper, we propose a user selection based low complexity relay protocol (US-LCRP) which is capable of achieving both CD and MUD under the TP-TRA framework. In each broadcast TS, the "best" user with the strongest direct link broadcasts its data block. Then in the relay phase, all the relays serve the transmission in a round-robin fashion. In each relay TS, instead of selecting relay, the destination selects a "worst" data block which most needs to be relayed according to the quality record of each block. Afterwards, a single relay transmits its observation of the selected data block, and the destination performs data combining and quality record updating. To show the effectiveness of the US-LCRP, the diversity-multiplexing tradeoff (DMT) performance is analyzed in this paper. The merit of the US-LCRP is twofold: 1. With the aid of good design, the US-LCRP achieves higher spectrum efficiency while obtaining both CD and MUD. To be more specific, let us suppose the broadcast phase and the relay phase last for \(L\) and \(M\) TSs, respectively. Then, the US-LCRP provides better DMT performance in the scenario of \(L>M\), which indicates that the proposed protocol achieves higher spectrum efficiency than the existing protocols while maintaining the same reliability performance, or it attains higher diversity gain than the existing protocols despite providing the same data rate. 2. The US-LCRP requires the CSI of only the \(N\) direct links for user selection in each broadcast TS. The data block selection in the relay phase is based on the quality record of the previously transmitted data blocks. This quality record may be simply characterized as the SNR of the corresponding signals received at the destination, and can be estimated by using some SNR estimation algorithms. Hence the US-LCRP imposes a significantly lower implementation complexity in practice compared with existing protocols. The rest of this paper is organized as follows. The system model and the proposed US-LCRP are described in Section II. The DMT performance of the US-LCRP is analyzed in Section III, and simulation results are provided in Section IV. Finally, conclusions are drawn in Section V. # System Model and The Proposed US-LCRP ## System Model We consider an AF cooperative network with \(N+M+1\) nodes, where \(N\) users (\(S_{n},1\leq n\leq N\) ) transmit individual information to one destination (\(D\)) with the aid of \(M\) relays (\(R_{m},1\leq m\leq M\)). The system model is shown in Fig. [\[fig:system_model\]](#fig:system_model){reference-type="ref" reference="fig:system_model"}. All the nodes are assumed to have single antenna and transmit with power \(E_{s}\), and operate in half-duplex mode. All the channels in the network are assumed to be independent flat Rayleigh block fading channels with additive white Gaussian noise (AWGN). We further assume that the variances of the channel coefficients of the \(S_{n}\rightarrow D\), \(S_{n}\rightarrow R_{m}\), and \(R_{m}\rightarrow D\) links are \(\gamma_{S_{n}D}\), \(\gamma_{S_{n}R_{m}}\), and \(\gamma_{R_{m}D}\), respectively, while the average noise power of each link in the network is \(N_{0}\). ## The Proposed US-LCRP In and, the relays serve one data block immediately after the data block's direct transmission has been finished. In this TRA, the relays are dedicated to assist one data block in each relay time slot and brings no benefits for the other data blocks. Differently, we exploit the TRA as in where all the relays are shared by all the sources. To achieve this effect, the relays do not participate in assisting the signal transmission until all the direct transmissions have been finished. Therefore, the whole transmission is divided into two phases: the broadcast phase and the relay phase. First, the sources transmit data blocks in the broadcast phase. Afterwards, the relays assist the transmissions in the relay phase. Studies show that with the aid of a well-designed protocol, it is attractive to achieve the effect of "relay sharing", which means that multiple sources are able to benefit from a single relay TS. The design objective of US-LCRP is to achieve both CD and MUD under the TP-TRA framework. Different from, in the proposed US-LCRP, we exploit the TRA framework as in. The whole transmission is divided into two phases: the broadcast phase and the relay phase. First, the sources transmit data blocks in the broadcast phase. Afterwards, the relays assist in transmissions in the broadcast phase. Studies show that with the aid of a well-designed protocol, it is attractive to achieve the effect of "relay sharing" which means that multiple sources are able to benefit from a single relay TS. The design objective of US-LCRP is to achieve both CD and MUD under the TP-TRA framework. We assume that the broadcast phase occupies \(L\) TSs. In each TS of the broadcast phase, the "best" user whose link towards the destination exhibits the highest SNR is selected as a candidate for transmission. Then in the relay phase, the relays assist in transmissions one by one, thus the relay phase lasts for \(M\) TSs. In each relay TS, a single relay aids the transmission of the "worst" data block which has the lowest quality record at \(D\). Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}(a) illustrates TP-TRA, and its details are presented as follows. ### Broadcast Phase In the broadcast phase, a greedy scheduler is employed to obtain MUD. In the \(l\)th broadcast TS, the scheduler chooses the "best" user \(S_{i_{l}}\) whose link towards the destination has the highest SNR. Then, \(i_{l}\) can be expressed as \[i_{l}=arg\underset{n=1,\ldots,N}{\max}\rho_{n}^{\left(l,\mathrm{BP}\right)},\label{eq:UserSelection}\] where "BP" is the abbreviation of "broadcast phase", \(\rho_{n}^{\left(l,\mathrm{BP}\right)}=\frac{E_{s}\left|h_{S_{n}D}^{\left(l,\mathrm{BP}\right)}\right|^{2}}{N_{0}}\) represents the instantaneous SNR of the link \(S_{n}\rightarrow D\) in the \(l\)th broadcast TS, and \(h_{S_{n}D}^{\left(l,\mathrm{BP}\right)}\) denotes the channel coefficient of this link in the \(l\)th broadcast TS. Due to the broadcast nature of wireless environment, all the relays and the destination can receive \(S_{i_{l}}\)'s signal. The received signal at \(D\) and \(R_{m}\) are \[\begin{aligned} y_{S_{i_{l}}D} & =h_{S_{i_{l}}D}^{\left(l,\mathrm{BP}\right)}x_{l}+n_{S_{i_{l}}D}^{\left(l,\mathrm{BP}\right)},\\ y_{S_{i_{l}}R_{m}} & =h_{S_{i_{l}}R_{m}}^{\left(l,\mathrm{BP}\right)}x_{l}+n_{S_{i_{l}}R_{m}}^{\left(l,\mathrm{BP}\right)}, \end{aligned}\] respectively, where \(h_{S_{i_{l}}R_{m}}^{\left(l,\mathrm{BP}\right)}\) is the channel coefficient of the link \(S_{i_{l}}\rightarrow R_{m}\) in the \(l\)th broadcast TS, \(x_{l}\) is the transmitted data block of \(S{}_{i_{l}}\) in the \(l\)th broadcast TS, \(n_{S_{i_{l}}D}^{\left(l,\mathrm{BP}\right)}\) and \(n_{S_{i_{l}}R_{m}}^{\left(l,\mathrm{BP}\right)}\) are the AWGN at \(D\) and \(R_{m}\), respectively. ### Relay Phase The relay phase lasts for \(M\) TSs, during which all the relays participate in the transmission one by one, i.e., in a round-robin fashion. In the first relay TS, \(R_{1}\) transmits, and then in the second relay TS, \(R_{2}\) transmits. This procedure goes on until all the relays have assisted the transmission. In addition, the US-LCRP employs data block selection to facilitate the transmission. Briefly speaking, a single relay assists the transmission of the "worst" data block in each relay TS. After relays' transmission, the destination performs data combining and then updates the quality record of the selected data block in order to prepare for the next relay TS. Since selective combining (SC) is capable of providing diversity order with rather low complexity, we focus on SC in this paper. It should be noted that other combining schemes such as maximum ratio combining (MRC) and equal gain combining (EGC) can be readily introduced into the US-LCRP in the same way. The details of relay phase operation are described as follows. We denote \(\rho_{l}^{\left(m\right)}\) as the SNR of the received signals at \(D\) related to \(x_{l}\) after combining and before the \(m\)th relay TS (i.e., the SNR of the combined signals from all the links over which \(x_{l}\) has been transmitted to \(D\) before the \(m\)th relay TS), \(\mathrm{\phi}_{m}\) as the set of \(\rho_{l}^{\left(m\right)}\), where \(1\leq l\leq L\). Suppose \(h_{S_{n}R_{m}}^{\left(m,\mathrm{RP}\right)}\), \(h_{S_{n}D}^{\left(m,\mathrm{RP}\right)}\), and \(h_{R_{m}D}^{\left(m,\mathrm{RP}\right)}\) are the channel fading coefficients of the links \(S_{n}\rightarrow R_{m},\) \(S_{n}\rightarrow D,\) and \(R_{m}\rightarrow D\) in the \(m\)th relay TS, respectively, where "RP" is the abbreviation of "relay phase". For the ease of exposition, the details of the calculation and the updating of \(\rho_{l}^{\left(m\right)}\) will be explained later. In the \(m\)th relay TS, \(D\) first selects the "worst" data block and broadcasts its index \(\theta_{m}\),

{'timestamp': '2013-10-14T02:11:11', 'yymm': '1310', 'arxiv_id': '1310.3248', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3248'}

# Introduction {#sec:Intro} The stellar content of galaxies represents the time integral, up to the epoch of observation, of both their star formation and chemical enrichment histories. This implies that an understanding of stellar populations is essential in the study of the galaxy formation problem. The nature of galaxies' stellar populations may be pursued through multi-band photometric and/or spectroscopic datasets. The latter tend to be more highly valued for their greater number of population tracers (i.e.  absorption line/band strengths or full spectra versus broad-filter fluxes), reduced sensitivity to dust and multiplexed ability to simultaneously constrain stellar ages, metallicities, chemical abundance patterns, dynamics and mass distributions . The present work concerns spectroscopic-based stellar population studies. Stellar population synthesis (SPS) models are the tools which connect observations of stellar systems to their physical parameters. In order to apply them with confidence, the accuracies of their predictions must be subjected to in-depth evaluations first. The standard approach to SPS model evaluations is to verify that the predictions uniquely reproduce, to within some desired tolerance, the benchmark values of various parameters for well-characterized stellar systems obtained by way of independent and trustworthy techniques ). The stellar systems best suited for evaluations of SPS models are the many star clusters found in the Milky Way and its nearby satellite galaxies. The reasons for this are two-fold. First, those clusters are defined by comparitively simple star formation and chemical enrichment histories, making them the closest tangible approximation to the most basic stellar system treated by SPS models. In other words, they best embody the concept of the so-called simple stellar population: a collection of stars born from an instantaneous burst of star formation and having a uniform chemical composition. Second, the relative proximity of these clusters allows their stellar content to be resolved well below their respective main sequence turn-offs. This makes it possible to accurately constrain their ages via isochrone fitting to colour-magnitude diagrams (CMDs) and abundance patterns from the synthesis of high-resolution spectra of individual members. Of all the star cluster systems found in the nearby Universe, the Galactic globular clusters (GGCs) are most valuable for SPS model evaluations, albeit in the regime of old ages and low to solar metallicities, akin to quiescent galaxies, spiral bulges, and extragalactic globular cluster systems. The value of GGCs in this respect is tied to the fact that their stellar contents have been the most extensively studied to date through CMDs and spectral syntheses. Somewhat ironically, it is because of the special attention paid to GGCs that we now know of many instances where they systematically deviate from the textbook definition of a simple stellar population, largely through inhomogeneities in the abundances of several light elements (see @Gr12a for a recent and detailed review). Given the necessity of model evaluations though, the emergent complexity of GGCs is insufficient grounds to void their status as the premier sample for such purposes. Instead, modellers should adjust the aim of their evaluations to reproducing the *mean* abundance patterns of those GGCs known to harbour multiple populations, but progress along these lines is only in its infancy. Amongst the many public SPS models, that of is one of three designed to recover the abundances of light elements for an observed stellar population, in addition to the usual diagnostics of age and metallicity; the other two models are from and. Of these three models, that of S07 stands out as the one whose abundance predictions have been the most rigorously tested thus far over a considerable metallicity range[^1] (\(\Delta\)\[Fe/H\] \(\sim\) 1.2 dex). found that the S07 model reproduces, to within \(\pm\)`<!-- -->`{=html}0.1 dex, the known ages, metallicities, and abundance ratios of the GGCs 47 Tuc, NGC 6441 and NGC 6528, as well as the Galactic open cluster M67. Despite their success, the sample analysed was rather small, totalling five clusters, implying that a more extensive evaluation covering a wider range in GGC properties, e.g.  horizontal branch morphology, is still needed to establish the ultimate robustness of this model. For instance, these authors noted that the calcium abundance recovered for the metal-poor GGC NGC 6121 was \(\sim\)`<!-- -->`{=html}0.3 dex lower than that measured by . This discrepancy led them to caution about the use of the S07 model in the regime \[Fe/H\] \(\lesssim\)-1.0, precisely where the reliability of this model has been poorly validated to this point. In their analysis, used the library of high-quality integrated blue spectra measured by to recover the ages, metallicities and abundance patterns for the four GGCs in their sample. Indeed, the work of S05 was inspired by the need for in-depth evaluations of the accuracies of spectroscopic-based SPS model predictions. As such, the authors selected their targets (41 in all) to be representative of the entire GGC system, spanning a wide range of metallicities, horizontal branch morphologies, concentrations, Galactocentric coordinates, and Galactocentric distances. To this day, the S05 library remains the only one of its kind. In addition to it, SPS model evaluations require a matching database of independent age, metallicity, and abundance pattern estimates for the 41 S05 GGCs against which the model predictions may be compared. Existing compilations of GGC parameters do not satisfy these needs, however, because they either provide metallicity information alone (@Ha96, 2010 edition[^2]; hereafter Ha10) or overlook the abundances of certain light elements (e.g.  C, N) and cover only a fraction of the full S05 sample. These shortcomings have arguably been at the heart of the statistically weak tests of the S07 and models to date[^3]. Stringent validations of such SPS models are thus stymied until a more complete database of independent stellar population information for this sample is assembled. In this paper, we wish to rectify this situation by presenting the most extensive combined compilation yet of available GGC ages, metallicities, and abundance patterns. To do so, we draw on existing compilations of GGC parameters, as well as the vast literature on the chemical compositions of the S05 clusters. The application of our data set to a statistically robust evaluation of the S07 and other SPS models will be presented elsewhere (Roediger et al., in prep.; hereafter Paper II). The layout of the present paper is as follows. In , we discuss the GGC sample, methodology and some of the principal data sources which underlie our compilation. Granted that our sample is representative of the GGC system as a whole, we use our compilation in to draw insights into a variety of topics related to the stellar populations of these objects. Comparisons of our work with other prior compilations of GGC stellar population data are presented in . Finally, we conclude and contemplate other possible uses of this compilation in . # Data & results {#sec:D&R} Spectroscopic-based SPS models are designed to predict the full spectra and/or strengths of absorption line/band indices over a wide range of ages and metallicities for simple stellar populations of any specified abundance pattern. The ability of the S07 model to fit for chemical abundances is realized by inverting its functionality, that is, by perturbing the specified abundance pattern until the same age and metallicity are obtained amongst all possible index-index pairs under consideration for an observed stellar system. Practically speaking, the model steps through the \(n\)-dimensional parameter space spanned by the available data in a hierarchical fashion, beginning with indices most sensitive to age and metallicity effects (e.g.  H\(\beta\), Fe5250, Fe5335) and ending with those that trace chemical abundances (e.g.  Mg*b*, Ca4227). In this way, the model simultaneously predicts the best-fit age, metallicity and light-element abundance pattern (\[Mg/Fe\], \[C/Fe\], \[N/Fe\], \[Ca/Fe\]) for a given system. While the S07 model can, in principle, fit for the ratios \[O/Fe\], \[Na/Fe\], \[Si/Fe\], \[Cr/Fe\], and \[Ti/Fe\] as well, their values are fixed at this time[^4] since they are not reliably traced by existing Lick indices. presented an efficient algorithm, "EZ_Ages", to carry out the required inversion of the S07 model so that it can be applied to the measured indices of any stellar system. Further details on either the S07 model or EZ_Ages are provided in those introductory papers, as well as in Paper II. For a robust evaluation of the S07 and other SPS models (Paper II), we will draw upon the S05 library of intermediate-dispersion, high-\(S/N\) integrated blue spectra for 41 GGCs, as do most other investigators for such purposes. An important aspect of this library is the sample's wide coverage of the known GGC parameter space (see and ), which makes it fairly representative of the entire GGC system as well as the stellar populations of whole galaxies (e.g.  early-types) or their sub-components (e.g.  bulges). Either a suite of absorption-line/band strengths measured from these data or the full spectra themselves may be fitted using one of several different SPS models to recover the ages, metallicities, and abundance patterns of these GGCs. The performance of any given model is then judged by comparing these fitted parameters against the most complete compilation yet of similar but independently-derived information for the S05 GGCs, the latter of which is the primary focus of this work. Although the preceding discussion has largely focussed on the S07 model, it must be appreciated that our compilation is perfectly general and can be applied to the evaluation of *any* spectroscopic-based SPS model. In fact, such an undertaking would undoubtedly help highlight the merits of the particular ingredients and/or fitting techniques adopted by different models. Our compilation may also be useful to any other field concerned with GGCs, such as the formation of the Milky Way, or stellar evolution and nucleosynthesis. In the following sub-section, we describe the methodology by which our compilation was assembled given the wealth of literature data pertaining to our needs. For reasons that will be made clear, this discussion will largely revolve around the abundance patterns of our clusters. ## Compilation methodology {#sec:D&R-CompMeth} The over-arching principle for our compilation of the available literature on the stellar populations of the S05 GGCs is that it be as comprehensive and complete of a record as possible. For each one of our clusters then, we have strived to obtain as many estimates as we could for its age, metallicity, and abundances of light elements, with the extent of the latter group being dictated by those elements currently treated within SPS models (i.e.  Mg, C, N, Ca, O, Na, Si, Cr, Ti)[^5]. While extensive and homogeneous compilations already exist with respect to GGC ages and metallicities, and cover large fractions of the S05 sample (e.g.  MF09, Ha10; see ), resources of similar quality on the individual chemistries of these clusters is more limited and heterogeneous. The premium compilation of GGC chemical compositions until now was presented by Pr05, who gathered \(\alpha\)-element abundances (Mg, Ca, Si, Ti) from high-resolution spectroscopic analyses in the literature for a sample of 45 GGCs. Their results prove less than ideal for the purposes of SPS model evaluations since the elements C, N, O, Na, and Cr, which SPS models now cover, were omitted and their sample has a rather small overlap with that of S05 (15 clusters only). To improve upon the shortcomings of Pr05, we extracted from the literature measurements of the relevant chemical abundance ratios for each S05 cluster from as many references as possible. In so doing, we have found that specifying a complete abundance pattern for any one S05 GGC often required data from at least two references; for example, the abundances of carbon and nitrogen are usually studied together but separately from those of the other elements listed earlier. Our desire for completeness therefore naturally encouraged us to draw on multiple sources when assembling the abundance patterns we are advocating for use here. In doing so, we have combined the results from *all* chemical composition studies on each cluster[^6]. The abundance pattern we adopt as the benchmark for each cluster was built by calculating the mean value and root-mean-square (rms) dispersion of the available independent measurements for each of the elements listed above. This aspect of our compilation embodies some noteworthy advantages. First, merging results as we do should reduce the *statistical* error in the value that we recommend for any given abundance ratio, albeit at the price of increasing its *systematic* error. We do not consider this a drawback but rather another advantage of our approach since systematic errors (e.g.  sample selection, solar abundances, atomic parameters) might be a significant source of discrepancy between abundance patterns predicted by SPS models and star-by-star spectral syntheses. Having a metric for the degree of systematic error involved in the latter, via the dispersions, will undoubtedly be very helpful for judging the reality of model predictions. The last advantage of our approach is that it should also naturally reflect the existence of putative multiple stellar populations when present within a given cluster (again, via the dispersions). A hallmark of the multiple population phenomenon is that, amongst the members of an affected cluster, the abundances of several elements either correlate (Al-Si) or anti-correlate (C-N, Na-O, Mg-Al) with one another. Modulo the particulars on sample selection, we then expect to find large spreads in the abundance ratios of these elements between the stars from either a single study or multiple ones[^7]. Moreover, by combining such scattered measurements into a single estimate for a cluster's abundance pattern, we can be assured that the corresponding dispersions reflect the presence of multiple populations by being comparitively large to those of species which are excluded from the above trends (e.g.  Ca). Note that any and all claims we make herein as to the causes of inflated rms dispersions (re: systematics versus multiple populations) are ultimately suggestive and not based on thorough quantitative analyses. Another major principle for our compilation involves concentrating, where possible, on studies whose results pertain solely to evolutionary stages from the main sequence (MS) through the asymptotic giant branch (AGB). We impose such a restriction because the onset of thermal flashes, third dredge-ups, dust-gas separation (winnowing) and mass loss during the final (post-AGB) stage in the evolution of low-mass stars can give rise to surface abundances which poorly reflect the chemical composition of the gas from which they formed. Third dredge-up episodes, in particular, would pollute the surfaces of such highly-evolved stars with CNO-processed material from the hydrogen-burning shell (if present), effectively lowering the abundance of carbon and raising that of nitrogen there, relative to those of MS stars. Stars ascending the sub-giant and red giant branches (SGB and RGB, respectively) can also have their surface chemistries of C, N, and O affected by mixing episodes (e.g.  @Ib64, @SM79; ), but given that most spectroscopic analyses of individual GGC members do not penetrate to fainter magnitudes than this phase precludes our rejection of such data. In other words, limiting our mean carbon and nitrogen abundances only to measurements obtained from MS stars would significantly thin out our compilation. Instead, we embrace such data and simply caution SPS modellers to consider the evolutionary stage down to which their predicted \[C/Fe\] and \[N/Fe\] values correspond. In we highlight the possibility that carbon depletion and nitrogen enhancement as a function of position along the SGB/RGB may be crudely quantified. In addition to highly-evolved stars, we also exclude from our compilation (again, where possible) data corresponding to "exotic" stages of stellar evolution, such as very hot (\(T_{eff} >\) 11 500 K) horizontal branch (HB) stars. In this case, the surface abundances of elements are often perturbed by effects like radiative levitation and gravitational sedimentation. Unlike the case of mixing along the SGB/RGB though, it is unclear that empirical corrections for these processes are forthcoming simply because it is rare to find individual GGCs with mixtures of exotic plus "normal" (\(T_{eff} <\) 11 500 K) HB stars, let alone homogeneous abundance analyses thereof. Thus, unless data from the MS, SGB, RGB, (cooler) HB, and AGB for a cluster are all not available, we deem abundance ratios based on the most advanced and exotic stages of stellar evolution unsuitable for our purposes and omit them from our compilation. In light of the above caveats, we wish to provide the exact rationale behind our compiled abundance pattern for each S05 cluster. We do just this, in brief and on a per cluster basis, in the Appendix, with attention being paid to the following related themes: adopted references, omitted data, systematic errors, and evidence of multiple populations from both our data and other methods (where applicable). also provide the relevant references from which our recommended ages, metallicities, and abundance patterns for the S05 GGCs were drawn. In the following section, we specifically address our sources and methodology used to arrive at the ages and metallicities of the S05 clusters. Much of that information will therefore not be repeated in the Appendix. ## Age and metallicity sources {#sec:D&R-A&ZSources} Our selection of sources for age and metallicity information on the S05 GGCs embraces similar principles as described above regarding their abundance patterns. In terms of their ages, a cursory review of the relevant literature reminds us of genuine discrepancies on a per cluster basis. While isochrone fitting of one form or another to CMDs has long been the standard by which GGC ages are obtained, the results therefrom appear to be plagued by rather large uncertainties. The origins of these discrepancies are most likely tied to the values of various parameters assumed by the scientist (i.e.  distance, reddening, metallicity, etc.) and/or each isochrone set (i.e.  mixing length, helium abundance, etc.). Since little is to be practically learned by merging together the available absolute age determinations for any given cluster (unlike the case with their chemical abundances), we prefer our compiled values of this parameter to come from a single source. The majority of the ages adopted in our compilation come from MF09. These authors have performed the most extensive and homogeneous age analysis of GGCs to date, totalling 64 clusters in all and using HST/ACS photometry plus many flavours of isochrones. The ages in this work from the Dotter et al.  isochrones were cast in terms of both the and metallicity scales. By normalizing their results from each isochrone set to the corresponding mean absolute age of their 13 lowest-metallicity GGCs, MF09 found that the *relative* ages were robust to the particular choice of isochrone (see their Table 4) and carried a formal precision between 2-7%. For the 25 S05 GGCs which overlap with the MF09 sample, we adopt their relative ages based on the Dotter et al.  isochrones and Carretta & Gratton scale. From their §6.1, we assume a normalization factor of 12.80 \(\pm\) 0.17 Gyr to put these ages on an absolute footing. Note that the uncertainty in the normalizing factor does not account for systematics, e.g.  bolometric corrections, but we anticipate this issue will be thoroughly addressed in forthcoming work on absolute GGC ages by the same group, as alluded to in MF09. To bolster the reliability of their *relative* ages, MF09 also compared them against those of, the formerly largest homogeneous GGC age compilation, totalling 55 clusters in all. In doing so, they found mutual consistency between the two datasets to within their own published error bars, where De Angeli et al.'s HST and ground-based sub-samples yielded mean residuals of-0.04 \(\pm\) 0.07 and-0.02 \(\pm\) 0.08, respectively. Furthermore, MF09 failed to detect any trends in the residuals as a function of metallicity. We conduct a similar comparison in , but in terms of *absolute* ages and with respect to multiple prior age compilations, where the relative ages from Rosenberg et al.  were transformed assuming a zeropoint of 13.2 Gyr (see their §4). To ensure the comparison is fair, we limited it to the twelve S05 GGCs common to all four compilations examined therein. The mutual overlap between the samples of MF09 and other age compilations in the literature is actually poorer than this and so we have omitted those results from . Instead, we compare in the Appendix (wherever possible) our adopted ages against *all* other estimates in the literature known to us, on a per cluster basis. Such an exercise provides us with a first-order impression of the systematic discrepancies involved between any two individual age determinations. There, we also provide the normalization factors or constants we have used to transform those ages originally expressed on relative scales into absolute terms. The most striking feature from is the presence of several clusters in each panel whose ages from prior compilations disagree egregiously with those of MF09. These disagreements are found in different age and metallicity regimes[^8], depending on the source under consideration: young ages and low metallicities for, old ages and high metallicities for, and at both age and metallicity extrema for . Of course, the identification of these inconsistencies hinges on how representative the published uncertainties are of the total error budget. The error bars shown in are largely statistical in nature and do not consider the systematic effects of uncertainties in, amongst many others, distance modulus, reddening and stellar evolution model. With exception to the results of Dotter et al., this criticism may be unwarranted or overstated since the investigators employed distance-and reddening-independent methods (Rosenberg et al.; Salaris & Weiss) and/or provided *relative* ages only (Rosenberg et al.; MF09). Overall then, leaves us with the impression that systematic uncertainties in absolute age dating of GGCs resides (at worst) at the 2-3 Gyr level. Neglecting ages which exceed that of the Universe, this estimate is corroborated by our cluster-by-cluster comparisons of individual age determinations in the Appendix. A more detailed examination than that of this issue lies beyond the scope of this paper. While the existence of significant systematics in absolute age determinations might suggest that, for now, it be best to evaluate age predictions from SPS models in a relative sense, it is not clear that this can be done in practice. One complication is that the S05 sample does not contain the same clusters upon which MF09 base their normalization. Another is that, for the sake of completeness of our compilation, we appeal to five other sources of age information for 13 of the 16 S05 GGCs which do not overlap with the MF09 sample. Based on , it is clear that transformations of these results onto the MF09 scale would at best be crudely defined. Therefore, until a more complete source of relative ages for the S05 GGCs becomes available, SPS modellers will have to bear in mind the systematics underpinning the absolute ages against which they test their predictions. To assist in this awareness, we explicitly caution the reader in the Appendix about those clusters in our sample for which their adopted age *does not* come from MF09. Our extensive list of references which provide independent age measurements for the S05 GGCs is summarized globally in and for individual clusters in . Note that the specific reference of our adopted age for each cluster is shown in boldface in the latter. Our knowledge of the metallicities of the S05 GGCs has greatly improved with the work of. Based on high-resolution optical spectra for about 2000 RGB stars belonging to 19 GGCs (13 of which are in the S05 sample), these authors have created the premier database of homogeneous and spectroscopic metallicities for GGCs. Through it, they have defined a new GGC metallicity scale spanning almost the full range of values exhibited by these systems in this parameter space, from \[Fe/H\] \(\sim\)-2.4 to-0.3 dex. Since Carretta et al.'s sample overlaps with those from previous metallicity scales by ten clusters or more, they were also able to derive transformations between those and their own. This enabled them to express the metallicities of all 133 GGCs from the 2003 version of the catalogue, which encompasses our whole sample, in terms of their own scale. Ha10 improves upon the work of by merging the latter's results with those of, after first transforming them to the Carretta et al.  scale, as well as metallicities for individual clusters from other studies. Given its complete coverage of the S05 sample and the sheer popularity of this database, we would ideally adopt the metallicities from Ha10 for our compilation. However, since Ha10 do not provide uncertainties on their values, we instead calculate the mean metallicities of our clusters using the same references and weighting scheme as Ha10. The bottom panel of shows the differences between the Ha10 metallicities and our replicas thereof. Considering the complete S05 sample, the agreement between the two datasets is superb, with 68% of the data points exhibiting differences of \(\pm\)`<!-- -->`{=html}0.01 dex or less. We also find that our adopted metallicities compare favorably with those from Carretta et al.  (, *top*), albeit with 68% of the data points showing differences up to \(\pm\)`<!-- -->`{=html}0.05 dex. This agreement is not all that surprising as the Ha10 metallicities, and thus our replicas as well, are weighted to the data of Carretta et al.  by a factor of three more than those from any other source. ## The stellar populations of Galactic globular clusters {#sec:D&R-SPGGC} The results of our literature compilation on the stellar population properties of the S05 GGCs are presented in and . In , we list the recommended age (column 2), mean metallicity (column 3), and mean Mg, C, N, and Ca abundances (columns 4-7) for each cluster, while gives their mean O, Na, Si, Cr, and Ti abundances (columns 2-6). Entries therein which we consider suspect with respect to our compilation principles (elucidated above) appear in boldface; the reader is referred to the Appendix for the rationale behind each of these flags. Recall that, because of practical limitations which bar the computation of relative ages from both literature data and SPS models for *all* of the S05 GGCs, we have cast all of our compiled ages into absolute terms. In light of the apparent systematics which afflict absolute age estimates, we look forward to future work from expert groups which properly address this issue. Until then, SPS modellers will simply have to be mindful of these uncertainties in our compiled data, which we optimistically gauge to be \(\sim\)`<!-- -->`{=html}2-3 Gyr, when evaluating age predictions. While the original intent of the present compilation was for evaluations of SPS models, we foresee its broad applicability to a variety of astrophysical endeavours since the S05 sample is representative of the whole GGC system (; ). To demonstrate this point, we use our compilation in the following sub-sections to garner some brief insight into the early formation and chemical evolution of the Milky Way, atmospheric mixing during stellar evolution and the sites of explosive stellar nucleosynthesis. ### Ages & metallicities {#sec:D&R-A&M} In we show the distributions of all the stellar population diagnostics presented in for the whole S05 sample. Referring to and the MF09 results shown therewithin (black histogram), it is seen that the S05 sample has an age distribution which is both strongly-peaked (between 12.5 and 13 Gyr) and skewed to very old ages. Those clusters whose ages come from other sources in the literature are represented by the gray histogram. These additional estimates tend to broaden the overall distribution to both younger and older ages, as well as the strong peak described by the MF09 results. The fact that over half of these other age determinations are found to have extreme values in relation to those from MF09 bolsters our position on treating them with caution. Note that the shaded region in demarcates ages which exceed that of the Universe and while some of our clusters are found there, the statistical errors on their values alone overlap with the allowed (unshaded) region. From the MF09 ages in , it is tempting to conclude that the S05 GGCs originated from a two-component star formation history. This history could be described as consisting of either a sharp burst superimposed upon a comparitively steady background (lasting \(\gtrsim\)`<!-- -->`{=html}4 Gyr) or a vigourous early episode which quickly peaked and was then regulated down to a more sustainable level. By broadening the overall distribution to more extreme ages, the additional literature data shown in seems to agree better with the first of these two scenarios. Modulo systematics, these data would temper this scenario though by spreading the burst component over a longer timescale (\(\sim\)`<!-- -->`{=html}1.0-1.5 Gyr). The metallicity distribution for the S05 GGCs () also appears to support the idea that these clusters arose from at least two distinct channels given its clear bimodal shape (in agreement with @Zi85), with a peak-to-peak separation of about \(\sim\)`<!-- -->`{=html}1.0 dex. When we examine the available MF09 ages of the clusters comprising each metallicity sub-group however (), there does not appear to be a strong correlation between the two parameters. The metal-poor and metal-rich GGCs of our sample, separated at \[Fe/H\] =-1.0 dex, which overlap with MF09 have a mean age and rms dispersion of 12.0 \(\pm\) 1.1 Gyr and 12.8 \(\pm\) 0.7 Gyr, respectively; the distinction of these two groups by age only worsens when we consider all of our adopted values. The situation seen in can be attributed to the presence of several old GGCs (\(>\)`<!-- -->`{=html}12 Gyr) in our metal-poor sub-sample, whereas only one of our metal-rich GGCs has an age of \(<\)`<!-- -->`{=html}12 Gyr. Moreover, our metal-rich sub-sample harbours a high incidence of very old clusters in that three (four, if the metal-rich/-poor boundary lay slightly lower; e.g.  \[Fe/H\] =-1.05) of our six oldest objects are contained therewithin. Since our sample is representative of the entire GGC system, these findings bear some implications with regards to the formation of the Milky Way, in particular its halo. For instance, the parameter spread in would be hard to explain within a scenario in which *all* of the S05 GGCs formed in situ since one would expect the metal-poor GGCs to be *older* than the metal-rich ones, not younger. Instead, this spread is consistent with a picture in which the GGC system arose from its members either forming in situ or being accreted from satellite galaxies. Although the possible correlation of these two channels with metallicity remains unclear, we interpret the older, metal-rich and younger, metal-poor clusters as the descendants of the former and latter, respectively. Similar conclusions have been reached by other analyses of the GGC age-metallicity relation using much larger samples. Further insight into this topic might be achieved by searching for correlations between the above parameters and those from the Ha10 catalogue, particularly velocity information, but such a task is beyond the scope of this work. and MF09 already investigated the relationship between age and galactocentric radius for GGCs, and found none. ### Chemical abundance distributions & atmospheric mixing {#sec:D&R-CAD&AM} Further critical insight into the stellar populations that comprise GGCs can be gleaned from the distributions of their mean chemical abundance ratios, as shown in . These, in turn, afford us with further information on the formation of the Milky Way, as well as certain aspects related to stellar evolution. From , we first note that the distributions of the mean abundances of the \(\alpha\)-elements (Mg, Ca, Si, Ti) amongst these systems all show sharp peaks towards super-solar values. The respective median values for the \[Mg/Fe\], \[Ca/Fe\], \[Si/Fe\] and \[Ti/Fe\] distributions are +0.38, +0.30, +0.36 and +0.24 dex, which implies that the GGC system, on the whole, formed over rapid timescales. On the other hand, slightly broader distributions are found for mean abundances of carbon, nitrogen, and oxygen amongst our GGCs, in that we obtain rms dispersions of \(\gtrsim\)`<!-- -->`{=html}0.15 dex for them compared to \(\lesssim\)`<!-- -->`{=html}0.1 dex for the \(\alpha\)-elements. Apart from possible undersampling effects[^9], we interpret the broader distributions of these elements as reflecting the combined and well-known phenomena of atmospheric mixing and multiple stellar populations within these clusters. We concentrate on the former for the remainder of this sub-section and take up the latter in the next. To bolster the above argument, we show with coloured points in the mean abundances and luminosities of stars from each individual study of the carbon and nitrogen abundances in the S05 GGCs. Unfortunately, not all studies we cite in this regard could be represented in these plots since luminosity information is not available to us in many cases. This may be responsible for the apparent gap in the data in the range-0.5 \(< M_V <\) +0.5 mag. The point types in reflect whether the assorted samples consist of CN-weak (triangles) or CN-strong (diamonds) stars; circles are used when CN strengths are unknown to us. Moreover, looking from bright to faint luminosities, the evolutionary status of the sample changes from predominantly RGB stars (leftmost-to-middle) to SGB/MS stars (rightmost). Concentrating only on the mean values for CN-weak and CN-strong stars, an evolutionary trend is apparent from whereby \[C/Fe\] tends to monotonically decrease as a given star leaves the MS and ascends the RGB. Over the same evolutionary path, \[N/Fe\] for the star will decrease to a minimum at about \(M_V \sim\) +1 mag and rise thereafter. These same trends are also conveyed by the coloured points in , where we circumvent the need for luminosity information and thus benefit from better statistics. The significant scatter amongst the data points representing both individual stars and mean values for unclassified samples may be due to mundane issues like systematics or more nuanced ones like mixtures of CN-weak and CN-strong stars within any one study. An example of each case might be the solar-like \[C/Fe\] value of upper-RGB stars in NGC 6121 at \(M_V \sim\)-2 mag for the former and the \[N/Fe\] value of lower-RGB stars in NGC 7078 at \(M_V \sim\) +2 mag for the latter. The mixture interpretation is supported by analyses of *single* clusters that have found large spreads in \[C/Fe\] and \[N/Fe\] of stars at any common evolutionary phase from the MS through the tip of the RGB (e.g.  NGC 0104, 6205, 6254, 6397, 6752, 7006; @Ca05, @SmG05, and references therein). The evolutionary trends seen in are commonly associated with a combination of the first dredge-up followed by deep atmospheric mixing that occur in the atmospheres of low-mass stars after they complete core-hydrogen burning. The first dredge-up is defined by the mixing of partially-processed material from the stellar interior into the outer atmosphere as the convective envelope grows in size during the star's SGB phase. It is responsible for the gentle decline observed in both \[C/Fe\] and \[N/Fe\] from the MS to about the midpoint of the RGB (\(M_V \sim\) +1 mag). Once the star reaches the RGB bump, deep mixing is thought to set in and create the rapid rate of depletion and (now) enhancement of atmospheric carbon and nitrogen, respectively. Deep mixing accomplishes this by bringing CN(O)-processed material from the hydrogen-burning shell into the outer atmosphere once the shell overcomes the molecular weight barrier left by the first dredge-up and expands into the outer convective envelope. Both the luminosity of the bump (i.e.  onset of mixing) and the depletion rate of carbon are known to decrease as a function of metallicity. These two dependencies may help explain some of the scatter seen in at high luminosities (\(M_V \gtrsim\) 0 mag). In light of the fact that evolved stars undergo episodes of atmospheric mixing, it seems conceivable that the distributions of \[C/Fe\] and \[N/Fe\] in our compilation would be somewhat broader than those of unaffected species, as seen in . The reason for this is that spectroscopic studies of resolved GGC members have historically measured \[C/Fe\] and \[N/Fe\] from a variety of evolutionary stages. Were measurements of these ratios for MS stars to become available for our entire sample, we would anticipate a tightening of the corresponding distributions. Until that time comes, the reality of these mixing episodes should compel SPS modellers to carefully consider the luminosity biases of published spectroscopic studies of individual GGC members when evaluating the accuracies of their \[C/Fe\] and \[N/Fe\] predictions[^10]. ### Multiple populations {#sec:D&R-MP} In addition to mixing phenomena, also shows that some intrinsic degree of broadening in the \[C/Fe\] and (especially) \[N/Fe\] distributions for the S05 GGCs is to be expected on account of the multiple populations found within many of them. From those plots, it is seen that the dichotomy in CN band strength (weak/strong) extends down to the MS, the CN band is much more sensitive to \[N/Fe\] than \[C/Fe\], and CN-weak stars are characterized by lower \[C/Fe\] and higher \[N/Fe\] values than CN-strong stars. The union of these facts then creates the potential for the mean values of \[C/Fe\] and \[N/Fe\] from any given study to be biased either low or high depending on how accurately the sample embodies the true CN distribution of the associated cluster. In fact, we might be able to infer the as yet unknown CN strengths of certain samples based on the relative proximity of circlular points to the triangular or diamond points in . Another hallmark of the multiple population phenomenon is the anti-correlation of \[O/Fe\] and \[Na/Fe\] ratios exhibited by stars from all major evolutionary stages within affected clusters. The existence of this anti-correlation would thus tend to bias estimates of a given cluster's mean abundances of these species if not accounted for during sample selection. We suspect that this effect may be at least a contributing factor to the relatively broader distribution for \[O/Fe\] seen in , compared to those of abundance ratios that are not known to vary from star-to-star. This suspicion could be tested by investigating whether the breadth of our \[O/Fe\] and \[Na/Fe\] distributions is *simultaneously* consistent with the observed ranges in these abundance ratios for individual stars from the large, homogeneous Na-O anti-correlation study of. This is beyond the scope of the present work however. It is worth mentioning at this juncture that the exact origins of the multiple populations observed in GGCs remains unknown. The perpetuation of chemical abundance variations down to unevolved, MS stars makes strongly certain that the existence of a second (and sometimes third; @Ca09a) generation of stars is tied to an external agent. However, at least three candidates could be responsible for the pattern of enhanced nitrogen, sodium, aluminum, and (possibly) helium abundances plus depleted carbon, oxygen, and magnesium abundances that typifies the younger generations: (i) massive AGB stars, (ii) massive rotating MS stars and (iii) as in (i) but for intermediate masses. One way to help distinguish between these candidates is to study whether the sum of the CNO elements varies between the populations in each affected cluster. Simply put, massive stars are expected to alter the individual abundances of these elements but leave their sum unchanged, while intermediate-mass AGB stars, by way of the third dredge-up, will not. Evidence thus far of variations in the CNO sum within individual clusters has been contentious. With some work, our extensive compilation of carbon, nitrogen, and oxygen abundances for individual GGC stars may be helpful in shedding further light on this issue, but will not be investigated further here. ### Source(s) of \(\alpha\)-elements {#sec:D&R-SAE} While point to the existence of atmospheric mixing episodes within the evolved stars of the S05 GGCs, shows that the mean abundances of \(\alpha\)-elements amongst these clusters remain more or less homogeneous. This homogeneity is accounted for within the context of mixing by the fact that \(\alpha\)-elements are exempt from the CN(O) cycle, such that their abundances likely reflect the chemistry of the gas from which these GGCs were born. Moreover, that the \(\alpha\)-element abundance ratios for the S05 GGCs are all greater than the solar value by factors of 1.7-2.4 implies that these systems must have formed quite rapidly, on timescales less than that of Type Ia supernovae (\(\sim\)`<!-- -->`{=html}1 Gyr). An interesting corollary on the chemical enrichment of GGCs, based on their \(\alpha\)-element abundances, is presented in . The left-hand panel shows \[Ca/Fe\] versus \[Mg/Fe\] for individual stars from our compilation and belonging to the 13 S05 clusters having the most measurements in this regard. These data are clearly uncorrelated and scatter about their centroid at (\[Mg/Fe\],\[Ca/Fe\]) \(\sim\) (+0.4,+0.3) with rms dispersions (0.13-0.14 dex) comparable to the median errors in the individual stellar abundances (0.12-0.14 dex). These evidences are enough to suggest that the production sites of these two chemical species are not one and the same. In the right-hand panel of , we show with open squares the mean magnesium and calcium abundances from Pr05 for the nine clusters in the left-hand panel that fall within their sample. Excluding the single outlier at (\[Mg/Fe\],\[Ca/Fe\]) \(\sim\) (+0.10,+0.24), we find the rather surprising result that these abundances are *anti*-correlated for this sub-sample and data. Not surprisingly, the corresponding values from our compilation (filled circles) exhibit no such correlation. This discrepant behaviour between our results and those of Pr05 could arise from the different approaches taken with respect to the following issues: (i) scope of input data, (ii) averaging method, and (iii) systematics. The latter refers to Pr05's attempt to homogenize all of their input data to the same log*gf* and solar abundance scale, something we neglect to do. To appreciate the potential role of systematics, we also show in this panel, with open circles, the mean abundances we derive based on the same references and averaging method used by Pr05, but without log*gf* and solar abundance corrections. For each cluster, we connect with a line the abundances from the three distinct methods. Comparing the positions of open squares and circles, it is clear that the choice of atomic constants and solar abundance pattern often plays a significant role in setting the values of \[X/Fe\] for any element X, by as much as \(\pm\)`<!-- -->`{=html}0.2 dex. On the other hand, the offsets between open and filled circles may be regarded as the effect of our using more references per cluster and straight averages, as well as neglecting systematic corrections. Our methodology clearly affects our adopted abundance patterns as well, but this should not reflect any fundamental flaws in our results. The prospect of anti-correlated magnesium and calcium abundances amongst the GGC population supports the inference that the production sites of these two species are not one and the same or, even more intriguing, that the yields of Type II supernovae fluctuate on an element-by-element level. Note that such a trend is also found amongst the points in the right-hand panel of which represent the results of our attempt to mimic the approach of Pr05 (open circles), again modulo a single outlier. When the full sample of either compilation is considered, however, this anti-correlation changes to a weak positive correlation. The fact that neither sample is complete though implies that further investigation of the ratios \[Mg/Fe\] and \[Ca/Fe\] amongst individual GGC stars on a larger, more homogeneous basis may be warranted. The suggestion that the abundances of magnesium and carbon do not track one another within any given stellar population is not new. For instance, several studies of the central stellar populations of early-type galaxies have concluded that \[Mg/Fe\] increases modestly with velocity dispersion amongst these systems, but that \[Ca/Fe\] is uniform, at about the solar value. casted doubt on the authenticity of these results by pointing out that the Ca4227 index, a popular tracer of calcium abundance, has its blue pseudocontinuum contaminated by a CN band. and have shown, however, that the \[Ca/Fe\] ratio remains uniform amongst red sequence galaxies even when the abundances of carbon and nitrogen are properly accounted for *a priori*. In the Milky Way, found that while \[Ca/Fe\] decreases with increasing \[Fe/H\] for RGB stars in the bulge, their \[Mg/Fe\] ratios remain more or less uniform at \(\sim\) +0.3 dex. Earlier claims to this same effect were made by, , and. These discrepant trends may only apply to the Galactic bulge though, since the calcium abundances of both metal-poor field and thick disk stars behave in a fashion consistent with that exhibited by the other \(\alpha\) elements. The sum of the above discrepancies results in a confusing picture, to say the least, of how stellar systems enrich themselves in the \(\alpha\)-elements. One possible solution is that the abundance pattern of a system, and thus the source(s) of its chemical evolution, depends sensitively on the intensity of the star formation from which it was created. For instance, the spheroidal systems which seem to exhibit genuine differences in the behaviours of their Mg and Ca abundances (i.e.  early-type galaxies, Milky Way bulge, and GGCs) likely formed most of their stellar mass over rapid timescales. In other systems, the star formation history could very well have been more protracted, leading to potentially different sources of chemical enrichment. A most useful test of this proposed solution would be to see if and how chemodynamical simulations could reproduce the precise pattern of \(\alpha\)-element abundances we find in our GGCs. ## Comparison with previous work {#sec:D&R-CPW} Having presented our compilation and advertised some of the immediate science that can be gleaned from it, it would also be prudent to assess the robustness of our results by comparing them to those from previous similar work. The comparison of our adopted ages and metallicities against other sources of such information has already been performed elsewhere (; Appendix) so that we need only focus here on the chemical abundance patterns of our clusters. As stated before, the most extensive survey of the literature on GGC chemical abundance patterns prior to our compilation was made by Pr05. We already compared their results against ours in terms of the ratios \[Mg/Fe\] and \[Ca/Fe\] in the right-hand panel of . In we expand on this comparison by plotting our estimates versus theirs of the metallicities and \(\alpha\)-element abundances (Mg, Ca, Si, Ti) for the 15 GGCs common to both samples. Despite several outliers, a good correspondence clearly exists between the Pr05 metallicities and our own. Conversely, the large scatter and low Pearson coefficients (shown at top-left) exhibited by the other sets of points in means we cannot reach the same conclusion regarding the abundances of \(\alpha\)-elements. Specifically, our results compare least favourably to those of Pr05 in terms of \[Ca/Fe\]. By inspection of Pr05's methodology, we find that the most egregious discrepancies between the two sets of abundance ratios can be explained by two effects. These are: (i) our inclusion of references that post-dated their work, and (ii) corrections that Pr05 implemented to standardize literature data to a common log *gf* system and solar abundance pattern. The latter corrections can often be significant in size (\(\sim\)`<!-- -->`{=html}0.2 dex), a point which was already hinted at in . As Pr05 point out though, abundance analyses are often published without specifying the assumed log *gf* values and solar abundances, making it difficult (if not impossible) to gauge what those corrections should be. We therefore abstain from attempting such corrections ourselves and instead embrace the fact that our adopted abundance patterns likely suffer from the full effect of the systematics which underpin spectral syntheses. # Conclusions {#sec:Concs} Drawing on a wealth of literature data up to mid-2012, we have assembled a new compilation of the known ages, metallicities, and chemical abundance patterns for the 41 GGCs studied by S05. This extensive compilation represents a singular expansion upon similar but more limited previous work on the stellar populations of these systems. We anticipate that it will prove to be a key ingredient for stringent evaluations of the absolute reality and robustness of predictions from the latest suite of SPS models designed to recover the above information for unresolved systems (Roediger et al., in prep.). Given the wide range of parameter space spanned by the S05 sample, our compilation should also benefit a wide range of other astrophysical interests. The age distribution for these clusters suggests that they arose from a star formation history that consisted of a strong peak (12.5-13.0 Gyr ago) superimposed upon a relatively smooth background. Combining this information with their metallicities and \(\alpha\)-element abundances, it appears that each GGC was formed rapidly either in situ or in a satellite galaxy and subsequently accreted onto the Milky Way. Furthermore, with our compiled abundance patterns we confirm previous claims that (i) the surface abundances of C and N in evolved stars are altered by mixing episodes as they ascend the SGB/RGB, (ii) many GGCs host at least two distinct stellar populations, and (iii) the enrichment of \(\alpha\)-elements in these systems, specifically Mg and Ca, likely occurred through multiple channels. The fact that the mean chemical abundance patterns of GGCs are sensitive to the first two phenomena are important caveats that must be considered during SPS model evaluations. Similarly, we estimate that absolute age determinations for GGCs are subject to systematic uncertainties on the order of \(\sim\)`<!-- -->`{=html}2-3 Gyr. While the above results are certainly of some value, it is our opinion that many other applications of our compilation have yet to be explored. To enable the community to further pursue such ancillary science or modify the results of our compilation as they see fit, we provide electronic tables of the input data online[^11] for each one of our clusters. We thank Bill Harris and Charlie Conroy for insightful discussions and detailed comments on an earlier version of this paper which led to valuable improvements. J. R. and S. C. acknowledge financial support from the National Science and Engineering Council of Canada in the form of an Alexander Graham Bell PGS D Fellowship and a Discovery Grant, respectively. [^1]: The models, however, are only intended for use on stellar populations of approximately solar metallicity. [^2]: [http://www.physics.mcmaster.ca/\\(\\sim\\)Harris/mwgc.dat](http://www.physics.mcmaster.ca/\(\sim\)Harris/mwgc.dat) [^3]: Recall that the model has a limited metallicity range, meaning that it can only be tested on metal-rich Galactic star clusters at this time. [^4]: \[O/Fe\] is fixed at 0.0 or +0.5, depending on the adopted isochrone (solar-scaled versus \(\alpha\)-enhanced), and \[Cr/Fe\] at 0.0, while \[Na/Fe\], \[Si/Fe\], and \[Ti/Fe\] track \[Mg/Fe\]. [^5]: We explicitly include the latter five elements to enable the most complete evaluations possible of *all* SPS models that predict abundance patterns. In future revisions to the present work, we envisage adding information on heavier elements which yield other unique insights from the perspective of galaxian stellar population analyses. [^6]: We have created an electronic data table which lists these results for each cluster in our sample. These data tables may be retrieved online at [www.astro.queensu.ca/people/Stephane_Courteau/roediger2013/index.html](www.astro.queensu.ca/people/Stephane_Courteau/roediger2013/index.html) [^7]: As discussed below and in , the C, N, and O abundances of individual GGC stars also depend on their evolutionary status, which results in an additional source of dispersion amongst these parameters [^8]: The age regimes quoted here refer, in a qualitative manner, to the values from MF09. [^9]: We possess \(\alpha\)-element abundances for 25 of our GGCs, while C, N, and O abundances are known for half of our sample, at best. [^10]: The depth of the data being modelled must also be considered. For instance, the luminosities of individual GGC members down to which the S05 spectra are sensitive has yet to be firmly established (but see Barber et al., in prep.). [^11]: <http://www.astro.queensu.ca/people/Stephane_Courteau/roediger2013/index.html>

{'timestamp': '2013-10-15T02:00:17', 'yymm': '1310', 'arxiv_id': '1310.3275', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3275'}

# Introduction {#sec:intro} We present and discuss data that are integral to two key unresolved questions regarding the stellar populations of early type galaxies. Do variations exist among the stellar initial mass functions of early-type galaxies (ETGs)? What is the physical mechanism for producing the stars that give rise to the UV flux in ETGs? A spate of recent results suggest that the stellar initial mass function (IMF) of giant ETGs has more low mass stars for a given total stellar mass than predicted even by the relatively extreme Salpeter mass function (\(dN/dM \propto M^{-2.35}\)). Specifically,, , and reach such a conclusion using spectral line indices that are sensitive to the dwarf-to-giant ratio. Independently, and find consistent results using dynamical models to measure the total mass. These results are to be contrasted with the extensive evidence in the local neighborhood for an IMF that turns over at sub-solar stellar masses. Although the evidence for variations in the bottom portion of the IMF now extends beyond studies of ETGs, with stellar clusters providing some of the most direct evidence, and direct imaging that resolves sub-solar mass stellar populations promising eventually to settle the matter, the analysis of ETGs is critical in our efforts to understand galaxy evolution. The second of our two questions traces its origin to the observation of an unexpectedly large UV flux from ETGs. Explanations proposed for this "UV-excess\" or "UV-upturn\" fall into two classes ever since the discovery paper: young stars and hot evolved stars. Although in some cases there may be a connection to recent, residual star formation, this explanation does not account for the majority of the UV-upturns even though it plays a more prominent role in intermediate and low mass ETGs and in lenticulars. Various scenarios involving evolved stars have been put forward including, post-asymptotic giant branch stars, hot horizontal branch stars, and accreting white dwarfs. Filling in the details of these populations has proved difficult, but a preliminary consensus is that hot, or extreme, horizontal branch (EHB) stars must be the key contributor to the UV fluxes of ETGs. Although models depend on numerous poorly constrained parameters (mass ratio distribution, tidal atmospheric stripping efficiency), they manage to reproduce several key observational characteristics of the UV-upturn population and have the benefit of relying on a population of objects that are directly observed to exist. The relevance of the UV-upturn extends beyond the nature of EHB stars. For example, models of EHB evolution may inform how the AGB is populated, which will affect how to compute the mass-to-light ratio in bands where the AGB contribution is significant. In general, investigators are focusing on the evolution of binary systems as a path to the formation of EHB stars. As expected, some of these models do predict a connection between UV excesses and AGB properties, thereby implying an effect on calculations of the mass-to-light ratio. Conversely, variations in the IMF could directly affect the UV-upturn population, particularly if binary systems with large mass ratios are an important progenitor class for EHBs. Our two questions are, therefore, intricately connected. UV-upturn properties vary widely among galaxies. Among the parameters that have been related to the strength of the UV-upturn are 1) stellar population age, which is particularly relevant if some fraction of the UV-bright population is relatively young, 2) galaxy mass, which is relevant if the galaxy's evolutionary history and total stellar mass are tied to each other, 3) stellar density, which is relevant if density-dependent binary processes play a role in forming EHB stars and densities are sufficiently large, and 4) metallicity, which is relevant because chemical abundance plays a role in stellar evolution, particularly in late-stage evolution where mass loss is significant. Interestingly, a promising correlation was found between UV properties and metallicity, with more prominent upturns seen in galaxies with greater Mg\(_2\) Lick index, an index which broadly tracks metallicity even though it does have some sensitivity to gravity. It is critical to continue the search for other patterns and investigate all such possibilities as a source of insight into the physical mechanisms at work to tie all of this together into a coherent framework. We examine a set of well-studied galaxies for which line-index based stellar mass-to-light ratios, \(\Upsilon_*\), are already available, as are metallicity diagnostics, internal velocity dispersions, and absolute luminosities. We present new measurements of the FUV and NUV fluxes from \(GALEX\) data for this set of galaxies. In §2 we discuss the data and in §3 we describe the findings of our study. # The Data and Measurements {#sec:data} The parent sample for this study is that of nearby ETGs studied by, which overlaps extensively with the sample defined by the SAURON team. Our addition to the existing data for this sample is the set of homogeneous measurements of the FUV (1350---1750Å) and NUV (1750---2750Å) photometry made possible by the Galaxy Evolution Explorer (\(GALEX\)) satellite. We analyze ultraviolet imaging data for the 34 ETGs and the bulge of M 31 in the Conroy & van Dokkum (2012) sample, where we obtained the most recently pipeline-processed (GR6/7 release) *GALEX* data from the MAST archive maintained by the Space Telescope Science Institute. For all 35 targets we have imaging data of MIS-depth (one *GALEX* orbit or more; mainly from the NGS, GI, MIS, and DIS surveys) in the *GALEX* NUV band and for 32 of them, all except NGC 3414, NGC 4382, and NGC 4564, we also have similarly deep FUV data. The first two of these three galaxies are excluded from our analysis because only shallow (\(\leq\)`<!-- -->`{=html}100 s-long exposures) FUV data from the All-sky Imaging Survey (AIS) exist. In the case of NGC 4564, the FUV total exposure time was 504 seconds, not quite as long as for the rest of the FUV imaging data but enough for the purpose of this study. We exclude the bulge of M 31 on the grounds that it is significantly different in nature than the other systems. We follow the procedure described by and. In summary, the analysis steps are (1) sky-background subtraction, using elliptical annuli centered on the galaxy that match the ellipticity and position angle (PA) of the galaxy and have major axes significantly larger than the isophotal diameter, D25, in all cases (the same region is used for each of the two UV bands), (2) interactive masking of foreground stars and background galaxies following an automated detection of all red (FUV-NUV \(>1\)) point sources as potential contaminants, 3) surface photometry within elliptical annuli with fixed center, ellipticity and PA (those of the D25 ellipse) and (4) calculation of the growth curve in both UV bands and the derivation of the corresponding asymptotic magnitudes. We use the FUV and NUV asymptotic magnitudes as the best measure of the total UV emission of our galaxies (Table [\[tab:dat\]](#tab:dat){reference-type="ref" reference="tab:dat"}). Distances are adopted from the on-line database NED. # Results We present the UV properties of the sample in Figure [\[fig:UV\]](#fig:UV){reference-type="ref" reference="fig:UV"} and Table [\[tab:dat\]](#tab:dat){reference-type="ref" reference="tab:dat"}. The top panel of the Figure shows the relation between \(M_I\) and \(M_{NUV}\), confirming that the UV flux is generally related to the properties of the galaxy as a whole and is not the product of UV "froth\" resulting from random low levels of recent star formation. We highlight the most striking outliers above the relationship (UV-bright) using blue triangles. In the bottom panel we explore the nature of these outliers in color-magnitude space. These outliers are clearly UV-bright and UV-blue and may be examples of galaxies where recent star formation is affecting the total photometry. We also note three galaxies that are redder than the red sequence in this panel, which are highlighted using red squares. These objects might represent cases where reddening is somewhat more prevalent, or alternatively galaxies with different mean stellar populations due to a factor such as mean metallicity. The latter supposition is not supported by the available data. The surprising result comes in the far right panel of Figure [\[fig:comp\]](#fig:comp){reference-type="ref" reference="fig:comp"}, where we show the relationship between UV color (FUV \(-\) NUV) and \(\Upsilon_*\). The relationship between FUV \(-\) NUV and \(\Upsilon_*\) is incredibly tight (Spearman \(R = 0.78\), excluding one outlier---see below), tighter even than that between \(\Upsilon_*\) and the velocity dispersion, \(\sigma_V\), found by and used to motivate the finding of a mass dependent IMF (see Table [\[tab:rank\]](#tab:rank){reference-type="ref" reference="tab:rank"} for a summary of the Spearman rank correlation coefficients in the lower half of the matrix and the probability that correlation coefficients that are at least as large as those measured arise by chance in the upper half). The one notable exception to the tight relationship between FUV \(-\) NUV and \(\Upsilon_*\) is NGC 2685, which is known as the "Spindle\" galaxy for its odd morphology and is modeled as a multiply-ringed polar-ring galaxy. We have removed this one object from all of the correlation calculations in Table [\[tab:rank\]](#tab:rank){reference-type="ref" reference="tab:rank"}, but include it in all of the Figures. The galaxies we identified as the blue outliers are NGC 474, 2685, 4262, 4660. These are all well-known S0 galaxies, three of which show clear interaction signatures. We discussed NGC 2685 above and NGC 4262 has an extended UV ring. NGC 474 is classified as peculiar on the basis of its shells and identified as having residual star formation by. The other three of our blue outliers are not among the galaxies classified by, but three other galaxies from our sample (NGC 1023, NGC 2974, and NGC 4459) are identified as having residual star formation, although these are manifestly not large outliers in Figure [\[fig:UV\]](#fig:UV){reference-type="ref" reference="fig:UV"}. Of the four blue outliers that we identified, NGC 4660, is the most normal in appearance and its large disk component was only discovered through its kinematics. We suspect that the unusual UV properties of these four are related to recent, and perhaps unusually strong, interactions rather than the S0 nature of the galaxies themselves because the three red outliers (NGC 524, NGC 4621, and NGC 5846) are also S0 galaxies and lack clear interaction signatures. Interestingly, "blue\" S0's are typically difficult to find in the types of environment where S0's are thought to be forming and may represent a stage following the E+A or post-starburst phase. Nevertheless, with the exception of NGC 2685, which is quite an unusual galaxy, the relationship between the UV-upturn color and \(\Upsilon_*\) holds. Of the other correlations explored in Table [\[tab:rank\]](#tab:rank){reference-type="ref" reference="tab:rank"} and Figure [\[fig:comp\]](#fig:comp){reference-type="ref" reference="fig:comp"}, only the ones with \(\sigma_V\) and \(\langle Mg/Fe \rangle\) are statistically significant, but they are both quantitatively somewhat weaker than that between FUV \(-\) NUV and \(\Upsilon_*\). We are concerned that the extremely strong correlation between FUV \(-\) NUV and \(\Upsilon_*\) results from a systematic error in the derivation of \(\Upsilon_*\) caused either directly or indirectly by the population responsible for the UV-upturn. A direct influence is unlikely as any black body responsible for the UV flux would provide negligible flux in the I-band where the IMF sensitive lines lie. More plausible is that the presence of UV emitting stars is related to a "distortion\" of the AGB or RGB population, either affecting the line indices or lowering the I-band flux below expectations. The former, an effect on the line indices, cannot be responsible because the variation in \(\Upsilon_*\) is also inferred from kinematic analyses. If, however, a lower-than-expected I\(-\)band flux is the rule in galaxies with large UV-upturns, then both the line-index and kinematic method will yield higher \(\Upsilon_*\) for those galaxies. In such a scenario, the measurement of \(\Upsilon_*\) would not be in error, but the inference that variations in \(\Upsilon_*\) are related to a change in the IMF would be. We now discuss two scenarios that attempt to explain the empirical findings with and without IMF variations. First, consider the situation where the UV upturn stars come directly from the population of AGB/RGB stars, which is to say that some process converts AGB/RGB stars into UV-upturn stars, and this phenomenon is not included in current spectral synthesis models. In such a scenario, proportionally more UV upturn stars in a certain galaxy means proportionally fewer AGB/RGB stars, and therefore a lower \(I-\)band luminosity. A relationship in this sense is found in at least one model of EHB evolution, although it is unclear whether that particular model will work quantitatively for the UV upturn-\(\Upsilon_*\) relationship we observe. Other possibilities, for example one in which chemical abundance plays a role through winds and mass transfer rates, would also help explain other observables such that of larger UV-upturn populations in galaxies with higher Mg\(_2\). Second, consider the situation where low mass stars in multiple star systems play a key role in transforming more massive evolved stars into the EHB stars of the UV-upturn population. In such a scenario, having more low mass stars in a bottom-heavy stellar population will lead to a larger UV-upturn population in the galaxies with higher \(\Upsilon_*\)---again in qualitative agreement with the observations. We cannot, with the data discussed so far, distinguish between these two possibilities. Photometric data in other passbands might help resolve the situation. In a scenario where the I-band luminosity is different than the model expectations---whether the origin of that discrepancy is physical or observational---we might expect to see significant deviations in colors involving the I-band. We obtain H-band data for much of the sample from 2MASS (collated from the NED database) and find no significant correlation between I-band and I\(-\)H. We conclude that variations in I-band luminosities alone are an unlikely origin for our findings. Given 1) that the \(\Upsilon_*\) correlation with \(\sigma_V\) is observed when \(\Upsilon_*\) is either measured spectroscopically or kinematically, 2) the lack of any signature of odd behavior in the I-band luminosities as reflected in I-H, 3) the lack of a dependence of the \(\Upsilon_*\) of old (log(age)\(>\)`<!-- -->`{=html}9.8) Local Group stellar clusters with \(\langle Mg/Fe \rangle\), and 4) that the strongest correlation we find is between \(\Upsilon_*\) and UV color, we conclude that the variation in \(\Upsilon_*\) is real, and that it is directly correlated to UV color. We suggest that IMF variations are responsible for the differences in the EHB populations of ETGs. Nevertheless, we need to remain cognizant of the possibility that other physical effects beside IMF variations are responsible for the various correlations. The possibility that metallicity is somehow responsible for both the IMF and UV-upturn variations remains a real possibility. # Conclusion Using *GALEX* data we have identified a strong correlation between UV color and the stellar mass to light ratio, \(\Upsilon_*\) in early type galaxies (ETGs). This correlation is stronger, within the same sample, than those previously identified between \(\Upsilon_*\) and either \(\sigma_V\), metallicity, or alpha enhancement. We are faced with that task of explaining an interrelated set of significant correlations between \(\sigma_V\), \(\langle Mg/Fe \rangle\), \(\Upsilon_*\), and UV color. It is inherently difficult to draw conclusions from a quantitative comparison among statistically significant correlations because the measured strength of the correlations depends on the magnitude of the observational uncertainties. We hypothesize that the variations in the IMF identified in previous investigations are driving the correlation between \(\Upsilon_*\) and UV color, but the role of metallicity in both is still unknown. As is usually the case with empirical correlations, understanding their origin and determining whether a correlation implies causality is a more difficult task than the discovery of the correlation. A final intriguing possibility is the use of the UV color as a tracer of IMF variations in galaxies without recent star formation, at least as a selection criteria with which to identify galaxies for further study.

{'timestamp': '2013-10-14T02:11:19', 'yymm': '1310', 'arxiv_id': '1310.3255', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3255'}

# Introduction Understanding entanglement in multipartite quantum states is one of the major goals in quantum information theory. While entanglement in bipartite pure states is well understood both in the local operations and classical communication (LOCC) and stochastic local operations and classical communication (SLOCC) paradigms at the single copy level and asymptotically, much less is known for three or more parties. Among the few results are the complete classification of pure three and four qubit states under SLOCC equivalence. Already for three qubits one finds two incomparable genuinely tripartite-entangled classes, indicating the complexity of the problem. Recently a connection between algebraic complexity theory and the study of SLOCC transformations in an asymptotic setting has been discovered, opening the possibility to transfer ideas from one field to the other. More precisely, it has been observed that finding the rate at which triples of EPR pairs shared among three parties can be extracted from GHZ states is the same as finding the exponent of matrix multiplication, commonly denoted by \(\omega\), a problem that has been studied for over 40 years by mathematicians. This number is the infimum of real numbers \(\tau\) such that \(n\times n\) matrices can be multiplied together using \(O(n^\tau)\) arithmetic operations. Later, bounds on the tensor rank of multiple copies of the W state have been found, as well as its generalization to more than three parties and other symmetric states. In this paper we further explore this connection and show that border rank and degeneration, two important concepts in algebraic complexity, can be used to prove nontrivial bounds on asymptotic conversion rates. These techniques are well-known in algebraic complexity theory, but seem not to have been applied so far in the present context. The structure of the paper is as follows. In section [2](#sec:aSLOCC){reference-type="ref" reference="sec:aSLOCC"} we provide the definition of asymptotic SLOCC conversion rates between two states and give some basic properties, including its connection to tensor rank and asymptotic rank. In section [3](#sec:border){reference-type="ref" reference="sec:border"} we introduce the concepts of degeneration and border rank into the study of asymptotic SLOCC transformations and illustrate their usefulness by computing the conversion rate from GHZ to (generalized) W states for any number of parties. Section [4](#sec:symm){reference-type="ref" reference="sec:symm"} extends the latter result to an upper bound on the conversion rate from GHZ states to certain families of symmetric states, uniformly in the number of subsystems. In section [5](#sec:WtoGHZ){reference-type="ref" reference="sec:WtoGHZ"} we compute the conversion rate from W to GHZ states. The construction follows an idea of Coppersmith and Winograd which was used to prove an upper bound on the exponent of matrix multiplication. Optimality is shown using the monotones introduced by Strassen. During the preparation of this manuscript we have learned about independent related work by Yu, Guo and Duan where they give a proof of theorem [\[thm:GHZtoW\]](#thm:GHZtoW){reference-type="ref" reference="thm:GHZtoW"}. # Asymptotic SLOCC transformations {#sec:aSLOCC} Given a pair of \(k\)-partite pure states \(\psi\in\mathcal{H}_1\otimes\cdots\otimes\mathcal{H}_k\) and \(\varphi\in\mathcal{K}_1\otimes\cdots\otimes\mathcal{K}_k\) we denote by \(\psi\slto\varphi\) the fact that there exist linear transformations \(A_i:\mathcal{H}_i\to\mathcal{K}_i\) such that \(\varphi=(A_1\otimes\cdots\otimes A_k)\psi\). This is the well-known mathematical condition for the ability of \(k\) parties to transform the state \(\psi\) into \(\varphi\) with nonzero probability, each being able to control one of the subsystems, while their actions are allowed to be coordinated via classical communication. It is clear that this definition completely ignores the normalization of the states (in fact, only the vector space structure is used and not the norm). For simplicity we therefore prefer to work with unnormalized states. We are interested in transformations of multiple copies where the number of initial and final copies may be different. The relevant quantity is \[\omega_n(\psi,\varphi):=\frac{1}{n}\inf\{m\in\mathbb{N}|\psi^{\otimes m}\slto\varphi^{\otimes n}\}\] where the infimum of the empty set is considered to be \(\infty\). It is easy to see that \((n_1+n_2)\omega_{n_1+n_2}(\psi,\varphi)\le n_1\omega_{n_1}(\psi,\varphi)+n_2\omega_{n_2}(\psi,\varphi)\), which implies that the limit \(\omega(\psi,\varphi):=\lim_{n\to\infty}\omega_n(\psi,\varphi)\) exists and is equal to \(\inf\omega_n(\psi,\varphi)\). It was observed by Chitambar et al. that when \(k=3\) and we let \(\psi=GHZ\) and \(\varphi\) be the triple of EPR pairs, one shared between each pair of subsystems, then \(\omega(\psi,\varphi)\) is precisely the exponent of matrix multiplication, the smallest real number \(\tau\) such that for any \(\varepsilon>0\) two \(n\times n\) matrices can be multiplied using \(O(n^{\tau+\varepsilon})\) arithmetic operations. This exponent is traditionally denoted by \(\omega\)--our notation is chosen so that it emphasizes this connection to algebraic complexity. SLOCC transformation rates have also been investigated in ref., the relation between our quantity and theirs is \[R(\psi,\varphi)=\frac{1}{\omega(\psi,\varphi)}\] For \(S\subseteq[k]\) let \(\rk_S\psi\) denote the Schmidt rank of \(\psi\), considered as a bipartite state on the subsystems \(S\) and \(\bar{S}=[k]\setminus S\). The following facts are simple consequences of the definition, and therefore we omit the proofs. By taking the limit \(n\to\infty\) we get the following useful asymptotic forms: We will make use of the \(a\)-level generalization of GHZ states: \[GHZ_a=\sum_{i=1}^a|ii\ldots i\rangle\] where the number of parties should always be clear from the context. We omit the subscript when \(a=2\). In algebraic complexity these are called unit tensors and have a special role due to their connection to tensor rank. Tensor rank itself can be seen as a generalization of matrix rank (or Schmidt rank), and is defined as the minimum number of product states spanning a subspace containing a given tensor. We denote tensor rank by \(\rk\). The quantity \(2^{\omega(GHZ,\psi)}\) is also known as the asymptotic rank of \(\psi\). Since the rank of a tensor is always finite, one can extract any state from GHZ states at a positive rate (\(\omega(GHZ,\psi)<\infty\)). In the other direction, a state clearly needs to be globally entangled if we are to distill GHZ states out of many copies of it, i.e. it cannot be biseparable across any bipartite cut. For two and three subsystems it is easy to see that this condition is also sufficient. It turns out that this is also true for more parties, as the following lemma shows, but the argument is more complicated in this case. A similar result is proved for exact LOCC transformations in ref.. Since \(EPR_{1,2}\otimes EPR_{1,3}\otimes\cdots\otimes EPR_{1,k}\) can be converted to a GHZ via SLOCC using teleportation, this implies that \[\begin{split} \omega(\psi,GHZ) & \le\omega(\psi,EPR_{1,2}\otimes EPR_{1,3}\otimes\cdots\otimes EPR_{1,k}) \\ & \le\omega(\psi,EPR_{1,2})+\omega(\psi,EPR_{1,3})+\cdots+\omega(\psi,EPR_{1,k})<\infty \end{split}\] # Degeneration and border rank {#sec:border} It is a standard fact that a GHZ state cannot be transformed into a W state by SLOCC, forming distinct entanglement classes of three qubits, but a W state can be approximated to arbitrary precision with states in the GHZ orbit. Even though \(\rk W=3\), it can be approximated by rank 2 GHZ states, and we say that its border rank is 2 (notation: \(\brk W=2\)). This phenomenon is known as degeneration in algebraic complexity, and is important in the study of the complexity of tensor powers. More generally, we say that \(\psi\) degenerates to \(\varphi\) iff \(\varphi\) is in the orbit closure of \(\psi\) under the action of SLOCC. We remark that the closure in the Zariski topology is the same as that in the Euclidean topology. The following alternative definition is more convenient for calculations: It can be shown that over algebraically closed base fields this algebraic definition is equivalent to the analytic one described above, see e.g.. Just as the rank of a state \(\psi\) can be characterized as the smallest \(a\) such that \(GHZ_a\slto\psi\), the border rank \(\brk\psi\) is the smallest \(a\) such that \(GHZ_a\) degenerates to \(\psi\). We illustrate the concept using the W state as an example. Consider the following equality: \[\label{eq:Wappr1} \frac{(|0\rangle+\varepsilon|1\rangle)\otimes(|0\rangle+\varepsilon|1\rangle)\otimes(|0\rangle+\varepsilon|1\rangle)-|000\rangle}{\varepsilon}=(|100\rangle+|010\rangle+|001\rangle)+\varepsilon(|011\rangle+|101\rangle+|110\rangle)+\varepsilon^2|111\rangle\] For any \(\varepsilon\neq 0\) the tensor on the right hand side has rank \(2\) and in the limit \(\varepsilon\to 0\) it becomes the W state, hence \(\brk W=2\). Note that the limit \(\varepsilon\to 0\) can be seen as the derivative of a polynomial at \(\varepsilon=0\). We can also understand the situation in a geometric way: states contained in *secants* to the set of separable states have rank at most 2, while states on a *tangent* to the set of separable states have border rank at most 2. The idea works for higher derivatives as well. The largest degree appearing on the right hand side (\(2\) in the example) plays a role later, and is called the error degree of the approximation. In algebraic complexity it is a well-known result that, asymptotically, degeneration and restriction (i.e. SLOCC convertibility) are equivalent. In the language of asymptotic SLOCC transformations the statement translates to the following: The most general proof uses some nontrivial algebraic facts, but works for arbitrary base fields. Here we present a simplified version of the argument from, which works over algebraically closed fields of characteristic \(0\). As the base field \(\mathbb{C}\) is the most important in quantum physics, this level of generality is more than enough for our purposes. With this powerful result at hand it is easy to prove that GHZ states can be transformed to W states by SLOCC asymptotically at rate \(1\): This result has also been obtained in, by showing that the tensor rank of \(W^{\otimes n}\) is \(O(n^{k-1}2^n)\) for fixed \(k\) as \(n\to\infty\). Our proof improves this bound to \((n(k-1)+1)2^n\). Note that the best lower bound found so far is \((k-1)2^n-k+2\). # Symmetric states {#sec:symm} In this section we generalize the result on the asymptotic rank of W states to certain symmetric states. In it was shown that the tensor rank of the Dicke state \(D_{m,n}\) that is the symmetrization of \(|00\ldots 011\ldots 1\rangle\) with \(m\) \(0\)-s and \(n\) \(1\)-s is \(\max\{n,m\}+1\). First we show that its border rank and its asymptotic rank are both \(\min\{n,m\}+1\). Without loss of generality we can suppose \(m\le n\). We can write \[\frac{1}{m+1}\sum_{j=1}^{m+1}e^{-\frac{2\pi i}{m+1}mj}\left(|0\rangle+\varepsilon e^{\frac{2\pi i}{m+1}j}|1\rangle\right)\otimes\cdots\otimes\left(|0\rangle+\varepsilon e^{\frac{2\pi i}{m+1}j}|1\rangle\right)=\varepsilon^nD_{m,n}+O(\varepsilon^{2m+1})\] which implies \(\brk D_{m,n}\le m+1\). For the lower bound consider the bipartite cut where the first \(m\) and the last \(n\) subsystems are the two parts. The rank across this cut is precisely \(m+1\), so \(\brk D_{m+n}\ge m+1\) and the same lower bound holds for the asymptotic rank. For the conversion rate this implies \(\omega(GHZ,D_{m,n})=\log_2(\min\{m,n\}+1)\). Remarkably, if we keep the number of "excitations" \(n\) fixed and let \(m\to\infty\) the asymptotic conversion rate remains bounded, while the rank grows linearly. In the following we investigate the conversion rates of more general symmetric states having similar extensions to more subsystems. We make the following definition: In an earlier draft of this paper we proved that \(\omega(GHZ,D_{\lambda,k})\) is bounded as \(k\to\infty\) by finding an explicit upper bound on \(\brk D_{\lambda,k}\) not depending on \(k\), but we could not say how tight that bound is. As R. Duan explained to us, the proof in the appendix of can be formulated in the present framework and gives a much better bound--in fact one that agrees with the exact value for large \(k\). Here we outline how their result translates to a bound on the border rank. First observe that \[\frac{1}{\lambda_1!\cdots\lambda_d!} \left(\frac{\partial}{\partial\varepsilon_1}\right)^{\lambda_1}\cdots\left(\frac{\partial}{\partial\varepsilon_d}\right)^{\lambda_d}\big(|0\rangle+\varepsilon_1|1\rangle+\ldots+\varepsilon_d|d\rangle\big)\otimes\cdots\otimes\big(|0\rangle+\varepsilon_1|1\rangle+\ldots+\varepsilon_d|d\rangle\big)\Big|_{\lambda_i=0}=D_{\lambda,k}\] where the number of tensor factors is \(k\). The left hand side can be realized as the limit of finite differences using \[\frac{\partial}{\partial x}f(x)\Big|_{x=0}=\lim_{h\to 0}\frac{1}{h^n}\sum_{i=0}^n(-1)^i\binom{n}{i}f\big((n-i)h\big)\] for each of the variables. Without the limit, this replacement results in a linear combination depending on \(\varepsilon_1,\ldots,\varepsilon_d\) with \((\lambda_1+1)\cdots(\lambda_d+1)\) terms, each of which has rank \(1\), therefore the rank of the sum is at most \((\lambda_1+1)\cdots(\lambda_d+1)\). Since this remains true no matter how small values we substitute for \(\varepsilon_i\), we can conclude that \(\brk D_{\lambda,k}\le(\lambda_1+1)\cdots(\lambda_d+1)\). In general, for small \(k\) we expect the asymptotic rank to become smaller than this upper bound, but using the bipartite rank ref. shows that the asymptotic rank and therefore also the border rank is at least \((\lambda_1+1)\cdots(\lambda_d+1)\) when \(k\ge\lambda_1+\cdots+\lambda_k+\prod_{i=1}^d(\lambda_i+1)\). This means that for large \(k\) the asymptotic rank stays constant and is equal to the above bound. Note that, in contrast, the tensor rank \(\rk D_{\lambda,k}^{\otimes n}\) grows at least linearly in \(k\) for fixed \(\lambda\) and \(n\), since \(D_{\lambda,k}^{\otimes n}\slto D_{|\lambda|,k-|\lambda|}\). It would be desirable to find the asymptotic rank of any Dicke state \(D_{\lambda,k}\). Unfortunately, this appears to be difficult. Coppersmith and Winograd conjecture that \(\omega(GHZ_3,D_{(1,1,1),3})=1\), but proving this would imply that the exponent of matrix multiplication is \(2\), a long-standing open problem in algebraic complexity theory. # Transforming W states into GHZ states {#sec:WtoGHZ} We turn to transformations in the opposite direction. It is well known that a W state cannot be transformed into a GHZ state, but it is easy to see that \(k-1\) W states are enough to create a GHZ state: if \(k-2\) of the parties perform the transformation \(|0\rangle\mapsto|0\rangle,|1\rangle\mapsto 0\), then the remaining two end up sharing an EPR pair. Thus \(k-1\) W states are enough to create EPR pairs between the first and each of the remaining \(k-1\) parties, and then they can use teleportation to produce a GHZ state. In the following we will see that \(\omega(W,GHZ)\) is in fact sublinear in the number of parties. For tripartite systems an upper bound on the conversion rate \(\omega(W,GHZ)\) has essentially been computed as a byproduct by Coppersmith and Winograd. Later Strassen introduced a family of monotones for any number of parties which show that this upper bound is in fact optimal. In this section we are going to generalize the construction in and use the monotones to prove optimality, thereby finding the exact values of \(\omega(W,GHZ)\) for any number of subsystems. First we summarize the relevant definitions and theorems by Strassen, formulated in terms of SLOCC conversion rates and for any number of parties. To this end we need to fix some more notations. The set \[\Theta:=\left\{(\theta_1,\ldots,\theta_k)\in\mathbb{R}^{k}\big|\theta_1+\ldots+\theta_k=1, \forall j\in[k]:\theta_j\ge 0\right\}\] is the standard \(k\)-simplex. Given finite sets \(I_1,\ldots,I_k\) and a probability measure on (a subset of) \(I_1\times\cdots\times I_k\), its marginals are denoted by \(P_1,\ldots,P_k\) and for \(\theta\in\Theta\) we set \[H_\theta(P)=\sum_{j=1}^k\theta_jH(P_j)\] with \(H(P_j)=-\sum_{i\in I_j}P_j(i)\log_2 P_j(i)\). For \(\emptyset\neq\Psi\subseteq I_1\times\cdots\times I_k\) we set \(H_\theta(\Psi)=\max_P H_\theta(P)\) where the maximization is over probability measures on \(\Psi\), and extend this as \(H_\theta(\emptyset)=-\infty\). For \(f\in \mathcal{H}_1\otimes\cdots\otimes \mathcal{H}_k\) is a tensor and a \(k\)-tuple of ordered bases \(C=((u_{j,i})_{i=1}^{d_j})_{j=1}^k\) (one for each vector space), we can form the coordinate array \((f_{i_1,\ldots,i_k})_{i_j=1}^{d_j}\in\mathbb{C}^{d_1\times\cdots\times d_k}\) of \(f\). We introduce the notation \[\supp_Cf=\{(i_1,\ldots,i_k)|f_{i_1,\ldots,i_k}\neq 0\}\subseteq[d_1]\times\cdots\times[d_k]\] for the support of \(f\) with respect to \(C\). The set of such \(k\)-tuples of ordered bases will be denoted by \(\mathcal{C}(\mathcal{H}_1,\ldots,\mathcal{H}_k)\) or simply \(\mathcal{C}\) when the vector spaces are clear from the context. For \(f\in \mathcal{H}_1\otimes\cdots\otimes \mathcal{H}_k\) and \(\theta\in\Theta\) as before, we set \[\label{eq:uppersf} \rho^\theta(f)=\min_{C\in\mathcal{C}}H_\theta(\supp_Cf)\qquad\text{and}\qquad\zeta^\theta(f)=2^{\rho^\theta(f)}\] and call \(\zeta^\theta\) the *upper support functional*. The proof can be found in (explicitly only for \(k=3\), but the generalization is straightforward). For \(f\in \mathcal{H}_1\otimes\cdots\otimes \mathcal{H}_k\) and \(\theta\in\Theta\) as before, we set \[\label{eq:lowersf} \rho_\theta(f)=\max_{C\in\mathcal{C}}H_\theta(\max\supp_Cf)\qquad\text{and}\qquad\zeta_\theta(f)=2^{\rho_\theta(f)}\] where \(\max\supp_Cf\) denotes the set of maximal points in the support with respect to the product partial order, and call \(\zeta_\theta\) the *lower support functional*. The proof can be found in. The two functionals are related as follows: The proof can be found in. Note that \(\rho_\theta(f)\) as a function of \(\theta\) is a maximum of affine functions, and hence convex. \(\zeta_\theta(f)\) is the composition of \(\rho_\theta(f)\) with the increasing convex function \(x\mapsto 2^x\) and hence also convex. We say that a tensor \(f\) is *\(\theta\)-robust* when \(\zeta_\theta(f)=\zeta^\theta(f)\) and *robust* when it is \(\theta\)-robust for all \(\theta\in\Theta\). By the properties of \(\zeta^\theta\) and \(\zeta_\theta\) we see that \(f\in \mathcal{H}_1\otimes\cdots\otimes \mathcal{H}_k\) is \(\theta\)-robust iff there are \(C,C'\in\mathcal{C}(\mathcal{H}_1,\ldots,\mathcal{H}_k)\) and a probability measure \(P\) on \(\max\supp_Cf\) such that \(H_\theta(P)\ge H_\theta(\supp_{C'}f)\). Let us call \(f\) *oblique* when there exists \(C\in\mathcal{C}(\mathcal{H}_1,\ldots,\mathcal{H}_k)\) such that \(\supp_Cf\) is an antichain in the product partial order (i.e. no two elements are comparable). An oblique tensor is also robust. The sets of \(\theta\)-robust, robust and oblique tensors are closed under direct sums and tensor products. This implies that for robust tensors the upper and lower support functionals are multiplicative, because one of them is always submultiplicative while the other is always supermultiplicative, and the two agree on tensor powers of a robust tensor. Because of this multiplicativity one can use the upper and lower support functionals to bound the SLOCC conversion rates for robust tensors as follows. Let \(\psi\) and \(\varphi\) be robust \(k\)-partite states. By definition, for each \(n\) one can transform \(n\omega_n(\psi,\varphi)\) copies of \(\psi\) to \(n\) copies of \(\varphi\), and therefore, for each \(\theta\in\Theta\) we have \[\zeta_\theta(\psi)^{n\omega_n(\psi,\varphi)}=\zeta_\theta(\psi^{\otimes n\omega_n(\psi,\varphi)})\ge \zeta_\theta(\varphi^{\otimes n})=\zeta_\theta(\varphi)^n\] After taking logarithms and letting \(n\to\infty\) we get \(\omega(\psi,\varphi)\rho_\theta(\psi)\ge\rho_\theta(\varphi)\). Since this holds for any \(\theta\in\Theta\) we can also write \[\omega(\psi,\varphi)\ge\max_{\theta\in\Theta}\frac{\rho_\theta(\varphi)}{\rho_\theta(\psi)}\] The construction in makes use of Salem-Spencer sets. These are "large" sets of numbers containing no nontrivial three-term arithmetic progressions. We introduce the following generalization of Salem-Spencer sets: Note that \(2\)-average-free sets are precisely sets without three-term arithmetic progressions. The next lemma extends the result of Salem and Spencer: The proof is very similar to that of, therefore we leave it to the appendix. Now we have everything at hand to find the conversion rate from W to GHZ states. Note that as \(k\to\infty\) the sequence \(\omega(W,GHZ)\) is asymptotically equal to \(\frac{k}{\log_2k}\). In contrast, the trivial upper bound was linear in \(k\). # Conclusion We used the concept of border rank from algebraic complexity theory to bound asymptotic SLOCC conversion rates from GHZ to symmetric states. In the case of multiqubit W and Dicke states this bound gives the exact values. In some cases this results in a dramatic improvement to the bound obtained using tensor rank of an arbitrary fixed number of copies. In the other direction, we prove that \(n\) copies of the \(k\)-qubit W state can be transformed into \(h(\frac{1}{k})n+o(n)\) GHZ states. This follows from a generalization of a construction by Coppersmith and Winograd, optimal as shown using the monotones introduced by Strassen. These monotones are functions \(\zeta^\theta\), \(\zeta_\theta\) of the states (see eqs. [\[eq:uppersf\]](#eq:uppersf){reference-type="eqref" reference="eq:uppersf"} and [\[eq:lowersf\]](#eq:lowersf){reference-type="eqref" reference="eq:lowersf"}), not increasing under SLOCC, and are not restricted to the asymptotic regime, although they become particularly powerful in this setting. Observe that a \(k\) qubit \(W\) state is the symmetrization of \(|100\ldots 0\rangle\) and the empirical distribution of this sequence of bits is \((\frac{1}{k},1-\frac{1}{k})\), the entropy of which gives the rate at which GHZ states can be extracted from W states. It seems likely that our proof can be extended to show that \[\omega(D_\lambda)=\frac{1}{H\left(\frac{\lambda_1}{|\lambda|},\frac{\lambda_2}{|\lambda|},\ldots,\frac{\lambda_k}{|\lambda|}\right)}\] with \(H(p_1,\ldots,p_k)=-\sum_i p_i\log_2p_i\). Simple calculation shows that here the denominator is equal to \(\rho_\theta(\omega(D_\lambda))\), so the right hand side is a lower bound on the left hand side. It would be desirable, but much more difficult to find the conversion rates from GHZ states to an arbitrary \(D_{\lambda,k}\), or at least the limit for large \(k\). We leave these problems as open questions for further study.

{'timestamp': '2014-08-22T02:10:21', 'yymm': '1310', 'arxiv_id': '1310.3244', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3244'}

# Introduction {#sec:intro} Several problems in top quark physics require a full reconstruction of the kinematics of the top quark--antiquark pair. For example, to measure the forward-backward (or charge) asymmetry in  production, it is essential to know the direction of both the top quark and the antiquark. We consider  events where each top quark decays into a \(b\) quark and a , and where one  decays hadronically (\(W\to q'\bar{q}\)) and one  decays leptonically (\(W\to l\nu\)). We classify top quarks as "leptonic" or "hadronic", based on the mode of the  decay. The final state contains a lepton, a neutrino and four quarks that subsequently shower and hadronize into jets. This channel is commonly referred to as "". The four final state quarks do not always yield four reconstructed jets, which is the case, for example, when one of the quarks is too soft or when the angular separation between two of them is small. Though the signal purity is lower in the sample of events with exactly three jets than in the sample with at least four jets, it is still useful for measuring top properties , effectively increasing the sample size by \(\approx55\%\). Furthermore, extending the event selection of a top property measurement in the  channel to include three-jet events can reduce the acceptance bias  and reduces systematic uncertainties related to jet reconstruction, as events with one unreconstructed jet are still used. The three-jet sample can also be interesting in its own right. For example, at the Large Hadron Collider (LHC)  production of events with at least four jets is dominated by initial states that contain gluons, while the three-jet sample is enriched in  pairs produced from  initial states. When one of the jets from top decay is lost, it is not possible to fully reconstruct the  decay chain, which has so far limited the use of lepton plus three jets () events to measurements of observables such as the production cross section  and the rapidity[^1] of the lepton . The kinematics of the  events with a lepton and at least four jets () is fully reconstructed by matching four of the jets to the four final state quarks from  decay (for example, see ). As for events with at least four jets, the main challenge in fully reconstructing three-jet events is to disentangle the two top-quark decay chains. That is, the main challenge is to match the observed jets with the quarks from  decay, though with only three jets available a perfect 1-to-1 correspondence is impossible and partial matchings are used instead. In this  we present a method to infer the direction and kinematics of the top quark and antiquark in  events where only three jets are reconstructed, and demonstrate the application of the method to simulated  events. We focus on \(\ppbar\to\ttbar\) production at a center of mass energy of \(1.96\TeV\), as in the Tevatron. About half of the   events produced at the Tevatron contain only three jets. The main steps of the method have been described in Ref. , where it is used to reconstruct the directions of the top quark and antiquark and the invariant mass of the  system. This  provides the details of the method and quantifies its performance. We discuss the selection of the events in Section [2](#sec:sel){reference-type="ref" reference="sec:sel"}. In Section [3](#sec:reco3j){reference-type="ref" reference="sec:reco3j"} we detail the method to partially reconstruct the  pair using the invariant mass of various combinations of jets and jet lifetime observables . We compare the performance of different reconstruction algorithms in Section [4](#sec:performance){reference-type="ref" reference="sec:performance"}, for which we introduce a new figure of merit (FOM). The reconstruction of  events at the LHC poses different problems. The typical jet-selection threshold for the LHC is transverse momentum \(\pt>30\GeV\), and only in \(40(2)\%\) [^2] of the  events do all four jets associated with -decay quarks pass this threshold. Yet, due to the initial state radiation (ISR) only a small fraction of  events produced at the LHC end up in the three-jet sample. For the LHC  events, a method similar to that presented here may suffice. Roughly \(40\%\) of the LHC  events contain only three jets associated with  decay quarks, with the other jets due to ISR. An extension of the algorithm discussed in this paper could be used to partially reconstruct these events and thus increase the number of reconstructed events by approximately a factor of two. # Samples and selection {#sec:sel} The results shown in this paper are based on simulated \(\ppbar\to\ttbar\) events with a collision center-of-mass energy of \(\sqrt{s}=1.96\TeV\). The events were simulated with the  event generator  and processed through a detector simulation and object reconstruction that largely correspond to but are not identical to that of the  experiment. In particular, some of the quality selection criteria are not applied since they are not relevant for the development of the method. Simulated energy deposits in the calorimetry are clustered into jets using the "Run II Midpoint cone algorithm"  with a cone radius of \(0.5\) in the \(y\)-\(\phi\) plane, where \(\phi\) is the azimuthal angle and \(y\) is the rapidity. We select jets with \(\pt>20\GeV\) and with pseudorapidity \(|\eta|<2.5\). We select leptons from electron and muon candidates with \(\pt>20\GeV\) and with \(|\eta_e|<1.1\) or \(|\eta_\mu|<2.0\). We then select events with exactly one lepton and exactly three jets. We require that the transverse momentum imbalance measured by the calorimetry, , is greater than \(20\GeV\). We reject events where the  is closely aligned with the lepton and events with \(>500\GeV\). These two cuts suppress multijet background and events with misreconstructed , respectively. Generally, the signal purity is lower in the sample of events with exactly three jets than in the sample with at least four jets. However additional selection criteria, e.g. identification of jets associated with \(b\)-quarks (\(b\)-tagging), can improve the situation, making the  sample useful for measuring top properties. In particular, in  it was shown that purity of  sample with two \(b\)-tags is similar to that of  with one \(b\)-tag. We further categorize the selected events by how well the reconstructed jets match the quarks from  decay, as that affects the quality of reconstruction. We consider a jet to be matched to a quark when their angular separation \(\Delta R=\sqrt{(\Delta y)^2+(\Delta\phi)^2}\) is less than 0.5. We classify an event as "matchable" if all  decay products assumed to be present by the reconstruction algorithm were matched to reconstructed objects. For the reconstruction of  events at the Tevatron , a matchable event is the one in which the four jets of highest  match the four final state quarks from  decay. Only \(55(1)\%\) of the  events at the Tevatron are matchable. In the context of this  a  event is considered matchable if one jet matches the \(b\) quark from of the leptonic top quark decay and the two other jets match two of the three quarks from the decay of the hadronically decaying top quark. \(20(1)\%\) of the  events are classified as unmatchable because the \(b\) jet from the leptonic top decay, which is essential to the described algorithm, is lost. In 4.0(2)% of the events two jets were lost, while an extra one was gained from initial or final state radiation. Thus, \(76(1)\%\) of the  events are considered matchable. # Reconstructing  in  events {#sec:reco3j} For almost half of the simulated \(\ppbar\to\ttbar\) pairs that decay in the  channel, only three jets are reconstructed. In our study scenario, two quarks yield a single jet due to an accidental overlap in \(\approx18\%\) of these  events. One of the quarks is too forward (high \(|\eta|\)) to yield a selected jet in \(\approx8\%\) of the events. In the remaining \(\approx74\%\) of the events, either one of the quarks was too soft (low ) to yield a selected jet or a jet was lost due to reconstruction and identification inefficiencies. In Fig. [\[fig:cartoon\]](#fig:cartoon){reference-type="ref" reference="fig:cartoon"} we show a schematic of a possible  decay process. Instead of trying to infer the kinematics of the missing or merged jet in a  event, we partially reconstruct the  system by neglecting this jet altogether. Though there is some experimental sensitivity to the presence of two quarks in a single jet, e.g., through the jet width and mass, we found it too weak to be useful. Thus we do not attempt to "unmerge" any of the jets and assign two quarks to it. Events in the  channel are often reconstructed using a "kinematic fit" algorithm, which modifies the measured momenta to satisfy the known resonance masses ( Ref. ). Given that we neglect the missing jet, such refinements are of little use for  events. Thus we employ a simpler approach to partially reconstruct the  system in  events. ## Reconstructing the leptonic {#sec:recoWj} We start by reconstructing the leptonically decaying  using the lepton momentum and the . The neutrino momentum in the plane transverse to the beam direction, , is initially set equal to the . The longitudinal component of neutrino momentum, , is calculated using a constraint on the  mass, . The resultant quadratic equation can have two solutions, which creates a two-fold ambiguity. Both solutions are considered. Following Ref. , when the discriminant of the quadratic equation for  is negative, we scale  to satisfy the  constraint with a discriminant equal to zero. This results in another quadratic equation which yields two solutions for the scale, at least one of which is positive. When both solutions are positive, we use the one that is closer to unity. ## Reconstructing the top-quark candidates {#sec:recotj} The next step is to form leptonic and hadronic top quark candidates. To do so, we assume that the lost jet is from the decay of the hadronic top quark. One of the jets is combined with the leptonic  to form a leptonic top candidate. The two remaining jets are combined to form a "proxy" for the hadronic top quark, which serves instead of a fully reconstructed candidate. The assignment is completely defined by the choice of leptonic \(b\) jet. If the previous step yielded two  solutions, for each assignment we choose the solution where the combination of the leptonic \(b\) jet, the lepton and the neutrino yields an invariant mass closer to the nominal top quark mass . ### \(\chi^2\) method {#sec:chi2j} Invariant mass distributions on both the leptonic and hadronic sides have characteristic shapes as shown in Fig. [\[fig:mass_fits\]](#fig:mass_fits){reference-type="ref" reference="fig:mass_fits"}. Both can be used to find the best jet assignment. The distributions were made using an adaptive kernel estimator . A simple way to choose an assignment is to use a  test statistic for the masses reconstructed for the leptonic top candidate (\(m_t\)) and for the proxy (\(m_p\)): \[\chisq=\left(\frac{m_t-m_t^0}{\sigma_t}\right)^2+\left(\frac{m_p-m_p^0}{\sigma_p}\right)^2, \label{eq:chisq}\] where \(m_t^0\) (\(m_p^0\)) and \(\sigma_t\) (\(\sigma_p\)) are the mean and width of the Gaussian fits for leptonic (proxy) masses shown in Fig. [\[fig:mass_fits\]](#fig:mass_fits){reference-type="ref" reference="fig:mass_fits"}. This approach picks the correct assignment in \(66.0(1)\%\) of the cases where such an assignment exists. Below we discuss more detailed treatments that improve upon this basic technique. ### Complete likelihood method {#sec:lhoodj} We improve the choice of the assignment by replacing the  with a likelihood function. The likelihood formalism allows us to take into account additional information. The use of the invariant masses of the incorrect assignments, which too have distinct shapes, is detailed below. The use of "\(b\)-tagging" observables that attempt to identify jets likely to arise from a \(b\) quark is detailed further on. Figure [\[fig:mass_lep\]](#fig:mass_lep){reference-type="ref" reference="fig:mass_lep"} shows the distributions in top candidate mass on the leptonic side for three situations: when the leptonic  is (correctly) combined with the \(b\) jet from leptonic top decay (\(P_{t:l}\)), when it is (wrongly) combined with the hadronic \(b\) jet (\(P_{t:h}\)), and when it is (wrongly) combined with a jet from hadronic  decay (\(P_{t:q}\)). Using the distinct shape of a presumably "incorrect" assignment means we need to keep track of two types of assignments which may disagree. We will introduce notation for the assignment used to combine the jets into the mass observables and for the assignment hypothesized to be correct. Depending on which jet is lost and which jet is picked to form the leptonic top candidate there are four possible two-jet combinations for the proxy side. The probability distributions for the invariant mass on the proxy side are shown in Fig. [\[fig:mass_had\]](#fig:mass_had){reference-type="ref" reference="fig:mass_had"} for hadronic and leptonic \(b\) jets (\(P_{p:hl}\)), leptonic \(b\) jet and a jet from  decay(\(P_{p:lq}\)), hadronic \(b\) jet and a jet from  decay(\(P_{p:hq}\)), and both jets from  decay(\(P_{p:qq}\)). The first two combinations are incorrect, as they include the leptonic \(b\) jet. The last two combinations are correct, and under the assumption that the leptonic \(b\) jet was reconstructed, they cannot both be available in the same event. These shapes can be used to maximize the probability \(P\) of selecting the correct assignment \(a\) given the data \(d\), which according to Bayes' theorem is: \[P\left(a\mid d\right) = \frac{P\left(d\mid a\right)P\left(a\right)}{\sum\limits_b P\left(d\mid b\right)P\left(b\right)} = \frac{P\left(d\mid a\right)}{\sum\limits_b P\left(d\mid b\right)}, \label{eq:prob}\] where \(b\) is any assignment and the second equality uses the fact that a priori all assignments are equally probable. There are three possible jet assignments per event (\(i=1,2,3\)), corresponding to the choice of the candidate for the leptonic \(b\) jet. Each event is characterized by three possible masses on the leptonic side (\(t_1,t_2,t_3\)) and three possible masses on the proxy side (\(p_1,p_2,p_3\)). In addition to this kinematic information, \(b\)-tagging algorithms  can also help to identify the origins of the jets. The results of the \(b\)-tagging algorithms can usually be expressed as a single continuous variable per jet, which discriminates between light and \(b\)-flavored jets. We label the \(b\)-tagging discriminant for the \(i\)-th jet as \(b_i\). Thus, data are presented by nine variables: \[d=(t_1,t_2,t_3;p_1,p_2,p_3;b_1,b_2,b_3)\] In matchable events the lost jet is either the hadronic \(b\) jet or a jet from hadronic  decay. We label the former as \(Q=b_l qq\) and the latter \(H=b_lb_h q\). For a matchable event, the probability for assignment \(a\) is a weighted sum of the probabilities of \(H\) and \(Q\) types: \[P\left(d\mid a\right)=(1-f_Q) P\left(d\mid a,H\right)+f_Q P\left(d\mid a,Q\right),\] where \(f_Q\) is the fraction of matchable events that are type \(Q\), which in our study scenario is \(20.5(2)\%\). Each jet assignment hypothesis specifies the type of each jet: either a \(b\) jet, or a jet from hadronic  decay. The latter category includes jets that arise from \(c\) quarks, and are somewhat similar to \(b\) jets . The correlations between the \(b\)-tagging discriminants (\(b_j\)) are small. Furthermore, these correlations are mostly independent of the true jet flavors, hence they are irrelevant for our purposes. Thus, the \(b\)-tagging probabilities can be factorized: \[\begin{aligned} P\left(d\mid a, C\right) &= P\left(t_1,t_2,t_3;p_1,p_2,p_3\mid a, C\right)P\left(b_1,b_2,b_3\mid a, C\right)\\ &= P\left(t_1,t_2,t_3;p_1,p_2,p_3\mid a, C\right)\prod_{j=1}^3 P\left(b_j\mid a,C\right) \label{eq:b_factorized} \end{aligned}\] where \(C=H\) or \(Q\) is the hypothesized class of the event. By neglecting the correlations between the remaining variables we can factorize the first two terms into six of the one-dimensional distributions shown in Figs. [\[fig:mass_lep\]](#fig:mass_lep){reference-type="ref" reference="fig:mass_lep"} and [\[fig:mass_had\]](#fig:mass_had){reference-type="ref" reference="fig:mass_had"} (\(P_{t:y}\) and \(P_{p:y}\)): \[P\left(d\mid a, C\right) = \prod_{j=1}^3 P_{t:f\left(j,a,C\right)}\prod_{j=1}^3 P_{p:g\left(j,a,C\right)}\prod_{j=1}^3 P\left(b_j\mid a,C\right) \label{eq:factorized}\] where \(f\left(j,a,C\right)\in\left\{l,h,q\right\}\) gives the type of the \(j\)-th jet (, the jet assumed to be the leptonic \(b\) jet when building the \(t_j\) observable) according to assignment \(a\) and event class \(C\), and \(g\left(j,a,C\right)\in\left\{hq,lq,hl,qq\right\}\) gives the types of the non-\(j\)-th jets (, the jets combined to form the proxy for the \(p_j\) observable) according to \(a\) and \(C\). Though we neglected some of the correlations between the observables in Eq. [\[eq:factorized\]](#eq:factorized){reference-type="ref" reference="eq:factorized"}, the structure of the likelihood preserves the dominant correlations, such as having at most one  resonance, and the correlation between the presence of a  resonance and the \(b\)-tagging variables. Using the described algorithm, the correct jet assignment is chosen for \(69.1(2)\%\) of the matchable events, which is to be compared to \(66.0(1)\%\) of correct assignments using a simple \(\chi^2\) method discussed in Section [3.2.1](#sec:chi2j){reference-type="ref" reference="sec:chi2j"}. Returning to the example of Fig. [\[fig:cartoon\]](#fig:cartoon){reference-type="ref" reference="fig:cartoon"}, the following terms help identify the correct event class (\(H\)) and assignment (\(a=3\), i.e. \(j_3\) is the leptonic \(b\) jet): - the invariant mass formed by combining the leptonic  candidate (\(W_l\)) and the jet \(j_1\), \(t_1=m(W_l+j_1)\), should be consistent with the \(P_{t:h}\) distribution from Fig. [\[fig:mass_lep\]](#fig:mass_lep){reference-type="ref" reference="fig:mass_lep"}; - \(t_2=m(W_l+j_2)\) should be consistent with \(P_{t:q}\) (same figure); - \(t_3=m(W_l+j_3)\) should be consistent with \(P_{t:l}\) (same figure); - the invariant mass formed by the jets \(j_2\) and \(j_3\), \(p_1=m(j_2+j_3)\), should be consistent with the \(P_{p:lq}\) distribution from Fig. [\[fig:mass_had\]](#fig:mass_had){reference-type="ref" reference="fig:mass_had"}; - \(p_2=m(j_1+j_3)\), invariant mass of leptonic \(b\) jet and a light jet should be consistent with \(P_{p:hl}\) (same figure); - \(p_3=m(j_1+j_2)\), invariant mass of leptonic and hadronic \(b\) jets should be consistent with \(P_{p:hq}\) (same figure); - \(b_1\), the \(b\)-tagging discriminant of \(j_1\), should be consistent with the distribution for a \(b\) jet; - \(b_2\) should be consistent with the distribution for a jet from hadronic  decay; - \(b_3\) should be consistent with the distribution for a \(b\) jet. The inclusion of the rarer \(Q\) events in the likelihood can distort the reconstruction of the more common case, the \(H\) events. But this risk is mitigated when the likelihood contains enough information to distinguish between the two cases on an event-by-event basis. To demonstrate that, we calculate the a posteriori probability that a matchable event is of type \(Q\) as: \[P_Q=\frac{f_Q P\left(d\mid a,Q\right)}{(1-f_Q) P\left(d\mid a,H\right)+f_Q P\left(d\mid a,Q\right)}\] As Fig. [\[fig:fQ\]](#fig:fQ){reference-type="ref" reference="fig:fQ"} demonstrates the separation between the two cases is quite good. This separation is mostly due to the \(b\)-tagging discriminants. It is also useful to check the modeling of \(P_Q\) against collider data, as all the terms in \(P\left(d\mid a\right)\) also appear in \(P_Q\). ### Scaling the proxy {#sec:alpha} Given a specific jet to quark assignment we have a candidate for the leptonic top \(t\) with the energy \(E_t\), momentum \(\vec{P_t}\) and invariant mass \(m_t=\sqrt{E_t^2-\vec{P_t}^2}\) and a proxy \(p\) for the hadronic top with the energy \(E_p\), momentum \(\vec{P_p}\) and invariant mass \(m_p=\sqrt{E_p^2-\vec{P_p}^2}\). Since the proxy tends to underestimate the 4-vector of the hadronic top quark, the invariant mass of these two objects, \(m(t+p)\), is likely to underestimate the generated invariant mass of the  system, , as shown in Fig. [\[fig:target\]](#fig:target){reference-type="ref" reference="fig:target"}. Additional scaling can be applied to the proxy 4-vector to partially correct for this underestimation. Furthermore, since the reconstructed proxy mass, , indicates the size of the underestimation in each event, this scaling can be parametrized as a function of . For each simulated event, we define the ideal scaling of the proxy 4-vector, \(\alpha\), as the scale that will bring the reconstructed \(m(t+p)\) to the peak position[^3] of the reconstructed mass, (see Fig. [\[fig:target\]](#fig:target){reference-type="ref" reference="fig:target"}). Since  is a function of , this scale is unavailable in collider data. Instead, we reconstruct events using a scale  which is an estimate of \(\alpha\) based on the observable . To derive this estimate, we solve for \(\alpha\) in simulated events, which results in a quadratic equation: \[\alpha^2m_p^2+2\alpha\left(E_tE_p-\vec{P_t}\vec{P_p}\right)+\left(m_t^2-\mpeak^2\right)=0.\] We then plot, in Fig. [\[fig:alpha_fit\]](#fig:alpha_fit){reference-type="ref" reference="fig:alpha_fit"}, the two-dimensional distribution of the proxy mass scaled by \(\alpha(\gmtt)\) and the unscaled . From this distribution we parametrize the most probable value of \(\alpha\) as a function of  to find our estimated . The parametrization of \(\ahat\left(\mproxy\right)\) was chosen from polynomial functions that were constrained so that the scaled mass, \(\ahat\mproxy\), is non-decreasing [^4]. Finally, we construct the invariant mass of the  system from the sum of the 4-vector of the proxy, scaled by \(\ahat\left(\mproxy\right)\), and the 4-vector of the leptonic top candidate. ### Averaging the assignments {#sec:averaging} The most significant improvement is from considering more than one jet assignment. The algorithms described so far considered only the most likely assignment, the one that minimizes the  in Eq. [\[eq:chisq\]](#eq:chisq){reference-type="ref" reference="eq:chisq"} or that maximizes \(P\left(a\mid d\right)\) in Eq. [\[eq:prob\]](#eq:prob){reference-type="ref" reference="eq:prob"}. But we can also use all the possible assignments weighted by their a posteriori probabilities. For example: \[\amtt=\sum\limits_a m_{t\bar{t}}^a P\left(a\mid d\right).\] These averaged reconstructions tend to have the advantage of a spread lower than that of the single-assignment reconstructions, and the disadvantage of a lower response. Here we define the "response" for an observable as the derivative of the average reconstructed value as a function of the true, generated value and the "spread" as the RMS of the distribution of the reconstructed value for a fixed true, generated value. # Performance {#sec:performance} ## Definition of the figure of merit {#sec:FOM} To compare the performance of different reconstruction algorithms, we require an appropriate figure of merit. Algorithm performance is usually quantified by summarizing the distribution of the difference (or the ratio) between the reconstructed and generated observable into its RMS, or into the width of a Gaussian fit to the core of the distribution. However, this quantification presumes that the reconstruction is unbiased and centered around the true value. For the reconstruction algorithms discussed here[^5] the difference distributions are intrinsically bimodal, since the performance differs for matchable and unmatchable events. For matchable events, the reconstruction typically has a response that is close to one and a narrow spread, while for the unmatchable events it typically has a low response and a wide spread. Hence the average reconstruction is biased, while the peak position is almost unbiased, and the reconstruction can not be calibrated so it is both unbiased and peaks at the generated value. To quantify the quality of the reconstruction without relying on the properties of its calibration, we contrast the reconstructed observable for two categories of events, defined by the quantiles of the generated observable. This is demonstrated in Fig. [\[fig:FOM\]](#fig:FOM){reference-type="ref" reference="fig:FOM"}. Each category contains 10% of the events, and they are defined according to an offset, \(s\), so that one category is generated between the \(s\) and \(s+0.1\) quantiles and the other between the \(0.9-s\) and \(1-s\) quantiles (see Fig. [\[subfig:fom_a\]](#subfig:fom_a){reference-type="ref" reference="subfig:fom_a"} where the 2nd and 9th deciles are used). The FOM quantifies how well the reconstruction separates these two categories. We denote the distributions of the reconstructed observable for these categories \(f_L\) and \(f_H\). An example is shown in Fig. [\[subfig:fom_b\]](#subfig:fom_b){reference-type="ref" reference="subfig:fom_b"}. Were these distributions Gaussian and identical, it would be natural to quantify the separation in terms of , the number of standard deviations between their peaks. To generalize this concept to arbitrary distributions and to focus on the possible misclassification of events between the two categories, we define \(T(x)\) as the overlap between these distributions at observable value \(x\) and the minimal overlap \(M\): \[\begin{aligned} M = \min_{x} T(x), && T(x) = \max\left( \int_x^{+\infty} f_L(x')\dd x', \int_{-\infty}^x f_H(x')\dd x' \right). \end{aligned}\] These too are shown in Fig. [\[subfig:fom_b\]](#subfig:fom_b){reference-type="ref" reference="subfig:fom_b"}. Smaller \(M\) values indicate less misclassification and hence better performance of the reconstruction algorithm. We can translate \(M\) to the more familiar "number of \(\sigma\)s" by considering \(M\) for two Gaussian distributions of width one, whose means are separated by : \[M\left(\Nsigma\right) = \int_{\frac{1}{2}\Nsigma}^{\infty} G(x)\dd x = \frac{1}{2}\left(1-\erf\left(\frac{\Nsigma}{2\sqrt{2}}\right)\right),\] where \(G\) is the normal distribution (see Fig. [\[subfig:sas_a\]](#subfig:sas_a){reference-type="ref" reference="subfig:sas_a"}). By inverting this relationship (see Fig. [\[subfig:sas_b\]](#subfig:sas_b){reference-type="ref" reference="subfig:sas_b"}), we can present the minimal overlap in terms of . This FOM has another, incidental advantage. Unlike RMS values, it can be interpreted without referring to the width and shape of the expected generated distribution. ## Comparison of the algorithms {#sec:comparison} Figures [\[fig:mtt_res\]](#fig:mtt_res){reference-type="ref" reference="fig:mtt_res"} and [\[fig:yres\]](#fig:yres){reference-type="ref" reference="fig:yres"} compare the reconstruction of different classes of events with the new algorithm. For ease of display, a rough linear calibration of  is used when displaying the resolutions of the partial reconstruction algorithm. Both classes of matchable events (case \(H\) and case \(Q\)) are reconstructed well, and the reconstruction of unmatchable events is not much worse. As \(76(1)\%\) of the events are matchable, the reconstruction for all events is almost as good as for matchable events. The reconstruction of the hadronic-top rapidity is especially weak for events of type \(Q\), indicating that a missing "hadronic" \(b\) jet is more problematic than a missing jet from  decay. The reconstruction of the leptonic-top rapidity is especially weak for unmatchable events, since for most of these events the "leptonic" \(b\) jet is lost. No partial reconstruction algorithm was previously applied to \(\ttbar\to\lpj\) events, so we choose to compare the performance of the algorithm described in this  to that of a kinematic fit algorithm that was used to fully reconstruct  events  in many top measurements (e.g. in Refs.  and ). As with the new algorithm, we can either use the most likely assignment from the kinematic fit algorithm or use a weighted average of all assignments. The relative weight of each assignment is \(\exp\left(-\chisq/2\right)\), as in Ref. . We compare the performance of the two algorithms for the ability to reconstruct the following observable: the invariant mass of the  system (), the rapidity of the leptonically decaying top quark (\(y_l\)), the rapidity of the hadronically decaying top quark (\(y_h\)) and the rapidity difference (\(=y_l-y_h\)). The distributions of the differences and ratio between reconstructed and generated observables for these two algorithms, shown in Fig. [\[fig:comp_3to4\]](#fig:comp_3to4){reference-type="ref" reference="fig:comp_3to4"}, illustrate that the partial reconstruction provides a performance similar in quality to that of the full reconstruction. Table [1](#tab:perf){reference-type="ref" reference="tab:perf"} uses the FOM introduced in Section [4.1](#sec:FOM){reference-type="ref" reference="sec:FOM"} to quantitatively compare the performance of the two algorithms. As the generated distributions differ between the  and the  samples, there is some arbitrariness in such a comparison. To quantify this arbitrariness, for the  samples each FOM was evaluated twice, once using the quantiles found in the  sample and once using the quantiles found in the  sample. Table [1](#tab:perf){reference-type="ref" reference="tab:perf"} also lists the performance of simpler versions of the new algorithm, corresponding to Sections [3.2.1](#sec:chi2j){reference-type="ref" reference="sec:chi2j"}, [3.2.2](#sec:lhoodj){reference-type="ref" reference="sec:lhoodj"}, [3.2.3](#sec:alpha){reference-type="ref" reference="sec:alpha"}, and [3.2.4](#sec:averaging){reference-type="ref" reference="sec:averaging"}. A constant offset, \(s\), was chosen for each observable (, \(y_l\), \(y_h\) and ). The offsets were chosen so the resulting  values are \(\approx 2\), a level of separation where further improvements are still useful (see Fig. [\[subfig:fom_b\]](#subfig:fom_b){reference-type="ref" reference="subfig:fom_b"}). Though the tail behavior of the reconstructions varies, the variations are limited to a fraction of events much smaller than the 10% we consider in each category. Thus the choice of offsets has little effect on the comparison of reconstruction techniques. We find that the partial reconstruction of  and  in  sample is fully competitive with that of the full reconstruction in the  events. ::: The  channel has the obvious disadvantage of missing a jet. On the other hand, it has the advantage of fewer jets from initial state radiation, and for the algorithm outlined here, of fewer unmatchable events. These advantages compensate quite well for the missing jet. It may be that the reconstruction of  can be improved by considering additional reconstruction hypotheses, in particular, events where one jet is lost and a jet from initial state radiation was selected. # Summary {#sec:summary} We present an algorithm that partially reconstructs  events in the  channel in the case when one of the jets is lost, resulting in a  topology. Probabilities for correct and incorrect jet assignment are formed based on \(b\)-tagging discriminants and on all possible mass combinations on the leptonic and hadronic sides. The algorithm can be applied to measure the forward-backward asymmetry in  production, the invariant mass spectrum of the  system and for a number of other analyses that require a full reconstruction. The performance of the partial reconstruction algorithm is competitive with that commonly achieved for fully reconstructed  events. The inclusion of  events can improve the statistical strength and reduce the systematic uncertainties of a top properties measurement. Gains equivalent to having 50% more data were achieved at the Tevatron .

{'timestamp': '2015-05-01T02:11:18', 'yymm': '1310', 'arxiv_id': '1310.3263', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3263'}

# Introduction {#skorik-skorik_section_1} Systems with controls appearing linearly are most close to linear systems. Such systems are well studied and various methods are developed, namely, differential-geometric methods, algebraic methods, and those commonly used for linear systems. In particular, the important role is played by the feedback linearization method. In this paper we consider systems for which just non-linearity with respect to a control allows to solve the controllability problem. Namely, we consider a class of systems which are equivalent to systems of differential equations with one dimensional control \[\label{skorik-intr_1} y_i^{(n_i)}=H_i(y_1,\ldots, y_1^{(n_1-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)},u),\quad i=1,\ldots, m,\quad u\in{\Bbb R},\] where \(y_i^{(s)}\) means the derivative of order \(s\) and the functions \(H_1,\) \(\ldots,\) \(H_m\) are non-linear with respect to \(u.\) The basic idea of our approach consists in the following. We solve the problem of controllability to a rest point of the system ([\[skorik-intr_1\]](#skorik-intr_1){reference-type="ref" reference="skorik-intr_1"}) step by step. On the first step we construct a positional control which depends on all state variables, i.e. a control of the form \[u=u_1\left(y_1,\ldots, y_1^{(n_1-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)}\right),\] transferring state coordinates of the first equation to the rest point in certain finite time \(T_1,\) i.e. \(y_1(T_1)=\ldots=y_1^{(n_1-1)}(T_1)=0.\) On the second step we construct a positional control \[u=u_2\left(y_2,\ldots, y_2^{(n_2-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)}\right)\] which transfers state variables of the second equation to the rest point in certain finite time \(T_2-T_1,\) i.e. \(y_2(T_2)=\ldots=y_2^{(n_2-1)}(T_2)=0,\) and keeps coordinates \(y_1,\) \(\ldots,\) \(y_1^{(n_1-1)}\) at the rest point, i.e. \(y_1(t)=\ldots=y_1^{(n_1-1)}(t)=0\) for \(T_1\le t\le T_2.\) This can be done only in the case when \(H_1\) depends on a control non-linearly. Analogously, on the \(i\)-th step we construct a positional control \[u=u_i\left(y_i,\ldots, y_i^{(n_i-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)}\right)\] which transfers state variables of the \(i\)-th equation to the rest point in certain finite time \(T_i-T_{i-1}\) and keeps coordinates \(y_1,\) \(\ldots,\) \(y_1^{(n_1-1)},\) \(\ldots,\) \(y_{i-1},\) \(\ldots,\) \(y_{i-1}^{(n_{i-1}-1)}\) at the rest point, i.e. \(y_1(t)=\ldots=y_1^{(n_1-1)}(t)=\ldots=y_{i-1}(t)=\ldots=y_{i-1}^{(n_{i-1}-1)}(t)= 0\) as \(T_{i-1}\le t\le T_i,\) and so on. After \(m\) such steps we obtain the control of the form \[\begin{array}{l} u(y_1,\ldots, y_1^{(n_1-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)};t)=\\ =u_i\left(y_i,\ldots, y_i^{(n_i-1)},\ldots, y_m,\ldots, y_m^{(n_m-1)}\right), \;\; t\in[T_{i-1},T_i], \; i=1...,m\;\; (T_0=0). \end{array}\] This control transfers the initial point to the rest point in the time \(T=T_m.\) Thus, on each step we choose a new positional control solving the positional synthesis problem. As a result of our approach, we construct a programming control which is a concatenation of a finite number of the positional controls. *We call it a "stepwise synthesis control"* which transfers an arbitrary initial point to the rest point in a certain finite time \(T.\) Let us explain our construction by the following example. Consider the system \[\label{skorik-r1_pr0_p1} \dot y_1=\sin u,\quad \dot y_2 = u\cos 2 u.\] This system has the form ([\[skorik-intr_1\]](#skorik-intr_1){reference-type="ref" reference="skorik-intr_1"}), where \(n_1=n_2=1.\) Note that this system is not controllable with respect to the first approximation. Suppose \((y_{10}, y_{20})\) is an arbitrary point. On the first step we choose the control \(u_1(y_1,y_2)=-(\pi/2){\rm sign}\, y_1.\) This control transfers the initial point to the point \(y(T_1)=(y_1(T_1), y_2(T_1))=(0,\pi y_{10}/2+y_{20})\) in the time \(T_1=|y_{10}|.\) On the second step we choose a control \(u_2(y_2)\) such that \(y_1(t)=0\) for \(t\ge T_1\) and \(y_2(T_2)=0\) for a certain finite \(T_2\ge T_1.\) This means that the corresponding trajectory \(y(t)=(y_1(t),y_2(t))\) of the system ([\[skorik-r1_pr0_p1\]](#skorik-r1_pr0_p1){reference-type="ref" reference="skorik-r1_pr0_p1"}) belongs to the subspace \(\{(y_1,y_2): y_1=0\}\) for \(t\in [T_1,T_2].\) This can be done by the control \(u_2(y_2)=-\pi\, {\rm sign}\, y_2.\) This control transfers the point \(y(T_1)\) to the origin in the time \((T_2-T_1)=|y_{10}/2+y_{20}/\pi|.\) Therefore, the point \((y_{10}, y_{20})\) is transferred to the origin by the control \[u(y_1,y_2;t)=\left\{\begin{array} {cll}-(\pi/2){\rm sign}\,y_1& \mbox{for}& 0\le t \le |y_{10}|,\\-\pi\, {\rm sign}\, y_2 & \mbox{for}& |y_{10}|<t\le |y_{10}|+|y_{10}/2+y_{20}/\pi|, \end{array}\right.\] along the trajectory of the system ([\[skorik-r1_pr0_p1\]](#skorik-r1_pr0_p1){reference-type="ref" reference="skorik-r1_pr0_p1"}) in the time \(T=|y_{10}|+|y_{10}/2+y_{20}/\pi|.\) Thus, we have a stepwise synthesis, i.e. on the segment \([0,T_1]\) we choose the position control \(-(\pi/2){\rm sign}\,y_1,\) and on the segment \([T_1, T_2]\) we choose another positional control \(-\pi\, {\rm sign}\, y_2.\) Note that the times \(T_1\) and \(T_2\) are not given in advance but depend on the initial point \((y_{10},y_{20}).\) In the paper we introduce a new class of nonlinear single input systems \[\label{skorik-r1_f2} \dot x =a(x)+\sum\limits_{i=1}^mb_i(x) \beta_i(x,u)\equiv a(x)+B(x) \beta(x,u),\quad x\in{\Bbb R}^n, \quad u\in {\Bbb R},\] where \(B(x)\) is a \((n{\times}m)\)-matrix \((2\le m\le n)\) with columns \(b_1(x),\) \(\ldots,\) \(b_m(x),\) \(\beta(x,u)\) is a \(m\)-dimensional vector-function with components \(\beta_1(x,u),\) \(\ldots,\) \(\beta_m(x,u),\) and \(u\) is a one-dimensional control. On the first glance the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) looks like an affine control system of the form \[\dot x=a(x)+B(x)u,\quad x\in{\Bbb R}^n,\; u\in {\Bbb R}^m.\label{skorik-affine_system}\] However, let us emphasize that in the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) the control \(u\) is only one-dimensional and, moreover, the nonlinearity of \(\beta(x,u)\) with respect to \(u\) plays the crucial role in our approach. Though an arbitrary single input nonlinear system \(\dot x=f(x,u)\) can be written in the form ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) in different ways, nevertheless, not every form is appropriate for the further analysis. Within our approach we deal with systems of the form ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) which can be mapped to systems of the form \[\label{skorik-r1_kns} \dot z=A_0z+B_0H(z,u),\quad z\in{\Bbb R}^n, u\in {\Bbb R},\] where \(A_0={\rm diag}(A_{1},\ldots,A_{m})\) is a constant \((n{\times}n)\)-matrix, \(B_0=(e_{s_1}, \ldots, e_{s_m})\) is a constant \((n{\times}m)\)-matrix (\(e_{s_i}\) is the \(s_i\)-th unit vector of the space \({\Bbb R}^n,\) \(i=1,\ldots, m\)), and \(H(z,u)\) is a \(m\)-dimensional vector function. The system ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}) is equivalent to the system ([\[skorik-intr_1\]](#skorik-intr_1){reference-type="ref" reference="skorik-intr_1"}). In Section [2](#skorik-skorik_mappability_of_nonlinear_systems){reference-type="ref" reference="skorik-skorik_mappability_of_nonlinear_systems"} we give conditions of the mappability of the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) on the system ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}). These conditions are similar to the linearizability conditions for affine systems ([\[skorik-affine_system\]](#skorik-affine_system){reference-type="ref" reference="skorik-affine_system"}). Notice that changes of variables can be used to increase the amount of rest points of considered systems (see Example [5.2](#skorik-example_3){reference-type="ref" reference="skorik-example_3"}). This is extremely important for our method. The problem of linearizability for affine systems is well studied and the conditions are well known--. However, generally these conditions are not easy for check. Therefore it is important to find classes of systems for which these conditions are automatically satisfied. The first such class of systems called "the class of triangular systems" was introduced in the paper, where the feedback linearization was given. In the paper global properties of the triangular systems in the singular case is considered. In the present paper we introduce the new class of nonlinear systems called "the class of staircase systems" which are mapped on the systems ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}) and give the corresponding changes of variables (Section [7](#skorik-skorik_section_7){reference-type="ref" reference="skorik-skorik_section_7"}). In Section [3](#skorik-skorik_controllability_on_a_subspace){reference-type="ref" reference="skorik-skorik_controllability_on_a_subspace"} we solve the problem of positional synthesis to a subspace for certain class of nonlinear systems. Our main tool is the controllability function method proposed in [@korobov1; @korobov_kniga] for solving the synthesis problem of admissible positional constrained control. Later it was developed for different classes of systems and different statements of the synthesis problem, for example, for infinite systems, for systems in a finite-dimensional space with constraint on a control and its derivatives which called inertial control in and so on. In Subsection [3.1](#skorik-skorik_application_of_the_controllability_function_method){reference-type="ref" reference="skorik-skorik_application_of_the_controllability_function_method"} we recall the application of the controllability function method for linear systems . The main result is given in Section [4](#skorik-skorik_main_result){reference-type="ref" reference="skorik-skorik_main_result"} (Theorem [\[skorik-skorik_theorem_2\]](#skorik-skorik_theorem_2){reference-type="ref" reference="skorik-skorik_theorem_2"}). Namely, we give conditions under which the application of the method of stepwise synthesis gives the solution of the controllability problem from an arbitrary point to the rest point of the system. The obtained results are illustrated by the examples in Section [5](#skorik-skorik_section_examples){reference-type="ref" reference="skorik-skorik_section_examples"}. In Section [6](#skorik-calming_of_vibrations_of_a_two-link_pendulum){reference-type="ref" reference="skorik-calming_of_vibrations_of_a_two-link_pendulum"} the problem of complete stoppage of a two-link pendulum is solved. # Mappability of nonlinear systems on nonlinear systems of a special form {#skorik-skorik_mappability_of_nonlinear_systems} We consider the problem of \(0\)-controllability for the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}). Suppose \(a(x),\) \(b_1(x),\) \(\ldots,\) \(b_m(x)\) are a \(n\) times continuously differentiable vector functions, \(\beta_1(x,u),\) \(\ldots,\) \(\beta_m(x,u)\) are continuously differentiable scalar functions with respect to \(x,\) \(u,\) and \[\label{skorik-r1_f3} a(0) = 0, \quad \beta(0,0) = 0.\] In this section we give sufficient conditions under which system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) is mapped on a system of the form ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}). Below we use the following standard notations: for a scalar continuously differentiable function \(\varphi(x) = \varphi(x_1,\ldots,x_n)\), denote by \(L_a\varphi\) the derivative of the function \(\varphi(x)\) along the vector field \(a(x)\), i.e. \(L_a\varphi(x) = \varphi_x(x)a(x),\) where \(\varphi_x(x) = \left(\varphi_{x_1}(x),\ldots,\varphi_{x_n}(x)\right).\) By \([a(x),b(x)]\) denote the Lie bracket of the vector fields \(a(x)\) and \(b(x),\) i.e. \([a(x),b(x)]=b_x(x)a(x)-a_x(x)b(x),\) where \(a_x(x),\) \(b_x(x)\) are matrices of the first derivatives of vector-functions \(a(x),\) \(b(x).\) Also put \({\rm ad}_a^0 b(x) =b(x),\) \({\rm ad}_a^k b(x)= [a(x),{\rm ad}_a^{k-1} b(x)],\) \(k\ge 1.\) Suppose for system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) the condition \[\label{skorik-r1_rang} {\rm rang}\,Q(x)=n\;\;\mbox{for all}\;\;x\in {\Bbb R}^n,\] holds, where \(Q(x)=(b_1(x), \ldots, b_m(x),\ldots,{\rm ad}_a^{n-1} b_1(x),\ldots,{\rm ad}_a^{n-1} b_m(x)).\) By \(q_i(x),\) \(i=1,\ldots,nm,\) denote the columns of the matrix \(Q(x).\) Without loss of generality assume \({\rm rang}\,B(x)=m\) for all \(x\in {\Bbb R}^n.\) Moreover, since we are interested in the global \(0\)-controllability we require that the vector fields \(a(x),\) \(b_1(x),\) \(\ldots,\) \(b_m(x)\) satisfy the following regularity property: for all \(j=1,\ldots, nm\) \[\label{skorik-blqvblycb} {\rm rang}(q_1(x),\ldots, q_j(x))=c_j \;\; \mbox{for all} \;\; x\in{\Bbb R}^n,\] where \(c_j\) are certain constants, \(1\le c_j\le n.\) Now we delete all columns of the matrix \(Q(x)\) that linearly depend on previous ones, i.e. columns such that \(q_i(x)\in {\rm Lin} \{q_1(x),\ldots,q_{i-1}(x)\}.\) It is convenient to examine the columns \(q_i(x),\) \(i=1,\ldots,nm\), one by one from left to right and take into account the following remark: if the column \(q_i(x)={\rm ad }_a^kb_j(x)\) is deleted then all columns of the form \(q_{i+ms}(x)={\rm ad }_a^{k+s}b_j(x)\) for all \(s\ge 1\) such that \(i+ms\le nm\) should be deleted as well. This algorithm is the same as the algorithm for linear controllable systems with a multidimensional control given in the paper and is analogous to the algorithm given in for linearization of affine systems with multidimensional control. As a result, we obtain the matrix consisting of the columns of \(Q(x)\) which are not deleted. It is convenient to permutate these columns and deal with the matrix \(K(x)\) of the form \[\label{skorik-r1_m_K} K(x){=}\left( b_1(x),\ldots,{\rm ad}_a^{n_1-1}b_1(x),\ldots, b_m(x), \ldots, {\rm ad}_a^{n_m-1}b_m(x)\right),\] where \(n_1+\ldots+n_m=n\) and \({\rm rang}\,K(x)=n\) for all \(x\in {\Bbb R}^n.\) Our main assumption is as follows: suppose there exist scalar functions \(\varphi_1(x),\) \(\ldots,\) \(\varphi_m(x)\), which are no less than *twice continuously differentiable* such that: (a) for each \(i=1,\ldots,m\) the conditions \[\label{skorik-r1_f4} \left\{\begin{array}{l} \left(\varphi_i(x)\right)_x{\rm ad}_a^k b_j(x) =0,\;\; k= 0,\ldots,\min\{n_i-2,n_j-1\},\;\;j=1,\ldots,m,\\ \left(\varphi_i(x)\right)_x{\rm ad}_a^{n_i-1} b_i(x) \ne 0, \;\; x\in{\Bbb R}^n,\\\varphi_i(0)=0, \end{array}\right.\] are satisfied; (b) the change of variables \(z=L(x)\in C^{(2)}({\Bbb R}^n)\) of the form \[\label{skorik-r1_f5} z_{s_{i-1}+j} = L_a^{j-1}\varphi_i(x), \quad j= 1,\ldots,n_i, \; i=1,\ldots,m,\] is non-singular, i.e. \({\rm det}\,L_x(x)\ne 0\) for all \(x\in {\Bbb R}^n,\) where \(s_0=0,\) \(s_i=n_1{+}\ldots{+}n_i,\) \(i=1,\ldots,r.\) Notice that the conditions of solvability of the system ([\[skorik-r1_f4\]](#skorik-r1_f4){reference-type="ref" reference="skorik-r1_f4"}) in the class of once continuously differentiable functions are well known, however, for our aim this is not sufficient. Then in these variables the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) takes the form \[\label{skorik-r1_fs} \dot z_{s_{i-1}+1}=z_{s_{i-1}+2},\; \ldots,\;\dot z_{s_i-1}=z_{s_i},\; \dot z_{s_i}=H_i(z,u), \quad i=1,\ldots,m,\] where \[\label{skorik-r1_H_i} H_i(z,u) =L_a^{n_i}\varphi_i\left(L^{-1}(z)\right)+\sum\limits_{k=1}^m\beta_k\left(L^{-1}(z),u\right) L_{b_k}L_a^{n_i-1}\varphi_i\left(L^{-1}z\right).\] For \(i=1,\ldots,m\) denote \(z^i=(z_{s_{i-1}+1},\ldots, z_{s_i})^*\) (the sign \* means the transposition). Then the system ([\[skorik-r1_fs\]](#skorik-r1_fs){reference-type="ref" reference="skorik-r1_fs"}) can be rewritten as \[\label{skorik-r1_kns_i} \dot z^i=A_iz^i+b_{0i}H_i(z,u),\quad z^i\in{\Bbb R}^{n_i},\quad i=1,\ldots,m,\] where \[A_{i}=\left(\begin{array}{cccccc} 0&1&0&\ldots&0&0\\ \ldots&\ldots&\ldots&\ldots&\ldots&\ldots\\ 0&0&0&\ldots&0&1\\ 0&0&0&\ldots&0&0\\ \end{array}\right),\quad b_{0i}=\left(\begin{array}{c} 0\\ \ldots\\0\\1\\\end{array}\right).\] The system ([\[skorik-r1_kns_i\]](#skorik-r1_kns_i){reference-type="ref" reference="skorik-r1_kns_i"}) obviously can be written in the form ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}). # Controllability to a subspace with respect to a part of variables {#skorik-skorik_controllability_on_a_subspace} In this section we construct a control which transfers any initial point to a subspace. This is done by use of the controllability function method. ## Application of the controllability function method for linear system {#skorik-skorik_application_of_the_controllability_function_method} The controllability function method gives a general approach for solving the problem of synthesis of positional constrained controls. We briefly recall the main ideas of this method. Consider the system \[\dot x=A_0x+b_0v,\quad x\in{\Bbb R}^{k_1},\;v\in{\Bbb R},\] with the constraint on a control of the form \(|v|\le d,\) where \(d>0\) is a given number and \[A_0=\left(\begin{array}{cccccc} 0&1&0&\ldots&0&0\\ \ldots&\ldots&\ldots&\ldots&\ldots&\ldots\\ 0&0&0&\ldots&0&1\\ 0&0&0&\ldots&0&0\\ \end{array}\right),\quad b_{0}=\left(\begin{array}{c} 0\\ \ldots\\0\\1\\\end{array}\right).\] Consider a nonsingular \((k_1{\times}k_1)\)-matrix \(N(\Theta) = \int\limits_0^\Theta\left(1-{\frac{t}{\Theta}}\right) e^{-A_0t}b_0b_0^*e^{-A_0^*t}dt\). Suppose that the number \(a_0\) satisfies the condition \[\label{skorik-uslovie_na_a_0} 0<a_0\le 2 d^2/ (N^{-1}(1)b_0,b_0).\] Define the controllability function \(\Theta(x)\) at \(x\ne 0\) as the unique positive solution of the equation \[\label{skorik-skoryk-r1_f13} 2a_0\Theta=\left(N^{-1}(\Theta)x,x\right)\] and put \(\Theta(0)=0.\) Then the function \(\Theta(x)\) is continuous and continuously differentiable for \(x\ne 0.\) Choose a control \(v=v(x)\) in the form \[\label{skorik-skoryk-r1_f14} v(x) =-\frac{1}{2}\, b_0^*N^{-1}(\Theta (x))x, \quad x \ne 0.\] It can be shown that this control \(v(x)\) satisfies the Lipschitz condition in each domain \(K(\rho_1,\rho_2)=\{x: 0<\rho_1\le \|x\|\le \rho_2\}\) with a Lipschitz constant \(L_v(\rho_1,\rho_2)\) such that \(L_v(\rho_1,\rho_2)\to+\infty\) as \(\rho_1\to 0.\) Put \(y(\Theta,x)=D(\Theta)x,\) where \(D(\Theta)= {\rm diag}\,\left(\Theta^{-\frac{2k_1-2j+1}{2}}\right)_{j=1}^{k_1}.\) Rewrite the control ([\[skorik-skoryk-r1_f14\]](#skorik-skoryk-r1_f14){reference-type="ref" reference="skorik-skoryk-r1_f14"}) in the form \(v(x) = a y(\Theta(x),x) \Theta^{-\frac{1}{2}}(x),\) where \(a=-\frac{1}{2}b_0^*N^{-1}(1).\) Let us show that the control satisfies the given constraint for any \(x\in {\Bbb R}^{k_1}.\) To this aim, for a fixed \(\Theta\) let us consider the extremal problem \[a y(\Theta,x) \Theta^{-\frac{1}{2}}\to {\rm extr},\quad (N^{-1}(1)y(\Theta,x),y(\Theta,x))-2a_0 \Theta=0.\] Using the Lagrange method we get \(y^0=\frac{1}{2\lambda} \Theta^{-\frac{1}{2}}N(1)a^*\) for an extremum point \(y^0.\) Since \(\frac{1}{2\lambda}=\sqrt{2a_0/(N(1)a^*,a^*)}\; \Theta\) then \(ay^0 \Theta^{-\frac{1}{2}}=\pm\sqrt{a_0(N^{-1}(1)b_0,b_0)/2}.\) Hence, the condition ([\[skorik-uslovie_na_a\_0\]](#skorik-uslovie_na_a_0){reference-type="ref" reference="skorik-uslovie_na_a_0"}) implies that the control ([\[skorik-skoryk-r1_f14\]](#skorik-skoryk-r1_f14){reference-type="ref" reference="skorik-skoryk-r1_f14"}) satisfies the constraint \(|v(x)| \le d\) for any \(x\in {\Bbb R}^{k_1}.\) Let us calculate the derivative of the controllability function by virtue of the system \[\label{skorik-skoryk-r1_zks} \dot x=A_{0}x+b_0v(x),\quad x\in{\Bbb R}^{k_1}.\] Substituting \(\Theta=\Theta (x)\) to ([\[skorik-skoryk-r1_f13\]](#skorik-skoryk-r1_f13){reference-type="ref" reference="skorik-skoryk-r1_f13"}) and differentiating we obtain \[\label{skorik-skoryk-r1_dottheta_1} \begin{array}{l} 2a_{0}\dot\Theta=-\left(N^{-1}(\Theta)\widetilde N(\Theta)N^{-1}(\Theta)x,x\right)\dot\Theta+\left((N^{-1}(\Theta)A_{0}+\right.\\[5pt] \qquad\qquad\;\;\left.+A_{0}^*N^{-1}(\Theta)x,x\right)-\left(N^{-1}(\Theta)b_0b_0^*N^{-1}(\Theta)x,x\right),\end{array}\] where \(\widetilde N(\Theta)=\frac{1}{\Theta^2}\int\limits_0^\Theta t e^{-A_{0}t}b_0b_0^*e^{-A_{0}^*t}dt.\) Since \[A_{0}N(\Theta)+N(\Theta)A_{0}^*=-\int\limits_0^\Theta\left(1-\frac{t}{\Theta}\right) d\left(e^{-A_{0}t}b_0b_0^*e^{-A_{0}^*t}\right)=b_0b_0^*-\widehat N(\Theta),\] where \(\widehat N(\Theta)=\frac{1}{\Theta}\int\limits_0^\Theta e^{-A_{0}t}b_0b_0^*e^{-A_{0}^*t}dt,\) hence, \[\label{skorik-skoryk-r1_dottheta_2} N^{-1}(\Theta)A_{0}+A_{0}^*N^{-1}(\Theta)=N^{-1}(\Theta)b_0b_0^*N^{-1}(\Theta)-N^{-1}(\Theta)\widehat N(\Theta)N^{-1}(\Theta).\] Then, denoting \(w=N^{-1}(\Theta)x,\) using ([\[skorik-skoryk-r1_f13\]](#skorik-skoryk-r1_f13){reference-type="ref" reference="skorik-skoryk-r1_f13"}), ([\[skorik-skoryk-r1_dottheta_2\]](#skorik-skoryk-r1_dottheta_2){reference-type="ref" reference="skorik-skoryk-r1_dottheta_2"}), ([\[skorik-skoryk-r1_dottheta_1\]](#skorik-skoryk-r1_dottheta_1){reference-type="ref" reference="skorik-skoryk-r1_dottheta_1"}), and taking into account the form of the matrices \(\widehat N(\Theta),\) \(\widetilde N(\Theta),\) \(N(\Theta)\) we get \[\dot{\Theta}(x)_{\bigl|(\ref{skorik-skoryk-r1_zks})} =-(\widehat N(\Theta)w,w)\Bigl/\left(\frac{1}{\Theta}( N(\Theta)w,w)+(\widetilde N(\Theta)w,w)\right)=-1.\] Thus, the time of motion \(T(x_0)\) from \(x_0\in {\Bbb R}^{k_1}\) to \(x_T=0\) equals \(\Theta(x_0),\) where \(\Theta(x_0)\) is the positive solution of the equation ([\[skorik-skoryk-r1_f13\]](#skorik-skoryk-r1_f13){reference-type="ref" reference="skorik-skoryk-r1_f13"}) at \(x=x_0\). ## Controllability to a subspace Solutions of all considered systems are understood in the sense of differential inclusions. At first, we consider the problem of controllability on a subspace with respect to a part of variables for the system \[\label{skorik-skoryk-skhg} \dot z=\left(\begin{array}{l} \dot x\\ \dot y\\ \end{array}\right)= \left(\begin{array}{c} A_0x+b_0 h(x,y,u)\\ g(x,y,u)\\ \end{array}\right),\quad z\in {\Bbb R}^k, x\in{\Bbb R}^{k_1}, y\in {\Bbb R}^{k_2}, u\in{\Bbb R},\] where \(h(z,u)=h(x,y,u)\) is a continuous scalar function, \(g(z,u)=g(x,y,u)\) is a continuous \(k_2\)-dimensional vector function which satisfy the Lipschitz condition with respect to \(z\) and \(u\) in each domain \(\{(z,u): 0<\rho_1\le \|z\|\le \rho_2, |u|\le \rho_3\}.\) Let us fix some number \(d>0.\) Choose \(a_0\) satisfying ([\[skorik-uslovie_na_a\_0\]](#skorik-uslovie_na_a_0){reference-type="ref" reference="skorik-uslovie_na_a_0"}) and define \(\Theta(x)\) as the unique positive solution of the equation ([\[skorik-skoryk-r1_f13\]](#skorik-skoryk-r1_f13){reference-type="ref" reference="skorik-skoryk-r1_f13"}) at \(x\ne 0\) and put \(\Theta (0)=0.\) Denote \(S^+=\{z\in{\Bbb R}^k: b_0^*N^{-1}(\Theta (x))x>0\},\) \(S^-=\{z\in{\Bbb R}^k: b_0^*N^{-1}(\Theta (x))x<0\},\) and \(S=\{z\in{\Bbb R}^k: b_0^*N^{-1}(\Theta (x))x=0\}.\) Further, we consider the problem of controllability to a subspace with respect to a part of variables for the system ([\[skorik-r1_kns_i\]](#skorik-r1_kns_i){reference-type="ref" reference="skorik-r1_kns_i"}). For any fixed \(i\in \{1,\ldots,m\}\) consider the nonsingular \((n_i{\times}n_i)\)-matrix \(N_i(\Theta) = \int\limits_0^\Theta\left(1-{\frac{t}{\Theta}}\right) e^{-A_{i}t}b_{0i}b_{0i}^*e^{-A_{i}^*t}dt\) and choose a number \(a_{0i}\) such that \(0<a_{0i}\le 2d_i^2/(N_i^{-1}(1)b_{0i},b_{0i})\) for a given \(d_i>0.\) Introduce the controllability function \(\Theta_i(z^i)\) as the unique positive solution of the equation \(2a_0\Theta=\left(N_i^{-1}(\Theta)z^i,z^i\right)\) at \(z^i\ne 0^i\) and put \(\Theta_i(0^i)=0.\) Lemma [\[skorik-skorik-lemma_1\]](#skorik-skorik-lemma_1){reference-type="ref" reference="skorik-skorik-lemma_1"} implies the following theorem. # Main result {#skorik-skorik_main_result} In this section we give sufficient conditions of \(0\)-controllability for system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) which is mapped on the system ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}) by the change of variables ([\[skorik-r1_f5\]](#skorik-r1_f5){reference-type="ref" reference="skorik-r1_f5"}). # Examples {#skorik-skorik_section_examples} In this section we give several examples illustrating Theorem  [\[skorik-skorik_theorem_2\]](#skorik-skorik_theorem_2){reference-type="ref" reference="skorik-skorik_theorem_2"}. ## Consider the system \[\label{skorik-r1_pr2_sys2} \dot x_1 = u^3+0.1\sin^2f_1(x_1,x_2,x_3,u), \quad \dot x_2 = u, \quad \dot x_3 = f_2(x_2),\quad |u|\le 2,\] where \(f_1\) and \(f_2\) are continuously differentiable functions such that \(f_1(0,0,0,0) = 0,\) \(f_2(0) = 0,\) \(\left| f_2'(x_2)\right|\ge \delta>0.\) We note that the system ([\[skorik-r1_pr2_sys2\]](#skorik-r1_pr2_sys2){reference-type="ref" reference="skorik-r1_pr2_sys2"}) is not controllable at the first approximation in a neighborhood of the stationary point \((x=0, u=0).\) We consider the \(0\)-controllability problem from any point \(x_0 =(x_{10},x_{20},x_{30})^*\) and construct a control transferring the point \(x_0\) to the origin. The system ([\[skorik-r1_pr2_sys2\]](#skorik-r1_pr2_sys2){reference-type="ref" reference="skorik-r1_pr2_sys2"}) can be rewritten in the form ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) with \(a(x)=(0,0,f_2(x_2))^*,\) \(a(0)=0,\) \(b_1(x)=(1,0,0)^*,\) \(b_2(x)=(0,1,0)^*,\) \(\beta_1(x,u)=u^3+0.1\sin^2f_1(x_1,x_2,x_3,u),\) \(\beta_2(x,u)=u,\) \(m=2.\) The matrix \(K(x)\) from ([\[skorik-r1_m\_K\]](#skorik-r1_m_K){reference-type="ref" reference="skorik-r1_m_K"}) has the form \[K(x)=(b_1(x), b_2(x),{\rm ad}_a b_2(x))=\left(\begin{array}{ccc} 1&0&0\\ 0&1&0\\ 0& 0&-f_2'(x_2)\end{array}\right),\quad {\rm rang}\, K(x)=3,\; x\in{\Bbb R}^3,\] hence, \(n_1=1,\) \(n_2=2.\) The conditions ([\[skorik-r1_f4\]](#skorik-r1_f4){reference-type="ref" reference="skorik-r1_f4"}) require that functions \(\varphi_1(x),\) \(\varphi_2(x)\) satisfy the condition \[\frac{\partial\varphi_1(x_1,x_2,x_3)}{\partial x_1}\ne 0;\; \frac{\partial\varphi_2(x_1,x_2,x_3)}{\partial x_1}= 0,\; \frac{\partial\varphi_2(x_1,x_2,x_3)}{\partial x_2}= 0,\; \frac{ \partial\varphi_2(x_1,x_2,x_3)}{\partial x_3} \ne 0.\] We choose \(\varphi_1(x_1,x_2,x_3)= x_1-x_2\) and \(\varphi_2(x_1,x_2,x_3)=x_3.\) Then the non-singular change of variables ([\[skorik-r1_f5\]](#skorik-r1_f5){reference-type="ref" reference="skorik-r1_f5"}) has the form \[\label{skorik-r1_pr2_zam} z_1 =x_1-x_2,\quad z_2 =x_3, \quad z_3 =f_2(x_2).\] We get \(x_3=z_2,\) \(x_2=f_2^{-1}(z_3),\) \(x_1=z_1+f_2^{-1}(z_3),\) then \[\label{skorik-r1_pr2_f2} \dot z_1 = u^3-u + 0.1\sin^2 \widetilde {f_1}(z_1,z_2,z_3,u),\quad \dot z _2 = z_3 ,\quad \dot z_3 =\widetilde {f_2}(z_3)u,\] where \(\widetilde {f_1}(z_1,z_2,z_3,u) = f_1\Bigl(z_1+f_2^{-1}(z_3),f_2^{-1}(z_3),z_2,u\Bigr),\) \(\widetilde {f_2}(z_3)=f_2'(x_2)_{\bigl| x_2=f_2^{-1}(z_3)}.\) The system ([\[skorik-r1_pr2_f2\]](#skorik-r1_pr2_f2){reference-type="ref" reference="skorik-r1_pr2_f2"}) has the form ([\[skorik-r1_fs\]](#skorik-r1_fs){reference-type="ref" reference="skorik-r1_fs"}), where \[H_1(z,u)=u^3-u+0.1\sin^2\widetilde {f_1}(z_1,z_2,z_3,u),\quad H_2(z,u)=\widetilde {f_2}(z_3)u.\] Now consider the \(0\)-controllability problem from the point \(z_0= (z_{10},z_{20},z_{30})^* = \left(x_{10}-x_{20}, x_{30},f_2(x_{20})\right)^*.\) On the first step of our approach we find controls \(u_1^+\) and \(u_1^-.\) Notice that the equation \[\label{skorik-r1_pr2_f3} u^3-u+0.1\sin^2\widetilde {f_1}(z_1,z_2,z_3,u) = v\] has three real roots on the segment \([-2,2]\) for all \(v\) such that \(|v|\le \frac{2}{3\sqrt{3}}-\frac{1}{10}.\) Put \(\varepsilon_1^\pm=0.2\) and choose controls \(u_1^+(z_1,z_2,z_3)\) and \(u_1^-(z_1,z_2,z_3)\) as the solutions of the equation ([\[skorik-r1_pr2_f3\]](#skorik-r1_pr2_f3){reference-type="ref" reference="skorik-r1_pr2_f3"}) with \(v=0.2\) and \(v=-0.2.\) It can be shown that \(u_1^-\in [-1.2,-0.8]\) and \(u_1^+\in [0.7, 1.1].\) Then we get \(H_1(z,u^+)=\varepsilon_1^+\) and \(H_1(z,u^-)=-\varepsilon_1^-.\) Put \[u_1(z_1,z_2,z_3)=\left\{\begin{array}{l} u_1^+(z_1,z_2,z_3)\;\; if \;\;z_{1}<0,\\ u_1^-(z_1,z_2,z_3)\;\; if \;\;z_{1}>0.\\ \end{array}\right.\] Then this control transfers any initial point \(z_0= (z_{10},z_{20},z_{30})^*\) to the point \(z_{T_1}=\left(0,z_{2T_1},z_{3T_1}\right)^*\) in the time \(T_1=5|z_{10}|.\) Therefore, the trajectory of the system ([\[skorik-r1_pr2_f2\]](#skorik-r1_pr2_f2){reference-type="ref" reference="skorik-r1_pr2_f2"}) with the control \(u=u_1(z_1,z_2,z_3)\) comes to the plane \(z_1=0.\) On the second step we choose the control \(u_2\) so that the trajectory of the system ([\[skorik-r1_pr2_f2\]](#skorik-r1_pr2_f2){reference-type="ref" reference="skorik-r1_pr2_f2"}) with the control \(u=u_2\) belongs to the plane \(z_1=0.\) This control should satisfy the equation ([\[skorik-r1_pr2_f3\]](#skorik-r1_pr2_f3){reference-type="ref" reference="skorik-r1_pr2_f3"}) with \(v=0\) and \(z_1=0.\) This equation has three real roots for any \(z_2,\) \(z_3.\) Moreover, it can be shown that two of them belong to the segments \([-1.1,-1]\) and \([0.9,1]\) respectively. Due to our assumption \(|\widetilde f_2(z_3)|\ge \delta>0\) for \(z_3\in {\Bbb R}.\) In the case \(\widetilde f_2(z_3)>0\) choose \(u_2^-(z_2,z_3)\in [-1.1,-1]\) and \(u_2^+(z_2,z_3)\in[0.9,1].\) In the case \(\widetilde f_2(z_3)<0\) choose \(u_2^+(z_2,z_3)\in [-1.1,-1]\) and \(u_2^-(z_2,z_3)\in[0.9,1].\) Then \(H_1(0,z_2,z_3,u_2^\pm(z_2,z_3))=0.\) Let \(\gamma^+\) and \(\gamma^-\) be the trajectories of the system \[\dot z_2 = z_3, \quad \dot z_3 =\widetilde {f_2}(z_3)u\] going to the origin and corresponding to the controls \(u=u_2^+(z_2,z_3)\) and \(u=u_2^-(z_2,z_3)\) respectively. The curve \(\gamma =\gamma^+\cup\gamma^-\) breaks the plane \(z_1=0\) in two parts. Put \[u_2(z_2,z_3)\!=\!\left\{\!\!\!\begin{array}{l} u_2^+(z_2,z_3)\; \mbox{if the point} \; (z_2,z_3)\;\mbox{lies below the curve}\; \gamma \;\mbox{or belongs to}\;\gamma^+,\\ u_2^-(z_2,z_3)\; \mbox{if the point} \; (z_2,z_3)\;\mbox{lies above the curve}\; \gamma \;\mbox{or belongs to}\;\gamma^-.\\ \end{array}\right.\] This control transfers the point \(z_{T_1}=\left(0,z_{2T_1},z_{3T_1}\right)^*\) to the origin in certain finite time \((T_2-T_1)\) and the corresponding trajectory belongs to the plane \(z_1=0.\) Returning to the initial variables we have that the control of the form \[u(x;t){=}\left\{\!\!\begin{array}{lcl} u_1\left(x_1-x_2,x_3,f_2(x_2)\right)& \mbox{as} & 0\le t<T_1, \\ u_2 \left(x_3,f_2(x_2)\right)& \mbox{as} & T_1\le t \le T_2, \\ \end{array}\right.\] transfers the initial point \((x_{10},x_{20},x_{30})\) to the origin by virtue of the initial system ([\[skorik-r1_pr2_sys2\]](#skorik-r1_pr2_sys2){reference-type="ref" reference="skorik-r1_pr2_sys2"}) in some finite time \(T_2,\) where \(T_1\) is defined by \(x_{10}-x_{20}\) and \(T_2-T_1\) is defined by \(x_{2T_1}\) and \(x_{3T_1}.\) ## {#skorik-example_3} Consider the system \[\label{skorik-r1_pr4_f1} \dot x_1=u,\quad \dot x_2=u^3, \quad \ldots,\quad \dot x_n=u^{2n-1},\] with constraints on a control of the form \(u\in \Omega=\{ u: \;|u|\le 1\}.\) The system ([\[skorik-r1_pr4_f1\]](#skorik-r1_pr4_f1){reference-type="ref" reference="skorik-r1_pr4_f1"}) can be written as \(\dot x=\Phi(u),\) \(x\in {\Bbb R}^n\), \(\Phi(u)=(u,u^3,\ldots,u^{2n-1})^*.\) This system is globally \(0\)-controllable due to the geometrical criterion since the origin \(x=0\) belongs to the interior of a convex span of the set \(\Phi(\Omega),\) i.e. \(0\in {\rm int\; co} \{\Phi(\Omega)\}.\) The system ([\[skorik-r1_pr4_f1\]](#skorik-r1_pr4_f1){reference-type="ref" reference="skorik-r1_pr4_f1"}) has the form ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) with \(a(x)=0,\) \(b_1(x)=e_n,\) \(\ldots,\) \(b_n(x)=e_1,\) where \(e_i\) is the \(i\)-th unit vector of the space \({\Bbb R}^n.\) By \(P_i(u)\) \((i=1,\ldots,n)\) denote the polynomial of degree \((2n{-}2i{+}1)\) of the form \[P_{i}(u) = u \prod\limits_{k = 1}^{n-i}\Bigl(u^2-\frac{k^2}{n^2}\Bigr)=u^{2n-2i+1}+\sum\limits_{k=1}^{n-i} c_k^{(i)} u^{2k-1},\; i=1,\ldots,n-1,\quad P_n(u)=u.\] Notice that all roots of the polynomial \(P_{i+1}(u)\) are the roots of polynomials \(P_1(u),\) \(\ldots,\) \(P_i(u).\) Put \[\varphi_i(x_1,\ldots,x_n)=x_{n-i+1}+\sum\limits_{k=1}^{n-i} c_k^{(i)}x_k,\;i=1,\ldots,n-1,\; \;\varphi_n(x_1,\ldots,x_n)=x_1,\] then the conditions ([\[skorik-r1_f4\]](#skorik-r1_f4){reference-type="ref" reference="skorik-r1_f4"}) are satisfied. Hence, the nonsingular change of variables \(z_i=\varphi_i(x_1,\ldots,x_n),\) \(i=1,\ldots,n,\) maps the system ([\[skorik-r1_pr4_f1\]](#skorik-r1_pr4_f1){reference-type="ref" reference="skorik-r1_pr4_f1"}) to the system \[\label{skorik-r1_pr4_f2} \dot z_i=P_i(u),\quad i=1,\ldots,n,\quad |u|\le 1,\] and an arbitrary point \(x_0=(x_{10},\ldots,x_{n0})^*\) is mapped to a point \(z_0=(z_{10},\ldots,z_{n0})^*,\) where \(z_{i0}=x_{n-i+1\,0}+\sum\limits_{k=1}^{n-i} c_k^{(i)} x_{k0}\) for \(i=1,\ldots,n-1\) and \(z_{n0}=x_{10}.\) We choose \(u_i(z_i)=-\frac{n-i+1}{n}\;{\rm sign}\,z_i,\) \(i=1,\ldots,n,\) and put \[\label{skorik-r1_pr4_upr} u(z;t)=u_i(z_i) \quad \mbox{as}\quad T_{i-1}\le t<T_i,\quad i=1,\ldots,n,\] where \(T_0=0,\) \(T_i=T_{i-1}+|z_{iT_{i-1}}/P_i(u_i)|\) (\(z_{iT_{i-1}}=z_{i}(T_{i-1})\)). The control ([\[skorik-r1_pr4_upr\]](#skorik-r1_pr4_upr){reference-type="ref" reference="skorik-r1_pr4_upr"}) transfers the point \(z_0\) to the origin by virtue of the system ([\[skorik-r1_pr4_f2\]](#skorik-r1_pr4_f2){reference-type="ref" reference="skorik-r1_pr4_f2"}) in some finite time \(T=T_n.\) Namely, on the first step the control \(u_1=-{\rm sign}\,z_{10}\) transfers the point \(z_0\) to the point \(z_{T_1}=(0,z_{2T_1},\ldots, z_{nT_1})^*,\) where \(z_{kT_1}=P_k(u_1)T_1+z_{k0},\) \(k=2,\ldots,n,\) in the time \(T_1=|z_{10}/P_1(u_1)|\) along the trajectory \(z(t)=\left(P_1(u_1)t+z_{10},\ldots, P_n(u_1)t+z_{n0}\right)^*.\) Since \(P_1(u_i)=0\) for \(i=2,\ldots,n\) then \(z_1(t)=0\) for \(t\ge T_1.\) Further, on the \(i\)-th step (\(i=2,\ldots,n\)) the control \(u=u_i\) transfers the point \(z_{T_{i-1}}=\left(0,\ldots,0,z_{iT_{i-1}},\ldots,z_{nT_{i-1}}\right)^*\) to the point \(z_{T_i}=(0,\ldots,0,z_{i+1T_i},\ldots, z_{nT_i})^*,\) where \(z_{kT_i}=P_k(u_i)(T_i-T_{i-1})+z_{kT_{i-1}},\) \(k=i+1,\ldots,n,\) in the time \(T_i-T_{i-1}=|z_{iT_{i-1}}/P_i(u_i)|\) along the trajectory \(z(t)=\left(0,\ldots,0,P_i(u_i)(t-T_{i-1})+z_{iT_{i-1}},\ldots, P_n(u_i)(t-T_{i-1})+z_{nT_{i-1}}\right)^*.\) Since \(P_i(u_k)=0\) for \(k=i+1,\ldots,n\) then \(z_i(t)=0\) as \(t\ge T_i.\) Returning to the initial variables we find \(x_1,\) \(\ldots,\) \(x_n\) successively from the equalities \(x_1=z_n,\) \(x_i=z_{n-i+1}-\sum\limits_{k=1}^{i-1} c_k^{(n-i+1)} x_k\) for \(i=2,\ldots,n.\) Thus, the control \[u(x;t)=-\frac{n{-}i{+}1}{n}{\rm sign}\left(x_{n{-}i{+}1}{+}\sum\limits_{k=1}^{n-i} c_k^{(i)} x_{k }\right)\quad\mbox{as}\quad T_{i-1}\le t<T_i,\quad i=1,\ldots,n,\] satisfies the preassigned constraint \(|u|\le 1\) and transfers an arbitrary point \(x_0\) to the origin in some finite time \(T(x_0)\) along the trajectory \(x(t)=(x_1(t),\ldots,x_n(t))^*\) of the system ([\[skorik-r1_pr4_f1\]](#skorik-r1_pr4_f1){reference-type="ref" reference="skorik-r1_pr4_f1"}), where \(x_1(t)=z_n(t),\) \(x_i(t)=z_{n-i+1}(t) -\sum\limits_{k=1}^{i-1} c_k^{(n-i+1)} x_k(t),\) \(i=2,\ldots,n.\) Notice that this construction admits an obvious generalization. Let us choose numbers \(\lambda_1,\) \(\ldots,\) \(\lambda_{n-1}\) such that \(\;0<\lambda_1<\ldots<\lambda_{n-1}< d\) and introduce the polynomials \[P_i(u) = u \prod\limits_{k = 1}^{n-i}\left(u^2-\lambda_k^2\right)= u^{2n-2i+1}+\sum\limits_{k=1}^{n-i} p_k^{(i)} u^{2k-1},\; i=1,\ldots,n-1,\quad P_n(u)=u.\] Consider the nonsingular change of variables \(z_i=x_{n-i+1}+\sum\limits_{k=1}^{n-i} p_k^{(i)} x_k,\) \(i=1,\ldots,n{-}1,\) \(z_n=x_1\) and choose the control ([\[skorik-r1_pr4_upr\]](#skorik-r1_pr4_upr){reference-type="ref" reference="skorik-r1_pr4_upr"}) with \(u_1(z)=-\alpha\, {\rm sign}\,z_1,\) \(\alpha{\in}(0,d]\setminus\{\lambda_1,\ldots,\lambda_{n-1}\},\) and \(u_i(z_i,\ldots,z_n)=-\lambda_{n+1-i}\;{\rm sign}\,z_{i}\) for \(i=2,\ldots,n,\) where \(T_i=T_{i-1}+|z_{iT_{i-1}}/P_i(u_i)|,\) \(i=1,\ldots,n.\) This control satisfies the constraint \(|u|\le d\) and transfers an arbitrary point \(z_0\) to the origin in the finite time \(T=T_n.\) # Calming of vibrations of a two-link pendulum {#skorik-calming_of_vibrations_of_a_two-link_pendulum} In this section we consider the model of a controllable two-link pendulum (see fig. [\[skorik-ndm-ris1\]](#skorik-ndm-ris1){reference-type="ref" reference="skorik-ndm-ris1"}). Namely, let a pendulum have two links of mass \(m_1,\) \(m_2\) and of lengths \(l_1,\) \(l_2\) respectively. Then the state of the pendulum is described by angles \(\varphi,\) \(\psi\) and angle velocities \(\dot \varphi,\) \(\dot \psi\) (\(\varphi\) is the angle between the first bar and the vertical axis; \(\psi\) is the angle between the second bar and the vertical axis). Let \(F_1,\) \(F_2\) be forces applied to the first and the second link respectively. Let \(g\) be the acceleration of the free fall. We consider the model of the pendulum with \(F_1=\alpha u^3,\) \(F_2=u,\) where \(u\) is a control, \(\alpha\in(0,(4/27)l_1^2/g^2].\) Suppose the initial state of the pendulum \((\varphi_0,\psi_0, \dot \varphi_0,\dot \psi_0)\) is given. We construct a control \(u=u(\varphi,\psi, \dot \varphi,\dot \psi,t)\) which calms the vibrations of the pendulum, that is transfers the initial state \((\varphi_0,\psi_0, \dot \varphi_0,\dot \psi_0)\) to the origin in some finite time \(T,\) i.e. \(\varphi(T)=0,\) \(\psi(T)=0,\) \(\dot \varphi(T)=0,\) \(\dot \psi(T)=0.\) The control motion of the two-link pendulum is described by the equations \[\label{skorik-syst_phi_psi} \begin{array}{l} \displaystyle \ddot\varphi=-\frac{gm_1\sin\varphi+m_2\sin(\varphi{-}\psi) \left(g\cos\psi+l_1\dot\varphi^2\cos(\varphi{-}\psi)+l_2\dot\psi^2\right)} {l_1\left(m_1+m_2\sin^2(\varphi{-}\psi)\right)}+\alpha u^3,\\[5pt] \displaystyle \ddot\psi=\frac{\sin(\varphi{-}\psi)\left((m_1+m_2)\left(g\cos\varphi+l_1\dot\varphi^2\right) +m_2l_2\dot\psi^2\cos(\varphi{-}\psi)\right)} {l_2\left(m_1{+}m_2\sin^2(\varphi{-}\psi)\right)}+u.\end{array}\] Put \(x_1=\varphi,\) \(x_2=\dot \varphi,\) \(x_3=\psi,\) \(x_4=\dot\psi,\) then we obtain the system \[\label{skorik-dnm_f1_x} \dot x_1=x_2,\quad \dot x_2=\beta_1(x,u),\quad \dot x_3=x_4,\quad \dot x_4=\beta_2(x,u),\] where \[\begin{array}{l} \displaystyle \beta_1(x,u)=-\frac{gm_1\sin x_1{+}m_2\sin(x_1{-}x_3)\left(g\cos x_3{+}l_1x_2^2\cos(x_1{-}x_3){+}l_2x_4^2)\right)}{l_1\left(m_1{+}m_2\sin^2(x_1{-}x_3)\right)}+\alpha u^3,\\[3pt] \displaystyle \beta_2(x,u)=\frac{\sin(x_1{-}x_3)\left((m_1{+}m_2)(g\cos x_1{+}l_1x_2^2){+}l_2m_2x_4^2\cos(x_1{-}x_3)\right)}{l_2\left(m_1+m_2\sin^2(x_1{-}x_3)\right)}+u, \end{array}\] The system ([\[skorik-dnm_f1_x\]](#skorik-dnm_f1_x){reference-type="ref" reference="skorik-dnm_f1_x"}) can be rewritten in the form ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) with \(a(x)=(x_2,0,x_4,0)^*,\) \(a(0)=0,\) \(b_1(x)=(0,1,0,0)^*,\) \(b_2(x)=(0,0,0,1)^*,\) \(m=2.\) The matrix \(K(x)\) from ([\[skorik-r1_m\_K\]](#skorik-r1_m_K){reference-type="ref" reference="skorik-r1_m_K"}) has the form \[K(x)=(b_1(x),{\rm ad}_a b_1(x), b_2(x),{\rm ad}_a b_2(x))= \left(\begin{array}{rrrr} 0&-1&0&0\\ 1&0&0&0\\ 0& 0& 0&-1\\ 0&0&1&0\end{array}\right)\] and \({\rm rang}\, K(x)=4,\) \(x\in{\Bbb R}^4,\) hence, \(n_1=2,\) \(n_2=2.\) The conditions ([\[skorik-r1_f4\]](#skorik-r1_f4){reference-type="ref" reference="skorik-r1_f4"}) imply \[\begin{array}{l} \displaystyle \frac{\partial\varphi_1(x_1,x_2,x_3)}{\partial x_2}= 0,\; \frac{\partial\varphi_1(x_1,x_2,x_3)}{\partial x_4}= 0,\; \frac{\partial\varphi_1(x_1,x_2,x_3)}{\partial x_1}\ne 0,\\ \displaystyle \frac{\partial\varphi_2(x_1,x_2,x_3)}{\partial x_2}= 0,\; \frac{\partial\varphi_2(x_1,x_2,x_3)}{\partial x_4}= 0,\; \frac{ \partial\varphi_2(x_1,x_2,x_3)}{\partial x_3} \ne 0, \end{array}\quad x\in{\Bbb R^4}.\] Choose \(\varphi_1(x_1,x_2,x_3)= x_1-x_2\) and \(\varphi_2(x_1,x_2,x_3)=x_3.\) Hence, the non-singular change of variables ([\[skorik-r1_f5\]](#skorik-r1_f5){reference-type="ref" reference="skorik-r1_f5"}) has the form \(z_1=x_1-x_3,\) \(z_2=x_2-x_4,\) \(z_3=x_3,\) \(z_4=x_4.\) Then ([\[skorik-dnm_f1_x\]](#skorik-dnm_f1_x){reference-type="ref" reference="skorik-dnm_f1_x"}) is mapped to the system \[\dot z_1=z_2,\quad\dot z_2=H_1(z_1,z_2,z_3,z_4,u),\quad \dot z_3=z_4,\quad \dot z_4=H_2(z_1,z_2,z_3,z_4,u), \label{skorik-dnm_f1_z}\] \[\begin{array}{l} \displaystyle H_1(z,u)=-\frac{1}{l_1 l_2 (m_1{+} m_2\sin^2 z_1)} (l_2 m_1 g \sin (z_1{+}z_3){+}l_2 m_2 \sin z_1 (g \cos z_3{+}l_2 z_4^2{+}\\[5pt] \qquad\qquad\quad + l_1 (z_2{+}z_4^2)\cos z_1){+} l_1\sin z_1 (l_2 m_2 z_4^2\cos z_1{+} (m_1 {+} m_2) (g \cos (z_1{+}z_3){+}\\[3pt] \qquad\qquad\quad + l_1 (z_2{+}z_4)^2))){+}\alpha u^3{-}u,\\[5pt] \displaystyle H_2(z,u)=\frac{\sin z_1(l_2m_2z_4^2\cos z_1{+}(m_1{+}m_2)(g\cos(z_1+z_3)+l_1(z_2{+}z_4)^2))} {l_2(m_1{+}m_2\sin^2z_1)}+u. \end{array}\] Notice that for a fixed \(z\) the function \(H_1(z,u)\) is a cubic polynomial with respect to \(u.\) We choose controls \(u_1^+(z)\) and \(u_1^-(z)\) as solutions of equations \(H_1(z,u)=\varepsilon_1^{+},\) \(H_1(z,u)=-\varepsilon_1^{-}\) respectively (we do not give the explicit formulas for these controls since they are too complicated). Put \(w_1(z_1)=-\sqrt{2\varepsilon_1^{+}z_1}\) for \(z_1\ge 0\) and \(w_1(z_1)=\sqrt{-2\varepsilon_1^{-}z_1}\) for \(z_1< 0,\) and define \[u_1(z)=\left\{\begin{array}{l} u_1^+(z_1,z_2,z_3,z_4)\;\; \mbox{if} \;\;z_2<w_1(z_1)\;\;\mbox{or} \;\; z_2=w_1(z_1)\;\; \mbox{and}\;\;z_1\ge 0,\\ u_1^-(z_1,z_2,z_3,z_4)\;\; \mbox{if} \;\;z_2>w_1(z_1)\;\;\mbox{or}\;\;z_2=w_1(z_1)\;\; \mbox{and}\;\;z_1\le 0.\\[3pt] \end{array}\right.\] If the initial point \(z_0=(z_{10},z_{20},z_{30},z_{40})^*\) satisfies the inequality \(z_{20}<w_1(z_{10})\) then the control \(u_1^+(z)\) transfers the point \(z_0\) to the point \(z_{T_{11}}=\bigl(z_{1T_{11}},z_{2T_{11}},\) \(z_{3T_{11}},z_{4T_{11}}\bigr)^*\) in the time \(T_{11}=\bigl(-z_{20}{+}\sqrt{(z_{20}^2{-}2z_{10}\varepsilon_1^{+}) \varepsilon_1^{-}/(\varepsilon_1^{+}{+}\varepsilon_1^{-})}\bigr)\bigl/\varepsilon_1^{+}\) and further the control \(u_1^-(z)\) transfers the point \(z_{T_{11}}\) to the point \(z_{T_1}=\left(0,0,z_{3T_1},z_{4T_1}\right)^*\) in the time \(T_{12}=\sqrt{\frac{z_{20}^2-2z_{10}\varepsilon_1^{+}}{\varepsilon_1^{-}(\varepsilon_1^{+}+ \varepsilon_1^{-})}}.\) Thus, the control \(u_1(z)\) transfers the initial point \(z_0\) to the point \(z_{T_1}\) in the time \(T_1=T_{11}+T_{12}=\bigl(-z_{20}{+}\sqrt{(z_{20}^2-2z_{10}\varepsilon_1^{+}) (\varepsilon_1^{+}+\varepsilon_1^{-})/\varepsilon_1^{-}}\bigr)\bigl/\varepsilon_1^{+}.\) If the initial point \(z_0\) satisfies the conditions \(z_{20}=w_1(z_{10})\) and \(z_{10}\ge 0\) then the control \(u_1^-(z)\) transfers the point \(z_0\) to the point \(z_{T_1}=\left(0,0,z_{3T_1},z_{4T_1}\right)^*\) in the time \(T_1=T_{12}.\) Analogously, in the case \(z_{20}>w_1(z_{10})\) the control \(u_1^-(z)\) transfers \(z_0\) to the point \(z_{T_{11}}\) in the time \(T_{11}=\bigl(z_{20}+\sqrt{(z_{20}^2+2z_{10}\varepsilon_1^{-}) \varepsilon_1^{+}/(\varepsilon_1^{+}+\varepsilon_1^{-})}\bigr)\bigl/\varepsilon_1^{-}\) and further the control \(u_1^+(z)\) transfers the point \(z_{T_{11}}\) to the point \(z_{T_1}=\left(0,0,z_{3T_1},z_{4T_1}\right)^*\) in the time \(T_{12}=\sqrt{\frac{z_{20}^2+2z_{10}\varepsilon_1^{-}}{\varepsilon_1^{+}(\varepsilon_1^{+}+ \varepsilon_1^{-})}}.\) Thus, the control \(u_1(z)\) transfers the point \(z_0\) to the point \(z_{T_1}\) in the time \(T_1=T_{11}+T_{12}=\bigl(z_{20}+\sqrt{(z_{20}^2+2z_{10}\varepsilon_1^{-}) (\varepsilon_1^{+}+\varepsilon_1^{-})/\varepsilon_1^{+}}\bigr) \bigl/\varepsilon_1^{-}.\) Finally, if \(z_{20}=w_1(z_{10})\) and \(z_{10}\le 0\) then the control \(u_1^+(z)\) transfers the point \(z_0\) to the point \(z_{T_1}=\left(0,0,z_{3T_1},z_{4T_1}\right)^*\) in the time \(T_1=T_{12}.\) Thus, on the first step the control \(u_1(z)\) transfers the point \(z_0\) to the point \(z_{T_1}=\left(0,0,z_{3T_1},z_{4T_1}\right)^*\) along the trajectory of the system ([\[skorik-dnm_f1_z\]](#skorik-dnm_f1_z){reference-type="ref" reference="skorik-dnm_f1_z"}) in the time \(T_1.\) On the second step the motion continues in the plane \(P=\{(z_1,z_2,z_3,z_4): z_1=z_2=0\}.\) To ensure this, we choose a control as a root of the equation \[\label{skorik-equation_H_1-0} H_1(0,0,z_3,z_4,u)=\alpha u^3-u-\frac{g}{l_1}\,\sin z_3=0.\] Notice that if \(\alpha \in (0,(4/27)l_1^2/g^2]\) then this equation has at least one positive root and one negative root for any \(z_3\in{\Bbb R}.\) Put \[\alpha_0(z_3){=}\sqrt[3]{\frac{g\sin z_3}{2\alpha l_1}{+}\frac{1}{\alpha}\sqrt{\frac{g^2\sin^2z_3}{4l_1^2}{-}\frac{1}{27\alpha}}}, \quad \beta_0(z_3){=}\sqrt[3]{\frac{g\sin z_3}{2\alpha l_1}{-}\frac{1}{\alpha}\sqrt{\frac{g^2\sin^2z_3}{4l_1^2}{-}\frac{1}{27\alpha}}}.\] Choose \(u_2^+(z_3)\) as the maximal root and \(u_2^-(z_3)\) as the minimal root of the equation ([\[skorik-equation_H\_1-0\]](#skorik-equation_H_1-0){reference-type="ref" reference="skorik-equation_H_1-0"}), i.e. \[u_2^+(z_3)=\alpha_0(z_3)+\beta_0(z_3),\quad u_2^-(z_3)=\frac{1}{2}\left(-1+i \sqrt{3}\right)\alpha_0(z_3)-\frac{1}{2}\left(1+i \sqrt{3}\right)\beta_0(z_3).\] Then for all \(z_3\in{\Bbb R}\) we have \[0<u_2^+(-\pi/2)\le u_2^+(z_3)\le u_2^+(\pi/2),\quad u_2^-(-\pi/2)\le u_2^-(z_3)\le u_2^-(\pi/2)<0.\] Since \(H_2(0,0,z_3,z_4,u)=u\) then \[H_2(0,0,z_3,z_4,u_2^+(z_3))\ge\varepsilon_2^{+},\quad H_2(0,0,z_3,z_4,u_2^-(z_3))\le-\varepsilon_2^{-},\] where \(\varepsilon_2^{+}=u_2^+(-\pi/2),\) \(\varepsilon_2^{-}=-u_2^+(\pi/2).\) Consider the trajectories of the system ([\[skorik-dnm_f1_z\]](#skorik-dnm_f1_z){reference-type="ref" reference="skorik-dnm_f1_z"}) with the controls \(u_2^\pm(z_3)\) which go to the origin, i.e. the trajectories of the system \[\dot z_1=z_2,\quad\dot z_2=0,\quad \dot z_3=z_4,\quad \dot z_4=u_2^\pm (z_3),\] which go to the origin. They belong to the plane \(P\) and, in addition, \(z_4=w_2(z_3),\) where \(w_2(z_3)=-\sqrt{2\int\limits_0^{z_3}u^+(\zeta)d\zeta}\) if \(z_3\ge 0\) and \(w_2(z_3)=\sqrt{-2\int\limits_{z_3}^0u^-(\zeta)d\zeta}\) if \(z_3\le 0.\) Define \[u_2(z_3,z_4)=\left\{\begin{array}{l} u_2^+(z_3)\;\; \mbox{if} \;\;z_4<w_2(z_3)\;\;\mbox{or} \;\; z_4=w_2(z_3)\;\; \mbox{and}\;\;z_3\ge 0,\\ u_2^-(z_3)\;\; \mbox{if} \;\;z_4>w_2(z_3)\;\;\mbox{or}\;\;z_4=w_2(z_3)\;\; \mbox{and}\;\;z_3\le 0.\\[3pt] \end{array}\right.\] Like the first step, this control transfers the point \(z_{T_1}\) to the origin in some finite time \(T_2-T_1.\) Therefore, returning to the initial variables we have that the control \[u(\varphi,\dot\varphi,\psi,\dot\psi;t)=\left\{\begin{array}{ll} u_1(\varphi{-}\psi,\dot\varphi{-}\dot\psi,\psi,\dot\psi),& 0\le t\le T_1,\\ u_2(\psi,\dot\psi),& T_1\le t\le T_2,\\ \end{array}\right.\] transfers the initial point \((\varphi(0),\dot\varphi(0),\psi(0),\dot\psi(0))\) to the origin along the trajectory of the system ([\[skorik-syst_phi_psi\]](#skorik-syst_phi_psi){reference-type="ref" reference="skorik-syst_phi_psi"}) in the finite time \(T=T_2.\) Let us summarize the results. We have proved that the stoppage problem of a controllable two-link pendulum can be solved in the following way. On the first step the control is chosen so that the angles \(\varphi,\) \(\psi\) and the angular speeds \(\dot\varphi,\) \(\dot\psi\) become equal in the finite time moment \(T_1,\) i.e. \(\varphi(T_1)=\psi_(T_1)\) and \(\dot\varphi(T_1)=\dot\psi(T_1)\) (see fig. [\[skorik-ndm-ris2\]](#skorik-ndm-ris2){reference-type="ref" reference="skorik-ndm-ris2"}). Roughly speaking, the two links of the pendulum form a one-link pendulum of length \(l=l_1+l_2.\) Further damping of vibrations of the two-link pendulum preserves this configuration of the links, i.e. we choose the control so that \(\varphi(t)=\psi(t)\) and \(\dot\varphi(t)= \dot\psi(t)\) for \(t\in [T_1,T_2]\) until the time moment \(T_2\) when the stoppage occurs. As an example, let us transfer the point \(x_0=(-2,1,-1,0.5)^*\) to the origin according to the algorithm described earlier for the system ([\[skorik-dnm_f1_x\]](#skorik-dnm_f1_x){reference-type="ref" reference="skorik-dnm_f1_x"}) with \(m_1=m_2,\) \(l_1=l_2=l,\) \(g/l=1,\) \(\alpha=1/9\) (see fig. [\[skorik-ndm-ft1\]](#skorik-ndm-ft1){reference-type="ref" reference="skorik-ndm-ft1"}--fig. [\[skorik-ndm-ut\]](#skorik-ndm-ut){reference-type="ref" reference="skorik-ndm-ut"}). In this case \(\varepsilon_1^{+}=20,\) \(\varepsilon_1^{-}=10\) and \(T_{11}\approx 0.15814,\) \(T_1\approx 0.52443,\) \(T_{21}\approx 2.64102,\) \(T=T_2\approx 3.53471.\) # Classes of staircase systems {#skorik-skorik_section_7} In this section we introduce the new classes of nonlinear systems which are mapped on the systems ([\[skorik-r1_kns\]](#skorik-r1_kns){reference-type="ref" reference="skorik-r1_kns"}). In addition, we give changes of variables satisfying ([\[skorik-r1_f4\]](#skorik-r1_f4){reference-type="ref" reference="skorik-r1_f4"})--([\[skorik-r1_f5\]](#skorik-r1_f5){reference-type="ref" reference="skorik-r1_f5"}). ## {#section-1} Let the system ([\[skorik-r1_f2\]](#skorik-r1_f2){reference-type="ref" reference="skorik-r1_f2"}) be of the form \[\label{skorik-r1_f20_obobshennaya_treugolnaya} \left\{\begin{array}{l} \dot x_1 \qquad\;\,= f_1(x_1,\ldots,x_{m{+}1}),\\ \dot x_2 \qquad\;\, = f_2(x_1,\ldots,x_{m{+}2}),\\.\qquad.\qquad.\qquad.\qquad.\qquad.\\ \dot x_{n{-}m}\quad = f_{n{-}m}(x_1,\ldots,x_n), \\ \dot x_{n{-}m{+}1} = f_{n{-}m{+}1}(x_1,\ldots,x_n,u),\\ .\qquad.\qquad.\qquad.\qquad.\qquad.\\ \dot x_n\qquad\;\, = f_n(x_1,\ldots,x_n,u), \end{array}\right. 1\le m<n.\] The system ([\[skorik-r1_f20_obobshennaya_treugolnaya\]](#skorik-r1_f20_obobshennaya_treugolnaya){reference-type="ref" reference="skorik-r1_f20_obobshennaya_treugolnaya"}) for \(m=1\) was introduced and considered in the paper and was named the triangular system. In this subsection we consider the case \(m=2\) in detail, i.e. we consider the system \[\label{skorik-r1_f20} \left\{\begin{array}{l} \dot x_1 \quad = f_1(x_1,x_2,x_3), \\ \dot x_2 \quad = f_2(x_1,x_2,x_3,x_4), \\.\qquad.\qquad.\qquad.\qquad.\qquad.\\ \dot x_{n-2} = f_{n-2}(x_1,x_2,\ldots,x_n), \\ \dot x_{n-1} = f_{n-1}(x_1,x_2,\ldots,x_n,u), \\ \dot x_n \quad = f_n(x_1,x_2,\ldots,x_n,u), \qquad\qquad n \ge 3. \\ \end{array}\right.\] Here and further \(a(x)=(f_1,\ldots,f_{n-2}),\) \(b_1 =(0,\ldots,0, 1, 0)^*,\) \(b_2 = (0,\ldots,0, 1)^*,\) \(\beta_1(x,u) = f_{n-1}(x_1,\ldots,x_n,u),\) \(\beta_2(x,u)=f_n(x_1,\ldots,x_n,u).\) Put \(\varphi_1(x) = x_1,\) \(\varphi_2(x) = x_2\) and consider the change of variables \[\label{skorik-r1_f22} \left\{\begin{array}{l} y_1= \varphi_1(x)=x_{1} \equiv F_{1}(x_{1}),\quad z_{1}=\varphi_2(x)=x_{2}\equiv \Phi_{1}(x_{1},x_{2}),\\ \displaystyle y_k= L_a^{k-1}\varphi_1(x) = \sum\limits_{i=1}^{2k-3}\frac {\partial F_{k-1}(x_{1},\ldots,x_{2k-3})}{\partial x_{i}}\ f_{i}(x_{1},\ldots,x_{i+2})\equiv \\ \quad\;\equiv F_{k}(x_{1},\ldots,x_{2k-1}), \quad 2\le k \le p\;\; \mbox{for}\;\; n=2p \;\; \mbox{or}\;\; n=2p-1,\\ \displaystyle z_k=L_{a}^{k-1}\varphi_2(x)=\sum\limits_{i=1}^{2k-2}\frac {\partial \Phi_{k-1}(x_{1},\ldots,x_{2k-2})}{\partial x_{i}}f_{i}(x_{1},\ldots,x_{i+2})\equiv\\ \quad\; \equiv \Phi_{k}(x_{1},\ldots,x_{2k}), \quad\left\{\begin{array}{lcl} 2\le k \le p & \mbox {if} & n= 2p,\\ 2\le k \le p-1&\mbox{if} & n=2p-1.\end{array}\right. \end{array} \right.\] In addition, if \(n=2p\) then put \[\begin{array}{l} F_{p+1}(x,u)=\sum\limits_{i=1}^{n-2}\frac {\partial F_{p}(x_{1},\ldots,x_{n-1})}{\partial x_{i}}f_{i}(x_{1},\ldots,x_{i+2}) + \frac {\partial F_p(x_1,\ldots,x_{n-1})}{\partial x_{n-1}}f_{n-1}(x,u),\\[5pt] \Phi_{p+1}(x,u)=\sum\limits_{i=1}^{n-2}\frac {\partial \Phi_{p}(x)}{\partial x_{i}}f_{i}(x_{1},\ldots,x_{i+2}){+} \frac {\partial \Phi_{p}(x)}{\partial x_{n-1}}f_{n-1}(x,u){+} \frac{\partial \Phi_{p}(x)}{\partial x_{n}}f_{n}(x,u). \end{array}\] If \(n=2p-1\) then put \[\begin{array}{l} F_{p+1}(x,u)= \sum\limits_{i=1}^{n-2}\frac {\partial F_{p}(x)}{\partial x_{i}}f_{i}(x_{1},\ldots,x_{i+2})+ \frac {\partial F_{p}(x)}{\partial x_{n-1}}f_{n-1}(x,u)+\frac {\partial F_p(x)}{\partial x_n}f_n(x,u),\\[5pt] \Phi_p(x,u)=\sum\limits_{i=1}^{n-2}\frac {\partial \Phi_{p-1}(x_1,\ldots,x_{n-1})}{\partial x_i}f_i(x_1,\ldots,x_{i+2}){+} \frac {\partial \Phi_{p-1}(x_1,\ldots,x_{n-1})}{\partial x_{n-1}}f_{n-1}(x,u).\end{array}\] For solvability of the system ([\[skorik-r1_f22\]](#skorik-r1_f22){reference-type="ref" reference="skorik-r1_f22"}) with respect to \(x_1,\) \(\ldots,\) \(x_n\) we require that \(\left|\frac{\partial f_i(x_1,\ldots,x_{i+2})}{\partial x_{i+2}}\right| \ge a>0\) for \(i=1,\ldots,n-2.\) Analogously to the paper, we prove the equalities \[\begin{array}{l} \displaystyle \frac{\partial F_k(x_1,\ldots,x_{2k-1})}{\partial x_{2k-1}} =\prod\limits_{i=1}^{k-1} \frac{\partial f_{2i-1}(x_1,\ldots,x_{2i+1})}{\partial x_{2i+1}},\quad k=2,\ldots, p, \\ \displaystyle \frac{\partial \Phi_k(x_1,\ldots,x_{2k})}{\partial x_{2k}} = \prod\limits_{i=1}^{k-1} \frac{\partial f_{2i}(x_1,\ldots,x_{2i+2})}{\partial x_{2i+2}},\quad k=\left\{\!\!\begin{array}{lcl} 2,\ldots, p{-}1&\mbox{if}& n=2p{-}1,\\ 2,\ldots, p&\mbox{if}& n=2p.\end{array}\right. \end{array}\] Hence, \[\left|\frac{\partial F_k(x_1,\ldots,x_{2k-1})}{\partial x_{2k-1}}\right|\ge a^{k-1}>0, \quad \left|\frac{\partial \Phi_k(x_1,\ldots,x_{2k})}{\partial x_{2k}}\right| \ge a^{k-1}>0.\] Thus, the change of variables ([\[skorik-r1_f22\]](#skorik-r1_f22){reference-type="ref" reference="skorik-r1_f22"}) is nonsingular. Let us explain how to solve the system ([\[skorik-r1_f22\]](#skorik-r1_f22){reference-type="ref" reference="skorik-r1_f22"}) with respect to \(x_1,\) \(\ldots,\) \(x_n.\) At the beginning we have \(x_1=y_1= h_1(y_1),\) \(x_2=z_1=h_2(z_1),\) according to the change of variables. Suppose that for certain \(k\ge 2\) the variables \(x_1,\) \(\ldots,\) \(x_{2k-2}\) are found and have the form \[\begin{array}{l} x_3=h_3(y_1,y_2,z_1),\\ x_4=h_4(y_1,y_2,z_1,z_2),\\.\qquad.\qquad.\qquad.\qquad.\\ x_{2k-3} = h_{2k-3}(y_1,\ldots,y_{k-1},z_1,\ldots,z_{k-2}),\\ x_{2k-2} = h_{2k-2}(y_1,\ldots,y_{k-1}, z_1,\ldots,z_{k-1}). \end{array}\] Consider the functions \(\widehat F(x_{2k-1})= F_{k}(x_1,\ldots,x_{2k-2},x_{2k-1}),\) \(\widehat\Phi(x_{2k})= \Phi_{k}(x_1,\ldots,\) \(x_{2k-1},x_{2k}).\) The functions \(\widehat F(x_{2k-1}),\) \(\widehat\Phi(x_{2k})\) are one-to-one mappings of \(\Bbb R\) to \(\Bbb R\). From the equation \[\begin{array}{c} y_k = F_k(x_1,\ldots,x_{2k-1}) =F_k\Bigl(h_1(y_1),h_2(z_1),\ldots, h_{2k-3}(y_1,\ldots\\ \ldots,y_{k-1},z_1,\ldots,z_{k-2}), h_{2k-2}(y_1,\ldots, y_{k-1},z_1,\ldots,z_{k-1}),x_{2k-1}\Bigr),\end{array}\] we find \(x_{2k-1}=h_{2k-1}(y_1,\ldots,y_k,z_1,\ldots,z_{k-1}).\) Substituting this expression to the equation \(z_k = \Phi_k(h_1(y_1),h_2(z_1),\ldots, h_{2k-2}(y_1,\ldots,y_{k-1},\ldots,z_1,\ldots,z_{k-1}), x_{2k-1}, x_{2k})\) we obtain \[\begin{array}{c} z_k =\Phi_k(x_1,\ldots,x_{2k}) = \Phi_k(h_1(y_1),h_2(z_1),\ldots, h_{2k-2}(y_1,\ldots\\ \ldots,y_{k-1},\ldots,z_1,\ldots,z_{k-1}),h_{2k-1}(y_1,\ldots,y_k,z_1,\ldots,z_{k-1}), x_{2k}).\end{array}\] From this equation we find \(x_{2k}= h_{2k}(y_1,\ldots,y_k,z_1,\ldots,z_k).\) Thus, if \(n=2p\) then the nonsingular change of variables ([\[skorik-r1_f22\]](#skorik-r1_f22){reference-type="ref" reference="skorik-r1_f22"}) maps the system ([\[skorik-r1_f20\]](#skorik-r1_f20){reference-type="ref" reference="skorik-r1_f20"}) to the system \[\label{skorik-chetnyi_slychai} \left\{\begin{array}{l} \dot y_i=y_{i{+}1},\; i=1,\ldots,p{-}1,\; \dot y_p=F_{p{+}1}\bigl(h_1(y_1),h_2(z_1)..., h_{2p-1}(y_1...,y_p,\\ \qquad z_1...,z_{p-1}), h_{2p}(y_1,\ldots, y_{p},z_1...,z_p),u\bigr)\equiv H_1(y_1...,y_p,z_1...,z_p,u),\\ \dot z_i=z_{i{+}1},\; i=1,\ldots,p{-}1,\; \dot z_p=\Phi_{p{+}1}(h_1(y_1),h_2(z_1)..., h_{2p-1}(y_1...,y_p,\\ \qquad z_1...,z_{p-1}),h_{2p}(y_1...,y_p,z_1...,z_p),u)\equiv H_2(y_1...,y_p,z_1...,z_p,u), \end{array}\right.\] and if \(n=2p-1\) then one maps the system ([\[skorik-r1_f20\]](#skorik-r1_f20){reference-type="ref" reference="skorik-r1_f20"}) to the system \[\label{skorik-kanonical_system_staicase} \left\{\begin{array}{l} \dot y_i=y_{i+1},\quad i=1,\ldots,p-1,\qquad \dot y_p\quad=H_1(y_1,\ldots,y_p,z_1,\ldots,z_{p-1},u),\\ \dot z_i=z_{i+1},\quad i=1,\ldots,p-2,\qquad \dot z_{p-1}=H_2(y_1,\ldots,y_p,z_1,\ldots,z_{p-1},u). \end{array}\right.\] In the partial case when the first \((n-2)\) equations of the system ([\[skorik-r1_f20\]](#skorik-r1_f20){reference-type="ref" reference="skorik-r1_f20"}) are linear with respect to the last argument, i.e. the system has the form \[\left\{ \begin{array}{lll} \displaystyle \dot x_i \quad\, = f_i(x_1,\ldots,x_{i+1})+c_ix_{i+2}, \qquad i=1,\ldots,n{-}2, \\ \dot x_{n-1}= f_{n-1}(x_1,\ldots,x_n,u), \quad \dot x_n = f_n(x_1,\ldots,x_n,u), \\ \end{array}\right.\quad \prod\limits_{i=1}^{n-2}c_i\ne 0,\] the system ([\[skorik-r1_f22\]](#skorik-r1_f22){reference-type="ref" reference="skorik-r1_f22"}) is solvable with respect to \(x_1,\ldots,x_n\) in an obvious way analogously to. For example, the nonsingular change of variables \(y_1=x_1,\) \(y_2=x_3,\) \(z_1=x_2,\) \(z_2=x_1^2+x_4\) maps the system \[\dot x_1=x_3,\,\dot x_2=x_1^2{+}x_4,\, \dot x_3=f_1(x_1,x_2,x_3,x_4)\cos u,\, \dot x_4=f_2(x_1,x_2,x_3,x_4)\sin u{-}2x_1x_3\] to the system \[\dot y_1=y_2,\; \dot y_2=f_1(y_1,z_1,y_2,z_2-y_1^2)\cos u,\; \dot z_1=z_2,\; \dot z_2=f_2(y_1,z_1,y_2,z_2-y_1^2)\sin u.\] ## {#section-2} For \(n=2p-1\ge 2\) we consider the system \[\label{skorik-r1_f23} \left\{ \begin{array}{l} \quad\;\dot x_1\quad \; = f_1(x_1,x_2,x_3), \\ \left\{\begin{array}{l}\dot x_{2i}\quad = f_{2i}(x_1,\ldots,x_{2i+3}), \\ \dot x_{2i+1}= f_{2i+1}(x_1,\ldots,x_{2i+3}), \end{array}\right.\quad i=1,\ldots,p-2, \\ \quad\; \dot x_{n-1} \, = f_{n-1}(x_1,\ldots,x_n,u), \\\quad\; \dot x_n \quad \, = f_n(x_1,\ldots,x_n,u). \end{array}\right.\] Put \(\varphi_1(x_1,\ldots,x_n)=x_1,\) \(\varphi_2(x_1,\ldots,x_n)=x_2.\) Then \[\label{skorik-r1_f30} \left\{\!\!\begin{array}{l} y_1=x_1\equiv F_1(x_1), \\ y_k= L_{a}^{k-1}\varphi_1 =\frac{d}{dt}F_{k-1}(x_{1},\ldots,x_{2k-3}) \equiv F_k(x_1,\ldots,x_{2k-1}), \; 2\le k\le p,\\ z_1=x_2\equiv \Phi_1(x_1,x_2),\\ z_k= L_{a}^{k-1}\varphi_2 = \frac{d}{dt} \Phi_{k-1}(x_{1},\ldots,x_{2k-1})\equiv \Phi_k(x_1,\ldots,x_{2k{+}1}),\; 2\le k\le p{-}1,\end{array}\right.\] and \(F_{p+1}(x,u)=\frac{d}{dt} F_p(x_1,\ldots,x_n),\) \(\Phi_p(x,u)=\frac{d}{dt} \Phi_{p-1}(x_1,\ldots,x_n).\) Following the paper, suppose that \[\Delta_0=\left|\frac{\partial f_1}{\partial x_3}\right|\ge \varepsilon_0,\quad \Delta_i= \left| \frac{\partial f_{2i}}{\partial x_{2i+2}}\frac{\partial f_{2i+1}}{\partial x_{2i+3}}-\frac{\partial f_{2i}}{\partial x_{2i+3}} \frac{\partial f_{2i+1}}{\partial x_{2i+2}} \right|\ge \varepsilon_i,\; i=1,\ldots,p-2.\] Then the change of variables ([\[skorik-r1_f30\]](#skorik-r1_f30){reference-type="ref" reference="skorik-r1_f30"}) is nonsingular and maps the system ([\[skorik-r1_f23\]](#skorik-r1_f23){reference-type="ref" reference="skorik-r1_f23"}) to the system ([\[skorik-kanonical_system_staicase\]](#skorik-kanonical_system_staicase){reference-type="ref" reference="skorik-kanonical_system_staicase"}). ## {#section-3} For \(n=2p\) we consider the system \[\label{skorik-r1_f41} \left\{ \begin{array}{l}\!\!\!\left\{ \begin{array}{l} \dot x_{2i-1} = f_{2i-1}(x_1,\ldots,x_{2i+2}), \\ \dot x_{2i}\quad = f_{2i}(x_1,\ldots,x_{2i+2}), \end{array}\right.\quad i=1,\ldots, p-1,\\ \;\;\,\dot x_{n-1}\,= f_{n-1}(x_{1},\ldots,x_n,u), \\ \;\;\, \dot x_n\quad \,= f_n(x_1,\ldots,x_n,u). \\ \end{array}\right.\] Put \(\varphi_1(x)=x_1,\) \(\varphi_2(x)=x_2\) and \[\label{skorik-r1_f42} y_k= L_{a}^{k-1}\varphi_1(x),\quad z_k= L^{k-1}_a\varphi_2(x),\quad k=1,\ldots,p.\] Suppose that \[\left|\! \displaystyle \frac{\partial f_{2i{-}3}(x_1,\ldots,x_{2i})}{\partial x_{2i-1}}\displaystyle \frac{\partial f_{2i{-}2}(x_1,\ldots,x_{2i})}{\partial x_{2i}}-\displaystyle \frac{\partial f_{2i{-}3}(x_1,\ldots,x_{2i})}{\partial x_{2i}}\displaystyle \frac{\partial f_{2i{-}2}(x_1,\ldots,x_{2i})}{\partial x_{2i-1}}\! \right|\ge \varepsilon\] for \(i=2,\ldots,p\) and \(\varepsilon > 0.\) Then the change of variables ([\[skorik-r1_f42\]](#skorik-r1_f42){reference-type="ref" reference="skorik-r1_f42"}) is nonsingular and maps the system ([\[skorik-r1_f41\]](#skorik-r1_f41){reference-type="ref" reference="skorik-r1_f41"}) to the system ([\[skorik-chetnyi_slychai\]](#skorik-chetnyi_slychai){reference-type="ref" reference="skorik-chetnyi_slychai"}). ## {#section-4} For a fixed \(k\) such that \(8\le 2k\le n{+}1\) we consider the system \[\label{skorik-r44_f1} \left\{\begin{array}{l} \dot x_i\quad\, = f_i(x_{1},\ldots,x_{k}), \qquad\;\;\; i=1,\ldots,k{-}1, \\ \dot x_i\quad\, = f _i(x_{1},\ldots,x_{i+1}), \qquad i=k,\ldots,n{-}3, \\ \dot x_{n-2}= f_{n-2}(x_{1},\ldots,x_{n}),\\ \dot x_{n-1} = f_{n-1}(x_{1},\ldots,x_{n},u), \\ \dot x_n \quad = f_{n}(x_{1},\ldots,x_{n},u). \\ \end{array}\right.\] Put \(\varphi_1(x)=x_1\) and \(\varphi_2(x)=x_{n-k+1}.\) Then the change of variables \[y_s=L^{s-1}_a \varphi_1(x),\;\; s=2,\ldots,n{-}k{+}1,\quad z_s=L^{s-1}_a \varphi_2(x),\;\; s=2,\ldots,k{-}1,\] maps the system ([\[skorik-r44_f1\]](#skorik-r44_f1){reference-type="ref" reference="skorik-r44_f1"}) to the system \[\left\{\begin{array}{lr} \dot y_i=y_{i+1},\quad i=1,\ldots,n{-}k,& \dot y_{n-k+1}=H_1(y_1,\ldots,y_{n-k+1},z_1,\ldots,z_{k-1},u),\\ \dot z_i=z_{i+1},\quad i=1,\ldots,k{-}2,& \dot z_{k-1}=H_2(y_1,\ldots,y_{n-k+1},z_1,\ldots,z_{k-1},u). \end{array}\right.\]

{'timestamp': '2013-10-14T02:09:23', 'yymm': '1310', 'arxiv_id': '1310.3194', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3194'}

# Introduction In thermal field theory, in order to deal with the infrared singularities which occur at finite temperature, it is necessary to put thermal masses into the zeroth order of a resummed perturbation theory. To this end, one must first calculate the hard thermal loops, where all the external energies and momenta are much smaller than the temperature \(T\). These loops yield gauge-invariant contributions, which are in general non-local functionals of the external fields. However, there are two special cases: the static and the long wave-length limits, when the hard thermal amplitudes become local functions, which are independent of the external energies and momenta. Nevertheless, these two limits give different functions. Of special interest is the evaluation in the above limits of causal thermal self-energy functions, which determine the high-temperature behavior of screening lengths and plasma frequencies. The purpose of this work is to derive a simple method for calculating the leading contributions of retarded thermal loops in static external bosonic fields. We will show that these contributions can be directly obtained by evaluating the hard thermal loops at zero external energies and momenta. This result has been previously derived in gauge theories at one-loop level and verified by explicit calculations at two-loops. Here, we present an argument which is valid to all orders in thermal perturbation theory. The above result may be more readily understood in the analytically continued imaginary-time formalism which is well suited for the study of retarded Green's functions. This formalism of thermal field theory defines the bosonic Green's functions at integral values of \({k_j}_0/2\pi\, i T\), where \({k_j}_0\) is the energy of the \(j\)-th external particle. After performing the sums over the integral values (half-integral for fermions) of \({Q_l}_0/ 2\pi\, i T\), where \({Q_l}_0\) is the energy of the \(l\)-th internal particle, one arrives at bosonic (fermionic) thermal distribution functions of the form \[\label{eq1} N(k_0+Q) = \frac{1}{{\rm e}^{(k_0+Q)/T} \mp 1}\] where \(k_0\) is some linear combination of external energies and \(Q\) is some combination of external and internal momenta. It is worthwhile to note that if one would now analytically continue the external energies, one would get an analytic behavior when all \({k_j}_o\) and \(\vec k_j\) become small, leading to a well defined result in the limit \({k_j}_\mu = 0\). However, the analytic continuation constructed in this manner would yield, after performing the integrations over internal momenta, factors like \(\exp{({k_0}/T)}\) which are exponentially increasing for large values of \({k_0}\). The proper procedure which avoids the appearance of such factors makes use of the relation \[\label{eq2} N(k_0+Q) = N(Q)\] which is valid before analytic continuation, since then, \({k_0}/2\pi\, i T\) is an integer. In this way, the Green's functions will be well behaved when \({k_j}_0\) are analytically continued to complex values, and various limits approaching the real axis from different directions may be taken. But this procedure, in contrast to the previous one, introduces non-analyticities in the thermal loops when \({k_j}_0\rightarrow 0\) and \({\vec k}_j\rightarrow 0\). Nevertheless, the leading contributions of hard thermal loops in the static case \({k_j}_0 = 0\), \({\vec k}_j\rightarrow 0\), still agree with those obtained by setting directly in the loops all external energies and momenta \({k_j}_\mu = 0\). This agreement occurs only in the static limit which entails that \(k_0=0\), since then the condition [\[eq2\]](#eq2){reference-type="eqref" reference="eq2"} reduces to an identity. Thus, in the static case, analytic continuation preserves the form of the original thermal distribution functions [\[eq1\]](#eq1){reference-type="eqref" reference="eq1"}, which lead to an analytic behavior when all \({\vec k}_j\rightarrow 0\). We will first exemplify the above argument in the case of a two-loop thermal amplitude. Next, we shall present a more general approach, based on a spectral representation of the thermal Green's functions, which allows to verify this argument to all orders. # Thermal self-energy at two-loops Let us consider, for example, a two-loop diagram in the scalar \(\lambda \phi^3\) theory, as shown in figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. The result for the one-loop sub-diagram in the renormalizable six-dimensional theory may be written in the form (\(p\equiv |\vec p|\))  \[\label{eq3} \Sigma^{(1)}(p_0,\vec p) = \frac{1}{2} (p_0^2-p^2) \Pi_T(p_0,p) + \sum_{n=0}^\infty \frac{P_{2(n+1)}(p_0,p)}{T^{2n}}\] where \(P_{2(n+1)}\) is a polynomial of degree \(2(n+1)\) in \(p_0\) and \(p\) (when \(n=0\), there appear additional logarithmic terms) and \(\Pi_T(p_0,p)\) is given by where \(\mu\) is a renormalization mass scale. The above one-loop result for \(\Pi_T(p_0,p)\) has branch points at \(p_0=\pm p\). When this term is inserted in the two-loop graph of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}, these branch points coincide with the poles at \(p_0=\pm p\) which arise from the other terms in this diagram. Therefore, in order to perform the \(p_0\) integration in a well defined manner, it is necessary to regularize \(\Pi_T\), which we do using dimensional regularization in \(6+2\epsilon\) dimensions. This leads to the regularized form of \(\Pi_T(p_0,p)\) given by where the anticlockwise contour \(C\), along the imaginary \(p_0\)-axis, is closed in the right half \(p_0\) plane. Evaluating the \(p_0\)-integral in terms of the poles inside \(C\), and using the fact that the leading contribution from the pole at \(p_0=p\cos\theta\) in \(\Pi_T(p_0,p,\epsilon)\) vanishes, leads to the result (\(\tilde p\equiv |\vec p + \vec k|\)) where the prime denotes a derivative with respect to \(p_0\). However, for the reasons mentioned previously, one must first use in [\[eq6\]](#eq6){reference-type="eqref" reference="eq6"} the relation \[\label{eq8} N(k_0+\tilde p) = N(\tilde p)\] and then make the analytic continuation of \(k_0\). This procedure will introduce a non-analyticity when \(k_\mu\rightarrow 0\). To study this, we remark that the leading contribution in \(T\) of the \(p\)-integral in [\[eq6\]](#eq6){reference-type="eqref" reference="eq6"} comes from the region where \(p\sim T\). Thus, in order to find the high-temperature behavior of the hard thermal loop, one may assume that \(|k_0|\), \(|\vec k| \ll p\). Then, it is easy to see that the leading contribution obtained in the static limit agrees with the result given in [\[eq7\]](#eq7){reference-type="eqref" reference="eq7"} \[\label{eq9} \Sigma_T^{(2)}(k_0=0,\vec k\rightarrow 0) = \Sigma_T^{(2)}(k_0=0,\vec k = 0).\] On the other hand, the leading contributions in the long wave-length limit would lead to a different result, namely Let us now evaluate the leading static thermal contribution which arises from [\[eq9\]](#eq9){reference-type="eqref" reference="eq9"}. In this case, there appear individual terms proportional to \(1/\epsilon\) which exhibit a collinear singularity in the region where \(p_0^2=p^2\), with \(\vec p\) and \(\vec q\) being nearly parallel (see Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}). However, such collinear singularities turn out to cancel so that in the present case, it is not necessary to resort to the KLN mechanism at finite temperature. Using Eqs. [\[eq4a\]](#eq4a){reference-type="eqref" reference="eq4a"} and [\[eq7\]](#eq7){reference-type="eqref" reference="eq7"}, together with the relations \[\label{eq11} \int_0^\infty dp N(p) p =-\frac{1}{2}\int_0^\infty dp N^\prime(p) p^2 = \frac{\pi^2 T^2}{6}\] one then gets a leading \(T^2 \log T\) contribution of the form \[\label{eq12} \Sigma_T^{(2)}(k_0=0,\vec k\rightarrow 0) = \frac{\lambda^4}{144}\frac{\pi}{4}\frac{1}{(2\pi)^5} T^2 \log \frac{T}{\mu}.\] # The self-energy to all orders In order to verify to all orders that the leading static contributions agree with the result obtained by setting directly, in the self-energy, the external energy-momentum equal to zero, it is convenient to use a spectral representation of the analytically continued bosonic self-energy (see, for example, chapter 3 in and references therein). One then obtains for the analytic retarded self-energy the spectral form \[\label{eq14} \Sigma(k_0,\vec k) = \Sigma(\infty,\vec k) + \int_{-\infty}^{\infty} \frac{d\kappa_0}{2\pi}\frac{\sigma(\kappa_0,\vec k)}{k_0-\kappa_0+i\epsilon},\] where \(k_0\) is a real energy and, for simplicity, only the energy-momentum dependence has been written explicitly. Here, the spectral density \(\sigma(\kappa_0,\vec k)=-\sigma(-\kappa_0,\vec k)\) is related to the discontinuity of \(\Sigma(\kappa_0,\vec k)\) across the real axis and the second term approaches zero when \(|k_0|\rightarrow\infty\). These features can be easily seen at one-loop in the scalar \(\lambda \phi^3_6\) model, where \(\Sigma(\infty,\vec k) =0\) and Here, \(\kappa_\mu=(\kappa_0,\vec k)\) and the free spectral density \(\rho_0(p)\) is given by \[\label{eq16} \rho_0(p) = 2 \pi \epsilon(p_0) \delta(p_0^2-p^2-m^2),\] where \(m\) is the ordinary mass of the scalar particles. One can check that, by substituting [\[eq15\]](#eq15){reference-type="eqref" reference="eq15"} into [\[eq14\]](#eq14){reference-type="eqref" reference="eq14"} and performing the \(\kappa_0\) integration, one gets the one-loop thermal contribution where \(k_0 \rightarrow k_0 + i\epsilon\) is to be understood and we have neglected \(m\) with respect to \(p\). This agrees with the result obtained from [\[eq6\]](#eq6){reference-type="eqref" reference="eq6"} by setting \(\Pi_T=1\) and employing the relation [\[eq8\]](#eq8){reference-type="eqref" reference="eq8"}. Evaluating [\[eq17\]](#eq17){reference-type="eqref" reference="eq17"} in the static and long wavelength limits, leads to distinct terms of order \(T^2\). Using the spectral representation [\[eq14\]](#eq14){reference-type="eqref" reference="eq14"}, where we take for definiteness \(\Sigma(\infty,\vec k)=0\), we now consider the leading thermal contributions which arise in the long wave-length and static limits of the retarded self-energy \[\label{eq18} \Sigma_T(k_0\rightarrow 0,\vec k=0)=\frac{1}{2\pi} \lim_{k_0\rightarrow 0}\int_{-\infty}^\infty \frac{d \kappa_0}{k_0-\kappa_0+i\epsilon}\sigma_T(\kappa_0,0),\] \[\label{eq19} \Sigma_T(k_0 = 0,\vec k\rightarrow 0)=-\frac{1}{2\pi} \lim_{\vec k \rightarrow 0} \int_{-\infty}^\infty \frac{d \kappa_0}{\kappa_0-i\epsilon}\sigma_T(\kappa_0,\vec k).\] Next, let us compare these contributions with the result obtained by setting directly \(k_0=0\), \(\vec k=0\) in the retarded thermal self-energy function \[\label{eq20} \Sigma_T(k_0 = 0,\vec k = 0)=-\frac{1}{2\pi} \int_{-\infty}^\infty \frac{d \kappa_0}{\kappa_0-i\epsilon}\sigma_T(\kappa_0,0).\] Since the integrand in \(\Sigma_T(k_0,\vec k = 0)\) is not a uniformly continous function of \(k_0\), we cannot take the limit \(k_0\rightarrow 0\) inside the integral [\[eq18\]](#eq18){reference-type="eqref" reference="eq18"}. Thus, we infer that the leading thermal contribution [\[eq18\]](#eq18){reference-type="eqref" reference="eq18"} got in the long wave-length limit, would generally differ from the result given in [\[eq20\]](#eq20){reference-type="eqref" reference="eq20"}. On the other hand, we will argue that in the static case, the limit \(\vec k\rightarrow 0\) can be taken inside the integral [\[eq19\]](#eq19){reference-type="eqref" reference="eq19"}. To this end, using the fact that \(\sigma_T(\kappa_0,\vec k)\) is an odd function of \(\kappa_0\), it is convenient to write [\[eq19\]](#eq19){reference-type="eqref" reference="eq19"} in the alternative form (where \({\cal P}\) denotes the principal value) \[\label{eq21} \Sigma_T(k_0 = 0,\vec k\rightarrow 0)=-\frac{1}{\pi} \lim_{\vec k \rightarrow 0} {\cal P} \int_{0}^\infty \frac{d \kappa_0}{\kappa_0}\sigma_T(\kappa_0,\vec k)\] Consider now the integral \[\label{eq22} I(\vec k)={\cal P} \int_{0}^\infty \frac{d \kappa_0}{\kappa_0}\sigma_T(\kappa_0,\vec k).\] It is well known that if \(I(\vec k)\) converges uniformly, then the limit \(\vec k\rightarrow 0\) can be taken inside the integral. Such a convergence may be shown by considering the physical meaning of \(\sigma_T(\kappa_0,\vec k)\), which gives the imaginary part of the retarded self-energy. It yields the rates of processes occurring in a thermal plasma, such as particle creation/annihilation or scattering, in the presence of an external field. At high temperatures, such that \(|\vec k | \ll T\), the leading contributions to these rates have a smooth behavior when \(\vec k\rightarrow 0\), in which case \(\sigma_T(\kappa_0,\vec k)\) would be a well behaved function in this limit. Furthermore, in consequence of unitarity (conservation of probability) such rates should decrease at large values of the energy \(\kappa_0\). Assuming, for example, that \(\sigma_T(\kappa_0,\vec k)\) behaves for large \(\kappa_0\) like \(\kappa_0^{2} \exp{(-C \kappa_0/T)}\), where \(C\) is a positive constant (which is consistent with [\[eq15\]](#eq15){reference-type="eqref" reference="eq15"}), one gets \[\label{eq23} \left |I(\vec k)-{\cal P} \int_{0}^E \frac{d \kappa_0}{\kappa_0}\sigma_T(\kappa_0,\vec k) \right| \propto \frac{E}{T}\exp{-\frac{CE}{T}} < \epsilon\] for every \(\epsilon\), provided \(E/T\) is sufficiently large. A similar condition is obtained also for more general forms of the spectral density at large \(\kappa_0\), which lead to a \(T^2\) behaviour of the self-energy at high temperatures. Thus, \(I(\vec k)\) will be uniformly convergent so that the limit \(\vec k\rightarrow 0\) can be taken inside [\[eq21\]](#eq21){reference-type="eqref" reference="eq21"} and, therefore, inside the integral \(\eqref{eq19}\). Consequently, it follows that the leading static contribution [\[eq19\]](#eq19){reference-type="eqref" reference="eq19"} will agree to all orders with the result [\[eq20\]](#eq20){reference-type="eqref" reference="eq20"}, got by calculating \(\Sigma_T\) at vanishing energy and momentum. This behavior is in agreement with the two-loop results given in Eqs. [\[eq9\]](#eq9){reference-type="eqref" reference="eq9"} and [\[eq12\]](#eq12){reference-type="eqref" reference="eq12"}. # generalization to \(n\)-point functions According to the simple arguments given following Eqs. [\[eq1\]](#eq1){reference-type="eqref" reference="eq1"} and [\[eq2\]](#eq2){reference-type="eqref" reference="eq2"}, the above result should hold as well in the case of higher point functions. We will now derive this property for the \(n\)-point Green's functions, using a treatment which generalizes the previous method. The spectral representation of the retarded \(n\)-point functions calculated in the imaginary time formalism may be written in the form where \({k_j}_0\) are real energy variables and \({\kappa_j}_\mu = ({\kappa_j}_o,{\vec k}_j)\). Setting in [\[eq24\]](#eq24){reference-type="eqref" reference="eq24"}, for example, \(\epsilon_l\) positive and all other epsilons negative such that \(\sum \epsilon_j = 0\), defines the analytic \(l\)-th retarded function. The spectral densities are the difference of two thermal Wightman functions which, in the case of pure bosonic fields, are given by We assume that to leading order at high temperature, when all \(|\vec k_j|\ll T\), these spectral densities (which may also depend on the particles masses, etc) are well behaved in the limit \(\vec k_j \rightarrow 0\). One may now consider the leading thermal contributions which arise from [\[eq24\]](#eq24){reference-type="eqref" reference="eq24"} in the static limit (when all \({k_j}_0=0\)) which is well defined due to the analyticity properties of \(\Gamma^{(n)}\). In this case, to leading order in \(T\), one may next take the limits \(\vec k_j\rightarrow 0\) and proceed similarly to the previous analysis. We then find a result which agrees with that obtained by setting in [\[eq24\]](#eq24){reference-type="eqref" reference="eq24"} all external energies and momenta equal to zero: We have explicitly verified this relation at two-loops order, by calculating the leading static thermal contributions of the three-point functions in gauge theories. Thus, we conclude that to all orders in the static limit, the leading thermal contributions of retarded bosonic Green's functions may be directly obtained by evaluating them at zero external energies and momenta. We would like to thank FAPESP and CNPq (Brazil) for a grant.

{'timestamp': '2013-10-14T02:11:09', 'yymm': '1310', 'arxiv_id': '1310.3246', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3246'}

# introduction {#sec1} Let \(G\) be finite, undirected, simple graph with \(n\) vertices and \(m\) edges having vertex set \(V(G)=\{v_1, v_2, \cdots, v_n\}\). Throughout this paper we denote such a graph by \(G(n,m)\). The adjacency matrix \(A=(a_{ij})\) of \(G\) is a \((0, 1)\)-square matrix of order \(n\) whose \((i,j)\)-entry is equal to one if \(v_i\) is adjacent to \(v_j\) and equal to zero, otherwise. The spectrum of the adjacency matrix is called the \(A\)-spectrum of \(G\). If \(\lambda_1, \lambda_2, \cdots, \lambda_n\) is the adjacency spectrum of \(G\), the energy of \(G\) is defined as \(E(G)=\sum_{i=1}^{n}|\lambda_i|\). This quantity introduced by I. Gutman has noteworthy chemical applications (see ).\ Let \(D(G)={diag}(d_1, d_2, \cdots, d_n)\) be the diagonal matrix associated to \(G\), where \(d_i\) is the degree of vertex \(v_i\). The matrices \(L(G)\)=\(D(G)\)-\(A(G)\) and \(L^+(G)\)=\(D(G)\)+\(A(G)\) are called Laplacian and signless Laplacian matrices and their spectras are respectively called Laplacian spectrum (\(L\)-spectrum) and signless Laplacian spectrum (\(Q\)-spectrum) of \(G\). Being real symmetric, positive semi-definite matrices, let \(0=\mu_n\leq\mu_{n-1}\leq\cdots\leq\mu_1\) and \(0\leq\mu^+_n\leq\mu^+_{n-1}\leq\cdots\leq\mu^+_1\) be respectively the \(L\)-spectrum and \(Q\)-spectrum of \(G\). It is well known that \(\mu_n\)=0 with multiplicity equal to the number of connected components of \(G\) (see ). In (see) Fiedler showed that a graph \(G\) is connected if and only if its second smallest Laplacian eigenvalue is positive and called it as the algebraic connectivity of the graph \(G\). Also it is well known that for a bipartite graph the \(L\)-spectra and \(Q\)-spectra are same (see ). The Laplacian energy of a graph \(G\) as put forward by Gutman and Zhou (see ) is defined as \(LE(G)=\sum\limits_{i=1}^{n}|\mu_i-\frac{2m}{n}|\). This quantity, which is an extension of graph-energy concept has found remarkable chemical applications beyond the molecular orbital theory of conjugated molecules (see ). Both energy and Laplacian energy have been extensively studied in the literature (see and the references therein). It is easy to see that \(tr(L(G))=\sum_{i=1}^{n}\mu_i=\sum_{i=1}^{n-1}\mu_i=2m\) and \(tr(LE^+(G))=\sum_{i=1}^{n}\mu^+_i=2m\).\ Two graphs \(G_1\) and \(G_2\) of same order are said to be equienergetic if \(E(G_1)\)=\(E(G_2)\), (see ). In analogy to this two graphs \(G_1\) and \(G_2\) of same order are said to \(L\)-equienergetic if \(LE(G_1)\)=\(LE(G_2)\) and \(Q\)-equienergetic if \(LE^+(G_1)\)=\(LE^+(G_2)\). Since cospectral (Laplacian cospectral) graphs are always equienergetic (\(L\)-equienergetic) the problem of constructing equienergetic (\(L\)-equienergetic) graphs is only considered for non-cospecral (non Laplacian cospectral) graphs.\ The extended double cover of the graph \(G(n,m)\) with vertex set \(V(G)=\{v_1, v_2, \cdots, v_n\}\) is a bipartite graph \(G^*\) with bipartition (X, Y), \(X=\{x_1, x_2, \cdots, x_n\}\) and \(Y=\{y_1, y_2, \cdots, y_n\}\), where two vertices \(x_i\) and \(y_j\) are adjacent if and only if \(i=j\) or \(v_i\) adjacent \(v_j\) in \(G\). It is easy to see that \(G^*\) is connected if and only if \(G\) is connected and a vertex \(v_i\) is of degree \(d_i\) in \(G\) if and only if it is of degree \(d_i+1\) in \(G^*\). Also the extended double cover \(G^*\) of the graph \(G\) always contains a perfect matching. The double graph \(D[G]\) of \(G\) is a graph obtained by taking two copies of \(G\) and joining each vertex in one copy with the neighbours of corresponding vertex in another copy. The \(k\)-fold graph \(D^k[G]\) of the graph \(G\) is obtained by taking \(k\) copies of the graph \(G\) and joining each vertex in one of the copy with the neighbours of the corresponding vertices in the other copies. If \(T_n\) is the graph obtained from the complete graph \(K_n\) by adding a loop at each of the vertex, it is easy to see that \(D^k[G]=G\otimes T_k\). In this paper we study energy, Laplacian energy of the graphs \(G^*\) and \(D[G]\), the \(L\)-spectra of \(G^{k*}\) the \(k\)-th iterated extended double cover of \(G\) and obtain a formula for the number of spanning trees of \(G^*\). We also obtain some new families of the equienergetic and \(L\)-equienergetic graphs.\ We denote the complement of graph \(G\) by \(\bar{G}\), the complete graph on \(n\) vertices by \(K_n\), the empty graph by \(\bar{K_n}\) and the complete bipartite graph with cardinalities of partite sets \(q\) and \(r\) by \(K_{q,r}\). The rest of the paper is organised as follows. In section 2, energy of the graphs \(G^*\) and \(D^k[G]\) are obtained and some new families of equienergetic graphs are given, in section 3 \(L\)-spectra of \(G^{k*}\) and a formula for the number of spanning tress of \(G^*\) is obtained and in section 4 Laplacian energy of the graphs \(G^*\) and \(D^k[G]\) and the construction of some new families of \(L\)-equienergetic graphs by using the graphs \(G^{k*}\) and \(D^k[G]\) is presented. # energy of double graphs In this section we find the energy of the graphs \(G^*\) and \(D^k[G]\). We also construct some new families of equienergetic graphs based on these graphs.\ For the graphs \(G_1\) and \(G_2\) with disjoint vertex sets \(V(G_1)\) and \(V(G_2)\), the *cartesian product* is a graph \(G=G_1\times G_2\) with vertex set \(V(G_1)\times V(G_2)\) and an edge \(((u_1, v_1), (u_2, v_2))\) if and only if \(u_1=u_2\) and \((v_1, v_2)\) is an edge of \(G_2\) or \(v_1=v_2\) and \((u_1, u_2)\) is an edge of \(G_1\). The following result gives the \(A\)-spectra (\(L\)-spectra) of the cartesian product of graphs. The *conjunction (Kronecker product)* of \(G_1\) and \(G_2\) is a graph \(G=G_1\otimes G_2\) with vertex set \(V(G_1)\times V(G_2)\) and an edge \(((u_1,v_1), (u_2, v_2))\) if and only if \((u_1, u_2)\) and \((v_1, v_2)\) are edges in \(G_1\) and \(G_2\), respectively. The following result gives the \(A\)-spectra (\(L\)-spectra) of the Kronecker product of graphs. The *join (complete product)* of \(G_1\) and \(G_2\) is a graph \(G=G_1\vee G_2\) with vertex set \(V(G_1)\cup V(G_2)\) and an edge set consisting of all the edges of \(G_1\) and \(G_2\) together with the edges joining each vertex of \(G_1\) with every vertex of \(G_2\). The \(L\)-spectra of join of graphs is given by the following result. The following result gives the \(A\)-spectra of \(G^*\), the extended double cover of the graph \(G\). If \(\lambda_1, \lambda_2, \cdots, \lambda_n\) is the \(A\)-spectra of the graph \(G\), then by Lemma 2.1, the \(A\)-spectra of the graph \(G\times K_2\) is \(\lambda_i+1, \lambda_i-1\) for \(1\leq i \leq n\). It is clear from Theorem 2.4, that the graphs \(G\times K_2\) and \(G^*\) are cospectral if and only if \(G\) is bipartite. If \(D^k[G]\) is the \(k\)-fold graph of the graph \(G\), the \(A\)-spectra of \(D^k[G]\) is given by the following result from. If \(\lambda_1, \lambda_2, \cdots, \lambda_n\) is the \(A\)-spectra of the graph \(G\), then by Theorem 2.4, the \(A\)-spectra of the graph \(G^*\) is \(\pm(\lambda_1+1), \pm(\lambda_2+1), \cdots, \pm(\lambda_n+1)\) and by Theorem 2.5, the \(A\)-spectra of \(D^k[G]\) is \(k\lambda_1, k\lambda_2, \cdots, k\lambda_n, 0\) (\((k-1)n\) times). Therefore, \[\begin{aligned} E(G^*)=\sum\limits_{i=1}^{n}|\lambda_i+1|+\sum\limits_{i=1}^{n}|-\lambda_i-1|=2\sum\limits_{i=1}^{n}|\lambda_i+1|, \end{aligned}\] and \[\begin{aligned} E(D^k[G])=\sum\limits_{i=1}^{n}|2\lambda_i|=2\sum\limits_{i=1}^{n}|\lambda_|=kE(G). \end{aligned}\] If \(\lambda_1, \lambda_2, \cdots, \lambda_n\) is the \(A\)-spectra of a graph \(G\), then the \(A\)-spectra of the graph \((G\otimes K_2)\times K_2\) is \(\lambda_i+1, \lambda_i-1,-\lambda_i+1,-\lambda_i-1\), \(1\leq i\leq n\). Therefore, \[\begin{aligned} E((G\otimes K_2)\times K_2)=2\sum_{i=1}^{n}|\lambda_i+1|+2\sum_{i=1}^{n}|\lambda_i-1|=2\left(\sum_{i=1}^{n}|\lambda_i+1|+\sum_{i=1}^{n}|\lambda_i-1| \right)\\=2E(G\times K_2)=E(2(G\times K_2))=E((G\times K_2)\cup (G\times K_2)). \end{aligned}\] From the above discussion, we observe that the graphs \((G\otimes K_2)\times K_2\) and \((G\times K_2)\cup (G\times K_2)\) are equienergetic. Moreover, if the graph \(G\) is a bipartite graph then the graphs \((G\otimes K_2)\times K_2\) and \(G^*\cup G^*\) are also equienergetic graphs.\ As seen above \(E(D^k[G])=k\sum_{i=1}^{n}|\lambda_i|=kE(G)=E(kG)\)=\(E(G\cup G\cup\cdots\) \(k\) copies). This shows that the graphs \(D^k[G]\) and \((G\cup G\cup\cdots\) \(k\) copies) are non-cospectral equienergetic. However, we show for any graph \(G\) the graphs \(D[G]\) and \(G\otimes K_2\) are always equienergetic non-cospectral graphs. In general, if \(D^k[G]\) be the \(k\)-fold graph of the graph \(G\), we have the following observation. Let \(G^{**}\) be the extended double cover of the graph \(G^*\). We have the following result. Let \(G\) be a bipartite graph. It is well known that the spectra of \(G\) is symmetric about the origin, so half of the non-zero eigenvalues of \(G\) lies to the left and half lies to the right of origin. Therefore if \(G\) is a bipartite graph having all its eigenvalues non-zero, the number of positive and negative eigenvalues of \(G\) are same. Keeping this in mind we have the following result. We can also prove Theorem 2.9 by using Theorem 2.6, the fact that the graphs \(G^*\) and \(G\times K_2\) are cospectral if \(G\) is bipartite (Theorem 2 in ) and the graphs \(G\times K_2\) and \(G\otimes K_2\) are equienergetic if an only if \(|\lambda_i|\geq 1\) (Theorem 8 in ). # The Laplacian spectra of \(G^{k*}\) Let \(G^*\) be the extended double cover of the graph \(G\), define \(G^{**}=(G^*)^*\), and in general \(G^{k*}=(G^{(k-1)*})^*\), \(k\geq 1\), called the \(k\)-th iterated double cover graph of \(G\). The \(A\)-spectra of \(G^{k*}\) is given in. Here we obtain the \(L\)-spectra of the \(k\)-th iterated extended double cover \(G^{k*}\) of the graph \(G\). Since the graph \(G^{k*}\) is always bipartite for \(k\geq 1\), therefore its Laplacian (\(L\)-spectra) and signless Laplacian (\(Q\)-spectra) spectra are same.\ For any complex square matrices \(A\) and \(B\) of same order, the following observation can be seen in ( p.no.41). We first obtain the \(L\)-spectra of \(G^*\), the extended double cover of the graph \(G\), in the following result. We now obtain the \(L\)-spectra of \(G^{k*}\) as follows. If \(G\) is a bipartite graph, it is easy to see that under elementary transformation the Laplacian characteristic polynomial of \(G\) coincides with the signless Laplacian characteristic polynomial of \(G\). Therefore the Laplacian and signless Laplacian spectra of \(G\) are same. We have the following observation. In three formulae are given for the number of spanning trees of \(G^*\) in terms of \(A\)-spectra of the corresponding graph \(G\). We now obtain a formula for the number of spanning trees in terms of the \(L\) and \(Q\)-spectra of \(G^*\). In case \(G\) is bipartite, \(\mu_i=\mu_i^+\), so we have \[\begin{aligned} \tau(G^*)=\frac{1}{2n}\prod\limits_{i=1}^{n-1}\mu_i\prod\limits_{i=1}^{n}(\mu_i+2)=\tau(G)\prod\limits_{i=1}^{n-1}(\mu_i+2). \end{aligned}\] In it is shown that the graphs \(G^*\) and \(G\times K_2\) are \(A\)-cospectral if and only if \(G=K_1\) or \(G\) is bipartite. An analogous result holds for the \(L\)-spectra and is given below. An *integral graph* is a graph all of whose eigenvalues are integers. Following observation is a consequence of Theorem 3.3. It is clear from Theorem 3.7, that given a Laplacian integral \(G\) it is always possible to construct an infinite sequence of Laplacian integral graphs. Indeed the graph \(G^{k*}\) is Laplacian integral for all \(k\geq 1\).\ Two graphs \(G_1\) and \(G_2\) are said to be co-spectral, if they are non-isomorphic and have the same spectra. We have the following result, which follows by Theorem 3.3. Thus given two Laplacian co-spectral graphs \(G_1\) and \(G_2\), it is always possible to construct an infinite sequence of Laplacian co-spectral graphs. Indeed the graphs \(G_1^{k*}\) and \(G_2^{k*}\) are Laplacian co-spectral for all \(k\geq 1\).\ Since the extended double cover \(G^*\) of the graph \(G\) is always bipartite, it follows by Theorem 3.6, the graphs \(G^{**}\) and \(G^*\times K_2\) are \(L\)-cospectral and in general the graphs \(G^{s*}\) and \(G^{(s-1)*}\times K_2\) are \(L\)-cospectral. Also it is easy to see that the graphs \((G\times K_2)^*\) and \(G^*\times K_2\) are \(L\)-cospectral and in general the graphs \((G\times K_2)^{s*}\) and \(G^{s*}\times K_2\) are both \(L\)-cospectral as well as \(Q\)-cospectral. Moreover, if \(G\) is bipartite then as seen in Theorem 3.6, the graphs \(G^*\) and \(G\times K_2\) are \(L\)-cospectral. Using the same argument it can be seen that the graphs \(G^{**}\) and \(G\times K_2\times K_2\) are \(L\)-cospectral if and only if \(G\) is bipartite. A repeated use of the argument as used in Theorem 3.6, gives the graphs \(G^{s*}\) and \((G\times K_2\times K_2\times\cdots s-times)=(G\times sK_2)=(G\times Q_s)\) are \(L\)-cospectral if and only if \(G\) is bipartite. From this discussion it follows that the graphs \(G^{s*}\), \(G^{(s-1)*}\times K_2\), \((G\times K_2)^{(s-1)*}\) and \(G\times Q_{s-1}\) are mutually non-isomorphic \(L\)-cospectral graphs if and only \(G\) is bipartite, where \(Q_n\) is the hypercube. # Laplacian energy of double graphs In this section, we study the Laplacian energy of the graphs \(D[G]\), \(D^k[G]\) and \(G^*\). Using these graphs we obtain some new families of non Laplacian cospectral \(L\)-equienergetic graphs. Let \(D[G]\) and \(G^*\) be respectively the double graph and the extended double cover of the graph \(G\). Then the Laplacian spectra of the graph \(G^*\) is given by Theorem 3.2, and the Laplacian spectra of \(D^k[G]\) is given by the following Theorem. Let \(\mu_i\) for \(1\leq i \leq n\) be the \(L\)-spectra of the graph \(G\). Then by Theorem 3.2, the \(L\)-spectra of the extended double cover \(G^*\) of the graph \(G\) is \(\mu_i, \mu_i^{+}+2\) for \(1\leq i\leq n\). Also the average vertex degree of \(G^*\) is \(\frac{2m}{n}+1\). Therefore, \[\begin{aligned} LE(G^*)=\sum\limits_{i=1}^{n}|\mu_i-\frac{2m}{n}-1|+\sum\limits_{i=1}^{n}|\mu_i^{+}-\frac{2m}{n}+1|. \end{aligned}\] Since average vertex degree of \(D^k[G]\) is \(k\frac{2m}{n}\), we have \[\begin{aligned} LE(D^k[G])&=\sum\limits_{i=1}^{n}|k\mu_i-k\frac{2m}{n}|+(k-1)\sum\limits_{i=1}^{n}|kd_i-k\frac{2m}{n}|\\& =k\sum\limits_{i=1}^{n}|\mu_i-\frac{2m}{n}|+k(k-1)\sum\limits_{i=1}^{n}|d_i-\frac{2m}{n}|\\& =kLE(G)+k(k-1)\sum\limits_{i=1}^{n}|d_i-\frac{2m}{n}|. \end{aligned}\] From this it is clear that \(LE(D^k[G])=k LE(G)\), if \(G\) is regular. Also, since the \(k\)-fold graph of a regular graph is regular, it follows, if \(G_1\) and \(G_2\) are \(r\)-regular \(L\)-equienergetic graphs then their \(k\)-fold graphs \(D^k[G_1]\) and \(D^k[G_2]\) are also \(L\)-equienergetic. Let \(\pounds(G)\) be the line graph of the graph \(G\). It is shown in that if \(G_1\) and \(G_2\) are \(r\)-regular graphs then their \(k\)-th, (\(k\geq 2\)) iterated line graphs \(\pounds^k(G_1)\) and \(\pounds^k(G_2)\) are always equienergetic and so \(L\)-equienergetic. Therefore it follows that given any two \(r\)-regular graphs, we can always construct an infinite family of \(L\)-equienergetic graphs.\ In case the given \(r\)-regular connected graphs are \(L\)-equienergetic, the \(k\)-fold graph forms a larger family of \(L\)-equienergetic graphs than the \(k\)-th iterated line graph. As an example, consider the \(4\)-regular graphs \(G_1\) and \(G_2\) shown in Figure 1 on \(9\)-vertices. It can be seen that the \(L\)-spectra of \(G_1\) and \(G_2\) are respectively as \(0, 3^4, 6^4\) and \(0, 2, 3^2, 5^2, 6^3\) (where \(a^s\) means \(a\) occurs \(s\) times in the spectrum). Therefore \(LE(G_1)=16=LE(G_2)\). This shows that the graphs \(G_1\) and \(G_2\) are regular \(L\)-equienergetic graphs, so their \(k\)-fold graphs \(D^k[G_1]\) and \(D^k[G_2]\) and their \(k\)-th, (\(k \geq 2\)) iterated line graphs are also \(L\)-equienergetic. Infact the \(k\)-fold graph gives an infinite family of \(L\)-equienergetic graph pairs of order \(n\equiv 0(\mod 9)\), whereas the \(k\)-th iterated line graph gives an infinite family of \(L\)-equienergetic graph pairs of orders \(n=54, 270, 2430\), and so on, from this the assertion follows.\ We have seen that the Laplacian energy of the graph \(D[G]\) is twice the Laplacian energy of \(G\) when \(G\) is regular. But this need not be true for the graph \(G^*\) as seen from the Laplacian energy of \(G^*\) given above. However we have the following observation. If \(G\) is a graph satisfying the conditions of Theorem 4.2, then clearly the graphs \(G^*\) and \(G\cup G\) are \(L\)-equienergetic. We now obtain some new families of \(L\)-equienergetic graphs by means of the graphs \(G^*\), \(G^{k*}\), \(D[G]\) and \(D^k[G]\). Let \(G^{t*}\) be the \(t\)-th iterated extended double cover of the graph \(G\). We have the following generalization of Theorem 4.3. Theorem 4.4 gives an infinite family of \(L\)-equienergetic graphs in various ways, firstly fix the value of \(t\) and allow \(p\) to vary we obtain families of \(L\)-equienergetic graphs with same \(t\), secondly fix the value of \(p\) and allow \(t\) to vary we obtain families of \(L\)-equienergetic graphs with same \(p\) and so on. From Theorem 4.4, it is clear if \(G_1\) and \(G_2\) are any two graphs with the same parameters, then we can always find tripartite graphs \((G_1^*\vee \bar{K_p})\) and \((G_2^*\vee \bar{K_p})\) having the same Laplacian energy. Next we show the construction of \(L\)-equienergetic graphs by means of graphs \(D[G]\) and \(D^k[G]\). If \(D^k[G]\) is the \(k\)-fold graph of the graph \(G\), we have the following generalization of Theorem 4.6. Theorem 4.7, generates families of \(L\)-equienergetic graphs in various ways. If we allow \(p\) to vary and keep \(k\) fixed, we obtain an infinite family of \(L\)-equienergetic graphs with same \(k\) and if we allow \(k\) to vary and keep \(p\) fixed, we obtain an infinite family of \(L\)-equienergetic graphs with same \(p\) and so on.\ If \(D[G]\) and \(G^*\) are respectively the double graph and the extended double cover of the graph \(G\), then the following result gives the construction of \(L\)-equienergetic graphs with different number of edges. Let \(D[G_1]\) be the double graph of the graph \(G_1(n,m_1)\) and let \(G_2^*\) be the extended double cover of the graph \(G_2(n,m_2)\), then for \(p\geq 2n+k\) and \(m_1\leq \frac{k(2n+k)}{8}\), \(k\geq 4\), we have from Theorem 4.6 \[\begin{aligned} LE(D[G_1]\vee \bar{K_p})=4n+8m_1+(p-2n)\frac{2m_1^{\prime}}{n^{\prime}}. \end{aligned}\] Also, for \(p\geq 2n+k\) and \(m_2\leq \frac{n(k-1)}{2}+\frac{k^2}{4}\), \(k\geq 4\), we have by Theorem 4.3 \[\begin{aligned} LE(G_2^*\vee \bar{K_p})=6n+4m_2+(p-2n)\frac{2m_2^{\prime}}{n^{\prime}}. \end{aligned}\] If we suppose that \(4m_1=2m_2+n\), then it follows from (4.1) and (4.2) that \[LE(D[G_1]\vee \bar{K_p})=LE(G_2^*\vee \bar{K_p}).\] This gives another construction of families of graphs with same Laplacian energy, same number of vertices but different number of edges. Next we give another way of constructing a family of graphs having same number of vertices, same Laplacian energy but different number of edges. Theorem 4.9 generates \(L\)-equienergetic graphs with same number of vertices but different number of edges, infact when one graph contains twice the number of edges as contained in other. Lastly we give the construction of family of graphs with same number of vertices, edges and Laplacian energy by means of cartesian product and extended double cover. Since \(G^*\) is always bipartite, Theorem 4.10 gives the construction of connected graphs from a given pair of \(L\)-equienergetic bipartite graphs having same number of vertices, edges and Laplacian energy. Moreover if \(t\) is the first value of \(p\) satisfying the conditions in Theorem 4.10, then every value greater than \(t\) also satisfies this condition, therefore we obtain an infinite family of \(L\)-equienergetic graph pairs.

{'timestamp': '2013-10-14T02:09:52', 'yymm': '1310', 'arxiv_id': '1310.3204', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3204'}

# Introduction Numerous spectroscopic studies of metal-poor (\[Fe/H\] \(<-1.0\)) candidates identified by the HK survey (Beers et al. 1985, 1992) and the Hamburg/ESO Survey (HES; Wisotzki et al. 1996; Christlieb et al. 2001, 2008; Christlieb 2003) have revealed that the frequency of carbon-enhanced stars increases strongly with decreasing \[Fe/H\]. These stars, now known as carbon-enhanced metal-poor (CEMP) stars, were originally defined as stars with metallicity \[Fe/H\] \(\leq-1.0\) and carbon-to-iron ratios \[C/Fe\] \(\geq +1.0\) (Beers & Christlieb 2005). [^1] Generally, the frequency of C-rich stars increases from a few percent at higher metallicity to on the order of 20% for \[Fe/H\] \(<-2.0\), 30% for \[Fe/H\] \(<-3.0\), 40% for \[Fe/H\] \(<-3.5\), and 75% for \[Fe/H\] \(<-4.0\) (Beers et al. 1992; Norris et al. 1997; Rossi et al. 1999; Beers & Christlieb 2005; Cohen et al. 2005; Marsteller et al. 2005; Rossi et al. 2005; Frebel et al. 2006; Lucatello et al. 2006; Norris et al. 2007; Carollo et al. 2012; Norris et al. 2013; Spite et al. 2013; Yong et al. 2013). This increasing trend of CEMP-star frequency with declining \[Fe/H\] is again confirmed from the many thousands of CEMP stars found among the several hundred thousand stars with available spectra from the Sloan Digital Sky Survey (SDSS; Fukugita et al. 1996; Gunn et al. 1998, 2006; York et al. 2000; Stoughton et al. 2002; Abazajian et al. 2003, 2004, 2005, 2009; Pier et al. 2003; Adelman-McCarthy et al. 2006, 2007, 2008; Aihara et al. 2011; Ahn et al. 2012) and the Sloan Extension for Galactic Understanding and Exploration (SEGUE-1; Yanny et al. 2009), and SEGUE-2 (C. Rockosi et al., in preparation) as described by Lee et al. (2013). There exist a number of subclasses within the CEMP classification, as originally defined by Beers & Christlieb (2005), which may provide direct clues to the nature of their likely progenitors. Stars in the CEMP-\(s\) subclass exhibit over-abundances of \(s\)(low)-process elements such as Ba and Sr, the CEMP-\(r\) subclass includes stars with enhanced \(r\)(apid)-process elements such as Eu, and the CEMP-\(r/s\) stars exhibit elemental abundance patterns associated with both the \(r\)-process and the \(s\)-process. The CEMP-no subclass exhibits no over-abundances of the neutron-capture elements. The CEMP-\(s\) (and CEMP-\(r/s\)) subclasses of CEMP stars are the most commonly found to date; high-resolution spectroscopic studies show that around 80% of the CEMP stars are categorized as CEMP-\(s\) (or CEMP-\(r/s\)) (Aoki et al. 2007, 2008). The favored mechanism for the production of the high \[C/Fe\] ratios found for CEMP-\(s\) (CEMP-\(r/s\)) stars is mass transfer of carbon-enhanced material from the envelope of a now-defunct asymptotic giant-branch (AGB) star to its (presently observed) binary companion (Suda et al. 2004; Herwig 2005; Komiya et al. 2007; Sneden et al. 2008; Masseron et al. 2010; Bisterzo et al. 2011, 2012). Observational evidence now exists to suggest that the CEMP-\(r/s\) stars (and other \(r\)-process-element rich stars) were enhanced in \(r\)-process elements in their natal gas clouds by previous generations of supernovae (SNe), and did not require a contribution of \(r\)-process elements from a binary companion (see Hansen et al. 2013). The limited amount of long-term radial-velocity monitoring available for CEMP stars indicates variations for almost all of the CEMP-\(s\) stars, confirming their binary status (Lucatello et al. 2005a). In addition, the CEMP-\(s\) stars are mostly, though not exclusively (e.g., Norris et al. 2013 and references therein), found among metal-poor stars with \[Fe/H\] \(>-3.0\). On the other hand, CEMP-no stars are found most commonly among the extremely metal-poor (EMP) stars, with \[Fe/H\] \(<-3.0\) (Aoki et al. 2007; Norris et al. 2013). Existing radial-velocity monitoring of these objects indicates that they are found in binary systems no more frequently than other metal-poor stars (T. Hansen et al., in preparation). Norris et al. (2013) found no CEMP-\(s\) stars among 18 CEMP stars with \[C/Fe\] \(\geq +0.7\) and \[Fe/H\] \(<-3.1\), as well as no discernible variations of their radial velocities. Although there is general consensus on the origin of CEMP-\(s\) stars, the likely progenitor or progenitors of the CEMP-no stars are still under discussion. Suggested models include massive, rapidly rotating, mega metal-poor (MMP; \[Fe/H\] \(<-6.0\)) stars, which produce large amounts of C, N, and O due to distinctive internal burning and mixing episodes (Meynet et al. 2006, 2010; Chiappini 2013), and faint (low-energy) SNe associated with the first generations of stars, which experience extensive mixing and fallback during their explosions, and eject large amounts of C and O, but not heavier metals (Umeda & Nomoto 2003, 2005; Tominaga et al. 2007, 2013; Ito et al. 2009, 2013; Nomoto et al. 2013). Nevertheless, the origin of the CEMP-no star phenomenon is yet to be fully resolved (see Norris et al. 2013, which summarizes other possible progenitors of the C-rich stars). Previous authors have attempted to understand the large fractions of CEMP stars at low metallicity, as well as the different subclasses of the CEMP stars, by invoking AGB models with different masses. Furthermore, there have been several efforts to constrain the form of the early initial mass function (IMF) by reproducing the observed frequencies of CEMP stars, as well as the number ratios of the different CEMP subclasses. Abia et al. (2001), for example, claimed that the large number of carbon-enhanced stars found among stars of very low metallicity could be accounted for if the IMF in the early history of the Galaxy was dominated by higher mass stars. Lucatello et al. (2005b) and Komiya et al. (2007) utilized population-synthesis models with an IMF biased towards massive stars to compare with the fractions of observed CEMP stars, and concluded that an IMF comprising a larger number of intermediate-to high-mass stars could reproduce the larger fraction of the CEMP stars among metal-poor stars (\[Fe/H\] \(<-2.5\)) better than the present-day (Salpeter) IMF. Recently, Suda et al. (2013) made use of the number ratios of CEMP/EMP, CEMP-no/CEMP, and NEMP/CEMP giant stars (where NEMP stands for nitrogen-enhanced metal-poor) from the Stellar Abundances for Galactic Archeology (SAGA; Suda et al. 2008) database to constrain the parameters in their binary population-synthesis model. They considered several IMFs, and proposed that the IMF changed from high-mass dominated in the early Galaxy to low-mass (\(M < 0.8\) \(M_\odot\)) dominated at present, and that this transition occurred around a metallicity of \[Fe/H\] \(\sim-2.0\). The above studies carried out comparisons of the CEMP fractions derived from small samples of stars comprising mostly giants. However, observational evidence indicates that the C-rich material at the surface of a giant could be easily depleted by extra mixing of CNO-processed material from its interior during the so-called first dredge-up episode (Spite et al. 2005, 2006; Lucatello et al. 2006; Aoki et al. 2007). It is also known that more luminous red giant-branch (RGB) stars are more affected by such mixing (Spite et al. 2005, 2006). Therefore, if such mixing does occur, the overall CEMP frequencies as estimated from giants are expected to be a lower limit. In order to avoid this complication, the best way forward would appear to be comparing model predictions with the observed CEMP frequencies based on unevolved stars, such as dwarfs or main-sequence turnoff stars. The current CEMP stars that have been studied with high-resolution spectroscopy are mostly giants (for which it is simpler to obtain high S/N spectra, due to their relative brightness and moderate temperatures, which allows for lines of interest to be measured with less uncertainty). The numbers of observed dwarf and turnoff stars with similar observations are in any case too small to derive statistically meaningful results for different subclasses of CEMP stars. In this study, we make use of stars with available carbon-to-iron ratios (\[C/Fe\]) and \[Fe/H\], based on medium-resolution (\(R \sim 2000\)) spectroscopy obtained during the course of the SDSS, SEGUE-1, and SEGUE-2, in order to derive accurate frequencies of CEMP stars among giants and turnoff stars as a function of \[Fe/H\]. The derived CEMP frequencies are then compared with the predictions from AGB binary-synthesis models that employ the two different IMFs explored by Suda et al. (2013). The results of these comparisons should provide more stringent constraints on the IMF of the Milky Way, and clues to the existence of progenitors *other than AGB stars* that are capable of producing large amounts of carbon-enhanced material in the early universe. This paper is outlined as follows. In Section 2, we describe the selection criteria used to assemble the sample for this study. Section 3 presents and discusses results of the comparison of the CEMP frequencies for giants and main-sequence turnoff stars with the binary-synthesis model predictions, and describes a procedure for correcting the anticipated undercounts of CEMP stars among warm, metal-poor turnoff stars. Our conclusions are presented in Section 4. # Carbon-enhanced SDSS/SEGUE Stars The SDSS, SEGUE-1, and SEGUE-2 surveys have produced an unprecedented number of high-quality medium-resolution stellar spectra, covering stars in various evolutionary stages, and spanning a wide range of metallicity (\(-\)`<!-- -->`{=html}4.0 \(<\) \[Fe/H\] \(< +0.5\)). A total of about 600,000 stars are potentially suitable for examination of the properties of the Milky Way's stellar populations. The resolving power of the spectra is \(R \sim 2000\), over the wavelength range 3820--9100 Å. Below we simply refer to these stars (spectra) as SDSS/SEGUE stars (spectra). Accurate estimates of the atmospheric parameters for most of the SDSS/SEGUE stars are derived using the latest version of the SEGUE Stellar Parameter Pipeline (SSPP; Lee et al. 2008a, 2008b, 2011; Allende Prieto et al. 2008; Smolinski et al. 2011). The typical external errors obtained by the SSPP are 180 K for \(T_{\rm eff}\), 0.24 dex for \(\log~g\), and 0.23 dex for \[Fe/H\], respectively (Smolinski et al. 2011). In addition, estimates of the carbonicity, \[C/Fe\], is obtained following the prescription of Lee et al. (2013), for stars with \(4400 \leq\) \(T_{\rm eff}\) \(\leq\) 6700 K, where accurate \[C/Fe\] can be determined. As reported by Lee et al., uncertainties in the determination of \[C/Fe\] are smaller than 0.35 dex for SDSS/SEGUE spectra with S/N \(\geq 15\) Å\(^{-1}\). In order to derive reliable frequencies of the CEMP stars among the field stellar populations, we follow the selection criteria of Lee et al. (2013). Briefly, we first exclude all stars located in the directions of known open and globular clusters. For stars that were observed more than once, we keep only the parameters derived from the highest S/N spectrum. We then restrict the sample to stars with spectra having S/N \(\geq 20\) Å\(^{-1}\), effective temperatures in the range 4400 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 6700 K, and metallicities in the range \(-\)`<!-- -->`{=html}4.0 \(\leq\) \[Fe/H\] \(\leq +0.5\), so that our estimates of \[C/Fe\] are as reliable as possible. We then visually inspect each spectrum with \[Fe/H\] \(\leq-2.0\), in order to reject spectra such as cool white dwarfs, or those with emission-line features in the cores of their Ca ii lines, or other spectral defects that could lead to spurious determinations of metallicity by the SSPP. Additionally, we visually examine the spectra for all stars with \[C/Fe\] \(\geq +0.7\), and exclude stars with poor estimates of \[Fe/H\] and/or \[C/Fe\]. Following Lee et al., for the purpose of deriving the CEMP frequencies we do not include stars with \[C/Fe\] \(\ge +0.7\) and indication of upper limit estimate, and consider these stars to have unknown carbon status. By application of the above procedures, we are left with a sample of about 247,350 stars. For the purpose of our analysis, we consider stars with 4400 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 5600 K and 1.0 \(\leq\) \(\log~g\) \(<\) 3.2 as giants, and stars with 5600 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 6600 K and 3.2 \(\leq\) \(\log~g\) \(<\) 4.5 as main-sequence turnoff stars. The distance to each star is estimated following the prescriptions of Beers et al. (2000, 2012), which obtains photometric distance estimates with errors on the order of 10%--20%. Note that we have added stars to our sample from Table 1 of Yong et al. (2013), and one object from Caffau et al. (2011), with determinations based on high-resolution spectroscopic analyses, in order to increase the number of stars with \[Fe/H\] \(<-3.0\). This results in better number statistics for the calculation of the CEMP frequencies in the extremely and ultra metal-poor regime. # Results and Discussion ## Differences in Average \[C/Fe\] Between Giants and Turnoff Stars Figure [\[fig:cfedist\]](#fig:cfedist){reference-type="ref" reference="fig:cfedist"} shows the distribution of \[C/Fe\] for SDSS/SEGUE stars in various bins of metallicity, decreasing from the upper to lower panels. Main-sequence turnoff stars are shown as solid histograms; giants are shown as dashed histograms. Inspection of this figure reveals that the overall distribution of \[C/Fe\] gradually shifts (for both turnoff stars and giants) to higher \[C/Fe\] with decreasing \[Fe/H\], with a tail extending toward higher \[C/Fe\] appearing as the metallicity decreases. As the metallicity decreases below \[Fe/H\] \(=-3.0\), this trend continues for the turnoff stars, but it is not as evident for the giants. Another interesting feature seen in Figure [\[fig:cfedist\]](#fig:cfedist){reference-type="ref" reference="fig:cfedist"} is that, for \[Fe/H\] \(<-3.0\), the turnoff stars are distributed over a wide range of \[C/Fe\], whereas the giants are mostly concentrated in the region of \[C/Fe\] \(< +1.0\). The red vertical lines in Figure [\[fig:cfedist\]](#fig:cfedist){reference-type="ref" reference="fig:cfedist"} indicate the mean values of \[C/Fe\] for the turnoff stars (solid line) and giants (dashed line), respectively. On average, the giants appear to exhibit lower carbonicity (by about 0.2 dex) than the turnoff stars, down to \[Fe/H\] = \(-2.5\). The mean value of \[C/Fe\] appears to increase with decreasing metallicity, as also found by Carollo et al. (2012, their Figure 11). They reported that the degree of carbon enhancement significantly increased from \[C/Fe\] \(\sim +1.0\) at \[Fe/H\] = \(-1.5\) to \[C/Fe\] \(\sim +1.7\) at \[Fe/H\] = \(-2.7\), somewhat higher than our values (it should be noted that a different sample of stars, as well as a different method for determination of \[C/Fe\], were employed by these authors). The difference in the distribution of \[C/Fe\] between the turnoff stars and the giants may be explained by the different masses of the convective envelopes between the two evolutionary stages. Because a giant has a much larger convective envelope, its surface material experiences more mixing, leading to reduction of the carbonicity. On the other hand, the turnoff stars have shallower convective envelopes, so that their surface abundances may not be expected to greatly change. As a result, the overall carbon abundance for the turnoff population is expected to be higher than that of the giant population, even if they were born with the same initial carbon abundance. Bonifacio et al. (2009) also noted a difference in the mean \[C/Fe\] between giants and turnoff stars of similar metallicities, finding a difference of about 0.2 dex (giants being lower) for stars with \[Fe/H\] \(<-2.5\). They argued that this difference arises because the giants suffer from extra mixing due to first dredge-up, and have their surface carbon abundance reduced. They also suggested the stellar models employed in the analysis could contribute to this discrepancy; they found a smaller difference between the giants and turnoff stars when deriving \[C/Fe\] from a 3D, rather than a 1D model atmosphere. ## Comparison with Model Predictions for the Frequency of C-rich Giants The black filled circles in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"} represent our derived differential frequencies for C-rich giants as a function of \[Fe/H\]. From inspection of this figure, it appears that the CEMP frequencies do not increase for \[Fe/H\] \(<-2.5\), but rather, remain relatively constant, in contrast to the results of previous studies. As discussed by Spite et al. (2005, 2006), Lucatello et al. (2006), and Aoki et al. (2007), this may be in part due to CN-processing, which converts carbon to nitrogen at the bottom of a star's convection zone, and in turn reduces the carbon abundance in the envelope at the time of first dredge-up. Whether or not a giant experiences such mixing can be identified by measuring its \(^{12}\)C/\(^{13}\)C or \[C/N\] ratios, as both will be lower for an object that has gone through such an event. Unfortunately, these ratios are difficult to assess from the SDSS/SEGUE spectra over the full range of metallicities we consider. In the metallicity regime of \[Fe/H\] \(>-1.5\), our derived CEMP frequencies of \(\sim\)`<!-- -->`{=html}1% agree with the previously claimed fraction of classical CH or Ba stars in the solar neighborhood (Luck & Bond 1991). Regarding the effect of the first dredge-up, it is worth mentioning the following arguments from a theoretical point of view. According to Suda et al. (2004), first dredge-up might not play a significant role in decreasing the carbon abundance on the surface of a giant, based on an 0.8 \(M_\odot\) model for HE 0107-5240 (a CEMP-no star with \[Fe/H\] = \(-5.3\); Christlieb et al. 2002). They examined the effect of the first dredge-up following the accretion of C-and O-rich matter onto the star (here assuming an AGB mass-transfer scenario), and found that if its envelope is significantly C-rich (\[C/Fe\] \(\gg +1.0\)), then after first dredge-up the surface carbon abundance changed little. This is because the surface carbon abundance is too large (\[C/H\] \(\sim-1\)) prior to the dredge-up to be significantly depleted by the relatively small amount of matter in the hydrogen-burning shell, \(M <\) 0.02 \(M_\odot\). Even in cases of EMP stars for which the initial carbon abundance is small, Suda & Fujimoto (2010) showed that the effect of the first dredge-up is also limited, due to the shallower convective envelopes in metal-poor (as compared to metal-rich) stars. According to their model calculation, the change of the CNO abundance before and after the first dredge-up was on the order of one percent. Therefore, they did not notice a large impact on the surface carbon abundances after first dredge-up for stars with \[Fe/H\] \(<-2.3\). The blue open circles in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"} are the predicted frequencies of CEMP giants, as a function of metallicity, from AGB binary-synthesis models with an IMF peaked at 10 \(M_\odot\), while the open squares are the predicted frequencies from models using a Salpeter IMF, adopted from Suda et al. (2013). In their model, they included a mechanism referred to as "pulsation-driven mass loss" (Wood 2011), which was argued to suppress the previously predicted over-production of NEMP stars by Izzard et al. (2009) and Pols et al. (2012). It appears that the predicted CEMP frequencies for the high-mass dominated IMF are in relatively good agreement with the observed CEMP giant frequencies for \[Fe/H\] \(<-2.5\), but the model predicts too many C-rich stars above \[Fe/H\] = \(-2.5\). The predicted CEMP frequencies from a Salpeter IMF are in good agreement with our derived frequencies for the metal-rich region (\[Fe/H\] \(>-1.5\)), while the predicted CEMP frequencies are far too low for stars with \[Fe/H\] \(<-2.5\). These are similar results to those found by Suda et al. (2013), who used the giants in the SAGA database to compare the observed CEMP frequencies with their model predictions. One of the reasons that Suda et al. (2013) employed giants to derive the CEMP frequency is that one can ignore effects such as atomic diffusion, which can alter the surface abundances of dwarfs and turnoff stars more significantly than in giants (e.g., Richard et al. 2002a,b; Korn et al. 2007; Lind et al. 2008). However, our derived frequencies show a much better agreement for the Salpeter IMF in the metallicity region \[Fe/H\] \(>-1.5\). In the study of Suda et al. (2013), the model-predicted frequency of the CEMP stars above \[Fe/H\] = \(-2.0\) was not well-constrained, most likely due to the selection biases associated with the assembly of their sample from previous high-resolution spectroscopic studies (which tended to emphasize the more metal-poor and/or carbon-enhanced stars). In contrast, the good agreement of the frequencies calculated from our considerably less-biased SDSS/SEGUE sample with the model prediction for \[Fe/H\] \(>-1.5\) suggests that the AGB binary-synthesis model with a Salpeter mass function used by Suda et al. (2013) works well, at least in this metallicity regime. Based on the results from the comparisons of the observed CEMP frequencies with model predictions from the two different IMFs, we conjecture that, for very low-metallicity (\[Fe/H\] \(<-2.5\)) stars, the distribution of the stellar masses was dominated by rather massive stars (\(\sim\)`<!-- -->`{=html}10 \(M_\odot\) or higher), while for the relatively more metal-rich stars (\[Fe/H\] \(>-1.5\)), it appears that the IMF did not much differ from a Salpeter IMF, which is biased towards low-mass progenitor stars (\(M < 0.8\) \(M_\odot\)). As previously claimed by Suda et al. (2013), our results also support the idea that there must exist a shift in the IMF from a high-mass dominated to low-mass dominated form in the early history of the Milky Way, corresponding to a "chemical time" between \[Fe/H\] \(=-2.5\) and \[Fe/H\] \(=-1.5\). By way of comparison, the binary population-synthesis model of Izzard et al. (2009) was able to reproduce the ratio of NEMP to very metal-poor (VMP; \[Fe/H\] \(<-2.0\)) stars (that is, C and N normal stars) *without* introducing an IMF dominated by higher mass stars, but not the high frequency of the CEMP stars. Pols et al. (2012) also argued, by comparing the observed number ratio of NEMP to CEMP stars with their model predictions, that they could derive a similar number ratio from a Salpeter IMF, and ruled out an IMF peaked at 10 \(M_\odot\) claimed by Komiya et al. (2007). ## Behavior of Derived CEMP Frequencies with Distance from the Galactic Plane Another interesting result emerges when one partitions the giant sample based on distance from the Galactic plane. The green squares in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"} are the CEMP frequencies for giants with distances from the Galactic mid-plane (\(|Z|\)) larger than 5 kpc, whereas the red triangles represent frequencies based on those with \(|Z| < 5\) kpc.[^2] The figure clearly indicates that the more distant halo giants exhibit higher frequencies of C-rich stars, while the stars closer to the Galactic plane tend to have lower frequencies of C-rich stars. This same trend with vertical distance was hinted at (due to small number statistics) in Frebel et al. (2006), and strongly confirmed in the much larger sample of SDSS/SEGUE calibration stars by Carollo et al. (2012). Carollo et al. argued that this result was likely due to the fact that the outer-halo population has about twice the frequency of CEMP stars, at a given low metallicity, as the inner-halo population. A few possible reasons for the observed differences in the CEMP frequencies between the two spatial regions might be suggested within the context of the AGB model predictions. First, the progenitors of the inner-halo population (which dominates for \(|Z| <\) 5 kpc) and the outer-halo population (which, at \(|Z| >\) 5 kpc, includes more outer-halo stars) might have formed their stars at different times, with different IMFs. Because the outer-halo population has more CEMP stars than the model prediction for \[Fe/H\] \(<-2.5\), it is possible that the outer-halo population might have had an IMF with more intermediate-mass stars than considered by the model. On the other hand, since the CEMP frequencies of the inner-halo stars are lower than the model estimate, the inner-halo population might have had an IMF with less intermediate-mass stars than the proposed IMF. Related ideas are discussed by Tumlinson et al. (2007). Another possibility is that Suda et al. (2013) assumed that all CEMP stars, including CEMP-no objects, formed from the AGB binary scenario. If there were to exist other channels of carbon production at \[Fe/H\] \(<-2.5\), such as faint SNe or rapidly rotating massive stars (producing CEMP-no stars in the subsequent generation), as suggested by several studies, we then might expect larger frequencies of CEMP stars than the AGB binary-synthesis model prediction (as seen in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}), even if the carbon dilution of the giants due to extra mixing is taken into account. A more detailed discussion of this is provided below. In any event, the frequency difference we find can be understood (as argued by Carollo et al. 2012) as the result of a change in the dominant population with distance above the plane, from the inner-halo population to the outer-halo population. Carollo et al. further argued that the inner halo is dominated by stars with modest carbon enhancement (\[C/Fe\] \(\sim +0.5\)), while the outer halo has a greater portion of stars with large carbon enhancements (\[C/Fe\] \(\sim +2.0\)), although considerable overlap still exists. They interpreted these results, as well as the increase in the global frequency of CEMP stars with distance from the Galactic plane, as evidence for the possible presence of additional astrophysical sources of carbon, beyond AGB production alone, associated with the progenitors of the outer-halo stars. It is difficult to separate CEMP-\(s\) and CEMP-no stars from our medium-resolution SDSS/SEGUE spectra, but determination of the ratio of CEMP-\(s\)/CEMP-no for stars in the inner-and outer-halo populations, based on high-resolution spectroscopy, will provide not only very strong constraints on the binary-synthesis model, but clues to the origin of the different CEMP frequencies between the inner-and outer-halo populations. ## Correcting the CEMP Frequencies for Main-Sequence Turnoff Stars Above we have compared the derived CEMP frequencies from our sample of giants with the predicted CEMP frequencies from AGB binary-synthesis models by Suda et al. (2013). However, as previously mentioned, observations suggest that a giant can suffer from dilution of the carbon-rich material in its envelope by mixing during the first dredge-up event, resulting in a lower overall carbon abundance and (as a population) lower frequencies of CEMP stars. Absent such dilution, we would expect that the actual frequencies of CEMP stars among giants would be higher than shown in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}. Stars near the main-sequence turnoff region do not experience dredge-up episodes; rather, they preserve unpolluted material on their surfaces. It might be possible that their surface abundances could be affected by atomic diffusion, but because the impact on the carbon abundance is not well known, we do not take this concern into consideration in this study. Thus, we expect that one could obtain a more valid estimate of the frequencies of CEMP stars in a given population by making this evaluation using main-sequence turnoff stars. Since turnoff stars evolve quickly into giants during their evolution, we might expect that the frequencies of CEMP stars inferred from stars near the main-sequence turnoff should be the same as for giants that have not yet mixed carbon-depleted material into their envelopes. Therefore, it is desirable to compare the predictions from the models to the frequencies of CEMP stars derived from the turnoff stars. However, additional complications exist. The stars located near the main-sequence turnoff are relatively warmer than the red giants, and as a result, for a given carbon abundance, the molecular CH \(G\)-band feature becomes significantly weaker. To make matters more difficult, at low metallicity (assuming carbon is not enriched) a star's CH \(G\)-band will also become lower in strength. Even with high-resolution spectroscopy Aoki et al. (2013) noted that, although they were able to detect the CH \(G\)-band for a star with \[C/Fe\] \(\geq +1.5\) and \[Fe/H\] \(\sim-3.0\) at \(T_{\rm eff}\) \(\sim 6000\) K, they failed to measure the CH \(G\)-band for \[C/Fe\] \(<+1.5\) in their sample of very metal-poor stars. These effects become even more prevalent for medium-resolution spectra, hence the calculated CEMP frequencies obtained from the turnoff stars may also be lower than the actual values. In order to address this difficulty, and to provide a check on just how many C-rich halo stars may have been misclassified as C-normal objects (\[C/Fe\] \(< +0.7\)) for stars around the turnoff region, we have performed the following experiment. Following the prescription by Lee et al. (2013), we inject artificial noise (with characteristics similar to that for a typical SDSS/SEGUE spectrum) into the grid of synthetic spectra that are used to estimate \[C/Fe\]. The noise-added synthetic spectra have S/N = 40, 45, and 50, which are typical of the quality of the SDSS/SEGUE spectra in our study with \(|Z| < 5\) kpc (justification for this choice is provided below). At each S/N there are 25 different realizations. These spectra are processed through the SSPP to determine estimates of \[C/Fe\]. With the measured \[C/Fe\] in hand, we then derive the CEMP frequencies of the spectra, as a function of \[Fe/H\], which have an *estimated* \[C/Fe\] less than \(+0.7\) among the spectra with \(+0.75 \leq\) \[C/Fe\] \(\leq +1.25\) for giants, turnoff stars, and dwarf stars (note that we employ discrete \[C/Fe\] values, \(+0.75, +1.0\), and \(+1.25\) for the synthetic spectra). For the purpose of this exercise, we define giants as models with parameters in the ranges 4500 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 5500 K and 1.0 \(\leq\) \(\log~g\) \(\leq 3.0\), turnoff stars for 5750 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 6500 K and 3.5 \(\leq\) \(\log~g\) \(\leq 4.0\), and dwarfs for 4500 K \(\leq\) \(T_{\rm eff}\) \(\leq\) 5500 K and 4.5 \(\leq\) \(\log~g\) \(\leq 5.0\). Figure [\[fig:corfun\]](#fig:corfun){reference-type="ref" reference="fig:corfun"} shows the results of this experiment. The plus signs represent the actual frequencies of CEMP stars, which are set to 1.0. The green squares indicate giants, the red circles the turnoff stars, and the blue triangles the dwarfs. The bottom plot exhibits the residuals in the derived fractions (simply 1--the fractions) in each category. Inspection of this figure reveals that some of the C-rich dwarfs and giants start to be classified as C-normal (\[C/Fe\] \(<+0.7\)) from around \[Fe/H\] \(<-2.0\); the number of the misclassified stars slowly increases with decreasing metallicity. By way of comparison, the C-rich turnoff stars begin to be misclassified as C-normal stars as metallicity drops below \[Fe/H\] = \(-1.0\); this misclassification rapidly increases with declining metallicity, as expected. We now derive a correction function for capturing the "true" CEMP frequency, as a function of \[Fe/H\], for the SDSS/SEGUE turnoff stars, based on the results of the test carried out above. We then use this correction function to adjust the frequency calculation, among the stars with \(+0.7 \leq\) \[C/Fe\] \(<+1.5\), taking the "missing" C-rich stars into account. ## Comparison with Model Predictions for the Frequencies of C-rich Turnoff Stars Since the giants are more luminous than the turnoff stars, they can probe to greater distances in the Galaxy. This increases the likelihood of introducing a greater number of giants than turnoff stars into a magnitude-limited sample (the frequency of giants can also be influenced by the luminosity function of the halo field stars, as well as by shifts in the mix of stellar populations between the nearby and more distant halo stars). This possible population transition has already been noted in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}, and discussed in detail in the previous section. Thus, in order to make sure we are sampling the giants and turnoff stars in similar regions of the Galaxy, we restrict our considerations to the stars with \(|Z| < 5\) kpc for the calculation of the CEMP frequencies. Figure [\[fig:cempcor\]](#fig:cempcor){reference-type="ref" reference="fig:cempcor"} shows the differential frequencies of C-rich stars for main-sequence turnoff stars, as a function of \[Fe/H\], as red open triangles; the corrected frequencies are indicated as green filled triangles. The blue open circles represent the predicted turnoff CEMP frequencies from an AGB binary-synthesis model with an IMF peaked at 10 \(M_\odot\), while the magenta open squares indicates the prediction obtained using a Salpeter IMF. For comparison, the filled black circles are the frequencies from the giants with \(|Z| <\) 5 kpc given in Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}. The behavior seen in this figure is consistent with our expectations, in that the CEMP frequencies for the giants are lower than that for the turnoff sample, at least at the lowest metallicities. Note that the corrected CEMP frequencies for the turnoff stars are, on average, higher than the uncorrected frequencies, by about 5%. The figure also shows that the models produce higher CEMP frequencies for the turnoff stars than for the giants for both IMFs (compare with Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}), which at least qualitatively agrees with the observations. The small difference in the model-predicted CEMP frequencies between the giants and turnoff stars (Figures [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"} and [\[fig:cempcor\]](#fig:cempcor){reference-type="ref" reference="fig:cempcor"}) may arise from the difference in the mass of the convective envelope; a star that evolves to the giant stage has a much deeper convective zone, and hence more effective dilution, resulting in lower CEMP frequencies derived for the giants. Our derived CEMP frequencies from the distance-restricted turnoff sample appears to be in good agreement with the model prediction based on a Salpeter IMF (magenta open squares), down to \[Fe/H\] = \(-2.5\). This reaffirms that the AGB model for progenitor stars in the low-mass range works well. However, the model estimate of the CEMP frequency does not reproduce the observed frequencies for \[Fe/H\] \(<-2.5\). Unlike the case for the giants (Figure [\[fig:cempcom\]](#fig:cempcom){reference-type="ref" reference="fig:cempcom"}), the observed CEMP frequencies from our turnoff sample do not agree with the model estimation from the top-heavy IMF for \[Fe/H\] \(<-2.5\) at all, as these remain roughly flat instead of growing dramatically with decreasing metallicity. One reason for the large discrepancy between the model estimates and the observed frequencies of CEMP stars may be uncertainties of the model parameters adopted for producing carbon in the AGB star, and subsequent processes that enrich (or deplete) the envelope with carbon. Below we discuss known sources of uncertainty associated with the AGB binary-synthesis model, which may result in changes of the predicted carbon abundance of the secondary star. Most AGB stars in the mass range of \(\sim\)`<!-- -->`{=html}1--8 \(M_\odot\) can produce carbon, but whether or not they develop carbon-enriched envelopes depends on the efficiency of the third dredge-up (TDU) and helium-flash driven deep mixing (He-FDDM; Fujimoto et al. 1990, 2000)[^3] events for \[Fe/H\] \(<-2.5\). At present, it is not fully understand how such episodes depend on the mass and metallicity of an AGB star. For example, Lau et al. (2009) claimed, in a study of the evolution of AGB stars with metallicity between Z = 10\(^{-8}\) and 10\(^{-4}\), that the He-FDDM did not take place for Z \(> 10^{-5}\) independent of the mass of a star, and that this event did not occur for a star with \(M > 2\) \(M_\odot\)  regardless of its metallicity. However, Suda & Fujimoto (2010) found, from an extensive set of stellar evolution models, that the He-FDDM event occurred for a star with \(M \leq 3\) \(M_\odot\) for zero metallicity, while it occurred for a star with \(M \leq 2\) \(M_\odot\) for \(-5 \leq\) \[Fe/H\] \(\leq-3\). They also found that the TDU episode was restricted to a mass range of \(M \sim\) 1.5--5 \(M_\odot\) for \[Fe/H\] = \(-3.0\), and that this mass range becomes smaller as the metallicity decreases. In the adopted models from Suda et al. (2013), the increasing fractions of CEMP stars comes from the He-FDDM, which can enhance the surface carbon abundance by a factor of 10, from \[C/H\] \(\sim-1\) to \(+0\), regardless of the initial metallicity of the models, for stars with masses of 0.8--3 \(M_\odot\). In this view, the value of \[C/Fe\] for the secondary component (the presently observed CEMP star) of a given binary increases with decreasing \[Fe/H\], so that a larger fraction of CEMP stars can be achieved at lower metallicity. It is also assumed that the efficiency of the binary mass transfer and the mass-loss rates do not depend on metallicity. In addition, among these AGB stars, the intermediate-mass (\(\sim\)`<!-- -->`{=html}3--8 \(M_\odot\)) objects can be enriched with nitrogen by operation of the hot-bottom burning (HBB) process, which converts carbon into nitrogen by CN processing, and predicts the production of NEMP stars (Johnson et al. 2007). However, the dependency of the HBB on the mass and metallicity of an AGB star is yet not well-established. Even taking into account the uncertainties in the parameters of the AGB models, a more plausible interpretation of our results may be the existence of additional (non-AGB) carbon-production mechanisms, as discussed in the Introduction, which result in large frequencies of CEMP stars in the metallicity regime \[Fe/H\] \(<-3.0\). The current observations certainly favor this interpretation, since most CEMP-no stars in the Galaxy appear at \[Fe/H\] \(<-3.0\), and these stars do not commonly exhibit the radial velocity variations that would be expected if membership in a binary system were required (as in the AGB mass-transfer scenario). The adopted models, however, assume that all CEMP-no stars form from the AGB binary mass-transfer scenario, rather than including additional sources that have been argued are likely to be present in the early universe. According to the AGB models, low-mass (\(M < 3.5\) \(M_\odot\)) AGB stars efficiently create \(s\)-process elements by generating extra neutrons via the \(^{13}\)C(\(\alpha\),n)\(^{16}\)O reaction, while a weak \(s\)-process (for light \(s\)-process elements) operates by \(^{22}\)Ne(\(\alpha\),n)\(^{25}\)Mg for the intermediate-mass stars. CEMP-no stars could form in the AGB mass-transfer scenario by suppressing the formation of the \(^{13}\)C pocket for intermediate mass stars (\(M > 3.5\) \(M_\odot\); Suda et al. 2013). Thus, in order to preferentially produce CEMP-no stars at low metallicity and maintain the observed ratio of CEMP-no/(CEMP-\(s\) \(+\) CEMP-no), which is close to 0.5 at \[Fe/H\] \(\sim-3.0\) (from Table 1 of Suda et al. 2013), the models have to assume that the \(^{13}\)C pocket does not form in stars with metallicity significantly below \[Fe/H\] = \(-2.5\). It has also been suggested (Komiya et al. 2007) that a secondary star with an AGB primary having 0.8 \(M_\odot\) \(< M < 3.5\) \(M_\odot\) could become a CEMP-\(s\) star following mass transfer, while systems that include an AGB primary with \(M > 3.5\) \(M_\odot\) could produce C without the enhancement of neutron-capture elements, leading to a CEMP-no star following mass transfer. However, current AGB models do not satisfactorily explain the absence of the \(^{13}\)C pocket at low metallicities, even though this assumption has been invoked to explain the observed decreasing trend of \[Pb/Ba\] for CEMP stars with \[Fe/H\] \(<-2.5\) (Aoki et al. 2002; Suda et al. 2004; Barbuy et al. 2005; Cohen et al. 2006; Aoki et al. 2008). It is also significant that Ito et al. (2013) obtained a rather low upper limit on the abundance of lead (log \(\epsilon\)(Pb) \(<-0.10\)) for the \[Fe/H\] \(=-3.8\) CEMP-no star BD\(+\)`<!-- -->`{=html}44\(^{\circ}\)`<!-- -->`{=html}493, while previous predictions called for log \(\epsilon\)(Pb) \(\sim +1.5\) at these low metallicities if the lead were produced by the \(s\)-process (Cohen et al. 2006). Chemical abundances of CEMP-no stars observed with high-resolution spectroscopy (e.g., Ito et al. 2009, 2013; Norris et al. 2013) support other scenarios for significant carbon production. Their abundance patterns are similar to the predictions from massive, rapidly rotating, MMP stars (Meynet et al. 2006, 2010) or faint SNe that experience mixing and fallback (Umeda & Nomoto 2003, 2005; Tominaga et al.2007, 2013; Kobayashi et al. 2011; Ito et al. 2013; Nomoto et al. 2013). If such mechanisms are the dominant sources of the large amounts of carbon produced at low metallicity, these scenarios also favor an IMF that preferentially produces massive stars. In fact, one might also expect a rather abrupt "break" in the CEMP frequencies when the primary carbon sources change in nature---such a sudden change can be seen for the turnoff stars in Figure [\[fig:cempcor\]](#fig:cempcor){reference-type="ref" reference="fig:cempcor"} below \[Fe/H\] \(=-3.0\). Although we are not able to distinguish CEMP-no stars from CEMP-\(s\) stars in our sample, given that the majority of the CEMP-no stars are found with \[Fe/H\] \(<-3.0\), our derived frequencies imply that non-AGB related phenomenon may be the dominant mechanisms for producing large carbon abundances at extremely low \[Fe/H\]. Finally, we see that a similar behavior in the CEMP frequencies with \[Fe/H\] applies to the turnoff stars with \(-2.5 <\) \[Fe/H\] \(<-1.5\) as for the giants, suggesting that the proposed shift in an IMF occurred over this chemical interval. # Conclusions We have compared our derived CEMP frequencies from the SDSS/SEGUE giant sample with that predicted by AGB binary-synthesis models with two different IMFs---a Salpeter IMF, and an IMF with a characteristic mass of 10 \(M_\odot\). Good agreement of the CEMP frequencies for \[Fe/H\] \(>-1.5\) with the Salpeter IMF indicates that the adopted AGB model works well for low-mass progenitor stars. Qualitatively, better agreement with an IMF biased to higher-mass progenitors is found for \[Fe/H\] \(<-2.5\), suggesting that the nature of the IMF shifted from one that is high-mass dominated in the early history of the Milky Way galaxy, to one that is now low-mass dominated. This transition appears to have occurred, in "chemical time", between \[Fe/H\] \(=-2.5\) and \[Fe/H\] \(=-1.5\), as other recent studies have argued (e.g., Suda et al. 2011, 2013; Yamada et al. 2013). As noted by other recent work, the more distant halo giants (those with \(|Z|\) \(> 5\) kpc) exhibit higher frequencies of CEMP stars compared to those closer to the Galactic plane. A plausible explanation for this difference is the expected change of the dominant stellar populations from the inner-halo to the outer-halo population, coupled with the assumption that the outer-halo stars are associated with progenitors capable of producing large amounts of carbon without the accompanying production of heavy metals. Thus, one might expect that the inner-halo population harbors a higher ratio of CEMP-\(s\)/CEMP-no stars, while the opposite may apply to the outer-halo population. Tests of this hypothesis are underway (D. Carollo et al., in preparation). The weak CH \(G\)-bands for moderately carbon-enhanced stars (\(+0.7 <\) \[C/Fe\] \(<+1.5\)) among warm, metal-poor main-sequence turnoff stars results in their likely having been undercounted by previous assessments of CEMP frequencies. We have derived a correction function to compensate for this, making use of noise-added synthetic spectra. The corrected CEMP frequencies for turnoff stars are, on average, higher by \(\sim 5\)% than with the uncorrected frequencies. Both the corrected and uncorrected CEMP frequencies derived from the turnoff sample exceed those of the giants for \[Fe/H\] \(<-3.0\). We have made use of main-sequence turnoff stars with \(|Z| < 5\) kpc to compute more realistic CEMP frequencies than obtained by using giants (or the combination of giants with other classes), corrected as mentioned above. For \[Fe/H\] \(>-2.5\), our corrected CEMP frequencies agree with the model predictions based on a Salpeter IMF, indicating that the AGB model used in this study is probably not far from reality, at least as applied to low-mass stellar progenitors. However, unlike the case for the giant sample, the top-heavy IMF model does not reproduce the observed trend of the CEMP frequencies for the turnoff stars at all. The combination of these results from the giant and turnoff samples suggests that the current AGB binary-synthesis model may not be suitable for creating carbon-enhanced envelopes for intermediate-to high-mass stars (3--8 \(M_\odot\)). As the AGB binary-synthesis model (using a Salpeter IMF or a top-heavy IMF) predict far too low frequencies of CEMP stars for our turnoff sample, there likely exists one or more additional mechanisms capable of producing carbon-rich stars below \[Fe/H\] = \(-3.0\), the metallicity regime where the CEMP-no stars dominate over the subclass of CEMP-\(s\) stars. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. 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, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, 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, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, 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. Y.S.L. is a Tombaugh Fellow. T.S. was supported by the JSPS Grants-in-Aid for Scientific Research (23224004). This work was supported in part by grant PHY 08-22648: Physics Frontiers Center/Joint Institute for Nuclear Astrophysics (JINA), awarded by the U.S. National Science Foundation. [^1]: Different criteria, such as \[C/Fe\] \(> +0.5\) and \[C/Fe\] \(> +0.7\) are used by a number of studies as well; most recent work assumes \[C/Fe\] \(> +0.7\). [^2]: A more quantitative analysis of the variation of CEMP frequencies with \(|Z|\) will be considered in an upcoming paper of this series. [^3]: Often referred to as a dual shell flash or carbon ingestion episode.

{'timestamp': '2013-10-15T02:00:19', 'yymm': '1310', 'arxiv_id': '1310.3277', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3277'}

# Introduction Measure equivalence is an equivalence relation on groups introduced by Gromov in . It has since become the object of considerable study: Furman's survey  provides a thorough overview. However, it is essentially trivial for countably infinite amenable groups. This is because two groups are measure equivalent whenever they have free orbit-equivalent probability-preserving ergodic actions. Such actions exist for any infinite group, since Bernoulli shifts give examples, and Ornstein and Weiss proved in  that any two such actions of any countably infinite amenable groups are orbit-equivalent, generalizing the classical theorems of Dye  about \(\mathbb{Z}\)-actions. A measure equivalence between two groups implicity defines a pair of (equivalence classes of) cocycles over probability-preserving actions of those groups. In , Bader, Furman and Sauer have sharpened measure equivalence to a finer equivalence relation by allowing only measure equivalences for which these cocycles satisfy an integrability condition. This sharper relation is called integrable measure equivalence, henceforth abbreviated to IME. Their focus is on applications to rigidity of hyperbolic lattices. The present paper considers instead finitely-generated groups of polynomial growth, and finds that these also exhibit considerable rigidity for IME, in sharp contrast to the original notion of measure equivalence. The rigidity for these 'small' groups is in terms of Gromov's notion of their asymptotic cones. Here the notation '\(\mathrm{Con}_\infty G\)' refers to the asymptotic cone of a group \(G\) with a given right-invariant word metric \(d_G\), as constructed in. By Gromov's Theorem in  that f.-g. groups of polynomial growth are virtually nilpotent, Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is effectively a theorem about nilpotent groups. For general groups, the construction of \(\mathrm{Con}_\infty G\) may depend on the choice of a non-principal ultrafilter (), but for nilpotent groups, and hence groups of polynomial growth, it is known to be independent of that choice. (We will later invoke more precise results of Pansu which imply this.) One can see Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} as a generalization to polynomial-growth groups of the result that an integrable measure equivalence between \(\mathbb{Z}^d\) and \(\mathbb{Z}^D\) must asymptotically define an isomorphism \(\mathbb{R}^d\to \mathbb{R}^D\), and hence requires that \(d = D\). This special case follows easily by applying the Norm Ergodic Theorem to the cocycles defining the measure equivalence. In the setting of more general groups, Lewis Bowen has shown that the growth function of a f.-g. group is an IME-invariant. His exposition is given as a self-contained appendix to the present paper. That result already implies that the amenable groups fall into many (indeed, uncountably many) distinct IME-classes, and that the subclass of groups of polynomial growth is IME-closed. However, it seems that more subtle arguments are needed, for example, to distinguish the discrete Heisenberg group from \(\mathbb{Z}^4\) up to IME, since both of these groups having quartic growth. Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} implies that they are not IME, because \[\begin{gathered} \mathrm{Con}_\infty(\hbox{discrete Heis},\hbox{word metric})\\ \cong_{\scriptsize{\hbox{bi-Lip}}} (\hbox{continuous Heis},\hbox{Carnot-Carath\'eodory metric}) \end{gathered}\] and \[\mathrm{Con}_\infty \mathbb{Z}^4 \cong_{\scriptsize{\hbox{bi-Lip}}} \mathbb{R}^4,\] and these are not bi-Lipschitz (for instance, because their dimensions as topological spaces do not match). More generally, Bowen's result implies that if \(G\) is IME to \(\mathbb{Z}^d\) then \(G\) must be of polynomial growth, and then Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} implies that \(\mathrm{Con}_\infty G \cong_{\scriptsize{\hbox{bi-Lip}}} \mathbb{R}^d\). It is known that \(\mathrm{Con}_\infty G\) is always a graded connected nilpotent Lie group, and it is a Euclidean space only if \(G\) was virtually Abelian (), so our remarks about the Heisenberg group generalize to the following. We will also need the invariance of the growth function for an auxiliary purpose during our proofs later. For nilpotent groups, the map \[G \mapsto \mathrm{Con}_\infty G\] seems to retain a great deal of large-scale geometric information about \(G\). The main result of Pansu's work  is a precise characterization of those pairs of f.-g. nilpotent groups whose asymptotic cones are bi-Lipschitz: this is equivalent to isomorphism of their associated graded Lie algebras. Moreover, for Carnot groups (that is, nilpotent groups which admit an endomorphism which enlarges all distances by a fixed factor), such as any \(\mathbb{Z}^d\) or the Heisenberg group, it is known that \(G\) is quasi-isometric to \(\mathrm{Con}_\infty G\) (see example 2.C\(_1\)(a) in ). For other f.-g. nilpotent groups, the issue of just what geometric information is retained by the construction of the asymptotic cone is still not completely understood. # Background and first steps ## Integrable measure equivalence {#subs:basics} This paper will largely assume the basic facts about measure equivalence and integrable measure equivalence: we will recall only a brief statement of them here. We essentially follow the treatment in Furman's survey  (which is also similar to Section 1.2 and Appendix A of ). Given countable discrete groups \(G\) and \(H\), a **measure coupling** between them is a nonzero \(\sigma\)-finite measure space \((\Omega,m)\) which admits commuting \(m\)-preserving actions of \(G\) and \(H\) which both have finite-measure fundamental domains. We denote the actions of both \(G\) and \(H\) on \(\Omega\) by \(\cdot\). By restricting attention to an ergodic component, one may always assume that \(m\) is ergodic for the resulting \(G\times H\)-action on \(\Omega\). The fundamental domains \(Y\) and \(X\) for the \(G\)-and \(H\)-actions give rise to functions \(\b:H\times Y \to G\) and \(\a:G\times X\to H\), defined uniquely by requiring that \[h\cdot y \in \b(h,y)^{-1}\cdot Y \quad \hbox{and} \quad g\cdot x \in \a(g,x)^{-1}\cdot X \quad \forall x \in X,\ y \in Y\] (the inverses are inserted so that some other calculations come out simpler later). This also defines auxiliary finite-measure-preserving actions \(S:H\curvearrowright (Y,m|_Y)\) and \(T:G\curvearrowright (X,m|_X)\) by requiring that \[h\cdot y = \b(h,y)^{-1}\cdot (S^hy) \quad \hbox{and} \quad g\cdot x = \a(g,x)^{-1}\cdot (T^gx).\] If \(m\) is ergodic for \(G\times H\), then \(m|_X\) is ergodic for \(T\) and \(m|_Y\) is ergodic for \(S\). These are both finite invariant measures, but at times it will be convenient to insist on probability measures: for those situations, we will set \[\mu_X := m(X)^{-1}\cdot m|_X \quad \hbox{and} \quad \mu_Y := m(Y)^{-1}\cdot m|_Y.\] Now a standard calculation shows that \(\b\) and \(\a\) are cocycles over \(S\) and \(T\) respectively: that is, \[\a(g_1g_2,x) = \a(g_1,T^{g_2}x)\a(g_2,x) \quad \forall g_1,g_2 \in G,\ x \in X,\] and similarly for \(\b\). In this construction, we may always replace the fundamental domain \(Y\) with one of its \(H\)-translates, and the cocycle \(\b\) will just be translated accordingly. Since countably many translates of \(Y\) cover \(\Omega\), we may therefore ensure that \(m(X \cap Y) > 0\). Now a simple calculation shows that if \[x \in X\cap Y \cap T^{g^{-1}}(X \cap Y) \quad \hbox{for some}\ g \in G,\] then we may write \[g^{-1}\cdot (T^gx) = \a(g^{-1},T^gx)^{-1}\cdot x = \b(\a(g^{-1},T^gx)^{-1},x)^{-1}\cdot (S^{\a(g^{-1},T^gx)^{-1}}x),\] where the first equality holds because \(T^gx \in X\), and the second because \(x \in Y\). Since we also assume that \(T^gx \in Y\), and the \(G\)-translates of \(Y\) are disjoint, this implies that \[\b(\a(g^{-1},T^gx)^{-1},x) = g \quad \hbox{and} \quad S^{\a(g^{-1},T^gx)^{-1}}x = T^g x.\] Finally, the cocycle equation for \(\a\) gives that \(\a(g^{-1},T^gx) = \a(g,x)^{-1}\), so these conclusions simplify to \[\begin{aligned} \label{eq:inv-reln} \b(\a(g,x),x) = g \quad \hbox{and} \quad S^{\a(g,x)}x = T^g x. \end{aligned}\] In particular, the orbit equivalence relations of \(T\) on \(X\) and \(S\) on \(Y\) have the same restriction to \(X\cap Y\). In the sequel, it will often be convenient to work instead with the functions \(\a_x := \a(\,\cdot\,,x):G\to H\) and \(\b_y := \b(\,\cdot\,,y):H\to G\). The cocycle equation for \(\a\) gives that \(x\mapsto \a_x\) is a map from \(X\) to \[[G,H] := \{f:G\to H\,|\ f(e_G) = e_H\}\] which intertwines the action \(T:G\curvearrowright X\) with the action of \(G\) on \([G,H]\) defined by \(g:f(x) \mapsto f(xg)f(g)^{-1}\). Similarly, \(\b\) is a map from \(Y\) to \([H,G]\) which intertwines \(S\) with the analagous action of \(H\) on \([H,G]\). With this interpretation, the pushforward of \(\mu_X\) under \(x\mapsto \a_x\) is an invariant probability on \([G,H]\): such objects are discussed by Monod in  under the term 'randomorphisms', and again by Furman . (Also, in the special case of \([\mathbb{Z}^2,\mathbb{Z}]\), they have a long history in statistical physics as models of random surfaces: see, for instance,  and the many references there.) Now, for \(x \in X\) and \(y \in Y\), let \[D_x := \{g \in G\,|\ T^gx \in X\cap Y\}\] and \[E_y := \{h \in H\,|\ S^hy \in X\cap Y\}.\] Then \(x\mapsto D_x\) is a map \[X\to \{\hbox{subsets of}\ G\}\] which is equivariant in the sense that \[\begin{aligned} \label{eq:D-equivt} D_{T^gx} = D_x\cdot g^{-1}, \end{aligned}\] and similarly for \(y\mapsto E_y\). Also, if \(m\) is ergodic for \(G\times H\), then \(m|_Y\) is ergodic for \(S\) and \(m|_X\) is ergodic for \(T\). Using this, we may extend the definitions of \(D_\bullet\) and \(E_\bullet\), \(\a_\bullet\) and \(\b_\bullet\) to almost all of \(X\cup Y\). By ergodicity, for \(m\)-a.e. \(y \in Y\) the set \(E_y\) is nonempty, so there is some \(h \in H\) such that \(S^hy \in X\cap Y\). This now gives \[y = S^{h^{-1}}S^hy = T^{\b(h^{-1},S^hy)}S^hy = T^{\b(h,y)^{-1}}S^hy,\] using ([\[eq:inv-reln\]](#eq:inv-reln){reference-type="ref" reference="eq:inv-reln"}) and the cocycle equation for \(\b\). Setting \[D_y := D_{S^hy}\cdot \b(h,y),\] this is independent from the choice of \(h\) by the cocycle relations. Similarly, for \(m\)-a.e. \(x \in X\) there is \(g \in G\) such that \(T^g x \in X\cap Y\), and now we may set \[E_x := E_{T^gx}\cdot \a(g,x).\] For the cocycles, if \(y \in Y\) and \(h\) is chosen as above, we set \[\a_y(g) := \a_{S^hy}(g\b(h,y)^{-1})\a_{S^hy}(\b(h,y)^{-1})^{-1},\] and similarly \[\b_x(h) := \b_{T^gx}(h\a(g,x)^{-1})\b_{T^gx}(\a(g,x)^{-1})^{-1}.\] Once again, the consistency of these definitions follows from the cocycle relations for \(\a\) and \(\b\). Having thus extended these objects, the relation ([\[eq:inv-reln\]](#eq:inv-reln){reference-type="ref" reference="eq:inv-reln"}) now asserts that \(\a_x|D_x\) is a bijection \(D_x\to E_x\) for every \(x \in X\cup Y\), and its inverse equals \(\b_x|E_x\). Our subsequent reasoning about measure equivalence will mostly be in terms of these equivariant maps \(x\mapsto (\a_x,D_x)\) and \(y\mapsto (\b_y,E_y)\). For any f.-g. groups \(G\) and \(H\) and a probability-preserving action \(T:G\curvearrowright (X,\mu)\), a cocycle \(\a:G\times X\to H\) is **integrable** if, for any choice of finite, symmetric generating set \(B_H\subseteq H\), we have \[\||\a(g,\,\cdot\,)|_H\|_1 = \int_X|\a(g,x)|_H\,\mu(\d x) < \infty \quad \forall g\in G,\] where \(|\cdot|_H\) is the length function on \(H\) associated to \(B_H\). Since the length functions arising from different choices of \(B_H\) are all equivalent up to constants, this notion does not depend on the choice of \(B_H\). Moreover, the subadditivity of \(|\cdot|_H\) gives \[\begin{aligned} \label{eq:cocyc-subadd} \||\a(g,\,\cdot\,)|_H\|_1 \leq |g|_G\cdot \max_{s \in B_G}\||\a(s,\,\cdot\,)|_H\|_1, \end{aligned}\] where \(B_G\) is a finite, symmetric generating set for \(G\), so it suffices to check integrability on this \(B_G\). A measure coupling as above is **integrable** if one can choose fundamental domains \(X\) and \(Y\) so that the cocycles \(\a\) and \(\b\) are integrable. Finally, f.-g. groups \(G\) and \(H\) are **integrably measure equivalent**, or **IME**, if they admit an integrable measure coupling. Standard arguments, given in , show that this defines an equivalence relation on f.-g. groups, independent of the choice of their generating sets. It will be denoted by \(\stackrel{\rm{IME}}{\sim}\). ## Initial simplification In our setting, standard properties of IME lead to an immediate, useful reduction of the task of proving Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}. According to Gromov's famous result from , any f.-g. group \(G\) of polynomial growth has a f.-g. nilpotent subgroup \(G_1\) of finite index. Letting \(\Omega_1:= G\) with counting measure, this defines a \((G_1,G)\)-coupling \[(g_1,g)\cdot \omega := g_1\omega g^{-1}.\] Since \(G_1\) has a finite fundamental domain in \(G\), this measure coupling is trivially integrable. The same reasoning holds for some finite-index nilpotent subgroup \(H_1 \leq H\), giving an integrable \((H,H_1)\)-measure equivalence. Therefore, in the setting of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, we obtain \[G_1 \stackrel{\rm{IME}}{\sim} G \stackrel{\rm{IME}}{\sim} H \stackrel{\rm{IME}}{\sim} H_1,\] and hence \(G_1 \stackrel{\rm{IME}}{\sim} H_1\), by transitivity. On the other hand, since asymptotic cones are insensitive to passage to finite-index subgroups, we have \[\mathrm{Con}_\infty G = \mathrm{Con}_\infty G_1 \quad \hbox{and} \quad \mathrm{Con}_\infty H = \mathrm{Con}_\infty H_1.\] It therefore suffices to prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} for the subgroups \(G_1\) and \(H_1\); equivalently, in the special case with \(G\) and \(H\) themselves nilpotent. This will simplify some calculations later. ## Asymptotic cones of nilpotent groups Henceforth \(G\) and \(H\) will be f.-g. nilpotent groups and \(B_G\) and \(B_H\) will be finite, symmetric generating sets for them. To the generating set \(B_G\) we associate the word-length function \(|\cdot|_G\) and the right-invariant word metric \(d_G\), and similarly for \(B_H\). It is known that all such groups \(G\) with right-invariant word metrics \(d_G\) have the following properties: 1. the asymptotic cone does not depend on the choice of ultrafilter \(\omega\) up to pointed isometry, and so may be written as \(\mathrm{Con}_\infty G\); 2. the sequence of re-scaled pointed metric spaces \((G,e_G,n^{-1}d_G)\) converges as \(n\to\infty\) in the local Gromov-Hausdorff sense to the pointed metric space \((\mathrm{Con}_\infty G,\ol{e}_G,d_G^\infty)\) for some limit metric \(d_G^\infty\) on \(\mathrm{Con}_\infty G\) (whereas for many groups \(\mathrm{Con}_\infty\) exists only as an ultralimit); 3. the asymptotic cone \(\mathrm{Con}_\infty G\) is a proper metric space under \(d_G^\infty\) (that is, all bounded sets are precompact). An element of \(\mathrm{Con}_\infty G\) will be signified by an overline, as in '\(\bar{g}\)'. Most of these properties follow from Pansu's results in ; the last already follows from the theory in . For the first, Pansu asserts only independence of the cone from \(\omega\) up to a pointed bi-Lipschitz map, but this is tightened to a pointed isometry in . On the other hand, in his discussion of asymptotic cones in Chapter 2 of , Gromov analyses more general groups for which these properties may fail, including (at least for the second property) some solvable examples. Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} will be deduced from the following. ## Invariance of growth Our approach to Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} will make use of the fact that the growth rate of a f.-g. group is an IME-invariant. This follows from a more general control of growth functions under 'integrable measure embeddings', proved by Lewis Bowen in his appendix to the present paper (Theorem [\[thm:growth2\]](#thm:growth2){reference-type="ref" reference="thm:growth2"}). The consequence that we will need is as follows. # A refined growth estimate for cocycles If \(G\) and \(H\) are f.-g. groups with word metrics \(d_G\) and \(d_H\), \((X,\mu,T)\) is a probability \(G\)-space and \(\sigma:G\times X\to H\) is an integrable cocycle, then the cocycle identity and an induction on word-length imply that \[\||\sigma(g,\cdot)|_H\|_1 \leq C|g|_G\] for some fixed constant \(C\), which may be taken to be \(\max_{s \in B_G}\||\sigma(s,\cdot)|_H\|_1\). Using Markov's Inequality, this implies that \[\mu\{|\sigma(g,x)|_H \geq MC|g|_G\} \leq 1/M \quad \forall M > 0.\] A key tool in proving Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} will be a small but crucial improvement on this estimate in the setting of nilpotent groups. This is most cleanly formulated in terms of the following abstract notion. This nomeclature is not completely standard. Setting \[\rho_x(g,h) := f(gh^{-1},T^hx),\] one can check that \(x \mapsto \rho_x\) is an equivariant map from \((X,T)\) to the space of pseudometrics on \(G\) with the action of \(G\) given by translation on the right (in particular, the sub-cocycle inequality becomes the triangle inequality). As with 'randomorphisms', important examples of such stationary random pseudometrics for \(G = \mathbb{Z}^d\) are classical objects in probability: in the study of first-passage percolation models, the first passage times between pairs of points define such a pseudometric. Classic references for the asymptotic behaviour of this pseudometric include , and a recent survey of this area can be found in . In a sense, the next proposition can be seen as very weak nilpotent-groups extension of the convergence of the reachable sets to the limit shape (that is, of these random pseudometrics to a deterministic limiting norm) in first-passage percolation. That is, as one considers increasingly large distances in \(G\), the function \(f\) is vanishingly unlikely to blow up those distances by any factor greater than \(M\). Note the convention that we always choose \(M \geq 1\), even if one could actually use a smaller \(M\) for some \(f\). The proof of Proposition [\[prop:as\]](#prop:as){reference-type="ref" reference="prop:as"} rests on two basic geometric facts about nilpotent groups. Intuitively, this asserts that 'any point in \((G,d_G)\) may be reached by a sequence of at most \(K\) straight-line segments of length not much greater than the distance to that point'. I have not been able to find a reference for Proposition [\[prop:sla\]](#prop:sla){reference-type="ref" reference="prop:sla"}, but it is a fairly routine exercise in nilpotency, so its proof is deferred to Appendix [\[app:nilpcalc\]](#app:nilpcalc){reference-type="ref" reference="app:nilpcalc"}. The second estimate we will need is the following. Proposition [\[prop:cc\]](#prop:cc){reference-type="ref" reference="prop:cc"} is a special instance of de Cornulier's Proposition 3.1, part (iii\('\)), and Corollary A.2 in . This is because, in his notation, the constant sequence \((h)\) is an element of \(\rm{Sublin}(G)\) (whose definition can be found in that paper). (Note that his Corollary A.2 seems to be mis-labelled as 'A.7' in some versions.) Assuming the above group-theoretic facts, the next step towards Proposition [\[prop:as\]](#prop:as){reference-type="ref" reference="prop:as"} is the following consequence of the Ergodic Theorem: It might be interesting to study the generalization of Proposition [\[prop:as\]](#prop:as){reference-type="ref" reference="prop:as"} to other groups. We will make use of Proposition [\[prop:as\]](#prop:as){reference-type="ref" reference="prop:as"} mostly through the following. At one point, it will be more convenient to use Proposition [\[prop:as\]](#prop:as){reference-type="ref" reference="prop:as"} through the following corollary. # Completion of the proof Now consider again two f.-g. nilpotent groups \(G,H\) and their asymptotic cones \(\mathrm{Con}_\infty G\) and \(\mathrm{Con}_\infty H\). It remains to prove Theorem [\[thm:main2\]](#thm:main2){reference-type="ref" reference="thm:main2"}: we must find some \(L > 0\) such that for each \(R > 0\) there are a set \(E\) and map \(\phi\) with the properties asserted there. This map \(\phi\) will be obtained from the restriction of the cocycle \(\a_x\) to a suitable finite subset of \(G\) for a 'typical' point \(x\). As usual, we fix generating sets \(B_G \subset G\) and \(B_H \subset H\), which will become the \(1\)-balls in the resulting metrics \(d_G\) and \(d_H\). The sequence of renormalized metric spaces \((G,n^{-1}d_G)\) converges in the local Gromov-Hausdorff sense to \((\mathrm{Con}_\infty G,d^G_\infty)\) as \(n\to\infty\), and similarly for \((H,n^{-1}d_H)\). This implies that for any finite subset \(E \subset \mathrm{Con}_\infty G\) we can find a sequence of finite subsets \(E_n \subset G\), \(|E_n| = |E|\), and bijections \(\phi_n:E_n\to E\) such that for any \(c > 1\) one has \[c^{-1}n^{-1}d_G(\phi_n(\ol{g}),\phi_n(\ol{g}')) \leq d^G_\infty(\ol{g},\ol{g}') \leq c n^{-1}d_G(\phi_n(\ol{g}),\phi_n(\ol{g}')) \quad \forall \ol{g},\ol{g}' \in E\] for all sufficiently large \(n\), and similarly for \(H\) and \(\mathrm{Con}_\infty H\). Let us refer to such a sequence of maps \(\phi_n\) as a sequence of **asymptotic copies** of \(E\). Since \(G\) and \(H\) are groups, by translating if necessary we may always assume that \(E \ni \ol{e}_G\), \(E_n \ni e_G\) for each \(n\), and \(\phi_n(e_G) = \ol{e}_G\); we will refer to such \(E\) and \(\phi_n\) as **pointed**. For the proof, fix \(R > 0\), and let \(E\) be a pointed \((1/R)\)-net in \(B_G^\infty(R)\) (that is, an inclusion-maximal \((1/R)\)-separated subset of this ball, which is therefore also \((1/R)\)-dense in the ball). Also, let \(\phi_n: E \to E_n\) be a pointed sequence of asymptotic copies of \(E\). Theorem [\[thm:main2\]](#thm:main2){reference-type="ref" reference="thm:main2"} will be a consequence of the following asymptotic behaviour of the cocycle \(\a\). Recall that a sequence of events \(X_n\) in a probability space \((X,\mu)\) is said to occur **with high probability** ('**w.h.p.**') in \(\mu\) if \(\mu(X_n) \to 1\). Property (i) of Theorem [\[thm:main3\]](#thm:main3){reference-type="ref" reference="thm:main3"} follows directly from Corollary [\[cor:interpt\]](#cor:interpt){reference-type="ref" reference="cor:interpt"}. Properties (ii) and (iii) will need a little more work. For these we will also need to use related estimates for the cocycle \(\b\) going in the other direction. In case our original measure coupling gives \(X = Y = X\cap Y\), so that \(\b_x = \a_x^{-1}\) for all \(x\), property (ii) looks very like property (i) with \(\a\) replaced by \(\b\). However, even in this special case, there is an extra subtlety here. Property (ii) is asserting that > \(\b_x|\a_x(E_n)\) is \((4M)\)-Lipschitz. This differs from property (i) in that the relevant domain, \(\a_x(E_n)\), now also depends on \(x\). This will force us to use a more careful argument than for Corollary [\[cor:interpt\]](#cor:interpt){reference-type="ref" reference="cor:interpt"}, because we must rule out the possibility that, as \(x\) varies, the set-valued function \(x\mapsto \a_x(E_n)\) always happens to choose a set on which \(\b_x\) behaves irregularly. To rule this out, we will choose a new fixed set \(F_n \subset H\) which is \((\delta n)\)-dense for some \(\delta \ll 1/R\), and show that w.h.p. the restriction \(\b_x|\a_x(E_n)\) stays very close to the restriction of \(\b_x\) to a set of points in \(F_n\) that lie nearby the points in \(\a_x(E_n)\). On the other hand, the analog of (i) will give that \(\b_x\) is \((2M)\)-Lipschitz on the whole of \(F_n\), and from this we can then gain control of the Lipschitz constant of its restriction to \(\a_x(E_n)\), notwithstanding that dependence on \(x\). At the end of this section we will present an example showing that cocycles such as \(\a_x\) can have occasional 'defects' where their behaviour is very far from Lipschitz, which suggests that this extra care is really needed. A similar comparison with \(\b_x|F_n:F_n\to G\) will also underly the proof of property (iii). The first step is the following. The next lemma asserts that once the radius \(R\) is sufficiently large, for most \(x\) the ball-image \(\a_x(B_G(g,R)) \subset H\) must be mostly contained inside the slightly larger ball \(B_H(\a_x(g),2MR)\). We will now combine the estimates of the previous two lemmas into the following conclusion. It will be the key to controlling both the typical co-Lipschitz constant of \(\a_x|E_n\) and the density of its image. This completes the proof of our main theorems. Before leaving this section, it is worth including an example of an IME in which the cocycle \(\a_x\) exhibits occasional bad behaviour at arbitrarily large scales for a.e. \(x\). This justified the care we have taken over the proofs of properties (ii) and (iii) above. # Remaining issues Most obviously, it would be interesting to know whether the results of this paper extend beyond the class of virtually nilpotent groups (I am confident that the methods do not). Among nilpotent groups, Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} suggests another interesting line of enquiry. For simplicity, consider a case in which \(G\) and \(H\) are both quasi-isometric to their asymptotic cones, say via maps \(\phi:\mathrm{Con}_\infty G\to G\) and \(\psi:H \to \mathrm{Con}_\infty H\). Recall () that the asymptotic cones are graded connected nilpotent Lie groups equipped with dilations \(\delta_{\mathrm{Con}_\infty G}\) and \(\delta_{\mathrm{Con}_\infty H}\). Given an integrable measure equivalence implemented by the cocycles \(\a\) and \(\b\) as before, for each \(x \in X\) and \(n \geq 1\) one can consider the map \[\kappa_{x,n}:\mathrm{Con}_\infty G\to \mathrm{Con}_\infty H:\bar{g} \mapsto \delta_{\mathrm{Con}_\infty H}^{1/n}(\psi(\a_x(\phi(\delta_{\mathrm{Con}_\infty G}^n(\bar{g}))))).\] If true, this would amount to a kind of 'nilpotent-valued' version of the Pointwise Ergodic Theorem. It has the flavour of a large-scale analog for cocycles of the problem of proving an analog of Rademacher's Theorem for Lipschitz maps between Carnot-Carathéodory metrics. Such a differentiation theorem has been studied by Pansu in  and Margulis and Mostow in .

{'timestamp': '2014-11-25T02:20:24', 'yymm': '1310', 'arxiv_id': '1310.3216', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3216'}

# Introduction The invariant theory of \(3 \times 3 \times 3\) arrays, or trilinear forms on a 3-dimensional vector space, has been studied by many authors; the earliest reference we have been able to find is the work of Aronhold (1850). A classification of the forms, and an application of classical invariant theory to this problem, appears in Chanler and Thrall (1938-39). These trilinear forms and their invariants are also closely tied to the classical study of the Hesse configuration of 12 lines in the projective plane, which was addressed with a modern view by Artebani and Dolgachev (2009). See Dolgachev (2012) for a comprehensive view. The first results using the representation theory of Lie groups were obtained by Vinberg (1976), who embedded the semisimple Lie algebra \(\mathfrak{sl}_3(\mathbb{C}) \oplus \mathfrak{sl}_3(\mathbb{C}) \oplus \mathfrak{sl}_3(\mathbb{C})\) into an exceptional Lie algebra of type \(E_6\), and deduced that the algebra of invariants is freely generated by homogeneous polynomials of degrees 6, 9 and 12, which we denote respectively by \(I_6\), \(I_9\) and \(I_{12}\). Strassen (1983) studied \(n \times n \times 3\) arrays (\(n\) odd), showed that the complement of the set of arrays of maximal border rank is a hypersurface, and determined its equation. For \(n = 3\), this polynomial is (up to a scalar multiple) the fundamental invariant of degree 9, which vanishes if and only if the array has rank \(\le 4\); see Table [\[ranktable\]](#ranktable){reference-type="ref" reference="ranktable"} below. Littelmann (1989) classified the irreducible representations of semisimple Lie groups for which the algebra of invariants is free. Gelfand et al. (1992) introduced a general theory of hyperdeterminants (modernizing a study initiated by Cayley (1845) almost 150 years earlier), and determined their degrees; in particular, the hyperdeterminant \(\Delta_{333}\) for \(3 \times 3 \times 3\) arrays is an invariant homogeneous polynomial of degree 36, and hence can be expressed in terms of the fundamental invariants as follows, for some \(a, b, c, d, e, f, g \in \mathbb{C}\): \[\label{hdformula} \Delta_{333} = a I_6^6 + b I_6^4 I_{12} + c I_6^3 I_9^2 + d I_6^2 I_{12}^2 + e I_6 I_9^2 I_{12} + f I_9^4 + g I_{12}^3. \tag{HD}\] (We note that \(I_6\) and \(I_9\) are uniquely determined up to nonzero scalar multiples, but \(I_{12}\) is only determined up to adding a scalar multiple of \(I_6^2\).) Ng (1995) described the orbits of \(PGL(\mathbb{C}^3) \times PGL(\mathbb{C}^3) \times PGL(\mathbb{C}^3)\) acting on the projective space \(\mathbb{P}( \mathbb{C}^3 \otimes \mathbb{C}^3 \otimes \mathbb{C}^3 )\), derived explicit matrix representations for the singular forms, and proved that the group quotient is a projective variety. Nurmiev (2000) obtained an implicit description of all three fundamental invariants in terms of convolutions of volume forms, and classified the normal forms (orbit representatives). Briand et al. (2004), motivated by a problem from quantum computing, used Cayley's \(\Omega\)-process from classical invariant theory to find explicit forms of the fundamental invariants in terms of concomitants. Duff and Ferrara (2007) presented an analogy between supersymmetric black holes in 5 dimensions and the bipartite entanglement of three qutrits from quantum information theory, where the common symmetry comes from Vinberg's embedding \(\mathfrak{sl}_3(\mathbb{C}) \oplus \mathfrak{sl}_3(\mathbb{C}) \oplus \mathfrak{sl}_3(\mathbb{C}) \hookrightarrow E_6\). Explicit computations of these invariants have also been attempted classically. For instance \(I_{6}\) has been attributed to Aronhold, \(I_{9}\) was computed by Strassen (1983) as a determinant of a certain commutator, and Ottaviani (2007) gave a determinantal formula. In theory Schläfli's method may be used to compute the \(3\times 3 \times 3\) hyperdeterminant, see Gelfand et al.  (1994); this computation turns out to require a large amount of memory, but it may still be useful for evaluating the hyperdeterminant and was used by one of us (2012). In general, these computations become very large very quickly; see Huggins et al.  (2008) for explicit results on the \(2 \times 2 \times 2 \times 2\) hyperdeterminant. Domokos and Drensky (2012) computed the single defining relation of the algebra of invariants for the action of \(SL(3,\mathbb{C}) \times SL(3,\mathbb{C})\) on triples of \(3 \times 3\) matrices, and provided an alternative proof of Vinberg's result. Allman et al.  (2013) gave a generalization of Cayley's hyperdeterminant for \(2 \times 2 \times 2\) arrays to a covariant for the action of \(SL_n(\mathbb{C}) \times SL_n(\mathbb{C}) \times SL_n(\mathbb{C})\) on \(n \times n \times n\) arrays. Two of the present authors recently used computer algebra (2013) to determine the fundamental invariants of degrees 6, 9 and 12 in terms of orbit sums for the symmetry group \(( S_3 \times S_3 \times S_3 ) \rtimes S_3\). Each invariant is a homogeneous polynomial in the variables \(x_{ijk}\) for \(1 \le i, j, k \le 3\): - the polynomial \(I_6\) of degree 6 from, which is a linear combination of 8 symmetric orbits, has 1152 terms; - the polynomial \(I_9\) of degree 9 from, which is the sum of 14 alternating orbits, has 9216 terms; - the polynomial \(I_{12}\) of degree 12 from, which is a linear combination of 235 symmetric orbits, has 209061 terms. In this paper we determine the coefficients \(a, b, c, d, e, f, g\) in equation [\[hdformula\]](#hdformula){reference-type="eqref" reference="hdformula"}. A similar explicit computation of the Lüroth invariant in terms of fundamental invariants was recently achieved by Basson et al.  (2013), although their task was more difficult: the algebra of invariants for homogeneous polynomials of degree 4 in three variables is not freely generated. Interest in Lüroth quartics was revived a few years ago by Ottaviani and Sernesi; see Ottaviani (2012) for the computational aspects. In §[2](#gaussianelimination){reference-type="ref" reference="gaussianelimination"} we recall a method for generating arrays of hyperdeterminant 0 which applies to any format \(m_1 \times \cdots \times m_k\). In §[3](#hdsection){reference-type="ref" reference="hdsection"} we repeatedly generate pseudorandom \(3 \times 3 \times 3\) arrays \(( x_{ijk} )\) of hyperdeterminant 0, substitute their entries for the variables in the fundamental invariants, and set the resulting polynomial to 0; in this way, we obtain linear equations satisfied by the coefficients \(a, b, c, d, e, f, g\) in [\[hdformula\]](#hdformula){reference-type="eqref" reference="hdformula"}. In §[4](#nurmiev){reference-type="ref" reference="nurmiev"} we apply our results to Nurmiev's classification of normal forms. A similar computation for the so-called trifocal variety was carried out by one of us for low degree covariants in (2012). In §[5](#ranksection){reference-type="ref" reference="ranksection"}, we test and determine which invariants vanish on which ranks. Our Maple worksheet for the computations described in the present paper is available as an ancillary file with the `arXiv` version. # Gaussian elimination: projective version {#gaussianelimination} We recall that a \(k\)-dimensional array \(M = ( m_{i_1 i_2 \cdots i_k} )\) is said to have rank 1 if it is the outer product of nonzero vectors \(a_1, a_2, \dots, a_k\) in the sense that \(m_{i_1 i_2 \cdots i_k} = a_{1,i_1} a_{2,i_2} \cdots a_{k,i_k}\). The array \(M\) has rank \(r\) if \(r\) is the least number such that \(M\) can be written as a sum of \(r\) arrays of rank 1. The hyperdeterminant \(\Delta = \Delta_{m_1+1, \dots, m_k+1}\) of a \(k\)-dimensional array of size \((m_1{+}1) \times \cdots \times (m_k{+}1)\) with \(\sum_{i}(m_{i}{+}1) \ge 2 m_{j}+1\) for \(j = 1, \dots, k\) is (by definition) the polynomial of minimal degree vanishing on the dual variety of the arrays of rank 1. (Limits of arrays of rank 1 still have rank \(\leq 1\), so the set of rank 1 tensors is closed in projective space.) If we assume that \(\Delta\) is primitive in the sense that its coefficients are integers with no common factors, then \(\Delta\) is unique up to a sign. Using basic multilinear algebra, we can apply a multilinear change of coordinates, so that the arrays in the dual variety have zeros on the edges emanating from one corner and arbitrary scalars in the other positions. With this description, we can easily generate pseudorandom arrays of hyperdeterminant 0 of any size \((m_1{+}1) \times \cdots \times (m_k{+}1)\). To remove dependence on our choice of coordinates, we can apply pseudorandom changes of basis along each of the \(k\) directions. However, by definition of invariant, these changes of basis will not affect the values of the fundamental invariants. So for the purposes of the present paper, this last step will not be necessary. In this section we review how to perform a multidimensional generalization (i.e. to tensors) of Gaussian elimination for matrices, using a projective point of view. Let \(A_1, \dots, A_k\) be vector spaces over \(\mathbb{C}\), with \(\dim A_i = m_i+1\) for \(i = 1, \dots, k\). The corresponding projective spaces \(\mathbb{P}(A_1), \dots, \mathbb{P}(A_k)\) have dimensions \(m_1, \dots, m_k\); \(\mathbb{P}(A_i)\) consists of all lines through the origin in \(A_i\). For a nonzero vector \(a_i \in A_i\) the corresponding element of \(\mathbb{P}(A_i)\) will be denoted \([ a_i ]\); this is the line with direction vector \(a_i\). The tensor product \(A_1 \otimes \cdots \otimes A_k\) has dimension \((m_1+1) \cdots (m_k+1)\), and the projective space \(\mathbb{P}( A_1 \otimes \cdots \otimes A_k )\) has dimension \((m_1+1) \cdots (m_k+1)-1\). An element of \(A_1 \otimes \cdots \otimes A_k\) is a sum of simple tensors \(a_1 \otimes \cdots \otimes a_k\). After a choice of basis, an array of size \((m_1+1) \times \cdots \times (m_k+1)\) can be identified with an element of \(A_1 \otimes \cdots \otimes A_k\). We denote the image of the Segre embedding by \(X\). Consider an array of rank 1 in \(X\), which by definition has the form \([ \, a_1 \otimes \cdots \otimes a_k \, ]\) for nonzero \(a_i \in A_i\). A parametrized curve in \(X\) through the point \([ \, a_1 \otimes \cdots \otimes a_k \, ]\) has the form \([ \, a_1(t) \otimes \cdots \otimes a_k(t) \, ]\) where \(a_i(0) = a_i\) for \(i = 1, \dots, k\). We differentiate with respect to \(t\) and then set \(t = 0\) to obtain \[\Big[ \, \sum_{i=1}^k a_1(t) \otimes \cdots \otimes a'_i(t) \otimes \cdots \otimes a_k(t) \, \Big] \Big[ \, \sum_{i=1}^k a_1 \otimes \cdots \otimes a'_i(0) \otimes \cdots \otimes a_k \, \Big].\] The cone over the tangent plane to \(X\) at the point \([ \, a_1 \otimes \cdots \otimes a_k \, ]\) is the span of these tangent vectors: \[T_{a_1 \cdots a_k}(X) = \sum_{i=1}^k a_1 \otimes \cdots \otimes A_i \otimes \cdots \otimes a_k.\] We convert this sum of vector spaces into a direct sum by separating the common 1-dimensional subspace: \[\label{directsum} \begin{array}{l} T_{a_1 \cdots a_k}(X) = \mathbb{C}(a_1 \otimes \cdots \otimes a_k) \; \oplus \\[3pt] \qquad\qquad\qquad \bigoplus_{i=1}^k ( a_1 \otimes \cdots \otimes A_i \otimes \cdots \otimes a_k ) / \mathbb{C}(a_1 \otimes \cdots \otimes a_k). \end{array}\] We want to construct the projective dual variety \(X^\vee\). By a result of Gelfand et al. , the dual of the Segre variety is a hypersurface if and only if \(\sum_i (m_i{+}1) \ge 2 m_j+1\) for \(j = 1, \dots, k\). Projective duality preserves irreducibility, so in this case the dual of the Segre variety is a hypersurface defined by a single equation \(\Delta = 0\), where the homogeneous polynomial \(\Delta\) is known as the hyperdeterminant of type \((m_1{+}1) \times \cdots \times (m_k{+}1)\); again see. A hyperplane \([H]\) in projective space \(\mathbb{P}(A_1 \otimes \cdots \otimes A_k)\) can be identified with a normal line \(N(H)\) in the dual projective space \(\mathbb{P}( A_1^\ast \otimes \cdots \otimes A_k^\ast )\). The condition that \(H\) contains the tangent plane \(T_{a_1 \cdots a_k}(X)\) is equivalent to the condition that the corresponding line \(N(H)\) annihilates \(T_{a_1 \cdots a_k}(X)\) in the natural pairing of a variety with its dual. To work in terms of coordinates, we choose bases: \[\{ \; a_{i,1} = a_i, \; a_{i,2}, \; \dots, \; a_{i,m_i+1} \; \} \subset A_i \quad ( i = 1, \dots, k ).\] The general element of \(A_1^\ast \otimes \cdots \otimes A_k^\ast\) then has the form \[\label{dualvector} \sum_{i_1=1}^{m_1+1} \cdots \sum_{i_k=1}^{m_k+1} \mu_{i_1 \dots i_k} \, a_{1,i_1}^\ast \otimes \cdots \otimes a_{k,i_k}^\ast \quad (\mu_{i_1 \dots i_k} \in \mathbb{C}).\] By equation [\[directsum\]](#directsum){reference-type="eqref" reference="directsum"}, the general element of \(T_{a_1 \cdots a_k}(X)\) has the following form where \(d, e_{i,j} \in \mathbb{C}\): \[\label{tangentvector} T_{a_1 \cdots a_k}(X) = d (a_1 \otimes \cdots \otimes a_k) + \sum_{i=1}^k \sum_{j=2}^{m_i+1} e_{i,j} ( a_1 \otimes \cdots \otimes a_{i,j} \otimes \cdots \otimes a_k ).\] We evaluate the dual vector [\[dualvector\]](#dualvector){reference-type="eqref" reference="dualvector"} at the tangent vector [\[tangentvector\]](#tangentvector){reference-type="eqref" reference="tangentvector"} and obtain \[d \mu_{1 \dots 1} + \sum_{i=1}^k \sum_{j=2}^{m_i+1} e_{i,j} \mu_{1 \dots j \dots 1},\] where the subscript \(j\) of \(\mu\) is in position \(i\). This must vanish for all \(d\) and \(e_{i,j}\) which gives the necessary and sufficient condition that \(\mu_{i_1 \dots i_k} = 0\) when \(k{-}1\) (or all \(k\)) of the subscripts \(i_1, \dots, i_k\) equal 1. # The \(3 \times 3 \times 3\) hyperdeterminant {#hdsection} # An application to Nurmiev's normal forms {#nurmiev} Nurmiev classified the normal forms of \(3 \times 3 \times 3\) arrays over \(\mathbb{C}\) with respect to the action of the Lie group \(G = SL_3(\mathbb{C}) \times SL_3(\mathbb{C}) \times SL_3(\mathbb{C})\), and in Nurmiev described the orbit closure poset. Every \(3 \times 3 \times 3\) array \(M\) can be written uniquely in the form \(M = S + N\) where \(S\) is semisimple, \(N\) is nilpotent, and \([S,N] = 0\) where the Lie bracket is taken from Vinberg's embedding of the vector space of \(3 \times 3 \times 3\) arrays into a simple Lie algebra of type \(E_6\). Nurmiev defines three basic semisimple arrays; we use the notation \(E_{ijk}\) for the array with 1 in position \((i,j,k)\) and 0 in the other positions: \[X_1 = E_{111} + E_{222} + E_{333}, \qquad X_2 = E_{123} + E_{231} + E_{312}, \qquad X_3 = E_{132} + E_{213} + E_{321}.\] Nurmiev shows that every semisimple array is equivalent to (i.e. lies in the \(G\)-orbit of) an array of the normal form \[u = a_1 X_1 + a_2 X_2 + a_3 X_3.\] For each type of orbit, non-vanishing of an invariant at a particular point is always a certainty, so evaluating at pseudorandom points is sufficient. To be certain of vanishing for all points on a parametrized orbit, we must check the expression for the invariant restricted to the parametrized normal form. Evaluating the three fundamental invariants and the hyperdeterminant on \(u\) gives the following results: \[\label{eq:array} \left. \begin{array}{rll} I_6 &= a_1^6 + a_2^6 + a_3^6-10a_1^3a_2^3-10a_1^3a_3^3-10a_2^3a_3^3, \\[4pt] I_9 &= -(a_1-a_2)(a_1-a_3)(a_2-a_3) (a_1^2 + a_1a_2 + a_2^2)(a_1^2 + a_1a_3 + a_3^2)(a_2^2 + a_2a_3 + a_3^2), \\[4pt] I_{12} &= a_1^3a_2^9 + a_1^9a_2^3 + a_1^9a_3^3 + a_1^3a_3^9 + a_2^3a_3^9 + a_2^9a_3^3 -4a_1^6a_2^6-4a_1^6a_3^6-4a_2^6a_3^6 \\[4pt] &\qquad + 2a_1^3a_2^3a_3^6 + 2a_1^3a_2^6a_3^3 + 2a_1^6a_2^3a_3^3, \\[4pt] \Delta &= -4a_1^3a_2^3a_3^3 (a_1 + a_2 + a_3)^3 (a_1^2 + a_2^2 + a_3^2 + 2a_1a_2-a_1a_3-a_2a_3)^3 \,\times \\[4pt] &\qquad (a_1^2 + a_2^2 + a_3^2-a_1a_2 + 2a_1a_3-a_2a_3)^3 (a_1^2 + a_2^2 + a_3^2-a_1a_2-a_1a_3 + 2a_2a_3)^3 \,\times \\[4pt] &\qquad (a_1^2 + a_2^2 + a_3^2-a_1a_2-a_2a_3-a_1a_3)^3. \end{array} \quad\right\}\] We note that \(I_9\) vanishes whenever two of the coefficients \(a_1, a_2, a_3\) are equal, and that \(\Delta\) vanishes whenever any coefficient is 0 or the sum \(a_1 + a_2 + a_3\) is 0. According to the Nurmiev classification, the normal forms belong to five families. ## First family The coefficients of the semisimple part satisfy \[a_1 a_2 a_3 \ne 0, \qquad ( a_1^3 + a_2^3 + a_3^3 )^3-( 3 a_1 a_2 a_3 )^3 \ne 0.\] The nilpotent part is 0, so the values of the invariants on this family are as in [\[eq:array\]](#eq:array){reference-type="eqref" reference="eq:array"}. ## Second family The coefficients of the semisimple part satisfy \[a_2 ( a_1^3 + a_2^3 ) \ne 0, \qquad a_3 = 0.\] The possible nilpotent parts are \[E_{132} + E_{213}, \qquad E_{132}, \qquad 0.\] It turns out that the values of the fundamental invariants and the hyperdeterminant do not depend on the nilpotent part, and for the second family we find that: \[\begin{aligned} & I_6 = a_1^6 + a_2^6-10a_1^3a_2^3, \qquad I_9 =-a_1^3a_2^3(a_1-a_2)(a_1^2 + a_1a_2 + a_2^2), \\ & I_{12} = a_1^3a_2^3(a_1^6 + a_2^6-4a_1^3a_2^3), \qquad \Delta = 0. \end{aligned}\] ## Third family The coefficients of the semisimple part satisfy \[a_1 \ne 0, \qquad a_2 = a_3 = 0.\] The possible nilpotent parts are \[\begin{aligned} & E_{123} + E_{132} + E_{213} + E_{231}, \quad E_{123} + E_{132} + E_{213}, \quad E_{123} + E_{132} + E_{231}, \\ & E_{123} + E_{132}, \quad E_{123} + E_{231}, \quad E_{132} + E_{213}, \quad E_{123}, \quad E_{132}, \quad 0. \end{aligned}\] We find that the values of the fundamental invariants and the hyperdeterminant do not depend on the nilpotent part; for all normal forms in this family we have \(I_6 = a_1^6\) and \(I_9 = I_{12} = \Delta = 0\). ## Fourth family The coefficients of the semisimple part satisfy \[a_1 = 0, \qquad a_3 =-a_2 \ne 0.\] The possible nilpotent parts are \[\begin{aligned} & E_{113} + E_{131} + E_{222} + E_{311}, \quad E_{113} + E_{122} + E_{131} + E_{212} + E_{221} + E_{311}, \\ & E_{111} + E_{222}, \quad E_{112} + E_{121} + E_{211}, \quad E_{111}, \quad 0. \end{aligned}\] The values of the fundamental invariants and the hyperdeterminant do not depend on the nilpotent part; for all normal forms in this family we have \[I_6 = 12 a_2^6, \qquad I_9 = 2 a_2^9, \qquad I_{12} =-6 a_2^{12}, \qquad \Delta = 0.\] ## Fifth family This family contains 24 nonzero nilpotent arrays (the semisimple part is 0); see Nurmiev. In every case all the fundamental invariants and the hyperdeterminant are 0. To put this another way, we have verified that all nilpotent orbits consist of singular arrays (from the point of view of hyperdeterminants). # Values of the invariants on arrays of known rank {#ranksection} In this section we determine which invariants vanish on arrays of each possible rank. In principle this could be deduced from our previous computations by determining the rank of each of Nurmiev's normal forms, but we prefer to argue directly. Kruskal's theorem states that every \(3 \times 3 \times 3\) array over \(\mathbb{R}\) or \(\mathbb{C}\) has rank at most 5. To parametrize an array of rank \(r = 1\) (strictly speaking, \(r \le 1\)) we take the outer product of three vectors: \(x_{ijk} = a_i b_j c_k\). To make the array pseudorandom, we assign pseudorandom values to each of the coordinates of each of the vectors in the outer product. Then to generate parametrized (respectively pseudorandom) arrays of rank \(r \le 5\) we take the sum of \(r\) parametrized (respectively pseudorandom) arrays of rank 1. Let \(\sigma_{r}^{o}\) denote the set of arrays of rank \(\leq r\) and let \(\sigma_{r}\) denote its (Zariski) closure. It is well known that \(\sigma_{r}\subset \sigma_{r+1}\) for all \(r\), and obviously \(\sigma_{r}^{o}\subset \sigma_{r+1}^{o}\); however it can be that \(\sigma_{r}^{o} \subsetneq \sigma_{r}\). The commonly used example is \(x_{211}+x_{121}+x_{112}\), which is in \(\sigma_{2}\) (and \(\sigma_{3}^{o}\)) but not in \(\sigma_{2}^{o}\). We also note that since the zero-locus of an invariant is both (Zariski) closed and invariant under the group action, if an invariant vanishes at one point of an orbit, then it does so for all points in the orbit, and for all points in the orbit closure. We use this fact implicitly in what follows. The arrays of rank 1 form a single orbit, and we may test our invariants on one representative, such as \([1,0,0]^{\otimes 3}\). For \(\sigma_{2}^{o}\), we may consider the representative \([1,0,0]^{\otimes 3} + [0,1,0]^{\otimes 3}\). All other arrays of rank 2 are in the closure of the orbit of this array. Our tests find that all three fundamental invariants vanish on these representatives. Similarly, for \(\sigma_{3}^{o}\) we may consider the representative \([1,0,0]^{\otimes 3} + [0,1,0]^{\otimes 3} + [0,0,1]^{\otimes 3}\). All other arrays of rank 3 are in the closure of the orbit of this array. We find that \(I_{6}\) does not vanish on this representative, but \(I_{9}, I_{12}\) and \(\Delta\) do. Experimental results for 10000 pseudorandom arrays with entries in \(\{-9, \dots, 9 \}\) and rank \(\le 3\) showed that \(I_6\) vanishes in only 188 cases, but \(I_9\) and \(I_{12}\) vanish in all 10000 cases. For \(\sigma_{4}^{o}\), the value of \(I_{6}\) was already non-zero for rank \(3\), so it must not vanish identically on rank 4 arrays. It was already determined by Strassen that \(I_{9}\) vanishes on arrays of rank \(\le 4\). Experimental results for 10000 pseudorandom arrays with entries in \(\{-9, \dots, 9 \}\) and rank \(\le 4\) showed that \(I_6\) vanishes in only 4 cases, \(I_9\) vanishes in all 10000 cases, and \(I_{12}\) vanishes in 573 cases. In fact, just one non-vanishing result is sufficient to indicate that \(I_{12}\) does not vanish identically on the locus of rank 4 arrays. A similar test also shows that \(\Delta\) does not vanish identically for rank 4 arrays. The non-vanishing of \(I_{12}\) and \(\Delta\) on \(\sigma_{4}^{o}\) can also be deduced from Strassen's result that \(\sigma_{4}\) is an irreducible hypersurface defined by \(I_{9}\) and that neither \(I_{12}\) nor \(\Delta\) are multiples of \(I_{9}\). Strassen showed that his degree 9 invariant was non-zero for rank 5 arrays, thus proving that his invariant was non-trivial. The closure of rank 5 arrays is all of the ambient projective space, so no non-trivial invariant vanishes identically on all rank 5 arrays. Experimental results for 10000 pseudorandom arrays with entries in \(\{-9, \dots, 9 \}\) and rank \(\le 5\) showed that \(I_6\) never vanishes, \(I_9\) vanishes in 47 cases, and \(I_{12}\) vanishes in 23 cases. Again, just one non-vanishing example is sufficient to prove that the invariant is not identically zero on \(\sigma_{5}^{o}\). We summarize these results in Table [\[ranktable\]](#ranktable){reference-type="ref" reference="ranktable"}. \[\begin{array}{l|rrrrr} &\quad r \le 1 &\quad r \le 2 &\quad r \le 3 &\quad r \le 4 &\quad r \le 5 \\ \midrule I_6 &\quad 0 &\quad 0 &\quad \ne 0 &\quad \ne 0 &\quad \ne 0 \\ I_9 &\quad 0 &\quad 0 &\quad 0 &\quad 0 &\quad \ne 0 \\ I_{12} &\quad 0 &\quad 0 &\quad 0 &\quad \ne 0 &\quad \ne 0 \\ \Delta &\quad 0 &\quad 0 &\quad 0 &\quad \ne 0 &\quad \ne 0 \end{array}\] We can use this method of pseudorandom arrays to reduce the size of the matrix entries when we perform the computation of §[3](#hdsection){reference-type="ref" reference="hdsection"} using rational arithmetic. There are 7 nonzero vectors of dimension 3 with entries in \(\{0,1\}\), and hence \(7^3 = 343\) nonzero \(3 \times 3 \times 3\) arrays of rank 1. We use a pseudorandom number generator to choose 5 of these arrays; their sum \(X = ( x_{ijk} )\) is an array of rank \(\le 5\), which has rank 5 with sufficiently high probability; in this case, all three fundamental invariants will have nonzero values on \(X\). We set \(x_{ijk} = 0\) whenever at least two of \(i,j,k\) equal 1, to ensure that \(X\) has hyperdeterminant 0. Following the same procedure as in §[3](#hdsection){reference-type="ref" reference="hdsection"}, we obtain the matrix in Table [\[char0matrix\]](#char0matrix){reference-type="ref" reference="char0matrix"}. The vector of equation [\[coef0\]](#coef0){reference-type="eqref" reference="coef0"} is a basis for the nullspace, confirming Theorem [\[maintheorem\]](#maintheorem){reference-type="ref" reference="maintheorem"} with rational arithmetic. \[\left[ \begin{array}{rrrrrrr} 12230590464 &\!-509607936 &\! 28311552 &\! 21233664 &\!-1179648 &\! 65536 &\!-884736 \\ 64000000 &\!-2560000 &\! 128000 &\! 102400 &\!-5120 &\! 256 &\!-4096 \\ 2985984 &\!-124416 &\! 6912 &\! 5184 &\!-288 &\! 16 &\!-216 \\ 75418890625 &\!-2427685000 &\! 17576000 &\! 78145600 &\!-565760 &\! 4096 &\!-2515456 \\ 12230590464 &\!-509607936 &\! 28311552 &\! 21233664 &\!-1179648 &\! 65536 &\!-884736 \\ 531441 &\!-19683 &\! 729 &\! 729 &\!-27 &\! 1 &\!-27 \\4096 &\! 512 &\!-256 &\! 64 &\!-32 &\! 16 &\! 8 \\ 13841287201 &\!-449654478 &\! 4235364 &\! 14607684 &\!-137592 &\! 1296 &\!-474552 \\ 2985984 &\!-124416 &\! 6912 &\! 5184 &\!-288 &\! 16 &\!-216 \\ 11390625 &\!-303750 &\!-13500 &\! 8100 &\! 360 &\! 16 &\!-216 \end{array} \right]\]

{'timestamp': '2014-02-18T02:15:07', 'yymm': '1310', 'arxiv_id': '1310.3257', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3257'}

# Introduction In the representation theory of Kac-Moody algebras, of great importance is the category of integrable modules, where classifying irreducible integrable modules has been an open problem (see ). For (finite rank) affine Kac-Moody algebras, irreducible integrable modules with finite-dimensional weight spaces were classified by Chari (see ). Integrable modules for affine Kac-Moody algebras were studied further in (cf., ). Integrable representations for infinite rank affine Kac-Moody algebras, including \(\gl_{\infty}\) (the Lie algebra of doubly infinite matrices with only finitely many nonzero entries), are also of great importance in many areas, especially in mathematical physics. Among the most important and interesting results is the remarkable relation of highest weight integrable representations with soliton equations, which was discovered and developed by Kyoto school (see, , ). Lie algebra \(\gl_{\infty}\) has also been used effectively to study \({\mathcal{W}}\)-algebras (see, ). In a previous study, we exhibited a natural association of quantum vertex algebras (see ) to \(\gl_{\infty}\). In that study we came across a category of \(\gl_{\infty}\)-modules \(W\) satisfying the condition that for any \(m\in \Z,\ w\in W\), \(E_{m,n}w=0\) for all but finitely many integers \(n\). (A canonical base of \(\gl_{\infty}\) consists of \(E_{m,n}\) \((m,n\in \Z)\), where \(E_{m,n}\) denotes the matrix whose only nonzero entry is the \((m,n)\)-entry which is \(1\).) This category contains the natural module \(\C^{\infty}\) and its tensor products, whereas it excludes nontrivial highest weight modules and lowest weight modules. It is our hope to classify the irreducible objects in this category. This is the main motivation for this current paper. In this paper, we focus on a smaller category of \(\gl_{\infty}\)-modules, for which we are able to determine and classify all the irreducible objects. Specifically, we study \(\gl_{\infty}\)-modules \(W\) satisfying the condition that for every \(w\in W\), there exists a finite subset \(S\) of \(\Z\) such that \[E_{m,n}w=0\ \ \ \mbox{ for }m,n\in \Z,\ n\notin S.\] Denote by \({\mathcal{C}}\) the category of such \(\gl_{\infty}\)-modules and by \({\mathcal{C}}_{int}\) the category of those integrable \(\gl_{\infty}\)-modules. The category \({\mathcal{C}}\) still contains the natural module \(\C^{\infty}\), and it is closed under tensor product. Note that for a finite-dimensional simple Lie algebra, or more generally for a Kac-Moody algebra (see ), one has the well known category \(o\) and its subcategory \(o_{int}\) of integrable modules. To a certain extent, this category \({\cal{C}}\) of \(\gl_{\infty}\)-modules is analogous to category \(o\). Especially, it is proved that every module in category \({\mathcal{C}}_{int}\) is completely reducible. By using certain generalized Verma modules, we can classify irreducible modules in categories \({\mathcal{C}}\) and \({\mathcal{C}}_{int}\). On the other hand, we can also determine the decomposition of tensor products of irreducible modules in category \({\mathcal{C}}_{int}\) in terms of the decomposition of tensor products of irreducible modules for \(\gl_{n}\) with sufficiently large \(n\). We now give a more detailed description of the main results. Let \(S\) be a finite and nonempty subset of \(\Z\). To \(S\) we associate a triangular decomposition \[\gl_{\infty}=\gl_{\infty}^{(S,+)}\oplus \gl_{\infty}^{(S,0)}\oplus \gl_{\infty}^{(S,-)},\] where \[\begin{aligned} &&\gl_{\infty}^{(S,+)}={\rm span}\{ E_{m,p}\ |\ m\in \Z,\ p\notin S\},\ \ \ \ \gl_{\infty}^{(S,-)}= {\rm span}\{ E_{p,n}\ |\ p\notin S,\ n\in S\},\\ &&\hspace{3cm} \gl_{\infty}^{(S,0)}={\rm span}\{ E_{m,n}\ |\ m,n\in S\}. \end{aligned}\] Alternatively, set \[\gl_{S}=\gl_{\infty}^{(S,0)}={\rm span}\{ E_{m,n}\ |\ m,n\in S\}.\] Given a \(\gl_{S}\)-module \(U\), using this particular triangular decomposition we define a generalized Verma module \(M(S,U)\), which is a \(\gl_{\infty}\)-module induced from \(\gl_{S}\)-module \(U\). When \(U\) is irreducible, \(M(S,U)\) has a unique irreducible quotient module, denoted by \(L(S,U)\). We show that \(M(S,U)\) and \(L(S,U)\) belong to category \({\cal{C}}\) and that every irreducible \(\gl_{\infty}\)-module in \({\cal{C}}\) is isomorphic to \(L(S,U)\) for some finite subset \(S\) of \(\Z\) and for some irreducible \(\gl_{S}\)-module \(U\). We also give a necessary and sufficient condition that \(L(S_{1},U_{1})\simeq L(S_{2},U_{2})\), where \(S_{1},S_{2}\) are finite subsets of \(\Z\) and \(U_{1},U_{2}\) are irreducible modules for \(\gl_{S_{1}}\) and \(\gl_{S_{2}}\), respectively. Set \[H={\rm span}\{ E_{n,n}\ |\ n\in \Z\},\] a Cartan subalgebra of \(\gl_{\infty}\). Let \(S\) be a finite nonempty subset of \(\Z\). Set \[H_{S}={\rm span}\{ E_{n,n}\ |\ n\in S\},\] which is a Cartan subalgebra of \(\gl_{S}\). Let \(\lambda\in H^{*}\) such that \(\lambda(E_{n,n})=0\) for \(n\notin S\). Define \(M(S,\lambda)\) to be the generalized Verma module \(M(S,U)\) with \(U=M(\lambda_{S})\), where \(M(\lambda_{S})\) denotes the Verma \(\gl_{S}\)-module with highest weight \(\lambda_{S}=\lambda_{H_{S}}\). A fact is that the weight-\(\lambda\) subspace of \(M(S,\lambda)\) is one-dimensional. We then define an irreducible module \(L(S,\lambda)\) as the quotient module of \(M(S,\lambda)\) by the maximal submodule. We show that \(L(S,\lambda)\) is integrable if and only if \(\lambda(E_{n,n})\in \N\) for \(n\in \Z\) and \(\lambda(E_{m,m})\ge \lambda(E_{n,n})\) for \(m,n\in S\) with \(m<n\). Furthermore, we show that every such irreducible integrable module belongs to category \({\cal{C}}_{int}\) and every irreducible \(\gl_{\infty}\)-module in \({\cal{C}}_{int}\) is isomorphic to such an irreducible module. We furthermore study the tensor products of irreducible modules in \({\mathcal{C}}_{int}\). It is shown that the tensor product of any two irreducible modules in \({\mathcal{C}}_{int}\) is always cyclic. Let \(W\) be any \(\gl_{\infty}\)-module in \({\mathcal{C}}\) and let \(S\) be a finite subset of \(\Z\). Set \[\Omega_{S}(W)=\{ w\in W\ |\ E_{p,q}w=0\ \ \mbox{ for any }p,q\in \Z\ \mbox{with }q\notin S\},\] which is a \(\gl_{S}\)-submodule of \(W\). It is proved that if \(W\) is irreducible, there exists a finite subset \(S'\) of \(\Z\) such that \(\Omega_{S}(W)\) is an irreducible \(\gl_{S}\)-module for any finite subset \(S\) of \(\Z\), containing \(S'\). Now, let \(W_{1},W_{2}\) be irreducible \(\gl_{\infty}\)-modules in \({\mathcal{C}}_{int}\). It is proved that for any sufficiently large \(S\), the decomposition of \(W_{1}\otimes W_{2}\) into irreducible \(\gl_{\infty}\)-modules is determined by the decomposition of \(\Omega_{S}(W_{1})\otimes \Omega_{S}(W_{2})\) into irreducible \(\gl_{S}\)-modules. Consequently, the tensor product of any two irreducible modules in category \({\mathcal{C}}_{int}\) always decomposes into a finite sum of irreducible modules, unlike the case with highest weight modules. After this paper was completed, we found a very interesting paper, in which Penkov and Serganova had studied the category of integrable modules with finite dimensional weight subspaces for \(sl(\infty)\), \(o(\infty)\), and \(sp(\infty)\). Among other results they proved that the category of integrable modules with finite dimensional weight subspaces is semisimple and they also identified each irreducible module in this category. Their method is somewhat different from that of this present paper. This paper is organized as follows: In Section 2, we define categories \({\cal{C}}\) and \({\cal{C}}_{int}\) of \(\gl_{\infty}\)-modules and we establish the complete reducibility. In Section 3, we classify irreducible modules in categories \({\mathcal{C}}\) and \({\mathcal{C}}_{int}\). In Section 4, we study the decomposition of the tensor product modules. # Categories \({\mathcal{C}}\) and \({\cal{C}}_{int}\) of \(\gl_{\infty}\)-modules In this section, we introduce a category \({\cal{C}}\) of \(\gl_{\infty}\)-modules and a subcategory \({\cal{C}}_{int}\) of integrable modules. As the main result of this section, we prove that every \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) is completely reducible. We begin with Lie algebra \(\gl_{\infty}\), which is the Lie algebra of doubly infinite matrices with only finitely many nonzero entries, under the commutator bracket. A canonical base consists of \(E_{m,n}\) \((m,n\in \Z)\), where \(E_{m,n}\) denotes the matrix whose only nonzero entry is the \((m,n)\)-entry which is \(1\), and we have \[\begin{aligned} =\delta_{n,r}E_{m,s}-\delta_{m,s}E_{r,n} \end{aligned}\] for \(m,n,r,s\in \Z\). Let \(\C^{\infty}\) denote the vector space of doubly infinite column vectors with only finitely many nonzero entries. Denote the standard unit base vectors by \(v_{n}\) for \(n\in \Z\). The natural action of \(\gl_{\infty}\) on \(\C^{\infty}\) is given by \[E_{i,j}v_{k}=\delta_{j,k}v_{i}\ \ \ \mbox{ for }i,j, k\in \Z.\] Define \(\deg E_{i,j}=j-i\) for \(i,j\in \Z\) to make \(\gl_{\infty}\) a \(\Z\)-graded Lie algebra, where the degree-\(n\) homogeneous subspace \((\gl_{\infty})_{(n)}\) for \(n\in \Z\) is linearly spanned by \(E_{m,m+n}\) for \(m\in \Z\). We have the standard triangular decomposition \[\gl_{\infty}=\gl_{\infty}^{+}\oplus \gl_{\infty}^{0}\oplus \gl_{\infty}^{-},\] where \(\gl_{\infty}^{\pm}=\sum_{\pm(j-i)>0}\C E_{i,j}\) and \(\gl_{\infty}^{0}=\sum_{n\in \Z}\C E_{n,n}\). Alternatively, set \[\begin{aligned} H=\gl_{\infty}^{0}={\rm span}\{ E_{n,n}\ |\ n\in \Z\}, \end{aligned}\] a Cartan subalgebra of \(\gl_{\infty}\). The formal completion \(\overline{\gl_{\infty}}\) of \(\gl_{\infty}\) is also a \(\Z\)-graded Lie algebra, where \[\overline{(\gl_{\infty})}_{(n)}=\left\{ \sum_{m\in \Z} a_{m}E_{m,m+n}\ |\; a_{m}\in \C\right\}.\] If \(W\) is a \(\gl_{\infty}\)-module such that for every \(w\in W\) and for every \(n\in \Z\), \(E_{m,m+n}w=0\) for all but finitely many integers \(m\), then \(W\) is naturally a \(\overline{\gl_{\infty}}\)-module. *A \(\gl_{\infty}\)-module \(W\) is said to be *integrable* if for every \(n\in \Z\), \(E_{n,n}\) is semi-simple on \(W\) and if for any \(p,q\in \Z\) with \(p\ne q\), \(E_{p,q}\) is locally nilpotent on \(W\).* A notion of integrable module for a general Lie algebra including Kac-Moody Lie algebras was introduced in. This definition of an integrable module for \(\gl_{\infty}\) is just a version of that. From (Proposition 3.8), a \(\gl_{\infty}\)-module \(W\) is integrable if \(E_{n,n}\) is semi-simple on \(W\) for *some* \(n\in \Z\) and if \(E_{j,j+1}\) and \(E_{j+1,j}\) for all \(j\in \Z\) are locally nilpotent on \(W\). *Let \(\g\) be a finite-dimensional simple Lie algebra with Chevalley generators \(e_{i},f_{i},h_{i}\) \((1\le i\le l)\). A \(\g\)-module \(W\) is *integrable* if \(e_{i},f_{i}\) \((1\le i\le l)\) are locally nilpotent on \(W\). From (Proposition 3.8), every integrable \(\g\)-module is \(\g\)-locally finite, hence a direct sum of finite-dimensional irreducible modules. Consequently, on an integrable \(\g\)-module, every root vector of \(\g\) is locally nilpotent and the Cartan algebra is semi-simple.* *Consider the three dimensional simple Lie algebra \(\g=\sl(2,\C)\) with the standard Chevalley generators \(e,f,h\). Suppose that \(V\) is an integrable \(\sl(2,\C)\)-module with a nonzero vector \(v\) satisfying the condition that \(ev=0\) and \(hv=kv\) for some \(k\in \C\). Then \(k\in \N\), \(f^{k+1}v=0\), and \(e^{k}f^{k}v=(k!)^{2}v\).* Recall that a highest weight \(\gl_{\infty}\)-module with highest weight \(\lambda\in H^{*}\) is a module \(W\) with a vector \(w\) such that \[\begin{aligned} &&E_{n,n}w=\lambda_{n}w,\ \ E_{n,n+1}w=0\ \ \ \mbox{ for }n\in \Z,\\ &&W=U(\gl_{\infty})w, \end{aligned}\] where \(\lambda_{n}=\lambda(E_{n,n})\). From, a highest weight integrable \(\gl_{\infty}\)-module is irreducible. *Denote by \({\cal{C}}\) the category of \(\gl_{\infty}\)-modules \(W\) such that for any \(w\in W\), there exists a finite subset \(S\) of \(\Z\) such that \(E_{m,n}w=0\) for all \(m,n\in \Z\) with \(n\notin S\). Furthermore, define \({\cal{C}}_{int}\) to be the subcategory consisting of integrable \(\gl_{\infty}\)-modules in \({\cal{C}}\).* It can be readily seen that every submodule of a \(\gl_{\infty}\)-module from \({\cal{C}}\) is in \({\cal{C}}\) and the tensor product of any (finitely many) \(\gl_{\infty}\)-modules from \({\cal{C}}\) is in \({\cal{C}}\). The same can be said for category \({\cal{C}}_{int}\). Set \[\begin{aligned} I_{\infty}=\sum_{n\in \Z}E_{n,n}, \end{aligned}\] which lies in the completion \(\overline{\gl_{\infty}}\) of \(\gl_{\infty}\). Notice that every \(\gl_{\infty}\)-module in category \({\cal{C}}\) is naturally a \(\overline{\gl_{\infty}}\)-module. Then for any \(\gl_{\infty}\)-module \(W\) in category \({\cal{C}}\), \(I_{\infty}\) is a well defined operator on \(W\) and \([I_{\infty},\gl_{\infty}]=0\). Let \(S\) be a finite and nonempty subset of \(\Z\). Set \[\begin{aligned} \gl_{S}={\rm span}\{ E_{m,n}\ |\ m,n\in S\}, \end{aligned}\] which is a subalgebra of \(\gl_{\infty}\), and set \[\begin{aligned} I_{S}=\sum_{n\in S}E_{n,n}\in \gl_{S}. \end{aligned}\] The Lie algebra \(\gl_{S}\) is reductive with one-dimensional center \(\C I_{S}\). Let \(W\) be any \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\). Then for every \(n\in \Z\), \(E_{n,n}\) is semi-simple on \(W\) with only nonnegative integer eigenvalues. Furthermore, \(I_{\infty}\) is semi-simple on \(W\) with only nonnegative integer eigenvalues and \(H\) is semi-simple. Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\). From Lemma [\[lenn-eigenvalues\]](#lenn-eigenvalues){reference-type="ref" reference="lenn-eigenvalues"} we have \[W=\oplus _{\ell \in \N}W[\ell],\] a direct sum of \(\gl_{\infty}\)-modules, where \(W[\ell]=\{ w\in W\;|\; I_{\infty}\cdot w=\ell w\}\) for \(\ell\in \N\). In view of this, it suffices to determine \(\gl_{\infty}\)-modules in category \({\cal{C}}_{int}\), on which \(I_{\infty}\) acts as a nonnegative integer scalar. As a refinement of Lemma [\[lenn-eigenvalues\]](#lenn-eigenvalues){reference-type="ref" reference="lenn-eigenvalues"} we have: Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) such that \(I_{\infty}\) acts as a nonnegative integer scalar \(\ell\). Then for every \(n\in \Z\), \(E_{n,n}\) is semi-simple with nonnegative integer eigenvalues not exceeding \(\ell\) and \((E_{m,n})^{\ell+1}=0\) for all \(m,n\in \Z\) with \(m\ne n\). The following is an immediate consequence of Lemma [\[lsecond\]](#lsecond){reference-type="ref" reference="lsecond"}: Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) such that \(I_{\infty}=0\). Then \(\gl_{\infty}\) acts trivially on \(W\). Next, we have: Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) such that \(I_{\infty}=1\) on \(W\). Then \(W\) is a direct sum of submodules isomorphic to the natural module \(\C^{\infty}\). Next, we shall prove that every \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) is completely reducible. First, we prove a technical result. Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\). Suppose that \(v\) is a nonzero \(H\)-eigenvector of weight \(\lambda\), satisfying the condition that \[\begin{aligned} \label{eassumption-2.28} &&E_{p,q}v=0\ \ \ \mbox{ for all }p,q\in \Z\ \mbox{ with }\ q\notin \{r,r+1,\dots,n\},\nonumber\\ &&E_{r,j+1}v=0\ \ \ \mbox{ for }r\le j\le n, \end{aligned}\] where \(r\) and \(n\) are some fixed integers with \(r<n\). Set \[v'=(E_{r-1,r})^{\lambda_{r}-\lambda_{r+1}}\cdots (E_{r-1,n-1})^{\lambda_{n-1}-\lambda_{n}}(E_{r-1,n})^{\lambda_{n}}v\in W.\] Then \[\begin{aligned} \label{eresult-2.29} &&E_{p,q}v'=0\ \ \ \mbox{ for all }p,q\in \Z\ \mbox{ with }\ q\notin \{r-1,r,\dots,n\},\nonumber\\ &&E_{r-1,j+1}v'=0\ \ \ \mbox{ for }r-1\le j\le n. \end{aligned}\] Furthermore, we have \[\begin{aligned} (E_{n,r-1})^{\lambda_{n}}(E_{n-1,r-1})^{\lambda_{n-1}-\lambda_{n}}\cdots (E_{r,r-1})^{\lambda_{r}-\lambda_{r+1}}v'=\alpha v, \end{aligned}\] where \(\alpha=(\lambda_{n}!)^{2}((\lambda_{n-1}-\lambda_{n})!)^{2}\cdots ((\lambda_{r}-\lambda_{r+1})!)^{2}\), a nonzero integer. On the basis of Lemma [\[lpreparation\]](#lpreparation){reference-type="ref" reference="lpreparation"} we have: Let \(W\) be an integrable \(\gl_{\infty}\)-module and let \(v\in W\) be a nonzero \(H\)-weight vector satisfying the condition that \[\begin{aligned} &&E_{p,q}v=0\ \ \ \mbox{ for }p,q\in \Z \ \mbox{ with }\ q\notin \{r,r+1,\dots,n\},\nonumber\\ &&E_{j,j+1}v=0\ \ \ \mbox{ for }r\le j\le n-1, \end{aligned}\] where \(r\) and \(n\) are some integers with \(r<n\). Then the submodule \(U(\gl_{\infty})v\) is irreducible. Now, we are in a position to present our main result. Every \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) is completely reducible. # Classification of irreducible \(\gl_{\infty}\)-modules in \({\cal{C}}\) and \({\cal{C}}_{int}\) In this section, we classify irreducible \(\gl_{\infty}\)-modules in categories \({\cal{C}}\) and \({\cal{C}}_{int}\). To achieve this goal, for any finite subset \(S\) of \(\Z\) and for any irreducible \(\gl_{S}\)-module \(U\), through a generalized Verma module construction we construct an irreducible \(\gl_{\infty}\)-module \(L(S,U)\) in category \({\cal{C}}\) and we show that any irreducible module in \({\cal{C}}\) is isomorphic to a module of this form. Furthermore, for a linear functional \(\lambda\) on \(H\) compatible with \(S\) in a certain sense we define a generalized Verma \(\gl_{\infty}\)-module \(M(S,\lambda)\) and construct an irreducible module \(L(S,\lambda)\). We then determine when \(L(S,\lambda)\) is integrable and we show that every irreducible module in \({\cal{C}}_{int}\) is isomorphic to such an integrable module \(L(S,\lambda)\). Let \(S\) be a *finite* and *nonempty* subset of \(\Z\), which is fixed temporarily. Recall \[\begin{aligned} \gl_{S}={\rm span}\{ E_{m,n}\ |\ m,n\in S\}\subset \gl_{\infty}. \end{aligned}\] To \(S\), we associate a triangular decomposition \[\begin{aligned} \gl_{\infty}=\gl_{\infty}^{(S,+)}\oplus \gl_{\infty}^{(S,0)}\oplus \gl_{\infty}^{(S,-)}, \end{aligned}\] where \(\gl_{\infty}^{(S,0)}=\gl_{S}\), \[\begin{aligned} \gl_{\infty}^{(S,+)}&=&\span\{ E_{m,n}\;|\; m,n\in \Z,\; n\notin S\},\nonumber\\ \gl_{\infty}^{(S,-)}&=&\span\{ E_{m,n}\;|\; m,n\in \Z,\; m\notin S,\; n\in S\}. \end{aligned}\] Notice that \(\gl_{\infty}^{(S,-)}\) is an abelian subalgebra and that \(\gl_{\infty}^{(S,+)}+\gl_{S}\) is a semi-product. Recall that \[\begin{aligned} I_{S}=\sum_{n\in S}E_{n,n}\in \gl_{S}. \end{aligned}\] We have \[\begin{aligned} &&[I_{S},E_{m,n}]=0\ \mbox{ for either }m,n\in S, \mbox{ or }m,n\notin S,\nonumber\\ &&[I_{S},E_{m,n}]=E_{m,n}\ \mbox{ for }m\in S,\ n\notin S,\nonumber\\ &&[I_{S},E_{m,n}]=-E_{m,n}\ \mbox{ for }m\notin S,\; n\in S. \end{aligned}\] Using \(\ad(I_{S})\), we make \(\gl_{\infty}\) a \(\Z\)-graded Lie algebra for which \[\begin{aligned} &&\gl_{\infty}^{(n)}=0\ \ \mbox{ if }|n|\ge 2,\nonumber\\ &&\gl_{\infty}^{(0)}={\rm span} \{E_{m,n}\ |\ \mbox{ either }m,n\in S, \mbox{ or }m,n\notin S\},\nonumber\\ &&\gl_{\infty}^{(1)}={\rm span} \{E_{m,n}\ |\ m\in S,\ n\notin S\},\nonumber\\ &&\gl_{\infty}^{(-1)}={\rm span} \{E_{m,n}\ |\ m\notin S,\ n\in S\}. \end{aligned}\] Set \[\begin{aligned} \gl_{S^{o}}={\rm span}\{ E_{m,n}\ |\ m,n\notin S\}. \end{aligned}\] Then \[\begin{aligned} \gl_{\infty}^{(0)}=\gl_{S}\oplus \gl_{S^{o}}, \end{aligned}\] a direct product. Notice that \(\gl_{\infty}^{(\pm 1)}\) both are abelian subalgebras. We see that \[\begin{aligned} \gl_{\infty}^{(S,+)}=\gl_{\infty}^{(1)}\oplus \gl_{S^{o}},\ \ \ \ \gl_{\infty}^{(S,-)}=\gl_{\infty}^{(-1)}. \end{aligned}\] Let \(U\) be a \(\gl_{S}\)-module. Let \(\gl_{\infty}^{(S,+)}\) act trivially on \(U\), to make \(U\) a \((\gl_{\infty}^{(S,+)}+\gl_{S})\)-module. Then form a generalized Verma \(\gl_{\infty}\)-module \[\begin{aligned} M(S,U)=U(\gl_{\infty})\otimes _{U\left(\gl_{\infty}^{(S,+)}+\gl_{S}\right)}U. \end{aligned}\] In view of the P-B-W theorem we have \[M(S,U)=U(\gl_{\infty}^{(-1)})\otimes U=S(\gl_{\infty}^{(-1)})\otimes U.\] By endowing \(U\) with degree \(0\), we make \(M(S,U)\) a \(\Z\)-graded \(\gl_{\infty}\)-module. (The homogeneous subspaces are infinite-dimensional in general.) The \(\gl_{\infty}\)-module \(M(S,U)\) belongs to the category \({\cal{C}}\). *Let \(U\) be a \(\gl_{S}\)-module as before. Denote by \(L(S,U)\) the quotient of the \(\Z\)-graded \(\gl_{\infty}\)-module \(M(S,U)\) by the maximal graded submodule with trivial degree-\(0\) homogeneous subspace.* *Assume that \(U\) is a \(\gl_{S}\)-module on which \(I_{S}\) acts as a scalar \(\alpha\in \C\). Then \(M(S,U)\) is a canonically graded \(\gl_{\infty}\)-module \[\begin{aligned} M(S,U)=\bigoplus _{n\in \N}M(S,U)_{\alpha-n}, \end{aligned}\] where \(M(S,U)_{\alpha-n}=\{ w\in M(S,U)\; |\; I_{S}\cdot w=(\alpha-n)w\}\) for \(n\in \N\).* Let \(W\) be a \(\gl_{\infty}\)-module and let \(S\) be a finite subset of \(\Z\) as before. Set \[\begin{aligned} \Omega_{S}(W)=\left\{ w\in W\; |\; E_{p,q}w=0\ \ \mbox{ for all }p,q\in \Z\ \mbox{ with }q\notin S\right\}. \end{aligned}\] It can be readily seen that \(\Omega_{S}(W)\) is a \(\gl_{S}\)-submodule of \(W\). Let \(S\) be a finite subset of \(\Z\) and let \(U\) be an irreducible \(\gl_{S}\)-module. Then \(L(S,U)\) is an irreducible \(\gl_{\infty}\)-module belong to category \({\cal{C}}\). On the other hand, every irreducible \(\gl_{\infty}\)-module in category \({\cal{C}}\) is isomorphic to \(L(S,U)\) for some finite subset \(S\) of \(\Z\) and for some irreducible \(\gl_{S}\)-module \(U\). From the second part of the proof of Proposition [\[panyC\]](#panyC){reference-type="ref" reference="panyC"} we immediately have: Let \(W\) be an irreducible \(\gl_{\infty}\)-module in \({\mathcal{C}}\). Then there exists a finite subset \(S\) of \(\Z\) such that \(\Omega_{S}(W)\ne 0\). Furthermore, for any such finite subset \(S\) of \(\Z\), \(\Omega_{S}(W)\) is an irreducible \(\gl_{S}\)-module and \(W\simeq L(S,\Omega_{S}(W))\). The following is also immediate: Let \(S\) be a finite subset of \(\Z\) and let \(U\) be an irreducible \(\gl_{S}\)-module. Then \(\Omega_{S}(L(S,U))=U\). We next determine the isomorphism classes of irreducible \(\gl_{\infty}\)-modules \(L(S,U)\). Let \(S\) be a finite subset of \(\Z\) and \(S_{1}\) a subset of \(S\). Assume that \(U_{1}\) is an irreducible \(\gl_{S_{1}}\)-module on which \(I_{S_{1}}\) acts as a scalar \(\alpha\in \C\). In the following we associate an irreducible \(\gl_{S}\)-module to \(U_{1}\). Set \[N={\rm span}\{ E_{p,q}\ |\ p,q\in S,\; q\notin S_{1}\}\ \ \mbox{ and }\ \ B=N+\gl_{S_{1}}.\] We see that both \(B\) and \(N\) are subalgebras of \(\gl_{S}\) and \(B\) contains \(N\) as an ideal. Letting \(N\) act on \(U_{1}\) trivially, we make \(U_{1}\) a \(B\)-module. Then form an induced module \[\begin{aligned} {\rm ind}_{\gl_{S_{1}}}^{\gl_{S}}(U_{1})=U(\gl_{S})\otimes_{U(B)}U_{1}. \end{aligned}\] As we have seen before, \(I_{S_{1}}\) gives rise to an \((\alpha+\Z)\)-grading on \({\rm ind}_{\gl_{S_{1}}}^{\gl_{S}}(U_{1})\) with \(U_{1}\) as the degree-\(\alpha\) subspace. It follows that \({\rm ind}_{\gl_{S_{1}}}^{\gl_{S}}(U_{1})\) has a unique maximal submodule. Then we define \(U_{1}^{S}\) to be the (unique) irreducible quotient \(\gl_{S}\)-module of \({\rm ind}_{\gl_{S_{1}}}^{\gl_{S}}(U_{1})\). Let \(S\) be a finite subset of \(\Z\) and \(S_{1}\) a subset of \(S\). Assume that \(U_{1}\) is an irreducible \(\gl_{S_{1}}\)-module. Then \(L(S_{1},U_{1})\simeq L(S,U_{1}^{S})\). As an immediate consequence of Lemmas [\[lvacuum-irred\]](#lvacuum-irred){reference-type="ref" reference="lvacuum-irred"} and [\[lu1us\]](#lu1us){reference-type="ref" reference="lu1us"} we have: Let \(S_{1}\) and \(S_{2}\) be finite subsets of \(\Z\) and let \(U_{1}\) and \(U_{2}\) be irreducible modules for \(\gl_{S_{1}}\) and \(\gl_{S_{2}}\), respectively. Set \(S=S_{1}\cup S_{2}\). Then \(L(S_{1},U_{1})\simeq L(S_{2},U_{2})\) if and only if \(U_{1}^{S}\simeq U_{2}^{S}\). Next, we study an analog of Verma module. For \(i\in \Z\), let \(\varepsilon_{i}\) denote the linear functional on \(H\) defined by \(\varepsilon_{i}(E_{j,j})=\delta_{i,j}\) for \(j\in \Z\). The root system of \(\gl_{\infty}\) with respect to Cartan subalgebra \(H\) is given by \[\begin{aligned} \Delta=\{ \varepsilon_{i}-\varepsilon_{j}\ |\ i,j\in \Z,\ i\ne j\}. \end{aligned}\] Note that the usual polarization is given by \(\Delta_{\pm}=\{\pm (\varepsilon_{i}-\varepsilon_{j})\ |\ i,j\in \Z,\ i<j\}\). Let \(S\) be a subset of \(\Z\). Recall \(\gl_{S}={\rm span}\{E_{m,n}\ |\ m,n\in S\}\). Set \[\begin{aligned} \gl_{S}^{\pm}&=&{\rm span}\{ E_{m,n}\ |\ m,n\in S,\ \pm (n-m)>0\},\nonumber\\ \gl_{S}^{0}&=&H_{S}={\rm span}\{ E_{n,n}\ |\ n\in S\}. \end{aligned}\] For \(\lambda\in H^{*}\), set \[{\rm supp}(\lambda)=\{ m\in \Z\ |\ \lambda_{m}\ (=\lambda(E_{m,m}))\ne 0\}.\] Let \(\lambda\in H^{*}\) with \({\rm supp}(\lambda)\subset S\). Denote by \(\lambda_{S}\) the restriction of \(\lambda\) on \(H_{S}\). Let \(M(\lambda_{S})\) and \(L(\lambda_{S})\) denote the Verma module and the irreducible quotient module for Lie algebra \(\gl_{S}\), respectively. Define \(M(S,\lambda)\) to be the generalized Verma \(\gl_{\infty}\)-module \(M(S,U)\) with \(U=M(\lambda_{S})\). On the other hand, we have a generalized Verma \(\gl_{\infty}\)-module \(M(S,L(\lambda_{S}))\), which is a quotient module of \(M(S,\lambda)\). Furthermore, denote by \(L(S,\lambda)\) the irreducible quotient module of \(M(S,L(\lambda_{S}))\). Set \[\begin{aligned} \Delta_{-}(S)=\{(\varepsilon_{i}-\varepsilon_{j})\ |\ i,j\in S,\ i>j\}\cup \{ (\varepsilon_{p}-\varepsilon_{i})\ |\ i\in S,\ p\notin S\}. \end{aligned}\] Furthermore, set \[\begin{aligned} Q_{-}(S)=\N\cdot \Delta_{-}(S)\subset H^{*}. \end{aligned}\] We have \[\begin{aligned} M(S,\lambda)=\oplus_{\alpha\in Q_{-}(S)}M(S,\lambda)_{\lambda+\alpha}, \end{aligned}\] where \(M(S,\lambda)_{\lambda}\) is \(1\)-dimensional. It is straightforward to show that every weight subspace is finite-dimensional. *Let \(W\) be a \(\gl_{\infty}\)-module and let \(S\) be a subset of \(\Z\). A nonzero vector \(v\in W\) is called an *\(S\)-singular vector* if \(v\) is an \(H\)-eigenvector such that \[\begin{aligned} &&E_{p,q}v=0\ \ \ \mbox{ for }p,q\in \Z \ \mbox{ with }\ q\notin S,\nonumber\\ &&E_{m,n}v=0\ \ \ \mbox{ for }m,n\in S\ \mbox{ with }m<n. \end{aligned}\]* The following universal property of \(M(S,\lambda)\) is straightforward to prove: Let \(W\) be a \(\gl_{\infty}\)-module and let \(w\) be an \(S\)-singular vector of weight \(\lambda\) in \(W\). Then there exists a \(\gl_{\infty}\)-module homomorphism \(\theta\) from \(M(S,\lambda)\) to \(W\), uniquely determined by \(\theta(v)=w\), where \(v\) is an \(S\)-singular vector in \(M(S,\lambda)\) of weight \(\lambda\). Furthermore, if \(W\) is irreducible, we have \(W\simeq L(S,\lambda)\).We also have the following result: Let \(S\) be a finite subset of \(\Z\) and let \(\lambda\in H^{*}\) with \({\rm supp}(\lambda)\subset S\). Then \(S\)-singular vectors in \(L(S,\lambda)\) are unique up to scalar multiples. As an immediate consequence we have: Let \(S\) be a finite subset of \(\Z\) and let \(\lambda,\mu\in H^{*}\) such that \({\rm supp}(\lambda)\), \({\rm supp}(\mu)\subset S\). Then \(L(S,\lambda)\simeq L(S,\mu)\) if and only if \(\lambda=\mu\). Let \(W\) be an irreducible \(\gl_{\infty}\)-module in category \({\cal{C}}\), satisfying the condition that for any vector \(w\in W\) and for any finite subset \(S\) of \(\Z\), \(U(\gl_{S})w\) is a \(\gl_{S}\)-module in category \(o\). Then \(W\simeq L(S,\lambda)\) where \(S=\{ r,r+1,\dots,n\}\) for some integers \(r\) and \(n\) with \(r<n\) and for some \(\lambda\in H^{*}\) with \({\rm supp}(\lambda)\subset S\). By Corollary [\[cu1u2\]](#cu1u2){reference-type="ref" reference="cu1u2"} we have: Let \(S_{1},S_{2}\) be finite subsets of \(\Z\) and let \(\lambda,\mu\in H^{*}\) be such that \({\rm supp}(\lambda)\subset S_{1}\) and \({\rm supp}(\mu)\subset S_{2}\). Then \(L(S_{1},\lambda)\simeq L(S_{2},\mu)\) if and only if \(L(\lambda_{S_{1}})^{S}\simeq L(\mu_{S_{2}})^{S}\), where \(S=S_{1}\cup S_{2}\), and \(L(\lambda_{S_{1}})\), \(L(\mu_{S_{2}})\) are the irreducible highest weight modules for \(\gl_{S_{1}}\) and \(\gl_{S_{2}}\), respectively. *Let \(\sigma\) be a permutation on \(\Z\). It can be readily seen that \(\sigma\) becomes an automorphism of the Lie algebra \(\gl_{\infty}\) by defining \[\begin{aligned} \sigma(E_{m,n})=E_{\sigma(m),\sigma(n)}\ \ \ \mbox{ for }m,n\in \Z. \end{aligned}\] For any \(\gl_{\infty}\)-module \(W\), we denote by \(W^{[\sigma]}\) the \(\gl_{\infty}\)-module with \(W\) as the underlying space and with the action given by \[a\cdot w=\sigma(a)w\ \ \ \mbox{ for }a\in \gl_{\infty},\ w\in W.\] We see that if \(W\) is an integrable \(\gl_{\infty}\)-module, then \(W^{[\sigma]}\) is still an integrable module. Furthermore, if \(W\) is in category \({\cal{C}}_{int}\), \(W^{[\sigma]}\) is still in category \({\cal{C}}_{int}\).* The following is straightforward to prove: Let \(\lambda\in H^{*}\) be such that \({\rm supp}(\lambda)\subset S\), and let \(\sigma\) be a permutation on \(\Z\) such that \(\sigma(m)<\sigma(n)\) for any \(m,n\in S\) with \(m<n\). Then \(L(S,\lambda)^{[\sigma]}\simeq L(\sigma^{-1}(S),\lambda\circ \sigma)\). Next, we determine when \(L(S,\lambda)\) is an integrable module. *Denote by \(P_{+}(S)\) the set of \(\lambda\in H^{*}\) such that \[\begin{aligned} &&\lambda(E_{i,i})\in \N\ \ \ \mbox{ for }i\in \Z,\nonumber\\ &&\lambda(E_{i,i})=0\ \ \ \mbox{ whenever }i\notin S,\nonumber\\ &&\lambda(E_{i,i})\ge \lambda(E_{j,j})\ \ \ \mbox{ for }i,j\in S\ \mbox{ with }i<j. \end{aligned}\]* Note that if \(\lambda \in P_{+}(S)\), then \({\rm supp}(\lambda)\subset S\). *We here mention a fact which we need in the proof of the next proposition. Let \(\lambda\) be a (dominant integral) weight for Lie algebra \(\gl_{n+1}\) such that \[\lambda_{j}\in \N\ \ \mbox{ for }1\le j\le n+1, \mbox{ and }\lambda_{1}\ge \lambda_{2}\ge \cdots \ge \lambda_{n+1}.\] Then the irreducible highest weight \(\gl_{n+1}\)-module \(L(\lambda)\) is finite-dimensional. Let \(\mu\) be the lowest weight of \(L(\lambda)\). We claim that \(\mu_{j}\in \N\) for \(1\le j\le n+1\). It was known (cf. ) that \(\mu=\sigma(\lambda)\) where \(\sigma\) is the longest Weyl group element. Suppose \(\gamma\) is any weight such that \(\gamma_{j}\in \N\) for \(1\le j\le n+1\). For \(1\le i\le n\), with the reflection \(r_{i}\), we have \[r_{i}(\gamma)=\gamma-\<\gamma,\alpha_{i}^{\vee}\>\alpha_{i} =\gamma-(\gamma_{i}-\gamma_{i+1})(\varepsilon_{i}-\varepsilon_{i+1}).\] Then \[\begin{aligned} r_{i}(\gamma)(E_{j,j})=\begin{cases}\gamma_{j} & \mbox{ if }j\ne i, i+1,\\ \gamma_{i+1}&\mbox{ if }j=i,\\ \gamma_{i}&\mbox{ if }j=i+1, \end{cases} \end{aligned}\] which implies \(r_{i}(\gamma)_{j}\in \N\) for \(1\le j\le n+1\). Then the claim follows from induction.* We have: Let \(S\) be a finite subset of \(\Z\) and let \(\lambda\in H^{*}\) be such that \({\rm supp}(\lambda)\subset S\). Then \(L(S,\lambda)\) is an integrable \(\gl_{\infty}\)-module if and only if \(\lambda\in P_{+}(S)\). With Proposition [\[pintegrability\]](#pintegrability){reference-type="ref" reference="pintegrability"} and Theorem [\[tmain-D\]](#tmain-D){reference-type="ref" reference="tmain-D"}, using a standard argument (see ) we obtain: Let \(S=\{r,r+1,\dots,n\}\) with \(r<n\) and let \(\lambda\in P_{+}(S)\). Then the maximal submodule of \(M(S,\lambda)\) is generated by \(E_{n+1,n}^{\lambda_{n}+1}v\), \(E_{r-1,r}^{\lambda_{r}+1}v\), and \(E_{i+1,i}^{\lambda_{i}-\lambda_{i+1}+1}v\) for \(r\le i\le n-1\), where \(v\) is a highest weight vector of weight \(\lambda\). To summarize we have: Let \(S=\{ r,r+1,\dots,n\}\) where \(r\) and \(s\) are integers with \(r<n\) and let \(\lambda\in H^{*}\) be such that \({\rm supp}(\lambda)\subset S\) and \(\lambda\in P_{+}(S)\). Then \(L(S,\lambda)\) belongs to category \({\cal{C}}_{int}\). On the other hand, every irreducible \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) is isomorphic to a module of this form. # Decomposition of tensor product modules in category \({\cal{C}}_{int}\) In this section, we construct the irreducible integrable \(\gl_{\infty}\)-modules \(L(S,\lambda)\) by using the natural module \(\C^{\infty}\), and we also determine the decomposition of tensor product modules in category \({\cal{C}}_{int}\). Let \(A\) denote the polynomial algebra \(\C[x_{m}\ |\ m\in \Z]\). Define \(\deg x_{m}=1\) for \(m\in \Z\) to make \(A\) a \(\Z\)-graded algebra \[A=\oplus _{r\ge 0}A_{r}.\] It was well known that Lie algebra \(\gl_{\infty}\) naturally acts on \(A\) with \[E_{m,n}=x_{m}\frac{\partial}{\partial x_{n}}\ \ \ \mbox{ for }m,n\in \Z.\] It can be readily seen that \(A\) is a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\). We see that \(A_{r}\) for \(r\ge 0\) are submodules with \(A_{0}=\C\) and \(A_{1}\simeq \C^{\infty}\). In fact, \(A\) is isomorphic to the symmetric algebra \(S(\C^{\infty})\) with \(v_{m}\) identified with \(x_{m}\) for \(m\in \Z\). One can show that for each \(r\ge 0\), \(A_{r}\) is an irreducible \(\gl_{\infty}\)-module. For each \(n\in \Z\), let \(\varepsilon_{n}\) be the linear functional on \(H\) defined by \[\varepsilon_{n}(E_{m,m})=\delta_{n,m}\ \ \ \mbox{ for }m\in \Z.\] Clearly, \(\varepsilon_{n}\) \((n\in \Z)\) are linearly independent. For \(i_{1},\dots,i_{r}\in \Z\), the \(H\)-weight of monomial \(x_{i_{1}}\cdots x_{i_{r}}\) is \(\varepsilon_{i_{1}}+\cdots +\varepsilon_{i_{r}}\). We see that every \(H\)-weight space of \(A_{r}\) is \(1\)-dimensional. Then any nonzero submodule of \(A_{r}\) must contain a monomial of degree \(r\). For \(j_{1},\dots,j_{k}\in \Z,\ n_{1},\dots,n_{k}\in \N\) with \(j_{1}<j_{2}<\cdots <j_{k}\) and \(n_{1}+\cdots +n_{k}=r\), we have \[(E_{j_{1},j_{2}})^{n_{2}}\cdots (E_{j_{1},j_{k}})^{n_{k}}\left(x_{j_{1}}^{n_{1}}\cdots x_{j_{k}}^{n_{k}}\right) =n_{2}!\cdots n_{k}!x_{j_{1}}^{r}.\] For any \(p,q\in \Z\) with \(p\ne q\), we have \[(E_{p,q})^{r}x_{q}^{r}=r!x_{p}^{r}.\] We also have \[\frac{1}{m_{1}!}\cdots \frac{1}{m_{s}!}(E_{i_{1},t})^{m_{1}}\cdots (E_{i_{s},t})^{m_{s}}\cdot x_{t}^{r} ={r\choose m_{1}}{r-m_{1}\choose m_{2}}\cdots {m_{s}\choose m_{s}}x_{i_{1}}^{m_{1}}\cdots x_{i_{s}}^{m_{s}}\] for \(i_{1},\dots,i_{s},t\in \Z,\ m_{1},\dots,m_{s}\in \N\) with \(i_{1}<i_{2}<\cdots <i_{s}\), \(m_{1}+\cdots +m_{s}=r\), and \(t\ne i_{1},\dots,i_{s}\). It then follows that \(A_{r}\) is an irreducible \(\gl_{\infty}\)-module. We see that \(A_{r}\simeq L(S,r\varepsilon_{1})\) with \(S=\{1\}\). On the other hand, \(\gl_{\infty}\) naturally acts on the exterior algebra \(\Lambda(\C^{\infty})\), which is also an \(\N\)-graded module \[\Lambda(\C^{\infty})=\bigoplus_{n\in \N}\Lambda^{n}.\] We have \(\Lambda^{0}=\C\) and \(\Lambda^{1}=\C^{\infty}\). For \(n\ge 2\), the submodule \(\Lambda^{n}\) has a basis consisting of vectors \[v_{i_{1}}\wedge\cdots \wedge v_{i_{n}}\] for \(i_{1},\dots,i_{n}\in \Z\) with \(i_{1}<i_{2}<\cdots <i_{n}\). One sees that the \(H\)-weight of vector \(v_{i_{1}}\wedge\cdots \wedge v_{i_{n}}\) is \(\varepsilon_{i_{1}}+\dots +\varepsilon_{i_{n}}\) of multiplicity one. Similarly, one can show that for every \(n\ge 0\), \(\Lambda^{n}\) is an irreducible \(\gl_{\infty}\)-module. For \(n\ge 1\), we have \(\Lambda^{n}\simeq L(S,\varepsilon_{1}+\dots+\varepsilon_{n})\) with \(S=\{ 1,2,\dots,n\}\). Note that for any permutation \(\sigma\) on \(\Z\), \(A_{m}\) and \(\Lambda^{n}\) are \(\sigma\)-invariant. *Let \(n\) be a positive integer and let \(\lambda\in H^{*}\) be such that \(\lambda_{m}=0\) for \(m\notin \{1,2,\dots,n\}\), \(\lambda_{i}\in \N\) for \(1\le i\le n\), and \[\lambda_{1}\ge \lambda_{2}\ge \cdots \ge \lambda_{n}.\] Note that \[\lambda=(\lambda_{1}-\lambda_{2})\varepsilon_{1}+(\lambda_{2}-\lambda_{3})(\varepsilon_{1}+\varepsilon_{2}) +\cdots +(\lambda_{n-1}-\lambda_{n})(\varepsilon_{1}+\cdots +\varepsilon_{n-1})+\lambda_{n}(\varepsilon_{1}+\varepsilon_{2}+\cdots +\varepsilon_{n}).\] For \(1\le k\le n\), set \[w_{k}= \Lambda^{k}.\] Furthermore, set \[w_{\lambda}=w_{1}^{\otimes (\lambda_{1}-\lambda_{2})}\otimes (w_{2})^{\otimes (\lambda_{2}-\lambda_{3})}\otimes\cdots \otimes (w_{n-1})^{\otimes (\lambda_{n-1}-\lambda_{n})} \otimes (w_{n})^{\otimes \lambda_{n}},\] which lies in the \(\gl_{\infty}\)-module \[(\Lambda^{1})^{\otimes (\lambda_{1}-\lambda_{2})}\otimes (\Lambda^{2})^{\otimes (\lambda_{2}-\lambda_{3})}\otimes\cdots \otimes (\Lambda^{n-1})^{\otimes (\lambda_{n-1}-\lambda_{n})} \otimes (\Lambda^{n})^{\otimes \lambda_{n}}.\] Set \(S=\{1,2,\dots,n\}\). It can be readily seen that \(w_{\lambda}\) is an \(S\)-singular vector of weight \(\lambda\). Then it follows that \(U(\gl_{\infty})w_{\lambda}\simeq L(S,\lambda)\).* For the rest of this section, we discuss the decomposition of tensor product modules in category \({\cal{C}}_{int}\). In view of Theorem [\[tmain-D\]](#tmain-D){reference-type="ref" reference="tmain-D"}, the tensor product of any two irreducible modules in category \({\cal{C}}_{int}\) is completely reducible. For example, we have \[A_{1}\otimes A_{1}=\C^{\infty}\otimes \C^{\infty}=S^{2}(\C^{\infty})\oplus \Lambda^{2}(\C^{\infty}) =A_{2}\oplus \Lambda^{2}.\] This typical example indicates that the tensor product of two irreducible modules in category \({\cal{C}}_{int}\) can be a finite sum of irreducible submodules. Next, we show that indeed this is the case. First, we establish a technical result. Let \(S\) be a finite subset of \(\Z\) and let \(\lambda\in P_{+}(S)\). Let \(v\in L(S,\lambda)_{\lambda}\) nonzero and set \(K_{S}^{+}={\rm span}\{ E_{p,i}\ |\ p>\max(S),\; i\in S\}\), an abelian subalgebra of \(\gl_{\infty}\). Then there exists a nonzero \(H\)-weight vector \(v'\in U(K_{S}^{+})v\subset L(S,\lambda)\) such that \[E_{p,i}v'=0\ \ \ \mbox{ for all }p\in \Z,\; i\in S.\] We shall also need the following simple fact: Let \(W\) be a \(\gl_{\infty}\)-module in category \({\cal{C}}_{int}\) and let \(S\) be a finite subset of \(\Z\). Suppose that \(U\) is a \(\gl_{S}\)-submodule of \(\Omega_{S}(W)\) such that \(U\) generates \(W\) as a \(\gl_{\infty}\)-module. If \(U=\coprod_{\alpha\in I}L(\lambda^{\alpha}_{S})\) as a \(\gl_{S}\)-module with \(\lambda^{\alpha}\in H^{*}\) such that \({\rm supp}(\lambda^{\alpha}) \subset S\) for \(\alpha\in I\), then \[W\simeq \coprod_{\alpha\in I} L(S,\lambda^{\alpha}).\] Furthermore, \(\Omega_{S}(W)=U\). Now, we give a decomposition into irreducible submodules of the tensor product of any two irreducible modules in category \({\cal{C}}_{int}\). Let \(S=\{ r,r+1,\dots,n\}\) be a finite subset of \(\Z\) with \(r\le n\), and let \(\lambda,\mu\in P_{+}(S)\). Then \(L(S,\lambda)\otimes L(S,\mu)\) is a cyclic \(\gl_{\infty}\)-module. Furthermore, there exists an integer \(k\) with \(k\ge n\) such that for any integer \(\bar{n}\ge k\), the decomposition of \(L(S,\lambda)\otimes L(S,\mu)\) into irreducible \(\gl_{\infty}\)-submodules agrees with the decomposition of \(L(\lambda_{\bar{S}})\otimes L(\mu_{\bar{S}})\) into irreducible \(\gl_{\bar{S}}\)-irreducible submodules where \(\bar{S}=\{r,r+1,\dots,\bar{n}\}\). *For an illustration, consider \(A_{m}\otimes A_{n}\) with \(m,n\) positive integers, where \[A_{m}=L(S,m\varepsilon_{1})\ \ \mbox{ and }\ A_{n}=L(S,n\varepsilon_{1})\] with \(S=\{ 1\}\). In this case, we can show that \(A_{m}\otimes A_{n}\) is cyclic on \(x_{1}^{m}\otimes x_{2}^{n}\). Set \(K={\rm span}\{E_{p,1}\ |\ p\ne 1\}\). Then \[A_{m}=U(\gl_{\infty})x_{1}^{m}=U(K)x_{1}^{m}\ \ \mbox{ and }\ \ K\cdot x_{2}^{n}=0.\] We have \[U(K)(x_{1}^{m}\otimes x_{2}^{n})=U(K)x_{1}^{m}\otimes x_{2}^{n}=A_{m}\otimes x_{2}^{n},\] from which it follows that \(U(\gl_{\infty})(x_{1}^{m}\otimes x_{2}^{n})=A_{m}\otimes A_{n}\). Set \(\bar{S}=\{1,2\}\). We have \(\gl_{\bar{S}}=\gl_{2}\), linearly spanned by \(E_{i,j}\) for \(1\le i,j\le 2\), and \[U(\gl_{\bar{S}})x_{1}^{m}=L(m\varepsilon_{1}), \ \ U(\gl_{\bar{S}})x_{2}^{n}=L(n\varepsilon_{1}).\] For \(\gl_{2}\)-modules, we have \[L(m\varepsilon_{1})\otimes L(n\varepsilon_{1})=\bigoplus_{j=0}^{|m-n|} L\left((m+n-j)\varepsilon_{1}+j\varepsilon_{2}\right).\] This gives rise to a decomposition of \(A_{m}\otimes A_{n}\) as a \(\gl_{\infty}\)-module \[A_{m}\otimes A_{n}=\bigoplus_{j=0}^{|m-n|} L\left(\bar{S},(m+n-j)\varepsilon_{1}+j\varepsilon_{2}\right).\]*

{'timestamp': '2013-10-14T02:11:14', 'yymm': '1310', 'arxiv_id': '1310.3251', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3251'}

# Introduction {#s:intro} In cold and dense regions of interstellar clouds where new stars are formed several molecules have been observed in the solid state, such as H\(_{2}\)O, CO, CO\(_{2}\), NH\(_{3}\), CH\(_{4}\), and CH\(_{3}\)OH (among others). Interstellar CO\(_{2}\) ice has been detected for the first time in 1989 by d'Hendecourt and de Muizon, confirmed by de Graauw \(et.~al.\), and it has been observed along many lines of sight with a constant CO\(_{2}\)/H\(_{2}\)O abundance ratio of about 0.3, with a high column density . The large amounts of CO\(_{2}\) ice observed cannot be explained by gas-phase reactions, and also because of the large amount of solid CO found in ice , it is believed that CO\(_{2}\) must be formed through solid phase reactions. However, its formation path is not at all understood yet, and in the literature different routes to form CO\(_{2}\) in ice have been proposed. Ground state CO can react with an electronically excited CO\(^{*}\) molecule leading to CO\(_{2}\) and atomic C. This reaction has been studied experimentally in different laboratories. CO\(_{2}\) can also be formed after irradiating CO ice with Ly\(\alpha\) photons, and subsequently bombarding the ice with energetic protons of 200 keV, or by irradiating the CO ice with 5 keV electrons in order to simulate the effect of cosmic ray particles, thereby breaking the CO bond and creating energetic atomic O. Another possible path to form ground state CO\(_{2}\) is by the reaction between CO with atomic O(\(^{3}\)P). Roser \(et~al.\) and Madzunkov \(et~al.\) demonstrated that this is a possible route to form CO\(_{2}\) using temperature programmed desorption (TPD) experiments, in one case using thermal O atoms below 160 K , and in another case using energetic O atoms (from 2 to 14 eV) in order to overcome the high reaction barrier. However, Grim \(et~al.\) demonstrated that the reaction CO(\(^{1}\Sigma\)) + O(\(^{3}\)P) \(\rightarrow\) CO\(_{2}\) does not take place on cold grain surfaces (10--20 K) due to its high activation energy (2970 K in the gas-phase ). A third mechanism is based on the reaction CO + OH \(\rightarrow\) CO\(_{2}\) + H, where the OH can result either from H\(_{2}\)O photodissociation (the photolysis mechanism) or thermal hydrogenation of oxygen species (the hydrogenation mechanism). Recent laboratory experiments show that both routes are efficient at low ice temperatures (10--20 K). In 2002, Watanabe \(et~al.\) measured the conversion rates of CO to CO\(_{2}\) irradiating the ice with photons with an energy close to Ly\(\alpha\), using H\(_{2}\)O and D\(_{2}\)O mixed with CO ices (with a ratio of H\(_{2}\)O/CO \(\approx\) 10) at 12 K and a photon flux of the order of 10\(^{14}\)--10\(^{15}\) photons s\(^{-1}\) cm\(^{-2}\). After two hours irradiation time, corresponding to a photon dose of about 10\(^{18}\) photons cm\(^{-2}\), most of the CO is converted to CO\(_{2}\) with a rate constant of about 2.5 \(\times\) 10\(^{-4}\) s\(^{-1}\) as obtained from fits of Fourier transform infrared spectroscopy (FTIR) measurements. They concluded that the rate constant is very small and suggested that CO is not immediately converted to CO\(_{2}\), probably because some intermediate like HOCO or DOCO is formed first. They did not detect HOCO, but in the end, over the long timescales of hours in the experiment, this may lead to CO\(_{2}\) and hydrogen or deuterium atom, respectively. In these experiments, no isotope effects were observed. In 2007, Watanabe \(et~al.\) also studied the formation of CO\(_{2}\) through the hydrogenation mechanism as well as the photolysis mechanism in two types of ices: a CO--H\(_{2}\)O mixture (with a ratio of H\(_{2}\)O/CO \(\approx\) 4), and a pure H\(_{2}\)O ice layer with a total thickness of about 30 monolayers with one monolayer of CO on top at low ice temperatures (10--50 K). The measured photon flux was 5.9 \(\times\) 10\(^{13}\) photons s\(^{-1}\) cm\(^{-2}\) with an energy in the range of the Lyman band and close to Ly\(\alpha\). After irradiating the ices for two hours, several products were identified in the spectra, such as CO\(_{2}\), which was the most abundant product, but also HCOOH, H\(_{2}\)CO, CH\(_{3}\)OH, as well as HCO and CH\(_{3}\)CHO, the presence of which was inferred from small peaks. The CO\(_{2}\) was assumed to be formed from the reaction of CO with OH, because CO cannot be dissociated at the UV photons energies used in these experiments. The experimentalists also fitted their data to a kinetic model in order to calculate the rate constants to form CO\(_{2}\), H\(_{2}\)CO, and HCOOH from CO, based on the column densities of these photoproducts and assuming that H\(_{2}\)O is constantly dissociated. According to their analysis, CO is converted to CO\(_{2}\) with a rate constant of 3.3 \(\pm\) 0.17 \(\times\) 10\(^{-4}\) s\(^{-1}\) at 10 K, in close agreement with the value reported previously by Watanabe \(et~al.\). However, after the bombardment of of the ice with H atoms, HCO was the most abundant product, and also H\(_{2}\)CO and CH\(_{3}\)OH, but no CO\(_{2}\) was detected in the infrared (IR) absorption spectrum. Another recent laboratory study on the formation of CO\(_{2}\) ice through surface reactions of CO molecules with OH radicals formed after H\(_{2}\)O photodissociation was done by Oba \(et~al.\). In these experiments, CO molecules were introduced in the main chamber, and also H\(_{2}\)O molecules that are dissociated by a microwave induced plasma. The experiments were carried out at 10 and 20 K for two hours. During the irradiation time the ratio OH/CO was about 0.8, and the products were recorded by infrared reflection absorption spectroscopy (RAIRS). From the IR spectra some lines were assigned to the following products: CO\(_{2}\), H\(_{2}\)CO\(_{3}\), and \(trans\)-HOCO and \(cis\)-HOCO, which disappear at higher temperatures. The experimentalists concluded that CO\(_{2}\) is formed by reaction of the OH radicals (that have low energies) with the CO matrix at low temperatures (below 20 K). However, it is expected that at higher temperatures the OH has enough energy to migrate far away from the CO, and the reaction may not be feasible. In a different kind of experiments, Ioppolo \(et~al.\) deposited a mixture of CO:O\(_{2}\) ice on a substrate at low temperatures (15 and 20 K) and bombarded the ice with a cold H atom beam, leading to the formation of thermal OH radicals, which after long timescales in the end reacted with CO to form CO\(_{2}\). The products formed in the ice were monitored by means of RAIRS. Noble \(et~al.\) studied the formation of CO\(_{2}\) ice in a nonporous H\(_{2}\)O ice and on an amorphous silicate surface through the reaction of CO with OH, where the OH radicals were formed after the hydrogenation of O\(_{2}\) and O\(_{3}\), and the formation of CO\(_{2}\) in ice was demonstrated by means of TPD experiments. Zins \(et~al.\) also studied the reactivity between CO and OH molecules at very low temperatures (3.5 K). The OH radicals were formed from discharged H\(_{2}\)O/He mixtures and they were mixed with a mixture of CO and CO\(_{2}\). A Fourier transform infrared spectrometer was used in order to monitor all the products that were formed in the ice, such as CO\(_{2}\), H\(_{2}\)CO, H\(_{2}\)O, and HO\(_{2}\), among others. The formation of the HOCO complex was also observed when the OH radicals and the CO molecules were injected at the same time, as an intermediate to the formation of CO\(_{2}\). Here the photo-induced pathway is studied, building on our previous molecular dynamics (MD) simulations of the photodissociation of water molecules in crystalline and amorphous water ice in order to better understand the formation of CO\(_{2}\) through the reaction between CO and OH from a fundamental molecular physics point of view. We also wanted to prove that CO\(_{2}\) can indeed be formed through this route even though the system has to pass through a deep well (the HOCO well) on the way to the exit channel barrier to reaction, so that the system could be trapped in the well if the energy of HOCO is efficiently dissipated to the surrounding ice. In our simulations, the OH + CO reaction is based on the gas-phase potential energy surface developed by Lakin \(et~al.\) that gives rise to low cross section values for the gas-phase reactions due to the barriers that the system has to overcome in order to eventually form CO\(_{2}\) and H as is described later. However, in our simulations the OH radical formed after H\(_{2}\)O photodissociation in many cases comes off vibrationally excited, and this might enhance the formation of CO\(_{2}\) compared to the gas-phase case. We report results for MD simulations on the photoinduced reaction of the photofragment OH radical with CO in different kinds of CO--H\(_{2}\)O ice conditions at 10 K, where a single H\(_{2}\)O molecule is photodissociated by a single photon. This low flux simulates the flux of UV photons that irradiate ice-coated grains deep inside interstellar clouds, which is of the order of 10\(^{3}\) photons cm\(^{-2}\) s\(^{-1}\), meaning one incident photon per month per grain, and therefore a photodissociation event will be finished by the time the next photon arrives. The methods employed in this study are explained in **Methods**, the main results are presented in **Results and discussion**, and the final conclusions are given in **Summary and Conclusion**. # Methods {#sec:methods} Classical MD methods and analytical potentials based on pair potential interactions have been used in order to describe the evolution of the interactions of all the molecules and fragments in a CO--H\(_{2}\)O ice system before and after the absorption of an UV photon by one of the H\(_{2}\)O molecules from the ice. All the specifics of the potentials and the switching functions are presented in the supporting material . ## Potentials {#ssec:Potentials} The total analytical potential energy surface (PES) for the CO--H\(_{2}\)O ice system used to describe the photodissociation of one of the water molecules and the subsequent interaction of the fragments with CO and the other H\(_{2}\)O molecules can be written as follows: \[\label{eq1} V_{\rm{tot}} =V_{\rm{ice}}+V_{\rm{H_{2}O^{*}-H_{2}O}}+V_{\rm{H_{2}O^{*}-CO}}+ V_{\rm{H_{2}O^{*}}}\]\ where \[\label{eq2} V_{\rm{ice}} =V_{\rm{H_{2}O-H_{2}O}}+ V_{\rm{H_{2}O-CO}}\] The first term of the total potential (Eq. [\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) is given by Eq. [\[eq2\]](#eq2){reference-type="ref" reference="eq2"}, which describes the intermolecular interactions between the H\(_{2}\)O molecules inside the ice excluding the H\(_{2}\)O molecule that is photoexcited (Eq. [\[eq2\]](#eq2){reference-type="ref" reference="eq2"}), which are described by the TIP4P potential. Eq. [\[eq2\]](#eq2){reference-type="ref" reference="eq2"} also contains the intermolecular interactions between those H\(_{2}\)O and the CO molecules, where all CO and H\(_{2}\)O molecules are kept rigid, and the CO--CO intermolecular interactions if more than one CO molecule is considered in the system. In this work, as expressed by Eq. [\[eq2\]](#eq2){reference-type="ref" reference="eq2"} only one CO molecule is taken into account, in order to limit the complexity of the total system. The V\(_{\rm{H_{2}O-CO}}\) potential is based on pair potentials between H\(_{2}\)O and CO molecules and consists of repulsion, dispersion and electrostatic terms based on CCSD(T) calculations using an AVDZ basis set and applying BSSE corrections for 2500 different configurations. The charges of the H\(_{2}\)O molecules are the same as those used in the TIP4P potential (H:0.52\(e\), O:0\(e\), and the additional charge site M:-1.04\(e\)). For the CO molecule we have used negative charges on C and on O (-0.47\(e\) and-0.615\(e\), respectively) and a compensating positive charge at the center of mass. The second term of the total potential (Eq. [\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) contains the intermolecular interactions of the photoexcited molecule, which is treated as fully flexible, with the rigid water molecules by means of a TIP3P-type potential. The photoexcited molecule will dissociate into H and OH. Thus, this term also takes into account the V\(_{\rm{H-H_{2}O}}\), V\(_{\rm{OH-H_{2}O}}\) interactions as fully described in our previous studies. The third term of Eq. [\[eq1\]](#eq1){reference-type="ref" reference="eq1"} covers the interactions between the photoexcited water molecule with CO, and the interactions between H with CO, and OH with CO. The V\(_{\rm{H_{2}O^{*}-CO}}\) potential is taken to be the same as for ground state H\(_{2}\)O interacting with CO because the lifetime of the excited water molecule is very short (\(\sim\) 0.2 fs ), leading to H and OH which will interact with CO by means of switching functions based on the OH bond distance as described elsewhere. For the V\(_{\rm{H-CO}}\) pair potential we fit an H--CO non reactive potential by using dispersion and repulsive terms (more details in Ref. ). We have used the well known LTSH potential, which is a six dimensional PES, to describe the V\(_{\rm{OH-CO}}\) interaction. To smoothly switch the intramolecular interaction of OH from the one valid when it interacts with H\(_{2}\)O to that valid when it interacts with CO, we have used a switching function based on the distance between the carbon atom with the oxygen atom of OH (more details can be found in the supporting material ). The minimum energy path of the OH + CO reaction according to the LTSH PES is plotted in Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}, showing a complex minimum energy path (MEP). In the entrance channel there is the OH--CO van der Waals minimum, followed by a small barrier (\(\sim\) 0.05 eV relative to the van der Waals minimum, the \(trans\) HO--CO saddle point) and the \(trans\)--HOCO well with an energy of about-1.3 eV relative to the gas-phase reactants. From the \(trans\)--HOCO minimum there are two possible paths that bring the system to the H + CO\(_{2}\) product. The first path (solid line in Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}) connects the \(trans\)--HOCO minimum to another minimum (of-1.2 eV), the \(cis\)--HOCO complex via a \(cis\)--\(trans\) barrier of \(\sim\) 0.4 eV relative to the \(trans\)--HOCO minimum. From the \(cis\)--HOCO minimum there is another saddle point (\(cis\)--H--OCO) with a much higher barrier of about 1.3 eV that connects the \(cis\)--HOCO complex with the final products. This exit channel barrier has a height of about 0.07 eV relative to the gas-phase reactants. The second MEP (dashed line in Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}) connects the \(trans\)--HOCO complex with another saddle point (HOCO--HCO\(_{2}\)) with a large barrier of about 1.6 eV that leads to the HCO\(_{2}\) minimum, and this barrier has a height of about 0.34 eV relative to the gas-phase reactants. Finally, the HCO\(_{2}\) minimum is connected with the final products through an H--CO\(_{2}\) stationary point that has an energy of 0.2 eV above the HCO\(_{2}\) minimum. Moreover, since the LTSH PES gives rise to two minima: HOCO and H + CO\(_{2}\), their interactions with the H\(_{2}\)O molecules must also be included in the total PES. The interaction between HOCO and H\(_{2}\)O has been taken simply as the sum of the OH--H\(_{2}\)O and the H\(_{2}\)O--CO interactions. However, we fit a new PES to describe the V\(_{\rm{H_{2}O-CO_{2}}}\) interactions, consisting of the dispersion, repulsive and electrostatic terms based on CCSD(T)/aug-cc-pVTZ calculations for 174 different configurations . The H--H\(_{2}\)O interactions were already described before. The last term of Eq. [\[eq1\]](#eq1){reference-type="ref" reference="eq1"} describes the intramolecular interactions in the water molecule that is photoexited to the first excited electronic state of gas-phase H\(_{2}\)O (\(\rm{\tilde{A}^{1}B_{1}}\)) by means of the Dobbyn and Knowles PES, which is repulsive and leads to dissociation into H and OH, as has already been explained elsewhere. ## Amorphous ice surface {#ssec:Amorphous} To study the UV photodissociation of H\(_{2}\)O ice in an amorphous CO--H\(_{2}\)O ice system followed by the reaction of OH + CO, we have used two different procedures to construct an amorphous CO--H\(_{2}\)O ice system. The first procedure of growing an ice surface is based on an 'hit and stick' method. First, a CO molecule is set up with center of mass coordinates fixed at the origin (\(x\), \(y\), \(z\)=0). A water molecule is generated with random position and orientation and it interacts with the CO molecule through the H\(_{2}\)O--CO PES described above. The energy of the H\(_{2}\)O--CO system is minimized by using the simplex algorithm. The CO--H\(_{2}\)O ice is grown by a consecutive addition of single H\(_{2}\)O molecules and minimizing the interaction energy which is described by the TIP4P PES for the newly arriving H\(_{2}\)O with the H\(_{2}\)O molecules already present in the ice, and by the H\(_{2}\)O--CO PES, using the simplex algorithm. After the amorphous ice is grown it is first thermalized, and next equilibrated at 10 K for 30 ps using MD with a thermostat switched on and off . The resulting ice physically looks like an ice ball made up of one CO and 50 H\(_{2}\)O molecules (Fig. [\[Figure3\]](#Figure3){reference-type="ref" reference="Figure3"}), and the closest H\(_{2}\)O molecule and the farthest one from the central CO molecule are located at a distance of 3.2 Å, and 7.2 Å, respectively. With this methodology, we have set up three different initial configurations of all the molecules in the ice ball, all of them thermalized at 10 K. The ice balls were used to explore the reactivity in the ice since they require little computer time. The second procedure is based on our previous amorphous water ice set up . Thus, first an hexagonal crystalline H\(_{2}\)O ice surface was modeled, and from its geometry an amorphous H\(_{2}\)O ice surface was obtained at 10 K using MD simulations and the fast-quenching method. The resulting ice can be divided in 16 monolayers (where the four bottom layers were not allowed to move during the dynamics in order to simulate the bulk), and in total the system contains 480 H\(_{2}\)O molecules with the system described by cell parameters of 22.4 Å, and 23.5 Å in \(x\) and \(y\), respectively. One water molecule was replaced by a CO molecule and its random coordinates and velocities were sampled according to a Maxwell-Boltzmann distribution of 10 K. We considered three different initial scenarios for the CO molecule: it was adsorbed on top of the first monolayer, it was absorbed in the second monolayer of the amorphous water ice surface, or it was absorbed in the fifth monolayer of the amorphous water ice surface. In the three cases, the total system was thermalized at 10 K for 20 ps using a thermostat, and later on it was equilibrated for 20 ps using microcanonical ensemble (NVE) MD simulations where the CO--H\(_{2}\)O and H\(_{2}\)O--H\(_{2}\)O interactions were taken into account. In Fig. [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}(a), the amorphous CO\(_{ad}\)--H\(_{2}\)O ice surface and in Fig. [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}(b), and (c), the amorphous CO\(_{ab}\)(ML2)--H\(_{2}\)O ice surface and CO\(_{ab}\)(ML5)--H\(_{2}\)O ice surface are represented, respectively. ## Initial conditions and dynamics {#ssec:Initial} For both types of CO--H\(_{2}\)O ice surfaces, only the H\(_{2}\)O molecules that are closest to the CO molecule (the distance between the center of mass of CO and the center of mass of water being smaller than 4.5 Å) were selected to be photodissociated, and between 200--2000 different initial coordinates and momenta were generated per excited H\(_{2}\)O molecule. As described elsewhere, in order to initialize the trajectories a Wigner phase-space distribution function fitted to the ground-state vibrational wavefunction of gas-phase water is used. The resulting coordinates and momenta were sampled using a Monte Carlo algorithm, and a vertical excitation is performed to put the system on the first electronically excited state, on the DK \(\rm{\tilde{A}^{1}B_{1}}\) PES . Thus, the molecules are excited with photon energies \(E_{\rm{exc}}\) in the range 7.5--9.5 eV, with a peak at 8.6 eV (see figure 3 in Ref. ). From the excitation energy and the dissociation energy of H\(_{2}\)O (\(E_{\rm{diss}}\rm{(H_{2}O)}\) \(\approx\) 5.4 eV ), we can estimate the initial energy of the photofragments, using that in the gas-phase (in the absence of the surrounding ice), the water photofragments (H and OH) have to obey momentum conservation (\(p_{\rm{H}}\)=\(-p_{\rm{OH}}\)) and energy conservation (\(E_{\rm{H}}\)+\(E_{\rm{OH}}\)=\(E_{\rm{H_{2}O}}\)=\(E\)). The initial available energy \(E\) can be calculated from \(E\)=\(E_{\rm{exc}}\[-\]E_{\rm{diss}}\rm{(H_{2}O)}\), which is in the range of 2.1--4.1 eV depending on the initial excitation energy. The initial energy that the fragments will have can then be estimated according to the following equation: \[\label{eq3} \frac{1}{2} m_{\rm{OH}} v^{2}_{\rm{OH}}=\frac{E}{(1 + \frac{m_{\rm{OH}}}{m_{\rm{H}}})}\]\ Therefore, OH radicals will be formed with approximately a maximum translational energy of E\(\rm_{OH}\)=\(E\)/18=0.2 eV, and approximately a maximum of vibrational energy E\(_{\rm{v}}\) of 2 eV . If the OH is formed with higher vibrational energy, the available total translational energy of OH is reduced by an 'equal' amount as follows: \[\label{EcolOHv} E\rm_{OH}(\upsilon)=E\rm_{OH}(\upsilon=0)-\frac{E_{\upsilon}-E_{\upsilon=0}}{18}\] During the photodissociation dynamics the CO molecule is treated in the same way as the water molecules that are not photodissociated, i.e., as a rigid rotor. Because the R\(_{\rm{C-O'}}\) bond is always fixed to the equilibrium distance of CO the system cannot follow the actual minimum energy path to H + CO\(_{2}\). However, the heights of the barriers change by less than 0.05 eV by fixing R\(_{\rm{C-O'}}\), so that this does not represent a too severe approximation. In line with this, due to its formation process OH may be vibrationally excited initially, and vibrational excitation of OH enhances the gas-phase OH + CO reaction more than vibrational excitation of CO. The initial momenta of CO are initialized from a Maxwell Boltzmann distribution of 10 K, and the whole system is equilibrated at the same temperature of 10 K. For each trajectory only one H\(_{2}\)O molecule is photodissociated and Newtons's equations of motion are integrated using a time step of 0.02 fs and a maximum time t\(_{\rm{max}}\) of 5 ps. The simulation stops if the CO\(_{2}\) molecule is formed, otherwise all the trajectories are run until t\(_{\rm{max}}\). The criterion to define the formation of the CO\(_{2}\) molecule after the interaction with OH (OH + CO' \(\rightarrow\) CO\(_{2}\) + H) is based on the intramolecular distances of HO--CO': \(R_{\rm{C-O}}\) \(\leq\) 1.3 Å, \(R_{\rm{C-O'}}\) \(\leq\) 1.3 Å, \(R_{\rm{O-O'}}\) \(\leq\) 3 Å, and the hydrogen atom originally forming OH is at a distance \(R_{\rm{O-H}}\) \(>\) 2 Å. The HOCO' complex is defined to be formed if the intramolecular distances are: \(R_{\rm{H-C}}\) \(\leq\) 2.3 Å, \(R_{\rm{H-O'}}\) \(\leq\) 3.3 Å, \(R_{\rm{C-O}}\) \(\leq\) 1.5 Å, \(R_{\rm{C-O'}}\) \(\leq\) 1.3 Å, and \(R_{\rm{O-O'}}\) \(\leq\) 3 Å. # Results and discussion {#sec:results} ## Amorphous ice ball {#ssec:Ball} We considered three different ice balls, and for each of the fourteen closest H\(_{2}\)O molecules to CO between 1000 and 2000 different initial configurations were sampled for each H\(_{2}\)O molecule in the three ice balls taking into account the excitation energies within the first UV absorption band of amorphous water ice. In our study, three possible outcomes have been observed. We calculated the formation probabilities and its standard errors (\(\epsilon\)=\(\sqrt{P\cdot(1-P)/N}\), where P is the probability and N is the total number of trajectories) for these channels for the three different ice balls, which lead to a total number of trajectories of \(\approx\) 48,000. The predominant outcome channel is the non reactive one with a probability of 0.970 \(\pm\) 0.001, and the second one the formation of the HOCO complex which stays trapped in the ice, with a probability of (2.98 \(\pm\) 1.07)\(\times\)`<!-- -->`{=html}10\(^{-2}\). The least probable channel is the formation of CO\(_{2}\) and a hydrogen atom that was always observed to desorb from the ice, with a probability of (3.6 \(\pm\) 0.9)\(\times\)`<!-- -->`{=html}10\(^{-4}\). The HCO van der Waals complex formation is also observed with an energy of about-18 meV, and a probability of (2.67 \(\pm\) 0.07)\(\times\)`<!-- -->`{=html}10\(^{-2}\), and this complex always dissociates to CO and H, which desorbs while CO remains trapped inside the ice. The HCO van der Waals complex is defined to be formed if the distance between the H atom formed after H\(_{2}\)O dissociation and the carbon atom from CO \(R_{\rm{H-C}}\) \(\leq\) 2.1 Å, and \(R_{\rm{C-O'}}\) is always \(<\) 1.3 Å because it is frozen. ## CO adsorbed and absorbed in amorphous water ice {#ssec:AWI} We have also studied the formation of CO\(_{2}\) and HOCO using a set up of amorphous ice similar to the one we have used in earlier MD studies of desorption of H, OH and H\(_{2}\)O from ice following photodissociation of an H\(_{2}\)O molecule in ice. In this Section we discuss an amorphous water ice surface set up at 10 K and a CO molecule that is either adsorbed on top of the first monolayer or located in the second or fifth monolayer (as we described above in the Methods Section). The closest H\(_{2}\)O molecules to the CO molecule were selected to be photoexcited and between 200--2000 different initial conditions were sampled per excited H\(_{2}\)O molecule. If the CO molecule was initially adsorbed on top of the first monolayer of an amorphous water ice (Fig. [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}(a)) no reaction was observed. The final outcome of the 1000 trajectories run in total was that CO stays adsorbed on top of the water ice surface, whereas the H atom always desorbs and the OH fragment either desorbs or stays trapped in the ice, as in our pure water photodissociation results. From the previous calculations on the ice ball system, we conclude that no more than 1000 trajectories were necessary in order to observe some reaction (HOCO that is a prerequisite for the CO\(_{2}\) formation). If the CO molecule was initially located in the second monolayer of the amorphous water ice surface (Fig. [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}(b)), 2000 trajectories were run in total and only one special event was observed to take place: the HCO van der Waals complex is formed --even though we have used a non reactive HCO PES--for a few femtoseconds (\(\sim\) 10 fs) after which the complex dissociates to H atom and CO molecule both absorbed in the ice at different locations, and the OH fragment remains trapped in the ice. This channel occurs with a probability of (3.15 \(\pm\) 0.39)\(\times\)`<!-- -->`{=html}10\(^{-2}\). Watanabe \(et~al.\) also observed HCO experimentally after irradiation of a mixed H\(_{2}\)O--CO ice with UV photons in the range of the Lyman band only for short irradiation times (\(\leq\) 10 min). Therefore, the photolysis of a mixed CO--H\(_{2}\)O ice seems to be another possible surface reaction route to HCO formation in interstellar ices. This can be investigated in the future with the use of a reactive H--CO potential. If the CO molecule was initially located in the fifth monolayer of the amorphous water ice surface (Fig. [\[Figure2\]](#Figure2){reference-type="ref" reference="Figure2"}(c)) three different outcomes were observed after running a total of 20,000 trajectories: the non reactive one with a probability of 0.970 \(\pm\) 0.001, the formation of the HOCO complex with a probability of (2.97 \(\pm\) 0.12)\(\times\)`<!-- -->`{=html}10\(^{-2}\), and also the formation of the CO\(_{2}\) molecule while the H atom desorbs from the ice surface with a probability of (3.5 \(\pm\) 1.3)\(\times\)`<!-- -->`{=html}10\(^{-4}\). Also in this case the transient formation of the HCO van der Waals complex has been observed, with a probability of (2.81 \(\pm\) 0.12)\(\times\)`<!-- -->`{=html}10\(^{-2}\). Therefore, these findings suggest that the initial location of the CO is very important to its subsequent chemistry. If the CO is at the surface or near the surface, a nearby OH photofragment formed upon H\(_{2}\)O photodissociation will usually be in the top two monolayers of the ice surface, and we know from previous MD simulations that these OH fragments tend to desorb to the gas-phase rather than stay trapped in the ice. However, if the OH fragment formed upon H\(_{2}\)O photodissociation was initially located more deeply in the ice it will stay trapped and it is more probable that it will find the right orientation towards the CO molecule to react. The CO molecule and the OH fragment must be located deeply in the water ice system to enable HOCO or CO\(_{2}\) formation from the reaction with an OH photofragment. ## HOCO and CO\(_{2}\) formation {#ssec:formation} The most important photoproduct formed in the ice after the reaction of the OH radical with the CO molecule for both types of ice is the HOCO complex, which loses its energy during the dynamics by dissipation to the surrounding H\(_{2}\)O molecules. In these cases, the HOCO complex is trapped in the \(trans\)-HOCO potential well (see Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}). After enough energy has been transferred to the surrounding ice at 10 K, it becomes nearly impossible to overcome the barriers and form CO\(_{2}\) and hydrogen, as demonstrated in Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}, where we have plotted the total HOCO energy as a function of the time for one particular trajectory. Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"} shows how fast (in less than 0.2 ps) the HOCO complex may lose its energy, after which it became trapped in the potential well with a total energy of about-1 eV in the case shown (which is close to the \(trans\)-HOCO minimum energy value, according to the LTSH PES, see also Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}). Goumans \(et~al.\) already proposed that the reaction between OH and CO may not lead to H + CO\(_{2}\) when the HOCO complex is formed, because its energy may be dissipated to the surrounding H\(_{2}\)O molecules, after which HOCO can react with atomic H to form CO\(_{2}\) + H\(_{2}\), or H\(_{2}\)O and CO, or HCOOH. The formation of the HOCO complex has been observed by Oba \(et~al.\) in their absorption spectra in both the \(trans\) and \(cis\) configurations through peaks at 1812 cm\(^{-1}\) and 1774  cm\(^{-1}\), respectively. These peaks were observed earlier at 1833 cm\(^{-1}\) and 1797 cm\(^{-1}\) by Milligan and Jacox during vacuum ultraviolet photolysis of H\(_{2}\)O in a CO matrix in the 200--300 nm range. Moreover, Watanabe \(et~al.\) assumed that the intermediate was formed even though they did not detect it, in order to explain the slow formation rate of CO\(_{2}\) ice from CO and OH. In another experiment, Noble \(et~al.\) measured the yield of \(^{13}\)CO\(_{2}\) to be only about 8 \(\%\) with respect to \(^{13}\)CO. This low yield could be explained assuming that the HOCO intermediate is formed as predicted by gas-phase theoretical and experimental studies, through the reaction of CO with OH. In an ice environment, the reactivity becomes even more complicated than in the gas-phase because it also depends on the initial orientation of the H\(_{2}\)O and CO reactants inside the ice before H\(_{2}\)O is photolysed, and on the resulting orientation of the OH photofragment relative to CO. Francisco \(et~al.\) studied the stability of the HOCO complex in the gas-phase, and they found that HOCO is stable, in agreement with the far infrared laser magnetic resonance experiment that confirms that HOCO is stable with a lifetime of about 10 ms. The HOCO complex has two conformers: \(trans\)-HOCO and \(cis\)-HOCO. The first one is about 0.1 eV more stable than the second one (Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}). The geometry of both conformers is rather similar: one of the C--O bonds is about 1.34--1.35 Å long, the other one is shorter at 1.18--1.19 Å, and the O--H distance is about 0.97 Å, with 107\(^{\circ}\)--108\(^{\circ}\), and 127\(^{\circ}\)--130\(^{\circ}\) HOC, and OCO bond angles, respectively. According to our simulations and taking into account the two types of ice, the probability to form the HOCO complex is (3.00 \(\pm\) 0.07)\(\times\)`<!-- -->`{=html}10\(^{-2}\). On average the HOCO complex that is observed trapped in the ices has the following geometry: one of the C--O bonds is about 1.37 Å long, and the other CO distance corresponding to the original CO molecule is always kept fixed, because in our simulations the CO is treated as a rigid rotor, at a bond length of 1.13 Å. The O--H distance is about 0.91 Å, the HOC angle is about 120\(^{\circ}\)--136\(^{\circ}\), this range being somewhat larger than in the gas-phase, the OCO bond angle is about 123\(^{\circ}\) on average, which is close to the gas-phase value, and the dihedral HO--CO angle is on average 175\(^{\circ}\). Therefore, the HOCO complex that is observed as trapped in the ice matrix has a geometry close to the \(trans\)-HOCO complex, and its energy (see Fig. [\[Figure4\]](#Figure4){reference-type="ref" reference="Figure4"}) agrees with the minimum of the \(trans\)-HOCO complex in the LTSH potential used (see Fig. [\[Figure1\]](#Figure1){reference-type="ref" reference="Figure1"}). The fixed CO bond distance in our simulations does not affect the minimum reaction barrier height relative to the gas-phase reactants (i.e., with this approximation the exit channel barrier height is only 0.02 eV higher). The total calculated probability to form CO\(_{2}\) + H for the two kinds of ice investigated is (3.6 \(\pm\) 0.7)\(\times\)`<!-- -->`{=html}10\(^{-4}\), which is very low and the reaction only occurs when the initial excitation energy of the photoexcited water molecule is on average (8.83 \(\pm\) 0.23) eV. At these energies usually vibrationally excited OH is produced on average at \(v\)=1 with a vibrational energy E\(_{\rm{v}}\) \(\approx\) 0.7 eV. We analyzed the initial configurations and orientations of the reactant molecules that lead to CO\(_{2}\) and the main conclusions are that the initial distance between the C atom from CO and the oxygen atom from the reacting OH fragment has to be within the range 3.1 Å \(\le\) R\(_{\rm.{CO}}\) \(\le\) 3.5 Å, and the dihedral angle between the OH bond and CO bond has to be in the range 41\(^{\circ}\)--178\(^{\circ}\) as is illustrated in Fig. [\[Figure6\]](#Figure6){reference-type="ref" reference="Figure6"} where the angle is 94.5\(^{\circ}\). During the dynamics OH approaches CO and orientates itself in a \(trans\) conformation relative to CO, and in less than 0.05 ps and without going through the \(cis\)-HOCO conformation the OH bond breaks and CO\(_{2}\) and H are formed. We compared the probability to form CO\(_{2}\) in the ice with gas-phase calculations performed using the VENUS96 code and using the LTSH PES implemented as described in Ref. . The collision energy of OH was set to 0.17 eV, which corresponds to the initial energy OH fragments would have on average in the ice if the water molecules were initially excited with an energy of 8.6 eV (corresponding to the peak of the first UV absorption band of amorphous water ice ), and the impact parameter range was between 0 and 2.4 Å. Fig. [\[Figure5\]](#Figure5){reference-type="ref" reference="Figure5"} presents the fitted reaction probability as a function of impact parameter for four values of \(\upsilon\). For each value of \(\upsilon\), the initial available total translational energy of OH is calculated according to Eq. [\[EcolOHv\]](#EcolOHv){reference-type="ref" reference="EcolOHv"} (e.g., for \(\upsilon\)=0, 1, 2, and 3, E\(\rm_{OH}\)(\(\upsilon\))=0.17, 0.15, 0.13, and 0.11 eV, respectively). From Fig. [\[Figure5\]](#Figure5){reference-type="ref" reference="Figure5"} it is clear that increasing the vibrational energy of the reactants increases the reaction probability as well even though the translational energy of OH is reduced, in agreement with very recent theoretical gas-phase calculations where a full-dimensional time-dependent wave packet study based on the LTSH potential showed how the reactivity is enhanced by initial OH vibrational excitation. The probabilities to form CO\(_{2}\) in the gas-phase for OH in \(\upsilon\)=0, 1, 2, and 3 are 0.01, 0.037, 0.063, and 0.082, respectively at \(b\)=0 Å (Fig. [\[Figure5\]](#Figure5){reference-type="ref" reference="Figure5"}). The most reactive trajectories are those that start with the OH radical more or less in line with the CO molecule, O of OH closes to C of CO. Therefore, in the gas-phase the probability to form CO\(_{2}\) and hydrogen from OH + CO is higher than in our calculations. But these probabilities are for impact parameter \(b\)=0 Å, and they decrease when the impact parameter increases up to a maximum impact parameter of about 2 Å (Fig. [\[Figure5\]](#Figure5){reference-type="ref" reference="Figure5"}), whereas in our simulations when water dissociates in ice into OH and H, the OH radical collides with CO at different impact parameters. In particular, the calculated impact parameter \(b\) for the reactive trajectories was within the range 1.27--4.4 Å, on average being (2.94 \(\pm\) 0.86) Å. Moreover, in the solid state the environment is also completely different than in the gas-phase because the OH and the CO molecules also interact with the other H\(_{2}\)O molecules in the ice and they may release their energy to the environment, which may also decrease the reactivity. We have analyzed our simulations for the two types of ice systems and found that initially the OH vibrational state distributions mostly correspond to \(\upsilon\)=0, 1, 2, and 3, using box quantization to assign vibrational states with inclusion of zero-point energy. The OH vibrational energy distribution agrees with the experimental gas-phase distribution for H\(_{2}\)O photodissociation at 7.9 eV (this value was chosen because the production of OH in ice is shifted by the same amount as the absorption spectrum). The experimental data show that when H\(_{2}\)O dissociates several vibrational states are assigned to the OH photoproduct with \(\upsilon\)=0, 1, 2, 3, and 4. ## Comparison with the experiments {#ssec:Exp} The comparison with the experiments that study the formation of CO\(_{2}\) ice through the reaction of CO with OH, where the OH comes from UV dissociation of H\(_{2}\)O ice, is not straightforward. The experiments only provide the rates to convert CO into CO\(_{2}\) based on the variation of the column densities as a function of irradiation time, and from these data a first-order rate equation law is fit. On the other hand, from our theoretical models we can provide the formation probabilities of CO\(_{2}\) and HOCO in ice per absorbed UV photon, which are (3.6 \(\pm\) 0.7)\(\times\)`<!-- -->`{=html}10\(^{-4}\), and (3.00 \(\pm\) 0.07)\(\times\)`<!-- -->`{=html}10\(^{-2}\) (Table [1](#table1){reference-type="ref" reference="table1"}), respectively, where it is assumed that, as in the gas-phase photoabsorption, each absorbed photon leads to dissociation into H and OH. In order to estimate the same probability from the experiments performed by Watanabe \(et~al.\), we need, besides the photon flux used in the experiments, the number of water and CO molecules in the sample, and the photon absorption cross section (\(\sigma\)) of amorphous ice in the Ly\(\alpha\) (10.21 eV) region. Unfortunately, \(\sigma\) is not accurately known. Ref.  claims it to be close to the absorption coefficient of liquid water and around 30\(\%\) of the gas-phase value. Accurate values of the gas-phase cross section are given in Ref.  at exactly Ly\(\alpha\) and in Ref.  for the whole VUV range. Here we use a gas-phase cross section, \(\sigma_{\rm{gas}}\), averaged over the 116--136 nm range used in the experiment, \(\sigma_{\rm{gas}}\)=7.1\(\times\)`<!-- -->`{=html}10\(^{-18}\) cm\(^{2}\). Let \(I\rm{_{0}}\) be the experimental photon flux per cm\(^{2}\) per second. Then the number of photons absorbed per second per cm\(^{2}\) follows from Beer's law as: \[\label{eq4} I_{0}-I=I_{0}(1-e^{-\sigma \cdot n \cdot l})\] where \(n\) is the number density (in cm\(^{-3}\)) of H\(_{2}\)O molecules in the sample and \(l\) the optical path length, which is twice the sample thickness because the light is reflected at the sample support. The number density of H\(_{2}\)O can be computed from the density \(d\) of amorphous ice (0.94 g cm\(^{-3}\)) taking into account that a 1 to 10 mixture of CO and H\(_{2}\)O contains 91\(\%\) H\(_{2}\)O, so that \(n\)=\(N_{\rm{av}}\) \(\cdot\) \(d\) \(\cdot\) 0.91/m\(_{\rm{H_{2}O}}\), where m\(_{\rm{H_{2}O}}\) is the molar mass of H\(_{2}\)O. Noting that the mixing ratio used is 10, we make the approximation that any H\(_{2}\)O molecule in the experimental sample that is photodissociated will be close to one CO molecule with which it can react, as in our model. Errors due to this approximation are not likely to exceed a factor 2. In Watanabe's experiment, the measured photon flux is \(I\rm{_{0}}\)=9.2\(\times\)`<!-- -->`{=html}10\(^{13}\) photons s\(^{-1}\) cm\(^{-2}\), and the estimated double path lenght is \(l\)=1.6\(\times\)`<!-- -->`{=html}10\(^{-6}\) cm. Thus, applying Eq. [\[eq4\]](#eq4){reference-type="ref" reference="eq4"} the probability that a UV photon will be absorbed in the experiments should be about 0.28, and per second the number of OH radicals formed in the experimental ice sample (\(l\)=1.6\(\times\)`<!-- -->`{=html}10\(^{-6}\) cm) per surface area of 1 cm\(^{2}\) should be 2.3\(\times\)`<!-- -->`{=html}10\(^{13}\) (\(N_{\rm{OH}}\)), which may react with CO molecules. The number of CO molecules (\(N_{\rm{CO}}\)) in a sample of 0.8\(\times\)`<!-- -->`{=html}10\(^{-6}\) cm is equal to \(N_{\rm{CO}}\)=0.1 \(\cdot\) \(n\) \(\cdot\) 0.8\(\times\)`<!-- -->`{=html}10\(^{-6}\) = 2.32\(\times\)`<!-- -->`{=html}10\(^{15}\). Watanabe \(et~al.\) measured the time variation of the integrated absorbance of CO molecules with respect to the initial amount of CO molecules (see figure 3 in Ref. ) and also the time variation of integrated absorbance of CO\(_{2}\) normalized to the initial integrated absorbance of CO (see figure 4 in Ref. ). From the fits of the two curve (figures 3 and 4 in Ref. ) they obtained rate constants equal to 6.5\(\times\)`<!-- -->`{=html}10\(^{-4}\) s\(^{-1}\), and 2.5\(\times\)`<!-- -->`{=html}10\(^{-4}\) s\(^{-1}\) for the time destruction of CO and formation of CO\(_{2}\), respectively. From the CO destruction rate 6.5\(\times\)`<!-- -->`{=html}10\(^{-4}\) s\(^{-1}\), we derive that per second 1.5\(\times\)`<!-- -->`{=html}10\(^{12}\) CO molecules are converted, either to form HOCO or CO\(_{2}\). Summarizing, in the experiment for each OH formed, 1.5\(\times\)`<!-- -->`{=html}10\(^{12}\)/2.3\(\times\)`<!-- -->`{=html}10\(^{13}\)= 0.07 molecule CO disappears. This number is significantly larger than the value obtained in the simulations (3.6 \(\pm\) 0.7)\(\times\)`<!-- -->`{=html}10\(^{-4}\). On the other hand, the experimentalists concluded that if CO converts immediately to CO\(_{2}\) both rate constants should be the same, but they are not. Therefore, in their experiments there must be some intermediate in the solid state before CO\(_{2}\) is formed. This intermediate could be the HOCO complex even though it was not observed in their FTIR measurements. In our simulations, the probability to destroy CO and form CO\(_{2}\) is much lower than the experimental value, if we assume that in the experiments all destroyed CO molecules correspond to CO\(_{2}\) formation (see Table [1](#table1){reference-type="ref" reference="table1"}). However, our results suggest that the HOCO complex is formed in the ice and that it is stabilized by energy dissipation. Therefore, if we compare the total calculated probability (Table [1](#table1){reference-type="ref" reference="table1"}) to form HOCO and CO\(_{2}\) with the experimental value assuming that all the CO that is destroyed results into HOCO or CO\(_{2}\) the agreement is rather good, even though there was no experimental evidence for large HOCO concentrations inside the ice in this particular experiment. But in a later experiment Watanabe \(et~al.\) found two peaks at 1808 and 1784 cm\(^{-1}\) that they decided not to investigate. However, we believe that these two peaks correspond to the \(trans\)-HOCO and \(cis\)-HOCO complex with frequencies of 1774 cm\(^{-1}\) and 1812 cm\(^{-1}\), respectively according to Oba \(et~al.\), and 1797 cm\(^{-1}\), and 1833 cm \(^{-1}\) (measured in a matrix), respectively, according to Milligan & Jacox. Thus, it is possible that these features were also present in the earlier experiments of Watanabe \(et~al.\) but not reported. ::: [\[table1\]]{#table1 label="table1"} A large difference between the theory and the experiments is that in the theory only one CO molecule and one nearby H\(_{2}\)O molecule excited by a single photon are considered, whereas in the laboratory conditions several OH radicals are produced that can interact with several CO molecules, and find the ideal orientation and react. Another important difference is that in our theoretical study we excite H\(_{2}\)O molecules to the first excited state (\(\rm{\tilde{A}^{1}B_{1}}\)) leading to H and OH, with the OH radical in its electronic ground state (\(X\)). However, in the experiments the ice is irradiated with photons with energies close to the Ly\(\alpha\) energy, which includes absorption into the \(\rm{\tilde{B}^{1}A_{1}}\) state and may lead to the formation of: OH(X) + H, OH(A) + H, O(\(^{3}\)P) + 2H, and O(\(^{1}\)D) + H\(_{2}\), with the corresponding experimental branching ratios of: 0.64, 0.14, 0.22, and 0, respectively, according to Mordaunt \(et~al.\). It is also known from gas-phase experiments and theoretical calculations that the OH(X) product has a vibrational state distribution mostly only populated at \(\upsilon\)=0, whereas the OH(A) product after H\(_{2}\)O photodissociation at Ly\(\alpha\) is formed not only in \(\upsilon\)=0, but also in \(\upsilon\)=1 and 2 in the gas-phase. Since the OH(X) radicals in \(\upsilon\)=0 are not so reactive (Fig. [\[Figure5\]](#Figure5){reference-type="ref" reference="Figure5"}), the OH radicals in the excited state (\(A\)) may well be responsible for the reaction with CO in the experiments. ## Improvements of the model {#ssec:Improve} In our previous studies we already discussed the most important approximations introduced in the study of molecular dynamics simulations of water ice photodissociation. Here, we discuss the ones that may effect our results regarding the formation of CO\(_{2}\) through the reaction of CO with OH formed upon H\(_{2}\)O photodissociation. We list them as follows: 1. The formation of CO\(_{2}\) ice has been studied only through the CO + OH(X) route. The photoexcitation of H\(_{2}\)O to the first electronic excited state only leads to OH(X) molecule in its ground state and H atom. In the experiments H\(_{2}\)O can also be photoexcited to the second excited electronic state and its dissociation can lead to the formation of OH(X) + H, OH(A) + H, O(\(^{3}\)P) + 2H, and O(\(^{1}\)D) + H. In principle, two additional routes may then lead to the formation of CO\(_{2}\): (1) CO + O(\(^{3}\)P) \(\rightarrow\) CO\(_{2}\), which can also take place when the O(\(^{3}\)P) is formed after the reaction of two OH radicals formed upon independent H\(_{2}\)O dissociation processes, (2) CO + OH(A) \(\rightarrow\) CO\(_{2}\) + H, which probably has a higher associated reaction probability than the CO + OH(X) reaction, as we already discussed in the previous section. Therefore, a more precise comparison with the experiments could be enabled by also modeling the excitation of H\(_{2}\)O to the \(\rm{\tilde{B}}\) state in our simulations. Also, experiments could be carried out in which H\(_{2}\)O ice is irradiated at lower excitation energies, so that only excitation to the \(\rm{\tilde{A}}\) state is possible, which would enable a better comparison to the present theoretical work. Another difference with the experiments is the CO/H\(_{2}\)O ratio being equal to 10 which is much larger than in our model. Therefore, in the experimental ice sample, the H\(_{2}\)O molecules can probably rotate rotate more easily and eventually find the best orientation to react with CO. On the other hand, our model is well suited to the description of interstellar chemistry in terms of photon fluxes (which are very low, meaning that the photodissociation by one incident photon is completely finished by the time the next photon arrives). But with the interstellar radiation field, excitation should also be possible to the \(\rm{\tilde{B}}\) state of H\(_{2}\)O. 2. The particular HOCO PES used in the model may introduce some errors in the reactivity. The HOCO PES that we considered is the LTSH PES. Gas-phase calculations computed for the HO + CO reaction by means of different PESs showed differences in the cross sections obtained, but only at collision energies E\(_{c}\) \(>\) 0.65 eV (15 kcal/mol). However, for the collision energies we are considering (E\(_{c}\) \(<\) 0.17 eV (3.9 kcal/mol)) the differences in the cross sections obtained with different PESs are almost negligible (see Fig.4(a) in Ref. ). 3. Not including quantum effects may be important at the low temperatures here considered. However, in the gas-phase calculations quantum effects are small, therefore we assume they will also be small in the solid state environment considered here. 4. In our model we have treated the interaction between H and CO through a non reactive HCO PES. Thus, no competition with HCO formation has been introduced, and in the future we are planning to also include a reactive HCO PES in order to quantify the probability of HCO ice formation in the interstellar medium. It should be noted that HCO formation would probably occur at the cost of the CO\(_{2}\) formation. # Summary and Conclusions {#sec:Conclusions} Molecular dynamics simulations have been performed for two different CO--H\(_{2}\)O ice systems in order to study the formation of CO\(_{2}\) in interstellar ices upon irradiation with UV photons. After the absorption of an UV photon H\(_{2}\)O dissociates into H and OH, and the OH radical can react with the CO present in the ice and form CO\(_{2}\). The first type of ice is based on the 'hit and stick' method with a single CO molecule in the center and 50 water molecules around it yielding an ice ball. Three different initial ice balls were set up at 10 K and several H\(_{2}\)O molecules were chosen to be photodissociated. We calculated the probabilities to form the HOCO complex and the CO\(_{2}\) molecule, and also the probability to form the transient HCO van der Waals complex. The second kind of ice was based on our previous MD simulations of amorphous water ice dissociation, and by using the same procedure an amorphous water ice at 10 K was set up with either a CO molecule adsorbed on the topmost ML or located in the second or fifth monolayer. Several initial conditions were sampled for the three cases, and no reaction was observed in the first case (CO\(_{\rm{ad}}\) adsorbed on the surface). In the second case (CO\(_{\rm{ab}}\) in ML 2) only the formation of the HCO van der Waals complex was registered (even though a non reactive HCO PES was used), and in the end this resulted in the H atom desorbing from the ice and the CO molecule being trapped inside the ice. However, if the CO was initially absorbed in ML 5, the HOCO complex and the CO\(_{2}\) molecule can be formed and the probabilities of these events were calculated. The formation of the HOCO complex is much more probable than the formation of CO\(_{2}\), because once the HOCO complex is formed it immediately starts losing its internal energy and it may become trapped inside the ice matrix. The HOCO complex has been detected experimentally by Oba \(et~al.\), and Watanabe \(et~al.\) also claimed that this complex must be formed as an intermediate that in the end may dissociate into CO\(_{2}\) and hydrogen atom. The comparison with the experiments is not straightforward mostly because the experimentalists irradiate the ice with a broadlamp peaking near Ly\(\alpha\), exciting the water molecules to the higher excited \(\rm{\tilde{B}}\) state. We have deduced from their data the number of photons that lead to conversion of CO into HOCO or further to CO\(_{2}\) after irradiating the CO--H\(_{2}\)O mixed ice with UV photons close to Ly\(\alpha\) energies. If we assume that all the CO is converted into CO\(_{2}\) in the experiments, the experimental probability is much higher than our probability to form CO\(_{2}\). However, if we assume that CO can be converted to HOCO and/or CO\(_{2}\), our total probability to form both molecules agrees reasonably well with their value. Overall, we can summarize the conclusions as follows: 1. The initial location of the CO molecule in the ice surface is very important as found from both types of ice simulated. This means that not only the environment (e.g., the number of nearest neighbor water molecules and their orientations relative to CO), but also the initial position of the CO molecule plays a big role. The CO must be located deep in the ice system. In conclusion, it is the initial position of the CO molecule and not the type of ice we modeled that appears to affect the reactivity of the OH photofragment with CO. 2. CO\(_{2}\) ice can be formed through the OH + CO reaction where the OH comes from H\(_{2}\)O photodissociation, if the water molecule has been excited with energies higher than 8.6 eV. However, the HOCO complex is the main product in our simulations. 3. HCO ice can probably also be formed through the H + CO reaction where the H comes from H\(_{2}\)O photodissociation. 4. Experiments regarding the OH + CO reaction upon excitation of H\(_{2}\)O in the first absorption band of amorphous water ice are necessary in order to confirm our theoretical results. If H\(_{2}\)O is excited in the second absorption band (as has been done in all existing experiments) other channels besides the OH(X) + H are possible after H\(_{2}\)O dissociation, and even the OH(X) formed will likely have a different reactivity with CO, due to a higher initial vibrational energy content. 5. A new theoretical model able to photoexcite H\(_{2}\)O to be \(\rm{\tilde{B}}\) state is needed in order to explore other possible routes to form CO\(_{2}\) ice. The authors would like to thank S. Andersson for providing the \(ab~initio\) data and the fitting of some of the potential energy surfaces employed in our model, and R. Valero for providing the implemented LTSH PES in the VENUS code. This project was funded with computer time by NCF/NWO, and by NWO astrochemistry grant No. 648.000.010. The parameters of the pair potentials and its analytical expressions, the analytical expressions of the switching funcions, and the operational definition of the outcome products (5 pages).

{'timestamp': '2013-10-14T02:09:18', 'yymm': '1310', 'arxiv_id': '1310.3189', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3189'}

# Introduction Atmospheric aerosols ranging in size from a few molecules to a 100\(\mu\)m and containing complex mixtures of soluble, insoluble, miscible and immiscible species from a variety of anthropological and natural sources, play a critical role in the microphysics of clouds. Soluble and insoluble particles can act as cloud condensation nuclei (CCN) by providing heterogeneous nucleation sites that lower the free energy barrier for droplet formation and ice nucleation. They also serve as reactants and catalysts in atmospheric chemical cycles. However, the ability of aerosol particles to initiate water uptake is dependent on both the chemical composition of the particle and its size.  Consequently, understanding how composition, the distribution of material inside a particle and particle size impacts its CCN affinity remain important questions. Köhler theory of activation  is based on the stability of the solutions to the Kelvin relation for a volatile solvent vapor in contact with a non-volatile drop and argues condensation begins at the vapor pressure where the drop becomes unstable with respect to particle growth. While nucleation theory provides a more detailed picture of condensation as a barrier crossing process, the difference in the predictions of the two approaches only becomes apparent for very small, non-volatile drops  and the simplicity of the activation theory is appealing in most practical cases. When the liquids in the drop are ideal, the Kelvin relation has a stable and unstable region in the coexisitence curve but Reiss and Koper found that the Kelvin relation involving a non-ideal liquid solution exhibits an additional unstable branch that is related to the mixing of the two components. Tanlanquer and Oxtoby used classical density functional theory (DFT) to explore the properties of miscible and partially miscible systems and showed that a solubility transition, between a drop with surface adsorbed solvent and a drop with solvent mixed into the core occurs before activation in the partially miscible case. Deliquescence, where solid particles of soluble salts absorb water directly from the surrounding vapor to form solution droplets, provides an alternative way for particles to grow in the atmosphere. Measurements of micro-sized levitated  particles showed deliquescence occurred at a well defined relative humidity corresponding to the vapor pressure where the activity of the water vapor equals the activity of the water in the bulk salt solution, but experiments on nanometer sized particles suggest surface effects can influence the transition for some systems.  Theoretical studies  show the properties of a surface film that partially dissolves the soluble core, and the nucleation barrier associated with the phase transition, play increasingly important roles in deliquescence as the particles become smaller, suggesting similar effects may become important for small droplets of partially miscible mixtures that can phase separate. In this paper, we use a combination of molecular dynamics (MD) simulations and thermodynamic theory to explore the kinetics and thermodynamics of the condensation of a vapor onto nanoscale, non-volatile liquid drops for systems made up of miscible and partially miscible solvent-solute mixtures. In particular, we are interested in observing features consistent with the presence of a solubility transition in nanoscale droplets. We use a binary mixture of Lennard-Jones particles in our MD simulations because this is a model where the volatility of the component in the drop and the energy of mixing for the two components are easily controlled by adjusting the well depth in the interaction potentials between particles. Measurements of the equilibrium drop size and density of the condensing solvent at the core of the droplet show that the miscible systems alway mix into the core of the drop, while still showing signs of surface enrichment of the volatile component. The partially miscible systems only begin to mix into the droplet core at an onset system volume, once there is a significant amount of vapor condensed onto the drop surface, suggesting the presence of a solubility transition. We also develop a simple capillarity based non-volatile liquid drop model that captures the free energy landscape for a transition between a small and large drop, characteristic of a solubility transition, in the partially miscible system. The paper is organized as follows: In Section II, we use molecular dynamics simulations to study the dynamics of how the vapor phase condenses onto the drop to form a film for both miscible and partially miscible mixtures. We also measure the equilibrium size and examine the structure of the resulting stable droplet. The non-volatile liquid drop model, based on simple capillarity-type approximations, is developed in Section III to examine the properties of solubility transition observed in partially miscible mixtures. The predictions of the model are compared to the results of our simulations and we also use the model to describe the correct normalization for free energy barriers used in classical nucleation involving nanoscale heterogeneities. Section IV contains our discussion and our conclusions are described in Section V. # Molecular Simulation Studies ## Methods We use MD simulations in the canonical ensemble to study the condensation of a vapor onto a liquid droplet. The composite system, vapor and droplet, is modelled using a binary mixture of Lennard-Jones particles interacting through the potential, \[U(r_{ij})=4\epsilon_{ij}\left[\left(\sigma_{ij}/r_{ij}\right)^{12}-\left(\sigma_{ij}/r_{ij}\right)^6\right]{ ,}\\ \label{eq:pot}\] where \(\epsilon_{ij}\) is the energy interaction parameter between species \(i\) and \(j\), \(\sigma_{ij}\) is the particle size interaction between species and the potential is cut, but not shifted, at half the box length. The simulation cell is cubic and we employ periodic boundary conditions. Denoting the volatile solvent and the non-volatile solute as components one and two respectively, we set \(\epsilon_{11}=1.0\), \(\epsilon_{22}=2.0\) and \(\sigma_{11}=\sigma_{22}=\sigma_{12}=1.0\), then vary \(\epsilon_{12}\) to control the miscibility of the components. The energy of mixing parameter, \[\Lambda^*=(\epsilon_{11}+\epsilon_{22}-2 \epsilon_{12})/\epsilon_{11}\mbox{ ,}\\ \label{eq:lam}\] provides a measure of the energetic drive force for mixing. When \(\Lambda^*<0\), particle interactions favor mixing, otherwise they promote phase separation and we study systems with \(\Lambda^*=-0.1\) and \(\Lambda^*=0.172\). The ratio \(\epsilon_{22}/ \epsilon_{11}\) controls the relative volatility of the components and we have chosen parameters consistent with the DFT model of Talanquer and Oxtoby  where the supersaturation of the non-volatile phase was \(10^5\) times smaller than that of the vapor. The number of volatile particles, \(N_1=300\), is maintained for all simulations and we study droplets with \(N_2=75,100\) and 150. Our simulations are carried out using the Gromacs package,  where the leap frog integration scheme, with a step size of \(\delta t^*=0.002\), is employed to evolve the equations of motion. The velocity rescaling thermostat is used to maintain the system at a reduced temperature, \(T^*=kT/\epsilon_{11}=0.8\), where \(k\) is the Boltzmann constant. All quantities are reported in reduced units. Cluster criteria that allow us to follow the evolution of the droplet as a function of time were developed by measuring the nearest neighbor distributions in both the pure volatile vapor system and in the isolated pure component two droplet phase. Particles of the vapor were initially placed randomly in the simulation cell, with the restriction that no two particles were closer than \(1.5\sigma_{11}\), then the system was equilibrated for \(10^6\) time steps before configurations were sampled every 10000 time steps for up to \(10^7\) time steps. Simulations of the isolated droplet were initialized using a cluster arranged in a compact, body-centered cubic structure that was allowed to equilibrate for \(10^7\) time steps. Configurations were then sampled in the same way as the vapor. Fig. [\[fig:nabs\]](#fig:nabs){reference-type="ref" reference="fig:nabs"} shows the fraction of particles with a given number of neighbors within \(1.5\sigma_{11}\) for both phases. The distribution of the droplet phase exhibits two distinct peaks that were decomposed into distributions associated with particles at the core of the droplet and those on the surface, using the cone  method to identify surface atoms. Most core particles have 12 neighbors, which is consistent with the nearest neighbor distribution of the bulk Lennard-Jones fluid with \(\epsilon_{ii}=1\) , but the core distribution measured here is narrower because our non-volatile component has a stronger interaction with \(\epsilon_{22}=2\). The distribution for the surface atoms peaks at seven nearest neighbors and marginally overlaps the distribution of the vapor phase, which has no particles with more than three neighbors. On the basis of these results, we identify liquid-like particles as those with three or more neighbors and consider two particles to be part of the same liquid cluster if they are neighbors. Throughout our simulations we follow three cluster based quantities: i) the size of the droplet, which is taken to be the largest cluster of liquid-like particles in the system and can contain both components, ii) the total number of component one particles in this largest cluster, and iii) the size and number of component one clusters that are part of the largest droplet. At each volume, \(V\), studied, we generate the starting configuration by initially equilibrating the isolated component two droplet in the container, then sequentially add the component one particles to the system in random locations, ensuring they are not closer than \(1.5\sigma_{11}\) to any other particle. Simulations were then run for \(10^7\) time steps. We also measure the equilibrium radial density distribution of each of the components in the droplet from the center of mass of the droplet. ## Results In the absence of the vapor phase, the pure component two droplets with \(N_2=100\) and 150 remain stable over the entire volume range studied and we only see the evaporation of five to six particles from the droplet at the largest volumes studied, confirming that the strong \(\epsilon_{22}\) interaction keeps the volatility of the droplet low. The \(N_2=75\) droplet begins to show significant evaporation above \(V/\sigma^3=3\times 10^5\) and these volumes are not included in our study. In the presence of vapor, the growth of the drop exhibits two distinct time trajectories, depending on the volume of the system. Figure [\[fig:trajsub\]](#fig:trajsub){reference-type="ref" reference="fig:trajsub"} shows the growth of the droplet, the growth of the largest component one cluster and the number of component one clusters on the drop as a function of time (Insert).When \(V\) is large, the size of the droplet only increases by a small amount and the vapor essentially remains stable. The growth occurs rapidly, then the droplet fluctuates around its equilibrium size, losing and gaining component one particles in a dynamic equilibrium with the vapor phase. The radial density distributions (Figures [\[fig:radial\]](#fig:radial){reference-type="ref" reference="fig:radial"}(a) and (c)) show that the component one particles are mainly located in the surface region of the droplet with a small amount of mixing into the core of the droplet when \(\Lambda^*=-0.1\). However, there are not enough particles to form a complete monolayer and we see an equilibrium number of component one liquid clusters distributed over the complete cluster surface. This is observed for both values of \(\Lambda^*\) studied. Clarke et al  also observed the submonolayer wetting of droplets in their study of the phase diagram for equimolar binary Lennard-Jones clusters for similar interaction parameters. The fluctuations in the size of the largest component one liquid clusters correlates with the fluctuations in the total size of the droplet, suggesting these clusters grow and shrink by gaining and losing particles to the vapor, although some coalescence between clusters on the droplet surface is expected to be occurring. When \(V\) is decreased, the vapor spontaneously condenses onto the droplet, causing it to grow. Figure [\[fig:trajmono\]](#fig:trajmono){reference-type="ref" reference="fig:trajmono"} shows that the initial condensation onto the droplet leads to a rapid increase in the number and size of component one liquid clusters on the surface of the droplet, but the limited surface area means that the clusters start to interact. In particular, we note that the large fluctuations in the size of the largest component one liquid cluster on the droplet are decoupled from the fluctuations in the total size of the droplet, indicating clusters on the surface are coalescing and breaking up again as the film grows. Eventually, the fluctuations decrease as all the clusters grow and coalesce to form a single cluster representing the completed film. The droplet is unable to grow indefinitely because the total number of component one particles in the simulation container remains fixed in the canonical ensemble and the vapor pressure necessarily decreases until a new equilibrium is established with the enlarged droplet. The value of \(\Lambda^*\) has a strong influence on the distribution of the components in the droplet with the radial density distributions (Figure [\[fig:radial\]](#fig:radial){reference-type="ref" reference="fig:radial"}) showing that lower values of the mixing parameter lead to greater mixing in the core of the droplet. However, even with \(\Lambda^* < 0\), which represents the point where mixing should be energetically favorable, we see a significant degree of surface enrichment of the volatile solvent at the drop-vapor interface. Figures [\[fig:film155\]](#fig:film155){reference-type="ref" reference="fig:film155"} and  [\[fig:film141\]](#fig:film141){reference-type="ref" reference="fig:film141"} show the number of component one particles contained in the droplet, \(n_1^d\), as a function of the volume of the system. The droplets formed from the miscible mixture generally grow larger than the partially miscible mixture, for drops of the same size, and the larger drops also grow more than the smaller ones, as we might expect. For all systems studied, \(n_1^d\) varies continuously over the full range of \(V\) studied. One of the key challenges we face in studying these systems is defining and identifying contributions to \(n_1^d\) that arise from adsorption of component one at the non-volatile droplet surface and the mixing of components in the drop. To make this distinction, we measure the extent of mixing into the core of the drop by plotting the density of component one at the core of the drop \(\rho_1(r=0)\), obtained from measuring the equilibrium density profiles at \(r=0\), as a function of \(V\) in Fig. [\[fig:cd\]](#fig:cd){reference-type="ref" reference="fig:cd"}. For \(\Lambda^*=-0.1\), \(\rho_1(r=0)>0\) at all \(V\) and varies continuously, indicating the components always mix to some degree. However, we see a change in the \(N_2\) dependence of \(\rho_1(r=0)\) compared to that observed for \(n_1^d\) in Figs. [\[fig:film155\]](#fig:film155){reference-type="ref" reference="fig:film155"} and [\[fig:film141\]](#fig:film141){reference-type="ref" reference="fig:film141"}. Larger non-volatile drops have a higher component one core density when \(V\) is large, but the core density of the smaller drops increases faster as \(V\) decreases, leading to an inversion of the \(\rho_1(r=0)\) dependence on \(N_2\), with small drops having greater core densities. The same trend is observed for cases with \(\Lambda^*=0.172\) at small system volumes, but at \(V/\sigma^3>1.5\times 10^5\), the core density goes to zero, which suggests that the \(n_1^d\) particles found in the drop at these volumes can be described as being surface adsorbed. The penetration of component one into the core of the particle, below a specific volume, occurs once more than a monolayer of material is condensed onto the drop and suggests the presence of a solubility transition as a function of volume of the system. # Non-volatile liquid drop model ## Model Details In this section, we develop the non-volatile liquid drop model, which is an extension of the liquid drop models used to study homogeneous nucleation , binary nucleation  and heterogeneous nucleation.  Preliminary results for this model were first published in Ref . The model consists of a canonical ensemble of \(N_1\) particles of the volatile species and \(N_2\) particles of the non-volatile species, contained in a fixed volume, \(V\), at a fixed temperature, \(T\). All \(N_2\) particles of the non-volatile species are contained in the spherical droplet phase, along with \(n_1^d\) particles of the volatile species. The remaining \(n_1^v=N_1-n_1^d\) particles of species one are in the vapor phase which is treated as an ideal gas. At constant \(N_1,N_2,V,T\), the Helmholtz free energy, \(F\), is the appropriate thermodynamic potential, with variations in \(F\) being given by \[dF=dU-TdS\mbox{,}\\ \label{eq:df0}\] where \(U=U^v+U^d\) is the internal energy, \(S=S^v+S^d\) is the entropy, and we have denoted quantities relating to the vapor phase and droplet phase with the superscripts \(v\) and \(d\), respectively. Variations in \(U\) are given by \[dU^v=TdS^v-p^vdV^v+\mu_1^vdn_1^v\mbox{,}\\ \label{eq:duv}\] \[dU^d=TdS^d-p^ddV^d+\mu_1^ddn_1^d+\gamma dA^d\mbox{,}\\ \label{eq:dud}\] where \(p^v\) and \(p^d\) are the pressures in the respective phases, \(\mu_1^v\) is the chemical potential of component one in the vapor phase, \(\mu_1^d\) is the chemical potential of component one in the drop phase, \(V^d=\nu_1n_1^d+\nu_2 N_2\) is the volume occupied by the drop and \(V^v=V-V^d\) is the volume accessible to the vapor. Here, \(\nu_i\) is the molecular volume of component \(i\) in the bulk liquid phase. We will assume that the drop-vapor interface is sharp, consistent with the capillarity approximation, so that the surface area of the drop is given by \(A^d=4\pi R^2\), where \(R\) is the radius of the drop, and \(\gamma\) is the bulk planar surface tension. It should be noted that Eq. [\[eq:dud\]](#eq:dud){reference-type="ref" reference="eq:dud"} does not contain any chemical potential terms for species two because the non-volatile component does not exchange particles with the vapor phase, yielding \(dn_2^d=0\). However, in principle, \(dU^d\) should include a term corresponding to the work required to transfer particles from the pure non-volatile drop to the mixed drop. Classical nucleation theory (CNT) generally assumes this term is large and negative but independent of the radius of the drop so that it is ignored as it does not affect the nucleation rate . As our derivation proceeds, we will see that neglecting this term in Eq. [\[eq:dud\]](#eq:dud){reference-type="ref" reference="eq:dud"} amounts to assuming that component two behaves ideally in solution, even if we have included non-ideal behavior for component one in the drop phase. Using the conservation conditions \(dV^v=-dV^d\) and \(dn_1^v=-dn_1^d\), along with the relation \(dA^d=2dV^d/R\), in Eqs. [\[eq:df0\]](#eq:df0){reference-type="ref" reference="eq:df0"}-[\[eq:dud\]](#eq:dud){reference-type="ref" reference="eq:dud"} yields \[dF=\left(p^v-p^d+\frac{2\gamma}{R}\right)dV^d+\left(\mu_1^d-\mu_1^v\right) dn_1^d\mbox{ ,}\\ \label{eq:df1}\] which gives the equilibrium conditions satisfying \(dF=0\) as \(\mu_1^d=\mu_1^v\) and \(p^d-p^v=2\gamma/R\). To obtain a more detailed model, we need to obtain relations describing the chemical potential terms. Integrating the Gibbs-Duhem relation for the vapor phase, \(S^vdT-V^vdp^v+n_1^vd\mu_1^v=0\), at constant \(T\) gives \[\mu_1^v(p^v)-\mu_1^{eq}(p_1^{eq})=kT\ln\frac{p^v}{p_1^{eq}}\mbox{,}\\ \label{eq:muideal}\] where we have used the coexistence pressure of the vapor in contact with the pure fluid via a planar interface, \(p_1^{eq}\), as the reference state. The Gibbs-Duhem relation for the drop and its associated surface is \(S_1^ddT-V^d dp^d+n_1^d d\mu_1^d+A^d d\gamma=0\). Assuming that the drop is incompressible, and that the surface tension is independent of both the pressure and the composition of the drop, yields \[\mu_1^d(p^d,T,x_1)-\mu_1^d(p_1^{eq},T,x_1)=\nu_1(p^d-p_1^{eq})\mbox{,}\\ \label{eq:mu1a}\] where \(x_1=n_1^d/(n_1^d+N_2)\) is the mole fraction of component one in the drop. One way to capture the non-ideal nature of the system studied in our simulations is to treat component one in the drop as a regular solution,  which is based on the Bragg-Williams lattice approximation. This assumes that the drop mixes uniformly so the entropy of mixing is the same as the ideal solution, but the enthalpy of mixing is dependent on \(x_1\). The chemical potential of component one at \(p_1^{eq},T\) and \(x_1\) is then described by \[\mu_1^d(p_1^{eq},T,x_1)=\mu^0+kT\ln x_1 + b_0(1-x_1)^2\mbox{,}\\ \label{eq:mureal}\] where \(\mu^0=\mu_1^d(p_1^{eq},T,x_1=1)\) is the chemical potential of pure component one liquid at \(p_1^{eq}\) and \(b_0\) accounts for the interaction between components. When \(b_0\) is set to zero, Eq. [\[eq:mureal\]](#eq:mureal){reference-type="ref" reference="eq:mureal"} reduces to the expression for an ideal solution, while positive and negative values correspond to repulsive and attractive interactions respectively. Using Eqs. [\[eq:muideal\]](#eq:muideal){reference-type="ref" reference="eq:muideal"}-[\[eq:mureal\]](#eq:mureal){reference-type="ref" reference="eq:mureal"} in Eq. [\[eq:df1\]](#eq:df1){reference-type="ref" reference="eq:df1"} and noting \(dV^d=\nu_1 dn_1^d\) yields \[dF=\left[\nu_1(p^v-p_1^{eq})-kT\ln\frac{p^v}{x_1 p_1^{eq}}+\frac{2\nu_1\gamma}{R}+b_0(1-x_1)^2\right]dn_1^d\mbox{.}\\ \label{eq:dfdn}\] Equating the term inside the brackets of Eq. [\[eq:dfdn\]](#eq:dfdn){reference-type="ref" reference="eq:dfdn"} to zero then gives us the Kelvin relation for the binary drop, \[\frac{p^v}{p_1^{eq}}=a_1\exp \left[\frac{2\nu_1\gamma}{kTR}\right]\mbox{,}\\ \label{eq:kelvin}\] where \(a_1=x_1\exp[b_0(1-x_1)^2/kT]\) is the activity of component one in the drop and the first term has been ignored because it is generally small. Reiss and Koper also obtained Eq. [\[eq:kelvin\]](#eq:kelvin){reference-type="ref" reference="eq:kelvin"} using a different approach in an open system. Finally, Eq. [\[eq:dfdn\]](#eq:dfdn){reference-type="ref" reference="eq:dfdn"} is integrated with respect to \(n_1^d\) to obtain the free energy of forming the drop, \[\begin{array}{lll} \Delta F&=&F(n_1^d)-F(0)\nonumber\\ &=& n_1^d\left[ kT-\nu_1^d p_1^{eq}+b_0(1-x_1)\right]+N_2 kT\ln (1-x_1)\nonumber\\ &+&N_1 kT\ln \frac{p^v}{p_1^0}-n_1^d kT\ln\frac{p_1^v}{x_1 P_1^{eq}}+\gamma[A^d(n_1^d)-A^d(0)]\nonumber \mbox{,} \label{eq:fefinal} \end{array}\] where \(p_1^0=N_1kT/(V-V^d)\) is the pressure of the vapor before any particles have condensed onto the drop and the second term of the right hand side of the equation is the entropy of mixing for component two. ## Results We begin this section by exploring the general features of the free energy surface described by Eq. [\[eq:fefinal\]](#eq:fefinal){reference-type="ref" reference="eq:fefinal"} as a function \(V\) and \(n_1^d\), using the thermodynamic parameters for argon  where appropriate and assuming \(\nu_1=\nu_2\). We also set \(N_1=300\) to be consistent with our simulations. When \(b_0\) is small or negative, the entropy of mixing for the two components dominates the free energy and we find there is a single free energy minimum, corresponding to the spontaneous formation of a mixed drop, for all system volumes (Figure [\[fig:feld\]](#fig:feld){reference-type="ref" reference="fig:feld"}(a)). As \(V\) decreases, the minimum simply moves to larger \(n_1^d\) as more vapor condenses and the drop grows. Figures [\[fig:feld\]](#fig:feld){reference-type="ref" reference="fig:feld"}(b)-(e) show that the evolution of the free energy surface as a function of volume, for a system with larger \(b_0\), traces out a hysteresis loop similar to the one observed in the deliquescence and efflorescence of soluble salt particles. At large \(V\), we see the spontaneous absorption of a few component one particles to form a small mixed drop. As \(V\) is decreased, a minimum appears at larger \(n_1^d\), corresponding to a drop that has absorbed a significant fraction of the volatile solvent component, dissolving the non-volatile solute. Initially the minimum is metastable relative to the small drop, but it eventually becomes the most stable state. The free energy minima for the two drops is separated by a free energy maximum that is an unstable equilibrium solution to the Kelvin equation and the size of the drop at the maximum is the critical nucleus for the solubility transition. At smaller \(V\), we reach the limit of stability of the non-volatile drop, which then dissolves spontaneously. This point represents the activated drop described in Köhler theory. As the initial size of the non-volatile drop decreases, the minimum in the free energy for the small drop moves to smaller \(n_1^d\) and becomes shallower. In the limit of \(N_2\rightarrow 0\) Eq. [\[eq:fefinal\]](#eq:fefinal){reference-type="ref" reference="eq:fefinal"}, reduces to the free energy expression for the original homogeneous liquid drop model . For large \(N_2\), the free energy surface is once again characterized by a single broad minimum (Figure [\[fig:feld\]](#fig:feld){reference-type="ref" reference="fig:feld"}(f)) as the entropy of mixing again becomes the most dominant term in the free energy. To study the effect of the size of the non-volatile droplet on the location of the transition between small and large drops, we assume, as in the case of Köhler activation theory, that the transition occurs at the point the small drop becomes unstable and satisfies the conditions \(\partial F/\partial n_1^d=0\), and \(\partial^2 F/\partial n_1^{d2}=0\). Figure [\[fig:stability\]](#fig:stability){reference-type="ref" reference="fig:stability"}(a) shows that the pressure of the vapor surrounding the unstable drop, relative to the bulk equilibrium vapor pressure, as given by Eq. [\[eq:kelvin\]](#eq:kelvin){reference-type="ref" reference="eq:kelvin"}, increases with the decreasing \(N_2\). We also see the that the growth factor of the drop, \(G_R=R/R_0\) where \(R_0\) is the size of the pure non-volatile drop, decreases with decreasing \(N_2\) (Figure [\[fig:stability\]](#fig:stability){reference-type="ref" reference="fig:stability"}(b)). The results of the model are mildly dependent on \(N_1\) because the vapor phase becomes depleted as the droplet grows, but these effects decrease as the system size increases. Figures [\[fig:film155\]](#fig:film155){reference-type="ref" reference="fig:film155"} and [\[fig:film141\]](#fig:film141){reference-type="ref" reference="fig:film141"} show fits of the liquid drop model to our simulation data, where we have used \(b_0\) and \(\gamma\) as fit parameters, while fixing the remaining parameters. The model fits the data well for both values of \(\Lambda^*\), and all non-volatile drop sizes, over the entire range of volumes studied. The values of \(\gamma^*\) range from 1.59--1.80 and 1.92--2.36 for \(\Lambda^*=-0.1\) and \(0.174\) respectively. These values are higher than the surface tension for the pure fluid of component one and are physically reasonable, but we would expect the true surface tension of the drop to be dependent on the mole fraction of the components and this effect has been ignored in our model. We also find that the fit values for \(b_0\) are negative. This reflects the fact that the model assumes all \(n_1^d\) condensed particles are uniformly distributed in the drop rather than having some partitioned to the surface, so it over--estimates the degree to which the particles like to mix. A key feature of the models is that it predicts the transition between small and large drops should be accompanied by a discontinuous increase in \(n_1^d\) as a function of \(V\). The simulation trajectories show that the thin films form spontaneously and there is no clear sign of nucleation like behavior or of the expected discontinuity in the equilibrium droplet size, but these may be obscured to some degree by surface absorption. We also compare the model predictions for the core density using \(\rho_1(r=0)=\rho_1=n_1^d/(\nu_1n_1^d+\nu_2N_2)\) because it is assumed the components are uniformly mixed. Figure [\[fig:cd\]](#fig:cd){reference-type="ref" reference="fig:cd"} shows that the model, assuming ideal mixing (\(b_0=0\)) correctly predicts the dependency of \(\rho_1(r=0)\) on \(N_2\), including the inversion of the trend as a function of \(V\). Fitting the model to the data using \(b_0\) and \(\gamma\) as fit parameters yield excellent looking curves, but the two fit parameters become highly anti-correlated and particle size dependent. These fits have not been included here as the values of the parameters they yield appear unphysical. ## Nucleation Köhler activation theory, as used in the last section, successfully describes the location of the transition in large drops but nucleation becomes an increasingly important mechanism for droplet growth as particle size decreases  and we would expect the transition to occur at lower vapor pressures because droplets can get over the barrier in an activated process before the limit of stability is reached. In this section, we will use the non-volatile liquid drop model to examine the main features of nucleation in systems that exhibit a free energy minimum along the reaction coordinate and the effect of particle size on the nucleation barrier. The nucleation process of interest is the dissolution of the small drop, which occurs when the drop absorbs enough of the solvent phase to grow larger than the critical drop size, \(n^*\), located at the maximum in the free energy curve. Figure [\[fig:nminnmax\]](#fig:nminnmax){reference-type="ref" reference="fig:nminnmax"} plots the size of the critical droplet and the size of the small droplet at the minimum over the range of volumes where the small drop is metastable. As the system volume decreases, \(n^*\) decreases and \(n_{min}\) increases until they converge at the limit of stability, which represents the activation point. These plots show that the range of \(V\) over which the small mixed drop is metastable increases for decreasing \(N_2\), which helps explain why nucleation becomes more important as the non-volatile particle gets smaller. When \(N_2\) is large, a small change in \(V\) will rapidly move the system beyond the limit of stability, while the smaller particles require much larger changes in \(V\). Classical nucleation theory (CNT), gives the rate of drops going over the barrier as \[J=A\exp(-\Delta F^*/kT)\mbox{,}\\ \label{eq:ratecnt}\] where \(\Delta F^*/kT\) is the height of the free energy barrier and \(A\) is a pre-exponential factor containing information about the dynamics. When the free energy surface contains a local minimum, the free energy barrier used in Eq. [\[eq:ratecnt\]](#eq:ratecnt){reference-type="ref" reference="eq:ratecnt"} is usually defined as the difference in free energy between the maximum and the minimum,  \(\Delta F_{mm}=\Delta F(n^*)-\Delta F(n_{min})\). However, Scheifele et al.  showed that \(\Delta F_{mm}/kT\) did not correctly predict the rate for the heterogeneous nucleation of the two dimensional Ising model onto a nanoscale impurity and highlighted the fact that the exponential term in the rate expression is really a surrogate for the probability of finding the metastable drop at the transition state, \(P(n^*)\). Our model exhibits a free energy minimum and should be extensive in the number of non-volatile particles, so we would anticipate that the same analysis is needed here to correctly predict the rate. The free energy that provides the probability of finding a drop containing \(n_1^d\) particles can be expressed \[\Delta F_0(n_1^d)/kT=-\ln P(n_1^d)=-\ln\frac{Q(n_1^d)}{Q(\mbox{met})}\mbox{,}\\ \label{eq:fnorm}\] where \(Q(n_1^d)\) is the partition function of the drop with \(n_1^d\) and \(Q(\mbox{met})\) is the partition function of the metastable system, \[Q(\mbox{met})=\sum_{n_1^d=0}^{n_1^d=n^*}Q(n_1^d)\mbox{.}\\ \label{eq:qmet}\] The free energy that should appear in the rate expression is then given by \(\Delta F_0(n^*)/kT\) and represents the work required to constrain the metastable droplet to its critical size. In the context of the thermodynamic, capillarity based model developed here, \(\Delta F_0(n_1^d)/kT\) can be calculated by renormalizing the free energy given by Eq. [\[eq:fefinal\]](#eq:fefinal){reference-type="ref" reference="eq:fefinal"} so that, \[P(n_1^d)=\frac{\exp(-\Delta F(n_1^d)/kT)}{\sum_{n_1^d=0}^{n_1^d=n^*}\exp(-\Delta F(n_1^d)/kT)}\mbox{,}\\ \label{eq:pnorm}\] which ensures \(\sum_{n_1^d=0}^{n_1^d=n^*}P(n_1^d)=1\). \(\Delta F_0(n_1^d)/kT\) can then be obtained from the left hand equality in Eq. [\[eq:fnorm\]](#eq:fnorm){reference-type="ref" reference="eq:fnorm"}. Implicit in Eq. [\[eq:pnorm\]](#eq:pnorm){reference-type="ref" reference="eq:pnorm"} is the assumption that the capillarity model, which gives rise to \(\Delta F(n_1^d)\), describes all the microscopic states of the partition function, \(Q(n_1^d)\), and that \(n_1^d\) serves as an order parameter that can be used to rigorously sum over all the possible states. Figure [\[fig:dfcomp\]](#fig:dfcomp){reference-type="ref" reference="fig:dfcomp"} shows that the effect of the renormalization is to shift the free energy curve vertically, without changing its shape, because the demoninator in Eq. [\[eq:pnorm\]](#eq:pnorm){reference-type="ref" reference="eq:pnorm"} results in a constant term that is applied to the free energy of all the states. \(\Delta F_0(n_1^d)/kT)\) is positive for all the drop sizes in the metastable region since it always takes work to constrain the system to the subset of states, but the values for \(n_1^d > n^*\) have no real meaning because they are connected to the stable state and have not been included in the metastable drop partition function. A comparison of the nucleation free energy barriers predicted by \(\Delta F_{mm}\) and Eqs. [\[eq:fnorm\]](#eq:fnorm){reference-type="ref" reference="eq:fnorm"}--[\[eq:pnorm\]](#eq:pnorm){reference-type="ref" reference="eq:pnorm"} (see Figure [\[fig:nucbar\]](#fig:nucbar){reference-type="ref" reference="fig:nucbar"}) shows that \(\Delta F_0^*/kT\) is generally \(2-3\mbox{ } kT\) higher than the minimum to maximum barrier which can lead to orders of magnitudes of difference in the nucleation rates predicted by the two definitions. \(\Delta F_0^*/kT\) also remains finite at the limit of stability. This seems counter intuitive because we associate a positive free energy barrier with an activated process, but at the limit of stability the free energy landscape is monotonically decreasing from \(n_1^d=0\) so the drop grows spontaneously. This remains true, even after renormalization. However, in reality, the metastable region ceases to be defined at the limit of stability and the barrier is only defined as the limit of stability is approached from above. The finite barrier is then a result of the fact that \(n^*\rightarrow n_n\) at a finite value of \(n_1^d\), because of the minimum in the free energy. Despite this unusual property, Scheifele et. al.  showed that the free energy given by Eq [\[eq:fnorm\]](#eq:fnorm){reference-type="ref" reference="eq:fnorm"} correctly predicts the rate of nucleation right down to the limit of stability for their heterogeneous nucleation case. We do not have any independent rate data to test in our current model but our analysis provides a useful example showing how the renormalization can be implemented for a capillarity--based model. # Discussion In this work, we have performed a series of molecular dynamics simulations to study the condensation of a vapor onto a non-volatile drop for both miscible and partially miscible binary Lennard-Jones mixtures. In the canonical ensemble, the drop grows spontaneously as the vapor condenses, but eventually it comes to equilibrium as the vapor phase is depleted. When \(V\) is large, a submonolayer amount of the vapor is adsorbed onto the drop with the particles being distributed over the droplet surface in small clusters. Some mixing into the core does occur for the miscible mixtures but no mixing is observed for the partially miscible systems. When \(V\) is small enough to cause a substantial number of vapor particles to condense, we see a film growth mechanism that is dominated by cluster-cluster coalescence due to the restricted surface area available on the nanoscale sized drop. This is likely to be a common feature of nucleation and growth mechanisms in nanoscale systems. It is in contrast to the usual mechanisms observed during film formation on macroscopic surfaces that usually occur through the addition and loss of individual particles to and from isolated clusters. Once a monolayer is formed, we also begin to see mixing into the core for the droplet of the partially miscible system, which is a sign that the droplet core has started to dissolve. We also developed the non-volatile liquid drop model, combined with elements of the regular solution theory, to described the general features of the free energy surface associated with droplet growth in nanoscale systems. The free energy landscape of the model for partially miscible components exhibits a hysteresis loop similar to that observed in deliquescence and efflorescence of small soluble salt particles, caused by the presence of a nucleation barrier between the small drop and large dissolved drop phases. This transition resembles elements of the solubility transition described by Talanquer and Oxtoby  using DFT. However, the DFT model directly includes effects due to surface adsorption where our model ignores this feature, even though our simulations show that these are important. A number of capillarity based models have been used to study the effects of surface films on the deliquescence of small particles. For example, introducing the disjoining pressure  to account for the interaction between vapor--film and film--solid interfaces is able to stabilize the solution film which would otherwise spontaneously dissolve the salt particle. A thin layer criterion was developed by McGraw and Lewis  that requires the equality of the chemical potentials between the solvent component in the vapor and the film, as well as the equality of the chemical potentials between the salt in the film and the solid particle. Our simulations clearly highlight the importance of surface films in the cases of miscible and partially miscible systems and it may be useful to develop these approaches in the context of the solubility transition in small systems. Finally, experiments  have shown that large atmospheric aerosols made from complex mixtures of soluble salts, organics and water lead to increasingly complicated cycles of structural transformations where solubility-like liquid-liquid transitions, involving phase separation of organics from inorganic salt solutions, and the deliquescence of the resulting salt solution all occur as a function of the relative humidity. These studies have focused on large particles, but it is likely surface effects will further complicate the nature of these transformations as the particles become smaller. # Conclusion Understanding nanoscale particle size effects on the dynamics, thermodynamics and physical structure of small atmospheric aerosol particles remains an important challenge. We have shown that molecular dynamics simulations of a simple binary mixture of Lennard-Jones particles are able to capture the key elements of the dynamics and thermodynamics of the condensation of a solvent vapor onto a non-volatile solute particle. In particular, we have shown that cluster-cluster coalescence plays an important role in film formation in nanoscale surfaces and that partially miscible droplets exhibit a solubility transition. We have also shown that a simple capillarity based model also captures the main features of the solubility transition, but more complex models are needed to account for surface adsorption.

{'timestamp': '2013-10-14T02:10:12', 'yymm': '1310', 'arxiv_id': '1310.3215', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3215'}

{'timestamp': '2013-10-14T02:09:56', 'yymm': '1310', 'arxiv_id': '1310.3211', 'language': 'en', 'url': 'https://arxiv.org/abs/1310.3211'}

# Introduction {#Introduction} The interactions between a pair of heavy-light hadrons can be fairly regarded as the extension of the pattern of nuclear forces. The theoretical tools that designed for the nucleon systems shall also be generalized to the heavy-light systems via including the restriction of extra symmetries, such as the heavy quark symmetry. Meanwhile, in recent years, the observations of many near-threshold exotic states provide golden platforms to test and redevelop these tools , in which the successful generalizations of the effective field theories (EFTs), e.g., the pionless and pionful EFTs, is a epitome of the intimate connection between the nuclear physics and the hadron physics . Based on the instructive works of Weinberg , in the past decades, the modern framework of nuclear forces was constructed upon the chiral effective field theory (\(\chi\)EFT) . In \(\chi\)EFT, the short-range part of the nuclear forces is parameterized as the four-fermion contact interactions through integrating out the heavy particle exchanging (e.g., the vector meson \(\rho\) and \(\omega\), etc.), while the long-and intermediate-range parts are presented by the one-pion exchange (OPE) and multi-pion exchange interactions, respectively . The latter can be derived from the chiral symmetry of QCD via a model-independent way. The study of nucleon-nucleon (\(NN\)) interactions indicates that the leading order (LO) two-pion exchange (TPE) potential is very weak and insufficient to provide the appropriate attractive force at the intermediate range, and which is in fact described by the subleading TPE potential with an insertion of the subleading pion-nucleon vertices . It was found that the large values of the low energy constants (LECs) in the subleading pion-nucleon Lagrangians leads to the attractive source. The values of these LECs can be quantitatively understood using the phenomenological resonance saturation model (RSM) . It was shown that these large value LECs in the \(\chi\)EFT without explicit \(\Delta\) resonance actually stem from the 'high' (note that \(m_\Delta-m_N\approx 2m_\pi<m_\rho\), where \(m_\rho\sim770\) MeV is usually regarded as the truly high energy scale in chiral perturbation theory) energy scale \(\Delta\) baryon as well as the pion-pion correlation (or the \(\sigma\) meson) . Epelbaum *et al*. noticed that the TPE loop diagrams calculated within the dimensional regularization accompanying with the large value LECs in subleading pion-nucleon vertices lead to unsatisfactory convergence of chiral expansion and uncertain consequences in few-nucleon systems, e.g., the unphysical deeply bound states in the low partial waves of isoscalar channel . The expediency is to use the small value LECs, but this is not compatible with the pion-nucleon scattering data . In order to cure this problem, Epelbaum *et al*. argued that one needs to suppress the high-momentum modes of the exchanged pions, since they cannot be suitably handled in an EFT who only properly works in the soft scales. They proposed to calculate the TPE loop diagrams using the cutoff regularization and combining the spectral function representation rather than the dimensional regularization . This is analogous to the means for improving the convergence of chiral expansion in the SU(3) case . Obviously, one needs to consider the possible emergence of the above mentioned problem when generalizing the \(\chi\)EFT to the heavy-light systems. The application of \(\chi\)EFT in heavy-light systems for dealing with the hadronic molecules has achieved much progress in recent years . In Ref. , Liu *et al*. first calculated the \(BB\) interactions with considering the leading TPE contributions. Along this line, Xu *et al*. studied the \(DD^\ast\) interactions and used the RSM to determine the contact LECs, in which they predicted a bound state in the isoscalar channel with \(J^P\) quantum numbers \(1^+\) . Four years latter, the LHCb Collaboration observed a state, the \({T_{cc}^+}\) in \(D^0D^0\pi^+\) invariant mass spectrum . The \({T_{cc}^+}\) is below the \(D^{\ast+}D^0\) threshold about \(300\) keV, thus it is the very good candidate of \(DD^\ast\) hadronic molecule. Similar to Ref. , Wang *et al*. studied the \(B^{(\ast)}B^{(\ast)}\) interactions and predicted the possible bound states in the isoscalar \(BB^\ast\) and \(B^\ast B^\ast\) systems with \(J^P=1^+\) . The same framework was also adopted to investigate the LHCb pentaquarks \(P_\psi^N(4312)\), \(P_\psi^N(4440)\) and \(P_\psi^N(4457)\)  (throughout this paper, we use the new naming scheme of the exotic states proposed by the LHCb ), as well as to predict the existence of molecular pentaquarks with strangeness in \(\Xi_c^{(\prime,\ast)}\bar{D}^{(\ast)}\) systems  (see also the recent experimental measurements for the \(P_{\psi s}^\Lambda\) states near the \(\Xi_c\bar{D}^\ast\)  and \(\Xi_c\bar{D}\)  thresholds), and the double-charm pentaquarks . For a review of this topic, we refer to Ref. . In Ref. , the study of \(\Sigma_c\Sigma_c\) interactions turns out that there results in bad convergence and unnaturally deep bound state in the lowest isospin channel if one calculates the leading TPE diagrams with dimensional regularization. This demands us to properly treat the TPE contributions for heavy-light systems as those in the \(NN\) case. In this work, we revisit the \(DD^\ast\) interactions within \(\chi\)EFT, and calculate the \(DD^\ast\) interactions up to the third order \[i.e. the next-to-next-to-leading order (N\(^2\)LO)\] for the first time. We construct the subleading \(\pi D^{(\ast)}\) Lagrangians and determine the corresponding LECs with the RSM. The TPE diagrams will be calculated with the cutoff regularization, but we use the fully local momentum-space regularization rather than the semi-local form as those in Ref. . The \(DD^\ast\) interactions shall strongly correlate to the \({T_{cc}^+}\) inner structures and its other properties. In contrast to the well-known \(X(3872)\), there is no coupling with the charmonia for \({T_{cc}^+}\). Thus it provides a clean environment for investigating the interactions between the charmed mesons. This is very similar to the \(NN\) interactions. The \({T_{cc}^+}\) state has been intensively studied from various aspects, such as the decay behaviors , the mass spectrum , the productions , the lineshapes , and the magnetic moment , etc.. In order to pin down the inner configuration of \({T_{cc}^+}\), a systematic study of the \(DD^\ast\) interactions is very necessary. This paper is organized as follows. The \(DD^\ast\) effective potentials within the local momentum-space regularization are shown in Sec. [2](#sec:effpot){reference-type="ref" reference="sec:effpot"}. The pole trajectory of \(DD^\ast\) bound state and related discussions are given in Sec. [3](#sec:numsanddis){reference-type="ref" reference="sec:numsanddis"}. A short summary is given in Sec. [4](#sec:sum){reference-type="ref" reference="sec:sum"}. The estimations of LECs within the RSM are listed in the Appendix [\[sec:app\]](#sec:app){reference-type="ref" reference="sec:app"}. # Effective chiral potentials up to the third order {#sec:effpot} The effective potential of \(DD^\ast\) can be extracted from their scattering amplitude. In \(\chi\)EFT, the scattering amplitude of \(DD^\ast\) is expanded in powers of the ratio \(\mathcal{Q}/\Lambda_{\rm{b}}\), where \(\mathcal{Q}\) represents the soft scale, which could be the pion mass or the external momenta of \(D^{(\ast)}\), while \(\Lambda_{\rm{b}}\) denotes the hard scale at which the \(\chi\)EFT breaks down. The relative importance of the terms in the expansion is weighed by the power \(\nu\) of \((\mathcal{Q}/\Lambda_{\rm{b}})^\nu\), this is known as the power counting scheme. According to the naive dimensional analysis , the power \(\nu\) for a system with two matter fields (charmed mesons) is measured as \[\begin{aligned} \label{eq:powercounting} \nu=2L+\sum_i V_i \Delta_i,\qquad \Delta_i=d_i+\frac{n_i}{2}-2, \end{aligned}\] with \(L\) the number of loops in a diagram, \(V_i\) the number of vertices of type-\(i\). The \(d_i\) is the number of derivatives (or the pion-mass insertions), and \(n_i\) is the number of charmed meson fields that involved in the vertex-\(i\). The \(DD^\ast\) interaction starts at \(\nu=0\) (first order, the LO), and the higher orders come as \(\nu=2\) \[second order, the next-to-leading order (NLO)\], \(\nu=3\) (third order, the N\(^2\)LO), etc.. At the given order, the number of the corresponding irreducible diagrams are limited. In Fig. [\[fig:loopdiagrams\]](#fig:loopdiagrams){reference-type="ref" reference="fig:loopdiagrams"}, we show the pertinent Feynmann diagrams for the LO, NLO and N\(^2\)LO interactions of the \(DD^\ast\) system. Then the effective potential of the \(DD^\ast\) system can be written as \[\begin{aligned} \label{eq:vctopetpe} V_{\rm eff}=V_{\rm ct}+V_{1\pi}+V_{2\pi}+\dots, \end{aligned}\] with \[\begin{aligned} \label{eq:vctvpi} V_{\rm ct}&=&V_{\rm ct}^{(0)}+V_{\rm ct}^{(2)}+\dots,\nonumber\\ V_{2\pi}&=&V_{2\pi}^{(2)}+V_{2\pi}^{(3)}+\dots, \end{aligned}\] where \(V_{\rm ct}\), \(V_{1\pi}\) and \(V_{2\pi}\) denote the contact, OPE and TPE potentials, respectively. The numbers in the parentheses of the superscripts represent the power \(\nu\) \[see Eq. [\[eq:powercounting\]](#eq:powercounting){reference-type="eqref" reference="eq:powercounting"}\]. Each piece of the right hand side of Eq. [\[eq:vctopetpe\]](#eq:vctopetpe){reference-type="eqref" reference="eq:vctopetpe"} can be further decomposed into the following form, \[\begin{aligned} V_i=\left[V_{i,c}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}W_{i,c}\right]\mathcal{O}_1+\left[V_{i,t}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}W_{i,t}\right]\mathcal{O}_2, \end{aligned}\] where \(i=\mathrm{ct},1\pi,2\pi\), and \(\bm{\tau}_{1}\cdot\bm{\tau}_{2}\) denotes the isospin-isospin interaction. The matrix element \(\langle\bm{\tau}_{1}\cdot\bm{\tau}_{2}\rangle=-3\) and \(1\) for the isoscalar and isovector channels, respectively. The operators \(\mathcal{O}_1\) and \(\mathcal{O}_2\) are given as \[\begin{aligned} \label{eq:operatoro1o2} \mathcal{O}_{1}=\bm{\varepsilon}^{\prime\dagger}\cdot\bm{\varepsilon},\qquad\qquad \mathcal{O}_{2}=(\bm{q}\cdot\bm{\varepsilon}^{\prime\dagger})(\bm{q}\cdot\bm{\varepsilon}), \end{aligned}\] where \(\bm q=\bm p^\prime-\bm p\) (\(\bm p\) and \(\bm p^\prime\) denote the initial and final state momenta in the center of mass system, respectively) is the transferred momentum between \(D\) and \(D^\ast\), \(\bm{\varepsilon}\) and \(\bm{\varepsilon}^{\prime\dagger}\) denote the polarization vectors of the initial and final \(D^\ast\) mesons, respectively. In the heavy quark limit, we will not consider the \(1/m\) (with \(m\) the mass of the charmed mesons) corrections of the charmed meson fields. Then only two pertinent operators survive in the effective potentials of \(DD^\ast\) (for the \(NN\) case, see ), i.e., the \(\mathcal{O}_1\) and \(\mathcal{O}_2\). In the following subsections, we will derive the \(V_{\rm ct}\), \(V_{1\pi}\) and \(V_{2\pi}\), respectively. ## Short-range contact interactions The contact potentials of \(DD^\ast\) system at the order \(\nu=0,2,4\) can be respectively parameterized as \[\begin{aligned} V_{\mathrm{ct}}^{(0)} &=& \left(C_{1}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}C_{2}\right)\mathcal{O}_{1},\label{eq:vcto0}\\ V_{\mathrm{ct}}^{(2)} &=& (C_{3}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}C_{4})\bm{q}^{2}\mathcal{O}_{1}+\left(C_{5}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}C_{6}\right)\mathcal{O}_{2},\label{eq:vcto2}\\ V_{\mathrm{ct}}^{(4)} &=& (C_{7}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}C_{8})\bm{q}^{4}\mathcal{O}_{1} +(C_{9}+\bm{\tau}_{1}\cdot\bm{\tau}_{2}C_{10})\bm{q}^{2}\mathcal{O}_{2},\label{eq:vcto4}\nonumber\\ \end{aligned}\] where \(C_{1,\dots,10}\) are the corresponding LECs. In the following calculations, we will take the \(V_{\mathrm{ct}}^{(4)}\) to test the convergence of the expansion in different isospin channels. In Eqs. [\[eq:vcto2\]](#eq:vcto2){reference-type="eqref" reference="eq:vcto2"} and [\[eq:vcto4\]](#eq:vcto4){reference-type="eqref" reference="eq:vcto4"}, we ignore the pion-mass dependent terms for which are of irrelevance in our studies. In calculations, the local form Gaussian regulator \(\exp{(-\bm q^2/\Lambda^2)}\) is multiplied to Eqs. [\[eq:vcto0\]](#eq:vcto0){reference-type="eqref" reference="eq:vcto0"}-[\[eq:vcto4\]](#eq:vcto4){reference-type="eqref" reference="eq:vcto4"} to ensure the convergence when they are inserted into the Lippmann-Schwinger equations (LSEs). In order to determine all the LECs in Eqs. [\[eq:vcto0\]](#eq:vcto0){reference-type="eqref" reference="eq:vcto0"}-[\[eq:vcto4\]](#eq:vcto4){reference-type="eqref" reference="eq:vcto4"}, we resort to the phenomenological RSM  (see also the applications in heavy-light systems ). Within the RSM, we consider the exchanging of the scalar, pseudoscalar, vector and axial-vector mesons \[the tensor exchanges (e.g., \(a_2,f_2\) mesons) are not considered, since their contributions start at least at the fourth order \]. The derivation details are given in appendix [\[sec:app1\]](#sec:app1){reference-type="ref" reference="sec:app1"}. Their numerical values are listed in Table [1](#tab:numsci){reference-type="ref" reference="tab:numsci"}. ## Long-range one-pion exchange interactions The \({T_{cc}^+}\) was observed in the \(D^0D^0\pi^+\) final state, and its signal is absent in \(D^+D^0\pi^+\) , which implies that the \({T_{cc}^+}\) is an isoscalar state rather than the isovector one. The flavor wave function of \(DD^\ast\) in the isoscalar and isovector channels read, respectively, \[\begin{aligned} |DD^\ast,I=0,I_{3}=0\rangle &=& \frac{1}{\sqrt{2}}\left[D^{0}D^{\ast+}-D^{\ast0}D^{+}\right],\\ |DD^\ast,I=1,I_{3}=0\rangle &=& \frac{1}{\sqrt{2}}\left[D^{0}D^{\ast+}+D^{\ast0}D^{+}\right]. \end{aligned}\] We consider the explicit chiral dynamics for the light pion and relegate the heavy \(\eta\) (\(m_\eta\simeq4m_\pi\)) contribution to the contact terms. In the following, we show the complete LO (\(\Delta_i=0\)) chiral Lagrangian of \(\varphi D^{(\ast)}\) (\(\varphi=\pi,\eta\)) coupling  for the latter convenience. \[\begin{aligned} \label{eq:mesonlagsf} \mathcal{L}_{\varphi\mathcal{H}}^{(0)}&=&i\langle \mathcal{H}v\cdot\mathcal{D}\bar{\mathcal{H}}\rangle-\frac{1}{8}\delta_b\langle\mathcal{H}\sigma^{\mu \nu}\bar{\mathcal{H}}\sigma_{\mu \nu}\rangle\nonumber\\ &&+g_\varphi\langle\mathcal{H} \slashed{u}\gamma_{5}\bar{\mathcal{H}}\rangle, \end{aligned}\] where \(v=(1,\bm 0)\) denotes the four-velocity of heavy mesons, and \(\mathcal{D}_\mu=\partial_\mu+\Gamma_\mu\), with \(\Gamma_\mu=[ \begin{eqnarray} \varphi=\left[\begin{array}{cc} \pi^{0}+\frac{1}{\sqrt{3}}\eta & \sqrt{2}\pi^{+}\\ \sqrt{2}\pi^{-} &-\pi^{0}+\frac{1}{\sqrt{3}}\eta \end{array}\right]. \end{eqnarray} The\[denotes the superfield of (\)D,D\^\() doublet in the heavy quark symmetry, which reads \begin{eqnarray} \mathcal{H}&=&\frac{1+\slashed{v}}{2}\left(P_\mu^\ast\gamma^\mu+iP\gamma_5\right),\quad\bar{\mathcal{H}}=\gamma^0\mathcal{H}^\dag\gamma^0, \end{eqnarray} with\)P=(D\^0,D\^+)\^T\(and\)P\^=(D\^,D\^+)\^T\(. With the OPE diagram in Fig.~\ref{fig:loopdiagrams} (\)\_2\() and the LO chiral Lagrangian in Eq.~\eqref{eq:mesonlagsf}, one can easily get the OPE potential, which reads \begin{eqnarray}\label{eq:opeporig} \mathcal{V}_{1\pi}(u_\pi,\bm{q})&=&\mathrm{sgn}(\bm{\tau}_{1}\cdot\bm{\tau}_{2})\frac{g_\varphi^{2}}{4f_\pi^{2}}\frac{\mathcal{O}_2}{\bm{q}^{2}-u_\pi^{2}-i\epsilon}, \end{eqnarray} where\)f\_\(MeV, and\)u\_=\((with\)m\_\(MeV the pion mass). Eq.~\eqref{eq:opeporig} contains two parts---the principle-value and the imaginary parts. Its principle-value corresponds to an oscillatory potential in the coordinate space, e.g., see Eq.~\eqref{eq:fourtrans}, while the imaginary part comes from the three-body (\)DD\() cut, it will contribute a finite width to the bound state of\)DD\^\(. We then separate the operator\)\_2\(into the `spin-spin' part and the tensor part via the equation \begin{eqnarray} \mathcal{O}_2=(\bm{q}\cdot\bm{\varepsilon}^{\prime\dagger})(\bm{q}\cdot\bm{\varepsilon})=\frac{1}{3}(\bm{\varepsilon}^{\prime\dagger}\cdot\bm{\varepsilon})\bm{q^{2}}+\bm{q}^{2}\mathcal{S}_{12}, \end{eqnarray} where\)\_12=(\^)()-\^\(, with\)=q/\|q\|\(. Then the principle-value part of Eq.~\eqref{eq:opeporig} can be transformed into \begin{eqnarray} \mathcal{V}_{1\pi}(u_\pi,\bm{q})&=&\mathrm{sgn}(\bm{\tau}_{1}\cdot\bm{\tau}_{2})\frac{g_\varphi^{2}}{4f_\pi^{2}}\bigg(\frac{1}{3}\mathcal{O}_{1}+\frac{1}{3}\frac{u_\pi^{2}}{\bm{q}^{2}-u_\pi^{2}}\mathcal{O}_{1}\nonumber\\ &&+\frac{\bm{q}^{2}}{\bm{q}^{2}-u_\pi^{2}}\mathcal{S}_{12}\bigg), \end{eqnarray} in which the first term corresponds to a\]-function in the coordinate space after the Fourier transform. It is an artefact arising from the idealized point-like\)D\^()\(coupling. In reality, the OPE dominates at the long-distance region, i.e.,\)r  m\_\^-1\(~. Therefore, it is better to subtract the unphysical\[-function part from the OPE potential. In Ref.~, Reinert \textit{et al}.~introduced a subtraction scheme for the\)NN\(interaction with the following form \begin{eqnarray} V_{1\pi,\Lambda}(u_\pi,\bm{q})&=&\mathrm{sgn}(\bm{\tau}_{1}\cdot\bm{\tau}_{2})\frac{g_\varphi^{2}}{4f_\pi^{2}}\Bigg\{\frac{u_\pi^{2}}{3}\frac{1}{\bm{q}^{2}-u_\pi^{2}}\mathcal{O}_{1}\nonumber\\ &&+\left[\frac{1}{3}+\mathcal{C}(u_\pi,\Lambda)\right]\mathcal{O}_{1}+ \frac{\bm{q}^{2}}{\bm{q}^{2}-u_\pi^{2}} \mathcal{S}_{12}\Bigg\}\nonumber\\ &&\times\exp\Big(-\frac{\bm{q}^{2}-u_\pi^{2}}{\Lambda^{2}}\Big), \end{eqnarray} where an\)u\_\(-dependent term in the Gaussian regulator is introduced to ensure the strength of OPE potential remains unchanged at the pion pole~. The subtraction term\)(u\_,)\(is determined by the requirement that the OPE potential vanishes at the origin, i.e., when\)r\(. With the following relations of Fourier transform, \begin{eqnarray}\label{eq:fourtrans} &&\int\frac{d^{3}q}{(2\pi)^{3}}e^{i\bm{q}\cdot\bm{r}}\frac{1}{\bm{q}^{2}-u_\pi^{2}}=\frac{1}{4\pi r}\cos (u_\pi r),\\ &&\mathcal{U}_{\Lambda}(u_\pi,r)=\int\frac{d^{3}q}{(2\pi)^{3}}e^{i\bm{q}\cdot\bm{r}}\frac{1}{\bm{q}^{2}-u_\pi^{2}}\exp\Big(-\frac{\bm{q}^{2}-u_\pi^{2}}{\Lambda^{2}}\Big)\nonumber\\ &&~~~~~=\frac{\cos(u_\pi r)}{8\pi r}\left[\mathrm{erfc}\Big(\frac{u_\pi}{\Lambda}-\frac{\Lambda r}{2}\Big)-\mathrm{erfc}\Big(\frac{u_\pi}{\Lambda}+\frac{\Lambda r}{2}\Big)\right],\nonumber\\ \\ &&\int\frac{d^{3}q}{(2\pi)^{3}}e^{i\bm{q}\cdot\bm{r}}\exp\Big(-\frac{\bm{q}^{2}-u_\pi^{2}}{\Lambda^{2}}\Big)\nonumber\\ &&\qquad\qquad=\Big(\frac{\Lambda^{2}}{4\pi}\Big)^{3/2}\exp\Big(\frac{u_\pi^{2}}{\Lambda^{2}}-\frac{\Lambda^{2}r^{2}}{4}\Big), \end{eqnarray} one easily obtains \begin{eqnarray}\label{eq:opesub} &&V_{1\pi,\Lambda}(u_\pi,r) = \int\frac{d^{3}q}{(2\pi)^{3}}e^{i\bm{q}\cdot\bm{r}}V_{1\pi,\Lambda}(u_\pi,\bm{q})\nonumber\\ &&\quad= -\mathrm{sgn}(\bm{\tau}_{1}\cdot\bm{\tau}_{2})\frac{g_\varphi^{2}}{4f_\pi^{2}}\Bigg\{ \mathcal{S}_{12}r\frac{\partial}{\partial r}\left(\frac{1}{r}\frac{\partial}{\partial r}\right)\mathcal{U}_{\Lambda}(u_\pi,r)\nonumber\\ &&\quad\quad-\mathcal{O}_{1}\bigg[\frac{u_\pi^{2}}{3}\mathcal{U}_{\Lambda}(u_\pi,r)+\Big(\frac{1}{3}+\mathcal{C}(u_\pi,\Lambda)\Big)\Big(\frac{\Lambda^{2}}{4\pi}\Big)^{3/2}\nonumber\\ &&\quad\quad\times\exp\Big(\frac{u_\pi^{2}}{\Lambda^{2}}-\frac{\Lambda^{2}r^{2}}{4}\Big)\bigg]\Bigg\}. \end{eqnarray} With the constraints\)V\_1,(u\_,r)=0\(, we get \begin{eqnarray} \mathcal{C}(u_\pi,\Lambda)=-\Big(\frac{1}{3}+\frac{2u_\pi^{2}}{3\Lambda^{2}}e^{-\frac{2u_\pi^{2}}{\Lambda^{2}}}\Big). \end{eqnarray} Note that, in Eq.~\eqref{eq:fourtrans} the\)(x)\(represents the complementary error function, i.e., \begin{eqnarray} \mathrm{erfc}(x)=\frac{2}{\sqrt{\pi}}\int_{x}^{\infty}dt e^{-t^2}. \end{eqnarray} In Fig.~\ref{fig:ope}, we show the behaviors of the central part [\)\_1\(related term in Eq.~\eqref{eq:opesub}] and tensor part [\)\_12\(related term in Eq.~\eqref{eq:opesub}] of the\]-function subtracted OPE potential for the\)I=0\(case (the behaviors of the\)I=1\(case are similar). One can see that both the central and tensor potentials vanish for\)r\(, and the strength of the central potential is much weaker than that of the tensor potential. Therefore, though the central potential is attractive, it is too weak to form bound state. However, if one does not subtract the\[-function, then there would result in very attractive central potential once using a large cutoff when making the Fourier transform. This may also lead to the bound state, but it is unreasonable. One also sees that the central potential can extend to large distances since the effective mass\)u\_\(in the pion propagator is much smaller than the\)m\_\(, this is a very typical feature of the\)DD\^\(system. \subsection{Intermediate-range two-pion exchange interactions} We first show the LO (\]) TPE contributions, which come from the diagrams in Figs.~\ref{fig:loopdiagrams} (\)\_2\()-(\)\_11\(). They can be obtained using the Lagrangian~\eqref{eq:mesonlagsf} and calculating the loop integrals. We adopt the spectral function representation for the TPE interactions. The long-range part of the TPE interactions is determined by the non-analytic terms in momentum-space. They have the following forms within the dimensional regularization, \begin{widetext} \begin{eqnarray} \mathcal{V}_{2\pi,c}^{(2)}&=&\frac{g_\varphi^{4}}{512\pi\delta_{b}f_{\pi}^{4}}\left[6\bm{q}^2\mathrm{sgn}(4m_{\pi}^{2}-4\delta_{b}^{2}+\bm{q}^2)A^{\prime}(q)-3(8m_{\pi}^{4}+10m_{\pi}^{2}\bm{q}^2+3\bm{q}^4)A(q)-\frac{8\delta_{b}}{\pi}(5\delta_{b}^{2}+3\bm{q}^2\mathrm{sgn})L(q)\right],\label{eq:v2pic}\\ \mathcal{W}_{2\pi,c}^{(2)}&=&\frac{g_{\varphi}^{4}}{256\pi\delta_{b}f_{\pi}^{4}}\bigg\{\frac{1}{g_{\varphi}^{2}}(4m_{\pi}^{2}-4\delta_{b}^{2}+3\bm{q}^2)\left[2\delta_{b}^{2}+g_{\varphi}^{2}(-2\delta_{b}^{2}+2m_{\pi}^{2}+\bm{q}^2)\right]A^{\prime}(q)-2\mathrm{sgn}(4m_{\pi}^{2}+\bm{q}^2)\bm{q}^2A(q)\nonumber\\ &&+\frac{2\delta_{b}}{3\pi g_{\varphi}^{4}}\left[16\delta_{b}^{2}g_{\varphi}^{2}(5g_{\varphi}^{2}-3)+4(-5g_{\varphi}^{4}+4g_{\varphi}^{2}+1)m_{\pi}^{2}+(-23g_{\varphi}^{4}+10g_{\varphi}^{2}+1)\bm{q}^2\right]L(q)\bigg\}, \\ \mathcal{V}_{2\pi,t}^{(2)}&=&\frac{g_\varphi^{4}}{512\pi\delta_{b}f_{\pi}^{4}}\left[\frac{3}{\bm{q}^2}(-8m_{\pi}^{4}-2m_{\pi}^{2}\bm{q}^2+\bm{q}^4)A(q)-6\mathrm{sgn}(-4\delta_{b}^{2}+4m_{\pi}^{2}+\bm{q}^2)A^{\prime}(q)+ \frac{24\delta_{b}}{\pi}\mathrm{sgn}L(q)\right],\\ \mathcal{W}_{2\pi,t}^{(2)}&=&\frac{g_{\varphi}^{4}}{256\pi\delta_{b}f_{\pi}^{4}}\left\{ 2\mathrm{sgn}(4m_{\pi}^{2}+\bm{q}^2)A(q)-\frac{1}{g_{\varphi}^{2}\bm{q}^2}(4\delta_{b}^{2}-4m_{\pi}^{2}+\bm{q}^2)\left[2\delta_{b}^{2}+g_{\varphi}^{2}(-2\delta_{b}^{2}+2m_{\pi}^{2}+\bm{q}^2)\right]A^{\prime}(q)\right\},\label{eq:w2pit} \end{eqnarray} \end{widetext} where the three non-analytic functions\)A(q)\(,\)A\^(q)\(and\)L(q)\(respectively read \begin{eqnarray} A(q)&=&\frac{1}{2q}\arctan\frac{q}{2m_\pi},\\ A^{\prime}(q)&=&\frac{1}{2q}\arctan\frac{q}{2m^{\prime}},\\ L(q)&=&\frac{\varpi}{q}\ln\frac{q+\varpi}{2m_\pi}, \end{eqnarray} with\)q=\|q\|\(,\)m\^=\[m\_\^2-\_b\^2-i\]\^1/2\(, and\)=\[\^2+4m\_\^2-i\]\^1/2\(. The terms containing the non-analytic functions\)(q,n,i)=\_-1/2\^1/2(y\^n\^i)dy\((with\)=q\(,\)a=\(, and\)\^=\[\^2+4(m\_\^2-\_b\^2)-i\]\^1/2\() and their derivatives with respect to\)\_b\(are ignored for simplicity since we noticed that their contributions are much smaller than those in Eqs.~\eqref{eq:v2pic}-\eqref{eq:w2pit}. In order to obtain the subleading (\[) TPE potential (see the diagrams in the third column of Fig.~\ref{fig:loopdiagrams}), one needs an insertion of the subleading (\)\_i=1\()\)D\^()\(Lagrangians. The Lagrangians read~ \begin{eqnarray}\label{eq:nlopid} \mathcal{L}_{\varphi\mathcal{H}}^{(1)}&=&\tilde{c}_{1}\langle\mathcal{H}\bar{\mathcal{H}}\rangle\mathrm{Tr}(\chi_{+})+\tilde{c}_{2}\langle\mathcal{H}v\cdot uv\cdot u\bar{\mathcal{H}}\rangle+\tilde{c}_{3}\langle\mathcal{H}u\cdot u\bar{\mathcal{H}}\rangle\nonumber\\ &&+i\tilde{c}_{4}\langle\mathcal{H}[u_{\mu},u_{\nu}]\sigma^{\mu\nu}\bar{\mathcal{H}}\rangle+\tilde{c}_{5}\langle\mathcal{H}\hat{\chi}_{+}\bar{\mathcal{H}}\rangle, \end{eqnarray} where\)\_+= In literature, only the LECs in partial terms in Eq. [\[eq:nlopid\]](#eq:nlopid){reference-type="eqref" reference="eq:nlopid"} were determined for certain problems (see Ref. ). Here, we again use the RSM to estimate the \(\tilde{c}_i\). One can consult appendix [\[sec:app2\]](#sec:app2){reference-type="ref" reference="sec:app2"} for details. The numerical values of the LECs \(\tilde{c}_i~(i=1,\dots,5)\) in Eq. [\[eq:nlopid\]](#eq:nlopid){reference-type="eqref" reference="eq:nlopid"} are summarized in Table [2](#tab:numscii){reference-type="ref" reference="tab:numscii"}. From Table [2](#tab:numscii){reference-type="ref" reference="tab:numscii"} one can see that the couplings of the subleading \(\pi D^{(\ast)}\) vertices are of natural size and are much smaller than those of the \(\pi N\) system . In contrast to the \(NN\) system, this makes the main contribution for the binding forces of \(DD^\ast\) come from the short-range contact interactions. The non-analytic terms of the subleading TPE potentials read # Numerical results and discussions {#sec:numsanddis} With the effective potentials, we solve the LSEs to analyze the pole distributions in the physical Riemann sheet. The LSE in the partial wave basis \(|\ell sj\rangle\) reads \]\begin{aligned} T_{\ell^\prime\ell sj}(p^\prime,p)&=&V_{\ell^\prime\ell sj}(p^\prime,p) +\sum_{\ell^{\prime\prime}}\int\frac{k^2dk}{(2\pi)^3}V_{\ell^\prime\ell^{\prime\prime} sj}(p^\prime,k)\nonumber\\ &&\times\frac{2\mu_m}{p^2-k^2+i\epsilon}T_{\ell^{\prime\prime}\ell sj}(k,p), \end{aligned}\[ where \(\mu_m\) denotes the reduced mass of \(DD^\ast\) system, and \(V_{\ell^\prime\ell sj}(p^\prime,p)=\langle \ell^\prime sj|V(\bm q)|\ell sj\rangle\) can be easily obtained with the approach in Ref. . The S-D wave coupling is considered in our calculations. Thus, the \(V_{\ell^\prime\ell sj}(p^\prime,p)\) is given with the \(2\times2\) matrix in the coupled-channel \(|\ell sj\rangle\) basis. We use the isospin average mass for \(DD^\ast\) system in our calculations. The threshold of \(DD^\ast\) and the experimentally measured mass of \({T_{cc}^+}\)  are given as \]\begin{aligned} m_{\mathrm{th}}&=&3875.8~\mathrm{MeV},\nonumber\\ m_{\mathrm{exp}}&=&(m_{D^{\ast+}}+m_{D^0})-0.36=3874.7~\mathrm{MeV}. \end{aligned}\(\) We first study the pole trajectory of \(DD^\ast\) scattering T-matrix in the isoscalar channel with the contact interactions being kept up to NLO \[including the Eqs. [\[eq:vcto0\]](#eq:vcto0){reference-type="eqref" reference="eq:vcto0"}-[\[eq:vcto2\]](#eq:vcto2){reference-type="eqref" reference="eq:vcto2"} in effective potentials\], and the cutoff is in the range \(520-600\) MeV. We notice that the binding solution begins to appear when \(\Lambda\approx600\) MeV in this case. The pole trajectory is very similar to the case in which the contact interaction is kept up to N\(^2\)LO \[including the Eqs. [\[eq:vcto0\]](#eq:vcto0){reference-type="eqref" reference="eq:vcto0"}-[\[eq:vcto4\]](#eq:vcto4){reference-type="eqref" reference="eq:vcto4"} in effective potentials, and the result in this case is shown in Fig. [\[fig:pole\]](#fig:pole){reference-type="ref" reference="fig:pole"}\]. In this case, the pole appears when \(\Lambda\approx520\) MeV, and the pole mass approaches to the experimental value when \(\Lambda\simeq560\) MeV. From Fig. [\[fig:pole\]](#fig:pole){reference-type="eqref" reference="fig:pole"}, one also sees that the binding becomes deeper with the increasing of cutoff, while the half width is almost insensitive to the cutoff. The width of \({T_{cc}^+}\) is around \(80\) keV in our calculations, which is about two times larger than the measured value \(47.8\pm1.9\) keV with an unitarized Breit-Wigner profile in Ref. . We also investigated the situation in isovector channel, but we did not find binding solutions in this channel. This is consistent with the experimental facts---there are no structures in the \(D^+D^0\pi^+\) invariant mass spectrum. # Summary {#sec:sum} We revisit the \(DD^\ast\) interactions within the \(\chi\)EFT up to the third order. The pion-exchanged interactions are carefully treated with the local momentum-space regularization, in which their short-range components are subtracted via demanding the pion-exchanged contributions vanish at the origin in the coordinate space. This is consistent with the new developments of nuclear forces in Ref. . The contact interactions and the subleading \(\pi D^{(\ast)}\) vertices are ascribed to the heavier meson exchanging, and consequently the LECs are estimated with the resonance saturation model. We notice that the subleading \(\pi D^{(\ast)}\) couplings are much smaller that those in the \(\pi N\) system, which makes the binding force of \(DD^\ast\) mainly come from the short-range part. This is very different with that of the \(NN\) system. We investigate the pole trajectory of \(DD^\ast\) scattering T-matrix in two cases---keeping the contact interactions up to NLO and N\(^2\)LO, respectively. The pole trajectory are very similar in these two cases, which implies the low-energy expansion works well in the isoscalar channel. The binding solution only exists in the isoscalar channel, and this is consistent with the experimental data. Our calculation favors the molecular explanation of \({T_{cc}^+}\).

{'timestamp': '2022-12-19T02:11:09', 'yymm': '2212', 'arxiv_id': '2212.08447', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08447'}